diff --git a/.gitignore b/.gitignore index 02f2a8e..c1d362c 100644 --- a/.gitignore +++ b/.gitignore @@ -8,4 +8,15 @@ build/ *.h5 *.tar.gz *.weights -.idea/ \ No newline at end of file +.idea/ +*.hdf5 +*.pk +*.table +cmake-build-release/ +demo/COCO_val2017 +demo/BDD100K_val +/.vs +cmake-build-minsizerel/* +scripts/COCO_val2017/* +scripts/COCO_val2017.zip +scripts/all_labels.txt \ No newline at end of file diff --git a/CMakeLists.txt b/CMakeLists.txt index 7af66c0..56d4fff 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -1,8 +1,18 @@ cmake_minimum_required(VERSION 3.5) - +set(PROJ_NAME tkDNN) project (tkDNN) set(CMAKE_MODULE_PATH ${CMAKE_MODULE_PATH} ${CMAKE_CURRENT_SOURCE_DIR}/cmake) -set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -std=c++11 -fPIC") +if(UNIX) + +#### +set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -std=c++11 -fPIC -Wno-deprecated-declarations -Wno-unused-variable ") +endif() +if(WIN32) +set(CMAKE_CXX_STANDARD 11) +set(CMAKE_CXX_FLAGS "/O2 /FS /EHsc") +set(CMAKE_WINDOWS_EXPORT_ALL_SYMBOLS ON) +#add extras for baggage +endif(WIN32) include_directories(${CMAKE_CURRENT_SOURCE_DIR}/include/tkDNN) # project specific flags @@ -10,6 +20,7 @@ if(DEBUG) add_definitions(-DDEBUG) endif() +add_definitions(-DTKDNN_PATH="${CMAKE_CURRENT_SOURCE_DIR}") #------------------------------------------------------------------------------- # CUDA @@ -17,97 +28,82 @@ endif() find_package(CUDA 9.0 REQUIRED) SET(CUDA_SEPARABLE_COMPILATION ON) #set(CUDA_NVCC_FLAGS "${CUDA_NVCC_FLAGS} -arch=sm_30 --compiler-options '-fPIC'") +set(CUDA_NVCC_FLAGS ${CUDA_NVCC_FLAGS} --maxrregcount=32) find_package(CUDNN REQUIRED) +include_directories(${CUDNN_INCLUDE_DIR}) + # compile -file(GLOB tkdnn_CUSRC "src/kernels/*.cu" "src/*.cu") +file(GLOB tkdnn_CUSRC "src/kernels/*.cu" "src/sorting.cu") cuda_include_directories(${CMAKE_CURRENT_SOURCE_DIR}/include ${CUDA_INCLUDE_DIRS} ${CUDNN_INCLUDE_DIRS}) cuda_add_library(kernels SHARED ${tkdnn_CUSRC}) +target_link_libraries(kernels ${CUDA_CUBLAS_LIBRARIES}) #------------------------------------------------------------------------------- # External Libraries #------------------------------------------------------------------------------- +find_package(Eigen3 REQUIRED) +include_directories(${EIGEN3_INCLUDE_DIR}) + find_package(OpenCV REQUIRED) set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -DOPENCV") +# gives problems in cross-compiling, probably malformed cmake config +find_package(yaml-cpp REQUIRED) #------------------------------------------------------------------------------- # Build Libraries #------------------------------------------------------------------------------- -file(GLOB tkdnn_SRC "src/*.cpp") -set(tkdnn_LIBS kernels ${CUDA_LIBRARIES} ${CUDA_CUBLAS_LIBRARIES} ${CUDNN_LIBRARIES} ${OpenCV_LIBS}) +file(GLOB tkdnn_SRC "src/*.cpp",src/*.c) +set(tkdnn_LIBS kernels ${CUDA_LIBRARIES} ${CUDA_CUBLAS_LIBRARIES} ${CUDNN_LIBRARIES} ${OpenCV_LIBS} yaml-cpp) -set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wall -std=c++11") +set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS}") include_directories(${CMAKE_CURRENT_SOURCE_DIR}/include ${CUDA_INCLUDE_DIRS} ${OPENCV_INCLUDE_DIRS} ${NVINFER_INCLUDES}) add_library(tkDNN SHARED ${tkdnn_SRC}) target_link_libraries(tkDNN ${tkdnn_LIBS}) +####compile +#set(PROJ_NAME BaggageAIApi) +# Path to BaggageAI project folder. +set(BAGGAGEAI_PATH /home/baggageai/files) +# Give a custom name to shared library which is provided by DIMENSIONLESS. +#set(BAGGAGEAI_LIB_NAME libBaggageAI) +# Define C++ level, could be 11 or 17 as well. +set(CMAKE_CXX_STANDARD 11) +set(CMAKE_CXX_STANDARD_REQUIRED TRUE) +# Define compiler optimization level. +set(CMAKE_CXX_FLAGS "-O3") +# Do print warnings uppon compilation, let's keep our code as clean as possible. +set(CMAKE_CXX_FLAGS "-Wall -Wextra") +# Apply flags. +set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -std=c++11 -DBOOST_LOG_DYN_LINK") +set(Casablanca_LIBRARIES "-lboost_log -lboost_log_setup -lboost_thread -lboost_system -lcrypto -lssl -lcpprest -lpthread") + + +# Note: We do not recommend using GLOB or GLOB_RECURSE to collect a list of source files from your source tree. +# If no CMakeLists.txt file changes when a source is added or removed then the generated build system cannot know +# when to ask CMake to regenerate. + +file(GLOB_RECURSE SOURCE_FILES "main.cpp" "handler.cpp" "src/*.cpp","src/*.c") + +add_executable(baggageAPI ${SOURCE_FILES}) +set(Casablanca_LIBRARIES "-lboost_log -lboost_log_setup -lboost_thread -lboost_system -lcrypto -lssl -lcpprest -lpthread" ) +set(tkdnn_LIBS kernels ${Casablanca_LIBRARIES} ${CUDA_LIBRARIES} ${CUDA_CUBLAS_LIBRARIES} ${CUDNN_LIBRARIES} ${OpenCV_LIBS} yaml-cpp) +set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS}") +# Link BaggageAI library' include folder. +include_directories(${CMAKE_CURRENT_SOURCE_DIR}/include ${CUDA_INCLUDE_DIRS} ${OPENCV_INCLUDE_DIRS} ${NVINFER_INCLUDES} ${Casablanca_LIBRARIES} ${CMAKE_CXX_FLAGS}) +# Define BaggageAI library' shared library. +#add_library(${BAGGAGEAI_LIB_NAME} SHARED IMPORTED) +# Set a path to BaggageAI library' shared library +#set_property(TARGET ${BAGGAGEAI_LIB_NAME} PROPERTY IMPORTED_LOCATION "${BAGGAGEAI_PATH}/libBaggageAI.so") + +# Link all libraries together. +target_link_libraries(baggageAPI ${tkdnn_LIBS}) #static -#add_library(tkDNN_static STATIC ${tkdnn_SRC}) -#target_link_libraries(tkDNN_static ${tkdnn_LIBS}) - -add_executable(test_simple tests/simple/test_simple.cpp) -target_link_libraries(test_simple tkDNN) - -add_executable(test_mnist tests/mnist/test_mnist.cpp) -target_link_libraries(test_mnist tkDNN) - -add_executable(test_mnistRT tests/mnist/test_mnistRT.cpp) -target_link_libraries(test_mnistRT tkDNN) - -## YOLO NETS -add_executable(test_yolo tests/yolo/yolo.cpp) -target_link_libraries(test_yolo tkDNN) - -add_executable(test_yolo_voc tests/yolo_voc/yolo_voc.cpp) -target_link_libraries(test_yolo_voc tkDNN) - -add_executable(test_yolo_tiny tests/yolo_tiny/yolo_tiny.cpp) -target_link_libraries(test_yolo_tiny tkDNN) - -add_executable(test_yolo_relu tests/yolo_relu/yolo_relu.cpp) -target_link_libraries(test_yolo_relu tkDNN) - - -add_executable(test_yolo_224 tests/yolo_224/yolo_224.cpp) -target_link_libraries(test_yolo_224 tkDNN) - -add_executable(test_yolo_berkeley tests/yolo_berkeley/yolo_berkeley.cpp) -target_link_libraries(test_yolo_berkeley tkDNN) - -add_executable(test_yolo3_coco4 tests/yolo3_coco4/yolo3_coco4.cpp) -target_link_libraries(test_yolo3_coco4 tkDNN) - -add_executable(test_yolo3 tests/yolo3/yolo3.cpp) -target_link_libraries(test_yolo3 tkDNN) - -add_executable(test_yolo3_tiny tests/yolo3_tiny/yolo3_tiny.cpp) -target_link_libraries(test_yolo3_tiny tkDNN) - -add_executable(test_yolo3_berkeley tests/yolo3_berkeley/yolo3_berkeley.cpp) -target_link_libraries(test_yolo3_berkeley tkDNN) - -add_executable(test_yolo3_flir tests/yolo3_flir/yolo3_flir.cpp) -target_link_libraries(test_yolo3_flir tkDNN) - - -add_executable(test_resnet101 tests/resnet101/resnet101.cpp) -target_link_libraries(test_resnet101 tkDNN) - -add_executable(test_resnet101_cnet tests/resnet101_cnet/resnet101_cnet.cpp) -target_link_libraries(test_resnet101_cnet tkDNN) - -################################################################################ - - -add_executable(test_rtinference tests/test_rtinference/rtinference.cpp) -target_link_libraries(test_rtinference tkDNN) - -add_executable(yolo3_demo demo/demo/demo.cpp) -target_link_libraries(yolo3_demo tkDNN) - +#add_executable(demo demo/inf.cpp) +#target_link_libraries(demo tkDNN) #------------------------------------------------------------------------------- # Install @@ -123,18 +119,3 @@ install(DIRECTORY "${CMAKE_CURRENT_SOURCE_DIR}/cmake/" # source directory DESTINATION "share/tkDNN/cmake/" # target directory ) - -#------------------------------------------------------------------------------- -# Prepare for test -#------------------------------------------------------------------------------- -set(TEST_DATA true CACHE BOOL "If true download deps") -if( ${TEST_DATA} ) - message("Launching pre-build dependency installer script...") - - execute_process (COMMAND bash -c "bash build_models.sh download" - WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}/tests) - - set(TEST_DATA false CACHE BOOL "If true download deps" FORCE) - message("Finished dowloading test weights") -endif() - diff --git a/README.md b/README.md index 3b16ec3..f64709d 100644 --- a/README.md +++ b/README.md @@ -1,51 +1,148 @@ -# tkDNN -tkDNN is a Deep Neural Network library built with cuDNN primitives specifically thought to work on NVIDIA TK1(and all successive) board.
-The main scope is to do high performance inference on already trained models. -this branch actually work on every NVIDIA GPU that support the dependencies: -* CUDA 10.0 -* CUDNN 7.603 -* TENSORRT 6.01 -* OPENCV 4.1 +# Steps to build docker image +## Docker +Docker version 19.03 will be required. -## Workflow -The recommended workflow follow these step: -* Build and train a model in Keras (on any PC) -* Export weights and bias -* Define the model on tkDNN -* Do inference (on TK1) +## For running with GPU +### NVIDIA Drivers +This drivers should be installed in the host system. +### For Ubuntu: +``` +$ sudo apt-get install linux-headers-$(uname -r) gcc g++ make +$ wget http://in.download.nvidia.com/tesla/418.67/NVIDIA-Linux-x86_64-418.67.run +$ chmod 777 NVIDIA-Linux-x86_64-418.67.run +$ bash NVIDIA-Linux-x86_64-418.67.run +``` -## Compile the library -Build with cmake +### For CentOS ``` -mkdir build -cd build -cmake .. -# use -DTEST_DATA=False to skip dataset download -make +$ sudo yum -y install kernel-devel-$(uname -r) kernel-header-$(uname -r) gcc make +$ wget http://in.download.nvidia.com/tesla/418.67/NVIDIA-Linux-x86_64-418.67.run +$ chmod 777 NVIDIA-Linux-x86_64-418.67.run +$ bash NVIDIA-Linux-x86_64-418.67.run ``` -during the cmake configuration it will be dowloaded the weights needed for running -the tests -## Test -Assumiung you have correctly builded the library these are the test ready to exec: -* test_simple: a simple convolutional and dense network (CUDNN only) -* test_mnist: the famous mnist netwok (CUDNN and TENSORRT) -* test_mnistRT: the mnist network hardcoded in using tensorRT apis (TENSORRT only) -* test_yolo: YOLO detection network (CUDNN and TENSORRT) -* test_yolo_tiny: smaller version of YOLO (CUDNN and TENSRRT) -* test_yolo3_berkeley: our yolo3 version trained with BDD100K dateset +### NVIDIA Container Toolkit +This toolkit should be installed in the host system. +### For Ubuntu +``` +$ distribution=$(. /etc/os-release;echo $ID$VERSION_ID) +$ curl -s -L https://nvidia.github.io/nvidia-docker/gpgkey | sudo apt-key add - +$ curl -s -L https://nvidia.github.io/nvidia-docker/$distribution/nvidia-docker.list | sudo tee /etc/apt/sources.list.d/nvidia-docker.list -## yolo3 berkeley demo detection -For the live detection you need to precompile the tensorRT file by luncing the desidered network test, this is the recommended process: +$ sudo apt-get update && sudo apt-get install -y nvidia-container-toolkit +$ sudo systemctl restart docker ``` -export TKDNN_MODE=FP16 # set the half floating point optimization -rm yolo3_berkeley.rt # be sure to delete(or move) old tensorRT files -./test_yolo3_berkeley # run the yolo test (is slow) -# with f16 inference the result will be a bit incorrect + +### For CentOS ``` -this will genereate a yolo3_berkeley.rt file that can be used for live detection: +$ distribution=$(. /etc/os-release;echo $ID$VERSION_ID) +$ curl -s -L https://nvidia.github.io/nvidia-docker/$distribution/nvidia-docker.repo | sudo tee /etc/yum.repos.d/nvidia-docker.repo + +$ sudo yum install -y nvidia-container-toolkit +$ sudo systemctl restart docker ``` -./yolo3_demo # launch detection on a demo video -./yolo3_demo yolo3_berkeley.rt /dev/video0 # launch detection on device 0 + +### Building docker image + +* Go to ``TKDNN/`` then run following command. + ```$ docker build -t baggageai:server -f docker/Dockerfile .``` + #### Note: +1. have to copy weights into a TKDNN/config/ (tkdnn converted weights) current support api (fp32X4 and fp16X1) +2. setup number of classes accrding to weights in TKDNN/config/config.yml +3. give path of this weights into handler.cpp line number 117-121. + +### Run docker image ``` +$ docker run --gpus all -p 8080:8080 -d +``` +Now server will be started in the container. You can check server is running or not using ``docker ps`` + + + + +### Run docker image +``` +$ docker run -p 8080:8080 -d +``` + + +# Using docker-compose file +## Docker Compose +Install docker-compose version 1.24.1 +[https://docs.docker.com/compose/install/](https://docs.docker.com/compose/install/) + +## Create Volume +Create a volume named ``BAI_logs`` using following command: +``` +$ docker volume create BAI_logs +``` +Change permission of the directory of docker volume, so that logs can be written to that directory. +``` +$ cd /var/lib/docker/volumes/BAI_logs +$ chmod 757 _data/ +``` +``` +## For running with GPU +Install nvidia-container-runtime: +``` +$ curl -s -L https://nvidia.github.io/nvidia-container-runtime/gpgkey | \ + sudo apt-key add - +$ distribution=$(. /etc/os-release;echo $ID$VERSION_ID) +$ curl -s -L https://nvidia.github.io/nvidia-container-runtime/$distribution/nvidia-container-runtime.list | \ + sudo tee /etc/apt/sources.list.d/nvidia-container-runtime.list +$ sudo apt-get update +$ sudo apt-get install nvidia-container-runtime +``` + +Add nvidia runtime in ``/etc/docker/daemon.json `` +``` +{ + "runtimes": { + "nvidia": { + "path": "/usr/bin/nvidia-container-runtime", + "runtimeArgs": [] + } + }, + "default-runtime": "nvidia" + +} +``` +After editing changes restart the docker. +``systemctl restart docker `` + +Go to ``BaggageAI-Darknet-API/baggageai/dist/server/with-gpu`` and run: +``` +$ docker-compose up -d +``` + +## Running containers in stack +First of all initialize a swarm. +``` +$ docker swarm init +``` +You can add a worker node using ``docker swarm join`` command displayed on terminal. + + +Check the statsus using +``` +$ docker service ls +$ docker stack ls +``` + +### With GPU +Go to ``TKDNN/docker/`` and run: +``` +$ docker stack deploy -c docker-compose.yml +``` + +# Calling the API +``` +curl -X POST http://localhost:8080?name= --data-binary "@" +``` +Example, +``` +curl -X POST http://localhost:8080?name=S0240628297_20180812164749_L-4_3.jpg \ +--data-binary "@/home/ubuntu/BaggageAI/S0240628297_20180812164749_L-4_3.jpg" +``` + diff --git a/bag_server.cpp b/bag_server.cpp new file mode 100644 index 0000000..b44d1f6 --- /dev/null +++ b/bag_server.cpp @@ -0,0 +1,248 @@ +#define STB_IMAGE_IMPLEMENTATION +#include +#include +#include /* srand, rand */ +#ifdef __linux__ +#include +#endif +#define STB_IMAGE_WRITE_IMPLEMENTATION +#include "stb_image_write.h" +#include "stb_image.h" +#include +#include "utils.h" +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include "Yolo3Detection.h" +//#include "CenternetDetection.h" +//#include "MobilenetDetection.h" +#include "evaluation.h" +#include +#include +#include +#include +#include +#include +#include +#include +#include "opencv2/core/core.hpp" +#include + +#include //socket +#include +#include + +using namespace std; +using namespace cv; + +#define PORT 8080 + +#define FRAME_WIDTH 640 +#define FRAME_HEIGHT 480 + +void error(const char *msg) +{ + perror(msg); + exit(1); +} int sockfd, newsockfd, portno, n, imgSize, bytes=0, IM_HEIGHT, IM_WIDTH;; + socklen_t clilen; + char buffer[256]; + // struct sockaddr_in serv_addr, cli_addr; +// sockfd=socket(AF_INET, SOCK_STREAM, 0); + cv::Mat img; + char ntype = 'y'; + const char *config_filename = "../demo/config.yaml"; + const char * net = "../demo/yolo4_fp32.rt"; +// const char * img_path = "../demo/demo.jpg"; + char * img_data; + bool show = false; + bool verbose; + int classes, map_points, map_levels; + float map_step, IoU_thresh, conf_thresh; + tk::dnn::Yolo3Detection yolo; + // tk::dnn::CenternetDetection cnet; +// tk::dnn::MobilenetDetection mbnet; + tk::dnn::DetectionNN *detNN; + int n_classes = classes; + std::vector images; + std::vector detected_bbox; + tk::dnn::Frame f; + + void init_bag(){tk::dnn::readmAPParams(config_filename, classes, map_points, map_levels, map_step, + IoU_thresh, conf_thresh, verbose); + + + //extract network name from rt path + std::string net_name; + removePathAndExtension(net, net_name); + std::cout<<"Network: "<init(net, n_classes, 1, conf_thresh); + + //read images + // std::ifstream all_labels(labels_path); +// std::cout << timeSinceEpochMillisec() << std::endl; + std::string l_filename; + + if(show) + cv::namedWindow("detection", cv::WINDOW_NORMAL); + return;} +// init_bag(); +// int images_done; + // for (images_done=0 ; std::getline(all_labels, l_filename) && images_done < n_images ; ++images_done) { + // std::cout < batch_frames; + batch_frames.push_back(frame); + int height = frame.rows; + int width = frame.cols; + std::cout< batch_dnn_input; + batch_dnn_input.push_back(frame.clone()); + std::cout<<"test1"<<"\n"; + //inference + detected_bbox.clear(); + detNN->update(batch_dnn_input,1); + detNN->draw(batch_frames); + detected_bbox = detNN->detected; + std::cout<<"test2"<<"\n"; + //try{ + //json::value response; + //vector jsonArray; + // save detections labels + for(auto d:detected_bbox){ + //convert detected bb in the same format as label + /// / / / + tk::dnn::BoundingBox b; + b.x = (d.x + d.w/2) / width; + b.y = (d.y + d.h/2) / height; + b.w = d.w / width; + b.h = d.h / height; + b.prob = d.prob; + b.cl = d.cl; + //f.det.push_back(b); + + /*json::value detection; + detection["label"] = json::value::number(b.cl); + detection["x"] = json::value::number(b.x); + detection["y"] = json::value::number(b.y); + detection["w"] = json::value::number(b.w); + detection["h"] = json::value::number(b.h); + detection["prob"] = json::value::number(b.prob); + jsonArray.push_back(detection);*/ + std::cout<< d.cl << " "<< d.prob << " "<< b.x << " "<< b.y << " "<< b.w << " "<< b.h <<"\n"; + + if(show)// draw rectangle for detection + cv::rectangle(batch_frames[0], cv::Point(d.x, d.y), cv::Point(d.x + d.w, d.y + d.h), cv::Scalar(0, 0, 255), 2); + } + //images.push_back(f); + + if(show){ + cv::imshow("detection", batch_frames[0]); + cv::waitKey(0); + } + // response["detections"] = json::value::array(jsonArray); //JSON Response + + // std::cout << timeSinceEpochMillisec() << std::endl; + return ; + } + +int main() +{ +// int sockfd, newsockfd, portno, n, imgSize, bytes=0, IM_HEIGHT, IM_WIDTH;; + // socklen_t clilen; + char buffer[256]; + struct sockaddr_in serv_addr, cli_addr; + //init_bag() +// cv::Mat img; + + sockfd=socket(AF_INET, SOCK_STREAM, 0); + if(sockfd<0) error("ERROR opening socket"); + + bzero((char*)&serv_addr, sizeof(serv_addr)); + portno = PORT; + + serv_addr.sin_family=AF_INET; + serv_addr.sin_addr.s_addr=INADDR_ANY; + serv_addr.sin_port=htons(portno); + + if(bind(sockfd, (struct sockaddr *) &serv_addr, + sizeof(serv_addr))<0) error("ERROR on binding"); + + listen(sockfd,5); + clilen=sizeof(cli_addr); + + newsockfd=accept(sockfd, (struct sockaddr *) &cli_addr, &clilen); + if(newsockfd<0) error("ERROR on accept"); + + uchar sock[3]; + cout << sock <(i,j) = Vec3b(sockData[ptr+0],sockData[ptr+1],sockData[ptr+2]); + ptr=ptr+3; + } + std::cout<<"t3"<<"\n"; + int height = img.cols; + std::cout< -#include -#include /* srand, rand */ -#include -#include -#include "utils.h" - -#include -#include -#include -#include - -// #include "Yolo3Detection.h" -#include "CenternetDetection.h" - -bool gRun; -bool SAVE_RESULT = false; - -void sig_handler(int signo) { - std::cout<<"request gateway stop\n"; - gRun = false; -} - -int main(int argc, char *argv[]) { - - std::cout<<"detection\n"; - signal(SIGINT, sig_handler); - - - char *net = "resnet101_cnet.rt"; - if(argc > 1) - net = argv[1]; - char *input = "../demo/yolo_test.mp4"; - if(argc > 2) - input = argv[2]; - - // tk::dnn::Yolo3Detection yolo; - tk::dnn::CenternetDetection yolo; - yolo.init(net); - - gRun = true; - - cv::VideoCapture cap(input); - if(!cap.isOpened()) - gRun = false; - else - std::cout<<"camera started\n"; - - - cv::VideoWriter resultVideo; - if(SAVE_RESULT) { - int w = cap.get(cv::CAP_PROP_FRAME_WIDTH); - int h = cap.get(cv::CAP_PROP_FRAME_HEIGHT); - resultVideo.open("result.mp4", cv::VideoWriter::fourcc('M','P','4','V'), 30, cv::Size(w, h)); - } - - cv::Mat frame; - cv::Mat dnn_input; - cv::namedWindow("detection", cv::WINDOW_NORMAL); - - while(gRun) { - cap >> frame; - if(!frame.data) { - break; - } - - // this will be resized to the net format - dnn_input = frame.clone(); - // TODO: async infer - yolo.update(dnn_input); - - frame = yolo.draw(dnn_input); - // // draw dets - // for(int i=0; iclassesNames[b.cl]; - // float prob = b.prob; - - // // std::cout< +#include +#include /* srand, rand */ +#ifdef __linux__ +#include +#endif +#include "stb_image.h" +#include +#include "utils.h" +#include "baggageDetect.hpp" +#include "handler.h" +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include "Yolo3Detection.h" +//#include "CenternetDetection.h" +//#include "MobilenetDetection.h" +#include "evaluation.h" +#include "tkdnn.h" +#include +#include +#include +using namespace std; +using namespace cv; +#include +#include +#include +#include "image.h" +void free_image(image m) +{ + if(m.data){ + free(m.data); + } +} +image make_empty_image(int w, int h, int c) +{ + image out; + out.data = 0; + out.h = h; + out.w = w; + out.c = c; + return out; +} + +image make_image(int w, int h, int c) +{ + image out = make_empty_image(w,h,c); + out.data = (float*)calloc(h * w * c, sizeof(float)); + return out; +} +int check_mistakes = 0; +image load_image_file(unsigned char *image_data, int channels, int antilog, int gray, int width, int height) +{ + int w, h, c; + unsigned char *data = image_data; + w = width; + h = height; + c = channels; + + if (!image_data) { + if (check_mistakes) getchar(); + return make_image(10, 10, 3); + + } + if (channels) c = channels; + + int i,j,k; + image im = make_image(w, h, c); + for(k = 0; k < c; ++k){ + for(j = 0; j < h; ++j){ + for(i = 0; i < w; ++i){ + int dst_index = i + w*j + w*h*k; + int src_index = k + c*i + c*w*j; + (im).data[dst_index] = (float)image_data[src_index]/255.; + } + } + } + //free(data); + return im; +} +cv::Mat image_to_mat(image img) +{ + int channels = img.c; + int width = img.w; + int height = img.h; + cv::Mat mat = cv::Mat(height, width, CV_8UC(channels)); + int step = mat.step; + + for (int y = 0; y < img.h; ++y) { + for (int x = 0; x < img.w; ++x) { + for (int c = 0; c < img.c; ++c) { + float val = img.data[c*img.h*img.w + y*img.w + x]; + mat.data[y*step + x*img.c + c] = (unsigned char)(val * 255); + } + } + } + return mat; +} + std::vector classesNames; + image im; + cv::Mat frame; + cv::Mat gray; + int h=0; + int w=0; + int channels; + const char *config_filename = "config/config.yaml"; + const char * net1="config/yolo4x_fp32.rt"; + const char * net2="config/yolo4x_fp32.rt"; + const char * net3="config/yolo4x_fp32.rt"; + const char * net4="config/yolo4x_fp32.rt"; + const char * net5="config/yolo4x_fp16.rt"; + int classes1 , classes2 , classes3 , classes4 , classes5,len; + char * img_data; + + string ustring; + float conf_thresh1 , conf_thresh2 , conf_thresh3 , conf_thresh4 , conf_thresh5; + tk::dnn::Yolo3Detection yolo1; + tk::dnn::DetectionNN *detNN1; + tk::dnn::Yolo3Detection yolo2; + tk::dnn::DetectionNN *detNN2; + tk::dnn::Yolo3Detection yolo3; + tk::dnn::DetectionNN *detNN3; + tk::dnn::Yolo3Detection yolo4; + tk::dnn::DetectionNN *detNN4; + tk::dnn::Yolo3Detection yolo5; + tk::dnn::DetectionNN *detNN5; + unsigned char * sockData; + + std::vector batch_frames; + std::vector batch_dnn_input; + std::vector classesNames1; + std::vector classesNames2; + std::vector classesNames3; + std::vector classesNames4; + std::vector classesNames5; + std::vector images; + std::vector detected_bbox1; + std::vector detected_bbox2; + std::vector detected_bbox3; + std::vector detected_bbox4; + std::vector detected_bbox5; + //read parametersi + handler::handler(utility::string_t url):m_listener(url) +{ + m_listener.support(methods::POST, bind(&handler::handle_post, this, placeholders::_1)); + +} + +string name_from_path(string path) +{ + return path.substr(path.find_last_of("/\\")+1); +} + void handler::init_bag(){tk::dnn::readmAPParams(config_filename, classes1,conf_thresh1, classes2,conf_thresh2 + , classes3,conf_thresh3, classes4,conf_thresh4,classes5,conf_thresh5); + + detNN1 = &yolo1; + detNN1->init(net1, classes1, 1, conf_thresh1); + classesNames1=detNN1-> classesNames; + detNN2 = &yolo2; + detNN2->init(net2, classes2, 1, conf_thresh2); + classesNames2=detNN2-> classesNames; + detNN3 = &yolo3; + detNN3->init(net3, classes3, 1, conf_thresh4); + classesNames3=detNN3-> classesNames; + detNN4 = &yolo4; + detNN4->init(net4, classes4, 1, conf_thresh4); + classesNames4=detNN4-> classesNames; + detNN5 = &yolo5; + detNN5->init(net5, classes5, 1, conf_thresh5); + classesNames5=detNN5-> classesNames; + return;} + + + void handler::handle_post(http_request request){ + BOOST_LOG_TRIVIAL(info) << "[" << name_from_path(string(__FILE__)) << " " << __LINE__ << "] " << request.to_string(); + + map http_get_vars = uri::split_query(request.request_uri().query()); + map::iterator it = http_get_vars.find("name"); + // int len; + if(it == http_get_vars.end()) + { + BOOST_LOG_TRIVIAL(error) << "[" << name_from_path(string(__FILE__)) << " " << __LINE__ << "] " << "Image name not passed in query."; + request.reply(status_codes::UnprocessableEntity,"Please pass image name in the query."); + return; + }https://github.com/baggageai/baggageai-code-one.git + // std::cout< v) { + ustring = {v.begin(),v.end()}; + len = ustring.size(); + }).wait(); + unsigned char *idata; +try { //printSize(ustring); + BOOST_LOG_TRIVIAL(info) << "[" << name_from_path(string(__FILE__)) << " " << __LINE__ << "] " << "Detection Started"; + sockData = (unsigned char *)ustring.c_str(); + idata = stbi_load_from_memory(sockData, len, &w, &h, &channels, 0); + im = load_image_file(idata, channels, 0, 0, w, h); + batch_dnn_input.clear(); + //batch_frames.clear(); + gray=image_to_mat(im); +// free(im); + cv::Mat in[] = {gray, gray,gray}; + cv::merge(in, 3, frame); + batch_dnn_input.push_back(frame.clone()); + json::value response; + vector jsonArray; + + detected_bbox1.clear(); + detNN1->update(batch_dnn_input,1); + detected_bbox1 = detNN1->detected; + + for(auto d1:detected_bbox1){ + json::value detection; + std::cout<<"1"<<" "<< d1.cl << " "<< d1.prob << " "<< d1.x << " "<< d1.y << " "<< d1.w << " "<< d1.h <<"\n"; + detection["label"] = json::value::string(classesNames1[d1.cl]); + detection["x"] = json::value::number(d1.x); + detection["y"] = json::value::number(d1.y); + detection["w"] = json::value::number(d1.w); + detection["h"] = json::value::number(d1.h); + detection["prob"] = json::value::number(d1.prob); + jsonArray.push_back(detection); +} + detected_bbox2.clear(); + batch_dnn_input.clear(); + batch_dnn_input.push_back(frame.clone()); + detNN2->update(batch_dnn_input,1); + // std::cout<detected; + + for(auto d2:detected_bbox2){ + std::cout<<"2"<<" "<< d2.cl << " "<< d2.prob << " "<< d2.x << " "<< d2.y << " "<< d2.w << " "<< d2.h <<"\n"; + json::value detection; + detection["label"] = json::value::string(classesNames2[d2.cl]); + detection["x"] = json::value::number(d2.x); + detection["y"] = json::value::number(d2.y); + detection["w"] = json::value::number(d2.w); + detection["h"] = json::value::number(d2.h); + detection["prob"] = json::value::number(d2.prob); + jsonArray.push_back(detection); +} + detected_bbox3.clear(); + batch_dnn_input.clear(); + batch_dnn_input.push_back(frame.clone()); + detNN3->update(batch_dnn_input,1); + detected_bbox3 = detNN3->detected; + for(auto d3:detected_bbox3){ + std::cout<< "3"<<" "<update(batch_dnn_input,1); + detected_bbox4 = detNN4->detected; + + for(auto d4:detected_bbox4){ + std::cout<<"4"<<" "<update(batch_dnn_input,1); + detected_bbox5 = detNN5->detected; + + for(auto d5:detected_bbox5){ + std::cout<<"5"<<" "< -#include -#include -#include -#include -#include -#include - - -#include "tkdnn.h" - -void sort(dnnType *src_begin, dnnType *src_end, int *idsrc); -void topk(dnnType *src_begin, int *idsrc, int K, float *topk_scores, - int *topk_inds, float *topk_ys, float *topk_xs); -void sortAndTopKonDevice(dnnType *src_begin, int *idsrc, float *topk_scores, int *topk_inds, float *topk_ys, float *topk_xs, const int size, const int K, const int n_classes); -void subtractWithThreshold(dnnType *src_begin, dnnType *src_end, dnnType *src2_begin, dnnType *src_out); -void topKxyclasses(int *ids_begin, int *ids_end, const int K, const int size, const int wh, int *clses, int *xs, int *ys); -void topKxyAddOffset(int * ids_begin, const int K, const int size, int *intxs_begin, int *intys_begin, float *xs_begin, float *ys_begin, dnnType *src_begin); -void bboxes(int * ids_begin, const int K, const int size, float *xs_begin, float *ys_begin, dnnType *src_begin, float *bbx0, float *bbx1, float *bby0, float *bby1); diff --git a/include/tkDNN/BoundingBox.h b/include/tkDNN/BoundingBox.h new file mode 100644 index 0000000..7f7449c --- /dev/null +++ b/include/tkDNN/BoundingBox.h @@ -0,0 +1,31 @@ +#ifndef BOUNDINGBOX_H +#define BOUNDINGBOX_H + +#include +#include "tkdnn.h" + +namespace tk { namespace dnn { +class BoundingBox : public tk::dnn::box +{ + float overlap(const float p1, const float l1, const float p2, const float l22); + float boxesIntersection(const BoundingBox &b); + float boxesUnion(const BoundingBox &b); + + public: + + int uniqueTruthIndex = -1; + int truthFlag = 0; + float maxIoU = 0; + + float IoU(const BoundingBox &b); + void clear(); + + friend std::ostream& operator<<(std::ostream& os, const BoundingBox& bb); +}; + +std::ostream& operator<<(std::ostream& os, const BoundingBox& bb); +bool boxComparison (const BoundingBox& a,const BoundingBox& b) ; + +}} +#endif /*BOUNDINGBOX_H*/ + diff --git a/include/tkDNN/CenternetDetection.h b/include/tkDNN/CenternetDetection.h index 8071112..3c8cfbb 100644 --- a/include/tkDNN/CenternetDetection.h +++ b/include/tkDNN/CenternetDetection.h @@ -1,112 +1,86 @@ -#include -#include -#include -#include /* srand, rand */ -#include -#include -#include "utils.h" -#include +#ifndef CENTERNETDETECTION_H +#define CENTERNETDETECTION_H + #include "kernels.h" +#include +#include "opencv2/opencv.hpp" +#include #include #include // std::iota #include // std::sort +#include "DetectionNN.h" -#include -#include -#include +#include "kernelsThrust.h" -#include "tkdnn.h" -#include "sorting.h" -namespace tk { namespace dnn { +namespace tk { namespace dnn { -/** - * - * @author Francesco Gatti - */ -class CenternetDetection { +class CenternetDetection : public DetectionNN +{ +private: + tk::dnn::dataDim_t dim; + tk::dnn::dataDim_t dim2; + tk::dnn::dataDim_t dim_hm; + tk::dnn::dataDim_t dim_wh; + tk::dnn::dataDim_t dim_reg; + float *topk_scores; + int *topk_inds_; + float *topk_ys_; + float *topk_xs_; + int *ids_d, *ids_, *ids_2, *ids_2d; - private: - tk::dnn::NetworkRT *netRT = nullptr; - dnnType *input_h, *input, *input_d; + float *scores, *scores_d; + int *clses, *clses_d; + int *topk_inds_d; + float *topk_ys_d; + float *topk_xs_d; + int *inttopk_xs_d, *inttopk_ys_d; - int ndets = 0; - // tk::dnn::Yolo::detection *dets = nullptr; - cv::Mat imageF; - cv::Mat bgr[3]; - - // variable to test cnet on dog pictures - tk::dnn::dataDim_t dim; - tk::dnn::dataDim_t dim2; - cv::Size sz; - const char *input_bin = "../tests/resnet101_cnet/debug/input.bin"; - - // pre-process - tk::dnn::dataDim_t dim_hm; - tk::dnn::dataDim_t dim_wh; - tk::dnn::dataDim_t dim_reg; - float *topk_scores; - int *topk_inds_; - float *topk_ys_; - float *topk_xs_; - int *ids_d, *ids_, *ids_2, *ids_2d; - - float *scores, *scores_d; - int *clses, *clses_d; - int *topk_inds_d; - float *topk_ys_d; - float *topk_xs_d; - int *inttopk_xs_d, *inttopk_ys_d; + float *bbx0, *bby0, *bbx1, *bby1; + float *bbx0_d, *bby0_d, *bbx1_d, *bby1_d; + float *target_coords; - float *bbx0, *bby0, *bbx1, *bby1; - float *bbx0_d, *bby0_d, *bbx1_d, *bby1_d; - - float *target_coords; - + #ifdef OPENCV_CUDACONTRIB + float *mean_d; + float *stddev_d; + #else cv::Vec mean; cv::Vec stddev; - cv::Mat src; - cv::Mat dst; - //processing - float toll = 0.000001; - int K = 100; - int width = 56; // TODO + dnnType *input; + #endif + + float *d_ptrs; + + cv::Mat src; + cv::Mat dst; + cv::Mat dst2; + cv::Mat trans, trans2; + //processing + float toll = 0.000001; + int K = 100; + int width = 128;//56; // TODO + + // pointer used in the kernels + float *src_out; + int *ids_out; + + struct threshold op; - - public: - dnnType *rt_out[4]; - - float inp_height = 224;//512; - float inp_width = 224;//512; - - int classes = 80; - int num = 0; - int n_masks = 0; - float thresh = 0.3; - cv::Scalar colors[256]; - - // this is filled with results - std::vector detected; - // draw - std::vector coco_class_name; - - CenternetDetection() {} - - virtual ~CenternetDetection() {} - - /** - * Method used for inizialize the class - * - * @return Success of the initialization - */ - bool init(std::string tensor_path); - void testdog(); - cv::Mat draw(cv::Mat &frame); - void update(cv::Mat &frame); +public: + CenternetDetection() {}; + ~CenternetDetection() {}; + bool init(const std::string& tensor_path, const int n_classes=80, const int n_batches=1, const float conf_thresh=0.3); + void preprocess(cv::Mat &frame, const int bi=0); + void postprocess(const int bi=0,const bool mAP=false); }; -}} + +} // namespace dnn +} // namespace tk + + +#endif /*CENTERNETDETECTION_H*/ \ No newline at end of file diff --git a/include/tkDNN/DarknetParser.h b/include/tkDNN/DarknetParser.h new file mode 100644 index 0000000..089c4d6 --- /dev/null +++ b/include/tkDNN/DarknetParser.h @@ -0,0 +1,51 @@ +#pragma once +#include +#include "tkDNN/tkdnn.h" + +namespace tk { namespace dnn { + + struct darknetFields_t{ + std::string type = ""; + int width = 0; + int height = 0; + int channels = 3; + int batch_normalize=0; + int groups = 1; + int group_id = 0; + int filters=1; + int size_x=1; + int size_y=1; + int stride_x=1; + int stride_y=1; + int padding_x = 0; + int padding_y = 0; + int n_mask = 0; + int classes = 20; + int num = 1; + int pad = 0; + int coords = 4; + int nms_kind = 0; + int new_coords= 0; + float scale_xy = 1; + float nms_thresh = 0.45; + std::vector layers; + std::string activation = "linear"; + + friend std::ostream& operator<<(std::ostream& os, const darknetFields_t& f){ + os << f.width << " " << f.height << " " << f.channels << " " << f.batch_normalize<< " " << f.filters << " " << f.activation<< " " << f.scale_xy; + return os; + } + }; + + std::string darknetParseType(const std::string& line); + bool divideNameAndValue(const std::string& line, std::string&name, std::string& value); + std::vector fromStringToIntVec(const std::string& line, const char delimiter); + + bool darknetParseFields(const std::string& line, darknetFields_t& fields); + tk::dnn::Network *darknetAddNet(darknetFields_t &fields); + void darknetAddLayer(tk::dnn::Network *net, darknetFields_t &f, std::string wgs_path, + std::vector &netLayers, const std::vector& names); + std::vector darknetReadNames(const std::string& names_file); + tk::dnn::Network* darknetParser(const std::string& cfg_file, const std::string& wgs_path, const std::string& names_file); + +}} diff --git a/include/tkDNN/DetectionNN.h b/include/tkDNN/DetectionNN.h new file mode 100644 index 0000000..a8c81f7 --- /dev/null +++ b/include/tkDNN/DetectionNN.h @@ -0,0 +1,185 @@ +#ifndef DETECTIONNN_H +#define DETECTIONNN_H + +#include +#include +#include +#ifdef __linux__ +#include +#endif + +#include +#include "utils.h" + +#include +#include +#include + +#include "tkdnn.h" + +//#define OPENCV_CUDACONTRIB //if OPENCV has been compiled with CUDA and contrib. + +#ifdef OPENCV_CUDACONTRIB +#include +#include +#endif + + +namespace tk { namespace dnn { + +class DetectionNN { + + protected: + tk::dnn::NetworkRT *netRT = nullptr; + dnnType *input_d; + + std::vector originalSize; + + cv::Scalar colors[256]; + + int nBatches = 1; + +#ifdef OPENCV_CUDACONTRIB + cv::cuda::GpuMat bgr[3]; + cv::cuda::GpuMat imagePreproc; +#else + cv::Mat bgr[3]; + cv::Mat imagePreproc; + dnnType *input; +#endif + + /** + * This method preprocess the image, before feeding it to the NN. + * + * @param frame original frame to adapt for inference. + * @param bi batch index + */ + virtual void preprocess(cv::Mat &frame, const int bi=0) = 0; + + /** + * This method postprocess the output of the NN to obtain the correct + * boundig boxes. + * + * @param bi batch index + * @param mAP set to true only if all the probabilities for a bounding + * box are needed, as in some cases for the mAP calculation + */ + virtual void postprocess(const int bi=0,const bool mAP=false) = 0; + + public: + int classes = 0; + float confThreshold = 0.3; /*threshold on the confidence of the boxes*/ + + std::vector detected; /*bounding boxes in output*/ + std::vector> batchDetected; /*bounding boxes in output*/ + std::vector stats; /*keeps track of inference times (ms)*/ + std::vector classesNames; + + DetectionNN() {}; + ~DetectionNN(){}; + + /** + * Method used to initialize the class, allocate memory and compute + * needed data. + * + * @param tensor_path path to the rt file of the NN. + * @param n_classes number of classes for the given dataset. + * @param n_batches maximum number of batches to use in inference + * @return true if everything is correct, false otherwise. + */ + virtual bool init(const std::string& tensor_path, const int n_classes=80, const int n_batches=1, const float conf_thresh=0.3) = 0; + + /** + * This method performs the whole detection of the NN. + * + * @param frames frames to run detection on. + * @param cur_batches number of batches to use in inference + * @param save_times if set to true, preprocess, inference and postprocess times + * are saved on a csv file, otherwise not. + * @param times pointer to the output stream where to write times + * @param mAP set to true only if all the probabilities for a bounding + * box are needed, as in some cases for the mAP calculation + */ + void update(std::vector& frames, const int cur_batches=1, bool save_times=false, std::ofstream *times=nullptr, const bool mAP=false){ + if(save_times && times==nullptr) + FatalError("save_times set to true, but no valid ofstream given"); + if(cur_batches > nBatches) + FatalError("A batch size greater than nBatches cannot be used"); + + originalSize.clear(); + if(TKDNN_VERBOSE) printCenteredTitle(" TENSORRT detection ", '=', 30); + { + TKDNN_TSTART + for(int bi=0; biinput_dim; + dim.n = cur_batches; + { + if(TKDNN_VERBOSE) dim.print(); + TKDNN_TSTART + netRT->infer(dim, input_d); + TKDNN_TSTOP + if(TKDNN_VERBOSE) dim.print(); + stats.push_back(t_ns); + if(save_times) *times<& frames) { + tk::dnn::box b; + int x0, w, x1, y0, h, y1; + int objClass; + std::string det_class; + int baseline = 0; + float font_scale = 0.5; + int thickness = 2; + + for(int bi=0; bi +#include +#include /* srand, rand */ + +#ifdef __linux__ +#include +#elif _WIN32 +#define _USE_MATH_DEFINES +#include +#endif + +#include +#include +#include "utils.h" +#include "tkdnn.h" + +namespace tk { namespace dnn { + +/** + * + * @author Francesco Gatti + */ +class ImuOdom { + + public: + tk::dnn::Network *net = nullptr; + + // Network input dim + tk::dnn::dataDim_t dim0; + tk::dnn::dataDim_t dim1; + tk::dnn::dataDim_t dim2; + + // Network output dim + tk::dnn::dataDim_t odim0; + tk::dnn::dataDim_t odim1; + + // input pointers + dnnType *i0_d, *i1_d, *i2_d; + // output pointers + dnnType *o0_d, *o1_d; + + // output eigen CPU + Eigen::MatrixXf deltaP, deltaQ; + + Eigen::MatrixXd odomPOS, odomEULER; + Eigen::Matrix3d odomROT; + Eigen::Isometry3f tf = Eigen::Isometry3f::Identity(); + + ImuOdom() {} + + virtual ~ImuOdom() {} + + /** + * Method used for initialize the class + * + * @return Success of the initialization + */ + bool init(std::string layers_path) { + + dim0 = tk::dnn::dataDim_t(1, 4, 1, 100); + dim1 = tk::dnn::dataDim_t(1, 3, 1, 100); + dim2 = tk::dnn::dataDim_t(1, 3, 1, 100); + + checkCuda( cudaMalloc(&i0_d, dim0.tot()*sizeof(dnnType)) ); + checkCuda( cudaMalloc(&i1_d, dim1.tot()*sizeof(dnnType)) ); + checkCuda( cudaMalloc(&i2_d, dim2.tot()*sizeof(dnnType)) ); + + std::string c0_bin = layers_path + "/conv1d_7.bin"; + std::string c1_bin = layers_path + "/conv1d_8.bin"; + std::string c2_bin = layers_path + "/conv1d_9.bin"; + std::string c3_bin = layers_path + "/conv1d_10.bin"; + std::string c4_bin = layers_path + "/conv1d_11.bin"; + std::string c5_bin = layers_path + "/conv1d_12.bin"; + std::string l0_bin = layers_path + "/bidirectional_3.bin"; + std::string l1_bin = layers_path + "/bidirectional_4.bin"; + std::string d0_bin = layers_path + "/dense_3.bin"; + std::string d1_bin = layers_path + "/dense_4.bin"; + + net = new tk::dnn::Network(dim0); + tk::dnn::Input *x0 = new tk::dnn::Input (net, dim0, i0_d); + tk::dnn::Conv2d *x0_0 = new tk::dnn::Conv2d (net, 128, 1, 11, 1, 1, 0, 0, c0_bin); + tk::dnn::Conv2d *x0_1 = new tk::dnn::Conv2d (net, 128, 1, 11, 1, 1, 0, 0, c1_bin); + tk::dnn::Pooling *x0_2 = new tk::dnn::Pooling(net, 1, 3, 1, 3 ,0, 0, tk::dnn::tkdnnPoolingMode_t::POOLING_MAX); + + tk::dnn::Input *x1 = new tk::dnn::Input (net, dim1, i1_d); + tk::dnn::Conv2d *x1_0 = new tk::dnn::Conv2d (net, 128, 1, 11, 1, 1, 0, 0, c2_bin); + tk::dnn::Conv2d *x1_1 = new tk::dnn::Conv2d (net, 128, 1, 11, 1, 1, 0, 0, c3_bin); + tk::dnn::Pooling *x1_2 = new tk::dnn::Pooling(net, 1, 3, 1, 3, 0, 0, tk::dnn::tkdnnPoolingMode_t::POOLING_MAX); + + tk::dnn::Input *x2 = new tk::dnn::Input (net, dim2, i2_d); + tk::dnn::Conv2d *x2_0 = new tk::dnn::Conv2d (net, 128, 1, 11, 1, 1, 0, 0, c4_bin); + tk::dnn::Conv2d *x2_1 = new tk::dnn::Conv2d (net, 128, 1, 11, 1, 1, 0, 0, c5_bin); + tk::dnn::Pooling *x2_2 = new tk::dnn::Pooling(net, 1, 3, 1, 3, 0, 0, tk::dnn::tkdnnPoolingMode_t::POOLING_MAX); + + tk::dnn::Layer *concat_l[3] = { x0_2, x1_2, x2_2 }; + tk::dnn::Route *concat = new tk::dnn::Route(net, concat_l, 3); + + tk::dnn::LSTM *lstm0 = new tk::dnn::LSTM(net, 128, true, l0_bin); + tk::dnn::LSTM *lstm1 = new tk::dnn::LSTM(net, 128, false, l1_bin); + + tk::dnn::Dense *d0 = new tk::dnn::Dense(net, 3, d0_bin); + + tk::dnn::Layer *lstm1_l[1] = { lstm1 }; + tk::dnn::Route *lstm1_link = new tk::dnn::Route(net, lstm1_l, 1); + tk::dnn::Dense *d1 = new tk::dnn::Dense(net, 4, d1_bin); + + net->print(); + + // output data + o0_d = d0->dstData; + o1_d = d1->dstData; + odim0 = d0->output_dim; + odim1 = d1->output_dim; + + deltaP.resize(odim0.tot(), 1); + deltaQ.resize(odim1.tot(), 1); + + odomPOS = Eigen::MatrixXd::Zero(3, 1); + odomROT = Eigen::MatrixXd::Identity(3, 3); + odomEULER = Eigen::MatrixXd::Zero(3, 1); + return true; + } + + void close() { + // TODO: dealloc :) + } + + void update(dnnType *x0, dnnType *x1, dnnType *x2) { + + checkCuda( cudaMemcpy(i0_d, x0, dim0.tot()*sizeof(dnnType), cudaMemcpyHostToDevice) ); + checkCuda( cudaMemcpy(i1_d, x1, dim1.tot()*sizeof(dnnType), cudaMemcpyHostToDevice) ); + checkCuda( cudaMemcpy(i2_d, x2, dim2.tot()*sizeof(dnnType), cudaMemcpyHostToDevice) ); + + // Inference + tk::dnn::dataDim_t dim; + net->infer(dim, nullptr); + + checkCuda( cudaMemcpy(deltaP.data(), o0_d, odim0.tot()*sizeof(dnnType), cudaMemcpyDeviceToHost) ); + checkCuda( cudaMemcpy(deltaQ.data(), o1_d, odim1.tot()*sizeof(dnnType), cudaMemcpyDeviceToHost) ); + + // compute odom + Eigen::Quaterniond q; + q.w() = deltaQ(0); + q.x() = deltaQ(1); + q.y() = deltaQ(2); + q.z() = deltaQ(3); + odomPOS = odomPOS + odomROT*deltaP.cast(); // V1 + //odomPOS = odomPOS + deltaP.cast(); // V2 + odomROT = odomROT * q.normalized().toRotationMatrix(); + + // compute Euler + auto newEULER = odomROT.eulerAngles(0, 1, 2); + for(int i=0; i<3; i++) { + while( fabs(newEULER(i) - odomEULER(i)) > M_PI_2 ) { + newEULER(i) += newEULER(i) - odomEULER(i) > 0 ? -M_PI : +M_PI; + //std::cout<(); + tf.matrix().block(0, 3, 3, 1) = odomPOS.cast(); + } + +}; + +}} diff --git a/include/tkDNN/Int8BatchStream.h b/include/tkDNN/Int8BatchStream.h new file mode 100644 index 0000000..c39a11c --- /dev/null +++ b/include/tkDNN/Int8BatchStream.h @@ -0,0 +1,72 @@ +#ifndef INT8BATCHSTREAM_H +#define INT8BATCHSTREAM_H + +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#ifdef __linux__ +#include +#endif + +#include + +#include "NvInfer.h" +#include "utils.h" +#include "tkdnn.h" + +/* + * BatchStream implements the stream for the INT8 calibrator. + * It reads the two files .txt with the list of image file names + * and the list of label file names. + * It then iterates on images and labels. + */ +class BatchStream { +public: + BatchStream(tk::dnn::dataDim_t dim, int batchSize, int maxBatches, const std::string& fileimglist, const std::string& filelabellist); + virtual ~BatchStream() { } + void reset(int firstBatch); + bool next(); + void skip(int skipCount); + float *getBatch() { return mBatch.data(); } + float *getLabels() { return mLabels.data(); } + int getBatchesRead() const { return mBatchCount; } + int getBatchSize() const { return mBatchSize; } + nvinfer1::DimsNCHW getDims() const { return mDims; } + float* getFileBatch() { return &mFileBatch[0]; } + float* getFileLabels() { return &mFileLabels[0]; } + void readInListFile(const std::string& dataFilePath, std::vector& mListIn); + void readCVimage(std::string inputFileName, std::vector& res, bool fixshape = true); + void readLabels(std::string inputFileName ,std::vector& ris); + bool update(); + +private: + int mBatchSize{ 0 }; + int mMaxBatches{ 0 }; + int mBatchCount{ 0 }; + int mFileCount{ 0 }; + int mFileBatchPos{ 0 }; + int mImageSize{ 0 }; + + nvinfer1::DimsNCHW mDims; + std::vector mBatch; + std::vector mLabels; + std::vector mFileBatch; + std::vector mFileLabels; + + int mHeight; + int mWidth; + std::string mFileImgList; + std::vector mListImg; + std::string mFileLabelList; + std::vector mListLabel; +}; + +#endif //INT8BATCHSTREAM \ No newline at end of file diff --git a/include/tkDNN/Int8Calibrator.h b/include/tkDNN/Int8Calibrator.h new file mode 100644 index 0000000..4a0ea47 --- /dev/null +++ b/include/tkDNN/Int8Calibrator.h @@ -0,0 +1,49 @@ +#ifndef INT8CALIBRATOR_H +#define INT8CALIBRATOR_H + +#include +#include +#include +#include +#include +#include +#include +#include "NvInfer.h" + +#include +#include + +#include "Int8BatchStream.h" + +#include "tkdnn.h" +#include "utils.h" + +/* + * Int8EntropyCalibrator implements the INT8 calibrator to achieve the + * INT8 quantization. It uses a BatchStream stream to scroll through + * images data. It also implements the calibration cache, a way to + * save the calibration process results to reduce the running time: + * the calibration process takes a long time. + */ +class Int8EntropyCalibrator : public nvinfer1::IInt8EntropyCalibrator { +public: + Int8EntropyCalibrator(BatchStream& stream, int firstBatch, const std::string& calibTableFilePath, + const std::string& inputBlobName, bool readCache = true); + virtual ~Int8EntropyCalibrator() { checkCuda(cudaFree(mDeviceInput)); } + int getBatchSize() const override { return mStream.getBatchSize(); } + bool getBatch(void* bindings[], const char* names[], int nbBindings) override; + const void* readCalibrationCache(size_t& length) override; + void writeCalibrationCache(const void* cache, size_t length) override; + +private: + BatchStream mStream; + const std::string mCalibTableFilePath{ nullptr }; + const std::string mInputBlobName; + bool mReadCache{ true }; + + size_t mInputCount; + void* mDeviceInput{ nullptr }; + std::vector mCalibrationCache; +}; + +#endif //INT8CALIBRATOR_H \ No newline at end of file diff --git a/include/tkDNN/Layer.h b/include/tkDNN/Layer.h index 2d84a0e..e097372 100644 --- a/include/tkDNN/Layer.h +++ b/include/tkDNN/Layer.h @@ -9,12 +9,19 @@ namespace tk { namespace dnn { enum layerType_t { + LAYER_INPUT, LAYER_DENSE, LAYER_CONV2D, LAYER_DECONV2D, LAYER_DEFORMCONV2D, + LAYER_LSTM, LAYER_ACTIVATION, + LAYER_ACTIVATION_CRELU, + LAYER_ACTIVATION_LEAKY, + LAYER_ACTIVATION_MISH, + LAYER_ACTIVATION_LOGISTIC, LAYER_FLATTEN, + LAYER_RESHAPE, LAYER_MULADD, LAYER_POOLING, LAYER_SOFTMAX, @@ -34,7 +41,7 @@ enum layerType_t { class Layer { public: - Layer(Network *net, bool final = false); + Layer(Network *net); virtual ~Layer(); virtual layerType_t getLayerType() = 0; @@ -42,9 +49,9 @@ public: std::cout<<"No infer action for this layer\n"; return NULL; } - + void setFinal() { this->final = true; } dataDim_t input_dim, output_dim; - dnnType *dstData; //where results will be putted + dnnType *dstData = nullptr; //where results will be putted int id = 0; bool final; //if the layer is the final one @@ -52,22 +59,29 @@ public: std::string getLayerName() { layerType_t type = getLayerType(); switch(type) { - case LAYER_DENSE: return "Dense"; - case LAYER_CONV2D: return "Conv2d"; - case LAYER_DECONV2D: return "DeConv2d"; - case LAYER_DEFORMCONV2D:return "DeformConv2d"; - case LAYER_ACTIVATION: return "Activation"; - case LAYER_FLATTEN: return "Flatten"; - case LAYER_MULADD: return "MulAdd"; - case LAYER_POOLING: return "Pooling"; - case LAYER_SOFTMAX: return "Softmax"; - case LAYER_ROUTE: return "Route"; - case LAYER_REORG: return "Reorg"; - case LAYER_SHORTCUT: return "Shortcut"; - case LAYER_UPSAMPLE: return "Upsample"; - case LAYER_REGION: return "Region"; - case LAYER_YOLO: return "Yolo"; - default: return "unknown"; + case LAYER_INPUT: return "Input"; + case LAYER_DENSE: return "Dense"; + case LAYER_CONV2D: return "Conv2d"; + case LAYER_DECONV2D: return "DeConv2d"; + case LAYER_DEFORMCONV2D: return "DeformConv2d"; + case LAYER_LSTM: return "LSTM"; + case LAYER_ACTIVATION: return "Activation"; + case LAYER_ACTIVATION_CRELU: return "ActivationCReLU"; + case LAYER_ACTIVATION_LEAKY: return "ActivationLeaky"; + case LAYER_ACTIVATION_MISH: return "ActivationMish"; + case LAYER_ACTIVATION_LOGISTIC: return "ActivationLogistic"; + case LAYER_FLATTEN: return "Flatten"; + case LAYER_RESHAPE: return "Reshape"; + case LAYER_MULADD: return "MulAdd"; + case LAYER_POOLING: return "Pooling"; + case LAYER_SOFTMAX: return "Softmax"; + case LAYER_ROUTE: return "Route"; + case LAYER_REORG: return "Reorg"; + case LAYER_SHORTCUT: return "Shortcut"; + case LAYER_UPSAMPLE: return "Upsample"; + case LAYER_REGION: return "Region"; + case LAYER_YOLO: return "Yolo"; + default: return "unknown"; } } @@ -85,7 +99,7 @@ class LayerWgs : public Layer { public: LayerWgs(Network *net, int inputs, int outputs, int kh, int kw, int kt, - std::string fname_weights, bool batchnorm = false, bool additional_bias = false, bool final = false); + std::string fname_weights, bool batchnorm = false, bool additional_bias = false, bool deConv = false, int groups = 1); virtual ~LayerWgs(); int inputs, outputs; @@ -96,23 +110,87 @@ public: // additional bias for DCN bool additional_bias; - dnnType *bias2_h, *bias2_d; + dnnType *bias2_h = nullptr, *bias2_d = nullptr; //batchnorm bool batchnorm; - dnnType *power_h; - dnnType *scales_h, *scales_d; - dnnType *mean_h, *mean_d; - dnnType *variance_h, *variance_d; + dnnType *power_h = nullptr; + dnnType *scales_h = nullptr, *scales_d = nullptr; + dnnType *mean_h = nullptr, *mean_d = nullptr; + dnnType *variance_h = nullptr, *variance_d = nullptr; //fp16 - __half *data16_h, *bias16_h; - __half *data16_d, *bias16_d; + __half *data16_h = nullptr, *bias16_h = nullptr; + __half *data16_d = nullptr, *bias16_d = nullptr; + __half *bias216_h = nullptr, *bias216_d = nullptr; - __half *power16_h, *power16_d; - __half *scales16_h, *scales16_d; - __half *mean16_h, *mean16_d; - __half *variance16_h, *variance16_d; + __half *power16_h = nullptr, *power16_d = nullptr; + __half *scales16_h = nullptr, *scales16_d = nullptr; + __half *mean16_h = nullptr, *mean16_d = nullptr; + __half *variance16_h = nullptr, *variance16_d = nullptr; + + void releaseHost(bool release32 = true, bool release16 = true) { + if(release32) { + if( data_h != nullptr) { delete [] data_h; data_h = nullptr; } + if( bias_h != nullptr) { delete [] bias_h; bias_h = nullptr; } + if( bias2_h != nullptr) { delete [] bias2_h; bias2_h = nullptr; } + if( scales_h != nullptr) { delete [] scales_h; scales_h = nullptr; } + if( mean_h != nullptr) { delete [] mean_h; mean_h = nullptr; } + if(variance_h != nullptr) { delete [] variance_h; variance_h = nullptr; } + if( power_h != nullptr) { delete [] power_h; power_h = nullptr; } + } + if(net->fp16 && release16) { + if( data16_h != nullptr) { delete [] data16_h; data16_h = nullptr; } + if( bias16_h != nullptr) { delete [] bias16_h; bias16_h = nullptr; } + if( bias216_h != nullptr) { delete [] bias216_h; bias216_h = nullptr; } + if( scales16_h != nullptr) { delete [] scales16_h; scales16_h = nullptr; } + if( mean16_h != nullptr) { delete [] mean16_h; mean16_h = nullptr; } + if(variance16_h != nullptr) { delete [] variance16_h; variance16_h = nullptr; } + if( power16_h != nullptr) { delete [] power16_h; power16_h = nullptr; } + + } + } + void releaseDevice(bool release32 = true, bool release16 = true) { + if(release32) { + if( data_d != nullptr) { cudaFree( data_d); data_d = nullptr; } + if( bias_d != nullptr) { cudaFree( bias_d); bias_d = nullptr; } + if( bias2_d != nullptr) { cudaFree( bias2_d); bias2_d = nullptr; } + if( scales_d != nullptr) { cudaFree( scales_d); scales_d = nullptr; } + if( mean_d != nullptr) { cudaFree( mean_d); mean_d = nullptr; } + if(variance_d != nullptr) { cudaFree(variance_d); variance_d = nullptr; } + } + if(net->fp16 && release16) { + if( data16_d != nullptr) { cudaFree( data16_d); data16_d = nullptr; } + if( bias16_d != nullptr) { cudaFree( bias16_d); bias16_d = nullptr; } + if( bias216_d != nullptr) { cudaFree( bias216_d); bias216_d = nullptr; } + if( scales16_d != nullptr) { cudaFree( scales16_d); scales16_d = nullptr; } + if( mean16_d != nullptr) { cudaFree( mean16_d); mean16_d = nullptr; } + if(variance16_d != nullptr) { cudaFree(variance16_d); variance16_d = nullptr; } + if( power16_d != nullptr) { cudaFree( power16_d); power16_d = nullptr; } + } + } +}; + + +/** + Input layer (it doesn't need weights) +*/ +class Input : public Layer { + +public: + + Input(Network *net, dataDim_t &dim, dnnType* srcData) : Layer(net) { + input_dim = dim; + output_dim = dim; + dstData = srcData; + } + virtual ~Input() {} + virtual layerType_t getLayerType() { return LAYER_INPUT; }; + + virtual dnnType* infer(dataDim_t &dim, dnnType* srcData) { + dim = output_dim; + return dstData; + } }; @@ -131,24 +209,38 @@ public: /** - Avaible activation functions + Available activation functions */ typedef enum { ACTIVATION_ELU = 100, - ACTIVATION_LEAKY = 101 + ACTIVATION_LEAKY = 101, + ACTIVATION_MISH = 102, + ACTIVATION_LOGISTIC = 103 } tkdnnActivationMode_t; /** - Activation layer (it doesnt need weigths) + Activation layer (it doesn't need weights) */ class Activation : public Layer { public: int act_mode; + float ceiling; - Activation(Network *net, int act_mode); + Activation(Network *net, int act_mode, const float ceiling=0.0); virtual ~Activation(); - virtual layerType_t getLayerType() { return LAYER_ACTIVATION; }; + virtual layerType_t getLayerType() { + if(act_mode == CUDNN_ACTIVATION_CLIPPED_RELU) + return LAYER_ACTIVATION_CRELU; + else if (act_mode == ACTIVATION_LEAKY) + return LAYER_ACTIVATION_LEAKY; + else if (act_mode == ACTIVATION_MISH) + return LAYER_ACTIVATION_MISH; + else if (act_mode == ACTIVATION_LOGISTIC) + return LAYER_ACTIVATION_LOGISTIC; + else + return LAYER_ACTIVATION; + }; virtual dnnType* infer(dataDim_t &dim, dnnType* srcData); @@ -159,26 +251,35 @@ protected: /** Convolutional 2D layer + + WEIGHTS shape: OUTCH, INCH, KH, KW ... + BIAS shape: OUTCH + + with BATCHNORM: + scales: OUTCH + means: OUTCH + variance: OUTCH */ class Conv2d : public LayerWgs { public: Conv2d( Network *net, int out_ch, int kernelH, int kernelW, int strideH, int strideW, int paddingH, int paddingW, - std::string fname_weights, bool batchnorm = false, bool deConv = false, bool final = false); + std::string fname_weights, bool batchnorm = false, bool deConv = false, int groups = 1, bool additional_bias=false); virtual ~Conv2d(); virtual layerType_t getLayerType() { return LAYER_CONV2D; }; virtual dnnType* infer(dataDim_t &dim, dnnType* srcData); int kernelH, kernelW, strideH, strideW, paddingH, paddingW; - bool deConv; + bool deConv, additional_bias; + int groups; protected: cudnnFilterDescriptor_t filterDesc; cudnnConvolutionDescriptor_t convDesc; - cudnnConvolutionFwdAlgo_t algo; - cudnnConvolutionBwdDataAlgo_t bwAlgo; + cudnnConvolutionFwdAlgoPerf_t algo; + cudnnConvolutionBwdDataAlgoPerf_t bwAlgo; cudnnTensorDescriptor_t biasTensorDesc; void initCUDNN(bool back = false); @@ -187,6 +288,71 @@ protected: size_t ws_sizeInBytes; }; +/** + Bidirectional LSTM layer + ONLY BIDIRECTIONAL (TODO: more configurable) + currently implemented as 2 inferences: forward and backward (TODO: only 1 cudnn inference) + + implementation info: + https://github.com/jiangnanhugo/seq2seq_cuda/blob/e4dbdcfa0517c972bfd4beea9f11a5233954093c/src/rnn.cpp + https://github.com/Jeffery-Song/mxnet-test/blob/aab666faad44011f7a67b527b5f6c960367d0422/src/operator/cudnn_rnn-inl.h + https://stackoverflow.com/a/38737941 + https://colah.github.io/posts/2015-08-Understanding-LSTMs/ + + PARAMS (numlayers*2): + layer0: + ( INCH, ? ) ??? + ( HIDDEN, ? ) ??? + ( HIDDEN * 8 ) ??? + layer2: + ( INCH, ? ) ??? + ( HIDDEN, ? ) ??? + ( HIDDEN * 8 ) ??? + + OUTPUT shape: + (N, C, 1, W) ---> LSTM(HIDDEN, returnSeq=True) ---> (N, 2*HIDDEN, 1, W) # W is seqLength + (N, C, 1, W) ---> LSTM(HIDDEN, returnSeq=False) ---> (N, 2*HIDDEN, 1, 1) +*/ +class LSTM : public Layer { + +public: + LSTM(Network *net, int hiddensize, bool returnSeq, std::string fname_weights); + virtual ~LSTM(); + virtual layerType_t getLayerType() { return LAYER_LSTM; }; + + virtual dnnType* infer(dataDim_t &dim, dnnType* srcData); + + const bool bidirectional = true; /**> is the net bidir */ + bool returnSeq = false; /**> if false return only the result of last timestamp */ + int stateSize = 0; /**> number of hidden states */ + int seqLen = 0; /**> number of timestamp */ + int numLayers = 1; /**> number of internal layers */ + +protected: + cudnnRNNDescriptor_t rnnDesc; + cudnnDropoutDescriptor_t dropoutDesc; + dnnType *dropout_states_, *work_space_; + + size_t workspace_byte_, dropout_byte_; + int workspace_size_, dropout_size_; + + std::vector x_desc_vec_, y_desc_vec_; + cudnnTensorDescriptor_t hx_desc_, cx_desc_; + cudnnTensorDescriptor_t hy_desc_, cy_desc_; + dnnType *hx_ptr, *cx_ptr, *hy_ptr, *cy_ptr; + int stateDataDim; + + cudnnFilterDescriptor_t w_desc_; + dnnType *w_ptr; + dnnType *w_h; + dnnType *wf_ptr, *wb_ptr; // params pointer forward and backward layer + + // used during inference + dataDim_t one_output_dim; // output dim of as single inference + dnnType *srcF, *srcB; // input of single inference + dnnType *dstF, *dstB_NR, *dstB; // output of single inference, dstB_NR = dstB not reversed +}; + /** Convolutional 2D layer @@ -196,8 +362,8 @@ class DeConv2d : public Conv2d { public: DeConv2d( Network *net, int out_ch, int kernelH, int kernelW, int strideH, int strideW, int paddingH, int paddingW, - std::string fname_weights, bool batchnorm = false) : - Conv2d(net, out_ch, kernelH, kernelW, strideH, strideW, paddingH, paddingW, fname_weights, batchnorm, true) {} + std::string fname_weights, bool batchnorm = false, int groups = 1) : + Conv2d(net, out_ch, kernelH, kernelW, strideH, strideW, paddingH, paddingW, fname_weights, batchnorm, true, groups) {} virtual ~DeConv2d() {} virtual layerType_t getLayerType() { return LAYER_DECONV2D; }; @@ -206,7 +372,7 @@ public: /** - Deformable Convolutionl 2d layer + Deformable Convolutional 2d layer */ class DeformConv2d : public LayerWgs { @@ -228,6 +394,9 @@ public: dnnType *offset, *mask; dnnType *output_conv; + cublasStatus_t stat; + cublasHandle_t handle; + protected: cudnnTensorDescriptor_t biasTensorDesc; @@ -249,6 +418,20 @@ public: virtual dnnType* infer(dataDim_t &dim, dnnType* srcData); }; +/** + Reshape layer +*/ +class Reshape : public Layer { + +public: + Reshape(Network *net, dataDim_t new_dim); + virtual ~Reshape(); + virtual layerType_t getLayerType() { return LAYER_RESHAPE; }; + + virtual dnnType* infer(dataDim_t &dim, dnnType* srcData); + +}; + /** MulAdd layer @@ -271,17 +454,18 @@ protected: /** - Avaible pooling functions (padding on tkDNN is not supported) + Available pooling functions (padding on tkDNN is not supported) */ typedef enum { POOLING_MAX = 0, - POOLING_AVERAGE = 1, // count for average includes padded values - POOLING_AVERAGE_EXCLUDE_PADDING = 2 // count for average does not include padded values + POOLING_AVERAGE = 1, // count for average includes padded values + POOLING_AVERAGE_EXCLUDE_PADDING = 2, // count for average does not include padded values + POOLING_MAX_FIXEDSIZE = 100 // max pool darknet fashion } tkdnnPoolingMode_t; /** Pooling layer - currenty supported only 2d pooing (also on 3d input) + currently supported only 2d pooing (also on 3d input) */ class Pooling : public Layer { @@ -289,12 +473,13 @@ public: int winH, winW; int strideH, strideW; int paddingH, paddingW; + bool size; tkdnnPoolingMode_t pool_mode; Pooling(Network *net, int winH, int winW, int strideH, int strideW, - int paddingH = 0, int paddingW = 0, - tkdnnPoolingMode_t pool_mode = POOLING_MAX, bool final = false); + int paddingH, int paddingW, + tkdnnPoolingMode_t pool_mode); virtual ~Pooling(); virtual layerType_t getLayerType() { return LAYER_POOLING; }; @@ -313,11 +498,13 @@ protected: class Softmax : public Layer { public: - Softmax(Network *net); + Softmax(Network *net, const tk::dnn::dataDim_t* dim=nullptr, const cudnnSoftmaxMode_t mode=CUDNN_SOFTMAX_MODE_CHANNEL); virtual ~Softmax(); virtual layerType_t getLayerType() { return LAYER_SOFTMAX; }; virtual dnnType* infer(dataDim_t &dim, dnnType* srcData); + dataDim_t dim; + cudnnSoftmaxMode_t mode; }; /** @@ -327,22 +514,24 @@ public: class Route : public Layer { public: - Route(Network *net, Layer **layers, int layers_n); + Route(Network *net, Layer **layers, int layers_n, int groups = 1, int group_id = 0); virtual ~Route(); virtual layerType_t getLayerType() { return LAYER_ROUTE; }; virtual dnnType* infer(dataDim_t &dim, dnnType* srcData); public: - static const int MAX_INPUT_LAYERS = 16; - Layer *layers[MAX_INPUT_LAYERS]; //ids of layers to be merged + static const int MAX_LAYERS = 32; + Layer *layers[MAX_LAYERS]; //ids of layers to be merged int layers_n; //number of layers + int groups; + int group_id; }; /** Reorg layer - Mantain same dimension but change C*H*W distribution + Maintains same dimension but change C*H*W distribution */ class Reorg : public Layer { @@ -375,7 +564,7 @@ public: /** Upsample layer - Mantain same dimension but change C*H*W distribution + Maintains same dimension but change C*H*W distribution */ class Upsample : public Layer { @@ -394,6 +583,12 @@ struct box { int cl; float x, y, w, h; float prob; + std::vector probs; + + void print() + { + std::cout<<"x: "< classesNames; virtual dnnType* infer(dataDim_t &dim, dnnType* srcData); - int computeDetections(Yolo::detection *dets, int &ndets, int netw, int neth, float thresh); + int computeDetections(Yolo::detection *dets, int &ndets, int netw, int neth, float thresh, int new_coords=0); dnnType *predictions; - static const int MAX_DETECTIONS = 256; + static const int MAX_DETECTIONS = 8192*2; static Yolo::detection *allocateDetections(int nboxes, int classes); - static void mergeDetections(Yolo::detection *dets, int ndets, int classes); + static void mergeDetections(Yolo::detection *dets, int ndets, int classes, double nms_thresh=0.45, nmsKind_t nsm_kind=GREEDY_NMS); }; /** diff --git a/include/tkDNN/MobilenetDetection.h b/include/tkDNN/MobilenetDetection.h new file mode 100644 index 0000000..9a5fedc --- /dev/null +++ b/include/tkDNN/MobilenetDetection.h @@ -0,0 +1,77 @@ +#ifndef MOBILENETDETECTION_H +#define MOBILENETDETECTION_H + +#include +#include "opencv2/opencv.hpp" + +#include "DetectionNN.h" + +#define N_COORDS 4 +#define N_SSDSPEC 6 + +namespace tk { namespace dnn { + +struct SSDSpec +{ + int featureSize = 0; + int shrinkage = 0; + int boxWidth = 0; + int boxHeight = 0; + int ratio1 = 0; + int ratio2 = 0; + + SSDSpec() {} + SSDSpec(int feature_size, int shrinkage, int box_width, int box_height, int ratio1, int ratio2) : + featureSize(feature_size), shrinkage(shrinkage), boxWidth(box_width), + boxHeight(box_height), ratio1(ratio1), ratio2(ratio2) {} + void setAll(int feature_size, int shrinkage, int box_width, int box_height, int ratio1, int ratio2) + { + this->featureSize = feature_size; + this->shrinkage = shrinkage; + this->boxWidth = box_width; + this->boxHeight = box_height; + this->ratio1 = ratio1; + this->ratio2 = ratio2; + } + void print() + { + std::cout << "fsize: " << featureSize << "\tshrinkage: " << shrinkage << + "\t box W:" << boxWidth << "\tbox H: " << boxHeight << + "\t x ratio:" << ratio1 << "\t y ratio:" << ratio2 << std::endl; + } +}; + +class MobilenetDetection : public DetectionNN +{ +private: + float IoUThreshold = 0.45; + float centerVariance = 0.1; + float sizeVariance = 0.2; + int imageSize; + + float *priors = nullptr; + int nPriors = 0; + float *locations_h, *confidences_h; + + + + void generate_ssd_priors(const SSDSpec *specs, const int n_specs, bool clamp = true); + void convert_locatios_to_boxes_and_center(); + float iou(const tk::dnn::box &a, const tk::dnn::box &b); + + + +public: + MobilenetDetection() {}; + ~MobilenetDetection() {}; + + bool init(const std::string& tensor_path, const int n_classes, const int n_batches=1, const float conf_thresh=0.3); + void preprocess(cv::Mat &frame, const int bi=0); + void postprocess(const int bi=0,const bool mAP=false); +}; + + +} // namespace dnn +} // namespace tk + +#endif /*MOBILENETDETECTION_H*/ \ No newline at end of file diff --git a/include/tkDNN/Network.h b/include/tkDNN/Network.h index b234f71..b78acff 100644 --- a/include/tkDNN/Network.h +++ b/include/tkDNN/Network.h @@ -1,17 +1,18 @@ #ifndef NETWORK_H #define NETWORK_H +#include #include "utils.h" namespace tk { namespace dnn { /** - Data rapresentation beetween layers + Data representation between layers n = batch size c = channels - h = heigth (lines) + h = height (lines) w = width (rows) - l = lenght (3rd dimension) + l = length (3rd dimension) */ struct dataDim_t { @@ -39,14 +40,16 @@ class Network { public: Network(dataDim_t input_dim); virtual ~Network(); + void releaseLayers(); /** - Do inferece for every added layer + Do inference for every added layer */ dnnType* infer(dataDim_t &dim, dnnType* data); bool addLayer(Layer *l); void print(); + const char *getNetworkRTName(const char *network_name); cudnnDataType_t dataType; cudnnTensorFormat_t tensorFormat; @@ -59,8 +62,14 @@ public: dataDim_t input_dim; dataDim_t getOutputDim(); - bool fp16, dla; + bool fp16, dla, int8; + int maxBatchSize; bool dontLoadWeights; + std::string fileImgList; + std::string fileLabelList; + std::string networkName; + std::string networkNameRT; + }; }} diff --git a/include/tkDNN/NetworkRT.h b/include/tkDNN/NetworkRT.h index 82d082a..29cacf5 100644 --- a/include/tkDNN/NetworkRT.h +++ b/include/tkDNN/NetworkRT.h @@ -6,6 +6,7 @@ #include "Network.h" #include "Layer.h" #include "NvInfer.h" +#include namespace tk { namespace dnn { @@ -24,15 +25,20 @@ template T readBUF(const char*& buffer) using namespace nvinfer1; #include "pluginsRT/ActivationLeakyRT.h" +#include "pluginsRT/ActivationLogisticRT.h" +#include "pluginsRT/ActivationReLUCeilingRT.h" +#include "pluginsRT/ActivationMishRT.h" #include "pluginsRT/ReorgRT.h" #include "pluginsRT/RegionRT.h" -//#include "pluginsRT/RouteRT.h" +#include "pluginsRT/RouteRT.h" #include "pluginsRT/ShortcutRT.h" #include "pluginsRT/YoloRT.h" #include "pluginsRT/UpsampleRT.h" #include "pluginsRT/ResizeLayerRT.h" -//#include "pluginsRT/Int8Calibrator.h" #include "pluginsRT/DeformableConvRT.h" +#include "pluginsRT/FlattenConcatRT.h" +#include "pluginsRT/ReshapeRT.h" +#include "pluginsRT/MaxPoolingFixedSizeRT.h" class PluginFactory : IPluginFactory { @@ -52,12 +58,16 @@ public: nvinfer1::IBuilder *builderRT; nvinfer1::IRuntime *runtimeRT; nvinfer1::INetworkDefinition *networkRT; +#if NV_TENSORRT_MAJOR >= 6 + nvinfer1::IBuilderConfig *configRT; +#endif nvinfer1::ICudaEngine *engineRT; nvinfer1::IExecutionContext *contextRT; const static int MAX_BUFFERS_RT = 10; void* buffersRT[MAX_BUFFERS_RT]; + dataDim_t buffersDIM[MAX_BUFFERS_RT]; int buf_input_idx, buf_output_idx; dataDim_t input_dim, output_dim; @@ -69,11 +79,25 @@ public: NetworkRT(Network *net, const char *name); virtual ~NetworkRT(); + int getMaxBatchSize() { + if(engineRT != nullptr) + return engineRT->getMaxBatchSize(); + else + return 0; + } + + int getBuffersN() { + if(engineRT != nullptr) + return engineRT->getNbBindings(); + else + return 0; + } + /** - Do inferece + Do inference */ dnnType* infer(dataDim_t &dim, dnnType* data); - void enqueue(); + void enqueue(int batchSize = 1); nvinfer1::ILayer* convert_layer(nvinfer1::ITensor *input, Layer *l); nvinfer1::ILayer* convert_layer(nvinfer1::ITensor *input, Conv2d *l); @@ -82,6 +106,8 @@ public: nvinfer1::ILayer* convert_layer(nvinfer1::ITensor *input, Pooling *l); nvinfer1::ILayer* convert_layer(nvinfer1::ITensor *input, Softmax *l); nvinfer1::ILayer* convert_layer(nvinfer1::ITensor *input, Route *l); + nvinfer1::ILayer* convert_layer(nvinfer1::ITensor *input, Flatten *l); + nvinfer1::ILayer* convert_layer(nvinfer1::ITensor *input, Reshape *l); nvinfer1::ILayer* convert_layer(nvinfer1::ITensor *input, Reorg *l); nvinfer1::ILayer* convert_layer(nvinfer1::ITensor *input, Region *l); nvinfer1::ILayer* convert_layer(nvinfer1::ITensor *input, Shortcut *l); @@ -91,6 +117,9 @@ public: bool serialize(const char *filename); bool deserialize(const char *filename); + + + }; }} diff --git a/include/tkDNN/NetworkViz.h b/include/tkDNN/NetworkViz.h new file mode 100644 index 0000000..c8b1bea --- /dev/null +++ b/include/tkDNN/NetworkViz.h @@ -0,0 +1,12 @@ +#pragma once +#include +#include +#include "tkdnn.h" + +namespace tk { namespace dnn { + +cv::Mat vizFloat2colorMap(cv::Mat map); +cv::Mat vizData2Mat(dnnType *dataInput, tk::dnn::dataDim_t dim, int imgdim); +cv::Mat vizLayer2Mat(tk::dnn::Network *net, int layer, int imgdim = 1000); + +}} diff --git a/include/tkDNN/Yolo3Detection.h b/include/tkDNN/Yolo3Detection.h index 6c873bc..100a720 100644 --- a/include/tkDNN/Yolo3Detection.h +++ b/include/tkDNN/Yolo3Detection.h @@ -1,65 +1,36 @@ -#include -#include -#include /* srand, rand */ -#include -#include -#include "utils.h" +#ifndef Yolo3Detection_H +#define Yolo3Detection_H +#include +#include "opencv2/opencv.hpp" -#include -#include -#include +#include "DetectionNN.h" -#include "tkdnn.h" +namespace tk { namespace dnn { -namespace tk { namespace dnn { +class Yolo3Detection : public DetectionNN +{ +private: + int num = 0; + int nMasks = 0; + int nDets = 0; + tk::dnn::Yolo::detection *dets = nullptr; + tk::dnn::Yolo* yolo[3]; -/** - * - * @author Francesco Gatti - */ -class Yolo3Detection { + tk::dnn::Yolo* getYoloLayer(int n=0); - private: - tk::dnn::NetworkRT *netRT = nullptr; - tk::dnn::Yolo* yolo[3]; - dnnType *input, *input_d; - - int ndets = 0; - tk::dnn::Yolo::detection *dets = nullptr; - - cv::Mat imageF; - cv::Mat bgr[3]; - - public: - int classes = 0; - int num = 0; - int n_masks = 0; - float thresh = 0.3; - cv::Scalar colors[256]; - - // this is filled with results - std::vector detected; - - Yolo3Detection() {} - - virtual ~Yolo3Detection() {} - - /** - * Method used for inizialize the class - * - * @return Success of the initialization - */ - bool init(std::string tensor_path); - - void update(cv::Mat &frame); - - tk::dnn::Yolo* getYoloLayer(int n=0) { - if(n<3) - return yolo[n]; - else - return nullptr; - } + cv::Mat bgr_h; + +public: + Yolo3Detection() {}; + ~Yolo3Detection() {}; + bool init(const std::string& tensor_path, const int n_classes=80, const int n_batches=1, const float conf_thresh=0.3); + void preprocess(cv::Mat &frame, const int bi=0); + void postprocess(const int bi=0,const bool mAP=false); }; -}} + +} // namespace dnn +} // namespace tk + +#endif /* Yolo3Detection_H*/ diff --git a/include/tkDNN/baggageDetect.hpp b/include/tkDNN/baggageDetect.hpp new file mode 100644 index 0000000..5a92d3b --- /dev/null +++ b/include/tkDNN/baggageDetect.hpp @@ -0,0 +1,59 @@ + +#ifdef OS_WIN + #pragma once + + #ifdef LIB_EXPORTS + #define LIB_API __declspec(dllexport) + #else + #define LIB_API __declspec(dllimport) + #endif +#endif + +#include + + +#include +#include + + +#ifdef OPENCV +#include +#include +using namespace cv; +#endif + + +using namespace std; + +struct baggagedetector { + int x,y,w,h,size; + char *label; + float prob; +}; + +#ifdef __cplusplus +class baggageAI +{ + //std::shared_ptr detector_gpu_ptr; + public: + //static LIB_API image_t image_load(std::string image_filename); + #ifdef OS_WIN + LIB_API baggageAI(); + //LIB_API ~baggageAI(); + LIB_API baggagedetector * baggageDetections(char *input); + LIB_API baggagedetector * baggageDetections(unsigned char *input, int len, int antiLog, int gray); + #ifdef OPENCV + LIB_API baggagedetector * baggageDetections(Mat m); + #endif + #else + baggageAI(); + //LIB_API ~baggageAI(); + baggagedetector * baggageDetections(char *input); + baggagedetector * baggageDetections(unsigned char *input, int len,int antiLog, int gray); + #ifdef OPENCV + baggagedetector * baggageDetections(Mat m); + #endif + #endif +}; + +#endif diff --git a/include/tkDNN/darknet.h b/include/tkDNN/darknet.h new file mode 100644 index 0000000..fb97cdf --- /dev/null +++ b/include/tkDNN/darknet.h @@ -0,0 +1,1032 @@ +#ifndef DARKNET_API +#define DARKNET_API + +#if defined(_MSC_VER) && _MSC_VER < 1900 +#define inline __inline +#endif + +#if defined(DEBUG) && !defined(_CRTDBG_MAP_ALLOC) +#define _CRTDBG_MAP_ALLOC +#endif + +#include +#include +#include +#include +#include +#include + +#ifndef LIB_API +#ifdef LIB_EXPORTS +#if defined(_MSC_VER) +#define LIB_API __declspec(dllexport) +#else +#define LIB_API __attribute__((visibility("default"))) +#endif +#else +#if defined(_MSC_VER) +#define LIB_API +#else +#define LIB_API +#endif +#endif +#endif + +#define SECRET_NUM -1234 + +typedef enum { UNUSED_DEF_VAL } UNUSED_ENUM_TYPE; + +#ifdef GPU + +#include +#include +#include + +#ifdef CUDNN +#include +#endif // CUDNN +#endif // GPU + +#ifdef __cplusplus +extern "C" { +#endif + +struct network; +typedef struct network network; + +struct network_state; +typedef struct network_state network_state; + +struct layer; +typedef struct layer layer; + +struct image; +typedef struct image image; + +struct detection; +typedef struct detection detection; + +struct load_args; +typedef struct load_args load_args; + +struct data; +typedef struct data data; + +struct metadata; +typedef struct metadata metadata; + +struct tree; +typedef struct tree tree; + +extern int gpu_index; + +// option_list.h +typedef struct metadata { + int classes; + char **names; +} metadata; + + +// tree.h +typedef struct tree { + int *leaf; + int n; + int *parent; + int *child; + int *group; + char **name; + + int groups; + int *group_size; + int *group_offset; +} tree; + + +// activations.h +typedef enum { + LOGISTIC, RELU, RELU6, RELIE, LINEAR, RAMP, TANH, PLSE, LEAKY, ELU, LOGGY, STAIR, HARDTAN, LHTAN, SELU, GELU, SWISH, MISH, NORM_CHAN, NORM_CHAN_SOFTMAX, NORM_CHAN_SOFTMAX_MAXVAL +}ACTIVATION; + +// parser.h +typedef enum { + IOU, GIOU, MSE, DIOU, CIOU +} IOU_LOSS; + +// parser.h +typedef enum { + DEFAULT_NMS, GREEDY_NMS, DIOU_NMS, CORNERS_NMS +} NMS_KIND; + +// parser.h +typedef enum { + YOLO_CENTER = 1 << 0, YOLO_LEFT_TOP = 1 << 1, YOLO_RIGHT_BOTTOM = 1 << 2 +} YOLO_POINT; + +// parser.h +typedef enum { + NO_WEIGHTS, PER_FEATURE, PER_CHANNEL +} WEIGHTS_TYPE_T; + +// parser.h +typedef enum { + NO_NORMALIZATION, RELU_NORMALIZATION, SOFTMAX_NORMALIZATION +} WEIGHTS_NORMALIZATION_T; + +// image.h +typedef enum{ + PNG, BMP, TGA, JPG +} IMTYPE; + +// activations.h +typedef enum{ + MULT, ADD, SUB, DIV +} BINARY_ACTIVATION; + +// layer.h +typedef enum { + CONVOLUTIONAL, + DECONVOLUTIONAL, + CONNECTED, + MAXPOOL, + LOCAL_AVGPOOL, + SOFTMAX, + DETECTION, + DROPOUT, + CROP, + ROUTE, + COST, + NORMALIZATION, + AVGPOOL, + LOCAL, + SHORTCUT, + SCALE_CHANNELS, + SAM, + ACTIVE, + RNN, + GRU, + LSTM, + CONV_LSTM, + CRNN, + BATCHNORM, + NETWORK, + XNOR, + REGION, + YOLO, + GAUSSIAN_YOLO, + ISEG, + REORG, + REORG_OLD, + UPSAMPLE, + LOGXENT, + L2NORM, + EMPTY, + BLANK +} LAYER_TYPE; + +// layer.h +typedef enum{ + SSE, MASKED, L1, SEG, SMOOTH,WGAN +} COST_TYPE; + +// layer.h +typedef struct update_args { + int batch; + float learning_rate; + float momentum; + float decay; + int adam; + float B1; + float B2; + float eps; + int t; +} update_args; + +// layer.h +struct layer { + LAYER_TYPE type; + ACTIVATION activation; + COST_TYPE cost_type; + void(*forward) (struct layer, struct network_state); + void(*backward) (struct layer, struct network_state); + void(*update) (struct layer, int, float, float, float); + void(*forward_gpu) (struct layer, struct network_state); + void(*backward_gpu) (struct layer, struct network_state); + void(*update_gpu) (struct layer, int, float, float, float, float); + layer *share_layer; + int train; + int avgpool; + int batch_normalize; + int shortcut; + int batch; + int dynamic_minibatch; + int forced; + int flipped; + int inputs; + int outputs; + float mean_alpha; + int nweights; + int nbiases; + int extra; + int truths; + int h, w, c; + int out_h, out_w, out_c; + int n; + int max_boxes; + int groups; + int group_id; + int size; + int side; + int stride; + int stride_x; + int stride_y; + int dilation; + int antialiasing; + int maxpool_depth; + int out_channels; + int reverse; + int flatten; + int spatial; + int pad; + int sqrt; + int flip; + int index; + int scale_wh; + int binary; + int xnor; + int peephole; + int use_bin_output; + int keep_delta_gpu; + int optimized_memory; + int steps; + int state_constrain; + int hidden; + int truth; + float smooth; + float dot; + int deform; + int sway; + int rotate; + int stretch; + int stretch_sway; + float angle; + float jitter; + float saturation; + float exposure; + float shift; + float ratio; + float learning_rate_scale; + float clip; + int focal_loss; + float *classes_multipliers; + float label_smooth_eps; + int noloss; + int softmax; + int classes; + int coords; + int background; + int rescore; + int objectness; + int does_cost; + int joint; + int noadjust; + int reorg; + int log; + int tanh; + int *mask; + int total; + float bflops; + + int adam; + float B1; + float B2; + float eps; + + int t; + + float alpha; + float beta; + float kappa; + + float coord_scale; + float object_scale; + float noobject_scale; + float mask_scale; + float class_scale; + int bias_match; + float random; + float ignore_thresh; + float truth_thresh; + float iou_thresh; + float thresh; + float focus; + int classfix; + int absolute; + int assisted_excitation; + + int onlyforward; + int stopbackward; + int train_only_bn; + int dont_update; + int burnin_update; + int dontload; + int dontsave; + int dontloadscales; + int numload; + + float temperature; + float probability; + float dropblock_size_rel; + int dropblock_size_abs; + int dropblock; + float scale; + + int receptive_w; + int receptive_h; + int receptive_w_scale; + int receptive_h_scale; + + char * cweights; + int * indexes; + int * input_layers; + int * input_sizes; + float **layers_output; + float **layers_delta; + WEIGHTS_TYPE_T weights_type; + WEIGHTS_NORMALIZATION_T weights_normalization; + int * map; + int * counts; + float ** sums; + float * rand; + float * cost; + float * state; + float * prev_state; + float * forgot_state; + float * forgot_delta; + float * state_delta; + float * combine_cpu; + float * combine_delta_cpu; + + float *concat; + float *concat_delta; + + float *binary_weights; + + float *biases; + float *bias_updates; + + float *scales; + float *scale_updates; + + float *weights; + float *weight_updates; + + float scale_x_y; + int objectness_smooth; + float max_delta; + float uc_normalizer; + float iou_normalizer; + float cls_normalizer; + IOU_LOSS iou_loss; + IOU_LOSS iou_thresh_kind; + NMS_KIND nms_kind; + float beta_nms; + YOLO_POINT yolo_point; + + char *align_bit_weights_gpu; + float *mean_arr_gpu; + float *align_workspace_gpu; + float *transposed_align_workspace_gpu; + int align_workspace_size; + + char *align_bit_weights; + float *mean_arr; + int align_bit_weights_size; + int lda_align; + int new_lda; + int bit_align; + + float *col_image; + float * delta; + float * output; + float * activation_input; + int delta_pinned; + int output_pinned; + float * loss; + float * squared; + float * norms; + + float * spatial_mean; + float * mean; + float * variance; + + float * mean_delta; + float * variance_delta; + + float * rolling_mean; + float * rolling_variance; + + float * x; + float * x_norm; + + float * m; + float * v; + + float * bias_m; + float * bias_v; + float * scale_m; + float * scale_v; + + + float *z_cpu; + float *r_cpu; + float *h_cpu; + float *stored_h_cpu; + float * prev_state_cpu; + + float *temp_cpu; + float *temp2_cpu; + float *temp3_cpu; + + float *dh_cpu; + float *hh_cpu; + float *prev_cell_cpu; + float *cell_cpu; + float *f_cpu; + float *i_cpu; + float *g_cpu; + float *o_cpu; + float *c_cpu; + float *stored_c_cpu; + float *dc_cpu; + + float *binary_input; + uint32_t *bin_re_packed_input; + char *t_bit_input; + + struct layer *input_layer; + struct layer *self_layer; + struct layer *output_layer; + + struct layer *reset_layer; + struct layer *update_layer; + struct layer *state_layer; + + struct layer *input_gate_layer; + struct layer *state_gate_layer; + struct layer *input_save_layer; + struct layer *state_save_layer; + struct layer *input_state_layer; + struct layer *state_state_layer; + + struct layer *input_z_layer; + struct layer *state_z_layer; + + struct layer *input_r_layer; + struct layer *state_r_layer; + + struct layer *input_h_layer; + struct layer *state_h_layer; + + struct layer *wz; + struct layer *uz; + struct layer *wr; + struct layer *ur; + struct layer *wh; + struct layer *uh; + struct layer *uo; + struct layer *wo; + struct layer *vo; + struct layer *uf; + struct layer *wf; + struct layer *vf; + struct layer *ui; + struct layer *wi; + struct layer *vi; + struct layer *ug; + struct layer *wg; + + tree *softmax_tree; + + size_t workspace_size; + +//#ifdef GPU + int *indexes_gpu; + + float *z_gpu; + float *r_gpu; + float *h_gpu; + float *stored_h_gpu; + + float *temp_gpu; + float *temp2_gpu; + float *temp3_gpu; + + float *dh_gpu; + float *hh_gpu; + float *prev_cell_gpu; + float *prev_state_gpu; + float *last_prev_state_gpu; + float *last_prev_cell_gpu; + float *cell_gpu; + float *f_gpu; + float *i_gpu; + float *g_gpu; + float *o_gpu; + float *c_gpu; + float *stored_c_gpu; + float *dc_gpu; + + // adam + float *m_gpu; + float *v_gpu; + float *bias_m_gpu; + float *scale_m_gpu; + float *bias_v_gpu; + float *scale_v_gpu; + + float * combine_gpu; + float * combine_delta_gpu; + + float * forgot_state_gpu; + float * forgot_delta_gpu; + float * state_gpu; + float * state_delta_gpu; + float * gate_gpu; + float * gate_delta_gpu; + float * save_gpu; + float * save_delta_gpu; + float * concat_gpu; + float * concat_delta_gpu; + + float *binary_input_gpu; + float *binary_weights_gpu; + float *bin_conv_shortcut_in_gpu; + float *bin_conv_shortcut_out_gpu; + + float * mean_gpu; + float * variance_gpu; + float * m_cbn_avg_gpu; + float * v_cbn_avg_gpu; + + float * rolling_mean_gpu; + float * rolling_variance_gpu; + + float * variance_delta_gpu; + float * mean_delta_gpu; + + float * col_image_gpu; + + float * x_gpu; + float * x_norm_gpu; + float * weights_gpu; + float * weight_updates_gpu; + float * weight_deform_gpu; + float * weight_change_gpu; + + float * weights_gpu16; + float * weight_updates_gpu16; + + float * biases_gpu; + float * bias_updates_gpu; + float * bias_change_gpu; + + float * scales_gpu; + float * scale_updates_gpu; + float * scale_change_gpu; + + float * input_antialiasing_gpu; + float * output_gpu; + float * output_avg_gpu; + float * activation_input_gpu; + float * loss_gpu; + float * delta_gpu; + float * rand_gpu; + float * drop_blocks_scale; + float * drop_blocks_scale_gpu; + float * squared_gpu; + float * norms_gpu; + + float *gt_gpu; + float *a_avg_gpu; + + int *input_sizes_gpu; + float **layers_output_gpu; + float **layers_delta_gpu; +#ifdef CUDNN + cudnnTensorDescriptor_t srcTensorDesc, dstTensorDesc; + cudnnTensorDescriptor_t srcTensorDesc16, dstTensorDesc16; + cudnnTensorDescriptor_t dsrcTensorDesc, ddstTensorDesc; + cudnnTensorDescriptor_t dsrcTensorDesc16, ddstTensorDesc16; + cudnnTensorDescriptor_t normTensorDesc, normDstTensorDesc, normDstTensorDescF16; + cudnnFilterDescriptor_t weightDesc, weightDesc16; + cudnnFilterDescriptor_t dweightDesc, dweightDesc16; + cudnnConvolutionDescriptor_t convDesc; + cudnnConvolutionFwdAlgo_t fw_algo, fw_algo16; + cudnnConvolutionBwdDataAlgo_t bd_algo, bd_algo16; + cudnnConvolutionBwdFilterAlgo_t bf_algo, bf_algo16; + cudnnPoolingDescriptor_t poolingDesc; +#else // CUDNN + void* srcTensorDesc, *dstTensorDesc; + void* srcTensorDesc16, *dstTensorDesc16; + void* dsrcTensorDesc, *ddstTensorDesc; + void* dsrcTensorDesc16, *ddstTensorDesc16; + void* normTensorDesc, *normDstTensorDesc, *normDstTensorDescF16; + void* weightDesc, *weightDesc16; + void* dweightDesc, *dweightDesc16; + void* convDesc; + UNUSED_ENUM_TYPE fw_algo, fw_algo16; + UNUSED_ENUM_TYPE bd_algo, bd_algo16; + UNUSED_ENUM_TYPE bf_algo, bf_algo16; + void* poolingDesc; +#endif // CUDNN +//#endif // GPU +}; + + +// network.h +typedef enum { + CONSTANT, STEP, EXP, POLY, STEPS, SIG, RANDOM, SGDR +} learning_rate_policy; + +// network.h +typedef struct network { + int n; + int batch; + uint64_t *seen; + int *cur_iteration; + float loss_scale; + int *t; + float epoch; + int subdivisions; + layer *layers; + float *output; + learning_rate_policy policy; + int benchmark_layers; + + float learning_rate; + float learning_rate_min; + float learning_rate_max; + int batches_per_cycle; + int batches_cycle_mult; + float momentum; + float decay; + float gamma; + float scale; + float power; + int time_steps; + int step; + int max_batches; + int num_boxes; + int train_images_num; + float *seq_scales; + float *scales; + int *steps; + int num_steps; + int burn_in; + int cudnn_half; + + int adam; + float B1; + float B2; + float eps; + + int inputs; + int outputs; + int truths; + int notruth; + int h, w, c; + int max_crop; + int min_crop; + float max_ratio; + float min_ratio; + int center; + int flip; // horizontal flip 50% probability augmentaiont for classifier training (default = 1) + int gaussian_noise; + int blur; + int mixup; + float label_smooth_eps; + int resize_step; + int attention; + int adversarial; + float adversarial_lr; + int letter_box; + float angle; + float aspect; + float exposure; + float saturation; + float hue; + int random; + int track; + int augment_speed; + int sequential_subdivisions; + int init_sequential_subdivisions; + int current_subdivision; + int try_fix_nan; + + int gpu_index; + tree *hierarchy; + + float *input; + float *truth; + float *delta; + float *workspace; + int train; + int index; + float *cost; + float clip; + +//#ifdef GPU + //float *input_gpu; + //float *truth_gpu; + float *delta_gpu; + float *output_gpu; + + float *input_state_gpu; + float *input_pinned_cpu; + int input_pinned_cpu_flag; + + float **input_gpu; + float **truth_gpu; + float **input16_gpu; + float **output16_gpu; + size_t *max_input16_size; + size_t *max_output16_size; + int wait_stream; + + float *global_delta_gpu; + float *state_delta_gpu; + size_t max_delta_gpu_size; +//#endif // GPU + int optimized_memory; + int dynamic_minibatch; + size_t workspace_size_limit; +} network; + +// network.h +typedef struct network_state { + float *truth; + float *input; + float *delta; + float *workspace; + int train; + int index; + network net; +} network_state; + +//typedef struct { +// int w; +// int h; +// float scale; +// float rad; +// float dx; +// float dy; +// float aspect; +//} augment_args; + +// image.h +typedef struct image { + int w; + int h; + int c; + float *data; +} image; + +//typedef struct { +// int w; +// int h; +// int c; +// float *data; +//} image; + +// box.h +typedef struct box { + float x, y, w, h; +} box; + +// box.h +typedef struct boxabs { + float left, right, top, bot; +} boxabs; + +// box.h +typedef struct dxrep { + float dt, db, dl, dr; +} dxrep; + +// box.h +typedef struct ious { + float iou, giou, diou, ciou; + dxrep dx_iou; + dxrep dx_giou; +} ious; + + +// box.h +typedef struct detection{ + box bbox; + int classes; + float *prob; + float *mask; + float objectness; + int sort_class; + float *uc; // Gaussian_YOLOv3 - tx,ty,tw,th uncertainty + int points; // bit-0 - center, bit-1 - top-left-corner, bit-2 - bottom-right-corner +} detection; + +// network.c -batch inference +typedef struct det_num_pair { + int num; + detection *dets; +} det_num_pair, *pdet_num_pair; + +// matrix.h +typedef struct matrix { + int rows, cols; + float **vals; +} matrix; + +// data.h +typedef struct data { + int w, h; + matrix X; + matrix y; + int shallow; + int *num_boxes; + box **boxes; +} data; + +// data.h +typedef enum { + CLASSIFICATION_DATA, DETECTION_DATA, CAPTCHA_DATA, REGION_DATA, IMAGE_DATA, COMPARE_DATA, WRITING_DATA, SWAG_DATA, TAG_DATA, OLD_CLASSIFICATION_DATA, STUDY_DATA, DET_DATA, SUPER_DATA, LETTERBOX_DATA, REGRESSION_DATA, SEGMENTATION_DATA, INSTANCE_DATA, ISEG_DATA +} data_type; + +// data.h +typedef struct load_args { + int threads; + char **paths; + char *path; + int n; + int m; + char **labels; + int h; + int w; + int c; // color depth + int out_w; + int out_h; + int nh; + int nw; + int num_boxes; + int min, max, size; + int classes; + int background; + int scale; + int center; + int coords; + int mini_batch; + int track; + int augment_speed; + int letter_box; + int show_imgs; + int dontuse_opencv; + float jitter; + int flip; + int gaussian_noise; + int blur; + int mixup; + float label_smooth_eps; + float angle; + float aspect; + float saturation; + float exposure; + float hue; + data *d; + image *im; + image *resized; + data_type type; + tree *hierarchy; +} load_args; + +// data.h +typedef struct box_label { + int id; + float x, y, w, h; + float left, right, top, bottom; +} box_label; + +// list.h +//typedef struct node { +// void *val; +// struct node *next; +// struct node *prev; +//} node; + +// list.h +//typedef struct list { +// int size; +// node *front; +// node *back; +//} list; + +// ----------------------------------------------------- + + +// parser.c +LIB_API network *load_network(char *cfg, char *weights, int clear); +LIB_API network *load_network_custom(char *cfg, char *weights, int clear, int batch); +LIB_API network *load_network(char *cfg, char *weights, int clear); +LIB_API void free_network(network net); + +// network.c +LIB_API load_args get_base_args(network *net); + +// box.h +LIB_API void do_nms_sort(detection *dets, int total, int classes, float thresh); +LIB_API void do_nms_obj(detection *dets, int total, int classes, float thresh); +LIB_API void diounms_sort(detection *dets, int total, int classes, float thresh, NMS_KIND nms_kind, float beta1); + +// network.h +LIB_API float *network_predict(network net, float *input); +LIB_API float *network_predict_ptr(network *net, float *input); +LIB_API detection *get_network_boxes(network *net, int w, int h, float thresh, float hier, int *map, int relative, int *num, int letter); +LIB_API det_num_pair* network_predict_batch(network *net, image im, int batch_size, int w, int h, float thresh, float hier, int *map, int relative, int letter); +LIB_API void free_detections(detection *dets, int n); +LIB_API void free_batch_detections(det_num_pair *det_num_pairs, int n); +LIB_API void fuse_conv_batchnorm(network net); +LIB_API void calculate_binary_weights(network net); +LIB_API char *detection_to_json(detection *dets, int nboxes, int classes, char **names, long long int frame_id, char *filename); + +LIB_API layer* get_network_layer(network* net, int i); +//LIB_API detection *get_network_boxes(network *net, int w, int h, float thresh, float hier, int *map, int relative, int *num, int letter); +LIB_API detection *make_network_boxes(network *net, float thresh, int *num); +LIB_API void reset_rnn(network *net); +LIB_API float *network_predict_image(network *net, image im); +LIB_API float *network_predict_image_letterbox(network *net, image im); +LIB_API float validate_detector_map(char *datacfg, char *cfgfile, char *weightfile, float thresh_calc_avg_iou, const float iou_thresh, const int map_points, int letter_box, network *existing_net); +LIB_API void train_detector(char *datacfg, char *cfgfile, char *weightfile, int *gpus, int ngpus, int clear, int dont_show, int calc_map, int mjpeg_port, int show_imgs, int benchmark_layers, char* chart_path); +LIB_API void test_detector(char *datacfg, char *cfgfile, char *weightfile, char *filename, float thresh, + float hier_thresh, int dont_show, int ext_output, int save_labels, char *outfile, int letter_box, int benchmark_layers); +LIB_API int network_width(network *net); +LIB_API int network_height(network *net); +LIB_API void optimize_picture(network *net, image orig, int max_layer, float scale, float rate, float thresh, int norm); + +// image.h +LIB_API void make_image_red(image im); +LIB_API image make_attention_image(int img_size, float *original_delta_cpu, float *original_input_cpu, int w, int h, int c); +LIB_API image resize_image(image im, int w, int h); +LIB_API void quantize_image(image im); +LIB_API void copy_image_from_bytes(image im, char *pdata); +LIB_API image letterbox_image(image im, int w, int h); +LIB_API void rgbgr_image(image im); +LIB_API image make_image(int w, int h, int c); +LIB_API image load_image_color(char *filename, int w, int h); +LIB_API void free_image(image m); +LIB_API image crop_image(image im, int dx, int dy, int w, int h); +LIB_API image resize_min(image im, int min); + +// layer.h +LIB_API void free_layer_custom(layer l, int keep_cudnn_desc); +LIB_API void free_layer(layer l); + +// data.c +LIB_API void free_data(data d); +LIB_API pthread_t load_data(load_args args); +LIB_API void free_load_threads(void *ptr); +LIB_API pthread_t load_data_in_thread(load_args args); +LIB_API void *load_thread(void *ptr); + +// dark_cuda.h +LIB_API void cuda_pull_array(float *x_gpu, float *x, size_t n); +LIB_API void cuda_pull_array_async(float *x_gpu, float *x, size_t n); +LIB_API void cuda_set_device(int n); +LIB_API void *cuda_get_context(); + +// utils.h +LIB_API void free_ptrs(void **ptrs, int n); +LIB_API void top_k(float *a, int n, int k, int *index); + +// tree.h +LIB_API tree *read_tree(char *filename); + +// option_list.h +LIB_API metadata get_metadata(char *file); + + +// http_stream.h +LIB_API void delete_json_sender(); +LIB_API void send_json_custom(char const* send_buf, int port, int timeout); +LIB_API double get_time_point(); +void start_timer(); +void stop_timer(); +double get_time(); +void stop_timer_and_show(); +void stop_timer_and_show_name(char *name); +void show_total_time(); + +// gemm.h +LIB_API void init_cpu(); + +#ifdef __cplusplus +} +#endif // __cplusplus +#endif // DARKNET_API + diff --git a/include/tkDNN/dimensionless.h b/include/tkDNN/dimensionless.h new file mode 100644 index 0000000..52eaa5e --- /dev/null +++ b/include/tkDNN/dimensionless.h @@ -0,0 +1,852 @@ +#ifndef DIMENSIONLESS_API +#define DIMENSIONLESS_API + +#if defined(_MSC_VER) && _MSC_VER < 1900 +#define inline __inline +#endif + +#if defined(DEBUG) && !defined(_CRTDBG_MAP_ALLOC) +#define _CRTDBG_MAP_ALLOC +#endif + +#include +#include +#include +#include +#include +#include + +#ifndef LIB_API +#ifdef LIB_EXPORTS +#if defined(_MSC_VER) +#define LIB_API __declspec(dllexport) +#else +#define LIB_API __attribute__((visibility("default"))) +#endif +#else +#if defined(_MSC_VER) +#define LIB_API +#else +#define LIB_API +#endif +#endif +#endif + +#define SECRET_NUM -1234 + +#ifdef GPU + +#include "cuda_runtime.h" +#include "curand.h" +#include "cublas_v2.h" + +#ifdef CUDNN +#include "cudnn.h" +#endif +#endif + +#ifdef __cplusplus +extern "C" { +#endif + +struct network; +typedef struct network network; + +struct network_state; +typedef struct network_state network_state; + +struct layer; +typedef struct layer layer; + +struct image; +typedef struct image image; + +struct detection; +typedef struct detection detection; + +struct load_args; +typedef struct load_args load_args; + +struct data; +typedef struct data data; + +struct metadata; +typedef struct metadata metadata; + +struct tree; +typedef struct tree tree; + +extern int gpu_index; + +// option_list.h +typedef struct metadata { + int classes; + char **names; +} metadata; + + +// tree.h +typedef struct tree { + int *leaf; + int n; + int *parent; + int *child; + int *group; + char **name; + + int groups; + int *group_size; + int *group_offset; +} tree; + + +// activations.h +typedef enum { + LOGISTIC, RELU, RELIE, LINEAR, RAMP, TANH, PLSE, LEAKY, ELU, LOGGY, STAIR, HARDTAN, LHTAN, SELU +}ACTIVATION; + +// image.h +typedef enum{ + PNG, BMP, TGA, JPG +} IMTYPE; + +// activations.h +typedef enum{ + MULT, ADD, SUB, DIV +} BINARY_ACTIVATION; + +// layer.h +typedef enum { + CONVOLUTIONAL, + DECONVOLUTIONAL, + CONNECTED, + MAXPOOL, + SOFTMAX, + DETECTION, + DROPOUT, + CROP, + ROUTE, + COST, + NORMALIZATION, + AVGPOOL, + LOCAL, + SHORTCUT, + ACTIVE, + RNN, + GRU, + LSTM, + CONV_LSTM, + CRNN, + BATCHNORM, + NETWORK, + XNOR, + REGION, + BAGGAGEAI, + ISEG, + REORG, + REORG_OLD, + UPSAMPLE, + LOGXENT, + L2NORM, + BLANK +} LAYER_TYPE; + +// layer.h +typedef enum{ + SSE, MASKED, L1, SEG, SMOOTH,WGAN +} COST_TYPE; + +// layer.h +typedef struct update_args { + int batch; + float learning_rate; + float momentum; + float decay; + int adam; + float B1; + float B2; + float eps; + int t; +} update_args; + +// layer.h +struct layer { + LAYER_TYPE type; + ACTIVATION activation; + COST_TYPE cost_type; + void(*forward) (struct layer, struct network_state); + void(*backward) (struct layer, struct network_state); + void(*update) (struct layer, int, float, float, float); + void(*forward_gpu) (struct layer, struct network_state); + void(*backward_gpu) (struct layer, struct network_state); + void(*update_gpu) (struct layer, int, float, float, float); + int batch_normalize; + int shortcut; + int batch; + int forced; + int flipped; + int inputs; + int outputs; + int nweights; + int nbiases; + int extra; + int truths; + int h, w, c; + int out_h, out_w, out_c; + int n; + int max_boxes; + int groups; + int size; + int side; + int stride; + int reverse; + int flatten; + int spatial; + int pad; + int sqrt; + int flip; + int index; + int binary; + int xnor; + int peephole; + int use_bin_output; + int steps; + int state_constrain; + int hidden; + int truth; + float smooth; + float dot; + float angle; + float jitter; + float saturation; + float exposure; + float shift; + float ratio; + float learning_rate_scale; + float clip; + int focal_loss; + int noloss; + int softmax; + int classes; + int coords; + int background; + int rescore; + int objectness; + int does_cost; + int joint; + int noadjust; + int reorg; + int log; + int tanh; + int *mask; + int total; + float bflops; + + int adam; + float B1; + float B2; + float eps; + + int t; + + float alpha; + float beta; + float kappa; + + float coord_scale; + float object_scale; + float noobject_scale; + float mask_scale; + float class_scale; + int bias_match; + int random; + float ignore_thresh; + float truth_thresh; + float thresh; + float focus; + int classfix; + int absolute; + + int onlyforward; + int stopbackward; + int dontload; + int dontsave; + int dontloadscales; + int numload; + + float temperature; + float probability; + float scale; + + char * cweights; + int * indexes; + int * input_layers; + int * input_sizes; + int * map; + int * counts; + float ** sums; + float * rand; + float * cost; + float * state; + float * prev_state; + float * forgot_state; + float * forgot_delta; + float * state_delta; + float * combine_cpu; + float * combine_delta_cpu; + + float *concat; + float *concat_delta; + + float *binary_weights; + + float *biases; + float *bias_updates; + + float *scales; + float *scale_updates; + + float *weights; + float *weight_updates; + + char *align_bit_weights_gpu; + float *mean_arr_gpu; + float *align_workspace_gpu; + float *transposed_align_workspace_gpu; + int align_workspace_size; + + char *align_bit_weights; + float *mean_arr; + int align_bit_weights_size; + int lda_align; + int new_lda; + int bit_align; + + float *col_image; + float * delta; + float * output; + int delta_pinned; + int output_pinned; + float * loss; + float * squared; + float * norms; + + float * spatial_mean; + float * mean; + float * variance; + + float * mean_delta; + float * variance_delta; + + float * rolling_mean; + float * rolling_variance; + + float * x; + float * x_norm; + + float * m; + float * v; + + float * bias_m; + float * bias_v; + float * scale_m; + float * scale_v; + + + float *z_cpu; + float *r_cpu; + float *h_cpu; + float *stored_h_cpu; + float * prev_state_cpu; + + float *temp_cpu; + float *temp2_cpu; + float *temp3_cpu; + + float *dh_cpu; + float *hh_cpu; + float *prev_cell_cpu; + float *cell_cpu; + float *f_cpu; + float *i_cpu; + float *g_cpu; + float *o_cpu; + float *c_cpu; + float *stored_c_cpu; + float *dc_cpu; + + float *binary_input; + uint32_t *bin_re_packed_input; + char *t_bit_input; + + struct layer *input_layer; + struct layer *self_layer; + struct layer *output_layer; + + struct layer *reset_layer; + struct layer *update_layer; + struct layer *state_layer; + + struct layer *input_gate_layer; + struct layer *state_gate_layer; + struct layer *input_save_layer; + struct layer *state_save_layer; + struct layer *input_state_layer; + struct layer *state_state_layer; + + struct layer *input_z_layer; + struct layer *state_z_layer; + + struct layer *input_r_layer; + struct layer *state_r_layer; + + struct layer *input_h_layer; + struct layer *state_h_layer; + + struct layer *wz; + struct layer *uz; + struct layer *wr; + struct layer *ur; + struct layer *wh; + struct layer *uh; + struct layer *uo; + struct layer *wo; + struct layer *vo; + struct layer *uf; + struct layer *wf; + struct layer *vf; + struct layer *ui; + struct layer *wi; + struct layer *vi; + struct layer *ug; + struct layer *wg; + + tree *softmax_tree; + + size_t workspace_size; + +#ifdef GPU + int *indexes_gpu; + + float *z_gpu; + float *r_gpu; + float *h_gpu; + float *stored_h_gpu; + + float *temp_gpu; + float *temp2_gpu; + float *temp3_gpu; + + float *dh_gpu; + float *hh_gpu; + float *prev_cell_gpu; + float *prev_state_gpu; + float *last_prev_state_gpu; + float *last_prev_cell_gpu; + float *cell_gpu; + float *f_gpu; + float *i_gpu; + float *g_gpu; + float *o_gpu; + float *c_gpu; + float *stored_c_gpu; + float *dc_gpu; + + // adam + float *m_gpu; + float *v_gpu; + float *bias_m_gpu; + float *scale_m_gpu; + float *bias_v_gpu; + float *scale_v_gpu; + + float * combine_gpu; + float * combine_delta_gpu; + + float * forgot_state_gpu; + float * forgot_delta_gpu; + float * state_gpu; + float * state_delta_gpu; + float * gate_gpu; + float * gate_delta_gpu; + float * save_gpu; + float * save_delta_gpu; + float * concat_gpu; + float * concat_delta_gpu; + + float *binary_input_gpu; + float *binary_weights_gpu; + float *bin_conv_shortcut_in_gpu; + float *bin_conv_shortcut_out_gpu; + + float * mean_gpu; + float * variance_gpu; + + float * rolling_mean_gpu; + float * rolling_variance_gpu; + + float * variance_delta_gpu; + float * mean_delta_gpu; + + float * col_image_gpu; + + float * x_gpu; + float * x_norm_gpu; + float * weights_gpu; + float * weight_updates_gpu; + float * weight_change_gpu; + + float * weights_gpu16; + float * weight_updates_gpu16; + + float * biases_gpu; + float * bias_updates_gpu; + float * bias_change_gpu; + + float * scales_gpu; + float * scale_updates_gpu; + float * scale_change_gpu; + + float * output_gpu; + float * loss_gpu; + float * delta_gpu; + float * rand_gpu; + float * squared_gpu; + float * norms_gpu; +#ifdef CUDNN + cudnnTensorDescriptor_t srcTensorDesc, dstTensorDesc; + cudnnTensorDescriptor_t srcTensorDesc16, dstTensorDesc16; + cudnnTensorDescriptor_t dsrcTensorDesc, ddstTensorDesc; + cudnnTensorDescriptor_t dsrcTensorDesc16, ddstTensorDesc16; + cudnnTensorDescriptor_t normTensorDesc, normDstTensorDesc, normDstTensorDescF16; + cudnnFilterDescriptor_t weightDesc, weightDesc16; + cudnnFilterDescriptor_t dweightDesc, dweightDesc16; + cudnnConvolutionDescriptor_t convDesc; + cudnnConvolutionFwdAlgo_t fw_algo, fw_algo16; + cudnnConvolutionBwdDataAlgo_t bd_algo, bd_algo16; + cudnnConvolutionBwdFilterAlgo_t bf_algo, bf_algo16; + cudnnPoolingDescriptor_t poolingDesc; +#endif // CUDNN +#endif // GPU +}; + + +// network.h +typedef enum { + CONSTANT, STEP, EXP, POLY, STEPS, SIG, RANDOM, SGDR +} learning_rate_policy; + +// network.h +typedef struct network { + int n; + int batch; + uint64_t *seen; + int *t; + float epoch; + int subdivisions; + layer *layers; + float *output; + learning_rate_policy policy; + + float learning_rate; + float learning_rate_min; + float learning_rate_max; + int batches_per_cycle; + int batches_cycle_mult; + float momentum; + float decay; + float gamma; + float scale; + float power; + int time_steps; + int step; + int max_batches; + float *seq_scales; + float *scales; + int *steps; + int num_steps; + int burn_in; + int cudnn_half; + float *pre_allocated_ptr; + int adam; + float B1; + float B2; + float eps; + + int inputs; + int outputs; + int truths; + int notruth; + int h, w, c; + int max_crop; + int min_crop; + float max_ratio; + float min_ratio; + int center; + int flip; // horizontal flip 50% probability augmentaiont for classifier training (default = 1) + int blur; + float angle; + float aspect; + float exposure; + float saturation; + float hue; + int random; + int track; + int augment_speed; + int sequential_subdivisions; + int init_sequential_subdivisions; + int current_subdivision; + int try_fix_nan; + + int gpu_index; + tree *hierarchy; + + float *input; + float *truth; + float *delta; + float *workspace; + int train; + int index; + float *cost; + float clip; + +#ifdef GPU + //float *input_gpu; + //float *truth_gpu; + float *delta_gpu; + float *output_gpu; + + float *input_state_gpu; + float *input_pinned_cpu; + int input_pinned_cpu_flag; + + float **input_gpu; + float **truth_gpu; + float **input16_gpu; + float **output16_gpu; + size_t *max_input16_size; + size_t *max_output16_size; + int wait_stream; +#endif +} network; + +// network.h +typedef struct network_state { + float *truth; + float *input; + float *delta; + float *workspace; + int train; + int index; + network net; +} network_state; + +//typedef struct { +// int w; +// int h; +// float scale; +// float rad; +// float dx; +// float dy; +// float aspect; +//} augment_args; + +// image.h +typedef struct image { + int w; + int h; + int c; + float *data; +} image; + +//typedef struct { +// int w; +// int h; +// int c; +// float *data; +//} image; + +// box.h +typedef struct box { + float x, y, w, h; +} box; + +// box.h +typedef struct detection{ + box bbox; + int classes; + float *prob; + float *mask; + float objectness; + int sort_class; +} detection; + +// matrix.h +typedef struct matrix { + int rows, cols; + float **vals; +} matrix; + +// data.h +typedef struct data { + int w, h; + matrix X; + matrix y; + int shallow; + int *num_boxes; + box **boxes; +} data; + +// data.h +typedef enum { + CLASSIFICATION_DATA, DETECTION_DATA, CAPTCHA_DATA, REGION_DATA, IMAGE_DATA, COMPARE_DATA, WRITING_DATA, SWAG_DATA, TAG_DATA, OLD_CLASSIFICATION_DATA, STUDY_DATA, DET_DATA, SUPER_DATA, LETTERBOX_DATA, REGRESSION_DATA, SEGMENTATION_DATA, INSTANCE_DATA, ISEG_DATA +} data_type; + +// data.h +typedef struct load_args { + int threads; + char **paths; + char *path; + int n; + int m; + char **labels; + int h; + int w; + int c; // color depth + int out_w; + int out_h; + int nh; + int nw; + int num_boxes; + int min, max, size; + int classes; + int background; + int scale; + int center; + int coords; + int mini_batch; + int track; + int augment_speed; + int show_imgs; + float jitter; + int flip; + int blur; + float angle; + float aspect; + float saturation; + float exposure; + float hue; + data *d; + image *im; + image *resized; + data_type type; + tree *hierarchy; +} load_args; + +// data.h +typedef struct box_label { + int id; + float x, y, w, h; + float left, right, top, bottom; +} box_label; + +// list.h +//typedef struct node { +// void *val; +// struct node *next; +// struct node *prev; +//} node; + +// list.h +//typedef struct list { +// int size; +// node *front; +// node *back; +//} list; + +// ----------------------------------------------------- + + +// parser.c +LIB_API network *load_network(char *cfg, char *weights, int clear); +LIB_API network *load_network_custom(char *cfg, char *weights, int clear, int batch); +LIB_API network *load_network(char *cfg, char *weights, int clear); + +// network.c +LIB_API load_args get_base_args(network *net); + +// box.h +LIB_API void do_nms_sort(detection *dets, int total, int classes, float thresh); +LIB_API void do_nms_obj(detection *dets, int total, int classes, float thresh); + +// network.h +LIB_API float *network_predict(network net, float *input); +LIB_API float *network_predict_ptr(network *net, float *input); +LIB_API detection *get_network_boxes(network *net, int w, int h, float thresh, float hier, int *map, int relative, int *num, int letter); +LIB_API void free_detections(detection *dets, int n); +LIB_API void fuse_conv_batchnorm(network net); +LIB_API void calculate_binary_weights(network net); +LIB_API char *detection_to_json(detection *dets, int nboxes, int classes, char **names, long long int frame_id, char *filename); + +LIB_API layer* get_network_layer(network* net, int i); +//LIB_API detection *get_network_boxes(network *net, int w, int h, float thresh, float hier, int *map, int relative, int *num, int letter); +LIB_API detection *make_network_boxes(network *net, float thresh, int *num); +LIB_API void reset_rnn(network *net); +LIB_API float *network_predict_image(network *net, image im); +LIB_API float validate_detector_map(char *datacfg, char *cfgfile, char *weightfile, float thresh_calc_avg_iou, const float iou_thresh, const int map_points, network *existing_net); +LIB_API void train_detector(char *datacfg, char *cfgfile, char *weightfile, int *gpus, int ngpus, int clear, int dont_show, int calc_map, int mjpeg_port, int show_imgs); +LIB_API void test_detector(char *datacfg, char *cfgfile, char *weightfile, char *filename, float thresh, + float hier_thresh, int dont_show, int ext_output, int save_labels, char *outfile, int letter_box); +LIB_API int network_width(network *net); +LIB_API int network_height(network *net); +LIB_API void optimize_picture(network *net, image orig, int max_layer, float scale, float rate, float thresh, int norm); + +// image.h +LIB_API image resize_image(image im, int w, int h); +LIB_API void copy_image_from_bytes(image im, char *pdata); +LIB_API image letterbox_image(image im, int w, int h); +LIB_API void rgbgr_image(image im); +LIB_API image make_image(int w, int h, int c); +LIB_API image load_image_color(char *filename, int w, int h); +LIB_API void free_image(image m); + +// layer.h +LIB_API void free_layer(layer); + +// data.c +LIB_API void free_data(data d); +LIB_API pthread_t load_data(load_args args); +LIB_API pthread_t load_data_in_thread(load_args args); + +// dark_cuda.h +LIB_API void cuda_pull_array(float *x_gpu, float *x, size_t n); +LIB_API void cuda_pull_array_async(float *x_gpu, float *x, size_t n); +LIB_API void cuda_set_device(int n); +LIB_API void *cuda_get_context(); + +// utils.h +LIB_API void free_ptrs(void **ptrs, int n); +LIB_API void top_k(float *a, int n, int k, int *index); + +// tree.h +LIB_API tree *read_tree(char *filename); + +// option_list.h +LIB_API metadata get_metadata(char *file); + + +// http_stream.h +LIB_API void delete_json_sender(); +LIB_API void send_json_custom(char const* send_buf, int port, int timeout); +LIB_API double get_time_point(); +void start_timer(); +void stop_timer(); +double get_time(); +void stop_timer_and_show(); +void stop_timer_and_show_name(char *name); +void show_total_time(); + +#ifdef __cplusplus +} +#endif // __cplusplus +#endif // DIMENSIONLESS_API diff --git a/include/tkDNN/evaluation.h b/include/tkDNN/evaluation.h new file mode 100644 index 0000000..1614d6b --- /dev/null +++ b/include/tkDNN/evaluation.h @@ -0,0 +1,120 @@ +#ifndef EVALUATION_H +#define EVALUATION_H + +#include +#include +#include + +#include + +#include "tkdnn.h" +#include "BoundingBox.h" + +namespace tk { namespace dnn { + +struct Frame +{ + std::string lFilename; + std::string iFilename; + std::vector gt; + std::vector det; + + void print() const; +}; + +struct PR +{ + double precision = 0; + double recall = 0; + int tp = 0, fp = 0, fn = 0; + + void print(); +}; + +void readmAPParams( const char* config_filename, int& classes1,float& conf_thresh1 + , int& classes2,float& conf_thresh2 + , int& classes3,float& conf_thresh3 + , int& classes4,float& conf_thresh4 + , int& classes5,float& conf_thresh5 + ); + +/** + * This method computes the mean Average Precision for a set of detections and + * groundtruths. It returns the mAP for a given IoU threshold, and a given + * confidence threshold over all the classes. + * + * @param images collection of frames on which to compute the metrics + * @param classes number of classes of the considered dataset + * @param IoU_thresh threshold used to compute Intersection over Union + * @param conf_thresh threshold used to filter bounding boxes based on their + * confidence (or probability) + * @param map_points number of point used to compute the mAP. if 0 is given, + * all the recall levels are evaluated, otherwise only + * map_point recall levels are used. For COCO evaluation + * 101 points are used. + * @param verbose is set to true, prints on screen additional info + * + * @return mAP computed + */ +double computeMap( std::vector &images,const int classes, + const float IoU_thresh, const float conf_thresh=0.3, + const int map_points=101, const bool verbose=false); + + +/** + * This method computes the mean Average Precision for a set of detections and + * groundtruths on several IoU thresholds. It is used to compute, for example, + * the most used metric in Object Detection, namely the mAP 0.5:0.95, which is + * the average among the mAP for IoU level from 0.5 to 0.95 with a step of 0.05. + * + * @param images collection of frames on which to compute the metrics + * @param classes number of classes of the considered dataset + * @param IoU_thresh starting threshold used to compute Intersection over Union + * @param conf_thresh threshold used to filter bounding boxes based on their + * confidence (or probability) + * @param map_points number of point used to compute the mAP. if 0 is given, + * all the recall levels are evaluated, otherwise only + * map_point recall levels are used. For COCO evaluation + * 101 points are used. + * @param map_step step used to increment IoU threshold + * @param map_levels number of IoU step to perform + * @param verbose is set to true, prints on screen additional info + * @param write_on_file if set to true, the results produced by this function + * are written on file + * @param net name of the considered neural network + * + * @return mAP IoU_tresh:IoU_tresh+map_step*map_levels (e.g. mAP 0.5:0.95 when + * map_step=0.05 and map_levels=10) + */ +double computeMapNIoULevels(std::vector &images,const int classes, + const float i_IoU_thresh=0.5, const float conf_thresh=0.3, + const int map_points=101, const float map_step=0.05, + const int map_levels=10, const bool verbose=false, + const bool write_on_file = false, std::string net = ""); +/** + * This method computes the number of True Positive (TP), False Positive (FP), + * False Negative (FN), precision, recall and f1-score. + * Those values are computer over all the detections, over all the classes. + * + * @param images collection of frames on which to compute the metrics + * @param classes number of classes of the considered dataset + * @param IoU_thresh threshold used to compute Intersection over Union + * @param conf_thresh threshold used to filter bounding boxes based on their + * confidence (or probability) + * @param verbose is set to true, prints on screen additional info + * @param write_on_file if set to true, the results produced by this function + * are written on file + * @param net name of the considered neural network + */ +void computeTPFPFN( std::vector &images,const int classes, + const float IoU_thresh=0.5, const float conf_thresh=0.3, + bool verbose=false, const bool write_on_file=false, + std::string net=""); + + +void printJsonCOCOFormat(std::ofstream *out_file, const std::string image_path, std::vector bbox, const int classes, const int w, const int h); + +}} +#endif /*EVALUATION_H*/ + + diff --git a/include/tkDNN/handler.h b/include/tkDNN/handler.h new file mode 100644 index 0000000..df91cd4 --- /dev/null +++ b/include/tkDNN/handler.h @@ -0,0 +1,28 @@ +#ifndef HANDLER_H +#define HANDLER_H +#include +#include "stdafx.h" + +using namespace std; +using namespace web; +using namespace http; +using namespace utility; +using namespace http::experimental::listener; + + +class handler +{ + public: + handler(utility::string_t url); + + pplx::taskopen(){return m_listener.open();} + pplx::taskclose(){return m_listener.close();} + static void init_bag(); + protected: + + private: + void handle_post(http_request message); + http_listener m_listener; +}; + +#endif // HANDLER_H diff --git a/include/tkDNN/image.h b/include/tkDNN/image.h new file mode 100644 index 0000000..2cd9c3a --- /dev/null +++ b/include/tkDNN/image.h @@ -0,0 +1,108 @@ +#ifndef IMAGE_H +#define IMAGE_H +#include "darknet.h" + +#include +#include +#include +#include +#include + +//#include "image_opencv.h" + +//#include "box.h" +#ifdef __cplusplus +extern "C" { +#endif +/* +typedef struct { + int w; + int h; + int c; + float *data; +} image; +*/ +float get_color(int c, int x, int max); +void flip_image(image a); +void draw_box(image a, int x1, int y1, int x2, int y2, float r, float g, float b); +void draw_box_width(image a, int x1, int y1, int x2, int y2, int w, float r, float g, float b); +void draw_bbox(image a, box bbox, int w, float r, float g, float b); +void draw_label(image a, int r, int c, image label, const float *rgb); +void write_label(image a, int r, int c, image *characters, char *string, float *rgb); +void draw_detections(image im, int num, float thresh, box *boxes, float **probs, char **names, image **labels, int classes); +void draw_detections_v3(image im, detection *dets, int num, float thresh, char **names, image **alphabet, int classes, int ext_output); +image image_distance(image a, image b); +void scale_image(image m, float s); +// image crop_image(image im, int dx, int dy, int w, int h); +image random_crop_image(image im, int w, int h); +image random_augment_image(image im, float angle, float aspect, int low, int high, int size); +void random_distort_image(image im, float hue, float saturation, float exposure); +//LIB_API image resize_image(image im, int w, int h); +//LIB_API void copy_image_from_bytes(image im, char *pdata); +void fill_image(image m, float s); +void letterbox_image_into(image im, int w, int h, image boxed); +//LIB_API image letterbox_image(image im, int w, int h); +// image resize_min(image im, int min); +image resize_max(image im, int max); +void translate_image(image m, float s); +void normalize_image(image p); +image rotate_image(image m, float rad); +void rotate_image_cw(image im, int times); +void embed_image(image source, image dest, int dx, int dy); +void saturate_image(image im, float sat); +void exposure_image(image im, float sat); +void distort_image(image im, float hue, float sat, float val); +void saturate_exposure_image(image im, float sat, float exposure); +void hsv_to_rgb(image im); +//LIB_API void rgbgr_image(image im); +void constrain_image(image im); +void composite_3d(char *f1, char *f2, char *out, int delta); +int best_3d_shift_r(image a, image b, int min, int max); + +image grayscale_image(image im); +image threshold_image(image im, float thresh); + +image collapse_image_layers(image source, int border); +image collapse_images_horz(image *ims, int n); +image collapse_images_vert(image *ims, int n); + +void show_image(image p, const char *name); +void show_image_normalized(image im, const char *name); +void save_image_png(image im, const char *name); +void save_image(image p, const char *name); +void show_images(image *ims, int n, char *window); +void show_image_layers(image p, char *name); +void show_image_collapsed(image p, char *name); + +void print_image(image m); + +//LIB_API image make_image(int w, int h, int c); +image make_random_image(int w, int h, int c); +image make_empty_image(int w, int h, int c); +image float_to_image_scaled(int w, int h, int c, float *data); +image float_to_image(int w, int h, int c, float *data); +image copy_image(image p); +void copy_image_inplace(image src, image dst); +image load_image(char *filename, int w, int h, int c); +image load_image_stb_resize(char *filename, int w, int h, int c); +image load_image_new(unsigned char *image_data, int len, int channels, int antiLog, int gray, int width, int height); +image load_image_file(unsigned char *image_data, int channels, int antilog, int gray, int width, int height); +//LIB_API image load_image_color(char *filename, int w, int h); +image **load_alphabet(); + +//float get_pixel(image m, int x, int y, int c); +//float get_pixel_extend(image m, int x, int y, int c); +//void set_pixel(image m, int x, int y, int c, float val); +//void add_pixel(image m, int x, int y, int c, float val); +float bilinear_interpolate(image im, float x, float y, int c); + +image get_image_layer(image m, int l); + +//LIB_API void free_image(image m); +void test_resize(char *filename); +#ifdef __cplusplus +} +#endif + +#endif + diff --git a/include/tkDNN/kernels.h b/include/tkDNN/kernels.h index d7d5d05..5d673c8 100644 --- a/include/tkDNN/kernels.h +++ b/include/tkDNN/kernels.h @@ -3,42 +3,38 @@ #include "utils.h" -void activationELUForward(dnnType* srcData, dnnType* dstData, int size, cudaStream_t stream = cudaStream_t(0)); -void activationLEAKYForward(dnnType* srcData, dnnType* dstData, int size, cudaStream_t stream = cudaStream_t(0)); -void activationLOGISTICForward(dnnType* srcData, dnnType* dstData, int size, cudaStream_t stream = cudaStream_t(0)); -void activationSIGMOIDForward(dnnType* srcData, dnnType* dstData, int size, cudaStream_t stream = cudaStream_t(0)); +void activationELUForward(dnnType *srcData, dnnType *dstData, int size, cudaStream_t stream = cudaStream_t(0)); +void activationLEAKYForward(dnnType *srcData, dnnType *dstData, int size, cudaStream_t stream = cudaStream_t(0)); +void activationReLUCeilingForward(dnnType *srcData, dnnType *dstData, int size, const float ceiling, cudaStream_t stream = cudaStream_t(0)); +void activationLOGISTICForward(dnnType *srcData, dnnType *dstData, int size, cudaStream_t stream = cudaStream_t(0)); +void activationSIGMOIDForward(dnnType *srcData, dnnType *dstData, int size, cudaStream_t stream = cudaStream_t(0)); +void activationMishForward(dnnType* srcData, dnnType* dstData, int size, cudaStream_t stream= cudaStream_t(0)); -void fill(dnnType* data, int size, dnnType val, cudaStream_t stream = cudaStream_t(0)); +void fill(dnnType *data, int size, dnnType val, cudaStream_t stream = cudaStream_t(0)); -void resizeForward( dnnType* srcData, dnnType* dstData, int n, int i_c, int i_h, int i_w, - int o_c, int o_h, int o_w, cudaStream_t stream = cudaStream_t(0)); +void resizeForward(dnnType *srcData, dnnType *dstData, int n, int i_c, int i_h, int i_w, + int o_c, int o_h, int o_w, cudaStream_t stream = cudaStream_t(0)); -void reorgForward( dnnType* srcData, dnnType* dstData, - int n, int c, int h, int w, int stride, cudaStream_t stream = cudaStream_t(0)); -void softmaxForward(float *input, int n, int batch, int batch_offset, +void reorgForward(dnnType *srcData, dnnType *dstData, + int n, int c, int h, int w, int stride, cudaStream_t stream = cudaStream_t(0)); + +void MaxPoolingForward(dnnType *srcData, dnnType *dstData, int n, int c, int h, int w, int stride_x, int stride_y, int size, int padding, cudaStream_t stream = cudaStream_t(0)); + +void softmaxForward(float *input, int n, int batch, int batch_offset, int groups, int group_offset, int stride, float temp, float *output, cudaStream_t stream = cudaStream_t(0)); - -void shortcutForward(dnnType* srcData, dnnType* dstData, int n1, int c1, int h1, int w1, int s1, - int n2, int c2, int h2, int w2, int s2, +void shortcutForward(dnnType *srcData, dnnType *dstData, int n1, int c1, int h1, int w1, int s1, + int n2, int c2, int h2, int w2, int s2, cudaStream_t stream = cudaStream_t(0)); -void upsampleForward(dnnType* srcData, dnnType* dstData, - int n, int c, int h, int w, int s, int forward, float scale, +void upsampleForward(dnnType *srcData, dnnType *dstData, + int n, int c, int h, int w, int s, int forward, float scale, cudaStream_t stream = cudaStream_t(0)); -void float2half(float* srcData, __half* dstData, int size, const cudaStream_t stream = cudaStream_t(0)); +void float2half(float *srcData, __half *dstData, int size, const cudaStream_t stream = cudaStream_t(0)); - -void modulated_deformable_im2col_cuda(cudaStream_t stream, - const float *data_im, const float *data_offset, const float *data_mask, - const int batch_size, const int channels, const int height_im, const int width_im, - const int height_col, const int width_col, const int kernel_h, const int kenerl_w, - const int pad_h, const int pad_w, const int stride_h, const int stride_w, - const int dilation_h, const int dilation_w, - const int deformable_group, float *data_col); - -void dcn_v2_cuda_forward(float *input, float *weight, +void dcnV2CudaForward(cublasStatus_t stat, cublasHandle_t handle, + float *input, float *weight, float *bias, float *ones, float *offset, float *mask, float *output, float *columns, @@ -46,8 +42,10 @@ void dcn_v2_cuda_forward(float *input, float *weight, const int stride_h, const int stride_w, const int pad_h, const int pad_w, const int dilation_h, const int dilation_w, - const int deformable_group, + const int deformable_group, const int batch_id, const int in_n, const int in_c, const int in_h, const int in_w, const int out_n, const int out_c, const int out_h, const int out_w, const int dst_dim, cudaStream_t stream = cudaStream_t(0)); + +void scalAdd(dnnType* dstData, int size, float alpha, float beta, int inc, cudaStream_t stream = cudaStream_t(0)); #endif //KERNELS_H diff --git a/include/tkDNN/kernelsThrust.h b/include/tkDNN/kernelsThrust.h new file mode 100644 index 0000000..a7b32e9 --- /dev/null +++ b/include/tkDNN/kernelsThrust.h @@ -0,0 +1,39 @@ +#ifndef KERNELSTHRUST_H +#define KERNELSTHRUST_H + + +#include +#include +#include +#include +#include +#include +#include + +#include "tkdnn.h" + +struct threshold : public thrust::binary_function +{ + __host__ __device__ + float operator()(float x, float y) { + double toll = 1e-6; + if(fabsf(x-y)>toll) + return 0.0f; + else + return x; + } +}; + +void sort(dnnType *src_begin, dnnType *src_end, int *idsrc); +void topk(dnnType *src_begin, int *idsrc, int K, float *topk_scores, + int *topk_inds, float *topk_ys, float *topk_xs); +// void sortAndTopKonDevice(dnnType *src_begin, int *idsrc, float *topk_scores, int *topk_inds, float *topk_ys, float *topk_xs, const int size, const int K, const int n_classes); +void normalize(float *bgr, const int ch, const int h, const int w, const float *mean, const float *stddev); +void subtractWithThreshold(dnnType *src_begin, dnnType *src_end, dnnType *src2_begin, dnnType *src_out, struct threshold op); +void topKxyclasses(int *ids_begin, int *ids_end, const int K, const int size, const int wh, int *clses, int *xs, int *ys); +void topKxyAddOffset(int * ids_begin, const int K, const int size, int *intxs_begin, int *intys_begin, + float *xs_begin, float *ys_begin, dnnType *src_begin, float *src_out, int *ids_out); +void bboxes(int * ids_begin, const int K, const int size, float *xs_begin, float *ys_begin, + dnnType *src_begin, float *bbx0, float *bbx1, float *bby0, float *bby1, float *src_out, int *ids_out); + +#endif //KERNELSTHRUST_H \ No newline at end of file diff --git a/include/tkDNN/models/Yolo3.h b/include/tkDNN/models/Yolo3.h deleted file mode 100644 index cd69b32..0000000 --- a/include/tkDNN/models/Yolo3.h +++ /dev/null @@ -1,289 +0,0 @@ -int preYoloFilters = (classes+5)*3; - -std::string input_bin = bin_path + "/layers/input.bin"; -std::vector output_bins = { - bin_path + "/debug/layer82_out.bin", - bin_path + "/debug/layer94_out.bin", - bin_path + "/debug/layer106_out.bin" -}; -std::string c0_bin = bin_path + "/layers/c0.bin"; -std::string c1_bin = bin_path + "/layers/c1.bin"; -std::string c2_bin = bin_path + "/layers/c2.bin"; -std::string c3_bin = bin_path + "/layers/c3.bin"; -std::string c5_bin = bin_path + "/layers/c5.bin"; -std::string c6_bin = bin_path + "/layers/c6.bin"; -std::string c7_bin = bin_path + "/layers/c7.bin"; -std::string c9_bin = bin_path + "/layers/c9.bin"; -std::string c10_bin = bin_path + "/layers/c10.bin"; -std::string c12_bin = bin_path + "/layers/c12.bin"; -std::string c13_bin = bin_path + "/layers/c13.bin"; -std::string c14_bin = bin_path + "/layers/c14.bin"; -std::string c16_bin = bin_path + "/layers/c16.bin"; -std::string c17_bin = bin_path + "/layers/c17.bin"; -std::string c19_bin = bin_path + "/layers/c19.bin"; -std::string c20_bin = bin_path + "/layers/c20.bin"; -std::string c22_bin = bin_path + "/layers/c22.bin"; -std::string c23_bin = bin_path + "/layers/c23.bin"; -std::string c25_bin = bin_path + "/layers/c25.bin"; -std::string c26_bin = bin_path + "/layers/c26.bin"; -std::string c28_bin = bin_path + "/layers/c28.bin"; -std::string c29_bin = bin_path + "/layers/c29.bin"; -std::string c31_bin = bin_path + "/layers/c31.bin"; -std::string c32_bin = bin_path + "/layers/c32.bin"; -std::string c34_bin = bin_path + "/layers/c34.bin"; -std::string c35_bin = bin_path + "/layers/c35.bin"; -std::string c37_bin = bin_path + "/layers/c37.bin"; -std::string c38_bin = bin_path + "/layers/c38.bin"; -std::string c39_bin = bin_path + "/layers/c39.bin"; -std::string c41_bin = bin_path + "/layers/c41.bin"; -std::string c42_bin = bin_path + "/layers/c42.bin"; -std::string c44_bin = bin_path + "/layers/c44.bin"; -std::string c45_bin = bin_path + "/layers/c45.bin"; -std::string c47_bin = bin_path + "/layers/c47.bin"; -std::string c48_bin = bin_path + "/layers/c48.bin"; -std::string c50_bin = bin_path + "/layers/c50.bin"; -std::string c51_bin = bin_path + "/layers/c51.bin"; -std::string c53_bin = bin_path + "/layers/c53.bin"; -std::string c54_bin = bin_path + "/layers/c54.bin"; -std::string c56_bin = bin_path + "/layers/c56.bin"; -std::string c57_bin = bin_path + "/layers/c57.bin"; -std::string c59_bin = bin_path + "/layers/c59.bin"; -std::string c60_bin = bin_path + "/layers/c60.bin"; -std::string c62_bin = bin_path + "/layers/c62.bin"; -std::string c63_bin = bin_path + "/layers/c63.bin"; -std::string c64_bin = bin_path + "/layers/c64.bin"; -std::string c66_bin = bin_path + "/layers/c66.bin"; -std::string c67_bin = bin_path + "/layers/c67.bin"; -std::string c69_bin = bin_path + "/layers/c69.bin"; -std::string c70_bin = bin_path + "/layers/c70.bin"; -std::string c72_bin = bin_path + "/layers/c72.bin"; -std::string c73_bin = bin_path + "/layers/c73.bin"; -std::string c75_bin = bin_path + "/layers/c75.bin"; -std::string c76_bin = bin_path + "/layers/c76.bin"; -std::string c77_bin = bin_path + "/layers/c77.bin"; -std::string c78_bin = bin_path + "/layers/c78.bin"; -std::string c79_bin = bin_path + "/layers/c79.bin"; -std::string c80_bin = bin_path + "/layers/c80.bin"; -std::string c81_bin = bin_path + "/layers/c81.bin"; -std::string g82_bin = bin_path + "/layers/g82.bin"; -std::string c84_bin = bin_path + "/layers/c84.bin"; -std::string c87_bin = bin_path + "/layers/c87.bin"; -std::string c88_bin = bin_path + "/layers/c88.bin"; -std::string c89_bin = bin_path + "/layers/c89.bin"; -std::string c90_bin = bin_path + "/layers/c90.bin"; -std::string c91_bin = bin_path + "/layers/c91.bin"; -std::string c92_bin = bin_path + "/layers/c92.bin"; -std::string c93_bin = bin_path + "/layers/c93.bin"; -std::string g94_bin = bin_path + "/layers/g94.bin"; -std::string c96_bin = bin_path + "/layers/c96.bin"; -std::string c99_bin = bin_path + "/layers/c99.bin"; -std::string c100_bin = bin_path + "/layers/c100.bin"; -std::string c101_bin = bin_path + "/layers/c101.bin"; -std::string c102_bin = bin_path + "/layers/c102.bin"; -std::string c103_bin = bin_path + "/layers/c103.bin"; -std::string c104_bin = bin_path + "/layers/c104.bin"; -std::string c105_bin = bin_path + "/layers/c105.bin"; -std::string g106_bin = bin_path + "/layers/g106.bin"; - -tk::dnn::Conv2d c0 (&net, 32, 3, 3, 1, 1, 1, 1, c0_bin, true); -tk::dnn::Activation a0 (&net, tk::dnn::ACTIVATION_LEAKY); -tk::dnn::Conv2d c1 (&net, 64, 3, 3, 2, 2, 1, 1, c1_bin, true); -tk::dnn::Activation a1 (&net, tk::dnn::ACTIVATION_LEAKY); -tk::dnn::Conv2d c2 (&net, 32, 1, 1, 1, 1, 0, 0, c2_bin, true); -tk::dnn::Activation a2 (&net, tk::dnn::ACTIVATION_LEAKY); -tk::dnn::Conv2d c3 (&net, 64, 3, 3, 1, 1, 1, 1, c3_bin, true); -tk::dnn::Activation a3 (&net, tk::dnn::ACTIVATION_LEAKY); -tk::dnn::Shortcut s4 (&net, &a1); -tk::dnn::Conv2d c5 (&net, 128, 3, 3, 2, 2, 1, 1, c5_bin, true); -tk::dnn::Activation a5 (&net, tk::dnn::ACTIVATION_LEAKY); -tk::dnn::Conv2d c6 (&net, 64, 1, 1, 1, 1, 0, 0, c6_bin, true); -tk::dnn::Activation a6 (&net, tk::dnn::ACTIVATION_LEAKY); -tk::dnn::Conv2d c7 (&net, 128, 3, 3, 1, 1, 1, 1, c7_bin, true); -tk::dnn::Activation a7 (&net, tk::dnn::ACTIVATION_LEAKY); -tk::dnn::Shortcut s8 (&net, &a5); -tk::dnn::Conv2d c9 (&net, 64, 1, 1, 1, 1, 0, 0, c9_bin, true); -tk::dnn::Activation a9 (&net, tk::dnn::ACTIVATION_LEAKY); -tk::dnn::Conv2d c10 (&net, 128, 3, 3, 1, 1, 1, 1, c10_bin, true); -tk::dnn::Activation a10 (&net, tk::dnn::ACTIVATION_LEAKY); -tk::dnn::Shortcut s11 (&net, &s8); - -tk::dnn::Conv2d c12 (&net, 256, 3, 3, 2, 2, 1, 1, c12_bin, true); -tk::dnn::Activation a12 (&net, tk::dnn::ACTIVATION_LEAKY); -tk::dnn::Conv2d c13 (&net, 128, 1, 1, 1, 1, 0, 0, c13_bin, true); -tk::dnn::Activation a13 (&net, tk::dnn::ACTIVATION_LEAKY); -tk::dnn::Conv2d c14 (&net, 256, 3, 3, 1, 1, 1, 1, c14_bin, true); -tk::dnn::Activation a14 (&net, tk::dnn::ACTIVATION_LEAKY); -tk::dnn::Shortcut s15 (&net, &a12); - -tk::dnn::Conv2d c16 (&net, 128, 1, 1, 1, 1, 0, 0, c16_bin, true); -tk::dnn::Activation a16 (&net, tk::dnn::ACTIVATION_LEAKY); -tk::dnn::Conv2d c17 (&net, 256, 3, 3, 1, 1, 1, 1, c17_bin, true); -tk::dnn::Activation a17 (&net, tk::dnn::ACTIVATION_LEAKY); -tk::dnn::Shortcut s18 (&net, &s15); -tk::dnn::Conv2d c19 (&net, 128, 1, 1, 1, 1, 0, 0, c19_bin, true); -tk::dnn::Activation a19 (&net, tk::dnn::ACTIVATION_LEAKY); -tk::dnn::Conv2d c20 (&net, 256, 3, 3, 1, 1, 1, 1, c20_bin, true); -tk::dnn::Activation a20 (&net, tk::dnn::ACTIVATION_LEAKY); -tk::dnn::Shortcut s21 (&net, &s18); -tk::dnn::Conv2d c22 (&net, 128, 1, 1, 1, 1, 0, 0, c22_bin, true); -tk::dnn::Activation a22 (&net, tk::dnn::ACTIVATION_LEAKY); -tk::dnn::Conv2d c23 (&net, 256, 3, 3, 1, 1, 1, 1, c23_bin, true); -tk::dnn::Activation a23 (&net, tk::dnn::ACTIVATION_LEAKY); -tk::dnn::Shortcut s24 (&net, &s21); -tk::dnn::Conv2d c25 (&net, 128, 1, 1, 1, 1, 0, 0, c25_bin, true); -tk::dnn::Activation a25 (&net, tk::dnn::ACTIVATION_LEAKY); -tk::dnn::Conv2d c26 (&net, 256, 3, 3, 1, 1, 1, 1, c26_bin, true); -tk::dnn::Activation a26 (&net, tk::dnn::ACTIVATION_LEAKY); -tk::dnn::Shortcut s27 (&net, &s24); -tk::dnn::Conv2d c28 (&net, 128, 1, 1, 1, 1, 0, 0, c28_bin, true); -tk::dnn::Activation a28 (&net, tk::dnn::ACTIVATION_LEAKY); -tk::dnn::Conv2d c29 (&net, 256, 3, 3, 1, 1, 1, 1, c29_bin, true); -tk::dnn::Activation a29 (&net, tk::dnn::ACTIVATION_LEAKY); -tk::dnn::Shortcut s30 (&net, &s27); -tk::dnn::Conv2d c31 (&net, 128, 1, 1, 1, 1, 0, 0, c31_bin, true); -tk::dnn::Activation a31 (&net, tk::dnn::ACTIVATION_LEAKY); -tk::dnn::Conv2d c32 (&net, 256, 3, 3, 1, 1, 1, 1, c32_bin, true); -tk::dnn::Activation a32 (&net, tk::dnn::ACTIVATION_LEAKY); -tk::dnn::Shortcut s33 (&net, &s30); -tk::dnn::Conv2d c34 (&net, 128, 1, 1, 1, 1, 0, 0, c34_bin, true); -tk::dnn::Activation a34 (&net, tk::dnn::ACTIVATION_LEAKY); -tk::dnn::Conv2d c35 (&net, 256, 3, 3, 1, 1, 1, 1, c35_bin, true); -tk::dnn::Activation a35 (&net, tk::dnn::ACTIVATION_LEAKY); -tk::dnn::Shortcut s36 (&net, &s33); - -tk::dnn::Conv2d c37 (&net, 512, 3, 3, 2, 2, 1, 1, c37_bin, true); -tk::dnn::Activation a37 (&net, tk::dnn::ACTIVATION_LEAKY); -tk::dnn::Conv2d c38 (&net, 256, 1, 1, 1, 1, 0, 0, c38_bin, true); -tk::dnn::Activation a38 (&net, tk::dnn::ACTIVATION_LEAKY); -tk::dnn::Conv2d c39 (&net, 512, 3, 3, 1, 1, 1, 1, c39_bin, true); -tk::dnn::Activation a39 (&net, tk::dnn::ACTIVATION_LEAKY); -tk::dnn::Shortcut s40 (&net, &a37); - -tk::dnn::Conv2d c41 (&net, 256, 1, 1, 1, 1, 0, 0, c41_bin, true); -tk::dnn::Activation a41 (&net, tk::dnn::ACTIVATION_LEAKY); -tk::dnn::Conv2d c42 (&net, 512, 3, 3, 1, 1, 1, 1, c42_bin, true); -tk::dnn::Activation a42 (&net, tk::dnn::ACTIVATION_LEAKY); -tk::dnn::Shortcut s43 (&net, &s40); -tk::dnn::Conv2d c44 (&net, 256, 1, 1, 1, 1, 0, 0, c44_bin, true); -tk::dnn::Activation a44 (&net, tk::dnn::ACTIVATION_LEAKY); -tk::dnn::Conv2d c45 (&net, 512, 3, 3, 1, 1, 1, 1, c45_bin, true); -tk::dnn::Activation a45 (&net, tk::dnn::ACTIVATION_LEAKY); -tk::dnn::Shortcut s46 (&net, &s43); -tk::dnn::Conv2d c47 (&net, 256, 1, 1, 1, 1, 0, 0, c47_bin, true); -tk::dnn::Activation a47 (&net, tk::dnn::ACTIVATION_LEAKY); -tk::dnn::Conv2d c48 (&net, 512, 3, 3, 1, 1, 1, 1, c48_bin, true); -tk::dnn::Activation a48 (&net, tk::dnn::ACTIVATION_LEAKY); -tk::dnn::Shortcut s49 (&net, &s46); -tk::dnn::Conv2d c50 (&net, 256, 1, 1, 1, 1, 0, 0, c50_bin, true); -tk::dnn::Activation a50 (&net, tk::dnn::ACTIVATION_LEAKY); -tk::dnn::Conv2d c51 (&net, 512, 3, 3, 1, 1, 1, 1, c51_bin, true); -tk::dnn::Activation a51 (&net, tk::dnn::ACTIVATION_LEAKY); -tk::dnn::Shortcut s52 (&net, &s49); -tk::dnn::Conv2d c53 (&net, 256, 1, 1, 1, 1, 0, 0, c53_bin, true); -tk::dnn::Activation a53 (&net, tk::dnn::ACTIVATION_LEAKY); -tk::dnn::Conv2d c54 (&net, 512, 3, 3, 1, 1, 1, 1, c54_bin, true); -tk::dnn::Activation a54 (&net, tk::dnn::ACTIVATION_LEAKY); -tk::dnn::Shortcut s55 (&net, &s52); -tk::dnn::Conv2d c56 (&net, 256, 1, 1, 1, 1, 0, 0, c56_bin, true); -tk::dnn::Activation a56 (&net, tk::dnn::ACTIVATION_LEAKY); -tk::dnn::Conv2d c57 (&net, 512, 3, 3, 1, 1, 1, 1, c57_bin, true); -tk::dnn::Activation a57 (&net, tk::dnn::ACTIVATION_LEAKY); -tk::dnn::Shortcut s58 (&net, &s55); -tk::dnn::Conv2d c59 (&net, 256, 1, 1, 1, 1, 0, 0, c59_bin, true); -tk::dnn::Activation a59 (&net, tk::dnn::ACTIVATION_LEAKY); -tk::dnn::Conv2d c60 (&net, 512, 3, 3, 1, 1, 1, 1, c60_bin, true); -tk::dnn::Activation a60 (&net, tk::dnn::ACTIVATION_LEAKY); -tk::dnn::Shortcut s61 (&net, &s58); - -tk::dnn::Conv2d c62 (&net,1024, 3, 3, 2, 2, 1, 1, c62_bin, true); -tk::dnn::Activation a62 (&net, tk::dnn::ACTIVATION_LEAKY); -tk::dnn::Conv2d c63 (&net, 512, 1, 1, 1, 1, 0, 0, c63_bin, true); -tk::dnn::Activation a63 (&net, tk::dnn::ACTIVATION_LEAKY); -tk::dnn::Conv2d c64 (&net,1024, 3, 3, 1, 1, 1, 1, c64_bin, true); -tk::dnn::Activation a64 (&net, tk::dnn::ACTIVATION_LEAKY); -tk::dnn::Shortcut s65 (&net, &a62); - -tk::dnn::Conv2d c66 (&net, 512, 1, 1, 1, 1, 0, 0, c66_bin, true); -tk::dnn::Activation a66 (&net, tk::dnn::ACTIVATION_LEAKY); -tk::dnn::Conv2d c67 (&net,1024, 3, 3, 1, 1, 1, 1, c67_bin, true); -tk::dnn::Activation a67 (&net, tk::dnn::ACTIVATION_LEAKY); -tk::dnn::Shortcut s68 (&net, &s65); - -tk::dnn::Conv2d c69 (&net, 512, 1, 1, 1, 1, 0, 0, c69_bin, true); -tk::dnn::Activation a69 (&net, tk::dnn::ACTIVATION_LEAKY); -tk::dnn::Conv2d c70 (&net,1024, 3, 3, 1, 1, 1, 1, c70_bin, true); -tk::dnn::Activation a70 (&net, tk::dnn::ACTIVATION_LEAKY); -tk::dnn::Shortcut s71 (&net, &s68); - -tk::dnn::Conv2d c72 (&net, 512, 1, 1, 1, 1, 0, 0, c72_bin, true); -tk::dnn::Activation a72 (&net, tk::dnn::ACTIVATION_LEAKY); -tk::dnn::Conv2d c73 (&net,1024, 3, 3, 1, 1, 1, 1, c73_bin, true); -tk::dnn::Activation a73 (&net, tk::dnn::ACTIVATION_LEAKY); -tk::dnn::Shortcut s74 (&net, &s71); - -tk::dnn::Conv2d c75 (&net, 512, 1, 1, 1, 1, 0, 0, c75_bin, true); -tk::dnn::Activation a75 (&net, tk::dnn::ACTIVATION_LEAKY); -tk::dnn::Conv2d c76 (&net,1024, 3, 3, 1, 1, 1, 1, c76_bin, true); -tk::dnn::Activation a76 (&net, tk::dnn::ACTIVATION_LEAKY); -tk::dnn::Conv2d c77 (&net, 512, 1, 1, 1, 1, 0, 0, c77_bin, true); -tk::dnn::Activation a77 (&net, tk::dnn::ACTIVATION_LEAKY); -tk::dnn::Conv2d c78 (&net,1024, 3, 3, 1, 1, 1, 1, c78_bin, true); -tk::dnn::Activation a78 (&net, tk::dnn::ACTIVATION_LEAKY); -tk::dnn::Conv2d c79 (&net, 512, 1, 1, 1, 1, 0, 0, c79_bin, true); -tk::dnn::Activation a79 (&net, tk::dnn::ACTIVATION_LEAKY); -tk::dnn::Conv2d c80 (&net,1024, 3, 3, 1, 1, 1, 1, c80_bin, true); -tk::dnn::Activation a80 (&net, tk::dnn::ACTIVATION_LEAKY); -tk::dnn::Conv2d c81 (&net, preYoloFilters, 1, 1, 1, 1, 0, 0, c81_bin, false); -tk::dnn::Yolo yolo0 (&net, classes, 3, g82_bin); - -tk::dnn::Layer *m83_layers[1] = { &a79 }; -tk::dnn::Route m83 (&net, m83_layers, 1); -tk::dnn::Conv2d c84 (&net, 256, 1, 1, 1, 1, 0, 0, c84_bin, true); -tk::dnn::Activation a84 (&net, tk::dnn::ACTIVATION_LEAKY); -tk::dnn::Upsample u85 (&net, 2); - -tk::dnn::Layer *m86_layers[2] = { &u85, &s61 }; -tk::dnn::Route m86 (&net, m86_layers, 2); -tk::dnn::Conv2d c87 (&net, 256, 1, 1, 1, 1, 0, 0, c87_bin, true); -tk::dnn::Activation a87 (&net, tk::dnn::ACTIVATION_LEAKY); -tk::dnn::Conv2d c88 (&net, 512, 3, 3, 1, 1, 1, 1, c88_bin, true); -tk::dnn::Activation a88 (&net, tk::dnn::ACTIVATION_LEAKY); -tk::dnn::Conv2d c89 (&net, 256, 1, 1, 1, 1, 0, 0, c89_bin, true); -tk::dnn::Activation a89 (&net, tk::dnn::ACTIVATION_LEAKY); -tk::dnn::Conv2d c90 (&net, 512, 3, 3, 1, 1, 1, 1, c90_bin, true); -tk::dnn::Activation a90 (&net, tk::dnn::ACTIVATION_LEAKY); -tk::dnn::Conv2d c91 (&net, 256, 1, 1, 1, 1, 0, 0, c91_bin, true); -tk::dnn::Activation a91 (&net, tk::dnn::ACTIVATION_LEAKY); - -tk::dnn::Conv2d c92 (&net, 512, 3, 3, 1, 1, 1, 1, c92_bin, true); -tk::dnn::Activation a92 (&net, tk::dnn::ACTIVATION_LEAKY); -tk::dnn::Conv2d c93 (&net, preYoloFilters, 1, 1, 1, 1, 0, 0, c93_bin, false); -tk::dnn::Yolo yolo1 (&net, classes, 3, g94_bin); - -tk::dnn::Layer *m95_layers[1] = { &a91 }; -tk::dnn::Route m95 (&net, m95_layers, 1); -tk::dnn::Conv2d c96 (&net, 128, 1, 1, 1, 1, 0, 0, c96_bin, true); -tk::dnn::Activation a96 (&net, tk::dnn::ACTIVATION_LEAKY); -tk::dnn::Upsample u97 (&net, 2); - -tk::dnn::Layer *m98_layers[2] = { &u97, &s36 }; -tk::dnn::Route m98 (&net, m98_layers, 2); -tk::dnn::Conv2d c99 (&net, 128, 1, 1, 1, 1, 0, 0, c99_bin, true); -tk::dnn::Activation a99 (&net, tk::dnn::ACTIVATION_LEAKY); -tk::dnn::Conv2d c100 (&net, 256, 3, 3, 1, 1, 1, 1, c100_bin, true); -tk::dnn::Activation a100 (&net, tk::dnn::ACTIVATION_LEAKY); -tk::dnn::Conv2d c101 (&net, 128, 1, 1, 1, 1, 0, 0, c101_bin, true); -tk::dnn::Activation a101 (&net, tk::dnn::ACTIVATION_LEAKY); -tk::dnn::Conv2d c102 (&net, 256, 3, 3, 1, 1, 1, 1, c102_bin, true); -tk::dnn::Activation a102 (&net, tk::dnn::ACTIVATION_LEAKY); -tk::dnn::Conv2d c103 (&net, 128, 1, 1, 1, 1, 0, 0, c103_bin, true); -tk::dnn::Activation a103 (&net, tk::dnn::ACTIVATION_LEAKY); - -tk::dnn::Conv2d c104 (&net, 256, 3, 3, 1, 1, 1, 1, c104_bin, true); -tk::dnn::Activation a104 (&net, tk::dnn::ACTIVATION_LEAKY); -tk::dnn::Conv2d c105 (&net, preYoloFilters, 1, 1, 1, 1, 0, 0, c105_bin, false); -tk::dnn::Yolo yolo2 (&net, classes, 3, g106_bin); - -yolo[0] = &yolo0; -yolo[1] = &yolo1; -yolo[2] = &yolo2; \ No newline at end of file diff --git a/include/tkDNN/pluginsRT/ActivationLeakyRT.h b/include/tkDNN/pluginsRT/ActivationLeakyRT.h index 1957e7c..9e26b2b 100644 --- a/include/tkDNN/pluginsRT/ActivationLeakyRT.h +++ b/include/tkDNN/pluginsRT/ActivationLeakyRT.h @@ -42,7 +42,7 @@ public: virtual int enqueue(int batchSize, const void*const * inputs, void** outputs, void* workspace, cudaStream_t stream) override { activationLEAKYForward((dnnType*)reinterpret_cast(inputs[0]), - reinterpret_cast(outputs[0]), size, stream); + reinterpret_cast(outputs[0]), batchSize*size, stream); return 0; } @@ -52,8 +52,9 @@ public: } virtual void serialize(void* buffer) override { - char *buf = reinterpret_cast(buffer); + char *buf = reinterpret_cast(buffer),*a=buf; tk::dnn::writeBUF(buf, size); + assert(buf == a + getSerializationSize()); } int size; diff --git a/include/tkDNN/pluginsRT/ActivationLogisticRT.h b/include/tkDNN/pluginsRT/ActivationLogisticRT.h new file mode 100644 index 0000000..063931f --- /dev/null +++ b/include/tkDNN/pluginsRT/ActivationLogisticRT.h @@ -0,0 +1,60 @@ +#include +#include "../kernels.h" + +class ActivationLogisticRT : public IPlugin { + +public: + ActivationLogisticRT() { + + + } + + ~ActivationLogisticRT(){ + + } + + int getNbOutputs() const override { + return 1; + } + + Dims getOutputDimensions(int index, const Dims* inputs, int nbInputDims) override { + return inputs[0]; + } + + void configure(const Dims* inputDims, int nbInputs, const Dims* outputDims, int nbOutputs, int maxBatchSize) override { + size = 1; + for(int i=0; i(inputs[0]), + reinterpret_cast(outputs[0]), batchSize*size, stream); + return 0; + } + + + virtual size_t getSerializationSize() override { + return 1*sizeof(int); + } + + virtual void serialize(void* buffer) override { + char *buf = reinterpret_cast(buffer); + tk::dnn::writeBUF(buf, size); + } + + int size; +}; diff --git a/include/tkDNN/pluginsRT/ActivationMishRT.h b/include/tkDNN/pluginsRT/ActivationMishRT.h new file mode 100644 index 0000000..5d660af --- /dev/null +++ b/include/tkDNN/pluginsRT/ActivationMishRT.h @@ -0,0 +1,61 @@ +#include +#include "../kernels.h" + +class ActivationMishRT : public IPlugin { + +public: + ActivationMishRT() { + + + } + + ~ActivationMishRT(){ + + } + + int getNbOutputs() const override { + return 1; + } + + Dims getOutputDimensions(int index, const Dims* inputs, int nbInputDims) override { + return inputs[0]; + } + + void configure(const Dims* inputDims, int nbInputs, const Dims* outputDims, int nbOutputs, int maxBatchSize) override { + size = 1; + for(int i=0; i(inputs[0]), + reinterpret_cast(outputs[0]), batchSize*size, stream); + return 0; + } + + + virtual size_t getSerializationSize() override { + return 1*sizeof(int); + } + + virtual void serialize(void* buffer) override { + char *buf = reinterpret_cast(buffer),*a=buf; + tk::dnn::writeBUF(buf, size); + assert(buf == a + getSerializationSize()); + } + + int size; +}; diff --git a/include/tkDNN/pluginsRT/ActivationReLUCeilingRT.h b/include/tkDNN/pluginsRT/ActivationReLUCeilingRT.h new file mode 100644 index 0000000..50ceb81 --- /dev/null +++ b/include/tkDNN/pluginsRT/ActivationReLUCeilingRT.h @@ -0,0 +1,63 @@ +#include +#include "../kernels.h" + +class ActivationReLUCeiling : public IPlugin { + +public: + ActivationReLUCeiling(const float ceiling) { + this->ceiling = ceiling; + } + + ~ActivationReLUCeiling(){ + + } + + int getNbOutputs() const override { + return 1; + } + + Dims getOutputDimensions(int index, const Dims* inputs, int nbInputDims) override { + return inputs[0]; + } + + void configure(const Dims* inputDims, int nbInputs, const Dims* outputDims, int nbOutputs, int maxBatchSize) override { + size = 1; + for(int i=0; i(inputs[0]), + reinterpret_cast(outputs[0]), batchSize*size, ceiling, stream); + return 0; + } + + + virtual size_t getSerializationSize() override { + return 1*sizeof(int) + 1*sizeof(float); + } + + virtual void serialize(void* buffer) override { + char *buf = reinterpret_cast(buffer),*a=buf; + tk::dnn::writeBUF(buf, ceiling); + tk::dnn::writeBUF(buf, size); + assert(buf = a + getSerializationSize()); + + } + + int size; + float ceiling; +}; diff --git a/include/tkDNN/pluginsRT/ActivationSigmoidRT.h b/include/tkDNN/pluginsRT/ActivationSigmoidRT.h new file mode 100644 index 0000000..bcc58c7 --- /dev/null +++ b/include/tkDNN/pluginsRT/ActivationSigmoidRT.h @@ -0,0 +1,61 @@ +#include +#include "../kernels.h" + +class ActivationSigmoidRT : public IPlugin { + +public: + ActivationSigmoidRT() { + + + } + + ~ActivationSigmoidRT(){ + + } + + int getNbOutputs() const override { + return 1; + } + + Dims getOutputDimensions(int index, const Dims* inputs, int nbInputDims) override { + return inputs[0]; + } + + void configure(const Dims* inputDims, int nbInputs, const Dims* outputDims, int nbOutputs, int maxBatchSize) override { + size = 1; + for(int i=0; i(inputs[0]), + reinterpret_cast(outputs[0]), batchSize*size, stream); + return 0; + } + + + virtual size_t getSerializationSize() override { + return 1*sizeof(int); + } + + virtual void serialize(void* buffer) override { + char *buf = reinterpret_cast(buffer),*a=buf; + tk::dnn::writeBUF(buf, size); + assert(buf == a + getSerializationSize()); + } + + int size; +}; diff --git a/include/tkDNN/pluginsRT/DeformableConvRT.h b/include/tkDNN/pluginsRT/DeformableConvRT.h index b236095..5cb2bab 100644 --- a/include/tkDNN/pluginsRT/DeformableConvRT.h +++ b/include/tkDNN/pluginsRT/DeformableConvRT.h @@ -12,12 +12,6 @@ public: int o_n, int o_c, int o_h, int o_w, tk::dnn::DeformConv2d *deformable = nullptr) { this->chunk_dim = chunk_dim; - // int dst_dim = conv_dim.tot(); - // std::cout<<"conv_dim: \n"; - // conv_dim.print(); - // if (dst_dim % 3 != 0 ) - // std::cout<<"take attention\n\n"; - // this->chunk_dim = dst_dim/3; this->kh = kh; this->kw = kw; this->sh = sh; @@ -52,10 +46,19 @@ public: checkCuda( cudaMemcpy(mask, deformable->mask, sizeof(dnnType)*chunk_dim, cudaMemcpyDeviceToDevice) ); checkCuda( cudaMemcpy(ones_d2, deformable->ones_d2, sizeof(dnnType)*dim_ones, cudaMemcpyDeviceToDevice) ); } + stat = cublasCreate(&handle); + if (stat != CUBLAS_STATUS_SUCCESS) + FatalError("CUBLAS initialization failed\n"); } - ~DeformableConvRT(){ - + ~DeformableConvRT() { + checkCuda( cudaFree(data_d) ); + checkCuda( cudaFree(bias2_d) ); + checkCuda( cudaFree(ones_d1) ); + checkCuda( cudaFree(offset) ); + checkCuda( cudaFree(mask) ); + checkCuda( cudaFree(ones_d2) ); + cublasDestroy(handle); } int getNbOutputs() const override { @@ -66,54 +69,43 @@ public: return DimsCHW{defRT->output_dim.c, defRT->output_dim.h, defRT->output_dim.w}; } - void configure(const Dims* inputDims, int nbInputs, const Dims* outputDims, int nbOutputs, int maxBatchSize) override { - // i_n = 1; - // i_c = inputDims[0].d[0]; - // i_h = inputDims[0].d[1]; - // i_w = inputDims[0].d[2]; - // o_n = 1; - // o_c = outputDims[0].d[0]; - // o_h = outputDims[0].d[1]; - // o_w = outputDims[0].d[2]; - } + void configure(const Dims* inputDims, int nbInputs, const Dims* outputDims, int nbOutputs, int maxBatchSize) override { } int initialize() override { - return 0; } - virtual void terminate() override { - } + virtual void terminate() override { } virtual size_t getWorkspaceSize(int maxBatchSize) const override { return 0; } virtual int enqueue(int batchSize, const void*const * inputs, void** outputs, void* workspace, cudaStream_t stream) override { -std::cout<<"LOL\n"; dnnType *srcData = (dnnType*)reinterpret_cast(inputs[0]); dnnType *output_conv = (dnnType*)reinterpret_cast(inputs[1]); // split conv2d outputs into offset to mask - checkCuda(cudaMemcpy(offset, output_conv, 2*chunk_dim*sizeof(dnnType), cudaMemcpyDeviceToDevice)); - checkCuda(cudaMemcpy(mask, output_conv + 2*chunk_dim, chunk_dim*sizeof(dnnType), cudaMemcpyDeviceToDevice)); - // kernel sigmoide - activationSIGMOIDForward(mask, mask, chunk_dim); - - // deformable convolution - dcn_v2_cuda_forward(srcData, data_d, - bias2_d, ones_d1, - offset, mask, - reinterpret_cast(outputs[0]), ones_d2, - kh, kw, - sh, sw, - ph, pw, - 1, 1, - deformableGroup, - i_n, i_c, i_h, i_w, - o_n, o_c, o_h, o_w, - chunk_dim); - + for(int b=0; b(outputs[0]), ones_d2, + kh, kw, + sh, sw, + ph, pw, + 1, 1, + deformableGroup, b, + i_n, i_c, i_h, i_w, + o_n, o_c, o_h, o_w, + chunk_dim); + } return 0; } @@ -124,7 +116,7 @@ std::cout<<"LOL\n"; } virtual void serialize(void* buffer) override { - char *buf = reinterpret_cast(buffer); + char *buf = reinterpret_cast(buffer),*a=buf; tk::dnn::writeBUF(buf, chunk_dim); tk::dnn::writeBUF(buf, kh); tk::dnn::writeBUF(buf, kw); @@ -171,8 +163,11 @@ std::cout<<"LOL\n"; for(int i=0; i + +class FlattenConcatRT : public IPlugin { + +public: + FlattenConcatRT() { + stat = cublasCreate(&handle); + if (stat != CUBLAS_STATUS_SUCCESS) { + printf ("CUBLAS initialization failed\n"); + return; + } + } + + ~FlattenConcatRT(){ + + } + + int getNbOutputs() const override { + return 1; + } + + Dims getOutputDimensions(int index, const Dims* inputs, int nbInputDims) override { + return DimsCHW{ inputs[0].d[0] * inputs[0].d[1] * inputs[0].d[2], 1, 1}; + } + + void configure(const Dims* inputDims, int nbInputs, const Dims* outputDims, int nbOutputs, int maxBatchSize) override { + assert(nbOutputs == 1 && nbInputs ==1); + rows = inputDims[0].d[0]; + cols = inputDims[0].d[1] * inputDims[0].d[2]; + c = inputDims[0].d[0] * inputDims[0].d[1] * inputDims[0].d[2]; + h = 1; + w = 1; + } + + int initialize() override { + return 0; + } + + virtual void terminate() override { + checkERROR(cublasDestroy(handle)); + } + + virtual size_t getWorkspaceSize(int maxBatchSize) const override { + return 0; + } + + virtual int enqueue(int batchSize, const void*const * inputs, void** outputs, void* workspace, cudaStream_t stream) override { + dnnType *srcData = (dnnType*)reinterpret_cast(inputs[0]); + dnnType *dstData = reinterpret_cast(outputs[0]); + checkCuda( cudaMemcpyAsync(dstData, srcData, batchSize*rows*cols*sizeof(dnnType), cudaMemcpyDeviceToDevice, stream)); + + checkERROR( cublasSetStream(handle, stream) ); + for(int i=0; i(buffer),*a = buf; + tk::dnn::writeBUF(buf, c); + tk::dnn::writeBUF(buf, h); + tk::dnn::writeBUF(buf, w); + tk::dnn::writeBUF(buf, rows); + tk::dnn::writeBUF(buf, cols); + assert(buf == a + getSerializationSize()); + } + + int c, h, w; + int rows, cols; + cublasStatus_t stat; + cublasHandle_t handle; +}; diff --git a/include/tkDNN/pluginsRT/MaxPoolingFixedSizeRT.h b/include/tkDNN/pluginsRT/MaxPoolingFixedSizeRT.h new file mode 100644 index 0000000..0899a34 --- /dev/null +++ b/include/tkDNN/pluginsRT/MaxPoolingFixedSizeRT.h @@ -0,0 +1,75 @@ +#include +#include "../kernels.h" + +class MaxPoolFixedSizeRT : public IPlugin { + +public: + MaxPoolFixedSizeRT(int c, int h, int w, int n, int strideH, int strideW, int winSize, int padding) { + this->c = c; + this->h = h; + this->w = w; + this->n = n; + this->stride_H = strideH; + this->stride_W = strideW; + this->winSize = winSize; + this->padding = padding; + } + + ~MaxPoolFixedSizeRT(){ + } + + int getNbOutputs() const override { + return 1; + } + + Dims getOutputDimensions(int index, const Dims* inputs, int nbInputDims) override { + return DimsCHW{this->c, this->h, this->w}; + } + + void configure(const Dims* inputDims, int nbInputs, const Dims* outputDims, int nbOutputs, int maxBatchSize) override { + } + + int initialize() override { + return 0; + } + + virtual void terminate() override { + } + + virtual size_t getWorkspaceSize(int maxBatchSize) const override { + return 0; + } + + virtual int enqueue(int batchSize, const void*const * inputs, void** outputs, void* workspace, cudaStream_t stream) override { + + //std::cout<n<<" "<c<<" "<h<<" "<w<<" "<stride_H<<" "<stride_W<<" "<winSize<<" "<padding<(inputs[0]); + dnnType *dstData = reinterpret_cast(outputs[0]); + MaxPoolingForward(srcData, dstData, batchSize, this->c, this->h, this->w, this->stride_H, this->stride_W, this->winSize, this->padding, stream); + return 0; + } + + + virtual size_t getSerializationSize() override { + return 8*sizeof(int); + } + + virtual void serialize(void* buffer) override { + char *buf = reinterpret_cast(buffer),*a=buf; + + tk::dnn::writeBUF(buf, this->c); + tk::dnn::writeBUF(buf, this->h); + tk::dnn::writeBUF(buf, this->w); + tk::dnn::writeBUF(buf, this->n); + tk::dnn::writeBUF(buf, this->stride_H); + tk::dnn::writeBUF(buf, this->stride_W); + tk::dnn::writeBUF(buf, this->winSize); + tk::dnn::writeBUF(buf, this->padding); + assert(buf == a + getSerializationSize()); + } + + int n, c, h, w; + int stride_H, stride_W; + int winSize; + int padding; +}; diff --git a/include/tkDNN/pluginsRT/RegionRT.h b/include/tkDNN/pluginsRT/RegionRT.h index 6331525..8487652 100644 --- a/include/tkDNN/pluginsRT/RegionRT.h +++ b/include/tkDNN/pluginsRT/RegionRT.h @@ -50,18 +50,18 @@ public: for (int b = 0; b < batchSize; ++b){ for(int n = 0; n < num; ++n){ - int index = entry_index(b, n*w*h, 0, batchSize); + int index = entry_index(b, n*w*h, 0); activationLOGISTICForward(srcData + index, dstData + index, 2*w*h, stream); - index = entry_index(b, n*w*h, coords, batchSize); + index = entry_index(b, n*w*h, coords); activationLOGISTICForward(srcData + index, dstData + index, w*h, stream); } } //softmax start - int index = entry_index(0, 0, coords + 1, batchSize); + int index = entry_index(0, 0, coords + 1); softmaxForward( srcData + index, classes, batchSize*num, - (batchSize*c*h*w)/num, + (c*h*w)/num, w*h, 1, w*h, 1, dstData + index, stream); return 0; @@ -73,22 +73,23 @@ public: } virtual void serialize(void* buffer) override { - char *buf = reinterpret_cast(buffer); + char *buf = reinterpret_cast(buffer),*a=buf; tk::dnn::writeBUF(buf, classes); tk::dnn::writeBUF(buf, coords); tk::dnn::writeBUF(buf, num); tk::dnn::writeBUF(buf, c); tk::dnn::writeBUF(buf, h); tk::dnn::writeBUF(buf, w); + assert(buf == a + getSerializationSize()); } int c, h, w; int classes, coords, num; - int entry_index(int batch, int location, int entry, int batchSize) { + int entry_index(int batch, int location, int entry) { int n = location / (w*h); int loc = location % (w*h); - return batch*c*h*w*batchSize + n*w*h*(coords+classes+1) + entry*w*h + loc; + return batch*c*h*w + n*w*h*(coords+classes+1) + entry*w*h + loc; } }; diff --git a/include/tkDNN/pluginsRT/ReorgRT.h b/include/tkDNN/pluginsRT/ReorgRT.h index ee85718..c1b529a 100644 --- a/include/tkDNN/pluginsRT/ReorgRT.h +++ b/include/tkDNN/pluginsRT/ReorgRT.h @@ -52,11 +52,12 @@ public: } virtual void serialize(void* buffer) override { - char *buf = reinterpret_cast(buffer); + char *buf = reinterpret_cast(buffer),*a=buf; tk::dnn::writeBUF(buf, stride); tk::dnn::writeBUF(buf, c); tk::dnn::writeBUF(buf, h); tk::dnn::writeBUF(buf, w); + assert(buf == a + getSerializationSize()); } int c, h, w, stride; diff --git a/include/tkDNN/pluginsRT/ReshapeRT.h b/include/tkDNN/pluginsRT/ReshapeRT.h new file mode 100644 index 0000000..37017c7 --- /dev/null +++ b/include/tkDNN/pluginsRT/ReshapeRT.h @@ -0,0 +1,62 @@ +#include + +class ReshapeRT : public IPlugin { + +public: + ReshapeRT(dataDim_t new_dim) { + n = new_dim.n; + c = new_dim.c; + h = new_dim.h; + w = new_dim.w; + } + + ~ReshapeRT(){ + + } + + int getNbOutputs() const override { + return 1; + } + + Dims getOutputDimensions(int index, const Dims* inputs, int nbInputDims) override { + return DimsCHW{ c,h,w}; + } + + void configure(const Dims* inputDims, int nbInputs, const Dims* outputDims, int nbOutputs, int maxBatchSize) override { + } + + int initialize() override { + return 0; + } + + virtual void terminate() override { + } + + virtual size_t getWorkspaceSize(int maxBatchSize) const override { + return 0; + } + + virtual int enqueue(int batchSize, const void*const * inputs, void** outputs, void* workspace, cudaStream_t stream) override { + dnnType *srcData = (dnnType*)reinterpret_cast(inputs[0]); + dnnType *dstData = reinterpret_cast(outputs[0]); + + checkCuda( cudaMemcpyAsync(dstData, srcData, batchSize*c*h*w*sizeof(dnnType), cudaMemcpyDeviceToDevice, stream)); + return 0; + } + + + virtual size_t getSerializationSize() override { + return 4*sizeof(int); + } + + virtual void serialize(void* buffer) override { + char *buf = reinterpret_cast(buffer),*a = buf; + tk::dnn::writeBUF(buf, n); + tk::dnn::writeBUF(buf, c); + tk::dnn::writeBUF(buf, h); + tk::dnn::writeBUF(buf, w); + assert(buf == a + getSerializationSize()); + } + + int n, c, h, w; +}; diff --git a/include/tkDNN/pluginsRT/ResizeLayerRT.h b/include/tkDNN/pluginsRT/ResizeLayerRT.h index ae87dbf..cde52bf 100644 --- a/include/tkDNN/pluginsRT/ResizeLayerRT.h +++ b/include/tkDNN/pluginsRT/ResizeLayerRT.h @@ -52,7 +52,7 @@ public: } virtual void serialize(void* buffer) override { - char *buf = reinterpret_cast(buffer); + char *buf = reinterpret_cast(buffer),*a=buf; tk::dnn::writeBUF(buf, o_c); tk::dnn::writeBUF(buf, o_h); @@ -61,6 +61,7 @@ public: tk::dnn::writeBUF(buf, i_c); tk::dnn::writeBUF(buf, i_h); tk::dnn::writeBUF(buf, i_w); + assert(buf == a + getSerializationSize()); } int i_c, i_h, i_w, o_c, o_h, o_w; diff --git a/include/tkDNN/pluginsRT/RouteRT.h b/include/tkDNN/pluginsRT/RouteRT.h index 9abcd7f..5a8c170 100644 --- a/include/tkDNN/pluginsRT/RouteRT.h +++ b/include/tkDNN/pluginsRT/RouteRT.h @@ -3,8 +3,14 @@ class RouteRT : public IPlugin { + /** + THIS IS NOT USED ANYMORE + */ + public: - RouteRT() { + RouteRT(int groups, int group_id) { + this->groups = groups; + this->group_id = group_id; } ~RouteRT(){ @@ -18,7 +24,7 @@ public: Dims getOutputDimensions(int index, const Dims* inputs, int nbInputDims) override { int out_c = 0; for(int i=0; i(outputs[0]); - int offset = 0; - for(int i=0; i(inputs[i]); - int in_dim = c_in[i]*h*w; - checkCuda( cudaMemcpyAsync(dstData + offset, input, in_dim*sizeof(dnnType), cudaMemcpyDeviceToDevice, stream) ); - offset += in_dim; + for(int b=0; b(inputs[i]); + int in_dim = c_in[i]*h*w; + int part_in_dim = in_dim / this->groups; + checkCuda( cudaMemcpyAsync(dstData + b*c*w*h + offset, input + b*c*w*h*groups + this->group_id*part_in_dim, part_in_dim*sizeof(dnnType), cudaMemcpyDeviceToDevice, stream) ); + offset += part_in_dim; + } } return 0; @@ -61,11 +71,13 @@ public: virtual size_t getSerializationSize() override { - return (4+MAX_INPUTS)*sizeof(int); + return (6+MAX_INPUTS)*sizeof(int); } virtual void serialize(void* buffer) override { - char *buf = reinterpret_cast(buffer); + char *buf = reinterpret_cast(buffer),*a=buf; + tk::dnn::writeBUF(buf, groups); + tk::dnn::writeBUF(buf, group_id); tk::dnn::writeBUF(buf, in); for(int i=0; ibc = bdim.c; + this->bh = bdim.h; + this->bw = bdim.w; } ~ShortcutRT(){ @@ -44,22 +47,29 @@ public: dnnType *dstData = reinterpret_cast(outputs[0]); checkCuda( cudaMemcpyAsync(dstData, srcData, batchSize*c*h*w*sizeof(dnnType), cudaMemcpyDeviceToDevice, stream)); - shortcutForward(srcDataBack, dstData, batchSize, c, h, w, 1, batchSize, c, h, w, 1, stream); + for(int b=0; b < batchSize; ++b) + shortcutForward(srcDataBack + b*bc*bh*bw, dstData + b*c*h*w, 1, c, h, w, 1, 1, bc, bh, bw, 1, stream); return 0; } virtual size_t getSerializationSize() override { - return 3*sizeof(int); + return 6*sizeof(int); } virtual void serialize(void* buffer) override { - char *buf = reinterpret_cast(buffer); + char *buf = reinterpret_cast(buffer),*a=buf; + tk::dnn::writeBUF(buf, bc); + tk::dnn::writeBUF(buf, bh); + tk::dnn::writeBUF(buf, bw); tk::dnn::writeBUF(buf, c); tk::dnn::writeBUF(buf, h); tk::dnn::writeBUF(buf, w); + assert(buf == a + getSerializationSize()); + } int c, h, w; + int bc, bh, bw; }; diff --git a/include/tkDNN/pluginsRT/UpsampleRT.h b/include/tkDNN/pluginsRT/UpsampleRT.h index 7a62abc..a11d7b4 100644 --- a/include/tkDNN/pluginsRT/UpsampleRT.h +++ b/include/tkDNN/pluginsRT/UpsampleRT.h @@ -54,11 +54,12 @@ public: } virtual void serialize(void* buffer) override { - char *buf = reinterpret_cast(buffer); + char *buf = reinterpret_cast(buffer),*a=buf; tk::dnn::writeBUF(buf, stride); tk::dnn::writeBUF(buf, c); tk::dnn::writeBUF(buf, h); tk::dnn::writeBUF(buf, w); + assert(buf == a + getSerializationSize()); } int c, h, w, stride; diff --git a/include/tkDNN/pluginsRT/YoloRT.h b/include/tkDNN/pluginsRT/YoloRT.h index a8b3b7d..5ffe39c 100644 --- a/include/tkDNN/pluginsRT/YoloRT.h +++ b/include/tkDNN/pluginsRT/YoloRT.h @@ -8,11 +8,15 @@ class YoloRT : public IPlugin { public: - YoloRT(int classes, int num, tk::dnn::Yolo *yolo = nullptr, int n_masks=3) { + YoloRT(int classes, int num, tk::dnn::Yolo *yolo = nullptr, int n_masks=3, float scale_xy=1, float nms_thresh=0.45, int nms_kind=0, int new_coords=0) { this->classes = classes; this->num = num; this->n_masks = n_masks; + this->scaleXY = scale_xy; + this->nms_thresh = nms_thresh; + this->nms_kind = nms_kind; + this->new_coords = new_coords; mask = new dnnType[n_masks]; bias = new dnnType[num*n_masks*2]; @@ -60,15 +64,23 @@ public: checkCuda( cudaMemcpyAsync(dstData, srcData, batchSize*c*h*w*sizeof(dnnType), cudaMemcpyDeviceToDevice, stream)); - for (int b = 0; b < batchSize; ++b){ - for(int n = 0; n < n_masks; ++n){ - int index = entry_index(b, n*w*h, 0, batchSize); - activationLOGISTICForward(srcData + index, dstData + index, 2*w*h, stream); - - index = entry_index(b, n*w*h, 4, batchSize); - activationLOGISTICForward(srcData + index, dstData + index, (1+classes)*w*h, stream); - } - } + + for (int b = 0; b < batchSize; ++b){ + for(int n = 0; n < n_masks; ++n){ + int index = entry_index(b, n*w*h, 0); + if (new_coords == 1){ + if (this->scaleXY != 1) scalAdd(dstData + index, 2 * w*h, this->scaleXY, -0.5*(this->scaleXY - 1), 1); + } + else{ + activationLOGISTICForward(srcData + index, dstData + index, 2*w*h, stream); //x,y + + if (this->scaleXY != 1) scalAdd(dstData + index, 2 * w*h, this->scaleXY, -0.5*(this->scaleXY - 1), 1); + + index = entry_index(b, n*w*h, 4); + activationLOGISTICForward(srcData + index, dstData + index, (1+classes)*w*h, stream); + } + } + } //std::cout<<"YOLO END\n"; return 0; @@ -76,21 +88,29 @@ public: virtual size_t getSerializationSize() override { - return 6*sizeof(int) + n_masks*sizeof(dnnType) + num*n_masks*2*sizeof(dnnType) + YOLORT_CLASSNAME_W*classes*sizeof(char); + return 8*sizeof(int) + 2*sizeof(float)+ n_masks*sizeof(dnnType) + num*n_masks*2*sizeof(dnnType) + YOLORT_CLASSNAME_W*classes*sizeof(char); } virtual void serialize(void* buffer) override { - char *buf = reinterpret_cast(buffer); - tk::dnn::writeBUF(buf, classes); - tk::dnn::writeBUF(buf, num); - tk::dnn::writeBUF(buf, n_masks); - tk::dnn::writeBUF(buf, c); - tk::dnn::writeBUF(buf, h); - tk::dnn::writeBUF(buf, w); - for(int i=0; i(buffer),*a=buf; + tk::dnn::writeBUF(buf, classes); //std::cout << "Classes :" << classes << std::endl; + tk::dnn::writeBUF(buf, num); //std::cout << "Num : " << num << std::endl; + tk::dnn::writeBUF(buf, n_masks); //std::cout << "N_Masks" << n_masks << std::endl; + tk::dnn::writeBUF(buf, scaleXY); //std::cout << "ScaleXY :" << scaleXY << std::endl; + tk::dnn::writeBUF(buf, nms_thresh); //std::cout << "nms_thresh :" << nms_thresh << std::endl; + tk::dnn::writeBUF(buf, nms_kind); //std::cout << "nms_kind : " << nms_kind << std::endl; + tk::dnn::writeBUF(buf, new_coords); //std::cout << "new_coords : " << new_coords << std::endl; + tk::dnn::writeBUF(buf, c); //std::cout << "C : " << c << std::endl; + tk::dnn::writeBUF(buf, h); //std::cout << "H : " << h << std::endl; + tk::dnn::writeBUF(buf, w); //std::cout << "C : " << c << std::endl; + for (int i = 0; i < n_masks; i++) + { + tk::dnn::writeBUF(buf, mask[i]); //std::cout << "mask[i] : " << mask[i] << std::endl; + } + for (int i = 0; i < n_masks * 2 * num; i++) + { + tk::dnn::writeBUF(buf, bias[i]); //std::cout << "bias[i] : " << bias[i] << std::endl; + } // save classes names for(int i=0; i classesNames; dnnType *mask; dnnType *bias; - int entry_index(int batch, int location, int entry, int batchSize) { + int entry_index(int batch, int location, int entry) { int n = location / (w*h); int loc = location % (w*h); - return batch*c*h*w*batchSize + n*w*h*(4+classes+1) + entry*w*h + loc; + return batch*c*h*w + n*w*h*(4+classes+1) + entry*w*h + loc; } }; diff --git a/include/tkDNN/stb_image.h b/include/tkDNN/stb_image.h new file mode 100644 index 0000000..e1b322f --- /dev/null +++ b/include/tkDNN/stb_image.h @@ -0,0 +1,7188 @@ +/* stb_image - v2.16 - public domain image loader - http://nothings.org/stb_image.h + no warranty implied; use at your own risk + + Do this: + #define STB_IMAGE_IMPLEMENTATION + before you include this file in *one* C or C++ file to create the implementation. + + // i.e. it should look like this: + #include ... + #include ... + #include ... + #define STB_IMAGE_IMPLEMENTATION + #include "stb_image.h" + + You can #define STBI_ASSERT(x) before the #include to avoid using assert.h. + And #define STBI_MALLOC, STBI_REALLOC, and STBI_FREE to avoid using malloc,realloc,free + + + QUICK NOTES: + Primarily of interest to game developers and other people who can + avoid problematic images and only need the trivial interface + + JPEG baseline & progressive (12 bpc/arithmetic not supported, same as stock IJG lib) + PNG 1/2/4/8/16-bit-per-channel + + TGA (not sure what subset, if a subset) + BMP non-1bpp, non-RLE + PSD (composited view only, no extra channels, 8/16 bit-per-channel) + + GIF (*comp always reports as 4-channel) + HDR (radiance rgbE format) + PIC (Softimage PIC) + PNM (PPM and PGM binary only) + + Animated GIF still needs a proper API, but here's one way to do it: + http://gist.github.com/urraka/685d9a6340b26b830d49 + + - decode from memory or through FILE (define STBI_NO_STDIO to remove code) + - decode from arbitrary I/O callbacks + - SIMD acceleration on x86/x64 (SSE2) and ARM (NEON) + + Full documentation under "DOCUMENTATION" below. + + +LICENSE + + See end of file for license information. + +RECENT REVISION HISTORY: + + 2.16 (2017-07-23) all functions have 16-bit variants; optimizations; bugfixes + 2.15 (2017-03-18) fix png-1,2,4; all Imagenet JPGs; no runtime SSE detection on GCC + 2.14 (2017-03-03) remove deprecated STBI_JPEG_OLD; fixes for Imagenet JPGs + 2.13 (2016-12-04) experimental 16-bit API, only for PNG so far; fixes + 2.12 (2016-04-02) fix typo in 2.11 PSD fix that caused crashes + 2.11 (2016-04-02) 16-bit PNGS; enable SSE2 in non-gcc x64 + RGB-format JPEG; remove white matting in PSD; + allocate large structures on the stack; + correct channel count for PNG & BMP + 2.10 (2016-01-22) avoid warning introduced in 2.09 + 2.09 (2016-01-16) 16-bit TGA; comments in PNM files; STBI_REALLOC_SIZED + + See end of file for full revision history. + + + ============================ Contributors ========================= + + Image formats Extensions, features + Sean Barrett (jpeg, png, bmp) Jetro Lauha (stbi_info) + Nicolas Schulz (hdr, psd) Martin "SpartanJ" Golini (stbi_info) + Jonathan Dummer (tga) James "moose2000" Brown (iPhone PNG) + Jean-Marc Lienher (gif) Ben "Disch" Wenger (io callbacks) + Tom Seddon (pic) Omar Cornut (1/2/4-bit PNG) + Thatcher Ulrich (psd) Nicolas Guillemot (vertical flip) + Ken Miller (pgm, ppm) Richard Mitton (16-bit PSD) + github:urraka (animated gif) Junggon Kim (PNM comments) + Daniel Gibson (16-bit TGA) + socks-the-fox (16-bit PNG) + Jeremy Sawicki (handle all ImageNet JPGs) + Optimizations & bugfixes + Fabian "ryg" Giesen + Arseny Kapoulkine + John-Mark Allen + + Bug & warning fixes + Marc LeBlanc David Woo Guillaume George Martins Mozeiko + Christpher Lloyd Jerry Jansson Joseph Thomson Phil Jordan + Dave Moore Roy Eltham Hayaki Saito Nathan Reed + Won Chun Luke Graham Johan Duparc Nick Verigakis + the Horde3D community Thomas Ruf Ronny Chevalier Baldur Karlsson + Janez Zemva John Bartholomew Michal Cichon github:rlyeh + Jonathan Blow Ken Hamada Tero Hanninen github:romigrou + Laurent Gomila Cort Stratton Sergio Gonzalez github:svdijk + Aruelien Pocheville Thibault Reuille Cass Everitt github:snagar + Ryamond Barbiero Paul Du Bois Engin Manap github:Zelex + Michaelangel007@github Philipp Wiesemann Dale Weiler github:grim210 + Oriol Ferrer Mesia Josh Tobin Matthew Gregan github:sammyhw + Blazej Dariusz Roszkowski Gregory Mullen github:phprus + Christian Floisand Kevin Schmidt github:poppolopoppo +*/ + +#ifndef STBI_INCLUDE_STB_IMAGE_H +#define STBI_INCLUDE_STB_IMAGE_H + +// DOCUMENTATION +// +// Limitations: +// - no 16-bit-per-channel PNG +// - no 12-bit-per-channel JPEG +// - no JPEGs with arithmetic coding +// - no 1-bit BMP +// - GIF always returns *comp=4 +// +// Basic usage (see HDR discussion below for HDR usage): +// int x,y,n; +// unsigned char *data = stbi_load(filename, &x, &y, &n, 0); +// // ... process data if not NULL ... +// // ... x = width, y = height, n = # 8-bit components per pixel ... +// // ... replace '0' with '1'..'4' to force that many components per pixel +// // ... but 'n' will always be the number that it would have been if you said 0 +// stbi_image_free(data) +// +// Standard parameters: +// int *x -- outputs image width in pixels +// int *y -- outputs image height in pixels +// int *channels_in_file -- outputs # of image components in image file +// int desired_channels -- if non-zero, # of image components requested in result +// +// The return value from an image loader is an 'unsigned char *' which points +// to the pixel data, or NULL on an allocation failure or if the image is +// corrupt or invalid. The pixel data consists of *y scanlines of *x pixels, +// with each pixel consisting of N interleaved 8-bit components; the first +// pixel pointed to is top-left-most in the image. There is no padding between +// image scanlines or between pixels, regardless of format. The number of +// components N is 'desired_channels' if desired_channels is non-zero, or +// *channels_in_file otherwise. If desired_channels is non-zero, +// *channels_in_file has the number of components that _would_ have been +// output otherwise. E.g. if you set desired_channels to 4, you will always +// get RGBA output, but you can check *channels_in_file to see if it's trivially +// opaque because e.g. there were only 3 channels in the source image. +// +// An output image with N components has the following components interleaved +// in this order in each pixel: +// +// N=#comp components +// 1 grey +// 2 grey, alpha +// 3 red, green, blue +// 4 red, green, blue, alpha +// +// If image loading fails for any reason, the return value will be NULL, +// and *x, *y, *channels_in_file will be unchanged. The function +// stbi_failure_reason() can be queried for an extremely brief, end-user +// unfriendly explanation of why the load failed. Define STBI_NO_FAILURE_STRINGS +// to avoid compiling these strings at all, and STBI_FAILURE_USERMSG to get slightly +// more user-friendly ones. +// +// Paletted PNG, BMP, GIF, and PIC images are automatically depalettized. +// +// =========================================================================== +// +// Philosophy +// +// stb libraries are designed with the following priorities: +// +// 1. easy to use +// 2. easy to maintain +// 3. good performance +// +// Sometimes I let "good performance" creep up in priority over "easy to maintain", +// and for best performance I may provide less-easy-to-use APIs that give higher +// performance, in addition to the easy to use ones. Nevertheless, it's important +// to keep in mind that from the standpoint of you, a client of this library, +// all you care about is #1 and #3, and stb libraries DO NOT emphasize #3 above all. +// +// Some secondary priorities arise directly from the first two, some of which +// make more explicit reasons why performance can't be emphasized. +// +// - Portable ("ease of use") +// - Small source code footprint ("easy to maintain") +// - No dependencies ("ease of use") +// +// =========================================================================== +// +// I/O callbacks +// +// I/O callbacks allow you to read from arbitrary sources, like packaged +// files or some other source. Data read from callbacks are processed +// through a small internal buffer (currently 128 bytes) to try to reduce +// overhead. +// +// The three functions you must define are "read" (reads some bytes of data), +// "skip" (skips some bytes of data), "eof" (reports if the stream is at the end). +// +// =========================================================================== +// +// SIMD support +// +// The JPEG decoder will try to automatically use SIMD kernels on x86 when +// supported by the compiler. For ARM Neon support, you must explicitly +// request it. +// +// (The old do-it-yourself SIMD API is no longer supported in the current +// code.) +// +// On x86, SSE2 will automatically be used when available based on a run-time +// test; if not, the generic C versions are used as a fall-back. On ARM targets, +// the typical path is to have separate builds for NEON and non-NEON devices +// (at least this is true for iOS and Android). Therefore, the NEON support is +// toggled by a build flag: define STBI_NEON to get NEON loops. +// +// If for some reason you do not want to use any of SIMD code, or if +// you have issues compiling it, you can disable it entirely by +// defining STBI_NO_SIMD. +// +// =========================================================================== +// +// HDR image support (disable by defining STBI_NO_HDR) +// +// stb_image now supports loading HDR images in general, and currently +// the Radiance .HDR file format, although the support is provided +// generically. You can still load any file through the existing interface; +// if you attempt to load an HDR file, it will be automatically remapped to +// LDR, assuming gamma 2.2 and an arbitrary scale factor defaulting to 1; +// both of these constants can be reconfigured through this interface: +// +// stbi_hdr_to_ldr_gamma(2.2f); +// stbi_hdr_to_ldr_scale(1.0f); +// +// (note, do not use _inverse_ constants; stbi_image will invert them +// appropriately). +// +// Additionally, there is a new, parallel interface for loading files as +// (linear) floats to preserve the full dynamic range: +// +// float *data = stbi_loadf(filename, &x, &y, &n, 0); +// +// If you load LDR images through this interface, those images will +// be promoted to floating point values, run through the inverse of +// constants corresponding to the above: +// +// stbi_ldr_to_hdr_scale(1.0f); +// stbi_ldr_to_hdr_gamma(2.2f); +// +// Finally, given a filename (or an open file or memory block--see header +// file for details) containing image data, you can query for the "most +// appropriate" interface to use (that is, whether the image is HDR or +// not), using: +// +// stbi_is_hdr(char *filename); +// +// =========================================================================== +// +// iPhone PNG support: +// +// By default we convert iphone-formatted PNGs back to RGB, even though +// they are internally encoded differently. You can disable this conversion +// by by calling stbi_convert_iphone_png_to_rgb(0), in which case +// you will always just get the native iphone "format" through (which +// is BGR stored in RGB). +// +// Call stbi_set_unpremultiply_on_load(1) as well to force a divide per +// pixel to remove any premultiplied alpha *only* if the image file explicitly +// says there's premultiplied data (currently only happens in iPhone images, +// and only if iPhone convert-to-rgb processing is on). +// +// =========================================================================== +// +// ADDITIONAL CONFIGURATION +// +// - You can suppress implementation of any of the decoders to reduce +// your code footprint by #defining one or more of the following +// symbols before creating the implementation. +// +// STBI_NO_JPEG +// STBI_NO_PNG +// STBI_NO_BMP +// STBI_NO_PSD +// STBI_NO_TGA +// STBI_NO_GIF +// STBI_NO_HDR +// STBI_NO_PIC +// STBI_NO_PNM (.ppm and .pgm) +// +// - You can request *only* certain decoders and suppress all other ones +// (this will be more forward-compatible, as addition of new decoders +// doesn't require you to disable them explicitly): +// +// STBI_ONLY_JPEG +// STBI_ONLY_PNG +// STBI_ONLY_BMP +// STBI_ONLY_PSD +// STBI_ONLY_TGA +// STBI_ONLY_GIF +// STBI_ONLY_HDR +// STBI_ONLY_PIC +// STBI_ONLY_PNM (.ppm and .pgm) +// +// - If you use STBI_NO_PNG (or _ONLY_ without PNG), and you still +// want the zlib decoder to be available, #define STBI_SUPPORT_ZLIB +// + + +#ifndef STBI_NO_STDIO +#include +#endif // STBI_NO_STDIO + +#define STBI_VERSION 1 + +enum +{ + STBI_default = 0, // only used for desired_channels + + STBI_grey = 1, + STBI_grey_alpha = 2, + STBI_rgb = 3, + STBI_rgb_alpha = 4 +}; + +typedef unsigned char stbi_uc; +typedef unsigned short stbi_us; + +#ifdef __cplusplus +extern "C" { +#endif + +#ifdef STB_IMAGE_STATIC +#define STBIDEF static +#else +#define STBIDEF extern +#endif + +////////////////////////////////////////////////////////////////////////////// +// +// PRIMARY API - works on images of any type +// + +// +// load image by filename, open file, or memory buffer +// + +typedef struct +{ + int (*read) (void *user,char *data,int size); // fill 'data' with 'size' bytes. return number of bytes actually read + void (*skip) (void *user,int n); // skip the next 'n' bytes, or 'unget' the last -n bytes if negative + int (*eof) (void *user); // returns nonzero if we are at end of file/data +} stbi_io_callbacks; + +//////////////////////////////////// +// +// 8-bits-per-channel interface +// + +STBIDEF stbi_uc *stbi_load_from_memory (stbi_uc const *buffer, int len , int *x, int *y, int *channels_in_file, int desired_channels); +STBIDEF stbi_uc *stbi_load_from_callbacks(stbi_io_callbacks const *clbk , void *user, int *x, int *y, int *channels_in_file, int desired_channels); + +#ifndef STBI_NO_STDIO +STBIDEF stbi_uc *stbi_load (char const *filename, int *x, int *y, int *channels_in_file, int desired_channels); +STBIDEF stbi_uc *stbi_load_from_file (FILE *f, int *x, int *y, int *channels_in_file, int desired_channels); +// for stbi_load_from_file, file pointer is left pointing immediately after image +#endif + +//////////////////////////////////// +// +// 16-bits-per-channel interface +// + +STBIDEF stbi_us *stbi_load_16_from_memory (stbi_uc const *buffer, int len, int *x, int *y, int *channels_in_file, int desired_channels); +STBIDEF stbi_us *stbi_load_16_from_callbacks(stbi_io_callbacks const *clbk, void *user, int *x, int *y, int *channels_in_file, int desired_channels); + +#ifndef STBI_NO_STDIO +STBIDEF stbi_us *stbi_load_16 (char const *filename, int *x, int *y, int *channels_in_file, int desired_channels); +STBIDEF stbi_us *stbi_load_from_file_16(FILE *f, int *x, int *y, int *channels_in_file, int desired_channels); +#endif + +//////////////////////////////////// +// +// float-per-channel interface +// +#ifndef STBI_NO_LINEAR + STBIDEF float *stbi_loadf_from_memory (stbi_uc const *buffer, int len, int *x, int *y, int *channels_in_file, int desired_channels); + STBIDEF float *stbi_loadf_from_callbacks (stbi_io_callbacks const *clbk, void *user, int *x, int *y, int *channels_in_file, int desired_channels); + + #ifndef STBI_NO_STDIO + STBIDEF float *stbi_loadf (char const *filename, int *x, int *y, int *channels_in_file, int desired_channels); + STBIDEF float *stbi_loadf_from_file (FILE *f, int *x, int *y, int *channels_in_file, int desired_channels); + #endif +#endif + +#ifndef STBI_NO_HDR + STBIDEF void stbi_hdr_to_ldr_gamma(float gamma); + STBIDEF void stbi_hdr_to_ldr_scale(float scale); +#endif // STBI_NO_HDR + +#ifndef STBI_NO_LINEAR + STBIDEF void stbi_ldr_to_hdr_gamma(float gamma); + STBIDEF void stbi_ldr_to_hdr_scale(float scale); +#endif // STBI_NO_LINEAR + +// stbi_is_hdr is always defined, but always returns false if STBI_NO_HDR +STBIDEF int stbi_is_hdr_from_callbacks(stbi_io_callbacks const *clbk, void *user); +STBIDEF int stbi_is_hdr_from_memory(stbi_uc const *buffer, int len); +#ifndef STBI_NO_STDIO +STBIDEF int stbi_is_hdr (char const *filename); +STBIDEF int stbi_is_hdr_from_file(FILE *f); +#endif // STBI_NO_STDIO + + +// get a VERY brief reason for failure +// NOT THREADSAFE +STBIDEF const char *stbi_failure_reason (void); + +// free the loaded image -- this is just free() +STBIDEF void stbi_image_free (void *retval_from_stbi_load); + +// get image dimensions & components without fully decoding +STBIDEF int stbi_info_from_memory(stbi_uc const *buffer, int len, int *x, int *y, int *comp); +STBIDEF int stbi_info_from_callbacks(stbi_io_callbacks const *clbk, void *user, int *x, int *y, int *comp); + +#ifndef STBI_NO_STDIO +STBIDEF int stbi_info (char const *filename, int *x, int *y, int *comp); +STBIDEF int stbi_info_from_file (FILE *f, int *x, int *y, int *comp); + +#endif + + + +// for image formats that explicitly notate that they have premultiplied alpha, +// we just return the colors as stored in the file. set this flag to force +// unpremultiplication. results are undefined if the unpremultiply overflow. +STBIDEF void stbi_set_unpremultiply_on_load(int flag_true_if_should_unpremultiply); + +// indicate whether we should process iphone images back to canonical format, +// or just pass them through "as-is" +STBIDEF void stbi_convert_iphone_png_to_rgb(int flag_true_if_should_convert); + +// flip the image vertically, so the first pixel in the output array is the bottom left +STBIDEF void stbi_set_flip_vertically_on_load(int flag_true_if_should_flip); + +// ZLIB client - used by PNG, available for other purposes + +STBIDEF char *stbi_zlib_decode_malloc_guesssize(const char *buffer, int len, int initial_size, int *outlen); +STBIDEF char *stbi_zlib_decode_malloc_guesssize_headerflag(const char *buffer, int len, int initial_size, int *outlen, int parse_header); +STBIDEF char *stbi_zlib_decode_malloc(const char *buffer, int len, int *outlen); +STBIDEF int stbi_zlib_decode_buffer(char *obuffer, int olen, const char *ibuffer, int ilen); + +STBIDEF char *stbi_zlib_decode_noheader_malloc(const char *buffer, int len, int *outlen); +STBIDEF int stbi_zlib_decode_noheader_buffer(char *obuffer, int olen, const char *ibuffer, int ilen); + + +#ifdef __cplusplus +} +#endif + +// +// +//// end header file ///////////////////////////////////////////////////// +#endif // STBI_INCLUDE_STB_IMAGE_H + +#ifdef STB_IMAGE_IMPLEMENTATION + +#if defined(STBI_ONLY_JPEG) || defined(STBI_ONLY_PNG) || defined(STBI_ONLY_BMP) \ + || defined(STBI_ONLY_TGA) || defined(STBI_ONLY_GIF) || defined(STBI_ONLY_PSD) \ + || defined(STBI_ONLY_HDR) || defined(STBI_ONLY_PIC) || defined(STBI_ONLY_PNM) \ + || defined(STBI_ONLY_ZLIB) + #ifndef STBI_ONLY_JPEG + #define STBI_NO_JPEG + #endif + #ifndef STBI_ONLY_PNG + #define STBI_NO_PNG + #endif + #ifndef STBI_ONLY_BMP + #define STBI_NO_BMP + #endif + #ifndef STBI_ONLY_PSD + #define STBI_NO_PSD + #endif + #ifndef STBI_ONLY_TGA + #define STBI_NO_TGA + #endif + #ifndef STBI_ONLY_GIF + #define STBI_NO_GIF + #endif + #ifndef STBI_ONLY_HDR + #define STBI_NO_HDR + #endif + #ifndef STBI_ONLY_PIC + #define STBI_NO_PIC + #endif + #ifndef STBI_ONLY_PNM + #define STBI_NO_PNM + #endif +#endif + +#if defined(STBI_NO_PNG) && !defined(STBI_SUPPORT_ZLIB) && !defined(STBI_NO_ZLIB) +#define STBI_NO_ZLIB +#endif + + +#include +#include // ptrdiff_t on osx +#include +#include +#include + +#if !defined(STBI_NO_LINEAR) || !defined(STBI_NO_HDR) +#include // ldexp +#endif + +#ifndef STBI_NO_STDIO +#include +#endif + +#ifndef STBI_ASSERT +#include +#define STBI_ASSERT(x) assert(x) +#endif + + +#ifndef _MSC_VER + #ifdef __cplusplus + #define stbi_inline inline + #else + #define stbi_inline + #endif +#else + #define stbi_inline __forceinline +#endif + + +#ifdef _MSC_VER +typedef unsigned short stbi__uint16; +typedef signed short stbi__int16; +typedef unsigned int stbi__uint32; +typedef signed int stbi__int32; +#else +#include +typedef uint16_t stbi__uint16; +typedef int16_t stbi__int16; +typedef uint32_t stbi__uint32; +typedef int32_t stbi__int32; +#endif + +// should produce compiler error if size is wrong +typedef unsigned char validate_uint32[sizeof(stbi__uint32)==4 ? 1 : -1]; + +#ifdef _MSC_VER +#define STBI_NOTUSED(v) (void)(v) +#else +#define STBI_NOTUSED(v) (void)sizeof(v) +#endif + +#ifdef _MSC_VER +#define STBI_HAS_LROTL +#endif + +#ifdef STBI_HAS_LROTL + #define stbi_lrot(x,y) _lrotl(x,y) +#else + #define stbi_lrot(x,y) (((x) << (y)) | ((x) >> (32 - (y)))) +#endif + +#if defined(STBI_MALLOC) && defined(STBI_FREE) && (defined(STBI_REALLOC) || defined(STBI_REALLOC_SIZED)) +// ok +#elif !defined(STBI_MALLOC) && !defined(STBI_FREE) && !defined(STBI_REALLOC) && !defined(STBI_REALLOC_SIZED) +// ok +#else +#error "Must define all or none of STBI_MALLOC, STBI_FREE, and STBI_REALLOC (or STBI_REALLOC_SIZED)." +#endif + +#ifndef STBI_MALLOC +#define STBI_MALLOC(sz) malloc(sz) +#define STBI_REALLOC(p,newsz) realloc(p,newsz) +#define STBI_FREE(p) free(p) +#endif + +#ifndef STBI_REALLOC_SIZED +#define STBI_REALLOC_SIZED(p,oldsz,newsz) STBI_REALLOC(p,newsz) +#endif + +// x86/x64 detection +#if defined(__x86_64__) || defined(_M_X64) +#define STBI__X64_TARGET +#elif defined(__i386) || defined(_M_IX86) +#define STBI__X86_TARGET +#endif + +#if defined(__GNUC__) && defined(STBI__X86_TARGET) && !defined(__SSE2__) && !defined(STBI_NO_SIMD) +// gcc doesn't support sse2 intrinsics unless you compile with -msse2, +// which in turn means it gets to use SSE2 everywhere. This is unfortunate, +// but previous attempts to provide the SSE2 functions with runtime +// detection caused numerous issues. The way architecture extensions are +// exposed in GCC/Clang is, sadly, not really suited for one-file libs. +// New behavior: if compiled with -msse2, we use SSE2 without any +// detection; if not, we don't use it at all. +#define STBI_NO_SIMD +#endif + +#if defined(__MINGW32__) && defined(STBI__X86_TARGET) && !defined(STBI_MINGW_ENABLE_SSE2) && !defined(STBI_NO_SIMD) +// Note that __MINGW32__ doesn't actually mean 32-bit, so we have to avoid STBI__X64_TARGET +// +// 32-bit MinGW wants ESP to be 16-byte aligned, but this is not in the +// Windows ABI and VC++ as well as Windows DLLs don't maintain that invariant. +// As a result, enabling SSE2 on 32-bit MinGW is dangerous when not +// simultaneously enabling "-mstackrealign". +// +// See https://github.com/nothings/stb/issues/81 for more information. +// +// So default to no SSE2 on 32-bit MinGW. If you've read this far and added +// -mstackrealign to your build settings, feel free to #define STBI_MINGW_ENABLE_SSE2. +#define STBI_NO_SIMD +#endif + +#if !defined(STBI_NO_SIMD) && (defined(STBI__X86_TARGET) || defined(STBI__X64_TARGET)) +#define STBI_SSE2 +#include + +#ifdef _MSC_VER + +#if _MSC_VER >= 1400 // not VC6 +#include // __cpuid +static int stbi__cpuid3(void) +{ + int info[4]; + __cpuid(info,1); + return info[3]; +} +#else +static int stbi__cpuid3(void) +{ + int res; + __asm { + mov eax,1 + cpuid + mov res,edx + } + return res; +} +#endif + +#define STBI_SIMD_ALIGN(type, name) __declspec(align(16)) type name + +static int stbi__sse2_available(void) +{ + int info3 = stbi__cpuid3(); + return ((info3 >> 26) & 1) != 0; +} +#else // assume GCC-style if not VC++ +#define STBI_SIMD_ALIGN(type, name) type name __attribute__((aligned(16))) + +static int stbi__sse2_available(void) +{ + // If we're even attempting to compile this on GCC/Clang, that means + // -msse2 is on, which means the compiler is allowed to use SSE2 + // instructions at will, and so are we. + return 1; +} +#endif +#endif + +// ARM NEON +#if defined(STBI_NO_SIMD) && defined(STBI_NEON) +#undef STBI_NEON +#endif + +#ifdef STBI_NEON +#include +// assume GCC or Clang on ARM targets +#define STBI_SIMD_ALIGN(type, name) type name __attribute__((aligned(16))) +#endif + +#ifndef STBI_SIMD_ALIGN +#define STBI_SIMD_ALIGN(type, name) type name +#endif + +/////////////////////////////////////////////// +// +// stbi__context struct and start_xxx functions + +// stbi__context structure is our basic context used by all images, so it +// contains all the IO context, plus some basic image information +typedef struct +{ + stbi__uint32 img_x, img_y; + int img_n, img_out_n; + + stbi_io_callbacks io; + void *io_user_data; + + int read_from_callbacks; + int buflen; + stbi_uc buffer_start[128]; + + stbi_uc *img_buffer, *img_buffer_end; + stbi_uc *img_buffer_original, *img_buffer_original_end; +} stbi__context; + + +static void stbi__refill_buffer(stbi__context *s); + +// initialize a memory-decode context +static void stbi__start_mem(stbi__context *s, stbi_uc const *buffer, int len) +{ + s->io.read = NULL; + s->read_from_callbacks = 0; + s->img_buffer = s->img_buffer_original = (stbi_uc *) buffer; + s->img_buffer_end = s->img_buffer_original_end = (stbi_uc *) buffer+len; +} + +// initialize a callback-based context +static void stbi__start_callbacks(stbi__context *s, stbi_io_callbacks *c, void *user) +{ + s->io = *c; + s->io_user_data = user; + s->buflen = sizeof(s->buffer_start); + s->read_from_callbacks = 1; + s->img_buffer_original = s->buffer_start; + stbi__refill_buffer(s); + s->img_buffer_original_end = s->img_buffer_end; +} + +#ifndef STBI_NO_STDIO + +static int stbi__stdio_read(void *user, char *data, int size) +{ + return (int) fread(data,1,size,(FILE*) user); +} + +static void stbi__stdio_skip(void *user, int n) +{ + fseek((FILE*) user, n, SEEK_CUR); +} + +static int stbi__stdio_eof(void *user) +{ + return feof((FILE*) user); +} + +static stbi_io_callbacks stbi__stdio_callbacks = +{ + stbi__stdio_read, + stbi__stdio_skip, + stbi__stdio_eof, +}; + +static void stbi__start_file(stbi__context *s, FILE *f) +{ + stbi__start_callbacks(s, &stbi__stdio_callbacks, (void *) f); +} + +//static void stop_file(stbi__context *s) { } + +#endif // !STBI_NO_STDIO + +static void stbi__rewind(stbi__context *s) +{ + // conceptually rewind SHOULD rewind to the beginning of the stream, + // but we just rewind to the beginning of the initial buffer, because + // we only use it after doing 'test', which only ever looks at at most 92 bytes + s->img_buffer = s->img_buffer_original; + s->img_buffer_end = s->img_buffer_original_end; +} + +enum +{ + STBI_ORDER_RGB, + STBI_ORDER_BGR +}; + +typedef struct +{ + int bits_per_channel; + int num_channels; + int channel_order; +} stbi__result_info; + +#ifndef STBI_NO_JPEG +static int stbi__jpeg_test(stbi__context *s); +static void *stbi__jpeg_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri); +static int stbi__jpeg_info(stbi__context *s, int *x, int *y, int *comp); +#endif + +#ifndef STBI_NO_PNG +static int stbi__png_test(stbi__context *s); +static void *stbi__png_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri); +static int stbi__png_info(stbi__context *s, int *x, int *y, int *comp); +#endif + +#ifndef STBI_NO_BMP +static int stbi__bmp_test(stbi__context *s); +static void *stbi__bmp_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri); +static int stbi__bmp_info(stbi__context *s, int *x, int *y, int *comp); +#endif + +#ifndef STBI_NO_TGA +static int stbi__tga_test(stbi__context *s); +static void *stbi__tga_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri); +static int stbi__tga_info(stbi__context *s, int *x, int *y, int *comp); +#endif + +#ifndef STBI_NO_PSD +static int stbi__psd_test(stbi__context *s); +static void *stbi__psd_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri, int bpc); +static int stbi__psd_info(stbi__context *s, int *x, int *y, int *comp); +#endif + +#ifndef STBI_NO_HDR +static int stbi__hdr_test(stbi__context *s); +static float *stbi__hdr_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri); +static int stbi__hdr_info(stbi__context *s, int *x, int *y, int *comp); +#endif + +#ifndef STBI_NO_PIC +static int stbi__pic_test(stbi__context *s); +static void *stbi__pic_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri); +static int stbi__pic_info(stbi__context *s, int *x, int *y, int *comp); +#endif + +#ifndef STBI_NO_GIF +static int stbi__gif_test(stbi__context *s); +static void *stbi__gif_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri); +static int stbi__gif_info(stbi__context *s, int *x, int *y, int *comp); +#endif + +#ifndef STBI_NO_PNM +static int stbi__pnm_test(stbi__context *s); +static void *stbi__pnm_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri); +static int stbi__pnm_info(stbi__context *s, int *x, int *y, int *comp); +#endif + +// this is not threadsafe +static const char *stbi__g_failure_reason; + +STBIDEF const char *stbi_failure_reason(void) +{ + return stbi__g_failure_reason; +} + +static int stbi__err(const char *str) +{ + stbi__g_failure_reason = str; + return 0; +} + +static void *stbi__malloc(size_t size) +{ + return STBI_MALLOC(size); +} + +// stb_image uses ints pervasively, including for offset calculations. +// therefore the largest decoded image size we can support with the +// current code, even on 64-bit targets, is INT_MAX. this is not a +// significant limitation for the intended use case. +// +// we do, however, need to make sure our size calculations don't +// overflow. hence a few helper functions for size calculations that +// multiply integers together, making sure that they're non-negative +// and no overflow occurs. + +// return 1 if the sum is valid, 0 on overflow. +// negative terms are considered invalid. +static int stbi__addsizes_valid(int a, int b) +{ + if (b < 0) return 0; + // now 0 <= b <= INT_MAX, hence also + // 0 <= INT_MAX - b <= INTMAX. + // And "a + b <= INT_MAX" (which might overflow) is the + // same as a <= INT_MAX - b (no overflow) + return a <= INT_MAX - b; +} + +// returns 1 if the product is valid, 0 on overflow. +// negative factors are considered invalid. +static int stbi__mul2sizes_valid(int a, int b) +{ + if (a < 0 || b < 0) return 0; + if (b == 0) return 1; // mul-by-0 is always safe + // portable way to check for no overflows in a*b + return a <= INT_MAX/b; +} + +// returns 1 if "a*b + add" has no negative terms/factors and doesn't overflow +static int stbi__mad2sizes_valid(int a, int b, int add) +{ + return stbi__mul2sizes_valid(a, b) && stbi__addsizes_valid(a*b, add); +} + +// returns 1 if "a*b*c + add" has no negative terms/factors and doesn't overflow +static int stbi__mad3sizes_valid(int a, int b, int c, int add) +{ + return stbi__mul2sizes_valid(a, b) && stbi__mul2sizes_valid(a*b, c) && + stbi__addsizes_valid(a*b*c, add); +} + +// returns 1 if "a*b*c*d + add" has no negative terms/factors and doesn't overflow +static int stbi__mad4sizes_valid(int a, int b, int c, int d, int add) +{ + return stbi__mul2sizes_valid(a, b) && stbi__mul2sizes_valid(a*b, c) && + stbi__mul2sizes_valid(a*b*c, d) && stbi__addsizes_valid(a*b*c*d, add); +} + +// mallocs with size overflow checking +static void *stbi__malloc_mad2(int a, int b, int add) +{ + if (!stbi__mad2sizes_valid(a, b, add)) return NULL; + return stbi__malloc(a*b + add); +} + +static void *stbi__malloc_mad3(int a, int b, int c, int add) +{ + if (!stbi__mad3sizes_valid(a, b, c, add)) return NULL; + return stbi__malloc(a*b*c + add); +} + +static void *stbi__malloc_mad4(int a, int b, int c, int d, int add) +{ + if (!stbi__mad4sizes_valid(a, b, c, d, add)) return NULL; + return stbi__malloc(a*b*c*d + add); +} + +// stbi__err - error +// stbi__errpf - error returning pointer to float +// stbi__errpuc - error returning pointer to unsigned char + +#ifdef STBI_NO_FAILURE_STRINGS + #define stbi__err(x,y) 0 +#elif defined(STBI_FAILURE_USERMSG) + #define stbi__err(x,y) stbi__err(y) +#else + #define stbi__err(x,y) stbi__err(x) +#endif + +#define stbi__errpf(x,y) ((float *)(size_t) (stbi__err(x,y)?NULL:NULL)) +#define stbi__errpuc(x,y) ((unsigned char *)(size_t) (stbi__err(x,y)?NULL:NULL)) + +STBIDEF void stbi_image_free(void *retval_from_stbi_load) +{ + STBI_FREE(retval_from_stbi_load); +} + +#ifndef STBI_NO_LINEAR +static float *stbi__ldr_to_hdr(stbi_uc *data, int x, int y, int comp); +#endif + +#ifndef STBI_NO_HDR +static stbi_uc *stbi__hdr_to_ldr(float *data, int x, int y, int comp); +#endif + +static int stbi__vertically_flip_on_load = 0; + +STBIDEF void stbi_set_flip_vertically_on_load(int flag_true_if_should_flip) +{ + stbi__vertically_flip_on_load = flag_true_if_should_flip; +} + +static void *stbi__load_main(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri, int bpc) +{ + memset(ri, 0, sizeof(*ri)); // make sure it's initialized if we add new fields + ri->bits_per_channel = 8; // default is 8 so most paths don't have to be changed + ri->channel_order = STBI_ORDER_RGB; // all current input & output are this, but this is here so we can add BGR order + ri->num_channels = 0; + + #ifndef STBI_NO_JPEG + if (stbi__jpeg_test(s)) return stbi__jpeg_load(s,x,y,comp,req_comp, ri); + #endif + #ifndef STBI_NO_PNG + if (stbi__png_test(s)) return stbi__png_load(s,x,y,comp,req_comp, ri); + #endif + #ifndef STBI_NO_BMP + if (stbi__bmp_test(s)) return stbi__bmp_load(s,x,y,comp,req_comp, ri); + #endif + #ifndef STBI_NO_GIF + if (stbi__gif_test(s)) return stbi__gif_load(s,x,y,comp,req_comp, ri); + #endif + #ifndef STBI_NO_PSD + if (stbi__psd_test(s)) return stbi__psd_load(s,x,y,comp,req_comp, ri, bpc); + #endif + #ifndef STBI_NO_PIC + if (stbi__pic_test(s)) return stbi__pic_load(s,x,y,comp,req_comp, ri); + #endif + #ifndef STBI_NO_PNM + if (stbi__pnm_test(s)) return stbi__pnm_load(s,x,y,comp,req_comp, ri); + #endif + + #ifndef STBI_NO_HDR + if (stbi__hdr_test(s)) { + float *hdr = stbi__hdr_load(s, x,y,comp,req_comp, ri); + return stbi__hdr_to_ldr(hdr, *x, *y, req_comp ? req_comp : *comp); + } + #endif + + #ifndef STBI_NO_TGA + // test tga last because it's a crappy test! + if (stbi__tga_test(s)) + return stbi__tga_load(s,x,y,comp,req_comp, ri); + #endif + + return stbi__errpuc("unknown image type", "Image not of any known type, or corrupt"); +} + +static stbi_uc *stbi__convert_16_to_8(stbi__uint16 *orig, int w, int h, int channels) +{ + int i; + int img_len = w * h * channels; + stbi_uc *reduced; + + reduced = (stbi_uc *) stbi__malloc(img_len); + if (reduced == NULL) return stbi__errpuc("outofmem", "Out of memory"); + + for (i = 0; i < img_len; ++i) + reduced[i] = (stbi_uc)((orig[i] >> 8) & 0xFF); // top half of each byte is sufficient approx of 16->8 bit scaling + + STBI_FREE(orig); + return reduced; +} + +static stbi__uint16 *stbi__convert_8_to_16(stbi_uc *orig, int w, int h, int channels) +{ + int i; + int img_len = w * h * channels; + stbi__uint16 *enlarged; + + enlarged = (stbi__uint16 *) stbi__malloc(img_len*2); + if (enlarged == NULL) return (stbi__uint16 *) stbi__errpuc("outofmem", "Out of memory"); + + for (i = 0; i < img_len; ++i) + enlarged[i] = (stbi__uint16)((orig[i] << 8) + orig[i]); // replicate to high and low byte, maps 0->0, 255->0xffff + + STBI_FREE(orig); + return enlarged; +} + +static void stbi__vertical_flip(void *image, int w, int h, int bytes_per_pixel) +{ + int row; + size_t bytes_per_row = (size_t)w * bytes_per_pixel; + stbi_uc temp[2048]; + stbi_uc *bytes = (stbi_uc *)image; + + for (row = 0; row < (h>>1); row++) { + stbi_uc *row0 = bytes + row*bytes_per_row; + stbi_uc *row1 = bytes + (h - row - 1)*bytes_per_row; + // swap row0 with row1 + size_t bytes_left = bytes_per_row; + while (bytes_left) { + size_t bytes_copy = (bytes_left < sizeof(temp)) ? bytes_left : sizeof(temp); + memcpy(temp, row0, bytes_copy); + memcpy(row0, row1, bytes_copy); + memcpy(row1, temp, bytes_copy); + row0 += bytes_copy; + row1 += bytes_copy; + bytes_left -= bytes_copy; + } + } +} + +static unsigned char *stbi__load_and_postprocess_8bit(stbi__context *s, int *x, int *y, int *comp, int req_comp) +{ + stbi__result_info ri; + void *result = stbi__load_main(s, x, y, comp, req_comp, &ri, 8); + + if (result == NULL) + return NULL; + + if (ri.bits_per_channel != 8) { + STBI_ASSERT(ri.bits_per_channel == 16); + result = stbi__convert_16_to_8((stbi__uint16 *) result, *x, *y, req_comp == 0 ? *comp : req_comp); + ri.bits_per_channel = 8; + } + + // @TODO: move stbi__convert_format to here + + if (stbi__vertically_flip_on_load) { + int channels = req_comp ? req_comp : *comp; + stbi__vertical_flip(result, *x, *y, channels * sizeof(stbi_uc)); + } + + return (unsigned char *) result; +} + +static stbi__uint16 *stbi__load_and_postprocess_16bit(stbi__context *s, int *x, int *y, int *comp, int req_comp) +{ + stbi__result_info ri; + void *result = stbi__load_main(s, x, y, comp, req_comp, &ri, 16); + + if (result == NULL) + return NULL; + + if (ri.bits_per_channel != 16) { + STBI_ASSERT(ri.bits_per_channel == 8); + result = stbi__convert_8_to_16((stbi_uc *) result, *x, *y, req_comp == 0 ? *comp : req_comp); + ri.bits_per_channel = 16; + } + + // @TODO: move stbi__convert_format16 to here + // @TODO: special case RGB-to-Y (and RGBA-to-YA) for 8-bit-to-16-bit case to keep more precision + + if (stbi__vertically_flip_on_load) { + int channels = req_comp ? req_comp : *comp; + stbi__vertical_flip(result, *x, *y, channels * sizeof(stbi__uint16)); + } + + return (stbi__uint16 *) result; +} + +#ifndef STBI_NO_HDR +static void stbi__float_postprocess(float *result, int *x, int *y, int *comp, int req_comp) +{ + if (stbi__vertically_flip_on_load && result != NULL) { + int channels = req_comp ? req_comp : *comp; + stbi__vertical_flip(result, *x, *y, channels * sizeof(float)); + } +} +#endif + +#ifndef STBI_NO_STDIO + +static FILE *stbi__fopen(char const *filename, char const *mode) +{ + FILE *f; +#if defined(_MSC_VER) && _MSC_VER >= 1400 + if (0 != fopen_s(&f, filename, mode)) + f=0; +#else + f = fopen(filename, mode); +#endif + return f; +} + + +STBIDEF stbi_uc *stbi_load(char const *filename, int *x, int *y, int *comp, int req_comp) +{ + FILE *f = stbi__fopen(filename, "rb"); + unsigned char *result; + if (!f) return stbi__errpuc("can't fopen", "Unable to open file"); + result = stbi_load_from_file(f,x,y,comp,req_comp); + fclose(f); + return result; +} + +STBIDEF stbi_uc *stbi_load_from_file(FILE *f, int *x, int *y, int *comp, int req_comp) +{ + unsigned char *result; + stbi__context s; + stbi__start_file(&s,f); + result = stbi__load_and_postprocess_8bit(&s,x,y,comp,req_comp); + if (result) { + // need to 'unget' all the characters in the IO buffer + fseek(f, - (int) (s.img_buffer_end - s.img_buffer), SEEK_CUR); + } + return result; +} + +STBIDEF stbi__uint16 *stbi_load_from_file_16(FILE *f, int *x, int *y, int *comp, int req_comp) +{ + stbi__uint16 *result; + stbi__context s; + stbi__start_file(&s,f); + result = stbi__load_and_postprocess_16bit(&s,x,y,comp,req_comp); + if (result) { + // need to 'unget' all the characters in the IO buffer + fseek(f, - (int) (s.img_buffer_end - s.img_buffer), SEEK_CUR); + } + return result; +} + +STBIDEF stbi_us *stbi_load_16(char const *filename, int *x, int *y, int *comp, int req_comp) +{ + FILE *f = stbi__fopen(filename, "rb"); + stbi__uint16 *result; + if (!f) return (stbi_us *) stbi__errpuc("can't fopen", "Unable to open file"); + result = stbi_load_from_file_16(f,x,y,comp,req_comp); + fclose(f); + return result; +} + + +#endif //!STBI_NO_STDIO + +STBIDEF stbi_us *stbi_load_16_from_memory(stbi_uc const *buffer, int len, int *x, int *y, int *channels_in_file, int desired_channels) +{ + stbi__context s; + stbi__start_mem(&s,buffer,len); + return stbi__load_and_postprocess_16bit(&s,x,y,channels_in_file,desired_channels); +} + +STBIDEF stbi_us *stbi_load_16_from_callbacks(stbi_io_callbacks const *clbk, void *user, int *x, int *y, int *channels_in_file, int desired_channels) +{ + stbi__context s; + stbi__start_callbacks(&s, (stbi_io_callbacks *)clbk, user); + return stbi__load_and_postprocess_16bit(&s,x,y,channels_in_file,desired_channels); +} + +STBIDEF stbi_uc *stbi_load_from_memory(stbi_uc const *buffer, int len, int *x, int *y, int *comp, int req_comp) +{ + stbi__context s; + stbi__start_mem(&s,buffer,len); + return stbi__load_and_postprocess_8bit(&s,x,y,comp,req_comp); +} + +STBIDEF stbi_uc *stbi_load_from_callbacks(stbi_io_callbacks const *clbk, void *user, int *x, int *y, int *comp, int req_comp) +{ + stbi__context s; + stbi__start_callbacks(&s, (stbi_io_callbacks *) clbk, user); + return stbi__load_and_postprocess_8bit(&s,x,y,comp,req_comp); +} + +#ifndef STBI_NO_LINEAR +static float *stbi__loadf_main(stbi__context *s, int *x, int *y, int *comp, int req_comp) +{ + unsigned char *data; + #ifndef STBI_NO_HDR + if (stbi__hdr_test(s)) { + stbi__result_info ri; + float *hdr_data = stbi__hdr_load(s,x,y,comp,req_comp, &ri); + if (hdr_data) + stbi__float_postprocess(hdr_data,x,y,comp,req_comp); + return hdr_data; + } + #endif + data = stbi__load_and_postprocess_8bit(s, x, y, comp, req_comp); + if (data) + return stbi__ldr_to_hdr(data, *x, *y, req_comp ? req_comp : *comp); + return stbi__errpf("unknown image type", "Image not of any known type, or corrupt"); +} + +STBIDEF float *stbi_loadf_from_memory(stbi_uc const *buffer, int len, int *x, int *y, int *comp, int req_comp) +{ + stbi__context s; + stbi__start_mem(&s,buffer,len); + return stbi__loadf_main(&s,x,y,comp,req_comp); +} + +STBIDEF float *stbi_loadf_from_callbacks(stbi_io_callbacks const *clbk, void *user, int *x, int *y, int *comp, int req_comp) +{ + stbi__context s; + stbi__start_callbacks(&s, (stbi_io_callbacks *) clbk, user); + return stbi__loadf_main(&s,x,y,comp,req_comp); +} + +#ifndef STBI_NO_STDIO +STBIDEF float *stbi_loadf(char const *filename, int *x, int *y, int *comp, int req_comp) +{ + float *result; + FILE *f = stbi__fopen(filename, "rb"); + if (!f) return stbi__errpf("can't fopen", "Unable to open file"); + result = stbi_loadf_from_file(f,x,y,comp,req_comp); + fclose(f); + return result; +} + +STBIDEF float *stbi_loadf_from_file(FILE *f, int *x, int *y, int *comp, int req_comp) +{ + stbi__context s; + stbi__start_file(&s,f); + return stbi__loadf_main(&s,x,y,comp,req_comp); +} +#endif // !STBI_NO_STDIO + +#endif // !STBI_NO_LINEAR + +// these is-hdr-or-not is defined independent of whether STBI_NO_LINEAR is +// defined, for API simplicity; if STBI_NO_LINEAR is defined, it always +// reports false! + +STBIDEF int stbi_is_hdr_from_memory(stbi_uc const *buffer, int len) +{ + #ifndef STBI_NO_HDR + stbi__context s; + stbi__start_mem(&s,buffer,len); + return stbi__hdr_test(&s); + #else + STBI_NOTUSED(buffer); + STBI_NOTUSED(len); + return 0; + #endif +} + +#ifndef STBI_NO_STDIO +STBIDEF int stbi_is_hdr (char const *filename) +{ + FILE *f = stbi__fopen(filename, "rb"); + int result=0; + if (f) { + result = stbi_is_hdr_from_file(f); + fclose(f); + } + return result; +} + +STBIDEF int stbi_is_hdr_from_file(FILE *f) +{ + #ifndef STBI_NO_HDR + stbi__context s; + stbi__start_file(&s,f); + return stbi__hdr_test(&s); + #else + STBI_NOTUSED(f); + return 0; + #endif +} +#endif // !STBI_NO_STDIO + +STBIDEF int stbi_is_hdr_from_callbacks(stbi_io_callbacks const *clbk, void *user) +{ + #ifndef STBI_NO_HDR + stbi__context s; + stbi__start_callbacks(&s, (stbi_io_callbacks *) clbk, user); + return stbi__hdr_test(&s); + #else + STBI_NOTUSED(clbk); + STBI_NOTUSED(user); + return 0; + #endif +} + +#ifndef STBI_NO_LINEAR +static float stbi__l2h_gamma=2.2f, stbi__l2h_scale=1.0f; + +STBIDEF void stbi_ldr_to_hdr_gamma(float gamma) { stbi__l2h_gamma = gamma; } +STBIDEF void stbi_ldr_to_hdr_scale(float scale) { stbi__l2h_scale = scale; } +#endif + +static float stbi__h2l_gamma_i=1.0f/2.2f, stbi__h2l_scale_i=1.0f; + +STBIDEF void stbi_hdr_to_ldr_gamma(float gamma) { stbi__h2l_gamma_i = 1/gamma; } +STBIDEF void stbi_hdr_to_ldr_scale(float scale) { stbi__h2l_scale_i = 1/scale; } + + +////////////////////////////////////////////////////////////////////////////// +// +// Common code used by all image loaders +// + +enum +{ + STBI__SCAN_load=0, + STBI__SCAN_type, + STBI__SCAN_header +}; + +static void stbi__refill_buffer(stbi__context *s) +{ + int n = (s->io.read)(s->io_user_data,(char*)s->buffer_start,s->buflen); + if (n == 0) { + // at end of file, treat same as if from memory, but need to handle case + // where s->img_buffer isn't pointing to safe memory, e.g. 0-byte file + s->read_from_callbacks = 0; + s->img_buffer = s->buffer_start; + s->img_buffer_end = s->buffer_start+1; + *s->img_buffer = 0; + } else { + s->img_buffer = s->buffer_start; + s->img_buffer_end = s->buffer_start + n; + } +} + +stbi_inline static stbi_uc stbi__get8(stbi__context *s) +{ + if (s->img_buffer < s->img_buffer_end) + return *s->img_buffer++; + if (s->read_from_callbacks) { + stbi__refill_buffer(s); + return *s->img_buffer++; + } + return 0; +} + +stbi_inline static int stbi__at_eof(stbi__context *s) +{ + if (s->io.read) { + if (!(s->io.eof)(s->io_user_data)) return 0; + // if feof() is true, check if buffer = end + // special case: we've only got the special 0 character at the end + if (s->read_from_callbacks == 0) return 1; + } + + return s->img_buffer >= s->img_buffer_end; +} + +static void stbi__skip(stbi__context *s, int n) +{ + if (n < 0) { + s->img_buffer = s->img_buffer_end; + return; + } + if (s->io.read) { + int blen = (int) (s->img_buffer_end - s->img_buffer); + if (blen < n) { + s->img_buffer = s->img_buffer_end; + (s->io.skip)(s->io_user_data, n - blen); + return; + } + } + s->img_buffer += n; +} + +static int stbi__getn(stbi__context *s, stbi_uc *buffer, int n) +{ + if (s->io.read) { + int blen = (int) (s->img_buffer_end - s->img_buffer); + if (blen < n) { + int res, count; + + memcpy(buffer, s->img_buffer, blen); + + count = (s->io.read)(s->io_user_data, (char*) buffer + blen, n - blen); + res = (count == (n-blen)); + s->img_buffer = s->img_buffer_end; + return res; + } + } + + if (s->img_buffer+n <= s->img_buffer_end) { + memcpy(buffer, s->img_buffer, n); + s->img_buffer += n; + return 1; + } else + return 0; +} + +static int stbi__get16be(stbi__context *s) +{ + int z = stbi__get8(s); + return (z << 8) + stbi__get8(s); +} + +static stbi__uint32 stbi__get32be(stbi__context *s) +{ + stbi__uint32 z = stbi__get16be(s); + return (z << 16) + stbi__get16be(s); +} + +#if defined(STBI_NO_BMP) && defined(STBI_NO_TGA) && defined(STBI_NO_GIF) +// nothing +#else +static int stbi__get16le(stbi__context *s) +{ + int z = stbi__get8(s); + return z + (stbi__get8(s) << 8); +} +#endif + +#ifndef STBI_NO_BMP +static stbi__uint32 stbi__get32le(stbi__context *s) +{ + stbi__uint32 z = stbi__get16le(s); + return z + (stbi__get16le(s) << 16); +} +#endif + +#define STBI__BYTECAST(x) ((stbi_uc) ((x) & 255)) // truncate int to byte without warnings + + +////////////////////////////////////////////////////////////////////////////// +// +// generic converter from built-in img_n to req_comp +// individual types do this automatically as much as possible (e.g. jpeg +// does all cases internally since it needs to colorspace convert anyway, +// and it never has alpha, so very few cases ). png can automatically +// interleave an alpha=255 channel, but falls back to this for other cases +// +// assume data buffer is malloced, so malloc a new one and free that one +// only failure mode is malloc failing + +static stbi_uc stbi__compute_y(int r, int g, int b) +{ + return (stbi_uc) (((r*77) + (g*150) + (29*b)) >> 8); +} + +static unsigned char *stbi__convert_format(unsigned char *data, int img_n, int req_comp, unsigned int x, unsigned int y) +{ + int i,j; + unsigned char *good; + + if (req_comp == img_n) return data; + STBI_ASSERT(req_comp >= 1 && req_comp <= 4); + + good = (unsigned char *) stbi__malloc_mad3(req_comp, x, y, 0); + if (good == NULL) { + STBI_FREE(data); + return stbi__errpuc("outofmem", "Out of memory"); + } + + for (j=0; j < (int) y; ++j) { + unsigned char *src = data + j * x * img_n ; + unsigned char *dest = good + j * x * req_comp; + + #define STBI__COMBO(a,b) ((a)*8+(b)) + #define STBI__CASE(a,b) case STBI__COMBO(a,b): for(i=x-1; i >= 0; --i, src += a, dest += b) + // convert source image with img_n components to one with req_comp components; + // avoid switch per pixel, so use switch per scanline and massive macros + switch (STBI__COMBO(img_n, req_comp)) { + STBI__CASE(1,2) { dest[0]=src[0], dest[1]=255; } break; + STBI__CASE(1,3) { dest[0]=dest[1]=dest[2]=src[0]; } break; + STBI__CASE(1,4) { dest[0]=dest[1]=dest[2]=src[0], dest[3]=255; } break; + STBI__CASE(2,1) { dest[0]=src[0]; } break; + STBI__CASE(2,3) { dest[0]=dest[1]=dest[2]=src[0]; } break; + STBI__CASE(2,4) { dest[0]=dest[1]=dest[2]=src[0], dest[3]=src[1]; } break; + STBI__CASE(3,4) { dest[0]=src[0],dest[1]=src[1],dest[2]=src[2],dest[3]=255; } break; + STBI__CASE(3,1) { dest[0]=stbi__compute_y(src[0],src[1],src[2]); } break; + STBI__CASE(3,2) { dest[0]=stbi__compute_y(src[0],src[1],src[2]), dest[1] = 255; } break; + STBI__CASE(4,1) { dest[0]=stbi__compute_y(src[0],src[1],src[2]); } break; + STBI__CASE(4,2) { dest[0]=stbi__compute_y(src[0],src[1],src[2]), dest[1] = src[3]; } break; + STBI__CASE(4,3) { dest[0]=src[0],dest[1]=src[1],dest[2]=src[2]; } break; + default: STBI_ASSERT(0); + } + #undef STBI__CASE + } + + STBI_FREE(data); + return good; +} + +static stbi__uint16 stbi__compute_y_16(int r, int g, int b) +{ + return (stbi__uint16) (((r*77) + (g*150) + (29*b)) >> 8); +} + +static stbi__uint16 *stbi__convert_format16(stbi__uint16 *data, int img_n, int req_comp, unsigned int x, unsigned int y) +{ + int i,j; + stbi__uint16 *good; + + if (req_comp == img_n) return data; + STBI_ASSERT(req_comp >= 1 && req_comp <= 4); + + good = (stbi__uint16 *) stbi__malloc(req_comp * x * y * 2); + if (good == NULL) { + STBI_FREE(data); + return (stbi__uint16 *) stbi__errpuc("outofmem", "Out of memory"); + } + + for (j=0; j < (int) y; ++j) { + stbi__uint16 *src = data + j * x * img_n ; + stbi__uint16 *dest = good + j * x * req_comp; + + #define STBI__COMBO(a,b) ((a)*8+(b)) + #define STBI__CASE(a,b) case STBI__COMBO(a,b): for(i=x-1; i >= 0; --i, src += a, dest += b) + // convert source image with img_n components to one with req_comp components; + // avoid switch per pixel, so use switch per scanline and massive macros + switch (STBI__COMBO(img_n, req_comp)) { + STBI__CASE(1,2) { dest[0]=src[0], dest[1]=0xffff; } break; + STBI__CASE(1,3) { dest[0]=dest[1]=dest[2]=src[0]; } break; + STBI__CASE(1,4) { dest[0]=dest[1]=dest[2]=src[0], dest[3]=0xffff; } break; + STBI__CASE(2,1) { dest[0]=src[0]; } break; + STBI__CASE(2,3) { dest[0]=dest[1]=dest[2]=src[0]; } break; + STBI__CASE(2,4) { dest[0]=dest[1]=dest[2]=src[0], dest[3]=src[1]; } break; + STBI__CASE(3,4) { dest[0]=src[0],dest[1]=src[1],dest[2]=src[2],dest[3]=0xffff; } break; + STBI__CASE(3,1) { dest[0]=stbi__compute_y_16(src[0],src[1],src[2]); } break; + STBI__CASE(3,2) { dest[0]=stbi__compute_y_16(src[0],src[1],src[2]), dest[1] = 0xffff; } break; + STBI__CASE(4,1) { dest[0]=stbi__compute_y_16(src[0],src[1],src[2]); } break; + STBI__CASE(4,2) { dest[0]=stbi__compute_y_16(src[0],src[1],src[2]), dest[1] = src[3]; } break; + STBI__CASE(4,3) { dest[0]=src[0],dest[1]=src[1],dest[2]=src[2]; } break; + default: STBI_ASSERT(0); + } + #undef STBI__CASE + } + + STBI_FREE(data); + return good; +} + +#ifndef STBI_NO_LINEAR +static float *stbi__ldr_to_hdr(stbi_uc *data, int x, int y, int comp) +{ + int i,k,n; + float *output; + if (!data) return NULL; + output = (float *) stbi__malloc_mad4(x, y, comp, sizeof(float), 0); + if (output == NULL) { STBI_FREE(data); return stbi__errpf("outofmem", "Out of memory"); } + // compute number of non-alpha components + if (comp & 1) n = comp; else n = comp-1; + for (i=0; i < x*y; ++i) { + for (k=0; k < n; ++k) { + output[i*comp + k] = (float) (pow(data[i*comp+k]/255.0f, stbi__l2h_gamma) * stbi__l2h_scale); + } + if (k < comp) output[i*comp + k] = data[i*comp+k]/255.0f; + } + STBI_FREE(data); + return output; +} +#endif + +#ifndef STBI_NO_HDR +#define stbi__float2int(x) ((int) (x)) +static stbi_uc *stbi__hdr_to_ldr(float *data, int x, int y, int comp) +{ + int i,k,n; + stbi_uc *output; + if (!data) return NULL; + output = (stbi_uc *) stbi__malloc_mad3(x, y, comp, 0); + if (output == NULL) { STBI_FREE(data); return stbi__errpuc("outofmem", "Out of memory"); } + // compute number of non-alpha components + if (comp & 1) n = comp; else n = comp-1; + for (i=0; i < x*y; ++i) { + for (k=0; k < n; ++k) { + float z = (float) pow(data[i*comp+k]*stbi__h2l_scale_i, stbi__h2l_gamma_i) * 255 + 0.5f; + if (z < 0) z = 0; + if (z > 255) z = 255; + output[i*comp + k] = (stbi_uc) stbi__float2int(z); + } + if (k < comp) { + float z = data[i*comp+k] * 255 + 0.5f; + if (z < 0) z = 0; + if (z > 255) z = 255; + output[i*comp + k] = (stbi_uc) stbi__float2int(z); + } + } + STBI_FREE(data); + return output; +} +#endif + +////////////////////////////////////////////////////////////////////////////// +// +// "baseline" JPEG/JFIF decoder +// +// simple implementation +// - doesn't support delayed output of y-dimension +// - simple interface (only one output format: 8-bit interleaved RGB) +// - doesn't try to recover corrupt jpegs +// - doesn't allow partial loading, loading multiple at once +// - still fast on x86 (copying globals into locals doesn't help x86) +// - allocates lots of intermediate memory (full size of all components) +// - non-interleaved case requires this anyway +// - allows good upsampling (see next) +// high-quality +// - upsampled channels are bilinearly interpolated, even across blocks +// - quality integer IDCT derived from IJG's 'slow' +// performance +// - fast huffman; reasonable integer IDCT +// - some SIMD kernels for common paths on targets with SSE2/NEON +// - uses a lot of intermediate memory, could cache poorly + +#ifndef STBI_NO_JPEG + +// huffman decoding acceleration +#define FAST_BITS 9 // larger handles more cases; smaller stomps less cache + +typedef struct +{ + stbi_uc fast[1 << FAST_BITS]; + // weirdly, repacking this into AoS is a 10% speed loss, instead of a win + stbi__uint16 code[256]; + stbi_uc values[256]; + stbi_uc size[257]; + unsigned int maxcode[18]; + int delta[17]; // old 'firstsymbol' - old 'firstcode' +} stbi__huffman; + +typedef struct +{ + stbi__context *s; + stbi__huffman huff_dc[4]; + stbi__huffman huff_ac[4]; + stbi__uint16 dequant[4][64]; + stbi__int16 fast_ac[4][1 << FAST_BITS]; + +// sizes for components, interleaved MCUs + int img_h_max, img_v_max; + int img_mcu_x, img_mcu_y; + int img_mcu_w, img_mcu_h; + +// definition of jpeg image component + struct + { + int id; + int h,v; + int tq; + int hd,ha; + int dc_pred; + + int x,y,w2,h2; + stbi_uc *data; + void *raw_data, *raw_coeff; + stbi_uc *linebuf; + short *coeff; // progressive only + int coeff_w, coeff_h; // number of 8x8 coefficient blocks + } img_comp[4]; + + stbi__uint32 code_buffer; // jpeg entropy-coded buffer + int code_bits; // number of valid bits + unsigned char marker; // marker seen while filling entropy buffer + int nomore; // flag if we saw a marker so must stop + + int progressive; + int spec_start; + int spec_end; + int succ_high; + int succ_low; + int eob_run; + int jfif; + int app14_color_transform; // Adobe APP14 tag + int rgb; + + int scan_n, order[4]; + int restart_interval, todo; + +// kernels + void (*idct_block_kernel)(stbi_uc *out, int out_stride, short data[64]); + void (*YCbCr_to_RGB_kernel)(stbi_uc *out, const stbi_uc *y, const stbi_uc *pcb, const stbi_uc *pcr, int count, int step); + stbi_uc *(*resample_row_hv_2_kernel)(stbi_uc *out, stbi_uc *in_near, stbi_uc *in_far, int w, int hs); +} stbi__jpeg; + +static int stbi__build_huffman(stbi__huffman *h, int *count) +{ + int i,j,k=0,code; + // build size list for each symbol (from JPEG spec) + for (i=0; i < 16; ++i) + for (j=0; j < count[i]; ++j) + h->size[k++] = (stbi_uc) (i+1); + h->size[k] = 0; + + // compute actual symbols (from jpeg spec) + code = 0; + k = 0; + for(j=1; j <= 16; ++j) { + // compute delta to add to code to compute symbol id + h->delta[j] = k - code; + if (h->size[k] == j) { + while (h->size[k] == j) + h->code[k++] = (stbi__uint16) (code++); + if (code-1 >= (1 << j)) return stbi__err("bad code lengths","Corrupt JPEG"); + } + // compute largest code + 1 for this size, preshifted as needed later + h->maxcode[j] = code << (16-j); + code <<= 1; + } + h->maxcode[j] = 0xffffffff; + + // build non-spec acceleration table; 255 is flag for not-accelerated + memset(h->fast, 255, 1 << FAST_BITS); + for (i=0; i < k; ++i) { + int s = h->size[i]; + if (s <= FAST_BITS) { + int c = h->code[i] << (FAST_BITS-s); + int m = 1 << (FAST_BITS-s); + for (j=0; j < m; ++j) { + h->fast[c+j] = (stbi_uc) i; + } + } + } + return 1; +} + +// build a table that decodes both magnitude and value of small ACs in +// one go. +static void stbi__build_fast_ac(stbi__int16 *fast_ac, stbi__huffman *h) +{ + int i; + for (i=0; i < (1 << FAST_BITS); ++i) { + stbi_uc fast = h->fast[i]; + fast_ac[i] = 0; + if (fast < 255) { + int rs = h->values[fast]; + int run = (rs >> 4) & 15; + int magbits = rs & 15; + int len = h->size[fast]; + + if (magbits && len + magbits <= FAST_BITS) { + // magnitude code followed by receive_extend code + int k = ((i << len) & ((1 << FAST_BITS) - 1)) >> (FAST_BITS - magbits); + int m = 1 << (magbits - 1); + if (k < m) k += (~0U << magbits) + 1; + // if the result is small enough, we can fit it in fast_ac table + if (k >= -128 && k <= 127) + fast_ac[i] = (stbi__int16) ((k << 8) + (run << 4) + (len + magbits)); + } + } + } +} + +static void stbi__grow_buffer_unsafe(stbi__jpeg *j) +{ + do { + int b = j->nomore ? 0 : stbi__get8(j->s); + if (b == 0xff) { + int c = stbi__get8(j->s); + while (c == 0xff) c = stbi__get8(j->s); // consume fill bytes + if (c != 0) { + j->marker = (unsigned char) c; + j->nomore = 1; + return; + } + } + j->code_buffer |= b << (24 - j->code_bits); + j->code_bits += 8; + } while (j->code_bits <= 24); +} + +// (1 << n) - 1 +static stbi__uint32 stbi__bmask[17]={0,1,3,7,15,31,63,127,255,511,1023,2047,4095,8191,16383,32767,65535}; + +// decode a jpeg huffman value from the bitstream +stbi_inline static int stbi__jpeg_huff_decode(stbi__jpeg *j, stbi__huffman *h) +{ + unsigned int temp; + int c,k; + + if (j->code_bits < 16) stbi__grow_buffer_unsafe(j); + + // look at the top FAST_BITS and determine what symbol ID it is, + // if the code is <= FAST_BITS + c = (j->code_buffer >> (32 - FAST_BITS)) & ((1 << FAST_BITS)-1); + k = h->fast[c]; + if (k < 255) { + int s = h->size[k]; + if (s > j->code_bits) + return -1; + j->code_buffer <<= s; + j->code_bits -= s; + return h->values[k]; + } + + // naive test is to shift the code_buffer down so k bits are + // valid, then test against maxcode. To speed this up, we've + // preshifted maxcode left so that it has (16-k) 0s at the + // end; in other words, regardless of the number of bits, it + // wants to be compared against something shifted to have 16; + // that way we don't need to shift inside the loop. + temp = j->code_buffer >> 16; + for (k=FAST_BITS+1 ; ; ++k) + if (temp < h->maxcode[k]) + break; + if (k == 17) { + // error! code not found + j->code_bits -= 16; + return -1; + } + + if (k > j->code_bits) + return -1; + + // convert the huffman code to the symbol id + c = ((j->code_buffer >> (32 - k)) & stbi__bmask[k]) + h->delta[k]; + STBI_ASSERT((((j->code_buffer) >> (32 - h->size[c])) & stbi__bmask[h->size[c]]) == h->code[c]); + + // convert the id to a symbol + j->code_bits -= k; + j->code_buffer <<= k; + return h->values[c]; +} + +// bias[n] = (-1<code_bits < n) stbi__grow_buffer_unsafe(j); + + sgn = (stbi__int32)j->code_buffer >> 31; // sign bit is always in MSB + k = stbi_lrot(j->code_buffer, n); + STBI_ASSERT(n >= 0 && n < (int) (sizeof(stbi__bmask)/sizeof(*stbi__bmask))); + j->code_buffer = k & ~stbi__bmask[n]; + k &= stbi__bmask[n]; + j->code_bits -= n; + return k + (stbi__jbias[n] & ~sgn); +} + +// get some unsigned bits +stbi_inline static int stbi__jpeg_get_bits(stbi__jpeg *j, int n) +{ + unsigned int k; + if (j->code_bits < n) stbi__grow_buffer_unsafe(j); + k = stbi_lrot(j->code_buffer, n); + j->code_buffer = k & ~stbi__bmask[n]; + k &= stbi__bmask[n]; + j->code_bits -= n; + return k; +} + +stbi_inline static int stbi__jpeg_get_bit(stbi__jpeg *j) +{ + unsigned int k; + if (j->code_bits < 1) stbi__grow_buffer_unsafe(j); + k = j->code_buffer; + j->code_buffer <<= 1; + --j->code_bits; + return k & 0x80000000; +} + +// given a value that's at position X in the zigzag stream, +// where does it appear in the 8x8 matrix coded as row-major? +static stbi_uc stbi__jpeg_dezigzag[64+15] = +{ + 0, 1, 8, 16, 9, 2, 3, 10, + 17, 24, 32, 25, 18, 11, 4, 5, + 12, 19, 26, 33, 40, 48, 41, 34, + 27, 20, 13, 6, 7, 14, 21, 28, + 35, 42, 49, 56, 57, 50, 43, 36, + 29, 22, 15, 23, 30, 37, 44, 51, + 58, 59, 52, 45, 38, 31, 39, 46, + 53, 60, 61, 54, 47, 55, 62, 63, + // let corrupt input sample past end + 63, 63, 63, 63, 63, 63, 63, 63, + 63, 63, 63, 63, 63, 63, 63 +}; + +// decode one 64-entry block-- +static int stbi__jpeg_decode_block(stbi__jpeg *j, short data[64], stbi__huffman *hdc, stbi__huffman *hac, stbi__int16 *fac, int b, stbi__uint16 *dequant) +{ + int diff,dc,k; + int t; + + if (j->code_bits < 16) stbi__grow_buffer_unsafe(j); + t = stbi__jpeg_huff_decode(j, hdc); + if (t < 0) return stbi__err("bad huffman code","Corrupt JPEG"); + + // 0 all the ac values now so we can do it 32-bits at a time + memset(data,0,64*sizeof(data[0])); + + diff = t ? stbi__extend_receive(j, t) : 0; + dc = j->img_comp[b].dc_pred + diff; + j->img_comp[b].dc_pred = dc; + data[0] = (short) (dc * dequant[0]); + + // decode AC components, see JPEG spec + k = 1; + do { + unsigned int zig; + int c,r,s; + if (j->code_bits < 16) stbi__grow_buffer_unsafe(j); + c = (j->code_buffer >> (32 - FAST_BITS)) & ((1 << FAST_BITS)-1); + r = fac[c]; + if (r) { // fast-AC path + k += (r >> 4) & 15; // run + s = r & 15; // combined length + j->code_buffer <<= s; + j->code_bits -= s; + // decode into unzigzag'd location + zig = stbi__jpeg_dezigzag[k++]; + data[zig] = (short) ((r >> 8) * dequant[zig]); + } else { + int rs = stbi__jpeg_huff_decode(j, hac); + if (rs < 0) return stbi__err("bad huffman code","Corrupt JPEG"); + s = rs & 15; + r = rs >> 4; + if (s == 0) { + if (rs != 0xf0) break; // end block + k += 16; + } else { + k += r; + // decode into unzigzag'd location + zig = stbi__jpeg_dezigzag[k++]; + data[zig] = (short) (stbi__extend_receive(j,s) * dequant[zig]); + } + } + } while (k < 64); + return 1; +} + +static int stbi__jpeg_decode_block_prog_dc(stbi__jpeg *j, short data[64], stbi__huffman *hdc, int b) +{ + int diff,dc; + int t; + if (j->spec_end != 0) return stbi__err("can't merge dc and ac", "Corrupt JPEG"); + + if (j->code_bits < 16) stbi__grow_buffer_unsafe(j); + + if (j->succ_high == 0) { + // first scan for DC coefficient, must be first + memset(data,0,64*sizeof(data[0])); // 0 all the ac values now + t = stbi__jpeg_huff_decode(j, hdc); + diff = t ? stbi__extend_receive(j, t) : 0; + + dc = j->img_comp[b].dc_pred + diff; + j->img_comp[b].dc_pred = dc; + data[0] = (short) (dc << j->succ_low); + } else { + // refinement scan for DC coefficient + if (stbi__jpeg_get_bit(j)) + data[0] += (short) (1 << j->succ_low); + } + return 1; +} + +// @OPTIMIZE: store non-zigzagged during the decode passes, +// and only de-zigzag when dequantizing +static int stbi__jpeg_decode_block_prog_ac(stbi__jpeg *j, short data[64], stbi__huffman *hac, stbi__int16 *fac) +{ + int k; + if (j->spec_start == 0) return stbi__err("can't merge dc and ac", "Corrupt JPEG"); + + if (j->succ_high == 0) { + int shift = j->succ_low; + + if (j->eob_run) { + --j->eob_run; + return 1; + } + + k = j->spec_start; + do { + unsigned int zig; + int c,r,s; + if (j->code_bits < 16) stbi__grow_buffer_unsafe(j); + c = (j->code_buffer >> (32 - FAST_BITS)) & ((1 << FAST_BITS)-1); + r = fac[c]; + if (r) { // fast-AC path + k += (r >> 4) & 15; // run + s = r & 15; // combined length + j->code_buffer <<= s; + j->code_bits -= s; + zig = stbi__jpeg_dezigzag[k++]; + data[zig] = (short) ((r >> 8) << shift); + } else { + int rs = stbi__jpeg_huff_decode(j, hac); + if (rs < 0) return stbi__err("bad huffman code","Corrupt JPEG"); + s = rs & 15; + r = rs >> 4; + if (s == 0) { + if (r < 15) { + j->eob_run = (1 << r); + if (r) + j->eob_run += stbi__jpeg_get_bits(j, r); + --j->eob_run; + break; + } + k += 16; + } else { + k += r; + zig = stbi__jpeg_dezigzag[k++]; + data[zig] = (short) (stbi__extend_receive(j,s) << shift); + } + } + } while (k <= j->spec_end); + } else { + // refinement scan for these AC coefficients + + short bit = (short) (1 << j->succ_low); + + if (j->eob_run) { + --j->eob_run; + for (k = j->spec_start; k <= j->spec_end; ++k) { + short *p = &data[stbi__jpeg_dezigzag[k]]; + if (*p != 0) + if (stbi__jpeg_get_bit(j)) + if ((*p & bit)==0) { + if (*p > 0) + *p += bit; + else + *p -= bit; + } + } + } else { + k = j->spec_start; + do { + int r,s; + int rs = stbi__jpeg_huff_decode(j, hac); // @OPTIMIZE see if we can use the fast path here, advance-by-r is so slow, eh + if (rs < 0) return stbi__err("bad huffman code","Corrupt JPEG"); + s = rs & 15; + r = rs >> 4; + if (s == 0) { + if (r < 15) { + j->eob_run = (1 << r) - 1; + if (r) + j->eob_run += stbi__jpeg_get_bits(j, r); + r = 64; // force end of block + } else { + // r=15 s=0 should write 16 0s, so we just do + // a run of 15 0s and then write s (which is 0), + // so we don't have to do anything special here + } + } else { + if (s != 1) return stbi__err("bad huffman code", "Corrupt JPEG"); + // sign bit + if (stbi__jpeg_get_bit(j)) + s = bit; + else + s = -bit; + } + + // advance by r + while (k <= j->spec_end) { + short *p = &data[stbi__jpeg_dezigzag[k++]]; + if (*p != 0) { + if (stbi__jpeg_get_bit(j)) + if ((*p & bit)==0) { + if (*p > 0) + *p += bit; + else + *p -= bit; + } + } else { + if (r == 0) { + *p = (short) s; + break; + } + --r; + } + } + } while (k <= j->spec_end); + } + } + return 1; +} + +// take a -128..127 value and stbi__clamp it and convert to 0..255 +stbi_inline static stbi_uc stbi__clamp(int x) +{ + // trick to use a single test to catch both cases + if ((unsigned int) x > 255) { + if (x < 0) return 0; + if (x > 255) return 255; + } + return (stbi_uc) x; +} + +#define stbi__f2f(x) ((int) (((x) * 4096 + 0.5))) +#define stbi__fsh(x) ((x) << 12) + +// derived from jidctint -- DCT_ISLOW +#define STBI__IDCT_1D(s0,s1,s2,s3,s4,s5,s6,s7) \ + int t0,t1,t2,t3,p1,p2,p3,p4,p5,x0,x1,x2,x3; \ + p2 = s2; \ + p3 = s6; \ + p1 = (p2+p3) * stbi__f2f(0.5411961f); \ + t2 = p1 + p3*stbi__f2f(-1.847759065f); \ + t3 = p1 + p2*stbi__f2f( 0.765366865f); \ + p2 = s0; \ + p3 = s4; \ + t0 = stbi__fsh(p2+p3); \ + t1 = stbi__fsh(p2-p3); \ + x0 = t0+t3; \ + x3 = t0-t3; \ + x1 = t1+t2; \ + x2 = t1-t2; \ + t0 = s7; \ + t1 = s5; \ + t2 = s3; \ + t3 = s1; \ + p3 = t0+t2; \ + p4 = t1+t3; \ + p1 = t0+t3; \ + p2 = t1+t2; \ + p5 = (p3+p4)*stbi__f2f( 1.175875602f); \ + t0 = t0*stbi__f2f( 0.298631336f); \ + t1 = t1*stbi__f2f( 2.053119869f); \ + t2 = t2*stbi__f2f( 3.072711026f); \ + t3 = t3*stbi__f2f( 1.501321110f); \ + p1 = p5 + p1*stbi__f2f(-0.899976223f); \ + p2 = p5 + p2*stbi__f2f(-2.562915447f); \ + p3 = p3*stbi__f2f(-1.961570560f); \ + p4 = p4*stbi__f2f(-0.390180644f); \ + t3 += p1+p4; \ + t2 += p2+p3; \ + t1 += p2+p4; \ + t0 += p1+p3; + +static void stbi__idct_block(stbi_uc *out, int out_stride, short data[64]) +{ + int i,val[64],*v=val; + stbi_uc *o; + short *d = data; + + // columns + for (i=0; i < 8; ++i,++d, ++v) { + // if all zeroes, shortcut -- this avoids dequantizing 0s and IDCTing + if (d[ 8]==0 && d[16]==0 && d[24]==0 && d[32]==0 + && d[40]==0 && d[48]==0 && d[56]==0) { + // no shortcut 0 seconds + // (1|2|3|4|5|6|7)==0 0 seconds + // all separate -0.047 seconds + // 1 && 2|3 && 4|5 && 6|7: -0.047 seconds + int dcterm = d[0] << 2; + v[0] = v[8] = v[16] = v[24] = v[32] = v[40] = v[48] = v[56] = dcterm; + } else { + STBI__IDCT_1D(d[ 0],d[ 8],d[16],d[24],d[32],d[40],d[48],d[56]) + // constants scaled things up by 1<<12; let's bring them back + // down, but keep 2 extra bits of precision + x0 += 512; x1 += 512; x2 += 512; x3 += 512; + v[ 0] = (x0+t3) >> 10; + v[56] = (x0-t3) >> 10; + v[ 8] = (x1+t2) >> 10; + v[48] = (x1-t2) >> 10; + v[16] = (x2+t1) >> 10; + v[40] = (x2-t1) >> 10; + v[24] = (x3+t0) >> 10; + v[32] = (x3-t0) >> 10; + } + } + + for (i=0, v=val, o=out; i < 8; ++i,v+=8,o+=out_stride) { + // no fast case since the first 1D IDCT spread components out + STBI__IDCT_1D(v[0],v[1],v[2],v[3],v[4],v[5],v[6],v[7]) + // constants scaled things up by 1<<12, plus we had 1<<2 from first + // loop, plus horizontal and vertical each scale by sqrt(8) so together + // we've got an extra 1<<3, so 1<<17 total we need to remove. + // so we want to round that, which means adding 0.5 * 1<<17, + // aka 65536. Also, we'll end up with -128 to 127 that we want + // to encode as 0..255 by adding 128, so we'll add that before the shift + x0 += 65536 + (128<<17); + x1 += 65536 + (128<<17); + x2 += 65536 + (128<<17); + x3 += 65536 + (128<<17); + // tried computing the shifts into temps, or'ing the temps to see + // if any were out of range, but that was slower + o[0] = stbi__clamp((x0+t3) >> 17); + o[7] = stbi__clamp((x0-t3) >> 17); + o[1] = stbi__clamp((x1+t2) >> 17); + o[6] = stbi__clamp((x1-t2) >> 17); + o[2] = stbi__clamp((x2+t1) >> 17); + o[5] = stbi__clamp((x2-t1) >> 17); + o[3] = stbi__clamp((x3+t0) >> 17); + o[4] = stbi__clamp((x3-t0) >> 17); + } +} + +#ifdef STBI_SSE2 +// sse2 integer IDCT. not the fastest possible implementation but it +// produces bit-identical results to the generic C version so it's +// fully "transparent". +static void stbi__idct_simd(stbi_uc *out, int out_stride, short data[64]) +{ + // This is constructed to match our regular (generic) integer IDCT exactly. + __m128i row0, row1, row2, row3, row4, row5, row6, row7; + __m128i tmp; + + // dot product constant: even elems=x, odd elems=y + #define dct_const(x,y) _mm_setr_epi16((x),(y),(x),(y),(x),(y),(x),(y)) + + // out(0) = c0[even]*x + c0[odd]*y (c0, x, y 16-bit, out 32-bit) + // out(1) = c1[even]*x + c1[odd]*y + #define dct_rot(out0,out1, x,y,c0,c1) \ + __m128i c0##lo = _mm_unpacklo_epi16((x),(y)); \ + __m128i c0##hi = _mm_unpackhi_epi16((x),(y)); \ + __m128i out0##_l = _mm_madd_epi16(c0##lo, c0); \ + __m128i out0##_h = _mm_madd_epi16(c0##hi, c0); \ + __m128i out1##_l = _mm_madd_epi16(c0##lo, c1); \ + __m128i out1##_h = _mm_madd_epi16(c0##hi, c1) + + // out = in << 12 (in 16-bit, out 32-bit) + #define dct_widen(out, in) \ + __m128i out##_l = _mm_srai_epi32(_mm_unpacklo_epi16(_mm_setzero_si128(), (in)), 4); \ + __m128i out##_h = _mm_srai_epi32(_mm_unpackhi_epi16(_mm_setzero_si128(), (in)), 4) + + // wide add + #define dct_wadd(out, a, b) \ + __m128i out##_l = _mm_add_epi32(a##_l, b##_l); \ + __m128i out##_h = _mm_add_epi32(a##_h, b##_h) + + // wide sub + #define dct_wsub(out, a, b) \ + __m128i out##_l = _mm_sub_epi32(a##_l, b##_l); \ + __m128i out##_h = _mm_sub_epi32(a##_h, b##_h) + + // butterfly a/b, add bias, then shift by "s" and pack + #define dct_bfly32o(out0, out1, a,b,bias,s) \ + { \ + __m128i abiased_l = _mm_add_epi32(a##_l, bias); \ + __m128i abiased_h = _mm_add_epi32(a##_h, bias); \ + dct_wadd(sum, abiased, b); \ + dct_wsub(dif, abiased, b); \ + out0 = _mm_packs_epi32(_mm_srai_epi32(sum_l, s), _mm_srai_epi32(sum_h, s)); \ + out1 = _mm_packs_epi32(_mm_srai_epi32(dif_l, s), _mm_srai_epi32(dif_h, s)); \ + } + + // 8-bit interleave step (for transposes) + #define dct_interleave8(a, b) \ + tmp = a; \ + a = _mm_unpacklo_epi8(a, b); \ + b = _mm_unpackhi_epi8(tmp, b) + + // 16-bit interleave step (for transposes) + #define dct_interleave16(a, b) \ + tmp = a; \ + a = _mm_unpacklo_epi16(a, b); \ + b = _mm_unpackhi_epi16(tmp, b) + + #define dct_pass(bias,shift) \ + { \ + /* even part */ \ + dct_rot(t2e,t3e, row2,row6, rot0_0,rot0_1); \ + __m128i sum04 = _mm_add_epi16(row0, row4); \ + __m128i dif04 = _mm_sub_epi16(row0, row4); \ + dct_widen(t0e, sum04); \ + dct_widen(t1e, dif04); \ + dct_wadd(x0, t0e, t3e); \ + dct_wsub(x3, t0e, t3e); \ + dct_wadd(x1, t1e, t2e); \ + dct_wsub(x2, t1e, t2e); \ + /* odd part */ \ + dct_rot(y0o,y2o, row7,row3, rot2_0,rot2_1); \ + dct_rot(y1o,y3o, row5,row1, rot3_0,rot3_1); \ + __m128i sum17 = _mm_add_epi16(row1, row7); \ + __m128i sum35 = _mm_add_epi16(row3, row5); \ + dct_rot(y4o,y5o, sum17,sum35, rot1_0,rot1_1); \ + dct_wadd(x4, y0o, y4o); \ + dct_wadd(x5, y1o, y5o); \ + dct_wadd(x6, y2o, y5o); \ + dct_wadd(x7, y3o, y4o); \ + dct_bfly32o(row0,row7, x0,x7,bias,shift); \ + dct_bfly32o(row1,row6, x1,x6,bias,shift); \ + dct_bfly32o(row2,row5, x2,x5,bias,shift); \ + dct_bfly32o(row3,row4, x3,x4,bias,shift); \ + } + + __m128i rot0_0 = dct_const(stbi__f2f(0.5411961f), stbi__f2f(0.5411961f) + stbi__f2f(-1.847759065f)); + __m128i rot0_1 = dct_const(stbi__f2f(0.5411961f) + stbi__f2f( 0.765366865f), stbi__f2f(0.5411961f)); + __m128i rot1_0 = dct_const(stbi__f2f(1.175875602f) + stbi__f2f(-0.899976223f), stbi__f2f(1.175875602f)); + __m128i rot1_1 = dct_const(stbi__f2f(1.175875602f), stbi__f2f(1.175875602f) + stbi__f2f(-2.562915447f)); + __m128i rot2_0 = dct_const(stbi__f2f(-1.961570560f) + stbi__f2f( 0.298631336f), stbi__f2f(-1.961570560f)); + __m128i rot2_1 = dct_const(stbi__f2f(-1.961570560f), stbi__f2f(-1.961570560f) + stbi__f2f( 3.072711026f)); + __m128i rot3_0 = dct_const(stbi__f2f(-0.390180644f) + stbi__f2f( 2.053119869f), stbi__f2f(-0.390180644f)); + __m128i rot3_1 = dct_const(stbi__f2f(-0.390180644f), stbi__f2f(-0.390180644f) + stbi__f2f( 1.501321110f)); + + // rounding biases in column/row passes, see stbi__idct_block for explanation. + __m128i bias_0 = _mm_set1_epi32(512); + __m128i bias_1 = _mm_set1_epi32(65536 + (128<<17)); + + // load + row0 = _mm_load_si128((const __m128i *) (data + 0*8)); + row1 = _mm_load_si128((const __m128i *) (data + 1*8)); + row2 = _mm_load_si128((const __m128i *) (data + 2*8)); + row3 = _mm_load_si128((const __m128i *) (data + 3*8)); + row4 = _mm_load_si128((const __m128i *) (data + 4*8)); + row5 = _mm_load_si128((const __m128i *) (data + 5*8)); + row6 = _mm_load_si128((const __m128i *) (data + 6*8)); + row7 = _mm_load_si128((const __m128i *) (data + 7*8)); + + // column pass + dct_pass(bias_0, 10); + + { + // 16bit 8x8 transpose pass 1 + dct_interleave16(row0, row4); + dct_interleave16(row1, row5); + dct_interleave16(row2, row6); + dct_interleave16(row3, row7); + + // transpose pass 2 + dct_interleave16(row0, row2); + dct_interleave16(row1, row3); + dct_interleave16(row4, row6); + dct_interleave16(row5, row7); + + // transpose pass 3 + dct_interleave16(row0, row1); + dct_interleave16(row2, row3); + dct_interleave16(row4, row5); + dct_interleave16(row6, row7); + } + + // row pass + dct_pass(bias_1, 17); + + { + // pack + __m128i p0 = _mm_packus_epi16(row0, row1); // a0a1a2a3...a7b0b1b2b3...b7 + __m128i p1 = _mm_packus_epi16(row2, row3); + __m128i p2 = _mm_packus_epi16(row4, row5); + __m128i p3 = _mm_packus_epi16(row6, row7); + + // 8bit 8x8 transpose pass 1 + dct_interleave8(p0, p2); // a0e0a1e1... + dct_interleave8(p1, p3); // c0g0c1g1... + + // transpose pass 2 + dct_interleave8(p0, p1); // a0c0e0g0... + dct_interleave8(p2, p3); // b0d0f0h0... + + // transpose pass 3 + dct_interleave8(p0, p2); // a0b0c0d0... + dct_interleave8(p1, p3); // a4b4c4d4... + + // store + _mm_storel_epi64((__m128i *) out, p0); out += out_stride; + _mm_storel_epi64((__m128i *) out, _mm_shuffle_epi32(p0, 0x4e)); out += out_stride; + _mm_storel_epi64((__m128i *) out, p2); out += out_stride; + _mm_storel_epi64((__m128i *) out, _mm_shuffle_epi32(p2, 0x4e)); out += out_stride; + _mm_storel_epi64((__m128i *) out, p1); out += out_stride; + _mm_storel_epi64((__m128i *) out, _mm_shuffle_epi32(p1, 0x4e)); out += out_stride; + _mm_storel_epi64((__m128i *) out, p3); out += out_stride; + _mm_storel_epi64((__m128i *) out, _mm_shuffle_epi32(p3, 0x4e)); + } + +#undef dct_const +#undef dct_rot +#undef dct_widen +#undef dct_wadd +#undef dct_wsub +#undef dct_bfly32o +#undef dct_interleave8 +#undef dct_interleave16 +#undef dct_pass +} + +#endif // STBI_SSE2 + +#ifdef STBI_NEON + +// NEON integer IDCT. should produce bit-identical +// results to the generic C version. +static void stbi__idct_simd(stbi_uc *out, int out_stride, short data[64]) +{ + int16x8_t row0, row1, row2, row3, row4, row5, row6, row7; + + int16x4_t rot0_0 = vdup_n_s16(stbi__f2f(0.5411961f)); + int16x4_t rot0_1 = vdup_n_s16(stbi__f2f(-1.847759065f)); + int16x4_t rot0_2 = vdup_n_s16(stbi__f2f( 0.765366865f)); + int16x4_t rot1_0 = vdup_n_s16(stbi__f2f( 1.175875602f)); + int16x4_t rot1_1 = vdup_n_s16(stbi__f2f(-0.899976223f)); + int16x4_t rot1_2 = vdup_n_s16(stbi__f2f(-2.562915447f)); + int16x4_t rot2_0 = vdup_n_s16(stbi__f2f(-1.961570560f)); + int16x4_t rot2_1 = vdup_n_s16(stbi__f2f(-0.390180644f)); + int16x4_t rot3_0 = vdup_n_s16(stbi__f2f( 0.298631336f)); + int16x4_t rot3_1 = vdup_n_s16(stbi__f2f( 2.053119869f)); + int16x4_t rot3_2 = vdup_n_s16(stbi__f2f( 3.072711026f)); + int16x4_t rot3_3 = vdup_n_s16(stbi__f2f( 1.501321110f)); + +#define dct_long_mul(out, inq, coeff) \ + int32x4_t out##_l = vmull_s16(vget_low_s16(inq), coeff); \ + int32x4_t out##_h = vmull_s16(vget_high_s16(inq), coeff) + +#define dct_long_mac(out, acc, inq, coeff) \ + int32x4_t out##_l = vmlal_s16(acc##_l, vget_low_s16(inq), coeff); \ + int32x4_t out##_h = vmlal_s16(acc##_h, vget_high_s16(inq), coeff) + +#define dct_widen(out, inq) \ + int32x4_t out##_l = vshll_n_s16(vget_low_s16(inq), 12); \ + int32x4_t out##_h = vshll_n_s16(vget_high_s16(inq), 12) + +// wide add +#define dct_wadd(out, a, b) \ + int32x4_t out##_l = vaddq_s32(a##_l, b##_l); \ + int32x4_t out##_h = vaddq_s32(a##_h, b##_h) + +// wide sub +#define dct_wsub(out, a, b) \ + int32x4_t out##_l = vsubq_s32(a##_l, b##_l); \ + int32x4_t out##_h = vsubq_s32(a##_h, b##_h) + +// butterfly a/b, then shift using "shiftop" by "s" and pack +#define dct_bfly32o(out0,out1, a,b,shiftop,s) \ + { \ + dct_wadd(sum, a, b); \ + dct_wsub(dif, a, b); \ + out0 = vcombine_s16(shiftop(sum_l, s), shiftop(sum_h, s)); \ + out1 = vcombine_s16(shiftop(dif_l, s), shiftop(dif_h, s)); \ + } + +#define dct_pass(shiftop, shift) \ + { \ + /* even part */ \ + int16x8_t sum26 = vaddq_s16(row2, row6); \ + dct_long_mul(p1e, sum26, rot0_0); \ + dct_long_mac(t2e, p1e, row6, rot0_1); \ + dct_long_mac(t3e, p1e, row2, rot0_2); \ + int16x8_t sum04 = vaddq_s16(row0, row4); \ + int16x8_t dif04 = vsubq_s16(row0, row4); \ + dct_widen(t0e, sum04); \ + dct_widen(t1e, dif04); \ + dct_wadd(x0, t0e, t3e); \ + dct_wsub(x3, t0e, t3e); \ + dct_wadd(x1, t1e, t2e); \ + dct_wsub(x2, t1e, t2e); \ + /* odd part */ \ + int16x8_t sum15 = vaddq_s16(row1, row5); \ + int16x8_t sum17 = vaddq_s16(row1, row7); \ + int16x8_t sum35 = vaddq_s16(row3, row5); \ + int16x8_t sum37 = vaddq_s16(row3, row7); \ + int16x8_t sumodd = vaddq_s16(sum17, sum35); \ + dct_long_mul(p5o, sumodd, rot1_0); \ + dct_long_mac(p1o, p5o, sum17, rot1_1); \ + dct_long_mac(p2o, p5o, sum35, rot1_2); \ + dct_long_mul(p3o, sum37, rot2_0); \ + dct_long_mul(p4o, sum15, rot2_1); \ + dct_wadd(sump13o, p1o, p3o); \ + dct_wadd(sump24o, p2o, p4o); \ + dct_wadd(sump23o, p2o, p3o); \ + dct_wadd(sump14o, p1o, p4o); \ + dct_long_mac(x4, sump13o, row7, rot3_0); \ + dct_long_mac(x5, sump24o, row5, rot3_1); \ + dct_long_mac(x6, sump23o, row3, rot3_2); \ + dct_long_mac(x7, sump14o, row1, rot3_3); \ + dct_bfly32o(row0,row7, x0,x7,shiftop,shift); \ + dct_bfly32o(row1,row6, x1,x6,shiftop,shift); \ + dct_bfly32o(row2,row5, x2,x5,shiftop,shift); \ + dct_bfly32o(row3,row4, x3,x4,shiftop,shift); \ + } + + // load + row0 = vld1q_s16(data + 0*8); + row1 = vld1q_s16(data + 1*8); + row2 = vld1q_s16(data + 2*8); + row3 = vld1q_s16(data + 3*8); + row4 = vld1q_s16(data + 4*8); + row5 = vld1q_s16(data + 5*8); + row6 = vld1q_s16(data + 6*8); + row7 = vld1q_s16(data + 7*8); + + // add DC bias + row0 = vaddq_s16(row0, vsetq_lane_s16(1024, vdupq_n_s16(0), 0)); + + // column pass + dct_pass(vrshrn_n_s32, 10); + + // 16bit 8x8 transpose + { +// these three map to a single VTRN.16, VTRN.32, and VSWP, respectively. +// whether compilers actually get this is another story, sadly. +#define dct_trn16(x, y) { int16x8x2_t t = vtrnq_s16(x, y); x = t.val[0]; y = t.val[1]; } +#define dct_trn32(x, y) { int32x4x2_t t = vtrnq_s32(vreinterpretq_s32_s16(x), vreinterpretq_s32_s16(y)); x = vreinterpretq_s16_s32(t.val[0]); y = vreinterpretq_s16_s32(t.val[1]); } +#define dct_trn64(x, y) { int16x8_t x0 = x; int16x8_t y0 = y; x = vcombine_s16(vget_low_s16(x0), vget_low_s16(y0)); y = vcombine_s16(vget_high_s16(x0), vget_high_s16(y0)); } + + // pass 1 + dct_trn16(row0, row1); // a0b0a2b2a4b4a6b6 + dct_trn16(row2, row3); + dct_trn16(row4, row5); + dct_trn16(row6, row7); + + // pass 2 + dct_trn32(row0, row2); // a0b0c0d0a4b4c4d4 + dct_trn32(row1, row3); + dct_trn32(row4, row6); + dct_trn32(row5, row7); + + // pass 3 + dct_trn64(row0, row4); // a0b0c0d0e0f0g0h0 + dct_trn64(row1, row5); + dct_trn64(row2, row6); + dct_trn64(row3, row7); + +#undef dct_trn16 +#undef dct_trn32 +#undef dct_trn64 + } + + // row pass + // vrshrn_n_s32 only supports shifts up to 16, we need + // 17. so do a non-rounding shift of 16 first then follow + // up with a rounding shift by 1. + dct_pass(vshrn_n_s32, 16); + + { + // pack and round + uint8x8_t p0 = vqrshrun_n_s16(row0, 1); + uint8x8_t p1 = vqrshrun_n_s16(row1, 1); + uint8x8_t p2 = vqrshrun_n_s16(row2, 1); + uint8x8_t p3 = vqrshrun_n_s16(row3, 1); + uint8x8_t p4 = vqrshrun_n_s16(row4, 1); + uint8x8_t p5 = vqrshrun_n_s16(row5, 1); + uint8x8_t p6 = vqrshrun_n_s16(row6, 1); + uint8x8_t p7 = vqrshrun_n_s16(row7, 1); + + // again, these can translate into one instruction, but often don't. +#define dct_trn8_8(x, y) { uint8x8x2_t t = vtrn_u8(x, y); x = t.val[0]; y = t.val[1]; } +#define dct_trn8_16(x, y) { uint16x4x2_t t = vtrn_u16(vreinterpret_u16_u8(x), vreinterpret_u16_u8(y)); x = vreinterpret_u8_u16(t.val[0]); y = vreinterpret_u8_u16(t.val[1]); } +#define dct_trn8_32(x, y) { uint32x2x2_t t = vtrn_u32(vreinterpret_u32_u8(x), vreinterpret_u32_u8(y)); x = vreinterpret_u8_u32(t.val[0]); y = vreinterpret_u8_u32(t.val[1]); } + + // sadly can't use interleaved stores here since we only write + // 8 bytes to each scan line! + + // 8x8 8-bit transpose pass 1 + dct_trn8_8(p0, p1); + dct_trn8_8(p2, p3); + dct_trn8_8(p4, p5); + dct_trn8_8(p6, p7); + + // pass 2 + dct_trn8_16(p0, p2); + dct_trn8_16(p1, p3); + dct_trn8_16(p4, p6); + dct_trn8_16(p5, p7); + + // pass 3 + dct_trn8_32(p0, p4); + dct_trn8_32(p1, p5); + dct_trn8_32(p2, p6); + dct_trn8_32(p3, p7); + + // store + vst1_u8(out, p0); out += out_stride; + vst1_u8(out, p1); out += out_stride; + vst1_u8(out, p2); out += out_stride; + vst1_u8(out, p3); out += out_stride; + vst1_u8(out, p4); out += out_stride; + vst1_u8(out, p5); out += out_stride; + vst1_u8(out, p6); out += out_stride; + vst1_u8(out, p7); + +#undef dct_trn8_8 +#undef dct_trn8_16 +#undef dct_trn8_32 + } + +#undef dct_long_mul +#undef dct_long_mac +#undef dct_widen +#undef dct_wadd +#undef dct_wsub +#undef dct_bfly32o +#undef dct_pass +} + +#endif // STBI_NEON + +#define STBI__MARKER_none 0xff +// if there's a pending marker from the entropy stream, return that +// otherwise, fetch from the stream and get a marker. if there's no +// marker, return 0xff, which is never a valid marker value +static stbi_uc stbi__get_marker(stbi__jpeg *j) +{ + stbi_uc x; + if (j->marker != STBI__MARKER_none) { x = j->marker; j->marker = STBI__MARKER_none; return x; } + x = stbi__get8(j->s); + if (x != 0xff) return STBI__MARKER_none; + while (x == 0xff) + x = stbi__get8(j->s); // consume repeated 0xff fill bytes + return x; +} + +// in each scan, we'll have scan_n components, and the order +// of the components is specified by order[] +#define STBI__RESTART(x) ((x) >= 0xd0 && (x) <= 0xd7) + +// after a restart interval, stbi__jpeg_reset the entropy decoder and +// the dc prediction +static void stbi__jpeg_reset(stbi__jpeg *j) +{ + j->code_bits = 0; + j->code_buffer = 0; + j->nomore = 0; + j->img_comp[0].dc_pred = j->img_comp[1].dc_pred = j->img_comp[2].dc_pred = j->img_comp[3].dc_pred = 0; + j->marker = STBI__MARKER_none; + j->todo = j->restart_interval ? j->restart_interval : 0x7fffffff; + j->eob_run = 0; + // no more than 1<<31 MCUs if no restart_interal? that's plenty safe, + // since we don't even allow 1<<30 pixels +} + +static int stbi__parse_entropy_coded_data(stbi__jpeg *z) +{ + stbi__jpeg_reset(z); + if (!z->progressive) { + if (z->scan_n == 1) { + int i,j; + STBI_SIMD_ALIGN(short, data[64]); + int n = z->order[0]; + // non-interleaved data, we just need to process one block at a time, + // in trivial scanline order + // number of blocks to do just depends on how many actual "pixels" this + // component has, independent of interleaved MCU blocking and such + int w = (z->img_comp[n].x+7) >> 3; + int h = (z->img_comp[n].y+7) >> 3; + for (j=0; j < h; ++j) { + for (i=0; i < w; ++i) { + int ha = z->img_comp[n].ha; + if (!stbi__jpeg_decode_block(z, data, z->huff_dc+z->img_comp[n].hd, z->huff_ac+ha, z->fast_ac[ha], n, z->dequant[z->img_comp[n].tq])) return 0; + z->idct_block_kernel(z->img_comp[n].data+z->img_comp[n].w2*j*8+i*8, z->img_comp[n].w2, data); + // every data block is an MCU, so countdown the restart interval + if (--z->todo <= 0) { + if (z->code_bits < 24) stbi__grow_buffer_unsafe(z); + // if it's NOT a restart, then just bail, so we get corrupt data + // rather than no data + if (!STBI__RESTART(z->marker)) return 1; + stbi__jpeg_reset(z); + } + } + } + return 1; + } else { // interleaved + int i,j,k,x,y; + STBI_SIMD_ALIGN(short, data[64]); + for (j=0; j < z->img_mcu_y; ++j) { + for (i=0; i < z->img_mcu_x; ++i) { + // scan an interleaved mcu... process scan_n components in order + for (k=0; k < z->scan_n; ++k) { + int n = z->order[k]; + // scan out an mcu's worth of this component; that's just determined + // by the basic H and V specified for the component + for (y=0; y < z->img_comp[n].v; ++y) { + for (x=0; x < z->img_comp[n].h; ++x) { + int x2 = (i*z->img_comp[n].h + x)*8; + int y2 = (j*z->img_comp[n].v + y)*8; + int ha = z->img_comp[n].ha; + if (!stbi__jpeg_decode_block(z, data, z->huff_dc+z->img_comp[n].hd, z->huff_ac+ha, z->fast_ac[ha], n, z->dequant[z->img_comp[n].tq])) return 0; + z->idct_block_kernel(z->img_comp[n].data+z->img_comp[n].w2*y2+x2, z->img_comp[n].w2, data); + } + } + } + // after all interleaved components, that's an interleaved MCU, + // so now count down the restart interval + if (--z->todo <= 0) { + if (z->code_bits < 24) stbi__grow_buffer_unsafe(z); + if (!STBI__RESTART(z->marker)) return 1; + stbi__jpeg_reset(z); + } + } + } + return 1; + } + } else { + if (z->scan_n == 1) { + int i,j; + int n = z->order[0]; + // non-interleaved data, we just need to process one block at a time, + // in trivial scanline order + // number of blocks to do just depends on how many actual "pixels" this + // component has, independent of interleaved MCU blocking and such + int w = (z->img_comp[n].x+7) >> 3; + int h = (z->img_comp[n].y+7) >> 3; + for (j=0; j < h; ++j) { + for (i=0; i < w; ++i) { + short *data = z->img_comp[n].coeff + 64 * (i + j * z->img_comp[n].coeff_w); + if (z->spec_start == 0) { + if (!stbi__jpeg_decode_block_prog_dc(z, data, &z->huff_dc[z->img_comp[n].hd], n)) + return 0; + } else { + int ha = z->img_comp[n].ha; + if (!stbi__jpeg_decode_block_prog_ac(z, data, &z->huff_ac[ha], z->fast_ac[ha])) + return 0; + } + // every data block is an MCU, so countdown the restart interval + if (--z->todo <= 0) { + if (z->code_bits < 24) stbi__grow_buffer_unsafe(z); + if (!STBI__RESTART(z->marker)) return 1; + stbi__jpeg_reset(z); + } + } + } + return 1; + } else { // interleaved + int i,j,k,x,y; + for (j=0; j < z->img_mcu_y; ++j) { + for (i=0; i < z->img_mcu_x; ++i) { + // scan an interleaved mcu... process scan_n components in order + for (k=0; k < z->scan_n; ++k) { + int n = z->order[k]; + // scan out an mcu's worth of this component; that's just determined + // by the basic H and V specified for the component + for (y=0; y < z->img_comp[n].v; ++y) { + for (x=0; x < z->img_comp[n].h; ++x) { + int x2 = (i*z->img_comp[n].h + x); + int y2 = (j*z->img_comp[n].v + y); + short *data = z->img_comp[n].coeff + 64 * (x2 + y2 * z->img_comp[n].coeff_w); + if (!stbi__jpeg_decode_block_prog_dc(z, data, &z->huff_dc[z->img_comp[n].hd], n)) + return 0; + } + } + } + // after all interleaved components, that's an interleaved MCU, + // so now count down the restart interval + if (--z->todo <= 0) { + if (z->code_bits < 24) stbi__grow_buffer_unsafe(z); + if (!STBI__RESTART(z->marker)) return 1; + stbi__jpeg_reset(z); + } + } + } + return 1; + } + } +} + +static void stbi__jpeg_dequantize(short *data, stbi__uint16 *dequant) +{ + int i; + for (i=0; i < 64; ++i) + data[i] *= dequant[i]; +} + +static void stbi__jpeg_finish(stbi__jpeg *z) +{ + if (z->progressive) { + // dequantize and idct the data + int i,j,n; + for (n=0; n < z->s->img_n; ++n) { + int w = (z->img_comp[n].x+7) >> 3; + int h = (z->img_comp[n].y+7) >> 3; + for (j=0; j < h; ++j) { + for (i=0; i < w; ++i) { + short *data = z->img_comp[n].coeff + 64 * (i + j * z->img_comp[n].coeff_w); + stbi__jpeg_dequantize(data, z->dequant[z->img_comp[n].tq]); + z->idct_block_kernel(z->img_comp[n].data+z->img_comp[n].w2*j*8+i*8, z->img_comp[n].w2, data); + } + } + } + } +} + +static int stbi__process_marker(stbi__jpeg *z, int m) +{ + int L; + switch (m) { + case STBI__MARKER_none: // no marker found + return stbi__err("expected marker","Corrupt JPEG"); + + case 0xDD: // DRI - specify restart interval + if (stbi__get16be(z->s) != 4) return stbi__err("bad DRI len","Corrupt JPEG"); + z->restart_interval = stbi__get16be(z->s); + return 1; + + case 0xDB: // DQT - define quantization table + L = stbi__get16be(z->s)-2; + while (L > 0) { + int q = stbi__get8(z->s); + int p = q >> 4, sixteen = (p != 0); + int t = q & 15,i; + if (p != 0 && p != 1) return stbi__err("bad DQT type","Corrupt JPEG"); + if (t > 3) return stbi__err("bad DQT table","Corrupt JPEG"); + + for (i=0; i < 64; ++i) + z->dequant[t][stbi__jpeg_dezigzag[i]] = (stbi__uint16)(sixteen ? stbi__get16be(z->s) : stbi__get8(z->s)); + L -= (sixteen ? 129 : 65); + } + return L==0; + + case 0xC4: // DHT - define huffman table + L = stbi__get16be(z->s)-2; + while (L > 0) { + stbi_uc *v; + int sizes[16],i,n=0; + int q = stbi__get8(z->s); + int tc = q >> 4; + int th = q & 15; + if (tc > 1 || th > 3) return stbi__err("bad DHT header","Corrupt JPEG"); + for (i=0; i < 16; ++i) { + sizes[i] = stbi__get8(z->s); + n += sizes[i]; + } + L -= 17; + if (tc == 0) { + if (!stbi__build_huffman(z->huff_dc+th, sizes)) return 0; + v = z->huff_dc[th].values; + } else { + if (!stbi__build_huffman(z->huff_ac+th, sizes)) return 0; + v = z->huff_ac[th].values; + } + for (i=0; i < n; ++i) + v[i] = stbi__get8(z->s); + if (tc != 0) + stbi__build_fast_ac(z->fast_ac[th], z->huff_ac + th); + L -= n; + } + return L==0; + } + + // check for comment block or APP blocks + if ((m >= 0xE0 && m <= 0xEF) || m == 0xFE) { + L = stbi__get16be(z->s); + if (L < 2) { + if (m == 0xFE) + return stbi__err("bad COM len","Corrupt JPEG"); + else + return stbi__err("bad APP len","Corrupt JPEG"); + } + L -= 2; + + if (m == 0xE0 && L >= 5) { // JFIF APP0 segment + static const unsigned char tag[5] = {'J','F','I','F','\0'}; + int ok = 1; + int i; + for (i=0; i < 5; ++i) + if (stbi__get8(z->s) != tag[i]) + ok = 0; + L -= 5; + if (ok) + z->jfif = 1; + } else if (m == 0xEE && L >= 12) { // Adobe APP14 segment + static const unsigned char tag[6] = {'A','d','o','b','e','\0'}; + int ok = 1; + int i; + for (i=0; i < 6; ++i) + if (stbi__get8(z->s) != tag[i]) + ok = 0; + L -= 6; + if (ok) { + stbi__get8(z->s); // version + stbi__get16be(z->s); // flags0 + stbi__get16be(z->s); // flags1 + z->app14_color_transform = stbi__get8(z->s); // color transform + L -= 6; + } + } + + stbi__skip(z->s, L); + return 1; + } + + return stbi__err("unknown marker","Corrupt JPEG"); +} + +// after we see SOS +static int stbi__process_scan_header(stbi__jpeg *z) +{ + int i; + int Ls = stbi__get16be(z->s); + z->scan_n = stbi__get8(z->s); + if (z->scan_n < 1 || z->scan_n > 4 || z->scan_n > (int) z->s->img_n) return stbi__err("bad SOS component count","Corrupt JPEG"); + if (Ls != 6+2*z->scan_n) return stbi__err("bad SOS len","Corrupt JPEG"); + for (i=0; i < z->scan_n; ++i) { + int id = stbi__get8(z->s), which; + int q = stbi__get8(z->s); + for (which = 0; which < z->s->img_n; ++which) + if (z->img_comp[which].id == id) + break; + if (which == z->s->img_n) return 0; // no match + z->img_comp[which].hd = q >> 4; if (z->img_comp[which].hd > 3) return stbi__err("bad DC huff","Corrupt JPEG"); + z->img_comp[which].ha = q & 15; if (z->img_comp[which].ha > 3) return stbi__err("bad AC huff","Corrupt JPEG"); + z->order[i] = which; + } + + { + int aa; + z->spec_start = stbi__get8(z->s); + z->spec_end = stbi__get8(z->s); // should be 63, but might be 0 + aa = stbi__get8(z->s); + z->succ_high = (aa >> 4); + z->succ_low = (aa & 15); + if (z->progressive) { + if (z->spec_start > 63 || z->spec_end > 63 || z->spec_start > z->spec_end || z->succ_high > 13 || z->succ_low > 13) + return stbi__err("bad SOS", "Corrupt JPEG"); + } else { + if (z->spec_start != 0) return stbi__err("bad SOS","Corrupt JPEG"); + if (z->succ_high != 0 || z->succ_low != 0) return stbi__err("bad SOS","Corrupt JPEG"); + z->spec_end = 63; + } + } + + return 1; +} + +static int stbi__free_jpeg_components(stbi__jpeg *z, int ncomp, int why) +{ + int i; + for (i=0; i < ncomp; ++i) { + if (z->img_comp[i].raw_data) { + STBI_FREE(z->img_comp[i].raw_data); + z->img_comp[i].raw_data = NULL; + z->img_comp[i].data = NULL; + } + if (z->img_comp[i].raw_coeff) { + STBI_FREE(z->img_comp[i].raw_coeff); + z->img_comp[i].raw_coeff = 0; + z->img_comp[i].coeff = 0; + } + if (z->img_comp[i].linebuf) { + STBI_FREE(z->img_comp[i].linebuf); + z->img_comp[i].linebuf = NULL; + } + } + return why; +} + +static int stbi__process_frame_header(stbi__jpeg *z, int scan) +{ + stbi__context *s = z->s; + int Lf,p,i,q, h_max=1,v_max=1,c; + Lf = stbi__get16be(s); if (Lf < 11) return stbi__err("bad SOF len","Corrupt JPEG"); // JPEG + p = stbi__get8(s); if (p != 8) return stbi__err("only 8-bit","JPEG format not supported: 8-bit only"); // JPEG baseline + s->img_y = stbi__get16be(s); if (s->img_y == 0) return stbi__err("no header height", "JPEG format not supported: delayed height"); // Legal, but we don't handle it--but neither does IJG + s->img_x = stbi__get16be(s); if (s->img_x == 0) return stbi__err("0 width","Corrupt JPEG"); // JPEG requires + c = stbi__get8(s); + if (c != 3 && c != 1 && c != 4) return stbi__err("bad component count","Corrupt JPEG"); + s->img_n = c; + for (i=0; i < c; ++i) { + z->img_comp[i].data = NULL; + z->img_comp[i].linebuf = NULL; + } + + if (Lf != 8+3*s->img_n) return stbi__err("bad SOF len","Corrupt JPEG"); + + z->rgb = 0; + for (i=0; i < s->img_n; ++i) { + static unsigned char rgb[3] = { 'R', 'G', 'B' }; + z->img_comp[i].id = stbi__get8(s); + if (s->img_n == 3 && z->img_comp[i].id == rgb[i]) + ++z->rgb; + q = stbi__get8(s); + z->img_comp[i].h = (q >> 4); if (!z->img_comp[i].h || z->img_comp[i].h > 4) return stbi__err("bad H","Corrupt JPEG"); + z->img_comp[i].v = q & 15; if (!z->img_comp[i].v || z->img_comp[i].v > 4) return stbi__err("bad V","Corrupt JPEG"); + z->img_comp[i].tq = stbi__get8(s); if (z->img_comp[i].tq > 3) return stbi__err("bad TQ","Corrupt JPEG"); + } + + if (scan != STBI__SCAN_load) return 1; + + if (!stbi__mad3sizes_valid(s->img_x, s->img_y, s->img_n, 0)) return stbi__err("too large", "Image too large to decode"); + + for (i=0; i < s->img_n; ++i) { + if (z->img_comp[i].h > h_max) h_max = z->img_comp[i].h; + if (z->img_comp[i].v > v_max) v_max = z->img_comp[i].v; + } + + // compute interleaved mcu info + z->img_h_max = h_max; + z->img_v_max = v_max; + z->img_mcu_w = h_max * 8; + z->img_mcu_h = v_max * 8; + // these sizes can't be more than 17 bits + z->img_mcu_x = (s->img_x + z->img_mcu_w-1) / z->img_mcu_w; + z->img_mcu_y = (s->img_y + z->img_mcu_h-1) / z->img_mcu_h; + + for (i=0; i < s->img_n; ++i) { + // number of effective pixels (e.g. for non-interleaved MCU) + z->img_comp[i].x = (s->img_x * z->img_comp[i].h + h_max-1) / h_max; + z->img_comp[i].y = (s->img_y * z->img_comp[i].v + v_max-1) / v_max; + // to simplify generation, we'll allocate enough memory to decode + // the bogus oversized data from using interleaved MCUs and their + // big blocks (e.g. a 16x16 iMCU on an image of width 33); we won't + // discard the extra data until colorspace conversion + // + // img_mcu_x, img_mcu_y: <=17 bits; comp[i].h and .v are <=4 (checked earlier) + // so these muls can't overflow with 32-bit ints (which we require) + z->img_comp[i].w2 = z->img_mcu_x * z->img_comp[i].h * 8; + z->img_comp[i].h2 = z->img_mcu_y * z->img_comp[i].v * 8; + z->img_comp[i].coeff = 0; + z->img_comp[i].raw_coeff = 0; + z->img_comp[i].linebuf = NULL; + z->img_comp[i].raw_data = stbi__malloc_mad2(z->img_comp[i].w2, z->img_comp[i].h2, 15); + if (z->img_comp[i].raw_data == NULL) + return stbi__free_jpeg_components(z, i+1, stbi__err("outofmem", "Out of memory")); + // align blocks for idct using mmx/sse + z->img_comp[i].data = (stbi_uc*) (((size_t) z->img_comp[i].raw_data + 15) & ~15); + if (z->progressive) { + // w2, h2 are multiples of 8 (see above) + z->img_comp[i].coeff_w = z->img_comp[i].w2 / 8; + z->img_comp[i].coeff_h = z->img_comp[i].h2 / 8; + z->img_comp[i].raw_coeff = stbi__malloc_mad3(z->img_comp[i].w2, z->img_comp[i].h2, sizeof(short), 15); + if (z->img_comp[i].raw_coeff == NULL) + return stbi__free_jpeg_components(z, i+1, stbi__err("outofmem", "Out of memory")); + z->img_comp[i].coeff = (short*) (((size_t) z->img_comp[i].raw_coeff + 15) & ~15); + } + } + + return 1; +} + +// use comparisons since in some cases we handle more than one case (e.g. SOF) +#define stbi__DNL(x) ((x) == 0xdc) +#define stbi__SOI(x) ((x) == 0xd8) +#define stbi__EOI(x) ((x) == 0xd9) +#define stbi__SOF(x) ((x) == 0xc0 || (x) == 0xc1 || (x) == 0xc2) +#define stbi__SOS(x) ((x) == 0xda) + +#define stbi__SOF_progressive(x) ((x) == 0xc2) + +static int stbi__decode_jpeg_header(stbi__jpeg *z, int scan) +{ + int m; + z->jfif = 0; + z->app14_color_transform = -1; // valid values are 0,1,2 + z->marker = STBI__MARKER_none; // initialize cached marker to empty + m = stbi__get_marker(z); + if (!stbi__SOI(m)) return stbi__err("no SOI","Corrupt JPEG"); + if (scan == STBI__SCAN_type) return 1; + m = stbi__get_marker(z); + while (!stbi__SOF(m)) { + if (!stbi__process_marker(z,m)) return 0; + m = stbi__get_marker(z); + while (m == STBI__MARKER_none) { + // some files have extra padding after their blocks, so ok, we'll scan + if (stbi__at_eof(z->s)) return stbi__err("no SOF", "Corrupt JPEG"); + m = stbi__get_marker(z); + } + } + z->progressive = stbi__SOF_progressive(m); + if (!stbi__process_frame_header(z, scan)) return 0; + return 1; +} + +// decode image to YCbCr format +static int stbi__decode_jpeg_image(stbi__jpeg *j) +{ + int m; + for (m = 0; m < 4; m++) { + j->img_comp[m].raw_data = NULL; + j->img_comp[m].raw_coeff = NULL; + } + j->restart_interval = 0; + if (!stbi__decode_jpeg_header(j, STBI__SCAN_load)) return 0; + m = stbi__get_marker(j); + while (!stbi__EOI(m)) { + if (stbi__SOS(m)) { + if (!stbi__process_scan_header(j)) return 0; + if (!stbi__parse_entropy_coded_data(j)) return 0; + if (j->marker == STBI__MARKER_none ) { + // handle 0s at the end of image data from IP Kamera 9060 + while (!stbi__at_eof(j->s)) { + int x = stbi__get8(j->s); + if (x == 255) { + j->marker = stbi__get8(j->s); + break; + } + } + // if we reach eof without hitting a marker, stbi__get_marker() below will fail and we'll eventually return 0 + } + } else if (stbi__DNL(m)) { + int Ld = stbi__get16be(j->s); + stbi__uint32 NL = stbi__get16be(j->s); + if (Ld != 4) stbi__err("bad DNL len", "Corrupt JPEG"); + if (NL != j->s->img_y) stbi__err("bad DNL height", "Corrupt JPEG"); + } else { + if (!stbi__process_marker(j, m)) return 0; + } + m = stbi__get_marker(j); + } + if (j->progressive) + stbi__jpeg_finish(j); + return 1; +} + +// static jfif-centered resampling (across block boundaries) + +typedef stbi_uc *(*resample_row_func)(stbi_uc *out, stbi_uc *in0, stbi_uc *in1, + int w, int hs); + +#define stbi__div4(x) ((stbi_uc) ((x) >> 2)) + +static stbi_uc *resample_row_1(stbi_uc *out, stbi_uc *in_near, stbi_uc *in_far, int w, int hs) +{ + STBI_NOTUSED(out); + STBI_NOTUSED(in_far); + STBI_NOTUSED(w); + STBI_NOTUSED(hs); + return in_near; +} + +static stbi_uc* stbi__resample_row_v_2(stbi_uc *out, stbi_uc *in_near, stbi_uc *in_far, int w, int hs) +{ + // need to generate two samples vertically for every one in input + int i; + STBI_NOTUSED(hs); + for (i=0; i < w; ++i) + out[i] = stbi__div4(3*in_near[i] + in_far[i] + 2); + return out; +} + +static stbi_uc* stbi__resample_row_h_2(stbi_uc *out, stbi_uc *in_near, stbi_uc *in_far, int w, int hs) +{ + // need to generate two samples horizontally for every one in input + int i; + stbi_uc *input = in_near; + + if (w == 1) { + // if only one sample, can't do any interpolation + out[0] = out[1] = input[0]; + return out; + } + + out[0] = input[0]; + out[1] = stbi__div4(input[0]*3 + input[1] + 2); + for (i=1; i < w-1; ++i) { + int n = 3*input[i]+2; + out[i*2+0] = stbi__div4(n+input[i-1]); + out[i*2+1] = stbi__div4(n+input[i+1]); + } + out[i*2+0] = stbi__div4(input[w-2]*3 + input[w-1] + 2); + out[i*2+1] = input[w-1]; + + STBI_NOTUSED(in_far); + STBI_NOTUSED(hs); + + return out; +} + +#define stbi__div16(x) ((stbi_uc) ((x) >> 4)) + +static stbi_uc *stbi__resample_row_hv_2(stbi_uc *out, stbi_uc *in_near, stbi_uc *in_far, int w, int hs) +{ + // need to generate 2x2 samples for every one in input + int i,t0,t1; + if (w == 1) { + out[0] = out[1] = stbi__div4(3*in_near[0] + in_far[0] + 2); + return out; + } + + t1 = 3*in_near[0] + in_far[0]; + out[0] = stbi__div4(t1+2); + for (i=1; i < w; ++i) { + t0 = t1; + t1 = 3*in_near[i]+in_far[i]; + out[i*2-1] = stbi__div16(3*t0 + t1 + 8); + out[i*2 ] = stbi__div16(3*t1 + t0 + 8); + } + out[w*2-1] = stbi__div4(t1+2); + + STBI_NOTUSED(hs); + + return out; +} + +#if defined(STBI_SSE2) || defined(STBI_NEON) +static stbi_uc *stbi__resample_row_hv_2_simd(stbi_uc *out, stbi_uc *in_near, stbi_uc *in_far, int w, int hs) +{ + // need to generate 2x2 samples for every one in input + int i=0,t0,t1; + + if (w == 1) { + out[0] = out[1] = stbi__div4(3*in_near[0] + in_far[0] + 2); + return out; + } + + t1 = 3*in_near[0] + in_far[0]; + // process groups of 8 pixels for as long as we can. + // note we can't handle the last pixel in a row in this loop + // because we need to handle the filter boundary conditions. + for (; i < ((w-1) & ~7); i += 8) { +#if defined(STBI_SSE2) + // load and perform the vertical filtering pass + // this uses 3*x + y = 4*x + (y - x) + __m128i zero = _mm_setzero_si128(); + __m128i farb = _mm_loadl_epi64((__m128i *) (in_far + i)); + __m128i nearb = _mm_loadl_epi64((__m128i *) (in_near + i)); + __m128i farw = _mm_unpacklo_epi8(farb, zero); + __m128i nearw = _mm_unpacklo_epi8(nearb, zero); + __m128i diff = _mm_sub_epi16(farw, nearw); + __m128i nears = _mm_slli_epi16(nearw, 2); + __m128i curr = _mm_add_epi16(nears, diff); // current row + + // horizontal filter works the same based on shifted vers of current + // row. "prev" is current row shifted right by 1 pixel; we need to + // insert the previous pixel value (from t1). + // "next" is current row shifted left by 1 pixel, with first pixel + // of next block of 8 pixels added in. + __m128i prv0 = _mm_slli_si128(curr, 2); + __m128i nxt0 = _mm_srli_si128(curr, 2); + __m128i prev = _mm_insert_epi16(prv0, t1, 0); + __m128i next = _mm_insert_epi16(nxt0, 3*in_near[i+8] + in_far[i+8], 7); + + // horizontal filter, polyphase implementation since it's convenient: + // even pixels = 3*cur + prev = cur*4 + (prev - cur) + // odd pixels = 3*cur + next = cur*4 + (next - cur) + // note the shared term. + __m128i bias = _mm_set1_epi16(8); + __m128i curs = _mm_slli_epi16(curr, 2); + __m128i prvd = _mm_sub_epi16(prev, curr); + __m128i nxtd = _mm_sub_epi16(next, curr); + __m128i curb = _mm_add_epi16(curs, bias); + __m128i even = _mm_add_epi16(prvd, curb); + __m128i odd = _mm_add_epi16(nxtd, curb); + + // interleave even and odd pixels, then undo scaling. + __m128i int0 = _mm_unpacklo_epi16(even, odd); + __m128i int1 = _mm_unpackhi_epi16(even, odd); + __m128i de0 = _mm_srli_epi16(int0, 4); + __m128i de1 = _mm_srli_epi16(int1, 4); + + // pack and write output + __m128i outv = _mm_packus_epi16(de0, de1); + _mm_storeu_si128((__m128i *) (out + i*2), outv); +#elif defined(STBI_NEON) + // load and perform the vertical filtering pass + // this uses 3*x + y = 4*x + (y - x) + uint8x8_t farb = vld1_u8(in_far + i); + uint8x8_t nearb = vld1_u8(in_near + i); + int16x8_t diff = vreinterpretq_s16_u16(vsubl_u8(farb, nearb)); + int16x8_t nears = vreinterpretq_s16_u16(vshll_n_u8(nearb, 2)); + int16x8_t curr = vaddq_s16(nears, diff); // current row + + // horizontal filter works the same based on shifted vers of current + // row. "prev" is current row shifted right by 1 pixel; we need to + // insert the previous pixel value (from t1). + // "next" is current row shifted left by 1 pixel, with first pixel + // of next block of 8 pixels added in. + int16x8_t prv0 = vextq_s16(curr, curr, 7); + int16x8_t nxt0 = vextq_s16(curr, curr, 1); + int16x8_t prev = vsetq_lane_s16(t1, prv0, 0); + int16x8_t next = vsetq_lane_s16(3*in_near[i+8] + in_far[i+8], nxt0, 7); + + // horizontal filter, polyphase implementation since it's convenient: + // even pixels = 3*cur + prev = cur*4 + (prev - cur) + // odd pixels = 3*cur + next = cur*4 + (next - cur) + // note the shared term. + int16x8_t curs = vshlq_n_s16(curr, 2); + int16x8_t prvd = vsubq_s16(prev, curr); + int16x8_t nxtd = vsubq_s16(next, curr); + int16x8_t even = vaddq_s16(curs, prvd); + int16x8_t odd = vaddq_s16(curs, nxtd); + + // undo scaling and round, then store with even/odd phases interleaved + uint8x8x2_t o; + o.val[0] = vqrshrun_n_s16(even, 4); + o.val[1] = vqrshrun_n_s16(odd, 4); + vst2_u8(out + i*2, o); +#endif + + // "previous" value for next iter + t1 = 3*in_near[i+7] + in_far[i+7]; + } + + t0 = t1; + t1 = 3*in_near[i] + in_far[i]; + out[i*2] = stbi__div16(3*t1 + t0 + 8); + + for (++i; i < w; ++i) { + t0 = t1; + t1 = 3*in_near[i]+in_far[i]; + out[i*2-1] = stbi__div16(3*t0 + t1 + 8); + out[i*2 ] = stbi__div16(3*t1 + t0 + 8); + } + out[w*2-1] = stbi__div4(t1+2); + + STBI_NOTUSED(hs); + + return out; +} +#endif + +static stbi_uc *stbi__resample_row_generic(stbi_uc *out, stbi_uc *in_near, stbi_uc *in_far, int w, int hs) +{ + // resample with nearest-neighbor + int i,j; + STBI_NOTUSED(in_far); + for (i=0; i < w; ++i) + for (j=0; j < hs; ++j) + out[i*hs+j] = in_near[i]; + return out; +} + +// this is a reduced-precision calculation of YCbCr-to-RGB introduced +// to make sure the code produces the same results in both SIMD and scalar +#define stbi__float2fixed(x) (((int) ((x) * 4096.0f + 0.5f)) << 8) +static void stbi__YCbCr_to_RGB_row(stbi_uc *out, const stbi_uc *y, const stbi_uc *pcb, const stbi_uc *pcr, int count, int step) +{ + int i; + for (i=0; i < count; ++i) { + int y_fixed = (y[i] << 20) + (1<<19); // rounding + int r,g,b; + int cr = pcr[i] - 128; + int cb = pcb[i] - 128; + r = y_fixed + cr* stbi__float2fixed(1.40200f); + g = y_fixed + (cr*-stbi__float2fixed(0.71414f)) + ((cb*-stbi__float2fixed(0.34414f)) & 0xffff0000); + b = y_fixed + cb* stbi__float2fixed(1.77200f); + r >>= 20; + g >>= 20; + b >>= 20; + if ((unsigned) r > 255) { if (r < 0) r = 0; else r = 255; } + if ((unsigned) g > 255) { if (g < 0) g = 0; else g = 255; } + if ((unsigned) b > 255) { if (b < 0) b = 0; else b = 255; } + out[0] = (stbi_uc)r; + out[1] = (stbi_uc)g; + out[2] = (stbi_uc)b; + out[3] = 255; + out += step; + } +} + +#if defined(STBI_SSE2) || defined(STBI_NEON) +static void stbi__YCbCr_to_RGB_simd(stbi_uc *out, stbi_uc const *y, stbi_uc const *pcb, stbi_uc const *pcr, int count, int step) +{ + int i = 0; + +#ifdef STBI_SSE2 + // step == 3 is pretty ugly on the final interleave, and i'm not convinced + // it's useful in practice (you wouldn't use it for textures, for example). + // so just accelerate step == 4 case. + if (step == 4) { + // this is a fairly straightforward implementation and not super-optimized. + __m128i signflip = _mm_set1_epi8(-0x80); + __m128i cr_const0 = _mm_set1_epi16( (short) ( 1.40200f*4096.0f+0.5f)); + __m128i cr_const1 = _mm_set1_epi16( - (short) ( 0.71414f*4096.0f+0.5f)); + __m128i cb_const0 = _mm_set1_epi16( - (short) ( 0.34414f*4096.0f+0.5f)); + __m128i cb_const1 = _mm_set1_epi16( (short) ( 1.77200f*4096.0f+0.5f)); + __m128i y_bias = _mm_set1_epi8((char) (unsigned char) 128); + __m128i xw = _mm_set1_epi16(255); // alpha channel + + for (; i+7 < count; i += 8) { + // load + __m128i y_bytes = _mm_loadl_epi64((__m128i *) (y+i)); + __m128i cr_bytes = _mm_loadl_epi64((__m128i *) (pcr+i)); + __m128i cb_bytes = _mm_loadl_epi64((__m128i *) (pcb+i)); + __m128i cr_biased = _mm_xor_si128(cr_bytes, signflip); // -128 + __m128i cb_biased = _mm_xor_si128(cb_bytes, signflip); // -128 + + // unpack to short (and left-shift cr, cb by 8) + __m128i yw = _mm_unpacklo_epi8(y_bias, y_bytes); + __m128i crw = _mm_unpacklo_epi8(_mm_setzero_si128(), cr_biased); + __m128i cbw = _mm_unpacklo_epi8(_mm_setzero_si128(), cb_biased); + + // color transform + __m128i yws = _mm_srli_epi16(yw, 4); + __m128i cr0 = _mm_mulhi_epi16(cr_const0, crw); + __m128i cb0 = _mm_mulhi_epi16(cb_const0, cbw); + __m128i cb1 = _mm_mulhi_epi16(cbw, cb_const1); + __m128i cr1 = _mm_mulhi_epi16(crw, cr_const1); + __m128i rws = _mm_add_epi16(cr0, yws); + __m128i gwt = _mm_add_epi16(cb0, yws); + __m128i bws = _mm_add_epi16(yws, cb1); + __m128i gws = _mm_add_epi16(gwt, cr1); + + // descale + __m128i rw = _mm_srai_epi16(rws, 4); + __m128i bw = _mm_srai_epi16(bws, 4); + __m128i gw = _mm_srai_epi16(gws, 4); + + // back to byte, set up for transpose + __m128i brb = _mm_packus_epi16(rw, bw); + __m128i gxb = _mm_packus_epi16(gw, xw); + + // transpose to interleave channels + __m128i t0 = _mm_unpacklo_epi8(brb, gxb); + __m128i t1 = _mm_unpackhi_epi8(brb, gxb); + __m128i o0 = _mm_unpacklo_epi16(t0, t1); + __m128i o1 = _mm_unpackhi_epi16(t0, t1); + + // store + _mm_storeu_si128((__m128i *) (out + 0), o0); + _mm_storeu_si128((__m128i *) (out + 16), o1); + out += 32; + } + } +#endif + +#ifdef STBI_NEON + // in this version, step=3 support would be easy to add. but is there demand? + if (step == 4) { + // this is a fairly straightforward implementation and not super-optimized. + uint8x8_t signflip = vdup_n_u8(0x80); + int16x8_t cr_const0 = vdupq_n_s16( (short) ( 1.40200f*4096.0f+0.5f)); + int16x8_t cr_const1 = vdupq_n_s16( - (short) ( 0.71414f*4096.0f+0.5f)); + int16x8_t cb_const0 = vdupq_n_s16( - (short) ( 0.34414f*4096.0f+0.5f)); + int16x8_t cb_const1 = vdupq_n_s16( (short) ( 1.77200f*4096.0f+0.5f)); + + for (; i+7 < count; i += 8) { + // load + uint8x8_t y_bytes = vld1_u8(y + i); + uint8x8_t cr_bytes = vld1_u8(pcr + i); + uint8x8_t cb_bytes = vld1_u8(pcb + i); + int8x8_t cr_biased = vreinterpret_s8_u8(vsub_u8(cr_bytes, signflip)); + int8x8_t cb_biased = vreinterpret_s8_u8(vsub_u8(cb_bytes, signflip)); + + // expand to s16 + int16x8_t yws = vreinterpretq_s16_u16(vshll_n_u8(y_bytes, 4)); + int16x8_t crw = vshll_n_s8(cr_biased, 7); + int16x8_t cbw = vshll_n_s8(cb_biased, 7); + + // color transform + int16x8_t cr0 = vqdmulhq_s16(crw, cr_const0); + int16x8_t cb0 = vqdmulhq_s16(cbw, cb_const0); + int16x8_t cr1 = vqdmulhq_s16(crw, cr_const1); + int16x8_t cb1 = vqdmulhq_s16(cbw, cb_const1); + int16x8_t rws = vaddq_s16(yws, cr0); + int16x8_t gws = vaddq_s16(vaddq_s16(yws, cb0), cr1); + int16x8_t bws = vaddq_s16(yws, cb1); + + // undo scaling, round, convert to byte + uint8x8x4_t o; + o.val[0] = vqrshrun_n_s16(rws, 4); + o.val[1] = vqrshrun_n_s16(gws, 4); + o.val[2] = vqrshrun_n_s16(bws, 4); + o.val[3] = vdup_n_u8(255); + + // store, interleaving r/g/b/a + vst4_u8(out, o); + out += 8*4; + } + } +#endif + + for (; i < count; ++i) { + int y_fixed = (y[i] << 20) + (1<<19); // rounding + int r,g,b; + int cr = pcr[i] - 128; + int cb = pcb[i] - 128; + r = y_fixed + cr* stbi__float2fixed(1.40200f); + g = y_fixed + cr*-stbi__float2fixed(0.71414f) + ((cb*-stbi__float2fixed(0.34414f)) & 0xffff0000); + b = y_fixed + cb* stbi__float2fixed(1.77200f); + r >>= 20; + g >>= 20; + b >>= 20; + if ((unsigned) r > 255) { if (r < 0) r = 0; else r = 255; } + if ((unsigned) g > 255) { if (g < 0) g = 0; else g = 255; } + if ((unsigned) b > 255) { if (b < 0) b = 0; else b = 255; } + out[0] = (stbi_uc)r; + out[1] = (stbi_uc)g; + out[2] = (stbi_uc)b; + out[3] = 255; + out += step; + } +} +#endif + +// set up the kernels +static void stbi__setup_jpeg(stbi__jpeg *j) +{ + j->idct_block_kernel = stbi__idct_block; + j->YCbCr_to_RGB_kernel = stbi__YCbCr_to_RGB_row; + j->resample_row_hv_2_kernel = stbi__resample_row_hv_2; + +#ifdef STBI_SSE2 + if (stbi__sse2_available()) { + j->idct_block_kernel = stbi__idct_simd; + j->YCbCr_to_RGB_kernel = stbi__YCbCr_to_RGB_simd; + j->resample_row_hv_2_kernel = stbi__resample_row_hv_2_simd; + } +#endif + +#ifdef STBI_NEON + j->idct_block_kernel = stbi__idct_simd; + j->YCbCr_to_RGB_kernel = stbi__YCbCr_to_RGB_simd; + j->resample_row_hv_2_kernel = stbi__resample_row_hv_2_simd; +#endif +} + +// clean up the temporary component buffers +static void stbi__cleanup_jpeg(stbi__jpeg *j) +{ + stbi__free_jpeg_components(j, j->s->img_n, 0); +} + +typedef struct +{ + resample_row_func resample; + stbi_uc *line0,*line1; + int hs,vs; // expansion factor in each axis + int w_lores; // horizontal pixels pre-expansion + int ystep; // how far through vertical expansion we are + int ypos; // which pre-expansion row we're on +} stbi__resample; + +// fast 0..255 * 0..255 => 0..255 rounded multiplication +static stbi_uc stbi__blinn_8x8(stbi_uc x, stbi_uc y) +{ + unsigned int t = x*y + 128; + return (stbi_uc) ((t + (t >>8)) >> 8); +} + +static stbi_uc *load_jpeg_image(stbi__jpeg *z, int *out_x, int *out_y, int *comp, int req_comp) +{ + int n, decode_n, is_rgb; + z->s->img_n = 0; // make stbi__cleanup_jpeg safe + + // validate req_comp + if (req_comp < 0 || req_comp > 4) return stbi__errpuc("bad req_comp", "Internal error"); + + // load a jpeg image from whichever source, but leave in YCbCr format + if (!stbi__decode_jpeg_image(z)) { stbi__cleanup_jpeg(z); return NULL; } + + // determine actual number of components to generate + n = req_comp ? req_comp : z->s->img_n >= 3 ? 3 : 1; + + is_rgb = z->s->img_n == 3 && (z->rgb == 3 || (z->app14_color_transform == 0 && !z->jfif)); + + if (z->s->img_n == 3 && n < 3 && !is_rgb) + decode_n = 1; + else + decode_n = z->s->img_n; + + // resample and color-convert + { + int k; + unsigned int i,j; + stbi_uc *output; + stbi_uc *coutput[4]; + + stbi__resample res_comp[4]; + + for (k=0; k < decode_n; ++k) { + stbi__resample *r = &res_comp[k]; + + // allocate line buffer big enough for upsampling off the edges + // with upsample factor of 4 + z->img_comp[k].linebuf = (stbi_uc *) stbi__malloc(z->s->img_x + 3); + if (!z->img_comp[k].linebuf) { stbi__cleanup_jpeg(z); return stbi__errpuc("outofmem", "Out of memory"); } + + r->hs = z->img_h_max / z->img_comp[k].h; + r->vs = z->img_v_max / z->img_comp[k].v; + r->ystep = r->vs >> 1; + r->w_lores = (z->s->img_x + r->hs-1) / r->hs; + r->ypos = 0; + r->line0 = r->line1 = z->img_comp[k].data; + + if (r->hs == 1 && r->vs == 1) r->resample = resample_row_1; + else if (r->hs == 1 && r->vs == 2) r->resample = stbi__resample_row_v_2; + else if (r->hs == 2 && r->vs == 1) r->resample = stbi__resample_row_h_2; + else if (r->hs == 2 && r->vs == 2) r->resample = z->resample_row_hv_2_kernel; + else r->resample = stbi__resample_row_generic; + } + + // can't error after this so, this is safe + output = (stbi_uc *) stbi__malloc_mad3(n, z->s->img_x, z->s->img_y, 1); + if (!output) { stbi__cleanup_jpeg(z); return stbi__errpuc("outofmem", "Out of memory"); } + + // now go ahead and resample + for (j=0; j < z->s->img_y; ++j) { + stbi_uc *out = output + n * z->s->img_x * j; + for (k=0; k < decode_n; ++k) { + stbi__resample *r = &res_comp[k]; + int y_bot = r->ystep >= (r->vs >> 1); + coutput[k] = r->resample(z->img_comp[k].linebuf, + y_bot ? r->line1 : r->line0, + y_bot ? r->line0 : r->line1, + r->w_lores, r->hs); + if (++r->ystep >= r->vs) { + r->ystep = 0; + r->line0 = r->line1; + if (++r->ypos < z->img_comp[k].y) + r->line1 += z->img_comp[k].w2; + } + } + if (n >= 3) { + stbi_uc *y = coutput[0]; + if (z->s->img_n == 3) { + if (is_rgb) { + for (i=0; i < z->s->img_x; ++i) { + out[0] = y[i]; + out[1] = coutput[1][i]; + out[2] = coutput[2][i]; + out[3] = 255; + out += n; + } + } else { + z->YCbCr_to_RGB_kernel(out, y, coutput[1], coutput[2], z->s->img_x, n); + } + } else if (z->s->img_n == 4) { + if (z->app14_color_transform == 0) { // CMYK + for (i=0; i < z->s->img_x; ++i) { + stbi_uc m = coutput[3][i]; + out[0] = stbi__blinn_8x8(coutput[0][i], m); + out[1] = stbi__blinn_8x8(coutput[1][i], m); + out[2] = stbi__blinn_8x8(coutput[2][i], m); + out[3] = 255; + out += n; + } + } else if (z->app14_color_transform == 2) { // YCCK + z->YCbCr_to_RGB_kernel(out, y, coutput[1], coutput[2], z->s->img_x, n); + for (i=0; i < z->s->img_x; ++i) { + stbi_uc m = coutput[3][i]; + out[0] = stbi__blinn_8x8(255 - out[0], m); + out[1] = stbi__blinn_8x8(255 - out[1], m); + out[2] = stbi__blinn_8x8(255 - out[2], m); + out += n; + } + } else { // YCbCr + alpha? Ignore the fourth channel for now + z->YCbCr_to_RGB_kernel(out, y, coutput[1], coutput[2], z->s->img_x, n); + } + } else + for (i=0; i < z->s->img_x; ++i) { + out[0] = out[1] = out[2] = y[i]; + out[3] = 255; // not used if n==3 + out += n; + } + } else { + if (is_rgb) { + if (n == 1) + for (i=0; i < z->s->img_x; ++i) + *out++ = stbi__compute_y(coutput[0][i], coutput[1][i], coutput[2][i]); + else { + for (i=0; i < z->s->img_x; ++i, out += 2) { + out[0] = stbi__compute_y(coutput[0][i], coutput[1][i], coutput[2][i]); + out[1] = 255; + } + } + } else if (z->s->img_n == 4 && z->app14_color_transform == 0) { + for (i=0; i < z->s->img_x; ++i) { + stbi_uc m = coutput[3][i]; + stbi_uc r = stbi__blinn_8x8(coutput[0][i], m); + stbi_uc g = stbi__blinn_8x8(coutput[1][i], m); + stbi_uc b = stbi__blinn_8x8(coutput[2][i], m); + out[0] = stbi__compute_y(r, g, b); + out[1] = 255; + out += n; + } + } else if (z->s->img_n == 4 && z->app14_color_transform == 2) { + for (i=0; i < z->s->img_x; ++i) { + out[0] = stbi__blinn_8x8(255 - coutput[0][i], coutput[3][i]); + out[1] = 255; + out += n; + } + } else { + stbi_uc *y = coutput[0]; + if (n == 1) + for (i=0; i < z->s->img_x; ++i) out[i] = y[i]; + else + for (i=0; i < z->s->img_x; ++i) *out++ = y[i], *out++ = 255; + } + } + } + stbi__cleanup_jpeg(z); + *out_x = z->s->img_x; + *out_y = z->s->img_y; + if (comp) *comp = z->s->img_n >= 3 ? 3 : 1; // report original components, not output + return output; + } +} + +static void *stbi__jpeg_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri) +{ + unsigned char* result; + stbi__jpeg* j = (stbi__jpeg*) stbi__malloc(sizeof(stbi__jpeg)); + STBI_NOTUSED(ri); + j->s = s; + stbi__setup_jpeg(j); + result = load_jpeg_image(j, x,y,comp,req_comp); + STBI_FREE(j); + return result; +} + +static int stbi__jpeg_test(stbi__context *s) +{ + int r; + stbi__jpeg* j = (stbi__jpeg*)stbi__malloc(sizeof(stbi__jpeg)); + j->s = s; + stbi__setup_jpeg(j); + r = stbi__decode_jpeg_header(j, STBI__SCAN_type); + stbi__rewind(s); + STBI_FREE(j); + return r; +} + +static int stbi__jpeg_info_raw(stbi__jpeg *j, int *x, int *y, int *comp) +{ + if (!stbi__decode_jpeg_header(j, STBI__SCAN_header)) { + stbi__rewind( j->s ); + return 0; + } + if (x) *x = j->s->img_x; + if (y) *y = j->s->img_y; + if (comp) *comp = j->s->img_n >= 3 ? 3 : 1; + return 1; +} + +static int stbi__jpeg_info(stbi__context *s, int *x, int *y, int *comp) +{ + int result; + stbi__jpeg* j = (stbi__jpeg*) (stbi__malloc(sizeof(stbi__jpeg))); + j->s = s; + result = stbi__jpeg_info_raw(j, x, y, comp); + STBI_FREE(j); + return result; +} +#endif + +// public domain zlib decode v0.2 Sean Barrett 2006-11-18 +// simple implementation +// - all input must be provided in an upfront buffer +// - all output is written to a single output buffer (can malloc/realloc) +// performance +// - fast huffman + +#ifndef STBI_NO_ZLIB + +// fast-way is faster to check than jpeg huffman, but slow way is slower +#define STBI__ZFAST_BITS 9 // accelerate all cases in default tables +#define STBI__ZFAST_MASK ((1 << STBI__ZFAST_BITS) - 1) + +// zlib-style huffman encoding +// (jpegs packs from left, zlib from right, so can't share code) +typedef struct +{ + stbi__uint16 fast[1 << STBI__ZFAST_BITS]; + stbi__uint16 firstcode[16]; + int maxcode[17]; + stbi__uint16 firstsymbol[16]; + stbi_uc size[288]; + stbi__uint16 value[288]; +} stbi__zhuffman; + +stbi_inline static int stbi__bitreverse16(int n) +{ + n = ((n & 0xAAAA) >> 1) | ((n & 0x5555) << 1); + n = ((n & 0xCCCC) >> 2) | ((n & 0x3333) << 2); + n = ((n & 0xF0F0) >> 4) | ((n & 0x0F0F) << 4); + n = ((n & 0xFF00) >> 8) | ((n & 0x00FF) << 8); + return n; +} + +stbi_inline static int stbi__bit_reverse(int v, int bits) +{ + STBI_ASSERT(bits <= 16); + // to bit reverse n bits, reverse 16 and shift + // e.g. 11 bits, bit reverse and shift away 5 + return stbi__bitreverse16(v) >> (16-bits); +} + +static int stbi__zbuild_huffman(stbi__zhuffman *z, const stbi_uc *sizelist, int num) +{ + int i,k=0; + int code, next_code[16], sizes[17]; + + // DEFLATE spec for generating codes + memset(sizes, 0, sizeof(sizes)); + memset(z->fast, 0, sizeof(z->fast)); + for (i=0; i < num; ++i) + ++sizes[sizelist[i]]; + sizes[0] = 0; + for (i=1; i < 16; ++i) + if (sizes[i] > (1 << i)) + return stbi__err("bad sizes", "Corrupt PNG"); + code = 0; + for (i=1; i < 16; ++i) { + next_code[i] = code; + z->firstcode[i] = (stbi__uint16) code; + z->firstsymbol[i] = (stbi__uint16) k; + code = (code + sizes[i]); + if (sizes[i]) + if (code-1 >= (1 << i)) return stbi__err("bad codelengths","Corrupt PNG"); + z->maxcode[i] = code << (16-i); // preshift for inner loop + code <<= 1; + k += sizes[i]; + } + z->maxcode[16] = 0x10000; // sentinel + for (i=0; i < num; ++i) { + int s = sizelist[i]; + if (s) { + int c = next_code[s] - z->firstcode[s] + z->firstsymbol[s]; + stbi__uint16 fastv = (stbi__uint16) ((s << 9) | i); + z->size [c] = (stbi_uc ) s; + z->value[c] = (stbi__uint16) i; + if (s <= STBI__ZFAST_BITS) { + int j = stbi__bit_reverse(next_code[s],s); + while (j < (1 << STBI__ZFAST_BITS)) { + z->fast[j] = fastv; + j += (1 << s); + } + } + ++next_code[s]; + } + } + return 1; +} + +// zlib-from-memory implementation for PNG reading +// because PNG allows splitting the zlib stream arbitrarily, +// and it's annoying structurally to have PNG call ZLIB call PNG, +// we require PNG read all the IDATs and combine them into a single +// memory buffer + +typedef struct +{ + stbi_uc *zbuffer, *zbuffer_end; + int num_bits; + stbi__uint32 code_buffer; + + char *zout; + char *zout_start; + char *zout_end; + int z_expandable; + + stbi__zhuffman z_length, z_distance; +} stbi__zbuf; + +stbi_inline static stbi_uc stbi__zget8(stbi__zbuf *z) +{ + if (z->zbuffer >= z->zbuffer_end) return 0; + return *z->zbuffer++; +} + +static void stbi__fill_bits(stbi__zbuf *z) +{ + do { + STBI_ASSERT(z->code_buffer < (1U << z->num_bits)); + z->code_buffer |= (unsigned int) stbi__zget8(z) << z->num_bits; + z->num_bits += 8; + } while (z->num_bits <= 24); +} + +stbi_inline static unsigned int stbi__zreceive(stbi__zbuf *z, int n) +{ + unsigned int k; + if (z->num_bits < n) stbi__fill_bits(z); + k = z->code_buffer & ((1 << n) - 1); + z->code_buffer >>= n; + z->num_bits -= n; + return k; +} + +static int stbi__zhuffman_decode_slowpath(stbi__zbuf *a, stbi__zhuffman *z) +{ + int b,s,k; + // not resolved by fast table, so compute it the slow way + // use jpeg approach, which requires MSbits at top + k = stbi__bit_reverse(a->code_buffer, 16); + for (s=STBI__ZFAST_BITS+1; ; ++s) + if (k < z->maxcode[s]) + break; + if (s == 16) return -1; // invalid code! + // code size is s, so: + b = (k >> (16-s)) - z->firstcode[s] + z->firstsymbol[s]; + STBI_ASSERT(z->size[b] == s); + a->code_buffer >>= s; + a->num_bits -= s; + return z->value[b]; +} + +stbi_inline static int stbi__zhuffman_decode(stbi__zbuf *a, stbi__zhuffman *z) +{ + int b,s; + if (a->num_bits < 16) stbi__fill_bits(a); + b = z->fast[a->code_buffer & STBI__ZFAST_MASK]; + if (b) { + s = b >> 9; + a->code_buffer >>= s; + a->num_bits -= s; + return b & 511; + } + return stbi__zhuffman_decode_slowpath(a, z); +} + +static int stbi__zexpand(stbi__zbuf *z, char *zout, int n) // need to make room for n bytes +{ + char *q; + int cur, limit, old_limit; + z->zout = zout; + if (!z->z_expandable) return stbi__err("output buffer limit","Corrupt PNG"); + cur = (int) (z->zout - z->zout_start); + limit = old_limit = (int) (z->zout_end - z->zout_start); + while (cur + n > limit) + limit *= 2; + q = (char *) STBI_REALLOC_SIZED(z->zout_start, old_limit, limit); + STBI_NOTUSED(old_limit); + if (q == NULL) return stbi__err("outofmem", "Out of memory"); + z->zout_start = q; + z->zout = q + cur; + z->zout_end = q + limit; + return 1; +} + +static int stbi__zlength_base[31] = { + 3,4,5,6,7,8,9,10,11,13, + 15,17,19,23,27,31,35,43,51,59, + 67,83,99,115,131,163,195,227,258,0,0 }; + +static int stbi__zlength_extra[31]= +{ 0,0,0,0,0,0,0,0,1,1,1,1,2,2,2,2,3,3,3,3,4,4,4,4,5,5,5,5,0,0,0 }; + +static int stbi__zdist_base[32] = { 1,2,3,4,5,7,9,13,17,25,33,49,65,97,129,193, +257,385,513,769,1025,1537,2049,3073,4097,6145,8193,12289,16385,24577,0,0}; + +static int stbi__zdist_extra[32] = +{ 0,0,0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7,8,8,9,9,10,10,11,11,12,12,13,13}; + +static int stbi__parse_huffman_block(stbi__zbuf *a) +{ + char *zout = a->zout; + for(;;) { + int z = stbi__zhuffman_decode(a, &a->z_length); + if (z < 256) { + if (z < 0) return stbi__err("bad huffman code","Corrupt PNG"); // error in huffman codes + if (zout >= a->zout_end) { + if (!stbi__zexpand(a, zout, 1)) return 0; + zout = a->zout; + } + *zout++ = (char) z; + } else { + stbi_uc *p; + int len,dist; + if (z == 256) { + a->zout = zout; + return 1; + } + z -= 257; + len = stbi__zlength_base[z]; + if (stbi__zlength_extra[z]) len += stbi__zreceive(a, stbi__zlength_extra[z]); + z = stbi__zhuffman_decode(a, &a->z_distance); + if (z < 0) return stbi__err("bad huffman code","Corrupt PNG"); + dist = stbi__zdist_base[z]; + if (stbi__zdist_extra[z]) dist += stbi__zreceive(a, stbi__zdist_extra[z]); + if (zout - a->zout_start < dist) return stbi__err("bad dist","Corrupt PNG"); + if (zout + len > a->zout_end) { + if (!stbi__zexpand(a, zout, len)) return 0; + zout = a->zout; + } + p = (stbi_uc *) (zout - dist); + if (dist == 1) { // run of one byte; common in images. + stbi_uc v = *p; + if (len) { do *zout++ = v; while (--len); } + } else { + if (len) { do *zout++ = *p++; while (--len); } + } + } + } +} + +static int stbi__compute_huffman_codes(stbi__zbuf *a) +{ + static stbi_uc length_dezigzag[19] = { 16,17,18,0,8,7,9,6,10,5,11,4,12,3,13,2,14,1,15 }; + stbi__zhuffman z_codelength; + stbi_uc lencodes[286+32+137];//padding for maximum single op + stbi_uc codelength_sizes[19]; + int i,n; + + int hlit = stbi__zreceive(a,5) + 257; + int hdist = stbi__zreceive(a,5) + 1; + int hclen = stbi__zreceive(a,4) + 4; + int ntot = hlit + hdist; + + memset(codelength_sizes, 0, sizeof(codelength_sizes)); + for (i=0; i < hclen; ++i) { + int s = stbi__zreceive(a,3); + codelength_sizes[length_dezigzag[i]] = (stbi_uc) s; + } + if (!stbi__zbuild_huffman(&z_codelength, codelength_sizes, 19)) return 0; + + n = 0; + while (n < ntot) { + int c = stbi__zhuffman_decode(a, &z_codelength); + if (c < 0 || c >= 19) return stbi__err("bad codelengths", "Corrupt PNG"); + if (c < 16) + lencodes[n++] = (stbi_uc) c; + else { + stbi_uc fill = 0; + if (c == 16) { + c = stbi__zreceive(a,2)+3; + if (n == 0) return stbi__err("bad codelengths", "Corrupt PNG"); + fill = lencodes[n-1]; + } else if (c == 17) + c = stbi__zreceive(a,3)+3; + else { + STBI_ASSERT(c == 18); + c = stbi__zreceive(a,7)+11; + } + if (ntot - n < c) return stbi__err("bad codelengths", "Corrupt PNG"); + memset(lencodes+n, fill, c); + n += c; + } + } + if (n != ntot) return stbi__err("bad codelengths","Corrupt PNG"); + if (!stbi__zbuild_huffman(&a->z_length, lencodes, hlit)) return 0; + if (!stbi__zbuild_huffman(&a->z_distance, lencodes+hlit, hdist)) return 0; + return 1; +} + +static int stbi__parse_uncompressed_block(stbi__zbuf *a) +{ + stbi_uc header[4]; + int len,nlen,k; + if (a->num_bits & 7) + stbi__zreceive(a, a->num_bits & 7); // discard + // drain the bit-packed data into header + k = 0; + while (a->num_bits > 0) { + header[k++] = (stbi_uc) (a->code_buffer & 255); // suppress MSVC run-time check + a->code_buffer >>= 8; + a->num_bits -= 8; + } + STBI_ASSERT(a->num_bits == 0); + // now fill header the normal way + while (k < 4) + header[k++] = stbi__zget8(a); + len = header[1] * 256 + header[0]; + nlen = header[3] * 256 + header[2]; + if (nlen != (len ^ 0xffff)) return stbi__err("zlib corrupt","Corrupt PNG"); + if (a->zbuffer + len > a->zbuffer_end) return stbi__err("read past buffer","Corrupt PNG"); + if (a->zout + len > a->zout_end) + if (!stbi__zexpand(a, a->zout, len)) return 0; + memcpy(a->zout, a->zbuffer, len); + a->zbuffer += len; + a->zout += len; + return 1; +} + +static int stbi__parse_zlib_header(stbi__zbuf *a) +{ + int cmf = stbi__zget8(a); + int cm = cmf & 15; + /* int cinfo = cmf >> 4; */ + int flg = stbi__zget8(a); + if ((cmf*256+flg) % 31 != 0) return stbi__err("bad zlib header","Corrupt PNG"); // zlib spec + if (flg & 32) return stbi__err("no preset dict","Corrupt PNG"); // preset dictionary not allowed in png + if (cm != 8) return stbi__err("bad compression","Corrupt PNG"); // DEFLATE required for png + // window = 1 << (8 + cinfo)... but who cares, we fully buffer output + return 1; +} + +static const stbi_uc stbi__zdefault_length[288] = +{ + 8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8, 8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8, + 8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8, 8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8, + 8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8, 8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8, + 8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8, 8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8, + 8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8, 9,9,9,9,9,9,9,9,9,9,9,9,9,9,9,9, + 9,9,9,9,9,9,9,9,9,9,9,9,9,9,9,9, 9,9,9,9,9,9,9,9,9,9,9,9,9,9,9,9, + 9,9,9,9,9,9,9,9,9,9,9,9,9,9,9,9, 9,9,9,9,9,9,9,9,9,9,9,9,9,9,9,9, + 9,9,9,9,9,9,9,9,9,9,9,9,9,9,9,9, 9,9,9,9,9,9,9,9,9,9,9,9,9,9,9,9, + 7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7, 7,7,7,7,7,7,7,7,8,8,8,8,8,8,8,8 +}; +static const stbi_uc stbi__zdefault_distance[32] = +{ + 5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5 +}; +/* +Init algorithm: +{ + int i; // use <= to match clearly with spec + for (i=0; i <= 143; ++i) stbi__zdefault_length[i] = 8; + for ( ; i <= 255; ++i) stbi__zdefault_length[i] = 9; + for ( ; i <= 279; ++i) stbi__zdefault_length[i] = 7; + for ( ; i <= 287; ++i) stbi__zdefault_length[i] = 8; + + for (i=0; i <= 31; ++i) stbi__zdefault_distance[i] = 5; +} +*/ + +static int stbi__parse_zlib(stbi__zbuf *a, int parse_header) +{ + int final, type; + if (parse_header) + if (!stbi__parse_zlib_header(a)) return 0; + a->num_bits = 0; + a->code_buffer = 0; + do { + final = stbi__zreceive(a,1); + type = stbi__zreceive(a,2); + if (type == 0) { + if (!stbi__parse_uncompressed_block(a)) return 0; + } else if (type == 3) { + return 0; + } else { + if (type == 1) { + // use fixed code lengths + if (!stbi__zbuild_huffman(&a->z_length , stbi__zdefault_length , 288)) return 0; + if (!stbi__zbuild_huffman(&a->z_distance, stbi__zdefault_distance, 32)) return 0; + } else { + if (!stbi__compute_huffman_codes(a)) return 0; + } + if (!stbi__parse_huffman_block(a)) return 0; + } + } while (!final); + return 1; +} + +static int stbi__do_zlib(stbi__zbuf *a, char *obuf, int olen, int exp, int parse_header) +{ + a->zout_start = obuf; + a->zout = obuf; + a->zout_end = obuf + olen; + a->z_expandable = exp; + + return stbi__parse_zlib(a, parse_header); +} + +STBIDEF char *stbi_zlib_decode_malloc_guesssize(const char *buffer, int len, int initial_size, int *outlen) +{ + stbi__zbuf a; + char *p = (char *) stbi__malloc(initial_size); + if (p == NULL) return NULL; + a.zbuffer = (stbi_uc *) buffer; + a.zbuffer_end = (stbi_uc *) buffer + len; + if (stbi__do_zlib(&a, p, initial_size, 1, 1)) { + if (outlen) *outlen = (int) (a.zout - a.zout_start); + return a.zout_start; + } else { + STBI_FREE(a.zout_start); + return NULL; + } +} + +STBIDEF char *stbi_zlib_decode_malloc(char const *buffer, int len, int *outlen) +{ + return stbi_zlib_decode_malloc_guesssize(buffer, len, 16384, outlen); +} + +STBIDEF char *stbi_zlib_decode_malloc_guesssize_headerflag(const char *buffer, int len, int initial_size, int *outlen, int parse_header) +{ + stbi__zbuf a; + char *p = (char *) stbi__malloc(initial_size); + if (p == NULL) return NULL; + a.zbuffer = (stbi_uc *) buffer; + a.zbuffer_end = (stbi_uc *) buffer + len; + if (stbi__do_zlib(&a, p, initial_size, 1, parse_header)) { + if (outlen) *outlen = (int) (a.zout - a.zout_start); + return a.zout_start; + } else { + STBI_FREE(a.zout_start); + return NULL; + } +} + +STBIDEF int stbi_zlib_decode_buffer(char *obuffer, int olen, char const *ibuffer, int ilen) +{ + stbi__zbuf a; + a.zbuffer = (stbi_uc *) ibuffer; + a.zbuffer_end = (stbi_uc *) ibuffer + ilen; + if (stbi__do_zlib(&a, obuffer, olen, 0, 1)) + return (int) (a.zout - a.zout_start); + else + return -1; +} + +STBIDEF char *stbi_zlib_decode_noheader_malloc(char const *buffer, int len, int *outlen) +{ + stbi__zbuf a; + char *p = (char *) stbi__malloc(16384); + if (p == NULL) return NULL; + a.zbuffer = (stbi_uc *) buffer; + a.zbuffer_end = (stbi_uc *) buffer+len; + if (stbi__do_zlib(&a, p, 16384, 1, 0)) { + if (outlen) *outlen = (int) (a.zout - a.zout_start); + return a.zout_start; + } else { + STBI_FREE(a.zout_start); + return NULL; + } +} + +STBIDEF int stbi_zlib_decode_noheader_buffer(char *obuffer, int olen, const char *ibuffer, int ilen) +{ + stbi__zbuf a; + a.zbuffer = (stbi_uc *) ibuffer; + a.zbuffer_end = (stbi_uc *) ibuffer + ilen; + if (stbi__do_zlib(&a, obuffer, olen, 0, 0)) + return (int) (a.zout - a.zout_start); + else + return -1; +} +#endif + +// public domain "baseline" PNG decoder v0.10 Sean Barrett 2006-11-18 +// simple implementation +// - only 8-bit samples +// - no CRC checking +// - allocates lots of intermediate memory +// - avoids problem of streaming data between subsystems +// - avoids explicit window management +// performance +// - uses stb_zlib, a PD zlib implementation with fast huffman decoding + +#ifndef STBI_NO_PNG +typedef struct +{ + stbi__uint32 length; + stbi__uint32 type; +} stbi__pngchunk; + +static stbi__pngchunk stbi__get_chunk_header(stbi__context *s) +{ + stbi__pngchunk c; + c.length = stbi__get32be(s); + c.type = stbi__get32be(s); + return c; +} + +static int stbi__check_png_header(stbi__context *s) +{ + static stbi_uc png_sig[8] = { 137,80,78,71,13,10,26,10 }; + int i; + for (i=0; i < 8; ++i) + if (stbi__get8(s) != png_sig[i]) return stbi__err("bad png sig","Not a PNG"); + return 1; +} + +typedef struct +{ + stbi__context *s; + stbi_uc *idata, *expanded, *out; + int depth; +} stbi__png; + + +enum { + STBI__F_none=0, + STBI__F_sub=1, + STBI__F_up=2, + STBI__F_avg=3, + STBI__F_paeth=4, + // synthetic filters used for first scanline to avoid needing a dummy row of 0s + STBI__F_avg_first, + STBI__F_paeth_first +}; + +static stbi_uc first_row_filter[5] = +{ + STBI__F_none, + STBI__F_sub, + STBI__F_none, + STBI__F_avg_first, + STBI__F_paeth_first +}; + +static int stbi__paeth(int a, int b, int c) +{ + int p = a + b - c; + int pa = abs(p-a); + int pb = abs(p-b); + int pc = abs(p-c); + if (pa <= pb && pa <= pc) return a; + if (pb <= pc) return b; + return c; +} + +static stbi_uc stbi__depth_scale_table[9] = { 0, 0xff, 0x55, 0, 0x11, 0,0,0, 0x01 }; + +// create the png data from post-deflated data +static int stbi__create_png_image_raw(stbi__png *a, stbi_uc *raw, stbi__uint32 raw_len, int out_n, stbi__uint32 x, stbi__uint32 y, int depth, int color) +{ + int bytes = (depth == 16? 2 : 1); + stbi__context *s = a->s; + stbi__uint32 i,j,stride = x*out_n*bytes; + stbi__uint32 img_len, img_width_bytes; + int k; + int img_n = s->img_n; // copy it into a local for later + + int output_bytes = out_n*bytes; + int filter_bytes = img_n*bytes; + int width = x; + + STBI_ASSERT(out_n == s->img_n || out_n == s->img_n+1); + a->out = (stbi_uc *) stbi__malloc_mad3(x, y, output_bytes, 0); // extra bytes to write off the end into + if (!a->out) return stbi__err("outofmem", "Out of memory"); + + img_width_bytes = (((img_n * x * depth) + 7) >> 3); + img_len = (img_width_bytes + 1) * y; + // we used to check for exact match between raw_len and img_len on non-interlaced PNGs, + // but issue #276 reported a PNG in the wild that had extra data at the end (all zeros), + // so just check for raw_len < img_len always. + if (raw_len < img_len) return stbi__err("not enough pixels","Corrupt PNG"); + + for (j=0; j < y; ++j) { + stbi_uc *cur = a->out + stride*j; + stbi_uc *prior; + int filter = *raw++; + + if (filter > 4) + return stbi__err("invalid filter","Corrupt PNG"); + + if (depth < 8) { + STBI_ASSERT(img_width_bytes <= x); + cur += x*out_n - img_width_bytes; // store output to the rightmost img_len bytes, so we can decode in place + filter_bytes = 1; + width = img_width_bytes; + } + prior = cur - stride; // bugfix: need to compute this after 'cur +=' computation above + + // if first row, use special filter that doesn't sample previous row + if (j == 0) filter = first_row_filter[filter]; + + // handle first byte explicitly + for (k=0; k < filter_bytes; ++k) { + switch (filter) { + case STBI__F_none : cur[k] = raw[k]; break; + case STBI__F_sub : cur[k] = raw[k]; break; + case STBI__F_up : cur[k] = STBI__BYTECAST(raw[k] + prior[k]); break; + case STBI__F_avg : cur[k] = STBI__BYTECAST(raw[k] + (prior[k]>>1)); break; + case STBI__F_paeth : cur[k] = STBI__BYTECAST(raw[k] + stbi__paeth(0,prior[k],0)); break; + case STBI__F_avg_first : cur[k] = raw[k]; break; + case STBI__F_paeth_first: cur[k] = raw[k]; break; + } + } + + if (depth == 8) { + if (img_n != out_n) + cur[img_n] = 255; // first pixel + raw += img_n; + cur += out_n; + prior += out_n; + } else if (depth == 16) { + if (img_n != out_n) { + cur[filter_bytes] = 255; // first pixel top byte + cur[filter_bytes+1] = 255; // first pixel bottom byte + } + raw += filter_bytes; + cur += output_bytes; + prior += output_bytes; + } else { + raw += 1; + cur += 1; + prior += 1; + } + + // this is a little gross, so that we don't switch per-pixel or per-component + if (depth < 8 || img_n == out_n) { + int nk = (width - 1)*filter_bytes; + #define STBI__CASE(f) \ + case f: \ + for (k=0; k < nk; ++k) + switch (filter) { + // "none" filter turns into a memcpy here; make that explicit. + case STBI__F_none: memcpy(cur, raw, nk); break; + STBI__CASE(STBI__F_sub) { cur[k] = STBI__BYTECAST(raw[k] + cur[k-filter_bytes]); } break; + STBI__CASE(STBI__F_up) { cur[k] = STBI__BYTECAST(raw[k] + prior[k]); } break; + STBI__CASE(STBI__F_avg) { cur[k] = STBI__BYTECAST(raw[k] + ((prior[k] + cur[k-filter_bytes])>>1)); } break; + STBI__CASE(STBI__F_paeth) { cur[k] = STBI__BYTECAST(raw[k] + stbi__paeth(cur[k-filter_bytes],prior[k],prior[k-filter_bytes])); } break; + STBI__CASE(STBI__F_avg_first) { cur[k] = STBI__BYTECAST(raw[k] + (cur[k-filter_bytes] >> 1)); } break; + STBI__CASE(STBI__F_paeth_first) { cur[k] = STBI__BYTECAST(raw[k] + stbi__paeth(cur[k-filter_bytes],0,0)); } break; + } + #undef STBI__CASE + raw += nk; + } else { + STBI_ASSERT(img_n+1 == out_n); + #define STBI__CASE(f) \ + case f: \ + for (i=x-1; i >= 1; --i, cur[filter_bytes]=255,raw+=filter_bytes,cur+=output_bytes,prior+=output_bytes) \ + for (k=0; k < filter_bytes; ++k) + switch (filter) { + STBI__CASE(STBI__F_none) { cur[k] = raw[k]; } break; + STBI__CASE(STBI__F_sub) { cur[k] = STBI__BYTECAST(raw[k] + cur[k- output_bytes]); } break; + STBI__CASE(STBI__F_up) { cur[k] = STBI__BYTECAST(raw[k] + prior[k]); } break; + STBI__CASE(STBI__F_avg) { cur[k] = STBI__BYTECAST(raw[k] + ((prior[k] + cur[k- output_bytes])>>1)); } break; + STBI__CASE(STBI__F_paeth) { cur[k] = STBI__BYTECAST(raw[k] + stbi__paeth(cur[k- output_bytes],prior[k],prior[k- output_bytes])); } break; + STBI__CASE(STBI__F_avg_first) { cur[k] = STBI__BYTECAST(raw[k] + (cur[k- output_bytes] >> 1)); } break; + STBI__CASE(STBI__F_paeth_first) { cur[k] = STBI__BYTECAST(raw[k] + stbi__paeth(cur[k- output_bytes],0,0)); } break; + } + #undef STBI__CASE + + // the loop above sets the high byte of the pixels' alpha, but for + // 16 bit png files we also need the low byte set. we'll do that here. + if (depth == 16) { + cur = a->out + stride*j; // start at the beginning of the row again + for (i=0; i < x; ++i,cur+=output_bytes) { + cur[filter_bytes+1] = 255; + } + } + } + } + + // we make a separate pass to expand bits to pixels; for performance, + // this could run two scanlines behind the above code, so it won't + // intefere with filtering but will still be in the cache. + if (depth < 8) { + for (j=0; j < y; ++j) { + stbi_uc *cur = a->out + stride*j; + stbi_uc *in = a->out + stride*j + x*out_n - img_width_bytes; + // unpack 1/2/4-bit into a 8-bit buffer. allows us to keep the common 8-bit path optimal at minimal cost for 1/2/4-bit + // png guarante byte alignment, if width is not multiple of 8/4/2 we'll decode dummy trailing data that will be skipped in the later loop + stbi_uc scale = (color == 0) ? stbi__depth_scale_table[depth] : 1; // scale grayscale values to 0..255 range + + // note that the final byte might overshoot and write more data than desired. + // we can allocate enough data that this never writes out of memory, but it + // could also overwrite the next scanline. can it overwrite non-empty data + // on the next scanline? yes, consider 1-pixel-wide scanlines with 1-bit-per-pixel. + // so we need to explicitly clamp the final ones + + if (depth == 4) { + for (k=x*img_n; k >= 2; k-=2, ++in) { + *cur++ = scale * ((*in >> 4) ); + *cur++ = scale * ((*in ) & 0x0f); + } + if (k > 0) *cur++ = scale * ((*in >> 4) ); + } else if (depth == 2) { + for (k=x*img_n; k >= 4; k-=4, ++in) { + *cur++ = scale * ((*in >> 6) ); + *cur++ = scale * ((*in >> 4) & 0x03); + *cur++ = scale * ((*in >> 2) & 0x03); + *cur++ = scale * ((*in ) & 0x03); + } + if (k > 0) *cur++ = scale * ((*in >> 6) ); + if (k > 1) *cur++ = scale * ((*in >> 4) & 0x03); + if (k > 2) *cur++ = scale * ((*in >> 2) & 0x03); + } else if (depth == 1) { + for (k=x*img_n; k >= 8; k-=8, ++in) { + *cur++ = scale * ((*in >> 7) ); + *cur++ = scale * ((*in >> 6) & 0x01); + *cur++ = scale * ((*in >> 5) & 0x01); + *cur++ = scale * ((*in >> 4) & 0x01); + *cur++ = scale * ((*in >> 3) & 0x01); + *cur++ = scale * ((*in >> 2) & 0x01); + *cur++ = scale * ((*in >> 1) & 0x01); + *cur++ = scale * ((*in ) & 0x01); + } + if (k > 0) *cur++ = scale * ((*in >> 7) ); + if (k > 1) *cur++ = scale * ((*in >> 6) & 0x01); + if (k > 2) *cur++ = scale * ((*in >> 5) & 0x01); + if (k > 3) *cur++ = scale * ((*in >> 4) & 0x01); + if (k > 4) *cur++ = scale * ((*in >> 3) & 0x01); + if (k > 5) *cur++ = scale * ((*in >> 2) & 0x01); + if (k > 6) *cur++ = scale * ((*in >> 1) & 0x01); + } + if (img_n != out_n) { + int q; + // insert alpha = 255 + cur = a->out + stride*j; + if (img_n == 1) { + for (q=x-1; q >= 0; --q) { + cur[q*2+1] = 255; + cur[q*2+0] = cur[q]; + } + } else { + STBI_ASSERT(img_n == 3); + for (q=x-1; q >= 0; --q) { + cur[q*4+3] = 255; + cur[q*4+2] = cur[q*3+2]; + cur[q*4+1] = cur[q*3+1]; + cur[q*4+0] = cur[q*3+0]; + } + } + } + } + } else if (depth == 16) { + // force the image data from big-endian to platform-native. + // this is done in a separate pass due to the decoding relying + // on the data being untouched, but could probably be done + // per-line during decode if care is taken. + stbi_uc *cur = a->out; + stbi__uint16 *cur16 = (stbi__uint16*)cur; + + for(i=0; i < x*y*out_n; ++i,cur16++,cur+=2) { + *cur16 = (cur[0] << 8) | cur[1]; + } + } + + return 1; +} + +static int stbi__create_png_image(stbi__png *a, stbi_uc *image_data, stbi__uint32 image_data_len, int out_n, int depth, int color, int interlaced) +{ + int bytes = (depth == 16 ? 2 : 1); + int out_bytes = out_n * bytes; + stbi_uc *final; + int p; + if (!interlaced) + return stbi__create_png_image_raw(a, image_data, image_data_len, out_n, a->s->img_x, a->s->img_y, depth, color); + + // de-interlacing + final = (stbi_uc *) stbi__malloc_mad3(a->s->img_x, a->s->img_y, out_bytes, 0); + for (p=0; p < 7; ++p) { + int xorig[] = { 0,4,0,2,0,1,0 }; + int yorig[] = { 0,0,4,0,2,0,1 }; + int xspc[] = { 8,8,4,4,2,2,1 }; + int yspc[] = { 8,8,8,4,4,2,2 }; + int i,j,x,y; + // pass1_x[4] = 0, pass1_x[5] = 1, pass1_x[12] = 1 + x = (a->s->img_x - xorig[p] + xspc[p]-1) / xspc[p]; + y = (a->s->img_y - yorig[p] + yspc[p]-1) / yspc[p]; + if (x && y) { + stbi__uint32 img_len = ((((a->s->img_n * x * depth) + 7) >> 3) + 1) * y; + if (!stbi__create_png_image_raw(a, image_data, image_data_len, out_n, x, y, depth, color)) { + STBI_FREE(final); + return 0; + } + for (j=0; j < y; ++j) { + for (i=0; i < x; ++i) { + int out_y = j*yspc[p]+yorig[p]; + int out_x = i*xspc[p]+xorig[p]; + memcpy(final + out_y*a->s->img_x*out_bytes + out_x*out_bytes, + a->out + (j*x+i)*out_bytes, out_bytes); + } + } + STBI_FREE(a->out); + image_data += img_len; + image_data_len -= img_len; + } + } + a->out = final; + + return 1; +} + +static int stbi__compute_transparency(stbi__png *z, stbi_uc tc[3], int out_n) +{ + stbi__context *s = z->s; + stbi__uint32 i, pixel_count = s->img_x * s->img_y; + stbi_uc *p = z->out; + + // compute color-based transparency, assuming we've + // already got 255 as the alpha value in the output + STBI_ASSERT(out_n == 2 || out_n == 4); + + if (out_n == 2) { + for (i=0; i < pixel_count; ++i) { + p[1] = (p[0] == tc[0] ? 0 : 255); + p += 2; + } + } else { + for (i=0; i < pixel_count; ++i) { + if (p[0] == tc[0] && p[1] == tc[1] && p[2] == tc[2]) + p[3] = 0; + p += 4; + } + } + return 1; +} + +static int stbi__compute_transparency16(stbi__png *z, stbi__uint16 tc[3], int out_n) +{ + stbi__context *s = z->s; + stbi__uint32 i, pixel_count = s->img_x * s->img_y; + stbi__uint16 *p = (stbi__uint16*) z->out; + + // compute color-based transparency, assuming we've + // already got 65535 as the alpha value in the output + STBI_ASSERT(out_n == 2 || out_n == 4); + + if (out_n == 2) { + for (i = 0; i < pixel_count; ++i) { + p[1] = (p[0] == tc[0] ? 0 : 65535); + p += 2; + } + } else { + for (i = 0; i < pixel_count; ++i) { + if (p[0] == tc[0] && p[1] == tc[1] && p[2] == tc[2]) + p[3] = 0; + p += 4; + } + } + return 1; +} + +static int stbi__expand_png_palette(stbi__png *a, stbi_uc *palette, int len, int pal_img_n) +{ + stbi__uint32 i, pixel_count = a->s->img_x * a->s->img_y; + stbi_uc *p, *temp_out, *orig = a->out; + + p = (stbi_uc *) stbi__malloc_mad2(pixel_count, pal_img_n, 0); + if (p == NULL) return stbi__err("outofmem", "Out of memory"); + + // between here and free(out) below, exitting would leak + temp_out = p; + + if (pal_img_n == 3) { + for (i=0; i < pixel_count; ++i) { + int n = orig[i]*4; + p[0] = palette[n ]; + p[1] = palette[n+1]; + p[2] = palette[n+2]; + p += 3; + } + } else { + for (i=0; i < pixel_count; ++i) { + int n = orig[i]*4; + p[0] = palette[n ]; + p[1] = palette[n+1]; + p[2] = palette[n+2]; + p[3] = palette[n+3]; + p += 4; + } + } + STBI_FREE(a->out); + a->out = temp_out; + + STBI_NOTUSED(len); + + return 1; +} + +static int stbi__unpremultiply_on_load = 0; +static int stbi__de_iphone_flag = 0; + +STBIDEF void stbi_set_unpremultiply_on_load(int flag_true_if_should_unpremultiply) +{ + stbi__unpremultiply_on_load = flag_true_if_should_unpremultiply; +} + +STBIDEF void stbi_convert_iphone_png_to_rgb(int flag_true_if_should_convert) +{ + stbi__de_iphone_flag = flag_true_if_should_convert; +} + +static void stbi__de_iphone(stbi__png *z) +{ + stbi__context *s = z->s; + stbi__uint32 i, pixel_count = s->img_x * s->img_y; + stbi_uc *p = z->out; + + if (s->img_out_n == 3) { // convert bgr to rgb + for (i=0; i < pixel_count; ++i) { + stbi_uc t = p[0]; + p[0] = p[2]; + p[2] = t; + p += 3; + } + } else { + STBI_ASSERT(s->img_out_n == 4); + if (stbi__unpremultiply_on_load) { + // convert bgr to rgb and unpremultiply + for (i=0; i < pixel_count; ++i) { + stbi_uc a = p[3]; + stbi_uc t = p[0]; + if (a) { + stbi_uc half = a / 2; + p[0] = (p[2] * 255 + half) / a; + p[1] = (p[1] * 255 + half) / a; + p[2] = ( t * 255 + half) / a; + } else { + p[0] = p[2]; + p[2] = t; + } + p += 4; + } + } else { + // convert bgr to rgb + for (i=0; i < pixel_count; ++i) { + stbi_uc t = p[0]; + p[0] = p[2]; + p[2] = t; + p += 4; + } + } + } +} + +#define STBI__PNG_TYPE(a,b,c,d) (((a) << 24) + ((b) << 16) + ((c) << 8) + (d)) + +static int stbi__parse_png_file(stbi__png *z, int scan, int req_comp) +{ + stbi_uc palette[1024], pal_img_n=0; + stbi_uc has_trans=0, tc[3]; + stbi__uint16 tc16[3]; + stbi__uint32 ioff=0, idata_limit=0, i, pal_len=0; + int first=1,k,interlace=0, color=0, is_iphone=0; + stbi__context *s = z->s; + + z->expanded = NULL; + z->idata = NULL; + z->out = NULL; + + if (!stbi__check_png_header(s)) return 0; + + if (scan == STBI__SCAN_type) return 1; + + for (;;) { + stbi__pngchunk c = stbi__get_chunk_header(s); + switch (c.type) { + case STBI__PNG_TYPE('C','g','B','I'): + is_iphone = 1; + stbi__skip(s, c.length); + break; + case STBI__PNG_TYPE('I','H','D','R'): { + int comp,filter; + if (!first) return stbi__err("multiple IHDR","Corrupt PNG"); + first = 0; + if (c.length != 13) return stbi__err("bad IHDR len","Corrupt PNG"); + s->img_x = stbi__get32be(s); if (s->img_x > (1 << 24)) return stbi__err("too large","Very large image (corrupt?)"); + s->img_y = stbi__get32be(s); if (s->img_y > (1 << 24)) return stbi__err("too large","Very large image (corrupt?)"); + z->depth = stbi__get8(s); if (z->depth != 1 && z->depth != 2 && z->depth != 4 && z->depth != 8 && z->depth != 16) return stbi__err("1/2/4/8/16-bit only","PNG not supported: 1/2/4/8/16-bit only"); + color = stbi__get8(s); if (color > 6) return stbi__err("bad ctype","Corrupt PNG"); + if (color == 3 && z->depth == 16) return stbi__err("bad ctype","Corrupt PNG"); + if (color == 3) pal_img_n = 3; else if (color & 1) return stbi__err("bad ctype","Corrupt PNG"); + comp = stbi__get8(s); if (comp) return stbi__err("bad comp method","Corrupt PNG"); + filter= stbi__get8(s); if (filter) return stbi__err("bad filter method","Corrupt PNG"); + interlace = stbi__get8(s); if (interlace>1) return stbi__err("bad interlace method","Corrupt PNG"); + if (!s->img_x || !s->img_y) return stbi__err("0-pixel image","Corrupt PNG"); + if (!pal_img_n) { + s->img_n = (color & 2 ? 3 : 1) + (color & 4 ? 1 : 0); + if ((1 << 30) / s->img_x / s->img_n < s->img_y) return stbi__err("too large", "Image too large to decode"); + if (scan == STBI__SCAN_header) return 1; + } else { + // if paletted, then pal_n is our final components, and + // img_n is # components to decompress/filter. + s->img_n = 1; + if ((1 << 30) / s->img_x / 4 < s->img_y) return stbi__err("too large","Corrupt PNG"); + // if SCAN_header, have to scan to see if we have a tRNS + } + break; + } + + case STBI__PNG_TYPE('P','L','T','E'): { + if (first) return stbi__err("first not IHDR", "Corrupt PNG"); + if (c.length > 256*3) return stbi__err("invalid PLTE","Corrupt PNG"); + pal_len = c.length / 3; + if (pal_len * 3 != c.length) return stbi__err("invalid PLTE","Corrupt PNG"); + for (i=0; i < pal_len; ++i) { + palette[i*4+0] = stbi__get8(s); + palette[i*4+1] = stbi__get8(s); + palette[i*4+2] = stbi__get8(s); + palette[i*4+3] = 255; + } + break; + } + + case STBI__PNG_TYPE('t','R','N','S'): { + if (first) return stbi__err("first not IHDR", "Corrupt PNG"); + if (z->idata) return stbi__err("tRNS after IDAT","Corrupt PNG"); + if (pal_img_n) { + if (scan == STBI__SCAN_header) { s->img_n = 4; return 1; } + if (pal_len == 0) return stbi__err("tRNS before PLTE","Corrupt PNG"); + if (c.length > pal_len) return stbi__err("bad tRNS len","Corrupt PNG"); + pal_img_n = 4; + for (i=0; i < c.length; ++i) + palette[i*4+3] = stbi__get8(s); + } else { + if (!(s->img_n & 1)) return stbi__err("tRNS with alpha","Corrupt PNG"); + if (c.length != (stbi__uint32) s->img_n*2) return stbi__err("bad tRNS len","Corrupt PNG"); + has_trans = 1; + if (z->depth == 16) { + for (k = 0; k < s->img_n; ++k) tc16[k] = (stbi__uint16)stbi__get16be(s); // copy the values as-is + } else { + for (k = 0; k < s->img_n; ++k) tc[k] = (stbi_uc)(stbi__get16be(s) & 255) * stbi__depth_scale_table[z->depth]; // non 8-bit images will be larger + } + } + break; + } + + case STBI__PNG_TYPE('I','D','A','T'): { + if (first) return stbi__err("first not IHDR", "Corrupt PNG"); + if (pal_img_n && !pal_len) return stbi__err("no PLTE","Corrupt PNG"); + if (scan == STBI__SCAN_header) { s->img_n = pal_img_n; return 1; } + if ((int)(ioff + c.length) < (int)ioff) return 0; + if (ioff + c.length > idata_limit) { + stbi__uint32 idata_limit_old = idata_limit; + stbi_uc *p; + if (idata_limit == 0) idata_limit = c.length > 4096 ? c.length : 4096; + while (ioff + c.length > idata_limit) + idata_limit *= 2; + STBI_NOTUSED(idata_limit_old); + p = (stbi_uc *) STBI_REALLOC_SIZED(z->idata, idata_limit_old, idata_limit); if (p == NULL) return stbi__err("outofmem", "Out of memory"); + z->idata = p; + } + if (!stbi__getn(s, z->idata+ioff,c.length)) return stbi__err("outofdata","Corrupt PNG"); + ioff += c.length; + break; + } + + case STBI__PNG_TYPE('I','E','N','D'): { + stbi__uint32 raw_len, bpl; + if (first) return stbi__err("first not IHDR", "Corrupt PNG"); + if (scan != STBI__SCAN_load) return 1; + if (z->idata == NULL) return stbi__err("no IDAT","Corrupt PNG"); + // initial guess for decoded data size to avoid unnecessary reallocs + bpl = (s->img_x * z->depth + 7) / 8; // bytes per line, per component + raw_len = bpl * s->img_y * s->img_n /* pixels */ + s->img_y /* filter mode per row */; + z->expanded = (stbi_uc *) stbi_zlib_decode_malloc_guesssize_headerflag((char *) z->idata, ioff, raw_len, (int *) &raw_len, !is_iphone); + if (z->expanded == NULL) return 0; // zlib should set error + STBI_FREE(z->idata); z->idata = NULL; + if ((req_comp == s->img_n+1 && req_comp != 3 && !pal_img_n) || has_trans) + s->img_out_n = s->img_n+1; + else + s->img_out_n = s->img_n; + if (!stbi__create_png_image(z, z->expanded, raw_len, s->img_out_n, z->depth, color, interlace)) return 0; + if (has_trans) { + if (z->depth == 16) { + if (!stbi__compute_transparency16(z, tc16, s->img_out_n)) return 0; + } else { + if (!stbi__compute_transparency(z, tc, s->img_out_n)) return 0; + } + } + if (is_iphone && stbi__de_iphone_flag && s->img_out_n > 2) + stbi__de_iphone(z); + if (pal_img_n) { + // pal_img_n == 3 or 4 + s->img_n = pal_img_n; // record the actual colors we had + s->img_out_n = pal_img_n; + if (req_comp >= 3) s->img_out_n = req_comp; + if (!stbi__expand_png_palette(z, palette, pal_len, s->img_out_n)) + return 0; + } else if (has_trans) { + // non-paletted image with tRNS -> source image has (constant) alpha + ++s->img_n; + } + STBI_FREE(z->expanded); z->expanded = NULL; + return 1; + } + + default: + // if critical, fail + if (first) return stbi__err("first not IHDR", "Corrupt PNG"); + if ((c.type & (1 << 29)) == 0) { + #ifndef STBI_NO_FAILURE_STRINGS + // not threadsafe + static char invalid_chunk[] = "XXXX PNG chunk not known"; + invalid_chunk[0] = STBI__BYTECAST(c.type >> 24); + invalid_chunk[1] = STBI__BYTECAST(c.type >> 16); + invalid_chunk[2] = STBI__BYTECAST(c.type >> 8); + invalid_chunk[3] = STBI__BYTECAST(c.type >> 0); + #endif + return stbi__err(invalid_chunk, "PNG not supported: unknown PNG chunk type"); + } + stbi__skip(s, c.length); + break; + } + // end of PNG chunk, read and skip CRC + stbi__get32be(s); + } +} + +static void *stbi__do_png(stbi__png *p, int *x, int *y, int *n, int req_comp, stbi__result_info *ri) +{ + void *result=NULL; + if (req_comp < 0 || req_comp > 4) return stbi__errpuc("bad req_comp", "Internal error"); + if (stbi__parse_png_file(p, STBI__SCAN_load, req_comp)) { + if (p->depth < 8) + ri->bits_per_channel = 8; + else + ri->bits_per_channel = p->depth; + result = p->out; + p->out = NULL; + if (req_comp && req_comp != p->s->img_out_n) { + if (ri->bits_per_channel == 8) + result = stbi__convert_format((unsigned char *) result, p->s->img_out_n, req_comp, p->s->img_x, p->s->img_y); + else + result = stbi__convert_format16((stbi__uint16 *) result, p->s->img_out_n, req_comp, p->s->img_x, p->s->img_y); + p->s->img_out_n = req_comp; + if (result == NULL) return result; + } + *x = p->s->img_x; + *y = p->s->img_y; + if (n) *n = p->s->img_n; + } + STBI_FREE(p->out); p->out = NULL; + STBI_FREE(p->expanded); p->expanded = NULL; + STBI_FREE(p->idata); p->idata = NULL; + + return result; +} + +static void *stbi__png_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri) +{ + stbi__png p; + p.s = s; + return stbi__do_png(&p, x,y,comp,req_comp, ri); +} + +static int stbi__png_test(stbi__context *s) +{ + int r; + r = stbi__check_png_header(s); + stbi__rewind(s); + return r; +} + +static int stbi__png_info_raw(stbi__png *p, int *x, int *y, int *comp) +{ + if (!stbi__parse_png_file(p, STBI__SCAN_header, 0)) { + stbi__rewind( p->s ); + return 0; + } + if (x) *x = p->s->img_x; + if (y) *y = p->s->img_y; + if (comp) *comp = p->s->img_n; + return 1; +} + +static int stbi__png_info(stbi__context *s, int *x, int *y, int *comp) +{ + stbi__png p; + p.s = s; + return stbi__png_info_raw(&p, x, y, comp); +} +#endif + +// Microsoft/Windows BMP image + +#ifndef STBI_NO_BMP +static int stbi__bmp_test_raw(stbi__context *s) +{ + int r; + int sz; + if (stbi__get8(s) != 'B') return 0; + if (stbi__get8(s) != 'M') return 0; + stbi__get32le(s); // discard filesize + stbi__get16le(s); // discard reserved + stbi__get16le(s); // discard reserved + stbi__get32le(s); // discard data offset + sz = stbi__get32le(s); + r = (sz == 12 || sz == 40 || sz == 56 || sz == 108 || sz == 124); + return r; +} + +static int stbi__bmp_test(stbi__context *s) +{ + int r = stbi__bmp_test_raw(s); + stbi__rewind(s); + return r; +} + + +// returns 0..31 for the highest set bit +static int stbi__high_bit(unsigned int z) +{ + int n=0; + if (z == 0) return -1; + if (z >= 0x10000) n += 16, z >>= 16; + if (z >= 0x00100) n += 8, z >>= 8; + if (z >= 0x00010) n += 4, z >>= 4; + if (z >= 0x00004) n += 2, z >>= 2; + if (z >= 0x00002) n += 1, z >>= 1; + return n; +} + +static int stbi__bitcount(unsigned int a) +{ + a = (a & 0x55555555) + ((a >> 1) & 0x55555555); // max 2 + a = (a & 0x33333333) + ((a >> 2) & 0x33333333); // max 4 + a = (a + (a >> 4)) & 0x0f0f0f0f; // max 8 per 4, now 8 bits + a = (a + (a >> 8)); // max 16 per 8 bits + a = (a + (a >> 16)); // max 32 per 8 bits + return a & 0xff; +} + +static int stbi__shiftsigned(int v, int shift, int bits) +{ + int result; + int z=0; + + if (shift < 0) v <<= -shift; + else v >>= shift; + result = v; + + z = bits; + while (z < 8) { + result += v >> z; + z += bits; + } + return result; +} + +typedef struct +{ + int bpp, offset, hsz; + unsigned int mr,mg,mb,ma, all_a; +} stbi__bmp_data; + +static void *stbi__bmp_parse_header(stbi__context *s, stbi__bmp_data *info) +{ + int hsz; + if (stbi__get8(s) != 'B' || stbi__get8(s) != 'M') return stbi__errpuc("not BMP", "Corrupt BMP"); + stbi__get32le(s); // discard filesize + stbi__get16le(s); // discard reserved + stbi__get16le(s); // discard reserved + info->offset = stbi__get32le(s); + info->hsz = hsz = stbi__get32le(s); + info->mr = info->mg = info->mb = info->ma = 0; + + if (hsz != 12 && hsz != 40 && hsz != 56 && hsz != 108 && hsz != 124) return stbi__errpuc("unknown BMP", "BMP type not supported: unknown"); + if (hsz == 12) { + s->img_x = stbi__get16le(s); + s->img_y = stbi__get16le(s); + } else { + s->img_x = stbi__get32le(s); + s->img_y = stbi__get32le(s); + } + if (stbi__get16le(s) != 1) return stbi__errpuc("bad BMP", "bad BMP"); + info->bpp = stbi__get16le(s); + if (info->bpp == 1) return stbi__errpuc("monochrome", "BMP type not supported: 1-bit"); + if (hsz != 12) { + int compress = stbi__get32le(s); + if (compress == 1 || compress == 2) return stbi__errpuc("BMP RLE", "BMP type not supported: RLE"); + stbi__get32le(s); // discard sizeof + stbi__get32le(s); // discard hres + stbi__get32le(s); // discard vres + stbi__get32le(s); // discard colorsused + stbi__get32le(s); // discard max important + if (hsz == 40 || hsz == 56) { + if (hsz == 56) { + stbi__get32le(s); + stbi__get32le(s); + stbi__get32le(s); + stbi__get32le(s); + } + if (info->bpp == 16 || info->bpp == 32) { + if (compress == 0) { + if (info->bpp == 32) { + info->mr = 0xffu << 16; + info->mg = 0xffu << 8; + info->mb = 0xffu << 0; + info->ma = 0xffu << 24; + info->all_a = 0; // if all_a is 0 at end, then we loaded alpha channel but it was all 0 + } else { + info->mr = 31u << 10; + info->mg = 31u << 5; + info->mb = 31u << 0; + } + } else if (compress == 3) { + info->mr = stbi__get32le(s); + info->mg = stbi__get32le(s); + info->mb = stbi__get32le(s); + // not documented, but generated by photoshop and handled by mspaint + if (info->mr == info->mg && info->mg == info->mb) { + // ?!?!? + return stbi__errpuc("bad BMP", "bad BMP"); + } + } else + return stbi__errpuc("bad BMP", "bad BMP"); + } + } else { + int i; + if (hsz != 108 && hsz != 124) + return stbi__errpuc("bad BMP", "bad BMP"); + info->mr = stbi__get32le(s); + info->mg = stbi__get32le(s); + info->mb = stbi__get32le(s); + info->ma = stbi__get32le(s); + stbi__get32le(s); // discard color space + for (i=0; i < 12; ++i) + stbi__get32le(s); // discard color space parameters + if (hsz == 124) { + stbi__get32le(s); // discard rendering intent + stbi__get32le(s); // discard offset of profile data + stbi__get32le(s); // discard size of profile data + stbi__get32le(s); // discard reserved + } + } + } + return (void *) 1; +} + + +static void *stbi__bmp_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri) +{ + stbi_uc *out; + unsigned int mr=0,mg=0,mb=0,ma=0, all_a; + stbi_uc pal[256][4]; + int psize=0,i,j,width; + int flip_vertically, pad, target; + stbi__bmp_data info; + STBI_NOTUSED(ri); + + info.all_a = 255; + if (stbi__bmp_parse_header(s, &info) == NULL) + return NULL; // error code already set + + flip_vertically = ((int) s->img_y) > 0; + s->img_y = abs((int) s->img_y); + + mr = info.mr; + mg = info.mg; + mb = info.mb; + ma = info.ma; + all_a = info.all_a; + + if (info.hsz == 12) { + if (info.bpp < 24) + psize = (info.offset - 14 - 24) / 3; + } else { + if (info.bpp < 16) + psize = (info.offset - 14 - info.hsz) >> 2; + } + + s->img_n = ma ? 4 : 3; + if (req_comp && req_comp >= 3) // we can directly decode 3 or 4 + target = req_comp; + else + target = s->img_n; // if they want monochrome, we'll post-convert + + // sanity-check size + if (!stbi__mad3sizes_valid(target, s->img_x, s->img_y, 0)) + return stbi__errpuc("too large", "Corrupt BMP"); + + out = (stbi_uc *) stbi__malloc_mad3(target, s->img_x, s->img_y, 0); + if (!out) return stbi__errpuc("outofmem", "Out of memory"); + if (info.bpp < 16) { + int z=0; + if (psize == 0 || psize > 256) { STBI_FREE(out); return stbi__errpuc("invalid", "Corrupt BMP"); } + for (i=0; i < psize; ++i) { + pal[i][2] = stbi__get8(s); + pal[i][1] = stbi__get8(s); + pal[i][0] = stbi__get8(s); + if (info.hsz != 12) stbi__get8(s); + pal[i][3] = 255; + } + stbi__skip(s, info.offset - 14 - info.hsz - psize * (info.hsz == 12 ? 3 : 4)); + if (info.bpp == 4) width = (s->img_x + 1) >> 1; + else if (info.bpp == 8) width = s->img_x; + else { STBI_FREE(out); return stbi__errpuc("bad bpp", "Corrupt BMP"); } + pad = (-width)&3; + for (j=0; j < (int) s->img_y; ++j) { + for (i=0; i < (int) s->img_x; i += 2) { + int v=stbi__get8(s),v2=0; + if (info.bpp == 4) { + v2 = v & 15; + v >>= 4; + } + out[z++] = pal[v][0]; + out[z++] = pal[v][1]; + out[z++] = pal[v][2]; + if (target == 4) out[z++] = 255; + if (i+1 == (int) s->img_x) break; + v = (info.bpp == 8) ? stbi__get8(s) : v2; + out[z++] = pal[v][0]; + out[z++] = pal[v][1]; + out[z++] = pal[v][2]; + if (target == 4) out[z++] = 255; + } + stbi__skip(s, pad); + } + } else { + int rshift=0,gshift=0,bshift=0,ashift=0,rcount=0,gcount=0,bcount=0,acount=0; + int z = 0; + int easy=0; + stbi__skip(s, info.offset - 14 - info.hsz); + if (info.bpp == 24) width = 3 * s->img_x; + else if (info.bpp == 16) width = 2*s->img_x; + else /* bpp = 32 and pad = 0 */ width=0; + pad = (-width) & 3; + if (info.bpp == 24) { + easy = 1; + } else if (info.bpp == 32) { + if (mb == 0xff && mg == 0xff00 && mr == 0x00ff0000 && ma == 0xff000000) + easy = 2; + } + if (!easy) { + if (!mr || !mg || !mb) { STBI_FREE(out); return stbi__errpuc("bad masks", "Corrupt BMP"); } + // right shift amt to put high bit in position #7 + rshift = stbi__high_bit(mr)-7; rcount = stbi__bitcount(mr); + gshift = stbi__high_bit(mg)-7; gcount = stbi__bitcount(mg); + bshift = stbi__high_bit(mb)-7; bcount = stbi__bitcount(mb); + ashift = stbi__high_bit(ma)-7; acount = stbi__bitcount(ma); + } + for (j=0; j < (int) s->img_y; ++j) { + if (easy) { + for (i=0; i < (int) s->img_x; ++i) { + unsigned char a; + out[z+2] = stbi__get8(s); + out[z+1] = stbi__get8(s); + out[z+0] = stbi__get8(s); + z += 3; + a = (easy == 2 ? stbi__get8(s) : 255); + all_a |= a; + if (target == 4) out[z++] = a; + } + } else { + int bpp = info.bpp; + for (i=0; i < (int) s->img_x; ++i) { + stbi__uint32 v = (bpp == 16 ? (stbi__uint32) stbi__get16le(s) : stbi__get32le(s)); + int a; + out[z++] = STBI__BYTECAST(stbi__shiftsigned(v & mr, rshift, rcount)); + out[z++] = STBI__BYTECAST(stbi__shiftsigned(v & mg, gshift, gcount)); + out[z++] = STBI__BYTECAST(stbi__shiftsigned(v & mb, bshift, bcount)); + a = (ma ? stbi__shiftsigned(v & ma, ashift, acount) : 255); + all_a |= a; + if (target == 4) out[z++] = STBI__BYTECAST(a); + } + } + stbi__skip(s, pad); + } + } + + // if alpha channel is all 0s, replace with all 255s + if (target == 4 && all_a == 0) + for (i=4*s->img_x*s->img_y-1; i >= 0; i -= 4) + out[i] = 255; + + if (flip_vertically) { + stbi_uc t; + for (j=0; j < (int) s->img_y>>1; ++j) { + stbi_uc *p1 = out + j *s->img_x*target; + stbi_uc *p2 = out + (s->img_y-1-j)*s->img_x*target; + for (i=0; i < (int) s->img_x*target; ++i) { + t = p1[i], p1[i] = p2[i], p2[i] = t; + } + } + } + + if (req_comp && req_comp != target) { + out = stbi__convert_format(out, target, req_comp, s->img_x, s->img_y); + if (out == NULL) return out; // stbi__convert_format frees input on failure + } + + *x = s->img_x; + *y = s->img_y; + if (comp) *comp = s->img_n; + return out; +} +#endif + +// Targa Truevision - TGA +// by Jonathan Dummer +#ifndef STBI_NO_TGA +// returns STBI_rgb or whatever, 0 on error +static int stbi__tga_get_comp(int bits_per_pixel, int is_grey, int* is_rgb16) +{ + // only RGB or RGBA (incl. 16bit) or grey allowed + if(is_rgb16) *is_rgb16 = 0; + switch(bits_per_pixel) { + case 8: return STBI_grey; + case 16: if(is_grey) return STBI_grey_alpha; + // else: fall-through + case 15: if(is_rgb16) *is_rgb16 = 1; + return STBI_rgb; + case 24: // fall-through + case 32: return bits_per_pixel/8; + default: return 0; + } +} + +static int stbi__tga_info(stbi__context *s, int *x, int *y, int *comp) +{ + int tga_w, tga_h, tga_comp, tga_image_type, tga_bits_per_pixel, tga_colormap_bpp; + int sz, tga_colormap_type; + stbi__get8(s); // discard Offset + tga_colormap_type = stbi__get8(s); // colormap type + if( tga_colormap_type > 1 ) { + stbi__rewind(s); + return 0; // only RGB or indexed allowed + } + tga_image_type = stbi__get8(s); // image type + if ( tga_colormap_type == 1 ) { // colormapped (paletted) image + if (tga_image_type != 1 && tga_image_type != 9) { + stbi__rewind(s); + return 0; + } + stbi__skip(s,4); // skip index of first colormap entry and number of entries + sz = stbi__get8(s); // check bits per palette color entry + if ( (sz != 8) && (sz != 15) && (sz != 16) && (sz != 24) && (sz != 32) ) { + stbi__rewind(s); + return 0; + } + stbi__skip(s,4); // skip image x and y origin + tga_colormap_bpp = sz; + } else { // "normal" image w/o colormap - only RGB or grey allowed, +/- RLE + if ( (tga_image_type != 2) && (tga_image_type != 3) && (tga_image_type != 10) && (tga_image_type != 11) ) { + stbi__rewind(s); + return 0; // only RGB or grey allowed, +/- RLE + } + stbi__skip(s,9); // skip colormap specification and image x/y origin + tga_colormap_bpp = 0; + } + tga_w = stbi__get16le(s); + if( tga_w < 1 ) { + stbi__rewind(s); + return 0; // test width + } + tga_h = stbi__get16le(s); + if( tga_h < 1 ) { + stbi__rewind(s); + return 0; // test height + } + tga_bits_per_pixel = stbi__get8(s); // bits per pixel + stbi__get8(s); // ignore alpha bits + if (tga_colormap_bpp != 0) { + if((tga_bits_per_pixel != 8) && (tga_bits_per_pixel != 16)) { + // when using a colormap, tga_bits_per_pixel is the size of the indexes + // I don't think anything but 8 or 16bit indexes makes sense + stbi__rewind(s); + return 0; + } + tga_comp = stbi__tga_get_comp(tga_colormap_bpp, 0, NULL); + } else { + tga_comp = stbi__tga_get_comp(tga_bits_per_pixel, (tga_image_type == 3) || (tga_image_type == 11), NULL); + } + if(!tga_comp) { + stbi__rewind(s); + return 0; + } + if (x) *x = tga_w; + if (y) *y = tga_h; + if (comp) *comp = tga_comp; + return 1; // seems to have passed everything +} + +static int stbi__tga_test(stbi__context *s) +{ + int res = 0; + int sz, tga_color_type; + stbi__get8(s); // discard Offset + tga_color_type = stbi__get8(s); // color type + if ( tga_color_type > 1 ) goto errorEnd; // only RGB or indexed allowed + sz = stbi__get8(s); // image type + if ( tga_color_type == 1 ) { // colormapped (paletted) image + if (sz != 1 && sz != 9) goto errorEnd; // colortype 1 demands image type 1 or 9 + stbi__skip(s,4); // skip index of first colormap entry and number of entries + sz = stbi__get8(s); // check bits per palette color entry + if ( (sz != 8) && (sz != 15) && (sz != 16) && (sz != 24) && (sz != 32) ) goto errorEnd; + stbi__skip(s,4); // skip image x and y origin + } else { // "normal" image w/o colormap + if ( (sz != 2) && (sz != 3) && (sz != 10) && (sz != 11) ) goto errorEnd; // only RGB or grey allowed, +/- RLE + stbi__skip(s,9); // skip colormap specification and image x/y origin + } + if ( stbi__get16le(s) < 1 ) goto errorEnd; // test width + if ( stbi__get16le(s) < 1 ) goto errorEnd; // test height + sz = stbi__get8(s); // bits per pixel + if ( (tga_color_type == 1) && (sz != 8) && (sz != 16) ) goto errorEnd; // for colormapped images, bpp is size of an index + if ( (sz != 8) && (sz != 15) && (sz != 16) && (sz != 24) && (sz != 32) ) goto errorEnd; + + res = 1; // if we got this far, everything's good and we can return 1 instead of 0 + +errorEnd: + stbi__rewind(s); + return res; +} + +// read 16bit value and convert to 24bit RGB +static void stbi__tga_read_rgb16(stbi__context *s, stbi_uc* out) +{ + stbi__uint16 px = (stbi__uint16)stbi__get16le(s); + stbi__uint16 fiveBitMask = 31; + // we have 3 channels with 5bits each + int r = (px >> 10) & fiveBitMask; + int g = (px >> 5) & fiveBitMask; + int b = px & fiveBitMask; + // Note that this saves the data in RGB(A) order, so it doesn't need to be swapped later + out[0] = (stbi_uc)((r * 255)/31); + out[1] = (stbi_uc)((g * 255)/31); + out[2] = (stbi_uc)((b * 255)/31); + + // some people claim that the most significant bit might be used for alpha + // (possibly if an alpha-bit is set in the "image descriptor byte") + // but that only made 16bit test images completely translucent.. + // so let's treat all 15 and 16bit TGAs as RGB with no alpha. +} + +static void *stbi__tga_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri) +{ + // read in the TGA header stuff + int tga_offset = stbi__get8(s); + int tga_indexed = stbi__get8(s); + int tga_image_type = stbi__get8(s); + int tga_is_RLE = 0; + int tga_palette_start = stbi__get16le(s); + int tga_palette_len = stbi__get16le(s); + int tga_palette_bits = stbi__get8(s); + int tga_x_origin = stbi__get16le(s); + int tga_y_origin = stbi__get16le(s); + int tga_width = stbi__get16le(s); + int tga_height = stbi__get16le(s); + int tga_bits_per_pixel = stbi__get8(s); + int tga_comp, tga_rgb16=0; + int tga_inverted = stbi__get8(s); + // int tga_alpha_bits = tga_inverted & 15; // the 4 lowest bits - unused (useless?) + // image data + unsigned char *tga_data; + unsigned char *tga_palette = NULL; + int i, j; + unsigned char raw_data[4] = {0}; + int RLE_count = 0; + int RLE_repeating = 0; + int read_next_pixel = 1; + STBI_NOTUSED(ri); + + // do a tiny bit of precessing + if ( tga_image_type >= 8 ) + { + tga_image_type -= 8; + tga_is_RLE = 1; + } + tga_inverted = 1 - ((tga_inverted >> 5) & 1); + + // If I'm paletted, then I'll use the number of bits from the palette + if ( tga_indexed ) tga_comp = stbi__tga_get_comp(tga_palette_bits, 0, &tga_rgb16); + else tga_comp = stbi__tga_get_comp(tga_bits_per_pixel, (tga_image_type == 3), &tga_rgb16); + + if(!tga_comp) // shouldn't really happen, stbi__tga_test() should have ensured basic consistency + return stbi__errpuc("bad format", "Can't find out TGA pixelformat"); + + // tga info + *x = tga_width; + *y = tga_height; + if (comp) *comp = tga_comp; + + if (!stbi__mad3sizes_valid(tga_width, tga_height, tga_comp, 0)) + return stbi__errpuc("too large", "Corrupt TGA"); + + tga_data = (unsigned char*)stbi__malloc_mad3(tga_width, tga_height, tga_comp, 0); + if (!tga_data) return stbi__errpuc("outofmem", "Out of memory"); + + // skip to the data's starting position (offset usually = 0) + stbi__skip(s, tga_offset ); + + if ( !tga_indexed && !tga_is_RLE && !tga_rgb16 ) { + for (i=0; i < tga_height; ++i) { + int row = tga_inverted ? tga_height -i - 1 : i; + stbi_uc *tga_row = tga_data + row*tga_width*tga_comp; + stbi__getn(s, tga_row, tga_width * tga_comp); + } + } else { + // do I need to load a palette? + if ( tga_indexed) + { + // any data to skip? (offset usually = 0) + stbi__skip(s, tga_palette_start ); + // load the palette + tga_palette = (unsigned char*)stbi__malloc_mad2(tga_palette_len, tga_comp, 0); + if (!tga_palette) { + STBI_FREE(tga_data); + return stbi__errpuc("outofmem", "Out of memory"); + } + if (tga_rgb16) { + stbi_uc *pal_entry = tga_palette; + STBI_ASSERT(tga_comp == STBI_rgb); + for (i=0; i < tga_palette_len; ++i) { + stbi__tga_read_rgb16(s, pal_entry); + pal_entry += tga_comp; + } + } else if (!stbi__getn(s, tga_palette, tga_palette_len * tga_comp)) { + STBI_FREE(tga_data); + STBI_FREE(tga_palette); + return stbi__errpuc("bad palette", "Corrupt TGA"); + } + } + // load the data + for (i=0; i < tga_width * tga_height; ++i) + { + // if I'm in RLE mode, do I need to get a RLE stbi__pngchunk? + if ( tga_is_RLE ) + { + if ( RLE_count == 0 ) + { + // yep, get the next byte as a RLE command + int RLE_cmd = stbi__get8(s); + RLE_count = 1 + (RLE_cmd & 127); + RLE_repeating = RLE_cmd >> 7; + read_next_pixel = 1; + } else if ( !RLE_repeating ) + { + read_next_pixel = 1; + } + } else + { + read_next_pixel = 1; + } + // OK, if I need to read a pixel, do it now + if ( read_next_pixel ) + { + // load however much data we did have + if ( tga_indexed ) + { + // read in index, then perform the lookup + int pal_idx = (tga_bits_per_pixel == 8) ? stbi__get8(s) : stbi__get16le(s); + if ( pal_idx >= tga_palette_len ) { + // invalid index + pal_idx = 0; + } + pal_idx *= tga_comp; + for (j = 0; j < tga_comp; ++j) { + raw_data[j] = tga_palette[pal_idx+j]; + } + } else if(tga_rgb16) { + STBI_ASSERT(tga_comp == STBI_rgb); + stbi__tga_read_rgb16(s, raw_data); + } else { + // read in the data raw + for (j = 0; j < tga_comp; ++j) { + raw_data[j] = stbi__get8(s); + } + } + // clear the reading flag for the next pixel + read_next_pixel = 0; + } // end of reading a pixel + + // copy data + for (j = 0; j < tga_comp; ++j) + tga_data[i*tga_comp+j] = raw_data[j]; + + // in case we're in RLE mode, keep counting down + --RLE_count; + } + // do I need to invert the image? + if ( tga_inverted ) + { + for (j = 0; j*2 < tga_height; ++j) + { + int index1 = j * tga_width * tga_comp; + int index2 = (tga_height - 1 - j) * tga_width * tga_comp; + for (i = tga_width * tga_comp; i > 0; --i) + { + unsigned char temp = tga_data[index1]; + tga_data[index1] = tga_data[index2]; + tga_data[index2] = temp; + ++index1; + ++index2; + } + } + } + // clear my palette, if I had one + if ( tga_palette != NULL ) + { + STBI_FREE( tga_palette ); + } + } + + // swap RGB - if the source data was RGB16, it already is in the right order + if (tga_comp >= 3 && !tga_rgb16) + { + unsigned char* tga_pixel = tga_data; + for (i=0; i < tga_width * tga_height; ++i) + { + unsigned char temp = tga_pixel[0]; + tga_pixel[0] = tga_pixel[2]; + tga_pixel[2] = temp; + tga_pixel += tga_comp; + } + } + + // convert to target component count + if (req_comp && req_comp != tga_comp) + tga_data = stbi__convert_format(tga_data, tga_comp, req_comp, tga_width, tga_height); + + // the things I do to get rid of an error message, and yet keep + // Microsoft's C compilers happy... [8^( + tga_palette_start = tga_palette_len = tga_palette_bits = + tga_x_origin = tga_y_origin = 0; + // OK, done + return tga_data; +} +#endif + +// ************************************************************************************************* +// Photoshop PSD loader -- PD by Thatcher Ulrich, integration by Nicolas Schulz, tweaked by STB + +#ifndef STBI_NO_PSD +static int stbi__psd_test(stbi__context *s) +{ + int r = (stbi__get32be(s) == 0x38425053); + stbi__rewind(s); + return r; +} + +static int stbi__psd_decode_rle(stbi__context *s, stbi_uc *p, int pixelCount) +{ + int count, nleft, len; + + count = 0; + while ((nleft = pixelCount - count) > 0) { + len = stbi__get8(s); + if (len == 128) { + // No-op. + } else if (len < 128) { + // Copy next len+1 bytes literally. + len++; + if (len > nleft) return 0; // corrupt data + count += len; + while (len) { + *p = stbi__get8(s); + p += 4; + len--; + } + } else if (len > 128) { + stbi_uc val; + // Next -len+1 bytes in the dest are replicated from next source byte. + // (Interpret len as a negative 8-bit int.) + len = 257 - len; + if (len > nleft) return 0; // corrupt data + val = stbi__get8(s); + count += len; + while (len) { + *p = val; + p += 4; + len--; + } + } + } + + return 1; +} + +static void *stbi__psd_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri, int bpc) +{ + int pixelCount; + int channelCount, compression; + int channel, i; + int bitdepth; + int w,h; + stbi_uc *out; + STBI_NOTUSED(ri); + + // Check identifier + if (stbi__get32be(s) != 0x38425053) // "8BPS" + return stbi__errpuc("not PSD", "Corrupt PSD image"); + + // Check file type version. + if (stbi__get16be(s) != 1) + return stbi__errpuc("wrong version", "Unsupported version of PSD image"); + + // Skip 6 reserved bytes. + stbi__skip(s, 6 ); + + // Read the number of channels (R, G, B, A, etc). + channelCount = stbi__get16be(s); + if (channelCount < 0 || channelCount > 16) + return stbi__errpuc("wrong channel count", "Unsupported number of channels in PSD image"); + + // Read the rows and columns of the image. + h = stbi__get32be(s); + w = stbi__get32be(s); + + // Make sure the depth is 8 bits. + bitdepth = stbi__get16be(s); + if (bitdepth != 8 && bitdepth != 16) + return stbi__errpuc("unsupported bit depth", "PSD bit depth is not 8 or 16 bit"); + + // Make sure the color mode is RGB. + // Valid options are: + // 0: Bitmap + // 1: Grayscale + // 2: Indexed color + // 3: RGB color + // 4: CMYK color + // 7: Multichannel + // 8: Duotone + // 9: Lab color + if (stbi__get16be(s) != 3) + return stbi__errpuc("wrong color format", "PSD is not in RGB color format"); + + // Skip the Mode Data. (It's the palette for indexed color; other info for other modes.) + stbi__skip(s,stbi__get32be(s) ); + + // Skip the image resources. (resolution, pen tool paths, etc) + stbi__skip(s, stbi__get32be(s) ); + + // Skip the reserved data. + stbi__skip(s, stbi__get32be(s) ); + + // Find out if the data is compressed. + // Known values: + // 0: no compression + // 1: RLE compressed + compression = stbi__get16be(s); + if (compression > 1) + return stbi__errpuc("bad compression", "PSD has an unknown compression format"); + + // Check size + if (!stbi__mad3sizes_valid(4, w, h, 0)) + return stbi__errpuc("too large", "Corrupt PSD"); + + // Create the destination image. + + if (!compression && bitdepth == 16 && bpc == 16) { + out = (stbi_uc *) stbi__malloc_mad3(8, w, h, 0); + ri->bits_per_channel = 16; + } else + out = (stbi_uc *) stbi__malloc(4 * w*h); + + if (!out) return stbi__errpuc("outofmem", "Out of memory"); + pixelCount = w*h; + + // Initialize the data to zero. + //memset( out, 0, pixelCount * 4 ); + + // Finally, the image data. + if (compression) { + // RLE as used by .PSD and .TIFF + // Loop until you get the number of unpacked bytes you are expecting: + // Read the next source byte into n. + // If n is between 0 and 127 inclusive, copy the next n+1 bytes literally. + // Else if n is between -127 and -1 inclusive, copy the next byte -n+1 times. + // Else if n is 128, noop. + // Endloop + + // The RLE-compressed data is preceeded by a 2-byte data count for each row in the data, + // which we're going to just skip. + stbi__skip(s, h * channelCount * 2 ); + + // Read the RLE data by channel. + for (channel = 0; channel < 4; channel++) { + stbi_uc *p; + + p = out+channel; + if (channel >= channelCount) { + // Fill this channel with default data. + for (i = 0; i < pixelCount; i++, p += 4) + *p = (channel == 3 ? 255 : 0); + } else { + // Read the RLE data. + if (!stbi__psd_decode_rle(s, p, pixelCount)) { + STBI_FREE(out); + return stbi__errpuc("corrupt", "bad RLE data"); + } + } + } + + } else { + // We're at the raw image data. It's each channel in order (Red, Green, Blue, Alpha, ...) + // where each channel consists of an 8-bit (or 16-bit) value for each pixel in the image. + + // Read the data by channel. + for (channel = 0; channel < 4; channel++) { + if (channel >= channelCount) { + // Fill this channel with default data. + if (bitdepth == 16 && bpc == 16) { + stbi__uint16 *q = ((stbi__uint16 *) out) + channel; + stbi__uint16 val = channel == 3 ? 65535 : 0; + for (i = 0; i < pixelCount; i++, q += 4) + *q = val; + } else { + stbi_uc *p = out+channel; + stbi_uc val = channel == 3 ? 255 : 0; + for (i = 0; i < pixelCount; i++, p += 4) + *p = val; + } + } else { + if (ri->bits_per_channel == 16) { // output bpc + stbi__uint16 *q = ((stbi__uint16 *) out) + channel; + for (i = 0; i < pixelCount; i++, q += 4) + *q = (stbi__uint16) stbi__get16be(s); + } else { + stbi_uc *p = out+channel; + if (bitdepth == 16) { // input bpc + for (i = 0; i < pixelCount; i++, p += 4) + *p = (stbi_uc) (stbi__get16be(s) >> 8); + } else { + for (i = 0; i < pixelCount; i++, p += 4) + *p = stbi__get8(s); + } + } + } + } + } + + // remove weird white matte from PSD + if (channelCount >= 4) { + if (ri->bits_per_channel == 16) { + for (i=0; i < w*h; ++i) { + stbi__uint16 *pixel = (stbi__uint16 *) out + 4*i; + if (pixel[3] != 0 && pixel[3] != 65535) { + float a = pixel[3] / 65535.0f; + float ra = 1.0f / a; + float inv_a = 65535.0f * (1 - ra); + pixel[0] = (stbi__uint16) (pixel[0]*ra + inv_a); + pixel[1] = (stbi__uint16) (pixel[1]*ra + inv_a); + pixel[2] = (stbi__uint16) (pixel[2]*ra + inv_a); + } + } + } else { + for (i=0; i < w*h; ++i) { + unsigned char *pixel = out + 4*i; + if (pixel[3] != 0 && pixel[3] != 255) { + float a = pixel[3] / 255.0f; + float ra = 1.0f / a; + float inv_a = 255.0f * (1 - ra); + pixel[0] = (unsigned char) (pixel[0]*ra + inv_a); + pixel[1] = (unsigned char) (pixel[1]*ra + inv_a); + pixel[2] = (unsigned char) (pixel[2]*ra + inv_a); + } + } + } + } + + // convert to desired output format + if (req_comp && req_comp != 4) { + if (ri->bits_per_channel == 16) + out = (stbi_uc *) stbi__convert_format16((stbi__uint16 *) out, 4, req_comp, w, h); + else + out = stbi__convert_format(out, 4, req_comp, w, h); + if (out == NULL) return out; // stbi__convert_format frees input on failure + } + + if (comp) *comp = 4; + *y = h; + *x = w; + + return out; +} +#endif + +// ************************************************************************************************* +// Softimage PIC loader +// by Tom Seddon +// +// See http://softimage.wiki.softimage.com/index.php/INFO:_PIC_file_format +// See http://ozviz.wasp.uwa.edu.au/~pbourke/dataformats/softimagepic/ + +#ifndef STBI_NO_PIC +static int stbi__pic_is4(stbi__context *s,const char *str) +{ + int i; + for (i=0; i<4; ++i) + if (stbi__get8(s) != (stbi_uc)str[i]) + return 0; + + return 1; +} + +static int stbi__pic_test_core(stbi__context *s) +{ + int i; + + if (!stbi__pic_is4(s,"\x53\x80\xF6\x34")) + return 0; + + for(i=0;i<84;++i) + stbi__get8(s); + + if (!stbi__pic_is4(s,"PICT")) + return 0; + + return 1; +} + +typedef struct +{ + stbi_uc size,type,channel; +} stbi__pic_packet; + +static stbi_uc *stbi__readval(stbi__context *s, int channel, stbi_uc *dest) +{ + int mask=0x80, i; + + for (i=0; i<4; ++i, mask>>=1) { + if (channel & mask) { + if (stbi__at_eof(s)) return stbi__errpuc("bad file","PIC file too short"); + dest[i]=stbi__get8(s); + } + } + + return dest; +} + +static void stbi__copyval(int channel,stbi_uc *dest,const stbi_uc *src) +{ + int mask=0x80,i; + + for (i=0;i<4; ++i, mask>>=1) + if (channel&mask) + dest[i]=src[i]; +} + +static stbi_uc *stbi__pic_load_core(stbi__context *s,int width,int height,int *comp, stbi_uc *result) +{ + int act_comp=0,num_packets=0,y,chained; + stbi__pic_packet packets[10]; + + // this will (should...) cater for even some bizarre stuff like having data + // for the same channel in multiple packets. + do { + stbi__pic_packet *packet; + + if (num_packets==sizeof(packets)/sizeof(packets[0])) + return stbi__errpuc("bad format","too many packets"); + + packet = &packets[num_packets++]; + + chained = stbi__get8(s); + packet->size = stbi__get8(s); + packet->type = stbi__get8(s); + packet->channel = stbi__get8(s); + + act_comp |= packet->channel; + + if (stbi__at_eof(s)) return stbi__errpuc("bad file","file too short (reading packets)"); + if (packet->size != 8) return stbi__errpuc("bad format","packet isn't 8bpp"); + } while (chained); + + *comp = (act_comp & 0x10 ? 4 : 3); // has alpha channel? + + for(y=0; ytype) { + default: + return stbi__errpuc("bad format","packet has bad compression type"); + + case 0: {//uncompressed + int x; + + for(x=0;xchannel,dest)) + return 0; + break; + } + + case 1://Pure RLE + { + int left=width, i; + + while (left>0) { + stbi_uc count,value[4]; + + count=stbi__get8(s); + if (stbi__at_eof(s)) return stbi__errpuc("bad file","file too short (pure read count)"); + + if (count > left) + count = (stbi_uc) left; + + if (!stbi__readval(s,packet->channel,value)) return 0; + + for(i=0; ichannel,dest,value); + left -= count; + } + } + break; + + case 2: {//Mixed RLE + int left=width; + while (left>0) { + int count = stbi__get8(s), i; + if (stbi__at_eof(s)) return stbi__errpuc("bad file","file too short (mixed read count)"); + + if (count >= 128) { // Repeated + stbi_uc value[4]; + + if (count==128) + count = stbi__get16be(s); + else + count -= 127; + if (count > left) + return stbi__errpuc("bad file","scanline overrun"); + + if (!stbi__readval(s,packet->channel,value)) + return 0; + + for(i=0;ichannel,dest,value); + } else { // Raw + ++count; + if (count>left) return stbi__errpuc("bad file","scanline overrun"); + + for(i=0;ichannel,dest)) + return 0; + } + left-=count; + } + break; + } + } + } + } + + return result; +} + +static void *stbi__pic_load(stbi__context *s,int *px,int *py,int *comp,int req_comp, stbi__result_info *ri) +{ + stbi_uc *result; + int i, x,y, internal_comp; + STBI_NOTUSED(ri); + + if (!comp) comp = &internal_comp; + + for (i=0; i<92; ++i) + stbi__get8(s); + + x = stbi__get16be(s); + y = stbi__get16be(s); + if (stbi__at_eof(s)) return stbi__errpuc("bad file","file too short (pic header)"); + if (!stbi__mad3sizes_valid(x, y, 4, 0)) return stbi__errpuc("too large", "PIC image too large to decode"); + + stbi__get32be(s); //skip `ratio' + stbi__get16be(s); //skip `fields' + stbi__get16be(s); //skip `pad' + + // intermediate buffer is RGBA + result = (stbi_uc *) stbi__malloc_mad3(x, y, 4, 0); + memset(result, 0xff, x*y*4); + + if (!stbi__pic_load_core(s,x,y,comp, result)) { + STBI_FREE(result); + result=0; + } + *px = x; + *py = y; + if (req_comp == 0) req_comp = *comp; + result=stbi__convert_format(result,4,req_comp,x,y); + + return result; +} + +static int stbi__pic_test(stbi__context *s) +{ + int r = stbi__pic_test_core(s); + stbi__rewind(s); + return r; +} +#endif + +// ************************************************************************************************* +// GIF loader -- public domain by Jean-Marc Lienher -- simplified/shrunk by stb + +#ifndef STBI_NO_GIF +typedef struct +{ + stbi__int16 prefix; + stbi_uc first; + stbi_uc suffix; +} stbi__gif_lzw; + +typedef struct +{ + int w,h; + stbi_uc *out, *old_out; // output buffer (always 4 components) + int flags, bgindex, ratio, transparent, eflags, delay; + stbi_uc pal[256][4]; + stbi_uc lpal[256][4]; + stbi__gif_lzw codes[4096]; + stbi_uc *color_table; + int parse, step; + int lflags; + int start_x, start_y; + int max_x, max_y; + int cur_x, cur_y; + int line_size; +} stbi__gif; + +static int stbi__gif_test_raw(stbi__context *s) +{ + int sz; + if (stbi__get8(s) != 'G' || stbi__get8(s) != 'I' || stbi__get8(s) != 'F' || stbi__get8(s) != '8') return 0; + sz = stbi__get8(s); + if (sz != '9' && sz != '7') return 0; + if (stbi__get8(s) != 'a') return 0; + return 1; +} + +static int stbi__gif_test(stbi__context *s) +{ + int r = stbi__gif_test_raw(s); + stbi__rewind(s); + return r; +} + +static void stbi__gif_parse_colortable(stbi__context *s, stbi_uc pal[256][4], int num_entries, int transp) +{ + int i; + for (i=0; i < num_entries; ++i) { + pal[i][2] = stbi__get8(s); + pal[i][1] = stbi__get8(s); + pal[i][0] = stbi__get8(s); + pal[i][3] = transp == i ? 0 : 255; + } +} + +static int stbi__gif_header(stbi__context *s, stbi__gif *g, int *comp, int is_info) +{ + stbi_uc version; + if (stbi__get8(s) != 'G' || stbi__get8(s) != 'I' || stbi__get8(s) != 'F' || stbi__get8(s) != '8') + return stbi__err("not GIF", "Corrupt GIF"); + + version = stbi__get8(s); + if (version != '7' && version != '9') return stbi__err("not GIF", "Corrupt GIF"); + if (stbi__get8(s) != 'a') return stbi__err("not GIF", "Corrupt GIF"); + + stbi__g_failure_reason = ""; + g->w = stbi__get16le(s); + g->h = stbi__get16le(s); + g->flags = stbi__get8(s); + g->bgindex = stbi__get8(s); + g->ratio = stbi__get8(s); + g->transparent = -1; + + if (comp != 0) *comp = 4; // can't actually tell whether it's 3 or 4 until we parse the comments + + if (is_info) return 1; + + if (g->flags & 0x80) + stbi__gif_parse_colortable(s,g->pal, 2 << (g->flags & 7), -1); + + return 1; +} + +static int stbi__gif_info_raw(stbi__context *s, int *x, int *y, int *comp) +{ + stbi__gif* g = (stbi__gif*) stbi__malloc(sizeof(stbi__gif)); + if (!stbi__gif_header(s, g, comp, 1)) { + STBI_FREE(g); + stbi__rewind( s ); + return 0; + } + if (x) *x = g->w; + if (y) *y = g->h; + STBI_FREE(g); + return 1; +} + +static void stbi__out_gif_code(stbi__gif *g, stbi__uint16 code) +{ + stbi_uc *p, *c; + + // recurse to decode the prefixes, since the linked-list is backwards, + // and working backwards through an interleaved image would be nasty + if (g->codes[code].prefix >= 0) + stbi__out_gif_code(g, g->codes[code].prefix); + + if (g->cur_y >= g->max_y) return; + + p = &g->out[g->cur_x + g->cur_y]; + c = &g->color_table[g->codes[code].suffix * 4]; + + if (c[3] >= 128) { + p[0] = c[2]; + p[1] = c[1]; + p[2] = c[0]; + p[3] = c[3]; + } + g->cur_x += 4; + + if (g->cur_x >= g->max_x) { + g->cur_x = g->start_x; + g->cur_y += g->step; + + while (g->cur_y >= g->max_y && g->parse > 0) { + g->step = (1 << g->parse) * g->line_size; + g->cur_y = g->start_y + (g->step >> 1); + --g->parse; + } + } +} + +static stbi_uc *stbi__process_gif_raster(stbi__context *s, stbi__gif *g) +{ + stbi_uc lzw_cs; + stbi__int32 len, init_code; + stbi__uint32 first; + stbi__int32 codesize, codemask, avail, oldcode, bits, valid_bits, clear; + stbi__gif_lzw *p; + + lzw_cs = stbi__get8(s); + if (lzw_cs > 12) return NULL; + clear = 1 << lzw_cs; + first = 1; + codesize = lzw_cs + 1; + codemask = (1 << codesize) - 1; + bits = 0; + valid_bits = 0; + for (init_code = 0; init_code < clear; init_code++) { + g->codes[init_code].prefix = -1; + g->codes[init_code].first = (stbi_uc) init_code; + g->codes[init_code].suffix = (stbi_uc) init_code; + } + + // support no starting clear code + avail = clear+2; + oldcode = -1; + + len = 0; + for(;;) { + if (valid_bits < codesize) { + if (len == 0) { + len = stbi__get8(s); // start new block + if (len == 0) + return g->out; + } + --len; + bits |= (stbi__int32) stbi__get8(s) << valid_bits; + valid_bits += 8; + } else { + stbi__int32 code = bits & codemask; + bits >>= codesize; + valid_bits -= codesize; + // @OPTIMIZE: is there some way we can accelerate the non-clear path? + if (code == clear) { // clear code + codesize = lzw_cs + 1; + codemask = (1 << codesize) - 1; + avail = clear + 2; + oldcode = -1; + first = 0; + } else if (code == clear + 1) { // end of stream code + stbi__skip(s, len); + while ((len = stbi__get8(s)) > 0) + stbi__skip(s,len); + return g->out; + } else if (code <= avail) { + if (first) return stbi__errpuc("no clear code", "Corrupt GIF"); + + if (oldcode >= 0) { + p = &g->codes[avail++]; + if (avail > 4096) return stbi__errpuc("too many codes", "Corrupt GIF"); + p->prefix = (stbi__int16) oldcode; + p->first = g->codes[oldcode].first; + p->suffix = (code == avail) ? p->first : g->codes[code].first; + } else if (code == avail) + return stbi__errpuc("illegal code in raster", "Corrupt GIF"); + + stbi__out_gif_code(g, (stbi__uint16) code); + + if ((avail & codemask) == 0 && avail <= 0x0FFF) { + codesize++; + codemask = (1 << codesize) - 1; + } + + oldcode = code; + } else { + return stbi__errpuc("illegal code in raster", "Corrupt GIF"); + } + } + } +} + +static void stbi__fill_gif_background(stbi__gif *g, int x0, int y0, int x1, int y1) +{ + int x, y; + stbi_uc *c = g->pal[g->bgindex]; + for (y = y0; y < y1; y += 4 * g->w) { + for (x = x0; x < x1; x += 4) { + stbi_uc *p = &g->out[y + x]; + p[0] = c[2]; + p[1] = c[1]; + p[2] = c[0]; + p[3] = 0; + } + } +} + +// this function is designed to support animated gifs, although stb_image doesn't support it +static stbi_uc *stbi__gif_load_next(stbi__context *s, stbi__gif *g, int *comp, int req_comp) +{ + int i; + stbi_uc *prev_out = 0; + + if (g->out == 0 && !stbi__gif_header(s, g, comp,0)) + return 0; // stbi__g_failure_reason set by stbi__gif_header + + if (!stbi__mad3sizes_valid(g->w, g->h, 4, 0)) + return stbi__errpuc("too large", "GIF too large"); + + prev_out = g->out; + g->out = (stbi_uc *) stbi__malloc_mad3(4, g->w, g->h, 0); + if (g->out == 0) return stbi__errpuc("outofmem", "Out of memory"); + + switch ((g->eflags & 0x1C) >> 2) { + case 0: // unspecified (also always used on 1st frame) + stbi__fill_gif_background(g, 0, 0, 4 * g->w, 4 * g->w * g->h); + break; + case 1: // do not dispose + if (prev_out) memcpy(g->out, prev_out, 4 * g->w * g->h); + g->old_out = prev_out; + break; + case 2: // dispose to background + if (prev_out) memcpy(g->out, prev_out, 4 * g->w * g->h); + stbi__fill_gif_background(g, g->start_x, g->start_y, g->max_x, g->max_y); + break; + case 3: // dispose to previous + if (g->old_out) { + for (i = g->start_y; i < g->max_y; i += 4 * g->w) + memcpy(&g->out[i + g->start_x], &g->old_out[i + g->start_x], g->max_x - g->start_x); + } + break; + } + + for (;;) { + switch (stbi__get8(s)) { + case 0x2C: /* Image Descriptor */ + { + int prev_trans = -1; + stbi__int32 x, y, w, h; + stbi_uc *o; + + x = stbi__get16le(s); + y = stbi__get16le(s); + w = stbi__get16le(s); + h = stbi__get16le(s); + if (((x + w) > (g->w)) || ((y + h) > (g->h))) + return stbi__errpuc("bad Image Descriptor", "Corrupt GIF"); + + g->line_size = g->w * 4; + g->start_x = x * 4; + g->start_y = y * g->line_size; + g->max_x = g->start_x + w * 4; + g->max_y = g->start_y + h * g->line_size; + g->cur_x = g->start_x; + g->cur_y = g->start_y; + + g->lflags = stbi__get8(s); + + if (g->lflags & 0x40) { + g->step = 8 * g->line_size; // first interlaced spacing + g->parse = 3; + } else { + g->step = g->line_size; + g->parse = 0; + } + + if (g->lflags & 0x80) { + stbi__gif_parse_colortable(s,g->lpal, 2 << (g->lflags & 7), g->eflags & 0x01 ? g->transparent : -1); + g->color_table = (stbi_uc *) g->lpal; + } else if (g->flags & 0x80) { + if (g->transparent >= 0 && (g->eflags & 0x01)) { + prev_trans = g->pal[g->transparent][3]; + g->pal[g->transparent][3] = 0; + } + g->color_table = (stbi_uc *) g->pal; + } else + return stbi__errpuc("missing color table", "Corrupt GIF"); + + o = stbi__process_gif_raster(s, g); + if (o == NULL) return NULL; + + if (prev_trans != -1) + g->pal[g->transparent][3] = (stbi_uc) prev_trans; + + return o; + } + + case 0x21: // Comment Extension. + { + int len; + if (stbi__get8(s) == 0xF9) { // Graphic Control Extension. + len = stbi__get8(s); + if (len == 4) { + g->eflags = stbi__get8(s); + g->delay = stbi__get16le(s); + g->transparent = stbi__get8(s); + } else { + stbi__skip(s, len); + break; + } + } + while ((len = stbi__get8(s)) != 0) + stbi__skip(s, len); + break; + } + + case 0x3B: // gif stream termination code + return (stbi_uc *) s; // using '1' causes warning on some compilers + + default: + return stbi__errpuc("unknown code", "Corrupt GIF"); + } + } + + STBI_NOTUSED(req_comp); +} + +static void *stbi__gif_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri) +{ + stbi_uc *u = 0; + stbi__gif* g = (stbi__gif*) stbi__malloc(sizeof(stbi__gif)); + memset(g, 0, sizeof(*g)); + STBI_NOTUSED(ri); + + u = stbi__gif_load_next(s, g, comp, req_comp); + if (u == (stbi_uc *) s) u = 0; // end of animated gif marker + if (u) { + *x = g->w; + *y = g->h; + if (req_comp && req_comp != 4) + u = stbi__convert_format(u, 4, req_comp, g->w, g->h); + } + else if (g->out) + STBI_FREE(g->out); + STBI_FREE(g); + return u; +} + +static int stbi__gif_info(stbi__context *s, int *x, int *y, int *comp) +{ + return stbi__gif_info_raw(s,x,y,comp); +} +#endif + +// ************************************************************************************************* +// Radiance RGBE HDR loader +// originally by Nicolas Schulz +#ifndef STBI_NO_HDR +static int stbi__hdr_test_core(stbi__context *s, const char *signature) +{ + int i; + for (i=0; signature[i]; ++i) + if (stbi__get8(s) != signature[i]) + return 0; + stbi__rewind(s); + return 1; +} + +static int stbi__hdr_test(stbi__context* s) +{ + int r = stbi__hdr_test_core(s, "#?RADIANCE\n"); + stbi__rewind(s); + if(!r) { + r = stbi__hdr_test_core(s, "#?RGBE\n"); + stbi__rewind(s); + } + return r; +} + +#define STBI__HDR_BUFLEN 1024 +static char *stbi__hdr_gettoken(stbi__context *z, char *buffer) +{ + int len=0; + char c = '\0'; + + c = (char) stbi__get8(z); + + while (!stbi__at_eof(z) && c != '\n') { + buffer[len++] = c; + if (len == STBI__HDR_BUFLEN-1) { + // flush to end of line + while (!stbi__at_eof(z) && stbi__get8(z) != '\n') + ; + break; + } + c = (char) stbi__get8(z); + } + + buffer[len] = 0; + return buffer; +} + +static void stbi__hdr_convert(float *output, stbi_uc *input, int req_comp) +{ + if ( input[3] != 0 ) { + float f1; + // Exponent + f1 = (float) ldexp(1.0f, input[3] - (int)(128 + 8)); + if (req_comp <= 2) + output[0] = (input[0] + input[1] + input[2]) * f1 / 3; + else { + output[0] = input[0] * f1; + output[1] = input[1] * f1; + output[2] = input[2] * f1; + } + if (req_comp == 2) output[1] = 1; + if (req_comp == 4) output[3] = 1; + } else { + switch (req_comp) { + case 4: output[3] = 1; /* fallthrough */ + case 3: output[0] = output[1] = output[2] = 0; + break; + case 2: output[1] = 1; /* fallthrough */ + case 1: output[0] = 0; + break; + } + } +} + +static float *stbi__hdr_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri) +{ + char buffer[STBI__HDR_BUFLEN]; + char *token; + int valid = 0; + int width, height; + stbi_uc *scanline; + float *hdr_data; + int len; + unsigned char count, value; + int i, j, k, c1,c2, z; + const char *headerToken; + STBI_NOTUSED(ri); + + // Check identifier + headerToken = stbi__hdr_gettoken(s,buffer); + if (strcmp(headerToken, "#?RADIANCE") != 0 && strcmp(headerToken, "#?RGBE") != 0) + return stbi__errpf("not HDR", "Corrupt HDR image"); + + // Parse header + for(;;) { + token = stbi__hdr_gettoken(s,buffer); + if (token[0] == 0) break; + if (strcmp(token, "FORMAT=32-bit_rle_rgbe") == 0) valid = 1; + } + + if (!valid) return stbi__errpf("unsupported format", "Unsupported HDR format"); + + // Parse width and height + // can't use sscanf() if we're not using stdio! + token = stbi__hdr_gettoken(s,buffer); + if (strncmp(token, "-Y ", 3)) return stbi__errpf("unsupported data layout", "Unsupported HDR format"); + token += 3; + height = (int) strtol(token, &token, 10); + while (*token == ' ') ++token; + if (strncmp(token, "+X ", 3)) return stbi__errpf("unsupported data layout", "Unsupported HDR format"); + token += 3; + width = (int) strtol(token, NULL, 10); + + *x = width; + *y = height; + + if (comp) *comp = 3; + if (req_comp == 0) req_comp = 3; + + if (!stbi__mad4sizes_valid(width, height, req_comp, sizeof(float), 0)) + return stbi__errpf("too large", "HDR image is too large"); + + // Read data + hdr_data = (float *) stbi__malloc_mad4(width, height, req_comp, sizeof(float), 0); + if (!hdr_data) + return stbi__errpf("outofmem", "Out of memory"); + + // Load image data + // image data is stored as some number of sca + if ( width < 8 || width >= 32768) { + // Read flat data + for (j=0; j < height; ++j) { + for (i=0; i < width; ++i) { + stbi_uc rgbe[4]; + main_decode_loop: + stbi__getn(s, rgbe, 4); + stbi__hdr_convert(hdr_data + j * width * req_comp + i * req_comp, rgbe, req_comp); + } + } + } else { + // Read RLE-encoded data + scanline = NULL; + + for (j = 0; j < height; ++j) { + c1 = stbi__get8(s); + c2 = stbi__get8(s); + len = stbi__get8(s); + if (c1 != 2 || c2 != 2 || (len & 0x80)) { + // not run-length encoded, so we have to actually use THIS data as a decoded + // pixel (note this can't be a valid pixel--one of RGB must be >= 128) + stbi_uc rgbe[4]; + rgbe[0] = (stbi_uc) c1; + rgbe[1] = (stbi_uc) c2; + rgbe[2] = (stbi_uc) len; + rgbe[3] = (stbi_uc) stbi__get8(s); + stbi__hdr_convert(hdr_data, rgbe, req_comp); + i = 1; + j = 0; + STBI_FREE(scanline); + goto main_decode_loop; // yes, this makes no sense + } + len <<= 8; + len |= stbi__get8(s); + if (len != width) { STBI_FREE(hdr_data); STBI_FREE(scanline); return stbi__errpf("invalid decoded scanline length", "corrupt HDR"); } + if (scanline == NULL) { + scanline = (stbi_uc *) stbi__malloc_mad2(width, 4, 0); + if (!scanline) { + STBI_FREE(hdr_data); + return stbi__errpf("outofmem", "Out of memory"); + } + } + + for (k = 0; k < 4; ++k) { + int nleft; + i = 0; + while ((nleft = width - i) > 0) { + count = stbi__get8(s); + if (count > 128) { + // Run + value = stbi__get8(s); + count -= 128; + if (count > nleft) { STBI_FREE(hdr_data); STBI_FREE(scanline); return stbi__errpf("corrupt", "bad RLE data in HDR"); } + for (z = 0; z < count; ++z) + scanline[i++ * 4 + k] = value; + } else { + // Dump + if (count > nleft) { STBI_FREE(hdr_data); STBI_FREE(scanline); return stbi__errpf("corrupt", "bad RLE data in HDR"); } + for (z = 0; z < count; ++z) + scanline[i++ * 4 + k] = stbi__get8(s); + } + } + } + for (i=0; i < width; ++i) + stbi__hdr_convert(hdr_data+(j*width + i)*req_comp, scanline + i*4, req_comp); + } + if (scanline) + STBI_FREE(scanline); + } + + return hdr_data; +} + +static int stbi__hdr_info(stbi__context *s, int *x, int *y, int *comp) +{ + char buffer[STBI__HDR_BUFLEN]; + char *token; + int valid = 0; + int dummy; + + if (!x) x = &dummy; + if (!y) y = &dummy; + if (!comp) comp = &dummy; + + if (stbi__hdr_test(s) == 0) { + stbi__rewind( s ); + return 0; + } + + for(;;) { + token = stbi__hdr_gettoken(s,buffer); + if (token[0] == 0) break; + if (strcmp(token, "FORMAT=32-bit_rle_rgbe") == 0) valid = 1; + } + + if (!valid) { + stbi__rewind( s ); + return 0; + } + token = stbi__hdr_gettoken(s,buffer); + if (strncmp(token, "-Y ", 3)) { + stbi__rewind( s ); + return 0; + } + token += 3; + *y = (int) strtol(token, &token, 10); + while (*token == ' ') ++token; + if (strncmp(token, "+X ", 3)) { + stbi__rewind( s ); + return 0; + } + token += 3; + *x = (int) strtol(token, NULL, 10); + *comp = 3; + return 1; +} +#endif // STBI_NO_HDR + +#ifndef STBI_NO_BMP +static int stbi__bmp_info(stbi__context *s, int *x, int *y, int *comp) +{ + void *p; + stbi__bmp_data info; + + info.all_a = 255; + p = stbi__bmp_parse_header(s, &info); + stbi__rewind( s ); + if (p == NULL) + return 0; + if (x) *x = s->img_x; + if (y) *y = s->img_y; + if (comp) *comp = info.ma ? 4 : 3; + return 1; +} +#endif + +#ifndef STBI_NO_PSD +static int stbi__psd_info(stbi__context *s, int *x, int *y, int *comp) +{ + int channelCount, dummy; + if (!x) x = &dummy; + if (!y) y = &dummy; + if (!comp) comp = &dummy; + if (stbi__get32be(s) != 0x38425053) { + stbi__rewind( s ); + return 0; + } + if (stbi__get16be(s) != 1) { + stbi__rewind( s ); + return 0; + } + stbi__skip(s, 6); + channelCount = stbi__get16be(s); + if (channelCount < 0 || channelCount > 16) { + stbi__rewind( s ); + return 0; + } + *y = stbi__get32be(s); + *x = stbi__get32be(s); + if (stbi__get16be(s) != 8) { + stbi__rewind( s ); + return 0; + } + if (stbi__get16be(s) != 3) { + stbi__rewind( s ); + return 0; + } + *comp = 4; + return 1; +} +#endif + +#ifndef STBI_NO_PIC +static int stbi__pic_info(stbi__context *s, int *x, int *y, int *comp) +{ + int act_comp=0,num_packets=0,chained,dummy; + stbi__pic_packet packets[10]; + + if (!x) x = &dummy; + if (!y) y = &dummy; + if (!comp) comp = &dummy; + + if (!stbi__pic_is4(s,"\x53\x80\xF6\x34")) { + stbi__rewind(s); + return 0; + } + + stbi__skip(s, 88); + + *x = stbi__get16be(s); + *y = stbi__get16be(s); + if (stbi__at_eof(s)) { + stbi__rewind( s); + return 0; + } + if ( (*x) != 0 && (1 << 28) / (*x) < (*y)) { + stbi__rewind( s ); + return 0; + } + + stbi__skip(s, 8); + + do { + stbi__pic_packet *packet; + + if (num_packets==sizeof(packets)/sizeof(packets[0])) + return 0; + + packet = &packets[num_packets++]; + chained = stbi__get8(s); + packet->size = stbi__get8(s); + packet->type = stbi__get8(s); + packet->channel = stbi__get8(s); + act_comp |= packet->channel; + + if (stbi__at_eof(s)) { + stbi__rewind( s ); + return 0; + } + if (packet->size != 8) { + stbi__rewind( s ); + return 0; + } + } while (chained); + + *comp = (act_comp & 0x10 ? 4 : 3); + + return 1; +} +#endif + +// ************************************************************************************************* +// Portable Gray Map and Portable Pixel Map loader +// by Ken Miller +// +// PGM: http://netpbm.sourceforge.net/doc/pgm.html +// PPM: http://netpbm.sourceforge.net/doc/ppm.html +// +// Known limitations: +// Does not support comments in the header section +// Does not support ASCII image data (formats P2 and P3) +// Does not support 16-bit-per-channel + +#ifndef STBI_NO_PNM + +static int stbi__pnm_test(stbi__context *s) +{ + char p, t; + p = (char) stbi__get8(s); + t = (char) stbi__get8(s); + if (p != 'P' || (t != '5' && t != '6')) { + stbi__rewind( s ); + return 0; + } + return 1; +} + +static void *stbi__pnm_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri) +{ + stbi_uc *out; + STBI_NOTUSED(ri); + + if (!stbi__pnm_info(s, (int *)&s->img_x, (int *)&s->img_y, (int *)&s->img_n)) + return 0; + + *x = s->img_x; + *y = s->img_y; + if (comp) *comp = s->img_n; + + if (!stbi__mad3sizes_valid(s->img_n, s->img_x, s->img_y, 0)) + return stbi__errpuc("too large", "PNM too large"); + + out = (stbi_uc *) stbi__malloc_mad3(s->img_n, s->img_x, s->img_y, 0); + if (!out) return stbi__errpuc("outofmem", "Out of memory"); + stbi__getn(s, out, s->img_n * s->img_x * s->img_y); + + if (req_comp && req_comp != s->img_n) { + out = stbi__convert_format(out, s->img_n, req_comp, s->img_x, s->img_y); + if (out == NULL) return out; // stbi__convert_format frees input on failure + } + return out; +} + +static int stbi__pnm_isspace(char c) +{ + return c == ' ' || c == '\t' || c == '\n' || c == '\v' || c == '\f' || c == '\r'; +} + +static void stbi__pnm_skip_whitespace(stbi__context *s, char *c) +{ + for (;;) { + while (!stbi__at_eof(s) && stbi__pnm_isspace(*c)) + *c = (char) stbi__get8(s); + + if (stbi__at_eof(s) || *c != '#') + break; + + while (!stbi__at_eof(s) && *c != '\n' && *c != '\r' ) + *c = (char) stbi__get8(s); + } +} + +static int stbi__pnm_isdigit(char c) +{ + return c >= '0' && c <= '9'; +} + +static int stbi__pnm_getinteger(stbi__context *s, char *c) +{ + int value = 0; + + while (!stbi__at_eof(s) && stbi__pnm_isdigit(*c)) { + value = value*10 + (*c - '0'); + *c = (char) stbi__get8(s); + } + + return value; +} + +static int stbi__pnm_info(stbi__context *s, int *x, int *y, int *comp) +{ + int maxv, dummy; + char c, p, t; + + if (!x) x = &dummy; + if (!y) y = &dummy; + if (!comp) comp = &dummy; + + stbi__rewind(s); + + // Get identifier + p = (char) stbi__get8(s); + t = (char) stbi__get8(s); + if (p != 'P' || (t != '5' && t != '6')) { + stbi__rewind(s); + return 0; + } + + *comp = (t == '6') ? 3 : 1; // '5' is 1-component .pgm; '6' is 3-component .ppm + + c = (char) stbi__get8(s); + stbi__pnm_skip_whitespace(s, &c); + + *x = stbi__pnm_getinteger(s, &c); // read width + stbi__pnm_skip_whitespace(s, &c); + + *y = stbi__pnm_getinteger(s, &c); // read height + stbi__pnm_skip_whitespace(s, &c); + + maxv = stbi__pnm_getinteger(s, &c); // read max value + + if (maxv > 255) + return stbi__err("max value > 255", "PPM image not 8-bit"); + else + return 1; +} +#endif + +static int stbi__info_main(stbi__context *s, int *x, int *y, int *comp) +{ + #ifndef STBI_NO_JPEG + if (stbi__jpeg_info(s, x, y, comp)) return 1; + #endif + + #ifndef STBI_NO_PNG + if (stbi__png_info(s, x, y, comp)) return 1; + #endif + + #ifndef STBI_NO_GIF + if (stbi__gif_info(s, x, y, comp)) return 1; + #endif + + #ifndef STBI_NO_BMP + if (stbi__bmp_info(s, x, y, comp)) return 1; + #endif + + #ifndef STBI_NO_PSD + if (stbi__psd_info(s, x, y, comp)) return 1; + #endif + + #ifndef STBI_NO_PIC + if (stbi__pic_info(s, x, y, comp)) return 1; + #endif + + #ifndef STBI_NO_PNM + if (stbi__pnm_info(s, x, y, comp)) return 1; + #endif + + #ifndef STBI_NO_HDR + if (stbi__hdr_info(s, x, y, comp)) return 1; + #endif + + // test tga last because it's a crappy test! + #ifndef STBI_NO_TGA + if (stbi__tga_info(s, x, y, comp)) + return 1; + #endif + return stbi__err("unknown image type", "Image not of any known type, or corrupt"); +} + +#ifndef STBI_NO_STDIO +STBIDEF int stbi_info(char const *filename, int *x, int *y, int *comp) +{ + FILE *f = stbi__fopen(filename, "rb"); + int result; + if (!f) return stbi__err("can't fopen", "Unable to open file"); + result = stbi_info_from_file(f, x, y, comp); + fclose(f); + return result; +} + +STBIDEF int stbi_info_from_file(FILE *f, int *x, int *y, int *comp) +{ + int r; + stbi__context s; + long pos = ftell(f); + stbi__start_file(&s, f); + r = stbi__info_main(&s,x,y,comp); + fseek(f,pos,SEEK_SET); + return r; +} +#endif // !STBI_NO_STDIO + +STBIDEF int stbi_info_from_memory(stbi_uc const *buffer, int len, int *x, int *y, int *comp) +{ + stbi__context s; + stbi__start_mem(&s,buffer,len); + return stbi__info_main(&s,x,y,comp); +} + +STBIDEF int stbi_info_from_callbacks(stbi_io_callbacks const *c, void *user, int *x, int *y, int *comp) +{ + stbi__context s; + stbi__start_callbacks(&s, (stbi_io_callbacks *) c, user); + return stbi__info_main(&s,x,y,comp); +} + +#endif // STB_IMAGE_IMPLEMENTATION + +/* + revision history: + 2.16 (2017-07-23) all functions have 16-bit variants; + STBI_NO_STDIO works again; + compilation fixes; + fix rounding in unpremultiply; + optimize vertical flip; + disable raw_len validation; + documentation fixes + 2.15 (2017-03-18) fix png-1,2,4 bug; now all Imagenet JPGs decode; + warning fixes; disable run-time SSE detection on gcc; + uniform handling of optional "return" values; + thread-safe initialization of zlib tables + 2.14 (2017-03-03) remove deprecated STBI_JPEG_OLD; fixes for Imagenet JPGs + 2.13 (2016-11-29) add 16-bit API, only supported for PNG right now + 2.12 (2016-04-02) fix typo in 2.11 PSD fix that caused crashes + 2.11 (2016-04-02) allocate large structures on the stack + remove white matting for transparent PSD + fix reported channel count for PNG & BMP + re-enable SSE2 in non-gcc 64-bit + support RGB-formatted JPEG + read 16-bit PNGs (only as 8-bit) + 2.10 (2016-01-22) avoid warning introduced in 2.09 by STBI_REALLOC_SIZED + 2.09 (2016-01-16) allow comments in PNM files + 16-bit-per-pixel TGA (not bit-per-component) + info() for TGA could break due to .hdr handling + info() for BMP to shares code instead of sloppy parse + can use STBI_REALLOC_SIZED if allocator doesn't support realloc + code cleanup + 2.08 (2015-09-13) fix to 2.07 cleanup, reading RGB PSD as RGBA + 2.07 (2015-09-13) fix compiler warnings + partial animated GIF support + limited 16-bpc PSD support + #ifdef unused functions + bug with < 92 byte PIC,PNM,HDR,TGA + 2.06 (2015-04-19) fix bug where PSD returns wrong '*comp' value + 2.05 (2015-04-19) fix bug in progressive JPEG handling, fix warning + 2.04 (2015-04-15) try to re-enable SIMD on MinGW 64-bit + 2.03 (2015-04-12) extra corruption checking (mmozeiko) + stbi_set_flip_vertically_on_load (nguillemot) + fix NEON support; fix mingw support + 2.02 (2015-01-19) fix incorrect assert, fix warning + 2.01 (2015-01-17) fix various warnings; suppress SIMD on gcc 32-bit without -msse2 + 2.00b (2014-12-25) fix STBI_MALLOC in progressive JPEG + 2.00 (2014-12-25) optimize JPG, including x86 SSE2 & NEON SIMD (ryg) + progressive JPEG (stb) + PGM/PPM support (Ken Miller) + STBI_MALLOC,STBI_REALLOC,STBI_FREE + GIF bugfix -- seemingly never worked + STBI_NO_*, STBI_ONLY_* + 1.48 (2014-12-14) fix incorrectly-named assert() + 1.47 (2014-12-14) 1/2/4-bit PNG support, both direct and paletted (Omar Cornut & stb) + optimize PNG (ryg) + fix bug in interlaced PNG with user-specified channel count (stb) + 1.46 (2014-08-26) + fix broken tRNS chunk (colorkey-style transparency) in non-paletted PNG + 1.45 (2014-08-16) + fix MSVC-ARM internal compiler error by wrapping malloc + 1.44 (2014-08-07) + various warning fixes from Ronny Chevalier + 1.43 (2014-07-15) + fix MSVC-only compiler problem in code changed in 1.42 + 1.42 (2014-07-09) + don't define _CRT_SECURE_NO_WARNINGS (affects user code) + fixes to stbi__cleanup_jpeg path + added STBI_ASSERT to avoid requiring assert.h + 1.41 (2014-06-25) + fix search&replace from 1.36 that messed up comments/error messages + 1.40 (2014-06-22) + fix gcc struct-initialization warning + 1.39 (2014-06-15) + fix to TGA optimization when req_comp != number of components in TGA; + fix to GIF loading because BMP wasn't rewinding (whoops, no GIFs in my test suite) + add support for BMP version 5 (more ignored fields) + 1.38 (2014-06-06) + suppress MSVC warnings on integer casts truncating values + fix accidental rename of 'skip' field of I/O + 1.37 (2014-06-04) + remove duplicate typedef + 1.36 (2014-06-03) + convert to header file single-file library + if de-iphone isn't set, load iphone images color-swapped instead of returning NULL + 1.35 (2014-05-27) + various warnings + fix broken STBI_SIMD path + fix bug where stbi_load_from_file no longer left file pointer in correct place + fix broken non-easy path for 32-bit BMP (possibly never used) + TGA optimization by Arseny Kapoulkine + 1.34 (unknown) + use STBI_NOTUSED in stbi__resample_row_generic(), fix one more leak in tga failure case + 1.33 (2011-07-14) + make stbi_is_hdr work in STBI_NO_HDR (as specified), minor compiler-friendly improvements + 1.32 (2011-07-13) + support for "info" function for all supported filetypes (SpartanJ) + 1.31 (2011-06-20) + a few more leak fixes, bug in PNG handling (SpartanJ) + 1.30 (2011-06-11) + added ability to load files via callbacks to accomidate custom input streams (Ben Wenger) + removed deprecated format-specific test/load functions + removed support for installable file formats (stbi_loader) -- would have been broken for IO callbacks anyway + error cases in bmp and tga give messages and don't leak (Raymond Barbiero, grisha) + fix inefficiency in decoding 32-bit BMP (David Woo) + 1.29 (2010-08-16) + various warning fixes from Aurelien Pocheville + 1.28 (2010-08-01) + fix bug in GIF palette transparency (SpartanJ) + 1.27 (2010-08-01) + cast-to-stbi_uc to fix warnings + 1.26 (2010-07-24) + fix bug in file buffering for PNG reported by SpartanJ + 1.25 (2010-07-17) + refix trans_data warning (Won Chun) + 1.24 (2010-07-12) + perf improvements reading from files on platforms with lock-heavy fgetc() + minor perf improvements for jpeg + deprecated type-specific functions so we'll get feedback if they're needed + attempt to fix trans_data warning (Won Chun) + 1.23 fixed bug in iPhone support + 1.22 (2010-07-10) + removed image *writing* support + stbi_info support from Jetro Lauha + GIF support from Jean-Marc Lienher + iPhone PNG-extensions from James Brown + warning-fixes from Nicolas Schulz and Janez Zemva (i.stbi__err. Janez (U+017D)emva) + 1.21 fix use of 'stbi_uc' in header (reported by jon blow) + 1.20 added support for Softimage PIC, by Tom Seddon + 1.19 bug in interlaced PNG corruption check (found by ryg) + 1.18 (2008-08-02) + fix a threading bug (local mutable static) + 1.17 support interlaced PNG + 1.16 major bugfix - stbi__convert_format converted one too many pixels + 1.15 initialize some fields for thread safety + 1.14 fix threadsafe conversion bug + header-file-only version (#define STBI_HEADER_FILE_ONLY before including) + 1.13 threadsafe + 1.12 const qualifiers in the API + 1.11 Support installable IDCT, colorspace conversion routines + 1.10 Fixes for 64-bit (don't use "unsigned long") + optimized upsampling by Fabian "ryg" Giesen + 1.09 Fix format-conversion for PSD code (bad global variables!) + 1.08 Thatcher Ulrich's PSD code integrated by Nicolas Schulz + 1.07 attempt to fix C++ warning/errors again + 1.06 attempt to fix C++ warning/errors again + 1.05 fix TGA loading to return correct *comp and use good luminance calc + 1.04 default float alpha is 1, not 255; use 'void *' for stbi_image_free + 1.03 bugfixes to STBI_NO_STDIO, STBI_NO_HDR + 1.02 support for (subset of) HDR files, float interface for preferred access to them + 1.01 fix bug: possible bug in handling right-side up bmps... not sure + fix bug: the stbi__bmp_load() and stbi__tga_load() functions didn't work at all + 1.00 interface to zlib that skips zlib header + 0.99 correct handling of alpha in palette + 0.98 TGA loader by lonesock; dynamically add loaders (untested) + 0.97 jpeg errors on too large a file; also catch another malloc failure + 0.96 fix detection of invalid v value - particleman@mollyrocket forum + 0.95 during header scan, seek to markers in case of padding + 0.94 STBI_NO_STDIO to disable stdio usage; rename all #defines the same + 0.93 handle jpegtran output; verbose errors + 0.92 read 4,8,16,24,32-bit BMP files of several formats + 0.91 output 24-bit Windows 3.0 BMP files + 0.90 fix a few more warnings; bump version number to approach 1.0 + 0.61 bugfixes due to Marc LeBlanc, Christopher Lloyd + 0.60 fix compiling as c++ + 0.59 fix warnings: merge Dave Moore's -Wall fixes + 0.58 fix bug: zlib uncompressed mode len/nlen was wrong endian + 0.57 fix bug: jpg last huffman symbol before marker was >9 bits but less than 16 available + 0.56 fix bug: zlib uncompressed mode len vs. nlen + 0.55 fix bug: restart_interval not initialized to 0 + 0.54 allow NULL for 'int *comp' + 0.53 fix bug in png 3->4; speedup png decoding + 0.52 png handles req_comp=3,4 directly; minor cleanup; jpeg comments + 0.51 obey req_comp requests, 1-component jpegs return as 1-component, + on 'test' only check type, not whether we support this variant + 0.50 (2006-11-19) + first released version +*/ + + +/* +------------------------------------------------------------------------------ +This software is available under 2 licenses -- choose whichever you prefer. +------------------------------------------------------------------------------ +ALTERNATIVE A - MIT License +Copyright (c) 2017 Sean Barrett +Permission is hereby granted, free of charge, to any person obtaining a copy of +this software and associated documentation files (the "Software"), to deal in +the Software without restriction, including without limitation the rights to +use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies +of the Software, and to permit persons to whom the Software is furnished to do +so, subject to the following conditions: +The above copyright notice and this permission notice shall be included in all +copies or substantial portions of the Software. +THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, +OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE +SOFTWARE. +------------------------------------------------------------------------------ +ALTERNATIVE B - Public Domain (www.unlicense.org) +This is free and unencumbered software released into the public domain. +Anyone is free to copy, modify, publish, use, compile, sell, or distribute this +software, either in source code form or as a compiled binary, for any purpose, +commercial or non-commercial, and by any means. +In jurisdictions that recognize copyright laws, the author or authors of this +software dedicate any and all copyright interest in the software to the public +domain. We make this dedication for the benefit of the public at large and to +the detriment of our heirs and successors. We intend this dedication to be an +overt act of relinquishment in perpetuity of all present and future rights to +this software under copyright law. +THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +AUTHORS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN +ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION +WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. +------------------------------------------------------------------------------ +*/ + diff --git a/include/tkDNN/stb_image_write.h b/include/tkDNN/stb_image_write.h new file mode 100644 index 0000000..95943eb --- /dev/null +++ b/include/tkDNN/stb_image_write.h @@ -0,0 +1,1690 @@ +/* stb_image_write - v1.15 - public domain - http://nothings.org/stb + writes out PNG/BMP/TGA/JPEG/HDR images to C stdio - Sean Barrett 2010-2015 + no warranty implied; use at your own risk + + Before #including, + + #define STB_IMAGE_WRITE_IMPLEMENTATION + + in the file that you want to have the implementation. + + Will probably not work correctly with strict-aliasing optimizations. + +ABOUT: + + This header file is a library for writing images to C stdio or a callback. + + The PNG output is not optimal; it is 20-50% larger than the file + written by a decent optimizing implementation; though providing a custom + zlib compress function (see STBIW_ZLIB_COMPRESS) can mitigate that. + This library is designed for source code compactness and simplicity, + not optimal image file size or run-time performance. + +BUILDING: + + You can #define STBIW_ASSERT(x) before the #include to avoid using assert.h. + You can #define STBIW_MALLOC(), STBIW_REALLOC(), and STBIW_FREE() to replace + malloc,realloc,free. + You can #define STBIW_MEMMOVE() to replace memmove() + You can #define STBIW_ZLIB_COMPRESS to use a custom zlib-style compress function + for PNG compression (instead of the builtin one), it must have the following signature: + unsigned char * my_compress(unsigned char *data, int data_len, int *out_len, int quality); + The returned data will be freed with STBIW_FREE() (free() by default), + so it must be heap allocated with STBIW_MALLOC() (malloc() by default), + +UNICODE: + + If compiling for Windows and you wish to use Unicode filenames, compile + with + #define STBIW_WINDOWS_UTF8 + and pass utf8-encoded filenames. Call stbiw_convert_wchar_to_utf8 to convert + Windows wchar_t filenames to utf8. + +USAGE: + + There are five functions, one for each image file format: + + int stbi_write_png(char const *filename, int w, int h, int comp, const void *data, int stride_in_bytes); + int stbi_write_bmp(char const *filename, int w, int h, int comp, const void *data); + int stbi_write_tga(char const *filename, int w, int h, int comp, const void *data); + int stbi_write_jpg(char const *filename, int w, int h, int comp, const void *data, int quality); + int stbi_write_hdr(char const *filename, int w, int h, int comp, const float *data); + + void stbi_flip_vertically_on_write(int flag); // flag is non-zero to flip data vertically + + There are also five equivalent functions that use an arbitrary write function. You are + expected to open/close your file-equivalent before and after calling these: + + int stbi_write_png_to_func(stbi_write_func *func, void *context, int w, int h, int comp, const void *data, int stride_in_bytes); + int stbi_write_bmp_to_func(stbi_write_func *func, void *context, int w, int h, int comp, const void *data); + int stbi_write_tga_to_func(stbi_write_func *func, void *context, int w, int h, int comp, const void *data); + int stbi_write_hdr_to_func(stbi_write_func *func, void *context, int w, int h, int comp, const float *data); + int stbi_write_jpg_to_func(stbi_write_func *func, void *context, int x, int y, int comp, const void *data, int quality); + + where the callback is: + void stbi_write_func(void *context, void *data, int size); + + You can configure it with these global variables: + int stbi_write_tga_with_rle; // defaults to true; set to 0 to disable RLE + int stbi_write_png_compression_level; // defaults to 8; set to higher for more compression + int stbi_write_force_png_filter; // defaults to -1; set to 0..5 to force a filter mode + + + You can define STBI_WRITE_NO_STDIO to disable the file variant of these + functions, so the library will not use stdio.h at all. However, this will + also disable HDR writing, because it requires stdio for formatted output. + + Each function returns 0 on failure and non-0 on success. + + The functions create an image file defined by the parameters. The image + is a rectangle of pixels stored from left-to-right, top-to-bottom. + Each pixel contains 'comp' channels of data stored interleaved with 8-bits + per channel, in the following order: 1=Y, 2=YA, 3=RGB, 4=RGBA. (Y is + monochrome color.) The rectangle is 'w' pixels wide and 'h' pixels tall. + The *data pointer points to the first byte of the top-left-most pixel. + For PNG, "stride_in_bytes" is the distance in bytes from the first byte of + a row of pixels to the first byte of the next row of pixels. + + PNG creates output files with the same number of components as the input. + The BMP format expands Y to RGB in the file format and does not + output alpha. + + PNG supports writing rectangles of data even when the bytes storing rows of + data are not consecutive in memory (e.g. sub-rectangles of a larger image), + by supplying the stride between the beginning of adjacent rows. The other + formats do not. (Thus you cannot write a native-format BMP through the BMP + writer, both because it is in BGR order and because it may have padding + at the end of the line.) + + PNG allows you to set the deflate compression level by setting the global + variable 'stbi_write_png_compression_level' (it defaults to 8). + + HDR expects linear float data. Since the format is always 32-bit rgb(e) + data, alpha (if provided) is discarded, and for monochrome data it is + replicated across all three channels. + + TGA supports RLE or non-RLE compressed data. To use non-RLE-compressed + data, set the global variable 'stbi_write_tga_with_rle' to 0. + + JPEG does ignore alpha channels in input data; quality is between 1 and 100. + Higher quality looks better but results in a bigger image. + JPEG baseline (no JPEG progressive). + +CREDITS: + + + Sean Barrett - PNG/BMP/TGA + Baldur Karlsson - HDR + Jean-Sebastien Guay - TGA monochrome + Tim Kelsey - misc enhancements + Alan Hickman - TGA RLE + Emmanuel Julien - initial file IO callback implementation + Jon Olick - original jo_jpeg.cpp code + Daniel Gibson - integrate JPEG, allow external zlib + Aarni Koskela - allow choosing PNG filter + + bugfixes: + github:Chribba + Guillaume Chereau + github:jry2 + github:romigrou + Sergio Gonzalez + Jonas Karlsson + Filip Wasil + Thatcher Ulrich + github:poppolopoppo + Patrick Boettcher + github:xeekworx + Cap Petschulat + Simon Rodriguez + Ivan Tikhonov + github:ignotion + Adam Schackart + +LICENSE + + See end of file for license information. + +*/ + +#ifndef INCLUDE_STB_IMAGE_WRITE_H +#define INCLUDE_STB_IMAGE_WRITE_H + +#include + +// if STB_IMAGE_WRITE_STATIC causes problems, try defining STBIWDEF to 'inline' or 'static inline' +#ifndef STBIWDEF +#ifdef STB_IMAGE_WRITE_STATIC +#define STBIWDEF static +#else +#ifdef __cplusplus +#define STBIWDEF extern "C" +#else +#define STBIWDEF extern +#endif +#endif +#endif + +#ifndef STB_IMAGE_WRITE_STATIC // C++ forbids static forward declarations +extern int stbi_write_tga_with_rle; +extern int stbi_write_png_compression_level; +extern int stbi_write_force_png_filter; +#endif + +#ifndef STBI_WRITE_NO_STDIO +STBIWDEF int stbi_write_png(char const *filename, int w, int h, int comp, const void *data, int stride_in_bytes); +STBIWDEF int stbi_write_bmp(char const *filename, int w, int h, int comp, const void *data); +STBIWDEF int stbi_write_tga(char const *filename, int w, int h, int comp, const void *data); +STBIWDEF int stbi_write_hdr(char const *filename, int w, int h, int comp, const float *data); +STBIWDEF int stbi_write_jpg(char const *filename, int x, int y, int comp, const void *data, int quality); + +#ifdef STBI_WINDOWS_UTF8 +STBIWDEF int stbiw_convert_wchar_to_utf8(char *buffer, size_t bufferlen, const wchar_t* input); +#endif +#endif + +typedef void stbi_write_func(void *context, void *data, int size); + +STBIWDEF int stbi_write_png_to_func(stbi_write_func *func, void *context, int w, int h, int comp, const void *data, int stride_in_bytes); +STBIWDEF int stbi_write_bmp_to_func(stbi_write_func *func, void *context, int w, int h, int comp, const void *data); +STBIWDEF int stbi_write_tga_to_func(stbi_write_func *func, void *context, int w, int h, int comp, const void *data); +STBIWDEF int stbi_write_hdr_to_func(stbi_write_func *func, void *context, int w, int h, int comp, const float *data); +STBIWDEF int stbi_write_jpg_to_func(stbi_write_func *func, void *context, int x, int y, int comp, const void *data, int quality); + +STBIWDEF void stbi_flip_vertically_on_write(int flip_boolean); + +#endif//INCLUDE_STB_IMAGE_WRITE_H + +#ifdef STB_IMAGE_WRITE_IMPLEMENTATION + +#ifdef _WIN32 + #ifndef _CRT_SECURE_NO_WARNINGS + #define _CRT_SECURE_NO_WARNINGS + #endif + #ifndef _CRT_NONSTDC_NO_DEPRECATE + #define _CRT_NONSTDC_NO_DEPRECATE + #endif +#endif + +#ifndef STBI_WRITE_NO_STDIO +#include +#endif // STBI_WRITE_NO_STDIO + +#include +#include +#include +#include + +#if defined(STBIW_MALLOC) && defined(STBIW_FREE) && (defined(STBIW_REALLOC) || defined(STBIW_REALLOC_SIZED)) +// ok +#elif !defined(STBIW_MALLOC) && !defined(STBIW_FREE) && !defined(STBIW_REALLOC) && !defined(STBIW_REALLOC_SIZED) +// ok +#else +#error "Must define all or none of STBIW_MALLOC, STBIW_FREE, and STBIW_REALLOC (or STBIW_REALLOC_SIZED)." +#endif + +#ifndef STBIW_MALLOC +#define STBIW_MALLOC(sz) malloc(sz) +#define STBIW_REALLOC(p,newsz) realloc(p,newsz) +#define STBIW_FREE(p) free(p) +#endif + +#ifndef STBIW_REALLOC_SIZED +#define STBIW_REALLOC_SIZED(p,oldsz,newsz) STBIW_REALLOC(p,newsz) +#endif + + +#ifndef STBIW_MEMMOVE +#define STBIW_MEMMOVE(a,b,sz) memmove(a,b,sz) +#endif + + +#ifndef STBIW_ASSERT +#include +#define STBIW_ASSERT(x) assert(x) +#endif + +#define STBIW_UCHAR(x) (unsigned char) ((x) & 0xff) + +#ifdef STB_IMAGE_WRITE_STATIC +static int stbi_write_png_compression_level = 8; +static int stbi_write_tga_with_rle = 1; +static int stbi_write_force_png_filter = -1; +#else +int stbi_write_png_compression_level = 8; +int stbi_write_tga_with_rle = 1; +int stbi_write_force_png_filter = -1; +#endif + +static int stbi__flip_vertically_on_write = 0; + +STBIWDEF void stbi_flip_vertically_on_write(int flag) +{ + stbi__flip_vertically_on_write = flag; +} + +typedef struct +{ + stbi_write_func *func; + void *context; + unsigned char buffer[64]; + int buf_used; +} stbi__write_context; + +// initialize a callback-based context +static void stbi__start_write_callbacks(stbi__write_context *s, stbi_write_func *c, void *context) +{ + s->func = c; + s->context = context; +} + +#ifndef STBI_WRITE_NO_STDIO + +static void stbi__stdio_write(void *context, void *data, int size) +{ + fwrite(data,1,size,(FILE*) context); +} + +#if defined(_MSC_VER) && defined(STBI_WINDOWS_UTF8) +#ifdef __cplusplus +#define STBIW_EXTERN extern "C" +#else +#define STBIW_EXTERN extern +#endif +STBIW_EXTERN __declspec(dllimport) int __stdcall MultiByteToWideChar(unsigned int cp, unsigned long flags, const char *str, int cbmb, wchar_t *widestr, int cchwide); +STBIW_EXTERN __declspec(dllimport) int __stdcall WideCharToMultiByte(unsigned int cp, unsigned long flags, const wchar_t *widestr, int cchwide, char *str, int cbmb, const char *defchar, int *used_default); + +STBIWDEF int stbiw_convert_wchar_to_utf8(char *buffer, size_t bufferlen, const wchar_t* input) +{ + return WideCharToMultiByte(65001 /* UTF8 */, 0, input, -1, buffer, (int) bufferlen, NULL, NULL); +} +#endif + +static FILE *stbiw__fopen(char const *filename, char const *mode) +{ + FILE *f; +#if defined(_MSC_VER) && defined(STBI_WINDOWS_UTF8) + wchar_t wMode[64]; + wchar_t wFilename[1024]; + if (0 == MultiByteToWideChar(65001 /* UTF8 */, 0, filename, -1, wFilename, sizeof(wFilename))) + return 0; + + if (0 == MultiByteToWideChar(65001 /* UTF8 */, 0, mode, -1, wMode, sizeof(wMode))) + return 0; + +#if _MSC_VER >= 1400 + if (0 != _wfopen_s(&f, wFilename, wMode)) + f = 0; +#else + f = _wfopen(wFilename, wMode); +#endif + +#elif defined(_MSC_VER) && _MSC_VER >= 1400 + if (0 != fopen_s(&f, filename, mode)) + f=0; +#else + f = fopen(filename, mode); +#endif + return f; +} + +static int stbi__start_write_file(stbi__write_context *s, const char *filename) +{ + FILE *f = stbiw__fopen(filename, "wb"); + stbi__start_write_callbacks(s, stbi__stdio_write, (void *) f); + return f != NULL; +} + +static void stbi__end_write_file(stbi__write_context *s) +{ + fclose((FILE *)s->context); +} + +#endif // !STBI_WRITE_NO_STDIO + +typedef unsigned int stbiw_uint32; +typedef int stb_image_write_test[sizeof(stbiw_uint32)==4 ? 1 : -1]; + +static void stbiw__writefv(stbi__write_context *s, const char *fmt, va_list v) +{ + while (*fmt) { + switch (*fmt++) { + case ' ': break; + case '1': { unsigned char x = STBIW_UCHAR(va_arg(v, int)); + s->func(s->context,&x,1); + break; } + case '2': { int x = va_arg(v,int); + unsigned char b[2]; + b[0] = STBIW_UCHAR(x); + b[1] = STBIW_UCHAR(x>>8); + s->func(s->context,b,2); + break; } + case '4': { stbiw_uint32 x = va_arg(v,int); + unsigned char b[4]; + b[0]=STBIW_UCHAR(x); + b[1]=STBIW_UCHAR(x>>8); + b[2]=STBIW_UCHAR(x>>16); + b[3]=STBIW_UCHAR(x>>24); + s->func(s->context,b,4); + break; } + default: + STBIW_ASSERT(0); + return; + } + } +} + +static void stbiw__writef(stbi__write_context *s, const char *fmt, ...) +{ + va_list v; + va_start(v, fmt); + stbiw__writefv(s, fmt, v); + va_end(v); +} + +static void stbiw__write_flush(stbi__write_context *s) +{ + if (s->buf_used) { + s->func(s->context, &s->buffer, s->buf_used); + s->buf_used = 0; + } +} + +static void stbiw__putc(stbi__write_context *s, unsigned char c) +{ + s->func(s->context, &c, 1); +} + +static void stbiw__write1(stbi__write_context *s, unsigned char a) +{ + if (s->buf_used + 1 > sizeof(s->buffer)) + stbiw__write_flush(s); + s->buffer[s->buf_used++] = a; +} + +static void stbiw__write3(stbi__write_context *s, unsigned char a, unsigned char b, unsigned char c) +{ + int n; + if (s->buf_used + 3 > sizeof(s->buffer)) + stbiw__write_flush(s); + n = s->buf_used; + s->buf_used = n+3; + s->buffer[n+0] = a; + s->buffer[n+1] = b; + s->buffer[n+2] = c; +} + +static void stbiw__write_pixel(stbi__write_context *s, int rgb_dir, int comp, int write_alpha, int expand_mono, unsigned char *d) +{ + unsigned char bg[3] = { 255, 0, 255}, px[3]; + int k; + + if (write_alpha < 0) + stbiw__write1(s, d[comp - 1]); + + switch (comp) { + case 2: // 2 pixels = mono + alpha, alpha is written separately, so same as 1-channel case + case 1: + if (expand_mono) + stbiw__write3(s, d[0], d[0], d[0]); // monochrome bmp + else + stbiw__write1(s, d[0]); // monochrome TGA + break; + case 4: + if (!write_alpha) { + // composite against pink background + for (k = 0; k < 3; ++k) + px[k] = bg[k] + ((d[k] - bg[k]) * d[3]) / 255; + stbiw__write3(s, px[1 - rgb_dir], px[1], px[1 + rgb_dir]); + break; + } + /* FALLTHROUGH */ + case 3: + stbiw__write3(s, d[1 - rgb_dir], d[1], d[1 + rgb_dir]); + break; + } + if (write_alpha > 0) + stbiw__write1(s, d[comp - 1]); +} + +static void stbiw__write_pixels(stbi__write_context *s, int rgb_dir, int vdir, int x, int y, int comp, void *data, int write_alpha, int scanline_pad, int expand_mono) +{ + stbiw_uint32 zero = 0; + int i,j, j_end; + + if (y <= 0) + return; + + if (stbi__flip_vertically_on_write) + vdir *= -1; + + if (vdir < 0) { + j_end = -1; j = y-1; + } else { + j_end = y; j = 0; + } + + for (; j != j_end; j += vdir) { + for (i=0; i < x; ++i) { + unsigned char *d = (unsigned char *) data + (j*x+i)*comp; + stbiw__write_pixel(s, rgb_dir, comp, write_alpha, expand_mono, d); + } + stbiw__write_flush(s); + s->func(s->context, &zero, scanline_pad); + } +} + +static int stbiw__outfile(stbi__write_context *s, int rgb_dir, int vdir, int x, int y, int comp, int expand_mono, void *data, int alpha, int pad, const char *fmt, ...) +{ + if (y < 0 || x < 0) { + return 0; + } else { + va_list v; + va_start(v, fmt); + stbiw__writefv(s, fmt, v); + va_end(v); + stbiw__write_pixels(s,rgb_dir,vdir,x,y,comp,data,alpha,pad, expand_mono); + return 1; + } +} + +static int stbi_write_bmp_core(stbi__write_context *s, int x, int y, int comp, const void *data) +{ + int pad = (-x*3) & 3; + return stbiw__outfile(s,-1,-1,x,y,comp,1,(void *) data,0,pad, + "11 4 22 4" "4 44 22 444444", + 'B', 'M', 14+40+(x*3+pad)*y, 0,0, 14+40, // file header + 40, x,y, 1,24, 0,0,0,0,0,0); // bitmap header +} + +STBIWDEF int stbi_write_bmp_to_func(stbi_write_func *func, void *context, int x, int y, int comp, const void *data) +{ + stbi__write_context s = { 0 }; + stbi__start_write_callbacks(&s, func, context); + return stbi_write_bmp_core(&s, x, y, comp, data); +} + +#ifndef STBI_WRITE_NO_STDIO +STBIWDEF int stbi_write_bmp(char const *filename, int x, int y, int comp, const void *data) +{ + stbi__write_context s = { 0 }; + if (stbi__start_write_file(&s,filename)) { + int r = stbi_write_bmp_core(&s, x, y, comp, data); + stbi__end_write_file(&s); + return r; + } else + return 0; +} +#endif //!STBI_WRITE_NO_STDIO + +static int stbi_write_tga_core(stbi__write_context *s, int x, int y, int comp, void *data) +{ + int has_alpha = (comp == 2 || comp == 4); + int colorbytes = has_alpha ? comp-1 : comp; + int format = colorbytes < 2 ? 3 : 2; // 3 color channels (RGB/RGBA) = 2, 1 color channel (Y/YA) = 3 + + if (y < 0 || x < 0) + return 0; + + if (!stbi_write_tga_with_rle) { + return stbiw__outfile(s, -1, -1, x, y, comp, 0, (void *) data, has_alpha, 0, + "111 221 2222 11", 0, 0, format, 0, 0, 0, 0, 0, x, y, (colorbytes + has_alpha) * 8, has_alpha * 8); + } else { + int i,j,k; + int jend, jdir; + + stbiw__writef(s, "111 221 2222 11", 0,0,format+8, 0,0,0, 0,0,x,y, (colorbytes + has_alpha) * 8, has_alpha * 8); + + if (stbi__flip_vertically_on_write) { + j = 0; + jend = y; + jdir = 1; + } else { + j = y-1; + jend = -1; + jdir = -1; + } + for (; j != jend; j += jdir) { + unsigned char *row = (unsigned char *) data + j * x * comp; + int len; + + for (i = 0; i < x; i += len) { + unsigned char *begin = row + i * comp; + int diff = 1; + len = 1; + + if (i < x - 1) { + ++len; + diff = memcmp(begin, row + (i + 1) * comp, comp); + if (diff) { + const unsigned char *prev = begin; + for (k = i + 2; k < x && len < 128; ++k) { + if (memcmp(prev, row + k * comp, comp)) { + prev += comp; + ++len; + } else { + --len; + break; + } + } + } else { + for (k = i + 2; k < x && len < 128; ++k) { + if (!memcmp(begin, row + k * comp, comp)) { + ++len; + } else { + break; + } + } + } + } + + if (diff) { + unsigned char header = STBIW_UCHAR(len - 1); + stbiw__write1(s, header); + for (k = 0; k < len; ++k) { + stbiw__write_pixel(s, -1, comp, has_alpha, 0, begin + k * comp); + } + } else { + unsigned char header = STBIW_UCHAR(len - 129); + stbiw__write1(s, header); + stbiw__write_pixel(s, -1, comp, has_alpha, 0, begin); + } + } + } + stbiw__write_flush(s); + } + return 1; +} + +STBIWDEF int stbi_write_tga_to_func(stbi_write_func *func, void *context, int x, int y, int comp, const void *data) +{ + stbi__write_context s = { 0 }; + stbi__start_write_callbacks(&s, func, context); + return stbi_write_tga_core(&s, x, y, comp, (void *) data); +} + +#ifndef STBI_WRITE_NO_STDIO +STBIWDEF int stbi_write_tga(char const *filename, int x, int y, int comp, const void *data) +{ + stbi__write_context s = { 0 }; + if (stbi__start_write_file(&s,filename)) { + int r = stbi_write_tga_core(&s, x, y, comp, (void *) data); + stbi__end_write_file(&s); + return r; + } else + return 0; +} +#endif + +// ************************************************************************************************* +// Radiance RGBE HDR writer +// by Baldur Karlsson + +#define stbiw__max(a, b) ((a) > (b) ? (a) : (b)) + +static void stbiw__linear_to_rgbe(unsigned char *rgbe, float *linear) +{ + int exponent; + float maxcomp = stbiw__max(linear[0], stbiw__max(linear[1], linear[2])); + + if (maxcomp < 1e-32f) { + rgbe[0] = rgbe[1] = rgbe[2] = rgbe[3] = 0; + } else { + float normalize = (float) frexp(maxcomp, &exponent) * 256.0f/maxcomp; + + rgbe[0] = (unsigned char)(linear[0] * normalize); + rgbe[1] = (unsigned char)(linear[1] * normalize); + rgbe[2] = (unsigned char)(linear[2] * normalize); + rgbe[3] = (unsigned char)(exponent + 128); + } +} + +static void stbiw__write_run_data(stbi__write_context *s, int length, unsigned char databyte) +{ + unsigned char lengthbyte = STBIW_UCHAR(length+128); + STBIW_ASSERT(length+128 <= 255); + s->func(s->context, &lengthbyte, 1); + s->func(s->context, &databyte, 1); +} + +static void stbiw__write_dump_data(stbi__write_context *s, int length, unsigned char *data) +{ + unsigned char lengthbyte = STBIW_UCHAR(length); + STBIW_ASSERT(length <= 128); // inconsistent with spec but consistent with official code + s->func(s->context, &lengthbyte, 1); + s->func(s->context, data, length); +} + +static void stbiw__write_hdr_scanline(stbi__write_context *s, int width, int ncomp, unsigned char *scratch, float *scanline) +{ + unsigned char scanlineheader[4] = { 2, 2, 0, 0 }; + unsigned char rgbe[4]; + float linear[3]; + int x; + + scanlineheader[2] = (width&0xff00)>>8; + scanlineheader[3] = (width&0x00ff); + + /* skip RLE for images too small or large */ + if (width < 8 || width >= 32768) { + for (x=0; x < width; x++) { + switch (ncomp) { + case 4: /* fallthrough */ + case 3: linear[2] = scanline[x*ncomp + 2]; + linear[1] = scanline[x*ncomp + 1]; + linear[0] = scanline[x*ncomp + 0]; + break; + default: + linear[0] = linear[1] = linear[2] = scanline[x*ncomp + 0]; + break; + } + stbiw__linear_to_rgbe(rgbe, linear); + s->func(s->context, rgbe, 4); + } + } else { + int c,r; + /* encode into scratch buffer */ + for (x=0; x < width; x++) { + switch(ncomp) { + case 4: /* fallthrough */ + case 3: linear[2] = scanline[x*ncomp + 2]; + linear[1] = scanline[x*ncomp + 1]; + linear[0] = scanline[x*ncomp + 0]; + break; + default: + linear[0] = linear[1] = linear[2] = scanline[x*ncomp + 0]; + break; + } + stbiw__linear_to_rgbe(rgbe, linear); + scratch[x + width*0] = rgbe[0]; + scratch[x + width*1] = rgbe[1]; + scratch[x + width*2] = rgbe[2]; + scratch[x + width*3] = rgbe[3]; + } + + s->func(s->context, scanlineheader, 4); + + /* RLE each component separately */ + for (c=0; c < 4; c++) { + unsigned char *comp = &scratch[width*c]; + + x = 0; + while (x < width) { + // find first run + r = x; + while (r+2 < width) { + if (comp[r] == comp[r+1] && comp[r] == comp[r+2]) + break; + ++r; + } + if (r+2 >= width) + r = width; + // dump up to first run + while (x < r) { + int len = r-x; + if (len > 128) len = 128; + stbiw__write_dump_data(s, len, &comp[x]); + x += len; + } + // if there's a run, output it + if (r+2 < width) { // same test as what we break out of in search loop, so only true if we break'd + // find next byte after run + while (r < width && comp[r] == comp[x]) + ++r; + // output run up to r + while (x < r) { + int len = r-x; + if (len > 127) len = 127; + stbiw__write_run_data(s, len, comp[x]); + x += len; + } + } + } + } + } +} + +static int stbi_write_hdr_core(stbi__write_context *s, int x, int y, int comp, float *data) +{ + if (y <= 0 || x <= 0 || data == NULL) + return 0; + else { + // Each component is stored separately. Allocate scratch space for full output scanline. + unsigned char *scratch = (unsigned char *) STBIW_MALLOC(x*4); + int i, len; + char buffer[128]; + char header[] = "#?RADIANCE\n# Written by stb_image_write.h\nFORMAT=32-bit_rle_rgbe\n"; + s->func(s->context, header, sizeof(header)-1); + +#ifdef __STDC_WANT_SECURE_LIB__ + len = sprintf_s(buffer, sizeof(buffer), "EXPOSURE= 1.0000000000000\n\n-Y %d +X %d\n", y, x); +#else + len = sprintf(buffer, "EXPOSURE= 1.0000000000000\n\n-Y %d +X %d\n", y, x); +#endif + s->func(s->context, buffer, len); + + for(i=0; i < y; i++) + stbiw__write_hdr_scanline(s, x, comp, scratch, data + comp*x*(stbi__flip_vertically_on_write ? y-1-i : i)); + STBIW_FREE(scratch); + return 1; + } +} + +STBIWDEF int stbi_write_hdr_to_func(stbi_write_func *func, void *context, int x, int y, int comp, const float *data) +{ + stbi__write_context s = { 0 }; + stbi__start_write_callbacks(&s, func, context); + return stbi_write_hdr_core(&s, x, y, comp, (float *) data); +} + +#ifndef STBI_WRITE_NO_STDIO +STBIWDEF int stbi_write_hdr(char const *filename, int x, int y, int comp, const float *data) +{ + stbi__write_context s = { 0 }; + if (stbi__start_write_file(&s,filename)) { + int r = stbi_write_hdr_core(&s, x, y, comp, (float *) data); + stbi__end_write_file(&s); + return r; + } else + return 0; +} +#endif // STBI_WRITE_NO_STDIO + + +////////////////////////////////////////////////////////////////////////////// +// +// PNG writer +// + +#ifndef STBIW_ZLIB_COMPRESS +// stretchy buffer; stbiw__sbpush() == vector<>::push_back() -- stbiw__sbcount() == vector<>::size() +#define stbiw__sbraw(a) ((int *) (void *) (a) - 2) +#define stbiw__sbm(a) stbiw__sbraw(a)[0] +#define stbiw__sbn(a) stbiw__sbraw(a)[1] + +#define stbiw__sbneedgrow(a,n) ((a)==0 || stbiw__sbn(a)+n >= stbiw__sbm(a)) +#define stbiw__sbmaybegrow(a,n) (stbiw__sbneedgrow(a,(n)) ? stbiw__sbgrow(a,n) : 0) +#define stbiw__sbgrow(a,n) stbiw__sbgrowf((void **) &(a), (n), sizeof(*(a))) + +#define stbiw__sbpush(a, v) (stbiw__sbmaybegrow(a,1), (a)[stbiw__sbn(a)++] = (v)) +#define stbiw__sbcount(a) ((a) ? stbiw__sbn(a) : 0) +#define stbiw__sbfree(a) ((a) ? STBIW_FREE(stbiw__sbraw(a)),0 : 0) + +static void *stbiw__sbgrowf(void **arr, int increment, int itemsize) +{ + int m = *arr ? 2*stbiw__sbm(*arr)+increment : increment+1; + void *p = STBIW_REALLOC_SIZED(*arr ? stbiw__sbraw(*arr) : 0, *arr ? (stbiw__sbm(*arr)*itemsize + sizeof(int)*2) : 0, itemsize * m + sizeof(int)*2); + STBIW_ASSERT(p); + if (p) { + if (!*arr) ((int *) p)[1] = 0; + *arr = (void *) ((int *) p + 2); + stbiw__sbm(*arr) = m; + } + return *arr; +} + +static unsigned char *stbiw__zlib_flushf(unsigned char *data, unsigned int *bitbuffer, int *bitcount) +{ + while (*bitcount >= 8) { + stbiw__sbpush(data, STBIW_UCHAR(*bitbuffer)); + *bitbuffer >>= 8; + *bitcount -= 8; + } + return data; +} + +static int stbiw__zlib_bitrev(int code, int codebits) +{ + int res=0; + while (codebits--) { + res = (res << 1) | (code & 1); + code >>= 1; + } + return res; +} + +static unsigned int stbiw__zlib_countm(unsigned char *a, unsigned char *b, int limit) +{ + int i; + for (i=0; i < limit && i < 258; ++i) + if (a[i] != b[i]) break; + return i; +} + +static unsigned int stbiw__zhash(unsigned char *data) +{ + stbiw_uint32 hash = data[0] + (data[1] << 8) + (data[2] << 16); + hash ^= hash << 3; + hash += hash >> 5; + hash ^= hash << 4; + hash += hash >> 17; + hash ^= hash << 25; + hash += hash >> 6; + return hash; +} + +#define stbiw__zlib_flush() (out = stbiw__zlib_flushf(out, &bitbuf, &bitcount)) +#define stbiw__zlib_add(code,codebits) \ + (bitbuf |= (code) << bitcount, bitcount += (codebits), stbiw__zlib_flush()) +#define stbiw__zlib_huffa(b,c) stbiw__zlib_add(stbiw__zlib_bitrev(b,c),c) +// default huffman tables +#define stbiw__zlib_huff1(n) stbiw__zlib_huffa(0x30 + (n), 8) +#define stbiw__zlib_huff2(n) stbiw__zlib_huffa(0x190 + (n)-144, 9) +#define stbiw__zlib_huff3(n) stbiw__zlib_huffa(0 + (n)-256,7) +#define stbiw__zlib_huff4(n) stbiw__zlib_huffa(0xc0 + (n)-280,8) +#define stbiw__zlib_huff(n) ((n) <= 143 ? stbiw__zlib_huff1(n) : (n) <= 255 ? stbiw__zlib_huff2(n) : (n) <= 279 ? stbiw__zlib_huff3(n) : stbiw__zlib_huff4(n)) +#define stbiw__zlib_huffb(n) ((n) <= 143 ? stbiw__zlib_huff1(n) : stbiw__zlib_huff2(n)) + +#define stbiw__ZHASH 16384 + +#endif // STBIW_ZLIB_COMPRESS + +STBIWDEF unsigned char * stbi_zlib_compress(unsigned char *data, int data_len, int *out_len, int quality) +{ +#ifdef STBIW_ZLIB_COMPRESS + // user provided a zlib compress implementation, use that + return STBIW_ZLIB_COMPRESS(data, data_len, out_len, quality); +#else // use builtin + static unsigned short lengthc[] = { 3,4,5,6,7,8,9,10,11,13,15,17,19,23,27,31,35,43,51,59,67,83,99,115,131,163,195,227,258, 259 }; + static unsigned char lengtheb[]= { 0,0,0,0,0,0,0, 0, 1, 1, 1, 1, 2, 2, 2, 2, 3, 3, 3, 3, 4, 4, 4, 4, 5, 5, 5, 5, 0 }; + static unsigned short distc[] = { 1,2,3,4,5,7,9,13,17,25,33,49,65,97,129,193,257,385,513,769,1025,1537,2049,3073,4097,6145,8193,12289,16385,24577, 32768 }; + static unsigned char disteb[] = { 0,0,0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7,8,8,9,9,10,10,11,11,12,12,13,13 }; + unsigned int bitbuf=0; + int i,j, bitcount=0; + unsigned char *out = NULL; + unsigned char ***hash_table = (unsigned char***) STBIW_MALLOC(stbiw__ZHASH * sizeof(unsigned char**)); + if (hash_table == NULL) + return NULL; + if (quality < 5) quality = 5; + + stbiw__sbpush(out, 0x78); // DEFLATE 32K window + stbiw__sbpush(out, 0x5e); // FLEVEL = 1 + stbiw__zlib_add(1,1); // BFINAL = 1 + stbiw__zlib_add(1,2); // BTYPE = 1 -- fixed huffman + + for (i=0; i < stbiw__ZHASH; ++i) + hash_table[i] = NULL; + + i=0; + while (i < data_len-3) { + // hash next 3 bytes of data to be compressed + int h = stbiw__zhash(data+i)&(stbiw__ZHASH-1), best=3; + unsigned char *bestloc = 0; + unsigned char **hlist = hash_table[h]; + int n = stbiw__sbcount(hlist); + for (j=0; j < n; ++j) { + if (hlist[j]-data > i-32768) { // if entry lies within window + int d = stbiw__zlib_countm(hlist[j], data+i, data_len-i); + if (d >= best) { best=d; bestloc=hlist[j]; } + } + } + // when hash table entry is too long, delete half the entries + if (hash_table[h] && stbiw__sbn(hash_table[h]) == 2*quality) { + STBIW_MEMMOVE(hash_table[h], hash_table[h]+quality, sizeof(hash_table[h][0])*quality); + stbiw__sbn(hash_table[h]) = quality; + } + stbiw__sbpush(hash_table[h],data+i); + + if (bestloc) { + // "lazy matching" - check match at *next* byte, and if it's better, do cur byte as literal + h = stbiw__zhash(data+i+1)&(stbiw__ZHASH-1); + hlist = hash_table[h]; + n = stbiw__sbcount(hlist); + for (j=0; j < n; ++j) { + if (hlist[j]-data > i-32767) { + int e = stbiw__zlib_countm(hlist[j], data+i+1, data_len-i-1); + if (e > best) { // if next match is better, bail on current match + bestloc = NULL; + break; + } + } + } + } + + if (bestloc) { + int d = (int) (data+i - bestloc); // distance back + STBIW_ASSERT(d <= 32767 && best <= 258); + for (j=0; best > lengthc[j+1]-1; ++j); + stbiw__zlib_huff(j+257); + if (lengtheb[j]) stbiw__zlib_add(best - lengthc[j], lengtheb[j]); + for (j=0; d > distc[j+1]-1; ++j); + stbiw__zlib_add(stbiw__zlib_bitrev(j,5),5); + if (disteb[j]) stbiw__zlib_add(d - distc[j], disteb[j]); + i += best; + } else { + stbiw__zlib_huffb(data[i]); + ++i; + } + } + // write out final bytes + for (;i < data_len; ++i) + stbiw__zlib_huffb(data[i]); + stbiw__zlib_huff(256); // end of block + // pad with 0 bits to byte boundary + while (bitcount) + stbiw__zlib_add(0,1); + + for (i=0; i < stbiw__ZHASH; ++i) + (void) stbiw__sbfree(hash_table[i]); + STBIW_FREE(hash_table); + + { + // compute adler32 on input + unsigned int s1=1, s2=0; + int blocklen = (int) (data_len % 5552); + j=0; + while (j < data_len) { + for (i=0; i < blocklen; ++i) { s1 += data[j+i]; s2 += s1; } + s1 %= 65521; s2 %= 65521; + j += blocklen; + blocklen = 5552; + } + stbiw__sbpush(out, STBIW_UCHAR(s2 >> 8)); + stbiw__sbpush(out, STBIW_UCHAR(s2)); + stbiw__sbpush(out, STBIW_UCHAR(s1 >> 8)); + stbiw__sbpush(out, STBIW_UCHAR(s1)); + } + *out_len = stbiw__sbn(out); + // make returned pointer freeable + STBIW_MEMMOVE(stbiw__sbraw(out), out, *out_len); + return (unsigned char *) stbiw__sbraw(out); +#endif // STBIW_ZLIB_COMPRESS +} + +static unsigned int stbiw__crc32(unsigned char *buffer, int len) +{ +#ifdef STBIW_CRC32 + return STBIW_CRC32(buffer, len); +#else + static unsigned int crc_table[256] = + { + 0x00000000, 0x77073096, 0xEE0E612C, 0x990951BA, 0x076DC419, 0x706AF48F, 0xE963A535, 0x9E6495A3, + 0x0eDB8832, 0x79DCB8A4, 0xE0D5E91E, 0x97D2D988, 0x09B64C2B, 0x7EB17CBD, 0xE7B82D07, 0x90BF1D91, + 0x1DB71064, 0x6AB020F2, 0xF3B97148, 0x84BE41DE, 0x1ADAD47D, 0x6DDDE4EB, 0xF4D4B551, 0x83D385C7, + 0x136C9856, 0x646BA8C0, 0xFD62F97A, 0x8A65C9EC, 0x14015C4F, 0x63066CD9, 0xFA0F3D63, 0x8D080DF5, + 0x3B6E20C8, 0x4C69105E, 0xD56041E4, 0xA2677172, 0x3C03E4D1, 0x4B04D447, 0xD20D85FD, 0xA50AB56B, + 0x35B5A8FA, 0x42B2986C, 0xDBBBC9D6, 0xACBCF940, 0x32D86CE3, 0x45DF5C75, 0xDCD60DCF, 0xABD13D59, + 0x26D930AC, 0x51DE003A, 0xC8D75180, 0xBFD06116, 0x21B4F4B5, 0x56B3C423, 0xCFBA9599, 0xB8BDA50F, + 0x2802B89E, 0x5F058808, 0xC60CD9B2, 0xB10BE924, 0x2F6F7C87, 0x58684C11, 0xC1611DAB, 0xB6662D3D, + 0x76DC4190, 0x01DB7106, 0x98D220BC, 0xEFD5102A, 0x71B18589, 0x06B6B51F, 0x9FBFE4A5, 0xE8B8D433, + 0x7807C9A2, 0x0F00F934, 0x9609A88E, 0xE10E9818, 0x7F6A0DBB, 0x086D3D2D, 0x91646C97, 0xE6635C01, + 0x6B6B51F4, 0x1C6C6162, 0x856530D8, 0xF262004E, 0x6C0695ED, 0x1B01A57B, 0x8208F4C1, 0xF50FC457, + 0x65B0D9C6, 0x12B7E950, 0x8BBEB8EA, 0xFCB9887C, 0x62DD1DDF, 0x15DA2D49, 0x8CD37CF3, 0xFBD44C65, + 0x4DB26158, 0x3AB551CE, 0xA3BC0074, 0xD4BB30E2, 0x4ADFA541, 0x3DD895D7, 0xA4D1C46D, 0xD3D6F4FB, + 0x4369E96A, 0x346ED9FC, 0xAD678846, 0xDA60B8D0, 0x44042D73, 0x33031DE5, 0xAA0A4C5F, 0xDD0D7CC9, + 0x5005713C, 0x270241AA, 0xBE0B1010, 0xC90C2086, 0x5768B525, 0x206F85B3, 0xB966D409, 0xCE61E49F, + 0x5EDEF90E, 0x29D9C998, 0xB0D09822, 0xC7D7A8B4, 0x59B33D17, 0x2EB40D81, 0xB7BD5C3B, 0xC0BA6CAD, + 0xEDB88320, 0x9ABFB3B6, 0x03B6E20C, 0x74B1D29A, 0xEAD54739, 0x9DD277AF, 0x04DB2615, 0x73DC1683, + 0xE3630B12, 0x94643B84, 0x0D6D6A3E, 0x7A6A5AA8, 0xE40ECF0B, 0x9309FF9D, 0x0A00AE27, 0x7D079EB1, + 0xF00F9344, 0x8708A3D2, 0x1E01F268, 0x6906C2FE, 0xF762575D, 0x806567CB, 0x196C3671, 0x6E6B06E7, + 0xFED41B76, 0x89D32BE0, 0x10DA7A5A, 0x67DD4ACC, 0xF9B9DF6F, 0x8EBEEFF9, 0x17B7BE43, 0x60B08ED5, + 0xD6D6A3E8, 0xA1D1937E, 0x38D8C2C4, 0x4FDFF252, 0xD1BB67F1, 0xA6BC5767, 0x3FB506DD, 0x48B2364B, + 0xD80D2BDA, 0xAF0A1B4C, 0x36034AF6, 0x41047A60, 0xDF60EFC3, 0xA867DF55, 0x316E8EEF, 0x4669BE79, + 0xCB61B38C, 0xBC66831A, 0x256FD2A0, 0x5268E236, 0xCC0C7795, 0xBB0B4703, 0x220216B9, 0x5505262F, + 0xC5BA3BBE, 0xB2BD0B28, 0x2BB45A92, 0x5CB36A04, 0xC2D7FFA7, 0xB5D0CF31, 0x2CD99E8B, 0x5BDEAE1D, + 0x9B64C2B0, 0xEC63F226, 0x756AA39C, 0x026D930A, 0x9C0906A9, 0xEB0E363F, 0x72076785, 0x05005713, + 0x95BF4A82, 0xE2B87A14, 0x7BB12BAE, 0x0CB61B38, 0x92D28E9B, 0xE5D5BE0D, 0x7CDCEFB7, 0x0BDBDF21, + 0x86D3D2D4, 0xF1D4E242, 0x68DDB3F8, 0x1FDA836E, 0x81BE16CD, 0xF6B9265B, 0x6FB077E1, 0x18B74777, + 0x88085AE6, 0xFF0F6A70, 0x66063BCA, 0x11010B5C, 0x8F659EFF, 0xF862AE69, 0x616BFFD3, 0x166CCF45, + 0xA00AE278, 0xD70DD2EE, 0x4E048354, 0x3903B3C2, 0xA7672661, 0xD06016F7, 0x4969474D, 0x3E6E77DB, + 0xAED16A4A, 0xD9D65ADC, 0x40DF0B66, 0x37D83BF0, 0xA9BCAE53, 0xDEBB9EC5, 0x47B2CF7F, 0x30B5FFE9, + 0xBDBDF21C, 0xCABAC28A, 0x53B39330, 0x24B4A3A6, 0xBAD03605, 0xCDD70693, 0x54DE5729, 0x23D967BF, + 0xB3667A2E, 0xC4614AB8, 0x5D681B02, 0x2A6F2B94, 0xB40BBE37, 0xC30C8EA1, 0x5A05DF1B, 0x2D02EF8D + }; + + unsigned int crc = ~0u; + int i; + for (i=0; i < len; ++i) + crc = (crc >> 8) ^ crc_table[buffer[i] ^ (crc & 0xff)]; + return ~crc; +#endif +} + +#define stbiw__wpng4(o,a,b,c,d) ((o)[0]=STBIW_UCHAR(a),(o)[1]=STBIW_UCHAR(b),(o)[2]=STBIW_UCHAR(c),(o)[3]=STBIW_UCHAR(d),(o)+=4) +#define stbiw__wp32(data,v) stbiw__wpng4(data, (v)>>24,(v)>>16,(v)>>8,(v)); +#define stbiw__wptag(data,s) stbiw__wpng4(data, s[0],s[1],s[2],s[3]) + +static void stbiw__wpcrc(unsigned char **data, int len) +{ + unsigned int crc = stbiw__crc32(*data - len - 4, len+4); + stbiw__wp32(*data, crc); +} + +static unsigned char stbiw__paeth(int a, int b, int c) +{ + int p = a + b - c, pa = abs(p-a), pb = abs(p-b), pc = abs(p-c); + if (pa <= pb && pa <= pc) return STBIW_UCHAR(a); + if (pb <= pc) return STBIW_UCHAR(b); + return STBIW_UCHAR(c); +} + +// @OPTIMIZE: provide an option that always forces left-predict or paeth predict +static void stbiw__encode_png_line(unsigned char *pixels, int stride_bytes, int width, int height, int y, int n, int filter_type, signed char *line_buffer) +{ + static int mapping[] = { 0,1,2,3,4 }; + static int firstmap[] = { 0,1,0,5,6 }; + int *mymap = (y != 0) ? mapping : firstmap; + int i; + int type = mymap[filter_type]; + unsigned char *z = pixels + stride_bytes * (stbi__flip_vertically_on_write ? height-1-y : y); + int signed_stride = stbi__flip_vertically_on_write ? -stride_bytes : stride_bytes; + + if (type==0) { + memcpy(line_buffer, z, width*n); + return; + } + + // first loop isn't optimized since it's just one pixel + for (i = 0; i < n; ++i) { + switch (type) { + case 1: line_buffer[i] = z[i]; break; + case 2: line_buffer[i] = z[i] - z[i-signed_stride]; break; + case 3: line_buffer[i] = z[i] - (z[i-signed_stride]>>1); break; + case 4: line_buffer[i] = (signed char) (z[i] - stbiw__paeth(0,z[i-signed_stride],0)); break; + case 5: line_buffer[i] = z[i]; break; + case 6: line_buffer[i] = z[i]; break; + } + } + switch (type) { + case 1: for (i=n; i < width*n; ++i) line_buffer[i] = z[i] - z[i-n]; break; + case 2: for (i=n; i < width*n; ++i) line_buffer[i] = z[i] - z[i-signed_stride]; break; + case 3: for (i=n; i < width*n; ++i) line_buffer[i] = z[i] - ((z[i-n] + z[i-signed_stride])>>1); break; + case 4: for (i=n; i < width*n; ++i) line_buffer[i] = z[i] - stbiw__paeth(z[i-n], z[i-signed_stride], z[i-signed_stride-n]); break; + case 5: for (i=n; i < width*n; ++i) line_buffer[i] = z[i] - (z[i-n]>>1); break; + case 6: for (i=n; i < width*n; ++i) line_buffer[i] = z[i] - stbiw__paeth(z[i-n], 0,0); break; + } +} + +STBIWDEF unsigned char *stbi_write_png_to_mem(const unsigned char *pixels, int stride_bytes, int x, int y, int n, int *out_len) +{ + int force_filter = stbi_write_force_png_filter; + int ctype[5] = { -1, 0, 4, 2, 6 }; + unsigned char sig[8] = { 137,80,78,71,13,10,26,10 }; + unsigned char *out,*o, *filt, *zlib; + signed char *line_buffer; + int j,zlen; + + if (stride_bytes == 0) + stride_bytes = x * n; + + if (force_filter >= 5) { + force_filter = -1; + } + + filt = (unsigned char *) STBIW_MALLOC((x*n+1) * y); if (!filt) return 0; + line_buffer = (signed char *) STBIW_MALLOC(x * n); if (!line_buffer) { STBIW_FREE(filt); return 0; } + for (j=0; j < y; ++j) { + int filter_type; + if (force_filter > -1) { + filter_type = force_filter; + stbiw__encode_png_line((unsigned char*)(pixels), stride_bytes, x, y, j, n, force_filter, line_buffer); + } else { // Estimate the best filter by running through all of them: + int best_filter = 0, best_filter_val = 0x7fffffff, est, i; + for (filter_type = 0; filter_type < 5; filter_type++) { + stbiw__encode_png_line((unsigned char*)(pixels), stride_bytes, x, y, j, n, filter_type, line_buffer); + + // Estimate the entropy of the line using this filter; the less, the better. + est = 0; + for (i = 0; i < x*n; ++i) { + est += abs((signed char) line_buffer[i]); + } + if (est < best_filter_val) { + best_filter_val = est; + best_filter = filter_type; + } + } + if (filter_type != best_filter) { // If the last iteration already got us the best filter, don't redo it + stbiw__encode_png_line((unsigned char*)(pixels), stride_bytes, x, y, j, n, best_filter, line_buffer); + filter_type = best_filter; + } + } + // when we get here, filter_type contains the filter type, and line_buffer contains the data + filt[j*(x*n+1)] = (unsigned char) filter_type; + STBIW_MEMMOVE(filt+j*(x*n+1)+1, line_buffer, x*n); + } + STBIW_FREE(line_buffer); + zlib = stbi_zlib_compress(filt, y*( x*n+1), &zlen, stbi_write_png_compression_level); + STBIW_FREE(filt); + if (!zlib) return 0; + + // each tag requires 12 bytes of overhead + out = (unsigned char *) STBIW_MALLOC(8 + 12+13 + 12+zlen + 12); + if (!out) return 0; + *out_len = 8 + 12+13 + 12+zlen + 12; + + o=out; + STBIW_MEMMOVE(o,sig,8); o+= 8; + stbiw__wp32(o, 13); // header length + stbiw__wptag(o, "IHDR"); + stbiw__wp32(o, x); + stbiw__wp32(o, y); + *o++ = 8; + *o++ = STBIW_UCHAR(ctype[n]); + *o++ = 0; + *o++ = 0; + *o++ = 0; + stbiw__wpcrc(&o,13); + + stbiw__wp32(o, zlen); + stbiw__wptag(o, "IDAT"); + STBIW_MEMMOVE(o, zlib, zlen); + o += zlen; + STBIW_FREE(zlib); + stbiw__wpcrc(&o, zlen); + + stbiw__wp32(o,0); + stbiw__wptag(o, "IEND"); + stbiw__wpcrc(&o,0); + + STBIW_ASSERT(o == out + *out_len); + + return out; +} + +#ifndef STBI_WRITE_NO_STDIO +STBIWDEF int stbi_write_png(char const *filename, int x, int y, int comp, const void *data, int stride_bytes) +{ + FILE *f; + int len; + unsigned char *png = stbi_write_png_to_mem((const unsigned char *) data, stride_bytes, x, y, comp, &len); + if (png == NULL) return 0; + + f = stbiw__fopen(filename, "wb"); + if (!f) { STBIW_FREE(png); return 0; } + fwrite(png, 1, len, f); + fclose(f); + STBIW_FREE(png); + return 1; +} +#endif + +STBIWDEF int stbi_write_png_to_func(stbi_write_func *func, void *context, int x, int y, int comp, const void *data, int stride_bytes) +{ + int len; + unsigned char *png = stbi_write_png_to_mem((const unsigned char *) data, stride_bytes, x, y, comp, &len); + if (png == NULL) return 0; + func(context, png, len); + STBIW_FREE(png); + return 1; +} + + +/* *************************************************************************** + * + * JPEG writer + * + * This is based on Jon Olick's jo_jpeg.cpp: + * public domain Simple, Minimalistic JPEG writer - http://www.jonolick.com/code.html + */ + +static const unsigned char stbiw__jpg_ZigZag[] = { 0,1,5,6,14,15,27,28,2,4,7,13,16,26,29,42,3,8,12,17,25,30,41,43,9,11,18, + 24,31,40,44,53,10,19,23,32,39,45,52,54,20,22,33,38,46,51,55,60,21,34,37,47,50,56,59,61,35,36,48,49,57,58,62,63 }; + +static void stbiw__jpg_writeBits(stbi__write_context *s, int *bitBufP, int *bitCntP, const unsigned short *bs) { + int bitBuf = *bitBufP, bitCnt = *bitCntP; + bitCnt += bs[1]; + bitBuf |= bs[0] << (24 - bitCnt); + while(bitCnt >= 8) { + unsigned char c = (bitBuf >> 16) & 255; + stbiw__putc(s, c); + if(c == 255) { + stbiw__putc(s, 0); + } + bitBuf <<= 8; + bitCnt -= 8; + } + *bitBufP = bitBuf; + *bitCntP = bitCnt; +} + +static void stbiw__jpg_DCT(float *d0p, float *d1p, float *d2p, float *d3p, float *d4p, float *d5p, float *d6p, float *d7p) { + float d0 = *d0p, d1 = *d1p, d2 = *d2p, d3 = *d3p, d4 = *d4p, d5 = *d5p, d6 = *d6p, d7 = *d7p; + float z1, z2, z3, z4, z5, z11, z13; + + float tmp0 = d0 + d7; + float tmp7 = d0 - d7; + float tmp1 = d1 + d6; + float tmp6 = d1 - d6; + float tmp2 = d2 + d5; + float tmp5 = d2 - d5; + float tmp3 = d3 + d4; + float tmp4 = d3 - d4; + + // Even part + float tmp10 = tmp0 + tmp3; // phase 2 + float tmp13 = tmp0 - tmp3; + float tmp11 = tmp1 + tmp2; + float tmp12 = tmp1 - tmp2; + + d0 = tmp10 + tmp11; // phase 3 + d4 = tmp10 - tmp11; + + z1 = (tmp12 + tmp13) * 0.707106781f; // c4 + d2 = tmp13 + z1; // phase 5 + d6 = tmp13 - z1; + + // Odd part + tmp10 = tmp4 + tmp5; // phase 2 + tmp11 = tmp5 + tmp6; + tmp12 = tmp6 + tmp7; + + // The rotator is modified from fig 4-8 to avoid extra negations. + z5 = (tmp10 - tmp12) * 0.382683433f; // c6 + z2 = tmp10 * 0.541196100f + z5; // c2-c6 + z4 = tmp12 * 1.306562965f + z5; // c2+c6 + z3 = tmp11 * 0.707106781f; // c4 + + z11 = tmp7 + z3; // phase 5 + z13 = tmp7 - z3; + + *d5p = z13 + z2; // phase 6 + *d3p = z13 - z2; + *d1p = z11 + z4; + *d7p = z11 - z4; + + *d0p = d0; *d2p = d2; *d4p = d4; *d6p = d6; +} + +static void stbiw__jpg_calcBits(int val, unsigned short bits[2]) { + int tmp1 = val < 0 ? -val : val; + val = val < 0 ? val-1 : val; + bits[1] = 1; + while(tmp1 >>= 1) { + ++bits[1]; + } + bits[0] = val & ((1<0)&&(DU[end0pos]==0); --end0pos) { + } + // end0pos = first element in reverse order !=0 + if(end0pos == 0) { + stbiw__jpg_writeBits(s, bitBuf, bitCnt, EOB); + return DU[0]; + } + for(i = 1; i <= end0pos; ++i) { + int startpos = i; + int nrzeroes; + unsigned short bits[2]; + for (; DU[i]==0 && i<=end0pos; ++i) { + } + nrzeroes = i-startpos; + if ( nrzeroes >= 16 ) { + int lng = nrzeroes>>4; + int nrmarker; + for (nrmarker=1; nrmarker <= lng; ++nrmarker) + stbiw__jpg_writeBits(s, bitBuf, bitCnt, M16zeroes); + nrzeroes &= 15; + } + stbiw__jpg_calcBits(DU[i], bits); + stbiw__jpg_writeBits(s, bitBuf, bitCnt, HTAC[(nrzeroes<<4)+bits[1]]); + stbiw__jpg_writeBits(s, bitBuf, bitCnt, bits); + } + if(end0pos != 63) { + stbiw__jpg_writeBits(s, bitBuf, bitCnt, EOB); + } + return DU[0]; +} + +static int stbi_write_jpg_core(stbi__write_context *s, int width, int height, int comp, const void* data, int quality) { + // Constants that don't pollute global namespace + static const unsigned char std_dc_luminance_nrcodes[] = {0,0,1,5,1,1,1,1,1,1,0,0,0,0,0,0,0}; + static const unsigned char std_dc_luminance_values[] = {0,1,2,3,4,5,6,7,8,9,10,11}; + static const unsigned char std_ac_luminance_nrcodes[] = {0,0,2,1,3,3,2,4,3,5,5,4,4,0,0,1,0x7d}; + static const unsigned char std_ac_luminance_values[] = { + 0x01,0x02,0x03,0x00,0x04,0x11,0x05,0x12,0x21,0x31,0x41,0x06,0x13,0x51,0x61,0x07,0x22,0x71,0x14,0x32,0x81,0x91,0xa1,0x08, + 0x23,0x42,0xb1,0xc1,0x15,0x52,0xd1,0xf0,0x24,0x33,0x62,0x72,0x82,0x09,0x0a,0x16,0x17,0x18,0x19,0x1a,0x25,0x26,0x27,0x28, + 0x29,0x2a,0x34,0x35,0x36,0x37,0x38,0x39,0x3a,0x43,0x44,0x45,0x46,0x47,0x48,0x49,0x4a,0x53,0x54,0x55,0x56,0x57,0x58,0x59, + 0x5a,0x63,0x64,0x65,0x66,0x67,0x68,0x69,0x6a,0x73,0x74,0x75,0x76,0x77,0x78,0x79,0x7a,0x83,0x84,0x85,0x86,0x87,0x88,0x89, + 0x8a,0x92,0x93,0x94,0x95,0x96,0x97,0x98,0x99,0x9a,0xa2,0xa3,0xa4,0xa5,0xa6,0xa7,0xa8,0xa9,0xaa,0xb2,0xb3,0xb4,0xb5,0xb6, + 0xb7,0xb8,0xb9,0xba,0xc2,0xc3,0xc4,0xc5,0xc6,0xc7,0xc8,0xc9,0xca,0xd2,0xd3,0xd4,0xd5,0xd6,0xd7,0xd8,0xd9,0xda,0xe1,0xe2, + 0xe3,0xe4,0xe5,0xe6,0xe7,0xe8,0xe9,0xea,0xf1,0xf2,0xf3,0xf4,0xf5,0xf6,0xf7,0xf8,0xf9,0xfa + }; + static const unsigned char std_dc_chrominance_nrcodes[] = {0,0,3,1,1,1,1,1,1,1,1,1,0,0,0,0,0}; + static const unsigned char std_dc_chrominance_values[] = {0,1,2,3,4,5,6,7,8,9,10,11}; + static const unsigned char std_ac_chrominance_nrcodes[] = {0,0,2,1,2,4,4,3,4,7,5,4,4,0,1,2,0x77}; + static const unsigned char std_ac_chrominance_values[] = { + 0x00,0x01,0x02,0x03,0x11,0x04,0x05,0x21,0x31,0x06,0x12,0x41,0x51,0x07,0x61,0x71,0x13,0x22,0x32,0x81,0x08,0x14,0x42,0x91, + 0xa1,0xb1,0xc1,0x09,0x23,0x33,0x52,0xf0,0x15,0x62,0x72,0xd1,0x0a,0x16,0x24,0x34,0xe1,0x25,0xf1,0x17,0x18,0x19,0x1a,0x26, + 0x27,0x28,0x29,0x2a,0x35,0x36,0x37,0x38,0x39,0x3a,0x43,0x44,0x45,0x46,0x47,0x48,0x49,0x4a,0x53,0x54,0x55,0x56,0x57,0x58, + 0x59,0x5a,0x63,0x64,0x65,0x66,0x67,0x68,0x69,0x6a,0x73,0x74,0x75,0x76,0x77,0x78,0x79,0x7a,0x82,0x83,0x84,0x85,0x86,0x87, + 0x88,0x89,0x8a,0x92,0x93,0x94,0x95,0x96,0x97,0x98,0x99,0x9a,0xa2,0xa3,0xa4,0xa5,0xa6,0xa7,0xa8,0xa9,0xaa,0xb2,0xb3,0xb4, + 0xb5,0xb6,0xb7,0xb8,0xb9,0xba,0xc2,0xc3,0xc4,0xc5,0xc6,0xc7,0xc8,0xc9,0xca,0xd2,0xd3,0xd4,0xd5,0xd6,0xd7,0xd8,0xd9,0xda, + 0xe2,0xe3,0xe4,0xe5,0xe6,0xe7,0xe8,0xe9,0xea,0xf2,0xf3,0xf4,0xf5,0xf6,0xf7,0xf8,0xf9,0xfa + }; + // Huffman tables + static const unsigned short YDC_HT[256][2] = { {0,2},{2,3},{3,3},{4,3},{5,3},{6,3},{14,4},{30,5},{62,6},{126,7},{254,8},{510,9}}; + static const unsigned short UVDC_HT[256][2] = { {0,2},{1,2},{2,2},{6,3},{14,4},{30,5},{62,6},{126,7},{254,8},{510,9},{1022,10},{2046,11}}; + static const unsigned short YAC_HT[256][2] = { + {10,4},{0,2},{1,2},{4,3},{11,4},{26,5},{120,7},{248,8},{1014,10},{65410,16},{65411,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0}, + {12,4},{27,5},{121,7},{502,9},{2038,11},{65412,16},{65413,16},{65414,16},{65415,16},{65416,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0}, + {28,5},{249,8},{1015,10},{4084,12},{65417,16},{65418,16},{65419,16},{65420,16},{65421,16},{65422,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0}, + {58,6},{503,9},{4085,12},{65423,16},{65424,16},{65425,16},{65426,16},{65427,16},{65428,16},{65429,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0}, + {59,6},{1016,10},{65430,16},{65431,16},{65432,16},{65433,16},{65434,16},{65435,16},{65436,16},{65437,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0}, + {122,7},{2039,11},{65438,16},{65439,16},{65440,16},{65441,16},{65442,16},{65443,16},{65444,16},{65445,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0}, + {123,7},{4086,12},{65446,16},{65447,16},{65448,16},{65449,16},{65450,16},{65451,16},{65452,16},{65453,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0}, + {250,8},{4087,12},{65454,16},{65455,16},{65456,16},{65457,16},{65458,16},{65459,16},{65460,16},{65461,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0}, + {504,9},{32704,15},{65462,16},{65463,16},{65464,16},{65465,16},{65466,16},{65467,16},{65468,16},{65469,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0}, + {505,9},{65470,16},{65471,16},{65472,16},{65473,16},{65474,16},{65475,16},{65476,16},{65477,16},{65478,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0}, + {506,9},{65479,16},{65480,16},{65481,16},{65482,16},{65483,16},{65484,16},{65485,16},{65486,16},{65487,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0}, + {1017,10},{65488,16},{65489,16},{65490,16},{65491,16},{65492,16},{65493,16},{65494,16},{65495,16},{65496,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0}, + {1018,10},{65497,16},{65498,16},{65499,16},{65500,16},{65501,16},{65502,16},{65503,16},{65504,16},{65505,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0}, + {2040,11},{65506,16},{65507,16},{65508,16},{65509,16},{65510,16},{65511,16},{65512,16},{65513,16},{65514,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0}, + {65515,16},{65516,16},{65517,16},{65518,16},{65519,16},{65520,16},{65521,16},{65522,16},{65523,16},{65524,16},{0,0},{0,0},{0,0},{0,0},{0,0}, + {2041,11},{65525,16},{65526,16},{65527,16},{65528,16},{65529,16},{65530,16},{65531,16},{65532,16},{65533,16},{65534,16},{0,0},{0,0},{0,0},{0,0},{0,0} + }; + static const unsigned short UVAC_HT[256][2] = { + {0,2},{1,2},{4,3},{10,4},{24,5},{25,5},{56,6},{120,7},{500,9},{1014,10},{4084,12},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0}, + {11,4},{57,6},{246,8},{501,9},{2038,11},{4085,12},{65416,16},{65417,16},{65418,16},{65419,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0}, + {26,5},{247,8},{1015,10},{4086,12},{32706,15},{65420,16},{65421,16},{65422,16},{65423,16},{65424,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0}, + {27,5},{248,8},{1016,10},{4087,12},{65425,16},{65426,16},{65427,16},{65428,16},{65429,16},{65430,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0}, + {58,6},{502,9},{65431,16},{65432,16},{65433,16},{65434,16},{65435,16},{65436,16},{65437,16},{65438,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0}, + {59,6},{1017,10},{65439,16},{65440,16},{65441,16},{65442,16},{65443,16},{65444,16},{65445,16},{65446,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0}, + {121,7},{2039,11},{65447,16},{65448,16},{65449,16},{65450,16},{65451,16},{65452,16},{65453,16},{65454,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0}, + {122,7},{2040,11},{65455,16},{65456,16},{65457,16},{65458,16},{65459,16},{65460,16},{65461,16},{65462,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0}, + {249,8},{65463,16},{65464,16},{65465,16},{65466,16},{65467,16},{65468,16},{65469,16},{65470,16},{65471,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0}, + {503,9},{65472,16},{65473,16},{65474,16},{65475,16},{65476,16},{65477,16},{65478,16},{65479,16},{65480,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0}, + {504,9},{65481,16},{65482,16},{65483,16},{65484,16},{65485,16},{65486,16},{65487,16},{65488,16},{65489,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0}, + {505,9},{65490,16},{65491,16},{65492,16},{65493,16},{65494,16},{65495,16},{65496,16},{65497,16},{65498,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0}, + {506,9},{65499,16},{65500,16},{65501,16},{65502,16},{65503,16},{65504,16},{65505,16},{65506,16},{65507,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0}, + {2041,11},{65508,16},{65509,16},{65510,16},{65511,16},{65512,16},{65513,16},{65514,16},{65515,16},{65516,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0}, + {16352,14},{65517,16},{65518,16},{65519,16},{65520,16},{65521,16},{65522,16},{65523,16},{65524,16},{65525,16},{0,0},{0,0},{0,0},{0,0},{0,0}, + {1018,10},{32707,15},{65526,16},{65527,16},{65528,16},{65529,16},{65530,16},{65531,16},{65532,16},{65533,16},{65534,16},{0,0},{0,0},{0,0},{0,0},{0,0} + }; + static const int YQT[] = {16,11,10,16,24,40,51,61,12,12,14,19,26,58,60,55,14,13,16,24,40,57,69,56,14,17,22,29,51,87,80,62,18,22, + 37,56,68,109,103,77,24,35,55,64,81,104,113,92,49,64,78,87,103,121,120,101,72,92,95,98,112,100,103,99}; + static const int UVQT[] = {17,18,24,47,99,99,99,99,18,21,26,66,99,99,99,99,24,26,56,99,99,99,99,99,47,66,99,99,99,99,99,99, + 99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99}; + static const float aasf[] = { 1.0f * 2.828427125f, 1.387039845f * 2.828427125f, 1.306562965f * 2.828427125f, 1.175875602f * 2.828427125f, + 1.0f * 2.828427125f, 0.785694958f * 2.828427125f, 0.541196100f * 2.828427125f, 0.275899379f * 2.828427125f }; + + int row, col, i, k, subsample; + float fdtbl_Y[64], fdtbl_UV[64]; + unsigned char YTable[64], UVTable[64]; + + if(!data || !width || !height || comp > 4 || comp < 1) { + return 0; + } + + quality = quality ? quality : 90; + subsample = quality <= 90 ? 1 : 0; + quality = quality < 1 ? 1 : quality > 100 ? 100 : quality; + quality = quality < 50 ? 5000 / quality : 200 - quality * 2; + + for(i = 0; i < 64; ++i) { + int uvti, yti = (YQT[i]*quality+50)/100; + YTable[stbiw__jpg_ZigZag[i]] = (unsigned char) (yti < 1 ? 1 : yti > 255 ? 255 : yti); + uvti = (UVQT[i]*quality+50)/100; + UVTable[stbiw__jpg_ZigZag[i]] = (unsigned char) (uvti < 1 ? 1 : uvti > 255 ? 255 : uvti); + } + + for(row = 0, k = 0; row < 8; ++row) { + for(col = 0; col < 8; ++col, ++k) { + fdtbl_Y[k] = 1 / (YTable [stbiw__jpg_ZigZag[k]] * aasf[row] * aasf[col]); + fdtbl_UV[k] = 1 / (UVTable[stbiw__jpg_ZigZag[k]] * aasf[row] * aasf[col]); + } + } + + // Write Headers + { + static const unsigned char head0[] = { 0xFF,0xD8,0xFF,0xE0,0,0x10,'J','F','I','F',0,1,1,0,0,1,0,1,0,0,0xFF,0xDB,0,0x84,0 }; + static const unsigned char head2[] = { 0xFF,0xDA,0,0xC,3,1,0,2,0x11,3,0x11,0,0x3F,0 }; + const unsigned char head1[] = { 0xFF,0xC0,0,0x11,8,(unsigned char)(height>>8),STBIW_UCHAR(height),(unsigned char)(width>>8),STBIW_UCHAR(width), + 3,1,(unsigned char)(subsample?0x22:0x11),0,2,0x11,1,3,0x11,1,0xFF,0xC4,0x01,0xA2,0 }; + s->func(s->context, (void*)head0, sizeof(head0)); + s->func(s->context, (void*)YTable, sizeof(YTable)); + stbiw__putc(s, 1); + s->func(s->context, UVTable, sizeof(UVTable)); + s->func(s->context, (void*)head1, sizeof(head1)); + s->func(s->context, (void*)(std_dc_luminance_nrcodes+1), sizeof(std_dc_luminance_nrcodes)-1); + s->func(s->context, (void*)std_dc_luminance_values, sizeof(std_dc_luminance_values)); + stbiw__putc(s, 0x10); // HTYACinfo + s->func(s->context, (void*)(std_ac_luminance_nrcodes+1), sizeof(std_ac_luminance_nrcodes)-1); + s->func(s->context, (void*)std_ac_luminance_values, sizeof(std_ac_luminance_values)); + stbiw__putc(s, 1); // HTUDCinfo + s->func(s->context, (void*)(std_dc_chrominance_nrcodes+1), sizeof(std_dc_chrominance_nrcodes)-1); + s->func(s->context, (void*)std_dc_chrominance_values, sizeof(std_dc_chrominance_values)); + stbiw__putc(s, 0x11); // HTUACinfo + s->func(s->context, (void*)(std_ac_chrominance_nrcodes+1), sizeof(std_ac_chrominance_nrcodes)-1); + s->func(s->context, (void*)std_ac_chrominance_values, sizeof(std_ac_chrominance_values)); + s->func(s->context, (void*)head2, sizeof(head2)); + } + + // Encode 8x8 macroblocks + { + static const unsigned short fillBits[] = {0x7F, 7}; + int DCY=0, DCU=0, DCV=0; + int bitBuf=0, bitCnt=0; + // comp == 2 is grey+alpha (alpha is ignored) + int ofsG = comp > 2 ? 1 : 0, ofsB = comp > 2 ? 2 : 0; + const unsigned char *dataR = (const unsigned char *)data; + const unsigned char *dataG = dataR + ofsG; + const unsigned char *dataB = dataR + ofsB; + int x, y, pos; + if(subsample) { + for(y = 0; y < height; y += 16) { + for(x = 0; x < width; x += 16) { + float Y[256], U[256], V[256]; + for(row = y, pos = 0; row < y+16; ++row) { + // row >= height => use last input row + int clamped_row = (row < height) ? row : height - 1; + int base_p = (stbi__flip_vertically_on_write ? (height-1-clamped_row) : clamped_row)*width*comp; + for(col = x; col < x+16; ++col, ++pos) { + // if col >= width => use pixel from last input column + int p = base_p + ((col < width) ? col : (width-1))*comp; + float r = dataR[p], g = dataG[p], b = dataB[p]; + Y[pos]= +0.29900f*r + 0.58700f*g + 0.11400f*b - 128; + U[pos]= -0.16874f*r - 0.33126f*g + 0.50000f*b; + V[pos]= +0.50000f*r - 0.41869f*g - 0.08131f*b; + } + } + DCY = stbiw__jpg_processDU(s, &bitBuf, &bitCnt, Y+0, 16, fdtbl_Y, DCY, YDC_HT, YAC_HT); + DCY = stbiw__jpg_processDU(s, &bitBuf, &bitCnt, Y+8, 16, fdtbl_Y, DCY, YDC_HT, YAC_HT); + DCY = stbiw__jpg_processDU(s, &bitBuf, &bitCnt, Y+128, 16, fdtbl_Y, DCY, YDC_HT, YAC_HT); + DCY = stbiw__jpg_processDU(s, &bitBuf, &bitCnt, Y+136, 16, fdtbl_Y, DCY, YDC_HT, YAC_HT); + + // subsample U,V + { + float subU[64], subV[64]; + int yy, xx; + for(yy = 0, pos = 0; yy < 8; ++yy) { + for(xx = 0; xx < 8; ++xx, ++pos) { + int j = yy*32+xx*2; + subU[pos] = (U[j+0] + U[j+1] + U[j+16] + U[j+17]) * 0.25f; + subV[pos] = (V[j+0] + V[j+1] + V[j+16] + V[j+17]) * 0.25f; + } + } + DCU = stbiw__jpg_processDU(s, &bitBuf, &bitCnt, subU, 8, fdtbl_UV, DCU, UVDC_HT, UVAC_HT); + DCV = stbiw__jpg_processDU(s, &bitBuf, &bitCnt, subV, 8, fdtbl_UV, DCV, UVDC_HT, UVAC_HT); + } + } + } + } else { + for(y = 0; y < height; y += 8) { + for(x = 0; x < width; x += 8) { + float Y[64], U[64], V[64]; + for(row = y, pos = 0; row < y+8; ++row) { + // row >= height => use last input row + int clamped_row = (row < height) ? row : height - 1; + int base_p = (stbi__flip_vertically_on_write ? (height-1-clamped_row) : clamped_row)*width*comp; + for(col = x; col < x+8; ++col, ++pos) { + // if col >= width => use pixel from last input column + int p = base_p + ((col < width) ? col : (width-1))*comp; + float r = dataR[p], g = dataG[p], b = dataB[p]; + Y[pos]= +0.29900f*r + 0.58700f*g + 0.11400f*b - 128; + U[pos]= -0.16874f*r - 0.33126f*g + 0.50000f*b; + V[pos]= +0.50000f*r - 0.41869f*g - 0.08131f*b; + } + } + + DCY = stbiw__jpg_processDU(s, &bitBuf, &bitCnt, Y, 8, fdtbl_Y, DCY, YDC_HT, YAC_HT); + DCU = stbiw__jpg_processDU(s, &bitBuf, &bitCnt, U, 8, fdtbl_UV, DCU, UVDC_HT, UVAC_HT); + DCV = stbiw__jpg_processDU(s, &bitBuf, &bitCnt, V, 8, fdtbl_UV, DCV, UVDC_HT, UVAC_HT); + } + } + } + + // Do the bit alignment of the EOI marker + stbiw__jpg_writeBits(s, &bitBuf, &bitCnt, fillBits); + } + + // EOI + stbiw__putc(s, 0xFF); + stbiw__putc(s, 0xD9); + + return 1; +} + +STBIWDEF int stbi_write_jpg_to_func(stbi_write_func *func, void *context, int x, int y, int comp, const void *data, int quality) +{ + stbi__write_context s = { 0 }; + stbi__start_write_callbacks(&s, func, context); + return stbi_write_jpg_core(&s, x, y, comp, (void *) data, quality); +} + + +#ifndef STBI_WRITE_NO_STDIO +STBIWDEF int stbi_write_jpg(char const *filename, int x, int y, int comp, const void *data, int quality) +{ + stbi__write_context s = { 0 }; + if (stbi__start_write_file(&s,filename)) { + int r = stbi_write_jpg_core(&s, x, y, comp, data, quality); + stbi__end_write_file(&s); + return r; + } else + return 0; +} +#endif + +#endif // STB_IMAGE_WRITE_IMPLEMENTATION + +/* Revision history + 1.14 (2020-02-02) updated JPEG writer to downsample chroma channels + 1.13 + 1.12 + 1.11 (2019-08-11) + + 1.10 (2019-02-07) + support utf8 filenames in Windows; fix warnings and platform ifdefs + 1.09 (2018-02-11) + fix typo in zlib quality API, improve STB_I_W_STATIC in C++ + 1.08 (2018-01-29) + add stbi__flip_vertically_on_write, external zlib, zlib quality, choose PNG filter + 1.07 (2017-07-24) + doc fix + 1.06 (2017-07-23) + writing JPEG (using Jon Olick's code) + 1.05 ??? + 1.04 (2017-03-03) + monochrome BMP expansion + 1.03 ??? + 1.02 (2016-04-02) + avoid allocating large structures on the stack + 1.01 (2016-01-16) + STBIW_REALLOC_SIZED: support allocators with no realloc support + avoid race-condition in crc initialization + minor compile issues + 1.00 (2015-09-14) + installable file IO function + 0.99 (2015-09-13) + warning fixes; TGA rle support + 0.98 (2015-04-08) + added STBIW_MALLOC, STBIW_ASSERT etc + 0.97 (2015-01-18) + fixed HDR asserts, rewrote HDR rle logic + 0.96 (2015-01-17) + add HDR output + fix monochrome BMP + 0.95 (2014-08-17) + add monochrome TGA output + 0.94 (2014-05-31) + rename private functions to avoid conflicts with stb_image.h + 0.93 (2014-05-27) + warning fixes + 0.92 (2010-08-01) + casts to unsigned char to fix warnings + 0.91 (2010-07-17) + first public release + 0.90 first internal release +*/ + +/* +------------------------------------------------------------------------------ +This software is available under 2 licenses -- choose whichever you prefer. +------------------------------------------------------------------------------ +ALTERNATIVE A - MIT License +Copyright (c) 2017 Sean Barrett +Permission is hereby granted, free of charge, to any person obtaining a copy of +this software and associated documentation files (the "Software"), to deal in +the Software without restriction, including without limitation the rights to +use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies +of the Software, and to permit persons to whom the Software is furnished to do +so, subject to the following conditions: +The above copyright notice and this permission notice shall be included in all +copies or substantial portions of the Software. +THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, +OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE +SOFTWARE. +------------------------------------------------------------------------------ +ALTERNATIVE B - Public Domain (www.unlicense.org) +This is free and unencumbered software released into the public domain. +Anyone is free to copy, modify, publish, use, compile, sell, or distribute this +software, either in source code form or as a compiled binary, for any purpose, +commercial or non-commercial, and by any means. +In jurisdictions that recognize copyright laws, the author or authors of this +software dedicate any and all copyright interest in the software to the public +domain. We make this dedication for the benefit of the public at large and to +the detriment of our heirs and successors. We intend this dedication to be an +overt act of relinquishment in perpetuity of all present and future rights to +this software under copyright law. +THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +AUTHORS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN +ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION +WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. +------------------------------------------------------------------------------ +*/ diff --git a/include/tkDNN/stdafx.h b/include/tkDNN/stdafx.h new file mode 100644 index 0000000..b765708 --- /dev/null +++ b/include/tkDNN/stdafx.h @@ -0,0 +1,43 @@ +#ifndef STDAFX_H_INCLUDED +#define STDAFX_H_INCLUDED +#define BOOST_LOG_DYN_LINK 1 + +#pragma once + +#include +#include +#include +#include +#include +#include +#include + +#ifdef _WIN32 +#define NOMINMAX +#include +#else +# include +#endif + +#include "cpprest/json.h" +#include "cpprest/http_listener.h" +#include "cpprest/uri.h" +#include "cpprest/asyncrt_utils.h" +#include "cpprest/json.h" +#include "cpprest/filestream.h" +#include "cpprest/containerstream.h" +#include "cpprest/producerconsumerstream.h" + +#include +#include +#include +#include +#include +#include + +#pragma warning ( push ) +#pragma warning ( disable : 4457 ) +#pragma warning ( pop ) +#include +#include +#endif // STDAFX_H_INCLUDED diff --git a/include/tkDNN/test.h b/include/tkDNN/test.h new file mode 100644 index 0000000..e842269 --- /dev/null +++ b/include/tkDNN/test.h @@ -0,0 +1,78 @@ + +#include +int testInference(std::vector input_bins, std::vector output_bins, + tk::dnn::Network *net, tk::dnn::NetworkRT *netRT = nullptr) { + + std::vector outputs; + for(int i=0; inum_layers; i++) { + if(net->layers[i]->final) + outputs.push_back(net->layers[i]); + } + // no final layers, set last as output + if(outputs.size() == 0) { + outputs.push_back(net->layers[net->num_layers-1]); + } + + + // check input + if(input_bins.size() != 1) { + FatalError("currently support only 1 input"); + } + if(output_bins.size() != outputs.size()) { + std::cout<input_dim.tot(), &input_h, &data); + + // outputs + //dnnType *cudnn_out[outputs.size()], *rt_out[outputs.size()]; + std::vector cudnn_out,rt_out; + + tk::dnn::dataDim_t dim1 = net->input_dim; //input dim + printCenteredTitle(" CUDNN inference ", '=', 30); { + dim1.print(); + TKDNN_TSTART + net->infer(dim1, data); + TKDNN_TSTOP + dim1.print(); + } + for(int i=0; idstData); + + if(netRT != nullptr) { + tk::dnn::dataDim_t dim2 = net->input_dim; + printCenteredTitle(" TENSORRT inference ", '=', 30); { + dim2.print(); + TKDNN_TSTART + netRT->infer(dim2, data); + TKDNN_TSTOP + dim2.print(); + } + for(int i=0; ibuffersRT[i+1]); + } + + int ret_cudnn = 0, ret_tensorrt = 0, ret_cudnn_tensorrt = 0; + for(int i=0; iinfer(dim2, input_d); - TIMER_STOP - dim2.print(); - } - // checkResult(dim2.tot(), input_h, input); -} - -cv::Mat CenternetDetection::draw(cv::Mat &imageORIG) { - - tk::dnn::box b; - int x0, w, x1, y0, h, y1; - int objClass; - std::string det_class; - int baseline = 0; - float fontScale = 0.5; - int thickness = 2; - - for(int c=0; c(0,0)=width * 0.5; + dst2.at(0,1)=width * 0.5; + dst2.at(1,0)=width * 0.5; + dst2.at(1,1)=width * 0.5 + width * -0.5; - if(!imageORIG.data) { - std::cout<<"YOLO: NO IMAGE DATA\n"; - return; - } - TIMER_START - auto start_t = std::chrono::steady_clock::now(); - auto step_t = std::chrono::steady_clock::now(); - auto end_t = std::chrono::steady_clock::now(); - // -----------------------------------pre-process ------------------------------------------ - // it will resize the images to `224 x 224` in GETTING_STARTED.md - cv::Size sz = imageORIG.size(); - std::cout<<"image: "<(2,0)=dst2.at(1,0) + (-dst2.at(0,1)+dst2.at(1,1) ); + dst2.at(2,1)=dst2.at(1,1) + (dst2.at(0,0)-dst2.at(1,0) ); +} + + +void CenternetDetection::preprocess(cv::Mat &frame, const int bi){ + // -----------------------------------pre-process ------------------------------------------ + + // auto start_t = std::chrono::steady_clock::now(); + // auto step_t = std::chrono::steady_clock::now(); + // auto end_t = std::chrono::steady_clock::now(); + cv::Size sz = originalSize[bi]; + // std::cout<<"image: "< sz.height){ - s[0] = sz.width * 1.0; - s[1] = sz.width * 1.0; - } - else{ - s[0] = sz.height * 1.0; - s[1] = sz.height * 1.0; - } + if(sz.width > sz.height){ + s[0] = sz.width * 1.0; + s[1] = sz.width * 1.0; + } + else{ + s[0] = sz.height * 1.0; + s[1] = sz.height * 1.0; + } - // ----------- get_affine_transform - // rot_rad = pi * 0 / 100 --> 0 - - src.at(0,0)=c[0]; - src.at(0,1)=c[1]; - src.at(1,0)=c[0]; - src.at(1,1)=c[1] + s[0] * -0.5; - dst.at(0,0)=inp_width * 0.5; - dst.at(0,1)=inp_height * 0.5; - dst.at(1,0)=inp_width * 0.5; - dst.at(1,1)=inp_height * 0.5 + inp_width * -0.5; + // ----------- get_affine_transform + // rot_rad = pi * 0 / 100 --> 0 + + src.at(0,0)=c[0]; + src.at(0,1)=c[1]; + src.at(1,0)=c[0]; + src.at(1,1)=c[1] + s[0] * -0.5; + dst.at(0,0)=netRT->input_dim.w * 0.5; + dst.at(0,1)=netRT->input_dim.h * 0.5; + dst.at(1,0)=netRT->input_dim.w * 0.5; + dst.at(1,1)=netRT->input_dim.h * 0.5 + netRT->input_dim.w * -0.5; + + src.at(2,0)=src.at(1,0) + (-src.at(0,1)+src.at(1,1) ); + src.at(2,1)=src.at(1,1) + (src.at(0,0)-src.at(1,0) ); + dst.at(2,0)=dst.at(1,0) + (-dst.at(0,1)+dst.at(1,1) ); + dst.at(2,1)=dst.at(1,1) + (dst.at(0,0)-dst.at(1,0) ); + + trans = cv::getAffineTransform( src, dst ); + // end_t = std::chrono::steady_clock::now(); + // std::cout << " TIME gett affine trans: " << std::chrono::duration_cast(end_t - step_t).count() << " us" << std::endl; + // step_t = end_t; + + trans2 = cv::getAffineTransform( dst2, src ); + // end_t = std::chrono::steady_clock::now(); + // std::cout << " TIME getAffineTrans 2: " << std::chrono::duration_cast(end_t - step_t).count() << " us" << std::endl; + // step_t = end_t; + } + sz_old = sz; +#ifdef OPENCV_CUDACONTRIB + cv::cuda::GpuMat im_Orig; + cv::cuda::GpuMat imageF1_d, imageF2_d; + + im_Orig = cv::cuda::GpuMat(frame); + cv::cuda::resize (im_Orig, imageF1_d, cv::Size(new_width, new_height)); + checkCuda( cudaDeviceSynchronize() ); - src.at(2,0)=src.at(1,0) + (-src.at(0,1)+src.at(1,1) ); - src.at(2,1)=src.at(1,1) + (src.at(0,0)-src.at(1,0) ); - dst.at(2,0)=dst.at(1,0) + (-dst.at(0,1)+dst.at(1,1) ); - dst.at(2,1)=dst.at(1,1) + (dst.at(0,0)-dst.at(1,0) ); - // std::cout<<"src: "<(end_t - step_t).count() << " us" << std::endl; + // step_t = end_t; + + cv::cuda::warpAffine(imageF1_d, imageF2_d, trans, cv::Size(netRT->input_dim.w, netRT->input_dim.h), cv::INTER_LINEAR ); + checkCuda( cudaDeviceSynchronize() ); + + imageF2_d.convertTo(imageF1_d, CV_32FC3, 1/255.0); + checkCuda( cudaDeviceSynchronize() ); + // end_t = std::chrono::steady_clock::now(); + // std::cout << " TIME convert: " << std::chrono::duration_cast(end_t - step_t).count() << " us" << std::endl; + // step_t = end_t; + + dim2 = dim; + cv::cuda::GpuMat bgr[3]; + cv::cuda::split(imageF1_d,bgr);//split source + // end_t = std::chrono::steady_clock::now(); + // std::cout << " TIME split: " << std::chrono::duration_cast(end_t - step_t).count() << " us" << std::endl; + // step_t = end_t; + + for(int i=0; i(end_t - step_t).count() << " us" << std::endl; + // step_t = end_t; + + checkCuda(cudaMemcpy(input_d+ netRT->input_dim.tot()*bi, d_ptrs, dim2.tot()*sizeof(dnnType), cudaMemcpyDeviceToDevice)); + + // end_t = std::chrono::steady_clock::now(); + // std::cout << " TIME Memcpy to input_d: " << std::chrono::duration_cast(end_t - step_t).count() << " us" << std::endl; + // step_t = end_t; +#else + + cv::Mat imageF; + resize(frame, imageF, cv::Size(new_width, new_height)); + sz = imageF.size(); + // std::cout<<"size: "<(end_t - step_t).count() << " us" << std::endl; + // step_t = end_t; + cv::Mat trans = cv::getAffineTransform( src, dst ); - end_t = std::chrono::steady_clock::now(); - std::cout << " TIME getAffinetr : " << std::chrono::duration_cast(end_t - step_t).count() << " ms" << std::endl; - step_t = end_t; - resize(imageORIG, imageF, cv::Size(new_width, new_height)); - sz = imageF.size(); - std::cout<<"size: "<(end_t - step_t).count() << " ms" << std::endl; - step_t = end_t; - cv::warpAffine(imageF, imageF, trans, cv::Size(inp_width, inp_height), cv::INTER_LINEAR ); - - end_t = std::chrono::steady_clock::now(); - std::cout << " TIME warpAffine: " << std::chrono::duration_cast(end_t - step_t).count() << " ms" << std::endl; - step_t = end_t; + cv::warpAffine(imageF, imageF, trans, cv::Size(netRT->input_dim.w, netRT->input_dim.h), cv::INTER_LINEAR ); + // end_t = std::chrono::steady_clock::now(); + // std::cout << " TIME warpAffine: " << std::chrono::duration_cast(end_t - step_t).count() << " us" << std::endl; + // step_t = end_t; sz = imageF.size(); - std::cout<<"size: "<(end_t - step_t).count() << " ms" << std::endl; - step_t = end_t; - std::cout<<"mean: "<(end_t - step_t).count() << " us" << std::endl; + // step_t = end_t; dim2 = dim; - end_t = std::chrono::steady_clock::now(); - std::cout << " TIME before split: " << std::chrono::duration_cast(end_t - step_t).count() << " us" << std::endl; - step_t = end_t; - //split channels + cv::Mat bgr[3]; cv::split(imageF,bgr);//split source - - end_t = std::chrono::steady_clock::now(); - std::cout << " TIME split: " << std::chrono::duration_cast(end_t - step_t).count() << " ms" << std::endl; - step_t = end_t; - for(int i=0; i<3; i++){ bgr[i] = bgr[i] - mean[i]; bgr[i] = bgr[i] / stddev[i]; } - end_t = std::chrono::steady_clock::now(); - std::cout << " TIME mean std: " << std::chrono::duration_cast(end_t - step_t).count() << " ms" << std::endl; - step_t = end_t; - //write channels for(int i=0; iinput_dim.tot()*bi], (void*)bgr[ch].data, imageF.rows*imageF.cols*sizeof(dnnType)); } + checkCuda(cudaMemcpyAsync(input_d+ netRT->input_dim.tot()*bi, input+ netRT->input_dim.tot()*bi, dim2.tot()*sizeof(dnnType), cudaMemcpyHostToDevice)); +#endif +} - checkCuda(cudaMemcpyAsync(input_d, input, dim2.tot()*sizeof(dnnType), cudaMemcpyHostToDevice)); +void CenternetDetection::postprocess(const int bi, const bool mAP){ + dnnType *rt_out[4]; + rt_out[0] = (dnnType *)netRT->buffersRT[1]+ netRT->buffersDIM[1].tot()*bi; + rt_out[1] = (dnnType *)netRT->buffersRT[2]+ netRT->buffersDIM[2].tot()*bi; + rt_out[2] = (dnnType *)netRT->buffersRT[3]+ netRT->buffersDIM[3].tot()*bi; + rt_out[3] = (dnnType *)netRT->buffersRT[4]+ netRT->buffersDIM[4].tot()*bi; - printCenteredTitle(" TENSORRT inference ", '=', 30); { - dim2.print(); - TIMER_START - netRT->infer(dim2, input_d); - TIMER_STOP - dim2.print(); - } - // checkResult(dim2.tot(), input_h, input); - std::cout<<" --- pre-process ---\n"; - end_t = std::chrono::steady_clock::now(); - std::cout << " TIME : " << std::chrono::duration_cast(end_t - step_t).count() << " ms" << std::endl; - step_t = end_t; + // auto start_t = std::chrono::steady_clock::now(); + // auto step_t = std::chrono::steady_clock::now(); + // auto end_t = std::chrono::steady_clock::now(); // ------------------------------------ process -------------------------------------------- - - rt_out[0] = (dnnType *)netRT->buffersRT[1]; - rt_out[1] = (dnnType *)netRT->buffersRT[2]; - rt_out[2] = (dnnType *)netRT->buffersRT[3]; - rt_out[3] = (dnnType *)netRT->buffersRT[4]; - activationSIGMOIDForward(rt_out[0], rt_out[0], dim_hm.tot()); checkCuda( cudaDeviceSynchronize() ); - end_t = std::chrono::steady_clock::now(); - std::cout << " TIME sigmoid : " << std::chrono::duration_cast(end_t - step_t).count() << " ms" << std::endl; - step_t = end_t; - subtractWithThreshold(rt_out[0], rt_out[0] + dim_hm.tot(), rt_out[1], rt_out[0]); - - float *prova; - checkCuda( cudaMallocHost(&prova, K*sizeof(float)) ); - checkCuda( cudaMemcpy(prova, rt_out[0], K*sizeof(float), cudaMemcpyDeviceToHost) ); - std::cout<<"heat:\n"; - for(int i=0; i toll || hm_h[i]-hmax_h[i] < -toll){ - // hm_h[i] = 0.0f; - // } - // } - // checkCuda( cudaFreeHost(hmax_h) ); - std::cout<<" --- hmax ---\n"; - end_t = std::chrono::steady_clock::now(); - std::cout << " TIME : " << std::chrono::duration_cast(end_t - step_t).count() << " ms" << std::endl; - step_t = end_t; + subtractWithThreshold(rt_out[0], rt_out[0] + dim_hm.tot(), rt_out[1], rt_out[0], op); + // end_t = std::chrono::steady_clock::now(); + // std::cout << " TIME threshold: " << std::chrono::duration_cast(end_t - step_t).count() << " us" << std::endl; + // step_t = end_t; // ----------- nms end // ----------- topk - - // thrust::device_vector ids_d; - // int ids[dim_hm.h * dim_hm.w]; - // for(int i=0; i ids2( dim_hm.h * dim_hm.w ); - // for(int i=0; i dim_hm.h * dim_hm.w){ printf ("Error topk (K is too large)\n"); return; } - checkCuda( cudaMemcpy(ids_d, ids_, dim_hm.c * dim_hm.h * dim_hm.w*sizeof(int), cudaMemcpyHostToDevice) ); - // checkCuda( cudaMemcpy(ids_2d, ids_2, dim_hm.h * dim_hm.w*sizeof(int), cudaMemcpyHostToDevice) ); - - - // sortAndTopKonDevice(rt_out[0], ids_2d, topk_scores, topk_inds_ , topk_ys_ , topk_xs_ ,dim_hm.h * dim_hm.w, K, dim_hm.c); - // checkCuda( cudaDeviceSynchronize() ); - // for(int i=0; ioutput_dim.h * hm->output_dim.w elements for each channel and sort it. Then find the first 100 elements - // // memcpy(ids2, ids, dim_hm.h * dim_hm.w); - // sort(rt_out[0]+ i * dim_hm.h * dim_hm.w, - // rt_out[0]+ i * dim_hm.h * dim_hm.w + dim_hm.h * dim_hm.w, - // ids_d); - // // end_t = std::chrono::steady_clock::now(); - // // std::cout << " TIME sort channel "<(end_t - step_t).count() << " ms" << std::endl; - // // step_t = end_t; - // topk(rt_out[0]+ i * dim_hm.h * dim_hm.w, ids_d, K, topk_scores + i*K, - // topk_inds_ + i*K, topk_ys_ + i*K, topk_xs_ + i*K); - // // checkCuda( cudaMemcpy(ids2, ids2_d, dim_hm.h * dim_hm.w*sizeof(int), cudaMemcpyDeviceToHost) ); - - // // for (int j=0; j(end_t - step_t).count() << " ms" << std::endl; - // // step_t = end_t; - - // } - // checkCuda( cudaFree(ids_d )); - std::cout<<" --- a 100 ---\n"; - end_t = std::chrono::steady_clock::now(); - std::cout << " TIME sort topk on 80 channel: " << std::chrono::duration_cast(end_t - step_t).count() << " ms" << std::endl; - step_t = end_t; - // final - - // sort(topk_scores, - // topk_scores + dim_hm.c * K, - // topk_inds_); - sort(rt_out[0], - rt_out[0]+dim_hm.tot(), - ids_d); + sort(rt_out[0],rt_out[0]+dim_hm.tot(),ids_d); checkCuda( cudaDeviceSynchronize() ); - int *topk_inds; - checkCuda( cudaMallocHost(&topk_inds, K*sizeof(int)) ); - // checkCuda( cudaMemcpy(topk_inds, ids_d, K*sizeof(int), cudaMemcpyDeviceToHost) ); - // for(int i=0; i(end_t - step_t).count() << " us" << std::endl; + // step_t = end_t; - end_t = std::chrono::steady_clock::now(); - std::cout << " TIME sort channel: " << std::chrono::duration_cast(end_t - step_t).count() << " ms" << std::endl; - step_t = end_t; + topk(rt_out[0], ids_d, K, scores_d, topk_inds_d, topk_ys_d, topk_xs_d); + checkCuda( cudaDeviceSynchronize() ); - // topk(topk_scores, topk_inds_, K, scores_d, - // topk_inds_d, topk_ys_d, topk_xs_d); - topk(rt_out[0], ids_d, K, scores_d, - topk_inds_d, topk_ys_d, topk_xs_d); - checkCuda( cudaDeviceSynchronize() ); - end_t = std::chrono::steady_clock::now(); - std::cout << " TIME topk channel: " << std::chrono::duration_cast(end_t - step_t).count() << " ms" << std::endl; - step_t = end_t; - - - checkCuda( cudaMemcpy(topk_inds, topk_inds_d, K*sizeof(int), cudaMemcpyDeviceToHost) ); - for(int i=0; i(end_t - step_t).count() << " us" << std::endl; + // step_t = end_t; + checkCuda( cudaMemcpy(scores, scores_d, K *sizeof(float), cudaMemcpyDeviceToHost) ); - std::cout<<"\n\nscores:\n"; - for(int i=0; i(end_t - step_t).count() << " us" << std::endl; + // step_t = end_t; + checkCuda( cudaMemcpy(topk_xs_d, (float *)inttopk_xs_d, K*sizeof(float), cudaMemcpyDeviceToDevice) ); checkCuda( cudaMemcpy(topk_ys_d, (float *)inttopk_ys_d, K*sizeof(float), cudaMemcpyDeviceToDevice) ); checkCuda( cudaMemcpy(clses, clses_d, K*sizeof(int), cudaMemcpyDeviceToHost) ); - std::cout<<"\ntopk_ids: \n"; - checkCuda( cudaMemcpy(topk_inds, topk_inds_d, K*sizeof(int), cudaMemcpyDeviceToHost) ); - for(int i=0; i(end_t - step_t).count() << " ms" << std::endl; - step_t = end_t; + // ----------- topk end - // dnnType *reg_aus; - // checkCuda( cudaMallocHost(®_aus, dim_reg.tot()*sizeof(dnnType)) ); - // checkCuda( cudaMemcpy(reg_aus, rt_out[3], dim_reg.tot()*sizeof(dnnType), cudaMemcpyDeviceToHost) ); - - // for(int i = 0; i < K; i++){ - // topk_xs[i] = topk_xs[i] + reg_aus[topk_inds[i]]; - // topk_ys[i] = topk_ys[i] + reg_aus[topk_inds[i]+dim_reg.h*dim_reg.w]; - // } - topKxyAddOffset(topk_inds_d, K, dim_reg.h*dim_reg.w, inttopk_xs_d, inttopk_ys_d, topk_xs_d, topk_ys_d, rt_out[3]); + topKxyAddOffset(topk_inds_d, K, dim_reg.h*dim_reg.w, inttopk_xs_d, inttopk_ys_d, topk_xs_d, topk_ys_d, rt_out[3], src_out, ids_out); // checkCuda( cudaDeviceSynchronize() ); - end_t = std::chrono::steady_clock::now(); - std::cout << " TIME add offset: " << std::chrono::duration_cast(end_t - step_t).count() << " ms" << std::endl; - step_t = end_t; - // checkCuda( cudaFreeHost(reg_aus) ); - - // dnnType *wh_aus; - // checkCuda( cudaMemcpy(wh_aus, rt_out[2], dim_wh.tot()*sizeof(dnnType), cudaMemcpyDeviceToHost) ); - bboxes(topk_inds_d, K, dim_wh.h*dim_wh.w, topk_xs_d, topk_ys_d, rt_out[2], bbx0_d, bbx1_d, bby0_d, bby1_d); + // end_t = std::chrono::steady_clock::now(); + // std::cout << " TIME add offset: " << std::chrono::duration_cast(end_t - step_t).count() << " us" << std::endl; + // step_t = end_t; + + bboxes(topk_inds_d, K, dim_wh.h*dim_wh.w, topk_xs_d, topk_ys_d, rt_out[2], bbx0_d, bbx1_d, bby0_d, bby1_d, src_out, ids_out); // checkCuda( cudaDeviceSynchronize() ); + checkCuda( cudaMemcpy(bbx0, bbx0_d, K * sizeof(float), cudaMemcpyDeviceToHost) ); checkCuda( cudaMemcpy(bby0, bby0_d, K * sizeof(float), cudaMemcpyDeviceToHost) ); checkCuda( cudaMemcpy(bbx1, bbx1_d, K * sizeof(float), cudaMemcpyDeviceToHost) ); checkCuda( cudaMemcpy(bby1, bby1_d, K * sizeof(float), cudaMemcpyDeviceToHost) ); - // for(int i = 0; i < K; i++){ - // bboxes[i * 4] = topk_xs[i] - wh_aus[topk_inds[i]] / 2; - // bboxes[i * 4 + 1] = topk_ys[i] - wh_aus[topk_inds[i]+dim_reg.h*dim_reg.w] / 2; - // bboxes[i * 4 + 2] = topk_xs[i] + wh_aus[topk_inds[i]] / 2; - // bboxes[i * 4 + 3] = topk_ys[i] + wh_aus[topk_inds[i]+dim_reg.h*dim_reg.w] / 2; - // } - // for(int i = 0; i < K; i++){ - // std::cout<<"-----\n(x0, y0) = ("<(end_t - step_t).count() << " us" << std::endl; + // step_t = end_t; - // checkCuda( cudaFreeHost(wh_aus) ); - // checkCuda( cudaFreeHost(topk_inds) ); - // checkCuda( cudaFreeHost(topk_ys) ); - // checkCuda( cudaFreeHost(topk_xs) ); - std::cout<<" --- bboxes ---\n"; - end_t = std::chrono::steady_clock::now(); - std::cout << " TIME : " << std::chrono::duration_cast(end_t - step_t).count() << " ms" << std::endl; - step_t = end_t; - // servono [bboxes, scores, clses] - // checkCuda( cudaDeviceSynchronize() ); - - std::cout<<" --- process ---\n"; - end_t = std::chrono::steady_clock::now(); - std::cout << " TIME : " << std::chrono::duration_cast(end_t - step_t).count() << " ms" << std::endl; - step_t = end_t; // ---------------------------------- post-process ----------------------------------------- // --------- ctdet_post_process // --------- transform_preds - src.at(0,0)=c[0]; - src.at(0,1)=c[1]; - src.at(1,0)=c[0]; - src.at(1,1)=c[1] + s[0] * -0.5; - dst.at(0,0)=width * 0.5; - dst.at(0,1)=width * 0.5; - dst.at(1,0)=width * 0.5; - dst.at(1,1)=width * 0.5 + width * -0.5; - - src.at(2,0)=src.at(1,0) + (-src.at(0,1)+src.at(1,1) ); - src.at(2,1)=src.at(1,1) + (src.at(0,0)-src.at(1,0) ); - dst.at(2,0)=dst.at(1,0) + (-dst.at(0,1)+dst.at(1,1) ); - dst.at(2,1)=dst.at(1,1) + (dst.at(0,0)-dst.at(1,0) ); - - - cv::Mat trans2(cv::Size(3,2), CV_32F); - trans2 = cv::getAffineTransform( dst, src ); - end_t = std::chrono::steady_clock::now(); - std::cout << " TIME getAffineTrans 2: " << std::chrono::duration_cast(end_t - step_t).count() << " ms" << std::endl; - step_t = end_t; cv::Mat new_pt1(cv::Size(1,2), CV_32F); - cv::Mat new_pt2(cv::Size(1,2), CV_32F); - + cv::Mat new_pt2(cv::Size(1,2), CV_32F); for(int i = 0; i(0,0)=static_cast(trans2.at(0,0))*bbx0[i] + @@ -570,29 +358,24 @@ void CenternetDetection::update(cv::Mat &imageORIG) { static_cast(trans2.at(1,1))*bby1[i] + static_cast(trans2.at(1,2))*1.0; - // std::cout<<"\n new: "<(0,0); target_coords[i*4+1] = new_pt1.at(0,1); target_coords[i*4+2] = new_pt2.at(0,0); target_coords[i*4+3] = new_pt2.at(0,1); - // std::cout<(0,0)<<", "<(0,1)<<", "<(0,0)<<", "<(0,1)< thresh){ - std::cout<<"th: "< confThreshold){ + // std::cout<<"th: "<(end_t - step_t).count() << " ms" << std::endl; - step_t = end_t; - std::cout<<"TOTAL: \n"; - TIMER_STOP + + batchDetected.push_back(detected); + // end_t = std::chrono::steady_clock::now(); + // std::cout << " TIME detections: " << std::chrono::duration_cast(end_t - step_t).count() << " us" << std::endl; + // step_t = end_t; } -}} \ No newline at end of file + + +}} + + diff --git a/src/Conv2d.cpp b/src/Conv2d.cpp index f984f67..b57cf58 100644 --- a/src/Conv2d.cpp +++ b/src/Conv2d.cpp @@ -17,8 +17,6 @@ void Conv2d::initCUDNN(bool back) { idim = output_dim; odim = input_dim; } - //idim.print(); - //odim.print(); checkCUDNN( cudnnCreateFilterDescriptor(&filterDesc) ); checkCUDNN( cudnnCreateConvolutionDescriptor(&convDesc) ); @@ -29,7 +27,7 @@ void Conv2d::initCUDNN(bool back) { net->tensorFormat, net->dataType, idim.n, idim.c, idim.h, idim.w) ); checkCUDNN( cudnnSetFilter4dDescriptor(filterDesc, - net->dataType, net->tensorFormat, odim.c, idim.c, + net->dataType, net->tensorFormat, odim.c, idim.c/groups, kernelH, kernelW) ); checkCUDNN( cudnnSetConvolution2dDescriptor(convDesc, @@ -38,16 +36,20 @@ void Conv2d::initCUDNN(bool back) { 1,1, // upscale CUDNN_CROSS_CORRELATION, CUDNN_DATA_FLOAT) ); + checkCUDNN( cudnnSetConvolutionGroupCount(convDesc, + groups) ); + // check dimension of convolution output dataDim_t tmpdim; checkCUDNN( cudnnGetConvolution2dForwardOutputDim( convDesc, srcTensor, filterDesc, &tmpdim.n, &tmpdim.c, &tmpdim.h, &tmpdim.w) ); + if(odim.n != tmpdim.n || odim.c != tmpdim.c || odim.h != tmpdim.h || odim.w != tmpdim.w) { std::cout<<"tkdim input: "; idim.print(); std::cout<<"tkdim output: "; odim.print(); std::cout<<"cudnndim: "; tmpdim.print(); - FatalError("Eror conv dimension mismatch"); + FatalError("Error conv dimension mismatch"); } checkCUDNN( cudnnSetTensor4dDescriptor(dstTensor, @@ -60,25 +62,30 @@ void Conv2d::initCUDNN(bool back) { // init workspace workSpace = NULL; ws_sizeInBytes = 0; + int algo_count = 0; if(back) { - checkCUDNN( cudnnGetConvolutionBackwardDataAlgorithm(net->cudnnHandle, - filterDesc, dstTensor, convDesc, srcTensor, - CUDNN_CONVOLUTION_BWD_DATA_PREFER_FASTEST, 0, &bwAlgo) ); + checkCUDNN( cudnnGetConvolutionBackwardDataAlgorithm_v7(net->cudnnHandle, + filterDesc, dstTensor, convDesc, srcTensor, 1, &algo_count, &bwAlgo) ); checkCUDNN(cudnnGetConvolutionBackwardDataWorkspaceSize(net->cudnnHandle, - filterDesc, dstTensor, convDesc, srcTensor, - bwAlgo, &ws_sizeInBytes)); + filterDesc, dstTensor, convDesc, srcTensor, + bwAlgo.algo, &ws_sizeInBytes)); + // invert tensors srcTensorDesc = dstTensor; dstTensorDesc = srcTensor; } else { - checkCUDNN( cudnnGetConvolutionForwardAlgorithm(net->cudnnHandle, - srcTensor, filterDesc, convDesc, dstTensor, - CUDNN_CONVOLUTION_FWD_PREFER_FASTEST, 0, &algo) ); - checkCUDNN(cudnnGetConvolutionForwardWorkspaceSize(net->cudnnHandle, - srcTensor, filterDesc, convDesc, dstTensor, - algo, &ws_sizeInBytes)); + + checkCUDNN( cudnnGetConvolutionForwardAlgorithm_v7(net->cudnnHandle, + srcTensor, filterDesc, convDesc, dstTensor, + 1, &algo_count, &algo) ); + checkCUDNN(cudnnGetConvolutionForwardWorkspaceSize(net->cudnnHandle, + srcTensor, filterDesc, convDesc, dstTensor, + algo.algo, &ws_sizeInBytes)); } + + if(algo_count < 1) + FatalError("Cannot retrieve convolutional algo"); } void Conv2d::inferCUDNN(dnnType* srcData, bool back) { @@ -89,16 +96,16 @@ void Conv2d::inferCUDNN(dnnType* srcData, bool back) { checkCUDNN(cudnnConvolutionBackwardData(net->cudnnHandle, &alpha, filterDesc, data_d, srcTensorDesc, srcData, - convDesc, bwAlgo, workSpace, ws_sizeInBytes, + convDesc, bwAlgo.algo, workSpace, ws_sizeInBytes, &beta, dstTensorDesc, dstData)); } else { checkCUDNN(cudnnConvolutionForward(net->cudnnHandle, &alpha, srcTensorDesc, srcData, filterDesc, - data_d, convDesc, algo, workSpace, ws_sizeInBytes, + data_d, convDesc, algo.algo, workSpace, ws_sizeInBytes, &beta, dstTensorDesc, dstData)); } - if(!batchnorm) { + if(!batchnorm && !additional_bias) { //CHECK WITH IF CORRECT // bias alpha = dnnType(1); beta = dnnType(1); @@ -106,24 +113,34 @@ void Conv2d::inferCUDNN(dnnType* srcData, bool back) { &alpha, biasTensorDesc, bias_d, &beta, dstTensorDesc, dstData) ); } else { - alpha = dnnType(1); - beta = dnnType(0); - checkCUDNN( cudnnBatchNormalizationForwardInference(net->cudnnHandle, - CUDNN_BATCHNORM_SPATIAL, &alpha, &beta, - dstTensorDesc, dstData, dstTensorDesc, - dstData, biasTensorDesc, //same tensor descriptor as bias - scales_d, bias_d, mean_d, variance_d, - TKDNN_BN_MIN_EPSILON) ); + if(additional_bias) + { + alpha = dnnType(1); + beta = dnnType(1); + checkCUDNN( cudnnAddTensor(net->cudnnHandle, + &alpha, biasTensorDesc, bias2_d, + &beta, dstTensorDesc, dstData) ); + } + if(batchnorm) + { + alpha = dnnType(1); + beta = dnnType(0); + checkCUDNN( cudnnBatchNormalizationForwardInference(net->cudnnHandle, + CUDNN_BATCHNORM_SPATIAL, &alpha, &beta, + dstTensorDesc, dstData, dstTensorDesc, + dstData, biasTensorDesc, //same tensor descriptor as bias + scales_d, bias_d, mean_d, variance_d, + TKDNN_BN_MIN_EPSILON) ); + } } } Conv2d::Conv2d( Network *net, int out_ch, int kernelH, int kernelW, int strideH, int strideW, int paddingH, int paddingW, - std::string fname_weights, bool batchnorm, bool deConv, bool final) : + std::string fname_weights, bool batchnorm, bool deConv, int groups, bool additional_bias) : LayerWgs(net, net->getOutputDim().c, out_ch, kernelH, kernelW, 1, - fname_weights, batchnorm, false, final) { - + fname_weights, batchnorm, additional_bias, deConv, groups) { this->kernelH = kernelH; this->kernelW = kernelW; this->strideH = strideH; @@ -131,6 +148,8 @@ Conv2d::Conv2d( Network *net, int out_ch, int kernelH, int kernelW, this->paddingH = paddingH; this->paddingW = paddingW; this->deConv = deConv; + this->groups = groups; + this->additional_bias = additional_bias; if(!deConv) { output_dim.n = input_dim.n; diff --git a/src/DarknetParser.cpp b/src/DarknetParser.cpp new file mode 100644 index 0000000..69b6b29 --- /dev/null +++ b/src/DarknetParser.cpp @@ -0,0 +1,274 @@ +#include "tkDNN/DarknetParser.h" + +namespace tk { namespace dnn { + + std::string darknetParseType(const std::string& line){ + size_t start = line.find("["); + size_t end = line.find("]"); + if( start == std::string::npos || end == std::string::npos) + return ""; + start++; + std::string type = line.substr(start, end-start); + return type; + } + + bool divideNameAndValue(const std::string& line, std::string&name, std::string& value){ + size_t sep = line.find("="); + if(sep == std::string::npos) + return false; + + name = line.substr(0, sep); + value = line.substr(sep+1, line.size() - (sep+1)); + return true; + } + + std::vector fromStringToIntVec(const std::string& line, const char delimiter){ + std::stringstream linestream(line); + std::string value; + std::vector values; + + while(getline(linestream,value,delimiter)) + values.push_back(std::stoi(value)); + return values; + } + + bool darknetParseFields(const std::string& line, darknetFields_t& fields){ + + std::string name,value; + if(!divideNameAndValue(line, name, value)) + return false; + + if(name.find("new_coords") != std::string::npos) + fields.new_coords = std::stoi(value); + else if(name.find("width") != std::string::npos) + fields.width = std::stoi(value); + else if(name.find("height") != std::string::npos) + fields.height = std::stoi(value); + else if(name.find("channels") != std::string::npos) + fields.channels = std::stoi(value); + else if(name.find("batch_normalize") != std::string::npos) + fields.batch_normalize = std::stoi(value); + else if(name.find("filters") != std::string::npos) + fields.filters = std::stoi(value); + else if(name.find("activation") != std::string::npos) + fields.activation = value; + else if(name.find("size") != std::string::npos){ + fields.size_x = std::stoi(value); + fields.size_y = std::stoi(value); + } + else if(name.find("size_x") != std::string::npos) + fields.size_x = std::stoi(value); + else if(name.find("size_y") != std::string::npos) + fields.size_y = std::stoi(value); + else if(name.find("stride") != std::string::npos){ + fields.stride_x = std::stoi(value); + fields.stride_y = std::stoi(value); + } + else if(name.find("stride_x") != std::string::npos) + fields.stride_x = std::stoi(value); + else if(name.find("stride_y") != std::string::npos) + fields.stride_y = std::stoi(value); + else if(name.find("pad") != std::string::npos) + fields.pad = std::stoi(value); + else if(name.find("classes") != std::string::npos) + fields.classes = std::stoi(value); + else if(name.find("num") != std::string::npos) + fields.num = std::stoi(value); + else if(name.find("coords") != std::string::npos) + fields.coords = std::stoi(value); + else if(name.find("groups") != std::string::npos) + fields.groups = std::stoi(value); + else if(name.find("group_id") != std::string::npos) + fields.group_id = std::stoi(value); + else if(name.find("scale_x_y") != std::string::npos) + fields.scale_xy = std::stof(value); + else if(name.find("beta_nms") != std::string::npos) + fields.nms_thresh = std::stof(value); + else if(name.find("nms_kind") != std::string::npos){ + if(value == "greedynms") fields.nms_kind = 0; + else if(value == "diounms") fields.nms_kind = 1; + else std::cout<<"Not supported nms_kind "< &netLayers, const std::vector& names) { + if(net == nullptr) + FatalError("Cant add a layer without a Net\n"); + + // padding compute + if(f.pad == 1) { + f.padding_x = f.padding_y = f.size_x /2; + } + //std::cout<<"Add layer: "<= netLayers.size()) FatalError("impossible to shortcut\n"); + //std::cout<<"shortcut to "<getLayerName()<<"\n"; + netLayers.push_back(new tk::dnn::Shortcut(net, netLayers[layerIdx])); + + } else if(f.type == "upsample") { + netLayers.push_back(new tk::dnn::Upsample(net, f.stride_x)); + + } else if(f.type == "route") { + if(f.layers.size() == 0) FatalError("no layers to Route\n"); + std::vector layers; + for(int i=0; i= netLayers.size()) FatalError("impossible to route\n"); + //std::cout<<"Route to "<getLayerName()<<"\n"; + layers.push_back(netLayers[layerIdx]); + } + netLayers.push_back(new tk::dnn::Route(net, layers.data(), layers.size(), f.groups, f.group_id)); + + } else if(f.type == "reorg") { + netLayers.push_back(new tk::dnn::Reorg(net, f.stride_x)); + + } else if(f.type == "region") { + netLayers.push_back(new tk::dnn::Region(net, f.classes, f.coords, f.num)); + + } else if(f.type == "yolo") { + std::string wgs = wgs_path + "/g" + std::to_string(netLayers.size()) + ".bin"; + //printf("%d %d %s %d %f\n", f.classes, f.num/f.n_mask, wgs.c_str(), f.n_mask, f.scale_xy); + tk::dnn::Yolo *l = new tk::dnn::Yolo(net, f.classes, f.num/f.n_mask, wgs, f.n_mask, f.scale_xy, f.nms_thresh, (tk::dnn::Yolo::nmsKind_t) f.nms_kind, f.new_coords); + if(names.size() != f.classes) + FatalError("Mismatch between number of classes and names"); + l->classesNames = names; + netLayers.push_back(l); + + } else{ + FatalError("layer not supported: " + f.type); + } + + // add activation + if(netLayers.size() > 0 && f.activation != "linear") { + tkdnnActivationMode_t act; + if(f.activation == "relu") act = tkdnnActivationMode_t(CUDNN_ACTIVATION_RELU); + else if(f.activation == "leaky") act = tk::dnn::ACTIVATION_LEAKY; + else if(f.activation == "mish") act = tk::dnn::ACTIVATION_MISH; + else if(f.activation == "logistic") act = tk::dnn::ACTIVATION_LOGISTIC; + else { FatalError("activation not supported: " + f.activation); } + netLayers[netLayers.size()-1] = new tk::dnn::Activation(net, act); + }; + } + + std::vector darknetReadNames(const std::string& names_file){ + std::ifstream if_names(names_file); + if(!if_names.is_open()) + FatalError("cloud not open names file: " + names_file); + + std::vector names; + std::string line; + while(std::getline(if_names, line)) + if(line != "") + names.push_back(line); + + if_names.close(); + return names; + } + + tk::dnn::Network* darknetParser(const std::string& cfg_file, const std::string& wgs_path, const std::string& names_file) { + + tk::dnn::Network *net = nullptr; + + // layers without activations to retrieve correct id number + std::vector netLayers; + + std::ifstream if_cfg(cfg_file); + if(!if_cfg.is_open()) + FatalError("cloud not open cfg file: " + cfg_file); + + std::vector names = darknetReadNames(names_file); + + darknetFields_t fields; // will be filled with layers fields + std::string line; + while(std::getline(if_cfg, line)) { + // remove comments + std::size_t found = line.find("#"); + if ( found != std::string::npos ) { + line = line.substr(0, found); + } + + // skip empty lines + if(line.size() == 0) + continue; + + std::string type = darknetParseType(line); + if(type.size() > 0) { + // end of filled type + if(fields.type != "") { + if(fields.type == "net") + net = darknetAddNet(fields); + else + darknetAddLayer(net, fields, wgs_path, netLayers, names); + } + + // new type + //std::cout<<"type: "<tensorFormat, net->dataType, @@ -22,25 +26,27 @@ void DeformConv2d::initCUDNN() { const int dim_ones = preconv->input_dim.c * this->kernelH * this->kernelW * 1 * height_ones * width_ones; int dst_dim = preconv->output_dim.tot(); - if (dst_dim % 3 != 0 ) - std::cout<<"take attention\n\n"; + if( dst_dim % 3 != 0 ) + FatalError("DeformConv2d: the Conv2d output is not divisible by three"); chunk_dim = dst_dim/3; checkCuda( cudaMalloc(&offset, 2*chunk_dim*sizeof(dnnType))); checkCuda( cudaMalloc(&mask, chunk_dim*sizeof(dnnType))); // kernel ones - checkCuda( cudaMalloc(&ones_d1, (height_ones*width_ones)*sizeof(dnnType)) ); - float aus1[height_ones*width_ones]; + dnnType *ones_h1; + checkCuda( cudaMallocHost(&ones_h1, (height_ones*width_ones)*sizeof(dnnType)) ); for(int i=0; igetOutputDim().c, out_ch, kernelH, kernelW, 1, - d_fname_weights, batchnorm, true){ - + d_fname_weights, batchnorm, true) { this->out_ch = out_ch; this->deformableGroup = deformable_group; this->kernelH = kernelH; @@ -73,36 +78,37 @@ DeformConv2d::DeformConv2d( Network *net, int out_ch, int deformable_group, int } DeformConv2d::~DeformConv2d() { - checkCUDNN( cudnnDestroyTensorDescriptor(biasTensorDesc) ); checkCuda( cudaFree(dstData) ); - checkCuda( cudaFreeHost(ones_d1) ); - checkCuda( cudaFreeHost(ones_d2) ); + checkCuda( cudaFree(ones_d1) ); + checkCuda( cudaFree(ones_d2) ); checkCuda( cudaFree(offset) ); checkCuda( cudaFree(mask) ); checkCuda( cudaFree(output_conv) ); + cublasDestroy(handle); } dnnType* DeformConv2d::infer(dataDim_t &dim, dnnType* srcData) { // conv2d output_conv = preconv->infer(dim, srcData); - // split conv2d outputs into offset to mask + // split conv2d outputs into offset and mask checkCuda(cudaMemcpy(offset, output_conv, 2*chunk_dim*sizeof(dnnType), cudaMemcpyDeviceToDevice)); checkCuda(cudaMemcpy(mask, output_conv + 2*chunk_dim, chunk_dim*sizeof(dnnType), cudaMemcpyDeviceToDevice)); - // kernel sigmoide + // kernel sigmoid activationSIGMOIDForward(mask, mask, chunk_dim); - + // deformable convolution - dcn_v2_cuda_forward(srcData, this->data_d, + dcnV2CudaForward(stat, handle, + srcData, this->data_d, this->bias2_d, ones_d1, - offset, mask, + offset, mask, dstData, ones_d2, this->kernelH, this->kernelW, this->strideH, this->strideW, this->paddingH, this->paddingW, 1, 1, - this->deformableGroup, + this->deformableGroup, 0, //batch_id for cudnn is set to 0 (no batch) preconv->input_dim.n, preconv->input_dim.c, preconv->input_dim.h, preconv->input_dim.w, this->output_dim.n, this->output_dim.c, this->output_dim.h, this->output_dim.w, chunk_dim); diff --git a/src/Dense.cpp b/src/Dense.cpp index b6a9af2..4371d06 100644 --- a/src/Dense.cpp +++ b/src/Dense.cpp @@ -37,7 +37,7 @@ dnnType* Dense::infer(dataDim_t &dim, dnnType* srcData) { // place bias into dstData checkCuda( cudaMemcpy(dstData, bias_d, dim_y*sizeof(dnnType), cudaMemcpyDeviceToDevice) ); - //do matrix moltiplication + //do matrix multiplication checkERROR( cublasSgemv(net->cublasHandle, CUBLAS_OP_T, dim_x, dim_y, &alpha, diff --git a/src/Int8BatchStream.cpp b/src/Int8BatchStream.cpp new file mode 100644 index 0000000..fdc1db2 --- /dev/null +++ b/src/Int8BatchStream.cpp @@ -0,0 +1,165 @@ +#include "Int8BatchStream.h" + +#include +#include +#include +#include + +BatchStream::BatchStream(tk::dnn::dataDim_t dim, int batchSize, int maxBatches, const std::string& fileimglist, const std::string& filelabellist) { + mBatchSize = batchSize; + mMaxBatches = maxBatches; + mDims = nvinfer1::DimsNCHW{ dim.n, dim.c, dim.h, dim.w }; + mHeight = dim.h; + mWidth = dim.w; + mImageSize = mDims.c()*mDims.h()*mDims.w(); + mBatch.resize(mBatchSize*mImageSize, 0); + mLabels.resize(mBatchSize, 0); + mFileBatch.resize(mDims.n()*mImageSize, 0); + mFileLabels.resize(mDims.n(), 0); + mFileImgList = fileimglist; + readInListFile(fileimglist, mListImg); + mFileLabelList = filelabellist; + readInListFile(filelabellist, mListLabel); + + reset(0); +} + +void BatchStream::reset(int firstBatch) { + mBatchCount = 0; + mFileCount = 0; + mFileBatchPos = mDims.n(); + skip(firstBatch); +} + +bool BatchStream::next() { + std::cout<<"Next batch: "< 0 && mFileBatchPos <= mDims.n()); + if (mFileBatchPos == mDims.n() && !update()) + return false; + + csize = std::min(mBatchSize - batchPos, mDims.n() - mFileBatchPos); + std::copy_n(getFileBatch() + mFileBatchPos * mImageSize, csize * mImageSize, getBatch() + batchPos * mImageSize); + std::copy_n(getFileLabels() + mFileBatchPos, csize, getLabels() + batchPos); + } + mBatchCount++; + return true; +} + +void BatchStream::skip(int skipCount) { + if (mBatchSize >= mDims.n() && mBatchSize%mDims.n() == 0 && mFileBatchPos == mDims.n()) { + mFileCount += skipCount * mBatchSize / mDims.n(); + return; + } + + int x = mBatchCount; + for (int i = 0; i < skipCount; i++) + next(); + mBatchCount = x; +} + +void BatchStream::readInListFile(const std::string& dataFilePath, std::vector& mListIn) { + // dataFilePath contains the list of image paths + int count = 0; + FILE* f = fopen(dataFilePath.c_str(), "r"); + if (!f) + FatalError("failed to open " + dataFilePath); + + char str[512]; + while (fgets(str, 512, f) != NULL) { + for (int i = 0; str[i] != '\0'; ++i) { + if (str[i] == '\n'){ + str[i] = '\0'; + break; + } + } + count ++; + mListIn.push_back(str); + if(count == mMaxBatches) + break; + } + fclose(f); +} + +void BatchStream::readCVimage(std::string inputFileName, std::vector& res, bool fixshape) { + // unaltered original DsImage + cv::Mat m_OrigImage; + // letterboxed DsImage given to the network as input + cv::Mat m_LetterboxImage; + m_OrigImage = cv::imread(inputFileName, cv::IMREAD_COLOR); + + if (!m_OrigImage.data || m_OrigImage.cols <= 0 || m_OrigImage.rows <= 0) + FatalError("Unable to open " + inputFileName); + + int m_Height = m_OrigImage.rows; + int m_Width = m_OrigImage.cols; + if(fixshape) { + m_Height = mHeight; + m_Width = mWidth; + } + std::cout<<"image is "<(resizeH) / static_cast(m_Height); + + // Additional checks for images with non even dims + if ((m_Width - resizeW) % 2) resizeW--; + if ((m_Height - resizeH) % 2) resizeH--; + assert((m_Width - resizeW) % 2 == 0); + assert((m_Height - resizeH) % 2 == 0); + + int m_XOffset = (m_Width - resizeW) / 2; + int m_YOffset = (m_Height - resizeH) / 2; + + assert(2 * m_XOffset + resizeW == m_Width); + assert(2 * m_YOffset + resizeH == m_Height); + + // resizing + cv::resize(m_OrigImage, m_LetterboxImage, cv::Size(resizeW, resizeH), 0, 0, cv::INTER_CUBIC); + // letterboxing + cv::copyMakeBorder(m_LetterboxImage, m_LetterboxImage, m_YOffset, m_YOffset, m_XOffset, + m_XOffset, cv::BORDER_CONSTANT, cv::Scalar(128, 128, 128)); + m_LetterboxImage.convertTo(m_LetterboxImage, CV_32FC3, 1 / 255.0); + // converting to RGB and NCHW format + m_LetterboxImage = cv::dnn::blobFromImage(m_LetterboxImage); + res.assign(m_LetterboxImage.begin(), m_LetterboxImage.end()); +} + +void BatchStream::readLabels(std::string inputFileName, std::vector& ris) { + std::ifstream is(inputFileName.c_str()); + + std::string line; + while (std::getline(is, line)) + { + std::istringstream iss(line); + float val; + if(!(iss >> val)) { break; } // error + ris.push_back(val); + } +} + +bool BatchStream::update() { + std::string imgFileName = mListImg[mFileCount]; + std::string labelFileName = mListLabel[mFileCount]; + mFileCount++; + + //read image + mFileBatch.clear(); + readCVimage(imgFileName, mFileBatch); + // std::transform( + // singleImg_rawData.begin(), singleImg_rawData.end(), mFileBatch.begin(), [](uint8_t val) { return static_cast(val); }); + + //read label + mFileLabels.clear(); + readLabels(labelFileName, mFileLabels); + // std::transform( + // singleLabels_rawData.begin(), singleLabels_rawData.end(), mFileLabels.begin(), [](uint8_t val) { return static_cast(val); }); + + mFileBatchPos = 0; + return true; +} diff --git a/src/Int8Calibrator.cpp b/src/Int8Calibrator.cpp new file mode 100644 index 0000000..773a9d8 --- /dev/null +++ b/src/Int8Calibrator.cpp @@ -0,0 +1,46 @@ +#include "Int8Calibrator.h" + +Int8EntropyCalibrator::Int8EntropyCalibrator(BatchStream& stream, int firstBatch, + const std::string& calibTableFilePath, + const std::string& inputBlobName, + bool readCache): + mStream(stream), + mCalibTableFilePath(calibTableFilePath), + mInputBlobName(inputBlobName.c_str()), + mReadCache(readCache) { + nvinfer1::DimsNCHW dims = mStream.getDims(); + mInputCount = mStream.getBatchSize() * dims.c() * dims.h() * dims.w(); + checkCuda(cudaMalloc(&mDeviceInput, mInputCount * sizeof(float))); + mStream.reset(firstBatch); +} + +bool Int8EntropyCalibrator::getBatch(void* bindings[], const char* names[], int nbBindings) { + if (!mStream.next()) + return false; + + checkCuda(cudaMemcpy(mDeviceInput, mStream.getBatch(), mInputCount * sizeof(float), cudaMemcpyHostToDevice)); + assert(!strcmp(names[0], mInputBlobName.c_str())); + bindings[0] = mDeviceInput; + return true; +} + +const void* Int8EntropyCalibrator::readCalibrationCache(size_t& length) { + mCalibrationCache.clear(); + assert(!mCalibTableFilePath.empty()); + std::ifstream input(mCalibTableFilePath, std::ios::binary); + input >> std::noskipws; + input >> std::noskipws; + if (mReadCache && input.good()) + std::copy(std::istream_iterator(input), std::istream_iterator(), + std::back_inserter(mCalibrationCache)); + + length = mCalibrationCache.size(); + return length ? &mCalibrationCache[0] : nullptr; +} + +void Int8EntropyCalibrator::writeCalibrationCache(const void* cache, size_t length) { + assert(!mCalibTableFilePath.empty()); + std::ofstream output(mCalibTableFilePath, std::ios::binary); + output.write(reinterpret_cast(cache), length); + output.close(); +} \ No newline at end of file diff --git a/src/LSTM.cpp b/src/LSTM.cpp new file mode 100644 index 0000000..d0429e0 --- /dev/null +++ b/src/LSTM.cpp @@ -0,0 +1,336 @@ +#include + +#include "Layer.h" + +namespace tk { namespace dnn { + +LSTM::LSTM( Network *net, int hiddensize, bool returnSeq, std::string fname_weights) : + Layer(net) { + + this->returnSeq = returnSeq; + int batchSize = input_dim.n; + int inputSize = input_dim.c; + seqLen = input_dim.w; + stateSize = hiddensize; + + // init Tensor Descriptors + std::vector x_vec(seqLen); + std::vector y_vec(seqLen); + + int dimA[3]; + int strideA[3]; + for (int i = 0; i < seqLen; i++) { + checkCUDNN(cudnnCreateTensorDescriptor(&x_vec[i])); + checkCUDNN(cudnnCreateTensorDescriptor(&y_vec[i])); + + dimA[0] = batchSize; + dimA[1] = inputSize; + dimA[2] = 1; + dimA[0] = batchSize; + dimA[1] = inputSize; + strideA[0] = dimA[2] * dimA[1]; + strideA[1] = dimA[2]; + strideA[2] = 1; + checkCUDNN(cudnnSetTensorNdDescriptor(x_vec[i], + net->dataType, 3, dimA, strideA)); + + dimA[0] = batchSize; + dimA[1] = stateSize; + dimA[2] = 1; + strideA[0] = dimA[2] * dimA[1]; + strideA[1] = dimA[2]; + strideA[2] = 1; + checkCUDNN(cudnnSetTensorNdDescriptor(y_vec[i], + net->dataType, 3, dimA, strideA)); + } + // apply tensordesc + x_desc_vec_ = x_vec; + y_desc_vec_ = y_vec; + + + // set the state tensors + dimA[0] = numLayers; + dimA[1] = batchSize; + dimA[2] = stateSize; + strideA[0] = dimA[2] * dimA[1]; + strideA[1] = dimA[2]; + strideA[2] = 1; + checkCUDNN(cudnnCreateTensorDescriptor(&hx_desc_)); + checkCUDNN(cudnnCreateTensorDescriptor(&cx_desc_)); + checkCUDNN(cudnnCreateTensorDescriptor(&hy_desc_)); + checkCUDNN(cudnnCreateTensorDescriptor(&cy_desc_)); + checkCUDNN(cudnnSetTensorNdDescriptor(hx_desc_, net->dataType, 3, dimA, strideA)); + checkCUDNN(cudnnSetTensorNdDescriptor(cx_desc_, net->dataType, 3, dimA, strideA)); + checkCUDNN(cudnnSetTensorNdDescriptor(hy_desc_, net->dataType, 3, dimA, strideA)); + checkCUDNN(cudnnSetTensorNdDescriptor(cy_desc_, net->dataType, 3, dimA, strideA)); + // allocate dnnType *hx_ptr, *cx_ptr, *hy_ptr, *cy_ptr; + stateDataDim = dimA[0]*dimA[1]*dimA[2]; + checkCuda( cudaMalloc(&hx_ptr, stateDataDim*sizeof(dnnType)) ); + checkCuda( cudaMalloc(&cx_ptr, stateDataDim*sizeof(dnnType)) ); + checkCuda( cudaMalloc(&hy_ptr, stateDataDim*sizeof(dnnType)) ); + checkCuda( cudaMalloc(&cy_ptr, stateDataDim*sizeof(dnnType)) ); + + + + // Create Dropout descriptors // TODO: ??? IS IT NECESSARY ??? + float dropoutprob = 0.1f; // random val ???? + checkCUDNN(cudnnCreateDropoutDescriptor(&dropoutDesc)); + checkCUDNN(cudnnDropoutGetStatesSize(net->cudnnHandle, &dropout_byte_)); + dropout_size_ = dropout_byte_ / sizeof(dnnType); + checkCuda( cudaMalloc(&dropout_states_, dropout_byte_) ); + uint64_t seed_ = 17 + rand() % 4096; // NOLINT(runtime/threadsafe_fn) + checkCUDNN(cudnnSetDropoutDescriptor(dropoutDesc, + net->cudnnHandle, dropoutprob, dropout_states_, dropout_byte_, seed_)); + + + // RNN descriptors + checkCUDNN(cudnnCreateRNNDescriptor(&rnnDesc)); + +#if CUDNN_MAJOR > 7 + checkCUDNN(cudnnSetRNNDescriptor_v6(net->cudnnHandle,rnnDesc, stateSize, numLayers, dropoutDesc, + cudnnRNNInputMode_t::CUDNN_LINEAR_INPUT, + //(bidirectional ? cudnnDirectionMode_t::CUDNN_BIDIRECTIONAL : cudnnDirectionMode_t::CUDNN_UNIDIRECTIONAL), + cudnnDirectionMode_t::CUDNN_UNIDIRECTIONAL, + cudnnRNNMode_t::CUDNN_LSTM, + cudnnRNNAlgo_t::CUDNN_RNN_ALGO_STANDARD, + net->dataType)); +#else + checkCUDNN(cudnnSetRNNDescriptor(net->cudnnHandle,rnnDesc, stateSize, numLayers, dropoutDesc, + cudnnRNNInputMode_t::CUDNN_LINEAR_INPUT, + //(bidirectional ? cudnnDirectionMode_t::CUDNN_BIDIRECTIONAL : cudnnDirectionMode_t::CUDNN_UNIDIRECTIONAL), + cudnnDirectionMode_t::CUDNN_UNIDIRECTIONAL, + cudnnRNNMode_t::CUDNN_LSTM, + cudnnRNNAlgo_t::CUDNN_RNN_ALGO_STANDARD, + net->dataType)); +#endif + + + // Get temp space sizes + checkCUDNN(cudnnGetRNNWorkspaceSize(net->cudnnHandle, + rnnDesc, seqLen, x_desc_vec_.data(), &workspace_byte_)); + workspace_size_ = workspace_byte_ / sizeof(dnnType); + checkCuda( cudaMalloc(&work_space_, workspace_byte_) ); + + + // Check that number of params are correct + size_t cudnn_param_size; + checkCUDNN(cudnnGetRNNParamsSize(net->cudnnHandle, + rnnDesc,x_desc_vec_[0], &cudnn_param_size, net->dataType)); + int cudnn_params = cudnn_param_size/sizeof(dnnType); + //std::cout<<"LSTM params size: "<dataType, net->tensorFormat, 3, dim_w)); + + // load params + std::cout<<"Reading weights: PARAMS="<cudnnHandle, rnnDesc, + i, x_desc_vec_[0], w_desc_, 0, j, m_desc, (void**)&p)); + + std::cout << "ptr: " << ((int64_t)(p - NULL))/sizeof(dnnType)<<"\n"; + + cudnnDataType_t t; + cudnnTensorFormat_t f; + int ndim = 5; + int dims[5] = {0, 0, 0, 0, 0}; + checkCUDNN(cudnnGetFilterNdDescriptor(m_desc, ndim, &t, &f, &ndim, &dims[0])); + std::cout << "(layer, linlayer): " << i << " " << j << "\n"; + + int tot = 1; + for (int i = 0; i < ndim; ++i) { + std::cout << dims[i] << " "; + tot *= dims[i]; + } + std::cout<<"\t-> "<cublasHandle, srcData, srcF, dim.c, dim.h*dim.w*dim.l); + + // build srcB as reversed srcF + for(int i=0; icudnnHandle, + rnnDesc, + seqLen, // number of time steps (nT) + x_desc_vec_.data(), // input array of desc (nT*nC_in) + srcF, // input pointer + hx_desc_, // initial hidden state desc + hx_ptr, // initial hidden state pointer + cx_desc_, // initial cell state desc + cx_ptr, // initial cell state pointer + w_desc_, // weights desc + wf_ptr, // weights pointer + y_desc_vec_.data(), // output desc (nT*nC_out) + dstF, // output pointer + hy_desc_, // final hidden state desc + hy_ptr, // final hidden state pointer + cy_desc_, // final cell state desc + cy_ptr, // final cell state pointer + work_space_, // workspace pointer + workspace_byte_)); // workspace size + } + + // backward + { + // reset states + checkCuda( cudaMemset(hx_ptr, 0, stateDataDim*sizeof(float)) ); + checkCuda( cudaMemset(cx_ptr, 0, stateDataDim*sizeof(float)) ); + + checkCUDNN(cudnnRNNForwardInference(net->cudnnHandle, + rnnDesc, + seqLen, // number of time steps (nT) + x_desc_vec_.data(), // input array of desc (nT*nC_in) + srcB, // input pointer + hx_desc_, // initial hidden state desc + hx_ptr, // initial hidden state pointer + cx_desc_, // initial cell state desc + cx_ptr, // initial cell state pointer + w_desc_, // weights desc + wb_ptr, // weights pointer + y_desc_vec_.data(), // output desc (nT*nC_out) + dstB_NR, // output pointer + hy_desc_, // final hidden state desc + hy_ptr, // final hidden state pointer + cy_desc_, // final cell state desc + cy_ptr, // final cell state pointer + work_space_, // workspace pointer + workspace_byte_)); // workspace size + } + + + // reverse order of dstB + for(int i=0; icublasHandle, dstF, dstData, + one_output_dim.h* one_output_dim.w*one_output_dim.l, one_output_dim.c); + // backward transpose + matrixTranspose(net->cublasHandle, dstB, dstData + one_output_dim.tot(), + one_output_dim.h* one_output_dim.w*one_output_dim.l, one_output_dim.c); + } else { + // copy last of forward + checkCuda( cudaMemcpy(dstData, dstF + one_output_dim.tot() - one_output_dim.c, + one_output_dim.c*sizeof(dnnType), cudaMemcpyDeviceToDevice)); + // copy first of backward + checkCuda( cudaMemcpy(dstData + one_output_dim.c, dstB, + one_output_dim.c*sizeof(dnnType), cudaMemcpyDeviceToDevice)); + } + + dim = output_dim; + return dstData; +} + +}} diff --git a/src/Layer.cpp b/src/Layer.cpp index 57ee000..a355b90 100644 --- a/src/Layer.cpp +++ b/src/Layer.cpp @@ -4,10 +4,10 @@ namespace tk { namespace dnn { -Layer::Layer(Network *net, bool final) { +Layer::Layer(Network *net) { this->net = net; - this->final = final; + this->final = false; if(net != nullptr) { this->input_dim = net->getOutputDim(); this->output_dim = input_dim; @@ -24,6 +24,11 @@ Layer::~Layer() { checkCUDNN( cudnnDestroyTensorDescriptor(srcTensorDesc) ); checkCUDNN( cudnnDestroyTensorDescriptor(dstTensorDesc) ); + + if(dstData != nullptr) { + cudaFree(dstData); + dstData = nullptr; + } } }} \ No newline at end of file diff --git a/src/LayerWgs.cpp b/src/LayerWgs.cpp index a18fd53..a761327 100644 --- a/src/LayerWgs.cpp +++ b/src/LayerWgs.cpp @@ -8,32 +8,33 @@ namespace tk { namespace dnn { LayerWgs::LayerWgs(Network *net, int inputs, int outputs, int kh, int kw, int kl, - std::string fname_weights, bool batchnorm, bool additional_bias, bool final) : Layer(net, final) { - + std::string fname_weights, bool batchnorm, bool additional_bias, bool deConv, int groups) : Layer(net) { + inputs = inputs/groups; + this->inputs = inputs; this->outputs = outputs; this->weights_path = std::string(fname_weights); std::cout<<"Reading weights: I="<dontLoadWeights); + readBinaryFile(weights_path.c_str(), inputs*outputs*kh*kw*kl, &data_h, &data_d, seek); seek += inputs*outputs*kh*kw*kl; this->additional_bias = additional_bias; if(additional_bias) { - readBinaryFile(weights_path.c_str(), outputs, &bias2_h, &bias2_d, seek, net->dontLoadWeights); + readBinaryFile(weights_path.c_str(), outputs, &bias2_h, &bias2_d, seek); seek += outputs; } - readBinaryFile(weights_path.c_str(), outputs, &bias_h, &bias_d, seek, net->dontLoadWeights); + readBinaryFile(weights_path.c_str(), outputs, &bias_h, &bias_d, seek); this->batchnorm = batchnorm; if(batchnorm) { seek += outputs; - readBinaryFile(weights_path.c_str(), outputs, &scales_h, &scales_d, seek, net->dontLoadWeights); + readBinaryFile(weights_path.c_str(), outputs, &scales_h, &scales_d, seek); seek += outputs; - readBinaryFile(weights_path.c_str(), outputs, &mean_h, &mean_d, seek, net->dontLoadWeights); + readBinaryFile(weights_path.c_str(), outputs, &mean_h, &mean_d, seek); seek += outputs; - readBinaryFile(weights_path.c_str(), outputs, &variance_h, &variance_d, seek, net->dontLoadWeights); + readBinaryFile(weights_path.c_str(), outputs, &variance_h, &variance_d, seek); float eps = TKDNN_BN_MIN_EPSILON; @@ -58,6 +59,14 @@ LayerWgs::LayerWgs(Network *net, int inputs, int outputs, float2half(data_d, data16_d, w_size); cudaMemcpy(data16_h, data16_d, w_size*sizeof(__half), cudaMemcpyDeviceToHost); + if(additional_bias){ + int b2_size = outputs; + bias216_h = new __half[b2_size]; + cudaMalloc(&bias216_d, w_size*sizeof(__half)); + float2half(bias2_d, bias216_d, b2_size); + cudaMemcpy(bias216_h, bias216_d, b2_size*sizeof(__half), cudaMemcpyDeviceToHost); + } + int b_size = outputs; bias16_h = new __half[b_size]; cudaMalloc(&bias16_d, w_size*sizeof(__half)); @@ -86,7 +95,6 @@ LayerWgs::LayerWgs(Network *net, int inputs, int outputs, cudaMemcpy(power16_h, power16_d, b_size*sizeof(__half), cudaMemcpyDeviceToHost); //mean array - cudaMemcpy(tmp_d, mean_h, b_size*sizeof(float), cudaMemcpyHostToDevice); float2half(tmp_d, mean16_d, b_size); cudaMemcpy(mean16_h, mean16_d, b_size*sizeof(__half), cudaMemcpyDeviceToHost); @@ -97,27 +105,17 @@ LayerWgs::LayerWgs(Network *net, int inputs, int outputs, float2half(tmp_d, variance16_d, b_size); cudaMemcpy(variance16_h, variance16_d, b_size*sizeof(__half), cudaMemcpyDeviceToHost); - //conver scales + //convert scales float2half(scales_d, scales16_d, b_size); cudaMemcpy(scales16_h, scales16_d, b_size*sizeof(__half), cudaMemcpyDeviceToHost); + + cudaFree(tmp_d); } } LayerWgs::~LayerWgs() { - - delete [] data_h; - delete [] bias_h; - checkCuda( cudaFree(data_d) ); - checkCuda( cudaFree(bias_d) ); - - if(batchnorm) { - delete [] scales_h; - delete [] mean_h; - delete [] variance_h; - checkCuda( cudaFree(scales_d) ); - checkCuda( cudaFree(mean_d) ); - checkCuda( cudaFree(variance_d) ); - } + releaseHost(); + releaseDevice(); } }} diff --git a/src/MobilenetDetection.cpp b/src/MobilenetDetection.cpp new file mode 100644 index 0000000..3c54e28 --- /dev/null +++ b/src/MobilenetDetection.cpp @@ -0,0 +1,312 @@ +#include "MobilenetDetection.h" + +bool boxProbCmp(const tk::dnn::box &a, const tk::dnn::box &b){ + return (a.prob > b.prob); +} + +namespace tk{ namespace dnn{ + +void MobilenetDetection::generate_ssd_priors(const SSDSpec *specs, const int n_specs, bool clamp){ + nPriors = 0; + for (int i = 0; i < n_specs; i++){ + nPriors += specs[i].featureSize * specs[i].featureSize * 6; + } + + priors = (float *)malloc(N_COORDS * nPriors * sizeof(float)); + + int i_prio = 0; + float scale, x_center, y_center, h, w, size, ratio; + int min, max; + for (int i = 0; i < n_specs; i++){ + scale = (float)imageSize / (float)specs[i].shrinkage; + min = specs[i].boxHeight > specs[i].boxWidth ? specs[i].boxWidth : specs[i].boxHeight; + max = specs[i].boxHeight < specs[i].boxWidth ? specs[i].boxWidth : specs[i].boxHeight; + for (int j = 0; j < specs[i].featureSize; j++){ + for (int k = 0; k < specs[i].featureSize; k++){ + //small sized square box + size = min; + x_center = (k + 0.5f) / scale; + y_center = (j + 0.5f) / scale; + h = w = (float)size / (float)imageSize; + + priors[i_prio * N_COORDS + 0] = x_center; + priors[i_prio * N_COORDS + 1] = y_center; + priors[i_prio * N_COORDS + 2] = w; + priors[i_prio * N_COORDS + 3] = h; + ++i_prio; + + //big sized square box + size = sqrt(max * min); + h = w = (float)size / (float)imageSize; + + priors[i_prio * N_COORDS + 0] = x_center; + priors[i_prio * N_COORDS + 1] = y_center; + priors[i_prio * N_COORDS + 2] = w; + priors[i_prio * N_COORDS + 3] = h; + ++i_prio; + + //change h/w ratio of the small sized box + size = min; + h = w = size / (float)imageSize; + ratio = sqrt(specs[i].ratio1); + priors[i_prio * N_COORDS + 0] = x_center; + priors[i_prio * N_COORDS + 1] = y_center; + priors[i_prio * N_COORDS + 2] = w * ratio; + priors[i_prio * N_COORDS + 3] = h / ratio; + ++i_prio; + + priors[i_prio * N_COORDS + 0] = x_center; + priors[i_prio * N_COORDS + 1] = y_center; + priors[i_prio * N_COORDS + 2] = w / ratio; + priors[i_prio * N_COORDS + 3] = h * ratio; + ++i_prio; + + ratio = sqrt(specs[i].ratio2); + priors[i_prio * N_COORDS + 0] = x_center; + priors[i_prio * N_COORDS + 1] = y_center; + priors[i_prio * N_COORDS + 2] = w * ratio; + priors[i_prio * N_COORDS + 3] = h / ratio; + ++i_prio; + + priors[i_prio * N_COORDS + 0] = x_center; + priors[i_prio * N_COORDS + 1] = y_center; + priors[i_prio * N_COORDS + 2] = w / ratio; + priors[i_prio * N_COORDS + 3] = h * ratio; + ++i_prio; + } + } + } + + if (clamp){ + for (int i = 0; i < nPriors * N_COORDS; i++){ + priors[i] = priors[i] > 1.0f ? 1.0f : priors[i]; + priors[i] = priors[i] < 0.0f ? 0.0f : priors[i]; + } + } +} + +void MobilenetDetection::convert_locatios_to_boxes_and_center(){ + float cur_x, cur_y; + for (int i = 0; i < nPriors; i++){ + locations_h[i * N_COORDS + 0] = locations_h[i * N_COORDS + 0] * centerVariance * priors[i * N_COORDS + 2] + priors[i * N_COORDS + 0]; + locations_h[i * N_COORDS + 1] = locations_h[i * N_COORDS + 1] * centerVariance * priors[i * N_COORDS + 3] + priors[i * N_COORDS + 1]; + locations_h[i * N_COORDS + 2] = exp(locations_h[i * N_COORDS + 2] * sizeVariance) * priors[i * N_COORDS + 2]; + locations_h[i * N_COORDS + 3] = exp(locations_h[i * N_COORDS + 3] * sizeVariance) * priors[i * N_COORDS + 3]; + + cur_x = locations_h[i * N_COORDS + 0]; + cur_y = locations_h[i * N_COORDS + 1]; + + locations_h[i * N_COORDS + 0] = cur_x - locations_h[i * N_COORDS + 2] / 2; + locations_h[i * N_COORDS + 1] = cur_y - locations_h[i * N_COORDS + 3] / 2; + locations_h[i * N_COORDS + 2] = cur_x + locations_h[i * N_COORDS + 2] / 2; + locations_h[i * N_COORDS + 3] = cur_y + locations_h[i * N_COORDS + 3] / 2; + } +} + +float MobilenetDetection::iou(const tk::dnn::box &a, const tk::dnn::box &b){ + float max_x = a.x > b.x ? a.x : b.x; + float max_y = a.y > b.y ? a.y : b.y; + float min_w = a.w < b.w ? a.w : b.w; + float min_h = a.h < b.h ? a.h : b.h; + + float ao_w = min_w - max_x > 0 ? min_w - max_x : 0; + float ao_h = min_h - max_y > 0 ? min_h - max_y : 0; + + float area_overlap = ao_w * ao_h; + float area_0_w = a.w - a.x > 0 ? a.w - a.x : 0; + float area_0_h = a.h - a.y > 0 ? a.h - a.y : 0; + + float area_1_w = b.w - b.x > 0 ? b.w - b.x : 0; + float area_1_h = b.h - b.y > 0 ? b.h - b.y : 0; + + float area_0 = area_0_h * area_0_w; + float area_1 = area_1_h * area_1_w; + + float iou = area_overlap / (area_0 + area_1 - area_overlap + 1e-5); + return iou; +} + +bool MobilenetDetection::init(const std::string& tensor_path, const int n_classes, const int n_batches, const float conf_thresh){ + std::cout<<(tensor_path).c_str()<<"\n"; + netRT = new tk::dnn::NetworkRT(NULL, (tensor_path).c_str()); + imageSize = netRT->input_dim.h; + classes = n_classes; + nBatches = n_batches; + confThreshold = conf_thresh; + + SSDSpec specs[N_SSDSPEC]; + + if(imageSize == 300){ + specs[0].setAll(19, 16, 60, 105, 2, 3); + specs[1].setAll(10, 32, 105, 150, 2, 3); + specs[2].setAll(5, 64, 150, 195, 2, 3); + specs[3].setAll(3, 100, 195, 240, 2, 3); + specs[4].setAll(2, 150, 240, 285, 2, 3); + specs[5].setAll(1, 300, 285, 330, 2, 3); + } + else if(imageSize == 512){ + specs[0].setAll(32, 16, 60, 105, 2, 3); + specs[1].setAll(16, 32, 105, 150, 2, 3); + specs[2].setAll(8, 64, 150, 195, 2, 3); + specs[3].setAll(4, 100, 195, 240, 2, 3); + specs[4].setAll(2, 150, 240, 285, 2, 3); + specs[5].setAll(1, 300, 285, 330, 2, 3); + } + else{ + FatalError("Input size for mobilenet not supported"); + } + + generate_ssd_priors(specs, N_SSDSPEC); + +#ifndef OPENCV_CUDACONTRIB + checkCuda(cudaMallocHost(&input, sizeof(dnnType) * netRT->input_dim.tot() * nBatches)); +#endif + checkCuda(cudaMalloc(&input_d, sizeof(dnnType) * netRT->input_dim.tot() * nBatches)); + + locations_h = (float *)malloc(N_COORDS * nPriors * sizeof(float)); + confidences_h = (float *)malloc(nPriors * classes * sizeof(float)); + + for (int c = 0; c < classes; c++){ + int offset = c * 123457 % classes; + float r = getColor(2, offset, classes); + float g = getColor(1, offset, classes); + float b = getColor(0, offset, classes); + colors[c] = cv::Scalar(int(255.0 * b), int(255.0 * g), int(255.0 * r)); + } + + if(classes == 11){ //BDD + const char *classes_names_[] = { + "person","car","truck","bus","motor","bike","rider","traffic light","traffic sign","train"}; + classesNames = std::vector(classes_names_, std::end(classes_names_)); + } + else if(classes == 21){ //VOC + const char *classes_names_[] = { + "aeroplane", "bicycle", "bird", "boat", "bottle", "bus", + "car", "cat", "chair", "cow", "diningtable", "dog", "horse", "motorbike", + "person", "pottedplant", "sheep", "sofa", "train", "tvmonitor"}; + classesNames = std::vector(classes_names_, std::end(classes_names_)); + + } + else if (classes == 81){ //COCO + const char *classes_names_[] = { + "person" , "bicycle" , "car" , "motorbike" , "aeroplane" , "bus" , + "train" , "truck" , "boat" , "traffic light" , "fire hydrant" , "stop sign" , + "parking meter" , "bench" , "bird" , "cat" , "dog" , "horse" , "sheep" , "cow" , + "elephant" , "bear" , "zebra" , "giraffe" , "backpack" , "umbrella" , "handbag" , + "tie" , "suitcase" , "frisbee" , "skis" , "snowboard" , "sports ball" , "kite" , + "baseball bat" , "baseball glove" , "skateboard" , "surfboard" , "tennis racket" , + "bottle" , "wine glass" , "cup" , "fork" , "knife" , "spoon" , "bowl" , "banana" , + "apple" , "sandwich" , "orange" , "broccoli" , "carrot" , "hot dog" , "pizza" , + "donut" , "cake" , "chair" , "sofa" , "pottedplant" , "bed" , "diningtable" , + "toilet" , "tvmonitor" , "laptop" , "mouse" , "remote" , "keyboard" , + "cell phone" , "microwave" , "oven" , "toaster" , "sink" , "refrigerator" , + "book" , "clock" , "vase" , "scissors" , "teddy bear" , "hair drier" , "toothbrush"}; + classesNames = std::vector(classes_names_, std::end(classes_names_)); + + } + else{ + FatalError("Number of classes not supported for mobilenet"); + } + return 1; +} + +void MobilenetDetection::preprocess(cv::Mat &frame, const int bi){ +#ifdef OPENCV_CUDACONTRIB + //move original image on GPU + cv::cuda::GpuMat orig_img, frame_nomean; + orig_img = cv::cuda::GpuMat(frame); + + //resize image, remove mean, divide by std + cv::cuda::resize (orig_img, orig_img, cv::Size(netRT->input_dim.w, netRT->input_dim.h)); + orig_img.convertTo(frame_nomean, CV_32FC3, 1, -127); + frame_nomean.convertTo(imagePreproc, CV_32FC3, 1 / 128.0, 0); + + //copy image into tensors + cv::cuda::split(imagePreproc, bgr); + + for(int i=0; i < netRT->input_dim.c; i++){ + int idx = i * imagePreproc.rows * imagePreproc.cols; + checkCuda( cudaMemcpy((void *)&input_d[idx + netRT->input_dim.tot()*bi], (void *)bgr[i].data, imagePreproc.rows * imagePreproc.cols* sizeof(float), cudaMemcpyDeviceToDevice) ); + } +#else + //resize image, remove mean, divide by std + cv::Mat frame_nomean; + resize(frame, frame, cv::Size(netRT->input_dim.w, netRT->input_dim.h)); + frame.convertTo(frame_nomean, CV_32FC3, 1, -127); + frame_nomean.convertTo(imagePreproc, CV_32FC3, 1 / 128.0, 0); + + //copy image into tensor and copy it into GPU + cv::split(imagePreproc, bgr); + for (int i = 0; i < netRT->input_dim.c; i++){ + int idx = i * imagePreproc.rows * imagePreproc.cols; + memcpy((void *)&input[idx + netRT->input_dim.tot()*bi], (void *)bgr[i].data, imagePreproc.rows * imagePreproc.cols * sizeof(dnnType)); + } + checkCuda(cudaMemcpyAsync(input_d+ netRT->input_dim.tot()*bi, input + netRT->input_dim.tot()*bi, netRT->input_dim.tot() * sizeof(dnnType), cudaMemcpyHostToDevice, netRT->stream)); +#endif +} + +void MobilenetDetection::postprocess(const int bi, const bool mAP){ + //get confidences and locations_h + dnnType *rt_out[2]; + rt_out[0] = (dnnType *)netRT->buffersRT[3]+ netRT->buffersDIM[3].tot()*bi; + rt_out[1] = (dnnType *)netRT->buffersRT[4]+ netRT->buffersDIM[4].tot()*bi; + + detected.clear(); + + checkCuda(cudaMemcpy(confidences_h, rt_out[0], nPriors * classes * sizeof(float), cudaMemcpyDeviceToHost)); + checkCuda(cudaMemcpy(locations_h, rt_out[1], N_COORDS * nPriors * sizeof(float), cudaMemcpyDeviceToHost)); + convert_locatios_to_boxes_and_center(); + + int width = originalSize[bi].width; + int height = originalSize[bi].height; + + float *conf_per_class; + for (int i = 1; i < classes; i++){ + conf_per_class = &confidences_h[i * nPriors]; + std::vector boxes; + for (int j = 0; j < nPriors; j++){ + + if (conf_per_class[j] > confThreshold){ + tk::dnn::box b; + b.cl = i; + b.prob = conf_per_class[j]; + b.x = locations_h[j * N_COORDS + 0]; + b.y = locations_h[j * N_COORDS + 1]; + b.w = locations_h[j * N_COORDS + 2]; + b.h = locations_h[j * N_COORDS + 3]; + + if(mAP) + for(int c=1; c remaining; + while (boxes.size() > 0){ + remaining.clear(); + + tk::dnn::box b; + b.cl = boxes[0].cl -1 ; //remove background class + b.prob = boxes[0].prob; + b.x = boxes[0].x * width; + b.y = boxes[0].y * height; + b.w = boxes[0].w * width - b.x; //convert from x1 to width + b.h = boxes[0].h * height - b.y; //convert from y1 to height + detected.push_back(b); + for (size_t j = 1; j < boxes.size(); j++){ + if (iou(boxes[0], boxes[j]) <= IoUThreshold){ + remaining.push_back(boxes[j]); + } + } + boxes = remaining; + } + } + batchDetected.push_back(detected); +} + + +} // namespace dnn +} // namespace tk \ No newline at end of file diff --git a/src/MulAdd.cpp b/src/MulAdd.cpp index 0c2a962..25cec8d 100644 --- a/src/MulAdd.cpp +++ b/src/MulAdd.cpp @@ -12,7 +12,7 @@ MulAdd::MulAdd(Network *net, dnnType mul, dnnType add) : Layer(net) { int size = input_dim.tot(); - // create a vector with all value setted to add + // create a vector with all value set to add dnnType *add_vector_h = new dnnType[size]; for(int i=0; iplatformHasFastFp16()<<"\n"; std::cout<<"Int8 support: "<platformHasFastInt8()<<"\n"; - //std::cout<<"DLAs: "<getNbDLACores()<<"\n"; +#if NV_TENSORRT_MAJOR >= 5 + std::cout<<"DLAs: "<getNbDLACores()<<"\n"; +#endif networkRT = builderRT->createNetwork(); - +#if NV_TENSORRT_MAJOR >= 6 + configRT = builderRT->createBuilderConfig(); +#endif + if(!fileExist(name)) { +#if NV_TENSORRT_MAJOR >= 6 + // Calibrator life time needs to last until after the engine is built. + std::unique_ptr calibrator; + configRT->setAvgTimingIterations(1); + configRT->setMinTimingIterations(1); + configRT->setMaxWorkspaceSize(1 << 30); + configRT->setFlag(BuilderFlag::kDEBUG); +#endif //input and dataType dataDim_t dim = net->layers[0]->input_dim; dtRT = DataType::kFLOAT; - builderRT->setMaxBatchSize(1); + builderRT->setMaxBatchSize(net->maxBatchSize); builderRT->setMaxWorkspaceSize(1 << 30); if(net->fp16 && builderRT->platformHasFastFp16()) { dtRT = DataType::kHALF; builderRT->setHalf2Mode(true); +#if NV_TENSORRT_MAJOR >= 6 + configRT->setFlag(BuilderFlag::kFP16); +#endif } - /* +#if NV_TENSORRT_MAJOR >= 5 if(net->dla && builderRT->getNbDLACores() > 0) { dtRT = DataType::kHALF; builderRT->setFp16Mode(true); @@ -59,9 +76,33 @@ NetworkRT::NetworkRT(Network *net, const char *name) { builderRT->setDefaultDeviceType(DeviceType::kDLA); builderRT->setDLACore(0); } - */ - - //add input layer +#endif +#if NV_TENSORRT_MAJOR >= 6 + if(net->int8 && builderRT->platformHasFastInt8()){ + // dtRT = DataType::kINT8; + // builderRT->setInt8Mode(true); + configRT->setFlag(BuilderFlag::kINT8); + BatchStream calibrationStream(dim, 1, 100, //TODO: check if 100 images are sufficient to the calibration (or 4951) + net->fileImgList, net->fileLabelList); + + /* The calibTableFilePath contains the path+filename of the calibration table. + * Each calibration table can be found in the corresponding network folder (../Test/*). + * Each network is located in a folder with the same name as the network. + * If the folder has a different name, the calibration table is saved in build/ folder. + */ + std::string calib_table_name = net->networkName + "/" + net->networkNameRT.substr(0, net->networkNameRT.find('.')) + "-calibration.table"; + std::string calib_table_path = net->networkName; + if(!fileExist((const char *)calib_table_path.c_str())) + calib_table_name = "./" + net->networkNameRT.substr(0, net->networkNameRT.find('.')) + "-calibration.table"; + + calibrator.reset(new Int8EntropyCalibrator(calibrationStream, 1, + calib_table_name, + "data")); + configRT->setInt8Calibrator(calibrator.get()); + } +#endif + + // add input layer ITensor *input = networkRT->addInput("data", DataType::kFLOAT, DimsCHW{ dim.c, dim.h, dim.w}); checkNULL(input); @@ -70,12 +111,18 @@ NetworkRT::NetworkRT(Network *net, const char *name) { for(int i=0; inum_layers; i++) { Layer *l = net->layers[i]; ILayer *Ilay = convert_layer(input, l); +#if NV_TENSORRT_MAJOR >= 6 + if(net->int8 && builderRT->platformHasFastInt8()) + { + Ilay->setPrecision(DataType::kINT8); + } +#endif Ilay->setName( (l->getLayerName() + std::to_string(i)).c_str() ); input = Ilay->getOutput(0); input->setName( (l->getLayerName() + std::to_string(i) + "_out").c_str() ); - if(l->getLayerType() == LAYER_YOLO || l->final) + if(l->final) networkRT->markOutput(*input); tensors[l] = input; } @@ -86,8 +133,15 @@ NetworkRT::NetworkRT(Network *net, const char *name) { input->setName("out"); networkRT->markOutput(*input); + std::cout<<"Selected maxBatchSize: "<getMaxBatchSize()<<"\n"; + printCudaMemUsage(); std::cout<<"Building tensorRT cuda engine...\n"; +#if NV_TENSORRT_MAJOR >= 6 + engineRT = builderRT->buildEngineWithConfig(*networkRT, *configRT); +#else engineRT = builderRT->buildCudaEngine(*networkRT); + //engineRT = std::shared_ptr(builderRT->buildCudaEngine(*networkRT)); +#endif if(engineRT == nullptr) FatalError("cloud not build cuda engine") // we don't need the network any more @@ -110,7 +164,7 @@ NetworkRT::NetworkRT(Network *net, const char *name) { // note that indices are guaranteed to be less than IEngine::getNbBindings() buf_input_idx = engineRT->getBindingIndex("data"); buf_output_idx = engineRT->getBindingIndex("out"); - std::cout<<"input idex = "< output index = "< output index = "<getBindingDimensions(buf_input_idx); @@ -130,9 +184,11 @@ NetworkRT::NetworkRT(Network *net, const char *name) { // create GPU buffers and a stream for(int i=0; igetNbBindings(); i++) { Dims dim = engineRT->getBindingDimensions(i); - checkCuda(cudaMalloc(&buffersRT[i], dim.d[0]*dim.d[1]*dim.d[2]*sizeof(dnnType))); + buffersDIM[i] = dataDim_t(1, dim.d[0], dim.d[1], dim.d[2]); + std::cout<<"RtBuffer "<getMaxBatchSize()*dim.d[0]*dim.d[1]*dim.d[2]*sizeof(dnnType))); } - checkCuda(cudaMalloc(&output, output_dim.tot()*sizeof(dnnType))); + checkCuda(cudaMalloc(&output, engineRT->getMaxBatchSize()*output_dim.tot()*sizeof(dnnType))); checkCuda(cudaStreamCreate(&stream)); } @@ -141,19 +197,24 @@ NetworkRT::~NetworkRT() { } dnnType* NetworkRT::infer(dataDim_t &dim, dnnType* data) { + int batches = dim.n; + if(batches > getMaxBatchSize()) { + FatalError("input batch size too large"); + } - checkCuda(cudaMemcpyAsync(buffersRT[buf_input_idx], data, input_dim.tot()*sizeof(dnnType), cudaMemcpyDeviceToDevice, stream)); - contextRT->enqueue(1, buffersRT, stream, nullptr); - checkCuda(cudaMemcpyAsync(output, buffersRT[buf_output_idx], output_dim.tot()*sizeof(dnnType), cudaMemcpyDeviceToDevice, stream)); - cudaStreamSynchronize(stream); + checkCuda(cudaMemcpyAsync(buffersRT[buf_input_idx], data, batches*input_dim.tot()*sizeof(dnnType), cudaMemcpyDeviceToDevice, stream)); + contextRT->enqueue(batches, buffersRT, stream, nullptr); + checkCuda(cudaMemcpyAsync(output, buffersRT[buf_output_idx], batches*output_dim.tot()*sizeof(dnnType), cudaMemcpyDeviceToDevice, stream)); + checkCuda(cudaStreamSynchronize(stream)); dim = output_dim; + dim.n = batches; return output; } -void NetworkRT::enqueue() { - contextRT->enqueue(1, buffersRT, stream, nullptr); +void NetworkRT::enqueue(int batchSize) { + contextRT->enqueue(batchSize, buffersRT, stream, nullptr); } ILayer* NetworkRT::convert_layer(ITensor *input, Layer *l) { @@ -166,12 +227,16 @@ ILayer* NetworkRT::convert_layer(ITensor *input, Layer *l) { return convert_layer(input, (Conv2d*) l); if(type == LAYER_POOLING) return convert_layer(input, (Pooling*) l); - if(type == LAYER_ACTIVATION) + if(type == LAYER_ACTIVATION || type == LAYER_ACTIVATION_CRELU || type == LAYER_ACTIVATION_LEAKY || type == LAYER_ACTIVATION_MISH || type == LAYER_ACTIVATION_LOGISTIC) return convert_layer(input, (Activation*) l); if(type == LAYER_SOFTMAX) return convert_layer(input, (Softmax*) l); if(type == LAYER_ROUTE) return convert_layer(input, (Route*) l); + if(type == LAYER_FLATTEN) + return convert_layer(input, (Flatten*) l); + if(type == LAYER_RESHAPE) + return convert_layer(input, (Reshape*) l); if(type == LAYER_REORG) return convert_layer(input, (Reorg*) l); if(type == LAYER_REGION) @@ -215,10 +280,11 @@ ILayer* NetworkRT::convert_layer(ITensor *input, Conv2d *l) { // printf("%d %d %d %d %d\n", l->kernelH, l->kernelW, l->inputs, l->outputs, l->batchnorm); - void *data_b, *bias_b, *power_b, *mean_b, *variance_b, *scales_b; + void *data_b, *bias_b, *bias2_b, *power_b, *mean_b, *variance_b, *scales_b; if(dtRT == DataType::kHALF) { data_b = l->data16_h; bias_b = l->bias16_h; + bias2_b = l->bias216_h; power_b = l->power16_h; mean_b = l->mean16_h; variance_b = l->variance16_h; @@ -226,6 +292,7 @@ ILayer* NetworkRT::convert_layer(ITensor *input, Conv2d *l) { } else { data_b = l->data_h; bias_b = l->bias_h; + bias2_b = l->bias2_h; power_b = l->power_h; mean_b = l->mean_h; variance_b = l->variance_h; @@ -237,8 +304,12 @@ ILayer* NetworkRT::convert_layer(ITensor *input, Conv2d *l) { Weights b; if(!l->batchnorm) b = { dtRT, bias_b, l->outputs}; - else - b = { dtRT, nullptr, 0}; //on batchnorm bias are added later + else{ + if (l->additional_bias) + b = { dtRT, bias2_b, l->outputs}; + else + b = { dtRT, nullptr, 0}; //on batchnorm bias are added later + } ILayer *lRT = nullptr; if(!l->deConv) { @@ -247,6 +318,7 @@ ILayer* NetworkRT::convert_layer(ITensor *input, Conv2d *l) { checkNULL(lRTconv); lRTconv->setStride(DimsHW{l->strideH, l->strideW}); lRTconv->setPadding(DimsHW{l->paddingH, l->paddingW}); + lRTconv->setNbGroups(l->groups); lRT = (ILayer*) lRTconv; } else { IDeconvolutionLayer *lRTconv = networkRT->addDeconvolution(*input, @@ -254,10 +326,11 @@ ILayer* NetworkRT::convert_layer(ITensor *input, Conv2d *l) { checkNULL(lRTconv); lRTconv->setStride(DimsHW{l->strideH, l->strideW}); lRTconv->setPadding(DimsHW{l->paddingH, l->paddingW}); + lRTconv->setNbGroups(l->groups); lRT = (ILayer*) lRTconv; Dims d = lRTconv->getOutput(0)->getDimensions(); - std::cout<<"DECONV: "<pool_mode == tkdnnPoolingMode_t::POOLING_AVERAGE) ptype = PoolingType::kAVERAGE; if(l->pool_mode == tkdnnPoolingMode_t::POOLING_AVERAGE_EXCLUDE_PADDING) ptype = PoolingType::kMAX_AVERAGE_BLEND; - - // if(l->input_dim.h % 2 == 1 && l->input_dim.w % 2 == 1) - if(l->input_dim.h == l->output_dim.h && l->input_dim.w == l->output_dim.w) + if(l->pool_mode == tkdnnPoolingMode_t::POOLING_MAX_FIXEDSIZE) { - IPlugin *plugin = new ResizeLayerRT( l->output_dim.c,l->output_dim.h+1,l->output_dim.w+1 ); - IPluginLayer *lRT = networkRT->addPlugin(&input, 1, *plugin); - checkNULL(lRT); - lRT->setName( "Resize" ); - - input = lRT->getOutput(0); + IPlugin *plugin = new MaxPoolFixedSizeRT(l->output_dim.c, l->output_dim.h, l->output_dim.w, l->output_dim.n, l->strideH, l->strideW, l->winH, l->winH-1); + IPluginLayer *lRT = networkRT->addPlugin(&input, 1, *plugin); + checkNULL(lRT); + return lRT; } + else + { + IPoolingLayer *lRT = networkRT->addPooling(*input, ptype, DimsHW{l->winH, l->winW}); + checkNULL(lRT); - IPoolingLayer *lRT = networkRT->addPooling(*input, ptype, DimsHW{l->winH, l->winW}); - checkNULL(lRT); - - lRT->setPadding(DimsHW{l->paddingH, l->paddingW}); - lRT->setStride(DimsHW{l->strideH, l->strideW}); - return lRT; + lRT->setPadding(DimsHW{l->paddingH, l->paddingW}); + lRT->setStride(DimsHW{l->strideH, l->strideW}); + return lRT; + } } ILayer* NetworkRT::convert_layer(ITensor *input, Activation *l) { @@ -316,10 +387,19 @@ ILayer* NetworkRT::convert_layer(ITensor *input, Activation *l) { if(l->act_mode == ACTIVATION_LEAKY) { //std::cout<<"New plugin LEAKY\n"; + +#if NV_TENSORRT_MAJOR < 6 + // plugin version IPlugin *plugin = new ActivationLeakyRT(); IPluginLayer *lRT = networkRT->addPlugin(&input, 1, *plugin); checkNULL(lRT); return lRT; +#else + IActivationLayer *lRT = networkRT->addActivation(*input, ActivationType::kLEAKY_RELU); + lRT->setAlpha(0.1); + checkNULL(lRT); + return lRT; +#endif } else if(l->act_mode == CUDNN_ACTIVATION_RELU) { IActivationLayer *lRT = networkRT->addActivation(*input, ActivationType::kRELU); @@ -329,8 +409,26 @@ ILayer* NetworkRT::convert_layer(ITensor *input, Activation *l) { IActivationLayer *lRT = networkRT->addActivation(*input, ActivationType::kSIGMOID); checkNULL(lRT); return lRT; - - } else { + } + else if(l->act_mode == CUDNN_ACTIVATION_CLIPPED_RELU) { + IPlugin *plugin = new ActivationReLUCeiling(l->ceiling); + IPluginLayer *lRT = networkRT->addPlugin(&input, 1, *plugin); + checkNULL(lRT); + return lRT; + } + else if(l->act_mode == ACTIVATION_MISH) { + IPlugin *plugin = new ActivationMishRT(); + IPluginLayer *lRT = networkRT->addPlugin(&input, 1, *plugin); + checkNULL(lRT); + return lRT; + } + else if(l->act_mode == ACTIVATION_LOGISTIC) { + IPlugin *plugin = new ActivationLogisticRT(); + IPluginLayer *lRT = networkRT->addPlugin(&input, 1, *plugin); + checkNULL(lRT); + return lRT; + } + else { FatalError("this Activation mode is not yet implemented"); return NULL; } @@ -358,12 +456,33 @@ ILayer* NetworkRT::convert_layer(ITensor *input, Route *l) { // } // std::cout<<"\n"; } - - IConcatenationLayer *lRT = networkRT->addConcatenation(tens, l->layers_n); - //IPlugin *plugin = new RouteRT(); - //IPluginLayer *lRT = networkRT->addPlugin(tens, l->layers_n, *plugin); - checkNULL(lRT); + if(l->groups > 1){ + IPlugin *plugin = new RouteRT(l->groups, l->group_id); + IPluginLayer *lRT = networkRT->addPlugin(tens, l->layers_n, *plugin); + checkNULL(lRT); + return lRT; + } + IConcatenationLayer *lRT = networkRT->addConcatenation(tens, l->layers_n); + checkNULL(lRT); + return lRT; +} + +ILayer* NetworkRT::convert_layer(ITensor *input, Flatten *l) { + + IPlugin *plugin = new FlattenConcatRT(); + IPluginLayer *lRT = networkRT->addPlugin(&input, 1, *plugin); + checkNULL(lRT); + return lRT; +} + +ILayer* NetworkRT::convert_layer(ITensor *input, Reshape *l) { + // std::cout<<"convert Reshape\n"; + + l->output_dim.print(); + IPlugin *plugin = new ReshapeRT(l->output_dim); + IPluginLayer *lRT = networkRT->addPlugin(&input, 1, *plugin); + checkNULL(lRT); return lRT; } @@ -391,22 +510,33 @@ ILayer* NetworkRT::convert_layer(ITensor *input, Shortcut *l) { //std::cout<<"convert Shortcut\n"; //std::cout<<"New plugin Shortcut\n"; + ITensor *back_tens = tensors[l->backLayer]; - IPlugin *plugin = new ShortcutRT(); - ITensor **inputs = new ITensor*[2]; - inputs[0] = input; - inputs[1] = back_tens; - IPluginLayer *lRT = networkRT->addPlugin(inputs, 2, *plugin); - checkNULL(lRT); - return lRT; + if(l->backLayer->output_dim.c == l->output_dim.c) + { + IElementWiseLayer *lRT = networkRT->addElementWise(*input, *back_tens, ElementWiseOperation::kSUM); + checkNULL(lRT); + return lRT; + } + else + { + // plugin version + IPlugin *plugin = new ShortcutRT(l->backLayer->output_dim); + ITensor **inputs = new ITensor*[2]; + inputs[0] = input; + inputs[1] = back_tens; + IPluginLayer *lRT = networkRT->addPlugin(inputs, 2, *plugin); + checkNULL(lRT); + return lRT; + } } ILayer* NetworkRT::convert_layer(ITensor *input, Yolo *l) { //std::cout<<"convert Yolo\n"; //std::cout<<"New plugin YOLO\n"; - IPlugin *plugin = new YoloRT(l->classes, l->num, l); + IPlugin *plugin = new YoloRT(l->classes, l->num, l, l->n_masks, l->scaleXY, l->nms_thresh, l->nsm_kind, l->new_coords); IPluginLayer *lRT = networkRT->addPlugin(&input, 1, *plugin); checkNULL(lRT); return lRT; @@ -423,7 +553,7 @@ ILayer* NetworkRT::convert_layer(ITensor *input, Upsample *l) { } ILayer* NetworkRT::convert_layer(ITensor *input, DeformConv2d *l) { - std::cout<<"convert DEFORMABLE\n"; + //std::cout<<"convert DEFORMABLE\n"; ILayer *preconv = convert_layer(input, l->preconv); checkNULL(preconv); @@ -431,14 +561,14 @@ ILayer* NetworkRT::convert_layer(ITensor *input, DeformConv2d *l) { inputs[0] = input; inputs[1] = preconv->getOutput(0); - std::cout<<"New plugin DEFORMABLE\n"; + //std::cout<<"New plugin DEFORMABLE\n"; IPlugin *plugin = new DeformableConvRT(l->chunk_dim, l->kernelH, l->kernelW, l->strideH, l->strideW, l->paddingH, l->paddingW, l->deformableGroup, l->input_dim.n, l->input_dim.c, l->input_dim.h, l->input_dim.w, l->output_dim.n, l->output_dim.c, l->output_dim.h, l->output_dim.w, l); IPluginLayer *lRT = networkRT->addPlugin(inputs, 2, *plugin); checkNULL(lRT); lRT->setName( ("Deformable" + std::to_string(l->id)).c_str() ); - + delete[](inputs); // batchnorm void *bias_b, *power_b, *mean_b, *variance_b, *scales_b; if(dtRT == DataType::kHALF) { @@ -458,7 +588,7 @@ ILayer* NetworkRT::convert_layer(ITensor *input, DeformConv2d *l) { Weights power{dtRT, power_b, l->outputs}; Weights shift{dtRT, mean_b, l->outputs}; Weights scale{dtRT, variance_b, l->outputs}; - std::cout<getNbOutputs()<getNbOutputs()<addScale(*lRT->getOutput(0), ScaleMode::kCHANNEL, shift, scale, power); @@ -475,7 +605,7 @@ ILayer* NetworkRT::convert_layer(ITensor *input, DeformConv2d *l) { bool NetworkRT::serialize(const char *filename) { - std::ofstream p(filename); + std::ofstream p(filename, std::ios::binary); if (!p) { FatalError("could not open plan output file"); return false; @@ -515,59 +645,145 @@ bool NetworkRT::deserialize(const char *filename) { IPlugin* PluginFactory::createPlugin(const char* layerName, const void* serialData, size_t serialLength) { - const char * buf = reinterpret_cast(serialData); + const char * buf = reinterpret_cast(serialData),*bufCheck = buf; std::string name(layerName); - std::cout<size = readBUF(buf); + assert(buf == bufCheck + serialLength); + return a; + } + if(name.find("ActivationMish") == 0) { + ActivationMishRT *a = new ActivationMishRT(); + a->size = readBUF(buf); + assert(buf == bufCheck + serialLength); + return a; + } + if(name.find("ActivationLogistic") == 0) { + ActivationLogisticRT *a = new ActivationLogisticRT(); + a->size = readBUF(buf); + return a; + } + if(name.find("ActivationLogistic") == 0) { + ActivationLogisticRT *a = new ActivationLogisticRT(); + a->size = readBUF(buf); + return a; + } + if(name.find("ActivationCReLU") == 0) { + float activationReluTemp = readBUF(buf); + ActivationReLUCeiling* a = new ActivationReLUCeiling(activationReluTemp); + a->size = readBUF(buf); + assert(buf == bufCheck + serialLength); return a; } if(name.find("Region") == 0) { - RegionRT *r = new RegionRT(readBUF(buf), //classes - readBUF(buf), //coords - readBUF(buf)); //num + int classesTemp = readBUF(buf); + int coordsTemp = readBUF(buf); + int numTemp = readBUF(buf); + RegionRT* r = new RegionRT(classesTemp, coordsTemp, numTemp); r->c = readBUF(buf); r->h = readBUF(buf); r->w = readBUF(buf); + assert(buf == bufCheck + serialLength); return r; } if(name.find("Reorg") == 0) { - ReorgRT *r = new ReorgRT(readBUF(buf)); //stride + int strideTemp = readBUF(buf); + ReorgRT *r = new ReorgRT(strideTemp); r->c = readBUF(buf); r->h = readBUF(buf); r->w = readBUF(buf); + assert(buf == bufCheck + serialLength); return r; } if(name.find("Shortcut") == 0) { - ShortcutRT *r = new ShortcutRT(); + tk::dnn::dataDim_t bdim; + bdim.c = readBUF(buf); + bdim.h = readBUF(buf); + bdim.w = readBUF(buf); + bdim.l = 1; + + ShortcutRT *r = new ShortcutRT(bdim); r->c = readBUF(buf); r->h = readBUF(buf); r->w = readBUF(buf); return r; + assert(buf == bufCheck + serialLength); } + if(name.find("Pooling") == 0) { + int cTemp = readBUF(buf); + int hTemp = readBUF(buf); + int wTemp = readBUF(buf); + int nTemp = readBUF(buf); + int strideHTemp = readBUF(buf); + int strideWTemp = readBUF(buf); + int winSizeTemp = readBUF(buf); + int paddingTemp = readBUF(buf); + + MaxPoolFixedSizeRT* r = new MaxPoolFixedSizeRT(cTemp, hTemp, wTemp, nTemp, strideHTemp, strideWTemp, winSizeTemp, paddingTemp); + assert(buf == bufCheck + serialLength); + return r; + } + if(name.find("Resize") == 0) { - ResizeLayerRT *r = new ResizeLayerRT(readBUF(buf), //o_c - readBUF(buf), //o_h - readBUF(buf)); //o_w + int o_cTemp = readBUF(buf); + int o_hTemp = readBUF(buf); + int o_wTemp = readBUF(buf); + ResizeLayerRT* r = new ResizeLayerRT(o_cTemp, o_hTemp, o_wTemp); + r->i_c = readBUF(buf); r->i_h = readBUF(buf); r->i_w = readBUF(buf); + assert(buf == bufCheck + serialLength); + return r; + } + + if(name.find("Flatten") == 0) { + FlattenConcatRT *r = new FlattenConcatRT(); + r->c = readBUF(buf); + r->h = readBUF(buf); + r->w = readBUF(buf); + r->rows = readBUF(buf); + r->cols = readBUF(buf); + assert(buf == bufCheck + serialLength); + return r; + } + + if(name.find("Reshape") == 0) { + + dataDim_t new_dim; + new_dim.n = readBUF(buf); + new_dim.c = readBUF(buf); + new_dim.h = readBUF(buf); + new_dim.w = readBUF(buf); + ReshapeRT *r = new ReshapeRT(new_dim); + assert(buf == bufCheck + serialLength); + return r; } if(name.find("Yolo") == 0) { - YoloRT *r = new YoloRT(readBUF(buf), //classes - readBUF(buf), //num - nullptr, - readBUF(buf)); //n_masks + + int classes_temp = readBUF(buf); + int num_temp = readBUF(buf); + int n_masks_temp = readBUF(buf); + float scale_xy_temp = readBUF(buf); + float nms_thresh_temp = readBUF(buf); + int nms_kind_temp = readBUF(buf); + int new_coords_temp = readBUF(buf); + + YoloRT *r = new YoloRT(classes_temp,num_temp,nullptr,n_masks_temp,scale_xy_temp,nms_thresh_temp,nms_kind_temp,new_coords_temp); + + + r->c = readBUF(buf); r->h = readBUF(buf); r->w = readBUF(buf); @@ -584,36 +800,54 @@ IPlugin* PluginFactory::createPlugin(const char* layerName, const void* serialDa tmp[j] = readBUF(buf); r->classesNames[i] = std::string(tmp); } + assert(buf == bufCheck + serialLength); yolos[n_yolos++] = r; return r; } if(name.find("Upsample") == 0) { - UpsampleRT *r = new UpsampleRT(readBUF(buf)); //stride + int strideTemp = readBUF(buf); + UpsampleRT* r = new UpsampleRT(strideTemp); r->c = readBUF(buf); r->h = readBUF(buf); r->w = readBUF(buf); + assert(buf == bufCheck + serialLength); return r; } -/* + if(name.find("Route") == 0) { - RouteRT *r = new RouteRT(); + int groupsTemp = readBUF(buf); + int group_idTemp = readBUF(buf); + RouteRT* r = new RouteRT(groupsTemp, group_idTemp); r->in = readBUF(buf); for(int i=0; ic_in[i] = readBUF(buf); r->c = readBUF(buf); r->h = readBUF(buf); r->w = readBUF(buf); + assert(buf == bufCheck + serialLength); return r; } -*/ + if(name.find("Deformable") == 0) { - DeformableConvRT *r = new DeformableConvRT(readBUF(buf), readBUF(buf), readBUF(buf), - readBUF(buf), readBUF(buf), readBUF(buf), - readBUF(buf), readBUF(buf), - readBUF(buf),readBUF(buf),readBUF(buf),readBUF(buf), - readBUF(buf),readBUF(buf),readBUF(buf),readBUF(buf), - nullptr); + int chuck_dimTemp = readBUF(buf); + int khTemp = readBUF(buf); + int kwTemp = readBUF(buf); + int shTemp = readBUF(buf); + int swTemp = readBUF(buf); + int phTemp = readBUF(buf); + int pwTemp = readBUF(buf); + int deformableGroupTemp = readBUF(buf); + int i_nTemp = readBUF(buf); + int i_cTemp = readBUF(buf); + int i_hTemp = readBUF(buf); + int i_wTemp = readBUF(buf); + int o_nTemp = readBUF(buf); + int o_cTemp = readBUF(buf); + int o_hTemp = readBUF(buf); + int o_wTemp = readBUF(buf); + + DeformableConvRT* r = new DeformableConvRT(chuck_dimTemp, khTemp, kwTemp, shTemp, swTemp, phTemp, pwTemp, deformableGroupTemp, i_nTemp, i_cTemp, i_hTemp, i_wTemp, o_nTemp, o_cTemp, o_hTemp, o_wTemp, nullptr); dnnType *aus = new dnnType[r->chunk_dim*2]; for(int i=0; ichunk_dim*2; i++) aus[i] = readBUF(buf); @@ -644,6 +878,7 @@ IPlugin* PluginFactory::createPlugin(const char* layerName, const void* serialDa aus[i] = readBUF(buf); checkCuda( cudaMemcpy(r->ones_d2, aus, sizeof(dnnType)*r->dim_ones, cudaMemcpyHostToDevice) ); free(aus); + assert(buf == bufCheck + serialLength); return r; } diff --git a/src/NetworkViz.cpp b/src/NetworkViz.cpp new file mode 100644 index 0000000..6ac274c --- /dev/null +++ b/src/NetworkViz.cpp @@ -0,0 +1,69 @@ +#include +#include +#include +#include +#include "tkDNN/NetworkViz.h" + +namespace tk { namespace dnn { + +cv::Mat vizFloat2colorMap(cv::Mat map) { + + double min; + double max; + cv::minMaxIdx(map, &min, &max); + cv::Mat adjMap; + // expand your range to 0..255. Similar to histEq(); + map.convertTo(adjMap,CV_8UC1, 255 / (max-min), -min); + //return adjMap; + + + cv::Mat falseColorsMap; + applyColorMap(adjMap, falseColorsMap, cv::COLORMAP_HOT); + return falseColorsMap; +} + +cv::Mat vizData2Mat(dnnType *dataInput, tk::dnn::dataDim_t dim, int imgdim) { + dnnType *data = nullptr; + + // copy to CPU + if(isCudaPointer(dataInput)) { + data = new dnnType[dim.tot()]; + checkCuda( cudaMemcpy(data, dataInput, dim.tot()*sizeof(dnnType), cudaMemcpyDeviceToHost) ); + } else { + data = dataInput; + } + + int gridDim = ceil(sqrt(dim.c)); + cv::Size gridSize(dim.w*gridDim, dim.h*gridDim); + cv::Mat grid = cv::Mat(gridSize, CV_8UC3, cv::Scalar(0)); + + for(int i=0; i= net->num_layers) + FatalError("Could not viz layer\n"); + return vizData2Mat(net->layers[layer]->dstData, net->layers[layer]->output_dim, imgdim); + + //cv::imwrite("viz/layer" + std::to_string(layer) + ".png", viz); + //cv::imshow("layer", viz); + //cv::waitKey(0); +} + +}} \ No newline at end of file diff --git a/src/Pooling.cpp b/src/Pooling.cpp index 1ad0664..0838806 100644 --- a/src/Pooling.cpp +++ b/src/Pooling.cpp @@ -7,8 +7,8 @@ namespace tk { namespace dnn { Pooling::Pooling( Network *net, int winH, int winW, int strideH, int strideW, int paddingH, int paddingW, - tkdnnPoolingMode_t pool_mode, bool final) : - Layer(net, final) { + tkdnnPoolingMode_t pool_mode) : + Layer(net) { this->winH = winH; this->winW = winW; @@ -39,9 +39,10 @@ Pooling::Pooling( Network *net, int winH, int winW, int strideH, int strideW, n = l; } - + cudnnPoolingMode_t cudnn_pool_mode = cudnnPoolingMode_t(pool_mode); + if(pool_mode == POOLING_MAX_FIXEDSIZE) cudnn_pool_mode = cudnnPoolingMode_t(tkdnnPoolingMode_t::POOLING_MAX); - checkCUDNN( cudnnSetPooling2dDescriptor(poolingDesc, cudnnPoolingMode_t(pool_mode), + checkCUDNN( cudnnSetPooling2dDescriptor(poolingDesc, cudnn_pool_mode, CUDNN_NOT_PROPAGATE_NAN, winH, winW, paddingH, paddingW, strideH, strideW) ); checkCUDNN( cudnnSetTensor4dDescriptor(srcTensorDesc, @@ -51,18 +52,16 @@ Pooling::Pooling( Network *net, int winH, int winW, int strideH, int strideW, // checkCUDNN( cudnnGetPooling2dForwardOutputDim(poolingDesc, srcTensorDesc, &n, &c, &h, &w)); //compute w and h as in darknet - int padH = paddingH == 0? winH -1 : paddingH; - int padW = paddingW == 0? winW -1 : paddingW; - - if(final){ - h = (h + padH - winH)/strideH +1 +1; - w = (w + padW - winW)/strideW +1 +1; - } - else{ + if(pool_mode == tkdnnPoolingMode_t::POOLING_MAX_FIXEDSIZE){ + int padH = paddingH == 0? winH -1 : paddingH; + int padW = paddingW == 0? winW -1 : paddingW; h = (h + padH - winH)/strideH +1; w = (w + padW - winW)/strideW +1; } - + else{ + h = (h + 2*paddingH - winH)/strideH +1 ; + w = (w + 2*paddingW - winW)/strideW +1; + } // h = (h + winH*this->paddingH)/strideH; // w = (w + winW*this->paddingW)/strideW; @@ -111,11 +110,16 @@ dnnType* Pooling::infer(dataDim_t &dim, dnnType* srcData) { poolDst = tmpOutputData; } - dnnType alpha = dnnType(1); - dnnType beta = dnnType(0); - checkCUDNN( cudnnPoolingForward(net->cudnnHandle, poolingDesc, - &alpha, srcTensorDesc, poolSrc, - &beta, dstTensorDesc, poolDst) ); + if(pool_mode == tkdnnPoolingMode_t::POOLING_MAX_FIXEDSIZE){ + MaxPoolingForward(poolSrc, poolDst, dim.n, dim.c, dim.h, dim.w, this->strideH, this->strideW, this->winH, this->winH-1); + } + else{ + dnnType alpha = dnnType(1); + dnnType beta = dnnType(0); + checkCUDNN( cudnnPoolingForward(net->cudnnHandle, poolingDesc, + &alpha, srcTensorDesc, poolSrc, + &beta, dstTensorDesc, poolDst) ); + } //update dim dim = output_dim; diff --git a/src/Region.cpp b/src/Region.cpp index 1b52375..7c26208 100644 --- a/src/Region.cpp +++ b/src/Region.cpp @@ -12,8 +12,7 @@ namespace tk { namespace dnn { Region::Region(Network *net, int classes, int coords, int num) : - Layer(net) { - + Layer(net) { this->classes = classes; this->coords = coords; this->num = num; @@ -64,7 +63,7 @@ dnnType* Region::infer(dataDim_t &dim, dnnType* srcData) { } -/* Intepret class */ +/* Interpret class */ RegionInterpret::RegionInterpret(dataDim_t input_dim, dataDim_t output_dim, int classes, int coords, int num, float thresh, std::string fname_weights) { diff --git a/src/Reshape.cpp b/src/Reshape.cpp new file mode 100644 index 0000000..c1d814a --- /dev/null +++ b/src/Reshape.cpp @@ -0,0 +1,34 @@ +#include + +#include "Layer.h" +#include "kernels.h" + +namespace tk { namespace dnn { + +Reshape::Reshape(Network *net, dataDim_t new_dim) : Layer(net) { + + checkCuda( cudaMalloc(&dstData, input_dim.tot()*sizeof(dnnType)) ); + + output_dim.n = new_dim.n; + output_dim.c = new_dim.c; + output_dim.h = new_dim.h; + output_dim.w = new_dim.w; + output_dim.l = new_dim.l; + +} + +Reshape::~Reshape() { + + checkCuda( cudaFree(dstData) ); +} + +dnnType* Reshape::infer(dataDim_t &dim, dnnType* srcData) { + + //just copies the data and changes the output dim + checkCuda( cudaMemcpy(dstData, srcData, dim.n*dim.c*dim.h*dim.w*sizeof(dnnType), cudaMemcpyDeviceToDevice)); + dim = output_dim; + + return dstData; +} + +}} \ No newline at end of file diff --git a/src/Route.cpp b/src/Route.cpp index 7a95532..816566e 100644 --- a/src/Route.cpp +++ b/src/Route.cpp @@ -5,13 +5,18 @@ namespace tk { namespace dnn { -Route::Route(Network *net, Layer **layers, int layers_n) : Layer(net) { +Route::Route(Network *net, Layer **layers, int layers_n, int groups, int group_id) : Layer(net) { - this->layers_n = layers_n; - if(layers_n > MAX_INPUT_LAYERS) - FatalError("Route: MAX INPUT LAYERS overload"); - for(int i=0; i MAX_LAYERS) { + FatalError("ROUTE: reached max number of input layers"); + } + for(int i=0; ilayers[i] = layers[i]; + } + this->layers_n = layers_n; + this->groups = groups; + this->group_id = group_id; //get dims output_dim.l = 1; @@ -29,6 +34,7 @@ Route::Route(Network *net, Layer **layers, int layers_n) : Layer(net) { output_dim.c += layers[i]->output_dim.c; } + output_dim.c /= this->groups; input_dim = output_dim; checkCuda( cudaMalloc(&dstData, output_dim.tot()*sizeof(dnnType)) ); @@ -46,8 +52,9 @@ dnnType* Route::infer(dataDim_t &dim, dnnType* srcData) { for(int i=0; idstData; int in_dim = layers[i]->output_dim.tot(); - checkCuda( cudaMemcpy(dstData + offset, input, in_dim*sizeof(dnnType), cudaMemcpyDeviceToDevice)); - offset += in_dim; + int part_in_dim = in_dim / this->groups; + checkCuda( cudaMemcpy(dstData + offset, input + this->group_id*part_in_dim, part_in_dim*sizeof(dnnType), cudaMemcpyDeviceToDevice)); + offset += part_in_dim; } //update data dimensions diff --git a/src/Shortcut.cpp b/src/Shortcut.cpp index bcd2a00..2c7a4f4 100644 --- a/src/Shortcut.cpp +++ b/src/Shortcut.cpp @@ -10,10 +10,10 @@ Shortcut::Shortcut(Network *net, Layer *backLayer) : Layer(net) { this->backLayer = backLayer; checkCuda( cudaMalloc(&dstData, output_dim.tot()*sizeof(dnnType)) ); - if( backLayer->output_dim.c != input_dim.c || + if( /*backLayer->output_dim.c != input_dim.c ||*/ backLayer->output_dim.w != input_dim.w || backLayer->output_dim.h != input_dim.h ) - FatalError("Shortcut dim missmatch"); + FatalError("Shortcut dim mismatch"); } Shortcut::~Shortcut() { diff --git a/src/Softmax.cpp b/src/Softmax.cpp index af08f7f..8652d93 100644 --- a/src/Softmax.cpp +++ b/src/Softmax.cpp @@ -5,22 +5,40 @@ namespace tk { namespace dnn { -Softmax::Softmax(Network *net) : Layer(net) { +Softmax::Softmax(Network *net, const tk::dnn::dataDim_t* dim, const cudnnSoftmaxMode_t mode) : Layer(net) { checkCuda( cudaMalloc(&dstData, input_dim.tot()*sizeof(dnnType)) ); + this->mode = mode; + if(dim == nullptr) + { + this->dim.n= input_dim.n; + this->dim.c= input_dim.c; + this->dim.h= input_dim.h; + this->dim.w= input_dim.w; + this->dim.l= input_dim.l; + } + else + { + this->dim.n= dim->n; + this->dim.c= dim->c; + this->dim.h= dim->h; + this->dim.w= dim->w; + this->dim.l= dim->l; + } + checkCUDNN( cudnnSetTensor4dDescriptor(srcTensorDesc, net->tensorFormat, net->dataType, - input_dim.n*input_dim.l, - input_dim.c, - input_dim.h, input_dim.w) ); + this->dim.n*this->dim.l, + this->dim.c, + this->dim.h, this->dim.w) ); checkCUDNN( cudnnSetTensor4dDescriptor(dstTensorDesc, net->tensorFormat, net->dataType, - input_dim.n*input_dim.l, - input_dim.c, - input_dim.h, input_dim.w) ); + this->dim.n*this->dim.l, + this->dim.c, + this->dim.h, this->dim.w) ); } Softmax::~Softmax() { @@ -34,7 +52,7 @@ dnnType* Softmax::infer(dataDim_t &dim, dnnType* srcData) { dnnType beta = dnnType(0); checkCUDNN( cudnnSoftmaxForward(net->cudnnHandle, CUDNN_SOFTMAX_ACCURATE , - CUDNN_SOFTMAX_MODE_CHANNEL, + this->mode, &alpha, srcTensorDesc, srcData, diff --git a/src/Yolo.cpp b/src/Yolo.cpp index 634f5ff..eaf1de7 100644 --- a/src/Yolo.cpp +++ b/src/Yolo.cpp @@ -9,14 +9,20 @@ #include "Layer.h" #include "kernels.h" + namespace tk { namespace dnn { -Yolo::Yolo(Network *net, int classes, int num, std::string fname_weights, int n_masks) : +Yolo::Yolo(Network *net, int classes, int num, std::string fname_weights, int n_masks, float scale_xy, double nms_thresh, nmsKind_t nsm_kind, int new_coords) : Layer(net) { - + this->final = true; + this->classes = classes; this->num = num; this->n_masks = n_masks; + this->scaleXY = scale_xy; + this->nms_thresh = nms_thresh; + this->nsm_kind = nsm_kind; + this->new_coords = new_coords; // load anchors if(fname_weights != "") { @@ -57,12 +63,21 @@ int entry_index(int batch, int location, int entry, entry*input_dim.w*input_dim.h + loc; } -Yolo::box get_yolo_box(float *x, float *biases, int n, int index, int i, int j, int lw, int lh, int w, int h, int stride) { +Yolo::box get_yolo_box(float *x, float *biases, int n, int index, int i, int j, int lw, int lh, int w, int h, int stride, int new_coords) { Yolo::box b; - b.x = (i + x[index + 0*stride]) / lw; - b.y = (j + x[index + 1*stride]) / lh; - b.w = exp(x[index + 2*stride]) * biases[2*n] / w; - b.h = exp(x[index + 3*stride]) * biases[2*n+1] / h; + + if(new_coords == 0){ + b.x = (i + x[index + 0*stride]) / lw; + b.y = (j + x[index + 1*stride]) / lh; + b.w = exp(x[index + 2*stride]) * biases[2*n] / w; + b.h = exp(x[index + 3*stride]) * biases[2*n+1] / h; + } + else{ + b.x = (i + x[index + 0 * stride] ) / lw; + b.y = (j + x[index + 1 * stride] ) / lh; + b.w = x[index + 2 * stride] * x[index + 2 * stride] * 4 * biases[2 * n] / w; + b.h = x[index + 3 * stride] * x[index + 3 * stride] * 4 * biases[2 * n + 1] / h; + } return b; } @@ -73,10 +88,17 @@ dnnType* Yolo::infer(dataDim_t &dim, dnnType* srcData) { for (int b = 0; b < dim.n; ++b){ for(int n = 0; n < n_masks; ++n){ int index = entry_index(b, n*dim.w*dim.h, 0, classes, input_dim, output_dim); - activationLOGISTICForward(srcData + index, dstData + index, 2*dim.w*dim.h); - - index = entry_index(b, n*dim.w*dim.h, 4, classes, input_dim, output_dim); - activationLOGISTICForward(srcData + index, dstData + index, (1+classes)*dim.w*dim.h); + std::cout<<"new_coords"<scaleXY != 1) scalAdd(dstData + index, 2 * dim.w*dim.h, this->scaleXY, -0.5*(this->scaleXY - 1), 1); + } + else{ + activationLOGISTICForward(srcData + index, dstData + index, 2*dim.w*dim.h); + + if (this->scaleXY != 1) scalAdd(dstData + index, 2 * dim.w*dim.h, this->scaleXY, -0.5*(this->scaleXY - 1), 1); + index = entry_index(b, n*dim.w*dim.h, 4, classes, input_dim, output_dim); + activationLOGISTICForward(srcData + index, dstData + index, (1+classes)*dim.w*dim.h); + } } } @@ -112,7 +134,7 @@ void correct_yolo_boxes(Yolo::detection *dets, int n, int w, int h, int netw, in } } -int Yolo::computeDetections(Yolo::detection *dets, int &ndets, int netw, int neth, float thresh) { +int Yolo::computeDetections(Yolo::detection *dets, int &ndets, int netw, int neth, float thresh, int new_coords) { if(predictions == nullptr) predictions = new dnnType[output_dim.tot()]; @@ -136,7 +158,7 @@ int Yolo::computeDetections(Yolo::detection *dets, int &ndets, int netw, int net if(objectness <= thresh) continue; int box_index = entry_index(0, n*lw*lh + i, 0, classes, input_dim, output_dim); - dets[count].bbox = get_yolo_box(predictions, bias_h, mask_h[n], box_index, col, row, lw, lh, netw, neth, lw*lh); + dets[count].bbox = get_yolo_box(predictions, bias_h, mask_h[n], box_index, col, row, lw, lh, netw, neth, lw*lh, new_coords); dets[count].objectness = objectness; dets[count].classes = classes; for(j = 0; j < classes; ++j){ @@ -189,6 +211,32 @@ float yolo_box_iou(Yolo::box a, Yolo::box b) return yolo_box_intersection(a, b)/yolo_box_union(a, b); } +void box_c(const Yolo::box a, const Yolo::box b, float& top, float& bot, float& left, float& right) { + top = (std::min)(a.y - a.h / 2, b.y - b.h / 2); + bot = (std::max)(a.y + a.h / 2, b.y + b.h / 2); + left = (std::min)(a.x - a.w / 2, b.x - b.w / 2); + right = (std::max)(a.x + a.w / 2, b.x + b.w / 2); +} + +// https://github.com/Zzh-tju/DIoU-darknet +// https://arxiv.org/abs/1911.08287 +float yolo_box_diou(const Yolo::box a, const Yolo::box b, const float nms_thresh=0.6) +{ + float top, bot, left, right; + box_c(a, b, top, bot, left, right); + float w = right - left; + float h = bot - top; + float c = w * w + h * h; + float iou = yolo_box_iou(a, b); + if (c == 0) + return iou; + + float d = (a.x - b.x) * (a.x - b.x) + (a.y - b.y) * (a.y - b.y); + float u = pow(d / c, nms_thresh); + float diou_term = u; + return iou - diou_term; +} + int yolo_nms_comparator(const void *pa, const void *pb) { Yolo::detection a = *(Yolo::detection *)pa; @@ -215,8 +263,7 @@ Yolo::detection *Yolo::allocateDetections(int nboxes, int classes) { return dets; } -void Yolo::mergeDetections(Yolo::detection *dets, int ndets, int classes) { - double nms_thresh = 0.45; +void Yolo::mergeDetections(Yolo::detection *dets, int ndets, int classes, double nms_thresh, nmsKind_t nsm_kind) { int total = ndets; int i, j, k; @@ -242,13 +289,13 @@ void Yolo::mergeDetections(Yolo::detection *dets, int ndets, int classes) { box a = dets[i].bbox; for(j = i+1; j < total; ++j){ box b = dets[j].bbox; - if (yolo_box_iou(a, b) > nms_thresh){ + if (nsm_kind == GREEDY_NMS && yolo_box_iou(a, b) > nms_thresh) + dets[j].prob[k] = 0; + else if (nsm_kind == DIOU_NMS && yolo_box_diou(a, b, nms_thresh) > nms_thresh) dets[j].prob[k] = 0; - } } } } - } }} diff --git a/src/Yolo3Detection.cpp b/src/Yolo3Detection.cpp index aae23b7..0c638e6 100644 --- a/src/Yolo3Detection.cpp +++ b/src/Yolo3Detection.cpp @@ -1,27 +1,18 @@ #include "Yolo3Detection.h" + namespace tk { namespace dnn { -float _colors[6][3] = { {1,0,1}, {0,0,1},{0,1,1},{0,1,0},{1,1,0},{1,0,0} }; -float get_color(int c, int x, int max) -{ - float ratio = ((float)x/max)*5; - int i = floor(ratio); - int j = ceil(ratio); - ratio -= i; - float r = (1-ratio) * _colors[i % 6][c % 3] + ratio*_colors[j % 6][c % 3]; - //printf("%f\n", r); - return r; -} - -bool Yolo3Detection::init(std::string tensor_path) { - - //const char *tensor_path = "../data/yolo3/yolo3_berkeley.rt"; +bool Yolo3Detection::init(const std::string& tensor_path, const int n_classes, const int n_batches, const float conf_thresh) { //convert network to tensorRT std::cout<<(tensor_path).c_str()<<"\n"; netRT = new tk::dnn::NetworkRT(NULL, (tensor_path).c_str() ); + nBatches = n_batches; + confThreshold = conf_thresh; + tk::dnn::dataDim_t idim = netRT->input_dim; + idim.n = nBatches; if(netRT->pluginFactory->n_yolos < 2 ) { FatalError("this is not yolo3"); @@ -31,122 +22,141 @@ bool Yolo3Detection::init(std::string tensor_path) { YoloRT *yRT = netRT->pluginFactory->yolos[i]; classes = yRT->classes; num = yRT->num; - n_masks = yRT->n_masks; + nMasks = yRT->n_masks; - // make a yolo layer for interpret predictions - yolo[i] = new tk::dnn::Yolo(nullptr, classes, n_masks, ""); // yolo without input and bias - yolo[i]->mask_h = new dnnType[n_masks]; - yolo[i]->bias_h = new dnnType[num*n_masks*2]; - memcpy(yolo[i]->mask_h, yRT->mask, sizeof(dnnType)*n_masks); - memcpy(yolo[i]->bias_h, yRT->bias, sizeof(dnnType)*num*n_masks*2); + // make a yolo layer to interpret predictions + yolo[i] = new tk::dnn::Yolo(nullptr, classes, nMasks, ""); // yolo without input and bias + yolo[i]->mask_h = new dnnType[nMasks]; + yolo[i]->bias_h = new dnnType[num*nMasks*2]; + memcpy(yolo[i]->mask_h, yRT->mask, sizeof(dnnType)*nMasks); + memcpy(yolo[i]->bias_h, yRT->bias, sizeof(dnnType)*num*nMasks*2); yolo[i]->input_dim = yolo[i]->output_dim = tk::dnn::dataDim_t(1, yRT->c, yRT->h, yRT->w); yolo[i]->classesNames = yRT->classesNames; + yolo[i]->nms_thresh = yRT->nms_thresh; + yolo[i]->nsm_kind = (tk::dnn::Yolo::nmsKind_t) yRT->nms_kind; + yolo[i]->new_coords = yRT->new_coords; } dets = tk::dnn::Yolo::allocateDetections(tk::dnn::Yolo::MAX_DETECTIONS, classes); - - checkCuda(cudaMallocHost(&input, sizeof(dnnType)*netRT->input_dim.tot())); - checkCuda(cudaMalloc(&input_d, sizeof(dnnType)*netRT->input_dim.tot())); - +#ifndef OPENCV_CUDACONTRIB + checkCuda(cudaMallocHost(&input, sizeof(dnnType)*idim.tot())); +#endif + checkCuda(cudaMalloc(&input_d, sizeof(dnnType)*idim.tot())); // class colors precompute for(int c=0; cclassesNames; return true; } +void Yolo3Detection::preprocess(cv::Mat &frame, const int bi){ +#ifdef OPENCV_CUDACONTRIB + cv::cuda::GpuMat orig_img, img_resized; + orig_img = cv::cuda::GpuMat(frame); + cv::cuda::resize(orig_img, img_resized, cv::Size(netRT->input_dim.w, netRT->input_dim.h)); -void Yolo3Detection::update(cv::Mat &imageORIG) { - - if(!imageORIG.data) { - std::cout<<"YOLO: NO IMAGE DATA\n"; - return; - } - float xRatio = float(imageORIG.cols) / float(netRT->input_dim.w); - float yRatio = float(imageORIG.rows) / float(netRT->input_dim.h); - - resize(imageORIG, imageORIG, cv::Size(netRT->input_dim.w, netRT->input_dim.h)); - - imageORIG.convertTo(imageF, CV_32FC3, 1/255.0); + img_resized.convertTo(imagePreproc, CV_32FC3, 1/255.0); //split channels - cv::split(imageF,bgr);//split source + cv::cuda::split(imagePreproc,bgr);//split source //write channels for(int i=0; iinput_dim.c; i++) { - int idx = i*imageF.rows*imageF.cols; + int size = imagePreproc.rows * imagePreproc.cols; int ch = netRT->input_dim.c-1 -i; - memcpy((void*)&input[idx], (void*)bgr[ch].data, imageF.rows*imageF.cols*sizeof(dnnType)); + bgr[ch].download(bgr_h); //TODO: don't copy back on CPU + checkCuda( cudaMemcpy(input_d + i*size + netRT->input_dim.tot()*bi, (float*)bgr_h.data, size*sizeof(dnnType), cudaMemcpyHostToDevice)); } +#else + cv::resize(frame, frame, cv::Size(netRT->input_dim.w, netRT->input_dim.h)); + frame.convertTo(imagePreproc, CV_32FC3, 1/255.0); + //split channels + cv::split(imagePreproc,bgr);//split source - //DO INFERENCE - dnnType *rt_out[netRT->pluginFactory->n_yolos]; - tk::dnn::dataDim_t dim = netRT->input_dim; - checkCuda(cudaMemcpyAsync(input_d, input, dim.tot()*sizeof(dnnType), cudaMemcpyHostToDevice, netRT->stream)); - - printCenteredTitle(" TENSORRT inference ", '=', 30); { - dim.print(); - TIMER_START - netRT->infer(dim, input_d); - TIMER_STOP - dim.print(); + //write channels + for(int i=0; iinput_dim.c; i++) { + int idx = i*imagePreproc.rows*imagePreproc.cols; + int ch = netRT->input_dim.c-1 -i; + memcpy((void*)&input[idx + netRT->input_dim.tot()*bi], (void*)bgr[ch].data, imagePreproc.rows*imagePreproc.cols*sizeof(dnnType)); } - - TIMER_START + checkCuda(cudaMemcpyAsync(input_d + netRT->input_dim.tot()*bi, input + netRT->input_dim.tot()*bi, netRT->input_dim.tot()*sizeof(dnnType), cudaMemcpyHostToDevice, netRT->stream)); +#endif +} + +void Yolo3Detection::postprocess(const int bi, const bool mAP){ + + //get yolo outputs + std::vector rt_out; + //dnnType *rt_out[netRT->pluginFactory->n_yolos]; + for(int i=0; ipluginFactory->n_yolos; i++) + rt_out.push_back((dnnType*)netRT->buffersRT[i+1] + netRT->buffersDIM[i+1].tot()*bi); + + float x_ratio = float(originalSize[bi].width) / float(netRT->input_dim.w); + float y_ratio = float(originalSize[bi].height) / float(netRT->input_dim.h); + // compute dets - ndets = 0; + nDets = 0; for(int i=0; ipluginFactory->n_yolos; i++) { - rt_out[i] = (dnnType*)netRT->buffersRT[i+1]; yolo[i]->dstData = rt_out[i]; - yolo[i]->computeDetections(dets, ndets, netRT->input_dim.w, netRT->input_dim.h, thresh); + yolo[i]->computeDetections(dets, nDets, netRT->input_dim.w, netRT->input_dim.h, confThreshold, yolo[i]->new_coords); } - tk::dnn::Yolo::mergeDetections(dets, ndets, classes); - TIMER_STOP + tk::dnn::Yolo::mergeDetections(dets, nDets, classes, yolo[0]->nms_thresh, yolo[0]->nsm_kind); // fill detected detected.clear(); - for(int j=0; j= thresh) { - obj_class = c; - prob = dets[j].prob[c]; + if(dets[j].prob[c] >= confThreshold) { + int obj_class = c; + float prob = dets[j].prob[c]; + + tk::dnn::box res; + res.cl = obj_class; + res.prob = prob; + res.x = x0; + res.y = y0; + res.w = x1 - x0; + res.h = y1 - y0; + + // FIXME: this shuld be useless + // if(mAP) + // for(int c=0; c= 0) { - //std::cout< +namespace tk { namespace dnn { + +void Frame::print() const{ + std::cout<<"labels filename: "<(); + conf_thresh1 = config["conf_thresh1"].as(); + classes2 = config["classes2"].as(); + conf_thresh2 = config["conf_thresh2"].as(); + classes3 = config["classes3"].as(); + conf_thresh3 = config["conf_thresh3"].as(); + classes4 = config["classes4"].as(); + conf_thresh4 = config["conf_thresh4"].as(); + classes5 = config["classes5"].as(); + conf_thresh5 = config["conf_thresh5"].as(); +} + +/* Credits to https://github.com/AlexeyAB/darknet/blob/master/src/detector.c*/ +double computeMap( std::vector &images,const int classes, + const float IoU_thresh, const float conf_thresh, + const int map_points, const bool verbose) { + if(verbose) + for(auto img:images) + img.print(); + + int detections_count = 0; + int groundtruths_count = 0; + int unique_truth_count = 0; + std::vector truth_classes_count(classes,0); + std::vector dets_classes_count(classes,0); + + //count groundtruth and detections in total and for each class + for(auto i:images){ + for(auto gt:i.gt) + truth_classes_count[gt.cl]++; + for(auto det:i.det) + dets_classes_count[det.cl]++; + detections_count += i.det.size(); + groundtruths_count += i.gt.size(); + } + + if(verbose){ + std::cout<<"gt_count: "< all_dets; + std::vector all_gts; + + int gt_checked = 0; + + // for each detection compute IoU with groundtruth and match detetcion and + // groundtruth with IoU greater than IoU_thresh + for(auto &img:images){ + for(size_t i=0; i conf_thresh){ + float maxIoU = 0; + int truth_index = -1; + for(size_t j=0; j maxIoU && img.det[i].cl == img.gt[j].cl){ + maxIoU = currentIoU; + truth_index = j; + } + } + if(truth_index > -1 && maxIoU > IoU_thresh){ + img.det[i].uniqueTruthIndex = truth_index + gt_checked; + img.det[i].truthFlag = 1; + img.det[i].maxIoU = maxIoU; + } + } + + all_dets.push_back(img.det[i]); + } + gt_checked += img.gt.size(); + } + + if(verbose){ + for(auto img:images) + img.print(); + std::cout<<"\n\n\n\n"; + } + + //sort all detections by descending value of confidence + std::sort(all_dets.begin(), all_dets.end(), boxComparison); + std::vector truth_flags(groundtruths_count,0); + + if(verbose) + for(auto d:all_dets) + std::cout<> pr( classes, std::vector(detections_count)); + for(int rank = 0; rank< detections_count; ++rank){ + if (rank > 0) { + for (int class_id = 0; class_id < classes; ++class_id) { + pr[class_id][rank].tp = pr[class_id][rank - 1].tp; + pr[class_id][rank].fp = pr[class_id][rank - 1].fp; + } + } + + //if it was detected and never detected before + if (all_dets[rank].truthFlag == 1 && truth_flags[all_dets[rank].uniqueTruthIndex] == 0) { + truth_flags[all_dets[rank].uniqueTruthIndex] = 1; + pr[all_dets[rank].cl][rank].tp++; // true-positive + } + else { + pr[all_dets[rank].cl][rank].fp++; // false-positive + } + + for (int i = 0; i < classes; ++i){ + const int tp = pr[i][rank].tp; + const int fp = pr[i][rank].fp; + const int fn = truth_classes_count[i] - tp; // false-negative = objects - true-positive + pr[i][rank].fn = fn; + + if ((tp + fp) > 0) + pr[i][rank].precision = (double)tp / (double)(tp + fp); + else + pr[i][rank].precision = 0; + + if ((tp + fn) > 0) + pr[i][rank].recall = (double)tp / (double)(tp + fn); + else + pr[i][rank].recall = 0; + + if (rank == (detections_count - 1) && dets_classes_count[i] != (tp + fp)) { + // check for last rank + printf(" class_id: %d - detections = %d, tp+fp = %d, tp = %d, fp = %d \n", i, dets_classes_count[i], tp+fp, tp, fp); + } + } + } + + if(verbose){ + for(int i=0; i < pr.size(); i++) { + std::cout<<"---------Class "<= 0; --rank){ + delta_recall = last_recall - pr[i][rank].recall; + last_recall = pr[i][rank].recall; + + if (pr[i][rank].precision > last_precision) + last_precision = pr[i][rank].precision; + + avg_precision += delta_recall * last_precision; + } + } + else {//MSCOCO - 101 Recall-points, PascalVOC - 11 Recall-points + for (int point = 0; point < map_points; ++point) { + cur_recall = point * 1.0 / ( map_points - 1 ); + cur_precision = 0; + for (int rank = 0; rank < detections_count; ++rank) + if (pr[i][rank].recall >= cur_recall && pr[i][rank].precision > cur_precision) + cur_precision = pr[i][rank].precision; + + avg_precision += cur_precision; + } + avg_precision = avg_precision / map_points; + } + + if(verbose) + std::cout<<"Class: "< &images,const int classes, + const float i_IoU_thresh, const float conf_thresh, + const int map_points, const float map_step, + const int map_levels, const bool verbose, + const bool write_on_file, std::string net) { + std::ofstream out_file; + if(write_on_file){ + out_file.open("map.csv", std::ios_base::app); + out_file< &images,const int classes, + const float IoU_thresh, const float conf_thresh, + bool verbose, const bool write_on_file, std::string net) { + + std::ofstream out_file; + if(write_on_file){ + out_file.open("pr.csv", std::ios_base::app); + out_file< truth_classes_count(classes,0); + std::vector dets_classes_count(classes,0); + std::vector pr(classes); + + //compute TP, FP, FN for each image, for each class + for(auto &img:images){ + for(auto& tc: truth_classes_count) tc = 0; + for(auto& dc: dets_classes_count) dc = 0; + + std::vector det_assigned(img.det.size(), false); + for(size_t j=0; j conf_thresh){ + float currentIoU = img.det[i].IoU(img.gt[j]); + if(currentIoU > maxIoU && img.det[i].cl == img.gt[j].cl && !det_assigned[i]){ + maxIoU = currentIoU; + det_index = i; + } + } + } + if(det_index > -1 && maxIoU > IoU_thresh && !det_assigned[det_index]){ + img.det[det_index].uniqueTruthIndex = j; + img.det[det_index].truthFlag = 1; + img.det[det_index].maxIoU = maxIoU; + det_assigned[det_index] = true; + dets_classes_count[img.det[det_index].cl]++; + } + } + + for(size_t i=0; i 0 ? (double)pr[i].tp / (double)(pr[i].tp +pr[i].fp) : 0; + pr[i].recall = (pr[i].tp + pr[i].fn) > 0 ? (double)pr[i].tp / (double)(pr[i].tp +pr[i].fn) : 0; + if(verbose) + std::cout<<"Class "< 0 ? 2 * ( avg_precision * avg_recall ) / ( avg_precision + avg_recall ) : 0; + + if(write_on_file){ + out_file< bbox, const int classes, const int w, const int h) +{ + int coco_ids[] = { 1,2,3,4,5,6,7,8,9,10,11,13,14,15,16,17,18,19,20,21,22,23,24,25,27,28,31,32,33,34,35,36,37,38,39,40,41,42,43,44,46,47,48,49,50,51,52,53,54,55,56,57,58,59,60,61,62,63,64,65,67,70,72,73,74,75,76,77,78,79,80,81,82,84,85,86,87,88,89,90 }; + std::string id = image_path.substr(image_path.find("images/")+7, image_path.find(".jpg") - image_path.find("images/") -7); + int image_id = std::stoi(id); + for (int i = 0; i < bbox.size(); ++i) { + float xmin = bbox[i].x ; + float xmax = bbox[i].x + float(bbox[i].w); + float ymin = bbox[i].y; + float ymax = bbox[i].y + float(bbox[i].h); + + //limit to image borders + if (xmin < 0) xmin = 0; + if (ymin < 0) ymin = 0; + if (xmax > w) xmax = w; + if (ymax > h) ymax = h; + + float bx = xmin; + float by = ymin; + float bw = xmax - xmin; + float bh = ymax - ymin; + + if(bbox[i].probs.size() == classes) + for (int j = 0; j < classes; ++j) { + //min threshold confidence is set in DetectionNN.h + if (bbox[i].probs[j] > 0) { + + *out_file << "{\"image_id\":" << image_id << + ", \"category_id\":" << coco_ids[j] << + ", \"bbox\":[" << bx << ", " << by << ", " << bw << ", " << bh << + "], \"score\":" << bbox[i].probs[j] << "},\n"; + } + } + else + *out_file << "{\"image_id\":" << image_id << + ", \"category_id\":" << coco_ids[bbox[i].cl] << + ", \"bbox\":[" << bx << ", " << by << ", " << bw << ", " << bh << + "], \"score\":" << bbox[i].prob << "},\n"; + } +} + +}} diff --git a/src/image.c b/src/image.c new file mode 100644 index 0000000..08dcbfb --- /dev/null +++ b/src/image.c @@ -0,0 +1,1749 @@ +#ifndef _GNU_SOURCE +#define _GNU_SOURCE +#endif +#include "image.h" +//#include "utils.h" +//#include "blas.h" +//#include "dark_cuda.h" +#include +#ifndef _USE_MATH_DEFINES +#define _USE_MATH_DEFINES +#endif +#include + +#ifndef STB_IMAGE_IMPLEMENTATION +#define STB_IMAGE_IMPLEMENTATION +#include "stb_image.h" +#endif +#ifndef STB_IMAGE_WRITE_IMPLEMENTATION +#define STB_IMAGE_WRITE_IMPLEMENTATION +#include "stb_image_write.h" +#endif + +extern int check_mistakes; +//int windows = 0; + +float colors[6][3] = { {1,0,1}, {0,0,1},{0,1,1},{0,1,0},{1,1,0},{1,0,0} }; + +float get_color(int c, int x, int max) +{ + float ratio = ((float)x/max)*5; + int i = floor(ratio); + int j = ceil(ratio); + ratio -= i; + float r = (1-ratio) * colors[i][c] + ratio*colors[j][c]; + //printf("%f\n", r); + return r; +} + +static float get_pixel(image m, int x, int y, int c) +{ + assert(x < m.w && y < m.h && c < m.c); + return m.data[c*m.h*m.w + y*m.w + x]; +} +static float get_pixel_extend(image m, int x, int y, int c) +{ + if (x < 0 || x >= m.w || y < 0 || y >= m.h) return 0; + /* + if(x < 0) x = 0; + if(x >= m.w) x = m.w-1; + if(y < 0) y = 0; + if(y >= m.h) y = m.h-1; + */ + if (c < 0 || c >= m.c) return 0; + return get_pixel(m, x, y, c); +} +static void set_pixel(image m, int x, int y, int c, float val) +{ + if (x < 0 || y < 0 || c < 0 || x >= m.w || y >= m.h || c >= m.c) return; + assert(x < m.w && y < m.h && c < m.c); + m.data[c*m.h*m.w + y*m.w + x] = val; +} +static void add_pixel(image m, int x, int y, int c, float val) +{ + assert(x < m.w && y < m.h && c < m.c); + m.data[c*m.h*m.w + y*m.w + x] += val; +} + +void composite_image(image source, image dest, int dx, int dy) +{ + int x,y,k; + for(k = 0; k < source.c; ++k){ + for(y = 0; y < source.h; ++y){ + for(x = 0; x < source.w; ++x){ + float val = get_pixel(source, x, y, k); + float val2 = get_pixel_extend(dest, dx+x, dy+y, k); + set_pixel(dest, dx+x, dy+y, k, val * val2); + } + } + } +} + +image border_image(image a, int border) +{ + image b = make_image(a.w + 2*border, a.h + 2*border, a.c); + int x,y,k; + for(k = 0; k < b.c; ++k){ + for(y = 0; y < b.h; ++y){ + for(x = 0; x < b.w; ++x){ + float val = get_pixel_extend(a, x - border, y - border, k); + if(x - border < 0 || x - border >= a.w || y - border < 0 || y - border >= a.h) val = 1; + set_pixel(b, x, y, k, val); + } + } + } + return b; +} + +image tile_images(image a, image b, int dx) +{ + if(a.w == 0) return copy_image(b); + image c = make_image(a.w + b.w + dx, (a.h > b.h) ? a.h : b.h, (a.c > b.c) ? a.c : b.c); + fill_cpu(c.w*c.h*c.c, 1, c.data, 1); + embed_image(a, c, 0, 0); + composite_image(b, c, a.w + dx, 0); + return c; +} + +image get_label(image **characters, char *string, int size) +{ + if(size > 7) size = 7; + image label = make_empty_image(0,0,0); + while(*string){ + image l = characters[size][(int)*string]; + image n = tile_images(label, l, -size - 1 + (size+1)/2); + free_image(label); + label = n; + ++string; + } + image b = border_image(label, label.h*.25); + free_image(label); + return b; +} + +image get_label_v3(image **characters, char *string, int size) +{ + size = size / 10; + if (size > 7) size = 7; + image label = make_empty_image(0, 0, 0); + while (*string) { + image l = characters[size][(int)*string]; + image n = tile_images(label, l, -size - 1 + (size + 1) / 2); + free_image(label); + label = n; + ++string; + } + image b = border_image(label, label.h*.25); + free_image(label); + return b; +} + +void draw_label(image a, int r, int c, image label, const float *rgb) +{ + int w = label.w; + int h = label.h; + if (r - h >= 0) r = r - h; + + int i, j, k; + for(j = 0; j < h && j + r < a.h; ++j){ + for(i = 0; i < w && i + c < a.w; ++i){ + for(k = 0; k < label.c; ++k){ + float val = get_pixel(label, i, j, k); + set_pixel(a, i+c, j+r, k, rgb[k] * val); + } + } + } +} + +void draw_box_bw(image a, int x1, int y1, int x2, int y2, float brightness) +{ + //normalize_image(a); + int i; + if (x1 < 0) x1 = 0; + if (x1 >= a.w) x1 = a.w - 1; + if (x2 < 0) x2 = 0; + if (x2 >= a.w) x2 = a.w - 1; + + if (y1 < 0) y1 = 0; + if (y1 >= a.h) y1 = a.h - 1; + if (y2 < 0) y2 = 0; + if (y2 >= a.h) y2 = a.h - 1; + + for (i = x1; i <= x2; ++i) { + a.data[i + y1*a.w + 0 * a.w*a.h] = brightness; + a.data[i + y2*a.w + 0 * a.w*a.h] = brightness; + } + for (i = y1; i <= y2; ++i) { + a.data[x1 + i*a.w + 0 * a.w*a.h] = brightness; + a.data[x2 + i*a.w + 0 * a.w*a.h] = brightness; + } +} + +void draw_box_width_bw(image a, int x1, int y1, int x2, int y2, int w, float brightness) +{ + int i; + for (i = 0; i < w; ++i) { + float alternate_color = (w % 2) ? (brightness) : (1.0 - brightness); + draw_box_bw(a, x1 + i, y1 + i, x2 - i, y2 - i, alternate_color); + } +} + +void draw_box(image a, int x1, int y1, int x2, int y2, float r, float g, float b) +{ + //normalize_image(a); + int i; + if(x1 < 0) x1 = 0; + if(x1 >= a.w) x1 = a.w-1; + if(x2 < 0) x2 = 0; + if(x2 >= a.w) x2 = a.w-1; + + if(y1 < 0) y1 = 0; + if(y1 >= a.h) y1 = a.h-1; + if(y2 < 0) y2 = 0; + if(y2 >= a.h) y2 = a.h-1; + + for(i = x1; i <= x2; ++i){ + a.data[i + y1*a.w + 0*a.w*a.h] = r; + a.data[i + y2*a.w + 0*a.w*a.h] = r; + + a.data[i + y1*a.w + 1*a.w*a.h] = g; + a.data[i + y2*a.w + 1*a.w*a.h] = g; + + a.data[i + y1*a.w + 2*a.w*a.h] = b; + a.data[i + y2*a.w + 2*a.w*a.h] = b; + } + for(i = y1; i <= y2; ++i){ + a.data[x1 + i*a.w + 0*a.w*a.h] = r; + a.data[x2 + i*a.w + 0*a.w*a.h] = r; + + a.data[x1 + i*a.w + 1*a.w*a.h] = g; + a.data[x2 + i*a.w + 1*a.w*a.h] = g; + + a.data[x1 + i*a.w + 2*a.w*a.h] = b; + a.data[x2 + i*a.w + 2*a.w*a.h] = b; + } +} + +void draw_box_width(image a, int x1, int y1, int x2, int y2, int w, float r, float g, float b) +{ + int i; + for(i = 0; i < w; ++i){ + draw_box(a, x1+i, y1+i, x2-i, y2-i, r, g, b); + } +} + +void draw_bbox(image a, box bbox, int w, float r, float g, float b) +{ + int left = (bbox.x-bbox.w/2)*a.w; + int right = (bbox.x+bbox.w/2)*a.w; + int top = (bbox.y-bbox.h/2)*a.h; + int bot = (bbox.y+bbox.h/2)*a.h; + + int i; + for(i = 0; i < w; ++i){ + draw_box(a, left+i, top+i, right-i, bot-i, r, g, b); + } +} + +image **load_alphabet() +{ + int i, j; + const int nsize = 8; + image** alphabets = (image**)xcalloc(nsize, sizeof(image*)); + for(j = 0; j < nsize; ++j){ + alphabets[j] = (image*)xcalloc(128, sizeof(image)); + for(i = 32; i < 127; ++i){ + char buff[256]; + sprintf(buff, "data/labels/%d_%d.png", i, j); + alphabets[j][i] = load_image_color(buff, 0, 0); + } + } + return alphabets; +} + + + +// Creates array of detections with prob > thresh and fills best_class for them +detection_with_class* get_actual_detections(detection *dets, int dets_num, float thresh, int* selected_detections_num, char **names) +{ + int selected_num = 0; + detection_with_class* result_arr = (detection_with_class*)xcalloc(dets_num, sizeof(detection_with_class)); + int i; + for (i = 0; i < dets_num; ++i) { + int best_class = -1; + float best_class_prob = thresh; + int j; + for (j = 0; j < dets[i].classes; ++j) { + int show = strncmp(names[j], "dont_show", 9); + if (dets[i].prob[j] > best_class_prob && show) { + best_class = j; + best_class_prob = dets[i].prob[j]; + } + } + if (best_class >= 0) { + result_arr[selected_num].det = dets[i]; + result_arr[selected_num].best_class = best_class; + ++selected_num; + } + } + if (selected_detections_num) + *selected_detections_num = selected_num; + return result_arr; +} + +// compare to sort detection** by bbox.x +int compare_by_lefts(const void *a_ptr, const void *b_ptr) { + const detection_with_class* a = (detection_with_class*)a_ptr; + const detection_with_class* b = (detection_with_class*)b_ptr; + const float delta = (a->det.bbox.x - a->det.bbox.w/2) - (b->det.bbox.x - b->det.bbox.w/2); + return delta < 0 ? -1 : delta > 0 ? 1 : 0; +} + +// compare to sort detection** by best_class probability +int compare_by_probs(const void *a_ptr, const void *b_ptr) { + const detection_with_class* a = (detection_with_class*)a_ptr; + const detection_with_class* b = (detection_with_class*)b_ptr; + float delta = a->det.prob[a->best_class] - b->det.prob[b->best_class]; + return delta < 0 ? -1 : delta > 0 ? 1 : 0; +} + +void draw_detections_v3(image im, detection *dets, int num, float thresh, char **names, image **alphabet, int classes, int ext_output) +{ + static int frame_id = 0; + frame_id++; + + int selected_detections_num; + detection_with_class* selected_detections = get_actual_detections(dets, num, thresh, &selected_detections_num, names); + + // text output + qsort(selected_detections, selected_detections_num, sizeof(*selected_detections), compare_by_lefts); + int i; + for (i = 0; i < selected_detections_num; ++i) { + const int best_class = selected_detections[i].best_class; + printf("%s: %.0f%%", names[best_class], selected_detections[i].det.prob[best_class] * 100); + if (ext_output) + printf("\t(left_x: %4.0f top_y: %4.0f width: %4.0f height: %4.0f)\n", + round((selected_detections[i].det.bbox.x - selected_detections[i].det.bbox.w / 2)*im.w), + round((selected_detections[i].det.bbox.y - selected_detections[i].det.bbox.h / 2)*im.h), + round(selected_detections[i].det.bbox.w*im.w), round(selected_detections[i].det.bbox.h*im.h)); + else + printf("\n"); + int j; + for (j = 0; j < classes; ++j) { + if (selected_detections[i].det.prob[j] > thresh && j != best_class) { + printf("%s: %.0f%%", names[j], selected_detections[i].det.prob[j] * 100); + + if (ext_output) + printf("\t(left_x: %4.0f top_y: %4.0f width: %4.0f height: %4.0f)\n", + round((selected_detections[i].det.bbox.x - selected_detections[i].det.bbox.w / 2)*im.w), + round((selected_detections[i].det.bbox.y - selected_detections[i].det.bbox.h / 2)*im.h), + round(selected_detections[i].det.bbox.w*im.w), round(selected_detections[i].det.bbox.h*im.h)); + else + printf("\n"); + } + } + } + + // image output + qsort(selected_detections, selected_detections_num, sizeof(*selected_detections), compare_by_probs); + for (i = 0; i < selected_detections_num; ++i) { + int width = im.h * .006; + if (width < 1) + width = 1; + + /* + if(0){ + width = pow(prob, 1./2.)*10+1; + alphabet = 0; + } + */ + + //printf("%d %s: %.0f%%\n", i, names[selected_detections[i].best_class], prob*100); + int offset = selected_detections[i].best_class * 123457 % classes; + float red = get_color(2, offset, classes); + float green = get_color(1, offset, classes); + float blue = get_color(0, offset, classes); + float rgb[3]; + + //width = prob*20+2; + + rgb[0] = red; + rgb[1] = green; + rgb[2] = blue; + box b = selected_detections[i].det.bbox; + //printf("%f %f %f %f\n", b.x, b.y, b.w, b.h); + + int left = (b.x - b.w / 2.)*im.w; + int right = (b.x + b.w / 2.)*im.w; + int top = (b.y - b.h / 2.)*im.h; + int bot = (b.y + b.h / 2.)*im.h; + + if (left < 0) left = 0; + if (right > im.w - 1) right = im.w - 1; + if (top < 0) top = 0; + if (bot > im.h - 1) bot = im.h - 1; + + //int b_x_center = (left + right) / 2; + //int b_y_center = (top + bot) / 2; + //int b_width = right - left; + //int b_height = bot - top; + //sprintf(labelstr, "%d x %d - w: %d, h: %d", b_x_center, b_y_center, b_width, b_height); + + // you should create directory: result_img + //static int copied_frame_id = -1; + //static image copy_img; + //if (copied_frame_id != frame_id) { + // copied_frame_id = frame_id; + // if (copy_img.data) free_image(copy_img); + // copy_img = copy_image(im); + //} + //image cropped_im = crop_image(copy_img, left, top, right - left, bot - top); + //static int img_id = 0; + //img_id++; + //char image_name[1024]; + //int best_class_id = selected_detections[i].best_class; + //sprintf(image_name, "result_img/img_%d_%d_%d_%s.jpg", frame_id, img_id, best_class_id, names[best_class_id]); + //save_image(cropped_im, image_name); + //free_image(cropped_im); + + if (im.c == 1) { + draw_box_width_bw(im, left, top, right, bot, width, 0.8); // 1 channel Black-White + } + else { + draw_box_width(im, left, top, right, bot, width, red, green, blue); // 3 channels RGB + } + if (alphabet) { + char labelstr[4096] = { 0 }; + strcat(labelstr, names[selected_detections[i].best_class]); + int j; + for (j = 0; j < classes; ++j) { + if (selected_detections[i].det.prob[j] > thresh && j != selected_detections[i].best_class) { + strcat(labelstr, ", "); + strcat(labelstr, names[j]); + } + } + image label = get_label_v3(alphabet, labelstr, (im.h*.03)); + draw_label(im, top + width, left, label, rgb); + free_image(label); + } + if (selected_detections[i].det.mask) { + image mask = float_to_image(14, 14, 1, selected_detections[i].det.mask); + image resized_mask = resize_image(mask, b.w*im.w, b.h*im.h); + image tmask = threshold_image(resized_mask, .5); + embed_image(tmask, im, left, top); + free_image(mask); + free_image(resized_mask); + free_image(tmask); + } + } + free(selected_detections); +} + +void draw_detections(image im, int num, float thresh, box *boxes, float **probs, char **names, image **alphabet, int classes) +{ + int i; + + for(i = 0; i < num; ++i){ + int class_id = max_index(probs[i], classes); + float prob = probs[i][class_id]; + if(prob > thresh){ + + //// for comparison with OpenCV version of DNN Darknet Yolo v2 + //printf("\n %f, %f, %f, %f, ", boxes[i].x, boxes[i].y, boxes[i].w, boxes[i].h); + // int k; + //for (k = 0; k < classes; ++k) { + // printf("%f, ", probs[i][k]); + //} + //printf("\n"); + + int width = im.h * .012; + + if(0){ + width = pow(prob, 1./2.)*10+1; + alphabet = 0; + } + + int offset = class_id*123457 % classes; + float red = get_color(2,offset,classes); + float green = get_color(1,offset,classes); + float blue = get_color(0,offset,classes); + float rgb[3]; + + //width = prob*20+2; + + rgb[0] = red; + rgb[1] = green; + rgb[2] = blue; + box b = boxes[i]; + + int left = (b.x-b.w/2.)*im.w; + int right = (b.x+b.w/2.)*im.w; + int top = (b.y-b.h/2.)*im.h; + int bot = (b.y+b.h/2.)*im.h; + + if(left < 0) left = 0; + if(right > im.w-1) right = im.w-1; + if(top < 0) top = 0; + if(bot > im.h-1) bot = im.h-1; + printf("%s: %.0f%%", names[class_id], prob * 100); + + //printf(" - id: %d, x_center: %d, y_center: %d, width: %d, height: %d", + // class_id, (right + left) / 2, (bot - top) / 2, right - left, bot - top); + + printf("\n"); + draw_box_width(im, left, top, right, bot, width, red, green, blue); + if (alphabet) { + image label = get_label(alphabet, names[class_id], (im.h*.03)/10); + draw_label(im, top + width, left, label, rgb); + } + } + } +} + +void transpose_image(image im) +{ + assert(im.w == im.h); + int n, m; + int c; + for(c = 0; c < im.c; ++c){ + for(n = 0; n < im.w-1; ++n){ + for(m = n + 1; m < im.w; ++m){ + float swap = im.data[m + im.w*(n + im.h*c)]; + im.data[m + im.w*(n + im.h*c)] = im.data[n + im.w*(m + im.h*c)]; + im.data[n + im.w*(m + im.h*c)] = swap; + } + } + } +} + +void rotate_image_cw(image im, int times) +{ + assert(im.w == im.h); + times = (times + 400) % 4; + int i, x, y, c; + int n = im.w; + for(i = 0; i < times; ++i){ + for(c = 0; c < im.c; ++c){ + for(x = 0; x < n/2; ++x){ + for(y = 0; y < (n-1)/2 + 1; ++y){ + float temp = im.data[y + im.w*(x + im.h*c)]; + im.data[y + im.w*(x + im.h*c)] = im.data[n-1-x + im.w*(y + im.h*c)]; + im.data[n-1-x + im.w*(y + im.h*c)] = im.data[n-1-y + im.w*(n-1-x + im.h*c)]; + im.data[n-1-y + im.w*(n-1-x + im.h*c)] = im.data[x + im.w*(n-1-y + im.h*c)]; + im.data[x + im.w*(n-1-y + im.h*c)] = temp; + } + } + } + } +} + +void flip_image(image a) +{ + int i,j,k; + for(k = 0; k < a.c; ++k){ + for(i = 0; i < a.h; ++i){ + for(j = 0; j < a.w/2; ++j){ + int index = j + a.w*(i + a.h*(k)); + int flip = (a.w - j - 1) + a.w*(i + a.h*(k)); + float swap = a.data[flip]; + a.data[flip] = a.data[index]; + a.data[index] = swap; + } + } + } +} + +image image_distance(image a, image b) +{ + int i,j; + image dist = make_image(a.w, a.h, 1); + for(i = 0; i < a.c; ++i){ + for(j = 0; j < a.h*a.w; ++j){ + dist.data[j] += pow(a.data[i*a.h*a.w+j]-b.data[i*a.h*a.w+j],2); + } + } + for(j = 0; j < a.h*a.w; ++j){ + dist.data[j] = sqrt(dist.data[j]); + } + return dist; +} + +void embed_image(image source, image dest, int dx, int dy) +{ + int x,y,k; + for(k = 0; k < source.c; ++k){ + for(y = 0; y < source.h; ++y){ + for(x = 0; x < source.w; ++x){ + float val = get_pixel(source, x,y,k); + set_pixel(dest, dx+x, dy+y, k, val); + } + } + } +} + +image collapse_image_layers(image source, int border) +{ + int h = source.h; + h = (h+border)*source.c - border; + image dest = make_image(source.w, h, 1); + int i; + for(i = 0; i < source.c; ++i){ + image layer = get_image_layer(source, i); + int h_offset = i*(source.h+border); + embed_image(layer, dest, 0, h_offset); + free_image(layer); + } + return dest; +} + +void constrain_image(image im) +{ + int i; + for(i = 0; i < im.w*im.h*im.c; ++i){ + if(im.data[i] < 0) im.data[i] = 0; + if(im.data[i] > 1) im.data[i] = 1; + } +} + +void normalize_image(image p) +{ + int i; + float min = 9999999; + float max = -999999; + + for(i = 0; i < p.h*p.w*p.c; ++i){ + float v = p.data[i]; + if(v < min) min = v; + if(v > max) max = v; + } + if(max - min < .000000001){ + min = 0; + max = 1; + } + for(i = 0; i < p.c*p.w*p.h; ++i){ + p.data[i] = (p.data[i] - min)/(max-min); + } +} + +void normalize_image2(image p) +{ + float* min = (float*)xcalloc(p.c, sizeof(float)); + float* max = (float*)xcalloc(p.c, sizeof(float)); + int i,j; + for(i = 0; i < p.c; ++i) min[i] = max[i] = p.data[i*p.h*p.w]; + + for(j = 0; j < p.c; ++j){ + for(i = 0; i < p.h*p.w; ++i){ + float v = p.data[i+j*p.h*p.w]; + if(v < min[j]) min[j] = v; + if(v > max[j]) max[j] = v; + } + } + for(i = 0; i < p.c; ++i){ + if(max[i] - min[i] < .000000001){ + min[i] = 0; + max[i] = 1; + } + } + for(j = 0; j < p.c; ++j){ + for(i = 0; i < p.w*p.h; ++i){ + p.data[i+j*p.h*p.w] = (p.data[i+j*p.h*p.w] - min[j])/(max[j]-min[j]); + } + } + free(min); + free(max); +} + +void copy_image_inplace(image src, image dst) +{ + memcpy(dst.data, src.data, src.h*src.w*src.c * sizeof(float)); +} + +image copy_image(image p) +{ + image copy = p; + copy.data = (float*)xcalloc(p.h * p.w * p.c, sizeof(float)); + memcpy(copy.data, p.data, p.h*p.w*p.c*sizeof(float)); + return copy; +} + +void rgbgr_image(image im) +{ + int i; + for(i = 0; i < im.w*im.h; ++i){ + float swap = im.data[i]; + im.data[i] = im.data[i+im.w*im.h*2]; + im.data[i+im.w*im.h*2] = swap; + } +} + +void show_image(image p, const char *name) +{ +#ifdef OPENCV + show_image_cv(p, name); +#else + fprintf(stderr, "Not compiled with OpenCV, saving to %s.png instead\n", name); + save_image(p, name); +#endif // OPENCV +} + +void save_image_png(image im, const char *name) +{ + char buff[256]; + //sprintf(buff, "%s (%d)", name, windows); + sprintf(buff, "%s.png", name); + unsigned char* data = (unsigned char*)xcalloc(im.w * im.h * im.c, sizeof(unsigned char)); + int i,k; + for(k = 0; k < im.c; ++k){ + for(i = 0; i < im.w*im.h; ++i){ + data[i*im.c+k] = (unsigned char) (255*im.data[i + k*im.w*im.h]); + } + } + int success = stbi_write_png(buff, im.w, im.h, im.c, data, im.w*im.c); + free(data); + if(!success) fprintf(stderr, "Failed to write image %s\n", buff); +} + +void save_image_options(image im, const char *name, IMTYPE f, int quality) +{ + char buff[256]; + //sprintf(buff, "%s (%d)", name, windows); + if (f == PNG) sprintf(buff, "%s.png", name); + else if (f == BMP) sprintf(buff, "%s.bmp", name); + else if (f == TGA) sprintf(buff, "%s.tga", name); + else if (f == JPG) sprintf(buff, "%s.jpg", name); + else sprintf(buff, "%s.png", name); + unsigned char* data = (unsigned char*)xcalloc(im.w * im.h * im.c, sizeof(unsigned char)); + int i, k; + for (k = 0; k < im.c; ++k) { + for (i = 0; i < im.w*im.h; ++i) { + data[i*im.c + k] = (unsigned char)(255 * im.data[i + k*im.w*im.h]); + } + } + int success = 0; + if (f == PNG) success = stbi_write_png(buff, im.w, im.h, im.c, data, im.w*im.c); + else if (f == BMP) success = stbi_write_bmp(buff, im.w, im.h, im.c, data); + else if (f == TGA) success = stbi_write_tga(buff, im.w, im.h, im.c, data); + else if (f == JPG) success = stbi_write_jpg(buff, im.w, im.h, im.c, data, quality); + free(data); + if (!success) fprintf(stderr, "Failed to write image %s\n", buff); +} + +void save_image(image im, const char *name) +{ + save_image_options(im, name, JPG, 80); +} + +void save_image_jpg(image p, const char *name) +{ + save_image_options(p, name, JPG, 80); +} + +void show_image_layers(image p, char *name) +{ + int i; + char buff[256]; + for(i = 0; i < p.c; ++i){ + sprintf(buff, "%s - Layer %d", name, i); + image layer = get_image_layer(p, i); + show_image(layer, buff); + free_image(layer); + } +} + +void show_image_collapsed(image p, char *name) +{ + image c = collapse_image_layers(p, 1); + show_image(c, name); + free_image(c); +} + +image make_empty_image(int w, int h, int c) +{ + image out; + out.data = 0; + out.h = h; + out.w = w; + out.c = c; + return out; +} + +image make_image(int w, int h, int c) +{ + image out = make_empty_image(w,h,c); + out.data = (float*)xcalloc(h * w * c, sizeof(float)); + return out; +} + +image make_random_image(int w, int h, int c) +{ + image out = make_empty_image(w,h,c); + out.data = (float*)xcalloc(h * w * c, sizeof(float)); + int i; + for(i = 0; i < w*h*c; ++i){ + out.data[i] = (rand_normal() * .25) + .5; + } + return out; +} + +image float_to_image_scaled(int w, int h, int c, float *data) +{ + image out = make_image(w, h, c); + int abs_max = 0; + int i = 0; + for (i = 0; i < w*h*c; ++i) { + if (fabs(data[i]) > abs_max) abs_max = fabs(data[i]); + } + for (i = 0; i < w*h*c; ++i) { + out.data[i] = data[i] / abs_max; + } + return out; +} + +image float_to_image(int w, int h, int c, float *data) +{ + image out = make_empty_image(w,h,c); + out.data = data; + return out; +} + + +image rotate_crop_image(image im, float rad, float s, int w, int h, float dx, float dy, float aspect) +{ + int x, y, c; + float cx = im.w/2.; + float cy = im.h/2.; + image rot = make_image(w, h, im.c); + for(c = 0; c < im.c; ++c){ + for(y = 0; y < h; ++y){ + for(x = 0; x < w; ++x){ + float rx = cos(rad)*((x - w/2.)/s*aspect + dx/s*aspect) - sin(rad)*((y - h/2.)/s + dy/s) + cx; + float ry = sin(rad)*((x - w/2.)/s*aspect + dx/s*aspect) + cos(rad)*((y - h/2.)/s + dy/s) + cy; + float val = bilinear_interpolate(im, rx, ry, c); + set_pixel(rot, x, y, c, val); + } + } + } + return rot; +} + +image rotate_image(image im, float rad) +{ + int x, y, c; + float cx = im.w/2.; + float cy = im.h/2.; + image rot = make_image(im.w, im.h, im.c); + for(c = 0; c < im.c; ++c){ + for(y = 0; y < im.h; ++y){ + for(x = 0; x < im.w; ++x){ + float rx = cos(rad)*(x-cx) - sin(rad)*(y-cy) + cx; + float ry = sin(rad)*(x-cx) + cos(rad)*(y-cy) + cy; + float val = bilinear_interpolate(im, rx, ry, c); + set_pixel(rot, x, y, c, val); + } + } + } + return rot; +} + +void translate_image(image m, float s) +{ + int i; + for(i = 0; i < m.h*m.w*m.c; ++i) m.data[i] += s; +} + +void scale_image(image m, float s) +{ + int i; + for(i = 0; i < m.h*m.w*m.c; ++i) m.data[i] *= s; +} + +image crop_image(image im, int dx, int dy, int w, int h) +{ + image cropped = make_image(w, h, im.c); + int i, j, k; + for(k = 0; k < im.c; ++k){ + for(j = 0; j < h; ++j){ + for(i = 0; i < w; ++i){ + int r = j + dy; + int c = i + dx; + float val = 0; + r = constrain_int(r, 0, im.h-1); + c = constrain_int(c, 0, im.w-1); + if (r >= 0 && r < im.h && c >= 0 && c < im.w) { + val = get_pixel(im, c, r, k); + } + set_pixel(cropped, i, j, k, val); + } + } + } + return cropped; +} + +int best_3d_shift_r(image a, image b, int min, int max) +{ + if(min == max) return min; + int mid = floor((min + max) / 2.); + image c1 = crop_image(b, 0, mid, b.w, b.h); + image c2 = crop_image(b, 0, mid+1, b.w, b.h); + float d1 = dist_array(c1.data, a.data, a.w*a.h*a.c, 10); + float d2 = dist_array(c2.data, a.data, a.w*a.h*a.c, 10); + free_image(c1); + free_image(c2); + if(d1 < d2) return best_3d_shift_r(a, b, min, mid); + else return best_3d_shift_r(a, b, mid+1, max); +} + +int best_3d_shift(image a, image b, int min, int max) +{ + int i; + int best = 0; + float best_distance = FLT_MAX; + for(i = min; i <= max; i += 2){ + image c = crop_image(b, 0, i, b.w, b.h); + float d = dist_array(c.data, a.data, a.w*a.h*a.c, 100); + if(d < best_distance){ + best_distance = d; + best = i; + } + printf("%d %f\n", i, d); + free_image(c); + } + return best; +} + +void composite_3d(char *f1, char *f2, char *out, int delta) +{ + if(!out) out = "out"; + image a = load_image(f1, 0,0,0); + image b = load_image(f2, 0,0,0); + int shift = best_3d_shift_r(a, b, -a.h/100, a.h/100); + + image c1 = crop_image(b, 10, shift, b.w, b.h); + float d1 = dist_array(c1.data, a.data, a.w*a.h*a.c, 100); + image c2 = crop_image(b, -10, shift, b.w, b.h); + float d2 = dist_array(c2.data, a.data, a.w*a.h*a.c, 100); + + if(d2 < d1 && 0){ + image swap = a; + a = b; + b = swap; + shift = -shift; + printf("swapped, %d\n", shift); + } + else{ + printf("%d\n", shift); + } + + image c = crop_image(b, delta, shift, a.w, a.h); + int i; + for(i = 0; i < c.w*c.h; ++i){ + c.data[i] = a.data[i]; + } +#ifdef OPENCV + save_image_jpg(c, out); +#else + save_image(c, out); +#endif +} + +void fill_image(image m, float s) +{ + int i; + for (i = 0; i < m.h*m.w*m.c; ++i) m.data[i] = s; +} + +void letterbox_image_into(image im, int w, int h, image boxed) +{ + int new_w = im.w; + int new_h = im.h; + if (((float)w / im.w) < ((float)h / im.h)) { + new_w = w; + new_h = (im.h * w) / im.w; + } + else { + new_h = h; + new_w = (im.w * h) / im.h; + } + image resized = resize_image(im, new_w, new_h); + embed_image(resized, boxed, (w - new_w) / 2, (h - new_h) / 2); + free_image(resized); +} + +image letterbox_image(image im, int w, int h) +{ + int new_w = im.w; + int new_h = im.h; + if (((float)w / im.w) < ((float)h / im.h)) { + new_w = w; + new_h = (im.h * w) / im.w; + } + else { + new_h = h; + new_w = (im.w * h) / im.h; + } + image resized = resize_image(im, new_w, new_h); + image boxed = make_image(w, h, im.c); + fill_image(boxed, .5); + //int i; + //for(i = 0; i < boxed.w*boxed.h*boxed.c; ++i) boxed.data[i] = 0; + embed_image(resized, boxed, (w - new_w) / 2, (h - new_h) / 2); + free_image(resized); + return boxed; +} + +image resize_max(image im, int max) +{ + int w = im.w; + int h = im.h; + if(w > h){ + h = (h * max) / w; + w = max; + } else { + w = (w * max) / h; + h = max; + } + if(w == im.w && h == im.h) return im; + image resized = resize_image(im, w, h); + return resized; +} + +image resize_min(image im, int min) +{ + int w = im.w; + int h = im.h; + if(w < h){ + h = (h * min) / w; + w = min; + } else { + w = (w * min) / h; + h = min; + } + if(w == im.w && h == im.h) return im; + image resized = resize_image(im, w, h); + return resized; +} + +image random_crop_image(image im, int w, int h) +{ + int dx = rand_int(0, im.w - w); + int dy = rand_int(0, im.h - h); + image crop = crop_image(im, dx, dy, w, h); + return crop; +} + +image random_augment_image(image im, float angle, float aspect, int low, int high, int size) +{ + aspect = rand_scale(aspect); + int r = rand_int(low, high); + int min = (im.h < im.w*aspect) ? im.h : im.w*aspect; + float scale = (float)r / min; + + float rad = rand_uniform(-angle, angle) * 2.0 * M_PI / 360.; + + float dx = (im.w*scale/aspect - size) / 2.; + float dy = (im.h*scale - size) / 2.; + if(dx < 0) dx = 0; + if(dy < 0) dy = 0; + dx = rand_uniform(-dx, dx); + dy = rand_uniform(-dy, dy); + + image crop = rotate_crop_image(im, rad, scale, size, size, dx, dy, aspect); + + return crop; +} + +float three_way_max(float a, float b, float c) +{ + return (a > b) ? ( (a > c) ? a : c) : ( (b > c) ? b : c) ; +} + +float three_way_min(float a, float b, float c) +{ + return (a < b) ? ( (a < c) ? a : c) : ( (b < c) ? b : c) ; +} + +// http://www.cs.rit.edu/~ncs/color/t_convert.html +void rgb_to_hsv(image im) +{ + assert(im.c == 3); + int i, j; + float r, g, b; + float h, s, v; + for(j = 0; j < im.h; ++j){ + for(i = 0; i < im.w; ++i){ + r = get_pixel(im, i , j, 0); + g = get_pixel(im, i , j, 1); + b = get_pixel(im, i , j, 2); + float max = three_way_max(r,g,b); + float min = three_way_min(r,g,b); + float delta = max - min; + v = max; + if(max == 0){ + s = 0; + h = 0; + }else{ + s = delta/max; + if(r == max){ + h = (g - b) / delta; + } else if (g == max) { + h = 2 + (b - r) / delta; + } else { + h = 4 + (r - g) / delta; + } + if (h < 0) h += 6; + h = h/6.; + } + set_pixel(im, i, j, 0, h); + set_pixel(im, i, j, 1, s); + set_pixel(im, i, j, 2, v); + } + } +} + +void hsv_to_rgb(image im) +{ + assert(im.c == 3); + int i, j; + float r, g, b; + float h, s, v; + float f, p, q, t; + for(j = 0; j < im.h; ++j){ + for(i = 0; i < im.w; ++i){ + h = 6 * get_pixel(im, i , j, 0); + s = get_pixel(im, i , j, 1); + v = get_pixel(im, i , j, 2); + if (s == 0) { + r = g = b = v; + } else { + int index = floor(h); + f = h - index; + p = v*(1-s); + q = v*(1-s*f); + t = v*(1-s*(1-f)); + if(index == 0){ + r = v; g = t; b = p; + } else if(index == 1){ + r = q; g = v; b = p; + } else if(index == 2){ + r = p; g = v; b = t; + } else if(index == 3){ + r = p; g = q; b = v; + } else if(index == 4){ + r = t; g = p; b = v; + } else { + r = v; g = p; b = q; + } + } + set_pixel(im, i, j, 0, r); + set_pixel(im, i, j, 1, g); + set_pixel(im, i, j, 2, b); + } + } +} + +image grayscale_image(image im) +{ + assert(im.c == 3); + int i, j, k; + image gray = make_image(im.w, im.h, 1); + float scale[] = {0.587, 0.299, 0.114}; + for(k = 0; k < im.c; ++k){ + for(j = 0; j < im.h; ++j){ + for(i = 0; i < im.w; ++i){ + gray.data[i+im.w*j] += scale[k]*get_pixel(im, i, j, k); + } + } + } + return gray; +} + +image threshold_image(image im, float thresh) +{ + int i; + image t = make_image(im.w, im.h, im.c); + for(i = 0; i < im.w*im.h*im.c; ++i){ + t.data[i] = im.data[i]>thresh ? 1 : 0; + } + return t; +} + +image blend_image(image fore, image back, float alpha) +{ + assert(fore.w == back.w && fore.h == back.h && fore.c == back.c); + image blend = make_image(fore.w, fore.h, fore.c); + int i, j, k; + for(k = 0; k < fore.c; ++k){ + for(j = 0; j < fore.h; ++j){ + for(i = 0; i < fore.w; ++i){ + float val = alpha * get_pixel(fore, i, j, k) + + (1 - alpha)* get_pixel(back, i, j, k); + set_pixel(blend, i, j, k, val); + } + } + } + return blend; +} + +void scale_image_channel(image im, int c, float v) +{ + int i, j; + for(j = 0; j < im.h; ++j){ + for(i = 0; i < im.w; ++i){ + float pix = get_pixel(im, i, j, c); + pix = pix*v; + set_pixel(im, i, j, c, pix); + } + } +} + +void translate_image_channel(image im, int c, float v) +{ + int i, j; + for(j = 0; j < im.h; ++j){ + for(i = 0; i < im.w; ++i){ + float pix = get_pixel(im, i, j, c); + pix = pix+v; + set_pixel(im, i, j, c, pix); + } + } +} + +image binarize_image(image im) +{ + image c = copy_image(im); + int i; + for(i = 0; i < im.w * im.h * im.c; ++i){ + if(c.data[i] > .5) c.data[i] = 1; + else c.data[i] = 0; + } + return c; +} + +void saturate_image(image im, float sat) +{ + rgb_to_hsv(im); + scale_image_channel(im, 1, sat); + hsv_to_rgb(im); + constrain_image(im); +} + +void hue_image(image im, float hue) +{ + rgb_to_hsv(im); + int i; + for(i = 0; i < im.w*im.h; ++i){ + im.data[i] = im.data[i] + hue; + if (im.data[i] > 1) im.data[i] -= 1; + if (im.data[i] < 0) im.data[i] += 1; + } + hsv_to_rgb(im); + constrain_image(im); +} + +void exposure_image(image im, float sat) +{ + rgb_to_hsv(im); + scale_image_channel(im, 2, sat); + hsv_to_rgb(im); + constrain_image(im); +} + +void distort_image(image im, float hue, float sat, float val) +{ + if (im.c >= 3) + { + rgb_to_hsv(im); + scale_image_channel(im, 1, sat); + scale_image_channel(im, 2, val); + int i; + for(i = 0; i < im.w*im.h; ++i){ + im.data[i] = im.data[i] + hue; + if (im.data[i] > 1) im.data[i] -= 1; + if (im.data[i] < 0) im.data[i] += 1; + } + hsv_to_rgb(im); + } + else + { + scale_image_channel(im, 0, val); + } + constrain_image(im); +} + +void random_distort_image(image im, float hue, float saturation, float exposure) +{ + float dhue = rand_uniform_strong(-hue, hue); + float dsat = rand_scale(saturation); + float dexp = rand_scale(exposure); + distort_image(im, dhue, dsat, dexp); +} + +void saturate_exposure_image(image im, float sat, float exposure) +{ + rgb_to_hsv(im); + scale_image_channel(im, 1, sat); + scale_image_channel(im, 2, exposure); + hsv_to_rgb(im); + constrain_image(im); +} + +float bilinear_interpolate(image im, float x, float y, int c) +{ + int ix = (int) floorf(x); + int iy = (int) floorf(y); + + float dx = x - ix; + float dy = y - iy; + + float val = (1-dy) * (1-dx) * get_pixel_extend(im, ix, iy, c) + + dy * (1-dx) * get_pixel_extend(im, ix, iy+1, c) + + (1-dy) * dx * get_pixel_extend(im, ix+1, iy, c) + + dy * dx * get_pixel_extend(im, ix+1, iy+1, c); + return val; +} + +void quantize_image(image im) +{ + int size = im.c * im.w * im.h; + int i; + for (i = 0; i < size; ++i) im.data[i] = (int)(im.data[i] * 255) / 255. + (0.5/255); +} + +void make_image_red(image im) +{ + int r, c, k; + for (r = 0; r < im.h; ++r) { + for (c = 0; c < im.w; ++c) { + float val = 0; + for (k = 0; k < im.c; ++k) { + val += get_pixel(im, c, r, k); + set_pixel(im, c, r, k, 0); + } + for (k = 0; k < im.c; ++k) { + //set_pixel(im, c, r, k, val); + } + set_pixel(im, c, r, 0, val); + } + } +} + +image make_attention_image(int img_size, float *original_delta_cpu, float *original_input_cpu, int w, int h, int c) +{ + image attention_img; + attention_img.w = w; + attention_img.h = h; + attention_img.c = c; + attention_img.data = original_delta_cpu; + make_image_red(attention_img); + + int k; + float min_val = 999999, mean_val = 0, max_val = -999999; + for (k = 0; k < img_size; ++k) { + if (original_delta_cpu[k] < min_val) min_val = original_delta_cpu[k]; + if (original_delta_cpu[k] > max_val) max_val = original_delta_cpu[k]; + mean_val += original_delta_cpu[k]; + } + mean_val = mean_val / img_size; + float range = max_val - min_val; + + for (k = 0; k < img_size; ++k) { + float val = original_delta_cpu[k]; + val = fabs(mean_val - val) / range; + original_delta_cpu[k] = val * 4; + } + + image resized = resize_image(attention_img, w / 4, h / 4); + attention_img = resize_image(resized, w, h); + free_image(resized); + for (k = 0; k < img_size; ++k) attention_img.data[k] += original_input_cpu[k]; + + //normalize_image(attention_img); + //show_image(attention_img, "delta"); + return attention_img; +} + +image resize_image(image im, int w, int h) +{ + if (im.w == w && im.h == h) return copy_image(im); + + image resized = make_image(w, h, im.c); + image part = make_image(w, im.h, im.c); + int r, c, k; + float w_scale = (float)(im.w - 1) / (w - 1); + float h_scale = (float)(im.h - 1) / (h - 1); + for(k = 0; k < im.c; ++k){ + for(r = 0; r < im.h; ++r){ + for(c = 0; c < w; ++c){ + float val = 0; + if(c == w-1 || im.w == 1){ + val = get_pixel(im, im.w-1, r, k); + } else { + float sx = c*w_scale; + int ix = (int) sx; + float dx = sx - ix; + val = (1 - dx) * get_pixel(im, ix, r, k) + dx * get_pixel(im, ix+1, r, k); + } + set_pixel(part, c, r, k, val); + } + } + } + for(k = 0; k < im.c; ++k){ + for(r = 0; r < h; ++r){ + float sy = r*h_scale; + int iy = (int) sy; + float dy = sy - iy; + for(c = 0; c < w; ++c){ + float val = (1-dy) * get_pixel(part, c, iy, k); + set_pixel(resized, c, r, k, val); + } + if(r == h-1 || im.h == 1) continue; + for(c = 0; c < w; ++c){ + float val = dy * get_pixel(part, c, iy+1, k); + add_pixel(resized, c, r, k, val); + } + } + } + + free_image(part); + return resized; +} + + +void test_resize(char *filename) +{ + image im = load_image(filename, 0,0, 3); + float mag = mag_array(im.data, im.w*im.h*im.c); + printf("L2 Norm: %f\n", mag); + image gray = grayscale_image(im); + + image c1 = copy_image(im); + image c2 = copy_image(im); + image c3 = copy_image(im); + image c4 = copy_image(im); + distort_image(c1, .1, 1.5, 1.5); + distort_image(c2, -.1, .66666, .66666); + distort_image(c3, .1, 1.5, .66666); + distort_image(c4, .1, .66666, 1.5); + + + show_image(im, "Original"); + show_image(gray, "Gray"); + show_image(c1, "C1"); + show_image(c2, "C2"); + show_image(c3, "C3"); + show_image(c4, "C4"); + +#ifdef OPENCV + while(1){ + image aug = random_augment_image(im, 0, .75, 320, 448, 320); + show_image(aug, "aug"); + free_image(aug); + + + float exposure = 1.15; + float saturation = 1.15; + float hue = .05; + + image c = copy_image(im); + + float dexp = rand_scale(exposure); + float dsat = rand_scale(saturation); + float dhue = rand_uniform(-hue, hue); + + distort_image(c, dhue, dsat, dexp); + show_image(c, "rand"); + printf("%f %f %f\n", dhue, dsat, dexp); + free_image(c); + wait_until_press_key_cv(); + } +#endif +} + +image load_image_file(unsigned char *image_data, int channels, int antilog, int gray, int width, int height) +{ + int w, h, c; + unsigned char *data = image_data; + w = width; + h = height; + c = channels; + + if (!image_data) { + if (check_mistakes) getchar(); + return make_image(10, 10, 3); + + } + if (channels) c = channels; + + int i,j,k; + image im = make_image(w, h, c); + for(k = 0; k < c; ++k){ + for(j = 0; j < h; ++j){ + for(i = 0; i < w; ++i){ + int dst_index = i + w*j + w*h*k; + int src_index = k + c*i + c*w*j; + (im).data[dst_index] = (float)image_data[src_index]/255.; + } + } + } + //free(data); + return im; +} + +image load_image_new(unsigned char *image_data, int len, int channels, int antiLog, int gray, int width, int height) +{ + int w, h, c; + unsigned char *data=NULL; + if(antiLog || gray){ + data = image_data; + w = width; + h = height; + c = channels; + } + else + data = image_data; + + if (!data) { + if (check_mistakes) getchar(); + return make_image(10, 10, 3); + //exit(EXIT_FAILURE); + } + if (channels) c = channels; + ////printf("c:%d\n",c); + int i, j, k; + image im = make_image(w, h, c); + for (k = 0; k < c; ++k) { + for (j = 0; j < h; ++j) { + for (i = 0; i < w; ++i) { + int dst_index = i + w * j + w * h*k; + int src_index = k + c * i + c * w*j; + im.data[dst_index] = (float)data[src_index] / 255.; + } + } + } + //free(data); + return im; +} + + +image load_image_stb(char *filename, int channels) +{ + int w, h, c; + unsigned char *data = stbi_load(filename, &w, &h, &c, channels); + if (!data) { + char shrinked_filename[1024]; + if (strlen(filename) >= 1024) sprintf(shrinked_filename, "name is too long"); + else sprintf(shrinked_filename, "%s", filename); + fprintf(stderr, "Cannot load image \"%s\"\nSTB Reason: %s\n", shrinked_filename, stbi_failure_reason()); + FILE* fw = fopen("bad.list", "a"); + fwrite(shrinked_filename, sizeof(char), strlen(shrinked_filename), fw); + char *new_line = "\n"; + fwrite(new_line, sizeof(char), strlen(new_line), fw); + fclose(fw); + if (check_mistakes) { + printf("\n Error in load_image_stb() \n"); + getchar(); + } + return make_image(10, 10, 3); + //exit(EXIT_FAILURE); + } + if(channels) c = channels; + int i,j,k; + image im = make_image(w, h, c); + for(k = 0; k < c; ++k){ + for(j = 0; j < h; ++j){ + for(i = 0; i < w; ++i){ + int dst_index = i + w*j + w*h*k; + int src_index = k + c*i + c*w*j; + im.data[dst_index] = (float)data[src_index]/255.; + } + } + } + free(data); + return im; +} + +image load_image_stb_resize(char *filename, int w, int h, int c) +{ + image out = load_image_stb(filename, c); // without OpenCV + + if ((h && w) && (h != out.h || w != out.w)) { + image resized = resize_image(out, w, h); + free_image(out); + out = resized; + } + return out; +} + +image load_image(char *filename, int w, int h, int c) +{ +#ifdef OPENCV + //image out = load_image_stb(filename, c); + image out = load_image_cv(filename, c); +#else + image out = load_image_stb(filename, c); // without OpenCV +#endif // OPENCV + + if((h && w) && (h != out.h || w != out.w)){ + image resized = resize_image(out, w, h); + free_image(out); + out = resized; + } + return out; +} + +image load_image_color(char *filename, int w, int h) +{ + return load_image(filename, w, h, 3); +} + +image get_image_layer(image m, int l) +{ + image out = make_image(m.w, m.h, 1); + int i; + for(i = 0; i < m.h*m.w; ++i){ + out.data[i] = m.data[i+l*m.h*m.w]; + } + return out; +} + +void print_image(image m) +{ + int i, j, k; + for(i =0 ; i < m.c; ++i){ + for(j =0 ; j < m.h; ++j){ + for(k = 0; k < m.w; ++k){ + printf("%.2lf, ", m.data[i*m.h*m.w + j*m.w + k]); + if(k > 30) break; + } + printf("\n"); + if(j > 30) break; + } + printf("\n"); + } + printf("\n"); +} + +image collapse_images_vert(image *ims, int n) +{ + int color = 1; + int border = 1; + int h,w,c; + w = ims[0].w; + h = (ims[0].h + border) * n - border; + c = ims[0].c; + if(c != 3 || !color){ + w = (w+border)*c - border; + c = 1; + } + + image filters = make_image(w, h, c); + int i,j; + for(i = 0; i < n; ++i){ + int h_offset = i*(ims[0].h+border); + image copy = copy_image(ims[i]); + //normalize_image(copy); + if(c == 3 && color){ + embed_image(copy, filters, 0, h_offset); + } + else{ + for(j = 0; j < copy.c; ++j){ + int w_offset = j*(ims[0].w+border); + image layer = get_image_layer(copy, j); + embed_image(layer, filters, w_offset, h_offset); + free_image(layer); + } + } + free_image(copy); + } + return filters; +} + +image collapse_images_horz(image *ims, int n) +{ + int color = 1; + int border = 1; + int h,w,c; + int size = ims[0].h; + h = size; + w = (ims[0].w + border) * n - border; + c = ims[0].c; + if(c != 3 || !color){ + h = (h+border)*c - border; + c = 1; + } + + image filters = make_image(w, h, c); + int i,j; + for(i = 0; i < n; ++i){ + int w_offset = i*(size+border); + image copy = copy_image(ims[i]); + //normalize_image(copy); + if(c == 3 && color){ + embed_image(copy, filters, w_offset, 0); + } + else{ + for(j = 0; j < copy.c; ++j){ + int h_offset = j*(size+border); + image layer = get_image_layer(copy, j); + embed_image(layer, filters, w_offset, h_offset); + free_image(layer); + } + } + free_image(copy); + } + return filters; +} + +void show_image_normalized(image im, const char *name) +{ + image c = copy_image(im); + normalize_image(c); + show_image(c, name); + free_image(c); +} + +void show_images(image *ims, int n, char *window) +{ + image m = collapse_images_vert(ims, n); + /* + int w = 448; + int h = ((float)m.h/m.w) * 448; + if(h > 896){ + h = 896; + w = ((float)m.w/m.h) * 896; + } + image sized = resize_image(m, w, h); + */ + normalize_image(m); + save_image(m, window); + show_image(m, window); + free_image(m); +} + +void free_image(image m) +{ + if(m.data){ + free(m.data); + } +} + +// Fast copy data from a contiguous byte array into the image. +LIB_API void copy_image_from_bytes(image im, char *pdata) +{ + unsigned char *data = (unsigned char*)pdata; + int i, k, j; + int w = im.w; + int h = im.h; + int c = im.c; + for (k = 0; k < c; ++k) { + for (j = 0; j < h; ++j) { + for (i = 0; i < w; ++i) { + int dst_index = i + w * j + w * h*k; + int src_index = k + c * i + c * w*j; + im.data[dst_index] = (float)data[src_index] / 255.; + } + } + } +} + diff --git a/src/kernels/activation_mish.cu b/src/kernels/activation_mish.cu new file mode 100644 index 0000000..8900061 --- /dev/null +++ b/src/kernels/activation_mish.cu @@ -0,0 +1,50 @@ +#include "kernels.h" +#include + +#define MISH_THRESHOLD 20 + +__device__ +float tanh_activate_kernel(float x){return (2/(1 + expf(-2*x)) - 1);} + +__device__ +float softplus_kernel(float x, float threshold = 20) { + if (x > threshold) return x; // too large + else if (x < -threshold) return expf(x); // too small + return logf(expf(x) + 1); +} + + + +__device__ +float mish_yashas(float x) { + float e = __expf(x); + if (x <= -18.0f) + return x * e; + + float n = e * e + 2 * e; + if (x <= -5.0f) + return x * __fdividef(n, n + 2); + + return x - 2 * __fdividef(x, n + 2); +} + +// https://github.com/digantamisra98/Mish +// https://github.com/AlexeyAB/darknet/blob/master/src/activation_kernels.cu +__global__ +void activation_mish(dnnType *input, dnnType *output, int size) { + int i = (blockIdx.x + blockIdx.y*gridDim.x) * blockDim.x + threadIdx.x; + if (i < size) + // output[i] = input[i] * tanh_activate_kernel( softplus_kernel(input[i], MISH_THRESHOLD)); + output[i] = mish_yashas(input[i]); +} + +/** + Mish activation function +*/ +void activationMishForward(dnnType* srcData, dnnType* dstData, int size, cudaStream_t stream) +{ + int blocks = (size+255)/256; + int threads = 256; + + activation_mish<<>>(srcData, dstData, size); +} \ No newline at end of file diff --git a/src/kernels/activation_relu_ceiling.cu b/src/kernels/activation_relu_ceiling.cu new file mode 100644 index 0000000..72ccf4f --- /dev/null +++ b/src/kernels/activation_relu_ceiling.cu @@ -0,0 +1,33 @@ +#include "kernels.h" + +__global__ +void activation_relu_ceiling(dnnType *input, dnnType *output, int size, const float ceiling) { + + int i = blockDim.x*blockIdx.x + threadIdx.x; + + if(i0) + { + if (input[i]>ceiling) + output[i] = ceiling; + else + output[i] = input[i]; + } + else + output[i] = 0.0f; + } + } + + +/** + Relu ceiling activation function +*/ +void activationReLUCeilingForward(dnnType* srcData, dnnType* dstData, int size, const float ceiling, cudaStream_t stream) +{ + int blocks = (size+255)/256; + int threads = 256; + + activation_relu_ceiling<<>>(srcData, dstData, size, ceiling); +} + + diff --git a/src/kernels/activation_sigmoid.cu b/src/kernels/activation_sigmoid.cu index ba6c997..400a948 100644 --- a/src/kernels/activation_sigmoid.cu +++ b/src/kernels/activation_sigmoid.cu @@ -1,21 +1,13 @@ #include "kernels.h" - -__device__ -__forceinline__ -double sigmoid (double a) -{ - return 1.0 / (1.0 + exp (-a)); -} +#include __global__ void activation_sigmoid(dnnType *input, dnnType *output, int size) { - int stride = gridDim.x * blockDim.x; - int tid = blockDim.x * blockIdx.x + threadIdx.x; - for (int i = tid; i < size; i += stride) { - output[i] = sigmoid (input[i]); - } + int i = blockDim.x * blockIdx.x + threadIdx.x; + if(i < size) + output[i] = 1.0f / (1.0f + exp (-input[i])); } diff --git a/src/kernels/deformable_conv.cu b/src/kernels/deformable_conv.cu index cb9ded0..4dbc552 100644 --- a/src/kernels/deformable_conv.cu +++ b/src/kernels/deformable_conv.cu @@ -1,6 +1,8 @@ #include #include #include +#include +#include #include "kernels.h" #include @@ -9,93 +11,93 @@ i < (n); \ i += blockDim.x * gridDim.x) -const int CUDA_NUM_THREADS = 1024; +const int CUDA_NUM_THREADS = 512; inline int GET_BLOCKS(const int N) { return (N + CUDA_NUM_THREADS - 1) / CUDA_NUM_THREADS; } -__device__ float dmcn_im2col_bilinear(const float *bottom_data, const int data_width, - const int height, const int width, float h, float w) -{ - int h_low = floor(h); - int w_low = floor(w); - int h_high = h_low + 1; - int w_high = w_low + 1; +__device__ __host__ float dmcn_im2col_bilinear(const float *bottom_data, const int data_width, + const int height, const int width, float h, float w) { +int h_low = floor(h); +int w_low = floor(w); +int h_high = h_low + 1; +int w_high = w_low + 1; - float lh = h - h_low; - float lw = w - w_low; - float hh = 1 - lh, hw = 1 - lw; +float lh = h - h_low; +float lw = w - w_low; +float hh = 1 - lh, hw = 1 - lw; - float v1 = 0; - if (h_low >= 0 && w_low >= 0) - v1 = bottom_data[h_low * data_width + w_low]; - float v2 = 0; - if (h_low >= 0 && w_high <= width - 1) - v2 = bottom_data[h_low * data_width + w_high]; - float v3 = 0; - if (h_high <= height - 1 && w_low >= 0) - v3 = bottom_data[h_high * data_width + w_low]; - float v4 = 0; - if (h_high <= height - 1 && w_high <= width - 1) - v4 = bottom_data[h_high * data_width + w_high]; +float v1 = ( (h_low >= 0 && w_low >= 0) ? bottom_data[h_low * data_width + w_low]:0); +float v2 = ( (h_low >= 0 && w_high <= width - 1) ? bottom_data[h_low * data_width + w_high]:0); +float v3 = ( (h_high <= height - 1 && w_low >= 0) ? bottom_data[h_high * data_width + w_low]:0); +float v4 = ( (h_high <= height - 1 && w_high <= width - 1) ? bottom_data[h_high * data_width + w_high]:0); - float w1 = hh * hw, w2 = hh * lw, w3 = lh * hw, w4 = lh * lw; +float w1 = hh * hw, w2 = hh * lw, w3 = lh * hw, w4 = lh * lw; - float val = (w1 * v1 + w2 * v2 + w3 * v3 + w4 * v4); - return val; +float val = (w1 * v1 + w2 * v2 + w3 * v3 + w4 * v4); +return val; } __global__ void modulated_deformable_im2col_gpu_kernel(const int n, - const float *data_im, const float *data_offset, const float *data_mask, - const int height, const int width, const int kernel_h, const int kernel_w, - const int pad_h, const int pad_w, - const int stride_h, const int stride_w, - const int dilation_h, const int dilation_w, - const int channel_per_deformable_group, - const int batch_size, const int num_channels, const int deformable_group, - const int height_col, const int width_col, - float *data_col) -{ + const float *data_im, const float *data_offset, const float *data_mask, + const int height, const int width, + const int batch_size, const int num_channels, const int deformable_group, + const int height_col, const int width_col, + float *data_col) { CUDA_KERNEL_LOOP(index, n) { + //If n is a power of 2, ( i / n ) is equivalent to ( i ≫ log2 n ) and ( i % n ) is equivalent to ( i & n - 1 ). + const int ind_on_w = index / width_col; + const int ind_on_w_on_h = ind_on_w / height_col; + const int kk = 3 * 3; // index index of output matrix const int w_col = index % width_col; - const int h_col = (index / width_col) % height_col; - const int b_col = (index / width_col / height_col) % batch_size; - const int c_im = (index / width_col / height_col) / batch_size; - const int c_col = c_im * kernel_h * kernel_w; + const int h_col = (ind_on_w) % height_col; + const int b_col = (ind_on_w_on_h) % batch_size; + const int c_im = (ind_on_w_on_h) / batch_size; + const int c_col = c_im * kk; // compute deformable group index - const int deformable_group_index = c_im / channel_per_deformable_group; + const int deformable_group_index = c_im / (int)(num_channels / deformable_group); - const int h_in = h_col * stride_h - pad_h; - const int w_in = w_col * stride_w - pad_w; + const int h_in = h_col - 1; + const int w_in = w_col - 1; + const int s_col = height_col * width_col; + const int s_col2 = 2 * s_col; - float *data_col_ptr = data_col + ((c_col * batch_size + b_col) * height_col + h_col) * width_col + w_col; + const int first_member = w_col + width_col * h_col; + // float *data_col_ptr = data_col + ((c_col * batch_size + b_col) * height_col + h_col) * width_col + w_col; + float *data_col_ptr = data_col + first_member + s_col * (c_col * batch_size + b_col); //const float* data_im_ptr = data_im + ((b_col * num_channels + c_im) * height + h_in) * width + w_in; const float *data_im_ptr = data_im + (b_col * num_channels + c_im) * height * width; - const float *data_offset_ptr = data_offset + (b_col * deformable_group + deformable_group_index) * 2 * kernel_h * kernel_w * height_col * width_col; + const int add_ptr = (b_col * deformable_group + deformable_group_index) * kk * s_col; + const float *data_offset_ptr = data_offset + add_ptr + add_ptr; + const float *data_mask_ptr = data_mask + add_ptr; - const float *data_mask_ptr = data_mask + (b_col * deformable_group + deformable_group_index) * kernel_h * kernel_w * height_col * width_col; + #pragma unroll + for (int i = 0; i < 3; ++i) { + #pragma unroll + for (int j = 0; j < 3; ++j) { + const int iter_member = (i * 3 + j); + // const int data_offset_h_ptr = ((2 * (i * kernel_w + j)) * height_col + h_col) * width_col + w_col; + const int data_offset_h_ptr = first_member + s_col2 * iter_member; + + // const int data_offset_w_ptr = ((2 * (i * kernel_w + j) + 1) * height_col + h_col) * width_col + w_col; + const int data_offset_w_ptr = s_col + first_member + s_col2 * iter_member; + + // const int data_mask_hw_ptr = ((i * kernel_w + j) * height_col + h_col) * width_col + w_col; + const int data_mask_hw_ptr = first_member + s_col * iter_member; - for (int i = 0; i < kernel_h; ++i) - { - for (int j = 0; j < kernel_w; ++j) - { - const int data_offset_h_ptr = ((2 * (i * kernel_w + j)) * height_col + h_col) * width_col + w_col; - const int data_offset_w_ptr = ((2 * (i * kernel_w + j) + 1) * height_col + h_col) * width_col + w_col; - const int data_mask_hw_ptr = ((i * kernel_w + j) * height_col + h_col) * width_col + w_col; const float offset_h = data_offset_ptr[data_offset_h_ptr]; const float offset_w = data_offset_ptr[data_offset_w_ptr]; const float mask = data_mask_ptr[data_mask_hw_ptr]; - float val = static_cast(0); - const float h_im = h_in + i * dilation_h + offset_h; - const float w_im = w_in + j * dilation_w + offset_w; + const float h_im = offset_h + h_in + i; + const float w_im = offset_w + w_in + j; //if (h_im >= 0 && w_im >= 0 && h_im < height && w_im < width) { - if (h_im > -1 && w_im > -1 && h_im < height && w_im < width) - { + float val = static_cast(0); + if (h_im < height && w_im < width && h_im > -1 && w_im > -1) { //const float map_h = i * dilation_h + offset_h; //const float map_w = j * dilation_w + offset_w; //const int cur_height = height - h_in; @@ -104,15 +106,111 @@ __global__ void modulated_deformable_im2col_gpu_kernel(const int n, val = dmcn_im2col_bilinear(data_im_ptr, width, height, width, h_im, w_im); } *data_col_ptr = val * mask; - data_col_ptr += batch_size * height_col * width_col; + data_col_ptr += batch_size * s_col; //data_col_ptr += height_col * width_col; } } } } +__global__ void modulated_deformable_im2col_gpu_kernel_general_version(const int n, + const float *data_im, const float *data_offset, const float *data_mask, + const int height, const int width, const int kernel_h, const int kernel_w, + const int pad_h, const int pad_w, + const int stride_h, const int stride_w, + const int dilation_h, const int dilation_w, + const int channel_per_deformable_group, + const int batch_size, const int num_channels, const int deformable_group, + const int height_col, const int width_col, + float *data_col) { + CUDA_KERNEL_LOOP(index, n) + { + //If n is a power of 2, ( i / n ) is equivalent to ( i ≫ log2 n ) and ( i % n ) is equivalent to ( i & n - 1 ). + const int ind_on_w = index / width_col; + const int ind_on_w_on_h = ind_on_w / height_col; + const int kk = kernel_h * kernel_w; + // index index of output matrix + const int w_col = index % width_col; + const int h_col = (ind_on_w) % height_col; + const int b_col = (ind_on_w_on_h) % batch_size; + const int c_im = (ind_on_w_on_h) / batch_size; + const int c_col = c_im * kk; -void modulated_deformable_im2col_cuda(cudaStream_t stream, + // compute deformable group index + const int deformable_group_index = c_im / channel_per_deformable_group; + + const int h_in = h_col * stride_h - pad_h; + const int w_in = w_col * stride_w - pad_w; + const int s_col = height_col * width_col; + const int s_col2 = 2 * s_col; + + const int first_member = w_col + width_col * h_col; + // float *data_col_ptr = data_col + ((c_col * batch_size + b_col) * height_col + h_col) * width_col + w_col; + float *data_col_ptr = data_col + first_member + s_col * (c_col * batch_size + b_col); + //const float* data_im_ptr = data_im + ((b_col * num_channels + c_im) * height + h_in) * width + w_in; + const float *data_im_ptr = data_im + (b_col * num_channels + c_im) * height * width; + const int add_ptr = (b_col * deformable_group + deformable_group_index) * kk * s_col; + const float *data_offset_ptr = data_offset + add_ptr + add_ptr; + + const float *data_mask_ptr = data_mask + add_ptr; + + #pragma unroll + for (int i = 0; i < kernel_h; ++i) { + #pragma unroll + for (int j = 0; j < kernel_w; ++j) { + const int iter_member = (i * kernel_w + j); + // const int data_offset_h_ptr = ((2 * (i * kernel_w + j)) * height_col + h_col) * width_col + w_col; + const int data_offset_h_ptr = first_member + s_col2 * iter_member; + + // const int data_offset_w_ptr = ((2 * (i * kernel_w + j) + 1) * height_col + h_col) * width_col + w_col; + const int data_offset_w_ptr = s_col + first_member + s_col2 * iter_member; + + // const int data_mask_hw_ptr = ((i * kernel_w + j) * height_col + h_col) * width_col + w_col; + const int data_mask_hw_ptr = first_member + s_col * iter_member; + + const float offset_h = data_offset_ptr[data_offset_h_ptr]; + const float offset_w = data_offset_ptr[data_offset_w_ptr]; + const float mask = data_mask_ptr[data_mask_hw_ptr]; + const float h_im = offset_h + h_in + i * dilation_h; + const float w_im = offset_w + w_in + j * dilation_w; + //if (h_im >= 0 && w_im >= 0 && h_im < height && w_im < width) { + float val = static_cast(0); + if (h_im < height && w_im < width && h_im > -1 && w_im > -1) { + //const float map_h = i * dilation_h + offset_h; + //const float map_w = j * dilation_w + offset_w; + //const int cur_height = height - h_in; + //const int cur_width = width - w_in; + //val = dmcn_im2col_bilinear(data_im_ptr, width, cur_height, cur_width, map_h, map_w); + val = dmcn_im2col_bilinear(data_im_ptr, width, height, width, h_im, w_im); + } + *data_col_ptr = val * mask; + data_col_ptr += batch_size * s_col; + //data_col_ptr += height_col * width_col; + } + } + } +} + +void modulatedDeformableIm2colCuda(cudaStream_t stream, + const float* data_im, const float* data_offset, const float* data_mask, + const int batch_size, const int channels, const int height_im, const int width_im, + const int height_col, const int width_col, + const int deformable_group, float* data_col) { + // num_axes should be smaller than block size + // const int channel_per_deformable_group = channels / deformable_group; + const int num_kernels = channels * batch_size * height_col * width_col; + modulated_deformable_im2col_gpu_kernel + <<>>( + num_kernels, data_im, data_offset, data_mask, height_im, width_im, + batch_size, channels, deformable_group, height_col, width_col, data_col); + + cudaError_t err = cudaGetLastError(); + if (err != cudaSuccess) + FatalError("error in modulatedDeformableIm2colCuda: " + std::string(cudaGetErrorString(err)) + "\n"); +} + +void modulatedDeformableIm2colCudaGeneralVersion(cudaStream_t stream, const float* data_im, const float* data_offset, const float* data_mask, const int batch_size, const int channels, const int height_im, const int width_im, const int height_col, const int width_col, const int kernel_h, const int kenerl_w, @@ -122,7 +220,7 @@ void modulated_deformable_im2col_cuda(cudaStream_t stream, // num_axes should be smaller than block size const int channel_per_deformable_group = channels / deformable_group; const int num_kernels = channels * batch_size * height_col * width_col; - modulated_deformable_im2col_gpu_kernel + modulated_deformable_im2col_gpu_kernel_general_version <<>>( num_kernels, data_im, data_offset, data_mask, height_im, width_im, kernel_h, kenerl_w, @@ -131,14 +229,11 @@ void modulated_deformable_im2col_cuda(cudaStream_t stream, cudaError_t err = cudaGetLastError(); if (err != cudaSuccess) - { - printf("error in modulated_deformable_im2col_cuda: %s\n", cudaGetErrorString(err)); - } - + FatalError("error in modulatedDeformableIm2colCudaGeneralVersion: " + std::string(cudaGetErrorString(err)) + "\n"); } - -void dcn_v2_cuda_forward(float *input, float *weight, +void dcnV2CudaForward(cublasStatus_t stat, cublasHandle_t handle, + float *input, float *weight, float *bias, float *ones, float *offset, float *mask, float *output, float *columns, @@ -146,24 +241,17 @@ void dcn_v2_cuda_forward(float *input, float *weight, const int stride_h, const int stride_w, const int pad_h, const int pad_w, const int dilation_h, const int dilation_w, - const int deformable_group, + const int deformable_group, const int batch_id, const int in_n, const int in_c, const int in_h, const int in_w, const int out_n, const int out_c, const int out_h, const int out_w, const int chunk_dim, cudaStream_t stream) { - cublasStatus_t stat; - cublasHandle_t handle; - stat = cublasCreate(&handle); - if (stat != CUBLAS_STATUS_SUCCESS) { - printf ("CUBLAS initialization failed\n"); - return; - } - + // stat and handle have be moved out to preserve 2 - 6 milliseconds every 100. + const int batch = batch_id; const int channels = in_c; const int height = in_h; const int width = in_w; - const int channels_out = out_c; const int height_out = (height + 2 * pad_h - (dilation_h * (kernel_h - 1) + 1)) / stride_h + 1; @@ -178,19 +266,23 @@ void dcn_v2_cuda_forward(float *input, float *weight, stat = cublasSgemm(handle, CUBLAS_OP_T, CUBLAS_OP_N, n, m, k, &alpha, ones, k, bias, k, - &beta, output, n); - if (stat != CUBLAS_STATUS_SUCCESS) { - printf ("CUBLAS initialization failed\n"); - return ; - } + &beta, output + batch * out_c * out_h * out_w, n); + if (stat != CUBLAS_STATUS_SUCCESS) + FatalError("CUBLAS initialization failed\n"); - modulated_deformable_im2col_cuda(stream, - input, offset, - mask, + modulatedDeformableIm2colCuda(stream, + input + batch * channels * height * width, + offset,// + b * 2 * int((float)chunk_dim / batch), + mask,// + b * int((float)chunk_dim / batch), 1, channels, height, width, - height_out, width_out, kernel_h, kernel_w, - pad_h, pad_w, stride_h, stride_w, dilation_h, dilation_w, - deformable_group, columns); + height_out, width_out, deformable_group, columns); + // modulatedDeformableIm2colCudaGeneralVersion(stream, + // input, offset, + // mask, + // 1, channels, height, width, + // height_out, width_out, kernel_h, kernel_w, + // pad_h, pad_w, stride_h, stride_w, dilation_h, dilation_w, + // deformable_group, columns); //(k * m) x (m * n) // Y = WC @@ -200,10 +292,9 @@ void dcn_v2_cuda_forward(float *input, float *weight, stat = cublasSgemm(handle, CUBLAS_OP_N, CUBLAS_OP_N, n, m, k, &alpha, columns, n, weight, k, - &beta, output, n); + &beta, output + batch * out_c * out_h * out_w, n); + + if (stat != CUBLAS_STATUS_SUCCESS) + FatalError("CUBLAS initialization failed\n"); - if (stat != CUBLAS_STATUS_SUCCESS) { - printf ("CUBLAS initialization failed\n"); - return ; - } } diff --git a/src/kernels/normalize.cu b/src/kernels/normalize.cu new file mode 100644 index 0000000..5206256 --- /dev/null +++ b/src/kernels/normalize.cu @@ -0,0 +1,16 @@ +#include "kernelsThrust.h" + +__global__ +void normalize_kernel(float *bgr, const int dim, const float *mean, const float *stddev){ + int i = blockDim.x*blockIdx.x + threadIdx.x; + int j = blockIdx.y; + bgr[j*(dim)+i] = bgr[j*(dim)+i] - mean[j]; + bgr[j*(dim)+i] = bgr[j*(dim)+i] / stddev[j]; + +} + +void normalize(float *bgr, const int ch, const int h, const int w, const float *mean, const float *stddev){ + int num_thread = 256; + dim3 dimBlock(h*w/num_thread, ch); + normalize_kernel<<>>(bgr, h*w, mean, stddev); +} \ No newline at end of file diff --git a/src/kernels/pooling.cu b/src/kernels/pooling.cu new file mode 100644 index 0000000..cd80a6d --- /dev/null +++ b/src/kernels/pooling.cu @@ -0,0 +1,52 @@ +#include "kernels.h" + +__global__ void forward_maxpool_layer_kernel(int n, int in_h, int in_w, int in_c, int stride_x, int stride_y, int size, int pad, float *input, float *output) +{ + int h = (in_h + pad - size) / stride_y + 1; + int w = (in_w + pad - size) / stride_x + 1; + int c = in_c; + + int id = (blockIdx.x + blockIdx.y*gridDim.x) * blockDim.x + threadIdx.x; + if(id >= n) return; + + int j = id % w; + id /= w; + int i = id % h; + id /= h; + int k = id % c; + id /= c; + int b = id; + + int w_offset = -pad / 2; + int h_offset = -pad / 2; + + int out_index = j + w*(i + h*(k + c*b)); + float max = -9999999; + int max_i = -1; + int l, m; + for(l = 0; l < size; ++l){ + for(m = 0; m < size; ++m){ + int cur_h = h_offset + i*stride_y + l; + int cur_w = w_offset + j*stride_x + m; + int index = cur_w + in_w*(cur_h + in_h*(k + b*in_c)); + int valid = (cur_h >= 0 && cur_h < in_h && + cur_w >= 0 && cur_w < in_w); + float val = (valid != 0) ? input[index] : -9999999; + max_i = (val > max) ? index : max_i; + max = (val > max) ? val : max; + } + } + output[out_index] = max; +} + +void MaxPoolingForward(dnnType* srcData, dnnType* dstData, int n, int c, int h, int w, int stride_x, int stride_y, int size, int padding, cudaStream_t stream) +{ + + int tot_size = n*c*h*w; + + int blocks = (tot_size+255)/256; + int threads = 256; + + forward_maxpool_layer_kernel<<>>(tot_size, h, w, c, stride_x, stride_y, size, padding, srcData, dstData); +} + diff --git a/src/sorting.cu b/src/kernels/postprocessing.cu similarity index 72% rename from src/sorting.cu rename to src/kernels/postprocessing.cu index e007be1..3510200 100644 --- a/src/sorting.cu +++ b/src/kernels/postprocessing.cu @@ -1,61 +1,37 @@ +#include "kernelsThrust.h" -#include "sorting.h" -void sort(dnnType *src_begin, dnnType *src_end, int *idsrc) -{ +void subtractWithThreshold(dnnType *src_begin, dnnType *src_end, dnnType *src2_begin, dnnType *src_out, struct threshold op){ + thrust::transform(thrust::device, src_begin, src_end, src2_begin, src_out, op); +} + +void sort(dnnType *src_begin, dnnType *src_end, int *idsrc){ thrust::sort_by_key(thrust::device, src_begin, src_end, idsrc, thrust::greater()); - // thrust::stable_sort_by_key(thrust::device, // src_begin, src_end, idsrc, // thrust::greater()); } void topk(dnnType *src_begin, int *idsrc, int K, float *topk_scores, - int *topk_inds, float *topk_ys, float *topk_xs) -{ + int *topk_inds, float *topk_ys, float *topk_xs){ checkCuda( cudaMemcpy(topk_scores, (float *)src_begin, K*sizeof(float), cudaMemcpyDeviceToDevice) ); - checkCuda( cudaMemcpy(topk_inds, idsrc, K*sizeof(int), cudaMemcpyDeviceToDevice) ); - // topk_ys_[i*K +count] = (int)(ids2[j] / width); - // topk_xs_[i*K +count] = (int)(ids2[j] % width); - + checkCuda( cudaMemcpy(topk_inds, idsrc, K*sizeof(int), cudaMemcpyDeviceToDevice) ); } __global__ void sortAndTopK_kernel(dnnType *src_begin, int *idsrc, float *topk_scores, int *topk_inds, float *topk_ys, float *topk_xs,const int size, const int K){ int i = blockDim.x*blockIdx.x + threadIdx.x; - thrust::sort_by_key(thrust::device, src_begin + i * size, src_begin + i * size + size, idsrc + i * size, thrust::greater()); thrust::copy_n(thrust::device, src_begin + i * size, K, topk_scores + i * K); - // thrust::copy_n(thrust::device, idsrc + i * size, K, topk_inds + i * K ); thrust::copy_n(thrust::device, idsrc + i * size, K, topk_inds + i * K ); } -void sortAndTopKonDevice(dnnType *src_begin, int *idsrc, float *topk_scores, int *topk_inds, float *topk_ys, float *topk_xs, const int size, const int K, const int n_classes) -{ +void sortAndTopKonDevice(dnnType *src_begin, int *idsrc, float *topk_scores, int *topk_inds, float *topk_ys, float *topk_xs, const int size, const int K, const int n_classes){ int blocks = n_classes; int threads = 1; - - sortAndTopK_kernel<<>>(src_begin, idsrc, topk_scores, topk_inds, topk_ys, topk_xs, size, K); - -} - -struct threshold : public thrust::binary_function -{ - __host__ __device__ - float operator()(float x, float y) { - float toll = 1e-6; - if(fabsf(x-y)>toll) - return 0.0f; - else - return x; - } -}; - -void subtractWithThreshold(dnnType *src_begin, dnnType *src_end, dnnType *src2_begin, dnnType *src_out){ - struct threshold op; - thrust::transform(thrust::device, src_begin, src_end, src2_begin, src_out, op); + sortAndTopK_kernel<<>>(src_begin, idsrc, topk_scores, topk_inds, topk_ys, topk_xs, size, K); } void topKxyclasses(int *ids_begin, int *ids_end, const int K, const int size, const int wh, int *clses, int *xs, int *ys){ @@ -63,34 +39,27 @@ void topKxyclasses(int *ids_begin, int *ids_end, const int K, const int size, co thrust::transform(thrust::device, ids_begin, ids_end, thrust::make_constant_iterator(wh), ids_begin, thrust::modulus()); thrust::transform(thrust::device, ids_begin, ids_end, thrust::make_constant_iterator(size), ys, thrust::divides()); thrust::transform(thrust::device, ids_begin, ids_end, thrust::make_constant_iterator(size), xs, thrust::modulus()); - } -void topKxyAddOffset(int * ids_begin, const int K, const int size, int *intxs_begin, int *intys_begin, float *xs_begin, float *ys_begin, dnnType *src_begin){ - float *src_out; - checkCuda( cudaMalloc(&src_out, K *sizeof(float)) ); +void topKxyAddOffset(int * ids_begin, const int K, const int size, + int *intxs_begin, int *intys_begin, float *xs_begin, + float *ys_begin, dnnType *src_begin, float *src_out, int *ids_out){ thrust::gather(thrust::device, ids_begin, ids_begin + K, src_begin, src_out); thrust::transform(thrust::device, intxs_begin, intxs_begin + K, src_out, xs_begin, thrust::plus()); - int *ids_out; - checkCuda( cudaMalloc(&ids_out, K *sizeof(int)) ); thrust::transform(thrust::device, ids_begin, ids_begin + K, thrust::make_constant_iterator(size), ids_out, thrust::plus()); thrust::gather(thrust::device, ids_out, ids_out+K, src_begin, src_out); thrust::transform(thrust::device, intys_begin, intys_begin + K, src_out, ys_begin, thrust::plus()); - checkCuda( cudaFree(src_out) ); - checkCuda( cudaFree(ids_out) ); } -void bboxes(int * ids_begin, const int K, const int size, float *xs_begin, float *ys_begin, dnnType *src_begin, float *bbx0, float *bbx1, float *bby0, float *bby1){ - float *src_out; - checkCuda( cudaMalloc(&src_out, K *sizeof(float)) ); +void bboxes(int * ids_begin, const int K, const int size, float *xs_begin, float *ys_begin, + dnnType *src_begin, float *bbx0, float *bbx1, float *bby0, float *bby1, + float *src_out, int *ids_out){ thrust::gather(thrust::device, ids_begin, ids_begin + K, src_begin, src_out); thrust::transform(thrust::device, src_out, src_out + K, thrust::make_constant_iterator(2), src_out, thrust::divides()); // x0 thrust::transform(thrust::device, xs_begin, xs_begin + K, src_out, bbx0, thrust::minus()); // x1 thrust::transform(thrust::device, xs_begin, xs_begin + K, src_out, bbx1, thrust::plus()); - int *ids_out; - checkCuda( cudaMalloc(&ids_out, K *sizeof(int)) ); thrust::transform(thrust::device, ids_begin, ids_begin + K, thrust::make_constant_iterator(size), ids_out, thrust::plus()); thrust::gather(thrust::device, ids_out, ids_out + K, src_begin, src_out); thrust::transform(thrust::device, src_out, src_out + K, thrust::make_constant_iterator(2), src_out, thrust::divides()); @@ -98,7 +67,5 @@ void bboxes(int * ids_begin, const int K, const int size, float *xs_begin, float thrust::transform(thrust::device, ys_begin, ys_begin + K, src_out, bby0, thrust::minus()); // y1 thrust::transform(thrust::device, ys_begin, ys_begin + K, src_out, bby1, thrust::plus()); - checkCuda( cudaFree(src_out) ); - checkCuda( cudaFree(ids_out) ); } diff --git a/src/kernels/scaladd.cu b/src/kernels/scaladd.cu new file mode 100644 index 0000000..53bcd7a --- /dev/null +++ b/src/kernels/scaladd.cu @@ -0,0 +1,16 @@ +#include "kernels.h" +#include + +__global__ void scal_add_kernel(dnnType* dstData, int size, float alpha, float beta, int inc) +{ + int i = (blockIdx.x + blockIdx.y*gridDim.x) * blockDim.x + threadIdx.x; + if (i < size) dstData[i*inc] = dstData[i*inc] * alpha + beta; +} + +void scalAdd(dnnType* dstData, int size, float alpha, float beta, int inc, cudaStream_t stream) +{ + int blocks = (size+255)/256; + int threads = 256; + + scal_add_kernel<<>>(dstData, size, alpha, beta, inc); +} \ No newline at end of file diff --git a/src/utils.cpp b/src/utils.cpp index deb4def..fa6458f 100644 --- a/src/utils.cpp +++ b/src/utils.cpp @@ -20,46 +20,65 @@ bool fileExist(const char *fname) { return true; } +void downloadWeightsifDoNotExist(const std::string& input_bin, const std::string& test_folder, const std::string& weights_url){ + if(!fileExist(input_bin.c_str())){ + std::string mkdir_cmd = "mkdir " + test_folder; + std::string wget_cmd = "curl " + weights_url + " --output " + test_folder + "/weights.zip"; +#ifdef __linux__ + std::string unzip_cmd = "unzip " + test_folder + "/weights.zip -d" + test_folder; + std::string rm_cmd = "rm " + test_folder + "/weights.zip"; -void readBinaryFile(std::string fname, int size, dnnType** data_h, dnnType** data_d, int seek, bool skipLoad) +#elif _WIN32 + + std::string unzip_cmd = "7z x " + test_folder + "/weights.zip -o" + test_folder; +#endif + int err = 0; + err = system(mkdir_cmd.c_str()); + err = system(wget_cmd.c_str()); + err = system(unzip_cmd.c_str()); +#ifdef __linux__ + err = system(rm_cmd.c_str()); +#endif + + } +} + + +void readBinaryFile(std::string fname, int size, dnnType** data_h, dnnType** data_d, int seek) { + std::ifstream dataFile (fname, std::ios::in | std::ios::binary); + std::stringstream error_s; + if (!dataFile) + { + error_s << "Error opening file " << fname; + FatalError(error_s.str()); + } + + if(seek != 0) { + dataFile.seekg(seek*sizeof(dnnType), dataFile.cur); + } + int size_b = size*sizeof(dnnType); *data_h = new dnnType[size]; - - if(!skipLoad) { - std::ifstream dataFile(fname, std::ios::in | std::ios::binary); - std::stringstream error_s; - if (!dataFile) { - error_s << "Error opening file " << fname; - FatalError(error_s.str()); - } - - if (seek != 0) { - dataFile.seekg(seek * sizeof(dnnType), dataFile.cur); - } - - // printf("data_h %d size_b %d\n", *data_h,size_b); - if (!dataFile.read((char *) *data_h, size_b)) { - - error_s << "Error reading file " << fname; - FatalError(error_s.str()); - } - } else { - std::cout<> pid >> comm >> state >> ppid >> pgrp >> session >> tty_nr + >> tpgid >> flags >> minflt >> cminflt >> majflt >> cmajflt + >> utime >> stime >> cutime >> cstime >> priority >> nice + >> O >> itrealvalue >> starttime >> vsize >> rss; + + stat_stream.close(); +#ifdef __linux__ + long page_size_kb = sysconf(_SC_PAGE_SIZE) / 1024; // in case x86-64 is configured to use 2MB pages +#elif _WIN32 + long page_size_kb = 4096/1024; +#endif + + vm_usage_kb = vsize / 1024.0; + resident_set_kb = rss * page_size_kb; +} + +void printCudaMemUsage() { + size_t free, total; + checkCuda( cudaMemGetInfo(&free, &total) ); + std::cout<<"GPU free memory: "< -#include "tkdnn.h" - -const char *input_bin = "../tests/mnist/input.bin"; -const char *c0_bin = "../tests/mnist/layers/c0.bin"; -const char *c1_bin = "../tests/mnist/layers/c1.bin"; -const char *d2_bin = "../tests/mnist/layers/d2.bin"; -const char *d3_bin = "../tests/mnist/layers/d3.bin"; -const char *output_bin = "../tests/mnist/output.bin"; - -int main() { - - // Network layout - tk::dnn::dataDim_t dim(1, 1, 28, 28, 1); - tk::dnn::Network net(dim); - tk::dnn::Conv2d l0(&net, 20, 5, 5, 1, 1, 0, 0, c0_bin); - tk::dnn::Pooling l1(&net, 2, 2, 2, 2, tk::dnn::POOLING_MAX); - tk::dnn::Conv2d l2(&net, 50, 5, 5, 1, 1, 0, 0, c1_bin); - tk::dnn::Pooling l3(&net, 2, 2, 2, 2, tk::dnn::POOLING_MAX); - tk::dnn::Dense l4(&net, 500, d2_bin); - tk::dnn::Activation l5(&net, tk::dnn::ACTIVATION_LEAKY); - tk::dnn::Dense l6(&net, 10, d3_bin); - tk::dnn::Softmax l7(&net); - - tk::dnn::NetworkRT netRT(&net, "mnist.rt"); - - // Load input - dnnType *data; - dnnType *input_h; - readBinaryFile(input_bin, dim.tot(), &input_h, &data); - - dnnType *out_data, *out_data2; - - std::cout<<"CUDNN inference:\n"; { - dim.print(); //print initial dimension - TIMER_START - out_data = net.infer(dim, data); - TIMER_STOP - dim.print(); - } - - // Print result - //std::cout<<"\n======= CUDNN RESULT =======\n"; - //printDeviceVector(10, out_data); - - tk::dnn::dataDim_t dim2(1, 1, 28, 28, 1); - - std::cout<<"TENSORRT inference:\n"; { - dim2.print(); - TIMER_START - out_data2 = netRT.infer(dim2, data); - TIMER_STOP - dim2.print(); - } - - // Print result - //std::cout<<"\n======= TENRT RESULT =======\n"; - //printDeviceVector(10, out_data); - - std::cout<<"\n======= CHECK RESULT =======\n"; - checkResult(dim.tot(), out_data, out_data2); - - /* - // Print real test - std::cout<<"\n==== CHECK RESULT ====\n"; - dnnType *out; - dnnType *out_h; - readBinaryFile(output_bin, dim.tot(), &out_h, &out); - printDeviceVector(dim.tot(), out); -*/ - return 0; -} diff --git a/tests/mnist/test_mnistRT.cpp b/tests/mnist/test_mnistRT.cpp deleted file mode 100644 index c337097..0000000 --- a/tests/mnist/test_mnistRT.cpp +++ /dev/null @@ -1,178 +0,0 @@ -#include -#include -#include "tkdnn.h" -#include "NvInfer.h" - -const char *input_bin = "../tests/mnist/input.bin"; -const char *c0_bin = "../tests/mnist/layers/c0.bin"; -const char *c1_bin = "../tests/mnist/layers/c1.bin"; -const char *d2_bin = "../tests/mnist/layers/d2.bin"; -const char *d3_bin = "../tests/mnist/layers/d3.bin"; -const char *output_bin = "../tests/mnist/output.bin"; - -using namespace nvinfer1; - -// Logger for info/warning/errors -class Logger : public ILogger -{ - void log(Severity severity, const char* msg) override - { - // suppress info-level messages - if (severity != Severity::kINFO) - std::cout << msg << std::endl; - } -} gLogger; - -int main() { - - std::cout<<"\n==== CUDNN ====\n"; - // Network layout - tk::dnn::dataDim_t dim(1, 1, 28, 28, 1); - tk::dnn::Network net(dim); - tk::dnn::Conv2d l0(&net, 20, 5, 5, 1, 1, 0, 0, c0_bin); - tk::dnn::Pooling l1(&net, 2, 2, 2, 2, tk::dnn::POOLING_MAX); - tk::dnn::Conv2d l2(&net, 50, 5, 5, 1, 1, 0, 0, c1_bin); - tk::dnn::Pooling l3(&net, 2, 2, 2, 2, tk::dnn::POOLING_MAX); - tk::dnn::Dense l4(&net, 500, d2_bin); - tk::dnn::Activation l5(&net, CUDNN_ACTIVATION_RELU); - tk::dnn::Dense l6(&net, 10, d3_bin); - tk::dnn::Softmax l7(&net); - - // Load input - dnnType *data; - dnnType *input_h; - readBinaryFile(input_bin, dim.tot(), &input_h, &data); - - dim.print(); //print initial dimension - - // Inference - { - TIMER_START - data = net.infer(dim, data); - TIMER_STOP - dim.print(); - } - - // Print real test - std::cout<<"\n==== CHECK CUDNN RESULT ====\n"; - dnnType *out; - dnnType *out_h; - readBinaryFile(output_bin, dim.tot(), &out_h, &out); - std::cout<<"Diff: "<createNetwork(); - - DataType dt = DataType::kFLOAT; - // Create input of shape { 1, 1, 28, 28 } with name referenced by "data" - auto input = network->addInput("data", dt, DimsCHW{ 1, 28, 28}); - assert(input != nullptr); - - tk::dnn::Conv2d *c0 = &l0; - Weights w { dt, c0->data_h, c0->inputs*c0->outputs*c0->kernelH*c0->kernelW}; - Weights b { dt, c0->bias_h, c0->outputs}; - // Add a convolution layer with 20 outputs and a 5x5 filter. - auto conv1 = network->addConvolution(*input, 20, DimsHW{5, 5}, w, b); - assert(conv1 != nullptr); - conv1->setStride(DimsHW{1, 1}); - - // Add a max pooling layer with stride of 2x2 and kernel size of 2x2. - auto pool1 = network->addPooling(*conv1->getOutput(0), PoolingType::kMAX, DimsHW{2, 2}); - assert(pool1 != nullptr); - pool1->setStride(DimsHW{2, 2}); - - tk::dnn::Conv2d *c1 = &l2; - Weights w1 { dt, c1->data_h, c1->inputs*c1->outputs*c1->kernelH*c1->kernelW}; - Weights b1 { dt, c1->bias_h, c1->outputs}; - // Add a second convolution layer with 50 outputs and a 5x5 filter. - auto conv2 = network->addConvolution(*pool1->getOutput(0), 50, DimsHW{5, 5}, w1, b1); - assert(conv2 != nullptr); - conv2->setStride(DimsHW{1, 1}); - - // Add a second max pooling layer with stride of 2x2 and kernel size of 2x3> - auto pool2 = network->addPooling(*conv2->getOutput(0), PoolingType::kMAX, DimsHW{2, 2}); - assert(pool2 != nullptr); - pool2->setStride(DimsHW{2, 2}); - - tk::dnn::Dense *d2 = &l4; - Weights w2 { dt, d2->data_h, d2->inputs*d2->outputs}; - Weights b2 { dt, d2->bias_h, d2->outputs}; - // Add a fully connected layer with 500 outputs. - auto ip1 = network->addFullyConnected(*pool2->getOutput(0), 500, w2, b2); - assert(ip1 != nullptr); - - // Add an activation layer using the ReLU algorithm. - auto relu1 = network->addActivation(*ip1->getOutput(0), ActivationType::kRELU); - assert(relu1 != nullptr); - - tk::dnn::Dense *d3 = &l6; - Weights w3 { dt, d3->data_h, d3->inputs*d3->outputs}; - Weights b3 { dt, d3->bias_h, d3->outputs}; - // Add a second fully connected layer with 20 outputs. - auto ip2 = network->addFullyConnected(*relu1->getOutput(0), 10, w3, b3); - assert(ip2 != nullptr); - - // Add a softmax layer to determine the probability. - auto prob = network->addSoftMax(*ip2->getOutput(0)); - assert(prob != nullptr); - prob->getOutput(0)->setName("out"); - - network->markOutput(*prob->getOutput(0)); - - // Build the engine - builder->setMaxBatchSize(1); - builder->setMaxWorkspaceSize(1 << 20); - - auto engine = builder->buildCudaEngine(*network); - // we don't need the network any more - network->destroy(); - - IExecutionContext *context = engine->createExecutionContext(); - - // run inference - // input and output buffer pointers that we pass to the engine - the engine requires exactly IEngine::getNbBindings(), - // of these, but in this case we know that there is exactly one input and one output. - assert(engine->getNbBindings() == 2); - void* buffers[2]; - - // In order to bind the buffers, we need to know the names of the input and output tensors. - // note that indices are guaranteed to be less than IEngine::getNbBindings() - int inputIndex = engine->getBindingIndex("data"); - int outputIndex = engine->getBindingIndex("out"); - - float output[10]; - // create GPU buffers and a stream - checkCuda(cudaMalloc(&buffers[inputIndex], 28*28*sizeof(float))); - checkCuda(cudaMalloc(&buffers[outputIndex], 10*sizeof(float))); - - cudaStream_t stream; - checkCuda(cudaStreamCreate(&stream)); - - // DMA the input to the GPU, execute the batch asynchronously, and DMA it back: - { - checkCuda(cudaMemcpyAsync(buffers[inputIndex], input_h, 1 * 28*28* sizeof(float), cudaMemcpyHostToDevice, stream)); - cudaStreamSynchronize(stream); //want to test only the inference time - TIMER_START - context->enqueue(1, buffers, stream, nullptr); - TIMER_STOP - checkCuda(cudaMemcpyAsync(output, buffers[outputIndex],10*sizeof(float), cudaMemcpyDeviceToHost, stream)); - cudaStreamSynchronize(stream); - } - - std::cout<<"\n==== CHECK CUDNN RESULT ====\n"; - std::cout<<"Diff: "<destroy(); - engine->destroy(); - - return 0; -} diff --git a/tests/resnet101/resnet101.cpp b/tests/resnet101/resnet101.cpp deleted file mode 100644 index 5fe6bbe..0000000 --- a/tests/resnet101/resnet101.cpp +++ /dev/null @@ -1,338 +0,0 @@ -#include -#include "tkdnn.h" - -const char *input_bin = "../tests/resnet101/debug/input.bin"; -const char *conv1_bin = "../tests/resnet101/layers/conv1.bin"; - -//layer1 -const char *layer1_bin[]={ -"../tests/resnet101/layers/layer1-0-conv1.bin", -"../tests/resnet101/layers/layer1-0-conv2.bin", -"../tests/resnet101/layers/layer1-0-conv3.bin", -"../tests/resnet101/layers/layer1-0-downsample-0.bin", - -"../tests/resnet101/layers/layer1-1-conv1.bin", -"../tests/resnet101/layers/layer1-1-conv2.bin", -"../tests/resnet101/layers/layer1-1-conv3.bin", - -"../tests/resnet101/layers/layer1-2-conv1.bin", -"../tests/resnet101/layers/layer1-2-conv2.bin", -"../tests/resnet101/layers/layer1-2-conv3.bin"}; - - -//layer2 -const char *layer2_bin[]={ -"../tests/resnet101/layers/layer2-0-conv1.bin", -"../tests/resnet101/layers/layer2-0-conv2.bin", -"../tests/resnet101/layers/layer2-0-conv3.bin", -"../tests/resnet101/layers/layer2-0-downsample-0.bin", - -"../tests/resnet101/layers/layer2-1-conv1.bin", -"../tests/resnet101/layers/layer2-1-conv2.bin", -"../tests/resnet101/layers/layer2-1-conv3.bin", - -"../tests/resnet101/layers/layer2-2-conv1.bin", -"../tests/resnet101/layers/layer2-2-conv2.bin", -"../tests/resnet101/layers/layer2-2-conv3.bin", - -"../tests/resnet101/layers/layer2-3-conv1.bin", -"../tests/resnet101/layers/layer2-3-conv2.bin", -"../tests/resnet101/layers/layer2-3-conv3.bin" -}; -//layer3 -const char *layer3_bin[]={ -"../tests/resnet101/layers/layer3-0-conv1.bin", -"../tests/resnet101/layers/layer3-0-conv2.bin", -"../tests/resnet101/layers/layer3-0-conv3.bin", -"../tests/resnet101/layers/layer3-0-downsample-0.bin", - -"../tests/resnet101/layers/layer3-1-conv1.bin", -"../tests/resnet101/layers/layer3-1-conv2.bin", -"../tests/resnet101/layers/layer3-1-conv3.bin", - -"../tests/resnet101/layers/layer3-2-conv1.bin", -"../tests/resnet101/layers/layer3-2-conv2.bin", -"../tests/resnet101/layers/layer3-2-conv3.bin", - -"../tests/resnet101/layers/layer3-3-conv1.bin", -"../tests/resnet101/layers/layer3-3-conv2.bin", -"../tests/resnet101/layers/layer3-3-conv3.bin", - -"../tests/resnet101/layers/layer3-4-conv1.bin", -"../tests/resnet101/layers/layer3-4-conv2.bin", -"../tests/resnet101/layers/layer3-4-conv3.bin", - -"../tests/resnet101/layers/layer3-5-conv1.bin", -"../tests/resnet101/layers/layer3-5-conv2.bin", -"../tests/resnet101/layers/layer3-5-conv3.bin", - -"../tests/resnet101/layers/layer3-6-conv1.bin", -"../tests/resnet101/layers/layer3-6-conv2.bin", -"../tests/resnet101/layers/layer3-6-conv3.bin", - -"../tests/resnet101/layers/layer3-7-conv1.bin", -"../tests/resnet101/layers/layer3-7-conv2.bin", -"../tests/resnet101/layers/layer3-7-conv3.bin", - -"../tests/resnet101/layers/layer3-8-conv1.bin", -"../tests/resnet101/layers/layer3-8-conv2.bin", -"../tests/resnet101/layers/layer3-8-conv3.bin", - -"../tests/resnet101/layers/layer3-9-conv1.bin", -"../tests/resnet101/layers/layer3-9-conv2.bin", -"../tests/resnet101/layers/layer3-9-conv3.bin", - -"../tests/resnet101/layers/layer3-10-conv1.bin", -"../tests/resnet101/layers/layer3-10-conv2.bin", -"../tests/resnet101/layers/layer3-10-conv3.bin", - -"../tests/resnet101/layers/layer3-11-conv1.bin", -"../tests/resnet101/layers/layer3-11-conv2.bin", -"../tests/resnet101/layers/layer3-11-conv3.bin", - -"../tests/resnet101/layers/layer3-12-conv1.bin", -"../tests/resnet101/layers/layer3-12-conv2.bin", -"../tests/resnet101/layers/layer3-12-conv3.bin", - -"../tests/resnet101/layers/layer3-13-conv1.bin", -"../tests/resnet101/layers/layer3-13-conv2.bin", -"../tests/resnet101/layers/layer3-13-conv3.bin", - -"../tests/resnet101/layers/layer3-14-conv1.bin", -"../tests/resnet101/layers/layer3-14-conv2.bin", -"../tests/resnet101/layers/layer3-14-conv3.bin", - -"../tests/resnet101/layers/layer3-15-conv1.bin", -"../tests/resnet101/layers/layer3-15-conv2.bin", -"../tests/resnet101/layers/layer3-15-conv3.bin", - -"../tests/resnet101/layers/layer3-16-conv1.bin", -"../tests/resnet101/layers/layer3-16-conv2.bin", -"../tests/resnet101/layers/layer3-16-conv3.bin", - -"../tests/resnet101/layers/layer3-17-conv1.bin", -"../tests/resnet101/layers/layer3-17-conv2.bin", -"../tests/resnet101/layers/layer3-17-conv3.bin", - -"../tests/resnet101/layers/layer3-18-conv1.bin", -"../tests/resnet101/layers/layer3-18-conv2.bin", -"../tests/resnet101/layers/layer3-18-conv3.bin", - -"../tests/resnet101/layers/layer3-19-conv1.bin", -"../tests/resnet101/layers/layer3-19-conv2.bin", -"../tests/resnet101/layers/layer3-19-conv3.bin", - -"../tests/resnet101/layers/layer3-20-conv1.bin", -"../tests/resnet101/layers/layer3-20-conv2.bin", -"../tests/resnet101/layers/layer3-20-conv3.bin", - -"../tests/resnet101/layers/layer3-21-conv1.bin", -"../tests/resnet101/layers/layer3-21-conv2.bin", -"../tests/resnet101/layers/layer3-21-conv3.bin", - -"../tests/resnet101/layers/layer3-22-conv1.bin", -"../tests/resnet101/layers/layer3-22-conv2.bin", -"../tests/resnet101/layers/layer3-22-conv3.bin"}; - - -//layer4 -const char *layer4_bin[]={ -"../tests/resnet101/layers/layer4-0-conv1.bin", -"../tests/resnet101/layers/layer4-0-conv2.bin", -"../tests/resnet101/layers/layer4-0-conv3.bin", -"../tests/resnet101/layers/layer4-0-downsample-0.bin", - -"../tests/resnet101/layers/layer4-1-conv1.bin", -"../tests/resnet101/layers/layer4-1-conv2.bin", -"../tests/resnet101/layers/layer4-1-conv3.bin", - -"../tests/resnet101/layers/layer4-2-conv1.bin", -"../tests/resnet101/layers/layer4-2-conv2.bin", -"../tests/resnet101/layers/layer4-2-conv3.bin"}; - -//final -const char *fc_bin = "../tests/resnet101/layers/fc.bin"; - -const char *output_bin = "../tests/resnet101/debug/fc.bin"; - -int main() -{ - - // Network layout - tk::dnn::dataDim_t dim(1, 3, 224, 224, 1); - tk::dnn::Network net(dim); - - tk::dnn::Conv2d conv1(&net, 64, 7, 7, 2, 2, 3, 3, conv1_bin, true); - tk::dnn::Activation relu3(&net, CUDNN_ACTIVATION_RELU); - tk::dnn::Pooling maxpool4(&net, 3, 3, 2, 2, 1, 1, tk::dnn::POOLING_MAX); - - //layer 1 - int id_layer1_bin = 0; - tk::dnn::Layer *last = &maxpool4; - for(int i=0; i<3;i++) - { - tk::dnn::Conv2d *layer1_0_conv1 = new tk::dnn::Conv2d(&net, 64, 1, 1, 1, 1, 0, 0, layer1_bin[id_layer1_bin++], true); - tk::dnn::Activation *relu1_0_1 = new tk::dnn::Activation(&net, CUDNN_ACTIVATION_RELU); - tk::dnn::Conv2d *layer1_0_conv2 = new tk::dnn::Conv2d(&net, 64, 3, 3, 1, 1, 1, 1, layer1_bin[id_layer1_bin++], true); - tk::dnn::Activation *relu1_0_2 = new tk::dnn::Activation(&net, CUDNN_ACTIVATION_RELU); - tk::dnn::Conv2d *layer1_0_conv3 = new tk::dnn::Conv2d(&net, 256, 1, 1, 1, 1, 0, 0, layer1_bin[id_layer1_bin++], true); - if(i==0) { - tk::dnn::Layer *route_1_0_layers[1] = { last }; - tk::dnn::Route *route_1_0 = new tk::dnn::Route(&net, route_1_0_layers, 1); - tk::dnn::Conv2d *layer1_0_downsample_0 = new tk::dnn::Conv2d(&net, 256, 1, 1, 1, 1, 0, 0, layer1_bin[id_layer1_bin++], true); - tk::dnn::Shortcut *s1_0 = new tk::dnn::Shortcut(&net, layer1_0_conv3); - } else { - tk::dnn::Shortcut *s1_0 = new tk::dnn::Shortcut(&net, last); - } - tk::dnn::Activation *layer1_0_relu = new tk::dnn::Activation(&net, CUDNN_ACTIVATION_RELU); - last = layer1_0_relu; - } - - // tk::dnn::Activation *last_activation = (tk::dnn::Activation *) net.layers[net.num_layers-1]; - // layer 2 - int id_layer2_bin = 0; - for(int i=0; i<4;i++) - { - tk::dnn::Conv2d *layer1_0_conv1 = new tk::dnn::Conv2d(&net, 128, 1, 1, 1, 1, 0, 0, layer2_bin[id_layer2_bin++], true); - tk::dnn::Activation *relu1_0_1 = new tk::dnn::Activation(&net, CUDNN_ACTIVATION_RELU); - tk::dnn::Conv2d *layer1_0_conv2; - if(i==0) - layer1_0_conv2 = new tk::dnn::Conv2d(&net, 128, 3, 3, 2, 2, 1, 1, layer2_bin[id_layer2_bin++], true); - else - layer1_0_conv2 = new tk::dnn::Conv2d(&net, 128, 3, 3, 1, 1, 1, 1, layer2_bin[id_layer2_bin++], true); - - tk::dnn::Activation *relu1_0_2 = new tk::dnn::Activation(&net, CUDNN_ACTIVATION_RELU); - tk::dnn::Conv2d *layer1_0_conv3 = new tk::dnn::Conv2d(&net, 512, 1, 1, 1, 1, 0, 0, layer2_bin[id_layer2_bin++], true); - if(i==0) - { - tk::dnn::Layer *route_1_0_layers[1] = { last }; - tk::dnn::Route *route_1_0 = new tk::dnn::Route(&net, route_1_0_layers, 1); - tk::dnn::Conv2d *layer1_0_downsample_0 = new tk::dnn::Conv2d(&net, 512, 1, 1, 2, 2, 0, 0, layer2_bin[id_layer2_bin++], true); - tk::dnn::Shortcut *s1_0 = new tk::dnn::Shortcut(&net, layer1_0_conv3); - } - else - { - tk::dnn::Shortcut *s1_0 = new tk::dnn::Shortcut(&net, last); - } - tk::dnn::Activation *layer1_0_relu = new tk::dnn::Activation(&net, CUDNN_ACTIVATION_RELU); - last = layer1_0_relu; - } - - // layer 3 - int id_layer3_bin = 0; - for(int i=0; i<23;i++) - { - tk::dnn::Conv2d *layer1_0_conv1 = new tk::dnn::Conv2d(&net, 256, 1, 1, 1, 1, 0, 0, layer3_bin[id_layer3_bin++], true); - tk::dnn::Activation *relu1_0_1 = new tk::dnn::Activation(&net, CUDNN_ACTIVATION_RELU); - tk::dnn::Conv2d *layer1_0_conv2; - if(i==0) - layer1_0_conv2 = new tk::dnn::Conv2d(&net, 256, 3, 3, 2, 2, 1, 1, layer3_bin[id_layer3_bin++], true); - else - layer1_0_conv2 = new tk::dnn::Conv2d(&net, 256, 3, 3, 1, 1, 1, 1, layer3_bin[id_layer3_bin++], true); - - tk::dnn::Activation *relu1_0_2 = new tk::dnn::Activation(&net, CUDNN_ACTIVATION_RELU); - tk::dnn::Conv2d *layer1_0_conv3 = new tk::dnn::Conv2d(&net, 1024, 1, 1, 1, 1, 0, 0, layer3_bin[id_layer3_bin++], true); - if(i==0) - { - tk::dnn::Layer *route_1_0_layers[1] = { last }; - tk::dnn::Route *route_1_0 = new tk::dnn::Route(&net, route_1_0_layers, 1); - tk::dnn::Conv2d *layer1_0_downsample_0 = new tk::dnn::Conv2d(&net, 1024, 1, 1, 2, 2, 0, 0, layer3_bin[id_layer3_bin++], true); - tk::dnn::Shortcut *s1_0 = new tk::dnn::Shortcut(&net, layer1_0_conv3); - } - else - { - tk::dnn::Shortcut *s1_0 = new tk::dnn::Shortcut(&net, last); - } - tk::dnn::Activation *layer1_0_relu = new tk::dnn::Activation(&net, CUDNN_ACTIVATION_RELU); - last = layer1_0_relu; - } - - // layer 4 - int id_layer4_bin = 0; - for(int i=0; i<3;i++) - { - tk::dnn::Conv2d *layer1_0_conv1 = new tk::dnn::Conv2d(&net, 512, 1, 1, 1, 1, 0, 0, layer4_bin[id_layer4_bin++], true); - tk::dnn::Activation *relu1_0_1 = new tk::dnn::Activation(&net, CUDNN_ACTIVATION_RELU); - tk::dnn::Conv2d *layer1_0_conv2; - if(i==0) - layer1_0_conv2 = new tk::dnn::Conv2d(&net, 512, 3, 3, 2, 2, 1, 1, layer4_bin[id_layer4_bin++], true); - else - layer1_0_conv2 = new tk::dnn::Conv2d(&net, 512, 3, 3, 1, 1, 1, 1, layer4_bin[id_layer4_bin++], true); - - tk::dnn::Activation *relu1_0_2 = new tk::dnn::Activation(&net, CUDNN_ACTIVATION_RELU); - tk::dnn::Conv2d *layer1_0_conv3 = new tk::dnn::Conv2d(&net, 2048, 1, 1, 1, 1, 0, 0, layer4_bin[id_layer4_bin++], true); - if(i==0) - { - tk::dnn::Layer *route_1_0_layers[1] = { last }; - tk::dnn::Route *route_1_0 = new tk::dnn::Route(&net, route_1_0_layers, 1); - tk::dnn::Conv2d *layer1_0_downsample_0 = new tk::dnn::Conv2d(&net, 2048, 1, 1, 2, 2, 0, 0, layer4_bin[id_layer4_bin++], true); - tk::dnn::Shortcut *s1_0 = new tk::dnn::Shortcut(&net, layer1_0_conv3); - } - else - { - tk::dnn::Shortcut *s1_0 = new tk::dnn::Shortcut(&net, last); - } - tk::dnn::Activation *layer1_0_relu = new tk::dnn::Activation(&net, CUDNN_ACTIVATION_RELU); - last = layer1_0_relu; - } - - //final - tk::dnn::Pooling avgpool(&net, 7, 7, 7, 7, 0, 0, tk::dnn::POOLING_AVERAGE); - tk::dnn::Dense fc(&net, 1000, fc_bin); - - // Load input - dnnType *data; - dnnType *input_h; - readBinaryFile(input_bin, dim.tot(), &input_h, &data); - //printDeviceVector(64, data, true); - - //print network model - net.print(); - - //convert network to tensorRT - tk::dnn::NetworkRT netRT(&net, "resnet101.rt"); - - - tk::dnn::dataDim_t out_dim; - out_dim = net.layers[net.num_layers-1]->output_dim; - dnnType *cudnn_out, *rt_out; - - tk::dnn::dataDim_t dim1 = dim; //input dim - printCenteredTitle(" CUDNN inference ", '=', 30); - { - dim1.print(); - TIMER_START - net.infer(dim1, data); - TIMER_STOP - dim1.print(); - } - cudnn_out = net.layers[net.num_layers-1]->dstData; - - //printDeviceVector(64, cudnn_out, true); - - tk::dnn::dataDim_t dim2 = dim; - printCenteredTitle(" TENSORRT inference ", '=', 30); - { - dim2.print(); - TIMER_START - netRT.infer(dim2, data); - TIMER_STOP - dim2.print(); - } - rt_out = (dnnType *)netRT.buffersRT[1]; - - - printCenteredTitle(std::string(" RESNET CHECK RESULTS ").c_str(), '=', 30); - dnnType *out, *out_h; - int odim = out_dim.tot(); - readBinaryFile(output_bin, odim, &out_h, &out); - std::cout << "CUDNN vs correct"; - checkResult(odim, cudnn_out, out); - - std::cout << "TRT vs correct"; - checkResult(odim, rt_out, out); - std::cout << "CUDNN vs TRT "; - checkResult(odim, cudnn_out, rt_out); - - return 0; -} diff --git a/tests/resnet101/resnet101_weightsexporter.py b/tests/resnet101/resnet101_weightsexporter.py deleted file mode 100644 index 140154e..0000000 --- a/tests/resnet101/resnet101_weightsexporter.py +++ /dev/null @@ -1,163 +0,0 @@ -import torch -import urllib -from PIL import Image -from torchvision import transforms -from torchsummary import summary -import numpy as np -import struct - -import torch.nn as nn - -def bin_write(f, data): - data =data.flatten() - # print(data) - fmt = 'f'*len(data) - bin = struct.pack(fmt, *data) - f.write(bin) - -def hook(module, input, output): - setattr(module, "_value_hook", output) - - - -def print_wb(model, folder): - for name, param in model.named_parameters(): - # print ("Layer", name) - t = name.split('.')[0:-1] - arg = name.split('.')[-1] - t = '-'.join(t) - print (" type: ", t) - - if arg == 'weight': - w = param.data.numpy() - print (" weights shape:", np.shape(w)) - w.tofile(folder + "/" + t + ".bin", format="f") - elif arg == 'bias': - b = param.data.numpy() - print (" bias shape:", np.shape(b)) - b.tofile(folder + "/" + t + ".bias.bin", format="f") - else: - print("Ops!") - - -def print_wb_output(model, input_batch): - for n, m in model.named_modules(): - m.register_forward_hook(hook) - - model(input_batch) - i = input_batch.data.numpy() - i = np.array(i, dtype=np.float32) - print(i.shape) - i.tofile("debug/input.bin", format="f") - - f = None - for n, m in model.named_modules(): - in_output = m._value_hook - o = in_output.data.numpy() - o = np.array(o, dtype=np.float32) - t = '-'.join(n.split('.')) - o.tofile("debug/" + t + ".bin", format="f") - - if not(' of Conv2d' in str(m.type) or ' of Linear' in str(m.type) or ' of BatchNorm2d' in str(m.type)): - continue - - if ' of Conv2d' in str(m.type) or ' of Linear' in str(m.type): - file_name = "layers/" + t + ".bin" - print("open file: ", file_name) - f = open(file_name, mode='wb') - - print(n, ' ----------------------------------------------------------------') - # print(m._parameters) - #print(m.type) - - w = np.array([]) - b = np.array([]) - - - if 'weight' in m._parameters and m._parameters['weight'] is not None: - w = m._parameters['weight'].data.numpy() - w = np.array(w, dtype=np.float32) - print (" weights shape:", np.shape(w)) - - if 'bias' in m._parameters and m._parameters['bias'] is not None: - b = m._parameters['bias'].data.numpy() - b = np.array(b, dtype=np.float32) - print (" bias shape:", np.shape(b)) - # else: - # b = np.zeros(w.shape[0], dtype=np.float32) - # print (" bias shape:", np.shape(b)) - - if 'BatchNorm2d' in str(m.type): - b = m._parameters['bias'].data.numpy() - b = np.array(b, dtype=np.float32) - s = m._parameters['weight'].data.numpy() - s = np.array(s, dtype=np.float32) - rm = m.running_mean.data.numpy() - rm = np.array(rm, dtype=np.float32) - rv = m.running_var.data.numpy() - rv = np.array(rv, dtype=np.float32) - #s.tofile(f, format="f") - bin_write(f,b) - bin_write(f,s) - bin_write(f,rm) - bin_write(f,rv) - print (" s shape:", np.shape(s)) - print (" rm shape:", np.shape(rm)) - print (" rv shape:", np.shape(rv)) - - else: - - # w.tofile(f, format="f") - bin_write(f,w) - # print("w- ",w) - if b.size > 0: - # b.tofile(f, format="f") - bin_write(f,b) - # print("b - ",b) - - if ' of BatchNorm2d' in str(m.type) or ' of Linear' in str(m.type): - f.close() - print("close file") - f = None - # return - - - - - -if __name__ == '__main__': - - model = torch.hub.load('pytorch/vision', 'resnet101', pretrained=True) - model.eval() - - # Download an example image from the pytorch website - url, filename = ("https://github.com/pytorch/hub/raw/master/dog.jpg", "dog.jpg") - try: urllib.URLopener().retrieve(url, filename) - except: urllib.request.urlretrieve(url, filename) - - # sample execution (requires torchvision) - input_image = Image.open(filename) - preprocess = transforms.Compose([ - transforms.Resize(256), - transforms.CenterCrop(224), - transforms.ToTensor(), - transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]), - ]) - input_tensor = preprocess(input_image) - input_batch = input_tensor.unsqueeze(0) # create a mini-batch as expected by the model - - # move the input and model to GPU for speed if available - if torch.cuda.is_available(): - input_batch = input_batch.to('cuda') - model.to('cuda') - - with torch.no_grad(): - output = model(input_batch) - - # Tensor of shape 1000, with confidence scores over Imagenet's 1000 classes - # print(output) - - - print_wb_output(model, input_batch) - - # print(list(model.children())) diff --git a/tests/resnet101_cnet/resnet101_cnet.cpp b/tests/resnet101_cnet/resnet101_cnet.cpp deleted file mode 100644 index 061e4b6..0000000 --- a/tests/resnet101_cnet/resnet101_cnet.cpp +++ /dev/null @@ -1,1086 +0,0 @@ -#include - -#include "kernels.h" -#include "Yolo3Detection.h" -#include "tkdnn.h" -#include -#include // std::iota -#include // std::sort -// #include "utils.h" - -const char *input_bin = "../tests/resnet101_cnet/debug/input.bin"; -const char *conv1_bin = "../tests/resnet101_cnet/layers/conv1.bin"; - -//layer1 -const char *layer1_bin[]={ -"../tests/resnet101_cnet/layers/layer1-0-conv1.bin", -"../tests/resnet101_cnet/layers/layer1-0-conv2.bin", -"../tests/resnet101_cnet/layers/layer1-0-conv3.bin", -"../tests/resnet101_cnet/layers/layer1-0-downsample-0.bin", - -"../tests/resnet101_cnet/layers/layer1-1-conv1.bin", -"../tests/resnet101_cnet/layers/layer1-1-conv2.bin", -"../tests/resnet101_cnet/layers/layer1-1-conv3.bin", - -"../tests/resnet101_cnet/layers/layer1-2-conv1.bin", -"../tests/resnet101_cnet/layers/layer1-2-conv2.bin", -"../tests/resnet101_cnet/layers/layer1-2-conv3.bin"}; - - -//layer2 -const char *layer2_bin[]={ -"../tests/resnet101_cnet/layers/layer2-0-conv1.bin", -"../tests/resnet101_cnet/layers/layer2-0-conv2.bin", -"../tests/resnet101_cnet/layers/layer2-0-conv3.bin", -"../tests/resnet101_cnet/layers/layer2-0-downsample-0.bin", - -"../tests/resnet101_cnet/layers/layer2-1-conv1.bin", -"../tests/resnet101_cnet/layers/layer2-1-conv2.bin", -"../tests/resnet101_cnet/layers/layer2-1-conv3.bin", - -"../tests/resnet101_cnet/layers/layer2-2-conv1.bin", -"../tests/resnet101_cnet/layers/layer2-2-conv2.bin", -"../tests/resnet101_cnet/layers/layer2-2-conv3.bin", - -"../tests/resnet101_cnet/layers/layer2-3-conv1.bin", -"../tests/resnet101_cnet/layers/layer2-3-conv2.bin", -"../tests/resnet101_cnet/layers/layer2-3-conv3.bin" -}; -//layer3 -const char *layer3_bin[]={ -"../tests/resnet101_cnet/layers/layer3-0-conv1.bin", -"../tests/resnet101_cnet/layers/layer3-0-conv2.bin", -"../tests/resnet101_cnet/layers/layer3-0-conv3.bin", -"../tests/resnet101_cnet/layers/layer3-0-downsample-0.bin", - -"../tests/resnet101_cnet/layers/layer3-1-conv1.bin", -"../tests/resnet101_cnet/layers/layer3-1-conv2.bin", -"../tests/resnet101_cnet/layers/layer3-1-conv3.bin", - -"../tests/resnet101_cnet/layers/layer3-2-conv1.bin", -"../tests/resnet101_cnet/layers/layer3-2-conv2.bin", -"../tests/resnet101_cnet/layers/layer3-2-conv3.bin", - -"../tests/resnet101_cnet/layers/layer3-3-conv1.bin", -"../tests/resnet101_cnet/layers/layer3-3-conv2.bin", -"../tests/resnet101_cnet/layers/layer3-3-conv3.bin", - -"../tests/resnet101_cnet/layers/layer3-4-conv1.bin", -"../tests/resnet101_cnet/layers/layer3-4-conv2.bin", -"../tests/resnet101_cnet/layers/layer3-4-conv3.bin", - -"../tests/resnet101_cnet/layers/layer3-5-conv1.bin", -"../tests/resnet101_cnet/layers/layer3-5-conv2.bin", -"../tests/resnet101_cnet/layers/layer3-5-conv3.bin", - -"../tests/resnet101_cnet/layers/layer3-6-conv1.bin", -"../tests/resnet101_cnet/layers/layer3-6-conv2.bin", -"../tests/resnet101_cnet/layers/layer3-6-conv3.bin", - -"../tests/resnet101_cnet/layers/layer3-7-conv1.bin", -"../tests/resnet101_cnet/layers/layer3-7-conv2.bin", -"../tests/resnet101_cnet/layers/layer3-7-conv3.bin", - -"../tests/resnet101_cnet/layers/layer3-8-conv1.bin", -"../tests/resnet101_cnet/layers/layer3-8-conv2.bin", -"../tests/resnet101_cnet/layers/layer3-8-conv3.bin", - -"../tests/resnet101_cnet/layers/layer3-9-conv1.bin", -"../tests/resnet101_cnet/layers/layer3-9-conv2.bin", -"../tests/resnet101_cnet/layers/layer3-9-conv3.bin", - -"../tests/resnet101_cnet/layers/layer3-10-conv1.bin", -"../tests/resnet101_cnet/layers/layer3-10-conv2.bin", -"../tests/resnet101_cnet/layers/layer3-10-conv3.bin", - -"../tests/resnet101_cnet/layers/layer3-11-conv1.bin", -"../tests/resnet101_cnet/layers/layer3-11-conv2.bin", -"../tests/resnet101_cnet/layers/layer3-11-conv3.bin", - -"../tests/resnet101_cnet/layers/layer3-12-conv1.bin", -"../tests/resnet101_cnet/layers/layer3-12-conv2.bin", -"../tests/resnet101_cnet/layers/layer3-12-conv3.bin", - -"../tests/resnet101_cnet/layers/layer3-13-conv1.bin", -"../tests/resnet101_cnet/layers/layer3-13-conv2.bin", -"../tests/resnet101_cnet/layers/layer3-13-conv3.bin", - -"../tests/resnet101_cnet/layers/layer3-14-conv1.bin", -"../tests/resnet101_cnet/layers/layer3-14-conv2.bin", -"../tests/resnet101_cnet/layers/layer3-14-conv3.bin", - -"../tests/resnet101_cnet/layers/layer3-15-conv1.bin", -"../tests/resnet101_cnet/layers/layer3-15-conv2.bin", -"../tests/resnet101_cnet/layers/layer3-15-conv3.bin", - -"../tests/resnet101_cnet/layers/layer3-16-conv1.bin", -"../tests/resnet101_cnet/layers/layer3-16-conv2.bin", -"../tests/resnet101_cnet/layers/layer3-16-conv3.bin", - -"../tests/resnet101_cnet/layers/layer3-17-conv1.bin", -"../tests/resnet101_cnet/layers/layer3-17-conv2.bin", -"../tests/resnet101_cnet/layers/layer3-17-conv3.bin", - -"../tests/resnet101_cnet/layers/layer3-18-conv1.bin", -"../tests/resnet101_cnet/layers/layer3-18-conv2.bin", -"../tests/resnet101_cnet/layers/layer3-18-conv3.bin", - -"../tests/resnet101_cnet/layers/layer3-19-conv1.bin", -"../tests/resnet101_cnet/layers/layer3-19-conv2.bin", -"../tests/resnet101_cnet/layers/layer3-19-conv3.bin", - -"../tests/resnet101_cnet/layers/layer3-20-conv1.bin", -"../tests/resnet101_cnet/layers/layer3-20-conv2.bin", -"../tests/resnet101_cnet/layers/layer3-20-conv3.bin", - -"../tests/resnet101_cnet/layers/layer3-21-conv1.bin", -"../tests/resnet101_cnet/layers/layer3-21-conv2.bin", -"../tests/resnet101_cnet/layers/layer3-21-conv3.bin", - -"../tests/resnet101_cnet/layers/layer3-22-conv1.bin", -"../tests/resnet101_cnet/layers/layer3-22-conv2.bin", -"../tests/resnet101_cnet/layers/layer3-22-conv3.bin"}; - - -//layer4 -const char *layer4_bin[]={ -"../tests/resnet101_cnet/layers/layer4-0-conv1.bin", -"../tests/resnet101_cnet/layers/layer4-0-conv2.bin", -"../tests/resnet101_cnet/layers/layer4-0-conv3.bin", -"../tests/resnet101_cnet/layers/layer4-0-downsample-0.bin", - -"../tests/resnet101_cnet/layers/layer4-1-conv1.bin", -"../tests/resnet101_cnet/layers/layer4-1-conv2.bin", -"../tests/resnet101_cnet/layers/layer4-1-conv3.bin", - -"../tests/resnet101_cnet/layers/layer4-2-conv1.bin", -"../tests/resnet101_cnet/layers/layer4-2-conv2.bin", -"../tests/resnet101_cnet/layers/layer4-2-conv3.bin"}; - -const char *d_conv1_bin = "../tests/resnet101_cnet/layers/deconv_layers-0-conv_offset_mask.bin"; -const char *deform1_bin = "../tests/resnet101_cnet/layers/deconv_layers-0.bin"; -const char *deconv1_bin = "../tests/resnet101_cnet/layers/deconv_layers-3.bin"; - -const char *d_conv2_bin = "../tests/resnet101_cnet/layers/deconv_layers-6-conv_offset_mask.bin"; -const char *deform2_bin = "../tests/resnet101_cnet/layers/deconv_layers-6.bin"; -const char *deconv2_bin = "../tests/resnet101_cnet/layers/deconv_layers-9.bin"; - -const char *d_conv3_bin = "../tests/resnet101_cnet/layers/deconv_layers-12-conv_offset_mask.bin"; -const char *deform3_bin = "../tests/resnet101_cnet/layers/deconv_layers-12.bin"; -const char *deconv3_bin = "../tests/resnet101_cnet/layers/deconv_layers-15.bin"; - -const char *hm_conv1_bin = "../tests/resnet101_cnet/layers/hm-0.bin"; -const char *hm_conv2_bin = "../tests/resnet101_cnet/layers/hm-2.bin"; -const char *wh_conv1_bin = "../tests/resnet101_cnet/layers/wh-0.bin"; -const char *wh_conv2_bin = "../tests/resnet101_cnet/layers/wh-2.bin"; -const char *reg_conv1_bin = "../tests/resnet101_cnet/layers/reg-0.bin"; -const char *reg_conv2_bin = "../tests/resnet101_cnet/layers/reg-2.bin"; -//final -const char *fc_bin = "../tests/resnet101_cnet/layers/fc.bin"; - -const char *output_bin[]={ -"../tests/resnet101_cnet/debug/hm.bin", -"../tests/resnet101_cnet/debug/wh.bin", -"../tests/resnet101_cnet/debug/reg.bin"}; - - - -std::vector sort_indexes(const std::vector &v) { - - // initialize original index locations - std::vector idx(v.size()); - iota(idx.begin(), idx.end(), 0); - - // sort indexes based on comparing values in v - sort(idx.begin(), idx.end(), - [&v](size_t i1, size_t i2) {return v[i1] > v[i2];}); - - return idx; -} - -float _colors[6][3] = { {1,0,1}, {0,0,1},{0,1,1},{0,1,0},{1,1,0},{1,0,0} }; -float get_color(int c, int x, int max) -{ - float ratio = ((float)x/max)*5; - int i = floor(ratio); - int j = ceil(ratio); - ratio -= i; - float r = (1-ratio) * _colors[i % 6][c % 3] + ratio*_colors[j % 6][c % 3]; - //printf("%f\n", r); - return r; -} - -int computeDetections(dnnType *hm_d, dnnType *wh_d, dnnType *reg_d, int hm_dim, int wh_dim, int reg_dim, bool cat_spec_wh, int k){ - // _nms - int kernel = 3; - int pad = (kernel - 1)/2; - std::cout<<"computeDetections\n"; - // dnnType *hmax; - // tk::dnn::Pooling maxpool(&hmax, 3, 3, 2, 2, 1, 1, tk::dnn::POOLING_MAX) - // = (dnnType *) - - // net.functional.max_pool2d( - // heat, (kernel, kernel), stride=1, padding=pad) - // keep = (hmax == heat).float() - // return heat * keep -} - -int process(dnnType *hm_d, dnnType *wh_d, dnnType *reg_d, int hm_dim, int wh_dim, int reg_dim){ - std::cout<<"process\n"; - // computeDetections(hm_d, wh_d, reg_d, hm_dim, wh_dim, reg_dim, false, 100); -} - -int main() -{ - - // Network layout - tk::dnn::dataDim_t dim(1, 3, 224, 224, 1); - tk::dnn::Network net(dim); - - tk::dnn::Conv2d conv1(&net, 64, 7, 7, 2, 2, 3, 3, conv1_bin, true); - tk::dnn::Activation relu3(&net, CUDNN_ACTIVATION_RELU); - - tk::dnn::Pooling maxpool4(&net, 3, 3, 2, 2, 1, 1, tk::dnn::POOLING_MAX); - - - //layer 1 - int id_layer1_bin = 0; - tk::dnn::Layer *last = &maxpool4; - for(int i=0; i<3;i++) - { - tk::dnn::Conv2d *layer1_0_conv1 = new tk::dnn::Conv2d(&net, 64, 1, 1, 1, 1, 0, 0, layer1_bin[id_layer1_bin++], true); - tk::dnn::Activation *relu1_0_1 = new tk::dnn::Activation(&net, CUDNN_ACTIVATION_RELU); - tk::dnn::Conv2d *layer1_0_conv2 = new tk::dnn::Conv2d(&net, 64, 3, 3, 1, 1, 1, 1, layer1_bin[id_layer1_bin++], true); - tk::dnn::Activation *relu1_0_2 = new tk::dnn::Activation(&net, CUDNN_ACTIVATION_RELU); - tk::dnn::Conv2d *layer1_0_conv3 = new tk::dnn::Conv2d(&net, 256, 1, 1, 1, 1, 0, 0, layer1_bin[id_layer1_bin++], true); - if(i==0) { - tk::dnn::Layer *route_1_0_layers[1] = { last }; - tk::dnn::Route *route_1_0 = new tk::dnn::Route(&net, route_1_0_layers, 1); - tk::dnn::Conv2d *layer1_0_downsample_0 = new tk::dnn::Conv2d(&net, 256, 1, 1, 1, 1, 0, 0, layer1_bin[id_layer1_bin++], true); - tk::dnn::Shortcut *s1_0 = new tk::dnn::Shortcut(&net, layer1_0_conv3); - } else { - tk::dnn::Shortcut *s1_0 = new tk::dnn::Shortcut(&net, last); - } - tk::dnn::Activation *layer1_0_relu = new tk::dnn::Activation(&net, CUDNN_ACTIVATION_RELU); - last = layer1_0_relu; - } - - // layer 2 - int id_layer2_bin = 0; - for(int i=0; i<4;i++) - { - tk::dnn::Conv2d *layer1_0_conv1 = new tk::dnn::Conv2d(&net, 128, 1, 1, 1, 1, 0, 0, layer2_bin[id_layer2_bin++], true); - tk::dnn::Activation *relu1_0_1 = new tk::dnn::Activation(&net, CUDNN_ACTIVATION_RELU); - tk::dnn::Conv2d *layer1_0_conv2; - if(i==0) - layer1_0_conv2 = new tk::dnn::Conv2d(&net, 128, 3, 3, 2, 2, 1, 1, layer2_bin[id_layer2_bin++], true); - else - layer1_0_conv2 = new tk::dnn::Conv2d(&net, 128, 3, 3, 1, 1, 1, 1, layer2_bin[id_layer2_bin++], true); - - tk::dnn::Activation *relu1_0_2 = new tk::dnn::Activation(&net, CUDNN_ACTIVATION_RELU); - tk::dnn::Conv2d *layer1_0_conv3 = new tk::dnn::Conv2d(&net, 512, 1, 1, 1, 1, 0, 0, layer2_bin[id_layer2_bin++], true); - if(i==0) - { - tk::dnn::Layer *route_1_0_layers[1] = { last }; - tk::dnn::Route *route_1_0 = new tk::dnn::Route(&net, route_1_0_layers, 1); - tk::dnn::Conv2d *layer1_0_downsample_0 = new tk::dnn::Conv2d(&net, 512, 1, 1, 2, 2, 0, 0, layer2_bin[id_layer2_bin++], true); - tk::dnn::Shortcut *s1_0 = new tk::dnn::Shortcut(&net, layer1_0_conv3); - } - else - { - tk::dnn::Shortcut *s1_0 = new tk::dnn::Shortcut(&net, last); - } - tk::dnn::Activation *layer1_0_relu = new tk::dnn::Activation(&net, CUDNN_ACTIVATION_RELU); - last = layer1_0_relu; - } - - // layer 3 - int id_layer3_bin = 0; - for(int i=0; i<23;i++) - { - tk::dnn::Conv2d *layer1_0_conv1 = new tk::dnn::Conv2d(&net, 256, 1, 1, 1, 1, 0, 0, layer3_bin[id_layer3_bin++], true); - tk::dnn::Activation *relu1_0_1 = new tk::dnn::Activation(&net, CUDNN_ACTIVATION_RELU); - tk::dnn::Conv2d *layer1_0_conv2; - if(i==0) - layer1_0_conv2 = new tk::dnn::Conv2d(&net, 256, 3, 3, 2, 2, 1, 1, layer3_bin[id_layer3_bin++], true); - else - layer1_0_conv2 = new tk::dnn::Conv2d(&net, 256, 3, 3, 1, 1, 1, 1, layer3_bin[id_layer3_bin++], true); - - tk::dnn::Activation *relu1_0_2 = new tk::dnn::Activation(&net, CUDNN_ACTIVATION_RELU); - tk::dnn::Conv2d *layer1_0_conv3 = new tk::dnn::Conv2d(&net, 1024, 1, 1, 1, 1, 0, 0, layer3_bin[id_layer3_bin++], true); - if(i==0) - { - tk::dnn::Layer *route_1_0_layers[1] = { last }; - tk::dnn::Route *route_1_0 = new tk::dnn::Route(&net, route_1_0_layers, 1); - tk::dnn::Conv2d *layer1_0_downsample_0 = new tk::dnn::Conv2d(&net, 1024, 1, 1, 2, 2, 0, 0, layer3_bin[id_layer3_bin++], true); - tk::dnn::Shortcut *s1_0 = new tk::dnn::Shortcut(&net, layer1_0_conv3); - } - else - { - tk::dnn::Shortcut *s1_0 = new tk::dnn::Shortcut(&net, last); - } - tk::dnn::Activation *layer1_0_relu = new tk::dnn::Activation(&net, CUDNN_ACTIVATION_RELU); - last = layer1_0_relu; - } - - // layer 4 - int id_layer4_bin = 0; - for(int i=0; i<3;i++) - { - tk::dnn::Conv2d *layer1_0_conv1 = new tk::dnn::Conv2d(&net, 512, 1, 1, 1, 1, 0, 0, layer4_bin[id_layer4_bin++], true); - tk::dnn::Activation *relu1_0_1 = new tk::dnn::Activation(&net, CUDNN_ACTIVATION_RELU); - tk::dnn::Conv2d *layer1_0_conv2; - if(i==0) - layer1_0_conv2 = new tk::dnn::Conv2d(&net, 512, 3, 3, 2, 2, 1, 1, layer4_bin[id_layer4_bin++], true); - else - layer1_0_conv2 = new tk::dnn::Conv2d(&net, 512, 3, 3, 1, 1, 1, 1, layer4_bin[id_layer4_bin++], true); - - tk::dnn::Activation *relu1_0_2 = new tk::dnn::Activation(&net, CUDNN_ACTIVATION_RELU); - tk::dnn::Conv2d *layer1_0_conv3 = new tk::dnn::Conv2d(&net, 2048, 1, 1, 1, 1, 0, 0, layer4_bin[id_layer4_bin++], true); - if(i==0) - { - tk::dnn::Layer *route_1_0_layers[1] = { last }; - tk::dnn::Route *route_1_0 = new tk::dnn::Route(&net, route_1_0_layers, 1); - tk::dnn::Conv2d *layer1_0_downsample_0 = new tk::dnn::Conv2d(&net, 2048, 1, 1, 2, 2, 0, 0, layer4_bin[id_layer4_bin++], true); - tk::dnn::Shortcut *s1_0 = new tk::dnn::Shortcut(&net, layer1_0_conv3); - } - else - { - tk::dnn::Shortcut *s1_0 = new tk::dnn::Shortcut(&net, last); - } - tk::dnn::Activation *layer1_0_relu = new tk::dnn::Activation(&net, CUDNN_ACTIVATION_RELU); - last = layer1_0_relu; - } - - tk::dnn::DeformConv2d *layer0_deform1 = new tk::dnn::DeformConv2d(&net, 256, 1, 3, 3, 1, 1, 1, 1, deform1_bin, d_conv1_bin, true); - tk::dnn::Activation *layer0_deform1_relu = new tk::dnn::Activation(&net, CUDNN_ACTIVATION_RELU); - tk::dnn::DeConv2d *layer0_deconv1 = new tk::dnn::DeConv2d(&net, 256, 4, 4, 2, 2, 1, 1, deconv1_bin, true); - tk::dnn::Activation *layer0_deconv1_relu = new tk::dnn::Activation(&net, CUDNN_ACTIVATION_RELU); - - tk::dnn::DeformConv2d *layer1_deform1 = new tk::dnn::DeformConv2d(&net, 128, 1, 3, 3, 1, 1, 1, 1, deform2_bin, d_conv2_bin, true); - tk::dnn::Activation *layer1_deform1_relu = new tk::dnn::Activation(&net, CUDNN_ACTIVATION_RELU); - tk::dnn::DeConv2d *layer1_deconv1 = new tk::dnn::DeConv2d(&net, 128, 4, 4, 2, 2, 1, 1, deconv2_bin, true); - tk::dnn::Activation *layer1_deconv1_relu = new tk::dnn::Activation(&net, CUDNN_ACTIVATION_RELU); - - tk::dnn::DeformConv2d *layer2_deform1 = new tk::dnn::DeformConv2d(&net, 64, 1, 3, 3, 1, 1, 1, 1, deform3_bin, d_conv3_bin, true); - tk::dnn::Activation *layer2_deform1_relu = new tk::dnn::Activation(&net, CUDNN_ACTIVATION_RELU); - tk::dnn::DeConv2d *layer2_deconv1 = new tk::dnn::DeConv2d(&net, 64, 4, 4, 2, 2, 1, 1, deconv3_bin, true); - tk::dnn::Activation *layer2_deconv1_relu = new tk::dnn::Activation(&net, CUDNN_ACTIVATION_RELU); - - tk::dnn::Layer *route_1_0_layers[1] = { layer2_deconv1_relu }; - tk::dnn::Conv2d *hm_conv1 = new tk::dnn::Conv2d(&net, 64, 3, 3, 1, 1, 1, 1, hm_conv1_bin, false); - tk::dnn::Activation *hm_relu1 = new tk::dnn::Activation(&net, CUDNN_ACTIVATION_RELU); - tk::dnn::Conv2d *hm = new tk::dnn::Conv2d(&net, 80, 1, 1, 1, 1, 0, 0, hm_conv2_bin, false, false, true); - int kernel = 3; - int pad = (kernel - 1)/2; - tk::dnn::Activation *hm_sig = new tk::dnn::Activation(&net, CUDNN_ACTIVATION_SIGMOID); - tk::dnn::Pooling *hmax = new tk::dnn::Pooling(&net, kernel, kernel, 1, 1, pad, pad, tk::dnn::POOLING_MAX, true); - - tk::dnn::Route *route_1_0 = new tk::dnn::Route(&net, route_1_0_layers, 1); - tk::dnn::Conv2d *wh_conv1 = new tk::dnn::Conv2d(&net, 64, 3, 3, 1, 1, 1, 1, wh_conv1_bin, false); - tk::dnn::Activation *wh_relu1 = new tk::dnn::Activation(&net, CUDNN_ACTIVATION_RELU); - tk::dnn::Conv2d *wh = new tk::dnn::Conv2d(&net, 2, 1, 1, 1, 1, 0, 0, wh_conv2_bin, false, false, true); - - tk::dnn::Route *route_2_0 = new tk::dnn::Route(&net, route_1_0_layers, 1); - tk::dnn::Conv2d *reg_conv1 = new tk::dnn::Conv2d(&net, 64, 3, 3, 1, 1, 1, 1, reg_conv1_bin, false); - tk::dnn::Activation *reg_relu1 = new tk::dnn::Activation(&net, CUDNN_ACTIVATION_RELU); - tk::dnn::Conv2d *reg = new tk::dnn::Conv2d(&net, 2, 1, 1, 1, 1, 0, 0, reg_conv2_bin, false, false, true); - - // Load input - dnnType *data; - dnnType *input_h; - readBinaryFile(input_bin, dim.tot(), &input_h, &data); - // printDeviceVector(64, data, true); - - //print network model - net.print(); - - //convert network to tensorRT - tk::dnn::NetworkRT netRT(&net, "resnet101_cnet.rt"); - - - tk::dnn::dataDim_t dim1 = dim; //input dim - printCenteredTitle(" CUDNN inference ", '=', 30); - { - dim1.print(); - TIMER_START - net.infer(dim1, data); - TIMER_STOP - dim1.print(); - } - - // printDeviceVector(64, cudnn_out, true); - - tk::dnn::dataDim_t dim2 = dim; - printCenteredTitle(" TENSORRT inference ", '=', 30); - { - dim2.print(); - TIMER_START - netRT.infer(dim2, data); - TIMER_STOP - dim2.print(); - } - - tk::dnn::Layer *outs[3] = { hm, wh, reg }; - int out_count = 1; - for(int i=0; i<3; i++) { - printCenteredTitle((std::string(" RESNET CHECK RESULTS ") + std::to_string(i) + " ").c_str(), '=', 30); - - outs[i]->output_dim.print(); - - dnnType *out, *out_h; - int odim = outs[i]->output_dim.tot(); - readBinaryFile(output_bin[i], odim, &out_h, &out); - // std::cout<<"OUTPUT BIN:\n"; - // printDeviceVector(odim, cudnn_out, true); - // std::cout<<"FILE BIN:\n"; - // printDeviceVector(odim, out, true); - - dnnType *cudnn_out, *rt_out; - cudnn_out = outs[i]->dstData; - rt_out = (dnnType *)netRT.buffersRT[i+out_count]; - // there is the maxpool. It isn't an output but it is necessary for the process section - if(i==0) - out_count ++; - - std::cout << "CUDNN vs correct"; - checkResult(odim, cudnn_out, out); - - std::cout << "TRT vs correct"; - checkResult(odim, rt_out, out); - std::cout << "CUDNN vs TRT "; - checkResult(odim, cudnn_out, rt_out); - } - - TIMER_START - - // -------- transofrm compose - cv::Mat imageOrig = cv::imread("/media/davide/DATA/shared_home/Projects/Professionale/repos/photo_2020-01-14_09-56-07.jpg"); - cv::Mat imageF; - imageOrig.convertTo(imageF, CV_32FC3, 1/255.0); - cv::Mat image; - cv::Size sz = imageF.size(); - std::cout<<"image: "<output_dim.tot()*sizeof(dnnType)) ); - - dnnType *rt_out[4]; - rt_out[0] = (dnnType *)netRT.buffersRT[1]; - rt_out[1] = (dnnType *)netRT.buffersRT[2]; - rt_out[2] = (dnnType *)netRT.buffersRT[3]; - rt_out[3] = (dnnType *)netRT.buffersRT[4]; - - // checkCuda( cudaMemcpy(hm_h, rt_out[0], hm->output_dim.tot()*sizeof(dnnType), cudaMemcpyDeviceToHost) ); - std::cout<<"hm\n"; - hm->output_dim.print(); - // for(int i=0; ioutput_dim.tot(); i++ ){ - // std::cout<output_dim.tot()); - checkCuda( cudaDeviceSynchronize() ); - - - checkCuda( cudaMemcpy(hm_h, rt_out[0], hm->output_dim.tot()*sizeof(dnnType), cudaMemcpyDeviceToHost) ); - std::cout<<"hm\n"; - hm->output_dim.print(); - // for(int i=0; ioutput_dim.tot(); i++ ){ - // std::cout<infer(hmax->input_dim.tot(), rt_out[0]); - // keep = (hmax == heat).float() - // return heat * keep - - dnnType *hmax_h; - checkCuda( cudaMallocHost(&hmax_h, hmax->output_dim.tot()*sizeof(dnnType)) ); - checkCuda( cudaMemcpy(hmax_h, rt_out[1], hmax->output_dim.tot()*sizeof(dnnType), cudaMemcpyDeviceToHost) ); - - std::cout<<"hmax\n"; - hmax->output_dim.print(); - // for(int i=0; ioutput_dim.tot(); i++ ){ - // std::cout<output_dim.print(); - std::cout<<"hmax:\n"; - hmax->output_dim.print(); - // return 0; - float toll = 0.000001; - for(int i=0; i < hm->output_dim.tot(); i++){ - if(hm_h[i]-hmax_h[i] > toll || hm_h[i]-hmax_h[i] < -toll){ - hm_h[i] = 0.0f; - } - } - // std::cout<<"\n"; - // for(int i=0; ioutput_dim.tot(); i++ ){ - // std::cout<dstData, hm_h, hm->output_dim.tot()*sizeof(dnnType), cudaMemcpyHostToDevice) ); - checkCuda( cudaFreeHost(hmax_h) ); - // ----------- nms end - // ----------- topk - int K = 100; - int width = 56; // TODO - float *topk_scores; - int *topk_inds_; - float *topk_ys_; - float *topk_xs_; - std::cout<<"mah: "<output_dim.c * K<output_dim.c * K *sizeof(float)) ); - checkCuda( cudaMallocHost(&topk_inds_, hm->output_dim.c * K *sizeof(int)) ); - checkCuda( cudaMallocHost(&topk_ys_, hm->output_dim.c * K *sizeof(float)) ); - checkCuda( cudaMallocHost(&topk_xs_, hm->output_dim.c * K *sizeof(float)) ); - std::cout<<"1\n"; - dnnType *hm_aus; - checkCuda( cudaMallocHost(&hm_aus, hm->output_dim.h * hm->output_dim.w *sizeof(dnnType)) ); - std::cout<<"2\n"; - int count; - std::vector v = {2.0, 3.0, 9.0}; - for (auto i: sort_indexes(v)) { - std::cout << i<< "--" <output_dim.c; i++){ - count = 0; - // get the hm->output_dim.h * hm->output_dim.w elements for each channel and sort it. Then find the first 100 elements - checkCuda( cudaMemcpy(hm_aus, hm_h + i * hm->output_dim.h * hm->output_dim.w, - hm->output_dim.h * hm->output_dim.w * sizeof(dnnType), cudaMemcpyHostToHost) ); - // std::cout<<"top scores: "<output_dim.h * hm->output_dim.w<<"\n"; - // for(int k=0; koutput_dim.h * hm->output_dim.w; k++) - // std::cout< my_vector {arr, arr + arr_length} - std::vector my_vector{hm_aus, hm_aus + hm->output_dim.h * hm->output_dim.w}; - for (auto j: sort_indexes(my_vector)) { - // std::cout <<"j: "< "<< hm_aus[j] << std::endl; - topk_scores[i*K + count] = hm_aus[j]; - topk_inds_[i*K +count] = j; - topk_ys_[i*K +count] = (int)(j / width); - topk_xs_[i*K +count] = (int)(j % width); - if(++count == K) - break; - } - } - std::cout<<"topk_xs_[0]: "<output_dim.c * K; i++) - std::cout< my_vector{topk_scores, topk_scores + hm->output_dim.c * K }; - for (auto j: sort_indexes(my_vector)) { - // std::cout <<"j: "< "<< hm_aus[j] << std::endl; - scores[count] = topk_scores[j]; - clses[count] = (int)(j / K); - topk_inds[count] = topk_inds_[j]; - topk_ys[count] = topk_ys_[j]; - topk_xs[count] = topk_xs_[j]; - if(++count == K) - break; - } - checkCuda( cudaFreeHost(topk_scores) ); - checkCuda( cudaFreeHost(topk_inds_) ); - checkCuda( cudaFreeHost(topk_ys_) ); - checkCuda( cudaFreeHost(topk_xs_) ); - std::cout<<"5\n"; - // ----------- topk end - std::cout<<"topk_xs[0]: "<output_dim.tot()*sizeof(dnnType)) ); - checkCuda( cudaMemcpy(reg_aus, rt_out[3], reg->output_dim.tot()*sizeof(dnnType), cudaMemcpyDeviceToHost) ); - std::cout<<"reg:\n"; - reg->output_dim.print(); - // for(int i=0; ioutput_dim.tot(); i++ ){ - // std::cout<output_dim.h*reg->output_dim.w]; - } - std::cout<<"topk_xs[0]: "<output_dim.tot()*sizeof(dnnType)) ); - checkCuda( cudaMallocHost(&bboxes, 4 * K *sizeof(dnnType)) ); - checkCuda( cudaMemcpy(wh_aus, rt_out[2], wh->output_dim.tot()*sizeof(dnnType), cudaMemcpyDeviceToHost) ); - std::cout<<"7\n"; - for(int i = 0; i< K; i++) - std::cout<output_dim.h*reg->output_dim.w] / 2; - bboxes[i * 4 + 2] = topk_xs[i] + wh_aus[topk_inds[i]] / 2; - bboxes[i * 4 + 3] = topk_ys[i] + wh_aus[topk_inds[i]+reg->output_dim.h*reg->output_dim.w] / 2; - } - ////////////////// fin qui ok - - checkCuda( cudaFreeHost(wh_aus) ); - checkCuda( cudaFreeHost(topk_inds) ); - checkCuda( cudaFreeHost(topk_ys) ); - checkCuda( cudaFreeHost(topk_xs) ); - - std::cout<<"8\n"; - float *detections; - std::cout<<"bboxes:\n"; - for(int i = 0; i < K+1; i++){ - std::cout<(0,0)=c[0]; - src.at(0,1)=c[1]; - src.at(1,0)=c[0]; - src.at(1,1)=c[1] + s[0] * -0.5; - dst.at(0,0)=width * 0.5; - dst.at(0,1)=width * 0.5; - dst.at(1,0)=width * 0.5; - dst.at(1,1)=width * 0.5 + width * -0.5; - - src.at(2,0)=src.at(1,0) + (-src.at(0,1)+src.at(1,1) ); - src.at(2,1)=src.at(1,1) + (src.at(0,0)-src.at(1,0) ); - dst.at(2,0)=dst.at(1,0) + (-dst.at(0,1)+dst.at(1,1) ); - dst.at(2,1)=dst.at(1,1) + (dst.at(0,0)-dst.at(1,0) ); - std::cout<<"src: "<(0,0)<<" - "<(0,1)<<" - "<(0,2)<<"\n"<(1,0)<<" - "<(1,1)<<" - "<(1,2)<(0,0)=detections[i*4]; - // new_pt1.at(0,1)=detections[i*4+1]; - // new_pt1.at(0,2)=1.0; - // new_pt1 << detections[i], detections[i+K], 1.0; - // std::cout<<"----\ni: "<(0,0)=static_cast(trans2.at(0,0))*detections[i*4] + - static_cast(trans2.at(0,1))*detections[i*4+1] + - static_cast(trans2.at(0,2))*1.0; - new_pt1.at(0,1)=static_cast(trans2.at(1,0))*detections[i*4] + - static_cast(trans2.at(1,1))*detections[i*4+1] + - static_cast(trans2.at(1,2))*1.0; - - new_pt2.at(0,0)=static_cast(trans2.at(0,0))*detections[i*4+2] + - static_cast(trans2.at(0,1))*detections[i*4+3] + - static_cast(trans2.at(0,2))*1.0; - new_pt2.at(0,1)=static_cast(trans2.at(1,0))*detections[i*4+2] + - static_cast(trans2.at(1,1))*detections[i*4+3] + - static_cast(trans2.at(1,2))*1.0; - - - // std::cout<<"\n new: "<(0,0); - target_coords[i*4+1] = new_pt1.at(0,1); - target_coords[i*4+2] = new_pt2.at(0,0); - target_coords[i*4+3] = new_pt2.at(0,1); - // std::cout<(0,0)<<", "<(0,1)<<", "<(0,0)<<", "<(0,1)< coco_class_name(coco_class_name_, std::end( coco_class_name_ )); - int num_classes = 80; - float vis_threshold = 0.3; - // int *classes; - std::vector detected; - // checkCuda( cudaMallocHost(&classes, K *sizeof(int)) ); - // checkCuda( cudaMemcpy(classes, detections + 5 * K *sizeof(dnnType), K *sizeof(dnnType), cudaMemcpyHostToHost) ); - for(int i = 0; i i+1 (1:80); - - if(scores[j] > vis_threshold){ - std::cout<<"th: "< -#include "tkdnn.h" - -const char *input_bin = "../tests/simple/input.bin"; -const char *c0_bin = "../tests/simple/layers/c0.bin"; -const char *c1_bin = "../tests/simple/layers/c1.bin"; -const char *d2_bin = "../tests/simple/layers/d2.bin"; -const char *output_bin = "../tests/simple/output.bin"; - -int main() { - - // Network layout - tk::dnn::dataDim_t dim(1, 1, 10, 10, 1); - tk::dnn::Network net(dim); - tk::dnn::Conv2d l0(&net, 2, 4, 4, 2, 2, 0, 0, c0_bin); - tk::dnn::Activation l1(&net, CUDNN_ACTIVATION_RELU); - tk::dnn::Conv2d l2(&net, 4, 2, 2, 1, 1, 0, 0, c1_bin); - tk::dnn::Activation l3(&net, CUDNN_ACTIVATION_RELU); - tk::dnn::Dense l5(&net, 4, d2_bin); - tk::dnn::Activation l6(&net, CUDNN_ACTIVATION_RELU); - - net.print(); - - // Load input - dnnType *data; - dnnType *input_h; - readBinaryFile(input_bin, dim.tot(), &input_h, &data); - - // Print input - std::cout<<"\n======= INPUT =======\n"; - printDeviceVector(dim.tot(), data); - std::cout<<"\n"; - - //convert network to tensorRT - tk::dnn::NetworkRT netRT(&net, "simple.rt"); - - dnnType *out_data, *out_data2; // cudnn output, tensorRT output - - tk::dnn::dataDim_t dim1 = dim; //input dim - printCenteredTitle(" CUDNN inference ", '=', 30); { - dim1.print(); - TIMER_START - out_data = net.infer(dim1, data); - TIMER_STOP - dim1.print(); - } - - tk::dnn::dataDim_t dim2 = dim; - printCenteredTitle(" TENSORRT inference ", '=', 30); { - dim2.print(); - TIMER_START - out_data2 = netRT.infer(dim2, data); - TIMER_STOP - dim2.print(); - } - - std::cout<<"\n======= CUDNN =======\n"; - printDeviceVector(dim.tot(), out_data); - std::cout<<"\n======= TENSORRT =======\n"; - printDeviceVector(dim.tot(), out_data2); - - printCenteredTitle(" CHECK RESULTS ", '=', 30); - dnnType *out, *out_h; - int out_dim = net.getOutputDim().tot(); - //readBinaryFile(output_bin, out_dim, &out_h, &out); - //std::cout<<"CUDNN vs correct"; checkResult(out_dim, out_data, out); - //std::cout<<"TRT vs correct"; checkResult(out_dim, out_data2, out); - std::cout<<"CUDNN vs TRT "; checkResult(out_dim, out_data, out_data2); - return 0; -} diff --git a/tests/test_rtinference/rtinference.cpp b/tests/test_rtinference/rtinference.cpp deleted file mode 100644 index f99d003..0000000 --- a/tests/test_rtinference/rtinference.cpp +++ /dev/null @@ -1,34 +0,0 @@ -#include -#include "tkdnn.h" -#include /* srand, rand */ - -int main(int argc, char *argv[]) { - - if(argc < 2 || !fileExist(argv[1])) - FatalError("unable to read serialRT file"); - - //always same test - srand (0); - - //convert network to tensorRT - tk::dnn::NetworkRT netRT(NULL, argv[1]); - - dnnType *input = new float[netRT.input_dim.tot()]; - dnnType *output = new float[netRT.input_dim.tot()]; - - printCenteredTitle(" TENSORRT inference ", '=', 30); - for(int i=0; i<100; i++) { - for(int j=0; j -#include "tkdnn.h" - -const char *input_bin = "../tests/yolo/layers/input.bin"; -const char *c0_bin = "../tests/yolo/layers/c0.bin"; -const char *c2_bin = "../tests/yolo/layers/c2.bin"; -const char *c4_bin = "../tests/yolo/layers/c4.bin"; -const char *c5_bin = "../tests/yolo/layers/c5.bin"; -const char *c6_bin = "../tests/yolo/layers/c6.bin"; -const char *c8_bin = "../tests/yolo/layers/c8.bin"; -const char *c9_bin = "../tests/yolo/layers/c9.bin"; -const char *c10_bin = "../tests/yolo/layers/c10.bin"; -const char *c12_bin = "../tests/yolo/layers/c12.bin"; -const char *c13_bin = "../tests/yolo/layers/c13.bin"; -const char *c14_bin = "../tests/yolo/layers/c14.bin"; -const char *c15_bin = "../tests/yolo/layers/c15.bin"; -const char *c16_bin = "../tests/yolo/layers/c16.bin"; -const char *c18_bin = "../tests/yolo/layers/c18.bin"; -const char *c19_bin = "../tests/yolo/layers/c19.bin"; -const char *c20_bin = "../tests/yolo/layers/c20.bin"; -const char *c21_bin = "../tests/yolo/layers/c21.bin"; -const char *c22_bin = "../tests/yolo/layers/c22.bin"; -const char *c23_bin = "../tests/yolo/layers/c23.bin"; -const char *c24_bin = "../tests/yolo/layers/c24.bin"; -const char *c26_bin = "../tests/yolo/layers/c26.bin"; -const char *c29_bin = "../tests/yolo/layers/c29.bin"; -const char *c30_bin = "../tests/yolo/layers/c30.bin"; -const char *g31_bin = "../tests/yolo/layers/g31.bin"; -const char *output_bin = "../tests/yolo/layers/output.bin"; - -int main() { - - // Network layout - tk::dnn::dataDim_t dim(1, 3, 608, 608, 1); - tk::dnn::Network net(dim); - - tk::dnn::Conv2d c0 (&net, 32, 3, 3, 1, 1, 1, 1, c0_bin, true); - tk::dnn::Activation a0 (&net, tk::dnn::ACTIVATION_LEAKY); - tk::dnn::Pooling p1 (&net, 2, 2, 2, 2, tk::dnn::POOLING_MAX); - - tk::dnn::Conv2d c2 (&net, 64, 3, 3, 1, 1, 1, 1, c2_bin, true); - tk::dnn::Activation a2 (&net, tk::dnn::ACTIVATION_LEAKY); - tk::dnn::Pooling p3 (&net, 2, 2, 2, 2, tk::dnn::POOLING_MAX); - - tk::dnn::Conv2d c4 (&net, 128, 3, 3, 1, 1, 1, 1, c4_bin, true); - tk::dnn::Activation a4 (&net, tk::dnn::ACTIVATION_LEAKY); - tk::dnn::Conv2d c5 (&net, 64, 1, 1, 1, 1, 0, 0, c5_bin, true); - tk::dnn::Activation a5 (&net, tk::dnn::ACTIVATION_LEAKY); - tk::dnn::Conv2d c6 (&net, 128, 3, 3, 1, 1, 1, 1, c6_bin, true); - tk::dnn::Activation a6 (&net, tk::dnn::ACTIVATION_LEAKY); - tk::dnn::Pooling p7 (&net, 2, 2, 2, 2, tk::dnn::POOLING_MAX); - - tk::dnn::Conv2d c8 (&net, 256, 3, 3, 1, 1, 1, 1, c8_bin, true); - tk::dnn::Activation a8 (&net, tk::dnn::ACTIVATION_LEAKY); - tk::dnn::Conv2d c9 (&net, 128, 1, 1, 1, 1, 0, 0, c9_bin, true); - tk::dnn::Activation a9 (&net, tk::dnn::ACTIVATION_LEAKY); - tk::dnn::Conv2d c10(&net, 256, 3, 3, 1, 1, 1, 1, c10_bin, true); - tk::dnn::Activation a10(&net, tk::dnn::ACTIVATION_LEAKY); - tk::dnn::Pooling p11(&net, 2, 2, 2, 2, tk::dnn::POOLING_MAX); - - tk::dnn::Conv2d c12(&net, 512, 3, 3, 1, 1, 1, 1, c12_bin, true); - tk::dnn::Activation a12(&net, tk::dnn::ACTIVATION_LEAKY); - tk::dnn::Conv2d c13(&net, 256, 1, 1, 1, 1, 0, 0, c13_bin, true); - tk::dnn::Activation a13(&net, tk::dnn::ACTIVATION_LEAKY); - tk::dnn::Conv2d c14(&net, 512, 3, 3, 1, 1, 1, 1, c14_bin, true); - tk::dnn::Activation a14(&net, tk::dnn::ACTIVATION_LEAKY); - tk::dnn::Conv2d c15(&net, 256, 1, 1, 1, 1, 0, 0, c15_bin, true); - tk::dnn::Activation a15(&net, tk::dnn::ACTIVATION_LEAKY); - tk::dnn::Conv2d c16(&net, 512, 3, 3, 1, 1, 1, 1, c16_bin, true); - tk::dnn::Activation a16(&net, tk::dnn::ACTIVATION_LEAKY); - tk::dnn::Pooling p17(&net, 2, 2, 2, 2, tk::dnn::POOLING_MAX); - - tk::dnn::Conv2d c18(&net, 1024, 3, 3, 1, 1, 1, 1, c18_bin, true); - tk::dnn::Activation a18(&net, tk::dnn::ACTIVATION_LEAKY); - tk::dnn::Conv2d c19(&net, 512, 1, 1, 1, 1, 0, 0, c19_bin, true); - tk::dnn::Activation a19(&net, tk::dnn::ACTIVATION_LEAKY); - tk::dnn::Conv2d c20(&net, 1024, 3, 3, 1, 1, 1, 1, c20_bin, true); - tk::dnn::Activation a20(&net, tk::dnn::ACTIVATION_LEAKY); - tk::dnn::Conv2d c21(&net, 512, 1, 1, 1, 1, 0, 0, c21_bin, true); - tk::dnn::Activation a21(&net, tk::dnn::ACTIVATION_LEAKY); - tk::dnn::Conv2d c22(&net, 1024, 3, 3, 1, 1, 1, 1, c22_bin, true); - tk::dnn::Activation a22(&net, tk::dnn::ACTIVATION_LEAKY); - tk::dnn::Conv2d c23(&net, 1024, 3, 3, 1, 1, 1, 1, c23_bin, true); - tk::dnn::Activation a23(&net, tk::dnn::ACTIVATION_LEAKY); - tk::dnn::Conv2d c24(&net, 1024, 3, 3, 1, 1, 1, 1, c24_bin, true); - tk::dnn::Activation a24(&net, tk::dnn::ACTIVATION_LEAKY); - - tk::dnn::Layer *m25_layers[1] = { &a16 }; - tk::dnn::Route m25(&net, m25_layers, 1); - tk::dnn::Conv2d c26(&net, 64, 1, 1, 1, 1, 0, 0, c26_bin, true); - tk::dnn::Activation a26(&net, tk::dnn::ACTIVATION_LEAKY); - tk::dnn::Reorg r27(&net, 2); - - tk::dnn::Layer *m28_layers[2] = { &r27, &a24 }; - tk::dnn::Route m28(&net, m28_layers, 2); - - tk::dnn::Conv2d c29(&net, 1024, 3, 3, 1, 1, 1, 1, c29_bin, true); - tk::dnn::Activation a29(&net, tk::dnn::ACTIVATION_LEAKY); - tk::dnn::Conv2d c30(&net, 425, 1, 1, 1, 1, 0, 0, c30_bin, false); - tk::dnn::Region g31(&net, 80, 4, 5); - - tk::dnn::RegionInterpret rI(dim, g31.output_dim, 80, 4, 5, 0.6f, g31_bin); - - // Load input - dnnType *data; - dnnType *input_h; - readBinaryFile(input_bin, dim.tot(), &input_h, &data); - - //print network model - net.print(); - - //convert network to tensorRT - tk::dnn::NetworkRT netRT(&net, "yolo.rt"); - - dnnType *out_data, *out_data2; // cudnn output, tensorRT output - - tk::dnn::dataDim_t dim1 = dim; //input dim - printCenteredTitle(" CUDNN inference ", '=', 30); { - dim1.print(); - TIMER_START - out_data = net.infer(dim1, data); - TIMER_STOP - dim1.print(); - } - - tk::dnn::dataDim_t dim2 = dim; - printCenteredTitle(" TENSORRT inference ", '=', 30); { - dim2.print(); - TIMER_START - out_data2 = netRT.infer(dim2, data); - TIMER_STOP - dim2.print(); - } - - printCenteredTitle(" CHECK RESULTS ", '=', 30); - dnnType *out, *out_h; - int out_dim = net.getOutputDim().tot(); - readBinaryFile(output_bin, out_dim, &out_h, &out); - std::cout<<"CUDNN vs correct"; checkResult(out_dim, out_data, out); - std::cout<<"TRT vs correct"; checkResult(out_dim, out_data2, out); - std::cout<<"CUDNN vs TRT "; checkResult(out_dim, out_data, out_data2); - - std::cout<<"\n\nDetected objects: \n"; - dnnType *output_h = new dnnType[rI.output_dim.tot()]; - checkCuda(cudaMemcpy(output_h, out_data2, - rI.output_dim.tot()*sizeof(dnnType), cudaMemcpyDeviceToHost)); - rI.interpretData(output_h); - rI.showImageResult(input_h); - return 0; -} diff --git a/tests/yolo3/yolo3.cpp b/tests/yolo3/yolo3.cpp deleted file mode 100644 index a3772e5..0000000 --- a/tests/yolo3/yolo3.cpp +++ /dev/null @@ -1,92 +0,0 @@ -#include -#include -#include "tkdnn.h" - -int main() { - - // Network layout - tk::dnn::dataDim_t dim(1, 3, 416, 416, 1); - tk::dnn::Network net(dim); - - // create yolo3 model - std::string bin_path = "../tests/yolo3"; - int classes = 80; - tk::dnn::Yolo *yolo [3]; - #include "models/Yolo3.h" - - // fill classes names - for(int i=0; i<3; i++) { - yolo[i]->classesNames = {"person" , "bicycle" , "car" , "motorbike" , "aeroplane" , "bus" , "train" , "truck" , "boat" , "traffic light" , "fire hydrant" , "stop sign" , "parking meter" , "bench" , "bird" , "cat" , "dog" , "horse" , "sheep" , "cow" , "elephant" , "bear" , "zebra" , "giraffe" , "backpack" , "umbrella" , "handbag" , "tie" , "suitcase" , "frisbee" , "skis" , "snowboard" , "sports ball" , "kite" , "baseball bat" , "baseball glove" , "skateboard" , "surfboard" , "tennis racket" , "bottle" , "wine glass" , "cup" , "fork" , "knife" , "spoon" , "bowl" , "banana" , "apple" , "sandwich" , "orange" , "broccoli" , "carrot" , "hot dog" , "pizza" , "donut" , "cake" , "chair" , "sofa" , "pottedplant" , "bed" , "diningtable" , "toilet" , "tvmonitor" , "laptop" , "mouse" , "remote" , "keyboard" , "cell phone" , "microwave" , "oven" , "toaster" , "sink" , "refrigerator" , "book" , "clock" , "vase" , "scissors" , "teddy bear" , "hair drier" , "toothbrush"}; - } - - // Load input - dnnType *data; - dnnType *input_h; - readBinaryFile(input_bin, dim.tot(), &input_h, &data); - - //print network model - net.print(); - - //convert network to tensorRT - tk::dnn::NetworkRT netRT(&net, "yolo3.rt"); - - // the network have 3 outputs - tk::dnn::dataDim_t out_dim[3]; - for(int i=0; i<3; i++) out_dim[i] = yolo[i]->output_dim; - dnnType *cudnn_out[3], *rt_out[3]; - - tk::dnn::dataDim_t dim1 = dim; //input dim - printCenteredTitle(" CUDNN inference ", '=', 30); { - dim1.print(); - TIMER_START - net.infer(dim1, data); - TIMER_STOP - dim1.print(); - } - for(int i=0; i<3; i++) cudnn_out[i] = yolo[i]->dstData; - - printCenteredTitle(" compute detections ", '=', 30); - TIMER_START - int ndets = 0; - tk::dnn::Yolo::detection *dets = tk::dnn::Yolo::allocateDetections(tk::dnn::Yolo::MAX_DETECTIONS, classes); - for(int i=0; i<3; i++) yolo[i]->computeDetections(dets, ndets, net.input_dim.w, net.input_dim.h, 0.5); - tk::dnn::Yolo::mergeDetections(dets, ndets, classes); - - for(int j=0; j 0) - cl = c; - } - std::cout< -#include -#include "tkdnn.h" - -int main() { - - // Network layout - tk::dnn::dataDim_t dim(1, 3, 320, 544, 1); - tk::dnn::Network net(dim); - - // create yolo3 model - std::string bin_path = "../tests/yolo3_berkeley"; - int classes = 10; - tk::dnn::Yolo *yolo [3]; - #include "models/Yolo3.h" - - // fill classes names - for(int i=0; i<3; i++) { - yolo[i]->classesNames = {"person", "car", "truck", "bus", "motor", "bike", "rider", "traffic light", "traffic sign", "train"}; - } - - // Load input - dnnType *data; - dnnType *input_h; - readBinaryFile(input_bin, dim.tot(), &input_h, &data); - - //print network model - net.print(); - - //convert network to tensorRT - tk::dnn::NetworkRT netRT(&net, "yolo3_berkeley.rt"); - - // the network have 3 outputs - tk::dnn::dataDim_t out_dim[3]; - for(int i=0; i<3; i++) out_dim[i] = yolo[i]->output_dim; - dnnType *cudnn_out[3], *rt_out[3]; - - tk::dnn::dataDim_t dim1 = dim; //input dim - printCenteredTitle(" CUDNN inference ", '=', 30); { - dim1.print(); - TIMER_START - net.infer(dim1, data); - TIMER_STOP - dim1.print(); - } - for(int i=0; i<3; i++) cudnn_out[i] = yolo[i]->dstData; - - printCenteredTitle(" compute detections ", '=', 30); - TIMER_START - int ndets = 0; - tk::dnn::Yolo::detection *dets = tk::dnn::Yolo::allocateDetections(tk::dnn::Yolo::MAX_DETECTIONS, classes); - for(int i=0; i<3; i++) yolo[i]->computeDetections(dets, ndets, net.input_dim.w, net.input_dim.h, 0.5); - tk::dnn::Yolo::mergeDetections(dets, ndets, classes); - - for(int j=0; j 0) - cl = c; - } - std::cout< -#include -#include "tkdnn.h" - -int main() { - - // Network layout - tk::dnn::dataDim_t dim(1, 3, 416, 416, 1); - tk::dnn::Network net(dim); - - // create yolo3 model - std::string bin_path = "../tests/yolo3_coco4"; - int classes = 4; - tk::dnn::Yolo *yolo [3]; - #include "models/Yolo3.h" - - // Load input - dnnType *data; - dnnType *input_h; - readBinaryFile(input_bin, dim.tot(), &input_h, &data); - - //print network model - net.print(); - - //convert network to tensorRT - tk::dnn::NetworkRT netRT(&net, "yolo3_coco4.rt"); - - // the network have 3 outputs - tk::dnn::dataDim_t out_dim[3]; - for(int i=0; i<3; i++) out_dim[i] = yolo[i]->output_dim; - dnnType *cudnn_out[3], *rt_out[3]; - - tk::dnn::dataDim_t dim1 = dim; //input dim - printCenteredTitle(" CUDNN inference ", '=', 30); { - dim1.print(); - TIMER_START - net.infer(dim1, data); - TIMER_STOP - dim1.print(); - } - for(int i=0; i<3; i++) cudnn_out[i] = yolo[i]->dstData; - - printCenteredTitle(" compute detections ", '=', 30); - TIMER_START - int ndets = 0; - tk::dnn::Yolo::detection *dets = tk::dnn::Yolo::allocateDetections(tk::dnn::Yolo::MAX_DETECTIONS, classes); - for(int i=0; i<3; i++) yolo[i]->computeDetections(dets, ndets, net.input_dim.w, net.input_dim.h, 0.5); - tk::dnn::Yolo::mergeDetections(dets, ndets, classes); - - for(int j=0; j 0) - cl = c; - } - std::cout< -#include -#include "tkdnn.h" - - -int main() { - - // Network layout - tk::dnn::dataDim_t dim(1, 1, 320, 544, 1); - tk::dnn::Network net(dim); - - // create yolo3 model - std::string bin_path = "../tests/yolo3_flir"; - int classes = 3; - tk::dnn::Yolo *yolo [3]; - #include "models/Yolo3.h" - - // fill classes names - for(int i=0; i<3; i++) { - yolo[i]->classesNames = {"person", "bike", "car"}; - } - - // Load input - dnnType *data; - dnnType *input_h; - readBinaryFile(input_bin, dim.tot(), &input_h, &data); - - //print network model - net.print(); - - //convert network to tensorRT - tk::dnn::NetworkRT netRT(&net, "yolo3_flir.rt"); - - // the network have 3 outputs - tk::dnn::dataDim_t out_dim[3]; - for(int i=0; i<3; i++) out_dim[i] = yolo[i]->output_dim; - dnnType *cudnn_out[3], *rt_out[3]; - - tk::dnn::dataDim_t dim1 = dim; //input dim - printCenteredTitle(" CUDNN inference ", '=', 30); { - dim1.print(); - TIMER_START - net.infer(dim1, data); - TIMER_STOP - dim1.print(); - } - for(int i=0; i<3; i++) cudnn_out[i] = yolo[i]->dstData; - - printCenteredTitle(" compute detections ", '=', 30); - TIMER_START - int ndets = 0; - tk::dnn::Yolo::detection *dets = tk::dnn::Yolo::allocateDetections(tk::dnn::Yolo::MAX_DETECTIONS, classes); - for(int i=0; i<3; i++) yolo[i]->computeDetections(dets, ndets, net.input_dim.w, net.input_dim.h, 0.5); - tk::dnn::Yolo::mergeDetections(dets, ndets, classes); - - for(int j=0; j 0) - cl = c; - } - std::cout< -#include "tkdnn.h" - -const char *input_bin = "../tests/yolo3_tiny/layers/input.bin"; -const char *c0_bin = "../tests/yolo3_tiny/layers/c0.bin"; -const char *c2_bin = "../tests/yolo3_tiny/layers/c2.bin"; -const char *c4_bin = "../tests/yolo3_tiny/layers/c4.bin"; -const char *c6_bin = "../tests/yolo3_tiny/layers/c6.bin"; -const char *c8_bin = "../tests/yolo3_tiny/layers/c8.bin"; -const char *c10_bin = "../tests/yolo3_tiny/layers/c10.bin"; -const char *c12_bin = "../tests/yolo3_tiny/layers/c12.bin"; -const char *c13_bin = "../tests/yolo3_tiny/layers/c13.bin"; -const char *c14_bin = "../tests/yolo3_tiny/layers/c14.bin"; -const char *c15_bin = "../tests/yolo3_tiny/layers/c15.bin"; -const char *c18_bin = "../tests/yolo3_tiny/layers/c18.bin"; -const char *c21_bin = "../tests/yolo3_tiny/layers/c21.bin"; -const char *c22_bin = "../tests/yolo3_tiny/layers/c22.bin"; -const char *g16_bin = "../tests/yolo3_tiny/layers/g16.bin"; -const char *g23_bin = "../tests/yolo3_tiny/layers/g23.bin"; -// const char *output_bin = "../tests/yolo3_tiny/layers/output.bin"; - -const char *output_bin = "../tests/yolo3_tiny/debug/layer23_out.bin"; - -int main() { - - int classes = 80; - - // Network layout - tk::dnn::dataDim_t dim(1, 3, 416, 416, 1); - tk::dnn::Network net(dim); - - - tk::dnn::Conv2d c0 (&net, 16, 3, 3, 1, 1, 1, 1, c0_bin, true); - tk::dnn::Activation a0 (&net, tk::dnn::ACTIVATION_LEAKY); - tk::dnn::Pooling p1 (&net, 2, 2, 2, 2, tk::dnn::POOLING_MAX); - - tk::dnn::Conv2d c2 (&net, 32, 3, 3, 1, 1, 1, 1, c2_bin, true); - tk::dnn::Activation a2 (&net, tk::dnn::ACTIVATION_LEAKY); - tk::dnn::Pooling p3 (&net, 2, 2, 2, 2, tk::dnn::POOLING_MAX); - - tk::dnn::Conv2d c4 (&net, 64, 3, 3, 1, 1, 1, 1, c4_bin, true); - tk::dnn::Activation a4 (&net, tk::dnn::ACTIVATION_LEAKY); - tk::dnn::Pooling p5 (&net, 2, 2, 2, 2, tk::dnn::POOLING_MAX); - - tk::dnn::Conv2d c6 (&net, 128, 3, 3, 1, 1, 1, 1, c6_bin, true); - tk::dnn::Activation a6 (&net, tk::dnn::ACTIVATION_LEAKY); - tk::dnn::Pooling p7(&net, 2, 2, 2, 2, tk::dnn::POOLING_MAX); - - tk::dnn::Conv2d c8(&net, 256, 3, 3, 1, 1, 1, 1, c8_bin, true); - tk::dnn::Activation a8(&net, tk::dnn::ACTIVATION_LEAKY); - tk::dnn::Pooling p9(&net, 2, 2, 2, 2, tk::dnn::POOLING_MAX); - - tk::dnn::Conv2d c10(&net, 512, 3, 3, 1, 1, 1, 1, c10_bin, true); - tk::dnn::Activation a10(&net, tk::dnn::ACTIVATION_LEAKY); - tk::dnn::Pooling p11(&net, 2, 2, 1, 1,0,0, tk::dnn::POOLING_MAX); - - tk::dnn::Conv2d c12(&net, 1024, 3, 3, 1, 1, 1, 1, c12_bin, true); - tk::dnn::Activation a12(&net, tk::dnn::ACTIVATION_LEAKY); - - tk::dnn::Conv2d c13(&net, 256, 1, 1, 1, 1, 0, 0, c13_bin, true); - tk::dnn::Activation a13(&net, tk::dnn::ACTIVATION_LEAKY); - tk::dnn::Conv2d c14(&net, 512, 3, 3, 1, 1, 1, 1, c14_bin, true); - tk::dnn::Activation a14(&net, tk::dnn::ACTIVATION_LEAKY); - tk::dnn::Conv2d c15(&net, 255, 1, 1, 1, 1, 0, 0, c15_bin, false); - - tk::dnn::Yolo yolo0 (&net, classes, 2, g16_bin); - - tk::dnn::Layer *m17_layers[1] = { &a13 }; - tk::dnn::Route m17 (&net, m17_layers, 1); - tk::dnn::Conv2d c18(&net, 128, 1, 1, 1, 1, 0, 0, c18_bin, true); - tk::dnn::Activation a18(&net, tk::dnn::ACTIVATION_LEAKY); - tk::dnn::Upsample u19 (&net, 2); - - tk::dnn::Layer *m20_layers[2] = { &u19, &a8 }; - tk::dnn::Route m20 (&net, m20_layers, 2); - - tk::dnn::Conv2d c21(&net, 256, 3, 3, 1, 1, 1, 1, c21_bin, true); - tk::dnn::Activation a21(&net, tk::dnn::ACTIVATION_LEAKY); - tk::dnn::Conv2d c22(&net, 255, 1, 1, 1, 1, 0, 0, c22_bin, false); - - tk::dnn::Yolo yolo1 (&net, classes, 2, g23_bin); - - // Load input - dnnType *data; - dnnType *input_h; - readBinaryFile(input_bin, dim.tot(), &input_h, &data); - - //print network model - net.print(); - - // convert network to tensorRT - tk::dnn::NetworkRT netRT(&net, "yolo3_tiny.rt"); - - dnnType *out_data, *out_data2; // cudnn output, tensorRT output - - tk::dnn::dataDim_t dim1 = dim; //input dim - printCenteredTitle(" CUDNN inference ", '=', 30); { - dim1.print(); - TIMER_START - out_data = net.infer(dim1, data); - TIMER_STOP - dim1.print(); - } - - tk::dnn::dataDim_t dim2 = dim; - printCenteredTitle(" TENSORRT inference ", '=', 30); { - dim2.print(); - TIMER_START - out_data2 = netRT.infer(dim2, data); - TIMER_STOP - dim2.print(); - } - - printCenteredTitle(" CHECK RESULTS ", '=', 30); - dnnType *out, *out_h; - int out_dim = net.getOutputDim().tot(); - readBinaryFile(output_bin, out_dim, &out_h, &out); - std::cout<<"CUDNN vs correct"; checkResult(out_dim, out_data, out); - std::cout<<"TRT vs correct"; checkResult(out_dim, out_data2, out); - std::cout<<"CUDNN vs TRT "; checkResult(out_dim, out_data, out_data2); - - return 0; -} diff --git a/tests/yolo3_tiny/yolov3-tiny.cfg b/tests/yolo3_tiny/yolov3-tiny.cfg deleted file mode 100644 index cfca3cf..0000000 --- a/tests/yolo3_tiny/yolov3-tiny.cfg +++ /dev/null @@ -1,182 +0,0 @@ -[net] -# Testing -batch=1 -subdivisions=1 -# Training -# batch=64 -# subdivisions=2 -width=416 -height=416 -channels=3 -momentum=0.9 -decay=0.0005 -angle=0 -saturation = 1.5 -exposure = 1.5 -hue=.1 - -learning_rate=0.001 -burn_in=1000 -max_batches = 500200 -policy=steps -steps=400000,450000 -scales=.1,.1 - -[convolutional] -batch_normalize=1 -filters=16 -size=3 -stride=1 -pad=1 -activation=leaky - -[maxpool] -size=2 -stride=2 - -[convolutional] -batch_normalize=1 -filters=32 -size=3 -stride=1 -pad=1 -activation=leaky - -[maxpool] -size=2 -stride=2 - -[convolutional] -batch_normalize=1 -filters=64 -size=3 -stride=1 -pad=1 -activation=leaky - -[maxpool] -size=2 -stride=2 - -[convolutional] -batch_normalize=1 -filters=128 -size=3 -stride=1 -pad=1 -activation=leaky - -[maxpool] -size=2 -stride=2 - -[convolutional] -batch_normalize=1 -filters=256 -size=3 -stride=1 -pad=1 -activation=leaky - -[maxpool] -size=2 -stride=2 - -[convolutional] -batch_normalize=1 -filters=512 -size=3 -stride=1 -pad=1 -activation=leaky - -[maxpool] -size=2 -stride=1 - -[convolutional] -batch_normalize=1 -filters=1024 -size=3 -stride=1 -pad=1 -activation=leaky - -########### - -[convolutional] -batch_normalize=1 -filters=256 -size=1 -stride=1 -pad=1 -activation=leaky - -[convolutional] -batch_normalize=1 -filters=512 -size=3 -stride=1 -pad=1 -activation=leaky - -[convolutional] -size=1 -stride=1 -pad=1 -filters=255 -activation=linear - - - -[yolo] -mask = 3,4,5 -anchors = 10,14, 23,27, 37,58, 81,82, 135,169, 344,319 -classes=80 -num=6 -jitter=.3 -ignore_thresh = .7 -truth_thresh = 1 -random=1 - -[route] -layers = -4 - -[convolutional] -batch_normalize=1 -filters=128 -size=1 -stride=1 -pad=1 -activation=leaky - -[upsample] -stride=2 - -[route] -layers = -1, 8 - -[convolutional] -batch_normalize=1 -filters=256 -size=3 -stride=1 -pad=1 -activation=leaky - -[convolutional] -size=1 -stride=1 -pad=1 -filters=255 -activation=linear - -[yolo] -mask = 0,1,2 -anchors = 10,14, 23,27, 37,58, 81,82, 135,169, 344,319 -classes=80 -num=6 -jitter=.3 -ignore_thresh = .7 -truth_thresh = 1 -random=1 diff --git a/tests/yolo_224/yolo_224.cfg b/tests/yolo_224/yolo_224.cfg deleted file mode 100644 index dd9206c..0000000 --- a/tests/yolo_224/yolo_224.cfg +++ /dev/null @@ -1,258 +0,0 @@ -[net] -# Testing -#batch=1 -#subdivisions=1 -# Training - batch=64 - subdivisions=16 -width=224 -height=224 -channels=3 -momentum=0.9 -decay=0.0005 -angle=0 -saturation = 1.5 -exposure = 1.5 -hue=.1 - -learning_rate=0.001 -burn_in=1000 -max_batches = 500200 -policy=steps -steps=400000,450000 -scales=.1,.1 - -[convolutional] -batch_normalize=1 -filters=32 -size=3 -stride=1 -pad=1 -activation=leaky - -[maxpool] -size=2 -stride=2 - -[convolutional] -batch_normalize=1 -filters=64 -size=3 -stride=1 -pad=1 -activation=leaky - -[maxpool] -size=2 -stride=2 - -[convolutional] -batch_normalize=1 -filters=128 -size=3 -stride=1 -pad=1 -activation=leaky - -[convolutional] -batch_normalize=1 -filters=64 -size=1 -stride=1 -pad=1 -activation=leaky - -[convolutional] -batch_normalize=1 -filters=128 -size=3 -stride=1 -pad=1 -activation=leaky - -[maxpool] -size=2 -stride=2 - -[convolutional] -batch_normalize=1 -filters=256 -size=3 -stride=1 -pad=1 -activation=leaky - -[convolutional] -batch_normalize=1 -filters=128 -size=1 -stride=1 -pad=1 -activation=leaky - -[convolutional] -batch_normalize=1 -filters=256 -size=3 -stride=1 -pad=1 -activation=leaky - -[maxpool] -size=2 -stride=2 - -[convolutional] -batch_normalize=1 -filters=512 -size=3 -stride=1 -pad=1 -activation=leaky - -[convolutional] -batch_normalize=1 -filters=256 -size=1 -stride=1 -pad=1 -activation=leaky - -[convolutional] -batch_normalize=1 -filters=512 -size=3 -stride=1 -pad=1 -activation=leaky - -[convolutional] -batch_normalize=1 -filters=256 -size=1 -stride=1 -pad=1 -activation=leaky - -[convolutional] -batch_normalize=1 -filters=512 -size=3 -stride=1 -pad=1 -activation=leaky - -[maxpool] -size=2 -stride=2 - -[convolutional] -batch_normalize=1 -filters=1024 -size=3 -stride=1 -pad=1 -activation=leaky - -[convolutional] -batch_normalize=1 -filters=512 -size=1 -stride=1 -pad=1 -activation=leaky - -[convolutional] -batch_normalize=1 -filters=1024 -size=3 -stride=1 -pad=1 -activation=leaky - -[convolutional] -batch_normalize=1 -filters=512 -size=1 -stride=1 -pad=1 -activation=leaky - -[convolutional] -batch_normalize=1 -filters=1024 -size=3 -stride=1 -pad=1 -activation=leaky - - -####### - -[convolutional] -batch_normalize=1 -size=3 -stride=1 -pad=1 -filters=1024 -activation=leaky - -[convolutional] -batch_normalize=1 -size=3 -stride=1 -pad=1 -filters=1024 -activation=leaky - -[route] -layers=-9 - -[convolutional] -batch_normalize=1 -size=1 -stride=1 -pad=1 -filters=64 -activation=leaky - -[reorg] -stride=2 - -[route] -layers=-1,-4 - -[convolutional] -batch_normalize=1 -size=3 -stride=1 -pad=1 -filters=1024 -activation=leaky - -[convolutional] -size=1 -stride=1 -pad=1 -filters=425 -activation=linear - - -[region] -anchors = 0.57273, 0.677385, 1.87446, 2.06253, 3.33843, 5.47434, 7.88282, 3.52778, 9.77052, 9.16828 -bias_match=1 -classes=80 -coords=4 -num=5 -softmax=1 -jitter=.3 -rescore=1 - -object_scale=5 -noobject_scale=1 -class_scale=1 -coord_scale=1 - -absolute=1 -thresh = .6 -random=1 diff --git a/tests/yolo_224/yolo_224.cpp b/tests/yolo_224/yolo_224.cpp deleted file mode 100644 index 9d6d28a..0000000 --- a/tests/yolo_224/yolo_224.cpp +++ /dev/null @@ -1,150 +0,0 @@ -#include -#include "tkdnn.h" - -const char *input_bin = "../tests/yolo_224/layers/input.bin"; -const char *c0_bin = "../tests/yolo_224/layers/c0.bin"; -const char *c2_bin = "../tests/yolo_224/layers/c2.bin"; -const char *c4_bin = "../tests/yolo_224/layers/c4.bin"; -const char *c5_bin = "../tests/yolo_224/layers/c5.bin"; -const char *c6_bin = "../tests/yolo_224/layers/c6.bin"; -const char *c8_bin = "../tests/yolo_224/layers/c8.bin"; -const char *c9_bin = "../tests/yolo_224/layers/c9.bin"; -const char *c10_bin = "../tests/yolo_224/layers/c10.bin"; -const char *c12_bin = "../tests/yolo_224/layers/c12.bin"; -const char *c13_bin = "../tests/yolo_224/layers/c13.bin"; -const char *c14_bin = "../tests/yolo_224/layers/c14.bin"; -const char *c15_bin = "../tests/yolo_224/layers/c15.bin"; -const char *c16_bin = "../tests/yolo_224/layers/c16.bin"; -const char *c18_bin = "../tests/yolo_224/layers/c18.bin"; -const char *c19_bin = "../tests/yolo_224/layers/c19.bin"; -const char *c20_bin = "../tests/yolo_224/layers/c20.bin"; -const char *c21_bin = "../tests/yolo_224/layers/c21.bin"; -const char *c22_bin = "../tests/yolo_224/layers/c22.bin"; -const char *c23_bin = "../tests/yolo_224/layers/c23.bin"; -const char *c24_bin = "../tests/yolo_224/layers/c24.bin"; -const char *c26_bin = "../tests/yolo_224/layers/c26.bin"; -const char *c29_bin = "../tests/yolo_224/layers/c29.bin"; -const char *c30_bin = "../tests/yolo_224/layers/c30.bin"; -const char *g31_bin = "../tests/yolo_224/layers/g31.bin"; -const char *output_bin = "../tests/yolo_224/layers/output.bin"; - -int main() { - - // Network layout - tk::dnn::dataDim_t dim(1, 3, 224, 224, 1); - tk::dnn::Network net(dim); - - tk::dnn::Conv2d c0 (&net, 32, 3, 3, 1, 1, 1, 1, c0_bin, true); - tk::dnn::Activation a0 (&net, tk::dnn::ACTIVATION_LEAKY); - tk::dnn::Pooling p1 (&net, 2, 2, 2, 2, tk::dnn::POOLING_MAX); - - tk::dnn::Conv2d c2 (&net, 64, 3, 3, 1, 1, 1, 1, c2_bin, true); - tk::dnn::Activation a2 (&net, tk::dnn::ACTIVATION_LEAKY); - tk::dnn::Pooling p3 (&net, 2, 2, 2, 2, tk::dnn::POOLING_MAX); - - tk::dnn::Conv2d c4 (&net, 128, 3, 3, 1, 1, 1, 1, c4_bin, true); - tk::dnn::Activation a4 (&net, tk::dnn::ACTIVATION_LEAKY); - tk::dnn::Conv2d c5 (&net, 64, 1, 1, 1, 1, 0, 0, c5_bin, true); - tk::dnn::Activation a5 (&net, tk::dnn::ACTIVATION_LEAKY); - tk::dnn::Conv2d c6 (&net, 128, 3, 3, 1, 1, 1, 1, c6_bin, true); - tk::dnn::Activation a6 (&net, tk::dnn::ACTIVATION_LEAKY); - tk::dnn::Pooling p7 (&net, 2, 2, 2, 2, tk::dnn::POOLING_MAX); - - tk::dnn::Conv2d c8 (&net, 256, 3, 3, 1, 1, 1, 1, c8_bin, true); - tk::dnn::Activation a8 (&net, tk::dnn::ACTIVATION_LEAKY); - tk::dnn::Conv2d c9 (&net, 128, 1, 1, 1, 1, 0, 0, c9_bin, true); - tk::dnn::Activation a9 (&net, tk::dnn::ACTIVATION_LEAKY); - tk::dnn::Conv2d c10(&net, 256, 3, 3, 1, 1, 1, 1, c10_bin, true); - tk::dnn::Activation a10(&net, tk::dnn::ACTIVATION_LEAKY); - tk::dnn::Pooling p11(&net, 2, 2, 2, 2, tk::dnn::POOLING_MAX); - - tk::dnn::Conv2d c12(&net, 512, 3, 3, 1, 1, 1, 1, c12_bin, true); - tk::dnn::Activation a12(&net, tk::dnn::ACTIVATION_LEAKY); - tk::dnn::Conv2d c13(&net, 256, 1, 1, 1, 1, 0, 0, c13_bin, true); - tk::dnn::Activation a13(&net, tk::dnn::ACTIVATION_LEAKY); - tk::dnn::Conv2d c14(&net, 512, 3, 3, 1, 1, 1, 1, c14_bin, true); - tk::dnn::Activation a14(&net, tk::dnn::ACTIVATION_LEAKY); - tk::dnn::Conv2d c15(&net, 256, 1, 1, 1, 1, 0, 0, c15_bin, true); - tk::dnn::Activation a15(&net, tk::dnn::ACTIVATION_LEAKY); - tk::dnn::Conv2d c16(&net, 512, 3, 3, 1, 1, 1, 1, c16_bin, true); - tk::dnn::Activation a16(&net, tk::dnn::ACTIVATION_LEAKY); - tk::dnn::Pooling p17(&net, 2, 2, 2, 2, tk::dnn::POOLING_MAX); - - tk::dnn::Conv2d c18(&net, 1024, 3, 3, 1, 1, 1, 1, c18_bin, true); - tk::dnn::Activation a18(&net, tk::dnn::ACTIVATION_LEAKY); - tk::dnn::Conv2d c19(&net, 512, 1, 1, 1, 1, 0, 0, c19_bin, true); - tk::dnn::Activation a19(&net, tk::dnn::ACTIVATION_LEAKY); - tk::dnn::Conv2d c20(&net, 1024, 3, 3, 1, 1, 1, 1, c20_bin, true); - tk::dnn::Activation a20(&net, tk::dnn::ACTIVATION_LEAKY); - tk::dnn::Conv2d c21(&net, 512, 1, 1, 1, 1, 0, 0, c21_bin, true); - tk::dnn::Activation a21(&net, tk::dnn::ACTIVATION_LEAKY); - tk::dnn::Conv2d c22(&net, 1024, 3, 3, 1, 1, 1, 1, c22_bin, true); - tk::dnn::Activation a22(&net, tk::dnn::ACTIVATION_LEAKY); - tk::dnn::Conv2d c23(&net, 1024, 3, 3, 1, 1, 1, 1, c23_bin, true); - tk::dnn::Activation a23(&net, tk::dnn::ACTIVATION_LEAKY); - tk::dnn::Conv2d c24(&net, 1024, 3, 3, 1, 1, 1, 1, c24_bin, true); - tk::dnn::Activation a24(&net, tk::dnn::ACTIVATION_LEAKY); - - tk::dnn::Layer *m25_layers[1] = { &a16 }; - tk::dnn::Route m25(&net, m25_layers, 1); - tk::dnn::Conv2d c26(&net, 64, 1, 1, 1, 1, 0, 0, c26_bin, true); - tk::dnn::Activation a26(&net, tk::dnn::ACTIVATION_LEAKY); - tk::dnn::Reorg r27(&net, 2); - - tk::dnn::Layer *m28_layers[2] = { &r27, &a24 }; - tk::dnn::Route m28(&net, m28_layers, 2); - - tk::dnn::Conv2d c29(&net, 1024, 3, 3, 1, 1, 1, 1, c29_bin, true); - tk::dnn::Activation a29(&net, tk::dnn::ACTIVATION_LEAKY); - tk::dnn::Conv2d c30(&net, 425, 1, 1, 1, 1, 0, 0, c30_bin, false); - tk::dnn::Region g31(&net, 80, 4, 5); - - tk::dnn::RegionInterpret rI(dim, g31.output_dim, 80, 4, 5, 0.6f, g31_bin); - - // Load input - dnnType *data; - dnnType *input_h; - readBinaryFile(input_bin, dim.tot(), &input_h, &data); - - //print network model - net.print(); - - //convert network to tensorRT - tk::dnn::NetworkRT netRT(&net, "yolo_224.rt"); - - dnnType *out_data, *out_data2; // cudnn output, tensorRT output - - tk::dnn::dataDim_t dim1 = dim; //input dim - printCenteredTitle(" CUDNN inference ", '=', 30); { - dim1.print(); - TIMER_START - out_data = net.infer(dim1, data); - TIMER_STOP - dim1.print(); - } - - tk::dnn::dataDim_t dim2 = dim; - printCenteredTitle(" TENSORRT inference ", '=', 30); { - dim2.print(); - TIMER_START - out_data2 = netRT.infer(dim2, data); - TIMER_STOP - dim2.print(); - } - - printCenteredTitle(" CHECK RESULTS ", '=', 30); - dnnType *out, *out_h; - int out_dim = net.getOutputDim().tot(); - readBinaryFile(output_bin, out_dim, &out_h, &out); - std::cout<<"CUDNN vs correct"; checkResult(out_dim, out_data, out); - std::cout<<"TRT vs correct"; checkResult(out_dim, out_data2, out); - std::cout<<"CUDNN vs TRT "; checkResult(out_dim, out_data, out_data2); - - std::cout<<"\n\nDetected objects: \n"; - dnnType *output_h = new dnnType[rI.output_dim.tot()]; - checkCuda(cudaMemcpy(output_h, out_data2, - rI.output_dim.tot()*sizeof(dnnType), cudaMemcpyDeviceToHost)); - rI.interpretData(output_h); - rI.showImageResult(input_h); - return 0; -} diff --git a/tests/yolo_berkeley/yolo_berkeley.cpp b/tests/yolo_berkeley/yolo_berkeley.cpp deleted file mode 100644 index dcb8d7d..0000000 --- a/tests/yolo_berkeley/yolo_berkeley.cpp +++ /dev/null @@ -1,150 +0,0 @@ -#include -#include "tkdnn.h" - -const char *input_bin = "../tests/yolo_berkeley/layers/input.bin"; -const char *c0_bin = "../tests/yolo_berkeley/layers/c0.bin"; -const char *c2_bin = "../tests/yolo_berkeley/layers/c2.bin"; -const char *c4_bin = "../tests/yolo_berkeley/layers/c4.bin"; -const char *c5_bin = "../tests/yolo_berkeley/layers/c5.bin"; -const char *c6_bin = "../tests/yolo_berkeley/layers/c6.bin"; -const char *c8_bin = "../tests/yolo_berkeley/layers/c8.bin"; -const char *c9_bin = "../tests/yolo_berkeley/layers/c9.bin"; -const char *c10_bin = "../tests/yolo_berkeley/layers/c10.bin"; -const char *c12_bin = "../tests/yolo_berkeley/layers/c12.bin"; -const char *c13_bin = "../tests/yolo_berkeley/layers/c13.bin"; -const char *c14_bin = "../tests/yolo_berkeley/layers/c14.bin"; -const char *c15_bin = "../tests/yolo_berkeley/layers/c15.bin"; -const char *c16_bin = "../tests/yolo_berkeley/layers/c16.bin"; -const char *c18_bin = "../tests/yolo_berkeley/layers/c18.bin"; -const char *c19_bin = "../tests/yolo_berkeley/layers/c19.bin"; -const char *c20_bin = "../tests/yolo_berkeley/layers/c20.bin"; -const char *c21_bin = "../tests/yolo_berkeley/layers/c21.bin"; -const char *c22_bin = "../tests/yolo_berkeley/layers/c22.bin"; -const char *c23_bin = "../tests/yolo_berkeley/layers/c23.bin"; -const char *c24_bin = "../tests/yolo_berkeley/layers/c24.bin"; -const char *c26_bin = "../tests/yolo_berkeley/layers/c26.bin"; -const char *c29_bin = "../tests/yolo_berkeley/layers/c29.bin"; -const char *c30_bin = "../tests/yolo_berkeley/layers/c30.bin"; -const char *g31_bin = "../tests/yolo_berkeley/layers/g31.bin"; -const char *output_bin = "../tests/yolo_berkeley/layers/output.bin"; - -int main() { - - // Network layout - tk::dnn::dataDim_t dim(1, 3, 416, 736, 1); - tk::dnn::Network net(dim); - - tk::dnn::Conv2d c0 (&net, 32, 3, 3, 1, 1, 1, 1, c0_bin, true); - tk::dnn::Activation a0 (&net, tk::dnn::ACTIVATION_LEAKY); - tk::dnn::Pooling p1 (&net, 2, 2, 2, 2, tk::dnn::POOLING_MAX); - - tk::dnn::Conv2d c2 (&net, 64, 3, 3, 1, 1, 1, 1, c2_bin, true); - tk::dnn::Activation a2 (&net, tk::dnn::ACTIVATION_LEAKY); - tk::dnn::Pooling p3 (&net, 2, 2, 2, 2, tk::dnn::POOLING_MAX); - - tk::dnn::Conv2d c4 (&net, 128, 3, 3, 1, 1, 1, 1, c4_bin, true); - tk::dnn::Activation a4 (&net, tk::dnn::ACTIVATION_LEAKY); - tk::dnn::Conv2d c5 (&net, 64, 1, 1, 1, 1, 0, 0, c5_bin, true); - tk::dnn::Activation a5 (&net, tk::dnn::ACTIVATION_LEAKY); - tk::dnn::Conv2d c6 (&net, 128, 3, 3, 1, 1, 1, 1, c6_bin, true); - tk::dnn::Activation a6 (&net, tk::dnn::ACTIVATION_LEAKY); - tk::dnn::Pooling p7 (&net, 2, 2, 2, 2, tk::dnn::POOLING_MAX); - - tk::dnn::Conv2d c8 (&net, 256, 3, 3, 1, 1, 1, 1, c8_bin, true); - tk::dnn::Activation a8 (&net, tk::dnn::ACTIVATION_LEAKY); - tk::dnn::Conv2d c9 (&net, 128, 1, 1, 1, 1, 0, 0, c9_bin, true); - tk::dnn::Activation a9 (&net, tk::dnn::ACTIVATION_LEAKY); - tk::dnn::Conv2d c10(&net, 256, 3, 3, 1, 1, 1, 1, c10_bin, true); - tk::dnn::Activation a10(&net, tk::dnn::ACTIVATION_LEAKY); - tk::dnn::Pooling p11(&net, 2, 2, 2, 2, tk::dnn::POOLING_MAX); - - tk::dnn::Conv2d c12(&net, 512, 3, 3, 1, 1, 1, 1, c12_bin, true); - tk::dnn::Activation a12(&net, tk::dnn::ACTIVATION_LEAKY); - tk::dnn::Conv2d c13(&net, 256, 1, 1, 1, 1, 0, 0, c13_bin, true); - tk::dnn::Activation a13(&net, tk::dnn::ACTIVATION_LEAKY); - tk::dnn::Conv2d c14(&net, 512, 3, 3, 1, 1, 1, 1, c14_bin, true); - tk::dnn::Activation a14(&net, tk::dnn::ACTIVATION_LEAKY); - tk::dnn::Conv2d c15(&net, 256, 1, 1, 1, 1, 0, 0, c15_bin, true); - tk::dnn::Activation a15(&net, tk::dnn::ACTIVATION_LEAKY); - tk::dnn::Conv2d c16(&net, 512, 3, 3, 1, 1, 1, 1, c16_bin, true); - tk::dnn::Activation a16(&net, tk::dnn::ACTIVATION_LEAKY); - tk::dnn::Pooling p17(&net, 2, 2, 2, 2, tk::dnn::POOLING_MAX); - - tk::dnn::Conv2d c18(&net, 1024, 3, 3, 1, 1, 1, 1, c18_bin, true); - tk::dnn::Activation a18(&net, tk::dnn::ACTIVATION_LEAKY); - tk::dnn::Conv2d c19(&net, 512, 1, 1, 1, 1, 0, 0, c19_bin, true); - tk::dnn::Activation a19(&net, tk::dnn::ACTIVATION_LEAKY); - tk::dnn::Conv2d c20(&net, 1024, 3, 3, 1, 1, 1, 1, c20_bin, true); - tk::dnn::Activation a20(&net, tk::dnn::ACTIVATION_LEAKY); - tk::dnn::Conv2d c21(&net, 512, 1, 1, 1, 1, 0, 0, c21_bin, true); - tk::dnn::Activation a21(&net, tk::dnn::ACTIVATION_LEAKY); - tk::dnn::Conv2d c22(&net, 1024, 3, 3, 1, 1, 1, 1, c22_bin, true); - tk::dnn::Activation a22(&net, tk::dnn::ACTIVATION_LEAKY); - tk::dnn::Conv2d c23(&net, 1024, 3, 3, 1, 1, 1, 1, c23_bin, true); - tk::dnn::Activation a23(&net, tk::dnn::ACTIVATION_LEAKY); - tk::dnn::Conv2d c24(&net, 1024, 3, 3, 1, 1, 1, 1, c24_bin, true); - tk::dnn::Activation a24(&net, tk::dnn::ACTIVATION_LEAKY); - - tk::dnn::Layer *m25_layers[1] = { &a16 }; - tk::dnn::Route m25(&net, m25_layers, 1); - tk::dnn::Conv2d c26(&net, 64, 1, 1, 1, 1, 0, 0, c26_bin, true); - tk::dnn::Activation a26(&net, tk::dnn::ACTIVATION_LEAKY); - tk::dnn::Reorg r27(&net, 2); - - tk::dnn::Layer *m28_layers[2] = { &r27, &a24 }; - tk::dnn::Route m28(&net, m28_layers, 2); - - tk::dnn::Conv2d c29(&net, 1024, 3, 3, 1, 1, 1, 1, c29_bin, true); - tk::dnn::Activation a29(&net, tk::dnn::ACTIVATION_LEAKY); - tk::dnn::Conv2d c30(&net, 75, 1, 1, 1, 1, 0, 0, c30_bin, false); - tk::dnn::Region g31(&net, 10, 4, 5); - - tk::dnn::RegionInterpret rI(dim, g31.output_dim, 10, 4, 5, 0.3f, g31_bin); - - // Load input - dnnType *data; - dnnType *input_h; - readBinaryFile(input_bin, dim.tot(), &input_h, &data); - - //print network model - net.print(); - - //convert network to tensorRT - tk::dnn::NetworkRT netRT(&net, "yolo_berkeley.rt"); - - dnnType *out_data, *out_data2; // cudnn output, tensorRT output - - tk::dnn::dataDim_t dim1 = dim; //input dim - printCenteredTitle(" CUDNN inference ", '=', 30); { - dim1.print(); - TIMER_START - out_data = net.infer(dim1, data); - TIMER_STOP - dim1.print(); - } - - tk::dnn::dataDim_t dim2 = dim; - printCenteredTitle(" TENSORRT inference ", '=', 30); { - dim2.print(); - TIMER_START - out_data2 = netRT.infer(dim2, data); - TIMER_STOP - dim2.print(); - } - - printCenteredTitle(" CHECK RESULTS ", '=', 30); - dnnType *out, *out_h; - int out_dim = net.getOutputDim().tot(); - readBinaryFile(output_bin, out_dim, &out_h, &out); - std::cout<<"CUDNN vs correct"; checkResult(out_dim, out_data, out); - std::cout<<"TRT vs correct"; checkResult(out_dim, out_data2, out); - std::cout<<"CUDNN vs TRT "; checkResult(out_dim, out_data, out_data2); - - std::cout<<"\n\nDetected objects: \n"; - dnnType *output_h = new dnnType[rI.output_dim.tot()]; - checkCuda(cudaMemcpy(output_h, out_data2, - rI.output_dim.tot()*sizeof(dnnType), cudaMemcpyDeviceToHost)); - rI.interpretData(output_h); - rI.showImageResult(input_h); - return 0; -} diff --git a/tests/yolo_berkeley/yolov2-voc-10-resize-test.cfg b/tests/yolo_berkeley/yolov2-voc-10-resize-test.cfg deleted file mode 100644 index 184b32c..0000000 --- a/tests/yolo_berkeley/yolov2-voc-10-resize-test.cfg +++ /dev/null @@ -1,259 +0,0 @@ -[net] -# Testing -batch=1 -subdivisions=1 -# Training -#batch=64 -#subdivisions=8 -height=416 -width=736 -channels=3 -momentum=0.9 -decay=0.0005 -angle=0 -saturation = 1.5 -exposure = 1.5 -hue=.1 - -learning_rate=0.001 -burn_in=1000 -max_batches = 80200 -policy=steps -steps=40000,60000 -scales=.1,.1 - -[convolutional] -batch_normalize=1 -filters=32 -size=3 -stride=1 -pad=1 -activation=leaky - -[maxpool] -size=2 -stride=2 - -[convolutional] -batch_normalize=1 -filters=64 -size=3 -stride=1 -pad=1 -activation=leaky - -[maxpool] -size=2 -stride=2 - -[convolutional] -batch_normalize=1 -filters=128 -size=3 -stride=1 -pad=1 -activation=leaky - -[convolutional] -batch_normalize=1 -filters=64 -size=1 -stride=1 -pad=1 -activation=leaky - -[convolutional] -batch_normalize=1 -filters=128 -size=3 -stride=1 -pad=1 -activation=leaky - -[maxpool] -size=2 -stride=2 - -[convolutional] -batch_normalize=1 -filters=256 -size=3 -stride=1 -pad=1 -activation=leaky - -[convolutional] -batch_normalize=1 -filters=128 -size=1 -stride=1 -pad=1 -activation=leaky - -[convolutional] -batch_normalize=1 -filters=256 -size=3 -stride=1 -pad=1 -activation=leaky - -[maxpool] -size=2 -stride=2 - -[convolutional] -batch_normalize=1 -filters=512 -size=3 -stride=1 -pad=1 -activation=leaky - -[convolutional] -batch_normalize=1 -filters=256 -size=1 -stride=1 -pad=1 -activation=leaky - -[convolutional] -batch_normalize=1 -filters=512 -size=3 -stride=1 -pad=1 -activation=leaky - -[convolutional] -batch_normalize=1 -filters=256 -size=1 -stride=1 -pad=1 -activation=leaky - -[convolutional] -batch_normalize=1 -filters=512 -size=3 -stride=1 -pad=1 -activation=leaky - -[maxpool] -size=2 -stride=2 - -[convolutional] -batch_normalize=1 -filters=1024 -size=3 -stride=1 -pad=1 -activation=leaky - -[convolutional] -batch_normalize=1 -filters=512 -size=1 -stride=1 -pad=1 -activation=leaky - -[convolutional] -batch_normalize=1 -filters=1024 -size=3 -stride=1 -pad=1 -activation=leaky - -[convolutional] -batch_normalize=1 -filters=512 -size=1 -stride=1 -pad=1 -activation=leaky - -[convolutional] -batch_normalize=1 -filters=1024 -size=3 -stride=1 -pad=1 -activation=leaky - - -####### - -[convolutional] -batch_normalize=1 -size=3 -stride=1 -pad=1 -filters=1024 -activation=leaky - -[convolutional] -batch_normalize=1 -size=3 -stride=1 -pad=1 -filters=1024 -activation=leaky - -[route] -layers=-9 - -[convolutional] -batch_normalize=1 -size=1 -stride=1 -pad=1 -filters=64 -activation=leaky - -[reorg] -stride=2 - -[route] -layers=-1,-4 - -[convolutional] -batch_normalize=1 -size=3 -stride=1 -pad=1 -filters=1024 -activation=leaky - -[convolutional] -size=1 -stride=1 -pad=1 -filters=75 -activation=linear - - -[region] -anchors = 0.4043,0.4167, 1.2109,1.1018, 2.7258,2.1215, 4.9477,3.9132, 7.9508,6.6806 -bias_match=1 -classes=10 -coords=4 -num=5 -softmax=1 -jitter=.3 -rescore=1 - -object_scale=5 -noobject_scale=1 -class_scale=1 -coord_scale=1 - -absolute=1 -thresh = .6 -random=0 -flip=1 diff --git a/tests/yolo_relu/yolo_relu.cfg b/tests/yolo_relu/yolo_relu.cfg deleted file mode 100644 index 0abab4e..0000000 --- a/tests/yolo_relu/yolo_relu.cfg +++ /dev/null @@ -1,258 +0,0 @@ -[net] -# Testing -#batch=1 -#subdivisions=1 -# Training - batch=64 - subdivisions=16 -width=608 -height=608 -channels=3 -momentum=0.9 -decay=0.0005 -angle=0 -saturation = 1.5 -exposure = 1.5 -hue=.1 - -learning_rate=0.001 -burn_in=1000 -max_batches = 500200 -policy=steps -steps=400000,450000 -scales=.1,.1 - -[convolutional] -batch_normalize=1 -filters=32 -size=3 -stride=1 -pad=1 -activation=relu - -[maxpool] -size=2 -stride=2 - -[convolutional] -batch_normalize=1 -filters=64 -size=3 -stride=1 -pad=1 -activation=relu - -[maxpool] -size=2 -stride=2 - -[convolutional] -batch_normalize=1 -filters=128 -size=3 -stride=1 -pad=1 -activation=relu - -[convolutional] -batch_normalize=1 -filters=64 -size=1 -stride=1 -pad=1 -activation=relu - -[convolutional] -batch_normalize=1 -filters=128 -size=3 -stride=1 -pad=1 -activation=relu - -[maxpool] -size=2 -stride=2 - -[convolutional] -batch_normalize=1 -filters=256 -size=3 -stride=1 -pad=1 -activation=relu - -[convolutional] -batch_normalize=1 -filters=128 -size=1 -stride=1 -pad=1 -activation=relu - -[convolutional] -batch_normalize=1 -filters=256 -size=3 -stride=1 -pad=1 -activation=relu - -[maxpool] -size=2 -stride=2 - -[convolutional] -batch_normalize=1 -filters=512 -size=3 -stride=1 -pad=1 -activation=relu - -[convolutional] -batch_normalize=1 -filters=256 -size=1 -stride=1 -pad=1 -activation=relu - -[convolutional] -batch_normalize=1 -filters=512 -size=3 -stride=1 -pad=1 -activation=relu - -[convolutional] -batch_normalize=1 -filters=256 -size=1 -stride=1 -pad=1 -activation=relu - -[convolutional] -batch_normalize=1 -filters=512 -size=3 -stride=1 -pad=1 -activation=relu - -[maxpool] -size=2 -stride=2 - -[convolutional] -batch_normalize=1 -filters=1024 -size=3 -stride=1 -pad=1 -activation=relu - -[convolutional] -batch_normalize=1 -filters=512 -size=1 -stride=1 -pad=1 -activation=relu - -[convolutional] -batch_normalize=1 -filters=1024 -size=3 -stride=1 -pad=1 -activation=relu - -[convolutional] -batch_normalize=1 -filters=512 -size=1 -stride=1 -pad=1 -activation=relu - -[convolutional] -batch_normalize=1 -filters=1024 -size=3 -stride=1 -pad=1 -activation=relu - - -####### - -[convolutional] -batch_normalize=1 -size=3 -stride=1 -pad=1 -filters=1024 -activation=relu - -[convolutional] -batch_normalize=1 -size=3 -stride=1 -pad=1 -filters=1024 -activation=relu - -[route] -layers=-9 - -[convolutional] -batch_normalize=1 -size=1 -stride=1 -pad=1 -filters=64 -activation=relu - -[reorg] -stride=2 - -[route] -layers=-1,-4 - -[convolutional] -batch_normalize=1 -size=3 -stride=1 -pad=1 -filters=1024 -activation=relu - -[convolutional] -size=1 -stride=1 -pad=1 -filters=425 -activation=linear - - -[region] -anchors = 0.57273, 0.677385, 1.87446, 2.06253, 3.33843, 5.47434, 7.88282, 3.52778, 9.77052, 9.16828 -bias_match=1 -classes=80 -coords=4 -num=5 -softmax=1 -jitter=.3 -rescore=1 - -object_scale=5 -noobject_scale=1 -class_scale=1 -coord_scale=1 - -absolute=1 -thresh = .6 -random=1 diff --git a/tests/yolo_relu/yolo_relu.cpp b/tests/yolo_relu/yolo_relu.cpp deleted file mode 100644 index ddb0802..0000000 --- a/tests/yolo_relu/yolo_relu.cpp +++ /dev/null @@ -1,151 +0,0 @@ -#include -#include "tkdnn.h" - -const char *input_bin = "../tests/yolo_relu/layers/input.bin"; -const char *c0_bin = "../tests/yolo_relu/layers/c0.bin"; -const char *c2_bin = "../tests/yolo_relu/layers/c2.bin"; -const char *c4_bin = "../tests/yolo_relu/layers/c4.bin"; -const char *c5_bin = "../tests/yolo_relu/layers/c5.bin"; -const char *c6_bin = "../tests/yolo_relu/layers/c6.bin"; -const char *c8_bin = "../tests/yolo_relu/layers/c8.bin"; -const char *c9_bin = "../tests/yolo_relu/layers/c9.bin"; -const char *c10_bin = "../tests/yolo_relu/layers/c10.bin"; -const char *c12_bin = "../tests/yolo_relu/layers/c12.bin"; -const char *c13_bin = "../tests/yolo_relu/layers/c13.bin"; -const char *c14_bin = "../tests/yolo_relu/layers/c14.bin"; -const char *c15_bin = "../tests/yolo_relu/layers/c15.bin"; -const char *c16_bin = "../tests/yolo_relu/layers/c16.bin"; -const char *c18_bin = "../tests/yolo_relu/layers/c18.bin"; -const char *c19_bin = "../tests/yolo_relu/layers/c19.bin"; -const char *c20_bin = "../tests/yolo_relu/layers/c20.bin"; -const char *c21_bin = "../tests/yolo_relu/layers/c21.bin"; -const char *c22_bin = "../tests/yolo_relu/layers/c22.bin"; -const char *c23_bin = "../tests/yolo_relu/layers/c23.bin"; -const char *c24_bin = "../tests/yolo_relu/layers/c24.bin"; -const char *c26_bin = "../tests/yolo_relu/layers/c26.bin"; -const char *c29_bin = "../tests/yolo_relu/layers/c29.bin"; -const char *c30_bin = "../tests/yolo_relu/layers/c30.bin"; -const char *g31_bin = "../tests/yolo_relu/layers/g31.bin"; -const char *output_bin = "../tests/yolo_relu/layers/output.bin"; - -int main() { - - // Network layout - tk::dnn::dataDim_t dim(1, 3, 608, 608, 1); - tk::dnn::Network net(dim); - - tk::dnn::Conv2d c0 (&net, 32, 3, 3, 1, 1, 1, 1, c0_bin, true); - tk::dnn::Activation a0 (&net, CUDNN_ACTIVATION_RELU); - tk::dnn::Pooling p1 (&net, 2, 2, 2, 2, tk::dnn::POOLING_MAX); - - tk::dnn::Conv2d c2 (&net, 64, 3, 3, 1, 1, 1, 1, c2_bin, true); - tk::dnn::Activation a2 (&net, CUDNN_ACTIVATION_RELU); - tk::dnn::Pooling p3 (&net, 2, 2, 2, 2, tk::dnn::POOLING_MAX); - - tk::dnn::Conv2d c4 (&net, 128, 3, 3, 1, 1, 1, 1, c4_bin, true); - tk::dnn::Activation a4 (&net, CUDNN_ACTIVATION_RELU); - tk::dnn::Conv2d c5 (&net, 64, 1, 1, 1, 1, 0, 0, c5_bin, true); - tk::dnn::Activation a5 (&net, CUDNN_ACTIVATION_RELU); - tk::dnn::Conv2d c6 (&net, 128, 3, 3, 1, 1, 1, 1, c6_bin, true); - tk::dnn::Activation a6 (&net, CUDNN_ACTIVATION_RELU); - tk::dnn::Pooling p7 (&net, 2, 2, 2, 2, tk::dnn::POOLING_MAX); - - tk::dnn::Conv2d c8 (&net, 256, 3, 3, 1, 1, 1, 1, c8_bin, true); - tk::dnn::Activation a8 (&net, CUDNN_ACTIVATION_RELU); - tk::dnn::Conv2d c9 (&net, 128, 1, 1, 1, 1, 0, 0, c9_bin, true); - tk::dnn::Activation a9 (&net, CUDNN_ACTIVATION_RELU); - tk::dnn::Conv2d c10(&net, 256, 3, 3, 1, 1, 1, 1, c10_bin, true); - tk::dnn::Activation a10(&net, CUDNN_ACTIVATION_RELU); - tk::dnn::Pooling p11(&net, 2, 2, 2, 2, tk::dnn::POOLING_MAX); - - tk::dnn::Conv2d c12(&net, 512, 3, 3, 1, 1, 1, 1, c12_bin, true); - tk::dnn::Activation a12(&net, CUDNN_ACTIVATION_RELU); - tk::dnn::Conv2d c13(&net, 256, 1, 1, 1, 1, 0, 0, c13_bin, true); - tk::dnn::Activation a13(&net, CUDNN_ACTIVATION_RELU); - tk::dnn::Conv2d c14(&net, 512, 3, 3, 1, 1, 1, 1, c14_bin, true); - tk::dnn::Activation a14(&net, CUDNN_ACTIVATION_RELU); - tk::dnn::Conv2d c15(&net, 256, 1, 1, 1, 1, 0, 0, c15_bin, true); - tk::dnn::Activation a15(&net, CUDNN_ACTIVATION_RELU); - tk::dnn::Conv2d c16(&net, 512, 3, 3, 1, 1, 1, 1, c16_bin, true); - tk::dnn::Activation a16(&net, CUDNN_ACTIVATION_RELU); - tk::dnn::Pooling p17(&net, 2, 2, 2, 2, tk::dnn::POOLING_MAX); - - tk::dnn::Conv2d c18(&net, 1024, 3, 3, 1, 1, 1, 1, c18_bin, true); - tk::dnn::Activation a18(&net, CUDNN_ACTIVATION_RELU); - tk::dnn::Conv2d c19(&net, 512, 1, 1, 1, 1, 0, 0, c19_bin, true); - tk::dnn::Activation a19(&net, CUDNN_ACTIVATION_RELU); - tk::dnn::Conv2d c20(&net, 1024, 3, 3, 1, 1, 1, 1, c20_bin, true); - tk::dnn::Activation a20(&net, CUDNN_ACTIVATION_RELU); - tk::dnn::Conv2d c21(&net, 512, 1, 1, 1, 1, 0, 0, c21_bin, true); - tk::dnn::Activation a21(&net, CUDNN_ACTIVATION_RELU); - tk::dnn::Conv2d c22(&net, 1024, 3, 3, 1, 1, 1, 1, c22_bin, true); - tk::dnn::Activation a22(&net, CUDNN_ACTIVATION_RELU); - tk::dnn::Conv2d c23(&net, 1024, 3, 3, 1, 1, 1, 1, c23_bin, true); - tk::dnn::Activation a23(&net, CUDNN_ACTIVATION_RELU); - tk::dnn::Conv2d c24(&net, 1024, 3, 3, 1, 1, 1, 1, c24_bin, true); - tk::dnn::Activation a24(&net, CUDNN_ACTIVATION_RELU); - - tk::dnn::Layer *m25_layers[1] = { &a16 }; - tk::dnn::Route m25(&net, m25_layers, 1); - tk::dnn::Conv2d c26(&net, 64, 1, 1, 1, 1, 0, 0, c26_bin, true); - tk::dnn::Activation a26(&net, CUDNN_ACTIVATION_RELU); - tk::dnn::Reorg r27(&net, 2); - - tk::dnn::Layer *m28_layers[2] = { &r27, &a24 }; - tk::dnn::Route m28(&net, m28_layers, 2); - - tk::dnn::Conv2d c29(&net, 1024, 3, 3, 1, 1, 1, 1, c29_bin, true); - tk::dnn::Activation a29(&net, CUDNN_ACTIVATION_RELU); - tk::dnn::Conv2d c30(&net, 425, 1, 1, 1, 1, 0, 0, c30_bin, false); - tk::dnn::Region g31(&net, 80, 4, 5); - - tk::dnn::RegionInterpret rI(dim, g31.output_dim, 80, 4, 5, 0.3f, g31_bin); - - // Load input - dnnType *data; - dnnType *input_h; - readBinaryFile(input_bin, dim.tot(), &input_h, &data); - - //print network model - net.print(); - - //convert network to tensorRT - tk::dnn::NetworkRT netRT(&net, "yolo_relu.rt"); - - dnnType *out_data, *out_data2; // cudnn output, tensorRT output - - tk::dnn::dataDim_t dim1 = dim; //input dim - printCenteredTitle(" CUDNN inference ", '=', 30); { - dim1.print(); - TIMER_START - out_data = net.infer(dim1, data); - TIMER_STOP - dim1.print(); - } - - tk::dnn::dataDim_t dim2 = dim; - printCenteredTitle(" TENSORRT inference ", '=', 30); { - dim2.print(); - TIMER_START - out_data2 = netRT.infer(dim2, data); - TIMER_STOP - dim2.print(); - } - - printCenteredTitle(" CHECK RESULTS ", '=', 30); - dnnType *out, *out_h; - int out_dim = net.getOutputDim().tot(); - readBinaryFile(output_bin, out_dim, &out_h, &out); - std::cout<<"CUDNN vs correct"; checkResult(out_dim, out_data, out); - std::cout<<"TRT vs correct"; checkResult(out_dim, out_data2, out); - std::cout<<"CUDNN vs TRT "; checkResult(out_dim, out_data, out_data2); - - std::cout<<"\n\nDetected objects: \n"; - dnnType *output_h = new dnnType[rI.output_dim.tot()]; - checkCuda(cudaMemcpy(output_h, out_data2, - rI.output_dim.tot()*sizeof(dnnType), cudaMemcpyDeviceToHost)); - rI.interpretData(output_h, 608, 608); - rI.showImageResult(input_h); - - return 0; -} diff --git a/tests/yolo_tiny/tiny-yolo.cfg b/tests/yolo_tiny/tiny-yolo.cfg deleted file mode 100644 index 630a209..0000000 --- a/tests/yolo_tiny/tiny-yolo.cfg +++ /dev/null @@ -1,139 +0,0 @@ -[net] - Training - batch=64 - subdivisions=8 -# Testing -# batch=1 -# subdivisions=1 -width=416 -height=416 -channels=3 -momentum=0.9 -decay=0.0005 -angle=0 -saturation = 1.5 -exposure = 1.5 -hue=.1 - -learning_rate=0.001 -burn_in=1000 -max_batches = 500200 -policy=steps -steps=400000,450000 -scales=.1,.1 - -[convolutional] -batch_normalize=1 -filters=16 -size=3 -stride=1 -pad=1 -activation=leaky - -[maxpool] -size=2 -stride=2 - -[convolutional] -batch_normalize=1 -filters=32 -size=3 -stride=1 -pad=1 -activation=leaky - -[maxpool] -size=2 -stride=2 - -[convolutional] -batch_normalize=1 -filters=64 -size=3 -stride=1 -pad=1 -activation=leaky - -[maxpool] -size=2 -stride=2 - -[convolutional] -batch_normalize=1 -filters=128 -size=3 -stride=1 -pad=1 -activation=leaky - -[maxpool] -size=2 -stride=2 - -[convolutional] -batch_normalize=1 -filters=256 -size=3 -stride=1 -pad=1 -activation=leaky - -[maxpool] -size=2 -stride=2 - -[convolutional] -batch_normalize=1 -filters=512 -size=3 -stride=1 -pad=1 -activation=leaky - -#[maxpool] -#size=2 -#stride=1 - -[convolutional] -batch_normalize=1 -filters=1024 -size=3 -stride=1 -pad=1 -activation=leaky - -########### - -[convolutional] -batch_normalize=1 -size=3 -stride=1 -pad=1 -filters=512 -activation=leaky - -[convolutional] -size=1 -stride=1 -pad=1 -filters=425 -activation=linear - -[region] -anchors = 0.57273, 0.677385, 1.87446, 2.06253, 3.33843, 5.47434, 7.88282, 3.52778, 9.77052, 9.16828 -bias_match=1 -classes=80 -coords=4 -num=5 -softmax=1 -jitter=.2 -rescore=0 - -object_scale=5 -noobject_scale=1 -class_scale=1 -coord_scale=1 - -absolute=1 -thresh = .6 -random=1 diff --git a/tests/yolo_tiny/yolo_tiny.cpp b/tests/yolo_tiny/yolo_tiny.cpp deleted file mode 100644 index 0255108..0000000 --- a/tests/yolo_tiny/yolo_tiny.cpp +++ /dev/null @@ -1,93 +0,0 @@ -#include -#include "tkdnn.h" - -const char *input_bin = "../tests/yolo_tiny/layers/input.bin"; -const char *c0_bin = "../tests/yolo_tiny/layers/c0.bin"; -const char *c2_bin = "../tests/yolo_tiny/layers/c2.bin"; -const char *c4_bin = "../tests/yolo_tiny/layers/c4.bin"; -const char *c5_bin = "../tests/yolo_tiny/layers/c5.bin"; -const char *c6_bin = "../tests/yolo_tiny/layers/c6.bin"; -const char *c8_bin = "../tests/yolo_tiny/layers/c8.bin"; -const char *c10_bin = "../tests/yolo_tiny/layers/c10.bin"; -const char *c11_bin = "../tests/yolo_tiny/layers/c11.bin"; -const char *c12_bin = "../tests/yolo_tiny/layers/c12.bin"; -const char *c13_bin = "../tests/yolo_tiny/layers/c13.bin"; -const char *g14_bin = "../tests/yolo_tiny/layers/g14.bin"; -const char *output_bin = "../tests/yolo_tiny/layers/output.bin"; - -int main() { - - // Network layout - tk::dnn::dataDim_t dim(1, 3, 416, 416, 1); - tk::dnn::Network net(dim); - - tk::dnn::Conv2d c0 (&net, 16, 3, 3, 1, 1, 1, 1, c0_bin, true); - tk::dnn::Activation a0 (&net, tk::dnn::ACTIVATION_LEAKY); - tk::dnn::Pooling p1 (&net, 2, 2, 2, 2, tk::dnn::POOLING_MAX); - - tk::dnn::Conv2d c2 (&net, 32, 3, 3, 1, 1, 1, 1, c2_bin, true); - tk::dnn::Activation a2 (&net, tk::dnn::ACTIVATION_LEAKY); - tk::dnn::Pooling p3 (&net, 2, 2, 2, 2, tk::dnn::POOLING_MAX); - - tk::dnn::Conv2d c4 (&net, 64, 3, 3, 1, 1, 1, 1, c4_bin, true); - tk::dnn::Activation a4 (&net, tk::dnn::ACTIVATION_LEAKY); - tk::dnn::Pooling p5 (&net, 2, 2, 2, 2, tk::dnn::POOLING_MAX); - - tk::dnn::Conv2d c6 (&net, 128, 3, 3, 1, 1, 1, 1, c6_bin, true); - tk::dnn::Activation a6 (&net, tk::dnn::ACTIVATION_LEAKY); - tk::dnn::Pooling p7(&net, 2, 2, 2, 2, tk::dnn::POOLING_MAX); - - tk::dnn::Conv2d c8(&net, 256, 3, 3, 1, 1, 1, 1, c8_bin, true); - tk::dnn::Activation a8(&net, tk::dnn::ACTIVATION_LEAKY); - tk::dnn::Pooling p9(&net, 2, 2, 2, 2, tk::dnn::POOLING_MAX); - - tk::dnn::Conv2d c10(&net, 512, 3, 3, 1, 1, 1, 1, c10_bin, true); - tk::dnn::Activation a10(&net, tk::dnn::ACTIVATION_LEAKY); - - tk::dnn::Conv2d c11(&net, 1024, 3, 3, 1, 1, 1, 1, c11_bin, true); - tk::dnn::Activation a11(&net, tk::dnn::ACTIVATION_LEAKY); - tk::dnn::Conv2d c12(&net, 512, 3, 3, 1, 1, 1, 1, c12_bin, true); - tk::dnn::Activation a12(&net, tk::dnn::ACTIVATION_LEAKY); - tk::dnn::Conv2d c13(&net, 425, 1, 1, 1, 1, 0, 0, c13_bin, false); - tk::dnn::Region g14(&net, 80, 4, 5); - - // Load input - dnnType *data; - dnnType *input_h; - readBinaryFile(input_bin, dim.tot(), &input_h, &data); - - //print network model - net.print(); - - //convert network to tensorRT - tk::dnn::NetworkRT netRT(&net, "yolo_tiny.rt"); - - dnnType *out_data, *out_data2; // cudnn output, tensorRT output - - tk::dnn::dataDim_t dim1 = dim; //input dim - printCenteredTitle(" CUDNN inference ", '=', 30); { - dim1.print(); - TIMER_START - out_data = net.infer(dim1, data); - TIMER_STOP - dim1.print(); - } - - tk::dnn::dataDim_t dim2 = dim; - printCenteredTitle(" TENSORRT inference ", '=', 30); { - dim2.print(); - TIMER_START - out_data2 = netRT.infer(dim2, data); - TIMER_STOP - dim2.print(); - } - - printCenteredTitle(" CHECK RESULTS ", '=', 30); - dnnType *out, *out_h; - int out_dim = net.getOutputDim().tot(); - readBinaryFile(output_bin, out_dim, &out_h, &out); - std::cout<<"CUDNN vs correct"; checkResult(out_dim, out_data, out); - std::cout<<"TRT vs correct"; checkResult(out_dim, out_data2, out); - std::cout<<"CUDNN vs TRT "; checkResult(out_dim, out_data, out_data2); - return 0; -} diff --git a/tests/yolo_voc/yolo_voc.cfg b/tests/yolo_voc/yolo_voc.cfg deleted file mode 100644 index dbf2de2..0000000 --- a/tests/yolo_voc/yolo_voc.cfg +++ /dev/null @@ -1,258 +0,0 @@ -[net] -# Testing -batch=1 -subdivisions=1 -# Training -# batch=64 -# subdivisions=8 -height=416 -width=416 -channels=3 -momentum=0.9 -decay=0.0005 -angle=0 -saturation = 1.5 -exposure = 1.5 -hue=.1 - -learning_rate=0.001 -burn_in=1000 -max_batches = 80200 -policy=steps -steps=40000,60000 -scales=.1,.1 - -[convolutional] -batch_normalize=1 -filters=32 -size=3 -stride=1 -pad=1 -activation=leaky - -[maxpool] -size=2 -stride=2 - -[convolutional] -batch_normalize=1 -filters=64 -size=3 -stride=1 -pad=1 -activation=leaky - -[maxpool] -size=2 -stride=2 - -[convolutional] -batch_normalize=1 -filters=128 -size=3 -stride=1 -pad=1 -activation=leaky - -[convolutional] -batch_normalize=1 -filters=64 -size=1 -stride=1 -pad=1 -activation=leaky - -[convolutional] -batch_normalize=1 -filters=128 -size=3 -stride=1 -pad=1 -activation=leaky - -[maxpool] -size=2 -stride=2 - -[convolutional] -batch_normalize=1 -filters=256 -size=3 -stride=1 -pad=1 -activation=leaky - -[convolutional] -batch_normalize=1 -filters=128 -size=1 -stride=1 -pad=1 -activation=leaky - -[convolutional] -batch_normalize=1 -filters=256 -size=3 -stride=1 -pad=1 -activation=leaky - -[maxpool] -size=2 -stride=2 - -[convolutional] -batch_normalize=1 -filters=512 -size=3 -stride=1 -pad=1 -activation=leaky - -[convolutional] -batch_normalize=1 -filters=256 -size=1 -stride=1 -pad=1 -activation=leaky - -[convolutional] -batch_normalize=1 -filters=512 -size=3 -stride=1 -pad=1 -activation=leaky - -[convolutional] -batch_normalize=1 -filters=256 -size=1 -stride=1 -pad=1 -activation=leaky - -[convolutional] -batch_normalize=1 -filters=512 -size=3 -stride=1 -pad=1 -activation=leaky - -[maxpool] -size=2 -stride=2 - -[convolutional] -batch_normalize=1 -filters=1024 -size=3 -stride=1 -pad=1 -activation=leaky - -[convolutional] -batch_normalize=1 -filters=512 -size=1 -stride=1 -pad=1 -activation=leaky - -[convolutional] -batch_normalize=1 -filters=1024 -size=3 -stride=1 -pad=1 -activation=leaky - -[convolutional] -batch_normalize=1 -filters=512 -size=1 -stride=1 -pad=1 -activation=leaky - -[convolutional] -batch_normalize=1 -filters=1024 -size=3 -stride=1 -pad=1 -activation=leaky - - -####### - -[convolutional] -batch_normalize=1 -size=3 -stride=1 -pad=1 -filters=1024 -activation=leaky - -[convolutional] -batch_normalize=1 -size=3 -stride=1 -pad=1 -filters=1024 -activation=leaky - -[route] -layers=-9 - -[convolutional] -batch_normalize=1 -size=1 -stride=1 -pad=1 -filters=64 -activation=leaky - -[reorg] -stride=2 - -[route] -layers=-1,-4 - -[convolutional] -batch_normalize=1 -size=3 -stride=1 -pad=1 -filters=1024 -activation=leaky - -[convolutional] -size=1 -stride=1 -pad=1 -filters=125 -activation=linear - - -[region] -anchors = 1.3221, 1.73145, 3.19275, 4.00944, 5.05587, 8.09892, 9.47112, 4.84053, 11.2364, 10.0071 -bias_match=1 -classes=20 -coords=4 -num=5 -softmax=1 -jitter=.3 -rescore=1 - -object_scale=5 -noobject_scale=1 -class_scale=1 -coord_scale=1 - -absolute=1 -thresh = .6 -random=1 diff --git a/tests/yolo_voc/yolo_voc.cpp b/tests/yolo_voc/yolo_voc.cpp deleted file mode 100644 index d0b9456..0000000 --- a/tests/yolo_voc/yolo_voc.cpp +++ /dev/null @@ -1,150 +0,0 @@ -#include -#include "tkdnn.h" - -const char *input_bin = "../tests/yolo_voc/layers/input.bin"; -const char *c0_bin = "../tests/yolo_voc/layers/c0.bin"; -const char *c2_bin = "../tests/yolo_voc/layers/c2.bin"; -const char *c4_bin = "../tests/yolo_voc/layers/c4.bin"; -const char *c5_bin = "../tests/yolo_voc/layers/c5.bin"; -const char *c6_bin = "../tests/yolo_voc/layers/c6.bin"; -const char *c8_bin = "../tests/yolo_voc/layers/c8.bin"; -const char *c9_bin = "../tests/yolo_voc/layers/c9.bin"; -const char *c10_bin = "../tests/yolo_voc/layers/c10.bin"; -const char *c12_bin = "../tests/yolo_voc/layers/c12.bin"; -const char *c13_bin = "../tests/yolo_voc/layers/c13.bin"; -const char *c14_bin = "../tests/yolo_voc/layers/c14.bin"; -const char *c15_bin = "../tests/yolo_voc/layers/c15.bin"; -const char *c16_bin = "../tests/yolo_voc/layers/c16.bin"; -const char *c18_bin = "../tests/yolo_voc/layers/c18.bin"; -const char *c19_bin = "../tests/yolo_voc/layers/c19.bin"; -const char *c20_bin = "../tests/yolo_voc/layers/c20.bin"; -const char *c21_bin = "../tests/yolo_voc/layers/c21.bin"; -const char *c22_bin = "../tests/yolo_voc/layers/c22.bin"; -const char *c23_bin = "../tests/yolo_voc/layers/c23.bin"; -const char *c24_bin = "../tests/yolo_voc/layers/c24.bin"; -const char *c26_bin = "../tests/yolo_voc/layers/c26.bin"; -const char *c29_bin = "../tests/yolo_voc/layers/c29.bin"; -const char *c30_bin = "../tests/yolo_voc/layers/c30.bin"; -const char *g31_bin = "../tests/yolo_voc/layers/g31.bin"; -const char *output_bin = "../tests/yolo_voc/layers/output.bin"; - -int main() { - - // Network layout - tk::dnn::dataDim_t dim(1, 3, 416, 416, 1); - tk::dnn::Network net(dim); - - tk::dnn::Conv2d c0 (&net, 32, 3, 3, 1, 1, 1, 1, c0_bin, true); - tk::dnn::Activation a0 (&net, tk::dnn::ACTIVATION_LEAKY); - tk::dnn::Pooling p1 (&net, 2, 2, 2, 2, tk::dnn::POOLING_MAX); - - tk::dnn::Conv2d c2 (&net, 64, 3, 3, 1, 1, 1, 1, c2_bin, true); - tk::dnn::Activation a2 (&net, tk::dnn::ACTIVATION_LEAKY); - tk::dnn::Pooling p3 (&net, 2, 2, 2, 2, tk::dnn::POOLING_MAX); - - tk::dnn::Conv2d c4 (&net, 128, 3, 3, 1, 1, 1, 1, c4_bin, true); - tk::dnn::Activation a4 (&net, tk::dnn::ACTIVATION_LEAKY); - tk::dnn::Conv2d c5 (&net, 64, 1, 1, 1, 1, 0, 0, c5_bin, true); - tk::dnn::Activation a5 (&net, tk::dnn::ACTIVATION_LEAKY); - tk::dnn::Conv2d c6 (&net, 128, 3, 3, 1, 1, 1, 1, c6_bin, true); - tk::dnn::Activation a6 (&net, tk::dnn::ACTIVATION_LEAKY); - tk::dnn::Pooling p7 (&net, 2, 2, 2, 2, tk::dnn::POOLING_MAX); - - tk::dnn::Conv2d c8 (&net, 256, 3, 3, 1, 1, 1, 1, c8_bin, true); - tk::dnn::Activation a8 (&net, tk::dnn::ACTIVATION_LEAKY); - tk::dnn::Conv2d c9 (&net, 128, 1, 1, 1, 1, 0, 0, c9_bin, true); - tk::dnn::Activation a9 (&net, tk::dnn::ACTIVATION_LEAKY); - tk::dnn::Conv2d c10(&net, 256, 3, 3, 1, 1, 1, 1, c10_bin, true); - tk::dnn::Activation a10(&net, tk::dnn::ACTIVATION_LEAKY); - tk::dnn::Pooling p11(&net, 2, 2, 2, 2, tk::dnn::POOLING_MAX); - - tk::dnn::Conv2d c12(&net, 512, 3, 3, 1, 1, 1, 1, c12_bin, true); - tk::dnn::Activation a12(&net, tk::dnn::ACTIVATION_LEAKY); - tk::dnn::Conv2d c13(&net, 256, 1, 1, 1, 1, 0, 0, c13_bin, true); - tk::dnn::Activation a13(&net, tk::dnn::ACTIVATION_LEAKY); - tk::dnn::Conv2d c14(&net, 512, 3, 3, 1, 1, 1, 1, c14_bin, true); - tk::dnn::Activation a14(&net, tk::dnn::ACTIVATION_LEAKY); - tk::dnn::Conv2d c15(&net, 256, 1, 1, 1, 1, 0, 0, c15_bin, true); - tk::dnn::Activation a15(&net, tk::dnn::ACTIVATION_LEAKY); - tk::dnn::Conv2d c16(&net, 512, 3, 3, 1, 1, 1, 1, c16_bin, true); - tk::dnn::Activation a16(&net, tk::dnn::ACTIVATION_LEAKY); - tk::dnn::Pooling p17(&net, 2, 2, 2, 2, tk::dnn::POOLING_MAX); - - tk::dnn::Conv2d c18(&net, 1024, 3, 3, 1, 1, 1, 1, c18_bin, true); - tk::dnn::Activation a18(&net, tk::dnn::ACTIVATION_LEAKY); - tk::dnn::Conv2d c19(&net, 512, 1, 1, 1, 1, 0, 0, c19_bin, true); - tk::dnn::Activation a19(&net, tk::dnn::ACTIVATION_LEAKY); - tk::dnn::Conv2d c20(&net, 1024, 3, 3, 1, 1, 1, 1, c20_bin, true); - tk::dnn::Activation a20(&net, tk::dnn::ACTIVATION_LEAKY); - tk::dnn::Conv2d c21(&net, 512, 1, 1, 1, 1, 0, 0, c21_bin, true); - tk::dnn::Activation a21(&net, tk::dnn::ACTIVATION_LEAKY); - tk::dnn::Conv2d c22(&net, 1024, 3, 3, 1, 1, 1, 1, c22_bin, true); - tk::dnn::Activation a22(&net, tk::dnn::ACTIVATION_LEAKY); - tk::dnn::Conv2d c23(&net, 1024, 3, 3, 1, 1, 1, 1, c23_bin, true); - tk::dnn::Activation a23(&net, tk::dnn::ACTIVATION_LEAKY); - tk::dnn::Conv2d c24(&net, 1024, 3, 3, 1, 1, 1, 1, c24_bin, true); - tk::dnn::Activation a24(&net, tk::dnn::ACTIVATION_LEAKY); - - tk::dnn::Layer *m25_layers[1] = { &a16 }; - tk::dnn::Route m25(&net, m25_layers, 1); - tk::dnn::Conv2d c26(&net, 64, 1, 1, 1, 1, 0, 0, c26_bin, true); - tk::dnn::Activation a26(&net, tk::dnn::ACTIVATION_LEAKY); - tk::dnn::Reorg r27(&net, 2); - - tk::dnn::Layer *m28_layers[2] = { &r27, &a24 }; - tk::dnn::Route m28(&net, m28_layers, 2); - - tk::dnn::Conv2d c29(&net, 1024, 3, 3, 1, 1, 1, 1, c29_bin, true); - tk::dnn::Activation a29(&net, tk::dnn::ACTIVATION_LEAKY); - tk::dnn::Conv2d c30(&net, 125, 1, 1, 1, 1, 0, 0, c30_bin, false); - tk::dnn::Region g31(&net, 20, 4, 5); - - tk::dnn::RegionInterpret rI(dim, g31.output_dim, 20, 4, 5, 0.6f, g31_bin); - - // Load input - dnnType *data; - dnnType *input_h; - readBinaryFile(input_bin, dim.tot(), &input_h, &data); - - //print network model - net.print(); - - //convert network to tensorRT - tk::dnn::NetworkRT netRT(&net, "yolo_voc.rt"); - - dnnType *out_data, *out_data2; // cudnn output, tensorRT output - - tk::dnn::dataDim_t dim1 = dim; //input dim - printCenteredTitle(" CUDNN inference ", '=', 30); { - dim1.print(); - TIMER_START - out_data = net.infer(dim1, data); - TIMER_STOP - dim1.print(); - } - - tk::dnn::dataDim_t dim2 = dim; - printCenteredTitle(" TENSORRT inference ", '=', 30); { - dim2.print(); - TIMER_START - out_data2 = netRT.infer(dim2, data); - TIMER_STOP - dim2.print(); - } - - printCenteredTitle(" CHECK RESULTS ", '=', 30); - dnnType *out, *out_h; - int out_dim = net.getOutputDim().tot(); - readBinaryFile(output_bin, out_dim, &out_h, &out); - std::cout<<"CUDNN vs correct"; checkResult(out_dim, out_data, out); - std::cout<<"TRT vs correct"; checkResult(out_dim, out_data2, out); - std::cout<<"CUDNN vs TRT "; checkResult(out_dim, out_data, out_data2); - - std::cout<<"\n\nDetected objects: \n"; - dnnType *output_h = new dnnType[rI.output_dim.tot()]; - checkCuda(cudaMemcpy(output_h, out_data2, - rI.output_dim.tot()*sizeof(dnnType), cudaMemcpyDeviceToHost)); - rI.interpretData(output_h); - rI.showImageResult(input_h); - return 0; -}