Merge with master, works with Jetpack 4.3

Signed-off-by: xavier <micaelaverucchi@gmail.com>

CMakeLists.txt

@@ -18,17 +18,11 @@ find_package(CUDA 9.0 REQUIRED)
 SET(CUDA_SEPARABLE_COMPILATION ON)
 #set(CUDA_NVCC_FLAGS "${CUDA_NVCC_FLAGS} -arch=sm_30 --compiler-options '-fPIC'")
 
-# compile Discovery only if TensorRT is installed
+find_package(CUDNN REQUIRED)
-find_library(NVINFER NAMES nvinfer)
-if(NVINFER STREQUAL "NVINFER-NOTFOUND")
-    set(NVINFER_INCLUDES "/usr/local/nvidia/tensorrt/include/")
-    link_directories(/usr/local/nvidia/tensorrt/targets/x86_64-linux-gnu/lib/ 
-		     /usr/local/cuda/targets/x86_64-linux/lib/)
-endif()
 
 # compile
 file(GLOB tkdnn_CUSRC "src/kernels/*.cu")
-cuda_include_directories(${CMAKE_CURRENT_SOURCE_DIR}/include ${CUDA_INCLUDE_DIRS} ${NVINFER_INCLUDES})
+cuda_include_directories(${CMAKE_CURRENT_SOURCE_DIR}/include ${CUDA_INCLUDE_DIRS} ${CUDNN_INCLUDE_DIRS})
 cuda_add_library(kernels SHARED ${tkdnn_CUSRC})
 
 
@@ -43,7 +37,7 @@ set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -DOPENCV")
 # Build Libraries
 #-------------------------------------------------------------------------------
 file(GLOB tkdnn_SRC "src/*.cpp")
-set(tkdnn_LIBS kernels ${CUDA_LIBRARIES} ${CUDA_CUBLAS_LIBRARIES} -lcudnn -lnvinfer ${OpenCV_LIBS})
+set(tkdnn_LIBS kernels ${CUDA_LIBRARIES} ${CUDA_CUBLAS_LIBRARIES} ${CUDNN_LIBRARIES} ${OpenCV_LIBS})
 
 set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wall -std=c++11")
 include_directories(${CMAKE_CURRENT_SOURCE_DIR}/include ${CUDA_INCLUDE_DIRS} ${OPENCV_INCLUDE_DIRS} ${NVINFER_INCLUDES})

README.md

@@ -3,9 +3,10 @@ tkDNN is a Deep Neural Network library built with cuDNN primitives specifically
 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 9
+* CUDA 10.0
-* CUDNN 7.105
+* CUDNN 7.603
-* TENSORRT 4.02
+* TENSORRT 6.01
+* OPENCV 4.1
 
 ## Workflow
 The recommended workflow follow these step:
@@ -48,5 +49,3 @@ this will genereate a yolo3_berkeley.rt file that can be used for live detection
 ./yolo3_demo                               # launch detection on a demo video
 ./yolo3_demo yolo3_berkeley.rt /dev/video0 # launch detection on device 0
 ```
-
-

cmake/FindCUDNN.cmake

@@ -0,0 +1,33 @@
+# Find the header files
+
+find_path(CUDNN_INCLUDE_DIR
+        ${CMAKE_SYSROOT}/usr/local/include
+        ${CMAKE_SYSROOT}/usr/include
+        /usr/local/nvidia/tensorrt/include/
+        NO_DEFAULT_PATH
+)
+
+set(OLD_ROOT ${CMAKE_FIND_ROOT_PATH})
+list(APPEND CMAKE_FIND_ROOT_PATH /)
+list(APPEND CMAKE_FIND_LIBRARY_SUFFIXES .so.7)
+list(APPEND CMAKE_FIND_LIBRARY_SUFFIXES .so.5)
+find_library(CUDNN_LIB
+    NAMES cudnn
+    PATHS
+    /usr/local/driveworks/targets/${CMAKE_SYSTEM_PROCESSOR}-Linux/lib
+    /usr/lib/${CMAKE_SYSTEM_PROCESSOR}-linux-gnu/
+    NO_DEFAULT_PATH
+)
+find_library(CUDNN_NVLIB
+    NAMES "nvinfer"
+    PATHS
+    /usr/local/driveworks/targets/${CMAKE_SYSTEM_PROCESSOR}-Linux/lib
+    /usr/lib/${CMAKE_SYSTEM_PROCESSOR}-linux-gnu/
+    NO_DEFAULT_PATH
+)
+set(CMAKE_FIND_ROOT_PATH ${OLD_ROOT})
+
+set(CUDNN_LIBRARIES ${CUDNN_LIB} ${CUDNN_NVLIB})
+message("-- Found CUDNN: "  ${CUDNN_LIB})
+message("-- Found NVINFER: "  ${CUDNN_NVLIB})
+set(CUDNN_FOUND true)

cmake/tkDNNConfig.cmake

@@ -4,27 +4,21 @@ set(CMAKE_CXX_FLAGS  "${CMAKE_CXX_FLAGS} --std=c++11 -fPIC")
 
 find_package(CUDA REQUIRED)
 find_package(OpenCV REQUIRED)
-find_library(NVINFER NAMES nvinfer)
+find_package(CUDNN REQUIRED)
-if(NVINFER STREQUAL "NVINFER-NOTFOUND")
-    set(NVINFER_INCLUDES "/usr/local/nvidia/tensorrt/include/")
-    link_directories(/usr/local/nvidia/tensorrt/targets/x86_64-linux-gnu/lib/ 
-		     /usr/local/cuda/targets/x86_64-linux/lib/)
-endif()
 
 set(tkDNN_INCLUDE_DIRS 
 	${CUDA_INCLUDE_DIRS} 
 	${OPENCV_INCLUDE_DIRS} 
-    ${NVINFER_INCLUDES}
+    ${CUDNN_INCLUDE_DIRS}
 )
 
 set(tkDNN_LIBRARIES 
     tkDNN 
     kernels 
     ${CUDA_LIBRARIES} 
-    ${CUDA_CUBLAS_LIBRARIES} 
+    ${CUDA_CUBLAS_LIBRARIES}
-    -lcudnn 
+	${CUDNN_LIBRARIES}
-    -lnvinfer 
+	${OpenCV_LIBS}
-    ${OpenCV_LIBS}
 )
 
 set(tkDNN_FOUND true)

demo/demo/demo.cpp

@@ -7,6 +7,7 @@
 
 #include <opencv2/core/core.hpp>
 #include <opencv2/highgui/highgui.hpp>
+#include <opencv2/videoio.hpp>
 #include <opencv2/imgproc/imgproc.hpp>
 
 #include "Yolo3Detection.h"
@@ -46,9 +47,9 @@ int main(int argc, char *argv[]) {
 
     cv::VideoWriter resultVideo;
     if(SAVE_RESULT) {
-        int w = cap.get(CV_CAP_PROP_FRAME_WIDTH);
+        int w = cap.get(cv::CAP_PROP_FRAME_WIDTH);
-        int h = cap.get(CV_CAP_PROP_FRAME_HEIGHT);
+        int h = cap.get(cv::CAP_PROP_FRAME_HEIGHT);
-        resultVideo.open("result.mp4", CV_FOURCC('M','P','4','V'), 30, cv::Size(w, h));
+        resultVideo.open("result.mp4", cv::VideoWriter::fourcc('M','P','4','V'), 30, cv::Size(w, h));
     }
 
     cv::Mat frame;

include/tkDNN/Layer.h

@@ -26,6 +26,8 @@ enum layerType_t {
     LAYER_YOLO
 };
 
+#define TKDNN_BN_MIN_EPSILON 1e-5
+
 /** 
     Simple layer Father class
 */

src/Conv2d.cpp

@@ -113,7 +113,7 @@ void Conv2d::inferCUDNN(dnnType* srcData, bool back) {
                                                 dstTensorDesc, dstData, dstTensorDesc,
                                                 dstData, biasTensorDesc, //same tensor descriptor as bias
                                                 scales_d, bias_d, mean_d, variance_d,
-                                                CUDNN_BN_MIN_EPSILON) );
+                                                TKDNN_BN_MIN_EPSILON) );
     }
 }
 

src/DeformConv2d.cpp

@@ -124,7 +124,7 @@ dnnType* DeformConv2d::infer(dataDim_t &dim, dnnType* srcData) {
                                                 dstTensorDesc, dstData, dstTensorDesc,
                                                 dstData, biasTensorDesc, //same tensor descriptor as bias
                                                 scales_d, bias_d, mean_d, variance_d,
-                                                CUDNN_BN_MIN_EPSILON) );
+                                                TKDNN_BN_MIN_EPSILON) );
     }
 
     //update data dimensions

src/LayerWgs.cpp

@@ -35,7 +35,7 @@ LayerWgs::LayerWgs(Network *net, int inputs, int outputs,
         seek += outputs;
         readBinaryFile(weights_path.c_str(), outputs, &variance_h, &variance_d, seek, net->dontLoadWeights);
 
-        float eps = CUDNN_BN_MIN_EPSILON;
+        float eps = TKDNN_BN_MIN_EPSILON;
 
         power_h = new dnnType[outputs];
         for(int i=0; i<outputs; i++) power_h[i] = 1.0f;

src/Network.cpp

@@ -18,10 +18,6 @@ Network::Network(dataDim_t input_dim) {
     dataType = CUDNN_DATA_FLOAT;
     tensorFormat = CUDNN_TENSOR_NCHW;
     dontLoadWeights = false;
-
-    checkCUDNN( cudnnCreate(&cudnnHandle) );
-    checkERROR( cublasCreate(&cublasHandle) );
-
     num_layers = 0;
 
     fp16 = false;
@@ -39,6 +35,11 @@ Network::Network(dataDim_t input_dim) {
         std::cout<<COL_REDB<<"!! FP16 INERENCE ENABLED !!"<<COL_END<<"\n";
     if(dla)
         std::cout<<COL_GREENB<<"!! DLA INERENCE ENABLED !!"<<COL_END<<"\n";
+
+
+    checkCUDNN( cudnnCreate(&cudnnHandle) );
+    checkERROR( cublasCreate(&cublasHandle) );
+
 }
 
 Network::~Network() {