mmr/tkDNN
Archived
1
0
Fork 0
Code Issues 46 Pull Requests 12 Releases 4 Activity

Merge branch 'cnet' of https://github.com/ceccocats/tkDNN into cnet

Browse Source
This commit is contained in:
xavier
2020-03-12 20:18:11 +01:00
parent 5c347b1e11 ef1ebea5e9
commit a65c8aa91e
15 changed files with 1816 additions and 65 deletions
Download Patch File Download Diff File Expand all files Collapse all files
include/tkDNN/Layer.h
+2 -1
View File
@@ -321,12 +321,13 @@ public:
int winH, winW;
int strideH, strideW;
int paddingH, paddingW;
bool maxpoolfixedsize;
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);
tkdnnPoolingMode_t pool_mode = POOLING_MAX, bool final = false, bool test=false);
virtual ~Pooling();
virtual layerType_t getLayerType() { return LAYER_POOLING; };
include/tkDNN/NetworkRT.h
+1
View File
@@ -36,6 +36,7 @@ using namespace nvinfer1;
#include "pluginsRT/DeformableConvRT.h"
#include "pluginsRT/FlattenConcatRT.h"
#include "pluginsRT/ReshapeRT.h"
#include "pluginsRT/MaxPoolingFixedSizeRT.h"
class PluginFactory : IPluginFactory
{
include/tkDNN/Yolo3Detection.h
+1 -1
View File
@@ -37,7 +37,7 @@ class Yolo3Detection {
int classes = 0;
int num = 0;
int n_masks = 0;
float thresh = 0.05;
float thresh = 0.3;
cv::Scalar colors[256];
// this is filled with results
include/tkDNN/kernels.h
+25 -25
View File
@@ -3,33 +3,34 @@
#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 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 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 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,
@@ -39,13 +40,12 @@ void float2half(float* srcData, __half* dstData, int size, const cudaStream_t st
// const int dilation_h, const int dilation_w,
// const int deformable_group, float *data_col);
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 deformable_group, float *data_col);
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);
void dcn_v2_cuda_forward(cublasStatus_t stat, cublasHandle_t handle,
void dcn_v2_cuda_forward(cublasStatus_t stat, cublasHandle_t handle,
float *input, float *weight,
float *bias, float *ones,
float *offset, float *mask,
@@ -54,7 +54,7 @@ void dcn_v2_cuda_forward(cublasStatus_t stat, cublasHandle_t handle,
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 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));
include/tkDNN/pluginsRT/MaxPoolingFixedSizeRT.h
+74
View File
@@ -0,0 +1,74 @@
#include<cassert>
#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<<this->n<<" "<<this->c<<" "<<this->h<<" "<<this->w<<" "<<this->stride_H<<" "<<this->stride_W<<" "<<this->winSize<<" "<<this->padding<<std::endl;
dnnType *srcData = (dnnType*)reinterpret_cast<const dnnType*>(inputs[0]);
dnnType *dstData = reinterpret_cast<dnnType*>(outputs[0]);
MaxPoolingForward(srcData, dstData, this->n, this->c, this->h, this->w, this->stride_H, this->stride_W, this->winSize, this->padding);
return 0;
}
virtual size_t getSerializationSize() override {
return 8*sizeof(int);
}
virtual void serialize(void* buffer) override {
char *buf = reinterpret_cast<char*>(buffer);
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);
}
int n, c, h, w;
int stride_H, stride_W;
int winSize;
int padding;
};
include/tkDNN/pluginsRT/ShortcutRT.h
+11 -3
View File
@@ -4,7 +4,10 @@
class ShortcutRT : public IPlugin {
public:
ShortcutRT() {
ShortcutRT(tk::dnn::dataDim_t bdim) {
this->bc = bdim.c;
this->bh = bdim.h;
this->bw = bdim.w;
}
~ShortcutRT(){
@@ -44,22 +47,27 @@ public:
dnnType *dstData = reinterpret_cast<dnnType*>(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);
shortcutForward(srcDataBack, dstData, batchSize, c, h, w, 1, batchSize, 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<char*>(buffer);
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);
}
int c, h, w;
int bc, bh, bw;
};