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:
nvidia
2019-12-20 15:27:55 +01:00
parent 854a4c316a d889ed385d
commit f0ea2c027f
12 changed files with 253 additions and 187 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/Conv2d.cpp
+2 -2
View File
@@ -119,10 +119,10 @@ void Conv2d::inferCUDNN(dnnType* srcData, bool back) {
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) :
std::string fname_weights, bool batchnorm, bool deConv, bool final) :
LayerWgs(net, net->getOutputDim().c, out_ch, kernelH, kernelW, 1,
fname_weights, batchnorm) {
fname_weights, batchnorm, false, final) {
this->kernelH = kernelH;
this->kernelW = kernelW;
src/DeformConv2d.cpp
+42 -112
View File
@@ -14,9 +14,34 @@ void DeformConv2d::initCUDNN() {
net->tensorFormat, net->dataType,
1, output_dim.c, 1, 1) );
checkCUDNN( cudnnSetTensor4dDescriptor(dstTensorDesc,
checkCUDNN( cudnnSetTensor4dDescriptor(dstTensorDesc,
net->tensorFormat, net->dataType, output_dim.n, output_dim.c, output_dim.h, output_dim.w));
const int height_ones = (preconv->input_dim.h + 2 * this->paddingH - (1 * (this->kernelH - 1) + 1)) / this->strideH + 1;
const int width_ones = (preconv->input_dim.w + 2 * this->paddingW - (1 * (this->kernelW - 1) + 1)) / this->strideW + 1;
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";
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];
for(int i=0; i<height_ones*width_ones; i++)
aus1[i]=1.0f;
checkCuda( cudaMemcpy(ones_d1, aus1, (height_ones*width_ones)*sizeof(dnnType), cudaMemcpyHostToDevice) );
checkCuda( cudaMalloc(&ones_d2, dim_ones*sizeof(dnnType)) );
float aus2[dim_ones];
for(int i=0; i<dim_ones; i++)
aus2[i]=1.0f;
checkCuda( cudaMemcpy(ones_d2, aus2, (dim_ones)*sizeof(dnnType), cudaMemcpyHostToDevice) );
checkCuda( cudaDeviceSynchronize() );
}
DeformConv2d::DeformConv2d( Network *net, int out_ch, int deformable_group, int kernelH, int kernelW,
@@ -51,93 +76,25 @@ DeformConv2d::~DeformConv2d() {
checkCUDNN( cudnnDestroyTensorDescriptor(biasTensorDesc) );
checkCuda( cudaFree(dstData) );
}
void Conv2dToChunk(int dim, dnnType* srcData, dnnType* offset, dnnType* mask)
{
// std::cout<<"9\n";
// cudaMemcpyFromArray(offset, (const struct cudaArray *)srcData, 0, 2*dim.tot()/3, dim.tot()/3, cudaMemcpyDeviceToHost);
// checkCuda(cudaMemcpyFromArray(offset, (const struct cudaArray *)srcData, 0, 0, 2*dim, cudaMemcpyDeviceToDevice));
checkCuda(cudaMemcpy(offset, srcData, 2*dim*sizeof(dnnType), cudaMemcpyDeviceToDevice));
// std::cout<<"9a\n";
cudaDeviceSynchronize();
// cudaMemcpyFromArray(mask, (const struct cudaArray *)srcData, 2*dim.tot()/3, dim.tot(), dim.tot()/3, cudaMemcpyDeviceToHost);
// checkCuda(cudaMemcpyFromArray(mask, (const struct cudaArray *)srcData, 0, 2*dim, dim, cudaMemcpyDeviceToDevice));
checkCuda(cudaMemcpy(mask, srcData + 2*dim, dim*sizeof(dnnType), cudaMemcpyDeviceToDevice));
// std::cout<<"9b\n";
checkCuda( cudaFreeHost(ones_d1) );
checkCuda( cudaFreeHost(ones_d2) );
checkCuda( cudaFree(offset) );
checkCuda( cudaFree(mask) );
checkCuda( cudaFree(output_conv) );
}
dnnType* DeformConv2d::infer(dataDim_t &dim, dnnType* srcData) {
dnnType *input;
checkCuda(cudaMalloc(&input, dim.tot()*sizeof(dnnType)));
checkCuda(cudaMemcpy(input, srcData, dim.tot()*sizeof(dnnType), cudaMemcpyDeviceToDevice));
cudaDeviceSynchronize();
srcData = preconv->infer(dim, srcData);
dim = preconv->output_dim;
//split to chank
dnnType *offset, *mask;
int dst_dim = dim.tot();
if (dst_dim % 3 != 0 )
std::cout<<"take attention\n\n";
int chunk_dim = dst_dim/3;
checkCuda(cudaMalloc(&offset, 2*chunk_dim*sizeof(dnnType)));
checkCuda(cudaMalloc(&mask, chunk_dim*sizeof(dnnType)));
cudaDeviceSynchronize();
Conv2dToChunk(chunk_dim, srcData, offset, mask);
// conv2d
output_conv = preconv->infer(dim, srcData);
// 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
dnnType *vec;
vec = new dnnType[chunk_dim];
cudaDeviceSynchronize();
cudaMemcpy(vec, mask, chunk_dim*sizeof(dnnType), cudaMemcpyDeviceToHost);
cudaDeviceSynchronize();
for(int i=0; i<chunk_dim; i++){
// std::cout<<i<<" -- "<<vec[i]<<", ";
vec[i] = 1.0f / (1.0f + exp(-vec[i]));
// std::cout<<vec[i]<<std::endl;
}
cudaDeviceSynchronize();
cudaMemcpy(mask, vec, chunk_dim*sizeof(dnnType), cudaMemcpyHostToDevice);
cudaDeviceSynchronize();
free(vec);
// dnnType *tmp;
// cudaMallocHost(&tmp, chunk_dim*sizeof(dnnType));
// cudaMemcpy(tmp, mask, chunk_dim*sizeof(dnnType), cudaMemcpyDeviceToHost);
// std::cout<<"conv2d output before chunking"<<std::endl;
// for (size_t i = 0; i < chunk_dim; i++)
// {
// std::cout<<i<<" -- "<<tmp[i]<<", ";
// }
// std::cout<<std::endl;
// cudaFreeHost(tmp);
activationSIGMOIDForward(mask, mask, chunk_dim);
const int height_ones = (preconv->input_dim.h + 2 * this->paddingH - (1 * (this->kernelH - 1) + 1)) / this->strideH + 1;
const int width_ones = (preconv->input_dim.w + 2 * this->paddingW - (1 * (this->kernelW - 1) + 1)) / this->strideW + 1;
const int dim_ones = preconv->input_dim.c * this->kernelH * this->kernelW * 1 * height_ones * width_ones;
// kernel ones
dnnType *ones_d1;
cudaMallocHost(&ones_d1, (height_ones*width_ones)*sizeof(dnnType));
float aus1[height_ones*width_ones];
for(int i=0; i<height_ones*width_ones; i++)
aus1[i]=1.0f;
cudaMemcpy(ones_d1, aus1, (height_ones*width_ones)*sizeof(dnnType), cudaMemcpyHostToDevice);
cudaDeviceSynchronize();
dnnType *ones_d2;
cudaMallocHost(&ones_d2, dim_ones*sizeof(dnnType));
float aus2[dim_ones];
for(int i=0; i<dim_ones; i++)
aus2[i]=1.0f;
cudaMemcpy(ones_d2, aus2, (dim_ones)*sizeof(dnnType), cudaMemcpyHostToDevice);
cudaDeviceSynchronize();
dcn_v2_cuda_forward(input, this->data_d,
// deformable convolution
dcn_v2_cuda_forward(srcData, this->data_d,
this->bias2_d, ones_d1,
offset, mask,
dstData, ones_d2,
@@ -148,44 +105,18 @@ dnnType* DeformConv2d::infer(dataDim_t &dim, dnnType* srcData) {
this->deformableGroup,
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,
dst_dim);
cudaFree(offset);
cudaFree(mask);
cudaFree(input);
cudaFreeHost(ones_d1);
cudaFreeHost(ones_d2);
// dnnType *aus3;
// cudaMallocHost(&aus3, 256*7*7*sizeof(dnnType));
// cudaMemcpy(aus3, dstData, (256*7*7)*sizeof(dnnType), cudaMemcpyDeviceToHost);
// checkCuda(cudaDeviceSynchronize());
// std::cout<<"OutDim:\n";
// this->output_dim.print();
// std::cout<<"\n\n\nprint dstData: \n";
// for (int i = 0 ; i < 256*7*7; i++){
// if(i==294)
// std::cout<<"\n\n\n";
// std::cout<<aus3[i]<<" ";
// }
// std::cout<<"\n";
// cudaFreeHost(aus3);
std::cout<<"srcData BN:\n";
printDeviceVector(64, dstData);
chunk_dim);
dnnType alpha = dnnType(1);
dnnType beta = dnnType(0);
if(!batchnorm) {
// // // bias
// bias
alpha = dnnType(1);
beta = dnnType(1);
checkCUDNN( cudnnAddTensor(net->cudnnHandle,
&alpha, biasTensorDesc, bias_d,
&beta, dstTensorDesc, dstData) );
} else {
std::cout<<"LOL\n";
alpha = dnnType(1);
beta = dnnType(0);
checkCUDNN( cudnnBatchNormalizationForwardInference(net->cudnnHandle,
@@ -195,9 +126,8 @@ dnnType* DeformConv2d::infer(dataDim_t &dim, dnnType* srcData) {
scales_d, bias_d, mean_d, variance_d,
CUDNN_BN_MIN_EPSILON) );
}
//update data dimensions
std::cout<<"dstData BN:\n";
printDeviceVector(64, dstData);
dim = output_dim;
return dstData;
}
src/Layer.cpp
+2 -2
View File
@@ -4,10 +4,10 @@
namespace tk { namespace dnn {
Layer::Layer(Network *net) {
Layer::Layer(Network *net, bool final) {
this->net = net;
this->final = final;
if(net != nullptr) {
this->input_dim = net->getOutputDim();
this->output_dim = input_dim;
src/LayerWgs.cpp
+2 -2
View File
@@ -8,12 +8,12 @@ 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) : Layer(net) {
std::string fname_weights, bool batchnorm, bool additional_bias, bool final) : Layer(net, final) {
this->inputs = inputs;
this->outputs = outputs;
this->weights_path = std::string(fname_weights);
std::cout<<"Reading weights: I="<<inputs<<" O="<<outputs<<" KERNEL="<<kh<<"x"<<kw<<"x"<<kl<<"\n";
int seek = 0;
readBinaryFile(weights_path.c_str(), inputs*outputs*kh*kw*kl, &data_h, &data_d, seek, net->dontLoadWeights);
src/NetworkRT.cpp
+52 -1
View File
@@ -75,7 +75,7 @@ NetworkRT::NetworkRT(Network *net, const char *name) {
input = Ilay->getOutput(0);
input->setName( (l->getLayerName() + std::to_string(i) + "_out").c_str() );
if(l->getLayerType() == LAYER_YOLO)
if(l->getLayerType() == LAYER_YOLO || l->final)
networkRT->markOutput(*input);
tensors[l] = input;
}
@@ -182,6 +182,8 @@ ILayer* NetworkRT::convert_layer(ITensor *input, Layer *l) {
return convert_layer(input, (Yolo*) l);
if(type == LAYER_UPSAMPLE)
return convert_layer(input, (Upsample*) l);
if(type == LAYER_DEFORMCONV2D)
return convert_layer(input, (DeformConv2d*) l);
std::cout<<l->getLayerName()<<"\n";
FatalError("Layer not implemented in tensorRT");
@@ -254,6 +256,8 @@ ILayer* NetworkRT::convert_layer(ITensor *input, Conv2d *l) {
lRTconv->setPadding(DimsHW{l->paddingH, l->paddingW});
lRT = (ILayer*) lRTconv;
Dims d = lRTconv->getOutput(0)->getDimensions();
std::cout<<"DECONV: "<<d.d[0]<<" "<<d.d[1]<<" "<<d.d[2]<<" "<<d.d[3]<<"\n";
}
checkNULL(lRT);
@@ -421,6 +425,53 @@ ILayer* NetworkRT::convert_layer(ITensor *input, Upsample *l) {
return lRT;
}
ILayer* NetworkRT::convert_layer(ITensor *input, DeformConv2d *l) {
//std::cout<<"convert DEFORMABLE\n";
ILayer *preconv = convert_layer(input, l->preconv);
ITensor **inputs = new ITensor*[2];
inputs[0] = input;
inputs[1] = preconv->getOutput(0);
//std::cout<<"New plugin DEFORMABLE\n";
IPlugin *plugin = new DeformableConvRT(l);
IPluginLayer *lRT = networkRT->addPlugin(&input, 1, *plugin);
checkNULL(lRT);
// batchnorm
void *bias_b, *power_b, *mean_b, *variance_b, *scales_b;
if(dtRT == DataType::kHALF) {
bias_b = l->bias16_h;
power_b = l->power16_h;
mean_b = l->mean16_h;
variance_b = l->variance16_h;
scales_b = l->scales16_h;
} else {
bias_b = l->bias_h;
power_b = l->power_h;
mean_b = l->mean_h;
variance_b = l->variance_h;
scales_b = l->scales_h;
}
Weights power{dtRT, power_b, l->outputs};
Weights shift{dtRT, mean_b, l->outputs};
Weights scale{dtRT, variance_b, l->outputs};
std::cout<<lRT->getNbOutputs()<<std::endl;
IScaleLayer *lRT2 = networkRT->addScale(*lRT->getOutput(0), ScaleMode::kCHANNEL,
shift, scale, power);
checkNULL(lRT2);
Weights shift2{dtRT, bias_b, l->outputs};
Weights scale2{dtRT, scales_b, l->outputs};
IScaleLayer *lRT3 = networkRT->addScale(*lRT2->getOutput(0), ScaleMode::kCHANNEL,
shift2, scale2, power);
checkNULL(lRT3);
return lRT3;
}
bool NetworkRT::serialize(const char *filename) {
std::ofstream p(filename);
src/kernels/activation_sigmoid.cu
+31
View File
@@ -0,0 +1,31 @@
#include "kernels.h"
__device__
__forceinline__
double sigmoid (double a)
{
return 1.0 / (1.0 + exp (-a));
}
__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]);
}
}
/**
ELU activation function
*/
void activationSIGMOIDForward(dnnType* srcData, dnnType* dstData, int size, cudaStream_t stream)
{
int blocks = (size+255)/256;
int threads = 256;
activation_sigmoid<<<blocks, threads, 0, stream>>>(srcData, dstData, size);
}
src/kernels/deformable_conv.cu
+16 -51
View File
@@ -149,10 +149,8 @@ void dcn_v2_cuda_forward(float *input, float *weight,
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)
{
checkCuda(cudaDeviceSynchronize());
cudaError_t cudaStat;
const int chunk_dim, cudaStream_t stream)
{
cublasStatus_t stat;
cublasHandle_t handle;
stat = cublasCreate(&handle);
@@ -160,83 +158,50 @@ void dcn_v2_cuda_forward(float *input, float *weight,
printf ("CUBLAS initialization failed\n");
return;
}
checkCuda(cudaDeviceSynchronize());
const int batch = in_n;
const int channels = in_c;
const int height = in_h;
const int width = in_w;
const int channels_out = out_c;
const int channels_kernel = in_c;
const int kernel_h_ = kernel_h;
const int kernel_w_ = kernel_w;
const int height_out = (height + 2 * pad_h - (dilation_h * (kernel_h - 1) + 1)) / stride_h + 1;
const int width_out = (width + 2 * pad_w - (dilation_w * (kernel_w - 1) + 1)) / stride_w + 1;
float *input_n;
cudaMalloc(&input_n, (in_n*in_c*in_h*in_w)*sizeof(float));
cudaMemcpy(input_n, input, (in_n*in_c*in_h*in_w)*sizeof(float), cudaMemcpyDeviceToDevice);
checkCuda(cudaDeviceSynchronize());
float *offset_n;
cudaMalloc(&offset_n, ((dst_dim/3)*2)*sizeof(float));
cudaMemcpy(offset_n, offset, ((dst_dim/3)*2)*sizeof(float), cudaMemcpyDeviceToDevice);
checkCuda(cudaDeviceSynchronize());
float *mask_n;
cudaMalloc(&mask_n, (dst_dim/3)*sizeof(float));
cudaMemcpy(mask_n, mask, (dst_dim/3)*sizeof(float), cudaMemcpyDeviceToDevice);
checkCuda(cudaDeviceSynchronize());
float *output_n;
checkCuda(cudaMalloc(&output_n, (channels_out*height_out*width_out)*sizeof(float)));
checkCuda(cudaDeviceSynchronize());
long m_ = channels_out;
long n_ = height_out * width_out;
long k_ = 1;
long m = channels_out;
long n = height_out * width_out;
long k = 1;
float alpha = 1.0;
float beta = 0.0;
checkCuda(cudaDeviceSynchronize());
stat = cublasSgemm(handle, CUBLAS_OP_T, CUBLAS_OP_N,
n_, m_, k_, &alpha,
ones, k_, bias, k_,
&beta, output_n, n_);
n, m, k, &alpha,
ones, k, bias, k,
&beta, output, n);
if (stat != CUBLAS_STATUS_SUCCESS) {
printf ("CUBLAS initialization failed\n");
return ;
}
checkCuda(cudaDeviceSynchronize());
modulated_deformable_im2col_cuda(stream,
input_n, offset_n,
mask_n,
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);
checkCuda(cudaDeviceSynchronize());
//(k * m) x (m * n)
// Y = WC
long m = channels_out;
long n = height_out * width_out;
long k = channels * kernel_h * kernel_w;
alpha = 1.0;
k = channels * kernel_h * kernel_w;
beta = 1.0;
stat = cublasSgemm(handle, CUBLAS_OP_N, CUBLAS_OP_N,
n, m, k, &alpha,
columns, n, weight, k,
&beta, output_n, n);
&beta, output, n);
cudaMemcpy(output, output_n, (n*m)*sizeof(float), cudaMemcpyDeviceToDevice);
checkCuda(cudaDeviceSynchronize());
if (stat != CUBLAS_STATUS_SUCCESS) {
printf ("CUBLAS initialization failed\n");
return ;