Skip to content

Commit

Permalink
Add files via upload
Browse files Browse the repository at this point in the history
  • Loading branch information
@dedoogong
dedoogong authored May 1, 2018
1 parent f139a98 commit 17618a9
Show file tree
Hide file tree
Showing 14 changed files with 1,641 additions and 0 deletions.
34 changes: 34 additions & 0 deletions CMakeLists.txt
View file
Original file line number Diff line number Diff line change
@@ -0,0 +1,34 @@
cmake_minimum_required (VERSION 2.6)
project (cudnn-training)

find_package(CUDA 6.5 REQUIRED)


if (CMAKE_BUILD_TYPE STREQUAL "Debug")
message("Debug mode")
set(CUDA_NVCC_FLAGS ${CUDA_NVCC_FLAGS};-gencode;arch=compute_61,code=sm_61;-gencode;arch=compute_52,code=sm_52;-gencode;arch=compute_50,code=compute_50;-std=c++11;-g;-lineinfo;-Xcompiler;-ggdb)
else()
set(CUDA_NVCC_FLAGS ${CUDA_NVCC_FLAGS};-gencode;arch=compute_61,code=sm_61;-gencode;arch=compute_52,code=sm_52;-gencode;arch=compute_50,code=compute_50;-std=c++11;-O3;-DNDEBUG;-Xcompiler;-DNDEBUG)
endif()

set(CUDA_PROPAGATE_HOST_FLAGS OFF)

set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -std=c++11")

if(USE_GFLAGS)
add_definitions(-DUSE_GFLAGS)
endif()

include_directories($ENV{CUDNN_PATH} $ENV{CUDNN_PATH}/include)
link_directories($ENV{CUDNN_PATH} $ENV{CUDNN_PATH}/lib $ENV{CUDNN_PATH}/lib64)

cuda_add_executable(trainlenet lenet.cu readubyte.cpp)
cuda_add_cublas_to_target(trainlenet)

if(USE_GFLAGS)
target_link_libraries(trainlenet gflags cudnn)
else()
target_link_libraries(trainlenet cudnn)
endif()


17 changes: 17 additions & 0 deletions Makefile
View file
Original file line number Diff line number Diff line change
@@ -0,0 +1,17 @@
CXX := nvcc
TARGET := cudnn_conv_test
CUDNN_PATH := cudnn
HEADERS := -I $(CUDNN_PATH)/include
LIBS := -L $(CUDNN_PATH)/lib64 -L/usr/local/lib
CXXFLAGS := -gencode=arch=compute_52,code=sm_52 -gencode=arch=compute_61,code=sm_61 -std=c++11 -O2

all: conv

conv: $(TARGET).cu
$(CXX) $(CXXFLAGS) $(HEADERS) $(LIBS) $(TARGET).cu -o $(TARGET) \
-lcudnn -lopencv_imgcodecs -lopencv_imgproc -lopencv_core

.phony: clean

clean:
rm $(TARGET) || echo -n ""
Binary file added cudnn-out.png
View file
Loading
Sorry, something went wrong. Reload?
Sorry, we cannot display this file.
Sorry, this file is invalid so it cannot be displayed.
Binary file added cudnn_conv_test
View file
Binary file not shown.
226 changes: 226 additions & 0 deletions cudnn_conv_test.cu
View file
Original file line number Diff line number Diff line change
@@ -0,0 +1,226 @@
#include <cudnn.h>
#include <cassert>
#include <cstdlib>
#include <iostream>
#include <opencv2/opencv.hpp>

#define checkCUDNN(expression) \
{ \
cudnnStatus_t status = (expression); \
if (status != CUDNN_STATUS_SUCCESS) { \
std::cerr << "Error on line " << __LINE__ << ": " \
<< cudnnGetErrorString(status) << std::endl; \
std::exit(EXIT_FAILURE); \
} \
}

cv::Mat load_image(const char* image_path) {
cv::Mat image = cv::imread(image_path, CV_LOAD_IMAGE_COLOR);
image.convertTo(image, CV_32FC3);
cv::normalize(image, image, 0, 1, cv::NORM_MINMAX);
std::cerr << "Input Image: " << image.rows << " x " << image.cols << " x "
<< image.channels() << std::endl;
return image;
}

void save_image(const char* output_filename,
float* buffer,
int height,
int width) {
cv::Mat output_image(height, width, CV_32FC3, buffer);
// Make negative values zero.
cv::threshold(output_image,
output_image,
/*threshold=*/0,
/*maxval=*/0,
cv::THRESH_TOZERO);
cv::normalize(output_image, output_image, 0.0, 255.0, cv::NORM_MINMAX);
output_image.convertTo(output_image, CV_8UC3);
cv::imwrite(output_filename, output_image);
std::cerr << "Wrote output to " << output_filename << std::endl;
}

int main(int argc, const char* argv[]) {
if (argc < 2) {
std::cerr << "usage: conv <image> [gpu=0] [sigmoid=0]" << std::endl;
std::exit(EXIT_FAILURE);
}

int gpu_id = (argc > 2) ? std::atoi(argv[2]) : 0;
std::cerr << "GPU: " << gpu_id << std::endl;

bool with_sigmoid = (argc > 3) ? std::atoi(argv[3]) : 0;
std::cerr << "With sigmoid: " << std::boolalpha << with_sigmoid << std::endl;

cv::Mat image = load_image(argv[1]);

cudaSetDevice(gpu_id);

cudnnHandle_t cudnn;
cudnnCreate(&cudnn);

cudnnTensorDescriptor_t input_descriptor;
checkCUDNN(cudnnCreateTensorDescriptor(&input_descriptor));
checkCUDNN(cudnnSetTensor4dDescriptor(input_descriptor,
/*format=*/CUDNN_TENSOR_NHWC,
/*dataType=*/CUDNN_DATA_FLOAT,
/*batch_size=*/1,
/*channels=*/3,
/*image_height=*/image.rows,
/*image_width=*/image.cols));

cudnnFilterDescriptor_t kernel_descriptor;
checkCUDNN(cudnnCreateFilterDescriptor(&kernel_descriptor));
checkCUDNN(cudnnSetFilter4dDescriptor(kernel_descriptor,
/*dataType=*/CUDNN_DATA_FLOAT,
/*format=*/CUDNN_TENSOR_NCHW,
/*out_channels=*/3,
/*in_channels=*/3,
/*kernel_height=*/3,
/*kernel_width=*/3));

cudnnConvolutionDescriptor_t convolution_descriptor;
checkCUDNN(cudnnCreateConvolutionDescriptor(&convolution_descriptor));
checkCUDNN(cudnnSetConvolution2dDescriptor(convolution_descriptor,
/*pad_height=*/1,
/*pad_width=*/1,
/*vertical_stride=*/1,
/*horizontal_stride=*/1,
/*dilation_height=*/1,
/*dilation_width=*/1,
/*mode=*/CUDNN_CROSS_CORRELATION,
/*computeType=*/CUDNN_DATA_FLOAT));

int batch_size{0}, channels{0}, height{0}, width{0};
checkCUDNN(cudnnGetConvolution2dForwardOutputDim(convolution_descriptor,
input_descriptor,
kernel_descriptor,
&batch_size,
&channels,
&height,
&width));

std::cerr << "Output Image: " << height << " x " << width << " x " << channels
<< std::endl;

cudnnTensorDescriptor_t output_descriptor;
checkCUDNN(cudnnCreateTensorDescriptor(&output_descriptor));
checkCUDNN(cudnnSetTensor4dDescriptor(output_descriptor,
/*format=*/CUDNN_TENSOR_NHWC,
/*dataType=*/CUDNN_DATA_FLOAT,
/*batch_size=*/1,
/*channels=*/3,
/*image_height=*/image.rows,
/*image_width=*/image.cols));

cudnnConvolutionFwdAlgo_t convolution_algorithm;
checkCUDNN(
cudnnGetConvolutionForwardAlgorithm(cudnn,
input_descriptor,
kernel_descriptor,
convolution_descriptor,
output_descriptor,
CUDNN_CONVOLUTION_FWD_PREFER_FASTEST,
/*memoryLimitInBytes=*/0,
&convolution_algorithm));

size_t workspace_bytes{0};
checkCUDNN(cudnnGetConvolutionForwardWorkspaceSize(cudnn,
input_descriptor,
kernel_descriptor,
convolution_descriptor,
output_descriptor,
convolution_algorithm,
&workspace_bytes));
std::cerr << "Workspace size: " << (workspace_bytes / 1048576.0) << "MB"
<< std::endl;
assert(workspace_bytes > 0);

void* d_workspace{nullptr};
cudaMalloc(&d_workspace, workspace_bytes);

int image_bytes = batch_size * channels * height * width * sizeof(float);

float* d_input{nullptr};
cudaMalloc(&d_input, image_bytes);
cudaMemcpy(d_input, image.ptr<float>(0), image_bytes, cudaMemcpyHostToDevice);

float* d_output{nullptr};
cudaMalloc(&d_output, image_bytes);
cudaMemset(d_output, 0, image_bytes);

// clang-format off
const float kernel_template[3][3] = {
{1, 1, 1},
{1, -8, 1},
{1, 1, 1}
};
// clang-format on

float h_kernel[3][3][3][3];
for (int kernel = 0; kernel < 3; ++kernel) {
for (int channel = 0; channel < 3; ++channel) {
for (int row = 0; row < 3; ++row) {
for (int column = 0; column < 3; ++column) {
h_kernel[kernel][channel][row][column] = kernel_template[row][column];
}
}
}
}

float* d_kernel{nullptr};
cudaMalloc(&d_kernel, sizeof(h_kernel));
cudaMemcpy(d_kernel, h_kernel, sizeof(h_kernel), cudaMemcpyHostToDevice);

const float alpha = 1.0f, beta = 0.0f;

checkCUDNN(cudnnConvolutionForward(cudnn,
&alpha,
input_descriptor,
d_input,
kernel_descriptor,
d_kernel,
convolution_descriptor,
convolution_algorithm,
d_workspace,
workspace_bytes,
&beta,
output_descriptor,
d_output));

if (with_sigmoid) {
cudnnActivationDescriptor_t activation_descriptor;
checkCUDNN(cudnnCreateActivationDescriptor(&activation_descriptor));
checkCUDNN(cudnnSetActivationDescriptor(activation_descriptor,
CUDNN_ACTIVATION_SIGMOID,
CUDNN_PROPAGATE_NAN,
/*relu_coef=*/0));
checkCUDNN(cudnnActivationForward(cudnn,
activation_descriptor,
&alpha,
output_descriptor,
d_output,
&beta,
output_descriptor,
d_output));
cudnnDestroyActivationDescriptor(activation_descriptor);
}

float* h_output = new float[image_bytes];
cudaMemcpy(h_output, d_output, image_bytes, cudaMemcpyDeviceToHost);

save_image("cudnn-out.png", h_output, height, width);

delete[] h_output;
cudaFree(d_kernel);
cudaFree(d_input);
cudaFree(d_output);
cudaFree(d_workspace);

cudnnDestroyTensorDescriptor(input_descriptor);
cudnnDestroyTensorDescriptor(output_descriptor);
cudnnDestroyFilterDescriptor(kernel_descriptor);
cudnnDestroyConvolutionDescriptor(convolution_descriptor);

cudnnDestroy(cudnn);
}
Loading

0 comments on commit 17618a9

Please sign in to comment.