cuDNN examples

AI & Data Science Deep Learning (Training & Inference) cuDNN
1

Where are the code examples ?

2

This compiles and runs, but still working on the data layout, etc. Some examples in that area would be helpful.

// cudNNTest.cpp : Defines the entry point for the console application.
//
// Warning: Use at your own risk.

#include “stdafx.h”
#include “\cudnn-6.5-win-R1\cudnn-6.5-win-R1\cudnn.h”

int _tmain(int argc, _TCHAR* argv)
{
cudnnHandle_t hCudNN = NULL;
cudnnTensor4dDescriptor_t pInputDesc = NULL;
cudnnFilterDescriptor_t pFilterDesc = NULL;
cudnnConvolutionDescriptor_t pConvDesc = NULL;
cudnnTensor4dDescriptor_t pOutputDesc = NULL;
cudnnStatus_t status;
cudaError_t err;
int n_in = 64; // Number of images - originally 128
int c_in = 96; // Number of feature maps per image - originally 96
int h_in = 221; // Height of each feature map - originally 221
int w_in = 221; // Width of each feature map - originally 221
int k_pFilter_in = 256; // Number of output feature maps - originally 256
int c_pFilter_in = c_in; // Number of input feature maps - originally 96
int h_pFilter_in = 7; // Height of each pFilter - originally 7
int w_pFilter_in = 7; // Width of each pFilter - originally 7
int n_out = 0; // Number of output images.
int c_out = 0; // Number of output feature maps per image.
int h_out = 0; // Height of each output feature map.
int w_out = 0; // Width of each output feature map.

/* to change to double, chance CUDNN_DATA_FLOAT to CUDNN_DATA_DOUBLE and change each float to double below */

cudnnDataType_t dataType = CUDNN_DATA_FLOAT;
int nDataTypeSize = (((int)dataType)+1) * sizeof(float);
float* pImageInBatch = NULL;
float* pFilter = NULL;
float* pImageOutBatch = NULL;


try
{
	//---------------------------------------
	// Create CudNN 
	//---------------------------------------

	status = cudnnCreate(&hCudNN);
	if (status != CUDNN_STATUS_SUCCESS)
		throw status;


	//---------------------------------------
	// Create Descriptors
	//---------------------------------------

	status = cudnnCreateTensor4dDescriptor(&pInputDesc);
	if (status != CUDNN_STATUS_SUCCESS)
		throw status;

	status = cudnnCreateTensor4dDescriptor(&pOutputDesc);
	if (status != CUDNN_STATUS_SUCCESS)
		throw status;

	status = cudnnCreateFilterDescriptor(&pFilterDesc);
	if (status != CUDNN_STATUS_SUCCESS)
		throw status;

	status = cudnnCreateConvolutionDescriptor(&pConvDesc);
	if (status != CUDNN_STATUS_SUCCESS)
		throw status;


	//---------------------------------------
	// Allocate memory for pFilter and ImageBatch 
	//---------------------------------------

	err = cudaMalloc((void**)&pImageInBatch, n_in*c_in*h_in*w_in * nDataTypeSize);
	if (err != cudaSuccess)
		throw err;

	err = cudaMalloc((void**)&pFilter , k_pFilter_in*c_pFilter_in*h_pFilter_in*w_pFilter_in * nDataTypeSize);
	if (err != cudaSuccess)
		throw err;


	//---------------------------------------
	// Fill the input image and pFilter data
	//---------------------------------------

//TODO: Still figuring this out

	//---------------------------------------
	// Set decriptors
	//---------------------------------------

	status = cudnnSetTensor4dDescriptor(pInputDesc, CUDNN_TENSOR_NCHW, dataType, n_in, c_in, h_in, w_in);
	if (status != CUDNN_STATUS_SUCCESS)
		throw status;

	status = cudnnSetFilterDescriptor(pFilterDesc, dataType, k_pFilter_in, c_pFilter_in, h_pFilter_in, w_pFilter_in);
	if (status != CUDNN_STATUS_SUCCESS)
		throw status;

	status = cudnnSetConvolutionDescriptor(pConvDesc, pInputDesc, pFilterDesc, 0, 0, 2, 2, 1, 1, CUDNN_CONVOLUTION);
	if (status != CUDNN_STATUS_SUCCESS)
		throw status;


	//---------------------------------------
	// Query output layout
	//---------------------------------------

	status = cudnnGetOutputTensor4dDim(pConvDesc, CUDNN_CONVOLUTION_FWD, &n_out, &c_out, &h_out, &w_out);
	if (status != CUDNN_STATUS_SUCCESS)
		throw status;


	//---------------------------------------
	// Set and allocate output tensor descriptor 
	//---------------------------------------

	status = cudnnSetTensor4dDescriptor(pOutputDesc, CUDNN_TENSOR_NCHW, dataType, n_out, c_out, h_out, w_out);
	if (status != CUDNN_STATUS_SUCCESS)
		throw status;


	err = cudaMalloc((void**)&pImageOutBatch, n_out*c_out*h_out*w_out * nDataTypeSize);
	if (err != cudaSuccess)
		throw err;


	//---------------------------------------
	// Launch convolution on GPU 
	//---------------------------------------

	status = cudnnConvolutionForward(hCudNN, pInputDesc, pImageInBatch, pFilterDesc, pFilter, pConvDesc, pOutputDesc, pImageOutBatch, CUDNN_RESULT_NO_ACCUMULATE);
	if (status != CUDNN_STATUS_SUCCESS)
		throw status;


	//---------------------------------------
	// Extract output data 
	//---------------------------------------

//TBD
}
catch (…)
{
}

//---------------------------------------
// Clean-up
//---------------------------------------

if (pImageInBatch != NULL)
	cudaFree(pImageInBatch);

if (pImageOutBatch != NULL)
	cudaFree((void*)pImageOutBatch);

if (pFilter != NULL)
	cudaFree((void*)pFilter);

if (pInputDesc != NULL)
	cudnnDestroyTensor4dDescriptor(pInputDesc);

if (pOutputDesc != NULL)
	cudnnDestroyTensor4dDescriptor(pOutputDesc);

if (pFilterDesc != NULL)
	cudnnDestroyFilterDescriptor(pFilterDesc);

if (pConvDesc != NULL)
	cudnnDestroyConvolutionDescriptor(pConvDesc);

if (hCudNN != NULL)
	cudnnDestroy(hCudNN);

return 0;

}

3

So, were either of you (or anyone else seeing this topic) able to complete the above or come up with another fully working example?

4

Trying hard but failing. I can build a fully connected layer from a convolution layer. But getting convolution to work in general is difficult. I find the documentation very unclear and sparse! I am just experimenting with different parameters hoping to get the thing working.

5

Same. I’m close to finishing a simple fully connected (no convolution) network, but wow is the documentation poor to get that last 5% working. If I get it running I’ll post as helpful an explanation as I can, since I’m sure the vast majority of people trying to use this library have failed. It really is simple, like they say, but they did the classic “wrote the code and made a javadoc job well done guys.” 10 or 15 paragraphs explaining how everything fits together would have been fantastic.

6

(deleted)

7

Yes, an example of a Fully connected network, say for MNist to keep it simple would be very welcome.
Also the dimension/tensor Descriptors could be illustrated by figures in the documentation, to make people see how things work…

8

Hello,

If anyone has built a network (MNIST,CIFAR) using cudnn convolution APIs as layer implementation yet?

Thanks

9

:Following: