Converting ConvolutionFFT2D to ConvolutionFFT3D

Accelerated Computing CUDA CUDA Programming and Performance
1

Hello, I want to convert the example code ConvolutionFFT2D to ConvolutionFFT3D, i.e. perform 3D FFT convolution in CUDA. This is the first time I program in CUDA.

Most of the code is straight forward to change to 3D from 2D, but I got some problems.

I’m a bit confused about the memory allocation, why is the memory for a_Kernel allocated with cudaMallocArray and d_PaddedKernel with cudaMalloc?

Is there a way to use cudaMallocArray with 3 dimensions instead of 2? I tried it but the compiler complained.

Why do I need to bind a texture to the memory allocated with cudaMallocArray? Because the texture memory is faster?

I changed tex2D to tex3D in the kernel-code but how do I bind to a 3D texture if I cannot use cudaMallocArray for 3D ?

Is the textures only used for the padding of the data, not for the actual FFT calculations?

printf("Allocating memory...\n");

		h_Kernel	   = (Complex *)malloc(KERNEL_SIZE);

		h_Data		 = (Complex *)malloc(DATA_SIZE);

		h_ResultCPU	= (Complex *)malloc(DATA_SIZE);

		h_ResultGPU	= (Complex *)malloc(FFT_SIZE);

		//cutilSafeCall( cudaMallocArray(&a_Kernel, &float2tex, KERNEL_W, KERNEL_H) );

		//cutilSafeCall( cudaMallocArray(&a_Data,   &float2tex,   DATA_W,   DATA_H) );

		cudaMalloc((void**)&a_Kernel, sizeof(cufftComplex)*KERNEL_W * KERNEL_H * KERNEL_D);

		cudaMalloc((void**)&a_Data, sizeof(cufftComplex)*DATA_W * DATA_H * DATA_D);

		cutilSafeCall( cudaMalloc((void **)&d_PaddedKernel, FFT_SIZE) );

		cutilSafeCall( cudaMalloc((void **)&d_PaddedData,   FFT_SIZE) );

	   

		printf("...copying input data and convolution kernel from host to CUDA arrays\n");

		cutilSafeCall( cudaMemcpyToArray(a_Kernel, 0, 0, h_Kernel, KERNEL_SIZE, cudaMemcpyHostToDevice) );

		cutilSafeCall( cudaMemcpyToArray(a_Data,   0, 0, h_Data,   DATA_SIZE,   cudaMemcpyHostToDevice) );

		printf("...binding CUDA arrays to texture references\n");

		cutilSafeCall( cudaBindTextureToArray(texKernel, a_Kernel) );

		cutilSafeCall( cudaBindTextureToArray(texData,   a_Data)   );