OpenCLFFTConvolution

This simple program does FFT based convolution using OpenCL. It adapts code from CUDA and C implementations found here: CUDA implementation: https://github.com/andersbll/theano_ops/blob/master/theano_ops/abll/src/abll/conv_bc01_fft.cu#L185 C implementation: https://github.com/jeremyfix/FFTConvolution

FFT based convolution algorithm: 1. find the factor size 2. pad both image and filter to the fatcor size 3. flip the filter 4. FFT on image and filter (using batched 2D FFT, batch size is n_imgn_channel for images and n_filtern_channel for filters) Loop through n_img * n_filter (the loop can be done usint batched gemm like cublasCgemmBatched, but it is not supported in clBLAS): 5. dot product on one image and one filter 6. sum across channels for dot product 7. iFFT on the dot product sum 8. extract "same" convolution results from "full" convolution results Loop End

To compile the code, modify the Makefile accordingly, and type "make" You can run the code by "make test" Or you can run the code with your input: ./main.exe n_img n_channel img_h img_w kernel_size stride n_filter