Skip to content
FFT based convolution using OpenCL
Branch: master
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Type Name Latest commit message Commit time
Failed to load latest commit information.
Makefile
README
conv_with_fft.c
factorize.h
kernel.cl
main.c
main.h
opencl_init.c

README

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_img*n_channel for images and n_filter*n_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
You can’t perform that action at this time.