C++/CUDA GPU-accelerated convolution in 2D and 3D.

Based on NVIDIA cuda-samples convolutionFFT2D combined with matlab mexGPUexample.m.

• https://github.com/NVIDIA/cuda-samples/tree/master/Samples/5_Domain_Specific/convolutionFFT2D
• https://uk.mathworks.com/help/parallel-computing/run-mex-functions-containing-cuda-code.html

I provide compiled .mexw64 files from a Windows 10 and compiled .mexa64 files from unix, which should run out of the box.
If this doesn't work for you due to different machine, a new mex compilation will be attempted and the NVIDIA CUDA toolbox - including an nvcc compiler, supported C++ compiler, and library cuFFT - must be installed.

Run functions CUDAconvolution(data, kernel) or CUDAconvolution3D(data, kernel) analogous to matlab conv2, convn.

The method is convolution by FFT, pointwise multiply, and inverse FFT.

This method is much faster in the case of medium to large kernels; outperforms matlab starting at kernel size ~12 x 12 x 12 and speedup is more than 1000x at convolution 900x900x200 with 100x100x100 kernel (test3d.mlx). Execution time should be constant and is <1s on my machine up to GPU memory limit.

Data should be the bigger array, as an array cut to original dimensions of data is returned.