DTolm
commented
3 years ago I haven't tried VkFFT on portable devices and I don't know much about the GPU in raspberry pi 4 (if you can provide some information similar to gpuinfo report, that can be immensely useful).
As I see it has only Vulkan 1.0 support, have you changed to it in CMakeLists (VK_API_VERSION=10)?
If you uncomment the following line: printf("%f %f - %f %f \n", output_FFTW[i + j dims[0] + l dims[0] dims[1]][0] / N, output_FFTW[i + j dims[0] + l dims[0] dims[1]][1] / N, output_VkFFT[(loc_i + loc_j dims[0] + loc_l dims[0] dims[1])][0], output_VkFFT[(loc_i + loc_j dims[0] + loc_l dims[0] dims[1])][1]); The code will print what it outputs comparing to fftw. We should start with the smallest failing system (32x1x1 in this case)
Best regards, Dmitrii
kleinl
The raspberry pi 4 got a vulkan driver some weeks ago and I wanted to try if I can get VkFFT running. The precision results of running
./Vulkan_FFT-vkfft 11` look as follows:On my other machines it is running fine but there seems to be a problem here. This might has to do with the 32-Bit architecture? Maybe I should add here that I checked already the results for correctness and the ones from fftw are correct. Best regards