program to time cufft for various sizes

testfft.cpp

#include <stdio.h>
#include <assert.h>

#include <vector>
#include <chrono>
#include <string>

#include <cufftXt.h>

int main(int argc, char **argv)
{
    std::vector<int> vec_nfft;
    for (int k=1; k < argc; k++) {
        int val = std::stoi(argv[k]);
        vec_nfft.push_back(val);
    }

    int niter = 10;

    size_t nbytes = 512*1048576;
    cuComplex *d_data;
    cudaMalloc((void **)&d_data, nbytes);
    cudaMemset(d_data, 0, nbytes);

    for (auto&& nfft : vec_nfft)
    {
        // compute batch size that uses up all of our buffer
        int nbatch = nbytes / (sizeof(cuComplex) * nfft);

        size_t workSize;
        cufftHandle plan_fft;
        cufftCreate(&plan_fft);
        cufftSetAutoAllocation(plan_fft, false);
        cufftMakePlanMany(plan_fft, 1, &nfft,
                    0, 0, 0, 0, 0, 0,
                    CUFFT_C2C, nbatch, &workSize);

        void *d_fftwork = nullptr;
        cudaMalloc(&d_fftwork, workSize);
        assert(d_fftwork != nullptr);
        cufftSetWorkArea(plan_fft, d_fftwork);

        cudaDeviceSynchronize();
        auto time_start = std::chrono::system_clock::now();

        for (int idx=0; idx < niter; idx++) {
            cufftExecC2C(plan_fft, d_data, d_data, CUFFT_FORWARD);
        }

        cudaDeviceSynchronize();
        auto time_stop = std::chrono::system_clock::now();
        std::chrono::duration<double> elapsed = time_stop - time_start;
        printf("%d %f usec\n", nfft, elapsed.count() / nbatch / niter * 1e6);

        cufftDestroy(plan_fft);
        cudaFree(d_fftwork);
    }

    cudaFree(d_data);
}