cufft concurrent streams

i found the following code on stackoverflow as an example how to run cuffts in different streams concurrent on the gpu.

#include <stdio.h>
#include <iostream>
#include <cufft.h>

#define NUM_STREAMS 3

#define gpuErrchk(ans) { gpuAssert((ans), __FILE__, __LINE__); }
inline void gpuAssert(cudaError_t code, char *file, int line, bool abort=true)
   if (code != cudaSuccess) 
      fprintf(stderr,"GPUassert: %s %s %d\n", cudaGetErrorString(code), file, line);
      if (abort) exit(code);

/* MAIN */
int main()
    const int N = 40000;

    // --- Host input data initialization
    float2 *h_in1 = new float2[N];
    float2 *h_in2 = new float2[N];
    float2 *h_in3 = new float2[N];
    for (int i = 0; i < N; i++) {
        h_in1[i].x = 1.f;
        h_in1[i].y = 0.f;
        h_in2[i].x = 1.f;
        h_in2[i].y = 0.f;
        h_in3[i].x = 1.f;
        h_in3[i].y = 0.f;

   // --- Host output data initialization
    float2 *h_out1 = new float2[N];
    float2 *h_out2 = new float2[N];
    float2 *h_out3 = new float2[N];
    for (int i = 0; i < N; i++) {
        h_out1[i].x = 0.f;
        h_out1[i].y = 0.f;
        h_out2[i].x = 0.f;
        h_out2[i].y = 0.f;
        h_out3[i].x = 0.f;
        h_out3[i].y = 0.f;

    // --- Registers host memory as page-locked (required for asynch cudaMemcpyAsync)
    gpuErrchk(cudaHostRegister(h_in1, N*sizeof(float2), cudaHostRegisterPortable));
    gpuErrchk(cudaHostRegister(h_in2, N*sizeof(float2), cudaHostRegisterPortable));
    gpuErrchk(cudaHostRegister(h_in3, N*sizeof(float2), cudaHostRegisterPortable));
    gpuErrchk(cudaHostRegister(h_out1, N*sizeof(float2), cudaHostRegisterPortable));
    gpuErrchk(cudaHostRegister(h_out2, N*sizeof(float2), cudaHostRegisterPortable));
    gpuErrchk(cudaHostRegister(h_out3, N*sizeof(float2), cudaHostRegisterPortable));

    // --- Device input data allocation
    float2 *d_in1;          gpuErrchk(cudaMalloc((void**)&d_in1, N*sizeof(float2)));
    float2 *d_in2;          gpuErrchk(cudaMalloc((void**)&d_in2, N*sizeof(float2)));
    float2 *d_in3;          gpuErrchk(cudaMalloc((void**)&d_in3, N*sizeof(float2)));
    float2 *d_out1;         gpuErrchk(cudaMalloc((void**)&d_out1, N*sizeof(float2)));
    float2 *d_out2;         gpuErrchk(cudaMalloc((void**)&d_out2, N*sizeof(float2)));
    float2 *d_out3;         gpuErrchk(cudaMalloc((void**)&d_out3, N*sizeof(float2)));

    // --- Creates CUDA streams
    cudaStream_t streams[NUM_STREAMS];
    for (int i = 0; i < NUM_STREAMS; i++) gpuErrchk(cudaStreamCreate(&streams[i]));

    // --- Creates cuFFT plans and sets them in streams
    cufftHandle* plans = (cufftHandle*) malloc(sizeof(cufftHandle)*NUM_STREAMS);
    for (int i = 0; i < NUM_STREAMS; i++) {
        cufftPlan1d(&plans[i], N, CUFFT_C2C, 1);
        cufftSetStream(plans[i], streams[i]);

    // --- Async memcopyes and computations
    gpuErrchk(cudaMemcpyAsync(d_in1, h_in1, N*sizeof(float2), cudaMemcpyHostToDevice, streams[0]));
    gpuErrchk(cudaMemcpyAsync(d_in2, h_in2, N*sizeof(float2), cudaMemcpyHostToDevice, streams[1]));
    gpuErrchk(cudaMemcpyAsync(d_in3, h_in3, N*sizeof(float2), cudaMemcpyHostToDevice, streams[2]));
    cufftExecC2C(plans[0], (cufftComplex*)d_in1, (cufftComplex*)d_out1, CUFFT_FORWARD);
    cufftExecC2C(plans[1], (cufftComplex*)d_in2, (cufftComplex*)d_out2, CUFFT_FORWARD);
    cufftExecC2C(plans[2], (cufftComplex*)d_in3, (cufftComplex*)d_out3, CUFFT_FORWARD);
    gpuErrchk(cudaMemcpyAsync(h_out1, d_out1, N*sizeof(float2), cudaMemcpyDeviceToHost, streams[0]));
    gpuErrchk(cudaMemcpyAsync(h_out2, d_out2, N*sizeof(float2), cudaMemcpyDeviceToHost, streams[1]));
    gpuErrchk(cudaMemcpyAsync(h_out3, d_out3, N*sizeof(float2), cudaMemcpyDeviceToHost, streams[2]));

    for(int i = 0; i < NUM_STREAMS; i++)

    // --- Releases resources

    for(int i = 0; i < NUM_STREAMS; i++) gpuErrchk(cudaStreamDestroy(streams[i]));

    delete[] h_in1;
    delete[] h_in2;
    delete[] h_in3;
    delete[] h_out1;
    delete[] h_out2;
    delete[] h_out3;


    return 0;

the code works but looking at the Nvidia Visual Profiler i can see that the execution is not parallel, only the memcopy is overlapped by computing. Is there something wrong with the code? I am using a GTX 750 Ti and compile with -arch=sm_50.

There’s nothing wrong with the code. The kernels don’t overlap because a CUFFT FFT kernel of any significant size will completely fill the GPU, so overlap is not observed. Rather than try to use streams with small FFTs, it’s probably a better idea to use the CUFFT batching mechanism. For large FFTs, just as would be the case for large kernels, you won’t typically see overlap because a single kernel occupies the GPU (entirely). However, even for large FFTs, the code from stackoverflow demonstrates that overlap of copy and compute can still be obtained to optimize the processing of multiple large FFTs.

Okay thanks for the explanation, i think i understand the point.