#include <stdio.h>
#include <stdlib.h>
#include <unistd.h>

#include <cuComplex.h>
#include <cufft.h>


int main()
{
  cufftHandle plan;

  #ifdef FAIL
  // fails
  int n = 1001472;
  #else
  // works
  int n = 1040384;
  #endif

  cuFloatComplex *in, *out;
  int loop=5;
  int i;

  cudaMallocManaged(&in, n*sizeof(cuFloatComplex));
  cudaMallocManaged(&out, n*sizeof(cuFloatComplex));

  cufftPlan1d(&plan, n, CUFFT_C2C, 1);
  if (plan== 0)
  {
    fprintf(stderr, "failed to create FFT plan.\n");
    return 1;
  }
  //cufftSetStream(plan, cudaStreamPerThread);


  // DC signal
  for (i=0; i<n; i++)
  {
    in[i].x = 1.0;
    in[i].y = 0.0;
  }
  // push to GPU
  cudaDeviceSynchronize();

  
  while (loop--)
  {
    cudaMemset(out, 0, n*sizeof(cuFloatComplex));
    cudaDeviceSynchronize();

    cufftExecC2C(plan, in, out, CUFFT_FORWARD);
    cudaDeviceSynchronize();

    for (i=0; i<8; i++)
    {
      fprintf(stderr, "(%4.1f, %4.1f) ", in[i].x, in[i].y);
    }
    fprintf(stderr, "\n");
    for (i=0; i<8; i++)
    {
      fprintf(stderr, "(%4.1f, %4.1f) ", out[i].x, out[i].y);
    }
    fprintf(stderr, "\n");
    fprintf(stderr, "\n");
   
    usleep(1000000);
  }

  cufftDestroy(plan);

  return 0;
}