问题
I write because I'm facing problems with the cufft 3D transform in-place, while I have no problems for the out-of-place version. I tried to follow Robert Crovella's answer here but I'm not obtaining the correct results when I make a FFT+IFT. This is my code:
#include <stdio.h>
#include <stdlib.h>
#include <cuda_runtime.h>
#include <complex.h>
#include <cuComplex.h>
#include <cufft.h>
// Main function
int main(int argc, char **argv){
int N = 4;
double *in = NULL, *d_in = NULL;
cuDoubleComplex *out = NULL, *d_out = NULL;
cufftHandle plan_r2c, plan_c2r;
unsigned int out_mem_size = sizeof(cuDoubleComplex) * N*N*(N/2 + 1);
unsigned int in_mem_size = out_mem_size;
in = (double *) malloc (in_mem_size);
out = (cuDoubleComplex *)in;
cudaMalloc((void **)&d_in, in_mem_size);
d_out = (cuDoubleComplex *)d_in;
cufftPlan3d(&plan_r2c, N, N, N, CUFFT_D2Z);
cufftPlan3d(&plan_c2r, N, N, N, CUFFT_Z2D);
memset(in, 0, in_mem_size);
unsigned int idx;
for (int z = 0; z < N; z++){
for (int y = 0; y < N; y++){
for (int x = 0; x < N; x++){
idx = z + N * ( y + x * N);
in[idx] = idx;
}
}
}
printf("\nStart: \n");
for (int z = 0; z < N; z++){
printf("plane = %d ----------------------------\n", z);
for (int x = 0; x < N; x++){
for (int y = 0; y < N; y++){
idx = z + N * ( y + x * N);
printf("%.3f \t", in[idx]);
}
printf("\n");
}
}
cudaMemcpy(d_in, in, in_mem_size, cudaMemcpyHostToDevice);
cufftExecD2Z(plan_r2c, (cufftDoubleReal *)d_in, (cufftDoubleComplex *)d_out);
cufftExecZ2D(plan_c2r, (cufftDoubleComplex *)d_out, (cufftDoubleReal *)d_in);
memset(in, 0, in_mem_size);
CU_ERR_CHECK( cudaMemcpy(in, d_in, in_mem_size, cudaMemcpyDeviceToHost) );
printf("\nAfter FFT+IFT: \n");
for (int z = 0; z < N; z++){
printf("plane = %d ----------------------------\n", z);
for (int x = 0; x < N; x++){
for (int y = 0; y < N; y++){
idx = z + N * ( y + x * N);
// Normalisation
in[idx] /= (N*N*N);
printf("%.3f \t", in[idx]);
}
printf("\n");
}
}
return 0;
}
The program outputs the following data:
Starting file
plane = 0 ----------------------------
0.000 4.000 8.000 12.000
16.000 20.000 24.000 28.000
32.000 36.000 40.000 44.000
48.000 52.000 56.000 60.000
plane = 1 ----------------------------
1.000 5.000 9.000 13.000
17.000 21.000 25.000 29.000
33.000 37.000 41.000 45.000
49.000 53.000 57.000 61.000
plane = 2 ----------------------------
2.000 6.000 10.000 14.000
18.000 22.000 26.000 30.000
34.000 38.000 42.000 46.000
50.000 54.000 58.000 62.000
plane = 3 ----------------------------
3.000 7.000 11.000 15.000
19.000 23.000 27.000 31.000
35.000 39.000 43.000 47.000
51.000 55.000 59.000 63.000
After FFT+IFT
plane = 0 ----------------------------
-0.000 -0.344 8.000 12.000
-0.031 20.000 24.000 -0.031
32.000 36.000 0.031 44.000
48.000 -0.094 56.000 60.000
plane = 1 ----------------------------
1.000 -0.000 9.000 13.000
-0.000 21.000 25.000 0.125
33.000 37.000 0.000 45.000
49.000 0.000 57.000 61.000
plane = 2 ----------------------------
2.000 6.000 -0.000 14.000
18.000 0.000 26.000 30.000
0.000 38.000 42.000 -0.000
50.000 54.000 -0.000 62.000
plane = 3 ----------------------------
3.000 7.000 0.031 15.000
19.000 -0.031 27.000 31.000
-0.031 39.000 43.000 0.031
51.000 55.000 0.031 63.000
I even tried to pad the data this way:
// With padding
unsigned int idx;
for (int x = 0; x < N; x++){
for (int y = 0; y < N; y++){
for (int z = 0; z < 2*(N/2+1); z++){
idx = z + N * ( y + x * N);
if (z < 4) in[idx] = idx;
else in[idx] = 0;
}
}
}
What am I doing wrong?
回答1:
As you already found out, you need padding if you use the CUFFT_COMPATIBILITY_FFTW_PADDINGcompatibility mode which is default. For your code to work you could use cufftSetCompatibilityMode() to set CUFFT_COMPATIBILITY_NATIVE. However, this mode is marked as deprecated in the current version of CUDA.
Therefore, I recommend to use the default compatibility mode and use padding. Your try to implement padding is wrong. The formula to calculate a linear index for 3 dimension x, y, z where z is the fastest running index is idx = z + Nz*(y + Ny*x). The size Nz of the z dimension including padding is Nz = (N/2+1)*2. Then, the correct initialization of the array is:
unsigned int idx;
for (int z = 0; z < N; z++){
for (int y = 0; y < N; y++){
for (int x = 0; x < N; x++){
idx = z + (N/2+1)*2 * ( y + x * N);
in[idx] = idx;
}
}
}
Accordingly for the print loops.
来源：https://stackoverflow.com/questions/33370015/wrong-results-cufft-3d-in-place