cufftPlanMany: 1D FFT on matrix columns

Accelerated Computing GPU-Accelerated Libraries
1

Hello,

In my matrix, each row is VEC_LEN long.
A row is consecutive in GPU’s RAM.
The matrix has N_VEC rows.

I have to run 1D FFT on VEC_LEN columns. Each column contains N_VEC complex elements.
So I called:

int nCol[1] = {N_VEC};

res=cufftPlanMany (&plan, 1, nCol, //plan, rank, n
NULL, VEC_LEN, 1, //inembed, istride, idist
NULL, VEC_LEN, 1, //oneembed, ostride, odist,
CUFFT_C2C, VEC_LEN); //type, n_batch

res = cufftExecC2C (plan, pSrc, pDest, CUFFT_FORWARD);

I checked the results.
Each row starts with (1+j0) so the output of column should be:
(8 + j0), 0,0,0 …

But I got:
(36 + j0), (36 + j0) …

Can you please help ?

Thank you,
Zvika

2

I suggest providing a complete example. Include the output you actually get as well as the output you expect to get. Don’t use or include extraneous libraries like OpenCV or anything else.

Include the OS, CUDA version, GPU you are running on, and compile command line.

Make sure you are doing proper error checking for both CUDA and CUFFT calls, and make sure your code reports no errors when run with cuda-memcheck

3

note that passing NULL for inembed and onembed disables advanced data layout:

[url]https://docs.nvidia.com/cuda/cufft/index.html#advanced-data-layout[/url]

“Passing inembed or onembed set to NULL is a special case and is equivalent to passing n for each. This is same as the basic data layout and other advanced parameters such as istride are ignored.”

4

Here’s a full test case built around what you have shown:

$ cat t13.cu
#include <cufft.h>
#include <stdio.h>
#include <assert.h>

const int N_VEC = 8;    // vertical   - # of rows
const int VEC_LEN = 4;  // horizontal - # of columns

int main(){

int nCol[1] = {N_VEC};
cufftHandle plan;
cufftResult res;
res=cufftPlanMany (&plan, 1, nCol, //plan, rank, n
                NULL, VEC_LEN, 1, //inembed, istride, idist
                NULL, VEC_LEN, 1, //oneembed, ostride, odist,
                CUFFT_C2C, VEC_LEN); //type, n_batch
assert(res == CUFFT_SUCCESS);
cufftComplex *pSrc, *pDest, *h;
cudaMalloc(&pSrc, N_VEC*VEC_LEN*sizeof(pSrc[0]));
cudaMalloc(&pDest, N_VEC*VEC_LEN*sizeof(pDest[0]));
h = new cufftComplex[N_VEC*VEC_LEN];
for (int i = 0; i < N_VEC*VEC_LEN; i++) {
  h[i].x = 1;
  h[i].y = 0;}
cudaMemcpy(pSrc, h, N_VEC*VEC_LEN*sizeof(pSrc[0]), cudaMemcpyHostToDevice);
res = cufftExecC2C (plan, pSrc, pDest, CUFFT_FORWARD);
assert(res == CUFFT_SUCCESS);
cudaMemcpy(h, pDest, N_VEC*VEC_LEN*sizeof(pDest[0]), cudaMemcpyDeviceToHost);
for (int i = 0; i < N_VEC; i++) {
  for (int j = 0; j < VEC_LEN; j++) printf("%f,%f  ", h[i*VEC_LEN+j].x, h[i*VEC_LEN+j].y);
  printf("\n");}
}
$ nvcc -o t13 t13.cu -lcufft
$ cuda-memcheck ./t13
========= CUDA-MEMCHECK
8.000000,0.000000  0.000000,0.000000  0.000000,0.000000  0.000000,0.000000
0.000000,0.000000  0.000000,0.000000  0.000000,0.000000  0.000000,0.000000
8.000000,0.000000  0.000000,0.000000  0.000000,0.000000  0.000000,0.000000
0.000000,0.000000  0.000000,0.000000  0.000000,0.000000  0.000000,0.000000
8.000000,0.000000  0.000000,0.000000  0.000000,0.000000  0.000000,0.000000
0.000000,0.000000  0.000000,0.000000  0.000000,0.000000  0.000000,0.000000
8.000000,0.000000  0.000000,0.000000  0.000000,0.000000  0.000000,0.000000
0.000000,0.000000  0.000000,0.000000  0.000000,0.000000  0.000000,0.000000
========= ERROR SUMMARY: 0 errors
$

note that the output doesn’t match what you indicated:

when I make the change indicated in comment #3 above, I get possibly the output you are expecting:

$ cat t13.cu
#include <cufft.h>
#include <stdio.h>
#include <assert.h>

const int N_VEC = 8;    // vertical   - # of rows
const int VEC_LEN = 4;  // horizontal - # of columns

int main(){

int nCol[1] = {N_VEC};
cufftHandle plan;
cufftResult res;
res=cufftPlanMany (&plan, 1, nCol, //plan, rank, n
                nCol, VEC_LEN, 1, //inembed, istride, idist
                nCol, VEC_LEN, 1, //oneembed, ostride, odist,
                CUFFT_C2C, VEC_LEN); //type, n_batch
assert(res == CUFFT_SUCCESS);
cufftComplex *pSrc, *pDest, *h;
cudaMalloc(&pSrc, N_VEC*VEC_LEN*sizeof(pSrc[0]));
cudaMalloc(&pDest, N_VEC*VEC_LEN*sizeof(pDest[0]));
h = new cufftComplex[N_VEC*VEC_LEN];
for (int i = 0; i < N_VEC*VEC_LEN; i++) {
  h[i].x = 1;
  h[i].y = 0;}
cudaMemcpy(pSrc, h, N_VEC*VEC_LEN*sizeof(pSrc[0]), cudaMemcpyHostToDevice);
res = cufftExecC2C (plan, pSrc, pDest, CUFFT_FORWARD);
assert(res == CUFFT_SUCCESS);
cudaMemcpy(h, pDest, N_VEC*VEC_LEN*sizeof(pDest[0]), cudaMemcpyDeviceToHost);
for (int i = 0; i < N_VEC; i++) {
  for (int j = 0; j < VEC_LEN; j++) printf("%f,%f  ", h[i*VEC_LEN+j].x, h[i*VEC_LEN+j].y);
  printf("\n");}
}
$ nvcc -o t13 t13.cu -lcufft
$ cuda-memcheck ./t13
========= CUDA-MEMCHECK
8.000000,0.000000  8.000000,0.000000  8.000000,0.000000  8.000000,0.000000
0.000000,0.000000  0.000000,0.000000  0.000000,0.000000  0.000000,0.000000
0.000000,0.000000  0.000000,0.000000  0.000000,0.000000  0.000000,0.000000
0.000000,0.000000  0.000000,0.000000  0.000000,0.000000  0.000000,0.000000
0.000000,0.000000  0.000000,0.000000  0.000000,0.000000  0.000000,0.000000
0.000000,0.000000  0.000000,0.000000  0.000000,0.000000  0.000000,0.000000
0.000000,0.000000  0.000000,0.000000  0.000000,0.000000  0.000000,0.000000
0.000000,0.000000  0.000000,0.000000  0.000000,0.000000  0.000000,0.000000
========= ERROR SUMMARY: 0 errors
$

The choice of pointer parameter here (nCol) for inembed and onembed does not matter for the 1D case (read the advanced layout section doc link already given in comment 2) but it still must be non-NULL if you expect the istride, idist, ostride, odist parameters to be accounted for.

5

Hello txbob,

Thank you very much !
Your help is highly appreciable.
Now it works.

Best regards,
Zvika