FFT exec error for batched 1D FFT

Hi,

I have a relatively simply program using batched, 1D real-to-complex FFTs of size 1024.
The plan is generated OK, but after calling cufftExecR2C the following is printed on stderr:

cufft: ERROR: execute.cu, line 992
cufft: ERROR: CUFFT_EXEC_FAILED
cufft: ERROR: cufft.cu, line 115
cufft: ERROR: CUFFT_EXEC_FAILED
cuda FFT failed with result code 6

My system is a linux CentOS 4.3 with 8800 GTX. The card is working fine- I have done a simple 2D FFT test on it with no problems.

I’m hoping this is some simple error I’ve made allocating memory or something. I have double-checked all the setup before doing the plan and exec and it all seems ok.

If someone can give me some more details on what exactly these errors mean it would really help me sort this out.

Cheers,
Randall.

Can you post your code?

Hi,

code attached (as a .txt file). It needs an input file to actually do anything, but that is much too large to attach. If you need that, let me know and I’ll make it available to download.

Cheers,

Randall.
corr_gpu.txt (9.25 KB)

I am not sure it is the only bug, but you are using cudaMemset on an array allocated with cudaMallocHost.
cudaMallocHost allocates a page-locked array on the host, you should use host calls on the array not cuda calls.

If you use a regular memset on that array, the code seems to work ( at least the cufftt errors are gone).

Indeed that was the problem. My sincerest thanks!

I hope to see a CUDA accelerated Lensview…