Need support from NVIDA Error that reported in cudaFFT

I’m using cudaFFT with multi-GPUs, I have trouble when i use cudaFFT, it shows runtime error in CUDAFFT. The same code work fine with single GPU. Can some one tell me what is the error about, or sheet a light for what cause the error, and what i should do

cufft: ERROR: /root/cuda-stuff/sw/rel/gpgpu/toolkit/r2.0/cufft/src/, line 41

I use CUDA 2.0b version. I use an iterative method that run GPU threads every iteration, inside the thread I create cudaFFT plan, and use that one to perform cudaFFT.


You might want to paste your code or attach the project.

Thank you for your suggestion. It is a big project. I just want if someone involves in cudaFFT project tell me what is the error reported so that i can track down the problem.

I’ve senn a similar error some time before. The bug was actually in the last kernel call before the FFT execution. Maybe you should query the cuda error before FFT execution.


Thank you. I finally found the problem. When the CUDA kernel quit the thread, all the memory is invalid however the pointer is not reset and memory is not released, that why next time the kernel called, the pointer is not null and still occupy the memory, the same thing happen with cuda Plan. So my solution is to release the memory, destroy the plan and set all pointer to null before exit the thread. I think it would be nice if the memory contain can stay the same so that in the next iteration i don’t have to reupload data from host