I repeatedly get the error message ‘CUDA_EXCEPTION_9: Warp Hardware Stack Overflow’ in one of my methods trying to use malloc on the device. Using a very basic test case, malloc seems to work but for my actual function it crashes. Unfortunatelly the cuda-gdb manual does not explain what may cause this error or how to fix it. The only description given is
I have tried to increase the possible malloc size using
and using the equivalent …GetLimit() confirms my changed size. Since the amount of data required in my test cases isn’t too big anyway, I doubt a lack of heap memory directly causes the problem. Using cuda-memcheck within cuda-gdb gave the same results.
I am using a GTX 460 with CUDA 4.0, OS: Ubuntu 10.04 LTS 64bits;
NVRM version: NVIDIA UNIX x86_64 Kernel Module 280.13 Wed Jul 27 16:53:56 PDT 2011
For the time being it would already be quite helpful for me to understand the precise meaning of the error message (including possible implications on blocks/threads,…). Why “This should be a rare occurrence”, what would be expected?!