cuda - How to best check for "too much static shared memory" in a PTX?

Question

I'm compiling kernel(s) using NVRTC. Suppose I wrote one which uses more static shared memory than my GPU supports. Thus, when I try to cuModuleLoadDataEx() the compiled module, I get an error.

The thing is, the error I get is CUDA_ERROR_INVALID_PTX. That's not technically true, and regardless - it is not a good indication of the lack of sufficient shared memory.

What can I do - not having a PTX parser library myself - to automagically figure out the out-of-shared-memory issue?

Note: This question is about static memory only. Assume there is no use of dynamic shared memory, so we can completely ignore it.

Answer 1

等待大神答复