New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
"access violation" problem #5774
Comments
Thanks for the report and nice reproducer! I can reproduce this. |
Hi, I have a similar problem and it is very annoying (OSError: exception: access violation reading ... in nvvm.py) see #4783 . I managed to upload a minimal reproducer (see https://gitlab.com/pittnerf1/test-project-for-numba-and-cuda-functions/-/tree/master ) which crashes on both Windows 10 and Google Colab, too. |
For me, the reproducer is intermittently segfaulting, and repeated runs reduce the likelihood of segfault. |
I believe #6030 will fix this. |
Closing now that #6030 is merged. |
Since these (in my opinion, non-descript) access violation errors still pop up on occasion (as it did for me), debugging with Cuda Simulator seems to help pinpoint the source of the error in the code. See my answer for Numba code compilation for CUDA GPU fails with OSError: exception: access violation reading |
Thanks for linking to the answer. This issue was caused by some transformations that were done to the IR that made it invalid when passing it to NVVM. However, for errors in user code using the simulator can also be a good option. If the simulator is too slow, then one can also add |
@gmarkall good distinction. Thank you! |
Reporting a bug
Thank you all for developing numba! I'm now using numba 0.49.1 with cuda 10.2.89 on python 3.7.5, windows 10. I have a problem in my code that it... sometimes fail to execute due to
OSError: exception: access violation reading 0x<non-zero-non-0xFF...address>
.cuda-memcheck.exe
shows nothing.The memory access violation occurs every time in the numba\cuda\cudadrv\nvvm.py in compile(self, **options), line 232, err = self.driver.nvvmCompileProgram(self._handle, len(opts), c_opts)
If run the program again, it shows
CudaAPIError: [500] Call to cuModuleGetFunction results in CUDA_ERROR_NOT_FOUND
.Here's a reproducer works on my machine:
The text was updated successfully, but these errors were encountered: