magma_zgetri_gpu segfaults in OpenMP parallel for

[mgates3]

Not yet. We're aware of the issue, but it will take a while to upgrade all the functions to use CUBLAS v2, which has better stream support for thread safety.
-mark

[jeremiahpalmer]

Thanks for the update. I'd be happy to help in alpha or beta testing, if needed. :)

[jeremiahpalmer]

I applied the fix Stan mentioned in http://icl.cs.utk.edu/magma/forum/viewtopic.php?f=2&t=520:

Also, there was a bug in file interface_cuda/alloc.cpp. Function magma_free_host should free memory with cudaFreeHost (instead of using cudaFree).

Now, my OpenMP parallel calls to magma_zgesv seem to be working properly.

[jeremiahpalmer]

Any time frame for when all of the functions will use CUBLAS v2, as Mark mentioned?

Not yet. We're aware of the issue, but it will take a while to upgrade all the functions to use CUBLAS v2, which has better stream support for thread safety.
-mark

[mgates3]

I think this will be a goal for the fall, but don't have a specific deadline. It will help with porting MAGMA code to OpenCL, so that the two codes use a queue / stream in the same ways. There are some CUDA synchronization issues to deal with, though.
-mark