Feature/mi300 a support #270
Conversation
reguly
left a comment
reguly
left a comment
There was a problem hiding this comment.
This looks good to me. However, please try and run the full test suite on an MI300A to make sure all the code paths are covered well
|
@Ashutosh-Londhe do you have access to an MI300A to run our usual test scripts for this ? |
Sorry i dont have access to MI300A. But seems like benny is trying on COSMA machine so i can request an access to that project which has MPI300A. I will check wtih Beniel about the details. |
I have access to MI300A and am testing it right now before finalizing the pull request. There are some minor bugs in the current development related to ROCM-SMI power profiling. I will add that change with this pull request as well. |
…MI300A_support
|
Added all the changes. With these changes, 10% perf improvement on the gnu_ompi combination. The MI300A platform we tested on Archer2 does have ROCM toolchain + OpenMPI, where we saw significant improvement. Unfortunately, I couldn't |
|
An additional fix commit has been added to resolve the linking error related to power profiling with the rocmSMI APIs |
|
Hi @gihanmudalige I have got the access to MI300A to perform the testing. For Laplace2d and Poisson, HIP version is failing to produce correct result, informed this to Beniel, he is looking into the issue. @reguly will similar changes will also require for SYCL backend if underline GPU support unified memory? or SYCL doesnt support this feature? |
|
Yes, ideally with all "offload"-style backends should support this. |
No description provided.