LeiWang1999 commented on PR #15106: URL: https://github.com/apache/tvm/pull/15106#issuecomment-1598482428
The performance problem was due to the usage of local memory scope instead of warp scope in tensor intrins. To address this, we need to switch to warp scope in tensorization and pass the "lower_warp_storage" optimization pass to convert warp memory to register files. Using local memory resulted in excessive redundant register file usage, leading to register spills and decreased performance. this issue is hard to analysis in llvm ir btw. I wrote anther HIP source codegen to address the bug more effectively, and which can offer similar performance as llvm ir does. maybe we can open another thread [LeiWang1999/tvm/lei/feat-hip](https://github.com/LeiWang1999/tvm/blob/lei/feat-hip/tests/python/unittest/test_tir_schedule_tensorize_hip_mfma.py#L112). -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected]
