[Fix][TransformTensor] Ignore output buffer when BETA=0 - Githubissues

ROCm / MIOpen

AMD's Machine Intelligence Library

https://rocm.docs.amd.com/projects/MIOpen/en/latest/

Other

1.09k stars 231 forks source link

[Fix][TransformTensor] Ignore output buffer when BETA=0 #3184

Closed atamazov closed 3 months ago

atamazov commented 4 months ago

The primitive produces invalid results when BETA=0 and output buffer contains junk (NaNs). This PR fixes the issue.

By-products:

The fix eliminates reads of output buffer when BETA=0, so some perf gain is expected.
Small perf gain of kernels for cases when ALPHA=1.
Removal of some unused code.

Related issue:

2828

[Attribution] @junliume @JehandadKhan

https://github.com/ROCm/MIOpen/labels/bug
https://github.com/ROCm/MIOpen/labels/correctness
https://github.com/ROCm/MIOpen/labels/urgency_blocker
Proposed reviewers:
- @JehandadKhan
- @junliume
- @t4c1
- @ShanoToni
Optional reviewers:
- @xdgarrido

atamazov commented 3 months ago

@CAHEK7 Maybe later (I do not have perf tests for this primitive on hand). This PR is about correctness and I am quite sure that is doesn't lead to perf degradations (hope you sure too). Also I suspect that the most of time on GPU is spent for address/offset computations, so the expected perf gain is pretty small.