Tensile icon indicating copy to clipboard operation
Tensile copied to clipboard

Published 20 hours ago • verified publisher icon ROCm

Add defineLocalSgpr

Open awhittle3 opened this issue 1 year ago • 5 comments

Added a new method to manage SGPR usage: defineLocalSgpr. This allows in-place defining/undefing of SGPRs.

Also added new parameter PrintRegisterDebug that informs the invoker if SGPRs have not been explicitly undefined, as a optimization hint.

The old defineSgpr function still works as expected and is in place elsewhere in the code.

awhittle3 avatar Dec 06 '23 20:12 awhittle3

CI tests failed due to incorrect asm code. Please fix them.

nakajee avatar Dec 07 '23 02:12 nakajee

I think it is better to run precheckin test locally before you commit code change.

nakajee avatar Dec 07 '23 02:12 nakajee

@awhittle3 Some CI tests failed, do we still need this change?

babakpst avatar Jan 17 '24 22:01 babakpst

@awhittle3 Some CI tests failed, do we still need this change?

This PR will conflict with PR #1843, which probably has higher priority. I'll address the issues here once that PR is merged into develop.

@AlexBrownAMD believes this change will be helpful in managing SGPR usage.

awhittle3 avatar Jan 18 '24 18:01 awhittle3

@awhittle3 Some CI tests failed, do we still need this change?

This PR will conflict with PR #1843, which probably has higher priority. I'll address the issues here once that PR is merged into develop.

@AlexBrownAMD believes this change will be helpful in managing SGPR usage.

#1843 is merged today. Would you please restart this PR?

nakajee avatar Feb 07 '24 19:02 nakajee

@awhittle3 we are approaching FC 6.2 and we want to clear all the requests. Please let me know if you are planning to proceed with this PR. Thanks.

babakpst avatar Mar 21 '24 22:03 babakpst

Out of date and would take a while to spin back up. Closing work on this for now.

awhittle3 avatar Mar 22 '24 16:03 awhittle3