Skip to content

Conversation

@mirza-halilcevic
Copy link
Contributor

@mirza-halilcevic mirza-halilcevic commented Oct 29, 2025

Motivation

Update gemm and conv quick-tune lists for:

Technical Details

Test Plan

Test Result

Submission Checklist

@mirza-halilcevic mirza-halilcevic force-pushed the quick_tune_code_review branch 2 times, most recently from 424cb38 to f78e931 Compare November 3, 2025 18:19
@mirza-halilcevic mirza-halilcevic force-pushed the quick_tune_code_review branch 2 times, most recently from 8fb43c1 to b249779 Compare November 4, 2025 10:10
{"gfx950_conv_i8", {PopulateParamsXDL::initParametersForwardI8ConvGfx950, PopulateParamsXDL::nInitParametersForwardI8ConvGfx950}},
{"gfx1000_gemm_f16", {PopulateParamsWmma::initParametersFp16GemmGfx1000, PopulateParamsWmma::nInitParametersFp16GemmGfx1000}},
{"gfx1100_gemm_f16", {PopulateParamsWmma::initParametersFp16GemmGfx1100, PopulateParamsWmma::nInitParametersFp16GemmGfx1100}},
{"gfx1200_gemm_f16", {PopulateParamsWmma::initParametersFp16GemmGfx1200, PopulateParamsWmma::nInitParametersFp16GemmGfx1200}},
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Does the gfx1200 fall back to gfx1201? The gfx1200 does exist and should be supported.

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Yes, it will fallback to gfx1201. The fallback logic is implemented in #1994, and it basically looks for the closest arch if it can't find an exact match. The old list was also based on gfx1201, it was just generically named gfx1200 because it would be used for any gfx12**, that's why it is redundant now.

We can tune for gfx1200 as well, if needed.

@mirza-halilcevic mirza-halilcevic marked this pull request as ready for review November 6, 2025 14:03
@mirza-halilcevic mirza-halilcevic changed the title Update quick-tune lists Update quick-tune lists for Strix and Navi48 Nov 6, 2025
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

4 participants