Daniel Hernandez-Juarez
Daniel Hernandez-Juarez
By the way, reduce f16 should not be enabled for navi. As it does not support atomic_add with packed_f16, so it would be slow. Leaving this comment here for whoever...
> What's the status on this PR? Is there additional work needed or is the removal of the allowed types the only thing that needs review? I thought this was...
@richagadgil we recently merged (rocmlir) support for bf16 atomic add for navi4 and gfx950. So, that means we can enable reductions for those to architectures for bf16. To sum up,...
@turneram I think this is still needed?
@richagadgil can we reopen this?
I've tried getting run-time for the same kernel as above but changing the output layout as well: ```mlir module { func.func @mfma_layout(%a: !migraphx.shaped, %b: !migraphx.shaped) -> !migraphx.shaped attributes{kernel, arch =...
> Can you post the data as a comma separated list or a link to the spreadsheet online? [AccelLayout_migraphx.csv](https://github.com/user-attachments/files/21719944/AccelLayout_migraphx.csv)
Same file but for gfx1201: [AccelLayout_gfx1201_migraphx.csv](https://github.com/user-attachments/files/22382297/AccelLayout_gfx1201_migraphx.csv)