AMD perf cherry picks for release/internal/3.2.x#767

Merged

jataylo merged 19 commits intoROCm:release/internal/3.2.xfrom

jataylo:27_perf_cp

Mar 27, 2025

Commits on Mar 27, 2025

[AMD] Emit vectorized 16-bit float LLVM atomic ops (#4925)

joviliast
authored and
jataylo
committed
[AMD] Restructure ReorderInstructions pass (#4998)

antiagainst
authored and
jataylo
committed
[AMD] Support warp-level reduction with DPP (#5019)

knwng
authored and
jataylo
committed
[AMD] Add missing dependency to TritonAMDGPUIR (#5053)

antiagainst
authored and
jataylo
committed
[AMD] Support warp-level reduction with DPP (#5019)

knwng
authored and
jataylo
committed
[AMD] Reland sinking the 2nd tt.load after local_load's (#4935)

zhanglx13
authored and
jataylo
committed
[BACKEND][NVIDIA] Add Lowering for Shared-to-MMAv3-DotOp Copy (#5009)

ggengnv
authored and
jataylo
committed
[AMD] Add initial support for scaled_dot(mxfp8, fp8) (#4994)

antiagainst
authored and
jataylo
committed
[Frontend][Backend] Implement support for scale_dot(-, bf16) (#4996)

lezcano
authored and
jataylo
committed
[BACKEND] Improve detection of register to register conversion (#4991)

Jokeren
authored and
jataylo
committed
[BACKEND] Replace `isMmaToDotShortcut` with linear layout based logic (#4951)

Jokeren
authored and
jataylo
committed
[BACKEND]Fix DotOperand(Ampere) LinearLayoutConversion (#5038)

lezcano
authored and
jataylo
committed
[BACKEND] Fix uses of getOrder(DotOperand(Nvidia) and MMA(Nvidia)) (#5055)

lezcano
authored and
jataylo
committed
[AMD] Reland instruction scheduling hint changes (#4940)

ravil-mobile
authored and
jataylo
committed
[AMD] Enable scaled_dot(-, bf16) (#5029)

antiagainst
authored and
jataylo
committed
[AMD] Add support for scaled_dot(mxfp4, -) (#5034)

antiagainst
authored and
jataylo
committed
[BACKEND] Minor Bugfixes for SharedToDotOperand MMAv3 (#5030)

ggengnv
authored and
jataylo
committed
[BACKEND] Get rid of unpack/pack I32 (#5044)

Jokeren
authored and
jataylo
committed
Consolidate `getOrder` as "element order" and implement `getRepOrder` for general and NVIDIA layouts (#5089)

lezcano
authored and
jataylo
committed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

AMD perf cherry picks for release/internal/3.2.x#767

AMD perf cherry picks for release/internal/3.2.x#767
jataylo merged 19 commits intoROCm:release/internal/3.2.xfrom
jataylo:27_perf_cp

Commits on Mar 27, 2025

[AMD] Emit vectorized 16-bit float LLVM atomic ops (#4925)

[AMD] Restructure ReorderInstructions pass (#4998)

[AMD] Support warp-level reduction with DPP (#5019)

[AMD] Add missing dependency to TritonAMDGPUIR (#5053)

[AMD] Support warp-level reduction with DPP (#5019)

[AMD] Reland sinking the 2nd tt.load after local_load's (#4935)

[BACKEND][NVIDIA] Add Lowering for Shared-to-MMAv3-DotOp Copy (#5009)

[AMD] Add initial support for scaled_dot(mxfp8, fp8) (#4994)

[Frontend][Backend] Implement support for scale_dot(-, bf16) (#4996)

[BACKEND] Improve detection of register to register conversion (#4991)

[BACKEND] Replace `isMmaToDotShortcut` with linear layout based logic (#4951)

[BACKEND]Fix DotOperand(Ampere) LinearLayoutConversion (#5038)

[BACKEND] Fix uses of getOrder(DotOperand(Nvidia) and MMA(Nvidia)) (#5055)

[AMD] Reland instruction scheduling hint changes (#4940)

[AMD] Enable scaled_dot(-, bf16) (#5029)

[AMD] Add support for scaled_dot(mxfp4, -) (#5034)

[BACKEND] Minor Bugfixes for SharedToDotOperand MMAv3 (#5030)

[BACKEND] Get rid of unpack/pack I32 (#5044)

Consolidate `getOrder` as "element order" and implement `getRepOrder` for general and NVIDIA layouts (#5089)