Alexander Efimov

[AMD] MFMA shortcut test

1

This PR: - moves shortcut check above allocation code, before any scratch buffer shape is computed - raise priority of AMD specific over common conversions

[WIP] Support small dots and optimization of dot operands

2

This PR - Introduces several fixes in FMA dot implementation - Enables support of small dots with MNK dimensions down to 1 - Introduces dot operand optimization for dots with...

[AMD] Refactor decompose-unsupported-amd-conversions pass

1

This PR introduces: - Use common code, simplify pass code - Support 3d tensors in mfma -> dot conversion(supported in common code from item above) - More tests for decompose-unsupported-amd-conversions...

[AMD] Support FP8E5M2 with MFMA FP16 instructions

1

Cast dot arguments from unsupported FP8 to supported FP16 in order to use MFMA instructions instead of FMA. This approach is expected to give better performance and be more stable...

[Backend] Improve dot support to target FMA

1

This PR: - Refactors FMA dot implementation - Supports dot3d in FMA path - Fixes several issues in operand offset computation - Enables small dot operands This PR is a...

[AMD] Fix shared layout order for batch dimension in pipeline passes

3

Batch dimension should be slowest one, other cases are not supported by MFMA/WMMA/MMA pipeline.

This PR extends shared memory bypass for blocked->dot operand conversions and adds bypass check in DecomposeUnsupportedConversions and ReduceDataDuplication. This PR is a part of PR series. Final goal is to...

[WIP] [AMD] Emit AMD specific intrinsics for dot

1

This PR: - Makes AccelerateAMDMatmul pass to emit FMA i8xi8->i32 and fp16xfp16->fp32 cases - Extends AMD FMA Dot code generation with new v_dot instructions for fp16xfp16 and int8 dtypes This...

Hoist reduction outside a loop

6

This PR introduces an optimization that hoists reduction operation of dot accumulator outside a loop over K dimension: %acc = for k tiles: %acc3d_input = reshape %acc %acc3d_out = dot3d(%x,...

[WIP] Optimize fma dot

1

The core Triton is a small number of people, and we receive many PRs (thank you!). To help us review your code more quickly, **if you are a new contributor...