[WIP][Compiler][gfx120] Add modC/reuseA/reuseB to MmaOpGFX1250_WMMAType

[jli-melchior]

The non-scale WMMA path was missing modC/reuseA/reuseB parameters that its sibling WMMAScaleType already had. The lowering hardcoded WMMACModifier::none and reuseA/reuseB=false, causing mma_atom_call to emit rocdl.wmma ops without an explicit modC attribute. Since the ROCDL intrinsic distinguishes absent modC from modC=0, this produced incorrect ISA encodings and numerical precision regression for bf16 WMMA kernels.

Add modC (int32_t), reuseA (bool), reuseB (bool) to the type with back-compat builders defaulting to 0/false/false, forward them through the Python binding and wrapper, and pass them to the ROCDL ops at emission time.

Motivation

Technical Details

Test Plan

Test Result

Submission Checklist

• Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.

[sjfeng1999]

reuse flags should not be modeled as static fields of the WMMA type, since programmers cannot conveniently vary them across a sequence of WMMA calls.

[sjfeng1999]

Maybe a better way is to add a separate pass to attach these reuse attrs automatically.

[jli-melchior]

reuse flags should not be modeled as static fields of the WMMA type, since programmers cannot conveniently vary them across a sequence of WMMA calls.

reuseA/reuseB are bit fields of the wmma instruction that hint the XDL engine to reuse operand data. Exposing them gives users more flexibility for different scheduling scenarios ?

[jli-melchior]

Maybe a better way is to add a separate pass to attach these reuse attrs automatically.

Yes it sounds a good idea for a follow-up optimization which programmers may not set these flags, And it also need these fields on the type to construct the updated atom. This PR provides the foundation whether the flags are set manually or by a future automatic pass.

[sjfeng1999]

I can't see how callers could conveniently control reuse flags within a single fx.gemm invocation, which may expand into multiple atom calls. Could you provide a concrete example?

[jli-melchior]

I can't see how callers could conveniently control reuse flags within a single fx.gemm invocation, which may expand into multiple atom calls. Could you provide a concrete example?

In mha kernel i haven‘t see use the fx.gemm to call mma atom, the gemm kernel can use this to get a perf improvement, but this need a special pattern for reuseA/reusesB. This may need different pattern according to different cases.