[HLSL] Add LinAlg OuterProduct and VectorAccumulate coverage

[JoeCitizen]

Summary

• add capability-gated non-uniform Thread OuterProduct coverage for F16 8x16 and F32 4x8, while retaining and activating the existing F16 16x16 case
• accumulate in DXIL OuterProductOptimal layout and use the preview D3D12 matrix-conversion API for tight RowMajor host readback, with exact checked CPU expectations
• migrate the existing F16 length-four VectorAccumulate case to the shared harness and add F16/F32 length-eight cases with non-zero destinations and checked host multiplication/addition oracles
• replace the released-SDK OuterProduct skip with SDK-neutral preview ABI mirrors, validated against the preview headers
• require OuterProduct cases to advertise both ThreadOuterProduct and RWByteAddressBuffer atomic accumulation, assert every field offset in the conversion ABI mirrors, and guard every vector destination with typed trailing sentinels

The shader and host APIs intentionally use their respective enum values for OuterProductOptimal: DXIL layout value 4 and D3D12 conversion layout value 3.

Validation

• built the Release ExecHLSLTests target
• compiled both changed implementation units against the preview D3D12 headers, activating exhaustive ABI size, alignment, field-offset and enum assertions
• ran the capability-policy host test plus all 3 OuterProduct methods on the compatible WARP/Agility runtime: 4 passed; each OuterProduct case positively queried both operation capabilities
• ran all 3 VectorAccumulate methods on the compatible WARP decoder: 3 passed, including exact verification of typed trailing guards
• ran the complete 60-case LinAlg selection with the compatible local runtime: 57 passed, the unchanged FP8 and ColumnMajor capability cases skipped, and only the already-documented mandatory native-F32/SInt8 MatVec WARP defect failed
• inspected emitted DXIL for Thread scope, Accumulator use, non-uniform shapes, layout value 4 and handle-first VectorAccumulate operands
• clang-format 17.0.1, git diff --check and a focused correctness review pass

Runtime compatibility notes

Public preview WARP currently decodes VectorAccumulate using the obsolete vector-first operand order, while current DXC emits the specified handle-first order. Agility SDK 1.721 also rejects ConvertLinearAlgebraMatrix on a compute command list; the local OuterProduct run used the documented DIRECT-list compatibility workaround. Neither disposable runtime change is included here, and the tests retain the current compiler/runtime contract rather than hiding either defect.

• after the MatVec, Wave, ThreadGroup, and outer-vector review corrections were propagated through the complete stack, the 66-method suite on the compatible local runtime reported 61 passed, 0 failed, and 5 capability-backed skips

No physical GPU or packaged-HLK qualification is claimed.

Stack

This draft is stacked on PR #8674, which is stacked on PR #8673, PR #8672, PR #8671, PR #8670, PR #8669, PR #8668, PR #8667, PR #8666, PR #8665 and PR #8662. Until those ancestors land, this diff contains their commits as well. The OuterProduct and VectorAccumulate implementation is commit 8ac77765e6; review correction 353ca9a02 adds complete applicability, ABI-offset, and trailing-guard validation.

This remains a draft for named human review. The reviewer should verify the complete preview ABI mirrors, distinct DXIL/host layout values, intersected OuterProduct/atomic capability tuples, conversion/readback path, VectorAccumulate operand contract, trailing guards, and independent exact oracles before requesting maintainer review.

Refs #7841
Refs #8563
Refs #8565

Assisted-by: GitHub Copilot