[HLSL] Add LinAlg accumulation contention coverage

[JoeCitizen]

Summary

• extend the descriptor, group-shared and vector accumulation harnesses to drive deterministic contention from multiple Waves, threads and dispatch groups while preserving the existing single-invocation defaults
• add a guarded F32 descriptor case with two Waves in each of two thread groups; four exactly representable 0.5 additions to initial 1.0 produce the unique exact result 3.0
• add a guarded F16 group-shared case with two Waves contending in one thread group and a guarded F16 vector case with eight threads in each of two thread groups
• retain capability-gated I32 descriptor-matrix, group-shared-matrix and descriptor-vector cases for implementations that advertise integer atomic support
• add a checked host oracle for repeated integer accumulation totals and a host self-test

The runnable F16/F32 controls ensure every destination path executes locally. The I32 cases are not success-shaped fallbacks: they remain selected only when the runtime reports the matching atomic destination support.

Validation

• built the Release ExecHLSLTests target
• compiled LinAlgTests.cpp directly against the preview D3D12 headers
• ran the 14-case accumulation selection with the ABI-corrected disposable WARP: 11 passed and the 3 I32 capability cases skipped because WARP reports both integer atomic destinations unsupported
• positively executed the F32 descriptor and F16 group-shared contention cases on unmodified public preview WARP
• positively executed the F16 vector contention case, plus all three existing vector accumulation regressions, on the disposable WARP decoder corrected for current handle-first DXIL operands
• reran the capability-policy host test and the non-accumulating descriptor round-trip regression
• inspected emitted DXIL for exact wave size 4, two active Waves, Wave-scope Accumulator matrices, guarded descriptor offsets/strides, group-shared addrspace(3) accumulation, eight-thread dispatches and handle-first vector operands
• clang-format 17.0.1, git diff --check and a focused graphics correctness review pass

Runtime compatibility notes

WARP advertises MatrixConstruction for I32 but reports UAV=0 and groupshared=0 for I32 AtomicAccumulateStore, so those three cases correctly report NotApplicable. Public preview WARP also retains the known obsolete vector-first VectorAccumulate decoder; vector execution used the previously documented disposable decoder correction. No runtime change is included here.

• after the MatVec, Wave, ThreadGroup, and outer-vector review corrections were propagated through this stack, the complete 66-method suite on the compatible local runtime reported 61 passed, 0 failed, and 5 capability-backed skips
• serial review confirmed the matrix paths issue one collective update per active Wave, the vector path issues one update per Thread, all group barriers are uniform, and every floating intermediate is exactly representable and order-independent
• after restoring the clang-format form disturbed by the K/L propagation conflict, the focused six contention methods reported 3 passed and 3 authoritative I32 skips; the full 66-method suite again reported 61 passed, 0 failed, and 5 skips

No physical GPU or packaged-HLK qualification is claimed.

Stack

This draft is stacked on PR #8675, which is stacked on PR #8674, 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 contention implementation is commit 45da2a4882; serial-review formatting correction 9452edd22 restores the repository-required clang-format form without changing behaviour.

This remains a draft for named human review. The reviewer should verify the atomic invocation counts, uniform Wave participation, group-shared barriers, exact/order-independent value choices, destination guards, capability gating and checked host totals before requesting maintainer review.

Refs #7841
Refs #8561
Refs #8562
Refs #8565
Refs #8654

Assisted-by: GitHub Copilot

[github-actions]

✅ With the latest revision this PR passed the C/C++ code formatter.