[HLSL] Add CopyConvert and Convert coverage for LinAlg

[Jack Elliott (JoeCitizen)]

Proposal 0035 specifies CopyConvert as a converting copy between matrix layouts and Convert as an element-wise conversion with RTNE and saturation. Neither had cross-component coverage.

These four tests convert between component types at one advertised wave size and read the source matrix back so a destructive implementation is caught. Rectangular and transposed cases make the destination shape and stride observable. Expected values are derived on the host: exact I16 widening, and hand-derived RTNE and saturation vectors for F32 to I16.

FP8 coverage is held back for a follow-up. The gate here now asks whether a destination type is advertised as a vector input rather than whether FP8 matrices can be constructed. That is the right question, but it also makes the FP8 cases executable on WARP, where they fail. That result deserves its own change rather than being buried in this one.

[Copilot]

Pull request overview

Adds cross-component LinAlg execution coverage for CopyConvert and Convert.

Changes:

• Tests F16/F32 matrix conversion, transpose, and source preservation.
• Tests integer RTNE/saturation and both FP8 formats.
• Adds capability selection and host-side conversion oracles.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[Copilot]

Pull request overview

Copilot reviewed 1 out of 1 changed files in this pull request and generated no new comments.

Suppressed comments (2)

tools/clang/unittests/HLSLExec/LinAlgTests.cpp:2339

• The PR description promises six cross-component tests covering both FP8 formats, but only four tests are registered here, and none of the added test shaders invokes Convert with F8_E4M3FN or F8_E5M2. The advertised FP8 encode/decode coverage is therefore missing; please add and register those cases with the independent host inputs described in the PR, or revise the stated scope if they were intentionally dropped.

  // CopyConvert / Convert coverage
  TEST_METHOD(CopyConvert_Wave_4x8_F16_ToF32);
  TEST_METHOD(CopyConvert_Wave_4x8_F32_ToF16_Transpose);
  TEST_METHOD(Convert_I16_ToI32_Exact);
  TEST_METHOD(Convert_F32_ToI16_RTNE_Saturate);

tools/clang/unittests/HLSLExec/LinAlgTests.cpp:5287

• These declarations mirror an unreleased D3D12 ABI outside the repository's dedicated linalg_abi compatibility layer, and the assertions below check only total size. An SDK change that preserves size but changes alignment or field offsets would compile and then pass malformed data to CheckFeatureSupport. HlslExecTestUtils.h:169-239 establishes the convention of pinning mirrored types by enum value, size, alignment, and every field offset; please add these feature/enumeration definitions and checks there as well.

struct ConvertThreadVectorMatrixMultiplyEnumeration {
  UINT OperationType;
  UINT NumEntries;
  ConvertThreadVectorMatrixMultiplyEntry *ThreadVectorMatrixMultiply;
};

[Damyan Pepper (damyanp)]

Mainly LGTM, but there are some coding guidelines not being followed that we should get addressed and teach the agent about before merging.

[Copilot]

Pull request overview

Copilot reviewed 1 out of 1 changed files in this pull request and generated no new comments.