[SM6.10] LinAlg spec review: identify signature gaps vs 0035-linalg-matrix spec

[Copilot]

Reviewed PR changes against the SM6.10 LinAlg spec (proposal 0035) and identified discrepancies between the builtin signatures and the specified DXIL ops.

Key findings

• MatrixMatrixMultiplyAccumulate missing input accumulator: Spec defines linAlgMatrixMultiplyAccumulate as R = C + (A*B) — four distinct matrix params. Current builtin has only three (identical shape to MatrixMatrixMultiply); the input accumulator C is absent.

• No linAlgMatrixMultiplyAccumulate DXIL op in hctdb.py: Only LinAlgMatrixMulOp ("o,o,o") exists, covering the pure-multiply case. A separate op with signature "o,o,o,o" is needed for the accumulate variant.

• LinAlgMatrixMulOp naming drift: Spec names the pure-multiply op linAlgMatrixMultiply; hctdb.py uses LinAlgMatrixMulOp. Aligning names would improve cross-referencing.

• Matrix Use discrepancy in MatrixVectorMultiply: HLSL-level linalg::Multiply(vector, Matrix) specifies a B matrix, but the DXIL op linAlgMatVecMul references "matrix A of Thread scope". Needs validation that the correct Use constraint is enforced during lowering.

• Missing align parameter: linAlgMatrixLoadFromDescriptor, linAlgMatrixStoreToDescriptor, and linAlgMatrixAccumulateToDescriptor all include an optional align param absent from the builtins. If intentionally defaulted, a comment would help.

---

✨ Let Copilot coding agent set things up for you — coding agent works faster and does higher quality work when set up for your repo.