[HLSL] Correct LinAlg OuterProduct accumulator use (#8662)

## Summary

`OuterProduct` produces an accumulator, not a matrix-A operand. Proposal
0035 defines `dx::linalg::OuterProduct` as returning `Matrix<...,
MatrixUse::Accumulator, MatrixScope::Thread>`, and DXC's public
`<dx/linalg.h>` API implements that contract. The execution test instead
hard-coded `USE_A`, so its emitted DXIL described the result as a
matrix-A value (`U0`) before passing it to the accumulator-store path.

This change:

- declares the result using the shared `USE` parameter;
- sets the host parameter to `MatrixUse::Accumulator`;
- asserts that invariant in the host runner; and
- changes the emitted matrix metadata from `U0` (`A`) to `U2`
(`Accumulator`) while preserving the existing arithmetic and exact
value-4 result.

## Why this is needed

The original test was out of the SM6.10 proposal and public API
contract.

The mismatch was easy to miss because `MatrixUse` is type metadata
rather than part of the numerical outer-product calculation. The current
DXIL validator checks that an OuterProduct result has well-formed matrix
metadata, dimensions, and component type, but does not enforce that its
use is `Accumulator`. Consequently, the old shader compiled and produced
the expected numbers while exercising the wrong matrix role. Leaving it
unchanged would let the execution/HLK test bless `U0` metadata even
though source compiled through `dx::linalg::OuterProduct` produces `U2`.

This is therefore a conformance correction, not a functional expansion.
It adds no new shapes, types, scopes, or capability handling.

Refs #7841
Refs #8563
Refs #8561

Closes #8682

Assisted-by: GitHub Copilot

Co-authored-by: Jack Elliott <jackell@ntdev.microsoft.com>
Copilot-Session: 83725f5d-8e98-4c1d-91ee-ad47629e007b

tools/clang/unittests/HLSLExec/LinAlgTests.cpp

diff --git a/tools/clang/unittests/HLSLExec/LinAlgTests.cpp b/tools/clang/unittests/HLSLExec/LinAlgTests.cpp
index 8db0c67..ea3391b 100644
--- a/tools/clang/unittests/HLSLExec/LinAlgTests.cpp
+++ b/tools/clang/unittests/HLSLExec/LinAlgTests.cpp
@@ -1395,7 +1395,6 @@
 }
 
 static const char OuterProductShader[] = R"(
-  #define USE_A 0
   #define SCOPE_THREAD 0
 
   RWByteAddressBuffer Input : register(u0);
@@ -1416,7 +1415,7 @@
     }
 
     __builtin_LinAlgMatrix
-      [[__LinAlgMatrix_Attributes(COMP_TYPE, M_DIM, N_DIM, USE_A, SCOPE_THREAD)]]
+      [[__LinAlgMatrix_Attributes(COMP_TYPE, M_DIM, N_DIM, USE, SCOPE_THREAD)]]
       Mat;
     __builtin_LinAlg_MatrixOuterProduct(Mat, VecA, VecB);
 
@@ -1434,6 +1433,8 @@
   VERIFY_IS_TRUE(
       Params.Layout == MatrixLayout::OuterProductOptimal,
       "Outer product must output its matrix in OuterProductOptimal layout");
+  VERIFY_IS_TRUE(Params.Use == MatrixUse::Accumulator,
+                 "Outer product must output an accumulator matrix");
   const size_t NumVecElements = Params.M + Params.N;
   const size_t InBuffSize = NumVecElements * elementSize(Params.CompType);
   const size_t NumMatElements = Params.totalElements();
@@ -1502,6 +1503,7 @@
   Params.CompType = ComponentType::F16;
   Params.M = 16;
   Params.N = 16;
+  Params.Use = MatrixUse::Accumulator;
   Params.Scope = MatrixScope::Thread;
   Params.Layout = MatrixLayout::OuterProductOptimal;
   Params.NumThreads = 1;