[0035] Dynamic indexing of LinAlg wave-scope/group-scope matrices

[jenatali]

Which document does this relate to?
https://github.com/microsoft/hlsl-specs/blob/main/proposals/0035-linalg-matrix.md

Describe the issue you see with the spec
Is it valid to create and index arrays of scope-uniform matrices? For example:

using MatrixATy = Matrix<ComponentType::F16, TILE_SIZE, TILE_SIZE, MatrixUse::A, MatrixScope::Wave>;
MatrixATy matrices[3];
matrices[someIndex] = matrices[someOtherIndex];

My read of the spec doesn't make it clear one way or another. If it is valid, this indicates that matrices will be passed to Load and Store operations. Can matrices also be passed to select operations? What are the uniformity requirements on the indices to load/store, or the selector for select? If it's valid for matrices to be passed straight to load/store, then it's a bit confusing why there's dedicated load/store from memory intrinsics.

[llvm-beanz]

Yes, dynamically indexing on arrays of matrix objects in local memory is legal. All of the operations on wave and thread group scope matrices need to be under uniform control flow for the wave or thread group respectively (and operating on the same matrix), so the indices would need to be appropriately uniform.

The reason we allow load and store instructions on thread-local memory for matrix objects but use an intrinsic for device and groupshared memory is because the matrices don't have defined size or memory-layouts in thread-local memory. They're opaque handles that we expect the backends to replace with per-thread partial matrices since the actual storage will be shuffled across the uniformity scope.

I can put a clarifying update into the spec.