Repository navigation
Enable creation of a LinAlg Matrix (type, intrinsic, dxil op) #7846
Description
This issue is the very first of the set of issues to support the Linear Algebra spec. Because of this it has to solve the chicken-or-the-egg problem.
We need the new dx.types.MatrixRef type to implement any new intrinsic, but to add the dx.types.MatrixRef type we need an intrinsic to actually observe it. Because of this the following tasks must be completed as an atomic unit.
- Introduce and test the
dx.types.MatrixReftype - Introduce, lower, and test the
__builtin_la_CreateMatrix()HLSL intrinsic - Introduce, lower, and test the
dx.op.createMatrixDXIL Op
Testing for the above should over at least the following
- Clang codegen for the intrinsic
- DXIL lowering
- Myriad of tests to ensure correct usage/lowering/validation/error messages for the new HLSL type
#7097 is a good example PR of the type of work that needs to be done.
Metadata
Metadata
Assignees
Labels
No labels
Activity
__builtininstrinsics and DXIL Ops #7838damyanp commentedon Oct 28, 2025
From refinement:
There's no mention on the new type in the issue, we want to add the new type simultaneously with adding this op.
[-]Implement Create Matrix `__builtin` and DXIL Op[/-][+]Enable creation of a LinAlg Matrix (type, intrinsic, dxil op)[/+]V-FEXrt commentedon Feb 2, 2026
__builtin_la_CreateMatrix()is no longer a part of the latest spec but the abilty to lower into a dx type (nowdx.types.LinAlgMatrix<mangle>) is still needed.That work is currently on a "feature" branch but but relies on some hacky throwaway code. It can't be merged until the robust version of the higher level is ready to go in
V-FEXrt commentedon Feb 13, 2026
Okay this is merged into main, going to close out this issue
__builtininstrinsics and DXIL Ops #7838