Allow lhs and rhs to have different type than accumulator/destination. Some hardware like GPUs support natively operations like uint8xuint8xuint32.
Details
Diff Detail
- Repository
- rG LLVM Github Monorepo
Event Timeline
mlir/include/mlir/Dialect/Vector/VectorOps.td | ||
---|---|---|
44 | should lhs/rhs still have same element type, at least? | |
145 | perhaps add a mixed element type example? | |
mlir/test/Dialect/Vector/invalid.mlir | ||
763 | please don't remove the whole test, perhaps now test that acc/result have same type | |
mlir/test/Dialect/Vector/ops.mlir | ||
179 | can't we just leave this out, it seemed unused in the original method already? |
mlir/include/mlir/Dialect/Vector/VectorOps.td | ||
---|---|---|
44 | Yes that makes sense. I added back this constraint. | |
mlir/test/Dialect/Vector/invalid.mlir | ||
763 | I had tried to test acc/result have the same type but it would fail in the parser in a first place as it thinks the acc is used with a different type than declared. I ended up changing this to test that rhs and lhs have the same element type. |
should lhs/rhs still have same element type, at least?