Skip to content

Dialect/GPU

-iree-gpu-combine-barrier-regionslink

Combines iree_gpu.barrier_region ops

-iree-gpu-concretize-mma-shapeslink

Expands the inner dimensions of iree_gpu.multi_mma ops to match the thread layout

Optionslink

-concretize-inputs : Expand the inner dimensions for the lhs and rhs operands of the multi_mma ops.
-concretize-result : Expand the inner dimensions for the acc operand of the multi_mma ops.

-iree-gpu-distribute-mma-to-laneslink

Distributes iree_gpu.multi_mma ops to lanes

-iree-gpu-lower-opslink

Post bufferization lowerings of iree_gpu ops before late lowerings

-iree-gpu-unroll-to-intrinsicslink

Unrolls iree_gpu.multi_mma ops to their inner vector size.

-iree-gpu-vectorize-opslink

Vectorizes then lowers a few iree_gpu ops before vectorization.