Skip to content

[VectorDistribute] LDS promotion with async DMA#24509

Open
sommerlukas wants to merge 2 commits into
iree-org:mainfrom
sommerlukas:dma-promotion
Open

[VectorDistribute] LDS promotion with async DMA#24509
sommerlukas wants to merge 2 commits into
iree-org:mainfrom
sommerlukas:dma-promotion

Conversation

@sommerlukas
Copy link
Copy Markdown
Contributor

Extends the GPUVectorAlloc pass to promote operands to LDS using asynchronous DMA operations (e.g., AMD GPU direct-to-LDS).

Based on propagation of the use_global_load_dma propagation type during analysis, the pass inserts LDS allocation and async_dma operations, if the necessary requirements are met. Otherwise, promotion falls back to load/store via registers.

The inserted async_dma operations are lowered and distributed by the pass landed in #24299 further down in the pipeline.

This is part of #23782.

Assisted-by: Claude Code and Codex

Signed-off-by: Lukas Sommer <lukas.sommer@amd.com>
Signed-off-by: Lukas Sommer <lukas.sommer@amd.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant