Skip to content

[x86] Register tiling for non-contraction reductions - #296

Merged
adam-smnk merged 1 commit into
llvm:mainfrom
adam-smnk:x86-register-reduction-tiling
Oct 15, 2026
Merged

adam-smnk merged 1 commit into
llvm:mainfrom
adam-smnk:x86-register-reduction-tiling

Conversation

@adam-smnk

Copy link
Copy Markdown
Member

Adds a target-aware heuristic used by the register parallel, reduction and unroll strategies to drive tiling on non-contraction reductions.

General reductions were never tiling anchors, so they only received propagated parallel tiles and vectorized into full-sized multi-reductions.

Row and column reductions now vectorize into register-sized chunks with independent accumulators, avoiding register spills.

Assisted-by: Claude

@adam-smnk
adam-smnk requested a review from rengolin October 8, 2026 13:56
@adam-smnk

Copy link
Copy Markdown
Member Author

@rengolin Any thoughts?

@adam-smnk
adam-smnk force-pushed the x86-register-reduction-tiling branch from 70edd3b to 8d2efe4 Compare October 14, 2026 12:34
Adds a target-aware heuristic used by the register parallel, reduction
and unroll strategies to drive tiling on non-contraction reductions.

General reductions were never tiling anchors, so they only received
propagated parallel tiles and vectorized into full-sized
multi-reductions.

Row and column reductions now vectorize into register-sized chunks
with independent accumulators, avoiding register spills.

Assisted-by: Claude
@adam-smnk
adam-smnk force-pushed the x86-register-reduction-tiling branch from 8d2efe4 to 9f27bf0 Compare October 15, 2026 08:12
@adam-smnk
adam-smnk merged commit da54528 into llvm:main Oct 15, 2026
3 checks passed
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.

2 participants