Skip to content

Conversation

@Groverkss
Copy link
Contributor

This PR adds a serial tiling level, meant to tile parallel iterator types with serial loops (as opposed to ones mapped to workgroup/thread/subgroup/lane/etc.). Previously, VectorDistribute was using "reduction" tiling level to map parallel iterators which cannot be tiled to workgroups to "reduction" tiling level and hacking around it.

The serial tiling level also uses a scf.forall without a mapping, which also hints that all writes are independent.

Comment on lines 65 to +69
if (tilingLevel != IREE::GPU::TilingLevel::Reduction &&
tilingLevel != IREE::GPU::TilingLevel::Thread &&
tilingLevel != IREE::GPU::TilingLevel::Subgroup &&
tilingLevel != IREE::GPU::TilingLevel::PartialReduction) {
tilingLevel != IREE::GPU::TilingLevel::PartialReduction &&
tilingLevel != IREE::GPU::TilingLevel::Serial) {
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

use llvm::is_contained

@kuhar kuhar requested a review from efric October 30, 2025 16:22
Copy link
Contributor

@qedawkins qedawkins left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

LGTM % Jakub's nit

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.

3 participants