[Codegen][CPU] Eliminate all-true vector masks after vectorization #18190
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
MacDue
force-pushed
the
eliminate_masks
branch
from
3e1fe9a
to
6dea91f
Compare
This enables an upstream transform that eliminates all true `vector.create_mask` ops. This is particularly beneficial for scalable vectors, which use dynamic tensor types, which results in masks that otherwise would not fold away till much later, preventing some optimizations. Depends on #99314. Signed-off-by: Benjamin Maxwell <benjamin.maxwell@arm.com>
MacDue
force-pushed
the
eliminate_masks
branch
from
6dea91f
to
9f25223
Compare
Abbreviated Benchmark Summary@ commit 06255e62aec4827e77b30d6aed67f1f158154310 (vs. base c71fe1a7aa8a1c360e6418043b2a1165861d72f9) Data-Tiling Comparison TableClick to show
No improved or regressed benchmarks 🏖️ No improved or regressed compilation metrics 🏖️ For more information: |
|
Thanks! LG, but a test would be nice :) |
MacDue
force-pushed
the
eliminate_masks
branch
from
28e0d74
to
09525ba
Compare
Done 👍 |
hanhanW
approved these changes
Aug 13, 2024
Comment on lines
+328
to
+334
| /// Converts from iree_compiler::VscaleRange to vector::VscaleRange. | ||
| static std::optional<vector::VscaleRange> | ||
| toVectorVscaleRange(std::optional<iree_compiler::VscaleRange> vscaleRange) { | ||
| if (!vscaleRange.has_value()) | ||
| return std::nullopt; | ||
| return vector::VscaleRange{vscaleRange->min, vscaleRange->max}; | ||
| } |
There was a problem hiding this comment.
I was confused that why can't we just use vector::VscaleRagnge, then I found that it's added to upstream recently (llvm/llvm-project@9b06e25). Can you help remove the IREE version, and switch to use upstream version? It can be done in a follow-up.
There was a problem hiding this comment.
I'll do this in a follow up PR 👍
c-rhodes
approved these changes
Aug 14, 2024
compiler/src/iree/compiler/Codegen/Common/test/generic_vectorization.mlir
Outdated
Show resolved
Hide resolved
Signed-off-by: Benjamin Maxwell <benjamin.maxwell@arm.com>
This enables an upstream transform that eliminates all true
vector.create_maskops. This is particularly beneficial for scalable vectors, which use dynamic tensor types, which results in masks that otherwise would not fold away till much later, preventing some optimizations.Depends on llvm/llvm-project#99314.