Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

[CPU][DT] Select proper vec/unroll sizes for vecmat/matvec codegen #15421

Merged
merged 2 commits into from
Dec 1, 2023
+18 −7
Merged

[CPU][DT] Select proper vec/unroll sizes for vecmat/matvec codegen #15421

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
46e4dd7
Rebase on top of Benoit's changes
dcaballe Nov 14, 2023
dcf3256
Review comments
dcaballe Dec 1, 2023
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
25 changes: 18 additions & 7 deletions compiler/src/iree/compiler/Codegen/Common/CPU/CPUMaterializeEncodingPass.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -161,13 +161,24 @@ enumerateMatmulTileX86_64(EncodingUser user, TypeRange elementTypes,
// reconsider when taking advantage of native f16/bf16 arithmetic when the
// accumulator itself is f16/bf16.
if (hasFeature(target, "+avx512f")) {
return {
TileMxNxK{16, 16, 1}, // Aim to use VFMADD* (zmm).
TileMxNxK{8, 16, 1}, // Truncation of the above.
TileMxNxK{4, 16, 1}, // Truncation of the above.
TileMxNxK{2, 16, 1}, // Truncation of the above.
TileMxNxK{1, 16, 1}, // Truncation of the above.
};
if (hasUkernel(target)) {
return {
TileMxNxK{16, 16, 1}, // Aim to use VFMADD* (zmm).
TileMxNxK{8, 16, 1}, // Truncation of the above.
TileMxNxK{4, 16, 1}, // Truncation of the above.
TileMxNxK{2, 16, 1}, // Truncation of the above.
TileMxNxK{1, 16, 1}, // Truncation of the above.
};
} else {
// Code generation tile sizes.
return {
TileMxNxK{16, 16, 1}, // Aim to use VFMADD* (zmm).
TileMxNxK{8, 32, 1}, // Use same number of accumulators.
TileMxNxK{4, 64, 1}, // Use same number of accumulators.
TileMxNxK{2, 64, 1}, // Use half the number of accumulators.
TileMxNxK{1, 128, 1}, // Use half the number of accumulators.
};
Comment on lines +173 to +180
Copy link
Contributor

Choose a reason for hiding this comment

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

It looks like the comment you added is gone after you rebase it?

Copy link
Contributor

Choose a reason for hiding this comment

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

I''m talking about the review comment before your force-push: #15421 (comment)

Copy link
Contributor Author

Choose a reason for hiding this comment

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

Not sure but I addressed the feedback provided.

Copy link
Contributor

Choose a reason for hiding this comment

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

I think it was not just // Code generation tile sizes. before force-push.. You added a comment about "why they have different tile sizes comparing to ukernels.", but it does not show up after force-push..

}
}
if (hasFeature(target, "+avx")) {
// Note: for good performance, most +avx users will also want to add
Expand Down
Loading