Skip to content

Use unroll_count(4) for the NAX attention Q@K.T loop #3843

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

Open
wyanzhao wants to merge 1 commit into
base: main
Choose a base branch
from
Open

Use unroll_count(4) for the NAX attention Q@K.T loop #3843

Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
7 changes: 6 additions & 1 deletion mlx/backend/metal/kernels/steel/attn/kernels/steel_attention_nax.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -207,7 +207,12 @@ template <
for (short iq = 0; iq < TQ; iq++) {
STEEL_PRAGMA_UNROLL
for (short ik = 0; ik < TK; ik += 2) {
STEEL_PRAGMA_UNROLL
// Unrolling the head-dim loop by 4 rather than fully lets the
// compiler interleave the next K-tile loads with the running mma
// chain instead of hoisting all TD loads up front; measures +11%
// at head_dim 128 on M5 Max and is a no-op at head_dim 64 (TD == 4,
// already a full unroll).
#pragma clang loop unroll_count(4)
for (short id = 0; id < TD; id++) {
NAXTile<T, 1, 1> Qtile;
NAXTile<T, 2, 1> Ktile;
Expand Down