Skip to content

[Kernel]Remove recast iter from sliding window attention kernel. #997

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

Merged
Phil-amd merged 15 commits into from
Aug 11, 2026
Merged

[Kernel]Remove recast iter from sliding window attention kernel. #997

Changes from all commits
Commits
Show all changes
15 commits
Select commit Hold shift + click to select a range
fcd7e15
add sliding window attention kernel
amd-nprotaso Jul 23, 2026
7ffebf8
fix style issues
amd-nprotaso Jul 23, 2026
8aa5718
Merge branch 'main' into main
amd-nprotaso Jul 23, 2026
7989ddf
add swa kernele unit tests
amd-nprotaso Jul 24, 2026
8563859
fix style
amd-nprotaso Jul 24, 2026
237c42f
fix comments
amd-nprotaso Jul 27, 2026
92df85b
fix pr comments
amd-nprotaso Jul 27, 2026
bee5092
remove excess comments
amd-nprotaso Jul 27, 2026
48e9d47
fix comments
amd-nprotaso Jul 27, 2026
a036cd3
fix style
amd-nprotaso Jul 27, 2026
b51e1a9
Performance improvement:
amd-nprotaso Jul 28, 2026
a31b058
Merge branch 'ROCm:main' into main
amd-nprotaso Jul 28, 2026
0dcc0eb
Merge branch 'ROCm:main' into main
amd-nprotaso Aug 7, 2026
2d0bd2d
remove using recast
amd-nprotaso Aug 7, 2026
fef78ce
remove recast
Aug 10, 2026
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
2 changes: 1 addition & 1 deletion kernels/attention/swa_gfx950.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -246,7 +246,7 @@ def load_v_regs(buf):
ST_PER_ROW = 4
crd = fx.make_int_tuple((i32(row_offset), i32(col_in_sub)))
lane_elems = i32(fx.get_scalar(fx.crd2idx(crd, _v_smem_layout)))
base_ptr = fx.recast_iter(bf16, fx.add_offset(smem_ptr, fx.make_int_tuple(lane_elems)))
base_ptr = fx.add_offset(smem_ptr, fx.make_int_tuple(lane_elems))
vreg = [[None] * 4 for _ in range_constexpr(4)]
for i in range_constexpr(4):
for j in range_constexpr(4):
Expand Down
Loading