Skip to content

[FlyDSL] [gfx950] Fix 8-wave a8w8 GEMM reading LDS before the last tile lands #6368

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
coderfeli merged 10 commits into from
Oct 26, 2026
Merged

[FlyDSL] [gfx950] Fix 8-wave a8w8 GEMM reading LDS before the last tile lands #6368

Changes from all commits
Commits
Show all changes
10 commits
Select commit Hold shift + click to select a range
2463e41
[FlyDSL] [gfx950] Fix 8-wave a8w8 GEMM dropping LDS writes at the K-s…
Oct 10, 2026
65d1037
[FlyDSL] [gfx950] Note the divergent prologue barrier in the 8-wave GEMM
Oct 10, 2026
9b0de95
[FlyDSL] [gfx950] Correct the note on the prologue's half-wave barrier
Oct 10, 2026
d7c2e39
Move the fix into the tail, where the dependency is
Oct 16, 2026
f6fa9ac
[FlyDSL] [gfx950] Trim comments in the 8-wave GEMM sync fix
Oct 19, 2026
e8a6929
[FlyDSL] [gfx950] Drop the dedicated 8-wave GEMM test
Oct 19, 2026
a9db887
Merge remote-tracking branch 'origin/main' into fix-8wave-tail
Oct 19, 2026
d2c1981
[FlyDSL] [gfx950] Use rocdl.s_waitcnt instead of inline asm for the S…
Oct 20, 2026
8842a0a
[FlyDSL] [gfx950] Close the 8-wave GEMM's half-wave barrier stagger b…
Oct 20, 2026
27d2909
[FlyDSL] [gfx950] Balance the epilogue barrier on every tile; drop th…
Oct 20, 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
6 changes: 4 additions & 2 deletions aiter/ops/flydsl/kernels/gemm_a8w8_8wave.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -446,6 +446,7 @@ def kernel_gemm(
b0_frag = b_s2r.load(b_cur0, preshuffled=b_preshuffled)
a0_frag = a_s2r.load(a_cur0)
a_g2s.load(a_next1, A1_gl_offset + (k + 1) * BLOCK_K)
rocdl.s_waitcnt(lgkmcnt=0)
rocdl.s_barrier()

c00_frag = mfma.call(a0_frag, b0_frag, c00_frag)
Expand Down Expand Up @@ -477,7 +478,7 @@ def kernel_gemm(
k = K_ITERS - 2
b0_frag = b_s2r.load(b_cur0, preshuffled=b_preshuffled)
a0_frag = a_s2r.load(a_cur0)
rocdl.s_barrier()
wait_barrier(0)

c00_frag = mfma.call(a0_frag, b0_frag, c00_frag)

Expand Down Expand Up @@ -523,7 +524,8 @@ def kernel_gemm(
c10_frag = mfma.call(a1_frag, b0_frag, c10_frag, set_prio=False)
c11_frag = mfma.call(a1_frag, b1_frag, c11_frag, set_prio=False)
rocdl.s_setprio(0)
rocdl.s_barrier()
if wave_m == 0:
rocdl.s_barrier()

# Scale and store back to gmem
wave_n_offset = wave_n * (N_TILES_B * 16)
Expand Down