[solinzby1]

…ilogues overlap (helps small shapes)

2,port aiter pr6368 to fix race issue

Shape (M×N×K)	HEAD µs	now µs	Delta
256×256×256	7.05	5.94	-15.7%
256×256×512	9.11	8.09	-11.2%
256×256×1024	12.93	12.86	-0.5%
256×256×2048	20.11	19.08	-5.1%
256×256×4096	34.42	33.53	-2.6%
512×256×512	9.38	8.36	-10.9%
512×512×512	9.66	8.67	-10.2%
512×512×1024	12.98	13.10	+0.9%
512×512×2048	20.27	19.25	-5.0%
512×512×4096	34.70	33.86	-2.4%
768×768×1024	13.24	12.55	-5.2%
1024×512×1024	13.12	12.68	-3.4%
1024×1024×512	9.80	9.48	-3.3%
1024×1024×1024	13.31	12.62	-5.2%
1024×1024×2048	20.64	19.62	-4.9%
1024×1024×4096	35.28	34.31	-2.7%
1536×1536×2048	21.40	20.49	-4.3%
2048×1024×2048	21.21	20.21	-4.7%
2048×2048×1024	14.81	14.09	-4.9%
2048×2048×2048	22.55	21.72	-3.7%
2048×2048×4096	38.09	37.31	-2.0%
256×4096×4096	35.27	34.37	-2.6%
256×7168×2048	20.96	19.95	-4.8%
512×4096×7168	60.29	59.00	-2.1%
geomean			-4.9%

[Copilot]

🔵 Needs a closer look

Cross-wave synchronization changes need device validation of the K=256 zero-main-loop path before safe approval.

1 open issue

• Add gfx950 K=256 regression for all layouts and repeated launches · New

What changed in this PR

Updates the MXFP8 eight-wave GEMM kernel’s synchronization to address shared-memory races and improve small-shape epilogue overlap.

Changes:

• Waits for shared-memory reads before the refill barrier.
• Drains outstanding loads at the tail.
• Balances the final barrier between wave groups.

File	Description
kernels/​gemm/​mxfp8_gemm_8wave.py	Adjusts waits and barriers in the GEMM pipeline.

_{Review effort: Balanced}