What problem does it solve? Porting CK a8w8 fp8 block-scale GEMM kernels to FlyDSL on AMD MI355X (gfx950) fails parity because the hardware scaled-MFMA path only supports E8M0 power-of-two scales, while CK uses arbitrary fp32 per-block scales. This Skill provides a validated recipe to restore bit-level parity and then pick the correct shape-dependent performance lever. ## Core Features & Use Cases - Universal parity fix: Apply fp32 block scales in software after an unscaled fp8 MFMA, achieving err=0 / cos=1.0 against the CK oracle instead of silently rounding scales through E8M0 hardware. - Shape-dependent kernel selection: Choose between a 4-wave blockscale-preshuffle core (wide-N with XCD swizzle, or K-light with tile_n=256/tile_k=128) and an 8-wave ping-pong cluster core (narrow-N deep-K), with measured speedups from 1.05x to 1.45x. - Validated bottleneck diagnosis method: Use ablation, promote-free ceiling twins, roofline checks, and K-sweep phase decomposition to identify the true limiting factor before tuning. - Use Case: A profiled model shows a gemm_a8w8_blockscale hotspot on MI355X; apply this recipe to select the right FlyDSL core by M/N/K shape, pass parity first, then measure same-session A/B speedup versus CK. ## Quick Start Ask the agent to rewrite the CK fp8 a8w8 blockscale GEMM hotspot to FlyDSL on gfx950 using the software fp32 post-MFMA scale path and the shape-matched core recipe.
