Add GEMM + AllScatter benchmark sweep with multi-shape tiling analysis, optimal tile guidance, iris vectorization hints, roofline analysis, kernel comm-compute overlap optimization, comprehensive MNK roofline sweep, strong/weak scaling analysis, deep G...#411

Draft

Copilot wants to merge 14 commits intomainfrom

copilot/run-gemm-allscatter-benchmarks

Commits on Mar 2, 2026

Commits on Mar 3, 2026

Add tiling parameter sweep script, results markdown, and chart for GEMM+AllScatter on MI300X

Copilot
and
mawad-amd
committed
Add multi-shape tiling sweep results, guidance charts, and optimal tiling analysis for MI300X

Copilot
and
mawad-amd
committed
Add iris vectorization hints to GEMM+AllScatter kernel and assembly/perf analysis

Copilot
and
mawad-amd
committed
Merge branch 'main' into copilot/run-gemm-allscatter-benchmarks
mawad-amd
authored
Apply Ruff auto-fixes
github-actions[bot]
committed
Add roofline analysis chart and markdown: SoL, assembly root-cause, and tuning experiments

Copilot
and
mawad-amd
committed
Optimize GEMM+AllScatter: scatter from accumulator registers (ctx.put → ctx.store)

Copilot
and
mawad-amd
committed
Add comprehensive roofline chart: 60-point sweep across 20 shapes × 3 configs with hardware limits

Copilot
and
mawad-amd
committed
Add scaling analysis: GEMM-only vs GEMM+scatter roofline with strong/weak scaling breakdown

Copilot
and
mawad-amd
committed
Add deep GEMM tuning sweep: BLK_K=128 gives +20% at M=256, expose num_warps/mfma/kpack knobs in matmul_wrapper

Copilot
and
mawad-amd
committed
Add 1200-point roofline scatter sweep: X=M*N*K, Y=TFLOPS, 40 kernel configs, world_size=8

Copilot
and
mawad-amd
committed