Pilot v2: bf16 tile GEMM (AMX/AVX-512) as an additional ndarray-backed candidate - #5

Merged

AdaWorldAPI merged 8 commits into

mainfrom

claude/ndarray-gemm-pilot-v2-b7f3

Sep 4, 2026

Commits on Sep 4, 2026

Add an ndarray-bf16_tile_gemm-backed AVX-512/AMX f32 GEMM kernel as a second additional candidate
claude
committed
65bd54e
Merge remote-tracking branch 'origin/main' into claude/ndarray-gemm-pilot-v2-b7f3
claude
committed
19f79ba
Gate the bf16 GEMM candidate's ndarray import to x86_64
claude
committed
be9c14d
Merge remote-tracking branch 'origin/main' into claude/ndarray-gemm-pilot-v2-b7f3
claude
committed
f123331
Exclude the bf16 GEMM candidate from automatic dispatch
claude
committed
ac7b332
Skip MMMKernel!'s auto-generated exact-bit tests for lossy kernels
claude
committed
fe9a74c
Merge remote-tracking branch 'origin/main' into claude/ndarray-gemm-pilot-v2-b7f3
claude
committed
61ec53c
Add standalone benchmarks decomposing the bf16-AMX GEMM kernel's gap against the asm baseline into raw tile-primitive throughput versus per-tile conversion/packing overhead.
claude
committed
eed2f66