x86_64 CPU parity: AVX2+FMA quantized matmul kernels and kernel-machine routing#89

Merged

gabewillen merged 10 commits into

mainfrom

x86-avx2-quant-kernels-v2

Jul 2, 2026

Commits on Jul 2, 2026

feat: ship v1.27 ryzen avx2 fma support
gabewillen
committed
feat: view-sliced parallel matmul cutover with cross-engine comparison lanes
gabewillen
committed
feat: x86_64 AVX2+FMA quantized matmul kernels (Q4_K, Q4_0/Q4_1/Q5_0/Q8_0) and FMA F32 GEMM
gabewillen
committed
feat: route domain matmuls through kernel machine; add f32 matrix-vector AVX2+FMA kernel
gabewillen
committed
test: cover all whisper linear-variant instantiations and embeddings matrix matmul path
gabewillen
committed
chore: refresh quality gates timing data
gabewillen
committed
fix: gate host AVX flags on build-host CPU probe; run rejected wavefront lanes inline
gabewillen
committed
docs: document sign_epi8 -128 precondition on dot_i8_pairs_i32x8_avx2
gabewillen
committed
fix: bind reserved computes to the stored reservation's tensor lifecycle
gabewillen
committed
fix: route duplicate graph actors through serial wavefront dispatch
gabewillen
committed