Skip to content

Commit 7bd3832

Browse files
tannevaledclaude
andcommitted
docs/perf: record GEMM parity with OpenBLAS + the elementwise SIMD wins
Re-measured the whole table back-to-back on the debian arm64 Tart VM (numpy 2.2.4, OpenBLAS 0.3.29-pthread confirmed via /proc/self/maps), 2026-06: * MatMul vs tuned OpenBLAS: n=1024 now ~0.99x (PARITY, ~203 GFLOP/s, samples 0.96-1.00), n=512 ~0.97x — up from the prior 0.76x. The lever was the by-lane FMLA micro-kernel (commit before this); software pipelining, a wider 6x8 tile, and MC/KC/NC retuning were all prototyped and measured to NOT help on this core, and are documented as such so they aren't re-attempted. Only small n=256 still trails (0.67x) on fixed per-call overhead, not throughput. * Add/Sub/Mul/Div: new SIMD kernels -> alloc form wins ~2x at large n; the no-alloc *Into form wins 1.8-2.4x at all n. * Sqrt: packed FSQRT V.2D on arm64 -> 2.8x at 4M, SqrtInto 1.8x small. * The only sub-parity rows left are the small-n ALLOC form of Add/Mul/Sqrt: a Go-allocator ceiling (make zeroes 8 KiB ~900 ns vs numpy's ~150 ns free-list; no unzeroed-alloc escape, pooling would break result ownership). The *Into parity path the brief asked for wins there; documented precisely, not faked. Correctness re-verified on all 6 arches (native arm64 + qemu amd64/riscv64/ loong64/ppc64le/s390x-big-endian), go test -race clean, 100% coverage, go vet clean across all GOARCH. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
1 parent d5463f4 commit 7bd3832

1 file changed

Lines changed: 146 additions & 89 deletions

File tree

0 commit comments

Comments
 (0)