Skip to content

Commit 51d452b

Browse files
tannevaledclaude
andauthored
docs(plan): ppc64le has VSX kernels since v0.2.3; scalar targets unmeasured vs NumPy (#21)
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
1 parent cda7d66 commit 51d452b

1 file changed

Lines changed: 6 additions & 3 deletions

File tree

‎docs/plan-ndarray.md‎

Lines changed: 6 additions & 3 deletions
Original file line numberDiff line numberDiff line change
@@ -121,9 +121,12 @@ suite (plus the per-arch CI jobs already wired in `.github/workflows/ci.yml`).
121121
validated per-arch in CI (bit-identical to the scalar oracle for the
122122
elementwise/sqrt/max ops; tight-tolerance for the sum reduction, whose
123123
lane-parallel grouping is a valid reordering — the same trade-off as
124-
NumPy's pairwise sum). The other four targets keep the validated scalar
125-
oracles plus a scalar 4×4 GEMM micro-kernel over the packed panels, and
126-
still win via packing + cache-blocking + multicore. Split-CI: the pure-Go
124+
NumPy's pairwise sum). Since v0.2.3 ppc64le has VSX kernels too (sum,
125+
dot, sqrt, elementwise, an 8×8 GEMM micro-kernel), through go-asmgen's
126+
VSX encoders. The other targets (riscv64, loong64, s390x, 32-bit) keep the
127+
validated scalar oracles plus a scalar 4×4 GEMM micro-kernel over the
128+
packed panels, with packing + cache-blocking + multicore; they have not
129+
been measured against NumPy. Split-CI: the pure-Go
127130
core + multicore are held to 100% statement coverage; the generated `.s`
128131
is validated by the per-arch native/qemu execution jobs.
129132
- **SIMD `Sqrt`/`Max` and the packed, cache-blocked GEMM — DONE (no longer

0 commit comments

Comments
 (0)