Repository navigation
perf: optional FFI-backed BLAS backend for multiply/sum #11
Description
The original README roadmap mentioned "native SSE/AVX assembly methods". A realistic modern take: bind libopenblas/cblas via FFI as an optional fast path for multiply() and sum() on large matrices, keeping the pure-PHP loops as the portable fallback (feature-detect the shared library at bootstrap).
Scope: experimental; should ship with a performance-group benchmark comparing both paths, and must not add a hard dependency — the library stays pure-PHP by default.
Activity
lisachenko commentedon Oct 19, 2026
Delivered by #19, which shipped this as a full driver architecture: OpenBLAS over FFI (
cblas_dgemm/daxpy/dscal) as the optional fast path for multiplication, addition, subtraction and scalar operations — plus a CLBlast/OpenCL GPU driver behind the same interface. Availability is feature-detected at runtime with a real 1×1 probe, the pure-PHP array loops remain the portable fallback (autorouting, with mid-operation fallback), system libraries aresuggest-only so there is no hard dependency, andcomposer benchcompares the paths — measured tables are in the README (up to ~×1,900 on 1024² GEMM).Generated by Claude Code