2 months ago
6ecbdfa
Davy* perf(config): measured per-target execution-provider defaults (#170) The five-target dual-preference bench (run 30164564615, spectral-core stable artifact, archived on #170) shows the previous XNNPACK-everywhere default was wrong on three targets: the ORT CPU EP is ~1.6-2.1x FASTER than XNNPACK on Linux x64 (warm 3.01s vs 6.27s), Linux arm64 (2.54s vs 4.61s), and Intel macOS (3.50s vs 5.52s). XNNPACK wins on Apple Silicon (2.15s vs 2.39s on CI; 2.67s vs 5.08s on an M-series dev machine). Windows keeps DirectML: it accelerates when a GPU exists and the measured no-GPU fallback is cost-free (identical numbers to CPU). The platform policy table gains an arch dimension (Apple Silicon vs Intel macOS); defaults become: Apple Silicon -> XNNPACK, Windows -> DirectML, everything else -> CPU. Saved user preferences are preserved unchanged; only the unset/stale-value fallback moves. StemMode and overlap untouched (#173). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * test(config): host-conditional default expectations for all three policy classes The platform-default test asserted the old XNNPACK-everywhere value on non-Windows hosts; it now mirrors the measured policy table (Windows -> DirectML, Apple Silicon -> XNNPACK, else CPU) so it passes on every CI host. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> --------- Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Parentdfcc34a