Skip to content

Luna CE v4.2.0

Latest

Choose a tag to compare

@Spunc595 Spunc595 released this 04 Nov 07:35
· 16 commits to main since this release

Luna CE v4.2.0

Performance release. Search and evaluation are unchanged: the
benchmark node counts are identical to v4.1.1 (depth 18, 1 thread:
590,820 and 980,165), and so are the node counts of every position
used in the speed measurements below.

Faster

  • The Windows executable is built with profile-guided optimisation
    (PGO). It is still a baseline x86-64 build (SSE3 and CMPXCHG16B; no
    POPCNT, LZCNT or BMI1/BMI2 instructions). The profile comes from a
    reproducible self-play workload (scripts/pgo_workload.py); see
    docs/BUILD_PGO.md.
  • On x86-64 CPUs with AVX2 the NNUE accumulator update uses AVX2
    (detected at runtime; CPUs without AVX2 use the SSE2 path, which now
    has its own tests).
  • Move generation writes into a fixed-capacity list on the stack
    instead of allocating vectors. This raises the stack used per search
    ply from about 1 KB to about 3.5 KB; search threads have 2 MiB and
    the highest use we measured was about 170 KB.

Measured speed

Single thread, fixed depth 14, 100 positions, paired and interleaved
runs on one Windows machine (not measured on Linux ARM64 or Android):

  • v4.2.0 Windows build vs v4.1.1: 1.113 times faster (geometric mean,
    95% confidence interval 1.097 to 1.129).
  • Of this, profile-guided optimisation alone: 1.086 (1.070 to 1.101)
    on the final code; the code changes without PGO: about 1.03 (measured
    on an earlier build of the same changes).
  • A build for x86-64-v3 was measured 0.9% slower than the baseline
    build and is not provided.

Notes

  • The Linux ARM64 and Android binaries contain the code changes but are
    not PGO builds; their speed was not measured.
  • docs/PERF_NOTES.md is the working log of the optimisation attempts;
    where it conflicts with the measurements above, the measurements are
    right.
  • Release binaries were built from commit f62c360; the tagged commit (80b1648) differs from it only in documentation files.
  • A 130-game safety match of the Windows executable against v4.1.1
    (100 games at 10+0.1 with one thread and 30 games with two
    threads) had no time losses, illegal moves, crashes or
    disconnections.

Files

  • luna-v4.2.0-windows-x86_64.exe (baseline: SSE3 + CMPXCHG16B, PGO)
  • luna-v4.2.0-linux-aarch64 (needs glibc 2.34 or newer)
  • luna-v4.2.0-android-arm64 (built with NDK r27d, API 21)
  • SHA256SUMS (all three files; verify with sha256sum -c SHA256SUMS)