a work-in-progress UCI chess and chess960 engine, with NNUE evaluation trained from zero knowledge starting with random weights

this project is a continuation of my HCE engine Polaris

• standard PVS with quiescence search and iterative deepening
  • aspiration windows
  • futility pruning
  • history
    • capture history
    • 1-ply continuation history (countermove history)
    • 2-ply continuation history (follow-up history)
    • 4-ply continuation history
  • history pruning
  • internal iterative reduction
  • killers (2 per ply)
  • late move reductions
  • late move pruning
  • mate distance pruning
  • multicut
  • nullmove pruning
  • reverse futility pruning
  • probcut
  • SEE move ordering and pruning
  • singular extensions
    • double extensions
    • triple extensions
    • various negative extensions
  • Syzygy tablebase support
• NNUE
  • (768x4->1536)x2->1x8 architecture, horizontally mirrored
  • trained from zero knowledge with reinforcement learning from a randomly-initialised network
• BMI2 attacks in the bmi2 build, otherwise fancy black magic
  • pext/pdep for rooks
  • pext for bishops
• lazy SMP
• static contempt

• make it stronger uwu

vnni512: requires BMI2, AVX-512 and VNNI (Zen 4/Cascade Lake-SP/Rocket Lake and up)
avx512: requires BMI2 and AVX-512 (Skylake-X, Cannon Lake)
avx2-bmi2: requires BMI2 and AVX2 and assumes fast pext and pdep (i.e. no Bulldozer, Piledriver, Steamroller, Excavator, Zen 1, Zen+ or Zen 2)
avx2: requires BMI and AVX2 - primarily useful for pre-Zen 3 AMD CPUs back to Excavator
sse41-popcnt: needs SSE 4.1 and popcnt - for older x64 CPUs

If in doubt, compare the avx2-bmi2 and avx2 binaries and pick the one that's faster. BMI2 will always be faster on Intel CPUs.

Alternatively, build the makefile target native for a binary tuned for your specific CPU (see below)

• If you have an AMD Zen 1 (Ryzen x 1xxx), Zen+ (Ryzen x 2xxx) or Zen 2 (Ryzen x 3xxx) CPU, use the avx2 build even though your CPU supports BMI2. These CPUs implement the BMI2 instructions pext and pdep in microcode, which makes them unusably slow for Stormphrax's purposes.

Requires Make and a competent C++20 compiler that supports LTO. GCC is not currently supported, so the usual compiler is Clang. MSVC explicitly does not work.

> make <BUILD> CXX=<COMPILER>

• replace <COMPILER> with your preferred compiler - for example, clang++ or icpx
  • if not specified, the compiler defaults to clang++
• replace <BUILD> with the binary you wish to build - native/vnni512/avx512/avx2-bmi2/avx2/sse41-popcnt
  • if not specified, the default build is native
• if you wish, you can have Stormphrax include the current git commit hash in its UCI version string - pass COMMIT_HASH=on

By default, the makefile builds binaries with profile-guided optimisation (PGO). To disable this, pass PGO=off. When using Clang with PGO enabled, llvm-profdata must be in your PATH.

Stormphrax uses Fathom for tablebase probing, licensed under the MIT license, and a slightly modified version of incbin for embedding neural network files, under the Unlicense.

Stormphrax is tested on this OpenBench instance - thanks to all the people there, SP would be much weaker without your support :3

In no particular order, these engines have been notable sources of ideas or inspiration:

• Viridithas
• Svart
• Stockfish
• Ethereal
• Berserk
• Weiss
• Koivisto

Stormphrax's current networks are trained with bullet. Previous networks were trained with Marlinflow.

The name "Stormphrax" is a reference to the excellent Edge Chronicles :)