v0.20.1-rc3

Change

• Remove ffast-math from the default flags (#661)

v0.20.1-rc2

Changes

• Don't use Winograd for 1x1 conv. (#659)
• Fix issues with pondering and search limits. (#658)
• Check for zero capacity in cache (#648)
• fix undefined behavior in DiscoverWeightsFile() (#650)
• fix fastmath.h undefined behavior and clean it up (#643)

v0.20.1-rc1

Changes

• Search algorithm performance optimizations.
• Time management logic has been also optimized.

Full commit log:

• Simplify movestogo approximator to use median residual time. (#634)
• Replace time curve logic with movestogo approximator. (#271)
• Cache best edge to improve PickNodeToExtend performance. (#619)
• fix building with tensorflow 1.12 (#626)
• Minor changes to src/chess (#606)
• make uci search parameters the defaults ones (#609)
• Preallocate nodes in advance of their need to avoid the allocation being behind a mutex. (#613)
• imrpove meson error when no backends enabled (#614)
• allow building with the mklml library as an mkl alternative (#612)
• Only build the history up if we are actually going to extend the position. (#607)
• fix warning (#604)

v0.20.0

Changes

vs v0.20.0-rc2

• no lto builds by default (#625)

vs 0.19.1

• Squeeze-and-Excitation Networks are now supported! (lc0.org/se)
• Older text network files are no longer supported.
• Various performance fixes (most major being having fast approximate math
  functions).
• For systems with multiple GPUs, in addition to "multiplexing" backend
  we now also have "demux" backend and "roundrobin" backend.
• Compiler settings tweaks (use VS2017 for windows builds, always have LTO
  enabled, windows releases have PGO enabled).
• Benchmark mode has more options now (e.g. movetime) and saner defaults.
• Added an option to prevent engine to resign too early (used in training).
• Fixed a bug when number of visits could be too high in collision nodes.
  The fix is pretty hacky, there will be better fix later.
• 32-bit version compiles again.

v0.20.0-rc2

Only one change in this release:

• Fix for demux backend to match cuda expected threading model for computations. (#605)

v0.20.0-rc1

Changes

• Squeeze-and-Excitation Networks are now supported! (lc0.org/se)
• Older text network files are no longer supported.
• Various performance fixes (most major being having fast approximate math
  functions).
• For systems with multiple GPUs, in addition to "multiplexing" backend
  we now also have "demux" backend and "roundrobin" backend.
• Compiler settings tweaks (use VS2017 for windows builds, always have LTO
  enabled, windows releases have PGO enabled).
• Benchmark mode has more options now (e.g. movetime) and saner defaults.
• Added an option to prevent engine to resign too early (used in training).
• Fixed a bug when number of visits could be too high in collision nodes.
  The fix is pretty hacky, there will be better fix later.
• 32-bit version compiles again.

v0.19.1.1

Changes

Relative to v0.19.0

• Parameters from AlphaZero paper were introduced
• Do not load network on isready
• Fixed non-working nodes limit.

Relative to v0.19.1

• Correct version displayed (v0.19.1)

v0.19.1

Changes

Relative to v0.19.0

• Parameters from AlphaZero paper were introduced
• Do not load network on isready
• Fixed non-working nodes limit.

Relative to v0.19.1-rc2

(no changes)

v0.19.1-rc2

Changes

• Temperature and FPU related params. (#568)
• Rework Cpuct related params. (#567)

See also blog post.

v0.19.1-rc1

Changelog

• Updated cpuct formula from alphazero paper. (#563)
• remove UpdateFromUciOptions() from EnsureReady() (#558)
• revert IsSearchActive() and better fix for one of #500 crashes (#555)