v0.21.2-rc2

Changes

• Add 320 and 352 channel support for fused SE layer (#855)
• SE layer fix when not using fused kernel (#852)
• Fp16 nchw for cudnn-fp16 backend (support GTX 16xx GPUs) (#849)
• Install lc0 on openSUSE - Full documentation, Install script and links to experimental RPM packages (#675)

v0.21.2-rc1

• Make --sticky-endgames on by default (still off in training) (#844)
• update download links in README (#842)
• Recalibrate centipawn formula (#841)
• Also make parents Terminal if any move is a win or all moves are loss or draw. (#822)
• Use parent Q as a default score instead of 0 for unvisited pv. (#828)
• Add stop command to selfplay interactive mode to allow for graceful exit. (#810)
• Increased hard limit on batch size in opencl backend to 32 (#807)

v0.21.1

• FPU can now be independently controlled at root vs in general using the new AtRoot variants. Default value for StrategyAtRoot is 'same' which means it uses the values from the normal parameters. The reduction parameter has been removed - strategy reduction now uses the value parameter as the reduction amount. NOTE: If you are using a UCI host that remembers settings over restarts (like cutechess) be careful to ensure that when upgrading to this new version you reset settings back to defaults - or you may get a bad combination of fpu settings.
• TempVisitOffset option now allows you to specify values less than -1, and its documentation has been fixed.
• A small performance improvement in movegen.
• Windows packages should include v22 of the Client.

v0.21.0

• Change version number to indicate final. No other changes since v0.21.0-rc2.

v0.21.0-rc2

• Add support for cudnn7.0 (#717)
• Informative Tournament Stats (#698)
• Memory leak fix cuda backend (#747)
• cudnn-fp16 fallback path for unusual se-ratios. (#739)
• Cudnn 7.4.2 in packaged binary and warning for using old cudnn with new gpu (#741)
• Move mode specific options to end of help. (#745)
• LogLiveStats hidden option (#754)
• Optional markdown support for help output (#769)
• Improved folding of batch norm into weights and biases - fixes negative gamma bug. (#779)

v0.21.0-rc1

Major new features in this release are support for WDL value head, and convolution direct output (AZ-style) policy head.

• Check Syzygy tablebase file sizes for corruption (#690)
• search for nvcc on the path first (#709)
• AZ-style policy head support (#712)
• Implement V4TrainingData (#722)
• WDL value head support (#635)
• Add option for doing kldgain thresholding rather than absolute visit limiting (#721)
• Easily run latest releases of lc0 and client using NVIDIA docker (#621)
• Add WDL style resign option. (#724)
• Add a uniform output option for random backend to support a0 seed data style (#725)
• Fix c hw switching in cudnn-fp16 mode with convolution policy head. (#729)
• misc (non-functional) changes to cudnn backend (#731)
• handle 64 filter SE networks (#624)

v0.20.2

Changes

(relative to v0.20.2-rc1)

• Favor winning moves that minimize DTZ to reduce shuffling by assuming repeated position by default (#708)
• Print cuda and gpu info, warn if mismatches are noticed (#711)

v0.20.2-rc1

Changes

• no terminal multivisits (#683)
• better fix for issue 651 (#693)
• Changed output of --help flag to stdout rather than stderr (#687)
• Movegen speedup via magic bitboards (#640)
• modify default benchmark setting to run for 10 seconds (#681)
• Fix incorrect index in OpenCL Winograd output transform (#676)
• Update OpenCL (#655)

v0.20.1

#Changes

vs v0.20.1-rc3:

• Change to atomic for cache capacity. (#665)

vs 0.20.0:

• Search algorithm performance optimizations.
• Time management logic has been also optimized.
• Fixed pondering with movetime bug.
• Fixed a few potential problems in source code.

v0.20.1-rc3

Change

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