changelog.txt

v0.31.1 (2024-08-11)
~~~~~~~
* Make WDL_mu score type work as intended.
* Fix macos CI builds.

v0.31.0 (2024-06-16)
~~~~~~~
* No changes from rc3.

v0.31.0-rc3 (2024-05-29)
~~~~~~~
* The `WDLDrawRateTarget` option now accepts the value 0 (new default) to retain
  raw WDL values if `WDLCalibrationElo` is set to 0 (default).
* Improvements to the verbose move stats if `WDLEvalObjectivity is used.
* The centipawn score is displayed by default for old nets without WDL output.
* Some build system improvements.

v0.31.0-rc2 (2024-04-16)
~~~~~~~
* Changed cuda compilation options to use `-arch=native` or `-arch=all-major`
  if no specific version is requested, with fallback for older cuda that don't
  support those options.
* Updated android builds to use openblas 0.3.27.
* A few small fixes.

v0.31.0-rc1 (2024-03-25)
~~~~~~~
* The blas, cuda, eigen, metal and onnx backends now have support for multihead
  network architecture and can run BT3/BT4 nets.
* Updated the internal Elo model to better align with regular Elo for human
  players.
* There is a new XLA backend that uses OpenXLA compiler to produce code to
  execute the neural network. See
  <https://github.com/LeelaChessZero/lc0/wiki/XLA-backend> for details. Related
  are new leela2onnx options to output the HLO format that XLA understands.
* There is a vastly simplified lc0 interface available by renaming the
  executable to `lc0simple`.
* The backends can now suggest a minibatch size to the search, this is enabled
  by `--minibatch-size=0` (the new default).
* If the cudnn backend detected an unsupported network architecture it will
  switch to the cuda backend.
* Two new selfplay options enable value and policy tournaments. A policy
  tournament is using a single node policy to select the move to play, while a
  value tournament searches all possible moves at depth 1 to select the one with
  the best q.
* While it is easy to get a single node policy evaluation (`go nodes 1` using
  uci), there was no simple way to get the effect of a value only evalaution, so
  the `--value-only` option was added.
* Button uci options were implemented and a button to clear the tree was added
  (as hidden option).
* Support for the uci `go mate` option was added.
* The rescorer can now be built from the lc0 code base instead of a separate
  branch.
* A dicrete onnx layernorm implementation was added to get around a onnxruntime
  bug with directml - this has some overhead so it is only enabled for onnx-dml
  and can be switched off with the `alt_layernorm=false` backend option.
* The `--onnx2pytoch` option was added to leela2onnx to generate pytorch
  compatible models.
* There is a cuda `min_batch` backend option to reduce non-determinism with
  small batches.
* New options were added to onnx2leela to fix tf exported onnx models.
* The onnx backend can now be built for amd's rocm.
* Fixed a bug where the Contempt effect on eval was too low for nets with
  natively higher draw rates.
* Made the WDL Rescale sharpness limit configurable via the `--wdl-max-s` hidden
  option.
* The search task workers can be set automatically, to either 0 for cpu backends
  or up to 4 depending on the number of cpu cores. This is enabled by
  `--task-workers=-1` (the new default).
* Several assorted fixes and code cleanups.

v0.30.0 (2023-07-21)
~~~~~~~
* WDL_mu score type is now the default and the mlh-threshold default was
  changed from 0 to 0.8.
* Fixes for contempt with infinite search and pondering and for the wdl display
  when pondering.

v0.30.0-rc2 (2023-06-15)
~~~~~~~
* WDL conversion for more realistic WDL score and contempt. Adds an Elo based
  WDL transformation of the NN value head output. Helps with more accurate play
  at high level (WDL sharpening), more aggressive play against weaker opponents
  and draw avoiding openings (contempt), piece odds play. There will be a blog
  post soon explaining in detail how it works.
* A new score type `WDL_mu` which follows the new eval convention, where +1.00
  means 50% white win chance.
* Simplified to a single `--draw-score` parameter, adjusting the draw score from
  white's perspective: 0 gives standard scoring, -1 gives Armageddon scoring.
* Updated describenet for new net architectures.
* Added a `first-move-bonus` option to the legacy time manager, to accompany
  `book-ply-bonus` for shallow openings.
* Changed mlh threshold effect to create a smooth transition.
* Revised 'simple' time manager.
* A new spinlock implementation (selected with `--search-spin-backoff`) to help
  with many cpu threads (e.g. 128 threads), obviously for cpu backends only.
* Some assorted fixes and code cleanups.

v0.30.0-rc1 (2023-04-24)
~~~~~~~
* Support for networks with attention body and smolgen added to blas, cuda,
  metal and onnx backends.
* Persistent L2 cache optimization for the cuda backend. Use the
  `cache_opt=true` backend option to turn it on.
* Some performance improvements for the cuda, onnx and blas backends.
* Added the `threads` backend option to onnx, defaults to 0 (let the
  onnxruntime decide) except for onnx-cpu that defaults to 1.
* The onnx-dml package now includes a `directml.dll` installation script.
* Some users experienced memory issues with onnx-dml, so the defaults were
  changed. This may affect performance, in which case you can use the `steps=8`
  backend option to get the old behavior.
* The Python bindings are available as a package, see the README for
  instructions.
* Some assorted fixes and code cleanups.

v0.29.0 (2022-12-13)
~~~~~~~
* Updated onednn version to the latest one.

v0.29.0-rc1 (2022-12-09)
~~~~~~~
* New metal backend for apple systems. This is now the default backend for
  macos builds.
* New onnx-dml backend to use DirectML under windows, has better net
  compatibility than dx12 and is faster than opencl. See the README for use
  instructions, a separate download of the DirectML dll is required.
* Full attention policy support in cuda, cudnn, metal, onnx, blas, dnnl, and
  eigen backends.
* Partial attention policy support in onednn backend (good enough for T79).
* Now the onnx backends can use fp16 when running with a network file (not with
  .onnx model files). This is the default for onnx-cuda and onnx-dml, can be
  switched on or off with by setting the `fp16` backend option to `true` or
  `false` respectively.
* The onednn package comes with a dnnl compiled to allow running on an intel gpu
  by adding `gpu=0` to the backend options.
* The default net is now 791556 for most backends except opencl and dx12 that
  get 753723 (as they lack attention policy support).
* Support for using pgn book with long lines in training: selfplay can start at
  a random point in the book.
* New "simple" time manager.
* Support for double Fischer random chess (dfrc).
* Added TC-dependent output to the backendbench assistant.
* Starting with this version, the check backend compares policy for valid moves
  after softmax.
* Some assorted fixes and code cleanups.

v0.29.0-rc0 (2022-04-03)
~~~~~~~
* Initial support for attention policy, only cuda backend and partially in
  blas/dnnl/eigen (good enough for T79).
* Non multigather (legacy) search code and `--multigather` option are removed.
* 15b default net is now 753723.
* The onnx backend now allows selecting gpu to use.
* Improved error messages for unsupported network files.
* Some assorted fixes.

v0.28.2 (2021-12-13)
~~~~~~~
* No changes from v0.28.1-rc1 as the v0.28.1 release was botched.

v0.28.1 (2021-12-12)
~~~~~~~
* No changes from rc1.

v0.28.1-rc1 (2021-12-05)
~~~~~~~
* Improved cuda performance for 512 filter networks on Amprere GPUs.
* Several fixes for the onnx backend.
* Command line options for network file conversion to/from onnx.
* Documentation updates.
* Correctness fixes for rescorer support functions.

v0.28.0 (2021-08-25)
~~~~~~~
* Fixed an issue with small third-party nets on the cuda/cudnn backends.
* Minor tweak to the default task-workers for the cpu packages.

v0.28.0-rc2 (2021-08-20)
~~~~~~~
* The cuda backend option multi_stream is now off by default. You should
  consider setting it to on if you have a recent gpu with a lot of vram.
* The default time manager is back to "legacy".
* Updated default parameters.
* Newer and stronger nets are included in the release packages.
* Added support for onnx network files and runtime with the "onnx" backend.
* Several bug and stability fixes.

v0.28.0-rc1 (2021-06-16)
~~~~~~~
* Multigather is now made the default (and also improved). Some search settings
  have changed meaning, so if you have modified values please discard them.
  Specifically, `max-collision-events`, `max-collision-visits` and
 `max-out-of-order-evals-factor` have changed default values, but other options
  also affect the search. Similarly, check that your gui is not caching the old
  values.
* Performance improvements for the cuda/cudnn backends.
* Support for policy focus during training.
* Larger/stronger 15b default net for all packages except android, blas and dnnl
  that get a new 10b network.
* The distributed binaries come with the mimalloc memory allocator for better
  performance when a large tree has to be destroyed (e.g. after an unexpected
  move).
* The `legacy` time manager will use more time for the first move after a long
  book line.
* The `--preload` command line flag will initialize the backend and load the
  network during startup.
* A 'fen' command was added as a UCI extension to print the current position.
* Experimental onednn backend for recent intel cpus and gpus.

v0.27.0 (2021-02-21)
~~~~~~~
* A better value for the backendbench Clippy threshold.

v0.27.0-rc2 (2021-02-18)
~~~~~~~
* Fix additional cases where 'invalid move' could be incorrectly reported.
* Replace WDL softmax in cudnn backend with same implementation as cuda
  backend. This fixes some inaccuracy issues that were causing training
  data to be rejected at a fairly low frequency.
* Ensure that training data Q/D pairs form valid WDL targets even if there
  is accumulated drift in calculation.
* Fix for the calculation of the 'best q is proven' bit in training data.
* Multiple fixes for timelosses and infinite instamoving in smooth time
  manager. Smooth time manager now made default after these fixes.

v0.27.0-rc1 (2021-02-06)
~~~~~~~
* Fix a bug which meant `position ... moves ...` didn't work if the moves
  went off the end of the existing tree.

v0.27.0-rc0 (2021-02-06)
~~~~~~~
* Multigather search inspired by Ceres.
* V6 training format with additional info for training experiments.
* Updated default search parameters.
* A better algorithm for the backendbench assistant.
* Terminate search early if only 1 move isn't a proven loss.
* Various build system changes.

v0.26.3 (2020-10-10)
~~~~~~~
* Increased maximum value of TempDecayMoves.

v0.26.3-rc2 (2020-10-03)
~~~~~~~
* Fix for uninitialized variable that led to crashes with the cudnn backend.
* Correct windows support for systems with more than 64 threads.
* A new package is built for the `cuda` backend with cuda 11.1. The old cuda
  package is renamed to `cudnn`.

v0.26.3-rc1 (2020-09-28)
~~~~~~~
* Residual block fusion optimization for cudnn backend, that depends on
  `custom_winograd=true`. Enabled by default only for networks with up to 384
  filters in fp16 mode and never in fp32 mode. Default can be overridden with
  `--backend-opts=res_block_fusing=false` to disable (or `=true` to enable).
* New experimental cuda backend without cudnn dependency (`cuda-auto`, `cuda`
  and `cuda-fp16` are available).

v0.26.2 (2020-08-31)
~~~~~~~
* No changes from rc1.

v0.26.2-rc1 (2020-08-28)
~~~~~~~~~~~
* Repetitions in the search tree are marked as draws, to explore more promising
  lines. Enabled by default (except in selfplay mode) use
  `--two-fold-draws=false` to disable.
* Syzygy tablebase files can now be used in selfplay. Still need to add
  adjudication support before we can consider using this for training.
* Default net updated to 703810.
* Fix for book with CR/LF line endings.
* Updated Eigen wrap to use new download link.

v0.26.1 (2020-07-15)
~~~~~~~
* Fix a bug where invalid openings-pgn settings would result in the book
  being ignored rather than used.
* Add support for compressed book files.

v0.26.0 (2020-07-03)
~~~~~~~
* No changes from rc1.

v0.26.0-rc1 (2020-06-29)
~~~~~~~~~~~

* Verbose move stats now includes a line for the root node itself.
* Added optional `alphazero` time manager type for fixed fraction of
  remaining time per move.
* The WL score is now tracked with double pecision to improve accuracy
  during very long search.
* Fix for a performance bug when playing from tablebase position with
  tablebases enabled and the PV move was changing frequently.
* Illegal searchmove restrictions will now be ignored rather than crash.
* Policy is cleared for terminal losses to encourage better quality MLH
  estimates by reducing how many visits a move that will not be selected
  (unless all other options are equally bad) receives.
* Smart pruning will now cause leela to play immediately once mate score has
  been declared.
* Fix an issue where sometimes the pv reported wouldn't match the move that
  would be selected at that moment.
* Improvement for logic for when to disable custom_winograd optimization to
  avoid running out of video ram.
* `--show-hidden` can now be specified after `--help` and still work.
* Performance tuning for populating the policy into nodes after nn eval
  completes.
* Enable custom optimized SE paths for nets with 384 filters when using the
  custom_winograd=false path.
* Updates to zlib/gtest/eigen when included via meson wrap.
* Added build option to build python bindings to the lc0 engine.
* Only show the git hash in uci name if not a release tag build.
* Add `--nps-limit` option to artificially reduce nps to make for easier
  opponent or whatever other reason you want.
* Fixed a bug where search tree shape could be affected even when the
  `--smart-pruning-factor` setting was 0.
* Changed the search logic to find the lc0.config file if left on the default
  value.
* Changed the search logic to find network files in autodiscover mode.
* Changed the logic to determine the default location for training games
  generated by selfplay in training mode.
* Changed the logic to decide where to look for the opencl backend tuning
  settings file.
* Android binaries published by appveyor are now stripped.
* Build can now use system installed eigen if available.
* When nodes in the tree get proven terminal, parents are updated as if they
  had always been terminal. This allows for faster convergence on more
  accurate MLH estimates amongst other details.
* Removed shortsightedness and logit-q options that have not found a reliable
  use case.
* Fixed a bug where m_effect calculated as part of S in verbose move stats was
  not consistent with the value used in search itself.
* Added 'pro' mode as an alternative to `--show-hidden` for UCI hosts that do
  not support command line arguments. Simply rename the lc0 binary to include
  'pro' in order to enable.
* `backendbench` now has a `--clippy` option to try and auto suggest which
  batch size is a good idea.
* The demux backend now splits the batch into equal sizes based on the number
  of threads that demux is using rather than number of backends. By default
  this is no change as usually there is 1 thread per backend. But it allows
  to more easily use demux against a blas backend sending one chunk per core.
* Added support for new training input variants canonical_hectoplies and
  canonical_hectoplies_armageddon.
* Fixed a bug where if the network search paths for autodiscover contain files
  which lc0 cannot open it would error out rather than continuing on to other
  files.
* Blas backends no longer have a `blas_cores` option, as it never seemed useful
  compared to running more threads at a higher level.
* `--help-md` option removed as it was deemed not very useful.
* Updated to the latest version of dnnl for the dnnl build.
* Selfplay mode now supports per color settings in addition to per player
  settings. Per player settings have higher priority if there is a conflict.
  This will be used as part of armageddon training.
* Added a new experimental backend type: `recordreplay`. This allows to
  record the output of a backend under a particular search and then replay it
  back again later. Theoretically this lets you simulate a CPU bottlenecked
  environment but still use a search tree that is a match for what might be a
  GPU bottlenecked environment. In practice there are a lot of corner cases
  where replay is not reliable yet.  At a minimum you must disable prefetch.
* During search the node tree is occasionally compacted to reduce cache misses
  during the search tree walk. New option `--solid-tree-threshold` can be used
  to adjust how aggressive this optimization is.  Note that very small values
  can cause very large growth in ram usage and are not a good idea. The default
  value is a little conservative, if you have plenty of spare ram it can be
  good to decrease it a bit.
* Small performance optimization for windows build with MLH enabled.
* Meson configuration changed to build with LTO by default. Note that meson
  does not always configure visual studio project files to apply this
  correctly on windows.
* The included net in appveyor builds is now 703350. This network supports MLH
  although the default MLH parameters are still threshold 1.0 which means it
  will not trigger without parameter adjustment.
* New backend option to explicitly override the net details and force MLH
  disabled. If you weren't going to use MLH anyway, this may give a tiny nps
  increase.
* New flag `--show-movesleft` (or `UCI_ShowMovesLeft` for UCI hosts that
  support it) will cause movesleft (in moves) to be reported in the uci info
  messages. Only works with networks that have MLH enabled.
* More sensible default values for MLH are in. Note that threshold is still
  1.0 by default, so that will still need to be configured to enable it.
* The `smooth-experimental` time manager has been renamed `smooth` and support
  added to increase search time whenever the best N does not correspond with
  the move with best utility estimate. `legacy` remains the default for now
  as `smooth` has only been tuned for short time controls and evidence suggests
  it doesn't scale with these defaults.
* Selfplay mode now supports a logfile parameter just like normal mode.
* Reinstated the 4 billion visit limit on search to avoid overflowing counters
  and causing very strange behavior to occur.
* Performance optimization to make tree walk faster by ensuring that node
  edges are always sorted by policy. This has some very small side effects to
  do with tiebreaks in search no longer always being dominated by movegen
  order.
* Appveyor built blas and Android binaries now default to minibatch size 1
  and prefetch 0, which should be much better than the normal GPU optimized
  defaults. Note this *only* affects Appveyor built binaries.
* The included client in Windows Appveyor releases is now v27 and is named
  `lc0-training-client.exe` instead of `client.exe`.

v0.25.1 (2020-04-30)
~~~~~~~

* Fixed some issues with cudnn backend on the 16xx GTX models and also for
  low memory devices with large network files where the new optimizations
  could result in out of memory errors.
* Added a workaround for a cutechess issue where reporting depth 0 during
  instamoves causes it to ignore our info message.

v0.25.0 (2020-04-28)
~~~~~~~

* Relax strictness for complete standard fens in uci and opening books. Fen
  must still be standard, but default values will be substituted for sections
  that are missing.
* Restore some backwards compatibility in cudnn backends that was lost with
  the addition of the new convolution implementation. It is also on by default
  for more scenarios, although still off for fp16 on RTX gpus.
* Small logic fix for nps smoothing in the new optional experimental time
  manager.

v0.25.0-rc2 (2020-04-23)
~~~~~~~~~~~

* Increased upper limit for maximum collision events.
* Allow negative values for some of the extended moves left head parameters.
* Fix a critical bug in training data generation for input type 3.
* Fix for switching between positions in uci mode that only differ by 50 move
  rule in initial fen.
* Some refinements of certainty propagation.
* Better support for c++17 implementations that are missing charconv.
* Option to more accurately apply time management for uci hosts using
  cuteseal or similar timing techniques.
* Fix for selfplay mode to allow exactly book length total games.
* Fix for selfplay opening books with castling moves starting from chess960 fens.
* Add build option to override nvcc compiler.
* Improved validity checking for some uci input parameters.
* Updated the Q to CP conversion formula to better fit recent T60 net outputs to
  expectations.
* Add a new experimental time manager.
* Bug fix for the Q+U in verbose move stats. It is now called S: and contains
  the total score, including any moves left based effect if applicable.
* New temperature decay option to allow to delay the start of decay.
* All temperature options have been hidden by default.
* New optional cuda backend convolution implementation. Off by default for
  cudnn-fp16 until an issue with cublas performance on some gpus is resolved.

v0.25.0-rc1 (2020-04-09)
~~~~~~~~~~~

* Now requires a c++17 supporting compilation environment to build.
* Support for Moves Left Head based networks. Includes options to adjust search
  to favour shorter/longer wins/losses based on the moves left head output.
* Mate score reporting is now possible, and move selection will prefer shorter
  mates over longer ones when they are proven.
* Training now outputs v5 format data. This passes the moves left information
  back to training. This also includes support for multiple sub formats,
  including the existing standard, a new variant which can encode FRC960
  castling, and also a further extension of that which tries to make training
  data cannonical, so there aren't multiple positions that are trivially
  equivalent with different network inputs.
* Benchmark now includes a suite of 34 positions to test by default instead of
  just start position.
* Tensorflow backend works once more, almost just as hard to compile as it used
  to be though.
* `--noise` flag is gone, use `--noise-epsilon=0.25` to get the old behavior.
* Some bug fixes related to drawscore.
* Selfplay mode now defaults to the same value as match play for
  `--root-has-own-cpuct-params` (true).
* Some advanced time management parameters are now accessed via the new
  `--time-manager` parameter instead of individual parameters.
* Windows build script has been modernized.
* Separate Eigen backend option for CPU.
* Random backend no longer requires a network.
* Random backend supports producing training data of any input format sub type.
* Integer parameters now give better error messages when given invalid values.

v0.24.1 (2020-03-15)
~~~~~~~

* Fix issues where logitq was being passed as drawscore and logitq wasn't
  passed to some GetQ calls. Causing major performance issues when either
  setting was non-default.

v0.24.0 (2020-03-11)
~~~~~~~

* New parameter `--max-out-of-order-evals-factor` replaces
  `--max-out-of-order-evals` that was introduced in v0.24.0-rc3 and provides
  the factor to multiply the maximum batch size to set maximum number
  out-of-order evals per batch. The default value of 1.0 keeps the behavior
  of previous releases.
* Bug fix for hangs with very early stop command from non-conforming UCI hosts.

v0.24.0-rc3 (2020-03-08)
~~~~~~~~~~~

* New parameter `--max-out-of-order-evals` to set maximum number out-of-order
  evals per batch (was equal to the batch size before).
* It's now possible to embed networks into the binary. It allows easier builds
  of .apk for Android.
* New parameter `--smart-pruning-minimum-batches` to only allow smart pruning
  to stop after at least k batches, preventing insta-moves on slow backends.

v0.24.0-rc2 (2020-03-01)
~~~~~~~~~~~

* All releases are now bundled with network id591226 (and the file date is old
  enough so it has a lower priority than networks that you already may have
  in your directory).
* Added a 'backendbench' mode to benchmark NN evaluation performance without
  search.
* Android builds are added to the official releases.

v0.24.0-rc1 (2020-02-23)
~~~~~~~~~~~

* Introduced DirectX12 backend.
* Optimized Cpuct/FPU parameters are now default.
* There is now a separate set of CPuct parameters for the root node.
* Support of running selfplay games from an opening book.
* It's possible to adjust draw score from 0 to something else.
* There is a new --max-concurrent-seachers parameter (default is 1) which
  helps with thread congestion at the beginning of the search.
* Cache fullness is not reported in UCI info line by default anymore.
* Removed libproto dependency.

v0.23.3 (2020-02-18)
~~~~~~~

* Fix a bug in time management which sometimes led to insta-moves in long time
  control.

v0.23.2 (2019-12-31)
~~~~~~~

* Fixed a bug where odd length openings had reversed training data results in
  selfplay.
* Fixed a bug where zero length training games could be generated due to
  discard pile containing positions that were already considered end of game.
* Add cudnn-auto backend.

v0.23.1 (2019-12-03)
~~~~~~~

* Fixed a bug with Lc0 crashing sometimes during match phase of training game
  generation.
* Release packages now include CUDNN version without DLLs bundled.

v0.23.0 (2019-12-01)
~~~~~~~

* Fixed the order of BLAS options so that Eigen is lower priority, to match
  assumption in check_opencl patch introduced in v0.23.0-rc2.

v0.23.0-rc2 (2019-11-27)
~~~~~~~~~~~

* Fixes in nps and time reporting during search.
* Introduced DNNL BLAS build for modern CPUs in addition to OpenBLAS.
* Build fixes on MacOS without OpenCL.
* Fixed smart pruning and KLDGain trying to stop search in `go infinite` mode.
* OpenCL package now has check_opencl tool to find computation behaves sanely.
* Fixed a bug in interoperation of shortsighteness and certainty propagation.

v0.23.0-rc1 (2019-11-21)
~~~~~~~~~~~

* Support for Fischer Random Chess (`UCI_Chess960` option to enable FRC-style
  castling). Also added support for FRC-compatible weight files, but no training
  code yet.
* New option `--logit-q` (UCI: `LogitQ`). Changes subtree selection algorithm a
  bit, possibly making it stronger (experimental, default off).
* Lc0 now reports WDL score. To enable it, use `--show-wdl` command-line
  argument or `UCI_ShowWdl` UCI option.
* Added "Badgame split" mode during the training. After the engine makes
  inferior move due to temperature, the game is branched and later the game is
  replayed from the position of the branch.
* Added experimental `--short-sightedness` (UCI: `ShortSightedness`) parameter.
  Treats longer variations as more "drawish".
* Lc0 can now open Fat Fritz weight files.
* Time management code refactoring. No functional changes, but will make time
  management changes easier.
* Lc0 logo is now printed in red! \o/
* Command line argument `-v` is now short for `--verbose-move-stats`.
* Errors in `--backend-opts` parameter syntax are now reported.
* The most basic version of "certainty propagation" feature (actually without
  "propagation"). If the engine sees checkmate, it plays it!
  (before it could play other good move).
* Benchmark mode no longer supports smart pruning.
* Various small changes: hidden options to control Dirichlet noise, floating
  point optimizations, Better error reporting if there is exception in worker
  thread, better error messages in CUDA backend.

v0.22.0 (2019-08-05)
~~~~~~~

(no changes)

v0.22.0-rc1 (2019-08-03)
~~~~~~~~~~~

* Remove softmax calculation from backends and apply it after filtering for
  illegal moves to ensure spurious outputs on illegal moves don't reduce (or
  entirely remove) the quality of the policy values on the legal moves.
* Fix for blas backend allocation bug with small network sizes.
* The blas backend can be built with eigen - the result is reasonably optimized
  for the build machine.
* Other small tweaks piled up in master branch.


v0.21.4 (2019-07-28)
~~~~~~~~~~~~~~~~~~~~

* A fix for crashes that can occur during use of sticky-endgames.
* Change the false positive value reported when in wdl style resign and display
  average nodes per move as part of tournament stats in selfplay mode.

v0.21.3 (2019-07-21)
~~~~~~~

* Fix for potential memory corruption/crash in using small networks or using the
  wdl head with cuda backends. (#892)
* Fix for building with newer versions of meson. (#904)

v0.21.2 (2019-06-09)
~~~~~~~

* Divide by a slightly smaller divisor to truncate to +/-12800. (#880)

v0.21.2-rc3 (2019-06-08)
~~~~~~~~~~~

* Centipawn conversion (#860)

v0.21.2-rc2 (2019-05-22)
~~~~~~~~~~~

* 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)

v0.21.2-rc1 (2019-05-05)
~~~~~~~~~~~

* 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)

v0.21.0-rc2 (2019-03-06)
~~~~~~~~~~~

* 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 (2019-02-16)
~~~~~~~~~~~

* 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 (2019-02-01)
~~~~~~~~~~~

* 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 (2019-01-27)
~~~~~~~~~~~

* 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 (2019-01-07)
~~~~~~~~~~~

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

v0.20.1-rc3 (2019-01-07)
~~~~~~~~~~~

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

v0.20.1-rc2 (2019-01-05)
~~~~~~~~~~~

* 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 (2019-01-01)
~~~~~~~~~~~

* 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 (2019-01-01)
~~~~~~~~~~~

* no lto builds by default (#625)

v0.20.0-rc2 (2018-12-24)
~~~~~~~~~~~

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

v0.20.0-rc1 (2018-12-22)
~~~~~~~~~~~

* 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 (2018-12-10)
~~~~~~~

(no changes relative to v0.19.1-rc2)

v0.19.1-rc2 (2018-12-07)
~~~~~~~~~~~

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

v0.19.1-rc1 (2018-12-06)
~~~~~~~~~~~

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

v0.19.0 (2018-11-19)
~~~~~~~

* remove Wait() from EngineController::Stop() (#522)

v0.19.0-rc5 (2018-11-17)
~~~~~~~~~~~

* OpenCL: replace thread_local with a resource pool. (#516)
* optional wtime and btime (#515)
* Make convolve1 work with workgroup size of 128 (#514)
* adjust average depth calculation for multivisits (#510)

v0.19.0-rc4 (2018-11-12)
~~~~~~~~~~~

* Microseconds have 6 digits, not 3! (#505)
* use bestmove_is_sent_ for Search::IsSearchActive() (#502)

v0.19.0-rc3 (2018-11-07)
~~~~~~~~~~~

* Fix OpenCL tuner always loading the first saved tuning (#491)
* Do not show warning when ComputeBlocking() takes too much time. (#494)
* Output microseconds in log rather than milliseconds. (#495)
* Add benchmark features (#483)
* Fix EncodePositionForNN test failure (#490)

v0.19.0-rc2 (2018-11-03)
~~~~~~~~~~~

* Version v0.19.0-rc1 reported it's version as v0.19.0-dev
  Therefore v0.19.0-rc2 is released with this issue fixed.

v0.19.0-rc1 (2018-11-03)
~~~~~~~~~~~

* Search algorithm changes

  When visiting terminal nodes and collisions, instead of counting that as one
  visit, estimate how many subsequent visits will also go to the same node, and
  do a batch update.

  That should slightly improve nps near terminal nodes and in multithread
  configurations. Command line parameters that control that:

  --max-collision-events – number of collision events allowed per batch.
    Default is 32. This parameter is roughly equivalent to
    --allowed-node-collisions in v0.18.

  --max-collision-visits – total number of estimated collisions per NN batch.
    Default is 9999.

* Time management

  Multiple changes have been done to make Leela track used time more precisely
  (particularly, the moment when to start timer is now much closer to the moment
  GUIs start timer).

  For smart pruning, Leela's timer only starts when the first batch comes from
  NN eval. That should help against instamoves, especially on non-even GPUs.

  Also Leela stops the search quicker now when it sees that time is up (it could
  continue the search for hundreds of milliseconds after that, which caused time
  trouble if opponent moves very fast).

  Those changes should help a lot in ultra-bullet configurations.

* Better logging

  Much more information is outputted now to the log file. That will allow us to
  easier diagnose problems if they occur. To have debug file written, add a
  command line option:

  --logfile=/path/to/logfile

  (or short option "-l /path/to/logfile", or corresponding UCI option "LogFile")

  It's recommended to always have logging on, to make it easier to report bugs
  when it happens.

* Configuration parameters change

  Large part of parameter handling has been reworked. As the result:

  All UCI parameters have been changed to have more "classical" look.
    E.g. was "Network weights file path", became "WeightsFile".

  Much more detailed help is shown than before when you run
    ./lc0 --help

  Some flags have been renamed, e.g.
    --futile-move-aversion
    is renamed back to
    --smart-pruning-factor.

  After setting a parameter (using command line parameter or uci setoption
    command), uci command "uci" shows updated result. That way you can check the
    current option values.

  Some command-line and UCI options are hidden now. Use --show-hidden command
    line parameter to unhide them. E.g.
    ./lc0 --show-hidden --help

  Also, in selfplay mode the per player configuration format has been changed
  (although probably noone knew that anyway):
    Was: ./lc0 selfplay player1: --movetime=14
    Became: ./lc0 selfplay --player1.movetime=14

* Other

  "go depth X" uci command now causes search to stop when depth information in
  uci info line reaches X. Not that it makes much sense for it to work this way,
  but at least it's better than noting.

  Network file size can now be larger than 64MB.

  There is now an experimental flag --ramlimit-mb. The engine tries to estimate
  how much memory it uses and stops search when tree size (plus cache size)
  reaches RAM limit. The estimation is very rough. We'll see how it performs and
  improve estimation later.
  In situations when search cannot be stopped (`go infinite` or ponder),
  `bestmove` is not automatically outputted. Instead, search stops progress and
  outputs warning.

  Benchmark mode has been implemented. Run run, use the following command line:
    ./lc0 benchmark
  This feature is pretty basic in the current version, but will be expanded later.

  As Leela plays much weaker in positions without history, it now is able to
  synthesize it and do not blunder in custom FEN positions. There is a
  --history-fill flag for it. Setting it to "no" disables the feature, setting
  to "fen_only" (default) enables it for all positions except chess start
  position, and setting it to "always" enables it even for startpos.

  Instead of output current win estimation as centipawn score approximation,
  Leela can how show it's raw score. A flag that controls that is --score-type.
  Possible values:
    - centipawn (default) – approximate the win rate in centipawns, like Leela
      always did.
    - win_percentage – value from 0 to 100.0 which represents expected score in
      percents.
    - Q – the same, but scales from -100.0 to 100.0 rather than from 0 to 100.0

v0.18.1 (2018-10-02)
~~~~~~~

* Fix for falling into threefold repetition in a winning endgame tablebase position.


v0.18.0 (2018-09-30)
~~~~~~~

* No changes from rc2 except the version.


v0.18.0-rc2 (2018-09-26)
~~~~~~~~~~~

* Severe bug fixed: Race condition when out-of-order-eval was enabled (and it
  was enabled by default)

* Windows 32-bit builds are now possible (CPU only for now)


v0.18.0-rc1 (2018-09-24)
~~~~~~~~~~~

KNOWN BUG!

* We have credible reports that in some rare cases Lc0 crashes!
  However, we were not able to reproduce it reliably. If you see the crash,
  please report to devs! What seems to increase crash probability:
  - Very short move time (milliseconds)
  - Proximity to a checkmate (happens 1-3 moves before the checkmate)


New features:

* Endgame tablebases support! Both WDL and DTZ now.

* Added MultiPv support.


Time management changes:

* Introduced --immediate-time-use flag. Yes, yet another time management
  flag. Posible values are between 0.0 and 1.0. Setting it closer to
  1.0 makes Leela use time saved from futile search aversion earlier.

* Some time management parameters were changed:
  - Slowmover is 1.0 now (was 2.4)
  - Immediate-time-use is 0.6 now (didn't exist before, so was 0.0)

* Fixed a bug, because of which futile search aversion tolerance was incorrectly
  applied, which resulted in instamoves.

* Now search stops immediately when it runs out of budgeted time.
  Should help against timeouts, especially on slow backends (e.g. BLAS).

* Move overhead now is a fixed time, doesn't depend on number of remaining
  moves.


Other:

* Out of order eval is on by default. That brings slight nps improvement.

* Default FPU reduction is 1.2 now (was 0.9)

* Cudnn backend now has max_batch parameter.
  (can be set for example like this --backend-opts=max_batch=100).
  This is needed for lower end GPUs that didn't have enough VRAM for a buffer
  of size 1024. Make sure that this setting is not lower than --minibatch-size.

* Small memory usage optimizations.

* Engine name in UCI response is shorter now. Fritz chess UI should be able
  to work with Leela now

* Added flag --temp-visit-offset, will allow to offset temperature during
  training.

* Command line and UCI parameter values are now checked for validity.

* You can now build for older processors that don't support the popcnt
  instruction by passing -Dpopcnt=false to meson when building.

* 32-bit build is possible now. CPU only and we were only able to build it
  in Linux for now, including Raspberry Pi.

* Threading issue which caused crash in heavily multithreaded environment
  with slow backends was fixed.


v0.17.0 (2018-08-27)
~~~~~~~

No changes from rc2 except the version.


v0.17.0-rc2 (2018-08-21)
~~~~~~~~~~~

* Fixed a bug, that rule50 value was located in wrong place in a training data.
* OpenCL uses much less VRAM now.
* Default OpenCL batch size is 16 now (was 1).
* Default time management related configuration was tweaked:
  --futile-move-aversion is 1.33 now (was 1.47)
  --slowmover is 2.4 now (was 2.6)


v0.17.0-rc1 (2018-08-19)
~~~~~~~~~~~

New visible features:
* Implemented ponder support.
* Tablebases are supported now (only WDL probe for now).
  Command line parameter is
  --syzygy-paths=/path/to/syzygy/
* Old smart pruning flag is gone. Instead there is
  --futile-search-aversion flag.
  --futile-search-aversion=0 is equivalent to old --no-smart-pruning.
  --futile-search-aversion=1 is equivalent to old --smart-pruning.
  Now default is 1.47, which means that engine will sometimes decide to
  stop search earlier even when there is theoretical chance (but not very
  probable) that best move decision could be changed if allowed to think more.
* Lc0 now supports configuration files. Options can be listed there instead of
  command line flags / uci params.
  Config should be named lc0.config and located in the same directory as lc0.
  Should list one command line option per line, with '--' in the beginning
  being optional, for example:

     syzygy-paths=/path/to/syzygy/

* In uci info, "depth" is now average depth rather than full depth
  (which was 4 all the time).
  Also, depth values do not include reused tree, only nodes visited during the
  current search session.
* --sticky-checkmates experimental flag (default off), supposed to find shorter
  checkmate sequences.
* More features in backend "check".


Performance optimizations:
* Release windows executables are built with "whole program optimization".
* Added --out-of-order-eval flag (default is off).
  Switching it on makes cached/terminal nodes higher priority, which increases
  nps.
* OpenCL backend now supports batches (up to 5x speedup!)
* Performance optimizations for BLAS backend.
* Total visited policy (for FPU reduction) is now cached.
* Values of priors (P) are stored now as 16-bit float rather than 32-bit float,
  that saves considerable amount of RAM.


Bugfixes:
* Fixed en passant detection bug which caused the position after pawn moving by
  two squares not counted towards threefold repetition even if en passant was
  not possible.
* Fixed the bug which caused --cache-history-length for values 2..7 work the
  same as --cache-history-length=1.
  This is fixed, but default is temporarily changed to --cache-history-length=1
  during play. (For training games, it's 7)


Removed features:
* Backpropagation beta / backpropagation gamma parameters have been removed.


Other changes:
* Release lc0-windows-cuda.zip package now contains NVdia CUDA and cuDNN .dlls.


v0.16.0 (2018-07-20)
~~~~~~~

* Fully switched to official releases! No more https://crem.xyz/lc0/
* Fixed a bug when pv display and smart pruning didn't sometimes work properly
  after tree reuse.
* Format of protobuf network files was changed.
* Autodiscovery of protobuf based network files works now.


lc0-win-20180715-cuda92-cudnn714
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

* Support of new format of network files (needed for lc0 launch on main
  training server)
* Fixed hang/poor performance in the beginning of search when there are many
  threads. (Happened on linux only though).
* Memory footprint is reduced a bit. (~-60 bytes per node)

lc0-win-20180711-cuda92-cudnn714
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

* Edge-node separation introduced a bug that smart pruning didn't work. That's
  fixed.
* Changed options parsing so that --backend-opts=cudnn-fp16 is now possible.
* Performance fixes (mostly for slowness introduced by edge-node separation).

lc0-win-20180708-cuda92-cudnn714
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

* Mutex contention has been reduced (but locking mutex more rarely).
  Helps a lot with many threads running. Especially recommended to check with
  multi-GPU configuration.
* Memory usage reduced at least 2x (probably more).
* cudnn backend crashed on large batches (>800) that's fixed.
  There is still a limit of batch size 1024 though.
* (not in cudnn build, but for completeness)
  Fixed NN computation with BLAS backend, it had up to 5% error before that.
* Default time budgeting params have been changed again! (not by mach this time)
  --slowmover=1.95
  --time-curve-peak=26.2
  --time-curve-left-width=82
  --time-curve-right-width=74

lc0-win-20180701-cuda92-cudnn714
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

* fp16-based computation for very modern NVidia GPUs!
  May reduce precision a bit, but should be compensated by nps boost.
  Enable with --backend=cudnn-fp16 flag
* V is now not stored in nodes (a bit less RAM used while thinking)
* (not in cudnn build, but listing for completeness) blas batching support.

lc0-win-20180629-cuda92-cudnn714
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

* Default time budgeting parameters have been changed (again!):
  --slowmover=1.93
  --time-curve-peak=26
  --time-curve-left-width=67
  --time-curve-right-width=76
* When generating training games, the engine could confuse client by sending
  corrupted output. That's fixed.

lc0-win-20180624-cuda92-cudnn714
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

* Default time budgeting parameters have been changed:
  --slowmover=2.13  (was 1.8)
  --time-curve-peak=22.0  (was 41.0)
  --time-curve-left-width=450.0  (was 1000.0)
  --time-curve-right-width=30.0  (was 39.5)
* During training game generation, the engine is able to send resign statistics.


lc0-win-20180622-cuda92-cudnn714
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

* Time budged allocation has been changed, it allocates more time to early
  stages of the game.
  Graphs are here: https://github.com/LeelaChessZero/lc0/pull/59
  Slowmover value has so be recalibrated, and default value was changed from 2.2 to 1.8.
* Fixed a race condition in cache prefetch code. Realistically it hardly every
  occured before though.

lc0-win-20180619-cuda92-cudnn714-00
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

* Fix a bug instroduced in version 20180609 which caused the engine to miss checkmates sometimes.

lc0-win-20180614-cuda92-cudnn714
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

* "go searchmoves" uci command is now supported
* It's possible now to disable tree reuse in training games
* Few improvements for random backend
* Lc0 now shows version in uci response
* Analyzer mode has been removed
* extra-virtual-loss has been removed
* Implemented resign (for training games)

lc0-win-20180609-cuda92-cudnn714-01
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

* In addition to --backpropagate-gamma, there is also --backpropagate-beta!
  Default is 1.0.

lc0-win-20180609-cuda92-cudnn714-00
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

Visible changes:
* Experimental changes from 20180604 are now default.
* Memory footprint is reduced by 8 bytes per visible node (+ ~240 bytes in
  invisible nodes per visible)
* Introduced --backpropagate-gamma flag.
  Default is 1.0. There are rumours that reducing it to 0.75 improves play.
* Extra-virtual-loss parameter has been removed.
* Quotes in backend-opts parameter were not parsed properly, that's fixed.


lc0-win-20180604-cuda92-cudnn714-experimental
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

Visible changes:
* Experimental default settings:
  cPUCT: 3.4
  FPU reduction: 0.9
  policy Softmax: 2.2

* Fix memory leak when GUI doesn't ever issue `isready` uci command.


lc0-win-20180602-cuda92-cudnn714-00
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

Visible changes:
* cPUCT is now 3.1 by default instead of 1.2 (or what it was before)
* Fixed Batch normalization epsilon in tensorflow backend (but noone uses tensorflow anyway)
* Periodically (every 5 seconds) output "uci info" even if bestmove/depth doesn't change.
* Memory management is redone so that node release happens after "bestmove" and "isready", rather than after "position" uci command.
  That garbage collection could take tens of milliseconds and chess GUI already started timer at that point.
  Memory management is always fragile, so fresh crashes and memory leaks are possible.

Invisible changes:
* Store castlings again as e1g1 and not e1h1. Fixes a bug that tree was not reused after castling.