changelog.txt

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.