Merge pull request LeelaChessZero#8 from Ergodice/uncertainty-weighting

Uncertainty weighting

Author: Ergodice

.circleci/config.yml

@@ -7,9 +7,7 @@ jobs:
       - checkout
       - run:
           name: "Pull Submodules"
-          command: |
-            git submodule init
-            git submodule update --remote
+          command: git submodule update --init
       - run:
           name: Update Meson
           command: pip3 install --upgrade meson==0.58.1
@@ -31,15 +29,15 @@ jobs:
       - checkout
       - run:
           name: "Pull Submodules"
-          command: |
-            git submodule init
-            git submodule update --remote
+          command: git submodule update --init
       - run:
           name: Install build tools
           command: |
             pip3 install meson==0.63
             pip3 install ninja
-            brew install ispc
+            curl -LJOs https://github.com/ispc/ispc/releases/download/v1.21.0/ispc-v1.21.0-macOS.universal.tar.gz
+            tar xf ispc-v1.21.0-macOS.universal.tar.gz
+            cp ispc-v1.21.0-macOS.universal/bin/ispc /usr/local/bin
       - run:
           name: Build lc0
           command: |

README.md

@@ -4,39 +4,83 @@
 # Lc0
 
 Lc0 is a UCI-compliant chess engine designed to play chess via neural network, specifically those of the [LeelaChessZero project](https://lczero.org).
-This is an experimental repository for testing new features to the Lc0 chess engine. Updates are made frequently so there may be bugs. Please report any issues you find to the Lc0 Discord. All elo measurements were calculated on A100.
+This is an experimental repository for testing new features to the Lc0 chess engine. Updates are made frequently so there may be bugs. Please report any issues you find to the Lc0 Discord. All elo measurements were calculated on A100 with unbalanced human openings.
 Many of these features are taken from the Katago engine. A detailed description of these methods can be found [here](https://github.com/lightvector/KataGo/blob/master/docs/KataGoMethods.md). A list of the improvements follows.
 
 
 ## New features in this fork
 
+### Node reporting
+
+The nps statistic used to represent the number of playouts per second, which may not be what the user wants. The new `--reported-nodes` option specifies what the node count and nps statistic represent. 
+The default is `nodes` for LowNodes. The other options are `queries` for neural network queries and `playouts` or `legacy` for playouts (what we used to use).
+
+
 ### 50 move rule caching
 
 The first 64 plies out of 100 are partitioned into 8 equally sized buckets. Before a position is queried for NN evaluation, the 50 move rule ply is checked. If the bucket containing the position already has nodes, the eval is copied from the one with the most visits.
-The speedup can be anywhere from 5% to 50% depending on how transposition-heavy the position is. The gain was measured at 20 elo on STC.
+The speedup can be anywhere from 5% to 50% depending on how transposition-heavy the position is. The gain was measured at 20 elo on STC. The nodes per second statistic is now calculated by the number of true nodes (called LowNode in the code) rather than edges (technically playouts, called Node in the code) so the reported value may be lower than on previous dag versions.
+
+This feature is enabled by default and can be disabled by specifying `--move-rule-bucketing=false` in the config.
+
+### Multiple output heads
+
+The BT3 generation of nets now has multiple value and policy heads that you can choose among. You always want the `winner` head for value. There are two policy heads, `vanilla` and `optimistic`.
+The vanilla head is similar to what we used before, and the optimistic head upweights moves that the net believes may be exceptionally strong, an idea taken from the Katago engine. The optimistic head is recommended as it adds roughly 20 elo at LTC. The Katago team observed that this improvement is similar to increasing the net size by roughly 40%.
+
+The head you choose is specified in the backend options. For example, if you have a single GPU and want to use the optimistic policy head and winner value head, then you can use this can easily be specified as `--backend-opts=policy_head=vanilla,value_head=winner`.
+
+If you have are using more than one GPU then you have to set up the backend for each GPU. An example is
+`--backend-opts=(gpu0,policy_head=optimistic,value_head=winner),(gpu1,policy_head=optimistic,value_head=winner)`.
+
+### Uncertainty Weighting
+
+Identical to the [Katago implementation](https://github.com/lightvector/KataGo/blob/master/docs/KataGoMethods.md#uncertainty-weighted-mcts-playouts), but the error uses only the winrate. The new parameters are
+
+```
+--uncertainty-weighting-cap
+--uncertainty-weighting-coefficient
+--uncertainty-weighting-exponent
+--use-uncertainty-weighting
+```
+
+
+You must set `--use-uncertainty-weighting=true` to turn it on. With BT3, the gain was 20 elo at LTC with
+
+```
+--uncertainty-weighting-cap=1.03
+--uncertainty-weighting-coefficient=0.13
+--uncertainty-weighting-exponent=-0.88
+--use-uncertainty-weighting=true
+```
+
+
+
 
 ### CPUCT Utility Variance Scaling
 
 Identical to the [Katago implementation](https://github.com/lightvector/KataGo/blob/master/docs/KataGoMethods.md#dynamic-variance-scaled-cpuct). The new parameters are
 
 ```
-CPuctUtilityStdevPrior 
-CPuctUtilityStdevScale 
-CPuctUtilityStdevPriorWeight
+--cpuct-utility-stdev-prior 
+--cpuct-utility-stdev-scale
+--cpuct-utility-stdev-prior-weight
+--use-variance-scaling
 ```
 
-and the tuned values are
+and the tuned values (excluding the prior weight) for masterkni's T1 are
 
 ```
-CPuct 2.3097
-FpuValue 0.5619
-CPuctUtilityStdevPrior 0.2289
-CPuctUtilityStdevScale 0.3437
-CPuctUtilityStdevPriorWeight 10.0 (default, not tuned)
-WDLCalibrationElo 3400
+--cpuct=2.3097
+--fpu-value=0.5619
+--cpuct-utility-stdev-prior=0.2289
+--cpuct-utility-stdev-scale=0.3437
+--cpuct-utility-stdev-prior-weight=10.0
+--wdl-calibration-elo=3400
+--use-variance-scaling=true
 ```
 
-The gain  is around 10 elo on STC and LTC.
+You must set `--use-variance-scaling=true` to turn it on. The gain is around 10 elo on STC and 5 elo on LTC, but more testing is needed. For now the setting is NOT RECOMMENDED except for personal use.
 
 
 

appveyor.yml

@@ -124,7 +124,7 @@ before_build:
 - cmd: SET EXTRA=
 - cmd: IF %ANDROID%==false SET EXTRA=-Db_vscrt=md
 - cmd: IF %ONNX_DML%==true SET EXTRA=-Db_vscrt=md -Donnx_libdir=C:\cache\%ONNX_NAME%\lib -Donnx_include=C:\cache\%ONNX_NAME%\include
-- cmd: IF %ANDROID%==false meson build --backend vs2017 --buildtype release -Dgtest=%GTEST% -Dopencl=%OPENCL% -Dblas=%BUILD_BLAS% -Ddnnl=true -Ddx=%DX% -Dcudnn=%CUDNN% -Donednn=%ONEDNN% -Dispc_native_only=false -Dpopcnt=%POPCNT% -Df16c=%F16C% -Dcudnn_include="%CUDA_PATH%\include","%CUDA_PATH%\cuda\include" -Dcudnn_libdirs="%CUDA_PATH%\lib\x64","%CUDA_PATH%\cuda\lib\x64" -Dopenblas_include="%PKG_FOLDER%\OpenBLAS\dist64\include" -Dopenblas_libdirs="%PKG_FOLDER%\OpenBLAS\dist64\lib" -Ddnnl_dir="%PKG_FOLDER%\%DNNL_NAME%" -Dopencl_include="%PKG_FOLDER%\opencl-nug.0.777.77\build\native\include" -Dopencl_libdirs="%PKG_FOLDER%\opencl-nug.0.777.77\build\native\lib\x64" -Ddefault_library=static -Dmalloc=mimalloc -Dmimalloc_libdir="%MIMALLOC_PATH%"\out\msvc-x64\Release %EXTRA%
+- cmd: IF %ANDROID%==false meson build --backend vs2017 --buildtype release -Dgtest=%GTEST% -Dopencl=%OPENCL% -Dblas=%BUILD_BLAS% -Ddnnl=true -Ddx=%DX% -Dcudnn=%CUDNN% -Donednn=%ONEDNN% -Dispc_native_only=false -Dnative_cuda=false -Dpopcnt=%POPCNT% -Df16c=%F16C% -Dcudnn_include="%CUDA_PATH%\include","%CUDA_PATH%\cuda\include" -Dcudnn_libdirs="%CUDA_PATH%\lib\x64","%CUDA_PATH%\cuda\lib\x64" -Dopenblas_include="%PKG_FOLDER%\OpenBLAS\dist64\include" -Dopenblas_libdirs="%PKG_FOLDER%\OpenBLAS\dist64\lib" -Ddnnl_dir="%PKG_FOLDER%\%DNNL_NAME%" -Dopencl_include="%PKG_FOLDER%\opencl-nug.0.777.77\build\native\include" -Dopencl_libdirs="%PKG_FOLDER%\opencl-nug.0.777.77\build\native\lib\x64" -Ddefault_library=static -Dmalloc=mimalloc -Dmimalloc_libdir="%MIMALLOC_PATH%"\out\msvc-x64\Release %EXTRA%
 - cmd: IF %ANDROID%==true meson arm64-v8a --buildtype release -Dgtest=false -Dopenblas_include="%PKG_FOLDER%\OpenBLAS\android-aarch64\include" -Dopenblas_libdirs="%PKG_FOLDER%\OpenBLAS\android-aarch64\lib" -Dembed=%EMBED% -Ddefault_library=static --cross-file crossfile-aarch64
 - cmd: IF %ANDROID%==true meson armeabi-v7a --buildtype release -Dgtest=false -Dopenblas_include="%PKG_FOLDER%\OpenBLAS\android-armv7a\include" -Dopenblas_libdirs="%PKG_FOLDER%\OpenBLAS\android-armv7a\lib" -Dembed=%EMBED% -Ddefault_library=static --cross-file crossfile-armv7a -Dispc=false -Dneon=false
 build_script:

build.cmd

@@ -13,7 +13,7 @@ set EIGEN=false
 set TEST=false
 
 rem 2. Edit the paths for the build dependencies.
-set CUDA_PATH=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.0
+set CUDA_PATH=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.2
 set CUDNN_PATH=%CUDA_PATH%
 set OPENBLAS_PATH=C:\OpenBLAS
 set MKL_PATH=C:\Program Files (x86)\IntelSWTools\compilers_and_libraries\windows\mkl
@@ -63,7 +63,7 @@ meson build --backend %backend% --buildtype release -Ddx=%DX12% -Dcudnn=%CUDNN%
 -Dmkl_include="%MKL_PATH%\include" -Dmkl_libdirs="%MKL_PATH%\lib\intel64" -Ddnnl_dir="%DNNL_PATH%" ^
 -Dopencl_libdirs="%OPENCL_LIB_PATH%" -Dopencl_include="%OPENCL_INCLUDE_PATH%" ^
 -Dopenblas_include="%OPENBLAS_PATH%\include" -Dopenblas_libdirs="%OPENBLAS_PATH%\lib" ^
--Ddefault_library=static
+-Ddefault_library=static 
 
 if errorlevel 1 exit /b
 

build.sh

@@ -28,4 +28,8 @@ else
 fi
 
 "${MESON}" compile -C "${BUILDDIR}"
-[ -n "${INSTALL_PREFIX}" ] && "${MESON}" install -C "${BUILDDIR}"
+
+if [ -n "${INSTALL_PREFIX}" ]
+then
+  "${MESON}" install -C "${BUILDDIR}"
+fi

libs/lczero-common

@@ -448,9 +448,9 @@ if get_option('build_backends')
     if get_option('cudnn') and cu_dnn.found()
       deps += cu_dnn
       cuda_files += 'src/neural/cuda/network_cudnn.cc'
+      cuda_files += 'src/neural/cuda/network_cuda.cc' # To support newer nets.
       add_project_arguments('-DUSE_CUDNN', language : 'cpp')
-    endif
-    if get_option('plain_cuda')
+    elif get_option('plain_cuda')
       cuda_files += 'src/neural/cuda/network_cuda.cc'
     endif
     foreach d : get_option('cudnn_include')
@@ -461,8 +461,9 @@ if get_option('build_backends')
     includes += include_directories('src/neural/cuda/')
 
     cuda_arguments = ['-c', '@INPUT@', '-o', '@OUTPUT@',
-                      '-I', meson.source_root() + '/subprojects/abseil-cpp-20211102.0',
+                      '-I', meson.source_root() + '/subprojects/abseil-cpp-20220623.0',
                       '-I', meson.current_source_dir() + '/src']
+    nvcc_help = run_command(nvcc, '-h').stdout()
     if host_machine.system() == 'windows'
       if get_option('b_vscrt') == 'mt'
         cuda_arguments += ['-Xcompiler', '-MT']
@@ -483,6 +484,8 @@ if get_option('build_backends')
     nvcc_extra_args = []
     if cuda_cc != ''
       nvcc_extra_args = ['-arch=compute_' + cuda_cc, '-code=sm_' + cuda_cc]
+    elif get_option('native_cuda') and nvcc_help.contains('-arch=native')
+      nvcc_extra_args = ['-arch=native']
     endif
     foreach x : get_option('cudnn_include')
       cuda_arguments += ['-I', x]
@@ -500,23 +503,24 @@ if get_option('build_backends')
       command : [nvcc, nvcc_extra_args, cuda_arguments]
     )
 
-    # Handling of fp16 cuda code.
-    nvcc_sm_list = ['80', '75', '86', '70', '89', '90']
-    if host_machine.system() != 'windows'
-      nvcc_sm_list += ['60']
-      if ['arm', 'aarch64'].contains(host_machine.cpu_family())
-        # Add Jetson versions to the list.
-        message('Jetson support enabled.')
-        nvcc_sm_list += ['72', '62', '53', '87']
+    # Handling of fp16 cuda code: If nvcc_extra_args is empty add options to
+    # generate code for the major fp16 capable architectures.
+    if nvcc_extra_args == []
+      nvcc_arch = '-arch=compute_70'
+      nvcc_sm_list = ['sm_70', 'sm_75', 'sm_80', 'sm_90']
+      if host_machine.system() != 'windows'
+        nvcc_arch = '-arch=compute_60'
+        nvcc_sm_list += ['sm_60']
+        if ['arm', 'aarch64'].contains(host_machine.cpu_family())
+          message('Compiling for Jetson.')
+          nvcc_arch = '-arch=compute_53'
+          nvcc_sm_list = ['sm_53', 'sm_62', 'sm_72', 'sm_87']
+        endif
       endif
-    endif
-    # Ignore the given CC for fp16 when it is not in the supported list.
-    if cuda_cc == '' or not nvcc_sm_list.contains('sm_' + cuda_cc)
-      nvcc_extra_args = []
-      nvcc_help = run_command(nvcc, '-h').stdout()
+      nvcc_extra_args = [nvcc_arch]
       foreach x : nvcc_sm_list
-        if nvcc_help.contains('sm_' + x)
-          nvcc_extra_args += '-gencode=arch=compute_' + x + ',code=sm_' + x
+        if nvcc_help.contains(x)
+          nvcc_extra_args += '-code=' + x
         endif
       endforeach
       # For forward compatibility.
@@ -573,19 +577,23 @@ if get_option('build_backends')
   ## ONNX
   ## ~~~~~~~~~~
   if get_option('onnx_libdir') != '' and get_option('onnx_include') != ''
-    onnx_lib = cc.find_library('onnxruntime',
-                               dirs: get_option('onnx_libdir'),
-                               required: true)
+    deps += cc.find_library('onnxruntime', dirs: get_option('onnx_libdir'),
+                            required: true)
     includes += include_directories(get_option('onnx_include'), is_system: true)
-    cc.has_header('onnxruntime_cxx_api.h',
-                  required: true,
+    cc.has_header('onnxruntime_cxx_api.h', required: true,
                   args: '-I' + get_option('onnx_include'))
-    deps += [onnx_lib]
-
-    files += [
-      'src/neural/onnx/network_onnx.cc',
-    ]
-
+    if not cc.has_header('cpu_provider_factory.h',
+                         args: '-I' + get_option('onnx_include'))
+      cc.has_header('../providers/cpu/cpu_provider_factory.h', required: true,
+                    args: '-I' + get_option('onnx_include'))
+      includes += include_directories(get_option('onnx_include') + '/../providers/cpu',
+                                      is_system: true)
+    endif
+    files += 'src/neural/onnx/network_onnx.cc'
+    if cc.find_library('onnxruntime_providers_rocm',
+                       dirs: get_option('onnx_libdir'), required: false).found()
+      add_project_arguments('-DUSE_ROCM', language : 'cpp')
+    endif
     has_backends = true
   endif
 

meson.build

@@ -63,6 +63,11 @@ option('ispc_native_only',
        value: true,
        description: 'use ispc and enable native arch only')
 
+option('native_cuda',
+       type: 'boolean',
+       value: true,
+       description: 'build cuda code for native arch only (if supported)')
+
 option('cudnn',
        type: 'boolean',
        value: true,

meson_options.txt

@@ -1,6 +1,6 @@
 7z a lc0-%APPVEYOR_REPO_TAG_NAME%-windows-%NAME%.zip %APPVEYOR_BUILD_FOLDER%\build\lc0.exe
-appveyor DownloadFile "https://ci.appveyor.com/api/projects/LeelaChessZero/lczero-client/artifacts/lc0-training-client.exe?branch=release&pr=false&job=Environment%%3A%%20NAME%%3D.exe%%2C%%20GOOS%%3Dwindows"
-7z a lc0-%APPVEYOR_REPO_TAG_NAME%-windows-%NAME%.zip lc0-training-client.exe
+IF %NAME%==gpu-nvidia-cuda appveyor DownloadFile "https://github.com/LeelaChessZero/lczero-client/releases/latest/download/lc0-training-client.exe"
+IF %NAME%==gpu-nvidia-cuda 7z a lc0-%APPVEYOR_REPO_TAG_NAME%-windows-%NAME%.zip lc0-training-client.exe
 type COPYING |more /P > dist\COPYING
 7z a lc0-%APPVEYOR_REPO_TAG_NAME%-windows-%NAME%.zip .\dist\COPYING
 7z a lc0-%APPVEYOR_REPO_TAG_NAME%-windows-%NAME%.zip c:\cache\%NET%.pb.gz

scripts/appveyor_win_package.cmd

@@ -106,8 +106,8 @@ void Benchmark::Run() {
           std::make_unique<CallbackUciResponder>(
               std::bind(&Benchmark::OnBestMove, this, std::placeholders::_1),
               std::bind(&Benchmark::OnInfo, this, std::placeholders::_1)),
-          MoveList(), start, std::move(stopper), false, option_dict, &cache,
-          nullptr);
+          MoveList(), start, std::move(stopper), false, false, option_dict,
+          &cache, nullptr);
       search->StartThreads(option_dict.Get<int>(kThreadsOptionId));
       search->Wait();
       const auto end = std::chrono::steady_clock::now();

src/benchmark/benchmark.cc

@@ -54,7 +54,7 @@ const std::unordered_map<std::string, std::unordered_set<std::string>>
         {{"position"}, {"fen", "startpos", "moves"}},
         {{"go"},
          {"infinite", "wtime", "btime", "winc", "binc", "movestogo", "depth",
-          "nodes", "movetime", "searchmoves", "ponder"}},
+          "mate", "nodes", "movetime", "searchmoves", "ponder"}},
         {{"start"}, {}},
         {{"stop"}, {}},
         {{"ponderhit"}, {}},
@@ -192,6 +192,7 @@ bool UciLoop::DispatchCommand(
     UCIGOOPTION(binc);
     UCIGOOPTION(movestogo);
     UCIGOOPTION(depth);
+    UCIGOOPTION(mate);
     UCIGOOPTION(nodes);
     UCIGOOPTION(movetime);
 #undef UCIGOOPTION

src/chess/uciloop.cc

@@ -45,6 +45,7 @@ struct GoParams {
   std::optional<std::int64_t> binc;
   std::optional<int> movestogo;
   std::optional<int> depth;
+  std::optional<int> mate;
   std::optional<int> nodes;
   std::optional<std::int64_t> movetime;
   bool infinite = false;

src/chess/uciloop.h

@@ -33,6 +33,7 @@
 
 #include "mcts/search.h"
 #include "mcts/stoppers/factory.h"
+#include "utils/commandline.h"
 #include "utils/configfile.h"
 #include "utils/logging.h"
 
@@ -51,7 +52,7 @@ const OptionId kSyzygyTablebaseId{
     "List of Syzygy tablebase directories, list entries separated by system "
     "separator (\";\" for Windows, \":\" for Linux).",
     's'};
-const OptionId kPonderId{"ponder", "Ponder",
+const OptionId kPonderId{"", "Ponder",
                          "This option is ignored. Here to please chess GUIs."};
 const OptionId kUciChess960{
     "chess960", "UCI_Chess960",
@@ -93,12 +94,21 @@ EngineController::EngineController(std::unique_ptr<UciResponder> uci_responder,
 
 void EngineController::PopulateOptions(OptionsParser* options) {
   using namespace std::placeholders;
-
+  const bool is_simple =
+      CommandLine::BinaryName().find("simple") != std::string::npos;
   NetworkFactory::PopulateOptions(options);
   options->Add<IntOption>(kThreadsOptionId, 1, 128) = kDefaultThreads;
   options->Add<IntOption>(kNNCacheSizeId, 0, 999999999) = 2000;
   SearchParams::Populate(options);
 
+  ConfigFile::PopulateOptions(options);
+  if (is_simple) {
+    options->HideAllOptions();
+    options->UnhideOption(kThreadsOptionId);
+    options->UnhideOption(NetworkFactory::kWeightsId);
+    options->UnhideOption(SearchParams::kContemptId);
+    options->UnhideOption(SearchParams::kMultiPvId);
+  }
   options->Add<StringOption>(kSyzygyTablebaseId);
   // Add "Ponder" option to signal to GUIs that we support pondering.
   // This option is currently not used by lc0 in any way.
@@ -107,8 +117,8 @@ void EngineController::PopulateOptions(OptionsParser* options) {
   options->Add<BoolOption>(kShowWDL) = false;
   options->Add<BoolOption>(kShowMovesleft) = false;
 
-  ConfigFile::PopulateOptions(options);
-  PopulateTimeManagementOptions(RunType::kUci, options);
+  PopulateTimeManagementOptions(is_simple ? RunType::kSimpleUci : RunType::kUci,
+                                options);
 
   options->Add<BoolOption>(kStrictUciTiming) = false;
   options->HideOption(kStrictUciTiming);
@@ -238,6 +248,7 @@ class PonderResponseTransformer : public TransformingUciResponder {
         if (ponder_info.score) ponder_info.score = -*ponder_info.score;
         if (ponder_info.depth > 1) ponder_info.depth--;
         if (ponder_info.seldepth > 1) ponder_info.seldepth--;
+        if (ponder_info.wdl) std::swap(ponder_info.wdl->w, ponder_info.wdl->l);
         ponder_info.pv.clear();
       }
       if (!info.pv.empty() && info.pv[0].as_string() == ponder_move_) {
@@ -298,7 +309,7 @@ void EngineController::Go(const GoParams& params) {
   search_ = std::make_unique<Search>(
       tree_.get(), network_.get(), std::move(responder),
       StringsToMovelist(params.searchmoves, tree_->HeadPosition().GetBoard()),
-      *move_start_time_, std::move(stopper), params.infinite || params.ponder,
+      *move_start_time_, std::move(stopper), params.infinite, params.ponder,
       options_, &cache_, syzygy_tb_.get());
 
   LOGFILE << "Timer started at "