Implement complete Lc0-compatible neural network inference with Metal backend for MCTS

.github/workflows/ci.yml

@@ -80,6 +80,9 @@ jobs:
           echo ""
           echo "Available Metal devices:"
           system_profiler SPDisplaysDataType | grep -A 5 "Metal" || echo "Metal info not available in CI"
+          echo ""
+          echo "Installing dependencies..."
+          brew install protobuf zlib abseil
 
       - name: Setup environment (Ubuntu)
         if: runner.os == 'Linux'
@@ -99,6 +102,10 @@ jobs:
           else
             echo "No NVIDIA GPU detected (nvidia-smi not found)"
           fi
+          echo ""
+          echo "Installing dependencies..."
+          sudo apt-get update
+          sudo apt-get install -y libprotobuf-dev protobuf-compiler zlib1g-dev
         shell: bash
 
       - name: Setup environment (Windows)
@@ -118,6 +125,16 @@ jobs:
           }
         shell: pwsh
 
+      - name: Install dependencies (Windows)
+        if: runner.os == 'Windows'
+        run: |
+          # Install vcpkg and protobuf
+          git clone https://github.com/Microsoft/vcpkg.git C:\vcpkg
+          C:\vcpkg\bootstrap-vcpkg.bat
+          C:\vcpkg\vcpkg install protobuf:x64-windows zlib:x64-windows
+          echo "CMAKE_TOOLCHAIN_FILE=C:\vcpkg\scripts\buildsystems\vcpkg.cmake" >> $env:GITHUB_ENV
+        shell: pwsh
+
       - name: Download NNUE files
         run: |
           mkdir -p src
@@ -135,7 +152,8 @@ jobs:
           restore-keys: |
             ${{ runner.os }}-${{ matrix.os }}-cmake-
 
-      - name: Configure CMake
+      - name: Configure CMake (Unix)
+        if: runner.os != 'Windows'
         run: |
           cmake -S . -B build \
             -DCMAKE_BUILD_TYPE=${{ env.BUILD_TYPE }} \
@@ -144,6 +162,17 @@ jobs:
             -DBUILD_TESTS=ON
         shell: bash
 
+      - name: Configure CMake (Windows)
+        if: runner.os == 'Windows'
+        run: |
+          cmake -S . -B build `
+            -DCMAKE_BUILD_TYPE=${{ env.BUILD_TYPE }} `
+            -DUSE_METAL=${{ matrix.use_metal }} `
+            -DUSE_CUDA=${{ matrix.use_cuda }} `
+            -DBUILD_TESTS=ON `
+            -DCMAKE_TOOLCHAIN_FILE=C:\vcpkg\scripts\buildsystems\vcpkg.cmake
+        shell: pwsh
+
       - name: Build (Unix)
         if: runner.os != 'Windows'
         run: |
@@ -246,6 +275,12 @@ jobs:
           nvcc --version
         shell: bash
 
+      - name: Install dependencies
+        run: |
+          sudo apt-get update
+          sudo apt-get install -y libprotobuf-dev protobuf-compiler zlib1g-dev
+        shell: bash
+
       - name: Download NNUE files
         run: |
           mkdir -p src
@@ -354,6 +389,10 @@ jobs:
         with:
           submodules: recursive
 
+      - name: Install dependencies
+        run: brew install protobuf zlib abseil
+        shell: bash
+
       - name: Download NNUE files
         run: |
           mkdir -p src
@@ -412,6 +451,12 @@ jobs:
           method: "network"
           sub-packages: '["nvcc", "cudart", "cudart-dev"]'
 
+      - name: Install dependencies
+        run: |
+          sudo apt-get update
+          sudo apt-get install -y libprotobuf-dev protobuf-compiler zlib1g-dev
+        shell: bash
+
       - name: Download NNUE files
         run: |
           mkdir -p src

.github/workflows/elo-tournament.yml

@@ -1,11 +1,12 @@
 name: Elo Tournament
 
 on:
-  pull_request:
-    branches: [main]
-    types: [opened, synchronize, reopened]
-  workflow_dispatch: # Allow manual trigger
+  workflow_dispatch: # Manual trigger only
     inputs:
+      pr_number:
+        description: "PR number to run tournament on (leave empty for current branch)"
+        required: false
+        default: ""
       games_per_match:
         description: "Number of games per match (should be even for color swap)"
         required: false
@@ -15,9 +16,9 @@ on:
         required: false
         default: "600+0.1"
 
-# Cancel in-progress runs for the same PR when a new push occurs
+# Cancel in-progress runs when a new run is triggered
 concurrency:
-  group: elo-tournament-${{ github.event.pull_request.number || github.run_id }}
+  group: elo-tournament-${{ github.event.inputs.pr_number || github.run_id }}
   cancel-in-progress: true
 
 env:
@@ -43,7 +44,7 @@ jobs:
 
       - name: Install build dependencies
         run: |
-          brew install cmake ninja meson qt@6 coreutils
+          brew install cmake ninja meson qt@6 coreutils protobuf zlib abseil
           pip3 install meson ninja chess
           # coreutils provides gtimeout which we alias to timeout
           echo "alias timeout=gtimeout" >> ~/.bashrc
@@ -942,20 +943,20 @@ jobs:
           cat results/pr_comment.md
 
       - name: Find existing comment
-        if: github.event_name == 'pull_request'
+        if: github.event.inputs.pr_number != ''
         uses: peter-evans/find-comment@v3
         id: find-comment
         with:
-          issue-number: ${{ github.event.pull_request.number }}
+          issue-number: ${{ github.event.inputs.pr_number }}
           comment-author: "github-actions[bot]"
           body-includes: "🏆 MetalFish Elo Tournament Results"
 
       - name: Post or update PR comment
-        if: github.event_name == 'pull_request'
+        if: github.event.inputs.pr_number != ''
         uses: peter-evans/create-or-update-comment@v4
         with:
           comment-id: ${{ steps.find-comment.outputs.comment-id }}
-          issue-number: ${{ github.event.pull_request.number }}
+          issue-number: ${{ github.event.inputs.pr_number }}
           body-path: ${{ steps.aggregate.outputs.comment_file }}
           edit-mode: replace
 

.gitignore

@@ -391,4 +391,6 @@ TSWLatexianTemp*
 #*Notes.bib
 *.pyc
 .DS_Store
+_codeql_build_dir/
+_codeql_detected_source_root
 networks/BT4-1024x15x32h-swa-6147500.pb

CMakeLists.txt

@@ -275,13 +275,59 @@ endif()
 # mcts/hybrid commands - apple_silicon_mcts: Apple Silicon specific
 # optimizations
 set(MCTS_SOURCES
-    src/mcts/position_adapter.cpp src/mcts/position_classifier.cpp
-    src/mcts/ab_integration.cpp src/mcts/thread_safe_mcts.cpp
-    src/mcts/parallel_hybrid_search.cpp src/mcts/apple_silicon_mcts.cpp)
+    src/mcts/position_classifier.cpp
+    src/mcts/ab_integration.cpp
+    src/mcts/thread_safe_mcts.cpp
+    src/mcts/nn_mcts_evaluator.cpp
+    src/mcts/position_adapter.cpp
+    src/mcts/parallel_hybrid_search.cpp
+    src/mcts/apple_silicon_mcts.cpp)
+
+# Find protobuf (minimum version 3.0) - must be before NN_SOURCES
+find_package(Protobuf 3.0 REQUIRED)
+include_directories(${Protobuf_INCLUDE_DIRS})
+
+# Generate protobuf files from .proto definition
+# This ensures compatibility with the installed protobuf version
+set(PROTO_FILE ${CMAKE_CURRENT_SOURCE_DIR}/src/nn/proto/net.proto)
+set(PROTO_OUTPUT_DIR ${CMAKE_CURRENT_BINARY_DIR}/proto)
+file(MAKE_DIRECTORY ${PROTO_OUTPUT_DIR})
+
+# Custom command to generate protobuf files in the correct location
+# The proto file outputs to PROTO_OUTPUT_DIR so includes like "proto/net.pb.h" work
+add_custom_command(
+    OUTPUT ${PROTO_OUTPUT_DIR}/net.pb.cc ${PROTO_OUTPUT_DIR}/net.pb.h
+    COMMAND ${Protobuf_PROTOC_EXECUTABLE}
+    ARGS --cpp_out=${PROTO_OUTPUT_DIR}
+         --proto_path=${CMAKE_CURRENT_SOURCE_DIR}/src/nn/proto
+         ${PROTO_FILE}
+    DEPENDS ${PROTO_FILE}
+    COMMENT "Generating protobuf files from net.proto"
+    VERBATIM
+)
+
+# Add generated directory to include path (so "proto/net.pb.h" works from build dir)
+# The include path should be the parent of proto/ so that "proto/net.pb.h" resolves
+include_directories(${CMAKE_CURRENT_BINARY_DIR})
+include_directories(${CMAKE_CURRENT_SOURCE_DIR}/src/nn)
+
+# Neural network source files (includes generated protobuf)
+set(NN_SOURCES
+    ${PROTO_OUTPUT_DIR}/net.pb.cc
+    src/nn/loader.cpp
+    src/nn/encoder.cpp
+    src/nn/policy_map.cpp
+    src/nn/network.cpp)
 
 # Metal GPU acceleration (macOS only)
 if(USE_METAL AND METAL_CPP_AVAILABLE)
   set(GPU_SOURCES ${GPU_SOURCES} src/gpu/metal/metal_backend.mm)
+  set(NN_SOURCES ${NN_SOURCES} src/nn/metal/metal_network.mm)
+
+  # Disable ARC for Metal network implementation (uses manual memory management)
+  set_source_files_properties(src/nn/metal/metal_network.mm 
+                              PROPERTIES COMPILE_FLAGS "-fno-objc-arc")
+
   set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -DUSE_METAL")
   message(STATUS "Metal GPU acceleration: ENABLED")
 
@@ -328,7 +374,8 @@ set(SOURCES
     ${UCI_SOURCES}
     ${SYZYGY_SOURCES}
     ${GPU_SOURCES}
-    ${MCTS_SOURCES})
+    ${MCTS_SOURCES}
+    ${NN_SOURCES})
 
 # Create executable
 if(USE_CUDA AND CUDA_AVAILABLE)
@@ -346,9 +393,33 @@ if(TARGET metal_shaders)
   add_dependencies(metalfish metal_shaders)
 endif()
 
+# Find zlib (for loading .pb.gz files)
+find_package(ZLIB REQUIRED)
+
+# Find absl (required by protobuf >= 22 on some platforms)
+# Initialize to empty - only set if found
+set(ABSL_LIBS "")
+find_package(absl CONFIG QUIET)
+if(absl_FOUND)
+  message(STATUS "Found abseil - linking absl::log")
+  set(ABSL_LIBS absl::log absl::log_internal_check_op absl::log_internal_message)
+else()
+  # Try pkg-config as fallback (for Linux)
+  find_package(PkgConfig QUIET)
+  if(PKG_CONFIG_FOUND)
+    pkg_check_modules(ABSL_PKG QUIET absl_log)
+    if(ABSL_PKG_FOUND)
+      message(STATUS "Found abseil via pkg-config")
+      set(ABSL_LIBS ${ABSL_PKG_LIBRARIES})
+      include_directories(${ABSL_PKG_INCLUDE_DIRS})
+      link_directories(${ABSL_PKG_LIBRARY_DIRS})
+    endif()
+  endif()
+endif()
+
 # Link pthread
 find_package(Threads REQUIRED)
-target_link_libraries(metalfish Threads::Threads)
+target_link_libraries(metalfish Threads::Threads ${Protobuf_LIBRARIES} ${ZLIB_LIBRARIES} ${ABSL_LIBS})
 
 # macOS specific
 if(APPLE)
@@ -357,12 +428,14 @@ if(APPLE)
   find_library(ACCELERATE_FRAMEWORK Accelerate)
   find_library(COREFOUNDATION_FRAMEWORK CoreFoundation)
   find_library(QUARTZCORE_FRAMEWORK QuartzCore)
+  find_library(MPS_FRAMEWORK MetalPerformanceShaders)
+  find_library(MPSGRAPH_FRAMEWORK MetalPerformanceShadersGraph)
 
   if(USE_METAL AND METAL_CPP_AVAILABLE)
-    target_link_libraries(
-      metalfish ${METAL_FRAMEWORK} ${FOUNDATION_FRAMEWORK}
-      ${COREFOUNDATION_FRAMEWORK} ${QUARTZCORE_FRAMEWORK}
-      ${ACCELERATE_FRAMEWORK})
+    target_link_libraries(metalfish ${METAL_FRAMEWORK} ${FOUNDATION_FRAMEWORK}
+                          ${COREFOUNDATION_FRAMEWORK} ${QUARTZCORE_FRAMEWORK}
+                          ${MPS_FRAMEWORK} ${MPSGRAPH_FRAMEWORK}
+                          ${ACCELERATE_FRAMEWORK})
   endif()
 endif()
 
@@ -409,6 +482,11 @@ if(BUILD_TESTS)
       tests/test_metal.cpp
       tests/test_gpu_nnue.cpp
       tests/test_cuda.cpp)
+
+  set(NN_TEST_SOURCES
+      tests/test_nn_comparison.cpp
+      ${CORE_SOURCES}
+      ${NN_SOURCES})
 
   if(USE_CUDA AND CUDA_AVAILABLE)
     # CUDA test executable
@@ -427,6 +505,7 @@ if(BUILD_TESTS)
       metalfish_tests PROPERTIES CUDA_SEPARABLE_COMPILATION ON
                                  CUDA_RESOLVE_DEVICE_SYMBOLS ON)
     target_link_libraries(metalfish_tests Threads::Threads
+                          ${Protobuf_LIBRARIES} ${ZLIB_LIBRARIES} ${ABSL_LIBS}
                           ${CUDA_LINK_LIBRARIES})
   else()
     add_executable(
@@ -438,8 +517,9 @@ if(BUILD_TESTS)
       ${UCI_SOURCES}
       ${SYZYGY_SOURCES}
       ${GPU_SOURCES}
-      ${MCTS_SOURCES})
-    target_link_libraries(metalfish_tests Threads::Threads)
+      ${MCTS_SOURCES}
+      ${NN_SOURCES})
+    target_link_libraries(metalfish_tests Threads::Threads ${Protobuf_LIBRARIES} ${ZLIB_LIBRARIES} ${ABSL_LIBS})
   endif()
 
   if(APPLE
@@ -448,10 +528,24 @@ if(BUILD_TESTS)
     target_link_libraries(
       metalfish_tests ${METAL_FRAMEWORK} ${FOUNDATION_FRAMEWORK}
       ${COREFOUNDATION_FRAMEWORK} ${QUARTZCORE_FRAMEWORK}
+      ${MPS_FRAMEWORK} ${MPSGRAPH_FRAMEWORK}
       ${ACCELERATE_FRAMEWORK})
   endif()
 
   add_test(NAME metalfish_tests COMMAND metalfish_tests)
+
+  # Neural network comparison test
+  add_executable(test_nn_comparison ${NN_TEST_SOURCES} ${MCTS_SOURCES})
+  target_link_libraries(test_nn_comparison Threads::Threads ${Protobuf_LIBRARIES} ${ZLIB_LIBRARIES} ${ABSL_LIBS})
+
+  if(APPLE AND USE_METAL AND METAL_CPP_AVAILABLE)
+    target_link_libraries(
+      test_nn_comparison ${METAL_FRAMEWORK} ${FOUNDATION_FRAMEWORK}
+      ${COREFOUNDATION_FRAMEWORK} ${QUARTZCORE_FRAMEWORK}
+      ${MPS_FRAMEWORK} ${MPSGRAPH_FRAMEWORK})
+  endif()
+
+  add_test(NAME test_nn_comparison COMMAND test_nn_comparison)
 endif()
 
 # ============================================================================
@@ -465,7 +559,7 @@ if(BUILD_GPU_BENCHMARK)
     target_link_libraries(
       metalfish_gpu_bench Threads::Threads ${METAL_FRAMEWORK}
       ${FOUNDATION_FRAMEWORK} ${COREFOUNDATION_FRAMEWORK}
-      ${QUARTZCORE_FRAMEWORK})
+      ${QUARTZCORE_FRAMEWORK} ${MPS_FRAMEWORK} ${MPSGRAPH_FRAMEWORK})
 
     # Paper benchmark with full NNUE support
     add_executable(
@@ -474,7 +568,7 @@ if(BUILD_GPU_BENCHMARK)
     target_link_libraries(
       metalfish_paper_bench Threads::Threads ${METAL_FRAMEWORK}
       ${FOUNDATION_FRAMEWORK} ${COREFOUNDATION_FRAMEWORK}
-      ${QUARTZCORE_FRAMEWORK})
+      ${QUARTZCORE_FRAMEWORK} ${MPS_FRAMEWORK} ${MPSGRAPH_FRAMEWORK})
   elseif(USE_CUDA AND CUDA_AVAILABLE)
     add_executable(metalfish_gpu_bench src/benchmark_gpu.cpp ${CORE_SOURCES}
                                        ${GPU_SOURCES} ${CUDA_SOURCES})