move bits functions to separate file + fix missing defaults with disabled program_options on windows

.gitignore

@@ -6,3 +6,6 @@ cmake-build-release
 /.gdb_history
 /valgrind.log
 .idea
+/perf.data
+/perf.data.old
+/clusters.puzzle

CMakeLists.txt

@@ -1,23 +1,60 @@
-cmake_minimum_required(VERSION 3.25)
+cmake_minimum_required(VERSION 3.28)
 project(MassSprings)
 
 set(CMAKE_CXX_STANDARD 26)
 set(CMAKE_EXPORT_COMPILE_COMMANDS ON)
 
+# Disable boost warning because our cmake/boost are recent enough
+if(POLICY CMP0167)
+    cmake_policy(SET CMP0167 NEW)
+endif()
+
+option(DISABLE_THREADPOOL "Disable additional physics threads" OFF)
+option(DISABLE_BACKWARD "Disable backward stacktrace printer" OFF)
+option(DISABLE_TRACY "Disable the Tracy profiler client" OFF)
+option(DISABLE_TESTS "Disable building tests" OFF)
+option(DISABLE_BENCH "Disable building benchmarks" OFF)
+
+# Headers + Sources (excluding main.cpp)
+set(SOURCES
+    src/backward.cpp
+    src/bits.cpp
+    src/graph_distances.cpp
+    src/input_handler.cpp
+    src/load_save.cpp
+    src/mass_spring_system.cpp
+    src/octree.cpp
+    src/orbit_camera.cpp
+    src/puzzle.cpp
+    src/renderer.cpp
+    src/state_manager.cpp
+    src/threaded_physics.cpp
+    src/user_interface.cpp
+)
+
 # Libraries
+include(FetchContent)
 find_package(raylib REQUIRED)
-set(LIBS raylib)
+find_package(Boost COMPONENTS program_options REQUIRED)
+set(LIBS raylib Boost::headers Boost::program_options)
 set(FLAGS "")
 
-if(NOT DEFINED DISABLE_BACKWARD)
+if(WIN32)
+    list(APPEND LIBS opengl32 gdi32 winmm)
+endif()
+
+if(NOT DISABLE_THREADPOOL)
+    list(APPEND FLAGS THREADPOOL)
+endif()
+
+if(NOT DISABLE_BACKWARD)
     find_package(Backward REQUIRED)
 
     list(APPEND LIBS Backward::Backward)
     list(APPEND FLAGS BACKWARD)
 endif()
 
-if(NOT DEFINED DISABLE_TRACY)
-    include(FetchContent)
+if(NOT DISABLE_TRACY)
     FetchContent_Declare(tracy
                          GIT_REPOSITORY https://github.com/wolfpld/tracy.git
                          GIT_TAG v0.11.1
@@ -32,14 +69,11 @@ if(NOT DEFINED DISABLE_TRACY)
     list(APPEND FLAGS TRACY)
 endif()
 
-if(WIN32)
-    list(APPEND LIBS opengl32 gdi32 winmm)
-endif()
-
-# Set this after fetching tracy to hide tracy's warnings
-set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wall -Wextra -Wfloat-equal -Wundef -Wshadow -Wpointer-arith -Wcast-align -Wno-unused-parameter -Wunreachable-code")
-set(CMAKE_CXX_FLAGS_DEBUG "${CMAKE_CXX_FLAGS_DEBUG} -O2 -ggdb") # -fsanitize=address already fails on InitWindow(), -fsanitize=undefined, -fsanitize=leak
-set(CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE} -Ofast")
+# Set this after fetching tracy to hide tracy's warnings.
+# We set -Wno-alloc-size-larger-than because it prevents BS::thread_pool from building with current gcc
+set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wall -Wextra -Wfloat-equal -Wundef -Wshadow -Wpointer-arith -Wcast-align -Wno-unused-parameter -Wunreachable-code -Wno-alloc-size-larger-than")
+set(CMAKE_CXX_FLAGS_DEBUG "${CMAKE_CXX_FLAGS_DEBUG} -ggdb -O0")
+set(CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE} -ggdb -O3 -ffast-math -march=native")
 
 message("-- CMAKE_C_FLAGS: ${CMAKE_C_FLAGS}")
 message("-- CMAKE_C_FLAGS_DEBUG: ${CMAKE_C_FLAGS_DEBUG}")
@@ -48,40 +82,56 @@ message("-- CMAKE_CXX_FLAGS: ${CMAKE_CXX_FLAGS}")
 message("-- CMAKE_CXX_FLAGS_DEBUG: ${CMAKE_CXX_FLAGS_DEBUG}")
 message("-- CMAKE_CXX_FLAGS_RELEASE: ${CMAKE_CXX_FLAGS_RELEASE}")
 
-# Headers + Sources
-include_directories(include)
-set(SOURCES
-    src/backward.cpp
-    src/graph_distances.cpp
-    src/input_handler.cpp
-    src/main.cpp
-    src/mass_spring_system.cpp
-    src/octree.cpp
-    src/orbit_camera.cpp
-    src/threaded_physics.cpp
-    src/puzzle.cpp
-    src/renderer.cpp
-    src/state_manager.cpp
-    src/user_interface.cpp
-)
-
 # Main target
-add_executable(masssprings ${SOURCES})
-target_include_directories(masssprings PRIVATE ${RAYLIB_CPP_INCLUDE_DIR})
+add_executable(masssprings src/main.cpp ${SOURCES})
+target_include_directories(masssprings PRIVATE include)
 target_link_libraries(masssprings PRIVATE ${LIBS})
 target_compile_definitions(masssprings PRIVATE ${FLAGS})
 
+# Testing
+if(NOT DISABLE_TESTS AND NOT WIN32)
+    enable_testing()
+
+    FetchContent_Declare(Catch2
+                         GIT_REPOSITORY https://github.com/catchorg/Catch2.git
+                         GIT_TAG v3.13.0
+    )
+    FetchContent_MakeAvailable(Catch2)
+
+    set(TEST_SOURCES
+        test/bits.cpp
+        test/bitmap.cpp
+        test/bitmap_find_first_empty.cpp
+        # test/puzzle.cpp
+    )
+
+    add_executable(tests ${TEST_SOURCES} ${SOURCES})
+    target_include_directories(tests PRIVATE include)
+    target_link_libraries(tests Catch2::Catch2WithMain raylib)
+
+    include(Catch)
+    catch_discover_tests(tests)
+endif()
+
+# Benchmarking
+if(NOT DISABLE_BENCH AND NOT WIN32)
+    find_package(benchmark REQUIRED)
+
+    set(BENCH_SOURCES
+        benchmark/state_space.cpp
+    )
+
+    add_executable(benchmarks ${BENCH_SOURCES} ${SOURCES})
+    target_include_directories(benchmarks PRIVATE include)
+    target_link_libraries(benchmarks benchmark raylib)
+endif()
+
 # LTO
-if(NOT WIN32)
-    include(CheckIPOSupported)
-    check_ipo_supported(RESULT supported OUTPUT error)
-    if(supported)
-        message(STATUS "IPO / LTO enabled")
-        set_property(TARGET masssprings PROPERTY INTERPROCEDURAL_OPTIMIZATION TRUE)
-        if(USE_TRACY)
-            set_property(TARGET masssprings_tracy PROPERTY INTERPROCEDURAL_OPTIMIZATION TRUE)
-        endif()
-    else()
-        message(STATUS "IPO / LTO not supported: <${error}>")
-    endif()
+include(CheckIPOSupported)
+check_ipo_supported(RESULT supported OUTPUT error)
+if(supported)
+    message(STATUS "IPO / LTO enabled")
+    set_property(TARGET masssprings PROPERTY INTERPROCEDURAL_OPTIMIZATION TRUE)
+else()
+    message(STATUS "IPO / LTO not supported")
 endif()

README.md

@@ -5,4 +5,12 @@ The graph layout is calculated iteratively using a mass-spring-system with addit
 
 ![](screenshot.png)
 
-Build and run on NixOS: `nix run git+https://gitea.local.chriphost.de/christoph/cpp-masssprings`.
+## Running
+
+Requirements:
+
+- Directory `fonts`
+- Directory `shader`
+- Preset file `default.puzzle` (optional)
+
+Run `nix run git+https://gitea.local.chriphost.de/christoph/cpp-masssprings` from the working directory containing the listed requirements.

benchmark/state_space.cpp

@@ -0,0 +1,91 @@
+#include "puzzle.hpp"
+
+#include <benchmark/benchmark.h>
+
+static std::vector<std::string> puzzles = {
+    // 0: RushHour 1
+    "S:[6x6] G:[4,2] M:[R] B:[{3x1 _ _ _ _ 1x3} {_ _ _ _ _ _} {_ _ 1x2 2X1 _ _} {_ _ _ 1x2 2x1 _} {1x2 _ 1x2 _ 2x1 _} {_ _ _ 3x1 _ _}]",
+    // 1: RushHour 2
+    "S:[6x6] G:[4,2] M:[R] B:[{1x2 3x1 _ _ 1x2 1x3} {_ 3x1 _ _ _ _} {2X1 _ 1x2 1x2 1x2 _} {2x1 _ _ _ _ _} {_ _ _ 1x2 2x1 _} {_ _ _ _ 2x1 _}]",
+    // 2: RushHour 3
+    "S:[6x6] G:[4,2] M:[R] B:[{3x1 _ _ 1x2 _ _} {1x2 2x1 _ _ _ 1x2} {_ 2X1 _ 1x2 1x2 _} {2x1 _ 1x2 _ _ 1x2} {_ _ _ 2x1 _ _} {_ 2x1 _ 2x1 _ _}]",
+    // 3: RushHour 4
+    "S:[6x6] G:[4,2] M:[R] B:[{1x3 2x1 _ _ 1x2 _} {_ 1x2 1x2 _ _ 1x3} {_ _ _ 2X1 _ _} {3x1 _ _ 1x2 _ _} {_ _ 1x2 _ 2x1 _} {2x1 _ _ 2x1 _ _}]",
+    // 4: RushHour + Walls 1
+    "S:[6x6] G:[4,2] M:[R] B:[{1x2 2x1 _ 1*1 _ _} {_ _ _ 1x2 2x1 _} {1x2 2X1 _ _ _ _} {_ _ 1x2 2x1 _ 1x3} {2x1 _ _ _ _ _} {2x1 _ 3x1 _ _ _}]",
+    // 5: RushHour + Walls 2
+    "S:[6x6] G:[4,2] M:[R] B:[{2x1 _ _ 1x2 1x2 1*1} {3x1 _ _ _ _ _} {1x2 2X1 _ 1x2 _ _} {_ _ 1x2 _ 2x1 _} {_ _ _ 2x1 _ 1x2} {_ _ 2x1 _ 1*1 _}]",
+    // 6: Dad's Puzzler
+    "S:[4x5] G:[0,3] M:[F] B:[{2X2 _ 2x1 _} {_ _ 2x1 _} {1x1 1x1 _ _} {1x2 1x2 2x1 _} {_ _ 2x1 _}]",
+    // 7: Nine Blocks
+    "S:[4x5] G:[0,3] M:[F] B:[{1x2 1x2 _ _} {_ _ 2x1 _} {1x2 1x2 2x1 _} {_ _ 2X2 _} {1x1 1x1 _ _}]",
+    // 8: Quzzle
+    "S:[4x5] G:[2,0] M:[F] B:[{2X2 _ 2x1 _} {_ _ 1x2 1x2} {_ _ _ _} {1x2 2x1 _ 1x1} {_ 2x1 _ 1x1}]",
+    // 9: Thin Klotski
+    "S:[4x5] G:[1,4] M:[F] B:[{1x2 _ 2X1 _} {_ 2x2 _ 1x1} {_ _ _ 1x1} {2x2 _ 1x1 1x1} {_ _ 1x1 1x1}]",
+    // 10: Fat Klotski
+    "S:[4x5] G:[1,3] M:[F] B:[{_ 2X2 _ 1x1} {1x1 _ _ 1x2} {1x1 2x2 _ _} {1x1 _ _ _} {1x1 1x1 2x1 _}]",
+    // 11: Klotski
+    "S:[4x5] G:[1,3] M:[F] B:[{1x2 2X2 _ 1x2} {_ _ _ _} {1x2 2x1 _ 1x2} {_ 1x1 1x1 _} {1x1 _ _ 1x1}]",
+    // 12: Century
+    "S:[4x5] G:[1,3] M:[F] B:[{1x1 2X2 _ 1x1} {1x2 _ _ 1x2} {_ 1x2 _ _} {1x1 _ _ 1x1} {2x1 _ 2x1 _}]",
+    // 13: Super Century
+    "S:[4x5] G:[1,3] M:[F] B:[{1x2 1x1 1x1 1x1} {_ 1x2 2X2 _} {1x2 _ _ _} {_ 2x1 _ 1x1} {_ 2x1 _ _}]",
+    // 14: Supercompo
+    "S:[4x5] G:[1,3] M:[F] B:[{_ 2X2 _ _} {1x1 _ _ 1x1} {1x2 2x1 _ 1x2} {_ 2x1 _ _} {1x1 2x1 _ 1x1}]",
+};
+
+static auto explore_state_space(benchmark::State& state) -> void
+{
+    const puzzle p = puzzle(puzzles[state.range(0)]);
+
+    for (auto _ : state) {
+        auto space = p.explore_state_space();
+
+        benchmark::DoNotOptimize(space);
+    }
+}
+
+static auto explore_rush_hour_puzzle_space(benchmark::State& state) -> void
+{
+    // ReSharper disable once CppTooWideScope
+    constexpr uint8_t max_blocks = 5;
+
+    constexpr uint8_t board_width = 4;
+    constexpr uint8_t board_height = 5;
+    constexpr uint8_t goal_x = board_width - 1;
+    constexpr uint8_t goal_y = 2;
+    constexpr bool restricted = true;
+
+    const boost::unordered_flat_set<puzzle::block, block_hasher2, block_equal2> permitted_blocks = {
+        puzzle::block(0, 0, 2, 1, false, false),
+        puzzle::block(0, 0, 3, 1, false, false),
+        puzzle::block(0, 0, 1, 2, false, false),
+        puzzle::block(0, 0, 1, 3, false, false)
+    };
+    const puzzle::block target_block = puzzle::block(0, 0, 2, 1, true, false);
+    constexpr std::tuple<uint8_t, uint8_t, uint8_t, uint8_t> target_block_pos_range = {
+        0,
+        goal_y,
+        board_width - 1,
+        goal_y
+    };
+
+    const puzzle p = puzzle(board_width, board_height, goal_x, goal_y, restricted, true);
+
+    for (auto _ : state) {
+        boost::unordered_flat_set<puzzle, puzzle_hasher> result = p.explore_puzzle_space(
+            permitted_blocks,
+            target_block,
+            target_block_pos_range,
+            max_blocks,
+            std::nullopt);
+
+        benchmark::DoNotOptimize(result);
+    }
+}
+
+BENCHMARK(explore_state_space)->DenseRange(0, puzzles.size() - 1)->Unit(benchmark::kMicrosecond);
+BENCHMARK(explore_rush_hour_puzzle_space)->Unit(benchmark::kSecond);
+
+BENCHMARK_MAIN();

default.puzzle

@@ -1,44 +1,30 @@
 # RushHour 1
-R664231........13................12ba..........1221..12..12..21........31....
-
+S:[6x6] G:[4,2] M:[R] B:[{3x1 _ _ _ _ 1x3} {_ _ _ _ _ _} {_ _ 1x2 2X1 _ _} {_ _ _ 1x2 2x1 _} {1x2 _ 1x2 _ 2x1 _} {_ _ _ 3x1 _ _}]
 # RushHour 2
-R66421231....1213..31........ba..121212..21................1221..........21..
-
+S:[6x6] G:[4,2] M:[R] B:[{1x2 3x1 _ _ 1x2 1x3} {_ 3x1 _ _ _ _} {2X1 _ 1x2 1x2 1x2 _} {2x1 _ _ _ _ _} {_ _ _ 1x2 2x1 _} {_ _ _ _ 2x1 _}]
 # RushHour 3
-R664231....12....1221......12..ba..1212..21..12....12......21......21..21....
-
+S:[6x6] G:[4,2] M:[R] B:[{3x1 _ _ 1x2 _ _} {1x2 2x1 _ _ _ 1x2} {_ 2X1 _ 1x2 1x2 _} {2x1 _ 1x2 _ _ 1x2} {_ _ _ 2x1 _ _} {_ 2x1 _ 2x1 _ _}]
 # RushHour 4
-R66421321....12....1212....13......ba....31....12........12..21..21....21....
-
+S:[6x6] G:[4,2] M:[R] B:[{1x3 2x1 _ _ 1x2 _} {_ 1x2 1x2 _ _ 1x3} {_ _ _ 2X1 _ _} {3x1 _ _ 1x2 _ _} {_ _ 1x2 _ 2x1 _} {2x1 _ _ 2x1 _ _}]
 # RushHour + Walls 1
-R66421221..AA..........1221..12ba............1221..1321..........21..31......
-
+S:[6x6] G:[4,2] M:[R] B:[{1x2 2x1 _ 1*1 _ _} {_ _ _ 1x2 2x1 _} {1x2 2X1 _ _ _ _} {_ _ 1x2 2x1 _ 1x3} {2x1 _ _ _ _ _} {2x1 _ 3x1 _ _ _}]
 # RushHour + Walls 2
-R664221....1212AA31..........12ba..12........12..21........21..12....21..AA..
-
+S:[6x6] G:[4,2] M:[R] B:[{2x1 _ _ 1x2 1x2 1*1} {3x1 _ _ _ _ _} {1x2 2X1 _ 1x2 _ _} {_ _ 1x2 _ 2x1 _} {_ _ _ 2x1 _ 1x2} {_ _ 2x1 _ 1*1 _}]
 # Dad's Puzzler
-F4503bb..21......21..1111....121221......21..
-
-# Nine Block (Worse)
-F45031212........21..121221......bb..1111....
-
+S:[4x5] G:[0,3] M:[F] B:[{2X2 _ 2x1 _} {_ _ 2x1 _} {1x1 1x1 _ _} {1x2 1x2 2x1 _} {_ _ 2x1 _}]
+# Nine Blocks
+S:[4x5] G:[0,3] M:[F] B:[{1x2 1x2 _ _} {_ _ 2x1 _} {1x2 1x2 2x1 _} {_ _ 2X2 _} {1x1 1x1 _ _}]
 # Quzzle
-F4520bb..21......1212........1221..11..21..11
-
+S:[4x5] G:[2,0] M:[F] B:[{2X2 _ 2x1 _} {_ _ 1x2 1x2} {_ _ _ _} {1x2 2x1 _ 1x1} {_ 2x1 _ 1x1}]
 # Thin Klotski
-F451412..ba....22..11......1122..1111....1111
-
-# Klotski
-F451312bb..12........1221..12..1111..11....11
-
+S:[4x5] G:[1,4] M:[F] B:[{1x2 _ 2X1 _} {_ 2x2 _ 1x1} {_ _ _ 1x1} {2x2 _ 1x1 1x1} {_ _ 1x1 1x1}]
 # Fat Klotski
-F4513..bb..1111....121122....11......111121..
-
+S:[4x5] G:[1,3] M:[F] B:[{_ 2X2 _ 1x1} {1x1 _ _ 1x2} {1x1 2x2 _ _} {1x1 _ _ _} {1x1 1x1 2x1 _}]
+# Klotski
+S:[4x5] G:[1,3] M:[F] B:[{1x2 2X2 _ 1x2} {_ _ _ _} {1x2 2x1 _ 1x2} {_ 1x1 1x1 _} {1x1 _ _ 1x1}]
 # Century
-F451311bb..1112....12..12....11....1121..21..
-
+S:[4x5] G:[1,3] M:[F] B:[{1x1 2X2 _ 1x1} {1x2 _ _ 1x2} {_ 1x2 _ _} {1x1 _ _ 1x1} {2x1 _ 2x1 _}]
 # Super Century
-F451312111111..12bb..12........21..11....21..
-
+S:[4x5] G:[1,3] M:[F] B:[{1x2 1x1 1x1 1x1} {_ 1x2 2X2 _} {1x2 _ _ _} {_ 2x1 _ 1x1} {_ 2x1 _ _}]
 # Supercompo
-F4513..bb....11....111221..12..21....1121..11
+S:[4x5] G:[1,3] M:[F] B:[{_ 2X2 _ _} {1x1 _ _ 1x1} {1x2 2x1 _ 1x2} {_ 2x1 _ _} {1x1 2x1 _ 1x1}]

flake.lock

@@ -20,11 +20,11 @@
     },
     "nixpkgs": {
       "locked": {
-        "lastModified": 1770843696,
-        "narHash": "sha256-LovWTGDwXhkfCOmbgLVA10bvsi/P8eDDpRudgk68HA8=",
+        "lastModified": 1772479524,
+        "narHash": "sha256-u7nCaNiMjqvKpE+uZz9hE7pgXXTmm5yvdtFaqzSzUQI=",
         "owner": "NixOS",
         "repo": "nixpkgs",
-        "rev": "2343bbb58f99267223bc2aac4fc9ea301a155a16",
+        "rev": "4215e62dc2cd3bc705b0a423b9719ff6be378a43",
         "type": "github"
       },
       "original": {

flake.nix

@@ -14,21 +14,123 @@ rec {
   # Create a shell (and possibly package) for each possible system, not only x86_64-linux
     flake-utils.lib.eachDefaultSystem (
       system: let
+        # =========================================================================================
+        # Define pkgs/stdenvs
+        # =========================================================================================
         pkgs = import nixpkgs {
           inherit system;
           config.allowUnfree = true;
           overlays = [];
         };
-        inherit (pkgs) lib stdenv;
+
+        clangPkgs = import nixpkgs {
+          inherit system;
+          config.allowUnfree = true;
+          overlays = [];
+
+          # Use this to change the compiler:
+          # - GCC: pkgs.stdenv
+          # - Clang: pkgs.clangStdenv
+          # NixOS packages are built using GCC by default. Using clang requires a full rebuild/redownload.
+          config.replaceStdenv = {pkgs}: pkgs.clangStdenv;
+        };
 
         windowsPkgs = import nixpkgs {
           inherit system;
+          config.allowUnfree = true;
+          overlays = [];
+
+          # Use this to cross compile to a different system
           crossSystem = {
             config = "x86_64-w64-mingw32";
           };
-          config.allowUnfree = true;
         };
 
+        inherit (pkgs) lib stdenv;
+
+        # =========================================================================================
+        # Define shell environment
+        # =========================================================================================
+
+        # Setup the shell when entering the "nix develop" environment (bash script).
+        shellHook = let
+          mkCmakeScript = type: let
+            typeLower = lib.toLower type;
+          in
+            pkgs.writers.writeFish "cmake-${typeLower}.fish" ''
+              cd $FLAKE_PROJECT_ROOT
+
+              echo "Removing build directory ./cmake-build-${typeLower}/"
+              rm -rf ./cmake-build-${typeLower}
+
+              echo "Creating build directory"
+              mkdir cmake-build-${typeLower}
+              cd cmake-build-${typeLower}
+
+              echo "Running cmake"
+              cmake -G "Ninja" \
+                    -DCMAKE_BUILD_TYPE="${type}" \
+                    ..
+
+              echo "Linking compile_commands.json"
+              cd ..
+              ln -sf ./cmake-build-${typeLower}/compile_commands.json ./compile_commands.json
+            '';
+
+          cmakeDebug = mkCmakeScript "Debug";
+          cmakeRelease = mkCmakeScript "Release";
+
+          mkBuildScript = type: let
+            typeLower = lib.toLower type;
+          in
+            pkgs.writers.writeFish "cmake-build.fish" ''
+              cd $FLAKE_PROJECT_ROOT/cmake-build-${typeLower}
+
+              echo "Running cmake"
+              NIX_ENFORCE_NO_NATIVE=0 cmake --build . -j$(nproc)
+            '';
+
+          buildDebug = mkBuildScript "Debug";
+          buildRelease = mkBuildScript "Release";
+
+          # Use this to specify commands that should be ran after entering fish shell
+          initProjectShell = pkgs.writers.writeFish "init-shell.fish" ''
+            echo "Entering \"${description}\" environment..."
+
+            # Determine the project root, used e.g. in cmake scripts
+            set -g -x FLAKE_PROJECT_ROOT (git rev-parse --show-toplevel)
+
+            # C/C++:
+            abbr -a cmake-debug "${cmakeDebug}"
+            abbr -a cmake-release "${cmakeRelease}"
+            abbr -a build-debug "${buildDebug}"
+            abbr -a build-release "${buildRelease}"
+            abbr -a debug-clean "${cmakeDebug} && ${buildDebug} && ./cmake-build-debug/masssprings"
+            abbr -a release-clean "${cmakeRelease} && ${buildRelease} && ./cmake-build-release/masssprings"
+            abbr -a debug "${buildDebug} && ./cmake-build-debug/masssprings"
+            abbr -a release "${buildRelease} && ./cmake-build-release/masssprings"
+
+            abbr -a run "${buildRelease} && ./cmake-build-release/masssprings"
+            abbr -a run-clusters "${buildRelease} && ./cmake-build-release/masssprings --output=clusters.puzzle --space=rh --w=6 --h=6 --gx=4 --gy=2 --blocks=4"
+            abbr -a runtests "${buildDebug} && ./cmake-build-debug/tests"
+            abbr -a runbenchs "${buildRelease} && sudo cpupower frequency-set --governor performance && ./cmake-build-release/benchmarks; sudo cpupower frequency-set --governor powersave"
+            abbr -a rungdb "${buildDebug} && gdb --tui ./cmake-build-debug/masssprings"
+            abbr -a runvalgrind "${buildDebug} && valgrind --leak-check=full --show-reachable=no --show-leak-kinds=definite,indirect,possible --track-origins=no --suppressions=valgrind.supp --log-file=valgrind.log ./cmake-build-debug/masssprings && cat valgrind.log"
+            abbr -a runperf "${buildRelease} && perf record -g ./cmake-build-release/masssprings && hotspot ./perf.data"
+            abbr -a runperf-graph "${buildRelease} && perf record -g ./cmake-build-release/benchmarks --benchmark_filter='explore_state_space' && hotspot ./perf.data"
+            abbr -a runperf-space "${buildRelease} && perf record -g ./cmake-build-release/benchmarks --benchmark_filter='explore_rush_hour_puzzle_space' && hotspot ./perf.data"
+            abbr -a runtracy "tracy -a 127.0.0.1 &; ${buildRelease} && sudo -E ./cmake-build-release/masssprings"
+
+            abbr -a runclion "clion ./CMakeLists.txt 2>/dev/null 1>&2 & disown;"
+          '';
+        in
+          builtins.concatStringsSep "\n" [
+            # Launch into pure fish shell
+            ''
+              exec "$(type -p fish)" -C "source ${initProjectShell}"
+            ''
+          ];
+
         # ===========================================================================================
         # Define custom dependencies
         # ===========================================================================================
@@ -98,26 +200,20 @@ rec {
         # - Those which are needed on $PATH during the build, for example cmake and pkg-config
         # - Setup hooks, for example makeWrapper
         # - Interpreters needed by patchShebangs for build scripts (with the --build flag), which can be the case for e.g. perl
+        # NOTE: Do not add compiler here, they are provided by the stdenv
         nativeBuildInputs = with pkgs; [
           # Languages:
-          binutils
-          gcc
+          # binutils
 
           # C/C++:
-          gdb
-          valgrind
-          # gnumake
           cmake
           ninja
-          # pkg-config
-          # clang-tools
-          # compdb
-          # pprof
-          # gprof2dot
+          gdb
+          valgrind
+          kdePackages.kcachegrind
           perf
           hotspot
-          kdePackages.kcachegrind
-          gdbgui
+          # heaptrack
           # renderdoc
         ];
 
@@ -129,15 +225,20 @@ rec {
           raylib
           raygui
           thread-pool
+          boost
 
-          # Debugging
+          # Debugging/Testing/Profiling
           tracy-wayland
           backward-cpp
           libbfd
+          catch2_3
+          gbenchmark
         ];
+
         # ===========================================================================================
         # Define buildable + installable packages
         # ===========================================================================================
+
         package = stdenv.mkDerivation rec {
           inherit buildInputs;
           pname = "masssprings";
@@ -147,19 +248,45 @@ rec {
           nativeBuildInputs = with pkgs; [
             gcc
             cmake
+
+            # Fix the working directory so the auxiliary files are always available
+            makeWrapper
           ];
 
           cmakeFlags = [
+            "-DDISABLE_THREADPOOL=Off"
             "-DDISABLE_TRACY=On"
             "-DDISABLE_BACKWARD=On"
+            "-DDISABLE_TESTS=On"
+            "-DDISABLE_BENCH=On"
           ];
 
+          hardeningDisable = ["all"];
+
+          preConfigure = ''
+            unset NIX_ENFORCE_NO_NATIVE
+          '';
+
           installPhase = ''
+            mkdir -p $out/lib
+            cp ./${pname} $out/lib/
+            cp $src/default.puzzle $out/lib/
+            cp -r $src/fonts $out/lib/fonts
+            cp -r $src/shader $out/lib/shader
+
+            # The wrapper enters the correct working dir, so fonts/shaders/presets are available
             mkdir -p $out/bin
-            cp ./${pname} $out/bin/
-            cp $src/default.puzzle $out/bin/
-            cp -r $src/fonts $out/bin/fonts
-            cp -r $src/shader $out/bin/shader
+            makeWrapper $out/lib/${pname} $out/bin/${pname} --chdir "$out/lib"
+
+            # Generate a .desktop file
+            mkdir -p $out/share/applications
+            cat <<INI > $out/share/applications/${pname}.desktop
+            [Desktop Entry]
+            Terminal=true
+            Name=PuzzleSpaces
+            Exec=$out/bin/${pname} %f
+            Type=Application
+            INI
           '';
         };
 
@@ -178,12 +305,24 @@ rec {
             raylib
             raygui
             thread-pool
+
+            # Disable stacktrace since that's platform dependant and won't cross compile to windows
+            # https://github.com/NixOS/nixpkgs/blob/master/pkgs/development/libraries/boost/generic.nix#L43
+            (boost.override {
+              enableShared = false;
+              extraB2Args = [
+                "--without-stacktrace"
+              ];
+            })
           ];
 
           cmakeFlags = [
             "-DCMAKE_SYSTEM_NAME=Windows"
+            "-DDISABLE_THREADPOOL=Off"
             "-DDISABLE_TRACY=On"
             "-DDISABLE_BACKWARD=On"
+            "-DDISABLE_TESTS=On"
+            "-DDISABLE_BENCH=On"
           ];
 
           installPhase = ''
@@ -205,153 +344,53 @@ rec {
         # Provide environment for "nix develop"
         devShells = {
           default = pkgs.mkShell {
-            inherit nativeBuildInputs buildInputs;
+            inherit nativeBuildInputs buildInputs shellHook;
             name = description;
 
             # =========================================================================================
             # Define environment variables
             # =========================================================================================
 
-            # Custom dynamic libraries:
-            # LD_LIBRARY_PATH = builtins.concatStringsSep ":" [
-            #   # Rust Bevy GUI app:
-            #   # "${pkgs.xorg.libX11}/lib"
-            #   # "${pkgs.xorg.libXcursor}/lib"
-            #   # "${pkgs.xorg.libXrandr}/lib"
-            #   # "${pkgs.xorg.libXi}/lib"
-            #   # "${pkgs.libGL}/lib"
-            #
-            #   # JavaFX app:
-            #   # "${pkgs.libGL}/lib"
-            #   # "${pkgs.gtk3}/lib"
-            #   # "${pkgs.glib.out}/lib"
-            #   # "${pkgs.xorg.libXtst}/lib"
-            # ];
-
             # Dynamic libraries from buildinputs:
             LD_LIBRARY_PATH = nixpkgs.lib.makeLibraryPath buildInputs;
-
-            # =========================================================================================
-            # Define shell environment
-            # =========================================================================================
-
-            # Setup the shell when entering the "nix develop" environment (bash script).
-            shellHook = let
-              mkCmakeScript = type: let
-                typeLower = lib.toLower type;
-              in
-                pkgs.writers.writeFish "cmake-${typeLower}.fish" ''
-                  cd $FLAKE_PROJECT_ROOT
-
-                  # set -g -x CC ${pkgs.clang}/bin/clang
-                  # set -g -x CXX ${pkgs.clang}/bin/clang++
-
-                  echo "Removing build directory ./cmake-build-${typeLower}/"
-                  rm -rf ./cmake-build-${typeLower}
-
-                  echo "Creating build directory"
-                  mkdir cmake-build-${typeLower}
-                  cd cmake-build-${typeLower}
-
-                  echo "Running cmake"
-                  cmake -G "Ninja" \
-                        -DCMAKE_BUILD_TYPE="${type}" \
-                        ..
-
-                  echo "Linking compile_commands.json"
-                  cd ..
-                  ln -sf ./cmake-build-${typeLower}/compile_commands.json ./compile_commands.json
-                '';
-
-              cmakeDebug = mkCmakeScript "Debug";
-              cmakeRelease = mkCmakeScript "Release";
-
-              mkBuildScript = type: let
-                typeLower = lib.toLower type;
-              in
-                pkgs.writers.writeFish "cmake-build.fish" ''
-                  cd $FLAKE_PROJECT_ROOT/cmake-build-${typeLower}
-
-                  echo "Running cmake"
-                  NIX_ENFORCE_NO_NATIVE=0 cmake --build . -j$(nproc)
-                '';
-
-              buildDebug = mkBuildScript "Debug";
-              buildRelease = mkBuildScript "Release";
-
-              # Use this to specify commands that should be ran after entering fish shell
-              initProjectShell = pkgs.writers.writeFish "init-shell.fish" ''
-                echo "Entering \"${description}\" environment..."
-
-                # Determine the project root, used e.g. in cmake scripts
-                set -g -x FLAKE_PROJECT_ROOT (git rev-parse --show-toplevel)
-
-                # C/C++:
-                abbr -a cmake-debug "${cmakeDebug}"
-                abbr -a cmake-release "${cmakeRelease}"
-                abbr -a build-debug "${buildDebug}"
-                abbr -a build-release "${buildRelease}"
-                abbr -a debug "${buildDebug} && ./cmake-build-debug/masssprings"
-                abbr -a release "${buildRelease} && ./cmake-build-release/masssprings"
-                abbr -a debug-clean "${cmakeDebug} && ${buildDebug} && ./cmake-build-debug/masssprings"
-                abbr -a release-clean "${cmakeRelease} && ${buildRelease} && ./cmake-build-release/masssprings"
-                abbr -a rungdb "${buildDebug} && gdb --tui ./cmake-build-debug/masssprings"
-                abbr -a runtracy "tracy -a 127.0.0.1 &; ${buildRelease} && sudo -E ./cmake-build-release/masssprings"
-                abbr -a runvalgrind "${buildDebug} && valgrind --leak-check=full --show-reachable=no --show-leak-kinds=definite,indirect,possible --track-origins=no --suppressions=valgrind.supp --log-file=valgrind.log ./cmake-build-debug/masssprings && cat valgrind.log"
-
-                abbr -a runclion "clion ./CMakeLists.txt 2>/dev/null 1>&2 & disown;"
-              '';
-            in
-              builtins.concatStringsSep "\n" [
-                # Launch into pure fish shell
-                ''
-                  exec "$(type -p fish)" -C "source ${initProjectShell} && abbr -a menu '${pkgs.bat}/bin/bat "${initProjectShell}"'"
-                ''
-              ];
           };
 
-          # TODO: Can't get renderdoc in FHS to work
+          # Provide environment with clang stdenv for "nix develop .#clang"
+          # TODO: Broken. Clang can't find stdlib headers or library headers (raylib, backward, ...).
+          #       Does the clangStdenv not automatically collect the include paths?
+          clang =
+            pkgs.mkShell.override {
+              stdenv = pkgs.clangStdenv;
+            } {
+              inherit shellHook;
+              name = description;
 
-          # FHS environment for renderdoc. Access with "nix develop .#renderdoc".
-          # https://ryantm.github.io/nixpkgs/builders/special/fhs-environments
-          # renderdoc =
-          #   (pkgs.buildFHSEnv {
-          #     name = "renderdoc-env";
-          #
-          #     targetPkgs = pkgs:
-          #       with pkgs; [
-          #         # RenderDoc
-          #         renderdoc
-          #
-          #         # Build tools
-          #         gcc
-          #         cmake
-          #
-          #         # Raylib
-          #         raylib
-          #         libGL
-          #         mesa
-          #
-          #         # X11
-          #         libx11
-          #         libxcursor
-          #         libxrandr
-          #         libxinerama
-          #         libxi
-          #         libxext
-          #         libxfixes
-          #
-          #         # Wayland
-          #         wayland
-          #         wayland-protocols
-          #         libxkbcommon
-          #       ];
-          #
-          #     runScript = "fish";
-          #
-          #     profile = ''
-          #     '';
-          #   }).env;
+              nativeBuildInputs = with pkgs; [
+                cmake
+                ninja
+              ];
+
+              buildInputs = with pkgs; [
+                # C/C++:
+                raylib
+                raygui
+                thread-pool
+                boost
+
+                # Debugging/Testing/Profiling
+                backward-cpp
+                libbfd
+                catch2_3
+                gbenchmark
+              ];
+
+              # =========================================================================================
+              # Define environment variables
+              # =========================================================================================
+
+              # Dynamic libraries from buildinputs:
+              LD_LIBRARY_PATH = nixpkgs.lib.makeLibraryPath buildInputs;
+            };
         };
       }
     );

include/bits.hpp

@@ -0,0 +1,69 @@
+#ifndef BITS_HPP_
+#define BITS_HPP_
+
+#include "util.hpp"
+
+#include <concepts>
+
+template <class T>
+    requires std::unsigned_integral<T>
+// ReSharper disable once CppRedundantInlineSpecifier
+INLINE inline auto create_mask(const uint8_t first, const uint8_t last) -> T
+{
+    // If the mask width is equal the type width return all 1s instead of shifting
+    // as shifting by type-width is undefined behavior.
+    if (static_cast<size_t>(last - first + 1) >= sizeof(T) * 8) {
+        return ~T{0};
+    }
+
+    // Example: first=4, last=7, 7-4+1=4
+    //           1 << 4 = 0b00010000
+    //          32  - 1 = 0b00001111
+    //          31 << 4 = 0b11110000
+    // Subtracting 1 generates a consecutive mask.
+    return ((T{1} << (last - first + 1)) - 1) << first;
+}
+
+template <class T>
+    requires std::unsigned_integral<T>
+// ReSharper disable once CppRedundantInlineSpecifier
+INLINE inline auto clear_bits(T& bits, const uint8_t first, const uint8_t last) -> void
+{
+    const T mask = create_mask<T>(first, last);
+
+    bits = bits & ~mask;
+}
+
+template <class T, class U>
+    requires std::unsigned_integral<T> && std::unsigned_integral<U>
+// ReSharper disable once CppRedundantInlineSpecifier
+INLINE inline auto set_bits(T& bits, const uint8_t first, const uint8_t last, const U value) -> void
+{
+    const T mask = create_mask<T>(first, last);
+
+    // Example: first=4, last=6, value=0b1110,   bits = 0b 01111110
+    //          mask                                  = 0b 01110000
+    //          bits & ~mask                          = 0b 00001110
+    //          value << 4                            = 0b 11100000
+    //          (value << 4) & mask                   = 0b 01100000
+    //          (bits & ~mask) | (value << 4) & mask  = 0b 01101110
+    //                                 Insert position:     ^^^
+    // First clear the bits, then | with the value positioned at the insertion point.
+    // The value may be larger than [first, last], extra bits are ignored.
+    bits = (bits & ~mask) | ((static_cast<T>(value) << first) & mask);
+}
+
+template <class T>
+    requires std::unsigned_integral<T>
+// ReSharper disable once CppRedundantInlineSpecifier
+INLINE inline auto get_bits(const T bits, const uint8_t first, const uint8_t last) -> T
+{
+    const T mask = create_mask<T>(first, last);
+
+    // We can >> without sign extension because T is unsigned_integral
+    return (bits & mask) >> first;
+}
+
+auto print_bitmap(uint64_t bitmap, uint8_t w, uint8_t h, const std::string& title) -> void;
+
+#endif

include/config.hpp

@@ -3,12 +3,31 @@
 
 #include <raylib.h>
 
-#define THREADPOOL // Enable physics threadpool
+
+// TODO: Using the octree from the last frame completely breaks the physics :/
+// #define ASYNC_OCTREE
 
 // Gets set by CMake
+// #define THREADPOOL // Enable physics threadpool
 // #define BACKWARD   // Enable pretty stack traces
 // #define TRACY      // Enable tracy profiling support
 
+#ifdef TRACY
+#include <tracy/Tracy.hpp>
+#endif
+
+#if defined(_WIN32)
+#define NOGDI  // All GDI defines and routines
+#define NOUSER // All USER defines and routines
+#endif
+#define BS_THREAD_POOL_NATIVE_EXTENSIONS
+// ReSharper disable once CppUnusedIncludeDirective
+#include <BS_thread_pool.hpp>
+#if defined(_WIN32) // raylib uses these names as function parameters
+#undef near
+#undef far
+#endif
+
 // Window
 constexpr int INITIAL_WIDTH = 600;
 constexpr int INITIAL_HEIGHT = 600;
@@ -29,7 +48,7 @@ constexpr float CAMERA_FOV = 90.0;
 constexpr float FOV_SPEED = 1.0;
 constexpr float MIN_FOV = 10.0;
 constexpr float MAX_FOV = 180.0;
-constexpr float CAMERA_DISTANCE = 20.0;
+constexpr float CAMERA_DISTANCE = 150.0;
 constexpr float ZOOM_SPEED = 2.5;
 constexpr float MIN_CAMERA_DISTANCE = 2.0;
 constexpr float MAX_CAMERA_DISTANCE = 2000.0;

include/graph_distances.hpp

@@ -21,4 +21,4 @@ public:
     [[nodiscard]] auto get_shortest_path(size_t source) const -> std::vector<size_t>;
 };
 
-#endif
+#endif

include/input_handler.hpp

@@ -1,5 +1,5 @@
-#ifndef INPUT_HPP_
-#define INPUT_HPP_
+#ifndef INPUT_HANDLER_HPP_
+#define INPUT_HANDLER_HPP_
 
 #include "orbit_camera.hpp"
 #include "state_manager.hpp"
@@ -19,7 +19,7 @@ struct show_yes_no_message
 {
     std::string title;
     std::string message;
-    std::function<void(void)> on_yes;
+    std::function<void()> on_yes;
 };
 
 struct show_save_preset_window
@@ -78,7 +78,7 @@ public:
 
     // Camera
     bool camera_lock = true;
-    bool camera_mass_center_lock = false;
+    bool camera_mass_center_lock = true;
     bool camera_panning = false;
     bool camera_rotating = false;
 

include/load_save.hpp

@@ -0,0 +1,15 @@
+#ifndef LOAD_SAVE_HPP_
+#define LOAD_SAVE_HPP_
+
+#include "puzzle.hpp"
+
+#include <string>
+
+auto parse_preset_file(const std::string& preset_file) -> std::pair<std::vector<puzzle>, std::vector<std::string>>;
+auto append_preset_file(const std::string& preset_file, const std::string& preset_name, const puzzle& p) -> bool;
+auto append_preset_file_quiet(const std::string& preset_file,
+                              const std::string& preset_name,
+                              const puzzle& p,
+                              bool validate) -> bool;
+
+#endif

include/mass_spring_system.hpp

@@ -2,49 +2,14 @@
 #define MASS_SPRING_SYSTEM_HPP_
 
 #include "octree.hpp"
-#include "util.hpp"
 #include "config.hpp"
 
+#include <optional>
 #include <raylib.h>
-#include <raymath.h>
-
-#ifdef THREADPOOL
-#if defined(_WIN32)
-#define NOGDI  // All GDI defines and routines
-#define NOUSER // All USER defines and routines
-#endif
-#define BS_THREAD_POOL_NATIVE_EXTENSIONS
-#include <BS_thread_pool.hpp>
-#if defined(_WIN32) // raylib uses these names as function parameters
-#undef near
-#undef far
-#endif
-#endif
 
 class mass_spring_system
 {
 public:
-    class mass
-    {
-    public:
-        Vector3 position = Vector3Zero();
-        Vector3 previous_position = Vector3Zero(); // for verlet integration
-        Vector3 velocity = Vector3Zero();
-        Vector3 force = Vector3Zero();
-
-    public:
-        mass() = delete;
-
-        explicit mass(const Vector3 _position)
-            : position(_position), previous_position(_position) {}
-
-    public:
-        auto clear_force() -> void;
-        auto calculate_velocity(float delta_time) -> void;
-        auto calculate_position(float delta_time) -> void;
-        auto verlet_update(float delta_time) -> void;
-    };
-
     class spring
     {
     public:
@@ -54,61 +19,45 @@ public:
     public:
         spring(const size_t _a, const size_t _b)
             : a(_a), b(_b) {}
-
-    public:
-        static auto calculate_spring_force(mass& _a, mass& _b) -> void;
     };
 
-private:
-    #ifdef THREADPOOL
-    BS::thread_pool<> threads;
-    #endif
-
 public:
+    static constexpr int SMALL_TASK_BLOCK_SIZE = 256;
+    static constexpr int LARGE_TASK_BLOCK_SIZE = 256;
+
     octree tree;
 
     // This is the main ownership of all the states/masses/springs.
-    std::vector<mass> masses;
+    std::vector<Vector3> positions;
+    std::vector<Vector3> previous_positions; // for verlet integration
+    std::vector<Vector3> velocities;
+    std::vector<Vector3> forces;
+
     std::vector<spring> springs;
 
 public:
-    mass_spring_system()
-    #ifdef THREADPOOL
-        : threads(std::thread::hardware_concurrency() - 1, set_thread_name)
-    #endif
-    {
-        infoln("Using Barnes-Hut + Octree repulsion force calculation.");
-
-        #ifdef THREADPOOL
-        infoln("Thread-pool: {} threads.", threads.get_thread_count());
-        #else
-        infoln("Thread-pool: Disabled.");
-        #endif
-    }
+    mass_spring_system() {}
 
     mass_spring_system(const mass_spring_system& copy) = delete;
     auto operator=(const mass_spring_system& copy) -> mass_spring_system& = delete;
     mass_spring_system(mass_spring_system& move) = delete;
     auto operator=(mass_spring_system&& move) -> mass_spring_system& = delete;
 
-private:
-    #ifdef THREADPOOL
-    static auto set_thread_name(size_t idx) -> void;
-    #endif
-
-    auto build_octree() -> void;
-
 public:
     auto clear() -> void;
     auto add_mass() -> void;
     auto add_spring(size_t a, size_t b) -> void;
 
     auto clear_forces() -> void;
-    auto calculate_spring_forces() -> void;
-    auto calculate_repulsion_forces() -> void;
-    auto verlet_update(float delta_time) -> void;
+    auto calculate_spring_force(size_t s) -> void;
+    auto calculate_spring_forces(std::optional<BS::thread_pool<>* const> thread_pool = std::nullopt) -> void;
+    auto calculate_repulsion_forces(std::optional<BS::thread_pool<>* const> thread_pool = std::nullopt) -> void;
+    auto integrate_velocity(size_t m, float dt) -> void;
+    auto integrate_position(size_t m, float dt) -> void;
+    auto verlet_update(size_t m, float dt) -> void;
+    auto update(float dt, std::optional<BS::thread_pool<>* const> thread_pool = std::nullopt) -> void;
 
-    auto center_masses() -> void;
+    auto center_masses(std::optional<BS::thread_pool<>* const> thread_pool = std::nullopt) -> void;
 };
 
 #endif

include/octree.hpp

@@ -2,9 +2,10 @@
 #define OCTREE_HPP_
 
 #include <array>
+#include <vector>
+
 #include <raylib.h>
 #include <raymath.h>
-#include <vector>
 
 class octree
 {
@@ -31,22 +32,20 @@ public:
 public:
     octree() = default;
 
-    octree(const octree& copy) = delete;
-    auto operator=(const octree& copy) -> octree& = delete;
-    octree(octree&& move) = delete;
-    auto operator=(octree&& move) -> octree& = delete;
+    // octree(const octree& copy) = delete;
+    // auto operator=(const octree& copy) -> octree& = delete;
+    // octree(octree&& move) = delete;
+    // auto operator=(octree&& move) -> octree& = delete;
 
 public:
-    auto create_empty_leaf(const Vector3& box_min, const Vector3& box_max) -> int;
-
     [[nodiscard]] auto get_octant(int node_idx, const Vector3& pos) const -> int;
-
     [[nodiscard]] auto get_child_bounds(int node_idx, int octant) const
         -> std::pair<Vector3, Vector3>;
-
+    auto create_empty_leaf(const Vector3& box_min, const Vector3& box_max) -> int;
     auto insert(int node_idx, int mass_id, const Vector3& pos, float mass, int depth) -> void;
+    static auto build_octree(octree& t, const std::vector<Vector3>& positions) -> void;
 
     [[nodiscard]] auto calculate_force(int node_idx, const Vector3& pos) const -> Vector3;
 };
 
-#endif
+#endif

include/orbit_camera.hpp

@@ -28,4 +28,4 @@ public:
                 bool mass_center_lock) -> void;
 };
 
-#endif
+#endif

include/renderer.hpp

@@ -1,8 +1,8 @@
 #ifndef RENDERER_HPP_
 #define RENDERER_HPP_
 
-#include "orbit_camera.hpp"
 #include "config.hpp"
+#include "orbit_camera.hpp"
 #include "input_handler.hpp"
 #include "state_manager.hpp"
 #include "user_interface.hpp"
@@ -18,7 +18,7 @@ private:
     user_interface& gui;
 
     const orbit_camera& camera;
-    RenderTexture render_target =
+    RenderTexture graph_target =
         LoadRenderTexture(GetScreenWidth() / 2, GetScreenHeight() - MENU_HEIGHT);
 
     // TODO: Those should be moved to the user_interface.h
@@ -81,7 +81,7 @@ public:
 
     ~renderer()
     {
-        UnloadRenderTexture(render_target);
+        UnloadRenderTexture(graph_target);
         UnloadRenderTexture(klotski_target);
         UnloadRenderTexture(menu_target);
 

include/state_manager.hpp

@@ -2,12 +2,12 @@
 #define STATE_MANAGER_HPP_
 
 #include "graph_distances.hpp"
+#include "load_save.hpp"
 #include "threaded_physics.hpp"
 #include "puzzle.hpp"
 
-#include <stack>
-#include <unordered_map>
-#include <unordered_set>
+#include <boost/unordered/unordered_flat_map.hpp>
+#include <boost/unordered/unordered_flat_set.hpp>
 
 class state_manager
 {
@@ -16,26 +16,23 @@ private:
 
     std::string preset_file;
     size_t current_preset = 0;
-    std::vector<puzzle> preset_states = {puzzle(4, 5, 9, 9, false)};
-    std::vector<std::string> preset_comments = {"Empty"};
+    std::vector<puzzle> preset_states = {puzzle(4, 5, 0, 0, true, false)};
+    std::vector<std::string> preset_comments = {"# Empty"};
 
     // State storage (store states twice for bidirectional lookup).
     // Everything else should only store indices to state_pool.
 
     std::vector<puzzle> state_pool; // Indices are equal to mass_springs mass indices
-    std::unordered_map<puzzle, size_t> state_indices; // Maps states to indices
-    std::vector<std::pair<size_t, size_t>>
-        links; // Indices are equal to mass_springs springs indices
+    boost::unordered_flat_map<puzzle, size_t, puzzle_hasher> state_indices; // Maps states to indices
+    std::vector<std::pair<size_t, size_t>> links; // Indices are equal to mass_springs springs indices
 
-    graph_distances node_target_distances;      // Buffered and reused if the graph doesn't change
-    std::unordered_set<size_t> winning_indices; // Indices of all states where the board is solved
-    std::vector<size_t>
-        winning_path; // Ordered list of node indices leading to the nearest solved state
-    std::unordered_set<size_t>
-        path_indices; // For faster lookup if a vertex is part of the path in renderer
+    graph_distances node_target_distances; // Buffered and reused if the graph doesn't change
+    boost::unordered_flat_set<size_t> winning_indices; // Indices of all states where the board is solved
+    std::vector<size_t> winning_path; // Ordered list of node indices leading to the nearest solved state
+    boost::unordered_flat_set<size_t> path_indices; // For faster lookup if a vertex is part of the path in renderer
 
-    std::vector<size_t> move_history; // Moves between the starting state and the current state
-    std::unordered_map<size_t, int> visit_counts; // How often each state was visited
+    std::vector<size_t> move_history;                    // Moves between the starting state and the current state
+    boost::unordered_flat_map<size_t, int> visit_counts; // How often each state was visited
 
     size_t starting_state_index = 0;
     size_t current_state_index = 0;
@@ -45,10 +42,10 @@ private:
     bool edited = false;
 
 public:
-    state_manager(threaded_physics& _physics, const std::string& _preset_file) : physics(_physics)
+    state_manager(threaded_physics& _physics, const std::string& _preset_file)
+        : physics(_physics), preset_file(_preset_file)
     {
-        parse_preset_file(_preset_file);
-        load_preset(0);
+        reload_preset_file();
     }
 
     state_manager(const state_manager& copy) = delete;
@@ -97,15 +94,13 @@ private:
 
 public:
     // Presets
-
-    auto parse_preset_file(const std::string& _preset_file) -> bool;
-    auto append_preset_file(const std::string& preset_name) -> bool;
+    auto save_current_to_preset_file(const std::string& preset_comment) -> void;
+    auto reload_preset_file() -> void;
     auto load_preset(size_t preset) -> void;
     auto load_previous_preset() -> void;
     auto load_next_preset() -> void;
 
     // Update current_state
-
     auto update_current_state(const puzzle& p) -> void;
     auto edit_starting_state(const puzzle& p) -> void;
     auto goto_starting_state() -> void;
@@ -115,7 +110,6 @@ public:
     auto goto_closest_target_state() -> void;
 
     // Update graph
-
     auto populate_graph() -> void;
     auto clear_graph_and_add_current(const puzzle& p) -> void;
     auto clear_graph_and_add_current() -> void;
@@ -124,7 +118,6 @@ public:
     auto populate_winning_path() -> void;
 
     // Index mapping
-
     [[nodiscard]] auto get_index(const puzzle& state) const -> size_t;
     [[nodiscard]] auto get_current_index() const -> size_t;
     [[nodiscard]] auto get_starting_index() const -> size_t;
@@ -138,10 +131,10 @@ public:
     [[nodiscard]] auto get_link_count() const -> size_t;
     [[nodiscard]] auto get_path_length() const -> size_t;
     [[nodiscard]] auto get_links() const -> const std::vector<std::pair<size_t, size_t>>&;
-    [[nodiscard]] auto get_winning_indices() const -> const std::unordered_set<size_t>&;
-    [[nodiscard]] auto get_visit_counts() const -> const std::unordered_map<size_t, int>&;
+    [[nodiscard]] auto get_winning_indices() const -> const boost::unordered_flat_set<size_t>&;
+    [[nodiscard]] auto get_visit_counts() const -> const boost::unordered_flat_map<size_t, int>&;
     [[nodiscard]] auto get_winning_path() const -> const std::vector<size_t>&;
-    [[nodiscard]] auto get_path_indices() const -> const std::unordered_set<size_t>&;
+    [[nodiscard]] auto get_path_indices() const -> const boost::unordered_flat_set<size_t>&;
     [[nodiscard]] auto get_current_visits() const -> int;
     [[nodiscard]] auto get_current_preset() const -> size_t;
     [[nodiscard]] auto get_preset_count() const -> size_t;

include/threaded_physics.hpp

@@ -1,6 +1,8 @@
 #ifndef PHYSICS_HPP_
 #define PHYSICS_HPP_
 
+#include "config.hpp"
+
 #include <atomic>
 #include <condition_variable>
 #include <mutex>
@@ -11,13 +13,10 @@
 #include <variant>
 #include <vector>
 
-#ifdef TRACY
-#include <tracy/Tracy.hpp>
-#endif
-
 class threaded_physics
 {
-    struct add_mass {};
+    struct add_mass
+    {};
 
     struct add_spring
     {
@@ -25,24 +24,25 @@ class threaded_physics
         size_t b;
     };
 
-    struct clear_graph {};
+    struct clear_graph
+    {};
 
     using command = std::variant<add_mass, add_spring, clear_graph>;
 
     struct physics_state
     {
-        #ifdef TRACY
+#ifdef TRACY
         TracyLockable(std::mutex, command_mtx);
-        #else
+#else
         std::mutex command_mtx;
-        #endif
+#endif
         std::queue<command> pending_commands;
 
-        #ifdef TRACY
+#ifdef TRACY
         TracyLockable(std::mutex, data_mtx);
-        #else
+#else
         std::mutex data_mtx;
-        #endif
+#endif
         std::condition_variable_any data_ready_cnd;
         std::condition_variable_any data_consumed_cnd;
         Vector3 mass_center = Vector3Zero();
@@ -57,14 +57,17 @@ class threaded_physics
     };
 
 private:
+    std::optional<BS::thread_pool<>* const> thread_pool;
     std::thread physics;
 
 public:
     physics_state state;
 
 public:
-    threaded_physics()
-        : physics(physics_thread, std::ref(state)) {}
+    explicit threaded_physics(
+        const std::optional<BS::thread_pool<>* const> _thread_pool = std::nullopt)
+        : thread_pool(_thread_pool), physics(physics_thread, std::ref(state), std::ref(thread_pool))
+    {}
 
     threaded_physics(const threaded_physics& copy) = delete;
     auto operator=(const threaded_physics& copy) -> threaded_physics& = delete;
@@ -80,16 +83,19 @@ public:
     }
 
 private:
-    static auto physics_thread(physics_state& state) -> void;
+#ifdef ASYNC_OCTREE
+    static auto set_octree_pool_thread_name(size_t idx) -> void;
+#endif
+
+    static auto physics_thread(physics_state& state,
+                               std::optional<BS::thread_pool<>* const> thread_pool) -> void;
 
 public:
-    auto add_mass_cmd() -> void;
-
-    auto add_spring_cmd(size_t a, size_t b) -> void;
-
     auto clear_cmd() -> void;
-
-    auto add_mass_springs_cmd(size_t num_masses, const std::vector<std::pair<size_t, size_t>>& springs) -> void;
+    auto add_mass_cmd() -> void;
+    auto add_spring_cmd(size_t a, size_t b) -> void;
+    auto add_mass_springs_cmd(size_t num_masses,
+                              const std::vector<std::pair<size_t, size_t>>& springs) -> void;
 };
 
-#endif
+#endif

include/user_interface.hpp

@@ -1,5 +1,5 @@
-#ifndef GUI_HPP_
-#define GUI_HPP_
+#ifndef USER_INTERFACE_HPP_
+#define USER_INTERFACE_HPP_
 
 #include "orbit_camera.hpp"
 #include "config.hpp"
@@ -89,7 +89,7 @@ private:
 
     grid board_grid =
         grid(0, MENU_HEIGHT, GetScreenWidth() / 2, GetScreenHeight() - MENU_HEIGHT,
-             state.get_current_state().width, state.get_current_state().height, BOARD_PADDING);
+             state.get_current_state().get_width(), state.get_current_state().get_height(), BOARD_PADDING);
 
     grid graph_overlay_grid = grid(GetScreenWidth() / 2, MENU_HEIGHT, 200, 100, 1, 4, MENU_PAD);
 
@@ -100,11 +100,11 @@ private:
 
     std::string message_title;
     std::string message_message;
-    std::function<void(void)> yes_no_handler;
+    std::function<void()> yes_no_handler;
     bool ok_message = false;
     bool yes_no_message = false;
     bool save_window = false;
-    std::array<char, 256> preset_name = {};
+    std::array<char, 256> preset_comment = {};
     bool help_window = false;
 
 public:

include/util.hpp

@@ -3,7 +3,77 @@
 
 #include <iostream>
 #include <raylib.h>
-#include <raymath.h>
+
+#define INLINE __attribute__((always_inline))
+#define PACKED __attribute__((packed))
+
+// std::variant visitor
+
+// https://en.cppreference.com/w/cpp/utility/variant/visit
+template <class... Ts>
+struct overloads : Ts...
+{
+    using Ts::operator()...;
+};
+
+// Enums
+
+enum dir : uint8_t
+{
+    nor = 1 << 0,
+    eas = 1 << 1,
+    sou = 1 << 2,
+    wes = 1 << 3,
+};
+
+// Ansi
+
+enum class ctrl : uint8_t
+{
+    reset = 0,
+    bold_bright = 1,
+    underline = 4,
+    inverse = 7,
+    bold_bright_off = 21,
+    underline_off = 24,
+    inverse_off = 27
+};
+
+enum class fg : uint8_t
+{
+    black = 30,
+    red = 31,
+    green = 32,
+    yellow = 33,
+    blue = 34,
+    magenta = 35,
+    cyan = 36,
+    white = 37
+};
+
+enum class bg : uint8_t
+{
+    black = 40,
+    red = 41,
+    green = 42,
+    yellow = 43,
+    blue = 44,
+    magenta = 45,
+    cyan = 46,
+    white = 47
+};
+
+inline auto ansi_bold_fg(const fg color) -> std::string
+{
+    return std::format("\033[{};{}m", static_cast<int>(ctrl::bold_bright), static_cast<int>(color));
+}
+
+inline auto ansi_reset() -> std::string
+{
+    return std::format("\033[{}m", static_cast<int>(ctrl::reset));
+}
+
+// Output
 
 inline auto operator<<(std::ostream& os, const Vector2& v) -> std::ostream&
 {
@@ -17,86 +87,33 @@ inline auto operator<<(std::ostream& os, const Vector3& v) -> std::ostream&
     return os;
 }
 
-// https://en.cppreference.com/w/cpp/utility/variant/visit
-template <class... Ts>
-struct overloads : Ts...
-{
-    using Ts::operator()...;
-};
-
-enum direction
-{
-    nor = 1 << 0,
-    eas = 1 << 1,
-    sou = 1 << 2,
-    wes = 1 << 3,
-};
-
-enum ctrl
-{
-    reset = 0,
-    bold_bright = 1,
-    underline = 4,
-    inverse = 7,
-    bold_bright_off = 21,
-    underline_off = 24,
-    inverse_off = 27
-};
-
-enum fg
-{
-    fg_black = 30,
-    fg_red = 31,
-    fg_green = 32,
-    fg_yellow = 33,
-    fg_blue = 34,
-    fg_magenta = 35,
-    fg_cyan = 36,
-    fg_white = 37
-};
-
-enum bg
-{
-    bg_black = 40,
-    bg_red = 41,
-    bg_green = 42,
-    bg_yellow = 43,
-    bg_blue = 44,
-    bg_magenta = 45,
-    bg_cyan = 46,
-    bg_white = 47
-};
-
-inline auto ansi_bold_fg(const fg color) -> std::string
-{
-    return std::format("\033[1;{}m", static_cast<int>(color));
-}
-
-inline auto ansi_reset() -> std::string
-{
-    return "\033[0m";
-}
-
 // std::println doesn't work with mingw
+template <typename... Args>
+auto traceln(std::format_string<Args...> fmt, Args&&... args) -> void
+{
+    std::cout << std::format("[{}TRACE{}]: ", ansi_bold_fg(fg::cyan), ansi_reset()) << std::format(
+        fmt, std::forward<Args>(args)...) << std::endl;
+}
+
 template <typename... Args>
 auto infoln(std::format_string<Args...> fmt, Args&&... args) -> void
 {
-    std::cout << std::format("[{}INFO{}]: ", ansi_bold_fg(fg_blue), ansi_reset())
-              << std::format(fmt, std::forward<Args>(args)...) << std::endl;
+    std::cout << std::format("[{}INFO{}]: ", ansi_bold_fg(fg::blue), ansi_reset()) << std::format(
+        fmt, std::forward<Args>(args)...) << std::endl;
 }
 
 template <typename... Args>
 auto warnln(std::format_string<Args...> fmt, Args&&... args) -> void
 {
-    std::cout << std::format("[{}WARNING{}]: ", ansi_bold_fg(fg_yellow), ansi_reset())
-              << std::format(fmt, std::forward<Args>(args)...) << std::endl;
+    std::cout << std::format("[{}WARNING{}]: ", ansi_bold_fg(fg::yellow), ansi_reset()) << std::format(
+        fmt, std::forward<Args>(args)...) << std::endl;
 }
 
 template <typename... Args>
 auto errln(std::format_string<Args...> fmt, Args&&... args) -> void
 {
-    std::cout << std::format("[{}ERROR{}]: ", ansi_bold_fg(fg_red), ansi_reset())
-              << std::format(fmt, std::forward<Args>(args)...) << std::endl;
+    std::cout << std::format("[{}ERROR{}]: ", ansi_bold_fg(fg::red), ansi_reset()) << std::format(
+        fmt, std::forward<Args>(args)...) << std::endl;
 }
 
 #endif

src/bits.cpp

@@ -0,0 +1,15 @@
+#include "bits.hpp"
+
+auto print_bitmap(const uint64_t bitmap, const uint8_t w, const uint8_t h, const std::string& title) -> void {
+    traceln("{}:", title);
+    traceln("{}", std::string(2 * w - 1, '='));
+    for (size_t y = 0; y < w; ++y) {
+        std::cout << "         ";
+        for (size_t x = 0; x < h; ++x) {
+            std::cout << static_cast<int>(get_bits(bitmap, y * w + x, y * h + x)) << " ";
+        }
+        std::cout << "\n";
+    }
+    std::cout << std::flush;
+    traceln("{}", std::string(2 * w - 1, '='));
+}

src/graph_distances.cpp

@@ -2,10 +2,6 @@
 
 #include <queue>
 
-#ifdef TRACY
-    #include <tracy/Tracy.hpp>
-#endif
-
 auto graph_distances::clear() -> void
 {
     distances.clear();

src/input_handler.cpp

@@ -162,11 +162,12 @@ auto input_handler::select_block() -> void
 auto input_handler::start_add_block() -> void
 {
     const puzzle& current = state.get_current_state();
-    if (!editing || current.try_get_block(hov_x, hov_y) || has_block_add_xy) {
+    if (!editing || current.try_get_block(hov_x, hov_y) || has_block_add_xy || current.block_count() >=
+        puzzle::MAX_BLOCKS) {
         return;
     }
 
-    if (hov_x >= 0 && hov_x < current.width && hov_y >= 0 && hov_y < current.height) {
+    if (hov_x >= 0 && hov_x < current.get_width() && hov_y >= 0 && hov_y < current.get_height()) {
         block_add_x = hov_x;
         block_add_y = hov_y;
         has_block_add_xy = true;
@@ -187,7 +188,8 @@ auto input_handler::clear_add_block() -> void
 auto input_handler::add_block() -> void
 {
     const puzzle& current = state.get_current_state();
-    if (!editing || current.try_get_block(hov_x, hov_y) || !has_block_add_xy) {
+    if (!editing || current.try_get_block(hov_x, hov_y) || !has_block_add_xy || current.block_count() >=
+        puzzle::MAX_BLOCKS) {
         return;
     }
 
@@ -323,7 +325,7 @@ auto input_handler::move_block_eas() -> void
 
 auto input_handler::print_state() const -> void
 {
-    infoln("State: \"{}\"", state.get_current_state().state);
+    infoln("State: \"{}\"", state.get_current_state().string_repr());
 }
 
 auto input_handler::load_previous_preset() -> void
@@ -332,7 +334,7 @@ auto input_handler::load_previous_preset() -> void
         return;
     }
 
-    const auto handler = [&]()
+    const auto handler = [&]
     {
         block_add_x = -1;
         block_add_y = -1;
@@ -353,7 +355,7 @@ auto input_handler::load_next_preset() -> void
         return;
     }
 
-    const auto handler = [&]()
+    const auto handler = [&]
     {
         block_add_x = -1;
         block_add_y = -1;
@@ -370,7 +372,7 @@ auto input_handler::load_next_preset() -> void
 
 auto input_handler::goto_starting_state() -> void
 {
-    const auto handler = [&]()
+    const auto handler = [&]
     {
         state.goto_starting_state();
         sel_x = 0;
@@ -387,7 +389,7 @@ auto input_handler::populate_graph() const -> void
 
 auto input_handler::clear_graph() -> void
 {
-    const auto handler = [&]()
+    const auto handler = [&]
     {
         state.clear_graph_and_add_current();
     };
@@ -433,7 +435,7 @@ auto input_handler::goto_previous_state() const -> void
 auto input_handler::toggle_restricted_movement() const -> void
 {
     const puzzle& current = state.get_current_state();
-    const std::optional<puzzle>& next = current.try_toggle_restricted();
+    const std::optional<puzzle>& next = current.toggle_restricted();
 
     if (!editing || !next) {
         return;
@@ -471,7 +473,7 @@ auto input_handler::remove_board_column() const -> void
     const puzzle& current = state.get_current_state();
     const std::optional<puzzle>& next = current.try_remove_column();
 
-    if (!editing || current.width <= puzzle::MIN_WIDTH || !next) {
+    if (!editing || current.get_width() <= puzzle::MIN_WIDTH || !next) {
         return;
     }
 
@@ -483,7 +485,7 @@ auto input_handler::add_board_column() const -> void
     const puzzle& current = state.get_current_state();
     const std::optional<puzzle>& next = current.try_add_column();
 
-    if (!editing || current.width >= puzzle::MAX_WIDTH || !next) {
+    if (!editing || current.get_width() >= puzzle::MAX_WIDTH || !next) {
         return;
     }
 
@@ -495,7 +497,7 @@ auto input_handler::remove_board_row() const -> void
     const puzzle& current = state.get_current_state();
     const std::optional<puzzle>& next = current.try_remove_row();
 
-    if (!editing || current.height <= puzzle::MIN_HEIGHT || !next) {
+    if (!editing || current.get_height() <= puzzle::MIN_HEIGHT || !next) {
         return;
     }
 
@@ -507,7 +509,7 @@ auto input_handler::add_board_row() const -> void
     const puzzle& current = state.get_current_state();
     const std::optional<puzzle>& next = current.try_add_row();
 
-    if (!editing || current.height >= puzzle::MAX_HEIGHT || !next) {
+    if (!editing || current.get_height() >= puzzle::MAX_HEIGHT || !next) {
         return;
     }
 
@@ -528,7 +530,7 @@ auto input_handler::toggle_editing() -> void
 auto input_handler::clear_goal() const -> void
 {
     const puzzle& current = state.get_current_state();
-    const std::optional<puzzle>& next = current.try_clear_goal();
+    const std::optional<puzzle>& next = current.clear_goal();
 
     if (!editing || !next) {
         return;

src/load_save.cpp

@@ -0,0 +1,79 @@
+#include "load_save.hpp"
+
+#include <fstream>
+
+auto parse_preset_file(const std::string& preset_file) -> std::pair<std::vector<puzzle>, std::vector<std::string>>
+{
+    std::fstream file(preset_file, std::ios::in);
+    if (!file) {
+        infoln("Preset file \"{}\" couldn't be opened.", preset_file);
+        return {};
+    }
+
+    std::string line;
+    std::vector<std::string> comment_lines;
+    std::vector<std::string> preset_lines;
+    while (std::getline(file, line)) {
+        if (line.starts_with("S")) {
+            preset_lines.push_back(line);
+        } else if (line.starts_with("#")) {
+            comment_lines.push_back(line);
+        }
+    }
+
+    if (preset_lines.empty() || comment_lines.size() != preset_lines.size()) {
+        infoln("Preset file \"{}\" couldn't be opened.", preset_file);
+        return {};
+    }
+
+    std::vector<puzzle> preset_states;
+    for (const auto& preset : preset_lines) {
+        // Each char is a bit
+        const puzzle& p = puzzle(preset);
+
+        if (const std::optional<std::string>& reason = p.try_get_invalid_reason()) {
+            infoln("Preset file \"{}\" contained invalid presets: {}", preset_file, *reason);
+            return {};
+        }
+        preset_states.emplace_back(p);
+    }
+
+    infoln("Loaded {} presets from \"{}\".", preset_lines.size(), preset_file);
+
+    return {preset_states, comment_lines};
+}
+
+auto append_preset_file(const std::string& preset_file, const std::string& preset_name, const puzzle& p) -> bool
+{
+    infoln(R"(Saving preset "{}" to "{}")", preset_name, preset_file);
+
+    if (p.try_get_invalid_reason()) {
+        return false;
+    }
+
+    std::fstream file(preset_file, std::ios_base::app | std::ios_base::out);
+    if (!file) {
+        infoln("Preset file \"{}\" couldn't be opened.", preset_file);
+        return false;
+    }
+
+    file << "\n# " << preset_name << "\n" << p.string_repr() << std::flush;
+
+    return true;
+}
+
+auto append_preset_file_quiet(const std::string& preset_file, const std::string& preset_name, const puzzle& p, const bool validate) -> bool
+{
+    if (validate && p.try_get_invalid_reason()) {
+        return false;
+    }
+
+    std::fstream file(preset_file, std::ios_base::app | std::ios_base::out);
+    if (!file) {
+        return false;
+    }
+
+    file << "\n# " << preset_name << "\n" << p.string_repr() << std::flush;
+
+    return true;
+}

src/main.cpp

@@ -4,62 +4,59 @@
 
 #include "config.hpp"
 #include "input_handler.hpp"
-#include "mass_spring_system.hpp"
 #include "threaded_physics.hpp"
 #include "renderer.hpp"
 #include "state_manager.hpp"
 #include "user_interface.hpp"
 
-#ifdef TRACY
-    #include <tracy/Tracy.hpp>
+#include <filesystem>
+
+#ifndef WIN32
+#include <boost/program_options.hpp>
+namespace po = boost::program_options;
 #endif
 
-// TODO: Add some popups (my split between input.cpp/gui.cpp makes this ugly)
-//       - Clear graph: Notify that this will clear the visited states and move
-//         history
-//       - Reset state: Notify that this will reset the move count
-
-// TODO: Reduce memory usage
-//       - The memory model of the puzzle board is terrible (bitboards?)
-
-// TODO: Improve solver
-//       - Move discovery is terrible
-//         - Instead of trying each direction for each block, determine the
-//           possible moves more efficiently (requires a different memory model)
-//       - Implement state discovery/enumeration
-//         - Find all possible initial board states (single one for each
-//           possible statespace). Currently wer're just finding all states
-//           given the initial state
-//         - Would allow to generate random puzzles with a certain move count
-
-// TODO: Move selection accordingly when undoing moves (need to diff two states
-//       and get the moved blocks)
+// TODO: Implement state discovery/enumeration
+//       - Find all possible initial board states (single one for each possible statespace).
+//         Currently wer're just finding all states given the initial state
+//       - Would allow to generate random puzzles with a certain move count
+// TODO: Export cluster to graphviz
+// TODO: Fix naming:
+//       - Target: The block that has to leave the board to win
+//       - Goal: The opening in the board for the target
+//       - Puzzle (not board or state): A puzzle configuration (width, height, goal_x, goal_y, restricted, goal)
+//       - Block: A puzzle block (x, y, width, height, target, immovable)
+//       - Puzzle State: A specific puzzle state (width, height, goal_x, goal_y, restricted, goal, blocks)
+//       - Cluster: A graph of puzzle states connected by moves, generated from a specific Puzzle State
+//       - Puzzle Space: A number of Clusters generated from a generic Puzzle
+// TODO: Add state space generation time to debug overlay
+// TODO: Move selection accordingly when undoing moves (need to diff two states and get the moved blocks)
 
 // TODO: Click states in the graph to display them in the board
 
-// NOTE: Tracy uses a huge amount of memory. For longer testing disable Tracy.
-
-auto main(int argc, char* argv[]) -> int
+#ifdef THREADPOOL
+auto set_pool_thread_name(size_t idx) -> void
 {
-    std::string preset_file;
-    if (argc != 2) {
-        preset_file = "default.puzzle";
-    } else {
-        preset_file = argv[1];
-    }
+    BS::this_thread::set_os_thread_name(std::format("worker-{}", idx));
+}
 
-#ifdef BACKWARD
-    infoln("Backward stack-traces enabled.");
+BS::thread_pool<> threads(std::thread::hardware_concurrency() - 2, set_pool_thread_name);
+constexpr std::optional<BS::thread_pool<>* const> thread_pool = &threads;
 #else
-    infoln("Backward stack-traces disabled.");
+constexpr std::optional<BS::thread_pool<>* const> thread_pool = std::nullopt;
 #endif
 
-#ifdef TRACY
-    infoln("Tracy adapter enabled.");
-#else
-    infoln("Tracy adapter disabled.");
-#endif
+std::string preset_file = "default.puzzle";
+std::string output_file = "clusters.puzzle";
+int max_blocks = 5;
+int board_width = 6;
+int board_height = 6;
+int goal_x = 4;
+int goal_y = 2;
+bool restricted = true;
 
+auto ui_mode() -> int
+{
     // RayLib window setup
     SetTraceLogLevel(LOG_ERROR);
     SetConfigFlags(FLAG_VSYNC_HINT);
@@ -69,7 +66,7 @@ auto main(int argc, char* argv[]) -> int
     InitWindow(INITIAL_WIDTH * 2, INITIAL_HEIGHT + MENU_HEIGHT, "MassSprings");
 
     // Game setup
-    threaded_physics physics;
+    threaded_physics physics(thread_pool);
     state_manager state(physics, preset_file);
     orbit_camera camera;
     input_handler input(state, camera);
@@ -89,9 +86,9 @@ auto main(int argc, char* argv[]) -> int
 
     // Game loop
     while (!WindowShouldClose()) {
-#ifdef TRACY
+        #ifdef TRACY
         FrameMarkStart("MainThread");
-#endif
+        #endif
 
         // Time tracking
         std::chrono::time_point now = std::chrono::high_resolution_clock::now();
@@ -102,16 +99,16 @@ auto main(int argc, char* argv[]) -> int
         // Input update
         input.handle_input();
 
-// Read positions from physics thread
-#ifdef TRACY
+        // Read positions from physics thread
+        #ifdef TRACY
         FrameMarkStart("MainThreadConsumeLock");
-#endif
+        #endif
         {
-#ifdef TRACY
+            #ifdef TRACY
             std::unique_lock<LockableBase(std::mutex)> lock(physics.state.data_mtx);
-#else
+            #else
             std::unique_lock<std::mutex> lock(physics.state.data_mtx);
-#endif
+            #endif
 
             ups = physics.state.ups;
             mass_center = physics.state.mass_center;
@@ -130,16 +127,15 @@ auto main(int argc, char* argv[]) -> int
                 physics.state.data_consumed_cnd.notify_all();
             }
         }
-#ifdef TRACY
+        #ifdef TRACY
         FrameMarkEnd("MainThreadConsumeLock");
-#endif
+        #endif
 
         // Update the camera after the physics, so target lock is smooth
         size_t current_index = state.get_current_index();
         if (masses.size() > current_index) {
-            const mass_spring_system::mass& current_mass = mass_spring_system::mass(masses.at(current_index));
-            camera.update(current_mass.position, mass_center, input.camera_lock,
-                          input.camera_mass_center_lock);
+            const Vector3& current_mass = masses[current_index];
+            camera.update(current_mass, mass_center, input.camera_lock, input.camera_mass_center_lock);
         }
 
         // Rendering
@@ -153,13 +149,200 @@ auto main(int argc, char* argv[]) -> int
         }
         ++loop_iterations;
 
-#ifdef TRACY
+        #ifdef TRACY
         FrameMark;
         FrameMarkEnd("MainThread");
-#endif
+        #endif
     }
 
     CloseWindow();
 
     return 0;
+}
+
+auto rush_hour_puzzle_space() -> int
+{
+    const boost::unordered_flat_set<puzzle::block, block_hasher2, block_equal2> permitted_blocks = {
+        puzzle::block(0, 0, 2, 1, false, false),
+        puzzle::block(0, 0, 3, 1, false, false),
+        puzzle::block(0, 0, 1, 2, false, false),
+        puzzle::block(0, 0, 1, 3, false, false)
+    };
+    const puzzle::block target_block = puzzle::block(0, 0, 2, 1, true, false);
+    const std::tuple<uint8_t, uint8_t, uint8_t, uint8_t> target_block_pos_range = {0, goal_y, board_width - 1, goal_y};
+
+    infoln("Exploring Rush-Hour puzzle space:");
+    infoln("- Size:       {}x{}", board_width, board_height);
+    infoln("- Goal:       {},{}", goal_x, goal_y);
+    infoln("- Restricted: {}", restricted);
+    infoln("- Max Blocks: {}", max_blocks);
+    infoln("- Target:     {}x{}", target_block.get_width(), target_block.get_height());
+    infoln("- Permitted block sizes:");
+    std::cout << "         ";
+    for (const puzzle::block b : permitted_blocks) {
+        std::cout << std::format(" {}x{},", b.get_width(), b.get_height());
+    }
+    std::cout << std::endl;
+    const std::chrono::high_resolution_clock::time_point start = std::chrono::high_resolution_clock::now();
+
+    const puzzle p = puzzle(board_width, board_height, goal_x, goal_y, restricted, true);
+    const boost::unordered_flat_set<puzzle, puzzle_hasher> result = p.explore_puzzle_space(
+        permitted_blocks,
+        target_block,
+        target_block_pos_range,
+        max_blocks,
+        thread_pool);
+
+    const std::chrono::high_resolution_clock::time_point end = std::chrono::high_resolution_clock::now();
+    infoln("Found {} different clusters. Took {}s.",
+           result.size(),
+           std::chrono::duration_cast<std::chrono::seconds>(end - start).count());
+
+    infoln("Sorting clusters...");
+    std::vector<puzzle> result_sorted{result.begin(), result.end()};
+    std::ranges::sort(result_sorted, std::ranges::greater{});
+    // for (const puzzle& _p : result_sorted) {
+    //     traceln("{}", _p.string_repr());
+    // }
+
+    size_t i = 0;
+    size_t success = 0;
+    std::filesystem::remove(output_file);
+    for (const puzzle& _p : result_sorted) {
+        if (append_preset_file_quiet(output_file, std::format("Cluster {}", i), _p, true)) {
+            ++success;
+        }
+        ++i;
+    }
+    if (success != result_sorted.size()) {
+        warnln("Saved {} of {} clusters", success, result_sorted.size());
+    } else {
+        infoln("Saved {} of {} clusters", success, result_sorted.size());
+    }
+
+    return 0;
+}
+
+enum class runmode
+{
+    USER_INTERFACE, RUSH_HOUR_PUZZLE_SPACE, EXIT,
+};
+
+auto argparse(const int argc, char* argv[]) -> runmode
+{
+    #ifndef WIN32
+    po::options_description desc("Allowed options");
+    desc.add_options()                                                                                            //
+        ("help", "produce help message")                                                                          //
+        ("presets", po::value<std::string>()->default_value(preset_file), "load presets from file")               //
+        ("output", po::value<std::string>()->default_value(output_file), "output file for generated clusters")    //
+        ("space", po::value<std::string>()->value_name("rh|klotski"), "generate puzzle space with ruleset")       //
+        ("w", po::value<int>()->default_value(board_width)->value_name("[3, 8]"), "board width")                  //
+        ("h", po::value<int>()->default_value(board_height)->value_name("[3, 8"), "board height")                 //
+        ("gx", po::value<int>()->default_value(goal_x)->value_name("[0, w-1]"), "board goal horizontal position") //
+        ("gy", po::value<int>()->default_value(goal_y)->value_name("[0, h-1]"), "board goal vertical position")   //
+        ("free", "allow free block movement")                                                                     //
+        ("blocks",
+         po::value<int>()->default_value(max_blocks)->value_name("[1, 15]"),
+         "block limit for puzzle space generation") //
+        ;
+
+    po::positional_options_description positional;
+    positional.add("presets", -1);
+
+    po::variables_map vm;
+    po::store(po::command_line_parser(argc, argv).options(desc).positional(positional).run(), vm);
+    po::notify(vm);
+
+    if (vm.contains("help")) {
+        std::cout << desc << std::endl;
+        return runmode::EXIT;
+    }
+
+    if (vm.contains("output")) {
+        output_file = vm["output"].as<std::string>();
+    }
+
+    if (vm.contains("w")) {
+        board_width = vm["w"].as<int>();
+        board_width = std::max(static_cast<int>(puzzle::MIN_WIDTH),
+                               std::min(board_width, static_cast<int>(puzzle::MAX_WIDTH)));
+    }
+
+    if (vm.contains("h")) {
+        board_height = vm["h"].as<int>();
+        board_height = std::max(static_cast<int>(puzzle::MIN_HEIGHT),
+                                std::min(board_height, static_cast<int>(puzzle::MAX_HEIGHT)));
+    }
+
+    if (vm.contains("gx")) {
+        goal_x = vm["gx"].as<int>();
+        goal_x = std::max(0, std::min(goal_x, static_cast<int>(puzzle::MAX_WIDTH) - 1));
+    }
+
+    if (vm.contains("gy")) {
+        goal_y = vm["gy"].as<int>();
+        goal_y = std::max(0, std::min(goal_y, static_cast<int>(puzzle::MAX_HEIGHT) - 1));
+    }
+
+    if (vm.contains("free")) {
+        restricted = false;
+    }
+
+    if (vm.contains("blocks")) {
+        max_blocks = vm["blocks"].as<int>();
+        max_blocks = std::max(1, std::min(max_blocks, static_cast<int>(puzzle::MAX_BLOCKS)));
+    }
+
+    if (vm.contains("space")) {
+        const std::string ruleset = vm["space"].as<std::string>();
+        if (ruleset == "rh") {
+            return runmode::RUSH_HOUR_PUZZLE_SPACE;
+        }
+        if (ruleset == "klotski") {
+            throw std::runtime_error("Not implemented");
+        }
+    }
+
+    if (vm.contains("presets")) {
+        preset_file = vm["presets"].as<std::string>();
+    }
+    #endif
+
+    return runmode::USER_INTERFACE;
+}
+
+auto main(const int argc, char* argv[]) -> int
+{
+    #ifdef BACKWARD
+    infoln("Backward stack-traces enabled.");
+    #else
+    infoln("Backward stack-traces disabled.");
+    #endif
+
+    #ifdef TRACY
+    infoln("Tracy adapter enabled.");
+    #else
+    infoln("Tracy adapter disabled.");
+    #endif
+
+    infoln("Using background thread for physics.");
+    infoln("Using octree-barnes-hut for graph layout.");
+
+    #ifdef THREADPOOL
+    infoln("Additional thread-pool enabled ({} threads).", threads.get_thread_count());
+    #else
+    infoln("Additional thread-pool disabled.");
+    #endif
+
+    switch (argparse(argc, argv)) {
+    case runmode::USER_INTERFACE:
+        return ui_mode();
+    case runmode::RUSH_HOUR_PUZZLE_SPACE:
+        return rush_hour_puzzle_space();
+    case runmode::EXIT:
+        return 0;
+    };
+
+    return 1;
 }

src/mass_spring_system.cpp

@@ -2,66 +2,63 @@
 #include "config.hpp"
 
 #include <cfloat>
+#include <cstring>
 
-#ifdef TRACY
-#include <tracy/Tracy.hpp>
-#endif
-
-auto mass_spring_system::mass::clear_force() -> void
+auto mass_spring_system::calculate_spring_force(const size_t s) -> void
 {
-    force = Vector3Zero();
+    const spring _s = springs[s];
+    const Vector3 a_pos = positions[_s.a];
+    const Vector3 b_pos = positions[_s.b];
+    const Vector3 a_vel = velocities[_s.a];
+    const Vector3 b_vel = velocities[_s.b];
+
+    const Vector3 delta_pos = a_pos - b_pos;
+    const Vector3 delta_vel = a_vel - b_vel;
+
+    const float sq_len = Vector3DotProduct(delta_pos, delta_pos);
+    const float inv_len = 1.0f / sqrt(sq_len);
+    const float len = sq_len * inv_len;
+
+    const float hooke = SPRING_CONSTANT * (len - REST_LENGTH);
+    const float dampening = DAMPENING_CONSTANT * Vector3DotProduct(delta_vel, delta_pos) * inv_len;
+
+    const Vector3 a_force = Vector3Scale(delta_pos, -(hooke + dampening) * inv_len);
+    const Vector3 b_force = a_force * -1.0f;
+
+    forces[_s.a] += a_force;
+    forces[_s.b] += b_force;
 }
 
-auto mass_spring_system::mass::calculate_velocity(const float delta_time) -> void
+auto mass_spring_system::integrate_velocity(const size_t m, const float dt) -> void
 {
-    const Vector3 acceleration = Vector3Scale(force, 1.0 / MASS);
-    const Vector3 temp = Vector3Scale(acceleration, delta_time);
-    velocity = Vector3Add(velocity, temp);
+    const Vector3 acc = forces[m] / MASS;
+    velocities[m] += acc * dt;
 }
 
-auto mass_spring_system::mass::calculate_position(const float delta_time) -> void
+auto mass_spring_system::integrate_position(const size_t m, const float dt) -> void
 {
-    previous_position = position;
-
-    const Vector3 temp = Vector3Scale(velocity, delta_time);
-    position = Vector3Add(position, temp);
+    previous_positions[m] = positions[m];
+    positions[m] += velocities[m] * dt;
 }
 
-auto mass_spring_system::mass::verlet_update(const float delta_time) -> void
+auto mass_spring_system::verlet_update(const size_t m, const float dt) -> void
 {
-    const Vector3 acceleration = Vector3Scale(force, 1.0 / MASS);
-    const Vector3 temp_position = position;
+    const Vector3 acc = (forces[m] / MASS) * dt * dt;
+    const Vector3 pos = positions[m];
 
-    Vector3 displacement = Vector3Subtract(position, previous_position);
-    const Vector3 accel_term = Vector3Scale(acceleration, delta_time * delta_time);
+    Vector3 delta_pos = pos - previous_positions[m];
+    delta_pos *= 1.0 - VERLET_DAMPENING; // Minimal dampening
 
-    // Minimal dampening
-    displacement = Vector3Scale(displacement, 1.0 - VERLET_DAMPENING);
-
-    position = Vector3Add(Vector3Add(position, displacement), accel_term);
-    previous_position = temp_position;
-}
-
-auto mass_spring_system::spring::calculate_spring_force(mass& _a, mass& _b) -> void
-{
-    // TODO: Use a bungee force here instead of springs, since we already have global repulsion?
-    const Vector3 delta_position = Vector3Subtract(_a.position, _b.position);
-    const float current_length = Vector3Length(delta_position);
-    const Vector3 delta_velocity = Vector3Subtract(_a.velocity, _b.velocity);
-
-    const float hooke = SPRING_CONSTANT * (current_length - REST_LENGTH);
-    const float dampening = DAMPENING_CONSTANT * Vector3DotProduct(delta_velocity, delta_position) / current_length;
-
-    const Vector3 force_a = Vector3Scale(delta_position, -(hooke + dampening) / current_length);
-    const Vector3 force_b = Vector3Scale(force_a, -1.0);
-
-    _a.force = Vector3Add(_a.force, force_a);
-    _b.force = Vector3Add(_b.force, force_b);
+    positions[m] += delta_pos + acc;
+    previous_positions[m] = pos;
 }
 
 auto mass_spring_system::clear() -> void
 {
-    masses.clear();
+    positions.clear();
+    previous_positions.clear();
+    velocities.clear();
+    forces.clear();
     springs.clear();
     tree.nodes.clear();
 }
@@ -77,24 +74,27 @@ auto mass_spring_system::add_mass() -> void
     // };
     // position = Vector3Scale(Vector3Normalize(position), REST_LENGTH * 2.0);
 
-    masses.emplace_back(Vector3Zero());
+    positions.emplace_back(Vector3Zero());
+    previous_positions.emplace_back(Vector3Zero());
+    velocities.emplace_back(Vector3Zero());
+    forces.emplace_back(Vector3Zero());
 }
 
 auto mass_spring_system::add_spring(size_t a, size_t b) -> void
 {
     // Update masses to be located along a random walk when adding the springs
-    const mass& mass_a = masses.at(a);
-    mass& mass_b = masses.at(b);
+    const Vector3& mass_a = positions[a];
+    const Vector3& mass_b = positions[b];
 
-    Vector3 offset{
-        static_cast<float>(GetRandomValue(-100, 100)), static_cast<float>(GetRandomValue(-100, 100)),
-        static_cast<float>(GetRandomValue(-100, 100))
-    };
+    Vector3 offset{static_cast<float>(GetRandomValue(-100, 100)),
+                   static_cast<float>(GetRandomValue(-100, 100)),
+                   static_cast<float>(GetRandomValue(-100, 100))};
     offset = Vector3Normalize(offset) * REST_LENGTH;
 
     // If the offset moves the mass closer to the current center of mass, flip it
     if (!tree.nodes.empty()) {
-        const Vector3 mass_center_direction = Vector3Subtract(mass_a.position, tree.nodes.at(0).mass_center);
+        const Vector3 mass_center_direction =
+            Vector3Subtract(positions[a], tree.nodes[0].mass_center);
         const float mass_center_distance = Vector3Length(mass_center_direction);
 
         if (mass_center_distance > 0 && Vector3DotProduct(offset, mass_center_direction) < 0.0f) {
@@ -102,8 +102,8 @@ auto mass_spring_system::add_spring(size_t a, size_t b) -> void
         }
     }
 
-    mass_b.position = mass_a.position + offset;
-    mass_b.previous_position = mass_b.position;
+    positions[b] = mass_a + offset;
+    previous_positions[b] = mass_b;
 
     // infoln("Adding spring: ({}, {}, {})->({}, {}, {})", mass_a.position.x, mass_a.position.y,
     // mass_a.position.z,
@@ -114,118 +114,92 @@ auto mass_spring_system::add_spring(size_t a, size_t b) -> void
 
 auto mass_spring_system::clear_forces() -> void
 {
-    #ifdef TRACY
+#ifdef TRACY
     ZoneScoped;
-    #endif
-
-    for (auto& m : masses) {
-        m.clear_force();
-    }
-}
-
-auto mass_spring_system::calculate_spring_forces() -> void
-{
-    #ifdef TRACY
-    ZoneScoped;
-    #endif
-
-    for (const auto s : springs) {
-        mass& a = masses.at(s.a);
-        mass& b = masses.at(s.b);
-        spring::calculate_spring_force(a, b);
-    }
-}
-
-#ifdef THREADPOOL
-auto mass_spring_system::set_thread_name(size_t idx) -> void
-{
-    BS::this_thread::set_os_thread_name(std::format("bh-worker-{}", idx));
-}
 #endif
 
-auto mass_spring_system::build_octree() -> void
+    memset(forces.data(), 0, forces.size() * sizeof(Vector3));
+}
+
+auto mass_spring_system::calculate_spring_forces(
+    const std::optional<BS::thread_pool<>* const> thread_pool) -> void
 {
-    #ifdef TRACY
+#ifdef TRACY
     ZoneScoped;
-    #endif
+#endif
 
-    tree.nodes.clear();
-    tree.nodes.reserve(masses.size() * 2);
+    const auto solve_spring_force = [&](const int i) { calculate_spring_force(i); };
 
-    // Compute bounding box around all masses
-    Vector3 min{FLT_MAX, FLT_MAX, FLT_MAX};
-    Vector3 max{-FLT_MAX, -FLT_MAX, -FLT_MAX};
-    for (const auto& m : masses) {
-        min.x = std::min(min.x, m.position.x);
-        max.x = std::max(max.x, m.position.x);
-        min.y = std::min(min.y, m.position.y);
-        max.y = std::max(max.y, m.position.y);
-        min.z = std::min(min.z, m.position.z);
-        max.z = std::max(max.z, m.position.z);
-    }
-
-    // Pad the bounding box
-    constexpr float pad = 1.0;
-    min = Vector3Subtract(min, Vector3Scale(Vector3One(), pad));
-    max = Vector3Add(max, Vector3Scale(Vector3One(), pad));
-
-    // Make it cubic (so subdivisions are balanced)
-    const float max_extent = std::max({max.x - min.x, max.y - min.y, max.z - min.z});
-    max = Vector3Add(min, Vector3Scale(Vector3One(), max_extent));
-
-    // Root node spans the entire area
-    const int root = tree.create_empty_leaf(min, max);
-
-    for (size_t i = 0; i < masses.size(); ++i) {
-        tree.insert(root, static_cast<int>(i), masses[i].position, MASS, 0);
+    if (thread_pool) {
+        (*thread_pool)
+            ->submit_loop(0, springs.size(), solve_spring_force, SMALL_TASK_BLOCK_SIZE)
+            .wait();
+    } else {
+        for (size_t i = 0; i < springs.size(); ++i) {
+            solve_spring_force(i);
+        }
     }
 }
 
-auto mass_spring_system::calculate_repulsion_forces() -> void
+auto mass_spring_system::calculate_repulsion_forces(
+    const std::optional<BS::thread_pool<>* const> thread_pool) -> void
 {
-    #ifdef TRACY
+#ifdef TRACY
     ZoneScoped;
-    #endif
+#endif
 
-    build_octree();
-
-    auto solve_octree = [&](const int i)
+    const auto solve_octree = [&](const int i)
     {
-        const Vector3 force = tree.calculate_force(0, masses[i].position);
-        masses[i].force = Vector3Add(masses[i].force, force);
+        const Vector3 force = tree.calculate_force(0, positions[i]);
+        forces[i] += force;
     };
 
     // Calculate forces using Barnes-Hut
-    #ifdef THREADPOOL
-    const BS::multi_future<void> loop_future = threads.submit_loop(0, masses.size(), solve_octree, 256);
-    loop_future.wait();
-    #else
-    for (size_t i = 0; i < masses.size(); ++i) {
-        solve_octree(i);
+    if (thread_pool) {
+        (*thread_pool)
+            ->submit_loop(0, positions.size(), solve_octree, LARGE_TASK_BLOCK_SIZE)
+            .wait();
+    } else {
+        for (size_t i = 0; i < positions.size(); ++i) {
+            solve_octree(i);
+        }
     }
-    #endif
 }
 
-auto mass_spring_system::verlet_update(const float delta_time) -> void
+auto mass_spring_system::update(const float dt,
+                                const std::optional<BS::thread_pool<>* const> thread_pool) -> void
 {
-    #ifdef TRACY
+#ifdef TRACY
     ZoneScoped;
-    #endif
+#endif
 
-    for (auto& m : masses) {
-        m.verlet_update(delta_time);
+    const auto update = [&](const int i) { verlet_update(i, dt); };
+
+    if (thread_pool) {
+        (*thread_pool)->submit_loop(0, positions.size(), update, SMALL_TASK_BLOCK_SIZE).wait();
+    } else {
+        for (size_t i = 0; i < positions.size(); ++i) {
+            update(i);
+        }
     }
 }
 
-auto mass_spring_system::center_masses() -> void
+auto mass_spring_system::center_masses(const std::optional<BS::thread_pool<>* const> thread_pool)
+    -> void
 {
     Vector3 mean = Vector3Zero();
-    for (const auto& m : masses) {
-        mean += m.position;
+    for (const Vector3& pos : positions) {
+        mean += pos;
     }
-    mean /= static_cast<float>(masses.size());
+    mean /= static_cast<float>(positions.size());
 
-    for (auto& m : masses) {
-        m.position -= mean;
+    const auto center_mass = [&](const int i) { positions[i] -= mean; };
+
+    if (thread_pool) {
+        (*thread_pool)->submit_loop(0, positions.size(), center_mass, SMALL_TASK_BLOCK_SIZE).wait();
+    } else {
+        for (size_t i = 0; i < positions.size(); ++i) {
+            center_mass(i);
+        }
     }
-}
+}

src/octree.cpp

@@ -1,13 +1,9 @@
 #include "octree.hpp"
 #include "config.hpp"
-#include "util.hpp"
 
+#include <cfloat>
 #include <raymath.h>
 
-#ifdef TRACY
-    #include <tracy/Tracy.hpp>
-#endif
-
 auto octree::node::child_count() const -> int
 {
     int child_count = 0;
@@ -19,22 +15,11 @@ auto octree::node::child_count() const -> int
     return child_count;
 }
 
-auto octree::create_empty_leaf(const Vector3& box_min, const Vector3& box_max) -> int
-{
-    node n;
-    n.box_min = box_min;
-    n.box_max = box_max;
-    nodes.emplace_back(n);
-
-    return static_cast<int>(nodes.size() - 1);
-}
-
 auto octree::get_octant(const int node_idx, const Vector3& pos) const -> int
 {
     const node& n = nodes[node_idx];
-    auto [cx, cy, cz] =
-        Vector3((n.box_min.x + n.box_max.x) / 2.0f, (n.box_min.y + n.box_max.y) / 2.0f,
-                (n.box_min.z + n.box_max.z) / 2.0f);
+    auto [cx, cy, cz] = Vector3((n.box_min.x + n.box_max.x) / 2.0f, (n.box_min.y + n.box_max.y) / 2.0f,
+                                (n.box_min.z + n.box_max.z) / 2.0f);
 
     // The octant is encoded as a 3-bit integer "zyx". The node area is split
     // along all 3 axes, if a position is right of an axis, this bit is set to 1.
@@ -54,13 +39,11 @@ auto octree::get_octant(const int node_idx, const Vector3& pos) const -> int
     return octant;
 }
 
-auto octree::get_child_bounds(const int node_idx, const int octant) const
-    -> std::pair<Vector3, Vector3>
+auto octree::get_child_bounds(const int node_idx, const int octant) const -> std::pair<Vector3, Vector3>
 {
     const node& n = nodes[node_idx];
-    auto [cx, cy, cz] =
-        Vector3((n.box_min.x + n.box_max.x) / 2.0f, (n.box_min.y + n.box_max.y) / 2.0f,
-                (n.box_min.z + n.box_max.z) / 2.0f);
+    auto [cx, cy, cz] = Vector3((n.box_min.x + n.box_max.x) / 2.0f, (n.box_min.y + n.box_max.y) / 2.0f,
+                                (n.box_min.z + n.box_max.z) / 2.0f);
 
     Vector3 min = Vector3Zero();
     Vector3 max = Vector3Zero();
@@ -78,19 +61,27 @@ auto octree::get_child_bounds(const int node_idx, const int octant) const
     return std::make_pair(min, max);
 }
 
+auto octree::create_empty_leaf(const Vector3& box_min, const Vector3& box_max) -> int
+{
+    node n;
+    n.box_min = box_min;
+    n.box_max = box_max;
+    nodes.emplace_back(n);
+
+    return static_cast<int>(nodes.size() - 1);
+}
+
 auto octree::insert(const int node_idx, const int mass_id, const Vector3& pos, const float mass,
                     const int depth) -> void
 {
     // infoln("Inserting position ({}, {}, {}) into octree at node {} (depth {})", pos.x, pos.y,
     // pos.z, node_idx, depth);
     if (depth > MAX_DEPTH) {
-        errln("MAX_DEPTH! node={} box_min=({},{},{}) box_max=({},{},{}) pos=({},{},{})", node_idx,
-              nodes[node_idx].box_min.x, nodes[node_idx].box_min.y, nodes[node_idx].box_min.z,
-              nodes[node_idx].box_max.x, nodes[node_idx].box_max.y, nodes[node_idx].box_max.z,
-              pos.x, pos.y, pos.z);
-
-        // This runs from inside the physics thread so it won't exit cleanly
-        exit(1);
+        throw std::runtime_error(std::format("MAX_DEPTH! node={} box_min=({},{},{}) box_max=({},{},{}) pos=({},{},{})",
+                                             node_idx, nodes[node_idx].box_min.x, nodes[node_idx].box_min.y,
+                                             nodes[node_idx].box_min.z, nodes[node_idx].box_max.x,
+                                             nodes[node_idx].box_max.y, nodes[node_idx].box_max.z, pos.x, pos.y,
+                                             pos.z));
     }
 
     // NOTE: Do not store a nodes[node_idx] reference as the nodes vector might reallocate during
@@ -152,15 +143,50 @@ auto octree::insert(const int node_idx, const int mass_id, const Vector3& pos, c
 
     // Update the center of mass
     const float new_mass = nodes[node_idx].mass_total + mass;
-    nodes[node_idx].mass_center.x =
-        (nodes[node_idx].mass_center.x * nodes[node_idx].mass_total + pos.x) / new_mass;
-    nodes[node_idx].mass_center.y =
-        (nodes[node_idx].mass_center.y * nodes[node_idx].mass_total + pos.y) / new_mass;
-    nodes[node_idx].mass_center.z =
-        (nodes[node_idx].mass_center.z * nodes[node_idx].mass_total + pos.z) / new_mass;
+    nodes[node_idx].mass_center.x = (nodes[node_idx].mass_center.x * nodes[node_idx].mass_total + pos.x) / new_mass;
+    nodes[node_idx].mass_center.y = (nodes[node_idx].mass_center.y * nodes[node_idx].mass_total + pos.y) / new_mass;
+    nodes[node_idx].mass_center.z = (nodes[node_idx].mass_center.z * nodes[node_idx].mass_total + pos.z) / new_mass;
     nodes[node_idx].mass_total = new_mass;
 }
 
+auto octree::build_octree(octree& t, const std::vector<Vector3>& positions) -> void
+{
+    #ifdef TRACY
+    ZoneScoped;
+    #endif
+
+    t.nodes.clear();
+    t.nodes.reserve(positions.size() * 2);
+
+    // Compute bounding box around all masses
+    Vector3 min{FLT_MAX, FLT_MAX, FLT_MAX};
+    Vector3 max{-FLT_MAX, -FLT_MAX, -FLT_MAX};
+    for (const auto& [x, y, z] : positions) {
+        min.x = std::min(min.x, x);
+        max.x = std::max(max.x, x);
+        min.y = std::min(min.y, y);
+        max.y = std::max(max.y, y);
+        min.z = std::min(min.z, z);
+        max.z = std::max(max.z, z);
+    }
+
+    // Pad the bounding box
+    constexpr float pad = 1.0;
+    min = Vector3Subtract(min, Vector3Scale(Vector3One(), pad));
+    max = Vector3Add(max, Vector3Scale(Vector3One(), pad));
+
+    // Make it cubic (so subdivisions are balanced)
+    const float max_extent = std::max({max.x - min.x, max.y - min.y, max.z - min.z});
+    max = Vector3Add(min, Vector3Scale(Vector3One(), max_extent));
+
+    // Root node spans the entire area
+    const int root = t.create_empty_leaf(min, max);
+
+    for (size_t i = 0; i < positions.size(); ++i) {
+        t.insert(root, static_cast<int>(i), positions[i], MASS, 0);
+    }
+}
+
 auto octree::calculate_force(const int node_idx, const Vector3& pos) const -> Vector3
 {
     if (node_idx < 0) {
@@ -198,4 +224,4 @@ auto octree::calculate_force(const int node_idx, const Vector3& pos) const -> Ve
     }
 
     return force;
-}
+}

src/orbit_camera.cpp

@@ -4,10 +4,6 @@
 #include <raylib.h>
 #include <raymath.h>
 
-#ifdef TRACY
-    #include <tracy/Tracy.hpp>
-#endif
-
 auto orbit_camera::rotate(const Vector2 last_mouse, const Vector2 mouse) -> void
 {
     const auto [dx, dy] = Vector2Subtract(mouse, last_mouse);
@@ -73,4 +69,4 @@ auto orbit_camera::update(const Vector3& current_target, const Vector3& mass_cen
     camera.target = target;
     camera.fovy = fov;
     camera.projection = projection;
-}
+}

src/renderer.cpp

@@ -5,33 +5,29 @@
 #include <raymath.h>
 #include <rlgl.h>
 
-#ifdef TRACY
-    #include <tracy/Tracy.hpp>
-#endif
-
 auto renderer::update_texture_sizes() -> void
 {
     if (!IsWindowResized()) {
         return;
     }
 
-    UnloadRenderTexture(render_target);
+    UnloadRenderTexture(graph_target);
     UnloadRenderTexture(klotski_target);
     UnloadRenderTexture(menu_target);
 
     const int width = GetScreenWidth() / 2;
     const int height = GetScreenHeight() - MENU_HEIGHT;
 
-    render_target = LoadRenderTexture(width, height);
+    graph_target = LoadRenderTexture(width, height);
     klotski_target = LoadRenderTexture(width, height);
     menu_target = LoadRenderTexture(width * 2, MENU_HEIGHT);
 }
 
 auto renderer::draw_mass_springs(const std::vector<Vector3>& masses) -> void
 {
-#ifdef TRACY
+    #ifdef TRACY
     ZoneScoped;
-#endif
+    #endif
 
     if (masses.size() != state.get_state_count()) {
         // Because the physics run in a different thread, it might need time to catch up
@@ -40,16 +36,16 @@ auto renderer::draw_mass_springs(const std::vector<Vector3>& masses) -> void
 
     // Prepare connection batching
     {
-#ifdef TRACY
-        ZoneNamedN(prepare_masses, "PrepareConnectionsBatching", true);
-#endif
+        #ifdef TRACY
+        ZoneNamedN(prepare_connections, "PrepareConnectionsBatching", true);
+        #endif
 
         connections.clear();
         connections.reserve(state.get_target_count());
         if (input.connect_solutions) {
             for (const size_t& _state : state.get_winning_indices()) {
-                const Vector3& current_mass = masses.at(state.get_current_index());
-                const Vector3& winning_mass = masses.at(_state);
+                const Vector3& current_mass = masses[state.get_current_index()];
+                const Vector3& winning_mass = masses[_state];
                 connections.emplace_back(current_mass, winning_mass);
                 DrawLine3D(current_mass, winning_mass, Fade(TARGET_BLOCK_COLOR, 0.5));
             }
@@ -58,9 +54,9 @@ auto renderer::draw_mass_springs(const std::vector<Vector3>& masses) -> void
 
     // Prepare cube instancing
     {
-#ifdef TRACY
+        #ifdef TRACY
         ZoneNamedN(prepare_masses, "PrepareMassInstancing", true);
-#endif
+        #endif
 
         if (masses.size() < DRAW_VERTICES_LIMIT) {
             // Don't have to reserve, capacity is already set to DRAW_VERTICES_LIMIT in constructor
@@ -73,12 +69,10 @@ auto renderer::draw_mass_springs(const std::vector<Vector3>& masses) -> void
 
                 // Normal vertex
                 Color c = VERTEX_COLOR;
-                if ((input.mark_solutions || input.mark_path) &&
-                    state.get_winning_indices().contains(mass)) {
+                if ((input.mark_solutions || input.mark_path) && state.get_winning_indices().contains(mass)) {
                     // Winning vertex
                     c = VERTEX_TARGET_COLOR;
-                } else if ((input.mark_solutions || input.mark_path) &&
-                           state.get_path_indices().contains(mass)) {
+                } else if ((input.mark_solutions || input.mark_path) && state.get_path_indices().contains(mass)) {
                     // Path vertex
                     c = VERTEX_PATH_COLOR;
                 } else if (mass == state.get_starting_index()) {
@@ -98,21 +92,21 @@ auto renderer::draw_mass_springs(const std::vector<Vector3>& masses) -> void
         rlUpdateVertexBuffer(color_vbo_id, colors.data(), colors.size() * sizeof(Color), 0);
     }
 
-    BeginTextureMode(render_target);
+    BeginTextureMode(graph_target);
     ClearBackground(RAYWHITE);
     BeginMode3D(camera.camera);
 
     // Draw springs (batched)
     {
-#ifdef TRACY
+        #ifdef TRACY
         ZoneNamedN(draw_springs, "DrawSprings", true);
-#endif
+        #endif
 
         rlBegin(RL_LINES);
         for (const auto& [from, to] : state.get_links()) {
             if (masses.size() > from && masses.size() > to) {
-                const auto& [ax, ay, az] = masses.at(from);
-                const auto& [bx, by, bz] = masses.at(to);
+                const auto& [ax, ay, az] = masses[from];
+                const auto& [bx, by, bz] = masses[to];
                 rlColor4ub(EDGE_COLOR.r, EDGE_COLOR.g, EDGE_COLOR.b, EDGE_COLOR.a);
                 rlVertex3f(ax, ay, az);
                 rlVertex3f(bx, by, bz);
@@ -123,17 +117,16 @@ auto renderer::draw_mass_springs(const std::vector<Vector3>& masses) -> void
 
     // Draw masses (instanced)
     {
-#ifdef TRACY
+        #ifdef TRACY
         ZoneNamedN(draw_masses, "DrawMasses", true);
-#endif
+        #endif
 
         if (masses.size() < DRAW_VERTICES_LIMIT) {
             // NOTE: I don't know if drawing all this inside a shader would make it
             //       much faster... The amount of data sent to the GPU would be
             //       reduced (just positions instead of matrices), but is this
             //       noticable for < 100000 cubes?
-            DrawMeshInstanced(cube_instance, vertex_mat, transforms.data(),
-                              masses.size()); // NOLINT(*-narrowing-conversions)
+            DrawMeshInstanced(cube_instance, vertex_mat, transforms.data(), masses.size()); // NOLINT(*-narrowing-conversions)
         }
     }
 
@@ -152,9 +145,8 @@ auto renderer::draw_mass_springs(const std::vector<Vector3>& masses) -> void
     // Mark current state
     const size_t current_index = state.get_current_index();
     if (masses.size() > current_index) {
-        const Vector3& current_mass = masses.at(current_index);
-        DrawCube(current_mass, VERTEX_SIZE * 2, VERTEX_SIZE * 2, VERTEX_SIZE * 2,
-                 VERTEX_CURRENT_COLOR);
+        const Vector3& current_mass = masses[current_index];
+        DrawCube(current_mass, VERTEX_SIZE * 2, VERTEX_SIZE * 2, VERTEX_SIZE * 2, VERTEX_CURRENT_COLOR);
     }
 
     EndMode3D();
@@ -163,9 +155,9 @@ auto renderer::draw_mass_springs(const std::vector<Vector3>& masses) -> void
 
 auto renderer::draw_klotski() const -> void
 {
-#ifdef TRACY
+    #ifdef TRACY
     ZoneScoped;
-#endif
+    #endif
 
     BeginTextureMode(klotski_target);
     ClearBackground(RAYWHITE);
@@ -177,9 +169,9 @@ auto renderer::draw_klotski() const -> void
 
 auto renderer::draw_menu() const -> void
 {
-#ifdef TRACY
+    #ifdef TRACY
     ZoneScoped;
-#endif
+    #endif
 
     BeginTextureMode(menu_target);
     ClearBackground(RAYWHITE);
@@ -194,32 +186,28 @@ auto renderer::draw_textures(const int fps, const int ups, const size_t mass_cou
 {
     BeginDrawing();
 
-    DrawTextureRec(menu_target.texture,
-                   Rectangle(0, 0, menu_target.texture.width, -menu_target.texture.height),
+    DrawTextureRec(menu_target.texture, Rectangle(0, 0, menu_target.texture.width, -menu_target.texture.height),
                    Vector2(0, 0), WHITE);
     DrawTextureRec(klotski_target.texture,
                    Rectangle(0, 0, klotski_target.texture.width, -klotski_target.texture.height),
                    Vector2(0, MENU_HEIGHT), WHITE);
-    DrawTextureRec(render_target.texture,
-                   Rectangle(0, 0, render_target.texture.width, -render_target.texture.height),
+    DrawTextureRec(graph_target.texture, Rectangle(0, 0, graph_target.texture.width, -graph_target.texture.height),
                    Vector2(GetScreenWidth() / 2.0f, MENU_HEIGHT), WHITE);
 
     // Draw borders
     DrawRectangleLinesEx(Rectangle(0, 0, GetScreenWidth(), MENU_HEIGHT), 1.0f, BLACK);
-    DrawRectangleLinesEx(
-        Rectangle(0, MENU_HEIGHT, GetScreenWidth() / 2.0f, GetScreenHeight() - MENU_HEIGHT), 1.0f,
-        BLACK);
+    DrawRectangleLinesEx(Rectangle(0, MENU_HEIGHT, GetScreenWidth() / 2.0f, GetScreenHeight() - MENU_HEIGHT), 1.0f,
+                         BLACK);
     DrawRectangleLinesEx(Rectangle(GetScreenWidth() / 2.0f, MENU_HEIGHT, GetScreenWidth() / 2.0f,
-                                   GetScreenHeight() - MENU_HEIGHT),
-                         1.0f, BLACK);
+                                   GetScreenHeight() - MENU_HEIGHT), 1.0f, BLACK);
 
     gui.draw(fps, ups, mass_count, spring_count);
 
     EndDrawing();
 }
 
-auto renderer::render(const std::vector<Vector3>& masses, const int fps, const int ups,
-                      const size_t mass_count, const size_t spring_count) -> void
+auto renderer::render(const std::vector<Vector3>& masses, const int fps, const int ups, const size_t mass_count,
+                      const size_t spring_count) -> void
 {
     update_texture_sizes();
 
@@ -227,4 +215,4 @@ auto renderer::render(const std::vector<Vector3>& masses, const int fps, const i
     draw_klotski();
     draw_menu();
     draw_textures(fps, ups, mass_count, spring_count);
-}
+}

src/state_manager.cpp

@@ -2,17 +2,10 @@
 #include "graph_distances.hpp"
 #include "util.hpp"
 
-#include <fstream>
-#include <ios>
-
-#ifdef TRACY
-#include <tracy/Tracy.hpp>
-#endif
-
 auto state_manager::synced_try_insert_state(const puzzle& state) -> size_t
 {
     if (state_indices.contains(state)) {
-        return state_indices.at(state);
+        return state_indices[state];
     }
 
     const size_t index = state_pool.size();
@@ -77,77 +70,27 @@ auto state_manager::synced_clear_statespace() -> void
     physics.clear_cmd();
 }
 
-auto state_manager::parse_preset_file(const std::string& _preset_file) -> bool
+auto state_manager::save_current_to_preset_file(const std::string& preset_comment) -> void
 {
-    preset_file = _preset_file;
-
-    std::ifstream file(preset_file);
-    if (!file) {
-        infoln("Preset file \"{}\" couldn't be loaded.", preset_file);
-        return false;
+    if (append_preset_file(preset_file, preset_comment, get_current_state())) {
+        current_preset = preset_states.size();
+        reload_preset_file();
     }
-
-    std::string line;
-    std::vector<std::string> comment_lines;
-    std::vector<std::string> preset_lines;
-    while (std::getline(file, line)) {
-        if (line.starts_with("F") || line.starts_with("R")) {
-            preset_lines.push_back(line);
-        } else if (line.starts_with("#")) {
-            comment_lines.push_back(line);
-        }
-    }
-
-    if (preset_lines.empty() || comment_lines.size() != preset_lines.size()) {
-        infoln("Preset file \"{}\" couldn't be loaded.", preset_file);
-        return false;
-    }
-
-    preset_states.clear();
-    for (const auto& preset : preset_lines) {
-        const puzzle& p = puzzle(preset);
-
-        if (const std::optional<std::string>& reason = p.try_get_invalid_reason()) {
-            preset_states = {puzzle(4, 5, 9, 9, false)};
-            infoln("Preset file \"{}\" contained invalid presets: {}", preset_file, *reason);
-            return false;
-        }
-        preset_states.emplace_back(p);
-    }
-    preset_comments = comment_lines;
-
-    infoln("Loaded {} presets from \"{}\".", preset_lines.size(), preset_file);
-
-    return true;
 }
 
-auto state_manager::append_preset_file(const std::string& preset_name) -> bool
+auto state_manager::reload_preset_file() -> void
 {
-    infoln(R"(Saving preset "{}" to "{}")", preset_name, preset_file);
-
-    if (get_current_state().try_get_invalid_reason()) {
-        return false;
+    const auto [presets, comments] = parse_preset_file(preset_file);
+    if (!presets.empty()) {
+        preset_states = presets;
+        preset_comments = comments;
     }
-
-    std::ofstream file(preset_file, std::ios_base::app | std::ios_base::out);
-    if (!file) {
-        infoln("Preset file \"{}\" couldn't be loaded.", preset_file);
-        return false;
-    }
-
-    file << "\n# " << preset_name << "\n" << get_current_state().state << std::flush;
-
-    infoln("Refreshing presets...");
-    if (parse_preset_file(preset_file)) {
-        load_preset(preset_states.size() - 1);
-    }
-
-    return true;
+    load_preset(current_preset);
 }
 
 auto state_manager::load_preset(const size_t preset) -> void
 {
-    clear_graph_and_add_current(preset_states.at(preset));
+    clear_graph_and_add_current(preset_states[preset]);
     current_preset = preset;
     edited = false;
 }
@@ -189,7 +132,7 @@ auto state_manager::update_current_state(const puzzle& p) -> void
         move_history.emplace_back(previous_state_index);
     }
 
-    if (p.won()) {
+    if (p.goal_reached()) {
         winning_indices.insert(current_state_index);
     }
 
@@ -269,8 +212,8 @@ auto state_manager::goto_most_distant_state() -> void
     int max_distance = 0;
     size_t max_distance_index = 0;
     for (size_t i = 0; i < node_target_distances.distances.size(); ++i) {
-        if (node_target_distances.distances.at(i) > max_distance) {
-            max_distance = node_target_distances.distances.at(i);
+        if (node_target_distances.distances[i] > max_distance) {
+            max_distance = node_target_distances.distances[i];
             max_distance_index = i;
         }
     }
@@ -284,7 +227,7 @@ auto state_manager::goto_closest_target_state() -> void
         return;
     }
 
-    update_current_state(get_state(node_target_distances.nearest_targets.at(current_state_index)));
+    update_current_state(get_state(node_target_distances.nearest_targets[current_state_index]));
 }
 
 auto state_manager::populate_graph() -> void
@@ -301,12 +244,19 @@ auto state_manager::populate_graph() -> void
     synced_clear_statespace();
 
     // Explore the entire statespace starting from the current state
+    const std::chrono::high_resolution_clock::time_point start = std::chrono::high_resolution_clock::now();
+
     const auto& [states, _links] = s.explore_state_space();
     synced_insert_statespace(states, _links);
 
-    current_state_index = state_indices.at(p);
+    const std::chrono::high_resolution_clock::time_point end = std::chrono::high_resolution_clock::now();
+
+    infoln("Explored puzzle. Took {}ms.", std::chrono::duration_cast<std::chrono::milliseconds>(end - start).count());
+    infoln("State space has size {} with {} transitions.", state_pool.size(), links.size());
+
+    current_state_index = state_indices[p];
     previous_state_index = current_state_index;
-    starting_state_index = state_indices.at(s);
+    starting_state_index = state_indices[s];
 
     // Search for cool stuff
     populate_winning_indices();
@@ -340,7 +290,7 @@ auto state_manager::populate_winning_indices() -> void
 {
     winning_indices.clear();
     for (const auto& [state, index] : state_indices) {
-        if (state.won()) {
+        if (state.goal_reached()) {
             winning_indices.insert(index);
         }
     }
@@ -391,7 +341,7 @@ auto state_manager::get_starting_index() const -> size_t
 
 auto state_manager::get_state(const size_t index) const -> const puzzle&
 {
-    return state_pool.at(index);
+    return state_pool[index];
 }
 
 auto state_manager::get_current_state() const -> const puzzle&
@@ -429,12 +379,12 @@ auto state_manager::get_links() const -> const std::vector<std::pair<size_t, siz
     return links;
 }
 
-auto state_manager::get_winning_indices() const -> const std::unordered_set<size_t>&
+auto state_manager::get_winning_indices() const -> const boost::unordered_flat_set<size_t>&
 {
     return winning_indices;
 }
 
-auto state_manager::get_visit_counts() const -> const std::unordered_map<size_t, int>&
+auto state_manager::get_visit_counts() const -> const boost::unordered_flat_map<size_t, int>&
 {
     return visit_counts;
 }
@@ -444,7 +394,7 @@ auto state_manager::get_winning_path() const -> const std::vector<size_t>&
     return winning_path;
 }
 
-auto state_manager::get_path_indices() const -> const std::unordered_set<size_t>&
+auto state_manager::get_path_indices() const -> const boost::unordered_flat_set<size_t>&
 {
     return path_indices;
 }
@@ -466,7 +416,7 @@ auto state_manager::get_preset_count() const -> size_t
 
 auto state_manager::get_current_preset_comment() const -> const std::string&
 {
-    return preset_comments.at(current_preset);
+    return preset_comments[current_preset];
 }
 
 auto state_manager::has_history() const -> bool

src/threaded_physics.cpp

@@ -1,6 +1,7 @@
 #include "threaded_physics.hpp"
 #include "config.hpp"
 #include "mass_spring_system.hpp"
+#include "util.hpp"
 
 #include <chrono>
 #include <raylib.h>
@@ -8,20 +9,30 @@
 #include <utility>
 #include <vector>
 
-#ifdef TRACY
-#include <tracy/Tracy.hpp>
+#ifdef ASYNC_OCTREE
+auto threaded_physics::set_octree_pool_thread_name(size_t idx) -> void
+{
+    BS::this_thread::set_os_thread_name(std::format("octree-{}", idx));
+    // traceln("Using thread \"{}\"", BS::this_thread::get_os_thread_name().value_or("INVALID NAME"));
+}
 #endif
 
-auto threaded_physics::physics_thread(physics_state& state) -> void
+auto threaded_physics::physics_thread(physics_state& state, const std::optional<BS::thread_pool<>* const> thread_pool) -> void
 {
-    #ifdef THREADPOOL
-    BS::this_thread::set_os_thread_name("physics");
-    #endif
-
     mass_spring_system mass_springs;
 
+    #ifdef ASYNC_OCTREE
+    BS::this_thread::set_os_thread_name("physics");
+
+    BS::thread_pool<> octree_thread(1, set_octree_pool_thread_name);
+    std::future<void> octree_future;
+    octree tree_buffer;
+    size_t last_mass_count = 0;
+    infoln("Using asynchronous octree builder.");
+    #endif
+
     const auto visitor = overloads{
-        [&](const struct add_mass& am)
+        [&](const struct add_mass&)
         {
             mass_springs.add_mass();
         },
@@ -29,7 +40,7 @@ auto threaded_physics::physics_thread(physics_state& state) -> void
         {
             mass_springs.add_spring(as.a, as.b);
         },
-        [&](const struct clear_graph& cg)
+        [&](const struct clear_graph&)
         {
             mass_springs.clear();
         },
@@ -66,26 +77,58 @@ auto threaded_physics::physics_thread(physics_state& state) -> void
             }
         }
 
-        if (mass_springs.masses.empty()) {
+        if (mass_springs.positions.empty()) {
             std::this_thread::sleep_for(std::chrono::milliseconds(1));
             continue;
         }
 
         // Physics update
         if (physics_accumulator.count() > TIMESTEP) {
+            #ifdef ASYNC_OCTREE
+            // Snapshot the positions so mass_springs is not mutating the vector while the octree is building
+            std::vector<Vector3> positions = mass_springs.positions;
+
+            // Start building the octree for the next physics update.
+            // Move the snapshot into the closure so it doesn't get captured by reference (don't use [&])
+            octree_future = octree_thread.submit_task([&tree_buffer, positions = std::move(positions)]()
+            {
+                octree::build_octree(tree_buffer, positions);
+            });
+
+            // Rebuild the tree synchronously if we changed the number of masses to not use
+            // an empty tree from the last frame in the frame where the graph was generated
+            if (last_mass_count != mass_springs.positions.size()) {
+                traceln("Rebuilding octree synchronously because graph size changed");
+                octree::build_octree(mass_springs.tree, mass_springs.positions);
+                last_mass_count = mass_springs.positions.size();
+            }
+            #else
+            octree::build_octree(mass_springs.tree, mass_springs.positions);
+            #endif
+
             mass_springs.clear_forces();
-            mass_springs.calculate_spring_forces();
-            mass_springs.calculate_repulsion_forces();
-            mass_springs.verlet_update(TIMESTEP * SIM_SPEED);
+            mass_springs.calculate_spring_forces(thread_pool);
+            mass_springs.calculate_repulsion_forces(thread_pool);
+            mass_springs.update(TIMESTEP * SIM_SPEED, thread_pool);
 
             // This is only helpful if we're drawing a grid at (0, 0, 0). Otherwise, it's just
             // expensive and yields no benefit since we can lock the camera to the center of mass
-            // cheaply. mass_springs.center_masses();
+            // cheaply.
+            // mass_springs.center_masses(thread_pool);
 
             ++loop_iterations;
             physics_accumulator -= std::chrono::duration<double>(TIMESTEP);
         }
 
+        #ifdef ASYNC_OCTREE
+        // Wait for the octree to be built
+        if (octree_future.valid()) {
+            octree_future.wait();
+            octree_future = std::future<void>{};
+            std::swap(mass_springs.tree, tree_buffer);
+        }
+        #endif
+
         // Publish the positions for the renderer (copy)
         #ifdef TRACY
         FrameMarkStart("PhysicsThreadProduceLock");
@@ -114,16 +157,16 @@ auto threaded_physics::physics_thread(physics_state& state) -> void
             if (mass_springs.tree.nodes.empty()) {
                 state.mass_center = Vector3Zero();
             } else {
-                state.mass_center = mass_springs.tree.nodes.at(0).mass_center;
+                state.mass_center = mass_springs.tree.nodes[0].mass_center;
             }
 
             state.masses.clear();
-            state.masses.reserve(mass_springs.masses.size());
-            for (const auto& mass : mass_springs.masses) {
-                state.masses.emplace_back(mass.position);
+            state.masses.reserve(mass_springs.positions.size());
+            for (const Vector3& pos : mass_springs.positions) {
+                state.masses.emplace_back(pos);
             }
 
-            state.mass_count = mass_springs.masses.size();
+            state.mass_count = mass_springs.positions.size();
             state.spring_count = mass_springs.springs.size();
 
             state.data_ready = true;
@@ -131,14 +174,25 @@ auto threaded_physics::physics_thread(physics_state& state) -> void
         }
         // Notify the rendering thread that new data is available
         state.data_ready_cnd.notify_all();
-        #ifdef TRACY
-        FrameMarkEnd("PhysicsThreadProduceLock");
 
-        FrameMarkEnd("PhysicsThread");
+        #ifdef TRACY
+        FrameMarkEnd("PhysicsThreadProduceLock"); FrameMarkEnd("PhysicsThread");
         #endif
     }
 }
 
+auto threaded_physics::clear_cmd() -> void
+{
+    {
+        #ifdef TRACY
+        std::lock_guard<LockableBase(std::mutex)> lock(state.command_mtx);
+        #else
+        std::lock_guard<std::mutex> lock(state.command_mtx);
+        #endif
+        state.pending_commands.emplace(clear_graph{});
+    }
+}
+
 auto threaded_physics::add_mass_cmd() -> void
 {
     {
@@ -163,18 +217,6 @@ auto threaded_physics::add_spring_cmd(const size_t a, const size_t b) -> void
     }
 }
 
-auto threaded_physics::clear_cmd() -> void
-{
-    {
-        #ifdef TRACY
-        std::lock_guard<LockableBase(std::mutex)> lock(state.command_mtx);
-        #else
-        std::lock_guard<std::mutex> lock(state.command_mtx);
-        #endif
-        state.pending_commands.emplace(clear_graph{});
-    }
-}
-
 auto threaded_physics::add_mass_springs_cmd(const size_t num_masses,
                                             const std::vector<std::pair<size_t, size_t>>& springs) -> void
 {

src/user_interface.cpp

@@ -7,13 +7,8 @@
 #define RAYGUI_IMPLEMENTATION
 #include <raygui.h>
 
-#ifdef TRACY
-    #include <tracy/Tracy.hpp>
-#endif
-
-auto user_interface::grid::update_bounds(const int _x, const int _y, const int _width,
-                                         const int _height, const int _columns, const int _rows)
-    -> void
+auto user_interface::grid::update_bounds(const int _x, const int _y, const int _width, const int _height,
+                                         const int _columns, const int _rows) -> void
 {
     x = _x;
     y = _y;
@@ -23,8 +18,7 @@ auto user_interface::grid::update_bounds(const int _x, const int _y, const int _
     rows = _rows;
 }
 
-auto user_interface::grid::update_bounds(const int _x, const int _y, const int _width,
-                                         const int _height) -> void
+auto user_interface::grid::update_bounds(const int _x, const int _y, const int _width, const int _height) -> void
 {
     x = _x;
     y = _y;
@@ -48,19 +42,17 @@ auto user_interface::grid::bounds() const -> Rectangle
     return bounds;
 }
 
-auto user_interface::grid::bounds(const int _x, const int _y, const int _width,
-                                  const int _height) const -> Rectangle
+auto user_interface::grid::bounds(const int _x, const int _y, const int _width, const int _height) const -> Rectangle
 {
     if (_x < 0 || _x + _width > columns || _y < 0 || _y + _height > rows) {
-        errln("Grid bounds are outside range.");
-        exit(1);
+        throw std::invalid_argument("Grid bounds out of range");
     }
 
     const int cell_width = (width - padding) / columns;
     const int cell_height = (height - padding) / rows;
 
-    return Rectangle(x + _x * cell_width + padding, y + _y * cell_height + padding,
-                     _width * cell_width - padding, _height * cell_height - padding);
+    return Rectangle(x + _x * cell_width + padding, y + _y * cell_height + padding, _width * cell_width - padding,
+                     _height * cell_height - padding);
 }
 
 auto user_interface::grid::square_bounds() const -> Rectangle
@@ -80,8 +72,7 @@ auto user_interface::grid::square_bounds(const int _x, const int _y, const int _
     // filled
 
     if (_x < 0 || _x + _width > columns || _y < 0 || _y + _height > rows) {
-        errln("Grid bounds are outside range.");
-        exit(1);
+        throw std::invalid_argument("Grid bounds out of range");
     }
 
     const int available_width = width - padding * (columns + 1);
@@ -93,10 +84,8 @@ auto user_interface::grid::square_bounds(const int _x, const int _y, const int _
     const int x_offset = (width - grid_width) / 2;
     const int y_offset = (height - grid_height) / 2;
 
-    return Rectangle(x_offset + _x * (cell_size + padding) + padding,
-                     y_offset + _y * (cell_size + padding) + padding,
-                     _width * cell_size + padding * (_width - 1),
-                     _height * cell_size + padding * (_height - 1));
+    return Rectangle(x_offset + _x * (cell_size + padding) + padding, y_offset + _y * (cell_size + padding) + padding,
+                     _width * cell_size + padding * (_width - 1), _height * cell_size + padding * (_height - 1));
 }
 
 auto user_interface::init() -> void
@@ -155,19 +144,19 @@ auto user_interface::get_default_style() -> default_style
 {
     // Could've iterated over the values, but then it wouldn't be as nice to
     // access...
-    return {{GuiGetStyle(DEFAULT, BORDER_COLOR_NORMAL), GuiGetStyle(DEFAULT, BASE_COLOR_NORMAL),
-             GuiGetStyle(DEFAULT, TEXT_COLOR_NORMAL), GuiGetStyle(DEFAULT, BORDER_COLOR_FOCUSED),
-             GuiGetStyle(DEFAULT, BASE_COLOR_FOCUSED), GuiGetStyle(DEFAULT, TEXT_COLOR_FOCUSED),
-             GuiGetStyle(DEFAULT, BORDER_COLOR_PRESSED), GuiGetStyle(DEFAULT, BASE_COLOR_PRESSED),
-             GuiGetStyle(DEFAULT, TEXT_COLOR_PRESSED), GuiGetStyle(DEFAULT, BORDER_COLOR_DISABLED),
-             GuiGetStyle(DEFAULT, BASE_COLOR_DISABLED), GuiGetStyle(DEFAULT, TEXT_COLOR_DISABLED)},
-            GuiGetStyle(DEFAULT, BACKGROUND_COLOR),
-            GuiGetStyle(DEFAULT, LINE_COLOR),
-            GuiGetStyle(DEFAULT, TEXT_SIZE),
-            GuiGetStyle(DEFAULT, TEXT_SPACING),
-            GuiGetStyle(DEFAULT, TEXT_LINE_SPACING),
-            GuiGetStyle(DEFAULT, TEXT_ALIGNMENT_VERTICAL),
-            GuiGetStyle(DEFAULT, TEXT_WRAP_MODE)};
+    return {
+        {
+            GuiGetStyle(DEFAULT, BORDER_COLOR_NORMAL), GuiGetStyle(DEFAULT, BASE_COLOR_NORMAL),
+            GuiGetStyle(DEFAULT, TEXT_COLOR_NORMAL), GuiGetStyle(DEFAULT, BORDER_COLOR_FOCUSED),
+            GuiGetStyle(DEFAULT, BASE_COLOR_FOCUSED), GuiGetStyle(DEFAULT, TEXT_COLOR_FOCUSED),
+            GuiGetStyle(DEFAULT, BORDER_COLOR_PRESSED), GuiGetStyle(DEFAULT, BASE_COLOR_PRESSED),
+            GuiGetStyle(DEFAULT, TEXT_COLOR_PRESSED), GuiGetStyle(DEFAULT, BORDER_COLOR_DISABLED),
+            GuiGetStyle(DEFAULT, BASE_COLOR_DISABLED), GuiGetStyle(DEFAULT, TEXT_COLOR_DISABLED)
+        },
+        GuiGetStyle(DEFAULT, BACKGROUND_COLOR), GuiGetStyle(DEFAULT, LINE_COLOR), GuiGetStyle(DEFAULT, TEXT_SIZE),
+        GuiGetStyle(DEFAULT, TEXT_SPACING), GuiGetStyle(DEFAULT, TEXT_LINE_SPACING),
+        GuiGetStyle(DEFAULT, TEXT_ALIGNMENT_VERTICAL), GuiGetStyle(DEFAULT, TEXT_WRAP_MODE)
+    };
 }
 
 auto user_interface::set_default_style(const default_style& style) -> void
@@ -201,15 +190,17 @@ auto user_interface::set_default_style(const default_style& style) -> void
 auto user_interface::get_component_style(const int component) -> component_style
 {
     return {
-        {GuiGetStyle(component, BORDER_COLOR_NORMAL), GuiGetStyle(component, BASE_COLOR_NORMAL),
-         GuiGetStyle(component, TEXT_COLOR_NORMAL), GuiGetStyle(component, BORDER_COLOR_FOCUSED),
-         GuiGetStyle(component, BASE_COLOR_FOCUSED), GuiGetStyle(component, TEXT_COLOR_FOCUSED),
-         GuiGetStyle(component, BORDER_COLOR_PRESSED), GuiGetStyle(component, BASE_COLOR_PRESSED),
-         GuiGetStyle(component, TEXT_COLOR_PRESSED), GuiGetStyle(component, BORDER_COLOR_DISABLED),
-         GuiGetStyle(component, BASE_COLOR_DISABLED), GuiGetStyle(component, TEXT_COLOR_DISABLED)},
-        GuiGetStyle(component, BORDER_WIDTH),
-        GuiGetStyle(component, TEXT_PADDING),
-        GuiGetStyle(component, TEXT_ALIGNMENT)};
+        {
+            GuiGetStyle(component, BORDER_COLOR_NORMAL), GuiGetStyle(component, BASE_COLOR_NORMAL),
+            GuiGetStyle(component, TEXT_COLOR_NORMAL), GuiGetStyle(component, BORDER_COLOR_FOCUSED),
+            GuiGetStyle(component, BASE_COLOR_FOCUSED), GuiGetStyle(component, TEXT_COLOR_FOCUSED),
+            GuiGetStyle(component, BORDER_COLOR_PRESSED), GuiGetStyle(component, BASE_COLOR_PRESSED),
+            GuiGetStyle(component, TEXT_COLOR_PRESSED), GuiGetStyle(component, BORDER_COLOR_DISABLED),
+            GuiGetStyle(component, BASE_COLOR_DISABLED), GuiGetStyle(component, TEXT_COLOR_DISABLED)
+        },
+        GuiGetStyle(component, BORDER_WIDTH), GuiGetStyle(component, TEXT_PADDING),
+        GuiGetStyle(component, TEXT_ALIGNMENT)
+    };
 }
 
 auto user_interface::set_component_style(const int component, const component_style& style) -> void
@@ -238,13 +229,11 @@ auto user_interface::set_component_style(const int component, const component_st
 auto user_interface::popup_bounds() -> Rectangle
 {
     return Rectangle(static_cast<float>(GetScreenWidth()) / 2.0f - POPUP_WIDTH / 2.0f,
-                     static_cast<float>(GetScreenHeight()) / 2.0f - POPUP_HEIGHT / 2.0f,
-                     POPUP_WIDTH, POPUP_HEIGHT);
+                     static_cast<float>(GetScreenHeight()) / 2.0f - POPUP_HEIGHT / 2.0f, POPUP_WIDTH, POPUP_HEIGHT);
 }
 
-auto user_interface::draw_button(const Rectangle bounds, const std::string& label,
-                                 const Color color, const bool enabled, const int font_size) const
-    -> int
+auto user_interface::draw_button(const Rectangle bounds, const std::string& label, const Color color,
+                                 const bool enabled, const int font_size) const -> int
 {
     // Save original styling
     const default_style original_default = get_default_style();
@@ -275,16 +264,16 @@ auto user_interface::draw_button(const Rectangle bounds, const std::string& labe
 }
 
 auto user_interface::draw_menu_button(const int x, const int y, const int width, const int height,
-                                      const std::string& label, const Color color,
-                                      const bool enabled, const int font_size) const -> int
+                                      const std::string& label, const Color color, const bool enabled,
+                                      const int font_size) const -> int
 {
     const Rectangle bounds = menu_grid.bounds(x, y, width, height);
     return draw_button(bounds, label, color, enabled, font_size);
 }
 
 auto user_interface::draw_toggle_slider(const Rectangle bounds, const std::string& off_label,
-                                        const std::string& on_label, int* active, Color color,
-                                        bool enabled, int font_size) const -> int
+                                        const std::string& on_label, int* active, Color color, bool enabled,
+                                        int font_size) const -> int
 {
     // Save original styling
     const default_style original_default = get_default_style();
@@ -319,18 +308,16 @@ auto user_interface::draw_toggle_slider(const Rectangle bounds, const std::strin
     return pressed;
 }
 
-auto user_interface::draw_menu_toggle_slider(const int x, const int y, const int width,
-                                             const int height, const std::string& off_label,
-                                             const std::string& on_label, int* active,
-                                             const Color color, const bool enabled,
-                                             const int font_size) const -> int
+auto user_interface::draw_menu_toggle_slider(const int x, const int y, const int width, const int height,
+                                             const std::string& off_label, const std::string& on_label, int* active,
+                                             const Color color, const bool enabled, const int font_size) const -> int
 {
     const Rectangle bounds = menu_grid.bounds(x, y, width, height);
     return draw_toggle_slider(bounds, off_label, on_label, active, color, enabled, font_size);
 }
 
-auto user_interface::draw_spinner(Rectangle bounds, const std::string& label, int* value, int min,
-                                  int max, Color color, bool enabled, int font_size) const -> int
+auto user_interface::draw_spinner(Rectangle bounds, const std::string& label, int* value, int min, int max, Color color,
+                                  bool enabled, int font_size) const -> int
 {
     // Save original styling
     const default_style original_default = get_default_style();
@@ -366,16 +353,15 @@ auto user_interface::draw_spinner(Rectangle bounds, const std::string& label, in
 }
 
 auto user_interface::draw_menu_spinner(const int x, const int y, const int width, const int height,
-                                       const std::string& label, int* value, const int min,
-                                       const int max, const Color color, const bool enabled,
-                                       const int font_size) const -> int
+                                       const std::string& label, int* value, const int min, const int max,
+                                       const Color color, const bool enabled, const int font_size) const -> int
 {
     const Rectangle bounds = menu_grid.bounds(x, y, width, height);
     return draw_spinner(bounds, label, value, min, max, color, enabled, font_size);
 }
 
-auto user_interface::draw_label(const Rectangle bounds, const std::string& text, const Color color,
-                                const bool enabled, const int font_size) const -> int
+auto user_interface::draw_label(const Rectangle bounds, const std::string& text, const Color color, const bool enabled,
+                                const int font_size) const -> int
 {
     // Save original styling
     const default_style original_default = get_default_style();
@@ -405,8 +391,8 @@ auto user_interface::draw_label(const Rectangle bounds, const std::string& text,
     return pressed;
 }
 
-auto user_interface::draw_board_block(const int x, const int y, const int width, const int height,
-                                      const Color color, const bool enabled) const -> bool
+auto user_interface::draw_board_block(const int x, const int y, const int width, const int height, const Color color,
+                                      const bool enabled) const -> bool
 {
     component_style s = get_component_style(BUTTON);
     apply_block_color(s, color);
@@ -452,18 +438,19 @@ auto user_interface::window_open() const -> bool
 auto user_interface::draw_menu_header(const Color color) const -> void
 {
     int preset = static_cast<int>(state.get_current_preset());
-    draw_menu_spinner(0, 0, 1, 1, "Preset: ", &preset, -1,
-                      static_cast<int>(state.get_preset_count()), color, !input.editing);
+    draw_menu_spinner(0, 0, 1, 1, "Preset: ", &preset, -1, static_cast<int>(state.get_preset_count()), color,
+                      !input.editing);
     if (preset > static_cast<int>(state.get_current_preset())) {
         input.load_next_preset();
     } else if (preset < static_cast<int>(state.get_current_preset())) {
         input.load_previous_preset();
     }
 
-    draw_menu_button(1, 0, 1, 1,
-                     std::format("Puzzle: \"{}\"{}", state.get_current_preset_comment().substr(2),
-                                 state.was_edited() ? " (Modified)" : ""),
-                     color);
+    draw_menu_button(1, 0, 1, 1, std::format("{}: {}/{} Blocks",
+                                             state.was_edited()
+                                                 ? "Modified"
+                                                 : std::format("\"{}\"", state.get_current_preset_comment().substr(2)),
+                                             state.get_current_state().block_count(), puzzle::MAX_BLOCKS), color);
 
     int editing = input.editing;
     draw_menu_toggle_slider(2, 0, 1, 1, "Puzzle Mode (Tab)", "Edit Mode (Tab)", &editing, color);
@@ -474,18 +461,13 @@ auto user_interface::draw_menu_header(const Color color) const -> void
 
 auto user_interface::draw_graph_info(const Color color) const -> void
 {
-    draw_menu_button(0, 1, 1, 1,
-                     std::format("Found {} States ({} Winning)", state.get_state_count(),
-                                 state.get_target_count()),
-                     color);
+    draw_menu_button(0, 1, 1, 1, std::format("Found {} States ({} Winning)", state.get_state_count(),
+                                             state.get_target_count()), color);
 
-    draw_menu_button(1, 1, 1, 1, std::format("Found {} Transitions", state.get_link_count()),
-                     color);
+    draw_menu_button(1, 1, 1, 1, std::format("Found {} Transitions", state.get_link_count()), color);
 
-    draw_menu_button(2, 1, 1, 1,
-                     std::format("{} Moves to Nearest Solution",
-                                 state.get_path_length() > 0 ? state.get_path_length() - 1 : 0),
-                     color);
+    draw_menu_button(2, 1, 1, 1, std::format("{} Moves to Nearest Solution",
+                                             state.get_path_length() > 0 ? state.get_path_length() - 1 : 0), color);
 }
 
 auto user_interface::draw_graph_controls(const Color color) const -> void
@@ -506,8 +488,7 @@ auto user_interface::draw_graph_controls(const Color color) const -> void
     }
 
     int mark_solutions = input.mark_solutions;
-    draw_menu_toggle_slider(2, 2, 1, 1, "Solution Hidden (I)", "Solution Shown (I)",
-                            &mark_solutions, color);
+    draw_menu_toggle_slider(2, 2, 1, 1, "Solution Hidden (I)", "Solution Shown (I)", &mark_solutions, color);
     if (mark_solutions != input.mark_solutions) {
         input.toggle_mark_solutions();
     }
@@ -517,22 +498,20 @@ auto user_interface::draw_graph_controls(const Color color) const -> void
 auto user_interface::draw_camera_controls(const Color color) const -> void
 {
     int lock_camera = input.camera_lock;
-    draw_menu_toggle_slider(0, 3, 1, 1, "Free Camera (L)", "Locked Camera (L)", &lock_camera,
-                            color);
+    draw_menu_toggle_slider(0, 3, 1, 1, "Free Camera (L)", "Locked Camera (L)", &lock_camera, color);
     if (lock_camera != input.camera_lock) {
         input.toggle_camera_lock();
     }
 
     int lock_camera_mass_center = input.camera_mass_center_lock;
-    draw_menu_toggle_slider(1, 3, 1, 1, "Current Block (U)", "Graph Center (U)",
-                            &lock_camera_mass_center, color, input.camera_lock);
+    draw_menu_toggle_slider(1, 3, 1, 1, "Current Block (U)", "Graph Center (U)", &lock_camera_mass_center, color,
+                            input.camera_lock);
     if (lock_camera_mass_center != input.camera_mass_center_lock) {
         input.toggle_camera_mass_center_lock();
     }
 
     int projection = camera.projection == CAMERA_ORTHOGRAPHIC;
-    draw_menu_toggle_slider(2, 3, 1, 1, "Perspective (Alt)", "Orthographic (Alt)", &projection,
-                            color);
+    draw_menu_toggle_slider(2, 3, 1, 1, "Perspective (Alt)", "Orthographic (Alt)", &projection, color);
     if (projection != (camera.projection == CAMERA_ORTHOGRAPHIC)) {
         input.toggle_camera_projection();
     }
@@ -558,10 +537,8 @@ auto user_interface::draw_puzzle_controls(const Color color) const -> void
     };
 
     const int visits = state.get_current_visits();
-    draw_menu_button(0, 4, 1, 1,
-                     std::format("{} Moves ({}{} Time at this State)", state.get_total_moves(),
-                                 visits, nth(visits)),
-                     color);
+    draw_menu_button(0, 4, 1, 1, std::format("{} Moves ({}{} Time at this State)", state.get_total_moves(), visits,
+                                             nth(visits)), color);
 
     if (draw_menu_button(1, 4, 1, 1, "Make Optimal Move (Space)", color, state.has_distances())) {
         input.goto_optimal_next_state();
@@ -600,43 +577,40 @@ auto user_interface::draw_edit_controls(const Color color) const -> void
     }
 
     // Toggle Restricted/Free Block Movement
-    int free = !current.restricted;
+    int free = !current.get_restricted();
     draw_menu_toggle_slider(1, 4, 1, 1, "Restricted (F)", "Free (F)", &free, color);
-    if (free != !current.restricted) {
+    if (free != !current.get_restricted()) {
         input.toggle_restricted_movement();
     }
 
     // Clear Goal
-    if (draw_menu_button(1, 5, 1, 1, "Clear Goal (X)", color)) {
-    }
+    if (draw_menu_button(1, 5, 1, 1, "Clear Goal (X)", color)) {}
 
     // Column Count Spinner
-    int columns = current.width;
+    int columns = current.get_width();
     draw_menu_spinner(2, 4, 1, 1, "Cols: ", &columns, puzzle::MIN_WIDTH, puzzle::MAX_WIDTH, color);
-    if (columns > current.width) {
+    if (columns > current.get_width()) {
         input.add_board_column();
-    } else if (columns < current.width) {
+    } else if (columns < current.get_width()) {
         input.remove_board_column();
     }
 
     // Row Count Spinner
-    int rows = current.height;
+    int rows = current.get_height();
     draw_menu_spinner(2, 5, 1, 1, "Rows: ", &rows, puzzle::MIN_WIDTH, puzzle::MAX_WIDTH, color);
-    if (rows > current.height) {
+    if (rows > current.get_height()) {
         input.add_board_row();
-    } else if (rows < current.height) {
+    } else if (rows < current.get_height()) {
         input.remove_board_row();
     }
 }
 
 auto user_interface::draw_menu_footer(const Color color) -> void
 {
-    draw_menu_button(0, 6, 2, 1, std::format("State: \"{}\"", state.get_current_state().state),
-                     color);
+    draw_menu_button(0, 6, 2, 1, state.get_current_state().string_repr().data(), color);
 
     if (draw_menu_button(2, 6, 1, 1, "Save as Preset", color)) {
-        if (const std::optional<std::string>& reason =
-                state.get_current_state().try_get_invalid_reason()) {
+        if (const std::optional<std::string>& reason = state.get_current_state().try_get_invalid_reason()) {
             message_title = "Can't Save Preset";
             message_message = std::format("Invalid Board: {}.", *reason);
             ok_message = true;
@@ -651,8 +625,7 @@ auto user_interface::get_background_color() -> Color
     return GetColor(GuiGetStyle(DEFAULT, BACKGROUND_COLOR));
 }
 
-auto user_interface::help_popup() -> void
-{}
+auto user_interface::help_popup() -> void {}
 
 auto user_interface::draw_save_preset_popup() -> void
 {
@@ -661,14 +634,14 @@ auto user_interface::draw_save_preset_popup() -> void
     }
 
     // Returns the pressed button index
-    const int button = GuiTextInputBox(popup_bounds(), "Save as Preset", "Enter Preset Name",
-                                       "Ok;Cancel", preset_name.data(), 255, nullptr);
+    const int button = GuiTextInputBox(popup_bounds(), "Save as Preset", "Enter Preset Name", "Ok;Cancel",
+                                       preset_comment.data(), 255, nullptr);
     if (button == 1) {
-        state.append_preset_file(preset_name.data());
+        state.save_current_to_preset_file(preset_comment.data());
     }
     if (button == 0 || button == 1 || button == 2) {
         save_window = false;
-        TextCopy(preset_name.data(), "\0");
+        TextCopy(preset_comment.data(), "\0");
     }
 }
 
@@ -678,8 +651,7 @@ auto user_interface::draw_ok_message_box() -> void
         return;
     }
 
-    const int button =
-        GuiMessageBox(popup_bounds(), message_title.data(), message_message.data(), "Ok");
+    const int button = GuiMessageBox(popup_bounds(), message_title.data(), message_message.data(), "Ok");
     if (button == 0 || button == 1) {
         message_title = "";
         message_message = "";
@@ -693,8 +665,7 @@ auto user_interface::draw_yes_no_message_box() -> void
         return;
     }
 
-    const int button =
-        GuiMessageBox(popup_bounds(), message_title.data(), message_message.data(), "Yes;No");
+    const int button = GuiMessageBox(popup_bounds(), message_title.data(), message_message.data(), "Yes;No");
     if (button == 1) {
         yes_no_handler();
     }
@@ -727,37 +698,36 @@ auto user_interface::draw_puzzle_board() -> void
 {
     const puzzle& current = state.get_current_state();
 
-    board_grid.update_bounds(0, MENU_HEIGHT, GetScreenWidth() / 2, GetScreenHeight() - MENU_HEIGHT,
-                             current.width, current.height);
+    board_grid.update_bounds(0, MENU_HEIGHT, GetScreenWidth() / 2, GetScreenHeight() - MENU_HEIGHT, current.get_width(),
+                             current.get_height());
 
     // Draw outer border
     const Rectangle bounds = board_grid.square_bounds();
-    DrawRectangleRec(bounds, current.won() ? BOARD_COLOR_WON : BOARD_COLOR_RESTRICTED);
+    DrawRectangleRec(bounds, current.goal_reached() ? BOARD_COLOR_WON : BOARD_COLOR_RESTRICTED);
 
     // Draw inner borders
-    DrawRectangle(bounds.x + BOARD_PADDING, bounds.y + BOARD_PADDING,
-                  bounds.width - 2 * BOARD_PADDING, bounds.height - 2 * BOARD_PADDING,
-                  current.restricted ? BOARD_COLOR_RESTRICTED : BOARD_COLOR_FREE);
+    DrawRectangle(bounds.x + BOARD_PADDING, bounds.y + BOARD_PADDING, bounds.width - 2 * BOARD_PADDING,
+                  bounds.height - 2 * BOARD_PADDING,
+                  current.get_restricted() ? BOARD_COLOR_RESTRICTED : BOARD_COLOR_FREE);
 
     // Draw target opening
     // TODO: Only draw single direction (in corner) if restricted (use target block principal
     // direction)
     const std::optional<puzzle::block> target_block = current.try_get_target_block();
-    const int target_x = current.target_x;
-    const int target_y = current.target_y;
-    if (current.has_win_condition() && target_block) {
-        auto [x, y, width, height] = board_grid.square_bounds(
-            target_x, target_y, target_block.value().width, target_block.value().height);
+    const int target_x = current.get_goal_x();
+    const int target_y = current.get_goal_y();
+    if (current.get_goal() && target_block) {
+        auto [x, y, width, height] = board_grid.square_bounds(target_x, target_y, target_block->get_width(),
+                                                              target_block->get_height());
 
-        const Color opening_color =
-            Fade(current.won() ? BOARD_COLOR_WON : BOARD_COLOR_RESTRICTED, 0.3);
+        const Color opening_color = Fade(current.goal_reached() ? BOARD_COLOR_WON : BOARD_COLOR_RESTRICTED, 0.3);
 
         if (target_x == 0) {
             // Left opening
             DrawRectangle(x - BOARD_PADDING, y, BOARD_PADDING, height, RAYWHITE);
             DrawRectangle(x - BOARD_PADDING, y, BOARD_PADDING, height, opening_color);
         }
-        if (target_x + target_block.value().width == current.width) {
+        if (target_x + target_block->get_width() == current.get_width()) {
             // Right opening
             DrawRectangle(x + width, y, BOARD_PADDING, height, RAYWHITE);
             DrawRectangle(x + width, y, BOARD_PADDING, height, opening_color);
@@ -767,7 +737,7 @@ auto user_interface::draw_puzzle_board() -> void
             DrawRectangle(x, y - BOARD_PADDING, width, BOARD_PADDING, RAYWHITE);
             DrawRectangle(x, y - BOARD_PADDING, width, BOARD_PADDING, opening_color);
         }
-        if (target_y + target_block.value().height == current.height) {
+        if (target_y + target_block->get_height() == current.get_height()) {
             // Bottom opening
             DrawRectangle(x, y + height, width, BOARD_PADDING, RAYWHITE);
             DrawRectangle(x, y + height, width, BOARD_PADDING, opening_color);
@@ -794,87 +764,87 @@ auto user_interface::draw_puzzle_board() -> void
     }
 
     // Draw blocks
-    for (const puzzle::block& b : current) {
+    for (const puzzle::block b : current.block_view()) {
         Color c = BLOCK_COLOR;
-        if (b.target) {
+        if (b.get_target()) {
             c = TARGET_BLOCK_COLOR;
-        } else if (b.immovable) {
+        } else if (b.get_immovable()) {
             c = WALL_COLOR;
         }
 
-        if (!b.valid() || b.x < 0 || b.y < 0 || b.width <= 0 || b.height <= 0) {
-            warnln("Iterator returned invalid block for state \"{}\".", current.state);
-            continue;
-        }
+        const auto [x, y, w, h, t, i] = b.unpack_repr();
 
-        draw_board_block(b.x, b.y, b.width, b.height, c, !b.immovable);
+        draw_board_block(x, y, w, h, c, !i);
     }
 
     // Draw block placing
     if (input.editing && input.has_block_add_xy) {
-        if (current.covers(input.block_add_x, input.block_add_y) &&
-            input.hov_x >= input.block_add_x && input.hov_y >= input.block_add_y) {
+        if (current.covers(input.block_add_x, input.block_add_y) && input.hov_x >= input.block_add_x && input.hov_y >=
+            input.block_add_y) {
             bool collides = false;
-            for (const puzzle::block& b : current) {
-                if (b.collides(puzzle::block(input.block_add_x, input.block_add_y,
-                                             input.hov_x - input.block_add_x + 1,
+            for (const puzzle::block b : current.block_view()) {
+                if (b.collides(puzzle::block(input.block_add_x, input.block_add_y, input.hov_x - input.block_add_x + 1,
                                              input.hov_y - input.block_add_y + 1, false))) {
                     collides = true;
                     break;
                 }
             }
             if (!collides) {
-                draw_board_block(input.block_add_x, input.block_add_y,
-                                 input.hov_x - input.block_add_x + 1,
+                draw_board_block(input.block_add_x, input.block_add_y, input.hov_x - input.block_add_x + 1,
                                  input.hov_y - input.block_add_y + 1, PURPLE);
             }
         }
     }
 
+    // TODO: In edit mode
+    //       - Clear Goal button doesn't work
+    //       - Toggle Target Block button throws "Grid bounds out of range"
+    //       - Clicking the goal to remove it throws "Grid bounds out of range"
     // Draw goal boundaries when editing
-    if (input.editing) {
-        DrawRectangleLinesEx(board_grid.square_bounds(target_x, target_y, target_block->width, target_block->height), 2.0, TARGET_BLOCK_COLOR);
+    if (input.editing && current.get_goal() && target_block) {
+        DrawRectangleLinesEx(
+            board_grid.square_bounds(target_x, target_y, target_block->get_width(), target_block->get_height()), 2.0,
+            TARGET_BLOCK_COLOR);
     }
 }
 
-auto user_interface::draw_graph_overlay(int fps, int ups, size_t mass_count, size_t spring_count)
-    -> void
+auto user_interface::draw_graph_overlay(int fps, int ups, size_t mass_count, size_t spring_count) -> void
 {
     graph_overlay_grid.update_bounds(GetScreenWidth() / 2, MENU_HEIGHT);
     debug_overlay_grid.update_bounds(GetScreenWidth() / 2, GetScreenHeight() - 75);
 
-    draw_label(graph_overlay_grid.bounds(0, 0, 1, 1),
-               std::format("Dist: {:0>7.2f}", camera.distance), BLACK);
-    draw_label(graph_overlay_grid.bounds(0, 1, 1, 1), std::format("FoV:  {:0>6.2f}", camera.fov),
-               BLACK);
+    draw_label(graph_overlay_grid.bounds(0, 0, 1, 1), std::format("Dist: {:0>7.2f}", camera.distance), BLACK);
+    draw_label(graph_overlay_grid.bounds(0, 1, 1, 1), std::format("FoV:  {:0>6.2f}", camera.fov), BLACK);
     draw_label(graph_overlay_grid.bounds(0, 2, 1, 1), std::format("FPS:  {:0>3}", fps), LIME);
     draw_label(graph_overlay_grid.bounds(0, 3, 1, 1), std::format("UPS:  {:0>3}", ups), ORANGE);
 
     // Debug
     draw_label(debug_overlay_grid.bounds(0, 0, 1, 1), std::format("Physics Debug:"), BLACK);
-    draw_label(debug_overlay_grid.bounds(0, 1, 1, 1), std::format("Masses:  {}", mass_count),
-               BLACK);
-    draw_label(debug_overlay_grid.bounds(0, 2, 1, 1), std::format("Springs: {}", spring_count),
-               BLACK);
+    draw_label(debug_overlay_grid.bounds(0, 1, 1, 1), std::format("Masses:  {}", mass_count), BLACK);
+    draw_label(debug_overlay_grid.bounds(0, 2, 1, 1), std::format("Springs: {}", spring_count), BLACK);
 }
 
-auto user_interface::draw(const int fps, const int ups, const size_t mass_count,
-                          const size_t spring_count) -> void
+auto user_interface::draw(const int fps, const int ups, const size_t mass_count, const size_t spring_count) -> void
 {
-    const auto visitor = overloads{[&](const show_ok_message& msg)
-                                   {
-                                       message_title = msg.title;
-                                       message_message = msg.message;
-                                       ok_message = true;
-                                   },
-                                   [&](const show_yes_no_message& msg)
-                                   {
-                                       message_title = msg.title;
-                                       message_message = msg.message;
-                                       yes_no_handler = msg.on_yes;
-                                       yes_no_message = true;
-                                   },
-                                   [&](const show_save_preset_window& msg) { save_window = true; }};
+    const auto visitor = overloads{
+        [&](const show_ok_message& msg)
+        {
+            message_title = msg.title;
+            message_message = msg.message;
+            ok_message = true;
+        },
+        [&](const show_yes_no_message& msg)
+        {
+            message_title = msg.title;
+            message_message = msg.message;
+            yes_no_handler = msg.on_yes;
+            yes_no_message = true;
+        },
+        [&](const show_save_preset_window& msg)
+        {
+            save_window = true;
+        }
+    };
 
     while (!input.ui_commands.empty()) {
         const ui_command& cmd = input.ui_commands.front();

test/bitmap.cpp

@@ -0,0 +1,74 @@
+// ReSharper disable CppLocalVariableMayBeConst
+#include "puzzle.hpp"
+
+#include <random>
+#include <catch2/catch_test_macros.hpp>
+
+TEST_CASE("bitmap_is_full all bits set", "[puzzle][board]")
+{
+    puzzle p1(5, 5);
+    puzzle p2(3, 4);
+    puzzle p3(5, 4);
+    puzzle p4(3, 7);
+    uint64_t bitmap = -1;
+
+    REQUIRE(p1.bitmap_is_full(bitmap));
+    REQUIRE(p2.bitmap_is_full(bitmap));
+    REQUIRE(p3.bitmap_is_full(bitmap));
+    REQUIRE(p4.bitmap_is_full(bitmap));
+}
+
+TEST_CASE("bitmap_is_full no bits set", "[puzzle][board]")
+{
+    puzzle p1(5, 5);
+    puzzle p2(3, 4);
+    puzzle p3(5, 4);
+    puzzle p4(3, 7);
+    uint64_t bitmap = 0;
+
+    REQUIRE_FALSE(p1.bitmap_is_full(bitmap));
+    REQUIRE_FALSE(p2.bitmap_is_full(bitmap));
+    REQUIRE_FALSE(p3.bitmap_is_full(bitmap));
+    REQUIRE_FALSE(p4.bitmap_is_full(bitmap));
+}
+
+TEST_CASE("bitmap_is_full necessary bits set", "[puzzle][board]")
+{
+    puzzle p1(5, 5);
+    puzzle p2(3, 4);
+    puzzle p3(5, 4);
+    puzzle p4(3, 7);
+
+    uint64_t bitmap1 = (1ull << 25) - 1; // 5 * 5
+    uint64_t bitmap2 = (1ull << 12) - 1;  // 3 * 4
+    uint64_t bitmap3 = (1ull << 20) - 1; // 5 * 4
+    uint64_t bitmap4 = (1ull << 21) - 1; // 3 * 7
+
+    REQUIRE(p1.bitmap_is_full(bitmap1));
+    REQUIRE(p2.bitmap_is_full(bitmap2));
+    REQUIRE(p3.bitmap_is_full(bitmap3));
+    REQUIRE(p4.bitmap_is_full(bitmap4));
+}
+
+TEST_CASE("bitmap_is_full necessary bits not set", "[puzzle][board]")
+{
+    puzzle p1(5, 5);
+    puzzle p2(3, 4);
+    puzzle p3(5, 4);
+    puzzle p4(3, 7);
+
+    uint64_t bitmap1 = (1ull << 25) - 1; // 5 * 5
+    uint64_t bitmap2 = (1ull << 12) - 1;  // 3 * 4
+    uint64_t bitmap3 = (1ull << 20) - 1; // 5 * 4
+    uint64_t bitmap4 = (1ull << 21) - 1; // 3 * 7
+
+    bitmap1 &= ~(1ull << 12);
+    bitmap2 &= ~(1ull << 6);
+    bitmap3 &= ~(1ull << 8);
+    bitmap4 &= ~(1ull << 18);
+
+    REQUIRE_FALSE(p1.bitmap_is_full(bitmap1));
+    REQUIRE_FALSE(p2.bitmap_is_full(bitmap2));
+    REQUIRE_FALSE(p3.bitmap_is_full(bitmap3));
+    REQUIRE_FALSE(p4.bitmap_is_full(bitmap4));
+}

test/bitmap_find_first_empty.cpp

@@ -0,0 +1,266 @@
+// ReSharper disable CppLocalVariableMayBeConst
+#include "puzzle.hpp"
+
+#include <random>
+#include <catch2/catch_test_macros.hpp>
+#include <catch2/generators/catch_generators.hpp>
+
+static auto board_mask(const int w, const int h) -> uint64_t
+{
+    const int cells = w * h;
+    if (cells == 64) {
+        return ~0ULL;
+    }
+    return (1ULL << cells) - 1ULL;
+}
+
+TEST_CASE("Empty board returns (0,0)", "[puzzle][board]")
+{
+    puzzle p(5, 5);
+
+    int x = -1, y = -1;
+    REQUIRE(p.bitmap_find_first_empty(0ULL, x, y));
+
+    REQUIRE(x == 0);
+    REQUIRE(y == 0);
+}
+
+TEST_CASE("Full board detection respects width*height only", "[puzzle][board]")
+{
+    auto [w, h] = GENERATE(std::tuple{3, 3}, std::tuple{4, 4}, std::tuple{5, 6}, std::tuple{8, 8});
+
+    puzzle p(w, h);
+
+    uint64_t mask = board_mask(w, h);
+
+    int x = -1, y = -1;
+
+    REQUIRE_FALSE(p.bitmap_find_first_empty(mask, x, y));
+
+    // Bits outside board should not affect fullness
+    REQUIRE_FALSE(p.bitmap_find_first_empty(mask | (~mask), x, y));
+}
+
+TEST_CASE("Single empty cell at various positions", "[puzzle][board]")
+{
+    auto [w, h] = GENERATE(std::tuple{3, 3}, std::tuple{4, 4}, std::tuple{5, 5}, std::tuple{8, 8});
+
+    puzzle p(w, h);
+
+    int cells = w * h;
+
+    auto empty_index = GENERATE_COPY(values<int>({ 0, cells / 2, cells - 1}));
+
+    uint64_t bitmap = board_mask(w, h);
+    bitmap &= ~(1ULL << empty_index);
+
+    int x = -1, y = -1;
+    REQUIRE(p.bitmap_find_first_empty(bitmap, x, y));
+
+    REQUIRE(x == empty_index % w);
+    REQUIRE(y == empty_index / w);
+}
+
+TEST_CASE("Bits outside board are ignored", "[puzzle][board]")
+{
+    puzzle p(3, 3); // 9 cells
+
+    uint64_t mask = board_mask(3, 3);
+
+    // Board is full
+    uint64_t bitmap = mask;
+
+    // Set extra bits outside 9 cells
+    bitmap |= (1ULL << 20);
+    bitmap |= (1ULL << 63);
+
+    int x = -1, y = -1;
+    REQUIRE_FALSE(p.bitmap_find_first_empty(bitmap, x, y));
+}
+
+TEST_CASE("First empty found in forward search branch", "[puzzle][branch]")
+{
+    puzzle p(4, 4); // 16 cells
+
+    // Only MSB (within board) set
+    uint64_t bitmap = (1ULL << 15);
+
+    int x = -1, y = -1;
+    REQUIRE(p.bitmap_find_first_empty(bitmap, x, y));
+
+    REQUIRE(x == 0);
+    REQUIRE(y == 0);
+}
+
+TEST_CASE("Backward search branch finds gap before MSB cluster", "[puzzle][branch]")
+{
+    puzzle p(4, 4); // 16 cells
+
+    // Set bits 15,14,13 but leave 12 empty
+    uint64_t bitmap = (1ULL << 15) | (1ULL << 14) | (1ULL << 13);
+
+    int x = -1, y = -1;
+    REQUIRE(p.bitmap_find_first_empty(bitmap, x, y));
+
+    REQUIRE(x == 0);
+    REQUIRE(y == 0);
+}
+
+TEST_CASE("Rectangular board coordinate mapping", "[puzzle][rect]")
+{
+    puzzle p(5, 3); // 15 cells
+
+    int empty_index = 11;
+
+    uint64_t bitmap = board_mask(5, 3);
+    bitmap &= ~(1ULL << empty_index);
+
+    int x = -1, y = -1;
+    REQUIRE(p.bitmap_find_first_empty(bitmap, x, y));
+
+    REQUIRE(x == empty_index % 5);
+    REQUIRE(y == empty_index / 5);
+}
+
+TEST_CASE("Non-64-sized board near limit", "[puzzle][limit]")
+{
+    puzzle p(7, 8); // 56 cells
+
+    uint64_t mask = board_mask(7, 8);
+
+    // Full board should return false
+    int x = -1, y = -1;
+    REQUIRE_FALSE(p.bitmap_find_first_empty(mask, x, y));
+
+    // Clear highest valid cell
+    int empty_index = 55;
+    mask &= ~(1ULL << empty_index);
+
+    REQUIRE(p.bitmap_find_first_empty(mask, x, y));
+    REQUIRE(x == empty_index % 7);
+    REQUIRE(y == empty_index / 7);
+}
+
+// --- Oracle: find first zero bit inside board ---
+static auto oracle_find_first_empty(uint64_t bitmap, int w, int h, int& x, int& y) -> bool
+{
+    int cells = w * h;
+
+    for (int i = 0; i < cells; ++i) {
+        if ((bitmap & (1ULL << i)) == 0) {
+            x = i % w;
+            y = i / w;
+            return true;
+        }
+    }
+
+    return false;
+}
+
+TEST_CASE("Oracle validation across board sizes 3x3 to 8x8", "[puzzle][oracle]")
+{
+    auto [w, h] = GENERATE(std::tuple{3, 3}, std::tuple{4, 4}, std::tuple{5, 5}, std::tuple{6, 6}, std::tuple{7, 7},
+                           std::tuple{8, 8}, std::tuple{3, 5}, std::tuple{5, 3}, std::tuple{7, 8}, std::tuple{8, 7});
+
+    puzzle p(w, h);
+
+    uint64_t mask = board_mask(w, h);
+
+    std::mt19937_64 rng(12345);
+    std::uniform_int_distribution<uint64_t> dist(0, UINT64_MAX);
+
+    for (int iteration = 0; iteration < 200; ++iteration) {
+        uint64_t bitmap = dist(rng);
+
+        int ox = -1, oy = -1;
+        bool oracle_result = oracle_find_first_empty(bitmap, w, h, ox, oy);
+
+        int x = -1, y = -1;
+        bool result = p.bitmap_find_first_empty(bitmap, x, y);
+
+        REQUIRE(result == oracle_result);
+
+        if (result) {
+            REQUIRE(x == ox);
+            REQUIRE(y == oy);
+        }
+    }
+}
+
+TEST_CASE("Bits set outside board only behaves as empty board", "[puzzle][outside]")
+{
+    puzzle p(3, 3); // 9 cells
+
+    uint64_t bitmap = (1ULL << 40) | (1ULL << 63);
+
+    int x = -1, y = -1;
+    REQUIRE(p.bitmap_find_first_empty(bitmap, x, y));
+
+    REQUIRE(x == 0);
+    REQUIRE(y == 0);
+}
+
+TEST_CASE("Last valid cell empty stresses upper bound", "[puzzle][boundary]")
+{
+    auto [w, h] = GENERATE(std::tuple{4, 4}, std::tuple{5, 6}, std::tuple{7, 8}, std::tuple{8, 8});
+
+    puzzle p(w, h);
+
+    int cells = w * h;
+    uint64_t bitmap = board_mask(w, h);
+
+    // Clear last valid bit
+    bitmap &= ~(1ULL << (cells - 1));
+
+    int x = -1, y = -1;
+    REQUIRE(p.bitmap_find_first_empty(bitmap, x, y));
+
+    REQUIRE(x == (cells - 1) % w);
+    REQUIRE(y == (cells - 1) / w);
+}
+
+TEST_CASE("Board sizes between 33 and 63 cells", "[puzzle][midrange]")
+{
+    auto [w, h] = GENERATE(std::tuple{6, 6}, // 36
+                           std::tuple{7, 6}, // 42
+                           std::tuple{7, 7}, // 49
+                           std::tuple{8, 7}, // 56
+                           std::tuple{7, 8}  // 56
+    );
+
+    puzzle p(w, h);
+
+    int cells = w * h;
+
+    for (int index : {31, 32, cells - 2}) {
+        if (index >= cells) continue;
+
+        uint64_t bitmap = board_mask(w, h);
+        bitmap &= ~(1ULL << index);
+
+        int x = -1, y = -1;
+        REQUIRE(p.bitmap_find_first_empty(bitmap, x, y));
+
+        REQUIRE(x == index % w);
+        REQUIRE(y == index / w);
+    }
+}
+
+TEST_CASE("Multiple holes choose lowest index", "[puzzle][multiple]")
+{
+    puzzle p(5, 5);
+
+    uint64_t bitmap = board_mask(5, 5);
+
+    // Clear several positions
+    bitmap &= ~(1ULL << 3);
+    bitmap &= ~(1ULL << 7);
+    bitmap &= ~(1ULL << 12);
+
+    int x = -1, y = -1;
+    REQUIRE(p.bitmap_find_first_empty(bitmap, x, y));
+
+    // Oracle expectation: index 3
+    REQUIRE(x == 3 % 5);
+    REQUIRE(y == 3 / 5);
+}

test/bits.cpp

@@ -0,0 +1,267 @@
+#include <catch2/catch_test_macros.hpp>
+#include <catch2/catch_template_test_macros.hpp>
+#include <cstdint>
+
+#include "bits.hpp"
+
+// =============================================================================
+// Catch2
+// =============================================================================
+//
+// 1. TEST_CASE(name, tags)
+//    The basic unit of testing in Catch2. Each TEST_CASE is an independent test
+//    function. The first argument is a descriptive name (must be unique), and
+//    the second is a string of tags in square brackets (e.g. "[set_bits]")
+//    used to filter and group tests when running.
+//
+// 2. SECTION(name)
+//    Sections allow multiple subtests within a single TEST_CASE. Each SECTION
+//    runs the TEST_CASE from the top, so any setup code before the SECTION is
+//    re-executed fresh for every section. This gives each section an isolated
+//    starting state without needing separate TEST_CASEs or explicit teardown.
+//    Sections can also be nested.
+//
+// 3. REQUIRE(expression)
+//    The primary assertion macro. If the expression evaluates to false, the
+//    test fails immediately and Catch2 reports the actual values of both sides
+//    of the comparison (e.g. "0xF5 == 0xF0" on failure). There is also
+//    CHECK(), which records a failure but continues executing the rest of the
+//    test; REQUIRE() aborts the current test on failure.
+//
+// 4. TEMPLATE_TEST_CASE(name, tags, Type1, Type2, ...)
+//    A parameterised test that is instantiated once for each type listed.
+//    Inside the test body, the alias `TestType` refers to the current type.
+//    This avoids duplicating identical logic for uint8_t, uint16_t, uint32_t,
+//    and uint64_t. Catch2 automatically appends the type name to the test name
+//    in the output so you can see which instantiation failed.
+//
+// 5. Tags (e.g. "[create_mask]", "[round-trip]")
+//    Tags let you selectively run subsets of tests from the command line.
+//    For example:
+//      ./tests "[set_bits]"        -- runs only tests tagged [set_bits]
+//      ./tests "~[round-trip]"     -- runs everything except [round-trip]
+//      ./tests "[get_bits],[set_bits]" -- runs tests matching either tag
+//
+// =============================================================================
+
+// ---------------------------------------------------------------------------
+// create_mask
+// ---------------------------------------------------------------------------
+
+TEMPLATE_TEST_CASE("create_mask produces correct masks", "[create_mask]",
+                   uint8_t, uint16_t, uint32_t, uint64_t)
+{
+    SECTION("single bit mask at bit 0") {
+        auto m = create_mask<TestType>(0, 0);
+        REQUIRE(m == TestType{0b1});
+    }
+
+    SECTION("single bit mask at bit 3") {
+        auto m = create_mask<TestType>(3, 3);
+        REQUIRE(m == TestType{0b1000});
+    }
+
+    SECTION("mask spanning bits 0..7 gives 0xFF") {
+        auto m = create_mask<TestType>(0, 7);
+        REQUIRE(m == TestType{0xFF});
+    }
+
+    SECTION("mask spanning bits 4..7") {
+        auto m = create_mask<TestType>(4, 7);
+        REQUIRE(m == TestType{0xF0});
+    }
+
+    SECTION("full-width mask returns all ones") {
+        constexpr uint8_t last = sizeof(TestType) * 8 - 1;
+        auto m = create_mask<TestType>(0, last);
+        REQUIRE(m == static_cast<TestType>(~TestType{0}));
+    }
+}
+
+TEST_CASE("create_mask 32-bit specific cases", "[create_mask]") {
+    REQUIRE(create_mask<uint32_t>(0, 15)  == 0x0000FFFF);
+    REQUIRE(create_mask<uint32_t>(0, 31)  == 0xFFFFFFFF);
+    REQUIRE(create_mask<uint32_t>(16, 31) == 0xFFFF0000);
+}
+
+// ---------------------------------------------------------------------------
+// clear_bits
+// ---------------------------------------------------------------------------
+
+TEMPLATE_TEST_CASE("clear_bits zeroes the specified range", "[clear_bits]",
+                   uint8_t, uint16_t, uint32_t, uint64_t)
+{
+    SECTION("clear all bits") {
+        TestType val = static_cast<TestType>(~TestType{0});
+        constexpr uint8_t last = sizeof(TestType) * 8 - 1;
+        clear_bits(val, 0, last);
+        REQUIRE(val == TestType{0});
+    }
+
+    SECTION("clear lower nibble") {
+        TestType val = static_cast<TestType>(0xFF);
+        clear_bits(val, 0, 3);
+        REQUIRE(val == static_cast<TestType>(0xF0));
+    }
+
+    SECTION("clear upper nibble") {
+        TestType val = static_cast<TestType>(0xFF);
+        clear_bits(val, 4, 7);
+        REQUIRE(val == static_cast<TestType>(0x0F));
+    }
+
+    SECTION("clear single bit") {
+        TestType val = static_cast<TestType>(0xFF);
+        clear_bits(val, 0, 0);
+        REQUIRE(val == static_cast<TestType>(0xFE));
+    }
+
+    SECTION("clearing already-zero bits is a no-op") {
+        TestType val = TestType{0};
+        clear_bits(val, 0, 3);
+        REQUIRE(val == TestType{0});
+    }
+}
+
+// ---------------------------------------------------------------------------
+// set_bits
+// ---------------------------------------------------------------------------
+
+TEMPLATE_TEST_CASE("set_bits writes value into the specified range", "[set_bits]",
+                   uint8_t, uint16_t, uint32_t, uint64_t)
+{
+    SECTION("set lower nibble on zero") {
+        TestType val = TestType{0};
+        set_bits(val, uint8_t{0}, uint8_t{3}, static_cast<TestType>(0xA));
+        REQUIRE(val == static_cast<TestType>(0x0A));
+    }
+
+    SECTION("set upper nibble on zero") {
+        TestType val = TestType{0};
+        set_bits(val, uint8_t{4}, uint8_t{7}, static_cast<TestType>(0xB));
+        REQUIRE(val == static_cast<TestType>(0xB0));
+    }
+
+    SECTION("set_bits replaces existing bits") {
+        TestType val = static_cast<TestType>(0xFF);
+        set_bits(val, uint8_t{0}, uint8_t{3}, static_cast<TestType>(0x5));
+        REQUIRE(val == static_cast<TestType>(0xF5));
+    }
+
+    SECTION("set single bit to 1") {
+        TestType val = TestType{0};
+        set_bits(val, uint8_t{3}, uint8_t{3}, static_cast<TestType>(1));
+        REQUIRE(val == static_cast<TestType>(0x08));
+    }
+
+    SECTION("set single bit to 0") {
+        TestType val = static_cast<TestType>(0xFF);
+        set_bits(val, uint8_t{3}, uint8_t{3}, static_cast<TestType>(0));
+        REQUIRE(val == static_cast<TestType>(0xF7));
+    }
+
+    SECTION("setting value 0 clears the range") {
+        TestType val = static_cast<TestType>(0xFF);
+        set_bits(val, uint8_t{0}, uint8_t{7}, static_cast<TestType>(0));
+        REQUIRE(val == TestType{0});
+    }
+}
+
+TEST_CASE("set_bits with different value type (U != T)", "[set_bits]") {
+    uint32_t val = 0;
+    constexpr uint8_t small_val = 0x3F;
+    set_bits(val, uint8_t{8}, uint8_t{13}, small_val);
+    REQUIRE(val == (uint32_t{0x3F} << 8));
+}
+
+TEST_CASE("set_bits preserves surrounding bits in 32-bit", "[set_bits]") {
+    uint32_t val = 0xDEADBEEF;
+    set_bits(val, uint8_t{8}, uint8_t{15}, uint32_t{0x42});
+    REQUIRE(val == 0xDEAD42EF);
+}
+
+// ---------------------------------------------------------------------------
+// get_bits
+// ---------------------------------------------------------------------------
+
+TEMPLATE_TEST_CASE("get_bits extracts the specified range", "[get_bits]",
+                   uint8_t, uint16_t, uint32_t, uint64_t)
+{
+    SECTION("get lower nibble") {
+        TestType val = static_cast<TestType>(0xAB);
+        auto result = get_bits(val, uint8_t{0}, uint8_t{3});
+        REQUIRE(result == TestType{0xB});
+    }
+
+    SECTION("get upper nibble") {
+        TestType val = static_cast<TestType>(0xAB);
+        auto result = get_bits(val, uint8_t{4}, uint8_t{7});
+        REQUIRE(result == TestType{0xA});
+    }
+
+    SECTION("get single bit that is set") {
+        TestType val = static_cast<TestType>(0x08);
+        auto result = get_bits(val, uint8_t{3}, uint8_t{3});
+        REQUIRE(result == TestType{1});
+    }
+
+    SECTION("get single bit that is clear") {
+        TestType val = static_cast<TestType>(0xF7);
+        auto result = get_bits(val, uint8_t{3}, uint8_t{3});
+        REQUIRE(result == TestType{0});
+    }
+
+    SECTION("get all bits") {
+        TestType val = static_cast<TestType>(~TestType{0});
+        constexpr uint8_t last = sizeof(TestType) * 8 - 1;
+        auto result = get_bits(val, uint8_t{0}, last);
+        REQUIRE(result == val);
+    }
+
+    SECTION("get from zero returns zero") {
+        TestType val = TestType{0};
+        auto result = get_bits(val, uint8_t{0}, uint8_t{7});
+        REQUIRE(result == TestType{0});
+    }
+}
+
+TEST_CASE("get_bits 32-bit specific extractions", "[get_bits]") {
+    constexpr uint32_t val = 0xDEADBEEF;
+
+    REQUIRE(get_bits(val, uint8_t{0},  uint8_t{7})  == 0xEF);
+    REQUIRE(get_bits(val, uint8_t{8},  uint8_t{15}) == 0xBE);
+    REQUIRE(get_bits(val, uint8_t{16}, uint8_t{23}) == 0xAD);
+    REQUIRE(get_bits(val, uint8_t{24}, uint8_t{31}) == 0xDE);
+}
+
+// ---------------------------------------------------------------------------
+// Round-trip: set then get
+// ---------------------------------------------------------------------------
+
+TEST_CASE("set_bits then get_bits round-trips correctly", "[round-trip]") {
+    uint32_t reg = 0;
+
+    set_bits(reg, uint8_t{4}, uint8_t{11}, uint32_t{0xAB});
+    REQUIRE(get_bits(reg, uint8_t{4}, uint8_t{11}) == 0xAB);
+
+    REQUIRE(get_bits(reg, uint8_t{0},  uint8_t{3})  == 0x0);
+    REQUIRE(get_bits(reg, uint8_t{12}, uint8_t{31}) == 0x0);
+}
+
+TEST_CASE("multiple set_bits on different ranges", "[round-trip]") {
+    uint32_t reg = 0;
+
+    set_bits(reg, uint8_t{0},  uint8_t{7},  uint32_t{0x01});
+    set_bits(reg, uint8_t{8},  uint8_t{15}, uint32_t{0x02});
+    set_bits(reg, uint8_t{16}, uint8_t{23}, uint32_t{0x03});
+    set_bits(reg, uint8_t{24}, uint8_t{31}, uint32_t{0x04});
+
+    REQUIRE(reg == 0x04030201);
+}
+
+TEST_CASE("64-bit round-trip", "[round-trip]") {
+    uint64_t reg = 0;
+    set_bits(reg, uint8_t{32}, uint8_t{63}, uint64_t{0xCAFEBABE});
+    REQUIRE(get_bits(reg, uint8_t{32}, uint8_t{63}) == uint64_t{0xCAFEBABE});
+    REQUIRE(get_bits(reg, uint8_t{0},  uint8_t{31}) == uint64_t{0});
+}