Add OpenMP support

CMakeLists.txt

@@ -2,6 +2,8 @@ cmake_minimum_required(VERSION 3.20)
 
 project(sand)
 
+find_package(OpenMP REQUIRED)
+
 find_package(PkgConfig REQUIRED)
 pkg_check_modules(LUAJIT REQUIRED luajit)
 
@@ -14,7 +16,7 @@ target_include_directories(sand
   PUBLIC include
   PRIVATE ${LUAJIT_INCLUDE_DIRS}
 )
-target_link_libraries(sand PRIVATE ${LUAJIT_LIBRARIES})
+target_link_libraries(sand PRIVATE ${LUAJIT_LIBRARIES} OpenMP::OpenMP_CXX)
 set_target_properties(sand PROPERTIES
   CXX_STANDARD 20
   CXX_STANDARD_REQUIRED ON

src/main.cpp

@@ -7,6 +7,12 @@
 #include <SDL3/SDL_video.h>
 #include <sand.h>
 
+constexpr uint16_t WIDTH = 128;
+constexpr uint16_t HEIGHT = 128;
+
+constexpr int WINDOW_WIDTH = 1024;
+constexpr int WINDOW_HEIGHT = 1024;
+
 uint8_t lookup[][3]
   = { { 0, 0, 0 }, { 50, 0, 0 }, { 100, 100, 100 }, { 200, 200, 200 } };
 
@@ -23,14 +29,15 @@ int main(int argc, char** argv)
     .top_middle([](const auto& t) { return t == "stone1"; });
 
   rb.add_rule("stone1", "air")
-    .top_middle([](const auto& t) { return t == "air" || t == "offgrid"; })
     .bottom_middle([](const auto& t) { return t == "air"; });
 
-  auto s = rb.build(64, 64, "air");
+  auto s = rb.build(WIDTH, HEIGHT, "air");
 
   SDL_Init(SDL_INIT_VIDEO);
-  SDL_Window* window = SDL_CreateWindow("sand test", 512, 512, 0);
+  SDL_Window* window
-  SDL_Surface* surface = SDL_CreateSurface(64, 64, SDL_PIXELFORMAT_RGB24);
+    = SDL_CreateWindow("sand test", WINDOW_WIDTH, WINDOW_HEIGHT, 0);
+  SDL_Surface* surface
+    = SDL_CreateSurface(WIDTH, HEIGHT, SDL_PIXELFORMAT_RGB24);
 
   int prev_ms = SDL_GetTicks();
   float time = 0;
@@ -55,16 +62,22 @@ int main(int argc, char** argv)
     int state;
     if (time > 0.008 && (state = SDL_GetMouseState(&x, &y)))
     {
-      time = 0;
+      int ix = x * WIDTH / WINDOW_WIDTH;
-      s.set(x * 64 / 512, y * 64 / 512, state == 1 ? "stone1" : "stone2");
+      int iy = y * HEIGHT / WINDOW_HEIGHT;
+
+      if (ix >= 0 && ix < WIDTH && iy >= 0 && iy < HEIGHT)
+      {
+        time = 0;
+        s.set(ix, iy, state == 1 ? "stone1" : "stone2");
+      }
     }
 
     SDL_LockSurface(surface);
-    for (int i = 0; i < 64; i++)
+    for (int i = 0; i < WIDTH; i++)
     {
-      for (int j = 0; j < 64; j++)
+      for (int j = 0; j < HEIGHT; j++)
       {
-        memcpy(&static_cast<uint8_t*>(surface->pixels)[3 * (i + j * 64)],
+        memcpy(&static_cast<uint8_t*>(surface->pixels)[3 * (i + j * WIDTH)],
                lookup[s.get(i, j)], 3);
       }
     }
@@ -76,6 +89,7 @@ int main(int argc, char** argv)
     s.update();
   }
 
+  SDL_DestroySurface(surface);
   SDL_DestroyWindow(window);
   SDL_Quit();
 

src/sand.cpp

@@ -267,77 +267,79 @@ uint16_t Sand::get(uint16_t x, uint16_t y)
 
 void Sand::update()
 {
-  for (uint32_t i = 0; i < width; i++)
+  int n;
+#pragma omp parallel for private(n)
+  for (n = 0; n < width * height; n++)
   {
-    for (uint32_t j = 0; j < height; j++)
+    int i = n % width;
+    int j = n / height;
+
+    auto& current_rule = rules[get(i, j)];
+
+    unsigned __int128 neighbors = 0;
+    if (i > 0 && j > 0)
     {
-      auto& current_rule = rules[get(i, j)];
+      neighbors += get(i - 1, j - 1);
-
-      unsigned __int128 neighbors = 0;
-      if (i > 0 && j > 0)
-      {
-        neighbors += get(i - 1, j - 1);
-      }
-      neighbors <<= 16;
-
-      if (j > 0)
-      {
-        neighbors += get(i, j - 1);
-      }
-      neighbors <<= 16;
-
-      if (i < width - 1 && j > 0)
-      {
-        neighbors += get(i + 1, j);
-      }
-      neighbors <<= 16;
-
-      if (i > 0)
-      {
-        neighbors += get(i - 1, j);
-      }
-      neighbors <<= 16;
-
-      if (i < width - 1)
-      {
-        neighbors += get(i + 1, j);
-      }
-      neighbors <<= 16;
-
-      if (i > 0 && j < height - 1)
-      {
-        neighbors += get(i - 1, j + 1);
-      }
-      neighbors <<= 16;
-
-      if (j < height - 1)
-      {
-        neighbors += get(i, j + 1);
-      }
-      neighbors <<= 16;
-
-      if (i < width - 1 && j < height - 1)
-      {
-        neighbors += get(i + 1, j + 1);
-      }
-
-      uint16_t result = default_conversions[get(i, j)];
-      if (!current_rule.empty())
-      {
-        std::pair<unsigned __int128, uint16_t> element = { neighbors, 0 };
-        bool (*fn)(decltype(element), decltype(element))
-          = [](decltype(element) a, decltype(element) b) -> bool
-        { return a.first > b.first; };
-        const auto& found = binary_search(element, current_rule.data(), 0,
-                                          current_rule.size(), fn);
-        if (found.first == neighbors)
-        {
-          result = found.second;
-        }
-      }
-
-      elements[i + j * width + !iter * width * height] = result;
     }
+    neighbors <<= 16;
+
+    if (j > 0)
+    {
+      neighbors += get(i, j - 1);
+    }
+    neighbors <<= 16;
+
+    if (i < width - 1 && j > 0)
+    {
+      neighbors += get(i + 1, j);
+    }
+    neighbors <<= 16;
+
+    if (i > 0)
+    {
+      neighbors += get(i - 1, j);
+    }
+    neighbors <<= 16;
+
+    if (i < width - 1)
+    {
+      neighbors += get(i + 1, j);
+    }
+    neighbors <<= 16;
+
+    if (i > 0 && j < height - 1)
+    {
+      neighbors += get(i - 1, j + 1);
+    }
+    neighbors <<= 16;
+
+    if (j < height - 1)
+    {
+      neighbors += get(i, j + 1);
+    }
+    neighbors <<= 16;
+
+    if (i < width - 1 && j < height - 1)
+    {
+      neighbors += get(i + 1, j + 1);
+    }
+
+    uint16_t result = default_conversions[get(i, j)];
+    if (!current_rule.empty())
+    {
+      std::pair<unsigned __int128, uint16_t> element = { neighbors, 0 };
+      bool (*fn)(decltype(element), decltype(element))
+        = [](decltype(element) a, decltype(element) b) -> bool
+      { return a.first > b.first; };
+      const auto& found = binary_search(element, current_rule.data(), 0,
+                                        current_rule.size(), fn);
+      if (found.first == neighbors)
+      {
+        result = found.second;
+      }
+    }
+
+    elements[i + j * width + !iter * width * height] = result;
   }
 
   iter = !iter;