improve rendering performance even more by using instanced rendering

include/config.hpp

@@ -4,6 +4,8 @@
 #include <raylib.h>
 
 #define PRINT_TIMINGS
+// #define BATCHING
+#define INSTANCING // Fastest (so far)
 // #define WEB
 
 // Window

include/renderer.hpp

@@ -49,9 +49,17 @@ private:
   RenderTexture klotski_target;
   RenderTexture menu_target;
 
+  Material vertex_mat;
+
+  // Batching
   float *cube;
   Mesh graph;
-  Material mat;
+
+  // Instancing
+  int transforms_size;
+  Matrix *transforms;
+  Mesh cube_instance;
+  Shader instancing_shader;
 
 public:
   bool mark_solutions;
@@ -60,13 +68,16 @@ public:
 public:
   Renderer()
       : camera(OrbitCamera3D(Vector3(0, 0, 0), CAMERA_DISTANCE)),
-        mark_solutions(false), connect_solutions(false) {
+        mark_solutions(false), connect_solutions(false), cube(nullptr),
+        transforms_size(0), transforms(nullptr) {
     render_target = LoadRenderTexture(GetScreenWidth() / 2.0,
                                       GetScreenHeight() - MENU_HEIGHT);
     klotski_target = LoadRenderTexture(GetScreenWidth() / 2.0,
                                        GetScreenHeight() - MENU_HEIGHT);
     menu_target = LoadRenderTexture(GetScreenWidth(), MENU_HEIGHT);
-    AllocateGraphMesh();
+
+    vertex_mat = LoadMaterialDefault();
+    vertex_mat.maps[MATERIAL_MAP_DIFFUSE].color = VERTEX_COLOR;
   }
 
   Renderer(const Renderer &copy) = delete;
@@ -78,9 +89,23 @@ public:
     UnloadRenderTexture(render_target);
     UnloadRenderTexture(klotski_target);
     UnloadRenderTexture(menu_target);
-    UnloadMesh(graph);
-    UnloadMaterial(mat);
+
+    UnloadMaterial(vertex_mat);
+
+    // Batching
+    if (cube != nullptr) {
       MemFree(cube);
+      UnloadMesh(graph);
+    }
+
+    // Instancing
+    if (transforms != nullptr) {
+      MemFree(transforms);
+      UnloadMesh(cube_instance);
+
+      // I think the shader already gets unloaded with the material?
+      // UnloadShader(instancing_shader);
+    }
   }
 
 public:
@@ -89,10 +114,20 @@ public:
 
   auto UpdateTextureSizes() -> void;
 
-  auto AllocateGraphMesh() -> void;
+#ifdef BATCHING
+  auto AllocateGraphBatching() -> void;
 
-  auto ReallocateGraphMeshIfNecessary(const MassSpringSystem &mass_springs)
+  auto ReallocateGraphBatchingIfNecessary(const MassSpringSystem &mass_springs)
       -> void;
+#endif
+
+#ifdef INSTANCING
+  auto AllocateGraphInstancing(const MassSpringSystem &mass_springs) -> void;
+
+  auto
+  ReallocateGraphInstancingIfNecessary(const MassSpringSystem &mass_springs)
+      -> void;
+#endif
 
   auto DrawMassSprings(const MassSpringSystem &mass_springs,
                        const State &current_state,

instancing_fragment.glsl

@@ -0,0 +1,8 @@
+#version 330
+
+uniform vec4 colDiffuse;
+out vec4 finalColor;
+
+void main() {
+    finalColor = colDiffuse;
+}

instancing_vertex.glsl

@@ -0,0 +1,13 @@
+#version 330
+
+in vec3 vertexPosition;
+in vec2 vertexTexCoord;
+in vec3 vertexNormal;
+in vec4 vertexColor;
+in mat4 instanceTransform;
+
+uniform mat4 mvp;
+
+void main() {
+    gl_Position = mvp * instanceTransform * vec4(vertexPosition, 1.0);
+}

src/main.cpp

@@ -43,8 +43,8 @@ auto main(int argc, char *argv[]) -> int {
 #endif
 
   // RayLib window setup
-  SetTraceLogLevel(LOG_ERROR);
-  SetConfigFlags(FLAG_VSYNC_HINT);
+  // SetTraceLogLevel(LOG_ERROR);
+  // SetConfigFlags(FLAG_VSYNC_HINT);
   SetConfigFlags(FLAG_MSAA_4X_HINT);
   SetConfigFlags(FLAG_WINDOW_RESIZABLE);
   SetConfigFlags(FLAG_WINDOW_ALWAYS_RUN);
@@ -105,7 +105,6 @@ auto main(int argc, char *argv[]) -> int {
 
     renderer.UpdateCamera(mass_springs, state.current_state);
     renderer.UpdateTextureSizes();
-    renderer.ReallocateGraphMeshIfNecessary(mass_springs);
     renderer.DrawMassSprings(mass_springs, state.current_state,
                              state.winning_states);
 

src/renderer.cpp

@@ -114,7 +114,8 @@ auto Renderer::UpdateTextureSizes() -> void {
   menu_target = LoadRenderTexture(width * 2, MENU_HEIGHT);
 }
 
-auto Renderer::AllocateGraphMesh() -> void {
+#ifdef BATCHING
+auto Renderer::AllocateGraphBatching() -> void {
   int vertices = 36;
   int max_masses = 100000;
 
@@ -137,13 +138,10 @@ auto Renderer::AllocateGraphMesh() -> void {
   }
   UnloadMesh(indexed_cube);
 
-  mat = LoadMaterialDefault();
-  mat.maps[MATERIAL_MAP_DIFFUSE].color = VERTEX_COLOR;
-
   std::cout << "Allocated graph mesh." << std::endl;
 }
 
-auto Renderer::ReallocateGraphMeshIfNecessary(
+auto Renderer::ReallocateGraphBatchingIfNecessary(
     const MassSpringSystem &mass_springs) -> void {
   if (graph.vertexCount / 3 > mass_springs.masses.size()) {
     return;
@@ -165,6 +163,35 @@ auto Renderer::ReallocateGraphMeshIfNecessary(
 
   std::cout << "Reallocated graph mesh." << std::endl;
 }
+#endif
+
+#ifdef INSTANCING
+auto Renderer::AllocateGraphInstancing(const MassSpringSystem &mass_springs)
+    -> void {
+  cube_instance = GenMeshCube(VERTEX_SIZE, VERTEX_SIZE, VERTEX_SIZE);
+
+  instancing_shader =
+      LoadShader("instancing_vertex.glsl", "instancing_fragment.glsl");
+  instancing_shader.locs[SHADER_LOC_MATRIX_MVP] =
+      GetShaderLocation(instancing_shader, "mvp");
+  instancing_shader.locs[SHADER_LOC_VECTOR_VIEW] =
+      GetShaderLocation(instancing_shader, "viewPos");
+
+  vertex_mat.shader = instancing_shader;
+
+  transforms = (Matrix *)MemAlloc(mass_springs.masses.size() * sizeof(Matrix));
+  transforms_size = mass_springs.masses.size();
+}
+
+auto Renderer::ReallocateGraphInstancingIfNecessary(
+    const MassSpringSystem &mass_springs) -> void {
+  if (transforms_size != mass_springs.masses.size()) {
+    transforms = (Matrix *)MemRealloc(transforms, mass_springs.masses.size() *
+                                                      sizeof(Matrix));
+    transforms_size = mass_springs.masses.size();
+  }
+}
+#endif
 
 auto Renderer::DrawMassSprings(const MassSpringSystem &mass_springs,
                                const State &current_state,
@@ -186,7 +213,30 @@ auto Renderer::DrawMassSprings(const MassSpringSystem &mass_springs,
   }
   rlEnd();
 
+  // Draw masses (instanced)
+#ifdef INSTANCING
+  if (transforms == nullptr) {
+    AllocateGraphInstancing(mass_springs);
+  }
+  ReallocateGraphInstancingIfNecessary(mass_springs);
+
+  int i = 0;
+  for (const auto &[state, mass] : mass_springs.masses) {
+    transforms[i] =
+        MatrixTranslate(mass.position.x, mass.position.y, mass.position.z);
+    ++i;
+  }
+  DrawMeshInstanced(cube_instance, vertex_mat, transforms,
+                    mass_springs.masses.size());
+#endif
+
   // Draw masses (batched)
+#ifdef BATCHING
+  if (cube == nullptr) {
+    AllocateGraphBatching();
+  }
+  ReallocateGraphBatchingIfNecessary(mass_springs);
+
   int vertices = 36;
   int current_size = mass_springs.masses.size();
   int i = 0;
@@ -208,11 +258,12 @@ auto Renderer::DrawMassSprings(const MassSpringSystem &mass_springs,
   graph.vertexCount = current_size * vertices;
   graph.triangleCount = current_size * 12;
 
-  DrawMesh(graph, mat, MatrixIdentity());
+  DrawMesh(graph, vertex_mat, MatrixIdentity());
 
   // Restore the vertex count
   graph.vertexCount = full_size;
   graph.triangleCount = full_size / 3;
+#endif
 
   // Mark current state
   const Mass &current_mass = mass_springs.GetMass(current_state);