Update perf counters

build.bat

@@ -6,4 +6,4 @@ rem assets.exe
 rem tracy/TracyClient.cpp -DTRACY_ENABLE
 
 
-clang -O2 main.cpp -Wall -Wno-unused-function -Wno-missing-braces -fno-exceptions -fdiagnostics-absolute-paths -g -I".." -o main.exe -Wl,user32.lib -Wl,optick\lib\x64\release\OptickCore.lib
+clang main.cpp -O2 -Wall -Wno-unused-function -Wno-missing-braces -fno-exceptions -fdiagnostics-absolute-paths -g -I".." -o main.exe -Wl,user32.lib -Wl,optick\lib\x64\release\OptickCore.lib

main.cpp

@@ -288,17 +288,18 @@ F32 edge_function(Vec4 vecp0, Vec4 vecp1, Vec4 p) {
   return result;
 }
 
+// #include "optimization_log.cpp"
 function
 void draw_triangle_nearest(Bitmap* dst, F32 *depth_buffer, Bitmap *src, Vec3 light_direction,
                            Vec4 p0,   Vec4 p1,   Vec4 p2,
                            Vec2 tex0, Vec2 tex1, Vec2 tex2,
                            Vec3 norm0, Vec3 norm1, Vec3 norm2) {
-  // if(os.frame > 10) PROFILE_BEGIN(draw_triangle);
+  PROFILE_SCOPE(draw_triangle);
-  // ZoneScopedN("draw_triangle");
   F32 min_x1 = (F32)(min(p0.x, min(p1.x, p2.x)));
   F32 min_y1 = (F32)(min(p0.y, min(p1.y, p2.y)));
   F32 max_x1 = (F32)(max(p0.x, max(p1.x, p2.x)));
   F32 max_y1 = (F32)(max(p0.y, max(p1.y, p2.y)));
+
   S64 min_x = (S64)max(0.f, floor(min_x1));
   S64 min_y = (S64)max(0.f, floor(min_y1));
   S64 max_x = (S64)min((F32)dst->x, ceil(max_x1));
@@ -328,7 +329,7 @@ void draw_triangle_nearest(Bitmap* dst, F32 *depth_buffer, Bitmap *src, Vec3 lig
     F32 Cx2 = Cy2;
     for (S64 x = min_x; x < max_x; x++) {
       if (Cx0 >= 0 && Cx1 >= 0 && Cx2 >= 0) {
-      // ZoneNamedN(fill, "fill_pixel", true);
+        PROFILE_SCOPE(fill_triangle);
         F32 w1 = Cx1 / area;
         F32 w2 = Cx2 / area;
         F32 w3 = Cx0 / area;
@@ -408,108 +409,15 @@ void draw_triangle_nearest(Bitmap* dst, F32 *depth_buffer, Bitmap *src, Vec3 lig
     Cy1 -= dx21;
     Cy2 -= dx02;
     destination += dst->x;
-  }
-
-     // if(os.frame > 10) PROFILE_END(draw_triangle);
-}
-
-function
-void draw_triangle_bilinear(Bitmap* dst, F32 *depth_buffer, Bitmap *src, Vec3 light_direction,
-                            Vec4 p0,   Vec4 p1,   Vec4 p2,
-                            Vec2 tex0, Vec2 tex1, Vec2 tex2,
-                            Vec3 norm0, Vec3 norm1, Vec3 norm2) {
-  F32 min_x1 = (F32)(min(p0.x, min(p1.x, p2.x)));
-  F32 min_y1 = (F32)(min(p0.y, min(p1.y, p2.y)));
-  F32 max_x1 = (F32)(max(p0.x, max(p1.x, p2.x)));
-  F32 max_y1 = (F32)(max(p0.y, max(p1.y, p2.y)));
-
-  S64 min_x = (S64)clamp_bot(0.f, floor(min_x1));
-  S64 min_y = (S64)clamp_bot(0.f, floor(min_y1));
-  S64 max_x = (S64)clamp_top((F32)dst->x, ceil(max_x1));
-  S64 max_y = (S64)clamp_top((F32)dst->y, ceil(max_y1));
-
-  F32 area = edge_function(p0, p1, p2);
-  for (S64 y = min_y; y < max_y; y++) {
-    for (S64 x = min_x; x < max_x; x++) {
-      F32 edge0 = edge_function(p0, p1, { (F32)x,(F32)y });
-      F32 edge1 = edge_function(p1, p2, { (F32)x,(F32)y });
-      F32 edge2 = edge_function(p2, p0, { (F32)x,(F32)y });
-
-      if (edge0 >= 0 && edge1 >= 0 && edge2 >= 0) {
-        F32 w1 = edge1 / area;
-        F32 w2 = edge2 / area;
-        F32 w3 = edge0 / area;
-        F32 interpolated_w = (1.f / p0.w) * w1 + (1.f / p1.w) * w2 + (1.f / p2.w) * w3;
-
-        // @Note: We could do: interpolated_w = 1.f / interpolated_w to get proper depth
-        // but why waste an instruction, the smaller the depth value the farther the object
-        F32* depth = depth_buffer + (x + y * dst->x);
-        if (*depth < interpolated_w) {
-          *depth = interpolated_w;
-          F32 invw0 = (w1 / p0.w);
-          F32 invw1 = (w2 / p1.w);
-          F32 invw2 = (w3 / p2.w);
-
-          Vec3 norm = (norm0 * invw0 + norm1 * invw1 + norm2 * invw2) / interpolated_w;
-          F32 u = tex0.x * invw0 + tex1.x * invw1 + tex2.x * invw2;
-          F32 v = tex0.y * invw0 + tex1.y * invw1 + tex2.y * invw2;
-          {
-            u /= interpolated_w;
-            v /= interpolated_w;
-            u = u - floor(u);
-            v = v - floor(v);
-            u = u * (src->x - 1);
-            v = v * (src->y - 1);
-          }
-
-          S64 ui = (S64)(u);
-          S64 vi = (S64)(v);
-          F32 udiff = u - (F32)ui;
-          F32 vdiff = v - (F32)vi;
-          // Origin UV (0,0) is in bottom left
-          U32 *pixel = src->pixels + (ui + (src->y - 1ll - vi) * src->x);
-          U32 *dst_pixel = dst->pixels + (x + y * dst->x);
-
-#if 0
-          Vec4 result_color = vec4abgr(*pixel);
-#else
-          Vec4 pixelx1y1 = vec4abgr(*pixel);
-          Vec4 pixelx2y1 = vec4abgr(*(pixel + 1));
-          Vec4 pixelx1y2 = vec4abgr(*(pixel - src->x));
-          Vec4 pixelx2y2 = vec4abgr(*(pixel + 1 - src->x));
-          pixelx1y1 = srgb_to_almost_linear(pixelx1y1);
-          pixelx2y1 = srgb_to_almost_linear(pixelx2y1);
-          pixelx1y2 = srgb_to_almost_linear(pixelx1y2);
-          pixelx2y2 = srgb_to_almost_linear(pixelx2y2);
-          Vec4 blendx1 = lerp(pixelx1y1, pixelx2y1, udiff);
-          Vec4 blendx2 = lerp(pixelx1y2, pixelx2y2, udiff);
-          Vec4 result_color = lerp(blendx1, blendx2, vdiff);
-#endif
-
-          Vec3 light_color = vec3(0.8,0.8,1);
-          constexpr F32 ambient_strength = 0.1f; {
-            Vec3 ambient = ambient_strength * light_color;
-            Vec3 diffuse = clamp_bot(0.f, -dot(norm, light_direction)) * light_color;
-            result_color.rgb *= (ambient+diffuse);
-          }
-
-          Vec4 dst_color = vec4abgr(*dst_pixel);
-          dst_color = srgb_to_almost_linear(dst_color);
-          result_color = premultiplied_alpha(dst_color, result_color);
-          result_color = almost_linear_to_srgb(result_color);
-          U32 color32 = vec4_to_u32abgr(result_color);
-
-          *dst_pixel = color32;
-        }
-      }
-    }
   }
 }
 
 function
 void draw_mesh(Render *r, String scene_name, Obj_Material *materials, Obj_Mesh *mesh, Vec3 *vertices, Vec2 *tex_coords, Vec3 *normals) {
   // ZoneNamedN(m, "draw_all_meshes", true);
+  PROFILE_SCOPE(draw_all_meshes);
   for (int i = 0; i < mesh->indices.len; i++) {
+    PROFILE_SCOPE(draw_set_of_mesh_indices);
     // ZoneNamedN(m, "draw_single_mesh", true);
     Obj_Index *index = mesh->indices.data + i;
     Bitmap *image = &r->img;
@@ -681,14 +589,19 @@ UI_SIGNAL_CALLBACK(scene_callback) {
   scene = (Scene)(((int)scene + 1) % Scene_Count);
 }
 
+FILE *global_file;
 function void
 windows_log(Log_Kind kind, String string, char *file, int line){
-  OutputDebugStringA((char *)string.str);
+  fprintf(global_file, "%s", string.str);
 }
 
 int
 main(int argc, char **argv) {
+  global_file = fopen("perfclocks.txt", "a");
   thread_ctx.log_proc = windows_log;
+  fprintf(global_file, "\n---------------------");
+
+
   os.window_size.x = 1280;
   os.window_size.y = 720;
   os.window_resizable = 1;
@@ -717,7 +630,7 @@ main(int argc, char **argv) {
   B32 ui_mouse_lock = true;
 
   while (os_game_loop()) {
-    // FrameMark;
+    PROFILE_SCOPE(main_loop);
     if (ui_mouse_lock == false) {
       r.camera_yaw.x += os.delta_mouse_pos.x * 0.01f;
       r.camera_yaw.y -= os.delta_mouse_pos.y * 0.01f;
@@ -763,6 +676,7 @@ main(int argc, char **argv) {
     r.transform = mat4_rotation_z(rotation);
     r.transform = r.transform * mat4_rotation_y(rotation);
     for (int i = 0; i < obj->mesh.len; i++) {
+      PROFILE_SCOPE(draw_all_meshes);
       Vec2* tex_coords = (Vec2*)obj->texture_coordinates.data;
       Vec3 *normals = (Vec3 *)obj->normals.data;
       Obj_Mesh *mesh = obj->mesh.data;
@@ -786,14 +700,17 @@ main(int argc, char **argv) {
     frame_data = string_fmt(os.frame_arena, "FPS:%f dt:%f frame:%u camera_pos: %f %f %f camera_yaw: %f %f", os.fps, os.delta_time, os.frame,
       r.camera_pos.x, r.camera_pos.y, r.camera_pos.z, r.camera_yaw.x, r.camera_yaw.y);
 
-    auto *scope = &profile_scopes[ProfileScopeName_draw_triangle];
+
-    if(scope->i != 0){
+    for(int i = 0; i < ProfileScopeName_Count; i++){
+      auto *scope = &profile_scopes[i];
+      if(scope->i == 0) continue;
+
       U64 total = 0;
       for(int i = 0; i < scope->i; i++){
         total += scope->samples[i];
       }
 
-      log_info("\nTotal: %llu Hits: %llu", total, (U64)scope->i);
+      log_info("\n%s :: Total: %llu Hits: %llu, Avg: %llu", profile_scope_names[i], total, (U64)scope->i, total / scope->i);
       scope->i = 0;
     }
   }

optimization_log.cpp

@@ -1,5 +1,5 @@
 function
-void draw_triangle_nearest1(Bitmap* dst, F32 *depth_buffer, Bitmap *src, Vec3 light_direction,
+void draw_triangle_nearest_a(Bitmap* dst, F32 *depth_buffer, Bitmap *src, Vec3 light_direction,
                            Vec4 p0,   Vec4 p1,   Vec4 p2,
                            Vec2 tex0, Vec2 tex1, Vec2 tex2,
                            Vec3 norm0, Vec3 norm1, Vec3 norm2) {
@@ -222,3 +222,97 @@ void draw_triangle_nearest_b(Bitmap* dst, F32 *depth_buffer, Bitmap *src, Vec3 l
 
      // if(os.frame > 10) PROFILE_END(draw_triangle);
 }
+
+
+function
+void draw_triangle_bilinear(Bitmap* dst, F32 *depth_buffer, Bitmap *src, Vec3 light_direction,
+                            Vec4 p0,   Vec4 p1,   Vec4 p2,
+                            Vec2 tex0, Vec2 tex1, Vec2 tex2,
+                            Vec3 norm0, Vec3 norm1, Vec3 norm2) {
+  F32 min_x1 = (F32)(min(p0.x, min(p1.x, p2.x)));
+  F32 min_y1 = (F32)(min(p0.y, min(p1.y, p2.y)));
+  F32 max_x1 = (F32)(max(p0.x, max(p1.x, p2.x)));
+  F32 max_y1 = (F32)(max(p0.y, max(p1.y, p2.y)));
+
+  S64 min_x = (S64)clamp_bot(0.f, floor(min_x1));
+  S64 min_y = (S64)clamp_bot(0.f, floor(min_y1));
+  S64 max_x = (S64)clamp_top((F32)dst->x, ceil(max_x1));
+  S64 max_y = (S64)clamp_top((F32)dst->y, ceil(max_y1));
+
+  F32 area = edge_function(p0, p1, p2);
+  for (S64 y = min_y; y < max_y; y++) {
+    for (S64 x = min_x; x < max_x; x++) {
+      F32 edge0 = edge_function(p0, p1, { (F32)x,(F32)y });
+      F32 edge1 = edge_function(p1, p2, { (F32)x,(F32)y });
+      F32 edge2 = edge_function(p2, p0, { (F32)x,(F32)y });
+
+      if (edge0 >= 0 && edge1 >= 0 && edge2 >= 0) {
+        F32 w1 = edge1 / area;
+        F32 w2 = edge2 / area;
+        F32 w3 = edge0 / area;
+        F32 interpolated_w = (1.f / p0.w) * w1 + (1.f / p1.w) * w2 + (1.f / p2.w) * w3;
+
+        // @Note: We could do: interpolated_w = 1.f / interpolated_w to get proper depth
+        // but why waste an instruction, the smaller the depth value the farther the object
+        F32* depth = depth_buffer + (x + y * dst->x);
+        if (*depth < interpolated_w) {
+          *depth = interpolated_w;
+          F32 invw0 = (w1 / p0.w);
+          F32 invw1 = (w2 / p1.w);
+          F32 invw2 = (w3 / p2.w);
+
+          Vec3 norm = (norm0 * invw0 + norm1 * invw1 + norm2 * invw2) / interpolated_w;
+          F32 u = tex0.x * invw0 + tex1.x * invw1 + tex2.x * invw2;
+          F32 v = tex0.y * invw0 + tex1.y * invw1 + tex2.y * invw2;
+          {
+            u /= interpolated_w;
+            v /= interpolated_w;
+            u = u - floor(u);
+            v = v - floor(v);
+            u = u * (src->x - 1);
+            v = v * (src->y - 1);
+          }
+
+          S64 ui = (S64)(u);
+          S64 vi = (S64)(v);
+          F32 udiff = u - (F32)ui;
+          F32 vdiff = v - (F32)vi;
+          // Origin UV (0,0) is in bottom left
+          U32 *pixel = src->pixels + (ui + (src->y - 1ll - vi) * src->x);
+          U32 *dst_pixel = dst->pixels + (x + y * dst->x);
+
+#if 0
+          Vec4 result_color = vec4abgr(*pixel);
+#else
+          Vec4 pixelx1y1 = vec4abgr(*pixel);
+          Vec4 pixelx2y1 = vec4abgr(*(pixel + 1));
+          Vec4 pixelx1y2 = vec4abgr(*(pixel - src->x));
+          Vec4 pixelx2y2 = vec4abgr(*(pixel + 1 - src->x));
+          pixelx1y1 = srgb_to_almost_linear(pixelx1y1);
+          pixelx2y1 = srgb_to_almost_linear(pixelx2y1);
+          pixelx1y2 = srgb_to_almost_linear(pixelx1y2);
+          pixelx2y2 = srgb_to_almost_linear(pixelx2y2);
+          Vec4 blendx1 = lerp(pixelx1y1, pixelx2y1, udiff);
+          Vec4 blendx2 = lerp(pixelx1y2, pixelx2y2, udiff);
+          Vec4 result_color = lerp(blendx1, blendx2, vdiff);
+#endif
+
+          Vec3 light_color = vec3(0.8,0.8,1);
+          constexpr F32 ambient_strength = 0.1f; {
+            Vec3 ambient = ambient_strength * light_color;
+            Vec3 diffuse = clamp_bot(0.f, -dot(norm, light_direction)) * light_color;
+            result_color.rgb *= (ambient+diffuse);
+          }
+
+          Vec4 dst_color = vec4abgr(*dst_pixel);
+          dst_color = srgb_to_almost_linear(dst_color);
+          result_color = premultiplied_alpha(dst_color, result_color);
+          result_color = almost_linear_to_srgb(result_color);
+          U32 color32 = vec4_to_u32abgr(result_color);
+
+          *dst_pixel = color32;
+        }
+      }
+    }
+  }
+}

profile.cpp

@@ -1,25 +1,48 @@
 enum ProfileScopeName {
   ProfileScopeName_draw_triangle,
+  ProfileScopeName_fill_triangle,
+  ProfileScopeName_draw_all_meshes,
+  ProfileScopeName_draw_mesh,
+  ProfileScopeName_draw_set_of_mesh_indices,
+  ProfileScopeName_main_loop,
   ProfileScopeName_Count,
 };
 
-struct ProfileScope {
+const char *profile_scope_names[] = {
-  U64 samples[5096*16];
+  "draw_triangle",
+  "fill_triangle",
+  "draw_all_meshes",
+  "draw_mesh",
+  "draw_set_of_mesh_indices",
+  "main_loop",
+};
+
+struct ProfileState {
+  U64 samples[5096*32];
   S64 i;
 };
 
-global ProfileScope profile_scopes[ProfileScopeName_Count];
+global ProfileState profile_scopes[ProfileScopeName_Count];
 
-#define PROFILE_BEGIN(name)                                                              \
+force_inline void
-  do {                                                                                   \
+profile_begin(ProfileScopeName name){
-    ProfileScope *__profile_scope = profile_scopes + ProfileScopeName_##name;            \
+  ProfileState *p = profile_scopes + name;
-    __profile_scope->samples[__profile_scope->i] = __rdtsc();                            \
+  p->samples[p->i] = __rdtsc();
-  } while (0)
+}
 
-#define PROFILE_END(name)                                                                \
+force_inline void
-  do {                                                                                   \
+profile_end(ProfileScopeName name){
-    ProfileScope *_profile_scope = profile_scopes + ProfileScopeName_##name;             \
+  ProfileState *p = profile_scopes + name;
-    _profile_scope->samples[_profile_scope->i] =                                         \
+  p->samples[p->i] = __rdtsc() - p->samples[p->i];
-        __rdtsc() - _profile_scope->samples[_profile_scope->i];                          \
+  p->i = (p->i + 1) % buff_cap(p->samples);
-    _profile_scope->i = (_profile_scope->i + 1) % 5096*16;                                  \
+}
-  } while (0)
+
+struct Profile_Scope{
+  ProfileScopeName n;
+  force_inline Profile_Scope(ProfileScopeName name){ profile_begin(name); n=name; }
+  force_inline ~Profile_Scope(){ profile_end(n); }
+};
+
+#define PROFILE_BEGIN(name) profile_begin(ProfileScopeName_##name)
+#define PROFILE_END(name) profile_end(ProfileScopeName_##name)
+#define PROFILE_SCOPE(name) Profile_Scope profile_scope_##__LINE__(ProfileScopeName_##name)