From: hgn <hgodden00@gmail.com>
Date: Fri, 8 Dec 2023 18:35:10 +0000 (+0000)
Subject: performance wins!
X-Git-Url: https://harrygodden.com/git/?p=carveJwlIkooP6JGAAIwe30JlM.git;a=commitdiff_plain;h=f97f731806ae51a3ebf5813fd304c0234676a10a

performance wins!
---

diff --git a/bvh.h b/bvh.h
index 945d3cb..b04c3de 100644
--- a/bvh.h
+++ b/bvh.h
@@ -55,6 +55,7 @@ struct bh_tree{
 };
 
 struct bh_system{
+   u32 system_type;
    void  (*expand_bound)( void *user, boxf bound, u32 item_index );
    float (*item_centroid)( void *user, u32 item_index, int axis );
    void  (*item_closest)( void *user, u32 item_index, v3f point, v3f closest );
@@ -71,13 +72,20 @@ struct bh_system{
    int   (*cast_ray)( void *user, u32 index, v3f co, v3f dir, ray_hit *hit );
 };
 
+static float scene_bh_centroid( void *user, u32 item_index, int axis );
+static void scene_bh_swap( void *user, u32 ia, u32 ib );
+static void scene_bh_expand_bound( void *user, boxf bound, u32 item_index );
+
 static void bh_update_bounds( bh_tree *bh, u32 inode ){
    bh_node *node = &bh->nodes[ inode ];
 
    box_init_inf( node->bbx );
    for( u32 i=0; i<node->count; i++ ){
       u32 idx = node->start+i;
-      bh->system->expand_bound( bh->user, node->bbx, idx );
+      if( bh->system->system_type == 0x1 )
+         scene_bh_expand_bound( bh->user, node->bbx, idx );
+      else
+         bh->system->expand_bound( bh->user, node->bbx, idx );
    }
 }
 
@@ -96,10 +104,13 @@ static void bh_subdivide( bh_tree *bh, u32 inode ){
 
    float split = node->bbx[0][axis] + extent[axis]*0.5f;
    float avg = 0.0;
-   for( u32 t=0; t<node->count; t++ )
-   {
+   for( u32 t=0; t<node->count; t++ ){
       u32 idx = node->start+t;
-      avg += bh->system->item_centroid( bh->user, idx, axis );
+
+      if( bh->system->system_type == 0x1 )
+         avg += scene_bh_centroid( bh->user, idx, axis );
+      else 
+         avg += bh->system->item_centroid( bh->user, idx, axis );
    }
    avg /= (float)node->count;
    split = avg;
@@ -109,10 +120,20 @@ static void bh_subdivide( bh_tree *bh, u32 inode ){
        j = i + node->count-1;
    
    while( i <= j ){
-      if( bh->system->item_centroid( bh->user, i, axis ) < split )
+      f32 centroid;
+      
+      if( bh->system->system_type == 0x1 )
+         centroid = scene_bh_centroid( bh->user, i, axis );
+      else
+         centroid = bh->system->item_centroid( bh->user, i, axis );
+
+      if( centroid < split )
          i ++;
       else{
-         bh->system->item_swap( bh->user, i, j );
+         if( bh->system->system_type == 0x1 )
+            scene_bh_swap( bh->user, i, j );
+         else
+            bh->system->item_swap( bh->user, i, j );
          j --;
       }
    }
diff --git a/scene.h b/scene.h
index 8aaafe9..ba23ffb 100644
--- a/scene.h
+++ b/scene.h
@@ -389,6 +389,7 @@ static bh_system bh_system_scene =
    .item_closest = scene_bh_closest,
    .item_swap = scene_bh_swap,
    .item_debug = scene_bh_debug,
+   .system_type = 0x1
 };
 
 /*
diff --git a/world_gen.c b/world_gen.c
index 543be6d..3d705ab 100644
--- a/world_gen.c
+++ b/world_gen.c
@@ -126,6 +126,9 @@ static void world_apply_procedural_foliage( world_instance *world,
    /* TODO: Quasirandom? */
    vg_info( "Map area: %f. Max particles: %u\n", area, particles );
 
+
+   u64 t0 = SDL_GetPerformanceCounter();
+#if 0
    for( u32 i=0; i<particles; i++ ){
       v3f pos;
       v3_mul( volume, (v3f){ vg_randf64(), 1000.0f, vg_randf64() }, pos );
@@ -144,8 +147,49 @@ static void world_apply_procedural_foliage( world_instance *world,
          }
       }
    }
+#else
+   
+   const f32 tile_scale = 16.0f;
+   v2i tiles = { volume[0]/tile_scale, volume[2]/tile_scale };
+
+   u32 per_tile = particles/(tiles[0]*tiles[1]);
+
+   for( i32 x=0; x<tiles[0]; x ++ ){
+      for( i32 z=0; z<tiles[1]; z ++ ){
+         for( u32 i=0; i<per_tile; i ++ ){
+            v3f co = { (f32)x+vg_randf64(), 0.0f, (f32)z+vg_randf64() };
+            v3_muls( co, tile_scale, co );
+            co[1] = 1000.0f;
+            v3_add( co, world->scene_geo.bbx[0], co );
+
+            ray_hit hit;
+            hit.dist = INFINITY;
+
+            if( ray_world( world, co, (v3f){0.0f,-1.0f,0.0f}, &hit, 
+                           k_material_flag_ghosts )){
+               struct world_surface *m1 = ray_hit_surface( world, &hit );
+               if((hit.normal[1] > 0.8f) && (m1 == mat) &&
+                  (hit.pos[1] > 0.0f+10.0f)){
+                  world_gen_add_blob( world, scene, &hit );
+                  count ++;
+               }
+            }
+
+         }
+      }
+   }
+
+#endif
+
+
+
+   u64 t1 = SDL_GetPerformanceCounter(),
+       utime_blobs = t1-t0,
+       ufreq = SDL_GetPerformanceFrequency();
+   f64 ftime_blobs = ((f64)utime_blobs / (f64)ufreq)*1000.0;
 
-   vg_info( "%d foliage models added\n", count );
+   vg_info( "%d foliage models added. %f%% (%fms)\n", count, 
+            100.0*((f64)count/(f64)particles), ftime_blobs);
 }
 
 static 
diff --git a/world_load.c b/world_load.c
index 5ab2dfb..4c768bf 100644
--- a/world_load.c
+++ b/world_load.c
@@ -98,16 +98,39 @@ static void world_instance_load_mdl( u32 instance_id, const char *path ){
    world->time += (world->info.timezone/24.0);
 
    /* process resources from pack */
+   u64 t4 = SDL_GetPerformanceCounter();
    world_gen_load_surfaces( world );
+   u64 t5 = SDL_GetPerformanceCounter();
    world_gen_routes_ent_init( world );
    world_gen_entities_init( world );
+   u64 t6 = SDL_GetPerformanceCounter();
    
    /* main bulk */
+   u64 t0 = SDL_GetPerformanceCounter();
    world_gen_generate_meshes( world );
+   u64 t1 = SDL_GetPerformanceCounter();
    world_gen_routes_generate( instance_id );
+   u64 t2 = SDL_GetPerformanceCounter();
    world_gen_compute_light_indices( world );
+   u64 t3 = SDL_GetPerformanceCounter();
    mdl_close( meta );
 
+   u64 utime_mesh = t1-t0,
+       utime_route = t2-t1,
+       utime_indices = t3-t2,
+       utime_tex = t5-t4,
+       utime_ent = t6-t5,
+       ufreq = SDL_GetPerformanceFrequency();
+
+   f64 ftime_mesh = ((f64)utime_mesh / (f64)ufreq)*1000.0,
+       ftime_route = ((f64)utime_route / (f64)ufreq)*1000.0,
+       ftime_ind = ((f64)utime_route / (f64)ufreq)*1000.0,
+       ftime_tex = ((f64)utime_tex / (f64)ufreq)*1000.0,
+       ftime_ent = ((f64)utime_ent / (f64)ufreq)*1000.0;
+
+   vg_info( "wtime:mesh %.2fms route %.2fms ind %.2fms tex %.2fms ent %.2fms\n",
+               ftime_mesh, ftime_route, ftime_ind, ftime_tex, ftime_ent );
+
    /* init player position.
     *   - this is overriden by the save state when(if) it loads */
    ent_spawn *rp = world_find_spawn_by_name( world, "start" );