fast sprite render

scripts/render.js

@@ -758,6 +758,7 @@ function insertion_sort(arr, cmp)
 
 function sprites_to_queue(sprites, ysort = false)
 {
+  return render._main.make_sprite_queue(allsprites, prosperon.camera, sprite_pipeline)
   var sprites = allsprites;
   for (var i = 0; i < sprites.length; i++)
     sprites[i].transform.clean();
@@ -769,7 +770,6 @@ function sprites_to_queue(sprites, ysort = false)
   var mesh = render._main.make_sprite_mesh(sprites);
 
   var queue = [];
-  var idx = 0;
   var image;
   var first_index = 0;
   var count = 0;

source/jsffi.c

@@ -1149,7 +1149,6 @@ int js_arrlen(JSContext *js,JSValue v) {
   len = js_getnum_str(js,v,"length");
   return len;
 }
-
 static inline int js_transform_dirty_chain(JSContext *js, JSValue v)
 {
   transform *t = js2transform(js, v);
@@ -1753,6 +1752,64 @@ static const JSCFunctionListEntry js_spline_funcs[] = {
   MIST_FUNC_DEF(spline, bezier, 2)
 };
 
+
+shader_globals camera_globals(JSContext *js, JSValue camera)
+{
+  shader_globals data = {0};
+  HMM_Vec2 size;
+  transform *transform;
+  double fov;
+  double aspect;
+  int ortho;
+  double near;
+  double far;
+  
+  JS_GETPROP(js, size, camera, size, vec2)
+  JS_GETPROP(js, transform, camera, transform, transform)
+  JS_GETPROP(js, fov, camera, fov, number)
+  JS_GETPROP(js, aspect, camera, aspect, number)
+  JS_GETPROP(js, ortho, camera,ortho,bool)
+  JS_GETPROP(js,near,camera,near,number)
+  JS_GETPROP(js,far,camera,far,number)
+
+  HMM_Mat4 proj;
+  HMM_Mat4 view;
+  
+  if (ortho) {
+    proj = HMM_Orthographic_RH_NO(
+      -size.x*0.5, 0.5*size.x,
+      -size.y*0.5, 0.5*size.y,
+      -1.0f, 1.0f
+    );
+    view = HMM_Translate((HMM_Vec3){ -transform->pos.x, -transform->pos.y, 0.0f });    
+  }
+  else {
+    proj = HMM_Perspective_RH_NO(fov, aspect,near,far);
+
+    HMM_Mat4 camera_transform = HMM_Translate(transform->pos);
+    camera_transform = HMM_MulM4(camera_transform, HMM_QToM4(transform->rotation));
+//    camera_transform = HMM_MulM4(camera_transform, HMM_Scale(transform->scale));  // don't bother w/ scale
+    view = HMM_InvGeneralM4(camera_transform);
+  }
+
+  // Update your shader globals
+  data.world_to_projection = HMM_MulM4(proj, view);
+  data.projection_to_world = HMM_InvGeneralM4(data.world_to_projection);
+  data.camera_pos_world = transform->pos;
+  data.viewport_min_z = near;
+  data.viewport_max_z = far;
+  data.render_size = size;
+  data.world_to_view = view;
+  data.view_to_projection = proj;
+  data.camera_dir_world = HMM_NormV3(HMM_QVRot((HMM_Vec3){0,0,-1},transform->rotation));
+  data.viewport_size = (HMM_Vec2){0.5,0.5};
+  data.viewport_offset = (HMM_Vec2){0,0};
+  data.time = SDL_GetTicksNS() / 1000000000.0f;
+  return data;
+}
+
+
+
 JSValue js_vector_dot(JSContext *js, JSValue self, int argc, JSValue *argv) {
   size_t alen, blen;
   float *a = js2floats(js,argv[0], &alen);
@@ -3986,7 +4043,7 @@ int sort_sprite(const sprite *a, const sprite *b)
 
   if (a->quad.vert[0].pos.y != b->quad.vert[0].pos.y) return b->quad.vert[0].pos.y - a->quad.vert[0].pos.y;
 
-//  if (!JS_SameValue(a->js, a->image, b->image)) return JS_VALUE_GET_PTR(a->image) < JS_VALUE_GET_PTR(b->image) ? -1 : 1;
+  if (!JS_SameValue(a->js, a->image, b->image)) return JS_VALUE_GET_PTR(a->image) < JS_VALUE_GET_PTR(b->image) ? -1 : 1;
   
   return 0;
 }
@@ -4014,33 +4071,58 @@ JSC_CCALL(gpu_sort_sprite,
   return number2js(js,0);
 )
 
+inline int sprite_in_view(HMM_Mat3 sprite, HMM_Mat4 camera)
+{
+  int outside = 0;
+  for (int j = 0; j < 4; j++) {
+    HMM_Vec3 corner = HMM_MulM3V3(sprite, (HMM_Vec3){base_quad[j].x, base_quad[j].y, 1.0});    
+    HMM_Vec4 clip = HMM_MulM4V4(camera, (HMM_Vec4){corner.x,corner.y,corner.z,1.0});
+    if (clip.w <= 0.0) {
+      outside++;
+      continue;
+    }
+
+    float nx = clip.x/clip.w;
+    float ny = clip.y/clip.w;
+    float nz = clip.z/clip.w;
+
+    if (nx < -1 || nx > 1) outside++;
+    else if (ny < -1 || ny > 1) outside++;
+    else if (nz < -1 || nz > 1) outside++;
+  }
+
+  return outside != 4;
+}
+
 JSC_CCALL(gpu_make_sprite_queue,
   size_t quads = js_arrlen(js, argv[0]);
 
-  // Reserve an array of 'sprites' if needed, but watch for out-of-bounds
+  shader_globals info = camera_globals(js,argv[1]);
+
   sprite *sprites = NULL;
-  arrsetlen(sprites, quads);
+  arrsetcap(sprites, quads);
 
   for (int i = 0; i < quads; i++) {
     JSValue sub = JS_GetPropertyUint32(js, argv[0], i);
+    JSValue jstransform = JS_GetPropertyStr(js, sub, "transform");
+    HMM_Mat3 trmat = js2transform_mat3(js,jstransform);
+    if (!sprite_in_view(trmat,info.world_to_projection)) {
+      JS_FreeValue(js,jstransform);
+      JS_FreeValue(js,sub);
+      continue;
+    }
+
     rect src;
     HMM_Vec4 color;
 
     JS_GETATOM(js, src,   sub, src_atom,   rect)
     JS_GETATOM(js, color, sub, color_atom, color)
 
-    JSValue jstransform = JS_GetPropertyStr(js, sub, "transform");
-
     quad sprite_quad;
-    transform *t;
-    JS_GETATOM(js,t,sub,transform_atom,transform)
-    HMM_Mat3 trmat = t->gcache3;
 
-    // Transform the base_quad's 4 points into sprite_quad
     for (int j = 0; j < 4; j++)
       sprite_quad.vert[j].pos = HMM_MulM3V3(trmat, base_quad[j]).xy;
 
-    // Use [0..3] for uv and color
     sprite_quad.vert[0].uv = (HMM_Vec2){ src.x,       src.y + src.h };
     sprite_quad.vert[1].uv = (HMM_Vec2){ src.x+src.w, src.y + src.h };
     sprite_quad.vert[2].uv = (HMM_Vec2){ src.x,       src.y          };
@@ -4051,31 +4133,55 @@ JSC_CCALL(gpu_make_sprite_queue,
     sprite_quad.vert[2].color = color;
     sprite_quad.vert[3].color = color;
 
-    // If you need to store into a bigger array:
-    // sprites[i].quad = sprite_quad;
-    // etc.
     sprite sp;
     sp.quad = sprite_quad;
     sp.image = JS_GetPropertyStr(js,sub,"image");
     sp.js = js;
     JS_GETPROP(js,sp.layer,sub,layer,number)
-    sprites[i] = sp;
+    arrput(sprites,sp);
-    sprites[i].sprite = JS_DupValue(js,sub);
 
     JS_FreeValue(js, sub);
     JS_FreeValue(js, jstransform);
-    JS_FreeValue(js, sp.image);
   }
 
   qsort(sprites, arrlen(sprites),sizeof(sprite),sort_sprite);
 
+  text_vert *buffer = NULL;
+  for (int i = 0; i < quads; i++)
+    for (int j = 0; j < 4; j++)
+      arrpush(buffer, sprites[i].quad.vert[j]);
+
+  JSValue mesh = quads_to_mesh(js, buffer);
+
+  arrfree(buffer);
 
   ret = JS_NewArray(js);
-
+  int first_index = 0;
-  for (int i = 0; i < quads; i++)
+  int count = 0;
-    JS_SetPropertyUint32(js,ret,i,sprites[i].sprite);
+  int n = 0;
+  JSValue img = JS_UNDEFINED;
+  for (int i = 0; i < quads; i++) {
+    if (!JS_SameValue(js,sprites[i].image, img)) {
+      if (count > 0) {
+        JSValue q = JS_NewObject(js);
+        JS_SetPropertyStr(js,q,"type", JS_NewString(js,"geometry"));
+        JS_SetPropertyStr(js,q,"mesh", JS_DupValue(js,mesh));
+        JS_SetPropertyStr(js,q,"pipeline", JS_DupValue(js,argv[2]));
+        JS_SetPropertyStr(js,q,"image", JS_DupValue(js,img));
+        JS_SetPropertyStr(js,q,"first_index", number2js(js,first_index));
+        JS_SetPropertyStr(js,q,"num_indices",number2js(js,count*6));
+        JS_SetPropertyUint32(js,ret,n++,q);
+      }
+      first_index = i*6;
+      count = 1;
+      JS_FreeValue(js,img);
+      img = JS_DupValue(js,sprites[i].image);
+    } else count++;
+    JS_FreeValue(js,sprites[i].image);
+  }
 
   arrfree(sprites);
+  JS_FreeValue(js,mesh);
 )
 
 JSC_CCALL(gpu_make_sprite_mesh,
@@ -4789,7 +4895,7 @@ static const JSCFunctionListEntry js_SDL_GPUDevice_funcs[] = {
   MIST_FUNC_DEF(gpu, load_texture, 2),
   MIST_FUNC_DEF(gpu, texture, 1),
   MIST_FUNC_DEF(gpu, make_sprite_mesh, 2),
-  MIST_FUNC_DEF(gpu, make_sprite_queue, 1),
+  MIST_FUNC_DEF(gpu, make_sprite_queue, 3),
   MIST_FUNC_DEF(gpu, make_quad, 0),
   MIST_FUNC_DEF(gpu, driver, 0),
   MIST_FUNC_DEF(gpu, make_shader, 1),
@@ -5057,61 +5163,6 @@ JSC_CCALL(cmd_hud,
   SDL_PushGPUVertexUniformData(cmds, js2number(js,argv[1]), &data, sizeof(data));  
 )
 
-shader_globals camera_globals(JSContext *js, JSValue camera)
-{
-  shader_globals data = {0};
-  HMM_Vec2 size;
-  transform *transform;
-  double fov;
-  double aspect;
-  int ortho;
-  double near;
-  double far;
-  
-  JS_GETPROP(js, size, camera, size, vec2)
-  JS_GETPROP(js, transform, camera, transform, transform)
-  JS_GETPROP(js, fov, camera, fov, number)
-  JS_GETPROP(js, aspect, camera, aspect, number)
-  JS_GETPROP(js, ortho, camera,ortho,bool)
-  JS_GETPROP(js,near,camera,near,number)
-  JS_GETPROP(js,far,camera,far,number)
-
-  HMM_Mat4 proj;
-  HMM_Mat4 view;
-  
-  if (ortho) {
-    proj = HMM_Orthographic_RH_NO(
-      -size.x*0.5, 0.5*size.x,
-      -size.y*0.5, 0.5*size.y,
-      -1.0f, 1.0f
-    );
-    view = HMM_Translate((HMM_Vec3){ -transform->pos.x, -transform->pos.y, 0.0f });    
-  }
-  else {
-    proj = HMM_Perspective_RH_NO(fov, aspect,near,far);
-
-    HMM_Mat4 camera_transform = HMM_Translate(transform->pos);
-    camera_transform = HMM_MulM4(camera_transform, HMM_QToM4(transform->rotation));
-//    camera_transform = HMM_MulM4(camera_transform, HMM_Scale(transform->scale));  // don't bother w/ scale
-    view = HMM_InvGeneralM4(camera_transform);
-  }
-
-  // Update your shader globals
-  data.world_to_projection = HMM_MulM4(proj, view);
-  data.projection_to_world = HMM_InvGeneralM4(data.world_to_projection);
-  data.camera_pos_world = transform->pos;
-  data.viewport_min_z = near;
-  data.viewport_max_z = far;
-  data.render_size = size;
-  data.world_to_view = view;
-  data.view_to_projection = proj;
-  data.camera_dir_world = HMM_NormV3(HMM_QVRot((HMM_Vec3){0,0,-1},transform->rotation));
-  data.viewport_size = (HMM_Vec2){0.5,0.5};
-  data.viewport_offset = (HMM_Vec2){0,0};
-  data.time = SDL_GetTicksNS() / 1000000000.0f;
-  return data;
-}
-
 JSC_CCALL(cmd_camera,
   SDL_GPUCommandBuffer *cmds = js2SDL_GPUCommandBuffer(js, self);
   shader_globals data = camera_globals(js, argv[0]);
@@ -6963,39 +7014,12 @@ JSC_CCALL(os_cull_sprites,
 
   int len = js_arrlen(js,sprites);
 
-  HMM_Vec2 corners[4] = {
-    (HMM_Vec2){0,0},
-    (HMM_Vec2){1,0},
-    (HMM_Vec2){0,1},
-    (HMM_Vec2){1,1},
-  };
-
   for (int i = 0; i < len; i++) {
     JSValue sub = JS_GetPropertyUint32(js,sprites, i);
     transform *t;
     JS_GETATOM(js,t,sub,transform_atom,transform)
     HMM_Mat3 trmat = t->gcache3;
-    int outside = 0;
+    if (sprite_in_view(trmat, info.world_to_projection))  {
-
-    for (int j = 0; j < 4; j++) {
-      Uint64 start = SDL_GetTicksNS();
-      HMM_Vec3 corner = HMM_MulM3V3(trmat, (HMM_Vec3){corners[j].x, corners[j].y, 1.0});    
-      HMM_Vec4 clip = HMM_MulM4V4(info.world_to_projection, (HMM_Vec4){corner.x,corner.y,corner.z,1.0});
-      if (clip.w <= 0.0) {
-        outside++;
-        continue;
-      }
-
-      float nx = clip.x/clip.w;
-      float ny = clip.y/clip.w;
-      float nz = clip.z/clip.w;
-
-      if (nx < -1 || nx > 1) outside++;
-      else if (ny < -1 || ny > 1) outside++;
-      else if (nz < -1 || nz > 1) outside++;
-    }
-
-    if (outside != 4) {
       JS_SetPropertyUint32(js,ret,n,JS_DupValue(js,sub));
       n++;
     }