Organize render surface sorting key for optimizing API performance.

Author: DarioSamo

servers/rendering/renderer_rd/forward_clustered/render_forward_clustered.cpp

@@ -844,7 +844,7 @@ void RenderForwardClustered::_fill_instance_data(RenderListType p_render_list, i
 
 		RenderElementInfo &element_info = rl->element_info[p_offset + i];
 
-		element_info.value = uint32_t((surface->sort.sort_key2 & 0x0FFF00000000) >> 32u);
+		element_info.value = uint32_t(surface->sort.sort_key1 & 0xFFF);
 
 		if (cant_repeat) {
 			prev_surface = nullptr;
@@ -4075,7 +4075,8 @@ void RenderForwardClustered::_geometry_instance_add_surface_with_material(Geomet
 	sdcache->sort.sort_key2 = 0;
 
 	sdcache->sort.surface_index = p_surface;
-	sdcache->sort.material_id = p_material_id;
+	sdcache->sort.material_id_hi = (p_material_id & 0xFF000000) >> 24;
+	sdcache->sort.material_id_lo = (p_material_id & 0x00FFFFFF);
 	sdcache->sort.shader_id = p_shader_id;
 	sdcache->sort.geometry_id = p_mesh.get_local_index(); //only meshes can repeat anyway
 	sdcache->sort.uses_forward_gi = ginstance->can_sdfgi;

servers/rendering/renderer_rd/forward_clustered/render_forward_clustered.h

@@ -498,17 +498,23 @@ class RenderForwardClustered : public RendererSceneRenderRD {
 				uint64_t sort_key2;
 			};
 			struct {
-				uint64_t geometry_id : 32;
-				uint64_t material_id : 32;
-
-				uint64_t shader_id : 32;
+				// Needs to be grouped together to be used in RenderElementInfo, as the value is masked directly.
 				uint64_t lod_index : 8;
 				uint64_t uses_softshadow : 1;
 				uint64_t uses_projector : 1;
 				uint64_t uses_forward_gi : 1;
 				uint64_t uses_lightmap : 1;
+
+				// Sorted based on optimal order for respecting priority and reducing the amount of rebinding of shaders, materials,
+				// and geometry. This current order was found to be the most optimal in large projects. If you wish to measure
+				// differences, refer to RenderingDeviceGraph and the methods available to print statistics for draw lists.
 				uint64_t depth_layer : 4;
 				uint64_t surface_index : 8;
+				uint64_t geometry_id : 32;
+				uint64_t material_id_hi : 8;
+
+				uint64_t material_id_lo : 24;
+				uint64_t shader_id : 32;
 				uint64_t priority : 8;
 			};
 		} sort;

servers/rendering/renderer_rd/forward_mobile/render_forward_mobile.cpp

@@ -1981,7 +1981,7 @@ void RenderForwardMobile::_fill_instance_data(RenderListType p_render_list, uint
 		RenderElementInfo &element_info = rl->element_info[p_offset + i];
 
 		// Sets lod_index and uses_lightmap at once.
-		element_info.value = uint32_t((surface->sort.sort_key2 & 0x01FF00000000) >> 32u);
+		element_info.value = uint32_t(surface->sort.sort_key1 & 0x1FF);
 	}
 
 	if (p_update_buffer) {
@@ -2764,7 +2764,8 @@ void RenderForwardMobile::_geometry_instance_add_surface_with_material(GeometryI
 	sdcache->sort.sort_key2 = 0;
 
 	sdcache->sort.surface_index = p_surface;
-	sdcache->sort.material_id = p_material_id;
+	sdcache->sort.material_id_hi = (p_material_id & 0xFF000000) >> 24;
+	sdcache->sort.material_id_lo = (p_material_id & 0x00FFFFFF);
 	sdcache->sort.shader_id = p_shader_id;
 	sdcache->sort.geometry_id = p_mesh.get_local_index();
 	sdcache->sort.priority = p_material->priority;

servers/rendering/renderer_rd/forward_mobile/render_forward_mobile.h

@@ -478,15 +478,21 @@ class RenderForwardMobile : public RendererSceneRenderRD {
 				uint64_t sort_key2;
 			};
 			struct {
-				uint64_t geometry_id : 32;
-				uint64_t material_id : 32;
-
-				uint64_t shader_id : 32;
+				// Needs to be grouped together to be used in RenderElementInfo, as the value is masked directly.
 				uint64_t lod_index : 8;
 				uint64_t uses_lightmap : 1;
 				uint64_t pad : 3;
+
+				// Sorted based on optimal order for respecting priority and reducing the amount of rebinding of shaders, materials,
+				// and geometry. This current order was found to be the most optimal in large projects. If you wish to measure
+				// differences, refer to RenderingDeviceGraph and the methods available to print statistics for draw lists.
 				uint64_t depth_layer : 4;
 				uint64_t surface_index : 8;
+				uint64_t geometry_id : 32;
+				uint64_t material_id_hi : 8;
+
+				uint64_t material_id_lo : 24;
+				uint64_t shader_id : 32;
 				uint64_t priority : 8;
 			};
 		} sort;

servers/rendering/rendering_device_graph.cpp

@@ -35,6 +35,9 @@
 #define PRINT_RESOURCE_TRACKER_TOTAL 0
 #define PRINT_COMMAND_RECORDING 0
 
+// Prints the total number of bytes used for draw lists in a frame.
+#define PRINT_DRAW_LIST_STATS 0
+
 RenderingDeviceGraph::RenderingDeviceGraph() {
 	driver_honors_barriers = false;
 	driver_clears_with_copy_engine = false;
@@ -835,7 +838,15 @@ void RenderingDeviceGraph::_get_draw_list_render_pass_and_framebuffer(const Reco
 	r_framebuffer = it->value.framebuffer;
 }
 
+#if PRINT_DRAW_LIST_STATS
+static uint32_t draw_list_total_size = 0;
+#endif
+
 void RenderingDeviceGraph::_run_draw_list_command(RDD::CommandBufferID p_command_buffer, const uint8_t *p_instruction_data, uint32_t p_instruction_data_size) {
+#if PRINT_DRAW_LIST_STATS
+	draw_list_total_size += p_instruction_data_size;
+#endif
+
 	uint32_t instruction_data_cursor = 0;
 	while (instruction_data_cursor < p_instruction_data_size) {
 		DEV_ASSERT((instruction_data_cursor + sizeof(DrawListInstruction)) <= p_instruction_data_size);
@@ -2366,6 +2377,10 @@ void RenderingDeviceGraph::end(bool p_reorder_commands, bool p_full_barriers, RD
 			workarounds_state.draw_list_found = false;
 		}
 
+#if PRINT_DRAW_LIST_STATS
+		draw_list_total_size = 0;
+#endif
+
 		if (p_reorder_commands) {
 #if PRINT_RENDER_GRAPH
 			print_line("BEFORE SORT");
@@ -2416,6 +2431,9 @@ void RenderingDeviceGraph::end(bool p_reorder_commands, bool p_full_barriers, RD
 
 		_run_label_command_change(r_command_buffer, -1, -1, false, false, nullptr, 0, current_label_index, current_label_level);
 
+#if PRINT_DRAW_LIST_STATS
+		print_line(vformat("Draw list %d bytes", draw_list_total_size));
+#endif
 #if PRINT_COMMAND_RECORDING
 		print_line(vformat("Recorded %d commands", command_count));
 #endif