Merge pull request #51486 from reduz/fixes-to-mobile-renderer

Fixes and optimizations to mobile renderer

5 files changed, 51 insertions, 46 deletions

diff --git a/servers/rendering/renderer_rd/forward_mobile/render_forward_mobile.cpp b/servers/rendering/renderer_rd/forward_mobile/render_forward_mobile.cpp
index dea2975660..f810ae2454 100644
--- a/servers/rendering/renderer_rd/forward_mobile/render_forward_mobile.cpp
+++ b/servers/rendering/renderer_rd/forward_mobile/render_forward_mobile.cpp
@@ -519,7 +519,7 @@ void RenderForwardMobile::_render_scene(RenderDataRD *p_render_data, const Color
 			using_subpass_post_process = false;
 		}
 
-		if (scene_state.used_screen_texture || scene_state.used_depth_texture) {
+		if (using_ssr || using_sss || scene_state.used_screen_texture || scene_state.used_depth_texture) {
 			// can't use our last two subpasses
 			using_subpass_transparent = false;
 			using_subpass_post_process = false;
@@ -679,17 +679,20 @@ void RenderForwardMobile::_render_scene(RenderDataRD *p_render_data, const Color
 
 		_setup_environment(p_render_data, p_render_data->reflection_probe.is_valid(), screen_size, !p_render_data->reflection_probe.is_valid(), p_default_bg_color, p_render_data->render_buffers.is_valid());
 
-		RENDER_TIMESTAMP("Render Opaque Subpass");
+		if (using_subpass_transparent && using_subpass_post_process) {
+			RENDER_TIMESTAMP("Render Opaque + Transparent + Tonemap");
+		} else if (using_subpass_transparent) {
+			RENDER_TIMESTAMP("Render Opaque + Transparent");
+		} else {
+			RENDER_TIMESTAMP("Render Opaque");
+		}
 
 		RID rp_uniform_set = _setup_render_pass_uniform_set(RENDER_LIST_OPAQUE, p_render_data, radiance_texture, true);
 
-		bool can_continue_color = !scene_state.used_screen_texture && !using_ssr && !using_sss;
-		bool can_continue_depth = !scene_state.used_depth_texture && !using_ssr && !using_sss;
+		bool can_continue_color = !using_subpass_transparent && !scene_state.used_screen_texture && !using_ssr && !using_sss;
+		bool can_continue_depth = !using_subpass_transparent && !scene_state.used_depth_texture && !using_ssr && !using_sss;
 
 		{
-			bool will_continue_color = (can_continue_color || draw_sky || draw_sky_fog_only);
-			bool will_continue_depth = (can_continue_depth || draw_sky || draw_sky_fog_only);
-
 			// regular forward for now
 			Vector<Color> c;
 			c.push_back(clear_color.to_linear()); // our render buffer
@@ -709,11 +712,11 @@ void RenderForwardMobile::_render_scene(RenderDataRD *p_render_data, const Color
 				// secondary command buffers need more testing at this time
 				//multi threaded
 				thread_draw_lists.resize(RendererThreadPool::singleton->thread_work_pool.get_thread_count());
-				RD::get_singleton()->draw_list_begin_split(framebuffer, thread_draw_lists.size(), thread_draw_lists.ptr(), keep_color ? RD::INITIAL_ACTION_KEEP : RD::INITIAL_ACTION_CLEAR, will_continue_color ? RD::FINAL_ACTION_CONTINUE : RD::FINAL_ACTION_READ, RD::INITIAL_ACTION_CLEAR, will_continue_depth ? RD::FINAL_ACTION_CONTINUE : RD::FINAL_ACTION_READ, c, 1.0, 0);
+				RD::get_singleton()->draw_list_begin_split(framebuffer, thread_draw_lists.size(), thread_draw_lists.ptr(), keep_color ? RD::INITIAL_ACTION_KEEP : RD::INITIAL_ACTION_CLEAR, can_continue_color ? RD::FINAL_ACTION_CONTINUE : RD::FINAL_ACTION_READ, RD::INITIAL_ACTION_CLEAR, can_continue_depth ? RD::FINAL_ACTION_CONTINUE : RD::FINAL_ACTION_READ, c, 1.0, 0);
 				RendererThreadPool::singleton->thread_work_pool.do_work(thread_draw_lists.size(), this, &RenderForwardMobile::_render_list_thread_function, &render_list_params);
 			} else {
 				//single threaded
-				RD::DrawListID draw_list = RD::get_singleton()->draw_list_begin(framebuffer, keep_color ? RD::INITIAL_ACTION_KEEP : RD::INITIAL_ACTION_CLEAR, will_continue_color ? RD::FINAL_ACTION_CONTINUE : RD::FINAL_ACTION_READ, RD::INITIAL_ACTION_CLEAR, will_continue_depth ? RD::FINAL_ACTION_CONTINUE : RD::FINAL_ACTION_READ, c, 1.0, 0);
+				RD::DrawListID draw_list = RD::get_singleton()->draw_list_begin(framebuffer, keep_color ? RD::INITIAL_ACTION_KEEP : RD::INITIAL_ACTION_CLEAR, can_continue_color ? RD::FINAL_ACTION_CONTINUE : RD::FINAL_ACTION_READ, RD::INITIAL_ACTION_CLEAR, can_continue_depth ? RD::FINAL_ACTION_CONTINUE : RD::FINAL_ACTION_READ, c, 1.0, 0);
 				_render_list(draw_list, fb_format, &render_list_params, 0, render_list_params.element_count);
 			}
 		}
@@ -721,8 +724,6 @@ void RenderForwardMobile::_render_scene(RenderDataRD *p_render_data, const Color
 		RD::get_singleton()->draw_command_end_label(); //Render Opaque Subpass
 
 		if (draw_sky || draw_sky_fog_only) {
-			RENDER_TIMESTAMP("Render Sky Subpass");
-
 			RD::get_singleton()->draw_command_begin_label("Draw Sky Subpass");
 
 			RD::DrawListID draw_list = RD::get_singleton()->draw_list_switch_to_next_pass();
@@ -752,7 +753,6 @@ void RenderForwardMobile::_render_scene(RenderDataRD *p_render_data, const Color
 		}
 
 		// transparent pass
-		RENDER_TIMESTAMP("Render Transparent Subpass");
 
 		RD::get_singleton()->draw_command_begin_label("Render Transparent Subpass");
 
@@ -789,6 +789,8 @@ void RenderForwardMobile::_render_scene(RenderDataRD *p_render_data, const Color
 
 			RD::get_singleton()->draw_list_end(RD::BARRIER_MASK_ALL);
 		} else {
+			RENDER_TIMESTAMP("Render Transparent");
+
 			framebuffer = render_buffer->color_fbs[FB_CONFIG_ONE_PASS];
 
 			// this may be needed if we re-introduced steps that change info, not sure which do so in the previous implementation
diff --git a/servers/rendering/renderer_rd/renderer_scene_render_rd.cpp b/servers/rendering/renderer_rd/renderer_scene_render_rd.cpp
index 5b734fc603..fd797b1cab 100644
--- a/servers/rendering/renderer_rd/renderer_scene_render_rd.cpp
+++ b/servers/rendering/renderer_rd/renderer_scene_render_rd.cpp
@@ -2010,7 +2010,6 @@ void RendererSceneRenderRD::_post_process_subpass(RID p_source_texture, RID p_fr
 
 	bool can_use_effects = rb->width >= 8 && rb->height >= 8;
 
-	RENDER_TIMESTAMP("Tonemap");
 	RD::DrawListID draw_list = RD::get_singleton()->draw_list_switch_to_next_pass();
 
 	EffectsRD::TonemapSettings tonemap;
diff --git a/servers/rendering/renderer_rd/shaders/light_data_inc.glsl b/servers/rendering/renderer_rd/shaders/light_data_inc.glsl
index 2fce258cff..9155216d7e 100644
--- a/servers/rendering/renderer_rd/shaders/light_data_inc.glsl
+++ b/servers/rendering/renderer_rd/shaders/light_data_inc.glsl
@@ -3,7 +3,7 @@
 #define LIGHT_BAKE_STATIC 2
 
 struct LightData { //this structure needs to be as packed as possible
-	vec3 position;
+	highp vec3 position;
 	float inv_radius;
 
 	vec3 direction;
@@ -17,8 +17,8 @@ struct LightData { //this structure needs to be as packed as possible
 	float specular_amount;
 	bool shadow_enabled;
 
-	vec4 atlas_rect; // rect in the shadow atlas
-	mat4 shadow_matrix;
+	highp vec4 atlas_rect; // rect in the shadow atlas
+	highp mat4 shadow_matrix;
 	float shadow_bias;
 	float shadow_normal_bias;
 	float transmittance_bias;
@@ -27,7 +27,7 @@ struct LightData { //this structure needs to be as packed as possible
 	uint mask;
 	float shadow_volumetric_fog_fade;
 	uint bake_mode;
-	vec4 projector_rect; //projector rect in srgb decal atlas
+	highp vec4 projector_rect; //projector rect in srgb decal atlas
 };
 
 #define REFLECTION_AMBIENT_DISABLED 0
@@ -69,13 +69,13 @@ struct DirectionalLightData {
 	vec4 shadow_bias;
 	vec4 shadow_normal_bias;
 	vec4 shadow_transmittance_bias;
-	vec4 shadow_z_range;
-	vec4 shadow_range_begin;
+	highp vec4 shadow_z_range;
+	highp vec4 shadow_range_begin;
 	vec4 shadow_split_offsets;
-	mat4 shadow_matrix1;
-	mat4 shadow_matrix2;
-	mat4 shadow_matrix3;
-	mat4 shadow_matrix4;
+	highp mat4 shadow_matrix1;
+	highp mat4 shadow_matrix2;
+	highp mat4 shadow_matrix3;
+	highp mat4 shadow_matrix4;
 	vec4 shadow_color1;
 	vec4 shadow_color2;
 	vec4 shadow_color3;
diff --git a/servers/rendering/renderer_rd/shaders/scene_forward_mobile.glsl b/servers/rendering/renderer_rd/shaders/scene_forward_mobile.glsl
index 681656e1d5..2babe92c1c 100644
--- a/servers/rendering/renderer_rd/shaders/scene_forward_mobile.glsl
+++ b/servers/rendering/renderer_rd/shaders/scene_forward_mobile.glsl
@@ -59,27 +59,27 @@ layout(location = 11) in vec4 weight_attrib;
 
 /* Varyings */
 
-layout(location = 0) out vec3 vertex_interp;
+layout(location = 0) highp out vec3 vertex_interp;
 
 #ifdef NORMAL_USED
-layout(location = 1) out vec3 normal_interp;
+layout(location = 1) mediump out vec3 normal_interp;
 #endif
 
 #if defined(COLOR_USED)
-layout(location = 2) out vec4 color_interp;
+layout(location = 2) mediump out vec4 color_interp;
 #endif
 
 #ifdef UV_USED
-layout(location = 3) out vec2 uv_interp;
+layout(location = 3) mediump out vec2 uv_interp;
 #endif
 
 #if defined(UV2_USED) || defined(USE_LIGHTMAP)
-layout(location = 4) out vec2 uv2_interp;
+layout(location = 4) mediump out vec2 uv2_interp;
 #endif
 
 #if defined(TANGENT_USED) || defined(NORMAL_MAP_USED) || defined(LIGHT_ANISOTROPY_USED)
-layout(location = 5) out vec3 tangent_interp;
-layout(location = 6) out vec3 binormal_interp;
+layout(location = 5) mediump out vec3 tangent_interp;
+layout(location = 6) mediump out vec3 binormal_interp;
 #endif
 
 #ifdef MATERIAL_UNIFORMS_USED
@@ -370,6 +370,10 @@ void main() {
 
 #VERSION_DEFINES
 
+//use medium precision for floats on mobile.
+
+precision mediump float;
+
 /* Specialization Constants */
 
 /* Specialization Constants (Toggles) */
@@ -395,32 +399,32 @@ layout(constant_id = 11) const bool sc_projector_use_mipmaps = true;
 
 /* Varyings */
 
-layout(location = 0) in vec3 vertex_interp;
+layout(location = 0) highp in vec3 vertex_interp;
 
 #ifdef NORMAL_USED
-layout(location = 1) in vec3 normal_interp;
+layout(location = 1) mediump in vec3 normal_interp;
 #endif
 
 #if defined(COLOR_USED)
-layout(location = 2) in vec4 color_interp;
+layout(location = 2) mediump in vec4 color_interp;
 #endif
 
 #ifdef UV_USED
-layout(location = 3) in vec2 uv_interp;
+layout(location = 3) mediump in vec2 uv_interp;
 #endif
 
 #if defined(UV2_USED) || defined(USE_LIGHTMAP)
-layout(location = 4) in vec2 uv2_interp;
+layout(location = 4) mediump in vec2 uv2_interp;
 #endif
 
 #if defined(TANGENT_USED) || defined(NORMAL_MAP_USED) || defined(LIGHT_ANISOTROPY_USED)
-layout(location = 5) in vec3 tangent_interp;
-layout(location = 6) in vec3 binormal_interp;
+layout(location = 5) mediump in vec3 tangent_interp;
+layout(location = 6) mediump in vec3 binormal_interp;
 #endif
 
 #ifdef MODE_DUAL_PARABOLOID
 
-layout(location = 8) in float dp_clip;
+layout(location = 8) highp in float dp_clip;
 
 #endif
 
diff --git a/servers/rendering/renderer_rd/shaders/scene_forward_mobile_inc.glsl b/servers/rendering/renderer_rd/shaders/scene_forward_mobile_inc.glsl
index d9682d7b23..f1e554d01c 100644
--- a/servers/rendering/renderer_rd/shaders/scene_forward_mobile_inc.glsl
+++ b/servers/rendering/renderer_rd/shaders/scene_forward_mobile_inc.glsl
@@ -16,12 +16,12 @@
 /* don't exceed 128 bytes!! */
 /* put instance data into our push content, not a array */
 layout(push_constant, binding = 0, std430) uniform DrawCall {
-	mat4 transform; // 64 - 64
+	highp mat4 transform; // 64 - 64
 	uint flags; // 04 - 68
 	uint instance_uniforms_ofs; //base offset in global buffer for instance variables	// 04 - 72
 	uint gi_offset; //GI information when using lightmapping (VCT or lightmap index)    // 04 - 76
 	uint layer_mask; // 04 - 80
-	vec4 lightmap_uv_scale; // 16 - 96 doubles as uv_offset when needed
+	highp vec4 lightmap_uv_scale; // 16 - 96 doubles as uv_offset when needed
 
 	uvec2 reflection_probes; // 08 - 104
 	uvec2 omni_lights; // 08 - 112
@@ -126,14 +126,14 @@ global_variables;
 /* Set 1: Render Pass (changes per render pass) */
 
 layout(set = 1, binding = 0, std140) uniform SceneData {
-	mat4 projection_matrix;
-	mat4 inv_projection_matrix;
-	mat4 camera_matrix;
-	mat4 inv_camera_matrix;
+	highp mat4 projection_matrix;
+	highp mat4 inv_projection_matrix;
+	highp mat4 camera_matrix;
+	highp mat4 inv_camera_matrix;
 
 	// only used for multiview
-	mat4 projection_matrix_view[MAX_VIEWS];
-	mat4 inv_projection_matrix_view[MAX_VIEWS];
+	highp mat4 projection_matrix_view[MAX_VIEWS];
+	highp mat4 inv_projection_matrix_view[MAX_VIEWS];
 
 	vec2 viewport_size;
 	vec2 screen_pixel_size;