Merge pull request #114175 from BastiaanOlij/opengl_split_motion_vector_ubos

OpenGL: Split the ubos for motion vectors into separate uniforms

drivers/gles3/rasterizer_scene_gles3.cpp

@@ -1477,6 +1477,19 @@ void RasterizerSceneGLES3::_fill_render_list(RenderListType p_render_list, const
 	}
 }
 
+void RasterizerSceneGLES3::_update_scene_ubo(GLuint &p_ubo_buffer, GLuint p_index, uint32_t p_size, const void *p_source_data, String p_name) {
+	if (p_ubo_buffer == 0) {
+		glGenBuffers(1, &p_ubo_buffer);
+		glBindBufferBase(GL_UNIFORM_BUFFER, p_index, p_ubo_buffer);
+		GLES3::Utilities::get_singleton()->buffer_allocate_data(GL_UNIFORM_BUFFER, p_ubo_buffer, p_size, p_source_data, GL_STREAM_DRAW, p_name);
+	} else {
+		glBindBufferBase(GL_UNIFORM_BUFFER, p_index, p_ubo_buffer);
+		glBufferData(GL_UNIFORM_BUFFER, p_size, p_source_data, GL_STREAM_DRAW);
+	}
+
+	glBindBuffer(GL_UNIFORM_BUFFER, 0);
+}
+
 // Needs to be called after _setup_lights so that directional_light_count is accurate.
 void RasterizerSceneGLES3::_setup_environment(const RenderDataGLES3 *p_render_data, bool p_no_fog, const Size2i &p_screen_size, bool p_flip_y, const Color &p_default_bg_color, bool p_pancake_shadows, float p_shadow_bias) {
 	Projection correction;
@@ -1619,29 +1632,19 @@ void RasterizerSceneGLES3::_setup_environment(const RenderDataGLES3 *p_render_da
 		scene_state.data.IBL_exposure_normalization = 1.0;
 	}
 
-	if (scene_state.ubo_buffer == 0) {
+	_update_scene_ubo(scene_state.ubo_buffer, SCENE_DATA_UNIFORM_LOCATION, sizeof(SceneState::UBO), &scene_state.data, "Scene state UBO");
-		glGenBuffers(1, &scene_state.ubo_buffer);
+	if (p_render_data->view_count > 1) {
-		glBindBufferBase(GL_UNIFORM_BUFFER, SCENE_DATA_UNIFORM_LOCATION, scene_state.ubo_buffer);
+		_update_scene_ubo(scene_state.multiview_buffer, SCENE_MULTIVIEW_UNIFORM_LOCATION, sizeof(SceneState::MultiviewUBO), &scene_state.multiview_data, "Multiview UBO");
-		GLES3::Utilities::get_singleton()->buffer_allocate_data(GL_UNIFORM_BUFFER, scene_state.ubo_buffer, sizeof(SceneState::UBO) * 2, &scene_state.data, GL_STREAM_DRAW, "Scene state UBO");
-		glBindBuffer(GL_UNIFORM_BUFFER, 0);
-	} else {
-		glBindBufferBase(GL_UNIFORM_BUFFER, SCENE_DATA_UNIFORM_LOCATION, scene_state.ubo_buffer);
-		glBufferData(GL_UNIFORM_BUFFER, sizeof(SceneState::UBO) * 2, &scene_state.data, GL_STREAM_DRAW);
 	}
 
-	glBindBuffer(GL_UNIFORM_BUFFER, 0);
+	if (scene_state.prev_data_state != 0) {
+		void *source_data = scene_state.prev_data_state == 1 ? &scene_state.data : &scene_state.prev_data;
+		_update_scene_ubo(scene_state.prev_ubo_buffer, SCENE_PREV_DATA_UNIFORM_LOCATION, sizeof(SceneState::UBO), source_data, "Previous scene state UBO");
 
-	if (p_render_data->view_count > 1) {
+		if (p_render_data->view_count > 1) {
-		if (scene_state.multiview_buffer == 0) {
+			source_data = scene_state.prev_data_state == 1 ? &scene_state.multiview_data : &scene_state.prev_multiview_data;
-			glGenBuffers(1, &scene_state.multiview_buffer);
+			_update_scene_ubo(scene_state.prev_multiview_buffer, SCENE_PREV_MULTIVIEW_UNIFORM_LOCATION, sizeof(SceneState::MultiviewUBO), source_data, "Previous multiview UBO");
-			glBindBufferBase(GL_UNIFORM_BUFFER, SCENE_MULTIVIEW_UNIFORM_LOCATION, scene_state.multiview_buffer);
-			GLES3::Utilities::get_singleton()->buffer_allocate_data(GL_UNIFORM_BUFFER, scene_state.multiview_buffer, sizeof(SceneState::MultiviewUBO) * 2, &scene_state.multiview_data, GL_STREAM_DRAW, "Multiview UBO");
-		} else {
-			glBindBufferBase(GL_UNIFORM_BUFFER, SCENE_MULTIVIEW_UNIFORM_LOCATION, scene_state.multiview_buffer);
-			glBufferData(GL_UNIFORM_BUFFER, sizeof(SceneState::MultiviewUBO) * 2, &scene_state.multiview_data, GL_STREAM_DRAW);
 		}
-
-		glBindBuffer(GL_UNIFORM_BUFFER, 0);
 	}
 }
 
@@ -2408,6 +2411,17 @@ void RasterizerSceneGLES3::render_scene(const Ref<RenderSceneBuffers> &p_render_
 
 	scene_state.data.emissive_exposure_normalization = -1.0; // Use default exposure normalization.
 
+	bool enough_vertex_attribs_for_motion_blue = GLES3::Config::get_singleton()->max_vertex_attribs >= 22;
+	if (rt && rt->overridden.velocity_fbo != 0 && enough_vertex_attribs_for_motion_blue) {
+		// First frame we render motion vectors? Use our current data!
+		if (scene_state.prev_data_state == 0) {
+			scene_state.prev_data_state = 1;
+		}
+	} else {
+		// Not using motion vectors? We don't need to load our data.
+		scene_state.prev_data_state = 0;
+	}
+
 	bool flip_y = !is_reflection_probe;
 
 	if (rt && rt->overridden.color.is_valid()) {
@@ -2520,19 +2534,13 @@ void RasterizerSceneGLES3::render_scene(const Ref<RenderSceneBuffers> &p_render_
 	scene_state.reset_gl_state();
 
 	GLuint motion_vectors_fbo = rt ? rt->overridden.velocity_fbo : 0;
-	if (motion_vectors_fbo != 0 && GLES3::Config::get_singleton()->max_vertex_attribs >= 22) {
+	if (motion_vectors_fbo != 0 && enough_vertex_attribs_for_motion_blue) {
 		RENDER_TIMESTAMP("Motion Vectors Pass");
 		glBindFramebuffer(GL_FRAMEBUFFER, motion_vectors_fbo);
 
 		Size2i motion_vectors_target_size = rt->velocity_target_size;
 		glViewport(0, 0, motion_vectors_target_size.x, motion_vectors_target_size.y);
 
-		if (!scene_state.is_prev_data_stored) {
-			scene_state.prev_data = scene_state.data;
-			scene_state.prev_multiview_data = scene_state.multiview_data;
-			scene_state.is_prev_data_stored = true;
-		}
-
 		scene_state.enable_gl_depth_test(true);
 		scene_state.enable_gl_depth_draw(true);
 		scene_state.enable_gl_blend(false);
@@ -2553,8 +2561,10 @@ void RasterizerSceneGLES3::render_scene(const Ref<RenderSceneBuffers> &p_render_
 		RenderListParameters render_list_params(render_list[RENDER_LIST_OPAQUE].elements.ptr(), render_list[RENDER_LIST_OPAQUE].elements.size(), reverse_cull, spec_constant, use_wireframe);
 		_render_list_template<PASS_MODE_MOTION_VECTORS>(&render_list_params, &render_data, 0, render_list[RENDER_LIST_OPAQUE].elements.size());
 
+		// Copy our current scene data to our previous scene data for use in the next frame.
 		scene_state.prev_data = scene_state.data;
 		scene_state.prev_multiview_data = scene_state.multiview_data;
+		scene_state.prev_data_state = 2;
 	}
 
 	GLuint fbo = 0;
@@ -4602,10 +4612,18 @@ RasterizerSceneGLES3::~RasterizerSceneGLES3() {
 		GLES3::Utilities::get_singleton()->buffer_free_data(scene_state.ubo_buffer);
 	}
 
+	if (scene_state.prev_ubo_buffer != 0) {
+		GLES3::Utilities::get_singleton()->buffer_free_data(scene_state.prev_ubo_buffer);
+	}
+
 	if (scene_state.multiview_buffer != 0) {
 		GLES3::Utilities::get_singleton()->buffer_free_data(scene_state.multiview_buffer);
 	}
 
+	if (scene_state.prev_multiview_buffer != 0) {
+		GLES3::Utilities::get_singleton()->buffer_free_data(scene_state.prev_multiview_buffer);
+	}
+
 	if (scene_state.tonemap_buffer != 0) {
 		GLES3::Utilities::get_singleton()->buffer_free_data(scene_state.tonemap_buffer);
 	}

drivers/gles3/rasterizer_scene_gles3.h

@@ -79,6 +79,8 @@ enum SceneUniformLocation {
 	SCENE_POSITIONAL_SHADOW_UNIFORM_LOCATION,
 	SCENE_DIRECTIONAL_SHADOW_UNIFORM_LOCATION,
 	SCENE_EMPTY2, // Unused, put here to avoid conflicts with SKY_MULTIVIEW_UNIFORM_LOCATION.
+	SCENE_PREV_DATA_UNIFORM_LOCATION,
+	SCENE_PREV_MULTIVIEW_UNIFORM_LOCATION,
 };
 
 enum SkyUniformLocation {
@@ -94,6 +96,8 @@ enum SkyUniformLocation {
 	SKY_EMPTY6, // Unused, put here to avoid conflicts with SCENE_POSITIONAL_SHADOW_UNIFORM_LOCATION.
 	SKY_EMPTY7, // Unused, put here to avoid conflicts with SCENE_DIRECTIONAL_SHADOW_UNIFORM_LOCATION.
 	SKY_MULTIVIEW_UNIFORM_LOCATION,
+	SKY_EMPTY8, // Unused, put here to avoid conflicts with SCENE_PREV_DATA_UNIFORM_LOCATION.
+	SKY_EMPTY9, // Unused, put here to avoid conflicts with SCENE_PREV_MULTIVIEW_UNIFORM_LOCATION.
 };
 
 struct RenderDataGLES3 {
@@ -463,12 +467,14 @@ private:
 		UBO data;
 		UBO prev_data;
 		GLuint ubo_buffer = 0;
+		GLuint prev_ubo_buffer = 0;
 		MultiviewUBO multiview_data;
 		MultiviewUBO prev_multiview_data;
 		GLuint multiview_buffer = 0;
+		GLuint prev_multiview_buffer = 0;
 		GLuint tonemap_buffer = 0;
 
-		bool is_prev_data_stored = false;
+		int prev_data_state = 0; // 0 = Motion vectors not used, 1 = use data (first frame only), 2 = use previous data
 
 		bool used_depth_prepass = false;
 
@@ -717,6 +723,8 @@ private:
 
 	RenderList render_list[RENDER_LIST_MAX];
 
+	void _update_scene_ubo(GLuint &p_ubo_buffer, GLuint p_index, uint32_t p_size, const void *p_source_data, String p_name = "");
+
 	void _setup_lights(const RenderDataGLES3 *p_render_data, bool p_using_shadows, uint32_t &r_directional_light_count, uint32_t &r_omni_light_count, uint32_t &r_spot_light_count, uint32_t &r_directional_shadow_count);
 	void _setup_environment(const RenderDataGLES3 *p_render_data, bool p_no_fog, const Size2i &p_screen_size, bool p_flip_y, const Color &p_default_bg_color, bool p_pancake_shadows, float p_shadow_bias = 0.0);
 	void _fill_render_list(RenderListType p_render_list, const RenderDataGLES3 *p_render_data, PassMode p_pass_mode, bool p_append = false);

drivers/gles3/shaders/scene.glsl

@@ -198,7 +198,7 @@ struct SceneData {
 	float fog_aerial_perspective;
 	float time;
 
-	mat3 radiance_inverse_xform;
+	mat3x4 radiance_inverse_xform;
 
 	uint directional_light_count;
 	float z_far;
@@ -224,12 +224,16 @@ struct SceneData {
 	bool pancake_shadows;
 };
 
+// The containing data block is for historic reasons.
 layout(std140) uniform SceneDataBlock { // ubo:2
 	SceneData data;
-	SceneData prev_data;
 }
 scene_data_block;
 
+#ifdef RENDER_MOTION_VECTORS
+layout(std140) uniform SceneData prev_scene_data; // ubo:12
+#endif
+
 #ifndef RENDER_MOTION_VECTORS
 #ifdef USE_ADDITIVE_LIGHTING
 
@@ -456,10 +460,14 @@ struct MultiviewData {
 
 layout(std140) uniform MultiviewDataBlock { // ubo:8
 	MultiviewData data;
-	MultiviewData prev_data;
 }
 multiview_data_block;
-#endif
+
+#ifdef RENDER_MOTION_VECTORS
+layout(std140) uniform MultiviewData prev_multiview_data; // ubo:13
+#endif // RENDER_MOTION_VECTORS
+
+#endif // USE_MULTIVIEW
 
 uniform highp mat4 world_transform;
 uniform highp vec3 compressed_aabb_position;
@@ -901,7 +909,7 @@ void main() {
 			compressed_aabb_position,
 			prev_world_transform,
 			model_flags,
-			scene_data_block.prev_data,
+			prev_scene_data.data,
 #ifdef USE_INSTANCING
 			input_instance_xform0, input_instance_xform1, input_instance_xform2,
 			input_instance_color_custom_data,
@@ -919,9 +927,9 @@ void main() {
 			uv2_attrib,
 #endif
 #ifdef USE_MULTIVIEW
-			multiview_data_block.prev_data.projection_matrix_view[ViewIndex],
+			prev_multiview_data.projection_matrix_view[ViewIndex],
-			multiview_data_block.prev_data.inv_projection_matrix_view[ViewIndex],
+			prev_multiview_data.inv_projection_matrix_view[ViewIndex],
-			multiview_data_block.prev_data.eye_offset[ViewIndex].xyz,
+			prev_multiview_data.eye_offset[ViewIndex].xyz,
 #endif
 			uv_scale,
 			prev_clip_position);
@@ -1146,7 +1154,7 @@ struct SceneData {
 	float fog_aerial_perspective;
 	float time;
 
-	mat3 radiance_inverse_xform;
+	mat3x4 radiance_inverse_xform;
 
 	uint directional_light_count;
 	float z_far;
@@ -1174,7 +1182,6 @@ struct SceneData {
 
 layout(std140) uniform SceneDataBlock { // ubo:2
 	SceneData data;
-	SceneData prev_data;
 }
 scene_data_block;
 
@@ -1187,7 +1194,6 @@ struct MultiviewData {
 
 layout(std140) uniform MultiviewDataBlock { // ubo:8
 	MultiviewData data;
-	MultiviewData prev_data;
 }
 multiview_data_block;
 #endif
@@ -2205,7 +2211,7 @@ void main() {
 #endif
 		ref_vec = mix(ref_vec, normal, roughness * roughness);
 		float horizon = min(1.0 + dot(ref_vec, normal), 1.0);
-		ref_vec = scene_data_block.data.radiance_inverse_xform * ref_vec;
+		ref_vec = mat3(scene_data_block.data.radiance_inverse_xform) * ref_vec;
 		specular_light = textureLod(radiance_map, ref_vec, sqrt(roughness) * RADIANCE_MAX_LOD).rgb;
 		specular_light = srgb_to_linear(specular_light);
 		specular_light *= horizon * horizon;
@@ -2250,7 +2256,7 @@ void main() {
 
 #ifdef USE_RADIANCE_MAP
 		if (scene_data_block.data.use_ambient_cubemap) {
-			vec3 ambient_dir = scene_data_block.data.radiance_inverse_xform * normal;
+			vec3 ambient_dir = mat3(scene_data_block.data.radiance_inverse_xform) * normal;
 			vec3 cubemap_ambient = textureLod(radiance_map, ambient_dir, RADIANCE_MAX_LOD).rgb;
 			cubemap_ambient = srgb_to_linear(cubemap_ambient);
 			ambient_light = mix(ambient_light, cubemap_ambient * scene_data_block.data.ambient_light_color_energy.a, scene_data_block.data.ambient_color_sky_mix);