Merge pull request #94368 from CrazyRoka/optimize-render-pass-uniform-set

Optimize `RenderForwardClustered::_setup_render_pass_uniform_set` by reducing Vector allocations

servers/rendering/renderer_rd/effects/fsr2.cpp

@@ -394,7 +394,9 @@ static FfxErrorCode execute_gpu_job_compute_rd(FSR2Context::Scratch &p_scratch,
 	FSR2Effect::Pipeline &backend_pipeline = *reinterpret_cast<FSR2Effect::Pipeline *>(p_job.pipeline.pipeline);
 	ERR_FAIL_COND_V(backend_pipeline.pipeline_rid.is_null(), FFX_ERROR_INVALID_ARGUMENT);
 
-	Vector<RD::Uniform> compute_uniforms;
+	thread_local LocalVector<RD::Uniform> compute_uniforms;
+	compute_uniforms.clear();
+
 	for (uint32_t i = 0; i < p_job.pipeline.srvCount; i++) {
 		RID texture_rid = p_scratch.resources.rids[p_job.srvs[i].internalIndex];
 		RD::Uniform texture_uniform(RD::UNIFORM_TYPE_TEXTURE, p_job.pipeline.srvResourceBindings[i].slotIndex, texture_rid);

servers/rendering/renderer_rd/effects/ss_effects.cpp

@@ -464,7 +464,8 @@ void SSEffects::downsample_depth(Ref<RenderSceneBuffersRD> p_render_buffers, uin
 	if (use_mips) {
 		// Grab our downsample uniform set from cache, these are automatically cleaned up if the depth textures are cleared.
 		// This also ensures we can switch between left eye and right eye uniform sets without recreating the uniform twice a frame.
-		Vector<RD::Uniform> u_depths;
+		thread_local LocalVector<RD::Uniform> u_depths;
+		u_depths.clear();
 
 		// Note, use_full_mips is true if either SSAO or SSIL uses half size, but the other full size and we're using mips.
 		// That means we're filling all 5 levels.

servers/rendering/renderer_rd/environment/fog.cpp

@@ -264,7 +264,7 @@ ALBEDO = vec3(1.0);
 			uniforms.push_back(u);
 		}
 
-		uniforms.append_array(material_storage->samplers_rd_get_default().get_uniforms(SAMPLERS_BINDING_FIRST_INDEX));
+		material_storage->samplers_rd_get_default().append_uniforms(uniforms, SAMPLERS_BINDING_FIRST_INDEX);
 
 		volumetric_fog.base_uniform_set = RD::get_singleton()->uniform_set_create(uniforms, volumetric_fog.default_shader_rd, VolumetricFogShader::FogSet::FOG_SET_BASE);
 	}

servers/rendering/renderer_rd/environment/sky.cpp

@@ -552,7 +552,9 @@ void SkyRD::Sky::free() {
 RID SkyRD::Sky::get_textures(SkyTextureSetVersion p_version, RID p_default_shader_rd, Ref<RenderSceneBuffersRD> p_render_buffers) {
 	RendererRD::TextureStorage *texture_storage = RendererRD::TextureStorage::get_singleton();
 
-	Vector<RD::Uniform> uniforms;
+	thread_local LocalVector<RD::Uniform> uniforms;
+	uniforms.clear();
+
 	{
 		RD::Uniform u;
 		u.uniform_type = RD::UNIFORM_TYPE_TEXTURE;
@@ -873,7 +875,7 @@ void sky() {
 			uniforms.push_back(u);
 		}
 
-		uniforms.append_array(material_storage->samplers_rd_get_default().get_uniforms(SAMPLERS_BINDING_FIRST_INDEX));
+		material_storage->samplers_rd_get_default().append_uniforms(uniforms, SAMPLERS_BINDING_FIRST_INDEX);
 
 		sky_scene_state.uniform_set = RD::get_singleton()->uniform_set_create(uniforms, sky_shader.default_shader_rd, SKY_SET_UNIFORMS);
 	}

servers/rendering/renderer_rd/forward_clustered/render_forward_clustered.cpp

@@ -3139,7 +3139,8 @@ RID RenderForwardClustered::_setup_render_pass_uniform_set(RenderListType p_rend
 
 	//default render buffer and scene state uniform set
 
-	Vector<RD::Uniform> uniforms;
+	thread_local LocalVector<RD::Uniform> uniforms;
+	uniforms.clear();
 
 	{
 		RD::Uniform u;
@@ -3324,7 +3325,7 @@ RID RenderForwardClustered::_setup_render_pass_uniform_set(RenderListType p_rend
 		uniforms.push_back(u);
 	}
 
-	uniforms.append_array(p_samplers.get_uniforms(12));
+	p_samplers.append_uniforms(uniforms, 12);
 
 	{
 		RD::Uniform u;
@@ -3459,7 +3460,8 @@ RID RenderForwardClustered::_setup_render_pass_uniform_set(RenderListType p_rend
 
 RID RenderForwardClustered::_setup_sdfgi_render_pass_uniform_set(RID p_albedo_texture, RID p_emission_texture, RID p_emission_aniso_texture, RID p_geom_facing_texture, const RendererRD::MaterialStorage::Samplers &p_samplers) {
 	RendererRD::TextureStorage *texture_storage = RendererRD::TextureStorage::get_singleton();
-	Vector<RD::Uniform> uniforms;
+	thread_local LocalVector<RD::Uniform> uniforms;
+	uniforms.clear();
 
 	{
 		RD::Uniform u;
@@ -3623,7 +3625,7 @@ RID RenderForwardClustered::_setup_sdfgi_render_pass_uniform_set(RID p_albedo_te
 		uniforms.push_back(u);
 	}
 
-	uniforms.append_array(p_samplers.get_uniforms(12));
+	p_samplers.append_uniforms(uniforms, 12);
 
 	// actual sdfgi stuff
 

servers/rendering/renderer_rd/forward_mobile/render_forward_mobile.cpp

@@ -605,7 +605,7 @@ RID RenderForwardMobile::_setup_render_pass_uniform_set(RenderListType p_render_
 		uniforms.push_back(u);
 	}
 
-	uniforms.append_array(p_samplers.get_uniforms(13));
+	p_samplers.append_uniforms(uniforms, 13);
 
 	if (p_index >= (int)render_pass_uniform_sets.size()) {
 		render_pass_uniform_sets.resize(p_index + 1);

servers/rendering/renderer_rd/renderer_canvas_render_rd.cpp

@@ -455,7 +455,7 @@ RID RendererCanvasRenderRD::_create_base_uniform_set(RID p_to_render_target, boo
 		uniforms.push_back(u);
 	}
 
-	uniforms.append_array(material_storage->samplers_rd_get_default().get_uniforms(SAMPLERS_BINDING_FIRST_INDEX));
+	material_storage->samplers_rd_get_default().append_uniforms(uniforms, SAMPLERS_BINDING_FIRST_INDEX);
 
 	RID uniform_set = RD::get_singleton()->uniform_set_create(uniforms, shader.default_version_rd_shader, BASE_UNIFORM_SET);
 	if (p_backbuffer) {

servers/rendering/renderer_rd/storage_rd/material_storage.cpp

@@ -1177,24 +1177,25 @@ void MaterialStorage::MaterialData::set_as_used() {
 ///////////////////////////////////////////////////////////////////////////
 // MaterialStorage::Samplers
 
-Vector<RD::Uniform> MaterialStorage::Samplers::get_uniforms(int p_first_index) const {
-	Vector<RD::Uniform> uniforms;
+template void MaterialStorage::Samplers::append_uniforms(LocalVector<RD::Uniform> &p_uniforms, int p_first_index) const;
+
+template void MaterialStorage::Samplers::append_uniforms(Vector<RD::Uniform> &p_uniforms, int p_first_index) const;
 
+template <typename Collection>
+void MaterialStorage::Samplers::append_uniforms(Collection &p_uniforms, int p_first_index) const {
 	// Binding ids are aligned with samplers_inc.glsl.
-	uniforms.push_back(RD::Uniform(RD::UNIFORM_TYPE_SAMPLER, p_first_index + 0, rids[RS::CANVAS_ITEM_TEXTURE_FILTER_NEAREST][RS::CANVAS_ITEM_TEXTURE_REPEAT_DISABLED]));
-	uniforms.push_back(RD::Uniform(RD::UNIFORM_TYPE_SAMPLER, p_first_index + 1, rids[RS::CANVAS_ITEM_TEXTURE_FILTER_LINEAR][RS::CANVAS_ITEM_TEXTURE_REPEAT_DISABLED]));
-	uniforms.push_back(RD::Uniform(RD::UNIFORM_TYPE_SAMPLER, p_first_index + 2, rids[RS::CANVAS_ITEM_TEXTURE_FILTER_NEAREST_WITH_MIPMAPS][RS::CANVAS_ITEM_TEXTURE_REPEAT_DISABLED]));
-	uniforms.push_back(RD::Uniform(RD::UNIFORM_TYPE_SAMPLER, p_first_index + 3, rids[RS::CANVAS_ITEM_TEXTURE_FILTER_LINEAR_WITH_MIPMAPS][RS::CANVAS_ITEM_TEXTURE_REPEAT_DISABLED]));
-	uniforms.push_back(RD::Uniform(RD::UNIFORM_TYPE_SAMPLER, p_first_index + 4, rids[RS::CANVAS_ITEM_TEXTURE_FILTER_NEAREST_WITH_MIPMAPS_ANISOTROPIC][RS::CANVAS_ITEM_TEXTURE_REPEAT_DISABLED]));
-	uniforms.push_back(RD::Uniform(RD::UNIFORM_TYPE_SAMPLER, p_first_index + 5, rids[RS::CANVAS_ITEM_TEXTURE_FILTER_LINEAR_WITH_MIPMAPS_ANISOTROPIC][RS::CANVAS_ITEM_TEXTURE_REPEAT_DISABLED]));
-	uniforms.push_back(RD::Uniform(RD::UNIFORM_TYPE_SAMPLER, p_first_index + 6, rids[RS::CANVAS_ITEM_TEXTURE_FILTER_NEAREST][RS::CANVAS_ITEM_TEXTURE_REPEAT_ENABLED]));
-	uniforms.push_back(RD::Uniform(RD::UNIFORM_TYPE_SAMPLER, p_first_index + 7, rids[RS::CANVAS_ITEM_TEXTURE_FILTER_LINEAR][RS::CANVAS_ITEM_TEXTURE_REPEAT_ENABLED]));
-	uniforms.push_back(RD::Uniform(RD::UNIFORM_TYPE_SAMPLER, p_first_index + 8, rids[RS::CANVAS_ITEM_TEXTURE_FILTER_NEAREST_WITH_MIPMAPS][RS::CANVAS_ITEM_TEXTURE_REPEAT_ENABLED]));
-	uniforms.push_back(RD::Uniform(RD::UNIFORM_TYPE_SAMPLER, p_first_index + 9, rids[RS::CANVAS_ITEM_TEXTURE_FILTER_LINEAR_WITH_MIPMAPS][RS::CANVAS_ITEM_TEXTURE_REPEAT_ENABLED]));
-	uniforms.push_back(RD::Uniform(RD::UNIFORM_TYPE_SAMPLER, p_first_index + 10, rids[RS::CANVAS_ITEM_TEXTURE_FILTER_NEAREST_WITH_MIPMAPS_ANISOTROPIC][RS::CANVAS_ITEM_TEXTURE_REPEAT_ENABLED]));
-	uniforms.push_back(RD::Uniform(RD::UNIFORM_TYPE_SAMPLER, p_first_index + 11, rids[RS::CANVAS_ITEM_TEXTURE_FILTER_LINEAR_WITH_MIPMAPS_ANISOTROPIC][RS::CANVAS_ITEM_TEXTURE_REPEAT_ENABLED]));
-
-	return uniforms;
+	p_uniforms.push_back(RD::Uniform(RD::UNIFORM_TYPE_SAMPLER, p_first_index + 0, rids[RS::CANVAS_ITEM_TEXTURE_FILTER_NEAREST][RS::CANVAS_ITEM_TEXTURE_REPEAT_DISABLED]));
+	p_uniforms.push_back(RD::Uniform(RD::UNIFORM_TYPE_SAMPLER, p_first_index + 1, rids[RS::CANVAS_ITEM_TEXTURE_FILTER_LINEAR][RS::CANVAS_ITEM_TEXTURE_REPEAT_DISABLED]));
+	p_uniforms.push_back(RD::Uniform(RD::UNIFORM_TYPE_SAMPLER, p_first_index + 2, rids[RS::CANVAS_ITEM_TEXTURE_FILTER_NEAREST_WITH_MIPMAPS][RS::CANVAS_ITEM_TEXTURE_REPEAT_DISABLED]));
+	p_uniforms.push_back(RD::Uniform(RD::UNIFORM_TYPE_SAMPLER, p_first_index + 3, rids[RS::CANVAS_ITEM_TEXTURE_FILTER_LINEAR_WITH_MIPMAPS][RS::CANVAS_ITEM_TEXTURE_REPEAT_DISABLED]));
+	p_uniforms.push_back(RD::Uniform(RD::UNIFORM_TYPE_SAMPLER, p_first_index + 4, rids[RS::CANVAS_ITEM_TEXTURE_FILTER_NEAREST_WITH_MIPMAPS_ANISOTROPIC][RS::CANVAS_ITEM_TEXTURE_REPEAT_DISABLED]));
+	p_uniforms.push_back(RD::Uniform(RD::UNIFORM_TYPE_SAMPLER, p_first_index + 5, rids[RS::CANVAS_ITEM_TEXTURE_FILTER_LINEAR_WITH_MIPMAPS_ANISOTROPIC][RS::CANVAS_ITEM_TEXTURE_REPEAT_DISABLED]));
+	p_uniforms.push_back(RD::Uniform(RD::UNIFORM_TYPE_SAMPLER, p_first_index + 6, rids[RS::CANVAS_ITEM_TEXTURE_FILTER_NEAREST][RS::CANVAS_ITEM_TEXTURE_REPEAT_ENABLED]));
+	p_uniforms.push_back(RD::Uniform(RD::UNIFORM_TYPE_SAMPLER, p_first_index + 7, rids[RS::CANVAS_ITEM_TEXTURE_FILTER_LINEAR][RS::CANVAS_ITEM_TEXTURE_REPEAT_ENABLED]));
+	p_uniforms.push_back(RD::Uniform(RD::UNIFORM_TYPE_SAMPLER, p_first_index + 8, rids[RS::CANVAS_ITEM_TEXTURE_FILTER_NEAREST_WITH_MIPMAPS][RS::CANVAS_ITEM_TEXTURE_REPEAT_ENABLED]));
+	p_uniforms.push_back(RD::Uniform(RD::UNIFORM_TYPE_SAMPLER, p_first_index + 9, rids[RS::CANVAS_ITEM_TEXTURE_FILTER_LINEAR_WITH_MIPMAPS][RS::CANVAS_ITEM_TEXTURE_REPEAT_ENABLED]));
+	p_uniforms.push_back(RD::Uniform(RD::UNIFORM_TYPE_SAMPLER, p_first_index + 10, rids[RS::CANVAS_ITEM_TEXTURE_FILTER_NEAREST_WITH_MIPMAPS_ANISOTROPIC][RS::CANVAS_ITEM_TEXTURE_REPEAT_ENABLED]));
+	p_uniforms.push_back(RD::Uniform(RD::UNIFORM_TYPE_SAMPLER, p_first_index + 11, rids[RS::CANVAS_ITEM_TEXTURE_FILTER_LINEAR_WITH_MIPMAPS_ANISOTROPIC][RS::CANVAS_ITEM_TEXTURE_REPEAT_ENABLED]));
 }
 
 bool MaterialStorage::Samplers::is_valid() const {

servers/rendering/renderer_rd/storage_rd/material_storage.h

@@ -128,7 +128,8 @@ public:
 			return rids[p_filter][p_repeat];
 		}
 
-		Vector<RD::Uniform> get_uniforms(int p_first_index) const;
+		template <typename Collection>
+		void append_uniforms(Collection &p_uniforms, int p_first_index) const;
 		bool is_valid() const;
 		bool is_null() const;
 	};

servers/rendering/renderer_rd/storage_rd/particles_storage.cpp

@@ -155,7 +155,7 @@ void process() {
 			uniforms.push_back(u);
 		}
 
-		uniforms.append_array(material_storage->samplers_rd_get_default().get_uniforms(SAMPLERS_BINDING_FIRST_INDEX));
+		material_storage->samplers_rd_get_default().append_uniforms(uniforms, SAMPLERS_BINDING_FIRST_INDEX);
 
 		particles_shader.base_uniform_set = RD::get_singleton()->uniform_set_create(uniforms, particles_shader.default_shader_rd, BASE_UNIFORM_SET);
 	}

servers/rendering/renderer_rd/uniform_set_cache_rd.cpp

@@ -37,7 +37,8 @@ void UniformSetCacheRD::_bind_methods() {
 }
 
 RID UniformSetCacheRD::get_cache_array(RID p_shader, uint32_t p_set, const TypedArray<RDUniform> &p_uniforms) {
-	Vector<RD::Uniform> uniforms;
+	thread_local LocalVector<RD::Uniform> uniforms;
+	uniforms.clear();
 
 	for (int i = 0; i < p_uniforms.size(); i++) {
 		Ref<RDUniform> uniform = p_uniforms[i];

servers/rendering/renderer_rd/uniform_set_cache_rd.h

@@ -125,7 +125,8 @@ class UniformSetCacheRD : public Object {
 	void _invalidate(Cache *p_cache);
 	static void _uniform_set_invalidation_callback(void *p_userdata);
 
-	RID _allocate_from_uniforms(RID p_shader, uint32_t p_set, uint32_t p_hash, uint32_t p_table_idx, const Vector<RD::Uniform> &p_uniforms) {
+	template <typename Collection>
+	RID _allocate_from_uniforms(RID p_shader, uint32_t p_set, uint32_t p_hash, uint32_t p_table_idx, const Collection &p_uniforms) {
 		RID rid = RD::get_singleton()->uniform_set_create(p_uniforms, p_shader, p_set);
 		ERR_FAIL_COND_V(rid.is_null(), rid);
 
@@ -183,10 +184,10 @@ public:
 	}
 
 	template <typename... Args>
-	RID get_cache_vec(RID p_shader, uint32_t p_set, const Vector<RD::Uniform> &p_uniforms) {
+	RID get_cache_vec(RID p_shader, uint32_t p_set, const LocalVector<RD::Uniform> &p_uniforms) {
 		uint32_t h = hash_murmur3_one_64(p_shader.get_id());
 		h = hash_murmur3_one_32(p_set, h);
-		for (int i = 0; i < p_uniforms.size(); i++) {
+		for (uint32_t i = 0; i < p_uniforms.size(); i++) {
 			h = _hash_uniform(p_uniforms[i], h);
 		}
 
@@ -199,7 +200,7 @@ public:
 			while (c) {
 				if (c->hash == h && c->set == p_set && c->shader == p_shader && (uint32_t)p_uniforms.size() == c->uniforms.size()) {
 					bool all_ok = true;
-					for (int i = 0; i < p_uniforms.size(); i++) {
+					for (uint32_t i = 0; i < p_uniforms.size(); i++) {
 						if (!_compare_uniform(p_uniforms[i], c->uniforms[i])) {
 							all_ok = false;
 							break;

servers/rendering/rendering_device.cpp

@@ -3081,7 +3081,12 @@ void RenderingDevice::_uniform_set_update_shared(UniformSet *p_uniform_set) {
 	}
 }
 
-RID RenderingDevice::uniform_set_create(const Vector<Uniform> &p_uniforms, RID p_shader, uint32_t p_shader_set) {
+template RID RenderingDevice::uniform_set_create(const LocalVector<RD::Uniform> &p_uniforms, RID p_shader, uint32_t p_shader_set);
+
+template RID RenderingDevice::uniform_set_create(const Vector<RD::Uniform> &p_uniforms, RID p_shader, uint32_t p_shader_set);
+
+template <typename Collection>
+RID RenderingDevice::uniform_set_create(const Collection &p_uniforms, RID p_shader, uint32_t p_shader_set) {
 	_THREAD_SAFE_METHOD_
 
 	ERR_FAIL_COND_V(p_uniforms.is_empty(), RID());

servers/rendering/rendering_device.h

@@ -963,7 +963,8 @@ private:
 	void _uniform_set_update_shared(UniformSet *p_uniform_set);
 
 public:
-	RID uniform_set_create(const Vector<Uniform> &p_uniforms, RID p_shader, uint32_t p_shader_set);
+	template <typename Collection>
+	RID uniform_set_create(const Collection &p_uniforms, RID p_shader, uint32_t p_shader_set);
 	bool uniform_set_is_valid(RID p_uniform_set);
 	void uniform_set_set_invalidation_callback(RID p_uniform_set, InvalidationCallback p_callback, void *p_userdata);