Merge pull request #109971 from clayjohn/hdr2d-mobile

Use half float precision buffer for 3D when HDR2D is enabled

Author: Repiteo

servers/rendering/renderer_rd/forward_clustered/render_forward_clustered.cpp

@@ -689,7 +689,9 @@ void RenderForwardClustered::_setup_environment(const RenderDataRD *p_render_dat
 		}
 	}
 
-	p_render_data->scene_data->update_ubo(scene_state.uniform_buffers[p_index], get_debug_draw_mode(), env, reflection_probe_instance, p_render_data->camera_attributes, p_pancake_shadows, p_screen_size, p_default_bg_color, _render_buffers_get_luminance_multiplier(), p_opaque_render_buffers, p_apply_alpha_multiplier);
+	float luminance_multiplier = rd.is_valid() ? rd->get_luminance_multiplier() : 1.0;
+
+	p_render_data->scene_data->update_ubo(scene_state.uniform_buffers[p_index], get_debug_draw_mode(), env, reflection_probe_instance, p_render_data->camera_attributes, p_pancake_shadows, p_screen_size, p_default_bg_color, luminance_multiplier, p_opaque_render_buffers, p_apply_alpha_multiplier);
 
 	// now do implementation UBO
 
@@ -1434,7 +1436,7 @@ void RenderForwardClustered::_process_ssr(Ref<RenderSceneBuffersRD> p_render_buf
 
 	RENDER_TIMESTAMP("Process SSR");
 
-	ss_effects->ssr_allocate_buffers(p_render_buffers, rb_data->ss_effects_data.ssr, _render_buffers_get_color_format());
+	ss_effects->ssr_allocate_buffers(p_render_buffers, rb_data->ss_effects_data.ssr, p_render_buffers->get_base_data_format());
 
 	Projection reprojections[RendererSceneRender::MAX_RENDER_VIEWS];
 
@@ -1946,7 +1948,7 @@ void RenderForwardClustered::_render_scene(RenderDataRD *p_render_data, const Co
 	bool draw_sky = false;
 	bool draw_sky_fog_only = false;
 	// We invert luminance_multiplier for sky so that we can combine it with exposure value.
-	float sky_luminance_multiplier = 1.0 / _render_buffers_get_luminance_multiplier();
+	float sky_luminance_multiplier = 1.0 / rb->get_luminance_multiplier();
 	float sky_brightness_multiplier = 1.0;
 
 	Color clear_color;
@@ -2472,7 +2474,7 @@ void RenderForwardClustered::_render_scene(RenderDataRD *p_render_data, const Co
 		} else if (using_taa) {
 			RD::get_singleton()->draw_command_begin_label("TAA");
 			RENDER_TIMESTAMP("TAA");
-			taa->process(rb, _render_buffers_get_color_format(), p_render_data->scene_data->z_near, p_render_data->scene_data->z_far);
+			taa->process(rb, rb->get_base_data_format(), p_render_data->scene_data->z_near, p_render_data->scene_data->z_far);
 			RD::get_singleton()->draw_command_end_label();
 		}
 	}
@@ -4505,7 +4507,7 @@ void RenderForwardClustered::_mesh_compile_pipelines_for_surface(const SurfacePi
 
 	// Retrieve from the scene shader which groups are currently enabled.
 	const bool multiview_enabled = p_global.use_multiview && scene_shader.is_multiview_shader_group_enabled();
-	const RD::DataFormat buffers_color_format = _render_buffers_get_color_format();
+	const RD::DataFormat buffers_color_format = _render_buffers_get_preferred_color_format();
 	const bool buffers_can_be_storage = _render_buffers_can_be_storage();
 
 	// Set the attributes common to all pipelines.

servers/rendering/renderer_rd/forward_mobile/render_forward_mobile.cpp

@@ -414,19 +414,12 @@ void RenderForwardMobile::update() {
 
 /* Render functions */
 
-float RenderForwardMobile::_render_buffers_get_luminance_multiplier() {
-	// On mobile renderer we need to multiply source colors by 2 due to using a UNORM buffer
-	// and multiplying by the output color during 3D rendering by 0.5
-	return 2.0;
-}
-
-RD::DataFormat RenderForwardMobile::_render_buffers_get_color_format() {
+RD::DataFormat RenderForwardMobile::_render_buffers_get_preferred_color_format() {
 	// Using 32bit buffers enables AFBC on mobile devices which should have a definite performance improvement (MALI G710 and newer support this on 64bit RTs)
 	return RD::DATA_FORMAT_A2B10G10R10_UNORM_PACK32;
 }
 
 bool RenderForwardMobile::_render_buffers_can_be_storage() {
-	// Using 32bit buffers enables AFBC on mobile devices which should have a definite performance improvement (MALI G710 and newer support this on 64bit RTs)
 	// Doesn't support storage
 	return false;
 }
@@ -1002,8 +995,8 @@ void RenderForwardMobile::_render_scene(RenderDataRD *p_render_data, const Color
 	bool draw_sky = false;
 	bool draw_sky_fog_only = false;
 	// We invert luminance_multiplier for sky so that we can combine it with exposure value.
-	float inverse_luminance_multiplier = 1.0 / _render_buffers_get_luminance_multiplier();
-	float sky_luminance_multiplier = inverse_luminance_multiplier;
+	float inverse_luminance_multiplier = 1.0 / rb->get_luminance_multiplier();
+	float sky_luminance_multiplier = 1.0 / 2.0; // Hardcoded since sky always uses LDR in the mobile renderer
 	float sky_brightness_multiplier = 1.0;
 
 	Color clear_color = p_default_bg_color;
@@ -1122,6 +1115,7 @@ void RenderForwardMobile::_render_scene(RenderDataRD *p_render_data, const Color
 		base_specialization.scene_use_ambient_cubemap = use_ambient_cubemap;
 		base_specialization.scene_use_reflection_cubemap = use_reflection_cubemap;
 		base_specialization.scene_roughness_limiter_enabled = p_render_data->render_buffers.is_valid() && screen_space_roughness_limiter_is_active();
+		base_specialization.luminance_multiplier = p_render_data->render_buffers.is_valid() ? p_render_data->render_buffers->get_luminance_multiplier() : 1.0;
 	}
 
 	{
@@ -2201,7 +2195,9 @@ void RenderForwardMobile::_setup_environment(const RenderDataRD *p_render_data,
 		}
 	}
 
-	p_render_data->scene_data->update_ubo(scene_state.uniform_buffers[p_index], get_debug_draw_mode(), env, reflection_probe_instance, p_render_data->camera_attributes, p_pancake_shadows, p_screen_size, p_default_bg_color, _render_buffers_get_luminance_multiplier(), p_opaque_render_buffers, false);
+	float luminance_multiplier = p_render_data->render_buffers.is_valid() ? p_render_data->render_buffers->get_luminance_multiplier() : 1.0;
+
+	p_render_data->scene_data->update_ubo(scene_state.uniform_buffers[p_index], get_debug_draw_mode(), env, reflection_probe_instance, p_render_data->camera_attributes, p_pancake_shadows, p_screen_size, p_default_bg_color, luminance_multiplier, p_opaque_render_buffers, false);
 }
 
 /// RENDERING ///
@@ -3141,7 +3137,6 @@ void RenderForwardMobile::_mesh_compile_pipelines_for_surface(const SurfacePipel
 	pipeline_key.wireframe = false;
 
 	const bool multiview_enabled = p_global.use_multiview && scene_shader.is_multiview_shader_group_enabled();
-	const RD::DataFormat buffers_color_format = _render_buffers_get_color_format();
 	const bool buffers_can_be_storage = _render_buffers_can_be_storage();
 	const uint32_t vrs_iterations = p_global.use_vrs ? 2 : 1;
 
@@ -3155,6 +3150,7 @@ void RenderForwardMobile::_mesh_compile_pipelines_for_surface(const SurfacePipel
 		for (uint32_t use_post_pass = post_pass_start; use_post_pass < post_pass_iterations; use_post_pass++) {
 			const uint32_t hdr_iterations = use_post_pass ? hdr_target_iterations : (hdr_start + 1);
 			for (uint32_t use_hdr = hdr_start; use_hdr < hdr_iterations; use_hdr++) {
+				const RD::DataFormat buffers_color_format = use_hdr ? RD::DATA_FORMAT_R16G16B16A16_SFLOAT : _render_buffers_get_preferred_color_format();
 				pipeline_key.version = SceneShaderForwardMobile::SHADER_VERSION_COLOR_PASS;
 				pipeline_key.framebuffer_format_id = _get_color_framebuffer_format_for_pipeline(buffers_color_format, buffers_can_be_storage, RD::TextureSamples(p_global.texture_samples), RD::TextureSamples(p_global.target_samples), use_vrs, use_post_pass, use_hdr, 1);
 				_mesh_compile_pipeline_for_surface(p_surface.shader, p_surface.mesh_surface, p_surface.instanced, p_source, pipeline_key, r_pipeline_pairs);
@@ -3340,7 +3336,7 @@ void RenderForwardMobile::_update_shader_quality_settings() {
 RenderForwardMobile::RenderForwardMobile() {
 	singleton = this;
 
-	sky.set_texture_format(_render_buffers_get_color_format());
+	sky.set_texture_format(_render_buffers_get_preferred_color_format());
 
 	String defines;
 

servers/rendering/renderer_rd/forward_mobile/render_forward_mobile.h

@@ -413,8 +413,7 @@ class RenderForwardMobile : public RendererSceneRenderRD {
 	/* setup */
 	virtual void _update_shader_quality_settings() override;
 
-	virtual float _render_buffers_get_luminance_multiplier() override;
-	virtual RD::DataFormat _render_buffers_get_color_format() override;
+	virtual RD::DataFormat _render_buffers_get_preferred_color_format() override;
 	virtual bool _render_buffers_can_be_storage() override;
 
 	virtual RID _render_buffers_get_normal_texture(Ref<RenderSceneBuffersRD> p_render_buffers) override;

servers/rendering/renderer_rd/renderer_scene_render_rd.cpp

@@ -242,7 +242,7 @@ Ref<RenderSceneBuffers> RendererSceneRenderRD::render_buffers_create() {
 
 	rb->set_can_be_storage(_render_buffers_can_be_storage());
 	rb->set_max_cluster_elements(max_cluster_elements);
-	rb->set_base_data_format(_render_buffers_get_color_format());
+	rb->set_preferred_data_format(_render_buffers_get_preferred_color_format());
 	if (vrs) {
 		rb->set_vrs(vrs);
 	}
@@ -587,7 +587,7 @@ void RendererSceneRenderRD::_render_buffers_post_process_and_tonemap(const Rende
 			}
 		}
 
-		float luminance_multiplier = _render_buffers_get_luminance_multiplier();
+		float luminance_multiplier = rb->get_luminance_multiplier();
 		for (uint32_t l = 0; l < rb->get_view_count(); l++) {
 			for (int i = 0; i < (max_glow_level + 1); i++) {
 				Size2i vp_size = rb->get_texture_slice_size(RB_SCOPE_BUFFERS, RB_TEX_BLUR_1, i);
@@ -693,7 +693,7 @@ void RendererSceneRenderRD::_render_buffers_post_process_and_tonemap(const Rende
 			}
 		}
 
-		tonemap.luminance_multiplier = _render_buffers_get_luminance_multiplier();
+		tonemap.luminance_multiplier = rb->get_luminance_multiplier();
 		tonemap.view_count = rb->get_view_count();
 
 		RID dest_fb;
@@ -702,7 +702,7 @@ void RendererSceneRenderRD::_render_buffers_post_process_and_tonemap(const Rende
 		if (spatial_upscaler != nullptr || use_smaa) {
 			// If we use a spatial upscaler to upscale or SMAA to antialias we need to write our result into an intermediate buffer.
 			// Note that this is cached so we only create the texture the first time.
-			dest_fb_format = _render_buffers_get_color_format();
+			dest_fb_format = rb->get_base_data_format();
 			RID dest_texture = rb->create_texture(SNAME("Tonemapper"), SNAME("destination"), dest_fb_format, RD::TEXTURE_USAGE_SAMPLING_BIT | RD::TEXTURE_USAGE_STORAGE_BIT | RD::TEXTURE_USAGE_COLOR_ATTACHMENT_BIT, RD::TEXTURE_SAMPLES_1, Size2i(), 0, 1, true, true);
 			dest_fb = FramebufferCacheRD::get_singleton()->get_cache(dest_texture);
 			if (use_smaa) {
@@ -758,7 +758,7 @@ void RendererSceneRenderRD::_render_buffers_post_process_and_tonemap(const Rende
 		bool using_hdr = texture_storage->render_target_is_using_hdr(render_target);
 		RID dest_fb;
 		if (spatial_upscaler) {
-			rb->create_texture(SNAME("SMAA"), SNAME("destination"), _render_buffers_get_color_format(), RD::TEXTURE_USAGE_SAMPLING_BIT | RD::TEXTURE_USAGE_STORAGE_BIT | RD::TEXTURE_USAGE_COLOR_ATTACHMENT_BIT, RD::TEXTURE_SAMPLES_1, Size2i(), 0, 1, true, true);
+			rb->create_texture(SNAME("SMAA"), SNAME("destination"), rb->get_base_data_format(), RD::TEXTURE_USAGE_SAMPLING_BIT | RD::TEXTURE_USAGE_STORAGE_BIT | RD::TEXTURE_USAGE_COLOR_ATTACHMENT_BIT, RD::TEXTURE_SAMPLES_1, Size2i(), 0, 1, true, true);
 		}
 		if (rb->get_view_count() > 1) {
 			for (uint32_t v = 0; v < rb->get_view_count(); v++) {
@@ -802,7 +802,7 @@ void RendererSceneRenderRD::_render_buffers_post_process_and_tonemap(const Rende
 			RD::DataFormat format_for_debanding;
 
 			if (spatial_upscaler) {
-				RID dest_texture = rb->create_texture(SNAME("SMAA"), SNAME("destination"), _render_buffers_get_color_format(), RD::TEXTURE_USAGE_SAMPLING_BIT | RD::TEXTURE_USAGE_STORAGE_BIT | RD::TEXTURE_USAGE_COLOR_ATTACHMENT_BIT, RD::TEXTURE_SAMPLES_1, Size2i(), 0, 1, true, true);
+				RID dest_texture = rb->create_texture(SNAME("SMAA"), SNAME("destination"), rb->get_base_data_format(), RD::TEXTURE_USAGE_SAMPLING_BIT | RD::TEXTURE_USAGE_STORAGE_BIT | RD::TEXTURE_USAGE_COLOR_ATTACHMENT_BIT, RD::TEXTURE_SAMPLES_1, Size2i(), 0, 1, true, true);
 				dest_fb = FramebufferCacheRD::get_singleton()->get_cache(dest_texture);
 				// Debanding is currently not supported when using spatial upscaling, so apply it before scaling.
 				// This produces suboptimal results because the image will be modified by spatial upscaling after
@@ -935,7 +935,7 @@ void RendererSceneRenderRD::_post_process_subpass(RID p_source_texture, RID p_fr
 
 	tonemap.texture_size = Vector2i(target_size.x, target_size.y);
 
-	tonemap.luminance_multiplier = _render_buffers_get_luminance_multiplier();
+	tonemap.luminance_multiplier = rb->get_luminance_multiplier();
 	tonemap.view_count = rb->get_view_count();
 
 	if (rb->get_use_debanding()) {
@@ -1115,11 +1115,7 @@ RID RendererSceneRenderRD::render_buffers_get_default_voxel_gi_buffer() {
 	return gi.default_voxel_gi_buffer;
 }
 
-float RendererSceneRenderRD::_render_buffers_get_luminance_multiplier() {
-	return 1.0;
-}
-
-RD::DataFormat RendererSceneRenderRD::_render_buffers_get_color_format() {
+RD::DataFormat RendererSceneRenderRD::_render_buffers_get_preferred_color_format() {
 	return RD::DATA_FORMAT_R16G16B16A16_SFLOAT;
 }
 

servers/rendering/renderer_rd/renderer_scene_render_rd.h

@@ -244,9 +244,7 @@ class RendererSceneRenderRD : public RendererSceneRender, public RenderingShader
 	virtual void voxel_gi_set_quality(RS::VoxelGIQuality p_quality) override { gi.voxel_gi_quality = p_quality; }
 
 	/* render buffers */
-
-	virtual float _render_buffers_get_luminance_multiplier();
-	virtual RD::DataFormat _render_buffers_get_color_format();
+	virtual RD::DataFormat _render_buffers_get_preferred_color_format();
 	virtual bool _render_buffers_can_be_storage();
 	virtual Ref<RenderSceneBuffers> render_buffers_create() override;
 	virtual void gi_set_use_half_resolution(bool p_enable) override;

servers/rendering/renderer_rd/shaders/forward_clustered/scene_forward_clustered_inc.glsl

@@ -142,6 +142,8 @@ float sc_luminance_multiplier() {
 	return 1.0;
 }
 
+#define REFLECTION_MULTIPLIER 1.0
+
 #define SDFGI_MAX_CASCADES 8
 
 /* Set 0: Base Pass (never changes) */

servers/rendering/renderer_rd/shaders/forward_mobile/scene_forward_mobile.glsl

@@ -1563,7 +1563,7 @@ void main() {
 		indirect_specular_light = hvec3(textureLod(samplerCube(radiance_cubemap, DEFAULT_SAMPLER_LINEAR_WITH_MIPMAPS_CLAMP), vec3(ref_vec), lod).rgb);
 
 #endif //USE_RADIANCE_CUBEMAP_ARRAY
-		indirect_specular_light *= sc_luminance_multiplier();
+		indirect_specular_light *= REFLECTION_MULTIPLIER;
 		indirect_specular_light *= half(scene_data.IBL_exposure_normalization);
 		indirect_specular_light *= horizon * horizon;
 		indirect_specular_light *= half(scene_data.ambient_light_color_energy.a);
@@ -1585,7 +1585,7 @@ void main() {
 #else
 			hvec3 cubemap_ambient = hvec3(textureLod(samplerCube(radiance_cubemap, DEFAULT_SAMPLER_LINEAR_WITH_MIPMAPS_CLAMP), ambient_dir, MAX_ROUGHNESS_LOD).rgb);
 #endif //USE_RADIANCE_CUBEMAP_ARRAY
-			cubemap_ambient *= sc_luminance_multiplier();
+			cubemap_ambient *= REFLECTION_MULTIPLIER;
 			cubemap_ambient *= half(scene_data.IBL_exposure_normalization);
 			ambient_light = mix(ambient_light, cubemap_ambient * half(scene_data.ambient_light_color_energy.a), half(scene_data.ambient_color_sky_mix));
 		}

servers/rendering/renderer_rd/shaders/forward_mobile/scene_forward_mobile_inc.glsl

@@ -215,6 +215,10 @@ half sc_luminance_multiplier() {
 	return half(sc_packed_2());
 }
 
+// Like the luminance multiplier, but it is only for sky and reflection probes
+// since they are always LDR.
+#define REFLECTION_MULTIPLIER half(2.0)
+
 /* Set 0: Base Pass (never changes) */
 
 #include "../light_data_inc.glsl"

servers/rendering/renderer_rd/shaders/scene_forward_lights_inc.glsl

@@ -957,7 +957,7 @@ void reflection_process(uint ref_index, vec3 vertex, hvec3 ref_vec, hvec3 normal
 		hvec4 reflection;
 		half reflection_blend = max(half(0.0), blend - reflection_accum.a);
 
-		reflection.rgb = hvec3(textureLod(samplerCubeArray(reflection_atlas, DEFAULT_SAMPLER_LINEAR_WITH_MIPMAPS_CLAMP), vec4(local_ref_vec, reflections.data[ref_index].index), sqrt(roughness) * MAX_ROUGHNESS_LOD).rgb) * sc_luminance_multiplier();
+		reflection.rgb = hvec3(textureLod(samplerCubeArray(reflection_atlas, DEFAULT_SAMPLER_LINEAR_WITH_MIPMAPS_CLAMP), vec4(local_ref_vec, reflections.data[ref_index].index), sqrt(roughness) * MAX_ROUGHNESS_LOD).rgb) * REFLECTION_MULTIPLIER;
 		reflection.rgb *= half(reflections.data[ref_index].exposure_normalization);
 		reflection.a = reflection_blend;
 

servers/rendering/renderer_rd/storage_rd/light_storage.cpp

@@ -1542,7 +1542,7 @@ bool LightStorage::reflection_probe_instance_begin_render(RID p_instance, RID p_
 		}
 		atlas->reflections.resize(atlas->count);
 		for (int i = 0; i < atlas->count; i++) {
-			atlas->reflections.write[i].data.update_reflection_data(atlas->size, mipmaps, false, atlas->reflection, i * 6, update_always, RendererSceneRenderRD::get_singleton()->get_sky()->roughness_layers, RendererSceneRenderRD::get_singleton()->_render_buffers_get_color_format());
+			atlas->reflections.write[i].data.update_reflection_data(atlas->size, mipmaps, false, atlas->reflection, i * 6, update_always, RendererSceneRenderRD::get_singleton()->get_sky()->roughness_layers, RendererSceneRenderRD::get_singleton()->_render_buffers_get_preferred_color_format());
 			for (int j = 0; j < 6; j++) {
 				atlas->reflections.write[i].fbs[j] = RendererSceneRenderRD::get_singleton()->reflection_probe_create_framebuffer(atlas->reflections.write[i].data.layers[0].mipmaps[0].views[j], atlas->depth_buffer);
 			}
@@ -1814,7 +1814,7 @@ void LightStorage::update_reflection_probe_buffer(RenderDataRD *p_render_data, c
 }
 
 RD::DataFormat LightStorage::get_reflection_probe_color_format() {
-	return RendererSceneRenderRD::get_singleton()->_render_buffers_get_color_format();
+	return RendererSceneRenderRD::get_singleton()->_render_buffers_get_preferred_color_format();
 }
 
 uint32_t LightStorage::get_reflection_probe_color_usage_bits() {