Merge pull request #36588 from clayjohn/VULKAN-env-filter

Basic fast filtering implementation

doc/classes/ProjectSettings.xml

@@ -1037,23 +1037,28 @@
 			Lower-end override for [member rendering/quality/intended_usage/framebuffer_allocation] on mobile devices, due to performance concerns or driver support.
 		</member>
 		<member name="rendering/quality/reflection_atlas/reflection_count" type="int" setter="" getter="" default="64">
+			Number of cubemaps to store in the reflection atlas. The number of [ReflectionProbe]s in a scene will be limited by this amount. A higher number requires more VRAM.
 		</member>
-		<member name="rendering/quality/reflection_atlas/reflection_size" type="int" setter="" getter="" default="256">
+		<member name="rendering/quality/reflection_atlas/reflection_size" type="int" setter="" getter="" default="128">
+			Size of cubemap faces for [ReflectionProbe]s. A higher number requires more VRAM and may make reflection probe updating slower.
 		</member>
 		<member name="rendering/quality/reflection_atlas/reflection_size.mobile" type="int" setter="" getter="" default="128">
+			Lower-end override for [member rendering/quality/reflection_atlas/reflection_size] on mobile devices, due to performance concerns or driver support.
+		</member>
+		<member name="rendering/quality/reflections/fast_filter_high_quality" type="bool" setter="" getter="" default="false">
+			Use a higher quality variant of the fast filtering algorithm. Significantly slower than using default quality, but results in smoother reflections. Should only be used when the scene is especially detailed.
 		</member>
 		<member name="rendering/quality/reflections/ggx_samples" type="int" setter="" getter="" default="1024">
+			Sets the number of samples to take when using importance sampling for [Sky]s and [ReflectionProbe]s. A higher value will result in smoother, higher quality reflections, but increases time to calculate radiance maps. In general, fewer samples are needed for simpler, low dynamic range environments while more samples are needed for HDR environments and environments with a high level of detail.
 		</member>
 		<member name="rendering/quality/reflections/ggx_samples.mobile" type="int" setter="" getter="" default="128">
-		</member>
+			Lower-end override for [member rendering/quality/reflections/ggx_samples] on mobile devices, due to performance concerns or driver support.
-		<member name="rendering/quality/reflections/ggx_samples_realtime" type="int" setter="" getter="" default="64">
-		</member>
-		<member name="rendering/quality/reflections/ggx_samples_realtime.mobile" type="int" setter="" getter="" default="16">
 		</member>
 		<member name="rendering/quality/reflections/roughness_layers" type="int" setter="" getter="" default="6">
+			Limits the number of layers to use in radiance maps when using importance sampling. A lower number will be slightly faster and take up less VRAM.
 		</member>
 		<member name="rendering/quality/reflections/texture_array_reflections" type="bool" setter="" getter="" default="true">
-			If [code]true[/code], uses texture arrays instead of mipmaps for reflection probes and panorama backgrounds (sky). This reduces jitter noise on reflections, but costs more performance and memory.
+			If [code]true[/code], uses texture arrays instead of mipmaps for reflection probes and panorama backgrounds (sky). This reduces jitter noise and upscaling artifacts on reflections, but is significantly slower to compute and uses [member rendering/quality/reflections/roughness_layers] times more memory.
 		</member>
 		<member name="rendering/quality/reflections/texture_array_reflections.mobile" type="bool" setter="" getter="" default="false">
 			Lower-end override for [member rendering/quality/reflections/texture_array_reflections] on mobile devices, due to performance concerns or driver support.

doc/classes/ReflectionProbe.xml

@@ -4,7 +4,7 @@
 		Captures its surroundings to create reflections.
 	</brief_description>
 	<description>
-		Capture its surroundings as a dual parabolid image, and stores versions of it with increasing levels of blur to simulate different material roughnesses.
+		Captures its surroundings as a cubemap, and stores versions of it with increasing levels of blur to simulate different material roughnesses.
 		The [ReflectionProbe] is used to create high-quality reflections at the cost of performance. It can be combined with [GIProbe]s and Screen Space Reflections to achieve high quality reflections. [ReflectionProbe]s render all objects within their [member cull_mask], so updating them can be quite expensive. It is best to update them once with the important static objects and then leave them.
 	</description>
 	<tutorials>
@@ -52,7 +52,7 @@
 	</members>
 	<constants>
 		<constant name="UPDATE_ONCE" value="0" enum="UpdateMode">
-			Update the probe once on the next frame.
+			Update the probe once on the next frame. The corresponding radiance map will be generated over the following six frames. This is slower to update than [constant UPDATE_ALWAYS] but can result in higher quality reflections.
 		</constant>
 		<constant name="UPDATE_ALWAYS" value="1" enum="UpdateMode">
 			Update the probe every frame. This is needed when you want to capture dynamic objects. However, it results in an increased render time. Use [constant UPDATE_ONCE] whenever possible.

doc/classes/Sky.xml

@@ -12,6 +12,7 @@
 	</methods>
 	<members>
 		<member name="process_mode" type="int" setter="set_process_mode" getter="get_process_mode" enum="Sky.ProcessMode" default="0">
+			Sets the method for generating the radiance map from the sky. The radiance map is a cubemap with increasingly blurry versions of the sky corresponding to different levels of roughness. Radiance maps can be expensive to calculate. See [enum ProcessMode] for options.
 		</member>
 		<member name="radiance_size" type="int" setter="set_radiance_size" getter="get_radiance_size" enum="Sky.RadianceSize" default="2">
 			The [Sky]'s radiance map size. The higher the radiance map size, the more detailed the lighting from the [Sky] will be.
@@ -45,8 +46,11 @@
 			Represents the size of the [enum RadianceSize] enum.
 		</constant>
 		<constant name="PROCESS_MODE_QUALITY" value="0" enum="ProcessMode">
+			Uses high quality importance sampling to process the radiance map. In general, this results in much higher quality than [constant PROCESS_MODE_REALTIME] but takes much longer to generate. This should not be used if you plan on changing the sky at runtime.
 		</constant>
 		<constant name="PROCESS_MODE_REALTIME" value="1" enum="ProcessMode">
+			Uses the fast filtering algorithm to process the radiance map. In general this results in lower quality, but substantially faster run times.
+			[b]Note:[/b] The fast filtering algorithm is limited to 128x128 cubemaps, so [member radiance_size] must be set to [constant RADIANCE_SIZE_128].
 		</constant>
 	</constants>
 </class>

servers/visual/rasterizer_rd/rasterizer_effects_rd.cpp

@@ -30,6 +30,8 @@
 
 #include "rasterizer_effects_rd.h"
 #include "core/os/os.h"
+#include "core/project_settings.h"
+#include "cubemap_coeffs.h"
 
 static _FORCE_INLINE_ void store_transform_3x3(const Basis &p_basis, float *p_array) {
 	p_array[0] = p_basis.elements[0][0];
@@ -750,6 +752,44 @@ void RasterizerEffectsRD::roughness_limit(RID p_source_normal, RID p_roughness,
 	RD::get_singleton()->compute_list_end();
 }
 
+void RasterizerEffectsRD::cubemap_downsample(RID p_source_cubemap, bool p_source_is_panorama, RID p_dest_cubemap, const Size2i &p_size) {
+
+	cubemap_downsampler.push_constant.face_size = p_size.x;
+
+	RD::ComputeListID compute_list = RD::get_singleton()->compute_list_begin();
+	RD::get_singleton()->compute_list_bind_compute_pipeline(compute_list, cubemap_downsampler.pipelines[p_source_is_panorama ? CUBEMAP_DOWNSAMPLER_SOURCE_PANORAMA : CUBEMAP_DOWNSAMPLER_SOURCE_CUBEMAP]);
+	RD::get_singleton()->compute_list_bind_uniform_set(compute_list, _get_compute_uniform_set_from_texture(p_source_cubemap), 0);
+	RD::get_singleton()->compute_list_bind_uniform_set(compute_list, _get_uniform_set_from_image(p_dest_cubemap), 1);
+
+	int x_groups = (p_size.x - 1) / 8 + 1;
+	int y_groups = (p_size.y - 1) / 8 + 1;
+
+	RD::get_singleton()->compute_list_set_push_constant(compute_list, &cubemap_downsampler.push_constant, sizeof(CubemapDownsamplerPushConstant));
+
+	RD::get_singleton()->compute_list_dispatch(compute_list, x_groups, y_groups, 6); // one z_group for each face
+
+	RD::get_singleton()->compute_list_end();
+}
+
+void RasterizerEffectsRD::cubemap_filter(RID p_source_cubemap, Vector<RID> p_dest_cubemap, bool p_use_array) {
+
+	int pipeline = p_use_array ? FILTER_MODE_HIGH_QUALITY_ARRAY : FILTER_MODE_HIGH_QUALITY;
+	pipeline = filter.use_high_quality ? pipeline : pipeline + 1;
+	RD::ComputeListID compute_list = RD::get_singleton()->compute_list_begin();
+	RD::get_singleton()->compute_list_bind_compute_pipeline(compute_list, filter.pipelines[pipeline]);
+	RD::get_singleton()->compute_list_bind_uniform_set(compute_list, _get_compute_uniform_set_from_texture(p_source_cubemap, true), 0);
+	RD::get_singleton()->compute_list_bind_uniform_set(compute_list, filter.uniform_set, 1);
+	for (int i = 0; i < p_dest_cubemap.size(); i++) {
+		RD::get_singleton()->compute_list_bind_uniform_set(compute_list, _get_uniform_set_from_image(p_dest_cubemap[i]), i + 2);
+	}
+
+	int x_groups = p_use_array ? 1792 : 342; // (128 * 128 * 7) / 64 : (128*128 + 64*64 + 32*32 + 16*16 + 8*8 + 4*4 + 2*2) / 64
+
+	RD::get_singleton()->compute_list_dispatch(compute_list, x_groups, 6, 1); // one y_group for each face
+
+	RD::get_singleton()->compute_list_end();
+}
+
 RasterizerEffectsRD::RasterizerEffectsRD() {
 
 	{
@@ -930,7 +970,7 @@ RasterizerEffectsRD::RasterizerEffectsRD() {
 	}
 
 	{
-		// Initialize copier
+		// Initialize roughness limiter
 		Vector<String> shader_modes;
 		shader_modes.push_back("");
 
@@ -941,6 +981,55 @@ RasterizerEffectsRD::RasterizerEffectsRD() {
 		roughness_limiter.pipeline = RD::get_singleton()->compute_pipeline_create(roughness_limiter.shader.version_get_shader(roughness_limiter.shader_version, 0));
 	}
 
+	{
+		//Initialize cubemap downsampler
+		Vector<String> cubemap_downsampler_modes;
+		cubemap_downsampler_modes.push_back("\n#define MODE_SOURCE_PANORAMA\n");
+		cubemap_downsampler_modes.push_back("\n#define MODE_SOURCE_CUBEMAP\n");
+		cubemap_downsampler.shader.initialize(cubemap_downsampler_modes);
+
+		cubemap_downsampler.shader_version = cubemap_downsampler.shader.version_create();
+
+		for (int i = 0; i < CUBEMAP_DOWNSAMPLER_SOURCE_MAX; i++) {
+			cubemap_downsampler.pipelines[i] = RD::get_singleton()->compute_pipeline_create(cubemap_downsampler.shader.version_get_shader(cubemap_downsampler.shader_version, i));
+		}
+	}
+
+	{
+		// Initialize cubemap filter
+		filter.use_high_quality = GLOBAL_GET("rendering/quality/reflections/fast_filter_high_quality");
+
+		Vector<String> cubemap_filter_modes;
+		cubemap_filter_modes.push_back("\n#define USE_HIGH_QUALITY\n");
+		cubemap_filter_modes.push_back("\n#define USE_LOW_QUALITY\n");
+		cubemap_filter_modes.push_back("\n#define USE_HIGH_QUALITY\n#define USE_TEXTURE_ARRAY\n");
+		cubemap_filter_modes.push_back("\n#define USE_LOW_QUALITY\n#define USE_TEXTURE_ARRAY\n");
+		filter.shader.initialize(cubemap_filter_modes);
+		filter.shader_version = filter.shader.version_create();
+
+		for (int i = 0; i < FILTER_MODE_MAX; i++) {
+			filter.pipelines[i] = RD::get_singleton()->compute_pipeline_create(filter.shader.version_get_shader(filter.shader_version, i));
+		}
+
+		if (filter.use_high_quality) {
+			filter.coefficient_buffer = RD::get_singleton()->storage_buffer_create(sizeof(high_quality_coeffs));
+			RD::get_singleton()->buffer_update(filter.coefficient_buffer, 0, sizeof(high_quality_coeffs), &high_quality_coeffs[0], false);
+		} else {
+			filter.coefficient_buffer = RD::get_singleton()->storage_buffer_create(sizeof(low_quality_coeffs));
+			RD::get_singleton()->buffer_update(filter.coefficient_buffer, 0, sizeof(low_quality_coeffs), &low_quality_coeffs[0], false);
+		}
+
+		Vector<RD::Uniform> uniforms;
+		{
+			RD::Uniform u;
+			u.type = RD::UNIFORM_TYPE_STORAGE_BUFFER;
+			u.binding = 0;
+			u.ids.push_back(filter.coefficient_buffer);
+			uniforms.push_back(u);
+		}
+		filter.uniform_set = RD::get_singleton()->uniform_set_create(uniforms, filter.shader.version_get_shader(filter.shader_version, filter.use_high_quality ? 0 : 1), 1);
+	}
+
 	RD::SamplerState sampler;
 	sampler.mag_filter = RD::SAMPLER_FILTER_LINEAR;
 	sampler.min_filter = RD::SAMPLER_FILTER_LINEAR;
@@ -987,4 +1076,7 @@ RasterizerEffectsRD::~RasterizerEffectsRD() {
 	ssao.gather_shader.version_free(ssao.gather_shader_version);
 	ssao.blur_shader.version_free(ssao.blur_shader_version);
 	roughness_limiter.shader.version_free(roughness_limiter.shader_version);
+	cubemap_downsampler.shader.version_free(cubemap_downsampler.shader_version);
+	filter.shader.version_free(filter.shader_version);
+	RD::get_singleton()->free(filter.coefficient_buffer);
 }

servers/visual/rasterizer_rd/rasterizer_effects_rd.h

@@ -36,6 +36,8 @@
 #include "servers/visual/rasterizer_rd/shaders/blur.glsl.gen.h"
 #include "servers/visual/rasterizer_rd/shaders/bokeh_dof.glsl.gen.h"
 #include "servers/visual/rasterizer_rd/shaders/copy.glsl.gen.h"
+#include "servers/visual/rasterizer_rd/shaders/cubemap_downsampler.glsl.gen.h"
+#include "servers/visual/rasterizer_rd/shaders/cubemap_filter.glsl.gen.h"
 #include "servers/visual/rasterizer_rd/shaders/cubemap_roughness.glsl.gen.h"
 #include "servers/visual/rasterizer_rd/shaders/luminance_reduce.glsl.gen.h"
 #include "servers/visual/rasterizer_rd/shaders/roughness_limiter.glsl.gen.h"
@@ -355,6 +357,45 @@ class RasterizerEffectsRD {
 
 	} roughness_limiter;
 
+	enum CubemapDownsamplerSource {
+		CUBEMAP_DOWNSAMPLER_SOURCE_PANORAMA,
+		CUBEMAP_DOWNSAMPLER_SOURCE_CUBEMAP,
+		CUBEMAP_DOWNSAMPLER_SOURCE_MAX
+	};
+
+	struct CubemapDownsamplerPushConstant {
+		uint32_t face_size;
+		float pad[3];
+	};
+
+	struct CubemapDownsampler {
+
+		CubemapDownsamplerPushConstant push_constant;
+		CubemapDownsamplerShaderRD shader;
+		RID shader_version;
+		RID pipelines[CUBEMAP_DOWNSAMPLER_SOURCE_MAX];
+
+	} cubemap_downsampler;
+
+	enum CubemapFilterMode {
+		FILTER_MODE_HIGH_QUALITY,
+		FILTER_MODE_LOW_QUALITY,
+		FILTER_MODE_HIGH_QUALITY_ARRAY,
+		FILTER_MODE_LOW_QUALITY_ARRAY,
+		FILTER_MODE_MAX,
+	};
+
+	struct CubemapFilter {
+
+		CubemapFilterShaderRD shader;
+		RID shader_version;
+		RID pipelines[FILTER_MODE_MAX];
+		RID uniform_set;
+		RID coefficient_buffer;
+		bool use_high_quality;
+
+	} filter;
+
 	RID default_sampler;
 	RID default_mipmap_sampler;
 	RID index_buffer;
@@ -424,6 +465,8 @@ public:
 	void generate_ssao(RID p_depth_buffer, RID p_normal_buffer, const Size2i &p_depth_buffer_size, RID p_depth_mipmaps_texture, const Vector<RID> &depth_mipmaps, RID p_ao1, bool p_half_size, RID p_ao2, RID p_upscale_buffer, float p_intensity, float p_radius, float p_bias, const CameraMatrix &p_projection, VS::EnvironmentSSAOQuality p_quality, VS::EnvironmentSSAOBlur p_blur, float p_edge_sharpness);
 
 	void roughness_limit(RID p_source_normal, RID p_roughness, const Size2i &p_size, float p_curve);
+	void cubemap_downsample(RID p_source_cubemap, bool p_source_is_panorama, RID p_dest_cubemap, const Size2i &p_size);
+	void cubemap_filter(RID p_source_cubemap, Vector<RID> p_dest_cubemap, bool p_use_array);
 
 	RasterizerEffectsRD();
 	~RasterizerEffectsRD();

servers/visual/rasterizer_rd/rasterizer_scene_rd.cpp

@@ -40,21 +40,29 @@ void RasterizerSceneRD::_clear_reflection_data(ReflectionData &rd) {
 
 	rd.layers.clear();
 	rd.radiance_base_cubemap = RID();
+	if (rd.downsampled_radiance_cubemap.is_valid()) {
+		RD::get_singleton()->free(rd.downsampled_radiance_cubemap);
+	}
+	rd.downsampled_radiance_cubemap = RID();
+	rd.downsampled_layer.mipmaps.clear();
+	rd.coefficient_buffer = RID();
 }
 
-void RasterizerSceneRD::_update_reflection_data(ReflectionData &rd, int p_size, int p_mipmaps, bool p_use_array, RID p_base_cube, int p_base_layer) {
+void RasterizerSceneRD::_update_reflection_data(ReflectionData &rd, int p_size, int p_mipmaps, bool p_use_array, RID p_base_cube, int p_base_layer, bool p_low_quality) {
 	//recreate radiance and all data
 
 	int mipmaps = p_mipmaps;
 	uint32_t w = p_size, h = p_size;
 
 	if (p_use_array) {
+		int layers = p_low_quality ? 7 : roughness_layers;
 
-		for (int i = 0; i < roughness_layers; i++) {
+		for (int i = 0; i < layers; i++) {
 			ReflectionData::Layer layer;
 			uint32_t mmw = w;
 			uint32_t mmh = h;
 			layer.mipmaps.resize(mipmaps);
+			layer.views.resize(mipmaps);
 			for (int j = 0; j < mipmaps; j++) {
 				ReflectionData::Layer::Mipmap &mm = layer.mipmaps.write[j];
 				mm.size.width = mmw;
@@ -66,6 +74,8 @@ void RasterizerSceneRD::_update_reflection_data(ReflectionData &rd, int p_size,
 					mm.framebuffers[k] = RD::get_singleton()->framebuffer_create(fbtex);
 				}
 
+				layer.views.write[j] = RD::get_singleton()->texture_create_shared_from_slice(RD::TextureView(), p_base_cube, p_base_layer + i * 6, j, RD::TEXTURE_SLICE_CUBEMAP);
+
 				mmw = MAX(1, mmw >> 1);
 				mmh = MAX(1, mmh >> 1);
 			}
@@ -74,12 +84,14 @@ void RasterizerSceneRD::_update_reflection_data(ReflectionData &rd, int p_size,
 		}
 
 	} else {
+		mipmaps = p_low_quality ? 7 : mipmaps;
 		//regular cubemap, lower quality (aliasing, less memory)
 		ReflectionData::Layer layer;
 		uint32_t mmw = w;
 		uint32_t mmh = h;
-		layer.mipmaps.resize(roughness_layers);
+		layer.mipmaps.resize(mipmaps);
-		for (int j = 0; j < roughness_layers; j++) {
+		layer.views.resize(mipmaps);
+		for (int j = 0; j < mipmaps; j++) {
 			ReflectionData::Layer::Mipmap &mm = layer.mipmaps.write[j];
 			mm.size.width = mmw;
 			mm.size.height = mmh;
@@ -90,6 +102,8 @@ void RasterizerSceneRD::_update_reflection_data(ReflectionData &rd, int p_size,
 				mm.framebuffers[k] = RD::get_singleton()->framebuffer_create(fbtex);
 			}
 
+			layer.views.write[j] = RD::get_singleton()->texture_create_shared_from_slice(RD::TextureView(), p_base_cube, p_base_layer, j, RD::TEXTURE_SLICE_CUBEMAP);
+
 			mmw = MAX(1, mmw >> 1);
 			mmh = MAX(1, mmh >> 1);
 		}
@@ -98,13 +112,35 @@ void RasterizerSceneRD::_update_reflection_data(ReflectionData &rd, int p_size,
 	}
 
 	rd.radiance_base_cubemap = RD::get_singleton()->texture_create_shared_from_slice(RD::TextureView(), p_base_cube, p_base_layer, 0, RD::TEXTURE_SLICE_CUBEMAP);
+
+	RD::TextureFormat tf;
+	tf.format = RD::DATA_FORMAT_R16G16B16A16_SFLOAT;
+	tf.width = 64; // Always 64x64
+	tf.height = 64;
+	tf.type = RD::TEXTURE_TYPE_CUBE;
+	tf.array_layers = 6;
+	tf.mipmaps = 7;
+	tf.usage_bits = RD::TEXTURE_USAGE_SAMPLING_BIT | RD::TEXTURE_USAGE_STORAGE_BIT | RD::TEXTURE_USAGE_COLOR_ATTACHMENT_BIT;
+
+	rd.downsampled_radiance_cubemap = RD::get_singleton()->texture_create(tf, RD::TextureView());
+	{
+		uint32_t mmw = 64;
+		uint32_t mmh = 64;
+		rd.downsampled_layer.mipmaps.resize(7);
+		for (int j = 0; j < rd.downsampled_layer.mipmaps.size(); j++) {
+			ReflectionData::DownsampleLayer::Mipmap &mm = rd.downsampled_layer.mipmaps.write[j];
+			mm.size.width = mmw;
+			mm.size.height = mmh;
+			mm.view = RD::get_singleton()->texture_create_shared_from_slice(RD::TextureView(), rd.downsampled_radiance_cubemap, 0, j, RD::TEXTURE_SLICE_CUBEMAP);
+
+			mmw = MAX(1, mmw >> 1);
+			mmh = MAX(1, mmh >> 1);
+		}
+	}
 }
 
 void RasterizerSceneRD::_create_reflection_from_panorama(ReflectionData &rd, RID p_panorama, bool p_quality) {
 
-#ifndef _MSC_VER
-#warning TODO, should probably use this algorithm instead. Volunteers? - https://www.ppsloan.org/publications/ggx_filtering.pdf	 / https://github.com/dariomanesku/cmft
-#endif
 	if (sky_use_cubemap_array) {
 
 		if (p_quality) {
@@ -115,19 +151,18 @@ void RasterizerSceneRD::_create_reflection_from_panorama(ReflectionData &rd, RID
 				}
 			}
 		} else {
-			//render to first mipmap
+			// Use fast filtering. Render directly to base mip levels
-			for (int j = 0; j < 6; j++) {
+			storage->get_effects()->cubemap_downsample(p_panorama, true, rd.downsampled_layer.mipmaps[0].view, rd.downsampled_layer.mipmaps[0].size);
-				storage->get_effects()->cubemap_roughness(p_panorama, true, rd.layers[0].mipmaps[0].framebuffers[j], j, sky_ggx_samples_realtime, 0.0);
+
+			for (int i = 1; i < rd.downsampled_layer.mipmaps.size(); i++) {
+				storage->get_effects()->cubemap_downsample(rd.downsampled_layer.mipmaps[i - 1].view, false, rd.downsampled_layer.mipmaps[i].view, rd.downsampled_layer.mipmaps[i].size);
 			}
-			//do the rest in other mipmaps and use cubemap itself as source
+			Vector<RID> views;
-			for (int i = 1; i < roughness_layers; i++) {
+			for (int i = 0; i < rd.layers.size(); i++) {
-				//render using a smaller mipmap, then copy to main layer
+				views.push_back(rd.layers[i].views[0]);
-				for (int j = 0; j < 6; j++) {
-					//storage->get_effects()->cubemap_roughness(rd.radiance_base_cubemap, false, rd.layers[0].mipmaps[i].framebuffers[0], j, sky_ggx_samples_realtime, float(i) / (rd.layers.size() - 1.0));
-					storage->get_effects()->cubemap_roughness(p_panorama, true, rd.layers[0].mipmaps[i].framebuffers[0], j, sky_ggx_samples_realtime, float(i) / (rd.layers.size() - 1.0));
-					storage->get_effects()->region_copy(rd.layers[0].mipmaps[i].views[0], rd.layers[i].mipmaps[0].framebuffers[j], Rect2());
-				}
 			}
+
+			storage->get_effects()->cubemap_filter(rd.downsampled_radiance_cubemap, views, true);
 		}
 	} else {
 
@@ -139,16 +174,13 @@ void RasterizerSceneRD::_create_reflection_from_panorama(ReflectionData &rd, RID
 				}
 			}
 		} else {
+			// Use fast filtering. Render directly to each mip level
+			storage->get_effects()->cubemap_downsample(p_panorama, true, rd.downsampled_layer.mipmaps[0].view, rd.downsampled_layer.mipmaps[0].size);
 
-			for (int j = 0; j < 6; j++) {
+			for (int i = 1; i < rd.downsampled_layer.mipmaps.size(); i++) {
-				storage->get_effects()->cubemap_roughness(p_panorama, true, rd.layers[0].mipmaps[0].framebuffers[j], j, sky_ggx_samples_realtime, 0);
+				storage->get_effects()->cubemap_downsample(rd.downsampled_layer.mipmaps[i - 1].view, false, rd.downsampled_layer.mipmaps[i].view, rd.downsampled_layer.mipmaps[i].size);
-			}
-
-			for (int i = 1; i < rd.layers[0].mipmaps.size(); i++) {
-				for (int j = 0; j < 6; j++) {
-					storage->get_effects()->cubemap_roughness(rd.radiance_base_cubemap, false, rd.layers[0].mipmaps[i].framebuffers[j], j, sky_ggx_samples_realtime, float(i) / (rd.layers[0].mipmaps.size() - 1.0));
-				}
 			}
+			storage->get_effects()->cubemap_filter(rd.downsampled_radiance_cubemap, rd.layers[0].views, false);
 		}
 	}
 }
@@ -163,12 +195,18 @@ void RasterizerSceneRD::_create_reflection_from_base_mipmap(ReflectionData &rd,
 				storage->get_effects()->cubemap_roughness(rd.radiance_base_cubemap, false, rd.layers[i].mipmaps[0].framebuffers[p_cube_side], p_cube_side, sky_ggx_samples_quality, float(i) / (rd.layers.size() - 1.0));
 			}
 		} else {
-			//do the rest in other mipmaps and use cubemap itself as source
+
-			for (int i = 1; i < roughness_layers; i++) {
+			storage->get_effects()->cubemap_downsample(rd.radiance_base_cubemap, false, rd.downsampled_layer.mipmaps[0].view, rd.downsampled_layer.mipmaps[0].size);
-				//render using a smaller mipmap, then copy to main layer
+
-				storage->get_effects()->cubemap_roughness(rd.radiance_base_cubemap, false, rd.layers[0].mipmaps[i].framebuffers[0], p_cube_side, sky_ggx_samples_realtime, float(i) / (rd.layers.size() - 1.0));
+			for (int i = 1; i < rd.downsampled_layer.mipmaps.size(); i++) {
-				storage->get_effects()->region_copy(rd.layers[0].mipmaps[i].views[0], rd.layers[i].mipmaps[0].framebuffers[p_cube_side], Rect2());
+				storage->get_effects()->cubemap_downsample(rd.downsampled_layer.mipmaps[i - 1].view, false, rd.downsampled_layer.mipmaps[i].view, rd.downsampled_layer.mipmaps[i].size);
 			}
+			Vector<RID> views;
+			for (int i = 0; i < rd.layers.size(); i++) {
+				views.push_back(rd.layers[i].views[0]);
+			}
+
+			storage->get_effects()->cubemap_filter(rd.downsampled_radiance_cubemap, views, true);
 		}
 	} else {
 
@@ -179,9 +217,12 @@ void RasterizerSceneRD::_create_reflection_from_base_mipmap(ReflectionData &rd,
 			}
 		} else {
 
-			for (int i = 1; i < rd.layers[0].mipmaps.size(); i++) {
+			storage->get_effects()->cubemap_downsample(rd.radiance_base_cubemap, false, rd.downsampled_layer.mipmaps[0].view, rd.downsampled_layer.mipmaps[0].size);
-				storage->get_effects()->cubemap_roughness(rd.radiance_base_cubemap, false, rd.layers[0].mipmaps[i].framebuffers[p_cube_side], p_cube_side, sky_ggx_samples_realtime, float(i) / (rd.layers[0].mipmaps.size() - 1.0));
+
+			for (int i = 1; i < rd.downsampled_layer.mipmaps.size(); i++) {
+				storage->get_effects()->cubemap_downsample(rd.downsampled_layer.mipmaps[i - 1].view, false, rd.downsampled_layer.mipmaps[i].view, rd.downsampled_layer.mipmaps[i].size);
 			}
+			storage->get_effects()->cubemap_filter(rd.downsampled_radiance_cubemap, rd.layers[0].views, false);
 		}
 	}
 }
@@ -224,6 +265,12 @@ void RasterizerSceneRD::sky_set_radiance_size(RID p_sky, int p_radiance_size) {
 		return;
 	}
 	sky->radiance_size = p_radiance_size;
+
+	if (sky->mode == VS::SKY_MODE_REALTIME && sky->radiance_size != 128) {
+		WARN_PRINT("Realtime Skies can only use a radiance size of 128. Radiance size will be set to 128 internally.");
+		sky->radiance_size = 128;
+	}
+
 	_sky_invalidate(sky);
 	if (sky->radiance.is_valid()) {
 		RD::get_singleton()->free(sky->radiance);
@@ -241,7 +288,18 @@ void RasterizerSceneRD::sky_set_mode(RID p_sky, VS::SkyMode p_mode) {
 	}
 
 	sky->mode = p_mode;
+
+	if (sky->mode == VS::SKY_MODE_REALTIME && sky->radiance_size != 128) {
+		WARN_PRINT("Realtime Skies can only use a radiance size of 128. Radiance size will be set to 128 internally.");
+		sky_set_radiance_size(p_sky, 128);
+	}
+
 	_sky_invalidate(sky);
+	if (sky->radiance.is_valid()) {
+		RD::get_singleton()->free(sky->radiance);
+		sky->radiance = RID();
+	}
+	_clear_reflection_data(sky->reflection);
 }
 
 void RasterizerSceneRD::sky_set_texture(RID p_sky, RID p_panorama) {
@@ -275,27 +333,24 @@ void RasterizerSceneRD::_update_dirty_skys() {
 
 		if (sky->radiance.is_null()) {
 			int mipmaps = Image::get_image_required_mipmaps(sky->radiance_size, sky->radiance_size, Image::FORMAT_RGBAH) + 1;
-			if (sky->mode != VS::SKY_MODE_QUALITY) {
-				//use less mipmaps
-				mipmaps = MIN(8, mipmaps);
-			}
 
 			uint32_t w = sky->radiance_size, h = sky->radiance_size;
+			int layers = sky->mode == VS::SKY_MODE_REALTIME ? 7 : roughness_layers;
 
 			if (sky_use_cubemap_array) {
 				//array (higher quality, 6 times more memory)
 				RD::TextureFormat tf;
-				tf.array_layers = roughness_layers * 6;
+				tf.array_layers = layers * 6;
 				tf.format = RD::DATA_FORMAT_R16G16B16A16_SFLOAT;
 				tf.type = RD::TEXTURE_TYPE_CUBE_ARRAY;
 				tf.mipmaps = mipmaps;
 				tf.width = w;
 				tf.height = h;
-				tf.usage_bits = RD::TEXTURE_USAGE_COLOR_ATTACHMENT_BIT | RD::TEXTURE_USAGE_SAMPLING_BIT;
+				tf.usage_bits = RD::TEXTURE_USAGE_COLOR_ATTACHMENT_BIT | RD::TEXTURE_USAGE_SAMPLING_BIT | RD::TEXTURE_USAGE_STORAGE_BIT;
 
 				sky->radiance = RD::get_singleton()->texture_create(tf, RD::TextureView());
 
-				_update_reflection_data(sky->reflection, sky->radiance_size, mipmaps, true, sky->radiance, 0);
+				_update_reflection_data(sky->reflection, sky->radiance_size, mipmaps, true, sky->radiance, 0, sky->mode == VS::SKY_MODE_REALTIME);
 
 			} else {
 				//regular cubemap, lower quality (aliasing, less memory)
@@ -303,14 +358,14 @@ void RasterizerSceneRD::_update_dirty_skys() {
 				tf.array_layers = 6;
 				tf.format = RD::DATA_FORMAT_R16G16B16A16_SFLOAT;
 				tf.type = RD::TEXTURE_TYPE_CUBE;
-				tf.mipmaps = roughness_layers;
+				tf.mipmaps = MIN(mipmaps, layers);
 				tf.width = w;
 				tf.height = h;
-				tf.usage_bits = RD::TEXTURE_USAGE_COLOR_ATTACHMENT_BIT | RD::TEXTURE_USAGE_SAMPLING_BIT;
+				tf.usage_bits = RD::TEXTURE_USAGE_COLOR_ATTACHMENT_BIT | RD::TEXTURE_USAGE_SAMPLING_BIT | RD::TEXTURE_USAGE_STORAGE_BIT;
 
 				sky->radiance = RD::get_singleton()->texture_create(tf, RD::TextureView());
 
-				_update_reflection_data(sky->reflection, sky->radiance_size, mipmaps, false, sky->radiance, 0);
+				_update_reflection_data(sky->reflection, sky->radiance_size, MIN(mipmaps, layers), false, sky->radiance, 0, sky->mode == VS::SKY_MODE_REALTIME);
 			}
 		}
 
@@ -591,6 +646,9 @@ void RasterizerSceneRD::reflection_atlas_set_size(RID p_ref_atlas, int p_reflect
 		return; //no changes
 	}
 
+	ra->size = p_reflection_size;
+	ra->count = p_reflection_count;
+
 	if (ra->reflection.is_valid()) {
 		//clear and invalidate everything
 		RD::get_singleton()->free(ra->reflection);
@@ -673,17 +731,27 @@ bool RasterizerSceneRD::reflection_probe_instance_begin_render(RID p_instance, R
 
 	ERR_FAIL_COND_V(!atlas, false);
 
+	ReflectionProbeInstance *rpi = reflection_probe_instance_owner.getornull(p_instance);
+	ERR_FAIL_COND_V(!rpi, false);
+
+	if (storage->reflection_probe_get_update_mode(rpi->probe) == VS::REFLECTION_PROBE_UPDATE_ALWAYS && atlas->reflection.is_valid() && atlas->size != 128) {
+		WARN_PRINT("ReflectionProbes set to UPDATE_ALWAYS must have an atlas size of 128. Please update the atlas size in the ProjectSettings.");
+		reflection_atlas_set_size(p_reflection_atlas, 128, atlas->count);
+	}
+
 	if (atlas->reflection.is_null()) {
+		int mipmaps = MIN(roughness_layers, Image::get_image_required_mipmaps(atlas->size, atlas->size, Image::FORMAT_RGBAH) + 1);
+		mipmaps = storage->reflection_probe_get_update_mode(rpi->probe) == VS::REFLECTION_PROBE_UPDATE_ALWAYS ? 7 : mipmaps; // always use 7 mipmaps with real time filtering
 		{
 			//reflection atlas was unused, create:
 			RD::TextureFormat tf;
 			tf.array_layers = 6 * atlas->count;
 			tf.format = RD::DATA_FORMAT_R16G16B16A16_SFLOAT;
 			tf.type = RD::TEXTURE_TYPE_CUBE_ARRAY;
-			tf.mipmaps = roughness_layers;
+			tf.mipmaps = mipmaps;
 			tf.width = atlas->size;
 			tf.height = atlas->size;
-			tf.usage_bits = RD::TEXTURE_USAGE_COLOR_ATTACHMENT_BIT | RD::TEXTURE_USAGE_SAMPLING_BIT;
+			tf.usage_bits = RD::TEXTURE_USAGE_COLOR_ATTACHMENT_BIT | RD::TEXTURE_USAGE_SAMPLING_BIT | RD::TEXTURE_USAGE_STORAGE_BIT;
 
 			atlas->reflection = RD::get_singleton()->texture_create(tf, RD::TextureView());
 		}
@@ -698,7 +766,7 @@ bool RasterizerSceneRD::reflection_probe_instance_begin_render(RID p_instance, R
 		}
 		atlas->reflections.resize(atlas->count);
 		for (int i = 0; i < atlas->count; i++) {
-			_update_reflection_data(atlas->reflections.write[i].data, atlas->size, roughness_layers, false, atlas->reflection, i * 6);
+			_update_reflection_data(atlas->reflections.write[i].data, atlas->size, mipmaps, false, atlas->reflection, i * 6, storage->reflection_probe_get_update_mode(rpi->probe) == VS::REFLECTION_PROBE_UPDATE_ALWAYS);
 			for (int j = 0; j < 6; j++) {
 				Vector<RID> fb;
 				fb.push_back(atlas->reflections.write[i].data.layers[0].mipmaps[0].views[j]);
@@ -712,9 +780,6 @@ bool RasterizerSceneRD::reflection_probe_instance_begin_render(RID p_instance, R
 		atlas->depth_fb = RD::get_singleton()->framebuffer_create(fb);
 	}
 
-	ReflectionProbeInstance *rpi = reflection_probe_instance_owner.getornull(p_instance);
-	ERR_FAIL_COND_V(!rpi, false);
-
 	if (rpi->atlas_index == -1) {
 		for (int i = 0; i < atlas->reflections.size(); i++) {
 			if (atlas->reflections[i].owner.is_null()) {
@@ -761,6 +826,13 @@ bool RasterizerSceneRD::reflection_probe_instance_postprocess_step(RID p_instanc
 
 	_create_reflection_from_base_mipmap(atlas->reflections.write[rpi->atlas_index].data, false, storage->reflection_probe_get_update_mode(rpi->probe) == VS::REFLECTION_PROBE_UPDATE_ONCE, rpi->processing_side);
 
+	if (storage->reflection_probe_get_update_mode(rpi->probe) == VS::REFLECTION_PROBE_UPDATE_ALWAYS) {
+		// Using real time reflections, all roughness is done in one step
+		rpi->rendering = false;
+		rpi->processing_side = 0;
+		return true;
+	}
+
 	rpi->processing_side++;
 
 	if (rpi->processing_side == 6) {
@@ -814,7 +886,7 @@ void RasterizerSceneRD::shadow_atlas_set_size(RID p_atlas, int p_size) {
 	ERR_FAIL_COND(!shadow_atlas);
 	ERR_FAIL_COND(p_size < 0);
 	p_size = next_power_of_2(p_size);
-	p_size = MAX(p_size, 1 << roughness_layers);
+	p_size = MAX(p_size, 1 << roughness_layers); // TODO: use a number related to shadows rather than reflections
 
 	if (p_size == shadow_atlas->size)
 		return;
@@ -3025,7 +3097,6 @@ RasterizerSceneRD::RasterizerSceneRD(RasterizerStorageRD *p_storage) {
 
 	roughness_layers = GLOBAL_GET("rendering/quality/reflections/roughness_layers");
 	sky_ggx_samples_quality = GLOBAL_GET("rendering/quality/reflections/ggx_samples");
-	sky_ggx_samples_realtime = GLOBAL_GET("rendering/quality/reflections/ggx_samples_realtime");
 	sky_use_cubemap_array = GLOBAL_GET("rendering/quality/reflections/texture_array_reflections");
 	//	sky_use_cubemap_array = false;
 

servers/visual/rasterizer_rd/rasterizer_scene_rd.h

@@ -85,15 +85,28 @@ private:
 				RID views[6];
 				Size2i size;
 			};
+			Vector<Mipmap> mipmaps; //per-face view
+			Vector<RID> views; // per-cubemap view
+		};
+
+		struct DownsampleLayer {
+			struct Mipmap {
+				RID view;
+				Size2i size;
+			};
 			Vector<Mipmap> mipmaps;
 		};
+
 		RID radiance_base_cubemap; //cubemap for first layer, first cubemap
+		RID downsampled_radiance_cubemap;
+		DownsampleLayer downsampled_layer;
+		RID coefficient_buffer;
 
 		Vector<Layer> layers;
 	};
 
 	void _clear_reflection_data(ReflectionData &rd);
-	void _update_reflection_data(ReflectionData &rd, int p_size, int p_mipmaps, bool p_use_array, RID p_base_cube, int p_base_layer);
+	void _update_reflection_data(ReflectionData &rd, int p_size, int p_mipmaps, bool p_use_array, RID p_base_cube, int p_base_layer, bool p_low_quality);
 	void _create_reflection_from_panorama(ReflectionData &rd, RID p_panorama, bool p_quality);
 	void _create_reflection_from_base_mipmap(ReflectionData &rd, bool p_use_arrays, bool p_quality, int p_cube_side);
 	void _update_reflection_mipmaps(ReflectionData &rd, bool p_quality);
@@ -102,7 +115,7 @@ private:
 	struct Sky {
 		RID radiance;
 		RID uniform_set;
-		int radiance_size = 256;
+		int radiance_size = 128;
 		VS::SkyMode mode = VS::SKY_MODE_QUALITY;
 		RID panorama;
 		ReflectionData reflection;
@@ -116,7 +129,6 @@ private:
 	void _update_dirty_skys();
 
 	uint32_t sky_ggx_samples_quality;
-	uint32_t sky_ggx_samples_realtime;
 	bool sky_use_cubemap_array;
 
 	mutable RID_Owner<Sky> sky_owner;

servers/visual/rasterizer_rd/shaders/SCsub

@@ -7,6 +7,8 @@ if 'RD_GLSL' in env['BUILDERS']:
     env.RD_GLSL('canvas_occlusion.glsl');
     env.RD_GLSL('blur.glsl');
     env.RD_GLSL('cubemap_roughness.glsl');
+    env.RD_GLSL('cubemap_downsampler.glsl');
+    env.RD_GLSL('cubemap_filter.glsl');
     env.RD_GLSL('scene_high_end.glsl');
     env.RD_GLSL('sky.glsl');
     env.RD_GLSL('tonemap.glsl');

servers/visual/rasterizer_rd/shaders/cubemap_downsampler.glsl

@@ -0,0 +1,220 @@
+// Copyright 2016 Activision Publishing, Inc.
+//
+// Permission is hereby granted, free of charge, to any person obtaining
+// a copy of this software and associated documentation files (the "Software"),
+// to deal in the Software without restriction, including without limitation
+// the rights to use, copy, modify, merge, publish, distribute, sublicense,
+// and/or sell copies of the Software, and to permit persons to whom the Software
+// is furnished to do so, subject to the following conditions:
+//
+// The above copyright notice and this permission notice shall be included in all
+// copies or substantial portions of the Software.
+//
+// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+// OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+// SOFTWARE.
+
+/* clang-format off */
+[compute]
+
+#version 450
+
+VERSION_DEFINES
+
+#define BLOCK_SIZE 8
+
+layout(local_size_x = BLOCK_SIZE, local_size_y = BLOCK_SIZE, local_size_z = 1) in;
+/* clang-format on */
+
+#ifdef MODE_SOURCE_PANORAMA
+layout(set = 0, binding = 0) uniform sampler2D source_panorama;
+#endif
+
+#ifdef MODE_SOURCE_CUBEMAP
+layout(set = 0, binding = 0) uniform samplerCube source_cubemap;
+#endif
+
+layout(rgba16f, set = 1, binding = 0) uniform restrict writeonly imageCube dest_cubemap;
+
+layout(push_constant, binding = 1, std430) uniform Params {
+	uint face_size;
+}
+params;
+
+#define M_PI 3.14159265359
+
+void get_dir_0(out vec3 dir, in float u, in float v) {
+	dir[0] = 1.0;
+	dir[1] = v;
+	dir[2] = -u;
+}
+void get_dir_1(out vec3 dir, in float u, in float v) {
+	dir[0] = -1.0;
+	dir[1] = v;
+	dir[2] = u;
+}
+void get_dir_2(out vec3 dir, in float u, in float v) {
+	dir[0] = u;
+	dir[1] = 1.0;
+	dir[2] = -v;
+}
+void get_dir_3(out vec3 dir, in float u, in float v) {
+	dir[0] = u;
+	dir[1] = -1.0;
+	dir[2] = v;
+}
+void get_dir_4(out vec3 dir, in float u, in float v) {
+	dir[0] = u;
+	dir[1] = v;
+	dir[2] = 1.0;
+}
+void get_dir_5(out vec3 dir, in float u, in float v) {
+	dir[0] = -u;
+	dir[1] = v;
+	dir[2] = -1.0;
+}
+
+float calcWeight(float u, float v) {
+	float val = u * u + v * v + 1.0;
+	return val * sqrt(val);
+}
+
+#ifdef MODE_SOURCE_PANORAMA
+
+vec4 texturePanorama(vec3 normal, sampler2D pano) {
+
+	vec2 st = vec2(
+			atan(normal.x, -normal.z),
+			acos(normal.y));
+
+	if (st.x < 0.0)
+		st.x += M_PI * 2.0;
+
+	st /= vec2(M_PI * 2.0, M_PI);
+
+	return textureLod(pano, st, 0.0);
+}
+
+#endif
+
+vec4 get_texture(vec3 p_dir) {
+#ifdef MODE_SOURCE_PANORAMA
+	return texturePanorama(normalize(p_dir), source_panorama);
+#else
+	return textureLod(source_cubemap, normalize(p_dir), 0.0);
+#endif
+}
+
+void main() {
+	uvec3 id = gl_GlobalInvocationID;
+	uint face_size = params.face_size;
+
+	if (id.x < face_size && id.y < face_size) {
+		float inv_face_size = 1.0 / float(face_size);
+
+		float u0 = (float(id.x) * 2.0 + 1.0 - 0.75) * inv_face_size - 1.0;
+		float u1 = (float(id.x) * 2.0 + 1.0 + 0.75) * inv_face_size - 1.0;
+
+		float v0 = (float(id.y) * 2.0 + 1.0 - 0.75) * -inv_face_size + 1.0;
+		float v1 = (float(id.y) * 2.0 + 1.0 + 0.75) * -inv_face_size + 1.0;
+
+		float weights[4];
+		weights[0] = calcWeight(u0, v0);
+		weights[1] = calcWeight(u1, v0);
+		weights[2] = calcWeight(u0, v1);
+		weights[3] = calcWeight(u1, v1);
+
+		const float wsum = 0.5 / (weights[0] + weights[1] + weights[2] + weights[3]);
+		for (int i = 0; i < 4; i++) {
+			weights[i] = weights[i] * wsum + .125;
+		}
+
+		vec3 dir;
+		vec4 color;
+		switch (id.z) {
+			case 0:
+				get_dir_0(dir, u0, v0);
+				color = get_texture(dir) * weights[0];
+
+				get_dir_0(dir, u1, v0);
+				color += get_texture(dir) * weights[1];
+
+				get_dir_0(dir, u0, v1);
+				color += get_texture(dir) * weights[2];
+
+				get_dir_0(dir, u1, v1);
+				color += get_texture(dir) * weights[3];
+				break;
+			case 1:
+				get_dir_1(dir, u0, v0);
+				color = get_texture(dir) * weights[0];
+
+				get_dir_1(dir, u1, v0);
+				color += get_texture(dir) * weights[1];
+
+				get_dir_1(dir, u0, v1);
+				color += get_texture(dir) * weights[2];
+
+				get_dir_1(dir, u1, v1);
+				color += get_texture(dir) * weights[3];
+				break;
+			case 2:
+				get_dir_2(dir, u0, v0);
+				color = get_texture(dir) * weights[0];
+
+				get_dir_2(dir, u1, v0);
+				color += get_texture(dir) * weights[1];
+
+				get_dir_2(dir, u0, v1);
+				color += get_texture(dir) * weights[2];
+
+				get_dir_2(dir, u1, v1);
+				color += get_texture(dir) * weights[3];
+				break;
+			case 3:
+				get_dir_3(dir, u0, v0);
+				color = get_texture(dir) * weights[0];
+
+				get_dir_3(dir, u1, v0);
+				color += get_texture(dir) * weights[1];
+
+				get_dir_3(dir, u0, v1);
+				color += get_texture(dir) * weights[2];
+
+				get_dir_3(dir, u1, v1);
+				color += get_texture(dir) * weights[3];
+				break;
+			case 4:
+				get_dir_4(dir, u0, v0);
+				color = get_texture(dir) * weights[0];
+
+				get_dir_4(dir, u1, v0);
+				color += get_texture(dir) * weights[1];
+
+				get_dir_4(dir, u0, v1);
+				color += get_texture(dir) * weights[2];
+
+				get_dir_4(dir, u1, v1);
+				color += get_texture(dir) * weights[3];
+				break;
+			default:
+				get_dir_5(dir, u0, v0);
+				color = get_texture(dir) * weights[0];
+
+				get_dir_5(dir, u1, v0);
+				color += get_texture(dir) * weights[1];
+
+				get_dir_5(dir, u0, v1);
+				color += get_texture(dir) * weights[2];
+
+				get_dir_5(dir, u1, v1);
+				color += get_texture(dir) * weights[3];
+				break;
+		}
+		imageStore(dest_cubemap, ivec3(id), color);
+	}
+}

servers/visual/rasterizer_rd/shaders/cubemap_filter.glsl

@@ -0,0 +1,289 @@
+// Copyright 2016 Activision Publishing, Inc.
+//
+// Permission is hereby granted, free of charge, to any person obtaining
+// a copy of this software and associated documentation files (the "Software"),
+// to deal in the Software without restriction, including without limitation
+// the rights to use, copy, modify, merge, publish, distribute, sublicense,
+// and/or sell copies of the Software, and to permit persons to whom the Software
+// is furnished to do so, subject to the following conditions:
+//
+// The above copyright notice and this permission notice shall be included in all
+// copies or substantial portions of the Software.
+//
+// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+// OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+// SOFTWARE.
+
+/* clang-format off */
+[compute]
+
+#version 450
+
+VERSION_DEFINES
+
+#define GROUP_SIZE 64
+
+layout(local_size_x = GROUP_SIZE, local_size_y = 1, local_size_z = 1) in;
+/* clang-format on */
+
+layout(set = 0, binding = 0) uniform samplerCube source_cubemap;
+layout(rgba16f, set = 2, binding = 0) uniform restrict writeonly imageCube dest_cubemap0;
+layout(rgba16f, set = 3, binding = 0) uniform restrict writeonly imageCube dest_cubemap1;
+layout(rgba16f, set = 4, binding = 0) uniform restrict writeonly imageCube dest_cubemap2;
+layout(rgba16f, set = 5, binding = 0) uniform restrict writeonly imageCube dest_cubemap3;
+layout(rgba16f, set = 6, binding = 0) uniform restrict writeonly imageCube dest_cubemap4;
+layout(rgba16f, set = 7, binding = 0) uniform restrict writeonly imageCube dest_cubemap5;
+layout(rgba16f, set = 8, binding = 0) uniform restrict writeonly imageCube dest_cubemap6;
+
+#ifdef USE_HIGH_QUALITY
+#define NUM_TAPS 32
+#else
+#define NUM_TAPS 8
+#endif
+
+#define BASE_RESOLUTION 128
+
+#ifdef USE_HIGH_QUALITY
+layout(set = 1, binding = 0, std430) buffer restrict readonly Data {
+	vec4[7][5][3][24] coeffs;
+}
+data;
+#else
+layout(set = 1, binding = 0, std430) buffer restrict readonly Data {
+	vec4[7][5][6] coeffs;
+}
+data;
+#endif
+
+void get_dir(out vec3 dir, in vec2 uv, in uint face) {
+	switch (face) {
+		case 0:
+			dir = vec3(1.0, uv[1], -uv[0]);
+			break;
+		case 1:
+			dir = vec3(-1.0, uv[1], uv[0]);
+			break;
+		case 2:
+			dir = vec3(uv[0], 1.0, -uv[1]);
+			break;
+		case 3:
+			dir = vec3(uv[0], -1.0, uv[1]);
+			break;
+		case 4:
+			dir = vec3(uv[0], uv[1], 1.0);
+			break;
+		default:
+			dir = vec3(-uv[0], uv[1], -1.0);
+			break;
+	}
+}
+
+void main() {
+	// INPUT:
+	// id.x = the linear address of the texel (ignoring face)
+	// id.y = the face
+	// -> use to index output texture
+	// id.x = texel x
+	// id.y = texel y
+	// id.z = face
+	uvec3 id = gl_GlobalInvocationID;
+
+	// determine which texel this is
+#ifndef USE_TEXTURE_ARRAY
+	int level = 0;
+	if (id.x < (128 * 128)) {
+		level = 0;
+	} else if (id.x < (128 * 128 + 64 * 64)) {
+		level = 1;
+		id.x -= (128 * 128);
+	} else if (id.x < (128 * 128 + 64 * 64 + 32 * 32)) {
+		level = 2;
+		id.x -= (128 * 128 + 64 * 64);
+	} else if (id.x < (128 * 128 + 64 * 64 + 32 * 32 + 16 * 16)) {
+		level = 3;
+		id.x -= (128 * 128 + 64 * 64 + 32 * 32);
+	} else if (id.x < (128 * 128 + 64 * 64 + 32 * 32 + 16 * 16 + 8 * 8)) {
+		level = 4;
+		id.x -= (128 * 128 + 64 * 64 + 32 * 32 + 16 * 16);
+	} else if (id.x < (128 * 128 + 64 * 64 + 32 * 32 + 16 * 16 + 8 * 8 + 4 * 4)) {
+		level = 5;
+		id.x -= (128 * 128 + 64 * 64 + 32 * 32 + 16 * 16 + 8 * 8);
+	} else if (id.x < (128 * 128 + 64 * 64 + 32 * 32 + 16 * 16 + 8 * 8 + 4 * 4 + 2 * 2)) {
+		level = 6;
+		id.x -= (128 * 128 + 64 * 64 + 32 * 32 + 16 * 16 + 8 * 8 + 4 * 4);
+	} else {
+		return;
+	}
+	int res = BASE_RESOLUTION >> level;
+#else // Using Texture Arrays so all levels are the same resolution
+	int res = BASE_RESOLUTION;
+	int level = int(id.x / (BASE_RESOLUTION * BASE_RESOLUTION));
+	id.x -= level * BASE_RESOLUTION * BASE_RESOLUTION;
+#endif
+
+	// determine dir / pos for the texel
+	vec3 dir, adir, frameZ;
+	{
+		id.z = id.y;
+		id.y = id.x / res;
+		id.x -= id.y * res;
+
+		vec2 uv;
+		uv.x = (float(id.x) * 2.0 + 1.0) / float(res) - 1.0;
+		uv.y = -(float(id.y) * 2.0 + 1.0) / float(res) + 1.0;
+
+		get_dir(dir, uv, id.z);
+		frameZ = normalize(dir);
+
+		adir = abs(dir);
+	}
+
+	// GGX gather colors
+	vec4 color = vec4(0.0);
+	for (int axis = 0; axis < 3; axis++) {
+		const int otherAxis0 = 1 - (axis & 1) - (axis >> 1);
+		const int otherAxis1 = 2 - (axis >> 1);
+
+		float frameweight = (max(adir[otherAxis0], adir[otherAxis1]) - .75) / .25;
+		if (frameweight > 0.0) {
+			// determine frame
+			vec3 UpVector;
+			switch (axis) {
+				case 0:
+					UpVector = vec3(1, 0, 0);
+					break;
+				case 1:
+					UpVector = vec3(0, 1, 0);
+					break;
+				default:
+					UpVector = vec3(0, 0, 1);
+					break;
+			}
+
+			vec3 frameX = normalize(cross(UpVector, frameZ));
+			vec3 frameY = cross(frameZ, frameX);
+
+			// calculate parametrization for polynomial
+			float Nx = dir[otherAxis0];
+			float Ny = dir[otherAxis1];
+			float Nz = adir[axis];
+
+			float NmaxXY = max(abs(Ny), abs(Nx));
+			Nx /= NmaxXY;
+			Ny /= NmaxXY;
+
+			float theta;
+			if (Ny < Nx) {
+				if (Ny <= -0.999)
+					theta = Nx;
+				else
+					theta = Ny;
+			} else {
+				if (Ny >= 0.999)
+					theta = -Nx;
+				else
+					theta = -Ny;
+			}
+
+			float phi;
+			if (Nz <= -0.999)
+				phi = -NmaxXY;
+			else if (Nz >= 0.999)
+				phi = NmaxXY;
+			else
+				phi = Nz;
+
+			float theta2 = theta * theta;
+			float phi2 = phi * phi;
+
+			// sample
+			for (int iSuperTap = 0; iSuperTap < NUM_TAPS / 4; iSuperTap++) {
+				const int index = (NUM_TAPS / 4) * axis + iSuperTap;
+
+#ifdef USE_HIGH_QUALITY
+				vec4 coeffsDir0[3];
+				vec4 coeffsDir1[3];
+				vec4 coeffsDir2[3];
+				vec4 coeffsLevel[3];
+				vec4 coeffsWeight[3];
+
+				for (int iCoeff = 0; iCoeff < 3; iCoeff++) {
+					coeffsDir0[iCoeff] = data.coeffs[level][0][iCoeff][index];
+					coeffsDir1[iCoeff] = data.coeffs[level][1][iCoeff][index];
+					coeffsDir2[iCoeff] = data.coeffs[level][2][iCoeff][index];
+					coeffsLevel[iCoeff] = data.coeffs[level][3][iCoeff][index];
+					coeffsWeight[iCoeff] = data.coeffs[level][4][iCoeff][index];
+				}
+
+				for (int iSubTap = 0; iSubTap < 4; iSubTap++) {
+					// determine sample attributes (dir, weight, level)
+					vec3 sample_dir = frameX * (coeffsDir0[0][iSubTap] + coeffsDir0[1][iSubTap] * theta2 + coeffsDir0[2][iSubTap] * phi2) + frameY * (coeffsDir1[0][iSubTap] + coeffsDir1[1][iSubTap] * theta2 + coeffsDir1[2][iSubTap] * phi2) + frameZ * (coeffsDir2[0][iSubTap] + coeffsDir2[1][iSubTap] * theta2 + coeffsDir2[2][iSubTap] * phi2);
+
+					float sample_level = coeffsLevel[0][iSubTap] + coeffsLevel[1][iSubTap] * theta2 + coeffsLevel[2][iSubTap] * phi2;
+
+					float sample_weight = coeffsWeight[0][iSubTap] + coeffsWeight[1][iSubTap] * theta2 + coeffsWeight[2][iSubTap] * phi2;
+#else
+				vec4 coeffsDir0 = data.coeffs[level][0][index];
+				vec4 coeffsDir1 = data.coeffs[level][1][index];
+				vec4 coeffsDir2 = data.coeffs[level][2][index];
+				vec4 coeffsLevel = data.coeffs[level][3][index];
+				vec4 coeffsWeight = data.coeffs[level][4][index];
+
+				for (int iSubTap = 0; iSubTap < 4; iSubTap++) {
+					// determine sample attributes (dir, weight, level)
+					vec3 sample_dir = frameX * coeffsDir0[iSubTap] + frameY * coeffsDir1[iSubTap] + frameZ * coeffsDir2[iSubTap];
+
+					float sample_level = coeffsLevel[iSubTap];
+
+					float sample_weight = coeffsWeight[iSubTap];
+#endif
+
+					sample_weight *= frameweight;
+
+					// adjust for jacobian
+					sample_dir /= max(abs(sample_dir[0]), max(abs(sample_dir[1]), abs(sample_dir[2])));
+					sample_level += 0.75 * log2(dot(sample_dir, sample_dir));
+#ifndef USE_TEXTURE_ARRAY
+					sample_level += float(level) / 6.0; // Hack to increase the perceived roughness and reduce upscaling artifacts
+#endif
+					// sample cubemap
+					color.xyz += textureLod(source_cubemap, normalize(sample_dir), sample_level).xyz * sample_weight;
+					color.w += sample_weight;
+				}
+			}
+		}
+	}
+	color /= color.w;
+
+	// write color
+	color.xyz = max(vec3(0.0), color.xyz);
+	color.w = 1.0;
+
+	switch (level) {
+		case 0:
+			imageStore(dest_cubemap0, ivec3(id), color);
+			break;
+		case 1:
+			imageStore(dest_cubemap1, ivec3(id), color);
+			break;
+		case 2:
+			imageStore(dest_cubemap2, ivec3(id), color);
+			break;
+		case 3:
+			imageStore(dest_cubemap3, ivec3(id), color);
+			break;
+		case 4:
+			imageStore(dest_cubemap4, ivec3(id), color);
+			break;
+		case 5:
+			imageStore(dest_cubemap5, ivec3(id), color);
+			break;
+		default:
+			imageStore(dest_cubemap6, ivec3(id), color);
+			break;
+	}
+}

servers/visual_server.cpp

@@ -2316,9 +2316,8 @@ VisualServer::VisualServer() {
 	GLOBAL_DEF("rendering/quality/reflections/texture_array_reflections.mobile", false);
 	GLOBAL_DEF("rendering/quality/reflections/ggx_samples", 1024);
 	GLOBAL_DEF("rendering/quality/reflections/ggx_samples.mobile", 128);
-	GLOBAL_DEF("rendering/quality/reflections/ggx_samples_realtime", 64);
+	GLOBAL_DEF("rendering/quality/reflections/fast_filter_high_quality", false);
-	GLOBAL_DEF("rendering/quality/reflections/ggx_samples_realtime.mobile", 16);
+	GLOBAL_DEF("rendering/quality/reflection_atlas/reflection_size", 128);
-	GLOBAL_DEF("rendering/quality/reflection_atlas/reflection_size", 256);
 	GLOBAL_DEF("rendering/quality/reflection_atlas/reflection_size.mobile", 128);
 	GLOBAL_DEF("rendering/quality/reflection_atlas/reflection_count", 64);