render: implement PCSS for directional lights

examples/assets/wood-platform/model.gltf

@@ -0,0 +1,142 @@
+{
+	"asset":{
+		"generator":"Khronos glTF Blender I/O v4.1.63",
+		"version":"2.0"
+	},
+	"scene":0,
+	"scenes":[
+		{
+			"name":"Scene",
+			"nodes":[
+				0
+			]
+		}
+	],
+	"nodes":[
+		{
+			"mesh":0,
+			"name":"Cube",
+			"scale":[
+				10,
+				0.25,
+				10
+			]
+		}
+	],
+	"materials":[
+		{
+			"doubleSided":true,
+			"name":"Material.001",
+			"pbrMetallicRoughness":{
+				"baseColorTexture":{
+					"index":0
+				},
+				"metallicFactor":0,
+				"roughnessFactor":0.5
+			}
+		}
+	],
+	"meshes":[
+		{
+			"name":"Cube.001",
+			"primitives":[
+				{
+					"attributes":{
+						"POSITION":0,
+						"NORMAL":1,
+						"TEXCOORD_0":2
+					},
+					"indices":3,
+					"material":0
+				}
+			]
+		}
+	],
+	"textures":[
+		{
+			"sampler":0,
+			"source":0
+		}
+	],
+	"images":[
+		{
+			"mimeType":"image/jpeg",
+			"name":"wood1",
+			"uri":"wood1.jpg"
+		}
+	],
+	"accessors":[
+		{
+			"bufferView":0,
+			"componentType":5126,
+			"count":24,
+			"max":[
+				1,
+				1,
+				1
+			],
+			"min":[
+				-1,
+				-1,
+				-1
+			],
+			"type":"VEC3"
+		},
+		{
+			"bufferView":1,
+			"componentType":5126,
+			"count":24,
+			"type":"VEC3"
+		},
+		{
+			"bufferView":2,
+			"componentType":5126,
+			"count":24,
+			"type":"VEC2"
+		},
+		{
+			"bufferView":3,
+			"componentType":5123,
+			"count":36,
+			"type":"SCALAR"
+		}
+	],
+	"bufferViews":[
+		{
+			"buffer":0,
+			"byteLength":288,
+			"byteOffset":0,
+			"target":34962
+		},
+		{
+			"buffer":0,
+			"byteLength":288,
+			"byteOffset":288,
+			"target":34962
+		},
+		{
+			"buffer":0,
+			"byteLength":192,
+			"byteOffset":576,
+			"target":34962
+		},
+		{
+			"buffer":0,
+			"byteLength":72,
+			"byteOffset":768,
+			"target":34963
+		}
+	],
+	"samplers":[
+		{
+			"magFilter":9729,
+			"minFilter":9987
+		}
+	],
+	"buffers":[
+		{
+			"byteLength":840,
+			"uri":"model.bin"
+		}
+	]
+}

examples/shadows/src/main.rs

@@ -128,15 +128,29 @@ fn setup_scene_plugin(game: &mut Game) {
     platform_gltf.wait_recurse_dependencies_load();
     let platform_mesh = &platform_gltf.data_ref().unwrap().scenes[0];
 
+    let palm_tree_platform_gltf = resman
+        .request::<Gltf>("../assets/shadows-platform-palmtree.glb")
+        .unwrap();
+
+    palm_tree_platform_gltf.wait_recurse_dependencies_load();
+    let palm_tree_platform_mesh = &palm_tree_platform_gltf.data_ref().unwrap().scenes[0];
+
     drop(resman);
 
     // cube in the air
-    world.spawn((
+    /* world.spawn((
         cube_mesh.clone(),
         WorldTransform::default(),
         Transform::from_xyz(0.0, -2.0, -5.0),
     ));
 
+    // cube really high in the air
+    world.spawn((
+        cube_mesh.clone(),
+        WorldTransform::default(),
+        Transform::from_xyz(-6.0, 0.0, -5.0),
+    ));
+
     // cube on the right, on the ground
     world.spawn((
         cube_mesh.clone(),
@@ -149,10 +163,19 @@ fn setup_scene_plugin(game: &mut Game) {
         WorldTransform::default(),
         //Transform::from_xyz(0.0, -5.0, -5.0),
         Transform::new(math::vec3(0.0, -5.0, -5.0), math::Quat::IDENTITY, math::vec3(5.0, 1.0, 5.0)),
+    )); */
+
+    world.spawn((
+        palm_tree_platform_mesh.clone(),
+        WorldTransform::default(),
+        Transform::from_xyz(5.0, -15.0, 0.0),
+        //Transform::new(math::vec3(0.0, -5.0, -5.0), math::Quat::IDENTITY, math::vec3(5.0, 1.0, 5.0)),
     ));
 
+    //shadows-platform-palmtree.glb
+
     {
-        let mut light_tran = Transform::from_xyz(0.0, 2.5, 0.0);
+        let mut light_tran = Transform::from_xyz(0.0, 0.0, 0.0);
         light_tran.scale = Vec3::new(0.5, 0.5, 0.5);
         light_tran.rotate_x(math::Angle::Degrees(-45.0));
         light_tran.rotate_y(math::Angle::Degrees(-35.0));
@@ -192,7 +215,9 @@ fn setup_scene_plugin(game: &mut Game) {
     }
 
     let mut camera = CameraComponent::new_3d();
-    camera.transform.translation += math::Vec3::new(0.0, 2.0, 10.5);
+    //camera.transform.translation += math::Vec3::new(0.0, 2.0, 10.5);
-    camera.transform.rotate_x(math::Angle::Degrees(-17.0));
+    camera.transform.translation = math::Vec3::new(-3.0, -8.0, -3.0);
+    camera.transform.rotate_x(math::Angle::Degrees(-27.0));
+    camera.transform.rotate_y(math::Angle::Degrees(-55.0));
     world.spawn((camera, FreeFlyCamera::default()));
 }

lyra-game/src/render/graph/node.rs

@@ -63,6 +63,14 @@ pub enum SlotValue {
 }
 
 impl SlotValue {
+    pub fn is_none(&self) -> bool {
+        matches!(self, Self::None)
+    }
+
+    pub fn is_lazy(&self) -> bool {
+        matches!(self, Self::Lazy)
+    }
+
     pub fn as_texture_view(&self) -> Option<&Arc<wgpu::TextureView>> {
         bind_match!(self, Self::TextureView(v) => v)
     }

lyra-game/src/render/graph/passes/meshes.rs

@@ -115,6 +115,11 @@ impl Node for MeshPass {
                 .expect("missing ShadowMapsPassSlots::ShadowAtlasSampler")
                 .as_sampler()
                 .unwrap();
+            let atlas_sampler_compare = graph
+                .slot_value(ShadowMapsPassSlots::ShadowAtlasSamplerComparison)
+                .expect("missing ShadowMapsPassSlots::ShadowAtlasSamplerComparison")
+                .as_sampler()
+                .unwrap();
             let shadow_settings_buf = graph
                 .slot_value(ShadowMapsPassSlots::ShadowSettingsUniform)
                 .expect("missing ShadowMapsPassSlots::ShadowSettingsUniform")
@@ -130,6 +135,11 @@ impl Node for MeshPass {
                 .expect("missing ShadowMapsPassSlots::PcfPoissonDiscBuffer")
                 .as_buffer()
                 .unwrap();
+            let pcss_poisson_disc = graph
+                .slot_value(ShadowMapsPassSlots::PcssPoissonDiscBuffer)
+                .expect("missing ShadowMapsPassSlots::PcssPoissonDiscBuffer")
+                .as_buffer()
+                .unwrap();
 
             let atlas_layout = device.create_bind_group_layout(&wgpu::BindGroupLayoutDescriptor {
                 label: Some("bgl_shadows_atlas"),
@@ -147,11 +157,17 @@ impl Node for MeshPass {
                     wgpu::BindGroupLayoutEntry {
                         binding: 1,
                         visibility: wgpu::ShaderStages::FRAGMENT,
-                        ty: wgpu::BindingType::Sampler(wgpu::SamplerBindingType::Comparison),
+                        ty: wgpu::BindingType::Sampler(wgpu::SamplerBindingType::Filtering),
                         count: None,
                     },
                     wgpu::BindGroupLayoutEntry {
                         binding: 2,
+                        visibility: wgpu::ShaderStages::FRAGMENT,
+                        ty: wgpu::BindingType::Sampler(wgpu::SamplerBindingType::Comparison),
+                        count: None,
+                    },
+                    wgpu::BindGroupLayoutEntry {
+                        binding: 3,
                         visibility: wgpu::ShaderStages::VERTEX_FRAGMENT,
                         ty: wgpu::BindingType::Buffer {
                             ty: wgpu::BufferBindingType::Uniform,
@@ -161,7 +177,7 @@ impl Node for MeshPass {
                         count: None,
                     },
                     wgpu::BindGroupLayoutEntry {
-                        binding: 3,
+                        binding: 4,
                         visibility: wgpu::ShaderStages::VERTEX_FRAGMENT,
                         ty: wgpu::BindingType::Buffer {
                             ty: wgpu::BufferBindingType::Storage { read_only: true },
@@ -171,7 +187,17 @@ impl Node for MeshPass {
                         count: None,
                     },
                     wgpu::BindGroupLayoutEntry {
-                        binding: 4,
+                        binding: 5,
+                        visibility: wgpu::ShaderStages::VERTEX_FRAGMENT,
+                        ty: wgpu::BindingType::Buffer {
+                            ty: wgpu::BufferBindingType::Storage { read_only: true },
+                            has_dynamic_offset: false,
+                            min_binding_size: None,
+                        },
+                        count: None,
+                    },
+                    wgpu::BindGroupLayoutEntry {
+                        binding: 6,
                         visibility: wgpu::ShaderStages::VERTEX_FRAGMENT,
                         ty: wgpu::BindingType::Buffer {
                             ty: wgpu::BufferBindingType::Storage { read_only: true },
@@ -197,6 +223,10 @@ impl Node for MeshPass {
                     },
                     wgpu::BindGroupEntry {
                         binding: 2,
+                        resource: wgpu::BindingResource::Sampler(atlas_sampler_compare),
+                    },
+                    wgpu::BindGroupEntry {
+                        binding: 3,
                         resource: wgpu::BindingResource::Buffer(wgpu::BufferBinding {
                             buffer: shadow_settings_buf,
                             offset: 0,
@@ -204,7 +234,7 @@ impl Node for MeshPass {
                         }),
                     },
                     wgpu::BindGroupEntry {
-                        binding: 3,
+                        binding: 4,
                         resource: wgpu::BindingResource::Buffer(wgpu::BufferBinding {
                             buffer: light_uniform_buf,
                             offset: 0,
@@ -212,13 +242,21 @@ impl Node for MeshPass {
                         }),
                     },
                     wgpu::BindGroupEntry {
-                        binding: 4,
+                        binding: 5,
                         resource: wgpu::BindingResource::Buffer(wgpu::BufferBinding {
                             buffer: pcf_poisson_disc,
                             offset: 0,
                             size: None,
                         }),
                     },
+                    wgpu::BindGroupEntry {
+                        binding: 6,
+                        resource: wgpu::BindingResource::Buffer(wgpu::BufferBinding {
+                            buffer: pcss_poisson_disc,
+                            offset: 0,
+                            size: None,
+                        }),
+                    },
                 ],
             });
 

lyra-game/src/render/graph/passes/shadows.rs

@@ -3,6 +3,7 @@ use std::{
 };
 
 use fast_poisson::Poisson2D;
+use glam::Vec2;
 use itertools::Itertools;
 use lyra_ecs::{
     query::{filter::Has, Entities},
@@ -25,18 +26,19 @@ use crate::render::{
 
 use super::{MeshBufferStorage, RenderAssets, RenderMeshes};
 
-const PCF_SAMPLES_NUM: u32 = 6;
+const SHADOW_SIZE: glam::UVec2 = glam::uvec2(4096, 4096);
-const SHADOW_SIZE: glam::UVec2 = glam::uvec2(1024, 1024);
 
 #[derive(Debug, Clone, Hash, PartialEq, RenderGraphLabel)]
 pub enum ShadowMapsPassSlots {
     ShadowAtlasTexture,
     ShadowAtlasTextureView,
     ShadowAtlasSampler,
+    ShadowAtlasSamplerComparison,
     ShadowAtlasSizeBuffer,
     ShadowLightUniformsBuffer,
     ShadowSettingsUniform,
     PcfPoissonDiscBuffer,
+    PcssPoissonDiscBuffer,
 }
 
 #[derive(Debug, Clone, Hash, PartialEq, RenderGraphLabel)]
@@ -74,6 +76,7 @@ pub struct ShadowMapsPass {
     atlas: LightShadowMapAtlas,
     /// The depth map atlas sampler
     atlas_sampler: Rc<wgpu::Sampler>,
+    atlas_sampler_compare: Rc<wgpu::Sampler>,
 }
 
 impl ShadowMapsPass {
@@ -98,7 +101,7 @@ impl ShadowMapsPass {
             device,
             wgpu::TextureFormat::Depth32Float,
             wgpu::TextureUsages::RENDER_ATTACHMENT | wgpu::TextureUsages::TEXTURE_BINDING,
-            SHADOW_SIZE * 8,
+            SHADOW_SIZE * 2,
         );
 
         let atlas_size_buffer = device.create_buffer_init(&wgpu::util::BufferInitDescriptor {
@@ -107,16 +110,28 @@ impl ShadowMapsPass {
             contents: bytemuck::bytes_of(&atlas.atlas_size()),
         });
 
+        let sampler_compare = device.create_sampler(&wgpu::SamplerDescriptor {
+            label: Some("compare_sampler_shadow_map_atlas"),
+            address_mode_u: wgpu::AddressMode::ClampToBorder,
+            address_mode_v: wgpu::AddressMode::ClampToBorder,
+            address_mode_w: wgpu::AddressMode::ClampToBorder,
+            mag_filter: wgpu::FilterMode::Nearest,
+            min_filter: wgpu::FilterMode::Nearest,
+            mipmap_filter: wgpu::FilterMode::Nearest,
+            border_color: Some(wgpu::SamplerBorderColor::OpaqueWhite),
+            compare: Some(wgpu::CompareFunction::LessEqual),
+            ..Default::default()
+        });
+
         let sampler = device.create_sampler(&wgpu::SamplerDescriptor {
             label: Some("sampler_shadow_map_atlas"),
             address_mode_u: wgpu::AddressMode::ClampToBorder,
             address_mode_v: wgpu::AddressMode::ClampToBorder,
             address_mode_w: wgpu::AddressMode::ClampToBorder,
-            mag_filter: wgpu::FilterMode::Linear,
+            mag_filter: wgpu::FilterMode::Nearest,
-            min_filter: wgpu::FilterMode::Linear,
+            min_filter: wgpu::FilterMode::Nearest,
-            mipmap_filter: wgpu::FilterMode::Linear,
+            mipmap_filter: wgpu::FilterMode::Nearest,
             border_color: Some(wgpu::SamplerBorderColor::OpaqueWhite),
-            compare: Some(wgpu::CompareFunction::LessEqual),
             ..Default::default()
         });
 
@@ -155,6 +170,7 @@ impl ShadowMapsPass {
             point_light_pipeline: None,
 
             atlas_sampler: Rc::new(sampler),
+            atlas_sampler_compare: Rc::new(sampler_compare),
             atlas: LightShadowMapAtlas(Arc::new(RwLock::new(atlas))),
         }
     }
@@ -183,6 +199,14 @@ impl ShadowMapsPass {
 
                 let projection =
                     glam::Mat4::orthographic_rh(-10.0, 10.0, -10.0, 10.0, NEAR_PLANE, FAR_PLANE);
+                
+                // honestly no clue why this works, but I got it from here and the results are good
+                // https://github.com/asylum2010/Asylum_Tutorials/blob/423e5edfaee7b5ea450a450e65f2eabf641b2482/ShaderTutors/43_ShadowMapFiltering/main.cpp#L323
+                let frustum_size = Vec2::new(0.5 * projection.col(0).x, 0.5 * projection.col(1).y);
+                // maybe its better to make this a vec2 on the gpu?
+                let size_avg = (frustum_size.x + frustum_size.y) / 2.0;
+                let light_size_uv = 0.2 * size_avg;
+                
                 let look_view = glam::Mat4::look_to_rh(
                     light_pos.translation,
                     light_pos.forward(),
@@ -196,7 +220,8 @@ impl ShadowMapsPass {
                     atlas_frame,
                     near_plane: NEAR_PLANE,
                     far_plane,
-                    _padding1: [0; 2],
+                    light_size_uv,
+                    _padding1: 0,
                     light_pos: light_pos.translation,
                     _padding2: 0,
                 };
@@ -284,7 +309,8 @@ impl ShadowMapsPass {
                             atlas_frame: frames[i],
                             near_plane: NEAR_PLANE,
                             far_plane,
-                            _padding1: [0; 2],
+                            light_size_uv: 0.0,
+                            _padding1: 0,
                             light_pos: light_trans,
                             _padding2: 0,
                         },
@@ -322,26 +348,29 @@ impl ShadowMapsPass {
     fn create_poisson_disc_buffer(&self, device: &wgpu::Device, label: &str, num_samples: u32) -> wgpu::Buffer {
         device.create_buffer(&wgpu::BufferDescriptor {
             label: Some(label),
-            size: mem::size_of::<glam::Vec2>() as u64 * (num_samples.pow(2)) as u64,
+            size: mem::size_of::<glam::Vec2>() as u64 * (num_samples * 2) as u64,
             usage: wgpu::BufferUsages::STORAGE | wgpu::BufferUsages::COPY_DST,
             mapped_at_creation: false,
         })
     }
 
     /// Generate and write a Poisson disc to `buffer` with `num_pcf_samples.pow(2)` amount of points.
-    fn write_poisson_disc(&self, queue: &wgpu::Queue, buffer: &wgpu::Buffer, num_pcf_samples: u32) {
+    fn write_poisson_disc(&self, queue: &wgpu::Queue, buffer: &wgpu::Buffer, num_samples: u32) {
-        let num_points = num_pcf_samples.pow(2);
+        //let num_points = num_samples.pow(2);
-        let num_floats = num_points * 2; // points are vec2f
+        let num_floats = num_samples * 2; // points are vec2f
         let min_dist = (num_floats as f32).sqrt() / num_floats as f32;
+        //let min_dist = (num_samples as f32).sqrt() / num_samples as f32;
         let mut points = vec![];
 
         // use a while loop to ensure that the correct number of floats is created
         while points.len() < num_floats as usize {
             let poisson = Poisson2D::new()
                 .with_dimensions([1.0, 1.0], min_dist)
-                .with_samples(num_pcf_samples);
+                .with_samples(num_samples);
 
-            points = poisson.iter().flatten().collect_vec();
+            points = poisson.iter().flatten()
+                .map(|p| p * 2.0 - 1.0)
+                .collect_vec();
             
         }
         points.truncate(num_floats as _);
@@ -377,6 +406,12 @@ impl Node for ShadowMapsPass {
             Some(SlotValue::Sampler(self.atlas_sampler.clone())),
         );
 
+        node.add_sampler_slot(
+            ShadowMapsPassSlots::ShadowAtlasSamplerComparison,
+            SlotAttribute::Output,
+            Some(SlotValue::Sampler(self.atlas_sampler_compare.clone())),
+        );
+
         node.add_buffer_slot(
             ShadowMapsPassSlots::ShadowLightUniformsBuffer,
             SlotAttribute::Output,
@@ -404,11 +439,20 @@ impl Node for ShadowMapsPass {
             Some(SlotValue::Buffer(Arc::new(settings_buffer))),
         );
 
+        let def_settings = ShadowSettings::default();
         node.add_buffer_slot(
             ShadowMapsPassSlots::PcfPoissonDiscBuffer,
             SlotAttribute::Output,
             Some(SlotValue::Buffer(Arc::new(
-                self.create_poisson_disc_buffer(device, "buffer_poisson_disc_pcf", PCF_SAMPLES_NUM),
+                self.create_poisson_disc_buffer(device, "buffer_poisson_disc_pcf", def_settings.pcf_samples_num),
+            ))),
+        );
+
+        node.add_buffer_slot(
+            ShadowMapsPassSlots::PcssPoissonDiscBuffer,
+            SlotAttribute::Output,
+            Some(SlotValue::Buffer(Arc::new(
+                self.create_poisson_disc_buffer(device, "buffer_poisson_disc_pcss", def_settings.pcss_blocker_search_samples),
             ))),
         );
 
@@ -424,9 +468,14 @@ impl Node for ShadowMapsPass {
         {
             // TODO: Update the poisson disc every time the PCF sampling point number changed
             if !world.has_resource::<ShadowSettings>() {
+                let def_settings = ShadowSettings::default();
                 let buffer = graph.slot_value(ShadowMapsPassSlots::PcfPoissonDiscBuffer)
                     .unwrap().as_buffer().unwrap();
-                self.write_poisson_disc(&context.queue, &buffer, ShadowSettings::default().pcf_samples_num);
+                self.write_poisson_disc(&context.queue, &buffer, def_settings.pcf_samples_num);
+
+                let buffer = graph.slot_value(ShadowMapsPassSlots::PcssPoissonDiscBuffer)
+                    .unwrap().as_buffer().unwrap();
+                self.write_poisson_disc(&context.queue, &buffer, def_settings.pcss_blocker_search_samples);
             }
 
             // TODO: only write buffer on changes to resource
@@ -717,7 +766,9 @@ pub struct LightShadowUniform {
     atlas_frame: AtlasFrame, // 2xUVec2 (4xf32), so no padding needed
     near_plane: f32,
     far_plane: f32,
-    _padding1: [u32; 2],
+    /// Light size in UV space (light_size / frustum_size)
+    light_size_uv: f32,
+    _padding1: u32,
     light_pos: glam::Vec3,
     _padding2: u32,
 }
@@ -758,13 +809,22 @@ impl LightShadowMapAtlas {
 
 #[derive(Debug, Copy, Clone)]
 pub struct ShadowSettings {
+    /// How many PCF filtering samples are used per dimension.
+    /// 
+    /// A value of 16 is common.
     pub pcf_samples_num: u32,
+    /// How many samples are used for the PCSS blocker search step.
+    /// 
+    /// Multiple samples are required to avoid holes int he penumbra due to missing blockers.
+    /// A value of 16 is common.
+    pub pcss_blocker_search_samples: u32,
 }
 
 impl Default for ShadowSettings {
     fn default() -> Self {
         Self {
-            pcf_samples_num: PCF_SAMPLES_NUM,
+            pcf_samples_num: 64,
+            pcss_blocker_search_samples: 36,
         }
     }
 }
@@ -774,12 +834,14 @@ impl Default for ShadowSettings {
 #[derive(Debug, Copy, Clone, bytemuck::Pod, bytemuck::Zeroable)]
 struct ShadowSettingsUniform {
     pcf_samples_num: u32,
+    pcss_blocker_search_samples: u32,
 }
 
 impl From<ShadowSettings> for ShadowSettingsUniform {
     fn from(value: ShadowSettings) -> Self {
         Self {
             pcf_samples_num: value.pcf_samples_num,
+            pcss_blocker_search_samples: value.pcss_blocker_search_samples,
         }
     }
 }

lyra-game/src/render/shaders/base.wgsl

@@ -117,11 +117,13 @@ struct LightShadowMapUniform {
     atlas_frame: TextureAtlasFrame,
     near_plane: f32,
     far_plane: f32,
+    light_size_uv: f32,
     light_pos: vec3<f32>,
 }
 
 struct ShadowSettingsUniform {
     pcf_samples_num: u32,
+    pcss_blocker_search_samples: u32,
 }
 
 @group(4) @binding(0)
@@ -132,13 +134,17 @@ var t_light_grid: texture_storage_2d<rg32uint, read_write>; // rg32uint = vec2<u
 @group(5) @binding(0)
 var t_shadow_maps_atlas: texture_depth_2d;
 @group(5) @binding(1)
-var s_shadow_maps_atlas: sampler_comparison;
+var s_shadow_maps_atlas: sampler;
 @group(5) @binding(2)
-var<uniform> u_shadow_settings: ShadowSettingsUniform;
+var s_shadow_maps_atlas_compare: sampler_comparison;
 @group(5) @binding(3)
-var<storage, read> u_light_shadow: array<LightShadowMapUniform>;
+var<uniform> u_shadow_settings: ShadowSettingsUniform;
 @group(5) @binding(4)
+var<storage, read> u_light_shadow: array<LightShadowMapUniform>;
+@group(5) @binding(5)
 var<storage, read> u_pcf_poisson_disc: array<vec2<f32>>;
+@group(5) @binding(6)
+var<storage, read> u_pcss_poisson_disc: array<vec2<f32>>;
 
 @fragment
 fn fs_main(in: VertexOutput) -> @location(0) vec4<f32> {
@@ -185,8 +191,12 @@ fn fs_main(in: VertexOutput) -> @location(0) vec4<f32> {
     return vec4<f32>(light_object_res, object_color.a);
 }
 
-/// Get the cube map side index of a 3d texture coord
+/// Convert 3d coords for an unwrapped cubemap to 2d coords and a side index of the cube map.
 ///
+/// The `xy` components are the 2d coordinates in the side of the cube, and `z` is the cube
+/// map side index.
+///
+/// Cube map index results:
 /// 0 -> UNKNOWN
 /// 1 -> right
 /// 2 -> left
@@ -194,7 +204,7 @@ fn fs_main(in: VertexOutput) -> @location(0) vec4<f32> {
 /// 4 -> bottom
 /// 5 -> near
 /// 6 -> far
-fn get_side_idx(tex_coord: vec3<f32>) -> vec3<f32> {
+fn coords_to_cube_atlas(tex_coord: vec3<f32>) -> vec3<f32> {
     let abs_x = abs(tex_coord.x);
     let abs_y = abs(tex_coord.y);
     let abs_z = abs(tex_coord.z);
@@ -234,14 +244,7 @@ fn get_side_idx(tex_coord: vec3<f32>) -> vec3<f32> {
     }
 
     res = (res / abs(major_axis) + 1.0) * 0.5;
-    //res = normalize(res);
-    //res.y = 1.0-res.y; // invert y because wgsl
-    //let t = res.x;
-    //res.x = res.y;
-    
-    //res.y = 1.0 - t;
     res.y = 1.0 - res.y;
-    //res.x = 1.0 - res.x;
 
     return vec3<f32>(res, f32(cube_idx));
 }
@@ -264,10 +267,11 @@ fn calc_shadow_dir_light(normal: vec3<f32>, light_dir: vec3<f32>, frag_pos_light
     );
     
     // use a bias to avoid shadow acne
-    let bias = max(0.05 * (1.0 - dot(normal, light_dir)), 0.005);  
+    let bias = 0.005;//max(0.05 * (1.0 - dot(normal, light_dir)), 0.005);  
     let current_depth = proj_coords.z - bias;
 
-    var shadow = pcf_dir_light(region_coords, current_depth, shadow_u);
+    //var shadow = pcf_dir_light(region_coords, current_depth, shadow_u, 1.0);
+    var shadow = pcss_dir_light(xy_remapped, current_depth, shadow_u);
 
     // dont cast shadows outside the light's far plane
     if (proj_coords.z > 1.0) {
@@ -282,35 +286,87 @@ fn calc_shadow_dir_light(normal: vec3<f32>, light_dir: vec3<f32>, frag_pos_light
     return shadow;
 }
 
+// Comes from https://developer.download.nvidia.com/whitepapers/2008/PCSS_Integration.pdf
+fn search_width(light_near: f32, uv_light_size: f32, receiver_depth: f32) -> f32 {
+    return uv_light_size * (receiver_depth - light_near) / receiver_depth;
+}
+
+/// Convert texture coords to be texture coords of an atlas frame.
+fn to_atlas_frame_coords(shadow_u: LightShadowMapUniform, coords: vec2<f32>) -> vec2<f32> {
+    let atlas_dimensions = textureDimensions(t_shadow_maps_atlas);
+    
+    // get the rect of the frame as a vec4
+    var region_rect = vec4<f32>(f32(shadow_u.atlas_frame.x), f32(shadow_u.atlas_frame.y), f32(shadow_u.atlas_frame.width), f32(shadow_u.atlas_frame.height));
+    // put the frame rect in atlas UV space
+    region_rect /= f32(atlas_dimensions.x);
+    
+    // lerp input coords 
+    let region_coords = vec2<f32>(
+        mix(region_rect.x, region_rect.x + region_rect.z, coords.x),
+        mix(region_rect.y, region_rect.y + region_rect.w, coords.y)
+    );
+    
+    return region_coords;
+}
+
+/// Find the average blocker distance for a directiona llight
+fn find_blocker_distance_dir_light(tex_coords: vec2<f32>, receiver_depth: f32, bias: f32, shadow_u: LightShadowMapUniform) -> vec2<f32> {
+    let search_width = search_width(shadow_u.near_plane, shadow_u.light_size_uv, receiver_depth);
+
+    var blockers = 0;
+    var avg_dist = 0.0;
+    let samples = i32(u_shadow_settings.pcss_blocker_search_samples);
+    for (var i = 0; i < samples; i++) {
+        let offset_coords = tex_coords + u_pcss_poisson_disc[i] * search_width;
+        let new_coords = to_atlas_frame_coords(shadow_u, offset_coords);
+        let z = textureSampleLevel(t_shadow_maps_atlas, s_shadow_maps_atlas, new_coords, 0.0);
+
+        if z < (receiver_depth - bias) {
+            blockers += 1;
+            avg_dist += z;
+        }
+    }
+
+    let b = f32(blockers);
+    return vec2<f32>(avg_dist / b, b);
+}
+
+fn pcss_dir_light(tex_coords: vec2<f32>, receiver_depth: f32, shadow_u: LightShadowMapUniform) -> f32 {
+    let blocker_search = find_blocker_distance_dir_light(tex_coords, receiver_depth, 0.0, shadow_u);
+
+    // If no blockers were found, exit now to save in filtering
+    if blocker_search.y == 0.0 {
+        return 1.0;
+    }
+    let blocker_depth = blocker_search.x;
+
+    // penumbra estimation
+    let penumbra_width = (receiver_depth - blocker_depth) / blocker_depth;
+
+    // PCF
+    let uv_radius = penumbra_width * shadow_u.light_size_uv * shadow_u.near_plane / receiver_depth;
+    return pcf_dir_light(tex_coords, receiver_depth, shadow_u, uv_radius);
+}
+
 /// Calculate the shadow coefficient using PCF of a directional light
-fn pcf_dir_light(tex_coords: vec2<f32>, test_depth: f32, shadow_u: LightShadowMapUniform) -> f32 {
+fn pcf_dir_light(tex_coords: vec2<f32>, test_depth: f32, shadow_u: LightShadowMapUniform, uv_radius: f32) -> f32 {
-    let half_filter_size = f32(u_shadow_settings.pcf_samples_num) / 2.0;
-    let texel_size = 1.0 / vec2<f32>(f32(shadow_u.atlas_frame.width), f32(shadow_u.atlas_frame.height));
-
-    // Sample PCF 
     var shadow = 0.0;
-    var i = 0;
+    let samples_num = i32(u_shadow_settings.pcf_samples_num);
-    for (var x = -half_filter_size; x <= half_filter_size; x += 1.0) {
+    for (var i = 0; i < samples_num; i++) {
-        for (var y = -half_filter_size; y <= half_filter_size; y += 1.0) {
+        let offset = tex_coords + u_pcf_poisson_disc[i] * uv_radius;
-            //let random = u_pcf_poisson_disc[i] * texel_size;
+        let new_coords = to_atlas_frame_coords(shadow_u, offset);
-            let offset = tex_coords + (u_pcf_poisson_disc[i] + vec2<f32>(x, y)) * texel_size;
         
-            let pcf_depth = textureSampleCompare(t_shadow_maps_atlas, s_shadow_maps_atlas, offset, test_depth);
+        shadow += textureSampleCompare(t_shadow_maps_atlas, s_shadow_maps_atlas_compare, new_coords, test_depth);
-            shadow += pcf_depth;
-
-            i++;
     }
-    }
+    shadow /= f32(samples_num);
-    shadow /= pow(f32(u_shadow_settings.pcf_samples_num), 2.0);
-    // ensure the shadow value does not go above 1.0
-    shadow = min(shadow, 1.0);
 
-    return shadow;
+    // clamp shadow to [0; 1]
+    return saturate(shadow);
 }
 
 fn calc_shadow_point(world_pos: vec3<f32>, world_normal: vec3<f32>, light_dir: vec3<f32>, light: Light, atlas_dimensions: vec2<i32>) -> f32 {
     var frag_to_light = world_pos - light.position;
-    let temp = get_side_idx(normalize(frag_to_light));
+    let temp = coords_to_cube_atlas(normalize(frag_to_light));
     var coords_2d = temp.xy;
     let cube_idx = i32(temp.z);
     
@@ -342,7 +398,7 @@ fn calc_shadow_point(world_pos: vec3<f32>, world_normal: vec3<f32>, light_dir: v
     var current_depth = length(frag_to_light) - bias;
     current_depth /= u.far_plane;
 
-    var shadow = textureSampleCompare(t_shadow_maps_atlas, s_shadow_maps_atlas, coords_2d, current_depth);
+    var shadow = textureSampleCompare(t_shadow_maps_atlas, s_shadow_maps_atlas_compare, coords_2d, current_depth);
 
     return shadow;
 }