re_renderer: implement depth cloud renderer

crates/re_renderer/examples/multiview.rs

@@ -207,7 +207,7 @@ impl Example for Multiview {
             .collect_vec();
 
         let model_mesh_instances = {
-            let reader = std::io::Cursor::new(include_bytes!("rerun.obj.zip"));
+            let reader = std::io::Cursor::new(include_bytes!("assets/rerun.obj.zip"));
             let mut zip = zip::ZipArchive::new(reader).unwrap();
             let mut zipped_obj = zip.by_name("rerun.obj").unwrap();
             let mut obj_data = Vec::new();

crates/re_renderer/shader/composite.wgsl

@@ -17,7 +17,7 @@ fn main(in: VertexOutput) -> @location(0) Vec4 {
     // but are about the location of the texel in the target texture.
     var input = textureSample(input_texture, nearest_sampler, in.texcoord).rgb;
     // TODO(andreas): Do something meaningful with values above 1
-    input = clamp(input, ZERO, ONE);
+    input = clamp(input, ZERO.xyz, ONE.xyz);
 
     // Convert to srgb - this is necessary since the final eframe output does *not* have an srgb format.
     // Note that the input here is assumed to be linear - if the input texture was an srgb texture it would have been converted on load.

crates/re_renderer/shader/depth_cloud.wgsl

@@ -0,0 +1,123 @@
+//! Renders a point cloud from a depth texture and a set of intrinsics.
+//!
+//! See `src/renderer/depth_cloud.rs` for more documentation.
+
+#import <./global_bindings.wgsl>
+#import <./types.wgsl>
+#import <./utils/camera.wgsl>
+#import <./utils/flags.wgsl>
+#import <./utils/size.wgsl>
+#import <./utils/sphere_quad.wgsl>
+#import <./utils/srgb.wgsl>
+
+// ---
+
+struct PointData {
+    pos_in_world: Vec3,
+    unresolved_radius: f32,
+    color: Vec4
+}
+
+// Backprojects the depth texture using the intrinsics passed in the uniform buffer.
+fn compute_point_data(quad_idx: i32) -> PointData {
+    let wh = textureDimensions(depth_texture);
+    let texcoords = IVec2(quad_idx % wh.x, quad_idx / wh.x);
+
+    // TODO(cmc): expose knobs to linearize/normalize/flip/cam-to-plane depth.
+    let norm_linear_depth = textureLoad(depth_texture, texcoords, 0).x;
+
+    // TODO(cmc): support color maps & albedo textures
+    let color = Vec4(linear_from_srgb(Vec3(norm_linear_depth)), 1.0);
+
+    // TODO(cmc): This assumes a pinhole camera; need to support other kinds at some point.
+    let intrinsics = transpose(depth_cloud_info.depth_camera_intrinsics);
+    let focal_length = Vec2(intrinsics[0][0], intrinsics[1][1]);
+    let offset = Vec2(intrinsics[2][0], intrinsics[2][1]);
+
+    let pos_in_obj = Vec3(
+        (Vec2(texcoords) - offset) * norm_linear_depth / focal_length,
+        norm_linear_depth,
+    );
+
+    let pos_in_world = depth_cloud_info.world_from_obj * Vec4(pos_in_obj, 1.0);
+
+    var data: PointData;
+    data.pos_in_world = pos_in_world.xyz;
+    data.unresolved_radius = norm_linear_depth * depth_cloud_info.radius_scale;
+    data.color = color;
+
+    return data;
+}
+
+// ---
+
+struct DepthCloudInfo {
+    world_from_obj: Mat4,
+
+    /// The intrinsics of the camera used for the projection.
+    ///
+    /// Only supports pinhole cameras at the moment.
+    depth_camera_intrinsics: Mat3,
+
+    /// The scale to apply to the radii of the backprojected points.
+    radius_scale: f32,
+};
+@group(1) @binding(0)
+var<uniform> depth_cloud_info: DepthCloudInfo;
+
+@group(1) @binding(1)
+var depth_texture: texture_2d<f32>;
+
+struct VertexOut {
+    @builtin(position) pos_in_clip: Vec4,
+    @location(0) pos_in_world: Vec3,
+    @location(1) point_pos_in_world: Vec3,
+    @location(2) point_color: Vec4,
+    @location(3) point_radius: f32,
+};
+
+@vertex
+fn vs_main(@builtin(vertex_index) vertex_idx: u32) -> VertexOut {
+    let quad_idx = sphere_quad_index(vertex_idx);
+
+    // Compute point data (valid for the entire quad).
+    let point_data = compute_point_data(quad_idx);
+
+    // Span quad
+    let quad = sphere_quad_span(vertex_idx, point_data.pos_in_world, point_data.unresolved_radius);
+
+    var out: VertexOut;
+    out.pos_in_clip = frame.projection_from_world * Vec4(quad.pos_in_world, 1.0);
+    out.pos_in_world = quad.pos_in_world;
+    out.point_pos_in_world = point_data.pos_in_world;
+    out.point_color = point_data.color;
+    out.point_radius = quad.point_resolved_radius;
+
+    return out;
+}
+
+@fragment
+fn fs_main(in: VertexOut) -> @location(0) Vec4 {
+    // There's easier ways to compute anti-aliasing for when we are in ortho mode since it's
+    // just circles.
+    // But it's very nice to have mostly the same code path and this gives us the sphere world
+    // position along the way.
+    let ray_in_world = camera_ray_to_world_pos(in.pos_in_world);
+
+    // Sphere intersection with anti-aliasing as described by Iq here
+    // https://www.shadertoy.com/view/MsSSWV
+    // (but rearranged and labled to it's easier to understand!)
+    let d = ray_sphere_distance(ray_in_world, in.point_pos_in_world, in.point_radius);
+    let smallest_distance_to_sphere = d.x;
+    let closest_ray_dist = d.y;
+    let pixel_world_size = approx_pixel_world_size_at(closest_ray_dist);
+    if smallest_distance_to_sphere > pixel_world_size {
+        discard;
+    }
+
+    // NOTE: We only want clipping, alpha coverage and shading don't look great for this use case.
+
+    // TODO(andreas): Do we want manipulate the depth buffer depth to actually render spheres?
+
+    return vec4(in.point_color.rgb, 1.0);
+}

crates/re_renderer/shader/global_bindings.wgsl

@@ -1,5 +1,5 @@
 struct FrameUniformBuffer {
-    view_from_world: mat4x3<f32>,
+    view_from_world: Mat4x3,
     projection_from_view: Mat4,
     projection_from_world: Mat4,
 

crates/re_renderer/shader/point_cloud.wgsl

@@ -3,6 +3,7 @@
 #import <./utils/camera.wgsl>
 #import <./utils/flags.wgsl>
 #import <./utils/size.wgsl>
+#import <./utils/sphere_quad.wgsl>
 
 @group(1) @binding(0)
 var position_data_texture: texture_2d<f32>;
@@ -52,103 +53,27 @@ fn read_data(idx: i32) -> PointData {
     return data;
 }
 
-fn span_quad_perspective(
-    point_pos: Vec3,
-    point_radius: f32,
-    top_bottom: f32,
-    left_right: f32,
-    to_camera: Vec3,
-    camera_distance: f32
-) -> Vec3 {
-    let distance_to_camera_sq = camera_distance * camera_distance; // (passing on micro-optimization here for splitting this out of earlier length calculation)
-    let distance_to_camera_inv = 1.0 / camera_distance;
-    let quad_normal = to_camera * distance_to_camera_inv;
-    let quad_right = normalize(cross(quad_normal, frame.view_from_world[1].xyz)); // It's spheres so any orthogonal vector would do.
-    let quad_up = cross(quad_right, quad_normal);
-    let pos_in_quad = top_bottom * quad_up + left_right * quad_right;
-
-    // But we want to draw pretend-spheres here!
-    // If camera gets close to a sphere (or the sphere is large) then outlines of the sphere would not fit on a quad with radius r!
-    // Enlarging the quad is one solution, but then Z gets tricky (== we need to write correct Z and not quad Z to depth buffer) since we may get
-    // "unnecessary" overlaps. So instead, we change the size _and_ move the sphere closer (using math!)
-    let radius_sq = point_radius * point_radius;
-    let camera_offset = radius_sq * distance_to_camera_inv;
-    var modified_radius = point_radius * distance_to_camera_inv * sqrt(distance_to_camera_sq - radius_sq);
-
-    // We're computing a coverage mask in the fragment shader - make sure the quad doesn't cut off our antialiasing.
-    // It's fairly subtle but if we don't do this our spheres look slightly squarish
-    modified_radius += frame.pixel_world_size_from_camera_distance * camera_distance;
-
-    return point_pos + pos_in_quad * modified_radius + camera_offset * quad_normal;
-
-    // normal billboard (spheres are cut off!):
-    //      pos = point_data.pos + pos_in_quad * point_radius;
-    // only enlarged billboard (works but requires z care even for non-overlapping spheres):
-    //      modified_radius = length(toCamera) * radius / sqrt(distance_to_camera_sq - radius_sq);
-    //      pos = particleCenter + quadPosition * modified_radius;
-}
-
-fn span_quad_orthographic(point_pos: Vec3, point_radius: f32, top_bottom: f32, left_right: f32) -> Vec3 {
-    let quad_normal = frame.camera_forward;
-    let quad_right = normalize(cross(quad_normal, frame.view_from_world[1].xyz)); // It's spheres so any orthogonal vector would do.
-    let quad_up = cross(quad_right, quad_normal);
-    let pos_in_quad = top_bottom * quad_up + left_right * quad_right;
-
-    // We're computing a coverage mask in the fragment shader - make sure the quad doesn't cut off our antialiasing.
-    // It's fairly subtle but if we don't do this our spheres look slightly squarish
-    let radius = point_radius + frame.pixel_world_size_from_camera_distance;
-
-    return point_pos + pos_in_quad * radius;
-}
-
 @vertex
 fn vs_main(@builtin(vertex_index) vertex_idx: u32) -> VertexOut {
-    // Basic properties of the vertex we're at.
-    let quad_idx = i32(vertex_idx) / 6;
-    let local_idx = vertex_idx % 6u;
-    let top_bottom = f32(local_idx <= 1u || local_idx == 5u) * 2.0 - 1.0; // 1 for a top vertex, -1 for a bottom vertex.
-    let left_right = f32(vertex_idx % 2u) * 2.0 - 1.0; // 1 for a right vertex, -1 for a left vertex.
+    let quad_idx = sphere_quad_index(vertex_idx);
 
     // Read point data (valid for the entire quad)
     let point_data = read_data(quad_idx);
-    // Resolve radius to a world size. We need the camera distance for this, which is useful later on.
-    let to_camera = frame.camera_position - point_data.pos;
-    let camera_distance = length(to_camera);
-    let radius = unresolved_size_to_world(point_data.unresolved_radius, camera_distance, frame.auto_size_points);
 
     // Span quad
-    var pos: Vec3;
-    if is_camera_perspective() {
-        pos = span_quad_perspective(point_data.pos, radius, top_bottom, left_right, to_camera, camera_distance);
-    } else {
-        pos = span_quad_orthographic(point_data.pos, radius, top_bottom, left_right);
-    }
+    let quad = sphere_quad_span(vertex_idx, point_data.pos, point_data.unresolved_radius);
 
     // Output, transform to projection space and done.
     var out: VertexOut;
-    out.position = frame.projection_from_world * Vec4(pos, 1.0);
+    out.position = frame.projection_from_world * Vec4(quad.pos_in_world, 1.0);
     out.color = point_data.color;
-    out.radius = radius;
-    out.world_position = pos;
+    out.radius = quad.point_resolved_radius;
+    out.world_position = quad.pos_in_world;
     out.point_center = point_data.pos;
 
     return out;
 }
 
-
-// Returns distance to sphere surface (x) and distance to of closest ray hit (y)
-// Via https://iquilezles.org/articles/spherefunctions/ but with more verbose names.
-fn sphere_distance(ray: Ray, sphere_origin: Vec3, sphere_radius: f32) -> Vec2 {
-    let sphere_radius_sq = sphere_radius * sphere_radius;
-    let sphere_to_origin = ray.origin - sphere_origin;
-    let b = dot(sphere_to_origin, ray.direction);
-    let c = dot(sphere_to_origin, sphere_to_origin) - sphere_radius_sq;
-    let h = b * b - c;
-    let d = sqrt(max(0.0, sphere_radius_sq - h)) - sphere_radius;
-    return Vec2(d, -b - sqrt(max(h, 0.0)));
-}
-
-
 @fragment
 fn fs_main(in: VertexOut) -> @location(0) Vec4 {
     // There's easier ways to compute anti-aliasing for when we are in ortho mode since it's just circles.
@@ -158,11 +83,11 @@ fn fs_main(in: VertexOut) -> @location(0) Vec4 {
     // Sphere intersection with anti-aliasing as described by Iq here
     // https://www.shadertoy.com/view/MsSSWV
     // (but rearranged and labled to it's easier to understand!)
-    let d = sphere_distance(ray, in.point_center, in.radius);
+    let d = ray_sphere_distance(ray, in.point_center, in.radius);
     let smallest_distance_to_sphere = d.x;
     let closest_ray_dist = d.y;
     let pixel_world_size = approx_pixel_world_size_at(closest_ray_dist);
-    if  smallest_distance_to_sphere > pixel_world_size {
+    if smallest_distance_to_sphere > pixel_world_size {
         discard;
     }
     let coverage = 1.0 - saturate(smallest_distance_to_sphere / pixel_world_size);

crates/re_renderer/shader/types.wgsl

@@ -8,6 +8,8 @@ type UVec4 = vec4<u32>;
 type IVec2 = vec2<i32>;
 type IVec3 = vec3<i32>;
 type IVec4 = vec4<i32>;
+type Mat3 = mat3x3<f32>;
+type Mat4x3 = mat4x3<f32>;
 type Mat4 = mat4x4<f32>;
 
 const f32min = -3.4028235e38;
@@ -21,5 +23,5 @@ const X = Vec3(1.0, 0.0, 0.0);
 const Y = Vec3(0.0, 1.0, 0.0);
 const Z = Vec3(0.0, 0.0, 1.0);
 
-const ZERO = Vec3(0.0, 0.0, 0.0);
-const ONE  = Vec3(1.0, 1.0, 1.0);
+const ZERO = Vec4(0.0, 0.0, 0.0, 0.0);
+const ONE  = Vec4(1.0, 1.0, 1.0, 1.0);

crates/re_renderer/shader/utils/camera.wgsl

@@ -57,11 +57,22 @@ fn camera_ray_direction_from_screenuv(texcoord: Vec2) -> Vec3 {
     return normalize(world_space_dir);
 }
 
+// Returns distance to sphere surface (x) and distance to closest ray hit (y)
+// Via https://iquilezles.org/articles/spherefunctions/ but with more verbose names.
+fn ray_sphere_distance(ray: Ray, sphere_origin: Vec3, sphere_radius: f32) -> Vec2 {
+    let sphere_radius_sq = sphere_radius * sphere_radius;
+    let sphere_to_origin = ray.origin - sphere_origin;
+    let b = dot(sphere_to_origin, ray.direction);
+    let c = dot(sphere_to_origin, sphere_to_origin) - sphere_radius_sq;
+    let h = b * b - c;
+    let d = sqrt(max(0.0, sphere_radius_sq - h)) - sphere_radius;
+    return Vec2(d, -b - sqrt(max(h, 0.0)));
+}
+
 // Returns the projected size of a pixel at a given distance from the camera.
 //
 // This is accurate for objects in the middle of the screen, (depending on the angle) less so at the corners
 // since an object parallel to the camera (like a conceptual pixel) has a bigger projected surface at higher angles.
 fn approx_pixel_world_size_at(camera_distance: f32) -> f32 {
-    return select(frame.pixel_world_size_from_camera_distance,
-        camera_distance * frame.pixel_world_size_from_camera_distance, is_camera_perspective());
+    return select(frame.pixel_world_size_from_camera_distance, camera_distance * frame.pixel_world_size_from_camera_distance, is_camera_perspective());
 }