Fix web depth/projection regression, causing incorrect rendering on a…

…ll 3D scenes (#2170)

* Revert "Fix depth precision issues on WebGL due to different NDC space (#2123)"

This reverts commit 4f60fd7.

* fudge depth offsets on Webgl

* hardware_tier global constant, better depth handling for gles

* Improve the depth offset for all platforms

---------

Co-authored-by: Emil Ernerfeldt <emil.ernerfeldt@gmail.com>

crates/re_renderer/shader/global_bindings.wgsl

@@ -27,7 +27,11 @@ struct FrameUniformBuffer {
 
     // Size used for all line radii given with Size::AUTO.
     auto_size_lines: f32,
+
+    /// re_renderer defined hardware tier.
+    hardware_tier: u32,
 };
+
 @group(0) @binding(0)
 var<uniform> frame: FrameUniformBuffer;
 
@@ -36,3 +40,7 @@ var nearest_sampler: sampler;
 
 @group(0) @binding(2)
 var trilinear_sampler: sampler;
+
+// See config.rs#HardwareTier
+const HARDWARE_TIER_GLES = 0u;
+const HARDWARE_TIER_WEBGPU = 1u;

crates/re_renderer/shader/utils/depth_offset.wgsl

@@ -1,35 +1,59 @@
 #import <../global_bindings.wgsl>
 #import <../types.wgsl>
 
+/*
+We use reverse infinite depth, as promoted by https://developer.nvidia.com/content/depth-precision-visualized
+
+The projection matrix (from `glam::Mat4::perspective_infinite_reverse_rh`) looks like this:
+
+f / aspect_ratio   0     0      0
+0                  f     0      0
+0                  0     0      z_near
+0                  0    -1      0
+
+This means after multiplication with xyzw (with w=1) we end up with:
+
+    x_proj: x * f / aspect_ratio,
+    y_proj: y * f,
+    z_proj: w * z_near,
+    w_proj: -z
+
+This is then projected by dividing with w, giving:
+
+    x_ndc: x_proj / w_proj
+    y_ndc: y_proj / w_proj
+    z_ndc: z_proj / w_proj
+
+Without z offset, we get this:
+
+    x_ndc: x * f / aspect_ratio / -z
+    y_ndc: y * f / -z
+    z_ndc: w * z_near / -z
+
+The negative -z axis is away from the camera, so with w=1 we get
+z_near mapping to z_ndc=1, and infinity mapping to z_ndc=0.
+
+The code below act on the *_proj values by adding a scale multiplier on `w_proj` resulting in:
+    x_ndc: x_proj / (-z * w_scale)
+    y_ndc: y_proj / (-z * w_scale)
+    z_ndc: z_proj / (-z * w_scale)
+*/
+
 fn apply_depth_offset(position: Vec4, offset: f32) -> Vec4 {
-    // We're using inverse z, i.e. 0.0 is far, 1.0 is near.
-    // We want a positive offset to move towards the viewer, so offset needs to be added.
-    //
-    // With this in place we still may cross over to 0.0 (the far plane) too early,
-    // making objects disappear into the far when they'd be otherwise still rendered.
-    // Since we're actually supposed to have an *infinite* far plane this should never happen!
-    // Therefore we simply dictacte a minimum z value.
-    // This ofc wrecks the depth offset and may cause z fighting with all very far away objects, but it's better than having things disappear!
-
-    if true {
-        // This path assumes a `f32` depth buffer!
-
-        // 1.0 * eps _should_ be enough, but in practice it causes Z-fighting for unknown reasons.
-        // Maybe because of GPU interpolation of vertex coordinates?
-        let eps = 5.0 * f32eps;
-
-        return Vec4(
-            position.xy,
-            max(position.z * (1.0 + eps * offset), f32eps),
-            position.w
-        );
-    } else {
-        // Causes Z-collision at far distances
-        let eps = f32eps;
-        return Vec4(
-            position.xy,
-            max(position.z + eps * offset * position.w, f32eps),
-            position.w
-        );
-    }
+    // On GLES/WebGL, the NDC clipspace range for depth is from -1 to 1 and y is flipped.
+    // wgpu/Naga counteracts this by patching all vertex shaders with:
+    // "gl_Position.yz = vec2(-gl_Position.y, gl_Position.z * 2.0 - gl_Position.w);",
+    // This doesn't matter for us though.
+
+    // This path assumes a `f32` depth buffer.
+
+    // We set up the depth comparison to Greater, so that large z means closer (overdraw).
+    // We want a greater offset to win over a smaller offset,
+    // so a great depth offset should result in a large z_ndc.
+    // How do we get there? We let large depth offset lead to a smaller divisor (w_proj):
+
+    return Vec4(
+        position.xyz,
+        position.w * (1.0 - f32eps * offset),
+    );
 }

crates/re_renderer/src/config.rs

@@ -7,16 +7,18 @@
 /// but choosing lower tiers is always possible.
 /// Tiers may loosely relate to quality settings, but their primary function is an easier way to
 /// do bundle feature *support* checks.
+///
+/// See also `global_bindings.wgsl`
 #[derive(Clone, Copy, Debug, PartialEq, Eq)]
 pub enum HardwareTier {
     /// Limited feature support as provided by WebGL and native GLES2/OpenGL3(ish).
-    Gles,
+    Gles = 0,
 
     /// Full support of WebGPU spec without additional feature requirements.
     ///
     /// Expecting to run either in a stable WebGPU implementation.
     /// I.e. either natively with Vulkan/Metal or in a browser with WebGPU support.
-    FullWebGpuSupport,
+    FullWebGpuSupport = 1,
     // Run natively with Vulkan/Metal and require additional features.
     //HighEnd
 }

crates/re_renderer/src/global_bindings.rs

@@ -45,8 +45,8 @@ pub struct FrameUniformBuffer {
     // Size used for all line radii given with Size::AUTO.
     pub auto_size_lines: f32,
 
-    /// Factor used to compute depth offsets, see `depth_offset.wgsl`.
-    pub end_padding: wgpu_buffer_types::PaddingRow,
+    /// re_renderer defined hardware tier.
+    pub hardware_tier: wgpu_buffer_types::U32RowPadded,
 }
 
 pub(crate) struct GlobalBindings {

crates/re_renderer/src/view_builder.rs

@@ -4,7 +4,6 @@ use std::sync::Arc;
 
 use crate::{
     allocator::{create_and_fill_uniform_buffer, GpuReadbackIdentifier},
-    config::HardwareTier,
     context::RenderContext,
     draw_phases::{
         DrawPhase, OutlineConfig, OutlineMaskProcessor, PickingLayerError, PickingLayerProcessor,
@@ -331,30 +330,11 @@ impl ViewBuilder {
                     // We use infinite reverse-z projection matrix
                     // * great precision both with floating point and integer: https://developer.nvidia.com/content/depth-precision-visualized
                     // * no need to worry about far plane
-                    //
-                    // There's a pretty big catch though:
-                    // When we're on GLES, the NDC coordinates go from -1 to 1 and are then mapped to 0 to 1 as part of the viewport transformation.
-                    // This means, that if we don't care about this, we'll not only throw away "half" of the space,
-                    // we also have the most precise region around 0 in the wrong spot initially 😱
-                    // Therefore, we pretty much *have* to use a different projection in this case.
-                    // Modern OpenGL has a way to change this to the more common 0 to 1 z-range for NDC
-                    // which is the standard in Metal/Vulkan/DirectX/WebGPU, but WebGL is lacking this.
-                    // see https://registry.khronos.org/OpenGL-Refpages/gl4/html/glClipControl.xhtml
-                    let projection_from_view =
-                        if ctx.shared_renderer_data.config.hardware_tier == HardwareTier::Gles {
-                            glam::Mat4::perspective_rh_gl(
-                                vertical_fov,
-                                aspect_ratio,
-                                near_plane_distance,
-                                100000.0,
-                            )
-                        } else {
-                            glam::Mat4::perspective_infinite_reverse_rh(
-                                vertical_fov,
-                                aspect_ratio,
-                                near_plane_distance,
-                            )
-                        };
+                    let projection_from_view = glam::Mat4::perspective_infinite_reverse_rh(
+                        vertical_fov,
+                        aspect_ratio,
+                        near_plane_distance,
+                    );
 
                     // Calculate ratio between screen size and screen distance.
                     // Great for getting directions from normalized device coordinates.
@@ -391,31 +371,26 @@ impl ViewBuilder {
                         config.resolution_in_pixel[0] as f32 / config.resolution_in_pixel[1] as f32;
                     let horizontal_world_size = vertical_world_size * aspect_ratio;
                     // Note that we inverse z (by swapping near and far plane) to be consistent with our perspective projection.
-                    let (left, right, bottom, top, near, far) = match camera_mode {
-                        OrthographicCameraMode::NearPlaneCenter => (
+                    let projection_from_view = match camera_mode {
+                        OrthographicCameraMode::NearPlaneCenter => glam::Mat4::orthographic_rh(
                             -0.5 * horizontal_world_size,
                             0.5 * horizontal_world_size,
                             -0.5 * vertical_world_size,
                             0.5 * vertical_world_size,
                             far_plane_distance,
                             0.0,
                         ),
-                        OrthographicCameraMode::TopLeftCornerAndExtendZ => (
-                            0.0,
-                            horizontal_world_size,
-                            vertical_world_size,
-                            0.0,
-                            far_plane_distance,
-                            -far_plane_distance,
-                        ),
+                        OrthographicCameraMode::TopLeftCornerAndExtendZ => {
+                            glam::Mat4::orthographic_rh(
+                                0.0,
+                                horizontal_world_size,
+                                vertical_world_size,
+                                0.0,
+                                far_plane_distance,
+                                -far_plane_distance,
+                            )
+                        }
                     };
-                    let projection_from_view =
-                        if ctx.shared_renderer_data.config.hardware_tier == HardwareTier::Gles {
-                            // See comment on perspective projection for why we need to care about Gles vs non-Gles.
-                            glam::Mat4::orthographic_rh_gl(left, right, bottom, top, near, far)
-                        } else {
-                            glam::Mat4::orthographic_rh(left, right, bottom, top, near, far)
-                        };
 
                     let tan_half_fov = glam::vec2(f32::MAX, f32::MAX);
                     let pixel_world_size_from_camera_distance = vertical_world_size
@@ -482,7 +457,7 @@ impl ViewBuilder {
             auto_size_points: auto_size_points.0,
             auto_size_lines: auto_size_lines.0,
 
-            end_padding: Default::default(),
+            hardware_tier: (ctx.shared_renderer_data.config.hardware_tier as u32).into(),
         };
         let frame_uniform_buffer = create_and_fill_uniform_buffer(
             ctx,