Hack depth offset in order to make objects not disappear (#990)

None of the things here is great, but it solves the problem at hand.
* smaller depth offset steps (experimented in tracking_hf & colmap with history on the image itself)
* bypass far z culling completely (far away layered images will "loose" more and more layers, but they won't disappear)
* limit image plane distance to not be too insane
* ensure use of f32 depth buffer for more consistency in this whole mess (precision was essentially unspecified so far)

crates/re_renderer/shader/utils/depth_offset.wgsl

@@ -4,13 +4,19 @@
 fn apply_depth_offset(position: Vec4, offset: f32) -> Vec4 {
     // Z buffer z is computed using position.z/position.w,
     // Therefore, to affect the final output by a given offset we need to multiply it with position.w.
-    // (This also means that we're loosing some precision!)
+    // This also means though that we're loosing a lot of precision
     //
     // We're using inverse z, i.e. 0.0 is far, 1.0 is near.
     // We want a positive depth_offset_factor 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 stilil 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!
     return Vec4(
         position.xy,
-        position.z + frame.depth_offset_factor * offset * position.w,
+        max(position.z + frame.depth_offset_factor * offset * position.w, f32eps),
         position.w
     );
 }

crates/re_renderer/src/view_builder.rs

@@ -172,11 +172,11 @@ impl ViewBuilder {
 
     /// Depth format used for the main target of the view builder.
     ///
-    /// Consider using 32bit float depth buffer in the future if depth issues come up.
-    /// It is widely supported and has the best possible precision (with reverse infinite z projection which we're already using).
-    /// However, performance advice is still to use 24 bit depth buffering, see [Nvidia's Vulkan dos and dont's](https://developer.nvidia.com/blog/vulkan-dos-donts/):
-    /// > Don’t use 32-bit floating point depth formats, due to the performance cost, unless improved precision is actually required.
-    pub const MAIN_TARGET_DEPTH_FORMAT: wgpu::TextureFormat = wgpu::TextureFormat::Depth24Plus;
+    /// [`wgpu::TextureFormat::Depth24Plus`] would be preferable for performance, see [Nvidia's Vulkan dos and dont's](https://developer.nvidia.com/blog/vulkan-dos-donts/).
+    /// However, the problem with being "24Plus" is that we no longer know what format we'll actually get, which is a problem e.g. for vertex shader determined depth offsets.
+    /// (This is a real concern - for example on Metal we always get a floating point target with this!)
+    /// [`wgpu::TextureFormat::Depth32Float`] on the other hand is widely supported and has the best possible precision (with reverse infinite z projection which we're already using).
+    pub const MAIN_TARGET_DEPTH_FORMAT: wgpu::TextureFormat = wgpu::TextureFormat::Depth32Float;
 
     /// Enable MSAA always. This makes our pipeline less variable as well, as we need MSAA resolve steps if we want any MSAA at all!
     ///
@@ -385,7 +385,7 @@ impl ViewBuilder {
 
         // See `depth_offset.wgsl`.
         // Factor applied to depth offsets.
-        let depth_offset_factor = f32::EPSILON;
+        let depth_offset_factor = 1.0e-08; // Value determined by experimentation. Quite close to the f32 machine epsilon but a bit lower.
 
         ctx.queue.write_buffer(
             ctx.gpu_resources

crates/re_viewer/src/ui/selection_panel.rs

@@ -356,7 +356,7 @@ fn obj_props_ui(
                     if ui
                         .add(
                             egui::DragValue::new(&mut distance)
-                                .clamp_range(0.0..=f32::INFINITY)
+                                .clamp_range(0.0..=1.0e8)
                                 .speed(speed),
                         )
                         .on_hover_text("Controls how far away the image plane is.")