Don't keep around additional cpu copy of loaded mesh files (#7824)

### What

This is mainly a refactor of `re_renderer`'s model loading pipeline, but
as the title (== changelog entry) points out, a nice side-effect that
arose culling unnecessary memory usage which may be important for large
meshes.

@EtaLoop's attempt to add a color option to `Asset3D` (see
#7458) made it clear that the
output of the mesh loaders is really hard to work with:
Prior to this PR, they would eagerly create gpu-sided meshes and then
store them alongside an optional (but always filled-out) "cpu mesh" (in
essence the unpacked model file).

Now instead all model loading goes to a intermediate `CpuModel` which
can be rather easily post processed.
Gpu resources are then created as a separate step, consuming the
`CpuModel` (it should be trivial to create a variant that doesn't
consume if we need this in the future)


### Checklist
* [x] I have read and agree to [Contributor
Guide](https://github.com/rerun-io/rerun/blob/main/CONTRIBUTING.md) and
the [Code of
Conduct](https://github.com/rerun-io/rerun/blob/main/CODE_OF_CONDUCT.md)
* [x] I've included a screenshot or gif (if applicable)
* [x] I have tested the web demo (if applicable):
* Using examples from latest `main` build:
[rerun.io/viewer](https://rerun.io/viewer/pr/7824?manifest_url=https://app.rerun.io/version/main/examples_manifest.json)
* Using full set of examples from `nightly` build:
[rerun.io/viewer](https://rerun.io/viewer/pr/7824?manifest_url=https://app.rerun.io/version/nightly/examples_manifest.json)
* [x] The PR title and labels are set such as to maximize their
usefulness for the next release's CHANGELOG
* [x] If applicable, add a new check to the [release
checklist](https://github.com/rerun-io/rerun/blob/main/tests/python/release_checklist)!
* [x] If have noted any breaking changes to the log API in
`CHANGELOG.md` and the migration guide

- [PR Build Summary](https://build.rerun.io/pr/7824)
- [Recent benchmark results](https://build.rerun.io/graphs/crates.html)
- [Wasm size tracking](https://build.rerun.io/graphs/sizes.html)

To run all checks from `main`, comment on the PR with `@rerun-bot
full-check`.

crates/viewer/re_renderer/src/importer/cpu_model.rs

@@ -0,0 +1,95 @@
+use std::sync::Arc;
+
+use slotmap::{SecondaryMap, SlotMap};
+
+use crate::{
+    mesh::{CpuMesh, GpuMesh, MeshError},
+    renderer::GpuMeshInstance,
+    RenderContext,
+};
+
+slotmap::new_key_type! {
+    /// Key for identifying a cpu mesh in a model.
+    pub struct CpuModelMeshKey;
+}
+
+/// Like [`GpuMeshInstance`], but for CPU sided usage in a [`CpuModel`] only.
+pub struct CpuMeshInstance {
+    pub mesh: CpuModelMeshKey,
+    pub world_from_mesh: glam::Affine3A,
+    // TODO(andreas): Expose other properties we have on [`GpuMeshInstance`].
+}
+
+/// A collection of meshes & mesh instances on the CPU.
+///
+/// Note that there is currently no `GpuModel` equivalent, since
+/// [`GpuMeshInstance`]es use shared ownership of [`GpuMesh`]es.
+///
+/// This is the output of a model loader and is ready to be converted into
+/// a series of [`GpuMeshInstance`]s that can be rendered.
+///
+/// This is meant as a useful intermediate structure for doing post-processing steps on the model prior to gpu upload.
+#[derive(Default)]
+pub struct CpuModel {
+    pub meshes: SlotMap<CpuModelMeshKey, CpuMesh>,
+    pub instances: Vec<CpuMeshInstance>,
+}
+
+impl CpuModel {
+    /// Creates a new [`CpuModel`] from a single [`CpuMesh`], creating a single instance with identity transform.
+    pub fn from_single_mesh(mesh: CpuMesh) -> Self {
+        let mut model = Self::default();
+        model.add_single_instance_mesh(mesh);
+        model
+    }
+
+    /// Adds a new [`CpuMesh`] to the model, creating a single instance with identity transform.
+    pub fn add_single_instance_mesh(&mut self, mesh: CpuMesh) {
+        let mesh_key = self.meshes.insert(mesh);
+        self.instances.push(CpuMeshInstance {
+            mesh: mesh_key,
+            world_from_mesh: glam::Affine3A::IDENTITY,
+        });
+    }
+
+    pub fn calculate_bounding_box(&self) -> re_math::BoundingBox {
+        re_math::BoundingBox::from_points(
+            self.instances
+                .iter()
+                .filter_map(|mesh_instance| {
+                    self.meshes.get(mesh_instance.mesh).map(|mesh| {
+                        mesh.vertex_positions
+                            .iter()
+                            .map(|p| mesh_instance.world_from_mesh.transform_point3(*p))
+                    })
+                })
+                .flatten(),
+        )
+    }
+
+    /// Converts the entire model into a serious of mesh instances that can be rendered.
+    ///
+    /// Silently ignores:
+    /// * instances with invalid mesh keys
+    /// * unreferenced meshes
+    pub fn into_gpu_meshes(self, ctx: &RenderContext) -> Result<Vec<GpuMeshInstance>, MeshError> {
+        let mut gpu_meshes = SecondaryMap::with_capacity(self.meshes.len());
+        for (mesh_key, mesh) in &self.meshes {
+            gpu_meshes.insert(mesh_key, Arc::new(GpuMesh::new(ctx, mesh)?));
+        }
+
+        Ok(self
+            .instances
+            .into_iter()
+            .filter_map(|instance| {
+                Some(GpuMeshInstance {
+                    gpu_mesh: gpu_meshes.get(instance.mesh)?.clone(),
+                    world_from_mesh: instance.world_from_mesh,
+                    additive_tint: Default::default(),
+                    outline_mask_ids: Default::default(),
+                    picking_layer_id: Default::default(),
+                })
+            })
+            .collect())
+    }
+}

crates/viewer/re_renderer/src/importer/gltf.rs

@@ -1,15 +1,12 @@
-use std::sync::Arc;
-
 use ahash::{HashMap, HashMapExt};
 use gltf::texture::WrappingMode;
 use itertools::Itertools;
 use smallvec::SmallVec;
 
 use crate::{
-    mesh::{GpuMesh, Material, Mesh, MeshError},
-    renderer::MeshInstance,
+    mesh::{CpuMesh, Material, MeshError},
     resource_managers::{GpuTexture2D, ImageDataDesc, TextureManager2D},
-    RenderContext, Rgba32Unmul,
+    CpuMeshInstance, CpuModel, CpuModelMeshKey, RenderContext, Rgba32Unmul,
 };
 
 #[derive(thiserror::Error, Debug)]
@@ -41,7 +38,7 @@ pub fn load_gltf_from_buffer(
     mesh_name: &str,
     buffer: &[u8],
     ctx: &RenderContext,
-) -> Result<Vec<MeshInstance>, GltfImportError> {
+) -> Result<CpuModel, GltfImportError> {
     re_tracing::profile_function!();
 
     let (doc, buffers, images) = {
@@ -105,25 +102,28 @@ pub fn load_gltf_from_buffer(
         });
     }
 
-    let mut meshes = HashMap::with_capacity(doc.meshes().len());
+    let mut re_model = CpuModel::default();
+    let mut mesh_keys = HashMap::with_capacity(doc.meshes().len());
     for ref mesh in doc.meshes() {
         re_tracing::profile_scope!("mesh");
 
         let re_mesh = import_mesh(mesh, &buffers, &images_as_textures, &ctx.texture_manager_2d)?;
-        meshes.insert(
-            mesh.index(),
-            (Arc::new(GpuMesh::new(ctx, &re_mesh)?), Arc::new(re_mesh)),
-        );
+        let re_mesh_key = re_model.meshes.insert(re_mesh);
+        mesh_keys.insert(mesh.index(), re_mesh_key);
     }
 
-    let mut instances = Vec::new();
     for scene in doc.scenes() {
         for node in scene.nodes() {
-            gather_instances_recursive(&mut instances, &node, &glam::Affine3A::IDENTITY, &meshes);
+            gather_instances_recursive(
+                &mut re_model.instances,
+                &node,
+                &glam::Affine3A::IDENTITY,
+                &mesh_keys,
+            );
         }
     }
 
-    Ok(instances)
+    Ok(re_model)
 }
 
 fn map_format(format: gltf::image::Format) -> Option<wgpu::TextureFormat> {
@@ -154,7 +154,7 @@ fn import_mesh(
     buffers: &[gltf::buffer::Data],
     gpu_image_handles: &[GpuTexture2D],
     texture_manager: &TextureManager2D, //imported_materials: HashMap<usize, Material>,
-) -> Result<Mesh, GltfImportError> {
+) -> Result<CpuMesh, GltfImportError> {
     re_tracing::profile_function!();
 
     let mesh_name = mesh.name().map_or("<unknown", |f| f).to_owned();
@@ -277,7 +277,7 @@ fn import_mesh(
         return Err(GltfImportError::NoTrianglePrimitives { mesh_name });
     }
 
-    let mesh = Mesh {
+    let mesh = CpuMesh {
         label: mesh.name().into(),
         triangle_indices,
         vertex_positions,
@@ -293,10 +293,10 @@ fn import_mesh(
 }
 
 fn gather_instances_recursive(
-    instances: &mut Vec<MeshInstance>,
+    instances: &mut Vec<CpuMeshInstance>,
     node: &gltf::Node<'_>,
     transform: &glam::Affine3A,
-    meshes: &HashMap<usize, (Arc<GpuMesh>, Arc<Mesh>)>,
+    meshes: &HashMap<usize, CpuModelMeshKey>,
 ) {
     let (scale, rotation, translation) = match node.transform() {
         gltf::scene::Transform::Matrix { matrix } => {
@@ -324,11 +324,11 @@ fn gather_instances_recursive(
     }
 
     if let Some(mesh) = node.mesh() {
-        if let Some((gpu_mesh, mesh)) = meshes.get(&mesh.index()) {
-            let mut gpu_mesh =
-                MeshInstance::new_with_cpu_mesh(gpu_mesh.clone(), Some(mesh.clone()));
-            gpu_mesh.world_from_mesh = transform;
-            instances.push(gpu_mesh);
+        if let Some(mesh_key) = meshes.get(&mesh.index()) {
+            instances.push(CpuMeshInstance {
+                mesh: *mesh_key,
+                world_from_mesh: transform,
+            });
         }
     }
 }

crates/viewer/re_renderer/src/importer/mod.rs

@@ -1,3 +1,5 @@
+mod cpu_model;
+
 #[cfg(feature = "import-obj")]
 pub mod obj;
 
@@ -7,31 +9,4 @@ pub mod gltf;
 #[cfg(feature = "import-stl")]
 pub mod stl;
 
-use re_math::Vec3Ext as _;
-
-use crate::renderer::MeshInstance;
-
-pub fn to_uniform_scale(scale: glam::Vec3) -> f32 {
-    if scale.has_equal_components(0.001) {
-        scale.x
-    } else {
-        let uniform_scale = (scale.x * scale.y * scale.z).cbrt();
-        re_log::warn!("mesh has non-uniform scale ({:?}). This is currently not supported. Using geometric mean {}", scale,uniform_scale);
-        uniform_scale
-    }
-}
-
-pub fn calculate_bounding_box(instances: &[MeshInstance]) -> re_math::BoundingBox {
-    re_math::BoundingBox::from_points(
-        instances
-            .iter()
-            .filter_map(|mesh_instance| {
-                mesh_instance.mesh.as_ref().map(|mesh| {
-                    mesh.vertex_positions
-                        .iter()
-                        .map(|p| mesh_instance.world_from_mesh.transform_point3(*p))
-                })
-            })
-            .flatten(),
-    )
-}
+pub use cpu_model::{CpuMeshInstance, CpuModel, CpuModelMeshKey};