Skinned extraction speedup (#4428)

# Objective

- While animating 501 https://github.com/KhronosGroup/glTF-Sample-Models/tree/master/2.0/BrainStem, I noticed things were getting a little slow
- Looking in tracy, the system `extract_skinned_meshes` is taking a lot of time, with a mean duration of 15.17ms

## Solution

- ~~Use `Vec` instead of a `SmallVec`~~
- ~~Don't use an temporary variable~~
- Compute the affine matrix as an `Affine3A` instead
- Remove the `temp` vec

| |mean|
|---|---|
|base|15.17ms|
|~~vec~~|~~9.31ms~~|
|~~no temp variable~~|~~11.31ms~~|
|removing the temp vector|8.43ms|
|affine|13.21ms|
|all together|7.23ms|

crates/bevy_pbr/Cargo.toml

@@ -29,4 +29,3 @@ bevy_window = { path = "../bevy_window", version = "0.7.0-dev" }
 bitflags = "1.2"
 # direct dependency required for derive macro
 bytemuck = { version = "1", features = ["derive"] }
-smallvec = "1.0"

crates/bevy_pbr/src/render/mesh.rs

@@ -26,7 +26,6 @@ use bevy_render::{
     RenderApp, RenderStage,
 };
 use bevy_transform::components::GlobalTransform;
-use smallvec::SmallVec;
 use std::num::NonZeroU64;
 
 #[derive(Default)]
@@ -192,15 +191,16 @@ impl SkinnedMeshJoints {
         let inverse_bindposes = inverse_bindposes.get(&skin.inverse_bindposes)?;
         let bindposes = inverse_bindposes.iter();
         let skin_joints = skin.joints.iter();
-        let mut temp =
+        let start = buffer.len();
-            SmallVec::<[Mat4; MAX_JOINTS]>::with_capacity(bindposes.len().min(MAX_JOINTS));
         for (inverse_bindpose, joint) in bindposes.zip(skin_joints).take(MAX_JOINTS) {
-            let joint_matrix = joints.get(*joint).ok()?.compute_matrix();
+            if let Ok(joint) = joints.get(*joint) {
-            temp.push(joint_matrix * *inverse_bindpose);
+                buffer.push(joint.compute_affine() * *inverse_bindpose);
+            } else {
+                buffer.truncate(start);
+                return None;
+            }
         }
 
-        let start = buffer.len();
-        buffer.extend(temp);
         // Pad to 256 byte alignment
         while buffer.len() % 4 != 0 {
             buffer.push(Mat4::ZERO);

crates/bevy_transform/src/components/global_transform.rs

@@ -1,6 +1,6 @@
 use super::Transform;
 use bevy_ecs::{component::Component, reflect::ReflectComponent};
-use bevy_math::{const_vec3, Mat3, Mat4, Quat, Vec3};
+use bevy_math::{const_vec3, Affine3A, Mat3, Mat4, Quat, Vec3};
 use bevy_reflect::Reflect;
 use std::ops::Mul;
 
@@ -128,6 +128,13 @@ impl GlobalTransform {
         Mat4::from_scale_rotation_translation(self.scale, self.rotation, self.translation)
     }
 
+    /// Returns the 3d affine transformation from this transforms translation,
+    /// rotation, and scale.
+    #[inline]
+    pub fn compute_affine(&self) -> Affine3A {
+        Affine3A::from_scale_rotation_translation(self.scale, self.rotation, self.translation)
+    }
+
     /// Get the unit vector in the local x direction
     #[inline]
     pub fn local_x(&self) -> Vec3 {