Improve dithering CPU usage (#866)

CHANGELOG.md

@@ -9,6 +9,7 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ### Changed
 - Include build profile in the displayed version information
+- [playback] Improve dithering CPU usage by about 33%
 
 ### Fixed
 - [connect] Partly fix behavior after last track of an album/playlist

Cargo.lock

@@ -1288,6 +1288,7 @@ dependencies = [
  "portaudio-rs",
  "rand",
  "rand_distr",
+ "rand_xoshiro",
  "rodio",
  "sdl2",
  "shell-words",
@@ -1881,9 +1882,9 @@ dependencies = [
 
 [[package]]
 name = "rand_distr"
-version = "0.4.1"
+version = "0.4.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "051b398806e42b9cd04ad9ec8f81e355d0a382c543ac6672c62f5a5b452ef142"
+checksum = "964d548f8e7d12e102ef183a0de7e98180c9f8729f555897a857b96e48122d2f"
 dependencies = [
  "num-traits",
  "rand",
@@ -1898,6 +1899,15 @@ dependencies = [
  "rand_core",
 ]
 
+[[package]]
+name = "rand_xoshiro"
+version = "0.6.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "6f97cdb2a36ed4183de61b2f824cc45c9f1037f28afe0a322e9fff4c108b5aaa"
+dependencies = [
+ "rand_core",
+]
+
 [[package]]
 name = "redox_syscall"
 version = "0.2.10"

playback/Cargo.toml

@@ -49,6 +49,7 @@ ogg = "0.8"
 # Dithering
 rand = "0.8"
 rand_distr = "0.4"
+rand_xoshiro = "0.6"
 
 [features]
 alsa-backend = ["alsa"]

playback/src/dither.rs

@@ -1,4 +1,4 @@
-use rand::rngs::ThreadRng;
+use rand::SeedableRng;
 use rand_distr::{Distribution, Normal, Triangular, Uniform};
 use std::fmt;
 
@@ -41,20 +41,36 @@ impl fmt::Display for dyn Ditherer {
     }
 }
 
-// Implementation note: we save the handle to ThreadRng so it doesn't require
+// `SmallRng` is 33% faster than `ThreadRng`, but we can do even better.
-// a lookup on each call (which is on each sample!). This is ~2.5x as fast.
+// `SmallRng` defaults to `Xoshiro256PlusPlus` on 64-bit platforms and
-// Downside is that it is not Send so we cannot move it around player threads.
+// `Xoshiro128PlusPlus` on 32-bit platforms. These are excellent for the
+// general case. In our case of just 64-bit floating points, we can make
+// some optimizations. Compared to `SmallRng`, these hand-picked generators
+// improve performance by another 9% on 64-bit platforms and 2% on 32-bit
+// platforms.
 //
+// For reference, see https://prng.di.unimi.it. Note that we do not use
+// `Xoroshiro128Plus` or `Xoshiro128Plus` because they display low linear
+// complexity in the lower four bits, which is not what we want:
+// linearization is the very point of dithering.
+#[cfg(target_pointer_width = "64")]
+type Rng = rand_xoshiro::Xoshiro256Plus;
+#[cfg(not(target_pointer_width = "64"))]
+type Rng = rand_xoshiro::Xoshiro128StarStar;
+
+fn create_rng() -> Rng {
+    Rng::from_entropy()
+}
 
 pub struct TriangularDitherer {
-    cached_rng: ThreadRng,
+    cached_rng: Rng,
     distribution: Triangular<f64>,
 }
 
 impl Ditherer for TriangularDitherer {
     fn new() -> Self {
         Self {
-            cached_rng: rand::thread_rng(),
+            cached_rng: create_rng(),
             // 2 LSB peak-to-peak needed to linearize the response:
             distribution: Triangular::new(-1.0, 1.0, 0.0).unwrap(),
         }
@@ -74,14 +90,14 @@ impl TriangularDitherer {
 }
 
 pub struct GaussianDitherer {
-    cached_rng: ThreadRng,
+    cached_rng: Rng,
     distribution: Normal<f64>,
 }
 
 impl Ditherer for GaussianDitherer {
     fn new() -> Self {
         Self {
-            cached_rng: rand::thread_rng(),
+            cached_rng: create_rng(),
             // 1/2 LSB RMS needed to linearize the response:
             distribution: Normal::new(0.0, 0.5).unwrap(),
         }
@@ -103,7 +119,7 @@ impl GaussianDitherer {
 pub struct HighPassDitherer {
     active_channel: usize,
     previous_noises: [f64; NUM_CHANNELS],
-    cached_rng: ThreadRng,
+    cached_rng: Rng,
     distribution: Uniform<f64>,
 }
 
@@ -112,7 +128,7 @@ impl Ditherer for HighPassDitherer {
         Self {
             active_channel: 0,
             previous_noises: [0.0; NUM_CHANNELS],
-            cached_rng: rand::thread_rng(),
+            cached_rng: create_rng(),
             distribution: Uniform::new_inclusive(-0.5, 0.5), // 1 LSB +/- 1 LSB (previous) = 2 LSB
         }
     }