Write textures in separate thread; fixes perf

build.rs

@@ -5,7 +5,6 @@ fn main() {
     cc::Build::new()
         .include(Path::new("shrooms-vb-core/core"))
         .opt_level(2)
-        .flag_if_supported("-flto")
         .flag_if_supported("-fno-strict-aliasing")
         .file(Path::new("shrooms-vb-core/core/vb.c"))
         .compile("vb");

src/app/game.rs

@@ -1,4 +1,4 @@
-use std::{sync::Arc, time::Instant};
+use std::time::Instant;
 use wgpu::util::DeviceExt as _;
 use winit::{
     dpi::LogicalSize,
@@ -9,7 +9,7 @@ use winit::{
 
 use crate::{
     emulator::{EmulatorClient, EmulatorCommand, SimId},
-    renderer::GameRenderer,
+    graphics::TextureSink,
 };
 
 use super::{
@@ -46,15 +46,8 @@ impl GameWindow {
             .build();
         let device = &window.device;
 
-        let eyes = Arc::new(GameRenderer::create_texture(device, "eye"));
+        let (sink, texture_view) = TextureSink::new(device, window.queue.clone());
-        client.send_command(EmulatorCommand::SetRenderer(
+        client.send_command(EmulatorCommand::SetRenderer(sim_id, sink));
-            sim_id,
-            GameRenderer {
-                queue: window.queue.clone(),
-                eyes: eyes.clone(),
-            },
-        ));
-        let eyes = eyes.create_view(&wgpu::TextureViewDescriptor::default());
         let sampler = device.create_sampler(&wgpu::SamplerDescriptor::default());
         let colors = Colors {
             left: [1.0, 0.0, 0.0, 1.0],
@@ -103,7 +96,7 @@ impl GameWindow {
             entries: &[
                 wgpu::BindGroupEntry {
                     binding: 0,
-                    resource: wgpu::BindingResource::TextureView(&eyes),
+                    resource: wgpu::BindingResource::TextureView(&texture_view),
                 },
                 wgpu::BindGroupEntry {
                     binding: 1,

src/audio.rs

@@ -1,3 +1,5 @@
+use std::time::Duration;
+
 use anyhow::{bail, Result};
 use cpal::traits::{DeviceTrait, HostTrait, StreamTrait};
 use itertools::Itertools;
@@ -95,7 +97,7 @@ impl Audio {
         }
 
         while self.sample_sink.slots() < self.sampler.output_frames_max() * 2 {
-            std::hint::spin_loop();
+            std::thread::sleep(Duration::from_micros(500));
         }
     }
 }

src/emulator.rs

@@ -11,7 +11,7 @@ use std::{
 
 use anyhow::Result;
 
-use crate::{audio::Audio, renderer::GameRenderer};
+use crate::{audio::Audio, graphics::TextureSink};
 use shrooms_vb_core::Sim;
 pub use shrooms_vb_core::VBKey;
 
@@ -30,7 +30,10 @@ pub enum SimId {
     Player2,
 }
 impl SimId {
-    pub fn to_index(self) -> usize {
+    pub const fn values() -> [Self; 2] {
+        [Self::Player1, Self::Player2]
+    }
+    pub const fn to_index(self) -> usize {
         match self {
             Self::Player1 => 0,
             Self::Player2 => 1,
@@ -75,7 +78,7 @@ pub struct Emulator {
     sims: Vec<Sim>,
     audio: Audio,
     commands: mpsc::Receiver<EmulatorCommand>,
-    renderers: HashMap<SimId, GameRenderer>,
+    renderers: HashMap<SimId, TextureSink>,
     running: Arc<AtomicBool>,
     has_game: Arc<AtomicBool>,
 }
@@ -140,11 +143,17 @@ impl Emulator {
                 idle = false;
                 Sim::emulate_many(&mut self.sims);
             }
-            for (sim_id, renderer) in self.renderers.iter_mut() {
+            for sim_id in SimId::values() {
-                if let Some(sim) = self.sims.get_mut(sim_id.to_index()) {
+                let Some(renderer) = self.renderers.get_mut(&sim_id) else {
-                    if sim.read_pixels(&mut eye_contents) {
+                    continue;
-                        idle = false;
+                };
-                        renderer.render(&eye_contents);
+                let Some(sim) = self.sims.get_mut(sim_id.to_index()) else {
+                    continue;
+                };
+                if sim.read_pixels(&mut eye_contents) {
+                    idle = false;
+                    if renderer.queue_render(&eye_contents).is_err() {
+                        self.renderers.remove(&sim_id);
                     }
                 }
             }
@@ -226,7 +235,7 @@ impl Emulator {
 
 #[derive(Debug)]
 pub enum EmulatorCommand {
-    SetRenderer(SimId, GameRenderer),
+    SetRenderer(SimId, TextureSink),
     LoadGame(SimId, PathBuf),
     StartSecondSim(Option<PathBuf>),
     StopSecondSim,

src/graphics.rs

@@ -0,0 +1,143 @@
+use std::{
+    sync::{
+        atomic::{AtomicU64, Ordering},
+        mpsc, Arc, Mutex, MutexGuard,
+    },
+    thread,
+};
+
+use anyhow::{bail, Result};
+use itertools::Itertools as _;
+use wgpu::{
+    Device, Extent3d, ImageCopyTexture, ImageDataLayout, Origin3d, Queue, Texture,
+    TextureDescriptor, TextureFormat, TextureUsages, TextureView, TextureViewDescriptor,
+};
+
+#[derive(Debug)]
+pub struct TextureSink {
+    buffers: Arc<BufferPool>,
+    sink: mpsc::Sender<u64>,
+}
+
+impl TextureSink {
+    pub fn new(device: &Device, queue: Arc<Queue>) -> (Self, TextureView) {
+        let texture = Self::create_texture(device);
+        let view = texture.create_view(&TextureViewDescriptor::default());
+        let buffers = Arc::new(BufferPool::new());
+        let (sink, source) = mpsc::channel();
+        let bufs = buffers.clone();
+        thread::spawn(move || {
+            let mut local_buf = vec![0; 384 * 224 * 2];
+            while let Ok(id) = source.recv() {
+                {
+                    let Some(bytes) = bufs.read(id) else {
+                        continue;
+                    };
+                    local_buf.copy_from_slice(bytes.as_slice());
+                }
+                Self::write_texture(&queue, &texture, local_buf.as_slice());
+            }
+        });
+        let sink = Self { buffers, sink };
+        (sink, view)
+    }
+
+    pub fn queue_render(&mut self, bytes: &[u8]) -> Result<()> {
+        let id = {
+            let (mut buf, id) = self.buffers.write()?;
+            buf.copy_from_slice(bytes);
+            id
+        };
+        self.sink.send(id)?;
+        Ok(())
+    }
+
+    fn create_texture(device: &Device) -> Texture {
+        let desc = TextureDescriptor {
+            label: Some("eyes"),
+            size: Extent3d {
+                width: 384,
+                height: 224,
+                depth_or_array_layers: 1,
+            },
+            mip_level_count: 1,
+            sample_count: 1,
+            dimension: wgpu::TextureDimension::D2,
+            format: TextureFormat::Rg8Unorm,
+            usage: TextureUsages::COPY_SRC
+                | TextureUsages::COPY_DST
+                | TextureUsages::TEXTURE_BINDING,
+            view_formats: &[TextureFormat::Rg8Unorm],
+        };
+        device.create_texture(&desc)
+    }
+
+    fn write_texture(queue: &Queue, texture: &Texture, bytes: &[u8]) {
+        let texture = ImageCopyTexture {
+            texture,
+            mip_level: 0,
+            origin: Origin3d::ZERO,
+            aspect: wgpu::TextureAspect::All,
+        };
+        let size = Extent3d {
+            width: 384,
+            height: 224,
+            depth_or_array_layers: 1,
+        };
+        let data_layout = ImageDataLayout {
+            offset: 0,
+            bytes_per_row: Some(384 * 2),
+            rows_per_image: Some(224),
+        };
+        queue.write_texture(texture, bytes, data_layout, size);
+    }
+}
+
+#[derive(Debug)]
+struct BufferPool {
+    buffers: [Buffer; 3],
+}
+impl BufferPool {
+    fn new() -> Self {
+        Self {
+            buffers: std::array::from_fn(|i| Buffer::new(i as u64)),
+        }
+    }
+
+    fn read(&self, id: u64) -> Option<MutexGuard<'_, Vec<u8>>> {
+        let buf = self
+            .buffers
+            .iter()
+            .find(|buf| buf.id.load(Ordering::Acquire) == id)?;
+        buf.data.lock().ok()
+    }
+
+    fn write(&self) -> Result<(MutexGuard<'_, Vec<u8>>, u64)> {
+        let (min, max) = self
+            .buffers
+            .iter()
+            .minmax_by_key(|buf| buf.id.load(Ordering::Acquire))
+            .into_option()
+            .unwrap();
+        let Ok(lock) = min.data.lock() else {
+            bail!("lock was poisoned")
+        };
+        let id = max.id.load(Ordering::Acquire) + 1;
+        min.id.store(id, Ordering::Release);
+        Ok((lock, id))
+    }
+}
+
+#[derive(Debug)]
+struct Buffer {
+    data: Mutex<Vec<u8>>,
+    id: AtomicU64,
+}
+impl Buffer {
+    fn new(id: u64) -> Self {
+        Self {
+            data: Mutex::new(vec![0; 384 * 224 * 2]),
+            id: AtomicU64::new(id),
+        }
+    }
+}

src/main.rs

@@ -11,8 +11,8 @@ mod app;
 mod audio;
 mod controller;
 mod emulator;
+mod graphics;
 mod input;
-mod renderer;
 
 #[derive(Parser)]
 struct Args {

src/renderer.rs

@@ -1,53 +0,0 @@
-use std::sync::Arc;
-
-use wgpu::{
-    Extent3d, ImageCopyTexture, ImageDataLayout, Origin3d, Queue, Texture, TextureDescriptor,
-    TextureFormat, TextureUsages,
-};
-
-#[derive(Debug)]
-pub struct GameRenderer {
-    pub queue: Arc<Queue>,
-    pub eyes: Arc<Texture>,
-}
-
-impl GameRenderer {
-    pub fn render(&self, buffer: &[u8]) {
-        let texture = ImageCopyTexture {
-            texture: &self.eyes,
-            mip_level: 0,
-            origin: Origin3d::ZERO,
-            aspect: wgpu::TextureAspect::All,
-        };
-        let size = Extent3d {
-            width: 384,
-            height: 224,
-            depth_or_array_layers: 1,
-        };
-        let data_layout = ImageDataLayout {
-            offset: 0,
-            bytes_per_row: Some(384 * 2),
-            rows_per_image: Some(224),
-        };
-        self.queue.write_texture(texture, buffer, data_layout, size);
-    }
-    pub fn create_texture(device: &wgpu::Device, name: &str) -> Texture {
-        let desc = TextureDescriptor {
-            label: Some(name),
-            size: Extent3d {
-                width: 384,
-                height: 224,
-                depth_or_array_layers: 1,
-            },
-            mip_level_count: 1,
-            sample_count: 1,
-            dimension: wgpu::TextureDimension::D2,
-            format: TextureFormat::Rg8Unorm,
-            usage: TextureUsages::COPY_SRC
-                | TextureUsages::COPY_DST
-                | TextureUsages::TEXTURE_BINDING,
-            view_formats: &[TextureFormat::Rg8Unorm],
-        };
-        device.create_texture(&desc)
-    }
-}