Merge 937ccf60c9 into fb3fd5e8bb

fixes performance regressions since microw8 0.2.1

Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "uw8"
-version = "0.3.0"
+version = "0.3.1"
 edition = "2021"
 
 # See more keys and their definitions at https://doc.rust-lang.org/cargo/reference/manifest.html
@@ -11,7 +11,7 @@ native = ["wasmtime", "uw8-window", "cpal", "rubato" ]
 browser = ["warp", "tokio", "tokio-stream", "webbrowser"]
 
 [dependencies]
-wasmtime = { version = "19.0.1", optional = true }
+wasmtime = { git = "https://github.com/bytecodealliance/wasmtime.git", rev = "0f48f939b9870036562ca02ff21253547a9f1a5c", optional = true }
 anyhow = "1"
 env_logger = "0.11.3"
 log = "0.4"

examples/curlywas/tunnel.cwa

@@ -3,14 +3,16 @@ include "../include/microw8-api.cwa"
 export fn upd() {
     let i: i32;
     loop pixels {
-        let inline t = time() * 63 as f32;
-        let lazy x = (i % 320 - 160) as f32;
-        let lazy y = (i / 320 - 120) as f32;
-        let inline d = 40000 as f32 / sqrt(x * x + y * y);
-        let inline u = atan2(x, y) * (512.0 / 3.141);
-        let inline c = ((i32.trunc_sat_f32_s(d + t * 2 as f32) ^ i32.trunc_sat_f32_s(u + t)) & 255) >> 4;
+        let inline t = 16!56;
+        let inline x = (i % 320 - 160) as f32;
+        let inline y = (i #/ 320 - 120) as f32;
+        let inline d = 0xa000 as f32 / sqrt(x * x + y * y);
+        let inline a = atan2(x, y) * 163_f; // (512 / pi)
+        let inline u = i32.trunc_sat_f32_s(a) + t;
+        let inline v = i32.trunc_sat_f32_s(d) + t * 2;
+        let inline c = ((v ^ u) #/ 16) % 16;
         i?FRAMEBUFFER = c;
 
         branch_if (i := i + 1) < 320*240: pixels;
     }
-}
+}

examples/tinygo/build.sh

@@ -0,0 +1,6 @@
+#!/bin/bash
+
+tinygo build -o cart.wasm -target target.json ./main.go
+uw8 filter-exports cart.wasm cart.wasm && \
+#wasm-opt -Oz --fast-math --strip-producers -o cart.wasm cart.wasm && \
+uw8 pack -l 9 cart.wasm cart.uw8

examples/tinygo/main.go

@@ -0,0 +1,38 @@
+package main
+
+import (
+	"math"
+	"unsafe"
+)
+
+//go:wasm-module env
+//export atan2
+func atan2(x, y float32) float32
+
+//go:wasm-module env
+//export time
+func time() float32
+
+func sqrt(v float32) float32 {
+	return float32(math.Sqrt(float64(v)))
+}
+
+var FRAMEBUFFER = (*[320 * 240]byte)(unsafe.Pointer(uintptr(120)))
+
+//export upd
+func upd() {
+	var i int
+	for i < 320*240 {
+		t := time() * 63.0
+		x := float32(i%320 - 160)
+		y := float32(i/320 - 120)
+		d := float32(40000.0) / sqrt(x*x+y*y)
+		u := atan2(x, y) * 512.0 / 3.141
+		c := uint8((int(d+t*2.0) ^ int(u+t)) >> 4)
+		FRAMEBUFFER[i] = c
+		i++
+	}
+}
+
+func main() {
+}

examples/tinygo/target.json

@@ -0,0 +1,23 @@
+{
+  "llvm-target": "wasm32--wasi",
+  "build-tags": [ "tinygo.wasm" ],
+  "goos": "js",
+  "goarch": "wasm",
+  "linker": "wasm-ld",
+  "libc": "wasi-libc",
+  "cflags": [
+    "--target=wasm32--wasi",
+    "--sysroot={root}/lib/wasi-libc/sysroot",
+    "-Oz"
+  ],
+  "ldflags": [
+    "--no-entry",
+    "--export-all",
+    "--import-memory",
+    "--initial-memory=262144",
+    "--global-base=81920",
+    "-zstack-size=4096",
+    "--strip-all"
+  ],
+  "wasm-abi": "js"
+}

site/content/docs.md

@@ -21,7 +21,7 @@ loaded.
 00000-00040: user memory
 00040-00044: time since module start in ms
 00044-0004c: gamepad state
-0004c-00050: reserved
+0004c-00050: number of frames since module start
 00050-00070: sound data (synced to sound thread)
 00070-00078: reserved
 00078-12c78: frame buffer

src/run_native.rs

@@ -2,7 +2,7 @@ use std::sync::{mpsc, Arc, Mutex};
 use std::time::Duration;
 use std::{thread, time::Instant};
 
-use anyhow::{anyhow, Result};
+use anyhow::{anyhow, bail, Result};
 use cpal::traits::*;
 use rubato::Resampler;
 use uw8_window::{Window, WindowConfig};
@@ -27,6 +27,7 @@ struct UW8Instance {
     end_frame: TypedFunc<(), ()>,
     update: Option<TypedFunc<(), ()>>,
     start_time: Instant,
+    frame_counter: u32,
     watchdog: Arc<Mutex<UW8WatchDog>>,
     sound_tx: Option<mpsc::SyncSender<RegisterUpdate>>,
 }
@@ -159,6 +160,7 @@ impl super::Runtime for MicroW8 {
             end_frame,
             update,
             start_time: Instant::now(),
+            frame_counter: 0,
             watchdog,
             sound_tx,
         });
@@ -191,8 +193,11 @@ impl super::Runtime for MicroW8 {
                 let mem = instance.memory.data_mut(&mut instance.store);
                 mem[64..68].copy_from_slice(&time.to_le_bytes());
                 mem[68..72].copy_from_slice(&input.gamepads);
+                mem[72..76].copy_from_slice(&instance.frame_counter.to_le_bytes());
             }
 
+            instance.frame_counter = instance.frame_counter.wrapping_add(1);
+
             instance.store.set_epoch_deadline(self.timeout as u64);
             if let Some(ref update) = instance.update {
                 if let Err(err) = update.call(&mut instance.store, ()) {
@@ -328,26 +333,37 @@ fn init_sound(
     let mut configs: Vec<_> = device
         .supported_output_configs()?
         .filter(|config| {
-            config.channels() == 2
-                && (config.sample_format() == cpal::SampleFormat::F32
-                    || config.sample_format() == cpal::SampleFormat::I16)
+            config.sample_format() == cpal::SampleFormat::F32
+                || config.sample_format() == cpal::SampleFormat::I16
         })
         .collect();
+
+    if configs.is_empty() {
+        eprintln!(
+            "No suitable audio output config found on device \"{}\", available configs:",
+            device.name()?
+        );
+        for config in device.supported_output_configs()? {
+            eprintln!("  {}ch {}", config.channels(), config.sample_format());
+        }
+        bail!("Failed to configure audio out");
+    }
+
     configs.sort_by_key(|config| {
         let rate = 44100
             .max(config.min_sample_rate().0)
             .min(config.max_sample_rate().0);
-        let prio = if rate >= 44100 {
+        let rate_prio = if rate >= 44100 {
             rate - 44100
         } else {
             (44100 - rate) * 1000
         };
-        prio + (config.sample_format() == cpal::SampleFormat::I16) as u32
+        let format_prio = (config.sample_format() == cpal::SampleFormat::I16) as u32;
+        let channels_prio = (config.channels() != 2) as u32 * 16777216;
+        rate_prio + format_prio + channels_prio
     });
-    let config = configs
-        .into_iter()
-        .next()
-        .ok_or_else(|| anyhow!("Could not find float or 16bit signed output config"))?;
+    let config = configs.into_iter().next().unwrap();
+
     let sample_rate = cpal::SampleRate(44100)
         .max(config.min_sample_rate())
         .min(config.max_sample_rate());
@@ -359,6 +375,7 @@ fn init_sound(
         }
     };
     let sample_format = config.sample_format();
+    let num_channels = config.channels();
     let config = cpal::StreamConfig {
         buffer_size,
         ..config.config()
@@ -392,7 +409,7 @@ fn init_sound(
     let mut pending_updates: Vec<RegisterUpdate> = Vec::with_capacity(30);
     let mut current_time = 0;
 
-    let mut callback = move |mut outer_buffer: &mut [f32], _: &_| {
+    let mut callback = move |mut outer_buffer: &mut [f32]| {
         let mut first_update = true;
         while let Ok(update) = rx.try_recv() {
             if first_update {
@@ -484,36 +501,126 @@ fn init_sound(
             current_time = current_time.wrapping_add((step_size * 500 / sample_rate).max(1) as i32);
         }
     };
-    let stream = if sample_format == cpal::SampleFormat::F32 {
-        device.build_output_stream(
-            &config,
-            callback,
-            move |err| {
-                dbg!(err);
-            },
-            None,
-        )?
-    } else {
-        device.build_output_stream(
-            &config,
-            move |mut buffer: &mut [i16], info| {
-                let mut float_buffer = [0f32; 256];
 
-                while !buffer.is_empty() {
-                    let step_size = buffer.len().min(float_buffer.len());
-                    let step_buffer = &mut float_buffer[..step_size];
-                    callback(step_buffer, info);
-                    for (dest, src) in buffer.iter_mut().take(step_size).zip(step_buffer.iter()) {
-                        *dest = (src.max(-1.0).min(1.0) * 32767.0) as i16;
-                    }
-                    buffer = &mut buffer[step_size..];
-                }
-            },
-            move |err| {
-                dbg!(err);
-            },
-            None,
-        )?
+    fn f32_to_i16<F>(mut buffer: &mut [i16], callback: &mut F)
+    where
+        F: FnMut(&mut [f32]),
+    {
+        let mut float_buffer = [0f32; 256];
+
+        while !buffer.is_empty() {
+            let step_size = buffer.len().min(float_buffer.len());
+            let step_buffer = &mut float_buffer[..step_size];
+            callback(step_buffer);
+            for (dest, src) in buffer.iter_mut().take(step_size).zip(step_buffer.iter()) {
+                *dest = (src.max(-1.0).min(1.0) * 32767.0) as i16;
+            }
+            buffer = &mut buffer[step_size..];
+        }
+    }
+
+    fn stereo_to_mono<F>(mut buffer: &mut [f32], callback: &mut F)
+    where
+        F: FnMut(&mut [f32]),
+    {
+        let mut in_buffer = [0f32; 256];
+
+        while !buffer.is_empty() {
+            let step_size = buffer.len().min(in_buffer.len() / 2);
+            let step_buffer = &mut in_buffer[..step_size * 2];
+            callback(step_buffer);
+            for (index, dest) in buffer.iter_mut().take(step_size).enumerate() {
+                *dest = (step_buffer[index * 2] + step_buffer[index * 2 + 1]) * 0.5;
+            }
+            buffer = &mut buffer[step_size..];
+        }
+    }
+
+    fn stereo_to_surround<F>(mut buffer: &mut [f32], num_channels: usize, callback: &mut F)
+    where
+        F: FnMut(&mut [f32]),
+    {
+        let mut in_buffer = [0f32; 256];
+        buffer.fill(0.);
+
+        while !buffer.is_empty() {
+            let step_size = (buffer.len() / num_channels).min(in_buffer.len() / 2);
+            let step_buffer = &mut in_buffer[..step_size * 2];
+            callback(step_buffer);
+            for index in 0..step_size {
+                buffer[index * num_channels + 0] = step_buffer[index * 2 + 0];
+                buffer[index * num_channels + 1] = step_buffer[index * 2 + 1];
+            }
+            buffer = &mut buffer[step_size * num_channels..];
+        }
+    }
+
+    let stream = if sample_format == cpal::SampleFormat::F32 {
+        if num_channels == 2 {
+            device.build_output_stream(
+                &config,
+                move |buffer: &mut [f32], _| callback(buffer),
+                move |err| {
+                    dbg!(err);
+                },
+                None,
+            )?
+        } else if num_channels == 1 {
+            device.build_output_stream(
+                &config,
+                move |buffer: &mut [f32], _| stereo_to_mono(buffer, &mut callback),
+                move |err| {
+                    dbg!(err);
+                },
+                None,
+            )?
+        } else {
+            device.build_output_stream(
+                &config,
+                move |buffer: &mut [f32], _| {
+                    stereo_to_surround(buffer, num_channels as usize, &mut callback)
+                },
+                move |err| {
+                    dbg!(err);
+                },
+                None,
+            )?
+        }
+    } else {
+        if num_channels == 2 {
+            device.build_output_stream(
+                &config,
+                move |buffer: &mut [i16], _| f32_to_i16(buffer, &mut callback),
+                move |err| {
+                    dbg!(err);
+                },
+                None,
+            )?
+        } else if num_channels == 1 {
+            device.build_output_stream(
+                &config,
+                move |buffer: &mut [i16], _| {
+                    f32_to_i16(buffer, &mut |b| stereo_to_mono(b, &mut callback))
+                },
+                move |err| {
+                    dbg!(err);
+                },
+                None,
+            )?
+        } else {
+            device.build_output_stream(
+                &config,
+                move |buffer: &mut [i16], _| {
+                    f32_to_i16(buffer, &mut |b| {
+                        stereo_to_surround(b, num_channels as usize, &mut callback)
+                    })
+                },
+                move |err| {
+                    dbg!(err);
+                },
+                None,
+            )?
+        }
     };
 
     Ok(Uw8Sound { stream, tx })

web/src/index.html

@@ -10,7 +10,7 @@
     </head>
     <body>
         <div id="uw8">
-            <a href="https://exoticorn.github.io/microw8">MicroW8</a> 0.3.0
+            <a href="https://exoticorn.github.io/microw8">MicroW8</a> 0.3.1
         </div>
         <div id="centered">
             <canvas class="screen" id="screen" width="320" height="240">

web/src/microw8.js

@@ -240,6 +240,7 @@ export default function MicroW8(screen, config = {}) {
             await audioReadyPromise;
     
             let startTime = Date.now();
+            let frameCounter = 0;
     
             const timePerFrame = 1000 / 60;
 
@@ -306,6 +307,7 @@ export default function MicroW8(screen, config = {}) {
                         let time = Date.now() - startTime;
                         u32Mem[16] = time;
                         u32Mem[17] = pad | gamepad;
+                        u32Mem[18] = frameCounter++;
                         if(instance.exports.upd) {
                             instance.exports.upd();
                         }