Merge 1d89ef8613 into 8258bc1854

Cargo.toml

@@ -11,21 +11,21 @@ native = ["wasmtime", "uw8-window", "cpal", "rubato" ]
 browser = ["warp", "tokio", "tokio-stream", "webbrowser"]
 
 [dependencies]
-wasmtime = { version = "5.0.0", optional = true }
+wasmtime = { version = "12.0.0", optional = true }
 anyhow = "1"
 env_logger = "0.10"
 log = "0.4"
 uw8-window = { path = "uw8-window", optional = true }
-notify-debouncer-mini = { version = "0.2.1", default-features = false }
+notify-debouncer-mini = { version = "0.4.1", default-features = false }
 pico-args = "0.5"
 curlywas = { git = "https://github.com/exoticorn/curlywas.git", rev = "0e7ea50" }
 wat = "1"
 uw8-tool = { path = "uw8-tool" }
 same-file = "1"
-warp = { version = "0.3.3", optional = true }
+warp = { version = "0.3.5", optional = true }
-tokio = { version = "1.24.0", features = ["sync", "rt"], optional = true }
+tokio = { version = "1.32.0", features = ["sync", "rt"], optional = true }
-tokio-stream = { version = "0.1.11", features = ["sync"], optional = true }
+tokio-stream = { version = "0.1.14", features = ["sync"], optional = true }
-webbrowser = { version = "0.8.6", optional = true }
+webbrowser = { version = "0.8.11", optional = true }
 ansi_term = "0.12.1"
-cpal = { version = "0.14.2", optional = true }
+cpal = { version = "0.15.2", optional = true }
-rubato = { version = "0.12.0", optional = true }
+rubato = { version = "0.14.0", optional = true }

platform/src/ges.cwa

@@ -62,7 +62,6 @@ export fn sndGes(t: i32) -> f32 {
             let phase = channelState!GesChannelState.Phase;
 
             let inline pulseWidth = channelReg?1;
-            let phaseShift = (pulseWidth - 128) * 255;
             let invPhaseInc = 1 as f32 / phaseInc as f32;
 
             i = 0;
@@ -131,7 +130,7 @@ export fn sndGes(t: i32) -> f32 {
                 let phaseInc = (freq * (65536.0 / 44100.0)) as i32;
 
                 let phase = channelState!GesChannelState.Phase;
-                if modSrc > ch {
+                if modSrc < ch {
                     phase = phase - (phaseInc << 6);
                 }
 

site/content/docs.md

@@ -5,7 +5,7 @@ description = "Docs"
 
 # Overview
 
-MicroW8 loads WebAssembly modules with a maximum size of 256kb. You module needs to export
+MicroW8 loads WebAssembly modules with a maximum size of 256kb. Your module needs to export
 a function `fn upd()` which will be called once per frame.
 After calling `upd` MicroW8 will display the 320x240 8bpp framebuffer located
 at offset 120 in memory with the 32bpp palette located at 0x13000.

src/filewatcher.rs

@@ -16,16 +16,14 @@ pub struct FileWatcher {
 impl FileWatcher {
     pub fn new() -> Result<FileWatcher> {
         let (tx, rx) = mpsc::channel();
-        let debouncer = new_debouncer(Duration::from_millis(100), None, move |res| match res {
+        let debouncer = new_debouncer(Duration::from_millis(100), move |res| match res {
             Ok(events) => {
                 for event in events {
                     let _ = tx.send(event);
                 }
             }
-            Err(errs) => {
+            Err(err) => {
-                for err in errs {
+                eprintln!("Error watching for file changes: {err}");
-                    eprintln!("Error watching for file changes: {err}");
-                }
             }
         })?;
         Ok(FileWatcher {

src/run_native.rs

@@ -7,7 +7,7 @@ use cpal::traits::*;
 use rubato::Resampler;
 use uw8_window::{Window, WindowConfig};
 use wasmtime::{
-    Engine, GlobalType, Memory, MemoryType, Module, Mutability, Store, TypedFunc, ValType,
+    Engine, Func, GlobalType, Memory, MemoryType, Module, Mutability, Store, TypedFunc, ValType,
 };
 
 pub struct MicroW8 {
@@ -90,7 +90,7 @@ impl super::Runtime for MicroW8 {
         let memory = wasmtime::Memory::new(&mut store, MemoryType::new(4, Some(4)))?;
 
         let mut linker = wasmtime::Linker::new(&self.engine);
-        linker.define("env", "memory", memory)?;
+        linker.define(&store, "env", "memory", memory)?;
 
         let loader_instance = linker.instantiate(&mut store, &self.loader_module)?;
         let load_uw8 = loader_instance.get_typed_func::<i32, i32>(&mut store, "load_uw8")?;
@@ -255,15 +255,12 @@ fn add_native_functions(
         }
     })?;
     for i in 0..16 {
-        linker.define(
+        let global = wasmtime::Global::new(
-            "env",
+            &mut *store,
-            &format!("g_reserved{}", i),
+            GlobalType::new(ValType::I32, Mutability::Const),
-            wasmtime::Global::new(
+            0.into(),
-                &mut *store,
-                GlobalType::new(ValType::I32, Mutability::Const),
-                0.into(),
-            )?,
         )?;
+        linker.define(&store, "env", &format!("g_reserved{}", i), global)?;
     }
 
     Ok(())
@@ -276,14 +273,18 @@ fn instantiate_platform(
 ) -> Result<wasmtime::Instance> {
     let platform_instance = linker.instantiate(&mut *store, &platform_module)?;
 
-    for export in platform_instance.exports(&mut *store) {
+    let exports: Vec<(String, Func)> = platform_instance
-        linker.define(
+        .exports(&mut *store)
-            "env",
+        .map(|e| {
-            export.name(),
+            (
-            export
+                e.name().to_owned(),
-                .into_func()
+                e.into_func()
-                .expect("platform surely only exports functions"),
+                    .expect("platform surely only exports functions"),
-        )?;
+            )
+        })
+        .collect();
+    for (name, func) in exports {
+        linker.define(&store, "env", &name, func)?;
     }
 
     Ok(platform_instance)
@@ -310,7 +311,7 @@ fn init_sound(
     let memory = wasmtime::Memory::new(&mut store, MemoryType::new(4, Some(4)))?;
 
     let mut linker = wasmtime::Linker::new(engine);
-    linker.define("env", "memory", memory)?;
+    linker.define(&store, "env", "memory", memory)?;
     add_native_functions(&mut linker, &mut store)?;
 
     let platform_instance = instantiate_platform(&mut linker, &mut store, platform_module)?;
@@ -327,23 +328,26 @@ fn init_sound(
     let mut configs: Vec<_> = device
         .supported_output_configs()?
         .filter(|config| {
-            config.channels() == 2 && config.sample_format() == cpal::SampleFormat::F32
+            config.channels() == 2
+                && (config.sample_format() == cpal::SampleFormat::F32
+                    || config.sample_format() == cpal::SampleFormat::I16)
         })
         .collect();
     configs.sort_by_key(|config| {
         let rate = 44100
             .max(config.min_sample_rate().0)
             .min(config.max_sample_rate().0);
-        if rate >= 44100 {
+        let prio = if rate >= 44100 {
             rate - 44100
         } else {
             (44100 - rate) * 1000
-        }
+        };
+        prio + (config.sample_format() == cpal::SampleFormat::I16) as u32
     });
     let config = configs
         .into_iter()
         .next()
-        .ok_or_else(|| anyhow!("Could not find float output config"))?;
+        .ok_or_else(|| anyhow!("Could not find float or 16bit signed output config"))?;
     let sample_rate = cpal::SampleRate(44100)
         .max(config.min_sample_rate())
         .min(config.max_sample_rate());
@@ -354,6 +358,7 @@ fn init_sound(
             cpal::BufferSize::Fixed(256.max(min).min(max))
         }
     };
+    let sample_format = config.sample_format();
     let config = cpal::StreamConfig {
         buffer_size,
         ..config.config()
@@ -373,8 +378,8 @@ fn init_sound(
         None
     } else {
         let rs = rubato::FftFixedIn::new(44100, sample_rate, 128, 1, 2)?;
-        let input_buffers = rs.input_buffer_allocate();
+        let input_buffers = rs.input_buffer_allocate(true);
-        let output_buffers = rs.output_buffer_allocate();
+        let output_buffers = rs.output_buffer_allocate(true);
         Some(Resampler {
             resampler: rs,
             input_buffers,
@@ -386,96 +391,130 @@ fn init_sound(
     let mut sample_index = 0;
     let mut pending_updates: Vec<RegisterUpdate> = Vec::with_capacity(30);
     let mut current_time = 0;
-    let stream = device.build_output_stream(
+
-        &config,
+    let mut callback = move |mut outer_buffer: &mut [f32], _: &_| {
-        move |mut outer_buffer: &mut [f32], _| {
+        let mut first_update = true;
-            let mut first_update = true;
+        while let Ok(update) = rx.try_recv() {
-            while let Ok(update) = rx.try_recv() {
+            if first_update {
-                if first_update {
+                current_time += update.time.wrapping_sub(current_time) / 8;
-                    current_time += update.time.wrapping_sub(current_time) / 8;
+                first_update = false;
-                    first_update = false;
+            }
-                }
+            pending_updates.push(update);
-                pending_updates.push(update);
+        }
+
+        while !outer_buffer.is_empty() {
+            store.set_epoch_deadline(30);
+            while pending_updates
+                .first()
+                .into_iter()
+                .any(|u| u.time.wrapping_sub(current_time) <= 0)
+            {
+                let update = pending_updates.remove(0);
+                memory.write(&mut store, 80, &update.data).unwrap();
             }
 
-            while !outer_buffer.is_empty() {
+            let duration = if let Some(update) = pending_updates.first() {
-                store.set_epoch_deadline(30);
+                ((update.time.wrapping_sub(current_time) as usize) * sample_rate + 999) / 1000
-                while pending_updates
+            } else {
-                    .first()
+                outer_buffer.len()
-                    .into_iter()
+            };
-                    .any(|u| u.time.wrapping_sub(current_time) <= 0)
+            let step_size = (duration.max(64) * 2).min(outer_buffer.len());
-                {
-                    let update = pending_updates.remove(0);
-                    memory.write(&mut store, 80, &update.data).unwrap();
-                }
 
-                let duration = if let Some(update) = pending_updates.first() {
+            let mut buffer = &mut outer_buffer[..step_size];
-                    ((update.time.wrapping_sub(current_time) as usize) * sample_rate + 999) / 1000
+
+            {
+                let mem = memory.data_mut(&mut store);
+                mem[64..68].copy_from_slice(&current_time.to_le_bytes());
+            }
+
+            fn clamp_sample(s: f32) -> f32 {
+                if s.is_nan() {
+                    0.0
                 } else {
-                    outer_buffer.len()
+                    s.max(-1.0).min(1.0)
-                };
-                let step_size = (duration.max(64) * 2).min(outer_buffer.len());
-
-                let mut buffer = &mut outer_buffer[..step_size];
-
-                {
-                    let mem = memory.data_mut(&mut store);
-                    mem[64..68].copy_from_slice(&current_time.to_le_bytes());
                 }
+            }
 
-                if let Some(ref mut resampler) = resampler {
+            if let Some(ref mut resampler) = resampler {
-                    while !buffer.is_empty() {
+                while !buffer.is_empty() {
-                        let copy_size = resampler.output_buffers[0]
+                    let copy_size = resampler.output_buffers[0]
-                            .len()
+                        .len()
-                            .saturating_sub(resampler.output_index)
+                        .saturating_sub(resampler.output_index)
-                            .min(buffer.len() / 2);
+                        .min(buffer.len() / 2);
-                        if copy_size == 0 {
+                    if copy_size == 0 {
-                            resampler.input_buffers[0].clear();
+                        resampler.input_buffers[0].clear();
-                            resampler.input_buffers[1].clear();
+                        resampler.input_buffers[1].clear();
-                            for _ in 0..resampler.resampler.input_frames_next() {
+                        for _ in 0..resampler.resampler.input_frames_next() {
-                                resampler.input_buffers[0]
+                            resampler.input_buffers[0].push(clamp_sample(
-                                    .push(snd.call(&mut store, (sample_index,)).unwrap_or(0.0));
+                                snd.call(&mut store, (sample_index,)).unwrap_or(0.0),
-                                resampler.input_buffers[1]
+                            ));
-                                    .push(snd.call(&mut store, (sample_index + 1,)).unwrap_or(0.0));
+                            resampler.input_buffers[1].push(clamp_sample(
-                                sample_index = sample_index.wrapping_add(2);
+                                snd.call(&mut store, (sample_index + 1,)).unwrap_or(0.0),
-                            }
+                            ));
-
+                            sample_index = sample_index.wrapping_add(2);
-                            resampler
-                                .resampler
-                                .process_into_buffer(
-                                    &resampler.input_buffers,
-                                    &mut resampler.output_buffers,
-                                    None,
-                                )
-                                .unwrap();
-                            resampler.output_index = 0;
-                        } else {
-                            for i in 0..copy_size {
-                                buffer[i * 2] =
-                                    resampler.output_buffers[0][resampler.output_index + i];
-                                buffer[i * 2 + 1] =
-                                    resampler.output_buffers[1][resampler.output_index + i];
-                            }
-                            resampler.output_index += copy_size;
-                            buffer = &mut buffer[copy_size * 2..];
                         }
-                    }
+
-                } else {
+                        resampler
-                    for v in buffer {
+                            .resampler
-                        *v = snd.call(&mut store, (sample_index,)).unwrap_or(0.0);
+                            .process_into_buffer(
-                        sample_index = sample_index.wrapping_add(1);
+                                &resampler.input_buffers,
+                                &mut resampler.output_buffers,
+                                None,
+                            )
+                            .unwrap();
+                        resampler.output_index = 0;
+                    } else {
+                        for i in 0..copy_size {
+                            buffer[i * 2] = resampler.output_buffers[0][resampler.output_index + i];
+                            buffer[i * 2 + 1] =
+                                resampler.output_buffers[1][resampler.output_index + i];
+                        }
+                        resampler.output_index += copy_size;
+                        buffer = &mut buffer[copy_size * 2..];
                     }
                 }
-
+            } else {
-                outer_buffer = &mut outer_buffer[step_size..];
+                for v in buffer {
-                current_time =
+                    *v = clamp_sample(snd.call(&mut store, (sample_index,)).unwrap_or(0.0));
-                    current_time.wrapping_add((step_size * 500 / sample_rate).max(1) as i32);
+                    sample_index = sample_index.wrapping_add(1);
+                }
             }
-        },
+
-        move |err| {
+            outer_buffer = &mut outer_buffer[step_size..];
-            dbg!(err);
+            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,
+        )?
+    };
 
     Ok(Uw8Sound { stream, tx })
 }

todo.txt

@@ -0,0 +1 @@
+* add support for 16bit sound (not just float)

uw8-tool/src/filter_exports.rs

@@ -1,13 +1,17 @@
-use std::path::Path;
 use anyhow::Result;
+use std::path::Path;
 
 pub fn filter_exports(in_path: &Path, out_path: &Path) -> Result<()> {
     let mut module = walrus::Module::from_file(in_path)?;
 
-    let exports_to_delete: Vec<_> = module.exports.iter().filter_map(|export| match export.name.as_str() {
+    let exports_to_delete: Vec<_> = module
-        "upd" => None,
+        .exports
-        _ => Some(export.id())
+        .iter()
-    }).collect();
+        .filter_map(|export| match export.name.as_str() {
+            "start" | "upd" | "snd" => None,
+            _ => Some(export.id()),
+        })
+        .collect();
 
     for id in exports_to_delete {
         module.exports.delete(id);
@@ -18,4 +22,4 @@ pub fn filter_exports(in_path: &Path, out_path: &Path) -> Result<()> {
     module.emit_wasm_file(out_path)?;
 
     Ok(())
-}
+}

uw8-tool/src/pack.rs

@@ -143,6 +143,7 @@ fn to_val_type(type_: &wasmparser::ValType) -> Result<ValType> {
         I64 => ValType::I64,
         F32 => ValType::F32,
         F64 => ValType::F64,
+        V128 => ValType::V128,
         _ => bail!("Type {:?} isn't a value type", type_),
     })
 }
@@ -967,6 +968,244 @@ fn remap_function(
             De::I64TruncSatF64U => En::I64TruncSatF64U,
             De::MemoryCopy { src_mem, dst_mem } => En::MemoryCopy { src_mem, dst_mem },
             De::MemoryFill { mem } => En::MemoryFill(mem),
+
+            De::V128Const { value } => En::V128Const(value.i128()),
+            De::V128Load { memarg } => En::V128Load(mem(memarg)),
+            De::V128Store { memarg } => En::V128Store(mem(memarg)),
+            De::V128Load8x8S { memarg } => En::V128Load8x8S(mem(memarg)),
+            De::V128Load8x8U { memarg } => En::V128Load8x8U(mem(memarg)),
+            De::V128Load16x4S { memarg } => En::V128Load16x4S(mem(memarg)),
+            De::V128Load16x4U { memarg } => En::V128Load16x4U(mem(memarg)),
+            De::V128Load32x2S { memarg } => En::V128Load32x2S(mem(memarg)),
+            De::V128Load32x2U { memarg } => En::V128Load32x2U(mem(memarg)),
+            De::V128Load8Splat { memarg } => En::V128Load8Splat(mem(memarg)),
+            De::V128Load16Splat { memarg } => En::V128Load16Splat(mem(memarg)),
+            De::V128Load32Splat { memarg } => En::V128Load32Splat(mem(memarg)),
+            De::V128Load64Splat { memarg } => En::V128Load64Splat(mem(memarg)),
+            De::V128Load32Zero { memarg } => En::V128Load32Zero(mem(memarg)),
+            De::V128Load64Zero { memarg } => En::V128Load64Zero(mem(memarg)),
+
+            De::V128Load8Lane { memarg, lane } => En::V128Load8Lane { memarg: mem(memarg), lane },
+            De::V128Load16Lane { memarg, lane } => En::V128Load16Lane { memarg: mem(memarg), lane },
+            De::V128Load32Lane { memarg, lane } => En::V128Load32Lane { memarg: mem(memarg), lane },
+            De::V128Load64Lane { memarg, lane } => En::V128Load64Lane { memarg: mem(memarg), lane },
+            De::V128Store8Lane { memarg, lane } => En::V128Store8Lane { memarg: mem(memarg), lane },
+            De::V128Store16Lane { memarg, lane } => En::V128Store16Lane { memarg: mem(memarg), lane },
+            De::V128Store32Lane { memarg, lane } => En::V128Store32Lane { memarg: mem(memarg), lane },
+            De::V128Store64Lane { memarg, lane } => En::V128Store64Lane { memarg: mem(memarg), lane },
+
+            De::I8x16ExtractLaneS { lane } => En::I8x16ExtractLaneS(lane),
+            De::I8x16ExtractLaneU { lane } => En::I8x16ExtractLaneU(lane),
+            De::I8x16ReplaceLane { lane } => En::I8x16ReplaceLane(lane),
+            De::I16x8ExtractLaneS { lane } => En::I16x8ExtractLaneS(lane),
+            De::I16x8ExtractLaneU { lane } => En::I16x8ExtractLaneU(lane),
+            De::I16x8ReplaceLane { lane } => En::I16x8ReplaceLane(lane),
+            De::I32x4ExtractLane { lane } => En::I32x4ExtractLane(lane),
+            De::I32x4ReplaceLane { lane } => En::I32x4ReplaceLane(lane),
+            De::I64x2ExtractLane { lane } => En::I64x2ExtractLane(lane),
+            De::I64x2ReplaceLane { lane } => En::I64x2ReplaceLane(lane),
+            De::F32x4ExtractLane { lane } => En::F32x4ExtractLane(lane),
+            De::F32x4ReplaceLane { lane } => En::F32x4ReplaceLane(lane),
+            De::F64x2ExtractLane { lane } => En::F64x2ExtractLane(lane),
+            De::F64x2ReplaceLane { lane } => En::F64x2ReplaceLane(lane),
+
+            De::I8x16Splat => En::I8x16Splat,
+            De::I16x8Splat => En::I16x8Splat,
+            De::I32x4Splat => En::I32x4Splat,
+            De::I64x2Splat => En::I64x2Splat,
+            De::F32x4Splat => En::F32x4Splat,
+            De::F64x2Splat => En::F64x2Splat,
+            De::I8x16Swizzle => En::I8x16Swizzle,
+            De::I8x16Add => En::I8x16Add,
+            De::I16x8Add => En::I16x8Add,
+            De::I32x4Add => En::I32x4Add,
+            De::I64x2Add => En::I64x2Add,
+            De::F32x4Add => En::F32x4Add,
+            De::F64x2Add => En::F64x2Add,
+            De::I8x16Sub => En::I8x16Sub,
+            De::I16x8Sub => En::I16x8Sub,
+            De::I32x4Sub => En::I32x4Sub,
+            De::I64x2Sub => En::I64x2Sub,
+            De::F32x4Sub => En::F32x4Sub,
+            De::F64x2Sub => En::F64x2Sub,
+            De::I16x8Mul => En::I16x8Mul,
+            De::I32x4Mul => En::I32x4Mul,
+            De::I64x2Mul => En::I64x2Mul,
+            De::F32x4Mul => En::F32x4Mul,
+            De::F64x2Mul => En::F64x2Mul,
+            De::I32x4DotI16x8S => En::I32x4DotI16x8S,
+            De::I8x16Neg => En::I8x16Neg,
+            De::I16x8Neg => En::I16x8Neg,
+            De::I32x4Neg => En::I32x4Neg,
+            De::I64x2Neg => En::I64x2Neg,
+            De::F32x4Neg => En::F32x4Neg,
+            De::F64x2Neg => En::F64x2Neg,
+            De::I16x8ExtMulLowI8x16S => En::I16x8ExtMulLowI8x16S,
+            De::I16x8ExtMulHighI8x16S => En::I16x8ExtMulHighI8x16S,
+            De::I16x8ExtMulLowI8x16U => En::I16x8ExtMulLowI8x16U,
+            De::I16x8ExtMulHighI8x16U => En::I16x8ExtMulHighI8x16U,
+            De::I32x4ExtMulLowI16x8S => En::I32x4ExtMulLowI16x8S,
+            De::I32x4ExtMulHighI16x8S => En::I32x4ExtMulHighI16x8S,
+            De::I32x4ExtMulLowI16x8U => En::I32x4ExtMulLowI16x8U,
+            De::I32x4ExtMulHighI16x8U => En::I32x4ExtMulHighI16x8U,
+            De::I64x2ExtMulLowI32x4S => En::I64x2ExtMulLowI32x4S,
+            De::I64x2ExtMulHighI32x4S => En::I64x2ExtMulHighI32x4S,
+            De::I64x2ExtMulLowI32x4U => En::I64x2ExtMulLowI32x4U,
+            De::I64x2ExtMulHighI32x4U => En::I64x2ExtMulHighI32x4U,
+            De::I16x8ExtAddPairwiseI8x16S => En::I16x8ExtAddPairwiseI8x16S,
+            De::I16x8ExtAddPairwiseI8x16U => En::I16x8ExtAddPairwiseI8x16U,
+            De::I32x4ExtAddPairwiseI16x8S => En::I32x4ExtAddPairwiseI16x8S,
+            De::I32x4ExtAddPairwiseI16x8U => En::I32x4ExtAddPairwiseI16x8U,
+            De::I8x16AddSatS => En::I8x16AddSatS,
+            De::I8x16AddSatU => En::I8x16AddSatU,
+            De::I16x8AddSatS => En::I16x8AddSatS,
+            De::I16x8AddSatU => En::I16x8AddSatU,
+            De::I8x16SubSatS => En::I8x16SubSatS,
+            De::I8x16SubSatU => En::I8x16SubSatU,
+            De::I16x8SubSatS => En::I16x8SubSatS,
+            De::I16x8SubSatU => En::I16x8SubSatU,
+            De::I16x8Q15MulrSatS => En::I16x8Q15MulrSatS,
+            De::I8x16MinS => En::I8x16MinS,
+            De::I8x16MinU => En::I8x16MinU,
+            De::I16x8MinS => En::I16x8MinS,
+            De::I16x8MinU => En::I16x8MinU,
+            De::I32x4MinS => En::I32x4MinS,
+            De::I32x4MinU => En::I32x4MinU,
+            De::F32x4Min => En::F32x4Min,
+            De::F64x2Min => En::F64x2Min,
+            De::F32x4PMin => En::F32x4PMin,
+            De::F64x2PMin => En::F64x2PMin,
+            De::I8x16MaxS => En::I8x16MaxS,
+            De::I8x16MaxU => En::I8x16MaxU,
+            De::I16x8MaxS => En::I16x8MaxS,
+            De::I16x8MaxU => En::I16x8MaxU,
+            De::I32x4MaxS => En::I32x4MaxS,
+            De::I32x4MaxU => En::I32x4MaxU,
+            De::F32x4Max => En::F32x4Max,
+            De::F64x2Max => En::F64x2Max,
+            De::F32x4PMax => En::F32x4PMax,
+            De::F64x2PMax => En::F64x2PMax,
+            De::I8x16AvgrU => En::I8x16AvgrU,
+            De::I16x8AvgrU => En::I16x8AvgrU,
+            De::I8x16Abs => En::I8x16Abs,
+            De::I16x8Abs => En::I16x8Abs,
+            De::I32x4Abs => En::I32x4Abs,
+            De::I64x2Abs => En::I64x2Abs,
+            De::F32x4Abs => En::F32x4Abs,
+            De::F64x2Abs => En::F64x2Abs,
+            De::I8x16Shl => En::I8x16Shl,
+            De::I16x8Shl => En::I16x8Shl,
+            De::I32x4Shl => En::I32x4Shl,
+            De::I64x2Shl => En::I64x2Shl,
+            De::I8x16ShrS => En::I8x16ShrS,
+            De::I8x16ShrU => En::I8x16ShrU,
+            De::I16x8ShrS => En::I16x8ShrS,
+            De::I16x8ShrU => En::I16x8ShrU,
+            De::I32x4ShrS => En::I32x4ShrS,
+            De::I32x4ShrU => En::I32x4ShrU,
+            De::I64x2ShrS => En::I64x2ShrS,
+            De::I64x2ShrU => En::I64x2ShrU,
+            De::V128And => En::V128And,
+            De::V128Or => En::V128Or,
+            De::V128Xor => En::V128Xor,
+            De::V128Not => En::V128Not,
+            De::V128AndNot => En::V128AndNot,
+            De::V128Bitselect => En::V128Bitselect,
+            De::I8x16Popcnt => En::I8x16Popcnt,
+            De::V128AnyTrue => En::V128AnyTrue,
+            De::I8x16AllTrue => En::I8x16AllTrue,
+            De::I16x8AllTrue => En::I16x8AllTrue,
+            De::I32x4AllTrue => En::I32x4AllTrue,
+            De::I64x2AllTrue => En::I64x2AllTrue,
+            De::I8x16Bitmask => En::I8x16Bitmask,
+            De::I16x8Bitmask => En::I16x8Bitmask,
+            De::I32x4Bitmask => En::I32x4Bitmask,
+            De::I64x2Bitmask => En::I64x2Bitmask,
+            De::I8x16Eq => En::I8x16Eq,
+            De::I16x8Eq => En::I16x8Eq,
+            De::I32x4Eq => En::I32x4Eq,
+            De::I64x2Eq => En::I64x2Eq,
+            De::F32x4Eq => En::F32x4Eq,
+            De::F64x2Eq => En::F64x2Eq,
+            De::I8x16Ne => En::I8x16Ne,
+            De::I16x8Ne => En::I16x8Ne,
+            De::I32x4Ne => En::I32x4Ne,
+            De::I64x2Ne => En::I64x2Ne,
+            De::F32x4Ne => En::F32x4Ne,
+            De::F64x2Ne => En::F64x2Ne,
+            De::I8x16LtS => En::I8x16LtS,
+            De::I8x16LtU => En::I8x16LtU,
+            De::I16x8LtS => En::I16x8LtS,
+            De::I16x8LtU => En::I16x8LtU,
+            De::I32x4LtS => En::I32x4LtS,
+            De::I32x4LtU => En::I32x4LtU,
+            De::F32x4Lt => En::F32x4Lt,
+            De::F64x2Lt => En::F64x2Lt,
+            De::I8x16LeS => En::I8x16LeS,
+            De::I8x16LeU => En::I8x16LeU,
+            De::I16x8LeS => En::I16x8LeS,
+            De::I16x8LeU => En::I16x8LeU,
+            De::I32x4LeS => En::I32x4LeS,
+            De::I32x4LeU => En::I32x4LeU,
+            De::F32x4Le => En::F32x4Le,
+            De::F64x2Le => En::F64x2Le,
+            De::I8x16GtS => En::I8x16GtS,
+            De::I8x16GtU => En::I8x16GtU,
+            De::I16x8GtS => En::I16x8GtS,
+            De::I16x8GtU => En::I16x8GtU,
+            De::I32x4GtS => En::I32x4GtS,
+            De::I32x4GtU => En::I32x4GtU,
+            De::F32x4Gt => En::F32x4Gt,
+            De::F64x2Gt => En::F64x2Gt,
+            De::I8x16GeS => En::I8x16GeS,
+            De::I8x16GeU => En::I8x16GeU,
+            De::I16x8GeS => En::I16x8GeS,
+            De::I16x8GeU => En::I16x8GeU,
+            De::I32x4GeS => En::I32x4GeS,
+            De::I32x4GeU => En::I32x4GeU,
+            De::F32x4Ge => En::F32x4Ge,
+            De::F64x2Ge => En::F64x2Ge,
+            De::F32x4Div => En::F32x4Div,
+            De::F64x2Div => En::F64x2Div,
+            De::F32x4Sqrt => En::F32x4Sqrt,
+            De::F64x2Sqrt => En::F64x2Sqrt,
+            De::F32x4Ceil => En::F32x4Ceil,
+            De::F64x2Ceil => En::F64x2Ceil,
+            De::F32x4Floor => En::F32x4Floor,
+            De::F64x2Floor => En::F64x2Floor,
+            De::F32x4Trunc => En::F32x4Trunc,
+            De::F64x2Trunc => En::F64x2Trunc,
+            De::F32x4Nearest => En::F32x4Nearest,
+            De::F64x2Nearest => En::F64x2Nearest,
+            De::F32x4ConvertI32x4S => En::F32x4ConvertI32x4S,
+            De::F32x4ConvertI32x4U => En::F32x4ConvertI32x4U,
+            De::F64x2ConvertLowI32x4S => En::F64x2ConvertLowI32x4S,
+            De::F64x2ConvertLowI32x4U => En::F64x2ConvertLowI32x4U,
+            De::I32x4TruncSatF32x4S => En::I32x4TruncSatF32x4S,
+            De::I32x4TruncSatF32x4U => En::I32x4TruncSatF32x4U,
+            De::I32x4TruncSatF64x2SZero => En::I32x4TruncSatF64x2SZero,
+            De::I32x4TruncSatF64x2UZero => En::I32x4TruncSatF64x2UZero,
+            De::F32x4DemoteF64x2Zero => En::F32x4DemoteF64x2Zero,
+            De::F64x2PromoteLowF32x4 => En::F64x2PromoteLowF32x4,
+            De::I8x16NarrowI16x8S => En::I8x16NarrowI16x8S,
+            De::I8x16NarrowI16x8U => En::I8x16NarrowI16x8U,
+            De::I16x8NarrowI32x4S => En::I16x8NarrowI32x4S,
+            De::I16x8NarrowI32x4U => En::I16x8NarrowI32x4U,
+            De::I16x8ExtendLowI8x16S => En::I16x8ExtendLowI8x16S,
+            De::I16x8ExtendHighI8x16S => En::I16x8ExtendHighI8x16S,
+            De::I16x8ExtendLowI8x16U => En::I16x8ExtendLowI8x16U,
+            De::I16x8ExtendHighI8x16U => En::I16x8ExtendHighI8x16U,
+            De::I32x4ExtendLowI16x8S => En::I32x4ExtendLowI16x8S,
+            De::I32x4ExtendHighI16x8S => En::I32x4ExtendHighI16x8S,
+            De::I32x4ExtendLowI16x8U => En::I32x4ExtendLowI16x8U,
+            De::I32x4ExtendHighI16x8U => En::I32x4ExtendHighI16x8U,
+            De::I64x2ExtendLowI32x4S => En::I64x2ExtendLowI32x4S,
+            De::I64x2ExtendHighI32x4S => En::I64x2ExtendHighI32x4S,
+            De::I64x2ExtendLowI32x4U => En::I64x2ExtendLowI32x4U,
+            De::I64x2ExtendHighI32x4U => En::I64x2ExtendHighI32x4U,
+
+            De::I8x16Shuffle { lanes } => En::I8x16Shuffle(lanes),
+
             other => bail!("Unsupported instruction {:?}", other),
         });
     }

uw8-window/Cargo.toml

@@ -6,13 +6,13 @@ edition = "2021"
 # See more keys and their definitions at https://doc.rust-lang.org/cargo/reference/manifest.html
 
 [dependencies]
-winit = "0.27.5"
+winit = "0.28.6"
 env_logger = "0.10"
 log = "0.4"
 pico-args = "0.5"
-wgpu = "0.15"
+wgpu = "0.17"
-pollster = "0.2.5"
+pollster = "0.3.0"
 bytemuck = { version = "1.13", features = [ "derive" ] }
 anyhow = "1"
-minifb = { version = "0.23.0", default-features = false, features = ["x11"] }
+minifb = { version = "0.25.0", default-features = false, features = ["x11"] }
 winapi = { version = "0.3.9", features = [ "timeapi" ] }

uw8-window/src/gpu/fast_crt.wgsl

@@ -30,8 +30,8 @@ fn row_factor(offset: f32) -> f32 {
 }
 
 fn col_factor(offset: f32) -> f32 {
-    let offset = max(0.0, abs(offset) - 0.4);
+    let o = max(0.0, abs(offset) - 0.4);
-    return 1.0 / (1.0 + offset * offset * 16.0);
+    return 1.0 / (1.0 + o * o * 16.0);
 }
 
 fn sample_screen(tex_coords: vec2<f32>) -> vec4<f32> {

uw8-window/src/gpu/mod.rs

@@ -1,6 +1,6 @@
 use crate::{Input, WindowConfig, WindowImpl};
 use anyhow::{anyhow, Result};
-use std::{num::NonZeroU32, time::Instant};
+use std::time::Instant;
 
 use winit::{
     dpi::PhysicalSize,
@@ -41,7 +41,10 @@ impl Window {
         async fn create(window_config: WindowConfig) -> Result<Window> {
             let event_loop = EventLoop::new();
             let window = WindowBuilder::new()
-                .with_inner_size(PhysicalSize::new(640u32, 480))
+                .with_inner_size(PhysicalSize::new(
+                    (320. * window_config.scale).round() as u32,
+                    (240. * window_config.scale).round() as u32,
+                ))
                 .with_min_inner_size(PhysicalSize::new(320u32, 240))
                 .with_title("MicroW8")
                 .with_fullscreen(if window_config.fullscreen {
@@ -72,7 +75,13 @@ impl Window {
 
             let surface_config = wgpu::SurfaceConfiguration {
                 present_mode: wgpu::PresentMode::AutoNoVsync,
-                ..surface.get_default_config(&adapter, window.inner_size().width, window.inner_size().height).expect("Surface incompatible with adapter")
+                ..surface
+                    .get_default_config(
+                        &adapter,
+                        window.inner_size().width,
+                        window.inner_size().height,
+                    )
+                    .expect("Surface incompatible with adapter")
             };
 
             let filter: Box<dyn Filter> = create_filter(
@@ -354,7 +363,7 @@ impl PaletteScreenMode {
             format: wgpu::TextureFormat::R8Uint,
             usage: wgpu::TextureUsages::TEXTURE_BINDING | wgpu::TextureUsages::COPY_DST,
             label: None,
-            view_formats: &[]
+            view_formats: &[],
         });
 
         let palette_texture = device.create_texture(&wgpu::TextureDescriptor {
@@ -369,7 +378,7 @@ impl PaletteScreenMode {
             format: wgpu::TextureFormat::Rgba8UnormSrgb,
             usage: wgpu::TextureUsages::TEXTURE_BINDING | wgpu::TextureUsages::COPY_DST,
             label: None,
-            view_formats: &[]
+            view_formats: &[],
         });
 
         let screen_texture = device.create_texture(&wgpu::TextureDescriptor {
@@ -384,7 +393,7 @@ impl PaletteScreenMode {
             format: wgpu::TextureFormat::Rgba8UnormSrgb,
             usage: wgpu::TextureUsages::TEXTURE_BINDING | wgpu::TextureUsages::RENDER_ATTACHMENT,
             label: None,
-            view_formats: &[]
+            view_formats: &[],
         });
 
         let framebuffer_texture_view =
@@ -491,7 +500,7 @@ impl PaletteScreenMode {
             &bytemuck::cast_slice(pixels),
             wgpu::ImageDataLayout {
                 offset: 0,
-                bytes_per_row: NonZeroU32::new(320),
+                bytes_per_row: Some(320),
                 rows_per_image: None,
             },
             wgpu::Extent3d {
@@ -513,7 +522,7 @@ impl PaletteScreenMode {
             &bytemuck::cast_slice(palette),
             wgpu::ImageDataLayout {
                 offset: 0,
-                bytes_per_row: NonZeroU32::new(256 * 4),
+                bytes_per_row: Some(256 * 4),
                 rows_per_image: None,
             },
             wgpu::Extent3d {

uw8-window/src/lib.rs

@@ -15,7 +15,7 @@ struct FpsCounter {
 }
 
 impl Window {
-    pub fn new(config: WindowConfig) -> Result<Window> {
+    pub fn new(mut config: WindowConfig) -> Result<Window> {
         let fps_counter = if config.fps_counter {
             Some(FpsCounter {
                 start: Instant::now(),
@@ -24,6 +24,7 @@ impl Window {
         } else {
             None
         };
+        config.scale = config.scale.max(1.).min(20.);
         if config.enable_gpu {
             match gpu::Window::new(config) {
                 Ok(window) => {
@@ -71,6 +72,7 @@ pub struct WindowConfig {
     filter: u32,
     fullscreen: bool,
     fps_counter: bool,
+    scale: f32,
 }
 
 impl Default for WindowConfig {
@@ -80,6 +82,7 @@ impl Default for WindowConfig {
             filter: 5,
             fullscreen: false,
             fps_counter: false,
+            scale: 2.,
         }
     }
 }
@@ -102,6 +105,10 @@ impl WindowConfig {
         }
         self.fullscreen = args.contains("--fullscreen");
         self.fps_counter = args.contains("--fps");
+        self.scale = args
+            .opt_value_from_str("--scale")
+            .unwrap()
+            .unwrap_or(self.scale);
     }
 }