LightCraft: An open-source non-destructive photo editor written by scratch in pure Rust
Faces: whole networks on the GPU, batched faces, CPU-only switch, paced GPU use
Andrew Frahnf625c74c7a87
diff --git a/crates/engine/src/cmd/face_detect.rs b/crates/engine/src/cmd/face_detect.rsindex 8e706c367..718420ffc 100644--- a/crates/engine/src/cmd/face_detect.rs+++ b/crates/engine/src/cmd/face_detect.rs@@ -17,7 +17,7 @@ use lightcraft_faces::yunet::{Detector, Face, Options};use lightcraft_meta::{Region, RegionKind};use serde_json::{Value, json};-use super::{CommandSpec, bad, cmd, has_selection};+use super::{CommandSpec, always, bad, cmd, has_selection};use crate::{Result, Session};const C: &str = "faces.detect";@@ -34,10 +34,12 @@ const EXISTING_IOU: f64 = 0.3;/// What to tell someone who has not downloaded the detector yet.const NOT_INSTALLED: &str = "finding faces needs the YuNet face detector (232 KB, MIT): download it in Settings > Faces";+/// The YuNet loaded last, and from where.+static LOADED: Mutex<Option<(PathBuf, Arc<Detector>)>> = Mutex::new(None);+/// The installed YuNet, loaded once and kept (a new one is loaded when the model folder changes)./// (also used by the background scan, which aligns faces by the detector's landmarks and finds faces in photos that have none)pub(crate) fn detector(s: &Session) -> std::result::Result<Arc<Detector>, String> {- static LOADED: Mutex<Option<(PathBuf, Arc<Detector>)>> = Mutex::new(None);let home = s.face_models_dir.as_deref().ok_or("this build has nowhere to keep face models (the desktop app does)")?.join(YUNET_ID);let path = home.join("model.onnx");// installing writes the manifest last, so a folder without one is not (yet) a model@@ -59,33 +61,128 @@ pub(crate) fn detector(s: &Session) -> std::result::Result<Arc<Detector>, String#[cfg(not(target_arch = "wasm32"))]{let net = d.clone();- offer_gpu("detection", move |gpu| net.accelerate(gpu, true));+ offer_gpu(s, "detection", move |gpu| net.accelerate(gpu, true));}*loaded = Some((path, d.clone()));Ok(d)}+/// Where a face model runs, for Settings ▸ Faces and `faces.gpu`.+#[derive(Clone, Debug, Default, serde::Serialize)]+#[serde(rename_all = "camelCase")]+pub(crate) struct GpuState {+ /// The background check is still running.+ pub checking: bool,+ /// The GPU it runs on; `None` when it runs on the CPU.+ pub device: Option<String>,+ /// Why it runs on the CPU.+ pub why: Option<String>,+}++/// Each model's [`GpuState`], by what it does ("detection", "recognition").+static GPU_STATE: Mutex<Vec<(&'static str, GpuState)>> = Mutex::new(Vec::new());++pub(crate) fn gpu_state(what: &str) -> Option<GpuState> {+ GPU_STATE.lock().unwrap_or_else(PoisonError::into_inner).iter().find(|(w, _)| *w == what).map(|(_, st)| st.clone())+}++fn set_gpu_state(what: &'static str, st: GpuState) {+ let mut all = GPU_STATE.lock().unwrap_or_else(PoisonError::into_inner);+ match all.iter_mut().find(|(w, _)| *w == what) {+ Some((_, old)) => *old = st,+ None => all.push((what, st)),+ }+}++/// The user chose the CPU for face models (Settings ▸ Faces).+pub(crate) fn cpu_only(s: &Session) -> bool {+ s.face_models_dir.as_deref().is_some_and(|d| super::face_models::read_settings(d).cpu_only)+}+/// Check in the background whether the GPU runs a face model faster than the CPU and gives its results; the model-/// runs on the CPU meanwhile, and stays there when the GPU is off, missing, wrong or slower.+/// runs on the CPU meanwhile, and stays there when the GPU is off, missing, wrong or slower, or the user chose the CPU.+/// The check shows in the activity stack while it runs.#[cfg(not(target_arch = "wasm32"))]pub(crate) fn offer_gpu(+ s: &Session,what: &'static str,attach: impl FnOnce(Arc<dyn lightcraft_faces::net::Accelerator>) -> std::result::Result<(), String> + Send + 'static,) {+ if cpu_only(s) {+ set_gpu_state(what, GpuState { why: Some("the CPU is chosen in Settings".into()), ..GpuState::default() });+ return;+ }+ let task = s.activity.start("facesGpu", "Checking the GPU for faces", crate::activity::Cancel::No);+ task.detail(what);+ set_gpu_state(what, GpuState { checking: true, ..GpuState::default() });let spawned = std::thread::Builder::new().name(format!("faces-gpu-{what}")).spawn(move || {+ let _task = task;let gpu = match lightcraft_gpu::faces::accelerator() {Ok(gpu) => gpu,- Err(why) => return log::info!("faces: {what} runs on the CPU ({why})"),+ Err(why) => {+ log::info!("faces: {what} runs on the CPU ({why})");+ return set_gpu_state(what, GpuState { why: Some(why), ..GpuState::default() });+ }};let device = lightcraft_faces::net::Accelerator::device(&*gpu);match attach(gpu) {- Ok(()) => log::info!("faces: {what} runs on {device}"),- Err(why) => log::info!("faces: {what} stays on the CPU: {why}"),+ Ok(()) => {+ log::info!("faces: {what} runs on {device}");+ set_gpu_state(what, GpuState { device: Some(device), ..GpuState::default() });+ }+ Err(why) => {+ log::info!("faces: {what} stays on the CPU: {why}");+ set_gpu_state(what, GpuState { why: Some(why), ..GpuState::default() });+ }}});if let Err(e) = spawned {log::warn!("faces: could not check the GPU for {what}: {e}");+ set_gpu_state(what, GpuState { why: Some(format!("could not check the GPU: {e}")), ..GpuState::default() });+ }+}++/// `faces.gpu`: where the face models run, and the switch that keeps them on the CPU.+fn gpu(s: &mut Session, p: &Value) -> Result<Value> {+ const C: &str = "faces.gpu";+ if let Some(cpu) = p.get("cpuOnly") {+ let cpu = cpu.as_bool().ok_or_else(|| bad(C, "`cpuOnly` is true or false"))?;+ let dir = s.face_models_dir.clone().ok_or_else(|| bad(C, "this build has nowhere to keep face models (the desktop app does)"))?;+ super::face_models::set_cpu_only(&dir, cpu)?;+ let detector = LOADED.lock().unwrap_or_else(PoisonError::into_inner).as_ref().map(|(_, d)| d.clone());+ #[cfg(not(target_arch = "wasm32"))]+ let embedder = s.faces.embedder.as_ref().map(|(_, e)| e.clone());+ #[cfg(target_arch = "wasm32")]+ let embedder: Option<Arc<lightcraft_faces::runtime::Embedder>> = None;+ if cpu {+ for (what, net) in [("detection", detector.as_ref().map(|d| d.net())), ("recognition", embedder.as_ref().map(|e| e.net()))] {+ if let Some(net) = net {+ net.detach();+ }+ set_gpu_state(what, GpuState { why: Some("the CPU is chosen in Settings".into()), ..GpuState::default() });+ }+ } else {+ #[cfg(not(target_arch = "wasm32"))]+ {+ if let Some(d) = detector {+ offer_gpu(s, "detection", move |gpu| d.accelerate(gpu, true));+ }+ if let Some(e) = embedder {+ offer_gpu(s, "recognition", move |gpu| e.accelerate(gpu, true).map_err(|e| e.to_string()));+ }+ }+ }}+ Ok(gpu_json(s))+}++/// What `faces.gpu` and `faces.models.list` say about the GPU.+pub(crate) fn gpu_json(s: &Session) -> Value {+ json!({+ "cpuOnly": cpu_only(s),+ "detection": gpu_state("detection"),+ "recognition": gpu_state("recognition"),+ })}fn face_json(f: &Face) -> Value {@@ -175,15 +272,26 @@ fn detect(s: &mut Session, p: &Value) -> Result<Value> {}pub fn specs() -> Vec<CommandSpec> {- vec![cmd!(- "faces.detect",- "Detect Faces",- ["Photo"],- None,- "{ids?, apply?: bool (default true), score?: 0.6, nmsIou?: 0.3, maxFaces?} → {photos: [{id, faces: [{rect, score, landmarks}]}]}. Finds faces with the YuNet detector (download it first: `faces.models.download {id: \"yunet-2023mar\"}`) and, unless `apply` is false, adds them as unnamed face regions (one undo step), replacing earlier detections but never regions from XMP or ones you named; the sidecar is not touched",- has_selection,- detect- )]+ vec![+ cmd!(+ "faces.detect",+ "Detect Faces",+ ["Photo"],+ None,+ "{ids?, apply?: bool (default true), score?: 0.6, nmsIou?: 0.3, maxFaces?} → {photos: [{id, faces: [{rect, score, landmarks}]}]}. Finds faces with the YuNet detector (download it first: `faces.models.download {id: \"yunet-2023mar\"}`) and, unless `apply` is false, adds them as unnamed face regions (one undo step), replacing earlier detections but never regions from XMP or ones you named; the sidecar is not touched",+ has_selection,+ detect+ ),+ cmd!(+ query "faces.gpu",+ "Face Models on the GPU",+ [],+ None,+ "{cpuOnly?: bool} → {cpuOnly, detection, recognition}: where the face models run, each as {checking, device, why} (null until it is loaded). The GPU is used only after a background check shows it gives the CPU's results faster; `cpuOnly: true` keeps every face model on the CPU (Settings ▸ Faces), `false` checks the GPU again",+ always,+ gpu+ ),+ ]}#[cfg(test)]diff --git a/crates/engine/src/cmd/face_models.rs b/crates/engine/src/cmd/face_models.rsindex 6033c4dcc..1ea39d04c 100644--- a/crates/engine/src/cmd/face_models.rs+++ b/crates/engine/src/cmd/face_models.rs@@ -27,6 +27,9 @@ pub(super) struct FaceSettings {pub(super) enabled: bool,/// The installed recogniser in use.pub(super) embedder: Option<String>,+ /// Keep every face model on the CPU, even when the GPU would be faster.+ #[serde(skip_serializing_if = "std::ops::Not::not")]+ pub(super) cpu_only: bool,}fn fail(what: &str, e: impl std::fmt::Display) -> EngineError {@@ -57,6 +60,13 @@ fn write_atomic(path: &Path, bytes: &[u8]) -> Result<()> {})}+/// Keep the face models on the CPU (or let them use the GPU again).+pub(super) fn set_cpu_only(dir: &Path, on: bool) -> Result<()> {+ let mut st = read_settings(dir);+ st.cpu_only = on;+ write_settings(dir, &st)+}+fn write_settings(dir: &Path, st: &FaceSettings) -> Result<()> {std::fs::create_dir_all(dir).map_err(|e| fail("could not create the face models folder", e))?;write_atomic(&dir.join("settings.json"), &serde_json::to_vec_pretty(st).map_err(|e| fail("settings", e))?)@@ -195,6 +205,8 @@ fn list(s: &mut Session, _: &Value) -> Result<Value> {"dir": dir.as_ref().map(|d| d.display().to_string()),"enabled": settings.enabled,"embedder": settings.embedder,+ // where the models run: {cpuOnly, detection, recognition} as `faces.gpu` gives it+ "gpu": super::face_detect::gpu_json(s),// Native builds always have the pure-Rust runner; model files/background scanning remain native-only."runtime": cfg!(not(target_arch = "wasm32")),"models": models,diff --git a/crates/engine/src/cmd/face_recognize.rs b/crates/engine/src/cmd/face_recognize.rsindex 094735c6d..509a5bc70 100644--- a/crates/engine/src/cmd/face_recognize.rs+++ b/crates/engine/src/cmd/face_recognize.rs@@ -235,7 +235,7 @@ mod imp {}let embedder = Arc::new(Embedder::load(&dir.join(&id).join("model.onnx"), &installed.manifest).map_err(|e| bad(cmd, e.to_string()))?);let net = embedder.clone();- super::super::face_detect::offer_gpu("recognition", move |gpu| net.accelerate(gpu, true).map_err(|e| e.to_string()));+ super::super::face_detect::offer_gpu(s, "recognition", move |gpu| net.accelerate(gpu, true).map_err(|e| e.to_string()));let dim = match installed.manifest.output {lightcraft_faces::OutputSpec::Embedding { dim } => dim as usize,_ => return Err(bad(cmd, "the chosen model is not a recognition model")),@@ -584,6 +584,9 @@ mod imp {if let Some(w) = &s.faces.worker {w.allow(target, pace);}+ // the GPU is held back the same way: a light pace leaves it mostly free for the window's own renders+ let (runs, share) = crate::faces_worker::gpu_pace(pace);+ lightcraft_gpu::faces::set_pace(runs, share);while target > 0 && s.faces.in_flight.len() < target + PREFETCH {let Some(id) = s.faces.queue.pop() else { break };let Some(job) = prepare(s, id, true) else { continue };diff --git a/crates/engine/src/faces_worker.rs b/crates/engine/src/faces_worker.rsindex 81ff6adae..e44ad77c9 100644--- a/crates/engine/src/faces_worker.rs+++ b/crates/engine/src/faces_worker.rs@@ -127,6 +127,17 @@ fn pool_threads_for(cores: usize, pace: Pace, cap: Option<usize>) -> usize {cap.map_or(n, |c| n.min(c.max(1)))}+/// How much of the GPU face models may take at `pace`: face model runs at once, and the share of the time it may be+/// busy. Paused or light: one run at a time, busy a quarter of the time (the user is working, and the window renders+/// on the same GPU); normal: two at once, busy most of the time; full: as fast as it goes.+pub(crate) fn gpu_pace(pace: Pace) -> (usize, f32) {+ match pace {+ Pace::Pause | Pace::Light => (1, 0.25),+ Pace::Normal => (2, 0.75),+ Pace::Full => (4, 1.0),+ }+}+/// One photo's faces to embed (and, if `detect`, to find first).pub(crate) struct Prepared {/// `<model id>@<sha256>` the embeddings are for.@@ -197,7 +208,8 @@ pub(crate) fn process(p: Prepared, embedder: &Embedder) -> Done {let img = Rgb { data: &rgb, width: w, height: h };let (iw, ih) = embedder.input_size();let scaled = |f: &Face| f.landmarks.map(|(x, y)| (x * w as f32, y * h as f32));- let mut results: Vec<(Rect, Option<Vec<f32>>, Rect)> = todo+ // every face's crop first, then all of them through the model at once (one GPU dispatch for a group photo)+ let mut faces: Vec<(Rect, Option<Vec<u8>>, Rect)> = todo.into_iter().map(|rect| {let best = found.iter().map(|f| (overlap(&rect, f), f)).filter(|(o, _)| *o >= MATCH_IOU).max_by(|a, b| a.0.total_cmp(&b.0));@@ -214,20 +226,18 @@ pub(crate) fn process(p: Prepared, embedder: &Embedder) -> Done {ih,),};- let v = crop.and_then(|c| embedder.embed(&c).ok());// shown by the detector's box where it found this face: every face then has the same tightnesslet view = best.and_then(|(_, f)| face_rect(f)).unwrap_or(rect);- (rect, v, view)+ (rect, crop, view)}).collect();// a photo with no face regions: what the detector is sure of becomes its faces, embedded in the same passlet new_faces = if detect {- detected.map(|faces| {+ detected.map(|found| {let mut rects = Vec::new();- for f in faces.iter().filter(|f| f.score >= REGION_SCORE) {+ for f in found.iter().filter(|f| f.score >= REGION_SCORE) {let Some(rect) = face_rect(f) else { continue };- let v = align_to_template(&img, &scaled(f), iw, ih).and_then(|c| embedder.embed(&c).ok());- results.push((rect, v, rect));+ faces.push((rect, align_to_template(&img, &scaled(f), iw, ih), rect));rects.push(rect);}rects@@ -235,6 +245,10 @@ pub(crate) fn process(p: Prepared, embedder: &Embedder) -> Done {} else {None};+ let crops: Vec<&[u8]> = faces.iter().filter_map(|(_, c, _)| c.as_deref()).collect();+ let mut embedded = embedder.embed_batch(&crops).into_iter();+ let results: Vec<(Rect, Option<Vec<f32>>, Rect)> =+ faces.into_iter().map(|(rect, crop, view)| (rect, crop.and_then(|_| embedded.next().and_then(Result::ok)), view)).collect();(results, new_faces)}));let (results, found) = run.unwrap_or_else(|_| (fallback(rects), None));diff --git a/crates/engine/src/tests_face_models.rs b/crates/engine/src/tests_face_models.rsindex f1c27c849..00d6dc0cc 100644--- a/crates/engine/src/tests_face_models.rs+++ b/crates/engine/src/tests_face_models.rs@@ -613,3 +613,17 @@ fn recognition_cannot_be_turned_on_without_a_chosen_model() {assert!(s.execute("faces.enable", &json!({"enabled": true})).is_err());let _ = std::fs::remove_dir_all(&d);}++#[test]+fn the_cpu_only_switch_is_kept_and_bad_values_are_refused() {+ let d = temp("cpu-only");+ let mut s = session(&d);+ assert_eq!(s.execute("faces.gpu", &json!({})).unwrap()["cpuOnly"], false);+ assert_eq!(s.execute("faces.gpu", &json!({"cpuOnly": true})).unwrap()["cpuOnly"], true);+ // kept with the models, so a new session starts on the CPU too+ let mut again = session(&d);+ assert_eq!(again.execute("faces.models.list", &json!({})).unwrap()["gpu"]["cpuOnly"], true);+ assert!(again.execute("faces.gpu", &json!({"cpuOnly": "yes"})).is_err());+ assert_eq!(again.execute("faces.gpu", &json!({"cpuOnly": false})).unwrap()["cpuOnly"], false);+ let _ = std::fs::remove_dir_all(d);+}diff --git a/crates/faces/src/net.rs b/crates/faces/src/net.rsindex 3983404c5..430335840 100644--- a/crates/faces/src/net.rs+++ b/crates/faces/src/net.rs@@ -12,6 +12,7 @@ use std::rc::Rc;use std::sync::{Arc, PoisonError, RwLock};mod kernels;+pub mod program;mod shape;use crate::graph::{Data, Graph, Node};@@ -111,6 +112,12 @@ pub trait Accelerator: Send + Sync {fn device(&self) -> String;/// `[cout][oh][ow]`, or `Ok(None)` to leave this one to the CPU (too small to be worth a dispatch, say).fn conv(&self, job: &ConvJob) -> std::result::Result<Option<Vec<f32>>, String>;+ /// The whole network on the device, keeping every value there between steps, for each of `inputs` (all of+ /// `program.lens[INPUT]` values) in one go: per input, each of `program.outputs` in order. `Ok(None)` leaves the+ /// run to the CPU path (which may still use [`Accelerator::conv`]).+ fn run_program(&self, _program: &program::Program, _inputs: &[&[f32]]) -> std::result::Result<Option<Vec<Vec<Vec<f32>>>>, String> {+ Ok(None)+ }}/// A loaded network, ready to run.@@ -190,7 +197,7 @@ impl Net {/// A convolution on the accelerator, if one is attached and takes it. A failing accelerator is dropped.fn accelerated_conv(&self, layer: usize, node: &Node, x: &Tensor) -> Result<Option<Tensor>> {let Some(accelerator) = self.accelerator() else { return Ok(None) };- let job = kernels::conv_job(layer, node, x, &self.graph)?;+ let job = kernels::conv_job(layer, node, &x.shape, &x.data, &self.graph)?;match accelerator.conv(&job) {Ok(Some(out)) if out.len() == job.cout * job.oh * job.ow => Ok(Some(Tensor::new(vec![1, job.cout, job.oh, job.ow], out)?)),Ok(None) => Ok(None),@@ -214,7 +221,67 @@ impl Net {/// Run the network on `input` (its single input); returns each output by name.pub fn run(&self, input: Tensor) -> Result<Vec<(String, Tensor)>> {- shape::tensor(&input.shape, input.data.len(), &input.data)?;+ self.run_batch(vec![input])?.pop().ok_or_else(|| NetError::Shape("the run gave nothing".into()))+ }++ /// Run the network on each of `inputs`; an accelerator that keeps the values on its device runs inputs of one shape+ /// together (one upload and one read-back for all of them).+ pub fn run_batch(&self, inputs: Vec<Tensor>) -> Result<Vec<Vec<(String, Tensor)>>> {+ for input in &inputs {+ shape::tensor(&input.shape, input.data.len(), &input.data)?;+ }+ if let Some(out) = self.run_on_device(&inputs)? {+ return Ok(out);+ }+ inputs.into_iter().map(|input| self.run_one(input)).collect()+ }++ /// The whole batch on the attached accelerator, or `None` to run it on the CPU path. A failing or misbehaving+ /// accelerator is dropped.+ fn run_on_device(&self, inputs: &[Tensor]) -> Result<Option<Vec<Vec<(String, Tensor)>>>> {+ let (Some(accelerator), Some(first)) = (self.accelerator(), inputs.first()) else { return Ok(None) };+ if inputs.iter().any(|t| t.shape != first.shape) {+ return Ok(None);+ }+ let Some(program) = self.program(&first.shape)? else { return Ok(None) };+ let slices: Vec<&[f32]> = inputs.iter().map(|t| t.data.as_slice()).collect();+ let started = web_time::Instant::now();+ let outs = match accelerator.run_program(&program, &slices) {+ Ok(Some(outs)) => outs,+ Ok(None) => return Ok(None),+ Err(_) => {+ self.detach();+ return Ok(None);+ }+ };+ let gathered: Option<Vec<Vec<(String, Tensor)>>> = (outs.len() == inputs.len())+ .then(|| {+ outs.into_iter()+ .map(|values| {+ (values.len() == program.outputs.len())+ .then(|| {+ values+ .into_iter()+ .zip(&program.outputs)+ .map(|(data, (name, _, shape))| Tensor::new(shape.clone(), data).ok().map(|t| (name.clone(), t)))+ .collect::<Option<Vec<_>>>()+ })+ .flatten()+ })+ .collect()+ })+ .flatten();+ if std::env::var_os("LIGHTCRAFT_PROFILE").is_some() {+ eprintln!("[faces-gpu] {} steps x {}: {:.3} ms", program.steps.len(), inputs.len(), started.elapsed().as_secs_f64() * 1000.0);+ }+ if gathered.is_none() {+ // wrong sizes or values that are not finite: the device is not trusted any more+ self.detach();+ }+ Ok(gathered)+ }++ fn run_one(&self, input: Tensor) -> Result<Vec<(String, Tensor)>> {self.validate_input(&input.shape)?;let mut scratch = kernels::Scratch::default();let profile = std::env::var_os("LIGHTCRAFT_PROFILE").is_some();@@ -745,6 +812,176 @@ mod tests {}}+ /// An accelerator that runs whole programs with plain loops (the definition of each step), counting its calls;+ /// `poison` makes it give a value that is not finite.+ struct ProgramFake {+ calls: std::sync::atomic::AtomicUsize,+ poison: bool,+ }++ fn run_steps(p: &program::Program, input: &[f32]) -> Vec<Vec<f32>> {+ use program::{Arg, Binary, Op, Unary};+ let mut values: Vec<Vec<f32>> = vec![Vec::new(); p.lens.len()];+ values[program::INPUT] = input.to_vec();+ for step in &p.steps {+ let arg = |i: usize| -> Vec<f32> {+ match &step.args[i] {+ Arg::Value(v) => values[*v].clone(),+ Arg::Const { data, .. } => data.to_vec(),+ }+ };+ let len = p.lens[step.out];+ let x = arg(0);+ let out: Vec<f32> = match &step.op {+ Op::Conv(j) => {+ let (m, cg) = (j.cout / j.groups, j.cin / j.groups);+ let mut out = vec![0f32; len];+ for co in 0..j.cout {+ for oy in 0..j.oh {+ for ox in 0..j.ow {+ let mut v = j.bias.map_or(0.0, |b| b[co]);+ for ci in 0..cg {+ for ky in 0..j.kh {+ for kx in 0..j.kw {+ let (iy, ix) = ((oy * j.sh + ky) as isize - j.pt as isize, (ox * j.sw + kx) as isize - j.pl as isize);+ if iy >= 0 && ix >= 0 && (iy as usize) < j.h && (ix as usize) < j.w {+ v += j.weights[((co * cg + ci) * j.kh + ky) * j.kw + kx]+ * x[((co / m * cg + ci) * j.h + iy as usize) * j.w + ix as usize];+ }+ }+ }+ }+ out[(co * j.oh + oy) * j.ow + ox] = v;+ }+ }+ }+ out+ }+ Op::Unary(Unary::Relu) => x.iter().map(|v| v.max(0.0)).collect(),+ Op::Unary(Unary::Sigmoid) => x.iter().map(|v| 1.0 / (1.0 + (-v).exp())).collect(),+ Op::ChannelAffine { channels, plane } => {+ let (s, o) = (arg(1), arg(2));+ x.iter().enumerate().map(|(i, v)| v * s[(i / plane) % channels] + o[(i / plane) % channels]).collect()+ }+ Op::ChannelPrelu { channels, plane } => {+ let a = arg(1);+ x.iter().enumerate().map(|(i, v)| if *v < 0.0 { v * a[(i / plane) % channels] } else { *v }).collect()+ }+ Op::Binary { op, dims, a, b } => {+ let y = arg(1);+ (0..len)+ .map(|i| {+ let (mut rest, mut ia, mut ib) = (i, 0, 0);+ for d in (0..4).rev() {+ let at = rest % dims[d];+ rest /= dims[d];+ ia += at * a[d];+ ib += at * b[d];+ }+ match op {+ Binary::Add => x[ia] + y[ib],+ Binary::Sub => x[ia] - y[ib],+ Binary::Mul => x[ia] * y[ib],+ }+ })+ .collect()+ }+ Op::Transpose { dims, strides } => (0..len)+ .map(|i| {+ let (mut rest, mut from) = (i, 0);+ for d in (0..4).rev() {+ from += (rest % dims[d]) * strides[d];+ rest /= dims[d];+ }+ x[from]+ })+ .collect(),+ Op::MaxPool { h, w, kh, kw, sh, sw, oh, ow, .. } => (0..len)+ .map(|i| {+ let (ox, oy, c) = (i % ow, (i / ow) % oh, i / (ow * oh));+ let mut m = f32::NEG_INFINITY;+ for ky in 0..*kh {+ for kx in 0..*kw {+ m = m.max(x[(c * h + oy * sh + ky) * w + ox * sw + kx]);+ }+ }+ m+ })+ .collect(),+ Op::Resize { h, w, sh, sw, .. } => (0..len)+ .map(|i| {+ let (ow, oh) = (w * sw, h * sh);+ let (ox, oy, c) = (i % ow, (i / ow) % oh, i / (ow * oh));+ x[(c * h + oy / sh) * w + ox / sw]+ })+ .collect(),+ Op::GlobalAverage { plane, .. } => x.chunks(*plane).map(|c| c.iter().sum::<f32>() / *plane as f32).collect(),+ Op::Gemv { k, n, trans_b, alpha } => {+ let (wt, bias) = (arg(1), arg(2));+ (0..*n).map(|o| alpha * (0..*k).map(|j| x[j] * wt[if *trans_b { o * k + j } else { j * n + o }]).sum::<f32>() + bias[o]).collect()+ }+ };+ assert_eq!(out.len(), len, "{:?}", step.op);+ values[step.out] = out;+ for f in &step.free {+ values[*f] = Vec::new();+ }+ }+ p.outputs.iter().map(|(_, v, _)| values[*v].clone()).collect()+ }++ impl Accelerator for ProgramFake {+ fn device(&self) -> String {+ "program test".into()+ }+ fn conv(&self, _: &ConvJob) -> std::result::Result<Option<Vec<f32>>, String> {+ Ok(None)+ }+ fn run_program(&self, p: &program::Program, inputs: &[&[f32]]) -> std::result::Result<Option<Vec<Vec<Vec<f32>>>>, String> {+ self.calls.fetch_add(1, std::sync::atomic::Ordering::Relaxed);+ let mut out: Vec<Vec<Vec<f32>>> = inputs.iter().map(|x| run_steps(p, x)).collect();+ if self.poison {+ out[0][0][0] = f32::NAN;+ }+ Ok(Some(out))+ }+ }++ fn every_operator_input(seed: usize) -> Tensor {+ input(4, 8, 8, |i| (((i * 31 + seed * 7) % 37) as f32 - 18.0) * 0.05)+ }++ #[test]+ fn a_program_of_every_operator_gives_what_the_interpreter_gives() {+ let net = Net::new(crate::synthetic::every_operator_graph()).unwrap();+ let x = every_operator_input(0);+ let cpu = net.run(x.clone()).unwrap();+ let fake = Arc::new(ProgramFake { calls: 0.into(), poison: false });+ net.attach_checked(fake.clone(), &x, 1e-5, false).unwrap();+ let calls = fake.calls.load(std::sync::atomic::Ordering::Relaxed);+ assert!(calls > 0, "the whole network ran as one program");+ // several inputs go to the device together, and each gets its own outputs+ let inputs: Vec<Tensor> = (0..3).map(every_operator_input).collect();+ let together = net.run_batch(inputs.clone()).unwrap();+ assert_eq!(fake.calls.load(std::sync::atomic::Ordering::Relaxed), calls + 1, "one dispatch for the batch");+ net.detach();+ for (x, got) in inputs.into_iter().zip(&together) {+ let want = net.run(x).unwrap();+ agree(&want, got, 1e-5).unwrap();+ }+ assert_eq!(cpu.len(), 2);+ }++ #[test]+ fn a_device_that_gives_values_that_are_not_finite_is_dropped() {+ let net = Net::new(crate::synthetic::every_operator_graph()).unwrap();+ let x = every_operator_input(1);+ let cpu = net.run(x.clone()).unwrap();+ assert!(net.attach_checked(Arc::new(ProgramFake { calls: 0.into(), poison: true }), &x, 1e-5, false).is_err());+ assert!(net.accelerator().is_none());+ assert_eq!(net.run(x).unwrap(), cpu);+ }+#[test]fn unsupported_and_inconsistent_models_are_errors() {let x = input(1, 3, 3, |i| i as f32);diff --git a/crates/faces/src/net/kernels.rs b/crates/faces/src/net/kernels.rsindex 6d7c1dbb9..71ae95874 100644--- a/crates/faces/src/net/kernels.rs+++ b/crates/faces/src/net/kernels.rs@@ -60,8 +60,8 @@ fn shifted_plane(plane: &[f32], width: usize, ky: usize, kx: usize, pt: usize, pOk(())}/// The checked geometry of a convolution node, for an accelerator.-pub(super) fn conv_job<'a>(layer: usize, n: &Node, x: &'a Tensor, g: &'a Graph) -> Result<ConvJob<'a>> {- let (cin, h, w) = chw(x)?;+pub(super) fn conv_job<'a>(layer: usize, n: &Node, shape: &[usize], input: &'a [f32], g: &'a Graph) -> Result<ConvJob<'a>> {+ let (cin, h, w) = program::chw_of(shape)?;let wt = shape::weight(n, g, 1)?;let weights = shape::floats(n, g, 1)?;let [cout, cg, kh, kw] = wt.dims.as_slice() else {@@ -84,7 +84,7 @@ pub(super) fn conv_job<'a>(layer: usize, n: &Node, x: &'a Tensor, g: &'a Graph)return shape_err("the convolution bias does not match");}let (oh, ow) = ((ph - kh) / sh + 1, (pw - kw) / sw + 1);- Ok(ConvJob { layer, cin, h, w, cout, groups, kh, kw, sh, sw, pt, pl, oh, ow, weights, bias, input: &x.data })+ Ok(ConvJob { layer, cin, h, w, cout, groups, kh, kw, sh, sw, pt, pl, oh, ow, weights, bias, input })}pub(super) fn conv(n: &Node, x: &Tensor, g: &Graph, scratch: &mut Scratch) -> Result<Tensor> {let (cin, h, w) = chw(x)?;@@ -208,7 +208,7 @@ pub(super) fn conv(n: &Node, x: &Tensor, g: &Graph, scratch: &mut Scratch) -> ReOk(out)}/// Flat offset under ONNX's trailing-dimension broadcasting; shapes are checked before execution.-fn broadcast_offset(mut at: usize, out: &[usize], input: &[usize]) -> usize {+pub(super) fn broadcast_offset(mut at: usize, out: &[usize], input: &[usize]) -> usize {let mut offset = 0;let mut stride = 1;for (axis, &size) in out.iter().enumerate().rev() {diff --git a/crates/faces/src/net/program.rs b/crates/faces/src/net/program.rsnew file mode 100644index 000000000..e00e51da1--- /dev/null+++ b/crates/faces/src/net/program.rs@@ -0,0 +1,315 @@+//! A network lowered to a flat list of steps over numbered values, for an [`Accelerator`] that keeps every activation+//! on its device between layers (one upload, one read-back per run instead of one per convolution).+//!+//! Shapes are resolved for one input shape; every step names the values it reads and the one it writes, and which+//! values nothing reads after it. Shape-only operators (`Reshape`, `Flatten`, `Identity`, `Dropout`) make no step: their+//! output is the same value under another name. Anything this cannot express (a broadcast past four dimensions, a+//! transposed or batched `Gemm`, a slope that is not per channel) makes the whole lowering `None`, and the net runs+//! the CPU path instead.++use std::borrow::Cow;++use super::*;++/// Value 0 is the network's input.+pub const INPUT: usize = 0;++/// An operand: a value computed on the device, or a constant (uploaded once and kept, by `id`).+#[derive(Clone, Debug)]+pub enum Arg<'a> {+ Value(usize),+ /// `id` is the same for every run of the same net, so the device can keep the data.+ Const {+ id: usize,+ data: Cow<'a, [f32]>,+ },+}++#[derive(Clone, Copy, Debug, PartialEq, Eq)]+pub enum Unary {+ Relu,+ Sigmoid,+}++#[derive(Clone, Copy, Debug, PartialEq, Eq)]+pub enum Binary {+ Add,+ Sub,+ Mul,+}++/// What a step computes. Element counts come from the program's `lens` of the step's output.+#[derive(Clone, Debug)]+pub enum Op<'a> {+ /// A convolution; its `input` is empty (it reads the step's first argument).+ Conv(ConvJob<'a>),+ /// `out[i] = f(x[i])`. Args: `[x]`.+ Unary(Unary),+ /// `out[i] = x[i] * scale[c] + offset[c]`, `c = (i / plane) % channels`. Args: `[x, scale, offset]`.+ ChannelAffine { channels: usize, plane: usize },+ /// `out[i] = x[i] < 0 ? x[i] * slope[c] : x[i]`. Args: `[x, slope]` (`channels` values).+ ChannelPrelu { channels: usize, plane: usize },+ /// Broadcasting over four dimensions: `dims` of the output, element strides of each operand (0 where it repeats).+ /// Args: `[a, b]`.+ Binary { op: Binary, dims: [usize; 4], a: [usize; 4], b: [usize; 4] },+ /// `out[n] = alpha · Σk x[k] · B(k, n) + bias[n]`, `B(k, n) = weights[n·k_len + k]` when `trans_b`, else+ /// `weights[k·n_len + n]`. Args: `[x, weights, bias]` (bias already times beta; zeros when the node has none).+ Gemv { k: usize, n: usize, trans_b: bool, alpha: f32 },+ /// The mean of each `plane`-sized run. Args: `[x]`.+ GlobalAverage { channels: usize, plane: usize },+ /// Unpadded max pooling. Args: `[x]`.+ MaxPool { c: usize, h: usize, w: usize, kh: usize, kw: usize, sh: usize, sw: usize, oh: usize, ow: usize },+ /// Nearest-neighbour upscale by whole numbers. Args: `[x]`.+ Resize { c: usize, h: usize, w: usize, sh: usize, sw: usize },+ /// `out` in its own order (`dims`), read from the input at `strides` (the input's strides, permuted). Args: `[x]`.+ Transpose { dims: [usize; 4], strides: [usize; 4] },+}++#[derive(Clone, Debug)]+pub struct Step<'a> {+ pub op: Op<'a>,+ pub args: Vec<Arg<'a>>,+ pub out: usize,+ /// Values nothing reads after this step (never an output): their memory can be reused.+ pub free: Vec<usize>,+}++/// A whole network for one input shape.+#[derive(Clone, Debug)]+pub struct Program<'a> {+ /// Elements of each value; `lens[INPUT]` is the input's.+ pub lens: Vec<usize>,+ pub steps: Vec<Step<'a>>,+ /// Each graph output: its name, the value that holds it and its shape.+ pub outputs: Vec<(String, usize, Vec<usize>)>,+}++/// Leading ones up to rank four, or `None` past four.+fn rank4(s: &[usize]) -> Option<[usize; 4]> {+ let pad = 4usize.checked_sub(s.len())?;+ let mut out = [1; 4];+ for (i, d) in s.iter().enumerate() {+ *out.get_mut(pad + i)? = *d;+ }+ Some(out)+}++/// Element strides of `input` read as `out` under trailing-dimension broadcasting (0 along repeated axes).+fn broadcast_strides(out: [usize; 4], input: &[usize]) -> Option<[usize; 4]> {+ let dims = rank4(input)?;+ let mut strides = [0; 4];+ let mut step = 1;+ for axis in (0..4).rev() {+ let (d, o) = (*dims.get(axis)?, *out.get(axis)?);+ if d == o && d != 1 {+ *strides.get_mut(axis)? = step;+ } else if d != 1 {+ return None;+ }+ step *= d;+ }+ Some(strides)+}++impl Net {+ /// This network lowered for `input` (its shape), or `None` when a step has no device form.+ pub fn program(&self, input: &[usize]) -> Result<Option<Program<'_>>> {+ let g = &self.graph;+ let (shapes, _) = shape::infer_all(g, input)?;+ let Some((input_name, _)) = g.inputs.first() else { return shape_err("the model has no input") };+ let mut lens = vec![shape::elements(input)?];+ let mut names: HashMap<&str, (usize, Vec<usize>)> = HashMap::new();+ names.insert(input_name.as_str(), (INPUT, input.to_vec()));+ let mut steps: Vec<Step> = Vec::new();+ for (i, node) in g.nodes.iter().enumerate() {+ let Some(out_shape) = shapes.get(i) else { return shape_err("a node has no shape") };+ let Some(out_name) = node.outputs.first() else { return shape_err("a node has no output") };+ let value = |idx: usize| -> Option<&(usize, Vec<usize>)> { node.inputs.get(idx).and_then(|n| names.get(n.as_str())) };+ let constant = |idx: usize| -> Option<(Cow<'_, [f32]>, &[usize])> {+ let w = node.inputs.get(idx).and_then(|n| g.weights.get(n))?;+ match &w.data {+ Data::F32(v) => Some((Cow::Borrowed(v.as_slice()), w.dims.as_slice())),+ _ => None,+ }+ };+ let id = |k: usize| i * 4 + k;+ let Some((x, x_shape)) = value(0).cloned() else { return Ok(None) };+ let (op, args) = match node.op.as_str() {+ "Reshape" | "Flatten" | "Identity" | "Dropout" => {+ names.insert(out_name.as_str(), (x, out_shape.clone()));+ continue;+ }+ "Conv" => (Op::Conv(kernels::conv_job(i, node, &x_shape, &[], g)?), vec![Arg::Value(x)]),+ "Relu" => (Op::Unary(Unary::Relu), vec![Arg::Value(x)]),+ "Sigmoid" => (Op::Unary(Unary::Sigmoid), vec![Arg::Value(x)]),+ "BatchNormalization" => {+ let channels = *x_shape.get(1).ok_or_else(|| NetError::Shape("BatchNormalization rank".into()))?;+ let plane = shape::elements(x_shape.get(2..).unwrap_or(&[]))?;+ let (s, b, m, v) =+ (shape::floats(node, g, 1)?, shape::floats(node, g, 2)?, shape::floats(node, g, 3)?, shape::floats(node, g, 4)?);+ let eps = shape::float(node, "epsilon", 1e-5);+ let mut scale = Vec::with_capacity(channels);+ let mut offset = Vec::with_capacity(channels);+ for c in 0..channels {+ let (Some(s), Some(b), Some(m), Some(v)) = (s.get(c), b.get(c), m.get(c), v.get(c)) else {+ return shape_err("short BatchNormalization parameters");+ };+ let k = s / (v + eps).sqrt();+ scale.push(k);+ offset.push(b - m * k);+ }+ (+ Op::ChannelAffine { channels, plane },+ vec![Arg::Value(x), Arg::Const { id: id(1), data: scale.into() }, Arg::Const { id: id(2), data: offset.into() }],+ )+ }+ "PRelu" => {+ let Some((slope, dims)) = constant(1) else { return Ok(None) };+ let channels = x_shape.get(1).copied().unwrap_or(1);+ let plane = shape::elements(x_shape.get(2..).unwrap_or(&[]))?;+ let mut expanded = vec![1; x_shape.len().saturating_sub(dims.len())];+ expanded.extend_from_slice(dims);+ let per_channel = expanded.iter().enumerate().all(|(i, &d)| d == 1 || i == 1) && slope.len() == channels;+ let slope: Cow<[f32]> = if per_channel {+ slope+ } else if let [one] = slope.as_ref() {+ vec![*one; channels].into()+ } else {+ return Ok(None);+ };+ (Op::ChannelPrelu { channels, plane }, vec![Arg::Value(x), Arg::Const { id: id(1), data: slope }])+ }+ "Add" | "Sub" | "Mul" => {+ let op = match node.op.as_str() {+ "Add" => Binary::Add,+ "Sub" => Binary::Sub,+ _ => Binary::Mul,+ };+ let (b, b_shape) = match (value(1), constant(1)) {+ (Some((v, s)), _) => (Arg::Value(*v), s.clone()),+ (None, Some((data, dims))) => (Arg::Const { id: id(1), data }, dims.to_vec()),+ _ => return Ok(None),+ };+ let Some(dims) = rank4(out_shape) else { return Ok(None) };+ let (Some(sa), Some(sb)) = (broadcast_strides(dims, &x_shape), broadcast_strides(dims, &b_shape)) else { return Ok(None) };+ (Op::Binary { op, dims, a: sa, b: sb }, vec![Arg::Value(x), b])+ }+ "Gemm" => {+ let Some((weights, wdims)) = constant(1) else { return Ok(None) };+ let ([1, k], [n_out]) = (x_shape.as_slice(), out_shape.get(1..).unwrap_or(&[])) else { return Ok(None) };+ if node.int("transA").unwrap_or(0) != 0 || out_shape.first() != Some(&1) {+ return Ok(None);+ }+ let trans_b = node.int("transB").unwrap_or(0) != 0;+ if wdims.len() != 2 || weights.len() != k * n_out {+ return shape_err("Gemm weights do not match");+ }+ let mut bias = vec![0.0f32; *n_out];+ if node.inputs.get(2).is_some_and(|s| !s.is_empty()) {+ let wt = shape::weight(node, g, 2)?;+ let b = shape::floats(node, g, 2)?;+ let beta = shape::float(node, "beta", 1.0);+ for (j, to) in bias.iter_mut().enumerate() {+ *to = beta+ * b.get(kernels::broadcast_offset(j, out_shape, &wt.dims))+ .copied()+ .ok_or_else(|| NetError::Shape("short matrix bias".into()))?;+ }+ }+ (+ Op::Gemv { k: *k, n: *n_out, trans_b, alpha: shape::float(node, "alpha", 1.0) },+ vec![Arg::Value(x), Arg::Const { id: id(1), data: weights }, Arg::Const { id: id(2), data: bias.into() }],+ )+ }+ "GlobalAveragePool" => {+ let (c, h, w) = chw_of(&x_shape)?;+ (Op::GlobalAverage { channels: c, plane: h * w }, vec![Arg::Value(x)])+ }+ "MaxPool" => {+ let (c, h, w) = chw_of(&x_shape)?;+ let (kh, kw) = pair(node, "kernel_shape", 1)?;+ let (sh, sw) = pair(node, "strides", 1)?;+ if pads(node)? != (0, 0, 0, 0) {+ return Ok(None);+ }+ let (_, oh, ow) = chw_of(out_shape)?;+ (Op::MaxPool { c, h, w, kh, kw, sh, sw, oh, ow }, vec![Arg::Value(x)])+ }+ "Resize" => {+ let (c, h, w) = chw_of(&x_shape)?;+ let (_, oh, ow) = chw_of(out_shape)?;+ if oh % h != 0 || ow % w != 0 {+ return shape_err("a Resize that is not by whole numbers");+ }+ (Op::Resize { c, h, w, sh: oh / h, sw: ow / w }, vec![Arg::Value(x)])+ }+ "Transpose" => {+ let rank = x_shape.len();+ let default_perm: Vec<i64> = (0..rank as i64).rev().collect();+ let perm = node.ints("perm").unwrap_or(&default_perm);+ if rank > 4 || perm.len() != rank {+ return Ok(None);+ }+ let mut in_strides = vec![1usize; rank];+ for a in (0..rank.saturating_sub(1)).rev() {+ let next = in_strides.get(a + 1).copied().unwrap_or(1) * x_shape.get(a + 1).copied().unwrap_or(1);+ if let Some(s) = in_strides.get_mut(a) {+ *s = next;+ }+ }+ let Some(dims) = rank4(out_shape) else { return Ok(None) };+ let mut strides = [0usize; 4];+ for (j, p) in perm.iter().enumerate() {+ let p = usize::try_from(*p).map_err(|_| NetError::Shape("a negative permutation".into()))?;+ let s = in_strides.get(p).copied().ok_or_else(|| NetError::Shape("a bad permutation".into()))?;+ *strides.get_mut(4 - rank + j).ok_or_else(|| NetError::Shape("a bad permutation".into()))? = s;+ }+ (Op::Transpose { dims, strides }, vec![Arg::Value(x)])+ }+ _ => return Ok(None),+ };+ let out = lens.len();+ lens.push(shape::elements(out_shape)?);+ names.insert(out_name.as_str(), (out, out_shape.clone()));+ steps.push(Step { op, args, out, free: Vec::new() });+ }+ let outputs = g+ .outputs+ .iter()+ .map(|name| {+ let (v, s) = names.get(name.as_str()).ok_or_else(|| NetError::Shape(format!("output `{name}` was never computed")))?;+ Ok((name.clone(), *v, s.clone()))+ })+ .collect::<Result<Vec<_>>>()?;+ // the last step that reads each value; a value nothing reads is free right after it is made+ let mut last: Vec<Option<usize>> = vec![None; lens.len()];+ for (at, step) in steps.iter().enumerate() {+ if let Some(l) = last.get_mut(step.out) {+ *l = Some(at);+ }+ for a in &step.args {+ if let Arg::Value(v) = a+ && let Some(l) = last.get_mut(*v)+ {+ *l = Some(at);+ }+ }+ }+ for (v, at) in last.iter().enumerate() {+ if outputs.iter().any(|(_, o, _)| *o == v) {+ continue;+ }+ if let Some(step) = at.and_then(|at| steps.get_mut(at)) {+ step.free.push(v);+ }+ }+ Ok(Some(Program { lens, steps, outputs }))+ }+}++pub(super) fn chw_of(s: &[usize]) -> Result<(usize, usize, usize)> {+ match s {+ [1, c, h, w] => Ok((*c, *h, *w)),+ s => unsupported(format!("a tensor of shape {s:?} (only [1, channels, height, width])")),+ }+}diff --git a/crates/faces/src/net/shape.rs b/crates/faces/src/net/shape.rsindex d76271bc7..5b9fc68b1 100644--- a/crates/faces/src/net/shape.rs+++ b/crates/faces/src/net/shape.rs@@ -261,7 +261,12 @@ fn image(s: &[usize]) -> Result<(usize, usize, usize)> {}}pub(super) fn infer(g: &Graph, input: &[usize]) -> Result<Vec<Vec<usize>>> {+ Ok(infer_all(g, input)?.1)+}+/// Each node's output shape, and the graph's outputs' shapes.+pub(super) fn infer_all(g: &Graph, input: &[usize]) -> Result<(Vec<Vec<usize>>, Vec<Vec<usize>>)> {elements(input)?;+ let mut per_node = Vec::with_capacity(g.nodes.len());let mut shapes: HashMap<&str, Vec<usize>> = g.weights.iter().map(|(n, w)| (n.as_str(), w.dims.clone())).collect();shapes.insert(g.inputs.first().ok_or_else(|| NetError::Shape("no input".into()))?.0.as_str(), input.to_vec());let mut live = elements(input)?;@@ -484,6 +489,7 @@ pub(super) fn infer(g: &Graph, input: &[usize]) -> Result<Vec<Vec<usize>>> {.checked_add(out_count).filter(|&n| n <= 128 * 1024 * 1024).ok_or_else(|| NetError::Shape("live tensors exceed the 512 MiB working limit".into()))?;+ per_node.push(out.clone());shapes.insert(n.outputs.first().ok_or_else(|| NetError::Shape("no output".into()))?, out);let mut removed = std::collections::HashSet::new();for name in &n.inputs {@@ -492,5 +498,7 @@ pub(super) fn infer(g: &Graph, input: &[usize]) -> Result<Vec<Vec<usize>>> {}}}- g.outputs.iter().map(|s| shapes.get(s.as_str()).cloned().ok_or_else(|| NetError::Shape("missing output".into()))).collect()+ let outputs =+ g.outputs.iter().map(|s| shapes.get(s.as_str()).cloned().ok_or_else(|| NetError::Shape("missing output".into()))).collect::<Result<_>>()?;+ Ok((per_node, outputs))}diff --git a/crates/faces/src/runtime.rs b/crates/faces/src/runtime.rsindex 7a8ab948a..92f9e8134 100644--- a/crates/faces/src/runtime.rs+++ b/crates/faces/src/runtime.rs@@ -40,6 +40,9 @@ pub struct SelfTest {type Plan = Net;+/// Most faces [`Embedder::embed_batch`] sends to a GPU in one go.+pub const BATCH: usize = 16;+/// A loaded recognition model.#[derive(Clone)]pub struct Embedder {@@ -135,6 +138,66 @@ impl Embedder {/// The face as a unit-length vector. `rgb` is an aligned face at exactly [`Self::input_size`], 8-bit RGB.pub fn embed(&self, rgb: &[u8]) -> Result<Vec<f32>, RuntimeError> {+ let input = self.input(rgb)?;+ let out = guarded("running it", || {+ let mut result = self.plan.run(input).map_err(|e| RuntimeError::Run(e.to_string()))?;+ if result.len() != 1 {+ return Err(RuntimeError::Run("the model must give one embedding".into()));+ }+ Ok(result.pop().ok_or_else(|| RuntimeError::Run("no output".into()))?.1.data)+ })?;+ self.unit(out)+ }++ /// Several faces at once, each as [`Self::embed`] gives it. On a GPU they run together, up to [`BATCH`] per+ /// dispatch (one upload and one read-back for all of them); on the CPU one after another.+ pub fn embed_batch(&self, faces: &[&[u8]]) -> Vec<Result<Vec<f32>, RuntimeError>> {+ if self.plan.accelerator().is_none() || faces.len() < 2 {+ return faces.iter().map(|f| self.embed(f)).collect();+ }+ let mut results: Vec<Option<Result<Vec<f32>, RuntimeError>>> = faces.iter().map(|_| None).collect();+ let mut ready: Vec<(usize, Tensor)> = Vec::new();+ for (i, face) in faces.iter().enumerate() {+ match self.input(face) {+ Ok(t) => ready.push((i, t)),+ Err(e) => {+ if let Some(r) = results.get_mut(i) {+ *r = Some(Err(e));+ }+ }+ }+ }+ while !ready.is_empty() {+ let chunk: Vec<(usize, Tensor)> = ready.drain(..ready.len().min(BATCH)).collect();+ let (at, inputs): (Vec<usize>, Vec<Tensor>) = chunk.into_iter().unzip();+ let ran = guarded("running it", || self.plan.run_batch(inputs).map_err(|e| RuntimeError::Run(e.to_string())));+ match ran {+ Ok(outs) if outs.len() == at.len() => {+ for (i, mut out) in at.into_iter().zip(outs) {+ let r = match (out.len(), out.pop()) {+ (1, Some((_, t))) => self.unit(t.data),+ _ => Err(RuntimeError::Run("the model must give one embedding".into())),+ };+ if let Some(slot) = results.get_mut(i) {+ *slot = Some(r);+ }+ }+ }+ // one bad face spoils a batch: run them one by one so the others still get theirs+ _ => {+ for i in at {+ if let (Some(slot), Some(face)) = (results.get_mut(i), faces.get(i)) {+ *slot = Some(self.embed(face));+ }+ }+ }+ }+ }+ results.into_iter().map(|r| r.unwrap_or_else(|| Err(RuntimeError::Run("the face was not embedded".into())))).collect()+ }++ /// The model's input for an aligned face.+ fn input(&self, rgb: &[u8]) -> Result<Tensor, RuntimeError> {if rgb.len() != self.width * self.height * 3 {return Err(RuntimeError::Input("the face picture is not the size the model wants".into()));}@@ -149,14 +212,11 @@ impl Embedder {}}}- let out = guarded("running it", || {- let input = Tensor::new(vec![1, 3, self.height, self.width], data).map_err(|e| RuntimeError::Input(e.to_string()))?;- let mut result = self.plan.run(input).map_err(|e| RuntimeError::Run(e.to_string()))?;- if result.len() != 1 {- return Err(RuntimeError::Run("the model must give one embedding".into()));- }- Ok(result.pop().ok_or_else(|| RuntimeError::Run("no output".into()))?.1.data)- })?;+ Tensor::new(vec![1, 3, self.height, self.width], data).map_err(|e| RuntimeError::Input(e.to_string()))+ }++ /// The model's output, checked and scaled to unit length.+ fn unit(&self, out: Vec<f32>) -> Result<Vec<f32>, RuntimeError> {if out.len() != self.dim {return Err(RuntimeError::Run(format!("the model gave {} numbers per face, not the {} its description says", out.len(), self.dim)));}diff --git a/crates/faces/src/synthetic.rs b/crates/faces/src/synthetic.rsindex 61010a243..6bb792a1f 100644--- a/crates/faces/src/synthetic.rs+++ b/crates/faces/src/synthetic.rs@@ -248,3 +248,63 @@ pub fn tiny_detector_manifest() -> crate::manifest::ModelManifest {m.size_bytes = Some(bytes.len() as u64);m}++/// A small network that uses every operator the interpreter runs (both kinds of convolution, batch norm, PRelu,+/// sigmoid, broadcasting add/sub/mul with values and constants, max pooling, nearest upscaling, transposes, reshape,+/// global pooling, flatten and a fully connected layer), with two outputs, for checking an accelerator's program+/// against the CPU. Its input is `[1, 4, 8, 8]`.+pub fn every_operator_graph() -> crate::graph::Graph {+ use crate::graph::{Attr, Data, Graph, Node, Weight};+ let node = |op: &str, inputs: &[&str], output: &str, attrs: &[(&str, Attr)]| Node {+ op: op.into(),+ inputs: inputs.iter().map(|s| s.to_string()).collect(),+ outputs: vec![output.into()],+ attrs: attrs.iter().map(|(k, v)| (k.to_string(), v.clone())).collect(),+ };+ let wave = |n: usize, k: usize| -> Vec<f32> { (0..n).map(|i| (((i * k) % 23) as f32 - 11.0) * 0.04).collect() };+ let w = |dims: &[usize], v: Vec<f32>| Weight { dims: dims.to_vec(), data: Data::F32(v) };+ let weights = [+ ("w1", w(&[8, 4, 3, 3], wave(288, 7))),+ ("b1", w(&[8], wave(8, 5))),+ ("bn_s", w(&[8], (0..8).map(|i| 0.5 + i as f32 * 0.1).collect())),+ ("bn_b", w(&[8], wave(8, 3))),+ ("bn_m", w(&[8], wave(8, 11))),+ ("bn_v", w(&[8], (0..8).map(|i| 0.2 + i as f32 * 0.05).collect())),+ ("slope", w(&[8, 1, 1], (0..8).map(|i| 0.05 * i as f32).collect())),+ ("w2", w(&[8, 1, 3, 3], wave(72, 13))),+ ("scales", w(&[4], vec![1.0, 1.0, 2.0, 2.0])),+ ("half", w(&[1], vec![0.5])),+ ("per_channel", w(&[1, 8, 1, 1], wave(8, 17))),+ ("wg", w(&[5, 8], wave(40, 19))),+ ("bg", w(&[5], wave(5, 2))),+ ("flat", Weight { dims: vec![2], data: Data::I64(vec![1, -1]) }),+ ];+ let pads = |p: i64| ("pads", Attr::Ints(vec![p; 4]));+ let nodes = vec![+ node("Conv", &["x", "w1", "b1"], "c1", &[pads(1)]),+ node("BatchNormalization", &["c1", "bn_s", "bn_b", "bn_m", "bn_v"], "bn", &[]),+ node("PRelu", &["bn", "slope"], "pr", &[]),+ node("Conv", &["pr", "w2"], "c2", &[("group", Attr::Int(8)), ("strides", Attr::Ints(vec![2, 2])), pads(1)]),+ node("Sigmoid", &["c2"], "sg", &[]),+ node("Mul", &["c2", "sg"], "swish", &[]),+ node("Resize", &["swish", "", "scales"], "up", &[]),+ node("Add", &["up", "pr"], "sum", &[]),+ node("Sub", &["sum", "half"], "less", &[]),+ node("MaxPool", &["less"], "pool", &[("kernel_shape", Attr::Ints(vec![2, 2])), ("strides", Attr::Ints(vec![2, 2]))]),+ node("Relu", &["pool"], "relu", &[]),+ node("Transpose", &["relu"], "nhwc", &[("perm", Attr::Ints(vec![0, 2, 3, 1]))]),+ node("Transpose", &["nhwc"], "nchw", &[("perm", Attr::Ints(vec![0, 3, 1, 2]))]),+ node("Mul", &["nchw", "per_channel"], "scaled", &[]),+ node("Identity", &["scaled"], "same", &[]),+ node("GlobalAveragePool", &["same"], "avg", &[]),+ node("Flatten", &["avg"], "row", &[]),+ node("Reshape", &["row", "flat"], "row2", &[]),+ node("Gemm", &["row2", "wg", "bg"], "y", &[("transB", Attr::Int(1)), ("alpha", Attr::Float(0.5))]),+ ];+ Graph {+ nodes,+ weights: weights.into_iter().map(|(k, v)| (k.to_string(), v)).collect(),+ inputs: vec![("x".into(), vec![1, 4, 8, 8])],+ outputs: vec!["y".into(), "nhwc".into()],+ }+}diff --git a/crates/gpu/src/faces.rs b/crates/gpu/src/faces.rsindex ae982d30b..bee9929ef 100644--- a/crates/gpu/src/faces.rs+++ b/crates/gpu/src/faces.rs@@ -14,6 +14,7 @@ use std::sync::{Arc, Mutex, OnceLock, PoisonError};use std::time::Duration;use bytemuck::{Pod, Zeroable};+use lightcraft_faces::net::program::{Arg, Binary, INPUT, Op, Program, Unary};use lightcraft_faces::net::{Accelerator, ConvJob};use wgpu::util::DeviceExt;@@ -35,6 +36,22 @@ struct Dev {tiled: wgpu::ComputePipeline,direct: wgpu::ComputePipeline,reduce: wgpu::ComputePipeline,+ ops: Ops,+ /// Bound where a kernel has fewer operands than bindings.+ dummy: wgpu::Buffer,+}++/// The kernels of `wgsl/faces_ops.wgsl`.+struct Ops {+ unary: wgpu::ComputePipeline,+ affine: wgpu::ComputePipeline,+ prelu: wgpu::ComputePipeline,+ binary: wgpu::ComputePipeline,+ transpose: wgpu::ComputePipeline,+ maxpool: wgpu::ComputePipeline,+ resize: wgpu::ComputePipeline,+ gavg: wgpu::ComputePipeline,+ gemv: wgpu::ComputePipeline,}static DEV: OnceLock<Result<Dev, String>> = OnceLock::new();@@ -146,20 +163,43 @@ fn make_device(backends: wgpu::Backends) -> Result<Dev, String> {source: wgpu::ShaderSource::Wgsl(include_str!("wgsl/faces_conv.wgsl").into()),});let pl = device.create_pipeline_layout(&wgpu::PipelineLayoutDescriptor { label: None, bind_group_layouts: &[Some(&layout)], immediate_size: 0 });- let pipe = |entry: &str| {+ let ops_module = device.create_shader_module(wgpu::ShaderModuleDescriptor {+ label: Some("faces ops"),+ source: wgpu::ShaderSource::Wgsl(include_str!("wgsl/faces_ops.wgsl").into()),+ });+ let pipe_in = |module: &wgpu::ShaderModule, entry: &str| {device.create_compute_pipeline(&wgpu::ComputePipelineDescriptor {label: Some(entry),layout: Some(&pl),- module: &module,+ module,entry_point: Some(entry),// both kernels write all of their workgroup memory before reading itcompilation_options: wgpu::PipelineCompilationOptions { zero_initialize_workgroup_memory: false, ..Default::default() },cache: None,})};+ let pipe = |entry: &str| pipe_in(&module, entry);+ let op = |entry: &str| pipe_in(&ops_module, entry);let (tiled, direct, reduce) = (pipe("tiled"), pipe("direct"), pipe("reduce"));+ let ops = Ops {+ unary: op("unary"),+ affine: op("affine"),+ prelu: op("prelu"),+ binary: op("binary"),+ transpose: op("transpose"),+ maxpool: op("maxpool"),+ resize: op("resize"),+ gavg: op("gavg"),+ gemv: op("gemv"),+ };+ let dummy = device.create_buffer(&wgpu::BufferDescriptor {+ label: Some("faces unused"),+ size: 16,+ usage: wgpu::BufferUsages::STORAGE,+ mapped_at_creation: false,+ });log::info!("gpu (faces): {} ({:?})", info.name, info.backend);- Ok(Dev { device, queue, info, limits, layout, tiled, direct, reduce })+ Ok(Dev { device, queue, info, limits, layout, tiled, direct, reduce, ops, dummy })}/// The adapter face scans would use ("AMD Radeon RX 7900 XTX (Vulkan)"), or why they cannot use one. Creates the@@ -203,53 +243,151 @@ struct Resident {pub struct FaceGpu {dev: &'static Dev,layers: Mutex<HashMap<usize, Arc<Resident>>>,+ consts: Mutex<HashMap<usize, wgpu::Buffer>>,resident: AtomicU64,}/// A GPU accelerator for one face model, or why there is none (GPU off, missing or broken).pub fn accelerator() -> Result<Arc<FaceGpu>, String> {- Ok(Arc::new(FaceGpu { dev: dev()?, layers: Mutex::new(HashMap::new()), resident: AtomicU64::new(0) }))+ Ok(Arc::new(FaceGpu { dev: dev()?, layers: Mutex::new(HashMap::new()), consts: Mutex::new(HashMap::new()), resident: AtomicU64::new(0) }))}fn u32_of(v: usize) -> Result<u32, String> {u32::try_from(v).ok().filter(|v| *v < (1 << 31)).ok_or_else(|| "a layer too large for the GPU kernel".to_string())}+/// How much of the GPU face work may take: at most `concurrent` runs at once, and busy for at most `share` of the time+/// (after a run of `t`, the next one waits `t · (1 − share) / share`). Set by the background scan's pace.+struct Pace {+ concurrent: usize,+ share: f32,+ running: usize,+}++static PACE: Mutex<Pace> = Mutex::new(Pace { concurrent: usize::MAX, share: 1.0, running: 0 });+static PACE_FREED: std::sync::Condvar = std::sync::Condvar::new();++/// Limit the GPU work face models do: `concurrent` runs at once (at least one) and `share` of the time (0.05 to 1).+/// The background face scan sets it from its pace; anything else that runs a face model shares the same limit.+pub fn set_pace(concurrent: usize, share: f32) {+ let mut p = PACE.lock().unwrap_or_else(PoisonError::into_inner);+ p.concurrent = concurrent.max(1);+ p.share = if share.is_finite() { share.clamp(0.05, 1.0) } else { 1.0 };+ PACE_FREED.notify_all();+}++/// A turn on the GPU under the pace; waits for one. Dropping it gives it back.+struct Turn {+ share: f32,+ started: std::time::Instant,+}++impl Turn {+ fn take() -> Turn {+ let mut p = PACE.lock().unwrap_or_else(PoisonError::into_inner);+ while p.running >= p.concurrent {+ p = PACE_FREED.wait_timeout(p, Duration::from_millis(200)).map(|(g, _)| g).unwrap_or_else(|e| e.into_inner().0);+ }+ p.running += 1;+ Turn { share: p.share, started: std::time::Instant::now() }+ }+}++impl Drop for Turn {+ fn drop(&mut self) {+ // rest while still holding the turn, so the GPU stays idle for the rest of the period+ if self.share < 1.0 {+ let rest = self.started.elapsed().mul_f32((1.0 - self.share) / self.share).min(Duration::from_secs(2));+ std::thread::sleep(rest);+ }+ let mut p = PACE.lock().unwrap_or_else(PoisonError::into_inner);+ p.running = p.running.saturating_sub(1);+ PACE_FREED.notify_one();+ }+}++/// One dispatch to record: a pipeline, its parameters, its four buffers (three read, the last written) and its+/// workgroup counts.+struct Dispatch {+ pipeline: &'static wgpu::ComputePipeline,+ params: Vec<u8>,+ buffers: [wgpu::Buffer; 4],+ groups: [u32; 3],+}++/// Workgroups for `count` items of 256 (or `count` workgroups when `per_group`): x, y and the row length the kernel+/// reads back as `q(1)`.+fn flat(count: usize, limit: usize) -> Option<[u32; 3]> {+ let blocks = count.max(1);+ let row = blocks.clamp(1, limit.min(65_535));+ let rows = blocks.div_ceil(row);+ (rows <= limit).then_some([u32::try_from(row).ok()?, u32::try_from(rows).ok()?, u32::try_from(row).ok()?])+}++fn params24(values: &[u32]) -> Vec<u8> {+ let mut all = [0u32; 24];+ for (to, v) in all.iter_mut().zip(values) {+ *to = *v;+ }+ bytemuck::cast_slice(&all).to_vec()+}+impl FaceGpu {/// Bytes of weights this model keeps on the device.pub fn resident_bytes(&self) -> u64 {self.resident.load(Ordering::Relaxed)}- fn resident(&self, job: &ConvJob) -> Result<Option<Arc<Resident>>, String> {+ fn upload(&self, label: &str, data: &[f32]) -> wgpu::Buffer {+ let zero = [0f32; 4];+ self.dev.device.create_buffer_init(&wgpu::util::BufferInitDescriptor {+ label: Some(label),+ contents: bytemuck::cast_slice(if data.is_empty() { &zero } else { data }),+ usage: wgpu::BufferUsages::STORAGE,+ })+ }++ /// Whether `bytes` more may stay on the device; counts them if so.+ fn keep(&self, bytes: u64) -> bool {+ if self.resident_bytes().saturating_add(bytes) > MAX_RESIDENT {+ return false;+ }+ self.resident.fetch_add(bytes, Ordering::Relaxed);+ true+ }++ /// A convolution's weights and bias: kept on the device while the budget lasts, uploaded for this run past it.+ fn resident(&self, job: &ConvJob) -> Arc<Resident> {let mut layers = self.layers.lock().unwrap_or_else(PoisonError::into_inner);if let Some(r) = layers.get(&job.layer) {- return Ok(Some(r.clone()));+ return r.clone();}- let bytes = (job.weights.len() as u64 + job.cout as u64) * 4;- if self.resident_bytes().saturating_add(bytes) > MAX_RESIDENT {- return Ok(None);+ let r = Arc::new(Resident { weights: self.upload("faces weights", job.weights), bias: self.upload("faces bias", job.bias.unwrap_or(&[])) });+ if self.keep((job.weights.len() as u64 + job.cout as u64) * 4) {+ layers.insert(job.layer, r.clone());}- let d = self.dev;- let usage = wgpu::BufferUsages::STORAGE;- let weights = d.device.create_buffer_init(&wgpu::util::BufferInitDescriptor {- label: Some("faces weights"),- contents: bytemuck::cast_slice(job.weights),- usage,- });- let zero = [0f32; 1];- let bias = d.device.create_buffer_init(&wgpu::util::BufferInitDescriptor {- label: Some("faces bias"),- contents: bytemuck::cast_slice(job.bias.unwrap_or(&zero)),- usage,- });- let r = Arc::new(Resident { weights, bias });- self.resident.fetch_add(bytes, Ordering::Relaxed);- layers.insert(job.layer, r.clone());- Ok(Some(r))+ r}- fn run(&self, job: &ConvJob) -> Result<Option<Vec<f32>>, String> {+ /// A program's constant, kept on the device like the weights.+ fn constant(&self, id: usize, data: &[f32]) -> wgpu::Buffer {+ let mut consts = self.consts.lock().unwrap_or_else(PoisonError::into_inner);+ if let Some(b) = consts.get(&id) {+ return b.clone();+ }+ let b = self.upload("faces constant", data);+ if self.keep(data.len() as u64 * 4) {+ consts.insert(id, b.clone());+ }+ b+ }++ fn largest_binding(&self) -> u64 {+ self.dev.limits.max_storage_buffer_binding_size.min(self.dev.limits.max_buffer_size)+ }++ /// The dispatches of one convolution reading `input` and writing `output`, or `None` when it exceeds the device.+ fn conv_dispatches(&self, job: &ConvJob, input: &wgpu::Buffer, output: &wgpu::Buffer) -> Result<Option<Vec<Dispatch>>, String> {let d = self.dev;let groups = job.groups.max(1);let (m, cg) = (job.cout / groups, job.cin / groups);@@ -264,22 +402,15 @@ impl FaceGpu {let kchunk = k.div_ceil(splits).div_ceil(16) * 16;let splits = k.div_ceil(kchunk.max(1)).max(1);let partial = total.checked_mul(splits).ok_or("a layer too large")?;- let largest = [job.input.len(), job.weights.len(), partial].into_iter().max().unwrap_or(0) as u64 * 4;- if largest > d.limits.max_storage_buffer_binding_size.min(d.limits.max_buffer_size) {- return Ok(None);- }- let flat = |count: usize| {- let blocks = count.div_ceil(256);- let row = blocks.clamp(1, 65_535);- (row, blocks.div_ceil(row), 1, row)- };- let (wx, wy, wz, row) = if direct { flat(total) } else { (n.div_ceil(64), m.div_ceil(64), groups * splits, flat(total).3) };- let (rx, ry, _, _) = flat(total);+ let cin_len = job.cin.checked_mul(job.h).and_then(|v| v.checked_mul(job.w)).ok_or("a layer too large")?;+ let largest = [cin_len, job.weights.len(), partial].into_iter().max().unwrap_or(0) as u64 * 4;let limit = d.limits.max_compute_workgroups_per_dimension as usize;- if [wx, wy, wz, rx, ry].into_iter().any(|v| v > limit) {+ let Some(sum) = flat(total.div_ceil(256), limit) else { return Ok(None) };+ let (wx, wy, wz) = if direct { (sum[0] as usize, sum[1] as usize, 1) } else { (n.div_ceil(64), m.div_ceil(64), groups * splits) };+ if largest > self.largest_binding() || [wx, wy, wz].into_iter().any(|v| v > limit) {return Ok(None);}- let Some(res) = self.resident(job)? else { return Ok(None) };+ let res = self.resident(job);let params = Params {m: u32_of(m)?,n: u32_of(n)?,@@ -296,82 +427,93 @@ impl FaceGpu {pl: u32_of(job.pl)?,bias: u32::from(job.bias.is_some()),total: u32_of(total)?,- row: u32_of(row)?,+ row: sum[2],splits: u32_of(splits)?,kchunk: u32_of(kchunk)?,_a: 0,_b: 0,};- let Some(_work) = crate::exit::enter() else { return Err("the app is closing".into()) };- let scopes = [d.device.push_error_scope(wgpu::ErrorFilter::OutOfMemory), d.device.push_error_scope(wgpu::ErrorFilter::Validation)];- let uniform = d.device.create_buffer_init(&wgpu::util::BufferInitDescriptor {- label: Some("faces params"),- contents: bytemuck::bytes_of(¶ms),- usage: wgpu::BufferUsages::UNIFORM,- });- let input = d.device.create_buffer_init(&wgpu::util::BufferInitDescriptor {- label: Some("faces input"),- contents: bytemuck::cast_slice(job.input),+ let params = bytemuck::bytes_of(¶ms).to_vec();+ let main = |out: &wgpu::Buffer| Dispatch {+ pipeline: if direct { &d.direct } else { &d.tiled },+ params: params.clone(),+ buffers: [input.clone(), res.weights.clone(), res.bias.clone(), out.clone()],+ groups: [wx as u32, wy as u32, wz as u32],+ };+ if splits == 1 {+ return Ok(Some(vec![main(output)]));+ }+ let partials = d.device.create_buffer(&wgpu::BufferDescriptor {+ label: Some("faces partial sums"),+ size: partial as u64 * 4,usage: wgpu::BufferUsages::STORAGE,- });- let bytes = total as u64 * 4;- let output = d.device.create_buffer(&wgpu::BufferDescriptor {- label: Some("faces output"),- size: bytes,- usage: wgpu::BufferUsages::STORAGE | wgpu::BufferUsages::COPY_SRC,- mapped_at_creation: false,- });- let partials = (splits > 1).then(|| {- d.device.create_buffer(&wgpu::BufferDescriptor {- label: Some("faces partial sums"),- size: partial as u64 * 4,- usage: wgpu::BufferUsages::STORAGE,- mapped_at_creation: false,- })- });- let staging = d.device.create_buffer(&wgpu::BufferDescriptor {- label: Some("faces read-back"),- size: bytes,- usage: wgpu::BufferUsages::MAP_READ | wgpu::BufferUsages::COPY_DST,mapped_at_creation: false,});- let bind = d.device.create_bind_group(&wgpu::BindGroupDescriptor {- label: None,- layout: &d.layout,- entries: &[- wgpu::BindGroupEntry { binding: 0, resource: uniform.as_entire_binding() },- wgpu::BindGroupEntry { binding: 1, resource: input.as_entire_binding() },- wgpu::BindGroupEntry { binding: 2, resource: res.weights.as_entire_binding() },- wgpu::BindGroupEntry { binding: 3, resource: res.bias.as_entire_binding() },- wgpu::BindGroupEntry { binding: 4, resource: partials.as_ref().unwrap_or(&output).as_entire_binding() },- ],+ Ok(Some(vec![+ main(&partials),+ Dispatch {+ pipeline: &d.reduce,+ params,+ buffers: [partials, res.weights.clone(), res.bias.clone(), output.clone()],+ groups: [sum[0], sum[1], 1],+ },+ ]))+ }++ /// Record `list` into `enc` as one compute pass, with all their parameters in one uniform buffer.+ fn record(&self, enc: &mut wgpu::CommandEncoder, list: &[Dispatch]) -> Result<(), String> {+ let d = self.dev;+ if list.is_empty() {+ return Ok(());+ }+ let align = (d.limits.min_uniform_buffer_offset_alignment as usize).max(256);+ let mut blob = vec![0u8; list.len() * align];+ for (i, x) in list.iter().enumerate() {+ blob.get_mut(i * align..i * align + x.params.len()).ok_or("parameters too large")?.copy_from_slice(&x.params);+ }+ let uniform = d.device.create_buffer_init(&wgpu::util::BufferInitDescriptor {+ label: Some("faces params"),+ contents: &blob,+ usage: wgpu::BufferUsages::UNIFORM,});- let summing = partials.as_ref().map(|partials| {- d.device.create_bind_group(&wgpu::BindGroupDescriptor {- label: None,- layout: &d.layout,- entries: &[- wgpu::BindGroupEntry { binding: 0, resource: uniform.as_entire_binding() },- wgpu::BindGroupEntry { binding: 1, resource: partials.as_entire_binding() },- wgpu::BindGroupEntry { binding: 2, resource: res.weights.as_entire_binding() },- wgpu::BindGroupEntry { binding: 3, resource: res.bias.as_entire_binding() },- wgpu::BindGroupEntry { binding: 4, resource: output.as_entire_binding() },- ],+ let binds = list+ .iter()+ .enumerate()+ .map(|(i, x)| {+ let [a, b, c, out] = &x.buffers;+ d.device.create_bind_group(&wgpu::BindGroupDescriptor {+ label: None,+ layout: &d.layout,+ entries: &[+ wgpu::BindGroupEntry {+ binding: 0,+ resource: wgpu::BindingResource::Buffer(wgpu::BufferBinding {+ buffer: &uniform,+ offset: (i * align) as u64,+ size: std::num::NonZeroU64::new(x.params.len() as u64),+ }),+ },+ wgpu::BindGroupEntry { binding: 1, resource: a.as_entire_binding() },+ wgpu::BindGroupEntry { binding: 2, resource: b.as_entire_binding() },+ wgpu::BindGroupEntry { binding: 3, resource: c.as_entire_binding() },+ wgpu::BindGroupEntry { binding: 4, resource: out.as_entire_binding() },+ ],+ })})- });- let mut enc = d.device.create_command_encoder(&wgpu::CommandEncoderDescriptor { label: None });- {- let mut pass = enc.begin_compute_pass(&wgpu::ComputePassDescriptor { label: Some("faces conv"), timestamp_writes: None });- pass.set_pipeline(if direct { &d.direct } else { &d.tiled });- pass.set_bind_group(0, &bind, &[]);- pass.dispatch_workgroups(u32_of(wx)?, u32_of(wy)?, u32_of(wz)?);- if let Some(summing) = &summing {- pass.set_pipeline(&d.reduce);- pass.set_bind_group(0, summing, &[]);- pass.dispatch_workgroups(u32_of(rx)?, u32_of(ry)?, 1);- }+ .collect::<Vec<_>>();+ let mut pass = enc.begin_compute_pass(&wgpu::ComputePassDescriptor { label: Some("faces"), timestamp_writes: None });+ for (x, bind) in list.iter().zip(&binds) {+ pass.set_pipeline(x.pipeline);+ pass.set_bind_group(0, bind, &[]);+ let [gx, gy, gz] = x.groups;+ pass.dispatch_workgroups(gx, gy, gz);}- enc.copy_buffer_to_buffer(&output, 0, &staging, 0, bytes);+ Ok(())+ }++ /// Submit `enc`, which ends by copying into `staging` (`bytes` long), wait for it and read `staging` back.+ fn finish(&self, enc: wgpu::CommandEncoder, staging: &wgpu::Buffer, scopes: [wgpu::ErrorScopeGuard; 2]) -> Result<Vec<f32>, String> {+ let d = self.dev;let index = d.queue.submit([enc.finish()]);let slice = staging.slice(..);let (tx, rx) = std::sync::mpsc::channel();@@ -389,7 +531,7 @@ impl FaceGpu {return Err(format!("the GPU reported an error: {e}"));}if let Err(e) = waited {- mark_broken(&format!("the GPU did not finish a convolution: {e}"));+ mark_broken(&format!("the GPU did not finish a face model: {e}"));return Err(format!("the GPU did not finish: {e}"));}match rx.recv_timeout(Duration::from_secs(10)) {@@ -402,10 +544,246 @@ impl FaceGpu {data.as_chunks::<4>().0.iter().map(|b| f32::from_le_bytes(*b)).collect::<Vec<f32>>()};staging.unmap();- Ok(Some(out))+ Ok(out)+ }++ fn scopes(&self) -> [wgpu::ErrorScopeGuard; 2] {+ [self.dev.device.push_error_scope(wgpu::ErrorFilter::OutOfMemory), self.dev.device.push_error_scope(wgpu::ErrorFilter::Validation)]+ }++ fn staging(&self, bytes: u64) -> wgpu::Buffer {+ self.dev.device.create_buffer(&wgpu::BufferDescriptor {+ label: Some("faces read-back"),+ size: bytes.max(4),+ usage: wgpu::BufferUsages::MAP_READ | wgpu::BufferUsages::COPY_DST,+ mapped_at_creation: false,+ })+ }++ fn scratch(&self, bytes: u64) -> wgpu::Buffer {+ self.dev.device.create_buffer(&wgpu::BufferDescriptor {+ label: Some("faces values"),+ size: bytes.max(4),+ usage: wgpu::BufferUsages::STORAGE | wgpu::BufferUsages::COPY_SRC,+ mapped_at_creation: false,+ })+ }++ /// One convolution, uploaded and read back on its own.+ fn run(&self, job: &ConvJob) -> Result<Option<Vec<f32>>, String> {+ let total = job.cout.checked_mul(job.oh).and_then(|v| v.checked_mul(job.ow)).ok_or("a layer too large")?;+ let bytes = total as u64 * 4;+ if bytes > self.largest_binding() || job.input.len() as u64 * 4 > self.largest_binding() {+ return Ok(None);+ }+ let Some(_work) = crate::exit::enter() else { return Err("the app is closing".into()) };+ let _turn = Turn::take();+ let d = self.dev;+ let scopes = self.scopes();+ let input = d.device.create_buffer_init(&wgpu::util::BufferInitDescriptor {+ label: Some("faces input"),+ contents: bytemuck::cast_slice(job.input),+ usage: wgpu::BufferUsages::STORAGE,+ });+ let output = self.scratch(bytes);+ let Some(list) = self.conv_dispatches(job, &input, &output)? else {+ // leave the scopes balanced+ let [oom, validation] = scopes;+ let _ = (pollster::block_on(validation.pop()), pollster::block_on(oom.pop()));+ return Ok(None);+ };+ let staging = self.staging(bytes);+ let mut enc = d.device.create_command_encoder(&wgpu::CommandEncoderDescriptor { label: None });+ self.record(&mut enc, &list)?;+ enc.copy_buffer_to_buffer(&output, 0, &staging, 0, bytes);+ self.finish(enc, &staging, scopes).map(Some)+ }++ /// The dispatch of one non-convolution step.+ fn op_dispatch(&self, op: &Op, out_len: usize, args: [wgpu::Buffer; 3], out: &wgpu::Buffer) -> Result<Option<Dispatch>, String> {+ let d = self.dev;+ let limit = d.limits.max_compute_workgroups_per_dimension as usize;+ let u = u32_of;+ let n = u(out_len)?;+ let per_value = |pipeline: &'static wgpu::ComputePipeline, rest: &[u32]| -> Option<Dispatch> {+ let groups = flat(out_len.div_ceil(256), limit)?;+ let mut all = vec![n, groups[2]];+ all.extend_from_slice(rest);+ Some(Dispatch {+ pipeline,+ params: params24(&all),+ buffers: [args[0].clone(), args[1].clone(), args[2].clone(), out.clone()],+ groups: [groups[0], groups[1], 1],+ })+ };+ let per_group = |pipeline: &'static wgpu::ComputePipeline, rest: &[u32]| -> Option<Dispatch> {+ let groups = flat(out_len, limit)?;+ let mut all = vec![n, groups[2]];+ all.extend_from_slice(rest);+ Some(Dispatch {+ pipeline,+ params: params24(&all),+ buffers: [args[0].clone(), args[1].clone(), args[2].clone(), out.clone()],+ groups: [groups[0], groups[1], 1],+ })+ };+ let four = |v: &[usize; 4]| -> Result<[u32; 4], String> { Ok([u(v[0])?, u(v[1])?, u(v[2])?, u(v[3])?]) };+ let k = &d.ops;+ Ok(match op {+ Op::Conv(_) => None,+ Op::Unary(f) => per_value(&k.unary, &[u32::from(*f == Unary::Sigmoid)]),+ Op::ChannelAffine { channels, plane } => per_value(&k.affine, &[u(*plane)?, u(*channels)?]),+ Op::ChannelPrelu { channels, plane } => per_value(&k.prelu, &[u(*plane)?, u(*channels)?]),+ Op::Binary { op, dims, a, b } => {+ let kind = match op {+ Binary::Add => 0,+ Binary::Sub => 1,+ Binary::Mul => 2,+ };+ let mut rest = vec![kind];+ rest.extend(four(dims)?);+ rest.extend(four(a)?);+ rest.extend(four(b)?);+ per_value(&k.binary, &rest)+ }+ Op::Transpose { dims, strides } => {+ let mut rest = vec![0];+ rest.extend(four(dims)?);+ rest.extend(four(strides)?);+ per_value(&k.transpose, &rest)+ }+ Op::MaxPool { h, w, kh, kw, sh, sw, oh, ow, .. } => {+ per_value(&k.maxpool, &[u(*h)?, u(*w)?, u(*kh)?, u(*kw)?, u(*sh)?, u(*sw)?, u(*oh)?, u(*ow)?])+ }+ Op::Resize { h, w, sh, sw, .. } => per_value(&k.resize, &[u(*h)?, u(*w)?, u(*sh)?, u(*sw)?]),+ Op::GlobalAverage { plane, .. } => per_group(&k.gavg, &[u(*plane)?]),+ Op::Gemv { k: kk, n: nn, trans_b, alpha } => per_group(&k.gemv, &[u(*kk)?, u(*nn)?, u32::from(*trans_b), alpha.to_bits()]),+ })+ }++ /// A whole program for each input, every value staying on the device; one submission and one read-back.+ fn program(&self, prog: &Program, inputs: &[&[f32]]) -> Result<Option<Vec<Vec<Vec<f32>>>>, String> {+ let d = self.dev;+ let largest = self.largest_binding();+ if prog.lens.iter().any(|l| *l as u64 * 4 > largest) || inputs.iter().any(|i| Some(&i.len()) != prog.lens.get(INPUT)) {+ return Ok(None);+ }+ let out_bytes: u64 = prog.outputs.iter().map(|(_, v, _)| prog.lens.get(*v).copied().unwrap_or(0) as u64 * 4).sum();+ let total_bytes = out_bytes.checked_mul(inputs.len() as u64).ok_or("too many faces at once")?;+ if total_bytes > d.limits.max_buffer_size {+ return Ok(None);+ }+ let Some(_work) = crate::exit::enter() else { return Err("the app is closing".into()) };+ let _turn = Turn::take();+ let scopes = self.scopes();+ let staging = self.staging(total_bytes);+ let mut enc = d.device.create_command_encoder(&wgpu::CommandEncoderDescriptor { label: Some("faces program") });+ // buffers nothing reads any more, reused for later values (best fit)+ let mut pool: Vec<(u64, wgpu::Buffer)> = Vec::new();+ let take = |pool: &mut Vec<(u64, wgpu::Buffer)>, bytes: u64| -> wgpu::Buffer {+ let best = pool+ .iter()+ .enumerate()+ .filter(|(_, (size, _))| *size >= bytes && *size <= bytes.saturating_mul(2))+ .min_by_key(|(_, (size, _))| *size)+ .map(|(i, _)| i);+ match best {+ Some(i) => pool.swap_remove(i).1,+ None => self.scratch(bytes),+ }+ };+ let mut at = 0u64;+ let mut fits = true;+ for input in inputs {+ let mut values: Vec<Option<(u64, wgpu::Buffer)>> = vec![None; prog.lens.len()];+ // a fresh buffer: queued writes land before any of the encoder's work, so a pooled one could still be in use+ let first = d.device.create_buffer_init(&wgpu::util::BufferInitDescriptor {+ label: Some("faces input"),+ contents: bytemuck::cast_slice(input),+ usage: wgpu::BufferUsages::STORAGE | wgpu::BufferUsages::COPY_SRC,+ });+ if let Some(v) = values.get_mut(INPUT) {+ *v = Some((input.len() as u64 * 4, first));+ }+ let mut list = Vec::new();+ for step in &prog.steps {+ let len = prog.lens.get(step.out).copied().ok_or("a step writes a value the program does not have")?;+ let out = take(&mut pool, len as u64 * 4);+ let mut args = Vec::with_capacity(3);+ for a in &step.args {+ args.push(match a {+ Arg::Value(v) => {+ values.get(*v).and_then(|v| v.as_ref()).map(|(_, b)| b.clone()).ok_or("a step reads a value before it is made")?+ }+ Arg::Const { id, data } => self.constant(*id, data),+ });+ }+ let dispatches = match &step.op {+ Op::Conv(job) => {+ let input = args.first().ok_or("a convolution without input")?;+ self.conv_dispatches(job, input, &out)?+ }+ op => {+ let arg = |i: usize| args.get(i).cloned().unwrap_or_else(|| d.dummy.clone());+ self.op_dispatch(op, len, [arg(0), arg(1), arg(2)], &out)?.map(|x| vec![x])+ }+ };+ let Some(dispatches) = dispatches else {+ fits = false;+ break;+ };+ list.extend(dispatches);+ if let Some(v) = values.get_mut(step.out) {+ *v = Some((len as u64 * 4, out));+ }+ for f in &step.free {+ if let Some(b) = values.get_mut(*f).and_then(Option::take) {+ pool.push(b);+ }+ }+ }+ if !fits {+ break;+ }+ self.record(&mut enc, &list)?;+ for (_, v, _) in &prog.outputs {+ let (bytes, buffer) = values.get(*v).and_then(|v| v.as_ref()).ok_or("an output was never made")?;+ let len = prog.lens.get(*v).copied().unwrap_or(0) as u64 * 4;+ enc.copy_buffer_to_buffer(buffer, 0, &staging, at, len.min(*bytes));+ at += len;+ }+ pool.extend(values.into_iter().flatten());+ }+ if !fits {+ let [oom, validation] = scopes;+ let _ = (pollster::block_on(validation.pop()), pollster::block_on(oom.pop()));+ return Ok(None);+ }+ let flat_out = self.finish(enc, &staging, scopes)?;+ let mut rest = flat_out.as_slice();+ let mut per_input = Vec::with_capacity(inputs.len());+ for _ in inputs {+ let mut outs = Vec::with_capacity(prog.outputs.len());+ for (_, v, _) in &prog.outputs {+ let len = prog.lens.get(*v).copied().unwrap_or(0);+ let (this, next) = rest.split_at_checked(len).ok_or("the read-back is short")?;+ outs.push(this.to_vec());+ rest = next;+ }+ per_input.push(outs);+ }+ Ok(Some(per_input))}}+/// Run `f`, turning a panic into a broken device.+fn guarded<T>(f: impl FnOnce() -> Result<T, String>) -> Result<T, String> {+ std::panic::catch_unwind(std::panic::AssertUnwindSafe(f)).unwrap_or_else(|_| {+ mark_broken("a face model on the GPU panicked");+ Err("the GPU runner gave up".into())+ })+}+impl Accelerator for FaceGpu {fn device(&self) -> String {format!("{} ({:?})", self.dev.info.name, self.dev.info.backend)@@ -418,13 +796,23 @@ impl Accelerator for FaceGpu {if crate::switched_off().is_some() || job.macs() < MIN_MACS {return Ok(None);}- std::panic::catch_unwind(std::panic::AssertUnwindSafe(|| self.run(job))).unwrap_or_else(|_| {- mark_broken("a face convolution panicked");- Err("the GPU runner gave up".into())- })+ guarded(|| self.run(job))+ }++ fn run_program(&self, program: &Program, inputs: &[&[f32]]) -> Result<Option<Vec<Vec<Vec<f32>>>>, String> {+ if let Some(why) = broken_reason() {+ return Err(format!("the GPU is not used any more: {why}"));+ }+ if crate::switched_off().is_some() || PER_CONV.load(Ordering::Relaxed) {+ return Ok(None);+ }+ guarded(|| self.program(program, inputs))}}+/// Tests and benchmarks only: run convolutions one by one (the first step's way) instead of whole programs.+static PER_CONV: AtomicBool = AtomicBool::new(false);+#[cfg(test)]mod tests {use super::*;@@ -513,6 +901,28 @@ mod tests {assert!(gpu.resident_bytes() > 0, "the weights stay on the device");}+ #[test]+ fn every_operator_on_the_gpu_gives_what_the_cpu_gives() {+ let Some(gpu) = gpu() else { return };+ let net = lightcraft_faces::net::Net::new(lightcraft_faces::synthetic::every_operator_graph()).unwrap();+ let inputs: Vec<lightcraft_faces::net::Tensor> = (0..3)+ .map(|s| {+ lightcraft_faces::net::Tensor::new(vec![1, 4, 8, 8], (0..256).map(|i| (((i * 31 + s * 7) % 37) as f32 - 18.0) * 0.05).collect())+ .unwrap()+ })+ .collect();+ let cpu: Vec<_> = inputs.iter().map(|x| net.run(x.clone()).unwrap()).collect();+ let program = net.program(&[1, 4, 8, 8]).unwrap().unwrap();+ let slices: Vec<&[f32]> = inputs.iter().map(|t| t.data.as_slice()).collect();+ let got = gpu.program(&program, &slices).unwrap().unwrap();+ for (want, got) in cpu.iter().zip(&got) {+ for ((name, want), got) in want.iter().zip(got) {+ let worst = want.data.iter().zip(got).fold(0f32, |m, (a, b)| m.max((a - b).abs()));+ assert!(worst < 1e-4, "{name}: off by {worst}");+ }+ }+ }+/// The real models on the GPU agree with the CPU interpreter: `LC_FACE_MODELS=<folder>` holding/// `face_detection_yunet_2023mar.onnx`, `face_recognition_sface_2021dec.onnx` and (optionally) `glintr100.onnx`.#[test]@@ -555,7 +965,8 @@ mod tests {}}- /// Timings, CPU against GPU: `LC_FACE_MODELS=<folder> cargo test --release -p lightcraft-gpu faces::tests::timings -- --ignored --nocapture`.+ /// Timings, CPU against the GPU one convolution at a time and as whole programs (one face, and a batch of 16):+ /// `LC_FACE_MODELS=<folder> cargo test --release -p lightcraft-gpu faces::tests::timings -- --ignored --nocapture`.#[test]#[ignore = "a benchmark: needs LC_FACE_MODELS and a GPU"]fn timings() {@@ -573,21 +984,35 @@ mod tests {let (w, h) = (1024, 768);let rgb: Vec<u8> = (0..w * h * 3).map(|i| (i * 7 % 251) as u8).collect();let opts = lightcraft_faces::yunet::Options::default();- let cpu = time(5, &mut || drop(det.detect(&rgb, w, h, &opts).unwrap()));+ let cpu = time(10, &mut || drop(det.detect(&rgb, w, h, &opts).unwrap()));+ PER_CONV.store(true, Ordering::Relaxed);+ det.accelerate(accelerator().unwrap(), false).unwrap();+ let per_conv = time(10, &mut || drop(det.detect(&rgb, w, h, &opts).unwrap()));+ PER_CONV.store(false, Ordering::Relaxed);det.accelerate(accelerator().unwrap(), false).unwrap();- let fast = time(5, &mut || drop(det.detect(&rgb, w, h, &opts).unwrap()));- eprintln!("YuNet 640: CPU {cpu:.1} ms, GPU {fast:.1} ms");+ let whole = time(10, &mut || drop(det.detect(&rgb, w, h, &opts).unwrap()));+ assert!(det.net().accelerator().is_some());+ eprintln!("YuNet 640: CPU {cpu:.1} ms, GPU per convolution {per_conv:.1} ms, GPU whole network {whole:.1} ms");for (file, sha) in [("face_recognition_sface_2021dec.onnx", lightcraft_faces::known::SFACE_SHA256),("glintr100.onnx", lightcraft_faces::known::AURAFACE_SHA256),] {let e = lightcraft_faces::runtime::Embedder::load(&dir.join(file), &lightcraft_faces::known::lookup(sha).unwrap()).unwrap();let (ew, eh) = e.input_size();- let face: Vec<u8> = (0..ew * eh * 3).map(|i| (i * 13 % 251) as u8).collect();- let cpu = time(3, &mut || drop(e.embed(&face).unwrap()));+ let faces: Vec<Vec<u8>> = (0..16).map(|s| (0..ew * eh * 3).map(|i| ((i + s * 101) * 13 % 251) as u8).collect()).collect();+ let refs: Vec<&[u8]> = faces.iter().map(Vec::as_slice).collect();+ let cpu = time(5, &mut || drop(e.embed(&faces[0]).unwrap()));+ PER_CONV.store(true, Ordering::Relaxed);+ e.accelerate(accelerator().unwrap(), false).unwrap();+ let per_conv = time(5, &mut || drop(e.embed(&faces[0]).unwrap()));+ PER_CONV.store(false, Ordering::Relaxed);e.accelerate(accelerator().unwrap(), false).unwrap();- let fast = time(3, &mut || drop(e.embed(&face).unwrap()));- eprintln!("{file}: CPU {cpu:.1} ms, GPU {fast:.1} ms");+ let whole = time(5, &mut || drop(e.embed(&faces[0]).unwrap()));+ let batch = time(3, &mut || drop(e.embed_batch(&refs))) / 16.0;+ assert!(e.net().accelerator().is_some());+ eprintln!(+ "{file}: CPU {cpu:.1} ms, GPU per convolution {per_conv:.1} ms, GPU whole network {whole:.1} ms, batch of 16 {batch:.1} ms a face"+ );}}}diff --git a/crates/gpu/src/wgsl/faces_ops.wgsl b/crates/gpu/src/wgsl/faces_ops.wgslnew file mode 100644index 000000000..aa7eba242--- /dev/null+++ b/crates/gpu/src/wgsl/faces_ops.wgsl@@ -0,0 +1,188 @@+// The face networks' other operators (lightcraft-faces `net::program`), so activations stay on the device between layers.+// Same bindings as faces_conv.wgsl: the parameters, three read-only operands and the output. Element kernels run one+// thread per output value; `gavg` and `gemv` one workgroup per output value.+//+// Parameters are 24 numbers: q(0) = values (or workgroups) to make, q(1) = workgroups along x, then per kernel.++struct P {+ v: array<vec4<u32>, 6>,+}++@group(0) @binding(0) var<uniform> p: P;+@group(0) @binding(1) var<storage, read> a: array<f32>;+@group(0) @binding(2) var<storage, read> b: array<f32>;+@group(0) @binding(3) var<storage, read> c: array<f32>;+@group(0) @binding(4) var<storage, read_write> dst: array<f32>;++fn q(i: u32) -> u32 {+ return p.v[i / 4u][i % 4u];+}++fn index(wid: vec3<u32>, lid: u32) -> u32 {+ return (wid.y * q(1) + wid.x) * 256u + lid;+}++// q(2): 0 relu, 1 sigmoid+@compute @workgroup_size(256)+fn unary(@builtin(workgroup_id) wid: vec3<u32>, @builtin(local_invocation_index) lid: u32) {+ let i = index(wid, lid);+ if (i >= q(0)) { return; }+ let x = a[i];+ if (q(2) == 0u) {+ dst[i] = max(x, 0.0);+ } else {+ dst[i] = 1.0 / (1.0 + exp(-x));+ }+}++// q(2) plane, q(3) channels: x * b[c] + c[c]+@compute @workgroup_size(256)+fn affine(@builtin(workgroup_id) wid: vec3<u32>, @builtin(local_invocation_index) lid: u32) {+ let i = index(wid, lid);+ if (i >= q(0)) { return; }+ let ch = (i / q(2)) % q(3);+ dst[i] = a[i] * b[ch] + c[ch];+}++// q(2) plane, q(3) channels: x < 0 ? x * b[c] : x+@compute @workgroup_size(256)+fn prelu(@builtin(workgroup_id) wid: vec3<u32>, @builtin(local_invocation_index) lid: u32) {+ let i = index(wid, lid);+ if (i >= q(0)) { return; }+ let x = a[i];+ let ch = (i / q(2)) % q(3);+ dst[i] = select(x, x * b[ch], x < 0.0);+}++// q(2): 0 add, 1 sub, 2 mul; q(3..7) output dims; q(7..11) a's strides; q(11..15) b's strides+@compute @workgroup_size(256)+fn binary(@builtin(workgroup_id) wid: vec3<u32>, @builtin(local_invocation_index) lid: u32) {+ let i = index(wid, lid);+ if (i >= q(0)) { return; }+ var rest = i;+ var ia = 0u;+ var ib = 0u;+ for (var d = 3; d >= 0; d = d - 1) {+ let du = u32(d);+ let size = q(3u + du);+ let at = rest % size;+ rest = rest / size;+ ia = ia + at * q(7u + du);+ ib = ib + at * q(11u + du);+ }+ let x = a[ia];+ let y = b[ib];+ if (q(2) == 0u) {+ dst[i] = x + y;+ } else if (q(2) == 1u) {+ dst[i] = x - y;+ } else {+ dst[i] = x * y;+ }+}++// q(3..7) output dims; q(7..11) the input's strides for each output axis+@compute @workgroup_size(256)+fn transpose(@builtin(workgroup_id) wid: vec3<u32>, @builtin(local_invocation_index) lid: u32) {+ let i = index(wid, lid);+ if (i >= q(0)) { return; }+ var rest = i;+ var at_in = 0u;+ for (var d = 3; d >= 0; d = d - 1) {+ let du = u32(d);+ let size = q(3u + du);+ at_in = at_in + (rest % size) * q(7u + du);+ rest = rest / size;+ }+ dst[i] = a[at_in];+}++// q(2) h, q(3) w, q(4) kh, q(5) kw, q(6) sh, q(7) sw, q(8) oh, q(9) ow+@compute @workgroup_size(256)+fn maxpool(@builtin(workgroup_id) wid: vec3<u32>, @builtin(local_invocation_index) lid: u32) {+ let i = index(wid, lid);+ if (i >= q(0)) { return; }+ let h = q(2);+ let w = q(3);+ let oh = q(8);+ let ow = q(9);+ let ox = i % ow;+ let oy = (i / ow) % oh;+ let ch = i / (ow * oh);+ var m = -3.402823e38;+ for (var ky = 0u; ky < q(4); ky = ky + 1u) {+ for (var kx = 0u; kx < q(5); kx = kx + 1u) {+ let y = oy * q(6) + ky;+ let x = ox * q(7) + kx;+ if (y < h && x < w) {+ m = max(m, a[(ch * h + y) * w + x]);+ }+ }+ }+ dst[i] = m;+}++// q(2) h, q(3) w, q(4) sh, q(5) sw: nearest upscale by whole numbers+@compute @workgroup_size(256)+fn resize(@builtin(workgroup_id) wid: vec3<u32>, @builtin(local_invocation_index) lid: u32) {+ let i = index(wid, lid);+ if (i >= q(0)) { return; }+ let h = q(2);+ let w = q(3);+ let ow = w * q(5);+ let oh = h * q(4);+ let ox = i % ow;+ let oy = (i / ow) % oh;+ let ch = i / (ow * oh);+ dst[i] = a[(ch * h + oy / q(4)) * w + ox / q(5)];+}++var<workgroup> partial: array<f32, 256>;++fn reduce(lid: u32) {+ for (var s = 128u; s > 0u; s = s >> 1u) {+ workgroupBarrier();+ if (lid < s) {+ partial[lid] = partial[lid] + partial[lid + s];+ }+ }+ workgroupBarrier();+}++// one workgroup per channel; q(2) plane+@compute @workgroup_size(256)+fn gavg(@builtin(workgroup_id) wid: vec3<u32>, @builtin(local_invocation_index) lid: u32) {+ let ch = wid.y * q(1) + wid.x;+ let plane = q(2);+ var sum = 0.0;+ if (ch < q(0)) {+ for (var j = lid; j < plane; j = j + 256u) {+ sum = sum + a[ch * plane + j];+ }+ }+ partial[lid] = sum;+ reduce(lid);+ if (lid == 0u && ch < q(0)) {+ dst[ch] = partial[0] / f32(plane);+ }+}++// one workgroup per output; q(2) k, q(3) n, q(4) 1 when B is [n][k], q(5) alpha's bits+@compute @workgroup_size(256)+fn gemv(@builtin(workgroup_id) wid: vec3<u32>, @builtin(local_invocation_index) lid: u32) {+ let o = wid.y * q(1) + wid.x;+ let k = q(2);+ let n = q(3);+ var sum = 0.0;+ if (o < q(0)) {+ for (var j = lid; j < k; j = j + 256u) {+ let wi = select(j * n + o, o * k + j, q(4) == 1u);+ sum = sum + a[j] * b[wi];+ }+ }+ partial[lid] = sum;+ reduce(lid);+ if (lid == 0u && o < q(0)) {+ dst[o] = bitcast<f32>(q(5)) * partial[0] + c[o];+ }+}diff --git a/crates/ui-egui/locales/de.json b/crates/ui-egui/locales/de.jsonindex 25c411ce0..bfaf57fdf 100644--- a/crates/ui-egui/locales/de.json+++ b/crates/ui-egui/locales/de.json@@ -2118,5 +2118,13 @@"Adding to Apple Photos": "Wird zu Apple Fotos hinzugefügt","No graphics processor found: LightCraft renders on the CPU, which is much slower. Details in Help → System Info.": "Kein Grafikprozessor gefunden: LightCraft rendert auf der CPU, was viel langsamer ist. Details unter Hilfe → Systeminformationen.","On NixOS, start it from `nix develop`.": "Unter NixOS starten Sie es aus `nix develop`.",- "Files with the same content as another file in this import: only the first is imported.": "Dateien mit demselben Inhalt wie eine andere Datei dieses Imports: Nur die erste wird importiert."+ "Files with the same content as another file in this import: only the first is imported.": "Dateien mit demselben Inhalt wie eine andere Datei dieses Imports: Nur die erste wird importiert.",+ "Run face models on the CPU only": "Gesichtsmodelle nur auf der CPU ausführen",+ "Detection": "Erkennung",+ "Recognition": "Wiedererkennung",+ "Not loaded yet": "Noch nicht geladen",+ "Checking the GPU…": "GPU wird geprüft …",+ "CPU": "CPU",+ "A face model uses the GPU only when a check shows it gives the same results faster.": "Ein Gesichtsmodell nutzt die GPU nur, wenn eine Prüfung zeigt, dass sie dieselben Ergebnisse schneller liefert.",+ "Checking the GPU for faces": "GPU für Gesichter wird geprüft"}diff --git a/crates/ui-egui/locales/es.json b/crates/ui-egui/locales/es.jsonindex ba3e1e331..d153c911e 100644--- a/crates/ui-egui/locales/es.json+++ b/crates/ui-egui/locales/es.json@@ -1995,5 +1995,13 @@"Adding to Apple Photos": "Añadiendo a Fotos de Apple","No graphics processor found: LightCraft renders on the CPU, which is much slower. Details in Help → System Info.": "No se encontró ningún procesador gráfico: LightCraft renderiza en la CPU, lo que es mucho más lento. Detalles en Ayuda → Información del sistema.","On NixOS, start it from `nix develop`.": "En NixOS, inícielo desde `nix develop`.",- "Files with the same content as another file in this import: only the first is imported.": "Archivos con el mismo contenido que otro archivo de esta importación: solo se importa el primero."+ "Files with the same content as another file in this import: only the first is imported.": "Archivos con el mismo contenido que otro archivo de esta importación: solo se importa el primero.",+ "Run face models on the CPU only": "Ejecutar los modelos de caras solo en la CPU",+ "Detection": "Detección",+ "Recognition": "Reconocimiento",+ "Not loaded yet": "Aún no cargado",+ "Checking the GPU…": "Comprobando la GPU…",+ "CPU": "CPU",+ "A face model uses the GPU only when a check shows it gives the same results faster.": "Un modelo de caras usa la GPU solo cuando una comprobación muestra que da los mismos resultados más rápido.",+ "Checking the GPU for faces": "Comprobando la GPU para las caras"}diff --git a/crates/ui-egui/locales/fr.json b/crates/ui-egui/locales/fr.jsonindex b0becb154..3827ce3ac 100644--- a/crates/ui-egui/locales/fr.json+++ b/crates/ui-egui/locales/fr.json@@ -1957,5 +1957,13 @@"Adding to Apple Photos": "Ajout à Apple Photos","No graphics processor found: LightCraft renders on the CPU, which is much slower. Details in Help → System Info.": "Aucun processeur graphique trouvé : LightCraft effectue le rendu sur le processeur, ce qui est beaucoup plus lent. Détails dans Aide → Informations système.","On NixOS, start it from `nix develop`.": "Sous NixOS, lancez-le depuis `nix develop`.",- "Files with the same content as another file in this import: only the first is imported.": "Fichiers au contenu identique à un autre fichier de cette importation : seul le premier est importé."+ "Files with the same content as another file in this import: only the first is imported.": "Fichiers au contenu identique à un autre fichier de cette importation : seul le premier est importé.",+ "Run face models on the CPU only": "Exécuter les modèles de visages sur le processeur uniquement",+ "Detection": "Détection",+ "Recognition": "Reconnaissance",+ "Not loaded yet": "Pas encore chargé",+ "Checking the GPU…": "Vérification du GPU…",+ "CPU": "Processeur",+ "A face model uses the GPU only when a check shows it gives the same results faster.": "Un modèle de visages n’utilise le GPU que si une vérification montre qu’il donne les mêmes résultats plus vite.",+ "Checking the GPU for faces": "Vérification du GPU pour les visages"}diff --git a/crates/ui-egui/locales/ja.json b/crates/ui-egui/locales/ja.jsonindex 7a282e676..d9d83d109 100644--- a/crates/ui-egui/locales/ja.json+++ b/crates/ui-egui/locales/ja.json@@ -1961,5 +1961,13 @@"Adding to Apple Photos": "Appleの写真に追加中","No graphics processor found: LightCraft renders on the CPU, which is much slower. Details in Help → System Info.": "グラフィックプロセッサが見つかりません。LightCraftはCPUで描画するため、かなり遅くなります。詳細は「ヘルプ → システム情報」を参照してください。","On NixOS, start it from `nix develop`.": "NixOSでは `nix develop` から起動してください。",- "Files with the same content as another file in this import: only the first is imported.": "この読み込み内の別のファイルと内容が同じファイルです。最初のファイルだけが読み込まれます。"+ "Files with the same content as another file in this import: only the first is imported.": "この読み込み内の別のファイルと内容が同じファイルです。最初のファイルだけが読み込まれます。",+ "Run face models on the CPU only": "顔モデルを CPU だけで実行",+ "Detection": "検出",+ "Recognition": "認識",+ "Not loaded yet": "未読み込み",+ "Checking the GPU…": "GPU を確認中…",+ "CPU": "CPU",+ "A face model uses the GPU only when a check shows it gives the same results faster.": "顔モデルが GPU を使うのは、同じ結果をより速く出せると確認できた場合だけです。",+ "Checking the GPU for faces": "顔用の GPU を確認中"}diff --git a/crates/ui-egui/locales/pt-br.json b/crates/ui-egui/locales/pt-br.jsonindex 2e21c727c..f626e6433 100644--- a/crates/ui-egui/locales/pt-br.json+++ b/crates/ui-egui/locales/pt-br.json@@ -1998,5 +1998,13 @@"Recognition models": "Modelos de reconhecimento","No graphics processor found: LightCraft renders on the CPU, which is much slower. Details in Help → System Info.": "Nenhum processador gráfico encontrado: o LightCraft renderiza na CPU, o que é muito mais lento. Detalhes em Ajuda → Informações do sistema.","On NixOS, start it from `nix develop`.": "No NixOS, inicie-o a partir de `nix develop`.",- "Files with the same content as another file in this import: only the first is imported.": "Arquivos com o mesmo conteúdo de outro arquivo desta importação: apenas o primeiro é importado."+ "Files with the same content as another file in this import: only the first is imported.": "Arquivos com o mesmo conteúdo de outro arquivo desta importação: apenas o primeiro é importado.",+ "Run face models on the CPU only": "Executar os modelos de rostos só na CPU",+ "Detection": "Detecção",+ "Recognition": "Reconhecimento",+ "Not loaded yet": "Ainda não carregado",+ "Checking the GPU…": "Verificando a GPU…",+ "CPU": "CPU",+ "A face model uses the GPU only when a check shows it gives the same results faster.": "Um modelo de rostos só usa a GPU quando uma verificação mostra que ela dá os mesmos resultados mais rápido.",+ "Checking the GPU for faces": "Verificando a GPU para rostos"}diff --git a/crates/ui-egui/locales/ru.json b/crates/ui-egui/locales/ru.jsonindex c840218a0..f08c3cb7d 100644--- a/crates/ui-egui/locales/ru.json+++ b/crates/ui-egui/locales/ru.json@@ -2151,5 +2151,13 @@"Adding to Apple Photos": "Добавление в Apple Фото","No graphics processor found: LightCraft renders on the CPU, which is much slower. Details in Help → System Info.": "Графический процессор не найден: LightCraft выполняет рендеринг на CPU, а это намного медленнее. Подробности: Справка → Сведения о системе.","On NixOS, start it from `nix develop`.": "В NixOS запускайте его из `nix develop`.",- "Files with the same content as another file in this import: only the first is imported.": "Файлы с тем же содержимым, что и другой файл этого импорта: импортируется только первый."+ "Files with the same content as another file in this import: only the first is imported.": "Файлы с тем же содержимым, что и другой файл этого импорта: импортируется только первый.",+ "Run face models on the CPU only": "Запускать модели лиц только на ЦП",+ "Detection": "Обнаружение",+ "Recognition": "Распознавание",+ "Not loaded yet": "Ещё не загружена",+ "Checking the GPU…": "Проверка ГП…",+ "CPU": "ЦП",+ "A face model uses the GPU only when a check shows it gives the same results faster.": "Модель лиц использует ГП, только если проверка показала, что он даёт те же результаты быстрее.",+ "Checking the GPU for faces": "Проверка ГП для лиц"}diff --git a/crates/ui-egui/locales/uk.json b/crates/ui-egui/locales/uk.jsonindex 85a838098..a0783aa5f 100644--- a/crates/ui-egui/locales/uk.json+++ b/crates/ui-egui/locales/uk.json@@ -2151,5 +2151,13 @@"Adding to Apple Photos": "Додавання до Apple Фото","No graphics processor found: LightCraft renders on the CPU, which is much slower. Details in Help → System Info.": "Графічний процесор не знайдено: LightCraft виконує рендеринг на ЦП, а це набагато повільніше. Докладніше: Довідка → Інформація про систему.","On NixOS, start it from `nix develop`.": "У NixOS запускайте його з `nix develop`.",- "Files with the same content as another file in this import: only the first is imported.": "Файли з тим самим вмістом, що й інший файл цього імпорту: імпортується лише перший."+ "Files with the same content as another file in this import: only the first is imported.": "Файли з тим самим вмістом, що й інший файл цього імпорту: імпортується лише перший.",+ "Run face models on the CPU only": "Запускати моделі облич лише на ЦП",+ "Detection": "Виявлення",+ "Recognition": "Розпізнавання",+ "Not loaded yet": "Ще не завантажено",+ "Checking the GPU…": "Перевірка ГП…",+ "CPU": "ЦП",+ "A face model uses the GPU only when a check shows it gives the same results faster.": "Модель облич використовує ГП, лише якщо перевірка показала, що він дає ті самі результати швидше.",+ "Checking the GPU for faces": "Перевірка ГП для облич"}diff --git a/crates/ui-egui/locales/zh-hans.json b/crates/ui-egui/locales/zh-hans.jsonindex ad379f1ef..c2ef31084 100644--- a/crates/ui-egui/locales/zh-hans.json+++ b/crates/ui-egui/locales/zh-hans.json@@ -1997,5 +1997,13 @@"Adding to Apple Photos": "正在添加到 Apple 照片","No graphics processor found: LightCraft renders on the CPU, which is much slower. Details in Help → System Info.": "未找到图形处理器:LightCraft 改用 CPU 渲染,速度会慢很多。详情见“帮助 → 系统信息”。","On NixOS, start it from `nix develop`.": "在 NixOS 上,请从 `nix develop` 启动。",- "Files with the same content as another file in this import: only the first is imported.": "与本次导入中另一个文件内容相同的文件:只导入第一个。"+ "Files with the same content as another file in this import: only the first is imported.": "与本次导入中另一个文件内容相同的文件:只导入第一个。",+ "Run face models on the CPU only": "仅在 CPU 上运行人脸模型",+ "Detection": "检测",+ "Recognition": "识别",+ "Not loaded yet": "尚未加载",+ "Checking the GPU…": "正在检查 GPU…",+ "CPU": "CPU",+ "A face model uses the GPU only when a check shows it gives the same results faster.": "只有检查表明 GPU 能更快给出相同结果时,人脸模型才会使用 GPU。",+ "Checking the GPU for faces": "正在检查用于人脸的 GPU"}diff --git a/crates/ui-egui/locales/zh-hant.json b/crates/ui-egui/locales/zh-hant.jsonindex 647d153d0..b830bf434 100644--- a/crates/ui-egui/locales/zh-hant.json+++ b/crates/ui-egui/locales/zh-hant.json@@ -1959,5 +1959,13 @@"Adding to Apple Photos": "正在加入 Apple 照片","No graphics processor found: LightCraft renders on the CPU, which is much slower. Details in Help → System Info.": "找不到圖形處理器:LightCraft 改用 CPU 算繪,速度會慢很多。詳情請見「說明 → 系統資訊」。","On NixOS, start it from `nix develop`.": "在 NixOS 上,請從 `nix develop` 啟動。",- "Files with the same content as another file in this import: only the first is imported.": "與本次匯入中另一個檔案內容相同的檔案:只匯入第一個。"+ "Files with the same content as another file in this import: only the first is imported.": "與本次匯入中另一個檔案內容相同的檔案:只匯入第一個。",+ "Run face models on the CPU only": "僅在 CPU 上執行臉孔模型",+ "Detection": "偵測",+ "Recognition": "辨識",+ "Not loaded yet": "尚未載入",+ "Checking the GPU…": "正在檢查 GPU…",+ "CPU": "CPU",+ "A face model uses the GPU only when a check shows it gives the same results faster.": "只有在檢查顯示 GPU 能更快得出相同結果時,臉孔模型才會使用 GPU。",+ "Checking the GPU for faces": "正在檢查用於臉孔的 GPU"}diff --git a/crates/ui-egui/src/panels/faces.rs b/crates/ui-egui/src/panels/faces.rsindex 2b3fe0329..27933f53d 100644--- a/crates/ui-egui/src/panels/faces.rs+++ b/crates/ui-egui/src/panels/faces.rs@@ -157,6 +157,42 @@ pub fn setup_banner(app: &mut LightcraftApp, ui: &mut egui::Ui) {});}+/// Where the face models run (the GPU only once a check shows it gives the CPU's results faster), and the switch that+/// keeps them on the CPU.+fn gpu_section(app: &mut LccApp, ui: &mut egui::Ui, t: &Tokens, gpu: &Value) {+ heading(ui, t, "Processor");+ let mut cpu = gpu["cpuOnly"].as_bool().unwrap_or(false);+ let r = ui.checkbox(&mut cpu, crate::i18n::tr("Run face models on the CPU only"));+ register(ui.ctx(), "check:settings.facesCpuOnly", r.rect);+ if r.changed() {+ match app.run("faces.gpu", json!({"cpuOnly": cpu})) {+ Ok(_) => app.caches.faces_epoch += 1,+ Err(e) => app.toast(ui.ctx(), e),+ }+ }+ for (what, label) in [("detection", "Detection"), ("recognition", "Recognition")] {+ let st = &gpu[what];+ let (text, why) = if st.is_null() {+ (crate::i18n::tr("Not loaded yet").to_string(), None)+ } else if st["checking"] == true {+ (crate::i18n::tr("Checking the GPU…").to_string(), None)+ } else if let Some(device) = st["device"].as_str() {+ (device.to_string(), None)+ } else {+ (crate::i18n::tr("CPU").to_string(), st["why"].as_str())+ };+ ui.horizontal(|ui| {+ ui.label(RichText::new(crate::i18n::tr(label)).color(t.text_dim));+ let r = ui.label(text);+ register(ui.ctx(), format!("faces:device:{what}"), r.rect);+ if let Some(why) = why {+ r.on_hover_text(sentence(why));+ }+ });+ }+ hint(ui, t, "A face model uses the GPU only when a check shows it gives the same results faster.");+}+/// A download-style size: decimal megabytes, as file managers and model pages show them.fn mb(bytes: Option<u64>) -> String {match bytes {@@ -231,6 +267,7 @@ pub fn settings_tab(app: &mut LightcraftApp, ui: &mut egui::Ui, t: &Tokens) {let all: Vec<Value> = list["models"].as_array().cloned().unwrap_or_default();let downloads =app.session.execute("faces.models.downloads", &json!({})).map(|v| v["downloads"].as_array().cloned().unwrap_or_default()).unwrap_or_default();+ gpu_section(app, ui, t, &list["gpu"]);heading(ui, t, "Face detector");for m in all.iter().filter(|m| m["role"] == "detector") {model_row(app, ui, t, m, downloads.iter().find(|d| d["id"] == m["id"]), runtime);diff --git a/docs/faces.md b/docs/faces.mdindex f4f4a4757..c767ce913 100644--- a/docs/faces.md+++ b/docs/faces.md@@ -145,31 +145,40 @@ names alone do not guarantee the same speed, and those additional model files we### On the GPU-On a machine with a usable GPU, the convolutions of the detector and the recogniser (most of their time) run on it-through wgpu (`crates/gpu/src/faces.rs`, `wgsl/faces_conv.wgsl`: an implicit matrix product that gathers the input-straight into workgroup memory, cuts K into slices when a layer's map is small, and a direct kernel for depthwise-layers). Everything else in the graph stays on the CPU interpreter, which remains the reference. After a model loads,-a background check runs it once on the CPU and twice through the GPU on a fixed test input: the GPU is used only when+On a machine with a usable GPU, the whole detector and recogniser run on it through wgpu (`crates/gpu/src/faces.rs`).+The interpreter lowers the graph to a flat program for the input's shape (`crates/faces/src/net/program.rs`): each+convolution (`wgsl/faces_conv.wgsl`: an implicit matrix product that gathers the input straight into workgroup memory,+cuts K into slices when a map is small, and a direct kernel for depthwise layers) and every other operator (activations,+batch norm, PRelu, broadcasting add/sub/mul, max pooling, nearest upscaling, transposes, global pooling and the fully+connected layer, `wgsl/faces_ops.wgsl`) is a kernel, every value stays on the device between layers in reused+buffers, and a run is one upload, one submission and one read-back. Shape-only operators make no step. A model that+needs something the program cannot express runs its convolutions one by one on the GPU instead. Several faces of one+photo go together (`Embedder::embed_batch`, up to 16 per submission).++The CPU interpreter remains the reference. After a model loads, a background check (a row in the activity stack, "Checking+the GPU for faces") runs it once on the CPU and twice through the GPU on a fixed test input: the GPU is used only whenevery output is within 1e-3 of the CPU's (relative to the output's largest value) **and** it was faster; until then,-and whenever the GPU is switched off (`LIGHTCRAFT_GPU=0`, `LIGHTCRAFT_GPU_BACKEND=off`, the GPU preference), missing,-wrong or slower, the model runs on the CPU. The first GPU error drops it for that model. The weights stay on a device of-the scan's own (separate from the editor's) while the model is loaded, at most 768 MB per model. The log says which-device each model runs on.+and whenever the GPU is switched off (`LIGHTCRAFT_GPU=0`, `LIGHTCRAFT_GPU_BACKEND=off`, the GPU preference, or **Settings ▸+Faces ▸ Run face models on the CPU only**), missing, wrong or slower, the model runs on the CPU. The first GPU error,+or a result of the wrong size or with values that are not finite, drops it for that model. Settings ▸ Faces shows+which device each model runs on (`faces.gpu`). The weights stay on a device of the scan's own (separate from the+editor's) while the model is loaded, at most 768 MB per model.-Measured on Linux with a Radeon (Vulkan) and a 32-thread CPU, one call at a time, machine under heavy load from other-jobs (so the absolute numbers are noisy; the ratios held across three runs):+The background scan's pace also limits the GPU: paused or light, one model run at a time and the GPU busy at most a+quarter of the time (the editor renders on the same GPU); normal, two at once and three quarters; full, as fast as it+goes.-| Model | CPU | GPU |-| --- | ---: | ---: |-| YuNet (640², one photo) | 33–34 ms | 35–39 ms |-| SFace 2021dec | 25–28 ms | 22–27 ms |-| AuraFace v1 | 204–208 ms | 81–103 ms |+Measured on Linux with a Radeon (Vulkan) and a 32-thread CPU, one call at a time, machine under very heavy load from+other jobs (load average 100–250, so the absolute numbers are noisy):++| Model | CPU | GPU, one layer at a time | GPU, whole network | Batch of 16, per face |+| --- | ---: | ---: | ---: | ---: |+| YuNet (640², one photo) | 58–60 ms | 85–89 ms | 21 ms | – |+| SFace 2021dec | 40–42 ms | 47–51 ms | 9.6 ms | 8.1 ms |+| AuraFace v1 | 295–308 ms | 165–172 ms | 70–73 ms | 71 ms |-Each layer still goes to the GPU and back: for YuNet and SFace that round trip costs about what the layer saves, so on-most machines they stay on the CPU by the speed check, while AuraFace runs 2–2.5× faster. Keeping activations on the GPU-between layers (the remaining operators as kernels) and batching faces per dispatch are the next steps.`LC_FACE_MODELS=<folder> cargo test --release -p lightcraft-gpu faces::tests -- --include-ignored --nocapture` checks-agreement on the real models and prints these timings.+agreement on the real models and every operator, and prints these timings.### Agreement and regression coverage