Files
makeCAD/crates/gpu/src/concurrent.rs
T
yhamil04 3574956d29 Graphics: the window no longer waits for Direct3D 12 shader compiles (#4860) (#7288)
* claim: #4860

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Graphics: the window no longer waits for Direct3D 12 shader compiles (#4860)

On Direct3D 12, wgpu compiles every pipeline's shaders with FXC when the
pipeline is made, and nothing is kept between launches, so the window
waited seconds for the viewport.

- The GPU tessellation shader, most of that time, compiles on a thread of
  its own. Bodies are built on the CPU until it is ready, then drawn from
  their plans; where it never compiles they stay on the CPU, as before.
- The viewport's pipelines build in parallel, one job per fragment shader
  (solidcraft_gpu::concurrent). A job that errs, panics or cannot start
  runs again on the calling thread, so errors reach the caller's scopes.
  Threads only on Vulkan, Metal and D3D12 with FXC.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-27 12:41:55 +00:00

296 lines
12 KiB
Rust

//! Pipelines built on several threads at once (#4860).
//!
//! Direct3D 12 compiles a pipeline's shaders when the pipeline is made (wgpu writes the WGSL out
//! as HLSL and compiles that with FXC), and nothing compiled is kept between launches, so the
//! viewport's pipelines, built one after another, held the window back for seconds. Work
//! handed to [`Jobs::spawn`] runs on a thread of its own, and [`Job::join`] collects it.
//!
//! - wgpu's error scopes belong to the thread that pushed them, so a job runs under scopes of its
//! own. A job that raised an error, or whose thread panicked or could not start, runs again on
//! the calling thread when joined: its errors reach the caller's scopes and the device's handler
//! as they did when every pipeline was built there.
//! - Threads only on Vulkan, Metal and Direct3D 12 with FXC, whose devices make pipelines from any
//! thread. wgpu hands every thread one DXC compiler, which is not documented as safe to share,
//! and OpenGL works through one context. There, and in the browser, which has no threads, each
//! job runs when it is joined, one after another, as before.
#[cfg(target_arch = "wasm32")]
use std::marker::PhantomData;
/// The name of the threads jobs run on (in crash reports and debuggers).
#[cfg(not(target_arch = "wasm32"))]
const WORKER: &str = "solidcraft pipelines";
/// Jobs on `device`, collected before [`scope`] returns.
pub struct Jobs<'scope, 'env> {
device: &'env wgpu::Device,
/// The threads' scope; `None` where jobs run on the calling thread.
#[cfg(not(target_arch = "wasm32"))]
threads: Option<&'scope std::thread::Scope<'scope, 'env>>,
#[cfg(target_arch = "wasm32")]
threads: PhantomData<&'scope ()>,
}
/// Work handed to [`Jobs::spawn`].
pub struct Job<'scope, T, F> {
work: F,
#[cfg(not(target_arch = "wasm32"))]
worker: Option<std::thread::ScopedJoinHandle<'scope, Option<T>>>,
#[cfg(target_arch = "wasm32")]
worker: PhantomData<(&'scope (), T)>,
}
/// Runs `f` with a [`Jobs`] on `device`; every job is joined before this returns.
pub fn scope<'env, R>(device: &'env wgpu::Device, f: impl for<'scope> FnOnce(&Jobs<'scope, 'env>) -> R) -> R {
#[cfg(not(target_arch = "wasm32"))]
{
let threaded = threads_allowed(device);
std::thread::scope(|s| f(&Jobs { device, threads: threaded.then_some(s) }))
}
#[cfg(target_arch = "wasm32")]
{
f(&Jobs { device, threads: PhantomData })
}
}
/// `device` makes pipelines on several threads at once (see the module); the shader of the GPU
/// plan evaluation compiles on a thread of its own by the same rule (`view::TessLater`).
#[cfg(not(target_arch = "wasm32"))]
pub(crate) fn threads_allowed(device: &wgpu::Device) -> bool {
threaded(device.adapter_info().backend, dx12_compiles_with_fxc)
}
/// Backends whose devices make pipelines on several threads at once; D3D12 when `fxc` says it
/// compiles with FXC.
#[cfg(not(target_arch = "wasm32"))]
fn threaded(backend: wgpu::Backend, fxc: impl FnOnce() -> bool) -> bool {
match backend {
wgpu::Backend::Vulkan | wgpu::Backend::Metal => true,
wgpu::Backend::Dx12 => fxc(),
_ => false,
}
}
/// D3D12 compiles with FXC: `WGPU_DX12_COMPILER` says `fxc`, or leaves wgpu's choice (`auto`),
/// which takes DXC when it can load a `dxcompiler.dll` (#6756) and FXC otherwise. Looked up
/// once: neither changes while the app runs.
#[cfg(not(target_arch = "wasm32"))]
fn dx12_compiles_with_fxc() -> bool {
static FXC: std::sync::OnceLock<bool> = std::sync::OnceLock::new();
*FXC.get_or_init(|| fxc_chosen(std::env::var("WGPU_DX12_COMPILER").ok().as_deref(), dxcompiler_found))
}
/// [`dx12_compiles_with_fxc`] for the variable's value `setting`, read as wgpu reads it;
/// `dxc_found` looks for the DLL.
#[cfg(not(target_arch = "wasm32"))]
fn fxc_chosen(setting: Option<&str>, dxc_found: impl FnOnce() -> bool) -> bool {
match setting.map(str::to_lowercase).as_deref() {
Some("fxc") => true,
Some("dxc" | "dynamicdxc" | "staticdxc") => false,
// `auto`, unset, or a value wgpu does not know (it keeps `auto` then).
_ => !dxc_found(),
}
}
/// A `dxcompiler.dll` where Windows looks for one: beside the program, in the system folders, the
/// working folder or `PATH`. Conservative: one found there may be one Windows would not load.
#[cfg(not(target_arch = "wasm32"))]
fn dxcompiler_found() -> bool {
let program = std::env::current_exe().ok().and_then(|p| p.parent().map(std::path::Path::to_path_buf));
let windows = std::env::var_os("SystemRoot").map(std::path::PathBuf::from);
let system = windows.as_ref().map(|w| w.join("System32"));
let path = std::env::var_os("PATH");
program
.into_iter()
.chain(system)
.chain(windows)
.chain(std::env::current_dir().ok())
.chain(path.iter().flat_map(std::env::split_paths))
.any(|folder| folder.join("dxcompiler.dll").is_file())
}
impl<'scope, 'env> Jobs<'scope, 'env> {
/// Start `work` on a thread of its own (or later, on the calling thread: see the module).
/// `work` may run twice, so it only builds what it returns.
#[cfg(not(target_arch = "wasm32"))]
pub fn spawn<T, F>(&self, work: F) -> Job<'scope, T, F>
where
T: Send + 'scope,
F: Fn() -> T + Copy + Send + 'scope,
{
let device = self.device;
let worker = self.threads.and_then(|s| std::thread::Builder::new().name(WORKER.into()).spawn_scoped(s, move || checked(device, work)).ok());
Job { work, worker }
}
/// Run `work` when it is joined: the browser has no threads.
#[cfg(target_arch = "wasm32")]
pub fn spawn<T, F: Fn() -> T>(&self, work: F) -> Job<'scope, T, F> {
let _ = self.device;
Job { work, worker: PhantomData }
}
}
impl<T, F: Fn() -> T> Job<'_, T, F> {
/// What the job built: on its thread, or now on this one (see the module).
#[cfg_attr(target_arch = "wasm32", allow(unused_mut))] // Only native jobs have a thread to take.
pub fn join(mut self) -> T {
#[cfg(not(target_arch = "wasm32"))]
if let Some(built) = self.worker.take().and_then(|w| w.join().ok().flatten()) {
return built;
}
(self.work)()
}
}
/// A job dropped unjoined still waits for its thread: the scope would otherwise panic at its end
/// if that thread had panicked.
#[cfg(not(target_arch = "wasm32"))]
impl<T, F> Drop for Job<'_, T, F> {
fn drop(&mut self) {
if let Some(worker) = self.worker.take() {
let _ = worker.join();
}
}
}
/// `work`'s result, or `None` when it raised a GPU error on this thread (logged: the job runs
/// again on the calling thread, where the error may not recur).
#[cfg(not(target_arch = "wasm32"))]
fn checked<T>(device: &wgpu::Device, work: impl Fn() -> T) -> Option<T> {
let scopes = crate::error_scopes(device);
let built = work();
match crate::caught(scopes) {
Some(e) => {
log::warn!("GPU pipelines built again on the calling thread: {e}");
None
}
None => Some(built),
}
}
#[cfg(all(test, not(target_arch = "wasm32")))]
mod tests {
use std::sync::atomic::{AtomicUsize, Ordering};
use super::*;
#[test]
fn only_backends_that_make_pipelines_on_any_thread_get_threads() {
for fxc in [true, false] {
assert!(threaded(wgpu::Backend::Vulkan, || fxc));
assert!(threaded(wgpu::Backend::Metal, || fxc));
assert!(!threaded(wgpu::Backend::Gl, || fxc));
assert!(!threaded(wgpu::Backend::BrowserWebGpu, || fxc));
assert!(!threaded(wgpu::Backend::Noop, || fxc));
}
assert!(threaded(wgpu::Backend::Dx12, || true), "D3D12 with FXC");
assert!(!threaded(wgpu::Backend::Dx12, || false), "D3D12 with DXC");
}
#[test]
fn d3d12_compiles_with_fxc_unless_dxc_is_chosen_or_found() {
for setting in [Some("fxc"), Some("FXC")] {
assert!(fxc_chosen(setting, || true), "{setting:?} keeps FXC even with a DLL around");
}
for setting in [Some("dxc"), Some("DynamicDxc"), Some("staticdxc")] {
assert!(!fxc_chosen(setting, || false), "{setting:?} chooses DXC");
}
// wgpu does not trim the value: " fxc " is a value it does not know, so `auto`.
for setting in [None, Some("auto"), Some(""), Some("unknown"), Some(" fxc ")] {
assert!(fxc_chosen(setting, || false), "{setting:?} without a DLL: FXC");
assert!(!fxc_chosen(setting, || true), "{setting:?} with a DLL: DXC");
}
}
/// A buffer wgpu refuses: mapped for reading and for writing at once.
fn refused_buffer(device: &wgpu::Device) -> wgpu::Buffer {
device.create_buffer(&wgpu::BufferDescriptor {
label: Some("refused"),
size: 16,
usage: wgpu::BufferUsages::MAP_READ | wgpu::BufferUsages::MAP_WRITE,
mapped_at_creation: false,
})
}
#[test]
fn jobs_hand_back_what_they_built() {
let Ok((_, device, _)) = crate::open_headless() else {
eprintln!("skipped: no GPU adapter");
return;
};
let runs = AtomicUsize::new(0);
let built = scope(&device, |jobs| {
let a = jobs.spawn(|| {
runs.fetch_add(1, Ordering::Relaxed);
1
});
let b = jobs.spawn(|| {
runs.fetch_add(1, Ordering::Relaxed);
2
});
(a.join(), b.join())
});
assert_eq!(built, (1, 2));
assert_eq!(runs.load(Ordering::Relaxed), 2, "a job that went well runs once");
}
#[test]
fn an_error_on_a_job_reaches_the_callers_error_scope() {
let Ok((_, device, _)) = crate::open_headless() else {
eprintln!("skipped: no GPU adapter");
return;
};
let runs = AtomicUsize::new(0);
let caller = device.push_error_scope(wgpu::ErrorFilter::Validation);
let device_ref = &device;
let runs_ref = &runs;
scope(&device, |jobs| {
let job = jobs.spawn(move || {
runs_ref.fetch_add(1, Ordering::Relaxed);
refused_buffer(device_ref)
});
drop(job.join());
});
let error = crate::popped(caller.pop());
assert!(error.is_some(), "the caller's scope catches the job's error");
let expected = if threaded(device.adapter_info().backend, dx12_compiles_with_fxc) { 2 } else { 1 };
assert_eq!(runs.load(Ordering::Relaxed), expected, "the job ran again on the calling thread");
}
#[test]
fn a_job_dropped_unjoined_whose_thread_panicked_does_not_panic_the_scope() {
let Ok((_, device, _)) = crate::open_headless() else {
eprintln!("skipped: no GPU adapter");
return;
};
let finished = scope(&device, |jobs| {
drop(jobs.spawn(|| {
if std::thread::current().name() == Some(WORKER) {
panic!("a job's thread panicked");
}
}));
true
});
assert!(finished);
}
#[test]
fn a_job_whose_thread_panicked_runs_again_on_the_calling_thread() {
let Ok((_, device, _)) = crate::open_headless() else {
eprintln!("skipped: no GPU adapter");
return;
};
let built = scope(&device, |jobs| {
let job = jobs.spawn(|| {
if std::thread::current().name() == Some(WORKER) {
panic!("a job's thread panicked");
}
7
});
job.join()
});
assert_eq!(built, 7);
}
}