cargo, wgpu: test the CI feature set without rex-gpu, and report why a submission was refused

.cargo/config.toml

@@ -2,7 +2,7 @@
 # these and list no features. Needs cargo-nextest.
 [alias]
 # The features CI adds to the default build; the one place they are listed.
-ci-nextest = "nextest run --workspace --features jitv2,rex-gpu,rex-jit"
+ci-nextest = "nextest run --workspace --features jitv2,rex-jit,gputrace"
 ci-test = "ci-nextest --profile ci"
 # The default features alone: the only run of the real-cache path that tcache
 # (implied by jitv2) compiles out.

src/wgpu/device.rs

@@ -506,11 +506,12 @@ impl Gpu {
 
     /// Block until `submission` has finished (or everything, with None).
     pub fn wait(&self, submission: Option<wgpu::SubmissionIndex>) -> Result<(), String> {
-        self.device
-            .poll(wgpu::PollType::Wait { submission_index: submission, timeout: Some(Duration::from_secs(30)) })
-            .map(|_| ())
-            .map_err(|e| format!("wait for GPU: {e}"))?;
-        self.check()
+        let waited = self.device.poll(wgpu::PollType::Wait { submission_index: submission, timeout: Some(Duration::from_secs(30)) });
+        // wgpu cannot wait for a submission it refused (a pipeline that
+        // failed to build, a lost device); the error that refused it is
+        // the one to report.
+        self.check()?;
+        waited.map(|_| ()).map_err(|e| format!("wait for GPU: {e}"))
     }
 
     /// Let callbacks run without blocking.

src/wgpu/tests/mod.rs

@@ -378,3 +378,32 @@ fn empty_flush_is_a_no_op() {
     assert!(board.batch.flush(&board.pipelines, &[&board.vram_group]).is_none());
     assert!(board.plane().iter().all(|&w| w == 0));
 }
+
+/// A pipeline that fails to build makes the submission that uses it fail,
+/// and wgpu cannot wait for a failed submission. The wait reports the build
+/// error, not the missing submission.
+#[test]
+fn a_refused_submission_reports_why() {
+    let Some(gpu) = gpu() else { return };
+    let module = gpu.device.create_shader_module(wgpu::ShaderModuleDescriptor {
+        label: Some("refused"),
+        source: wgpu::ShaderSource::Wgsl("@compute @workgroup_size(1) fn main() { let x: u32 = 1.5; }".into()),
+    });
+    let pipeline = gpu.device.create_compute_pipeline(&wgpu::ComputePipelineDescriptor {
+        label: Some("refused"),
+        layout: None,
+        module: &module,
+        entry_point: Some("main"),
+        compilation_options: Default::default(),
+        cache: None,
+    });
+    let mut encoder = gpu.encoder("refused");
+    {
+        let mut pass = encoder.begin_compute_pass(&wgpu::ComputePassDescriptor::default());
+        pass.set_pipeline(&pipeline);
+        pass.dispatch_workgroups(1, 1, 1);
+    }
+    let index = gpu.submit(encoder);
+    let err = gpu.wait(Some(index)).unwrap_err();
+    assert!(err.starts_with("wgpu error"), "{err}");
+}