Fix #651: a dual-GPU Mac falls back to its other GPU

Metal reports no PCI ids, so both GPUs of a dual-GPU Mac had the adapter key
Metal:0000:0000. When the integrated GPU failed at start-up, the restart left
out both and the app quit, although the discrete GPU works. An adapter without
ids is now keyed by its name, so the restart leaves out only the one that
failed.

The failure itself was wgpu's indirect-call validation compute shader, which
OCLP-patched Metal on an Iris Pro can't compile. Nothing in VectorCraft or
egui draws indirectly, so that validation is off (WGPU_VALIDATION_INDIRECT_CALL=1
turns it back on).

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

apps/vectorcraft/src/gpu.rs

@@ -22,8 +22,9 @@ use std::sync::{Arc, OnceLock};
 use eframe::egui_wgpu::NativeAdapterSelectorMethod;
 use eframe::wgpu::{self, Backend, DeviceType, PowerPreference};
 
-/// The adapters a restart leaves out, those the window failed on, as comma-separated
-/// `backend:vendor:device` keys (`Vulkan:1002:164e`). Set by the app itself when it starts again.
+/// The adapters a restart leaves out, those the window failed on, as comma-separated [`key`]s
+/// (`Vulkan:1002:164e`, `Metal:Intel Iris Pro Graphics`). Set by the app itself when it starts
+/// again.
 pub const SKIP_ENV: &str = "VECTORCRAFT_GPU_SKIP";
 
 /// wgpu's variable for choosing an adapter by (part of) its name, any case.
@@ -60,7 +61,7 @@ pub fn power_preference(pref: Option<&str>, env: Option<PowerPreference>) -> Pow
 /// One adapter, as [`adapter_order`] sees it.
 #[derive(Clone, Debug)]
 pub struct Candidate {
-    /// Identifies the adapter across a restart: `backend:vendor:device` ([`key`]).
+    /// Identifies the adapter across a restart ([`key`]).
     pub key: String,
     pub name: String,
     pub backend: Backend,
@@ -73,7 +74,7 @@ impl Candidate {
     fn new(adapter: &wgpu::Adapter, surface: Option<&wgpu::Surface<'_>>) -> Self {
         let info = adapter.get_info();
         Self {
-            key: key(info.backend, info.vendor, info.device),
+            key: key(info.backend, info.vendor, info.device, &info.name),
             presents: surface.is_none_or(|s| adapter.is_surface_supported(s)),
             name: info.name,
             backend: info.backend,
@@ -83,8 +84,15 @@ impl Candidate {
 }
 
 /// `backend:vendor:device`, e.g. `Vulkan:1002:164e` (PCI ids, as `MESA_VK_DEVICE_SELECT` takes them).
-fn key(backend: Backend, vendor: u32, device: u32) -> String {
-    format!("{backend:?}:{vendor:04x}:{device:04x}")
+/// Metal reports no ids (`0000:0000` for every GPU), so an adapter without them is `backend:name`
+/// (`Metal:Intel Iris Pro Graphics`): otherwise leaving out the one that failed would leave out
+/// every GPU of a dual-GPU Mac (#651). Commas, which separate [`SKIP_ENV`]'s keys, become spaces.
+fn key(backend: Backend, vendor: u32, device: u32, name: &str) -> String {
+    if vendor == 0 && device == 0 {
+        format!("{backend:?}:{}", name.replace(',', " ").trim())
+    } else {
+        format!("{backend:?}:{vendor:04x}:{device:04x}")
+    }
 }
 
 /// The order to try `candidates` in (their indices): only those that present to the window and
@@ -307,6 +315,27 @@ mod tests {
 
     /// A restart leaves out the adapter the window failed on, so Power Saving on the machine of
     /// #502 ends up on the NVIDIA GPU, and once every adapter failed there is nothing to try.
+    /// A dual-GPU Mac reports both GPUs under Metal with no PCI ids (#651): their keys still
+    /// differ, so when the integrated GPU fails, the restart renders on the discrete one.
+    #[test]
+    fn gpus_without_pci_ids_are_told_apart_by_name() {
+        let mac = |name: &str, device_type| Candidate {
+            key: key(Backend::Metal, 0, 0, name),
+            name: name.into(),
+            backend: Backend::Metal,
+            device_type,
+            presents: true,
+        };
+        let c = vec![mac("NVIDIA GeForce GT 750M", DeviceType::DiscreteGpu), mac("Intel Iris Pro Graphics", DeviceType::IntegratedGpu)];
+        assert_eq!((c[0].key.as_str(), c[1].key.as_str()), ("Metal:NVIDIA GeForce GT 750M", "Metal:Intel Iris Pro Graphics"));
+        let first = adapter_order(&c, PowerPreference::LowPower, None, &[]);
+        assert_eq!(names(&c, &first)[0], "Intel Iris Pro Graphics");
+        let skip = parse_skip(&c[first[0]].key);
+        assert_eq!(names(&c, &adapter_order(&c, PowerPreference::LowPower, None, &skip)), ["NVIDIA GeForce GT 750M"]);
+        // A comma in a name can't split the list.
+        assert_eq!(parse_skip(&key(Backend::Gl, 0, 0, "Mesa, llvmpipe")), ["Gl:Mesa  llvmpipe"]);
+    }
+
     #[test]
     fn a_restart_tries_the_next_adapter() {
         let c = issue_502();
@@ -433,8 +462,8 @@ mod tests {
     fn skipped_adapters_parse_from_a_comma_separated_list() {
         assert_eq!(parse_skip("Vulkan:1002:164e, Gl:10de:2f04,,"), ["Vulkan:1002:164e", "Gl:10de:2f04"]);
         assert!(parse_skip("").is_empty());
-        assert_eq!(key(Backend::Vulkan, 0x1002, 0x164e), "Vulkan:1002:164e");
-        assert_eq!(key(Backend::Gl, 0x10de, 0x2f04), "Gl:10de:2f04");
+        assert_eq!(key(Backend::Vulkan, 0x1002, 0x164e, "AMD Radeon"), "Vulkan:1002:164e");
+        assert_eq!(key(Backend::Gl, 0x10de, 0x2f04, "NVIDIA"), "Gl:10de:2f04");
     }
 
     #[test]

apps/vectorcraft/src/main.rs

@@ -423,6 +423,11 @@ fn main() -> std::process::ExitCode {
         // Only adapters that can show the window, in the order `gpu` gives (#306, #502).
         if let eframe::egui_wgpu::WgpuSetup::CreateNew(create) = &mut options.wgpu_options.wgpu_setup {
             create.native_adapter_selector = Some(gpu::selector(power, startup.clone()));
+            // Nothing draws or dispatches indirectly, so wgpu's check of indirect arguments only
+            // costs a compute shader at start-up, one some drivers can't compile (OCLP-patched
+            // Metal on an Iris Pro, #651). `WGPU_VALIDATION_INDIRECT_CALL=1` turns it back on.
+            create.instance_descriptor.flags =
+                (eframe::wgpu::InstanceFlags::from_build_config() - eframe::wgpu::InstanceFlags::VALIDATION_INDIRECT_CALL).with_env();
         }
         options
     };

docs/development.md

@@ -43,11 +43,12 @@ The canvas is rasterized on the CPU (`vectorcraft-render`, vello_cpu); the GPU (
   - **Automatic on Windows and macOS** is power saving: the system shows frames from any GPU, and presenting frames rendered on a discrete GPU through the integrated one made the window flicker on some hybrid laptops (#306).
   - **Automatic on Linux and the BSDs** keeps the system's order: Mesa's Vulkan device-select layer puts the GPU the desktop runs on first (the integrated one on hybrid laptops; `DRI_PRIME` and `MESA_VK_DEVICE_SELECT` steer it). A Wayland compositor may not accept frames from another GPU: on a desktop whose compositor ran on an NVIDIA GPU, rendering on the Ryzen's integrated GPU made KWin end the window's connection ("importing the supplied dmabufs failed") and 0.5.0 crashed at startup (#502).
   - 0.5.0 saved `powerSaving`, its default, for everyone; it reads as `automatic`.
-- **If the window fails on its adapter while starting up** (an error from wgpu, or a panic inside wgpu or egui-wgpu during the first frames), the app starts again without that adapter and tries the next one, telling which in the status bar; the run that failed keeps its log as `vectorcraft.1.log`. On Unix the new app replaces the process (an AppImage stays mounted); on Windows it starts beside it. The adapters left out travel in `VECTORCRAFT_GPU_SKIP` (`backend:vendor:device`, such as `Vulkan:1002:164e`), one more on each restart, so the restarts end. When no adapter is left, the app logs why and exits with an error instead of panicking.
+- **If the window fails on its adapter while starting up** (an error from wgpu, or a panic inside wgpu or egui-wgpu during the first frames), the app starts again without that adapter and tries the next one, telling which in the status bar; the run that failed keeps its log as `vectorcraft.1.log`. On Unix the new app replaces the process (an AppImage stays mounted); on Windows it starts beside it. Adapters are told apart by backend and PCI ids, or by name where the backend reports no ids (Metal lists every GPU of a dual-GPU Mac as `0000:0000`, #651). The adapters left out travel in `VECTORCRAFT_GPU_SKIP` (`backend:vendor:device`, such as `Vulkan:1002:164e`), one more on each restart, so the restarts end. When no adapter is left, the app logs why and exits with an error instead of panicking.
 - **Environment variables** for when the choice still goes wrong, all read at startup and stronger than the preference:
   - `WGPU_POWER_PREF`: `low`, `high` or `none` (the system's order).
   - `WGPU_ADAPTER_NAME`: part of an adapter's name, any case (`WGPU_ADAPTER_NAME=nvidia`, `=radv`, `=intel`); the names are in the log.
   - `WGPU_BACKEND`: the backends to use, such as `vulkan`, `dx12`, `metal` or `gl`.
+  - `WGPU_VALIDATION_INDIRECT_CALL=1`: turns wgpu's check of indirect draw arguments back on. VectorCraft draws nothing indirectly and turns it off, since the compute shader it needs fails to compile on some drivers (#651).
   - On Linux with Mesa, `MESA_VK_DEVICE_SELECT=10de:2f04!` (vendor:device as `vulkaninfo --summary` or `lspci -nn` show it) leaves only that GPU to Vulkan, and `DRI_PRIME=1` picks the other GPU.
 - Help › About and the control channel's `ui.inspect` (`graphicsAdapter`) show the adapter in use, and the app logs it at startup.