mirror of
https://github.com/storytold/vectorcraft.git
synced 2026-11-02 06:07:06 -06:00
Windows: after a graphics driver crash, the next start leaves that adapter out
When a graphics driver crashed the process while the window started, as Intel's Direct3D 12 driver (igd12umd64.dll) did on an Intel UHD Graphics 605 (#2243), the next start picked the same adapter again. The restart without a failing adapter and the first-frame watch run inside the process, so the crash ended them along with it. Each start now writes graphics-start.json in the log folder before the graphics start, adds the adapter once one is picked, and clears both when the first frame reaches the screen or the app ends without crashing. A start that finds an earlier start still recorded there leaves that adapter out, as a restart does, logs it and shows a note in the status bar. When the crash came while the adapters were being listed (wgpu's Direct3D 12 backend creates a device on each GPU to list it), the next starts list OpenGL's adapters alone, except on macOS. The adapters stay left out until the app version or the Graphics Processor preference changes, a start shows its first frame on one, or no adapter is left. WGPU_BACKEND and WGPU_ADAPTER_NAME choose regardless of the record. Docs: docs/development.md, docs/target-app-parity.md.
This commit is contained in:
+2
-1
@@ -50,7 +50,7 @@ How the web shell (`apps/vectorcraft-web/src/web.rs`) differs from desktop:
|
||||
The canvas is rasterized on the CPU (`vectorcraft-render`, vello_cpu); the GPU (wgpu, through eframe) only composites the canvas texture and draws the UI, so any GPU that can show the window will do. The desktop app picks the window's adapter itself (`apps/vectorcraft/src/gpu.rs`, eframe's `native_adapter_selector`) and logs every adapter it found and the one it uses.
|
||||
|
||||
- **Order:** adapters that report they can't present to the window are never tried; then the one `WGPU_ADAPTER_NAME` names; hardware before software (llvmpipe, WARP); the native backends (Vulkan, Metal, DX12) before OpenGL; with Automatic, the GPU the Wayland compositor draws with, then the GPU that drives a display, and among those the one that drives the primary display; then the power preference; then a GPU's DX12 adapter before its Vulkan one (Intel's Vulkan driver made the whole window flicker black, #545), keeping the system's order among equals.
|
||||
- **Windows backends:** Direct3D 12, falling back to OpenGL. Vulkan isn't started at all unless `WGPU_BACKEND` asks for it (`WGPU_BACKEND=vulkan`): creating a Vulkan instance loads every installed Vulkan driver into the process, and a faulty one (Intel's `igvk64.dll`, #806) crashed the app before its window appeared.
|
||||
- **Windows backends:** Direct3D 12, falling back to OpenGL. Vulkan isn't started at all unless `WGPU_BACKEND` asks for it (`WGPU_BACKEND=vulkan`): creating a Vulkan instance loads every installed Vulkan driver into the process, and a faulty one (Intel's `igvk64.dll`, #806) crashed the app before its window appeared. After a start that crashed while it listed the Direct3D 12 adapters, later starts list OpenGL's alone (see the crash bullet below).
|
||||
- **The GPU that drives the display.** A GPU without a monitor is the wrong one even when it says it can show the window: on a desktop with a Ryzen's integrated GPU (no monitor) and an NVIDIA card driving both monitors, power saving picked the integrated GPU, Windows had to present every frame across adapters, and the integrated GPU's driver reset under that: the graphics device was lost, or the desktop's compositor went down and every monitor stayed black for minutes (pdfcraft#378). So Automatic asks the system which GPU drives which display and puts that GPU first; the power preference only orders the rest, and decides alone where the system can't say.
|
||||
- **Windows:** `EnumDisplayDevices` (the display devices attached to the desktop, and the primary one), through the `winsafe` crate; the PCI ids in each device id are matched against wgpu's adapters. The registry doesn't tell a user which adapter drives a display, and WMI reports a resolution for a GPU without a monitor, so neither is used.
|
||||
- **Linux:** the connected connectors in `/sys/class/drm` and their cards' PCI ids; a built-in panel (`eDP`, `LVDS`, `DSI`) counts as the primary display.
|
||||
@@ -63,6 +63,7 @@ The canvas is rasterized on the CPU (`vectorcraft-render`, vello_cpu); the GPU (
|
||||
- 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. 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.
|
||||
- **If the window's first frame never reaches the screen** within 20 seconds of the app being created (the files named at the start opened), the adapter counts as hanging and the app starts again without it in the same way (#964: High Performance on a hybrid laptop left the app running in the background with no window and nothing in the log). A window minimized at the start isn't counted. When the processor that hung was chosen in Preferences, the status bar points to Automatic.
|
||||
- **If the app crashes while the window starts**, as when Intel's Direct3D 12 driver (`igd12umd64.dll`) ended 0.9.0 before its window appeared on an Intel UHD Graphics 605 (#2243), no code in the process runs to start it again, so the next start leaves the adapter out. Each start writes `graphics-start.json` in the log folder before the graphics start and adds the adapter once one is picked; when the first frame reaches the screen or the app ends without crashing, the start is cleared from the file. A start that finds an earlier start still recorded there logs the adapter, leaves it out and shows a note in the status bar. When the crash came while the adapters were being listed (wgpu's Direct3D 12 backend creates a device on each GPU to list it), the next start lists OpenGL's adapters alone; on macOS, where wgpu has no OpenGL adapters, it lists them all. Later starts keep leaving them out until the app version or Preferences › Performance › Graphics Processor changes, a start shows its first frame on the adapter that crashed, or no adapter is left that can show the window. `WGPU_BACKEND` and `WGPU_ADAPTER_NAME` choose regardless of the file: a start with `WGPU_BACKEND=dx12` tries Direct3D 12 again, and when its first frame appears, the crash is cleared. A process ended from outside during the start (Task Manager, `kill`) counts as a crash. Runs without preferences (`VECTORCRAFT_NO_PREFS`) record nothing, and a start made while another copy of the app is starting neither reads nor writes the file.
|
||||
- **To recover from a Graphics Processor choice that doesn't work** without editing files: start once with `WGPU_POWER_PREF=low` (or `WGPU_ADAPTER_NAME=…`), which overrides the preference for that start while settings still save, and set Graphics Processor back to Automatic. `VECTORCRAFT_NO_PREFS=1` also starts without it, but saves nothing.
|
||||
- **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).
|
||||
|
||||
@@ -614,7 +614,7 @@ Moved from the ROADMAP's "Shipped so far" on 2026-10-10. Update it when a featur
|
||||
When the desktop app opens a document on a thread of its own (#1245), that thread also reads the document's linked files and lays out its type, and the UI thread only adds the finished document: in release builds, the step left on the UI thread takes 0.02 ms instead of 1.04 s for a document with 12 linked 3000 × 2000 TIFFs, and it takes 0.1 ms instead of 0.85 s for one with 300 Auto Size area type objects, every second one in three columns.
|
||||
The Layers panel's Delete Selection, Duplicate Selection, Collect in New Layer, Merge Selected, Flatten Artwork, Release to Layers (Sequence) and row drags also take time in proportion to the size of the document, and so do Cut Artboard and Duplicate Artboard with their art and each move of an Alt-drag of one artboard with the Artboard tool while Move/Copy Artwork with Artboard is on (#1429).
|
||||
In Outline view the renderer draws paths and compound paths from the geometry it keeps for them between frames, and it culls the members of clip groups by the bounds it keeps for them. Layers with Preview off and the outlines of Image Trace objects are drawn the same way.
|
||||
- **Graphics processor:** the canvas is rasterized on the CPU and the GPU only composites it, so any GPU that can show the window will do. Preferences › Performance › Graphics Processor (`gpuPreference`) is Automatic by default: the power-saving GPU on Windows and macOS (the integrated one on hybrid-graphics laptops, where presenting from the discrete GPU made the window flicker, #306), and the GPU the desktop runs on under Linux, where KWin on an NVIDIA GPU refused frames from a Ryzen's integrated GPU and 0.5.0 crashed at startup (#502). Power Saving and High Performance pick a kind of GPU; `WGPU_POWER_PREF`, `WGPU_ADAPTER_NAME` and `WGPU_BACKEND` override them. The app picks among the adapters that can present to the window (native backends before OpenGL, hardware before software, and on Windows a GPU's DX12 adapter before its Vulkan one: Intel's Vulkan driver made the whole window flicker black, #545), and if the window fails on one while starting up it starts again without it instead of crashing; Help › About, `ui.inspect` and the log name the adapter in use. When the browser drops the page's WebGPU device or WebGL context, the web build writes recovery copies, starts again on new graphics with the same app and uploads its textures again; if it can't, it removes the stale canvas and says to reload, which offers the copies back. The desktop app writes the copies and says to restart (#369). On Windows and Linux the tools' cursors are OS cursors (bitmaps rasterized once per display scale), which the system moves itself: painted into the window, they trailed the pointer by the frames waiting to be shown, and the window now queues one frame instead of two (#444). macOS and the web still paint them (macOS sizes cursor bitmaps in points, eframe has no bitmap cursors on the web). Pens and touch drive the tools as the mouse does, and a tap pressed and lifted within one frame clicks where it tapped (#491); pen pressure reaches the Liquify tools' Use Pressure Pen on Windows, while winit 0.30 reports none on macOS or Linux and egui none on the web. On Linux, see `docs/development.md` for pens under Wayland.
|
||||
- **Graphics processor:** the canvas is rasterized on the CPU and the GPU only composites it, so any GPU that can show the window will do. Preferences › Performance › Graphics Processor (`gpuPreference`) is Automatic by default: the power-saving GPU on Windows and macOS (the integrated one on hybrid-graphics laptops, where presenting from the discrete GPU made the window flicker, #306), and the GPU the desktop runs on under Linux, where KWin on an NVIDIA GPU refused frames from a Ryzen's integrated GPU and 0.5.0 crashed at startup (#502). Power Saving and High Performance pick a kind of GPU; `WGPU_POWER_PREF`, `WGPU_ADAPTER_NAME` and `WGPU_BACKEND` override them. The app picks among the adapters that can present to the window (native backends before OpenGL, hardware before software, and on Windows a GPU's DX12 adapter before its Vulkan one: Intel's Vulkan driver made the whole window flicker black, #545), and if the window fails on one while starting up it starts again without it instead of crashing; after a start that a graphics driver crashed before its first frame, the next starts leave that adapter out, or list OpenGL's adapters alone when the crash came while the adapters were being listed (#2243; verified by unit tests, not on the reporter's Intel UHD Graphics 605); Help › About, `ui.inspect` and the log name the adapter in use. When the browser drops the page's WebGPU device or WebGL context, the web build writes recovery copies, starts again on new graphics with the same app and uploads its textures again; if it can't, it removes the stale canvas and says to reload, which offers the copies back. The desktop app writes the copies and says to restart (#369). On Windows and Linux the tools' cursors are OS cursors (bitmaps rasterized once per display scale), which the system moves itself: painted into the window, they trailed the pointer by the frames waiting to be shown, and the window now queues one frame instead of two (#444). macOS and the web still paint them (macOS sizes cursor bitmaps in points, eframe has no bitmap cursors on the web). Pens and touch drive the tools as the mouse does, and a tap pressed and lifted within one frame clicks where it tapped (#491); pen pressure reaches the Liquify tools' Use Pressure Pen on Windows, while winit 0.30 reports none on macOS or Linux and egui none on the web. On Linux, see `docs/development.md` for pens under Wayland.
|
||||
- **Control transport:** idle sockets expire after 5 minutes, partial lines after an absolute 30 seconds from their first byte, and reply writes after 60 seconds (#1726). Real TCP regressions cover stalled peers and slot recovery; command execution has its separate existing timeout. MCP limits: 16 MiB lines, 64-message batches, 256-step `command_batch` with 8 MiB of kept results and 256 MiB control replies (#1764, fixes #1727). MCP requires the JSON-RPC 2.0 envelope and matches control replies by id, and an app refusal without an id reports the app's reason (#1768, fixes #1728). The MCP server reconnects to the app before a call when the connection was closed or holds stray bytes, after an app restart or an idle expiry (#1739).
|
||||
- **MCP vertical type:** `add_text {vertical: true}` forwards to `text.create` or `text.createInPath`. Regression coverage checks point type, rectangular area type, type in a path and type on a path with true, false and omitted direction. This exposes an existing engine option; it does not change glyph orientation, CJK composition or the type/automation scores.
|
||||
- **Malformed MCP replies:** invalid JSON from the app drops the connection, reports that the operation may have completed and leaves the next call to reconnect; it never replays the edit (#1761). TCP regressions cover malformed and uncorrelated replies, connection closure and a subsequent inspection. Envelope regressions also verify zero backend calls and unchanged initialization state for invalid versions.
|
||||
|
||||
Reference in New Issue
Block a user