flutter3d_wgpu 1.0.0-rc.2

The compute sidecar for flutter3d: a compute-only GraphicsDevice over wgpu-native through dart:ffi, so compute steps run on the GPU when the drawing device has none.

1.0.0-rc.2

Since the 2026-10-14 upload, the sidecar says which backend it chose. A readback no longer holds the caller, staging buffers and bind groups are kept, the encoder checks ranges and dispatches as every backend's does, and what it cannot draw is refused in the hardware layer's words. Nothing that went up on 2026-10-14 broke.

• A device wgpu loses is told on lost. wgpu-native's device-lost callback is wired through the C surface, and the sidecar hears it at beginFrame, at a failed read and at a failed submit: one DeviceLoss with reason unknown and wgpu's words, and isLost turns true. Before, only the sidecar's own dispose was told.

• The device request reaches the adapter. WgpuSidecar.open takes a DeviceRequest: its power preference picks the adapter, validation switches on the API's validation layer and label names the device. With no request, or PowerPreference.none, the API's default adapter is asked for; the sidecar asked for the discrete GPU whatever was wanted. The C surface is at ABI 4.

• The sidecar decides its backend and says which. It opened wgpu with every backend enabled, so a Windows machine computed on Direct3D 12 or on Vulkan depending on its drivers, and nothing recorded which. Now the platform's own backends are tried in order, each alone (Direct3D 12 then Vulkan on Windows, Metal on Apple platforms, Vulkan then OpenGL elsewhere), WgpuSidecar.open takes a backend and a fallbackAdapter (WARP on Direct3D 12, what CI computes on), and WgpuSidecar.adapterInfo (new) answers a WgpuAdapterInfo with the WgpuBackend it runs on. A backend asked for and missing is refused, never replaced. The C side is ABI 3 (f3d_wgpu_open_on, f3d_wgpu_adapter_backend, f3d_wgpu_adapter_is_fallback).

• A readback that does not hold the caller. readBuffer starts the copy, maps its staging buffer without waiting and polls the device on later turns of the event loop, where it waited on the queue in the calling turn: a frame's read now holds the UI isolate for its encoding, 0.09 ms in place of 2.19 ms for 2 MiB on an M-series Mac, and lands in 0.74 ms in place of 2.20 ms. readBufferSync still waits. The C side is ABI 2 (f3d_wgpu_read_begin, f3d_wgpu_read_poll).

• Staging buffers and bind groups are kept. A read copies through a staging buffer from a pool the device keeps, and a kernel keeps the last eight bind groups it made, keyed by resource ids that are never reused, and lets go of those of a buffer or texture released; each was made per read and per dispatch.

• The encoder checks ranges and dispatches, as every backend's does: a storage offset off the 256-byte alignment, which wgpu would have refused late, is a RangeError at the call.

• Its own dispose is a loss. The sidecar reported no loss at all; its dispose reports DeviceLossReason.destroyed on lost (new) and isLost follows.

• What it cannot draw is refused in the HAL's words: a render pass, pipeline or geometry is a DeviceResourceException, where it was an UnsupportedError. Its registration takes the HAL's DeviceRequest.

• Breaking: wgpu_native_hook.dart is no library of this package. The fetch of the pinned wgpu-native is the build hook's own file, hook/wgpu_native.dart; nothing an application imports changes.

• The limits say the sidecar draws nothing: no colour attachment, one sample, no vertex buffer, attribute or inter-stage variable, where it answered WebGPU's defaults for all of them. A sampler handed to bindTexture is checked by its features.

• backendId is wgpu. It followed the display name, the wgpu sidecar; an id is fixed for good.

• The compute sidecar. WgpuSidecar.open makes a compute-only GraphicsDevice over wgpu-native through dart:ffi: storage buffers, RGBA8 textures and storage textures, a bundle's .f3dmat compute kernels loaded through flutter3d_shaders' WGSL, compute passes and readback. Everything that draws is refused by name. registerWgpuSidecar adds it to a DeviceRegistry as ComputePlacement.sidecar. The four conformance kernels give the CPU's bytes on it.

• The pinned wgpu-native, in one file. hook/wgpu_native.dart is the fetch this package's build hook uses: the release, its checksums and the unpacking. flutter3d_physics_native keeps a copy under its own hook/, held to the same bytes, and does not depend on this package.

• Zero copy's first half. WgpuSidecar.metalTextureOf answers the MTLTexture behind a sidecar texture on Metal. The sidecar does not report DeviceFeature.sharedTextures, since it cannot import one back, so a SharedTexture between it and Impeller copies through host memory.