wgpu_native_zig

Zig bindings for wgpu-native

Requires Zig 0.16.x.

This package exposes two modules: wgpu-c and wgpu.

wgpu-c is wgpu.h (and therefore webgpu.h) translated directly by Zig, so it tracks wgpu-native's C API without a second handwritten declaration layer.

wgpu is a Zig-friendly wrapper over the bindings generated directly from wgpu.h and webgpu.h. The generated declarations are also available through wgpu.raw; wrapper methods and raw calls therefore share the headers as their single ABI source of truth.

Binding coverage

The pinned wgpu-native v29.0.0.0 headers currently declare 226 functions. The wgpu wrapper exposes every function with a usable v29 implementation, while wgpu.raw and wgpu-c expose all function and type declarations produced by translate-c. A compile-time audit checks this partition, rejects handwritten extern fn wgpu... declarations, and fails when a future header update adds an unclassified function or type.

The following v29 API groups intentionally remain raw-only because their upstream implementations panic, are blocked, or always return an unavailable result:

• wgpuGetProcAddress and the currently unimplemented *SetLabel functions.
• Async pipeline creation, shader compilation info, device-lost futures, waitAny, WGSL-language feature queries, and global instance-feature enumeration.
• Buffer map-state and mapped-range copy helpers.
• External-texture lifecycle functions.
• Device adapter-info lookup and texture binding-view-dimension lookup.
• The native Metal command-queue accessor, which always returns null in v29.

Supported wgpu-native extensions are available as regular Zig methods, including Queue.getTimestampPeriod, graphics-debugger capture control, and the borrowed Metal device/texture accessors. Platform-native pointers are optional and must not be released by the caller.

Adding this package to your build

Add the package to your dependencies, either with:

zig fetch --save https://github.com/openharmony-zig/wgpu_native_zig/archive/refs/tags/v7.0.0.tar.gz

or by manually adding to your build.zig.zon:

.{
// ...other stuff
.dependencies= .{
// ...other dependencies
.wgpu_native_zig= .{
// You can either use a commit hash:
.url="https://github.com/openharmony-zig/wgpu_native_zig/archive/<commit_hash>.tar.gz",
// or a tagged release:// .url = "https://github.com/openharmony-zig/wgpu_native_zig/archive/refs/tags/v7.0.0.tar.gz`
.hash="<dependency hash>"
}
}
}

Then, in build.zig add:

constwgpu_native_dep=b.dependency("wgpu_native_zig", .{});
// Add module to your exe (wgpu-c can also be added like this, just pass in "wgpu-c" instead)exe.root_module.addImport("wgpu", wgpu_native_dep.module("wgpu"));
// Or, add to your lib similarly:lib.root_module.addImport("wgpu", wgpu_native_dep.module("wgpu"));

Building on Windows

Windows x86_64 has two options for ABI: GNU and MSVC. For i686 and aarch64, only the MSVC option is available. If you need to specify the build target, you can do that with:

consttarget=b.standardTargetOptions(.{
.default_target= .{
// If not specified, defaults to the GNU abi
.abi=.msvc,
}
});

Or, specify it with your build command. For example, the triangle example in this repository can be run like so:

zig build --build-file build.examples.zig run-triangle-example -Dtarget=x86_64-windows-msvc

Either way, pass the resolved target to the dependency like so:

constwgpu_native_dep=b.dependency("wgpu_native_zig", .{
.target=target
});

When using static linking with MSVC, you might encounter duplicate symbol errors. If so, try

if (target.result.abi==.msvc) {
// "exe" here is the *std.Build.Step.Compile from b.addExecutable() (or b.addTest())exe.bundle_compiler_rt=false;
exe.bundle_ubsan_rt=false;
}

An example of using wgpu-native-zig with static linking on Windows can be found at wgpu-native-zig-windows-test.

Dynamic linking

Dynamic linking can be made to work, though it is a bit messy to use. When you initialize your wgpu_native_dep, add the option for dynamic linking like so:

constwgpu_native_dep=b.dependency("wgpu_native_zig", .{
// Defaults to .static if you don't specify
.link_mode=.dynamic
});

Then add the following with your install step dependencies:

constlib_dir=wgpu_native_dep.namedWriteFiles("lib").getDirectory();
// This would also work with .so files on linuxconstdll_path=lib_dir.join(b.allocator, "wgpu_native.dll") catchreturn;
// addInstallBinFile puts the dll in the same directory as your executableconstinstall_dll=b.addInstallBinFile(dll_path, "wgpu_native.dll");
// Make sure that the dll is installed when the install step is runb.getInstallStep().dependOn(&install_dll.step);

Building wgpu-native

wgpu-native v29.0.0.0 is built from its pinned source commit by default. The matching webgpu-headers commit is pinned separately, so the generated C ABI does not drift when an upstream branch changes.

The source build requires Cargo, a Rust toolchain with the selected target installed, and the platform SDK normally required by that target. For example:

# Native source build. Installs the static library, dynamic library, and C headers.
zig build -Doptimize=ReleaseFast
# Compile the bindings and link probes without running them.
zig build --build-file build.tests.zig check -Doptimize=ReleaseFast

Targets that cannot legally be built on every host still require their native toolchain: iOS requires Xcode on macOS, MSVC requires Windows, and Android/OpenHarmony require their respective NDKs.

The root build.zig only builds wgpu-native and exposes the wgpu/wgpu-c binding modules. build/Library.zig is the shared library entry point, while build/platform/root.zig dispatches to the Android, Apple, Linux, OpenHarmony, or Windows build implementation. Tests and examples are isolated behind build.tests.zig and build.examples.zig; their implementation stays in the corresponding directory.

Using published prebuilt libraries

Set WGPU_NATIVE_USE_PREBUILT=1 to skip the Cargo source build and use the published wgpu-native archive for the selected target:

WGPU_NATIVE_USE_PREBUILT=1 zig build --build-file build.tests.zig check \
-Dtarget=x86_64-linux-gnu

The equivalent Zig build option is -Duse_prebuilt=true. Downstream packages can pass it while resolving this dependency:

constwgpu_native_dep=b.dependency("wgpu_native_zig", .{
.target=target,
.optimize=optimize,
.use_prebuilt=true,
});

WGPU_NATIVE_PREBUILT_DIR=/path/to/prefix uses a local artifact directory and also implies prebuilt mode. The prefix must use the layout produced by this package:

prefix/
├── include/webgpu/{webgpu.h,wgpu.h}
└── lib/
├── libwgpu_native.a
└── libwgpu_native.so

Use the platform-specific dynamic and import-library names on Apple and Windows. OpenHarmony currently uses this local-directory mechanism when consuming CI artifacts, because upstream wgpu-native does not publish OpenHarmony archives.

Supported artifact targets

Platform	Architectures / ABIs	Source build	Published prebuilt
Android	arm64-v8a, armeabi-v7a, x86, x86_64	Yes, with ANDROID_NDK_HOME	Yes
iOS	arm64 device, arm64 simulator, x86_64 simulator	Yes, on macOS	Yes
Linux	aarch64, x86_64 (GNU); aarch64, x86_64 (musl)	Yes	GNU targets
macOS	aarch64, x86_64	Yes, on macOS	Yes
Windows	aarch64/x86/x86_64 MSVC, x86/x86_64 GNU	Yes, on Windows	All except x86 GNU
OpenHarmony	arm64-v8a, armeabi-v7a, x86_64	Yes, with OHOS_NDK_HOME	Local CI artifact

The OpenHarmony commands are:

rustup target add \
aarch64-unknown-linux-ohos \
armv7-unknown-linux-ohos \
x86_64-unknown-linux-ohos
zig build --build-file build.tests.zig check \
-Dtarget=aarch64-linux-ohos -Doptimize=ReleaseFast
zig build --build-file build.tests.zig check \
-Dtarget=arm-linux-ohoseabi -Doptimize=ReleaseFast
zig build --build-file build.tests.zig check \
-Dtarget=x86_64-linux-ohos -Doptimize=ReleaseFast

The target-artifact workflow builds the complete matrix on Linux x86_64/aarch64, macOS arm64/Intel, Windows, Android, and OpenHarmony runners. Every target runs the binding/ABI audit before packaging. Desktop targets compile all wrapper tests, while Android, OpenHarmony, and arm64 iOS targets build a Zig link probe. Zig 0.16 cannot link the historical x86_64-apple-ios Mach-O platform marker as an explicit simulator dylib, so that one target runs the binding/ABI audit without the link probe. Every artifact prefix contains both link modes and the matching headers.

How the wgpu module differs from wgpu-c

• Names are shortened to remove redundancy.
  • For example wgpu.WGPUSurfaceDescriptor becomes wgpu.SurfaceDescriptor
• C pointers ([*c]) are replaced with more specific pointer types.
  • For example [*c]const u8 is replaced with ?[*:0]const u8.
• Pointers to opaque structs are made explicit (and only optional when they need to be).
  • For example wgpu.WGPUAdapter from webgpu.h would instead be expressed as *wgpu.Adapter or ?*wgpu.Adapter, depending on the context.
• Methods are expressed as decls inside of structs
  • For example

    wgpu.wgpuInstanceCreateSurface(instance: WGPUInstance, descriptor: [*c]constWGPUSurfaceDescriptor) WGPUSurface

    becomes

    Instance.createSurface(self: *Instance, descriptor: *constSurfaceDescriptor) ?*Surface

• Callback-based operations provide synchronous helpers for adapter/device requests, buffer mapping, submitted queue work, and device error scopes.
  • For example, requesting an adapter with a callback looks something like

    fnhandleRequestAdapter(
    status: Instance.RequestAdapterStatus,
    adapter: ?*Adapter,
    message: StringView,
    userdata1: ?*anyopaque,
    userdata2: ?*anyopaque
    ) callconv(.c) void {
    switch(status) {
    .success=> {
    constud_adapter: **Adapter=@ptrCast(@alignCast(userdata1));
    ud_adapter.*=adapter.?;
    },
    else=> {
    std.log.err("{s}\n", .{message.toSlice()});
    }
    }
    constcompleted: *bool=@ptrCast(@alignCast(userdata2));
    completed.*=true;
    }
    varadapter_ptr: ?*Adapter=null;
    varcompleted=false;
    constrequest_adapter_info=Instance.RequestAdapterCallbackInfo {
    .callback=handleRequestAdapter,
    .userdata1=@ptrCast(&adapter_ptr),
    .userdata2=@ptrCast(&completed),
    }
    constra_future=instance.requestAdapter(null, request_adapter_info);
    // wgpu-native v29 does not implement Instance.waitAny(), so drive// allow_process_events callbacks with Instance.processEvents()._=ra_future;
    instance.processEvents();
    while(!completed) {
    tryio.sleep(.fromNanoseconds(200_000_000), .awake);
    instance.processEvents();
    }

    whereas the non-callback version looks like

    // The wrapper methods use polling, so 200_000_000 is the polling interval in nanoseconds.varresponse=tryinstance.requestAdapterSync(allocator, io, null, 200_000_000);
    deferresponse.deinit(allocator);
    constadapter_ptr: ?*Adapter=switch (response.status) {
    .success=>response.takeAdapter(),
    else=>blk: {
    std.log.err("{s}\n", .{response.messageorelse"adapter request failed"});
    break :blknull;
    }
    };

    The synchronous responses own their copied callback messages and any returned handle. Call deinit() on every response, and use takeAdapter() or takeDevice() to transfer a successful handle out of it. The other helpers follow the same lifetime model: Buffer.mapSync(), Queue.onSubmittedWorkDoneSync(), and Device.popErrorScopeSync(). If the supplied Io is cancelled, the synchronous wrapper finishes draining the native callback before returning the cancellation error, yielding the polling thread between processEvents() calls. This is required because wgpu-native v29 has no waitAny() implementation and the callback userdata must remain alive until completion.
• Output structures whose members are allocated by wgpu-native use pointer-based deinit() methods. In particular, SupportedFeatures, AdapterInfo, and SurfaceCapabilities clear their owned pointers, counts, and strings after releasing them, so a deferred deinit() cannot observe stale members. SupportedFeatures.slice() and the SurfaceCapabilities.*Slice() accessors expose their pointer/count arrays as bounded read-only slices.
• Instance.enumerateAdapters() allocates an AdapterList that owns every returned adapter. Deinitialize the list to release all remaining adapters, or transfer one:

  varadapters=tryinstance.enumerateAdapters(allocator, null);
  deferadapters.deinit(allocator);
  constadapter=adapters.takeAdapter(0).?;
  deferadapter.release();

• SurfaceTexture owns the texture returned by Surface.getCurrentTexture(). Use defer surface_texture.deinit() to release it automatically, or call takeTexture() to transfer the texture to code that will release it.
• Wrapper methods use slices where the C API uses a pointer/count pair. Queue.writeBuffer() and Queue.writeTexture() take byte slices, setBindGroup() takes a dynamic-offset slice, and setImmediates() takes a byte slice. Buffer mapped-range accessors return bounded byte slices and resolve WGPU_WHOLE_MAP_SIZE against the buffer size.
• ABI-compatible descriptors retain their C pointer/count fields, while init() and chainable withXxx() helpers accept borrowed slices and synchronize both fields. The slice must remain alive until the native call returns.
• Chained structs are provided with inline functions for constructing them, which come in two forms depending on whether or not the chained struct is likely to always be required.
  • For required chained structs, you can either write them explicitely:

    constsource=SurfaceSourceXlibWindow{
    .display=display,
    .window=window,
    };
    constdescriptor=SurfaceDescriptor{
    .next_in_chain=&source.chain,
    .label=StringView.fromSlice("xlib_surface_descriptor"),
    };

    or use a function to construct the root descriptor:

    constsource=SurfaceSourceXlibWindow{
    .display=display,
    .window=window,
    };
    constdescriptor=surfaceDescriptorFromXlibWindow(
    &source,
    "xlib_surface_descriptor",
    );

  • For optional chained structs, create the extension separately and attach it with withExtras():

    constextras=SurfaceConfigurationExtras{
    .desired_maximum_frame_latency=2,
    };
    constconfiguration= (SurfaceConfiguration{
    .device=device,
    // other fields
    }).withExtras(&extras);

    The source or extras value must remain alive until the native API call that consumes the descriptor has returned.
• WGPUBool is replaced with bool whenever possible.
  • This means it is replaced with bool in wrapper method parameters and return values, but not in structs that preserve the C ABI.
• Callback types that belong to one handle are scoped under that handle. For example, use Instance.RequestAdapterCallbackInfo, Adapter.RequestDeviceCallbackInfo, Buffer.MapCallbackInfo, and Device.PopErrorScopeCallbackInfo.
• Wrapper names retain meaningful WebGPU prefixes. For example, WGPUTextureSampleType is exposed as TextureSampleType.

TODO

• Expand headless coverage for limit queries, textures, samplers, query sets, and render bundles.
• Port wgpu-native-examples using wrapper code, as a basic form of documentation.