With wgpu-native (🟠WIP)¶
Resulting code: wgpu, wgpu-native and step030-test-foo-extension
Setup¶
Before modifying any code, we need to set up the 2 repositories that wgpu-native relies on.
Building wgpu-native¶
Start by building a “regular” wgpu-native using the instructions from their repository. You need in particular rust and LLVM/Clang. With these installed, building looks like this:
git clone --recurse-submodules https://github.com/gfx-rs/wgpu-native
cd wgpu-native
set LIBCLANG_PATH=E:/Libraries/libclang/bin
cargo build --release
Important
Adapt the value of LIBCLANG_PATH above to your actual installation of LLVM/Clang.
You then find the binaries in target\release. If you use the wgpu-native based WebGPU-distribution in your end project, simply replace the relevant files in webgpu/bin.
Building wgpu¶
The wgpu-native repository is a thin layer exposing as a C interface the actual wgpu backend. When creating a custom extension, we need to change the backend, and instruct the wgpu-native layer to use our custom wgpu branch.
git clone https://github.com/gfx-rs/wgpu
cd wgpu
cargo build --release
To point wgpu-native to our custom wgpu, we can modify its wgpu-native/Cargo.toml and add:
[patch."https://github.com/gfx-rs/wgpu"]
wgpu-core = { path = "../wgpu/wgpu-core" }
wgpu-types = { path = "../wgpu/wgpu-types" }
wgpu-hal = { path = "../wgpu/wgpu-hal" }
naga = { path = "../wgpu/naga" }
Note
To make sure to reproduce the very same binaries, check out in the wgpu directory the commit specified after rev = in the Cargo.toml of wgpu-native.
cd wgpu
git checkout 011a4e26d04f388ef40e3baee3f19a255b9b5148
But since you are writing an extension you may want to use the last version of wgpu instead.
You may also need to update the rev hash of [dependencies.naga] to match what your version of wgpu uses.
The Foo Extension¶
Extending types¶
We can copy our webgpu-ext-foo.h header in the wgpu-native/ffi/ directory, next to wgpu.h. In order to have rust’s build system parse these files, we add our custom header to build.rs:
let mut builder = bindgen::Builder::default()
.header("ffi/webgpu-headers/webgpu.h")
.header("ffi/wgpu.h")
.header("ffi/webgpu-ext-foo.h") // <-- Here is our custom file
// [...]
This defines in rust’s native namespace symbols that are equivalent to what the C header files expose. We must also modify some existing types in wgpu-types/src/lib.rs. At least to add our new adapter feature:
bitflags::bitflags! {
// [...]
pub struct Features: u64 {
// [...]
const SHADER_EARLY_DEPTH_TEST = 1 << 62;
// Our feature as a bitflag
const FOO = 1 << 63;
}
}
In our running example, we want to add a field foo to the render pipeline. We thus modify the RenderPipelineDescriptor in wgpu/wgpu-core/src/pipeline.rs:
pub struct RenderPipelineDescriptor<'a> {
// [...]
/// our custom new field
pub foo: Option<u32>,
}
Note
When trying to build from the wgpu root, do not try to build the whole project, we only use wgpu-core and wgpu-types (which the former depends on), so try building with:
wgpu$ cargo build --package wgpu-core
But more likely you will build from the wgpu-native projects:
wgpu-native$ cargo build
If you want your extension to also be available to rust users, you must also adapt the wgpu package, but I will not cover it here.
Extending native wrapper¶
Conversion utilities to link rust-side types defined in wgpu/wgpu-types with the native defines that follow the C header. Add your feature name to features_to_native and map_feature in wgpu-native/src/conv.rs:
pub fn features_to_native(features: wgt::Features) -> Vec<native::WGPUFeatureName> {
// [...]
if features.contains(wgt::Features::FOO) {
temp.push(native::WGPUFeatureName_Foo);
}
temp
}
pub fn map_feature(feature: native::WGPUFeatureName) -> Option<wgt::Features> {
use wgt::Features;
match feature {
// [...]
native::WGPUFeatureName_Foo => Some(Features::FOO),
// [...]
}
}
In order to recognize our SType when it is passed in the extension chain of a RenderPipelineDescriptor, we modify the wgpuDeviceCreateRenderPipeline procedure in wgpu-native/src/lib.rs:
pub unsafe extern "C" fn wgpuDeviceCreateRenderPipeline(
device: native::WGPUDevice,
descriptor: Option<&native::WGPURenderPipelineDescriptor>,
) -> native::WGPURenderPipeline {
let (device, context) = device.unwrap_handle();
let descriptor = descriptor.expect("invalid descriptor");
let desc = wgc::pipeline::RenderPipelineDescriptor {
// [...]
// Iteratively explore the extension chain until it finds an extension
// of our type, cast to WGPUFooRenderPipelineDescriptor and retrieve
// the value of foo.
foo: follow_chain!(
map_render_pipeline_descriptor(
descriptor,
WGPUFooSType_FooRenderPipelineDescriptor => native::WGPUFooRenderPipelineDescriptor)
),
};
}
This calls a map_render_pipeline_descriptor that we create in conv.rs:
pub unsafe fn map_render_pipeline_descriptor<'a>(
_: &native::WGPURenderPipelineDescriptor,
foo: Option<&native::WGPUFooRenderPipelineDescriptor>,
) -> Option<u32> {
foo.map(|x| x.foo)
}
Note
Do not forget to add map_render_pipeline_descriptor to the use conv::{...} line at the beginning of lib.rs.
Extending core¶
If your extension is shallow enough not to affect the backends, you should only have to modify wgpu/wgpu-core. But if you take the time to write a custom extension, it likely requires to modify one or multiple backends.
Extending backends¶
Warning
TODO: I still need to learn better how the code is organized. So far I noticed that:
wgpu-nativemaps C entry points to the rust API ofwgpu-corewgpu-coremaintains the common user API, that application based on wgpu use, either through the native wrapper or throughwgpu-rs(a.k.a. justwgpu). Behinds the scenes, it maps instructions towgpu-halwgpu-halis the backend/hardware abstraction layer, it defines the internal API that each backend (Vulkan, Metal, DX12, etc.) must implementwgpu-hal/vulkanis the Vulkan backend, that implements all of the HAL’s requirements
The file wgpu/wgpu-hal/src/lib.rs defines the interface that each backend must implement. Backends are the subdirectories of wgpu/wgpu-hal/src as well as the empty.rs file that defines a default behavior that does nothing.
We implement backends one by one, maybe only for the ones that interest us in practice. We must thus make sure that the Foo feature is advertised by the adapter only for the backend that we implemented.
Let’s start with the Vulkan backend. We first advertise that the adapter (a.k.a. physical device in Vulkan wording) supports our feature.
Note
Of course you may inspect the actual physical device properties to conditionally list the FOO feature only if the mechanism you want to implement is indeed supported.
// in wgpu/wgpu-hal/src/vulkan/adapter.rs
impl PhysicalDeviceFeatures {
// [...]
fn to_wgpu(
/* [...] */
) -> (wgt::Features, wgt::DownlevelFlags) {
// [...]
let mut features = F::empty()
| F::SPIRV_SHADER_PASSTHROUGH
| F::MAPPABLE_PRIMARY_BUFFERS
// [...]
| F::FOO; // our extension here!
}
// [...]
}
Now in DirectX 12 backend:
// in wgpu/wgpu-hal/src/dx12/adapter.rs
impl super::Adapter {
// [...]
pub(super) fn expose(
/* [...] */
) -> Option<crate::ExposedAdapter<super::Api>> {
// [...]
let mut features = wgt::Features::empty()
| wgt::Features::DEPTH_CLIP_CONTROL
| wgt::Features::DEPTH32FLOAT_STENCIL8
// [...]
| wgt::Features::FOO; // our extension here!
}
// [...]
}
Note
We focus only on the Vulkan backend in the remainder of this section.
We can now check that the feature is correctly made available in our application code. I start from the step030-headless branch.
if (wgpuAdapterHasFeature(adapter, (WGPUFeatureName)WGPUFeatureName_Foo)) {
std::cout << "Feature 'Foo' supported by adapter" << std::endl;
}
else {
std::cout << "Feature 'Foo' NOT supported by adapter" << std::endl;
}
After you copy the DLL and headers of your custom wgpu-native, you should see the Foo feature supported.
Note
You can force a particular backend by playing with the instance extras extension of wgpu-native in the instance descriptor:
WGPUInstanceExtras instanceExtras = {};
instanceExtras.chain.next = nullptr;
instanceExtras.chain.sType = (WGPUSType)WGPUSType_InstanceExtras;
instanceExtras.backends = WGPUInstanceBackend_Vulkan; // We only accept Vulkan adapters here!
WGPUInstanceDescriptor instanceDesc = {};
instanceDesc.nextInChain = &instanceExtras.chain;
Instance instance = createInstance(instanceDesc);
Do not forget to also request the feature when creating your device:
WGPUFeatureName featureFoo = (WGPUFeatureName)WGPUFeatureName_Foo;
deviceDesc.requiredFeaturesCount = 1;
deviceDesc.requiredFeatures = &featureFoo;
// [...]
if (wgpuDeviceHasFeature(device, (WGPUFeatureName)WGPUFeatureName_Foo)) {
std::cout << "Feature 'Foo' supported by device" << std::endl;
}
else {
std::cout << "Feature 'Foo' NOT supported by device" << std::endl;
}
We can then try extending the render pipeline:
RenderPipelineDescriptor pipelineDesc;
// [...]
WGPUFooRenderPipelineDescriptor fooRenderPipelineDesc;
fooRenderPipelineDesc.chain.next = nullptr;
fooRenderPipelineDesc.chain.sType = (WGPUSType)WGPUFooSType_FooRenderPipelineDescriptor;
fooRenderPipelineDesc.foo = 42; // some test value here.
pipelineDesc.nextInChain = &fooRenderPipelineDesc.chain;
RenderPipeline pipeline = device.createRenderPipeline(pipelineDesc);
In order to test that the value is correctly propagated, we just print a log line whenever setPipeline is called for this pipeline:
// In wgpu/wgpu-hal/src/vulkan/command.rs
unsafe fn set_render_pipeline(&mut self, pipeline: &super::RenderPipeline) {
// Debug print
if let Some(foo) = pipeline.foo {
println!("DEBUG of FOO feature: foo={:?}", foo);
}
// [...]
}
We need for this to add a foo field to the RenderPipeline (we only added it to the descriptor for now):
// In wgpu/wgpu-hal/src/vulkan/mod.rs
#[derive(Debug)]
pub struct RenderPipeline {
raw: vk::Pipeline,
foo: Option<u32>, // add this
}
And we propagate from the descriptor when creating the render pipeline:
// In wgpu/wgpu-hal/src/vulkan/device.rs
unsafe fn create_render_pipeline(
&self,
desc: &crate::RenderPipelineDescriptor<super::Api>,
) -> Result<super::RenderPipeline, crate::PipelineError> {
// [...]
let foo = desc.foo;
Ok(super::RenderPipeline { raw, foo })
}
But, as you may notice, this is yet another taste of RenderPipelineDescriptor type (defined in wgpu-hal/lib.rs). As this is getting quite confusing, let me summarize with a figure:
The architecture of the wgpu-native project.
// In wgpu/wgpu-hal/lib.rs
#[derive(Clone, Debug)]
pub struct RenderPipelineDescriptor<'a, A: Api> {
// [...]
/// Our custom FOO extension
pub foo: Option<u32>,
}
// In wgpu/wgpu-core/src/device/resource.rs
pub(super) fn create_render_pipeline<G: GlobalIdentityHandlerFactory>(
/* [...] */
) -> Result<pipeline::RenderPipeline<A>, pipeline::CreateRenderPipelineError> {
// [...]
let pipeline_desc = hal::RenderPipelineDescriptor {
// [...]
multiview: desc.multiview,
foo: desc.foo, // forward our custom member here
};
// [...]
}
At this stage, if you build wgpu-native, update the DLL in your C++ application and run that application, it should eventually display the test log line:
DEBUG of FOO feature: foo=42
Congratulation, you have your first extension of wgpu-native! Of course it does not do much, and it only does so on the Vulkan backend. But we have explored the architecture of the project. What remains now depends highly on what exact extension you want to implement!
Resulting code: wgpu, wgpu-native and step030-test-foo-extension