Architecture

Here's an overview of wgpu's architecture:

Working through this diagram from the bottom up:

• Each operating system provides its own API (or APIs) for getting at the GPU:

  • On Windows, Direct3D is the primary GPU programming API, but Vulkan and OpenGL are often also available.

  • On macOS, Metal is the official GPU programming API, although OpenGL is still around for legacy programs, and the MoltenVK open source project implements the Vulkan API on top of Metal.

  • On Linux, the Mesa library supports Vulkan and OpenGL.

  If you want to write an application that runs on all three of these platforms, you have a few options:

  • You could target Vulkan, and for macOS support embed MoltenVK, or require your users to install it.

  • You could target OpenGL, which is no longer being developed and is missing many modern features, and worry that macOS will drop support for it altogether.

  • You could port it to all three native platform APIs: Direct3D, Vulkan, and Metal. This would be a lot of code.

  • Or, you could just use wgpu!

• The wgpu_hal crate implements a portable Rust API that can use any of the platform-specific APIs mentioned above as a backend: a program that uses wgpu_hal correctly should behave consistently regardless of what platform you're running on.

  However, wgpu_hal's interface is completely unsafe; you must follow all of its safety requirements to the letter to avoid provoking undefined behavior from the underlying platforms. These requirements are complex, and not well-documented. And wgpu_hal performs almost no validation beyond the minimum necessary to ensure portability.

• The wgpu_core crate builds on wgpu_hal to provide a API with a similar flavor, preserving its portability, and adding full bullet-proof validation. By "bullet-proof", we mean that wgpu_core is intended to be driven by untrusted code, like web content using the WebGPU API. wgpu_core only assumes that Rust's safety rules are respected. Safe Rust code should not be able to cause a crash using wgpu_core.

  This means that wgpu_core is fully responsible for things like tracking resource lifetimes, generating barriers for usage transitions, checking parameters, and so on. We cover its duties in more detail below.

  The wgpu_core crate's API is designed to be easy to use from other languages via foreign function interfaces (FFI). This restricts the Rust features we can use in the API somewhat. But this makes wgpu_core the best dependency for crates like wgpu_native and applications like Deno and Firefox that have their own binding systems.

• The wgpu crate provides an idiomatic Rust API on top of wgpu_core, inheriting its validation and portability. This is what most users in the Rust ecosystem will want.

  You can also compile wgpu to WebAssembly and run it in a web browser, having it drive the browser's WebGPU implementation. In this mode of use, wgpu_core and wgpu_hal are not present. This can help you share code between in-browser and native versions of your app, by providing a single API for both versions to use.

wgpu

The wgpu crate (docs) provides an idiomatic Rust API for cross-platform GPU-based graphics and computation, modeled after the WebGPU JavaScript API.

The wgpu crate itself doesn't contain much interesting graphics-related code. It is mostly concerned with providing a consistent, idiomatic API that applications can use to drive any one of several implementations:

• The wgpu::backend::wgpu_core module is available in native or Emscripten environments. It forwards all operations to wgpu_core. Most wgpu::Backend values, like Vulkan or Metal, go through this wgpu module.

• The wgpu::backend::webgpu module, available when wgpu has been compiled to WebAssembly, forwards calls to a browser's WebGPU implementation. The wgpu::Backend::BrowserWebGPU backend selects this module.

• The wgpu::backend::custom module uses dynamic dispatch to let the user supply their own implementation. These are created via a separate mechanism, so they don't use any wgpu::Backend value.

wgpu_core

The wgpu_core crate implements a safe, cross-platform API on top of wgpu_hal's unsafe, cross-platform API. Almost every method in wgpu_hal's API is marked unsafe; it is wgpu_core's job to satisfy or enforce every requirement in wgpu_hal's safety contracts. Then, wgpu_hal's only responsibility is to get consistent behavior across all its backends. This relationship is explained more in wgpu_hal's documentation.

Beyond safety, wgpu_core also implements a few convenience APIs (for example, mapped-at-creation buffers) on top of wgpu_hal's more primitive features.

Lifetime tracking

The wgpu_hal API requires some objects to outlive others. For example, a wgpu_hal::Device must outlive all resources (wgpu_hal::Buffer, wgpu_hal::CommandEncoder, etc) created from it. The wgpu_core types ensure that these requirements are upheld, mostly just by using Arc.

TODO: Command buffer resource usage through submission probably counts as "Lifetime tracking" too.

This wgpu_hal requirement arises mostly from the Vulkan backend. Direct3D and Metal both use reference counting internally for most operations, but Vulkan foists the entire problem off on its users.

Synchronization

Some types in the wgpu_hal API cannot be accessed simultaneously by multiple threads. For example, most wgpu_hal::CommandEncoder methods take &mut self. wgpu_core protects these hal types with Mutex as appropriate.

There are a few other interesting requirements; for example, calls to wgpu_hal::Queue::submit on a given Queue that could occur simultaneously must use different Fences. wgpu_core addresses this by simply wrapping wgpu_hal::Queue in a Mutex.

Barrier generation

Vulkan and Direct3D require the user to include barriers in the command stream to notify the driver when a particular buffer or texture is transitioning from one kind of use to another. For example, if a texture is used as a render target and then sampled by a shader, Direct3D 12 requires a ResourceBarrier command between the two uses. Similarly, Vulkan requires VkCmdPipelineBarrier commands in some cases to ensure that values written to a resource by one operation will be seen by reads in subsequent operation.

The wgpu_hal API passes these backends' requirements through to its own users, requiring calls to transition_buffers and transition_textures between different kinds of access to such resources. But, following the lead of WebGPU, the wgpu API does not require its users to record barriers. This means that wgpu_core must track how each resource was last used, and generate the barriers itself.

Since wgpu encourages users to record CommandBuffers in advance (ideally in multiple threads), and then submit them in whatever order they please, wgpu_core cannot know what states the resources that a CommandBuffer uses will be in until it is actually submitted. Thus, each wgpu_core::CommandBuffer merely records what states it expects the resources it uses to be in initially, and which states it will leave those resources in when it is done. Submission is then responsible for comparing resources' actual states with those expected by each wgpu_core::CommandBuffer, and recording the necessary barriers in a fresh wgpu_hal::CommandBuffer submitted ahead of it.

When recording a command buffer, since the commands are only being recorded for later execution, and not executed immediately, a CommandEncoder tracks the state each resource will be in when the given commands are executed, and adds barriers to the recorded stream as appropriate. Naturally, the CommandEncoder must treat each resource's initial state must be treated as unknown, as it has no way to know what other commands will be submitted ahead of it.

Note that, within a single wgpu_hal::Texture, individual array elements and mip levels can be in different states. wgpu_hal::Buffers, however, have only a single state.

Within wgpu_core, the Tracker type records a set of resources, their current states, and (for command buffers being recorded) the initial state they should be in. Tracker also notes which stateless resources (samplers; render pipelines) a CommandBuffer uses, in order to keep them alive until the commands have been submitted and finished running.

MemoryInitTracker

Tracks memory initialization state of Buffer resources, i.e. if they have been written to either by the user or by wgpu previously zero-initializing it (WebGPU requires all buffers to behave as-if they are zero inititialized). A MemoryInitTracker tracks a single buffer and allows to insert zero initialization lazily on use. Zero init is inserted if necessary at:

• memory mapping
• queue_submit (for all bindings)

Device maintain

TODO (what/why)

Cross-Device use

The wgpu_hal API requires that a resource created with one wgpu_hal::Device may not be used with a different Device. wgpu_core tracks the Device to which each resource belongs, and checks for cross-device use where necessary.

General parameter validation

The wgpu_core crate is responsible for all the usual parameter validation for graphics operations: a texture's format must be supported by the device; copies must not exceed the bounds of the buffers or textures involved; bind groups need to have the right layout for the pipeline; and so on.

API Tracing

Enabled via feature flag. Allows to record all usage (WebGPU api functions essentially) to a .ron file. Any additional data (uploaded data to buffer etc.) is stored in additional files. A trace can be replayed with the Player compiled from the same wgpu revision. Used for testing and bug reporting.

Other WebGPU implementations

Note that wgpu represents only one possible way to implement the WebGPU API; Google's Dawn and WebKit's WebGPU module are other WebGPU implementations used in production.