This document describes the wgpu API. It is basically a Pythonic version of the WebGPU API. It exposes an API for performing operations, such as rendering and computation, on a Graphics Processing Unit.

The WebGPU API is still being developed and occasionally there are backwards incompatible changes. Since we mostly follow the WebGPU API, there may be backwards incompatible changes to wgpu-py too. This will be so until the WebGPU API settles as a standard.

How to read this API

The classes in this API all have a name staring with “GPU”, this helps discern them from flags and enums. These classes are never instantiated directly; new objects are returned by certain methods.

Most methods in this API have no positional arguments; each argument must be referenced by name. Some argument values must be a dict, these can be thought of as “nested” arguments.

Many arguments (and dict fields) must be a flags or enums. Flags are integer bitmasks that can be orred together. Enum values are strings in this API.

Some arguments have a default value. Most do not.

Selecting the backend

Before you can use this API, you have to select a backend. Eventually there may be multiple backends, but at the moment there is only one backend, which is based on the Rust libary wgpu-native. You select the backend by importing it:

import wgpu.backends.rs

The wgpu-py package comes with the wgpu-native library. If you want to use your own version of that library instead, set the WGPU_LIB_PATH environment variable.

Differences from WebGPU

This API is derived from the WebGPU spec, but differs in a few ways. For example, methods that in WebGPU accept a descriptor/struct/dict, here accept the fields in that struct as keyword arguments.

wgpu.base.apidiff Differences of base API:
  • Adds GPUAdapter.properties() - useful for desktop
  • Adds GPUBuffer.map_read() - Alternative to mapping API
  • Adds GPUBuffer.map_write() - Alternative to mapping API
  • Adds GPUBuffer.size() - Too useful to not-have
  • Adds GPUBuffer.usage() - Too useful to not-have
  • Adds GPUCanvasContext.present() - Present method is exposed
  • Adds GPUDevice.adapter() - Too useful to not-have
  • Adds GPUDevice.create_buffer_with_data() - replaces WebGPU’s mapping API
  • Adds GPUQueue.read_buffer() - replaces WebGPU’s mapping API
  • Adds GPUQueue.read_texture() - For symmetry, and to help work around the bytes_per_row constraint
  • Adds GPUTexture.dimension() - Too useful to not-have
  • Adds GPUTexture.format() - Too useful to not-have
  • Adds GPUTexture.mip_level_count() - Too useful to not-have
  • Adds GPUTexture.sample_count() - Too useful to not-have
  • Adds GPUTexture.size() - Too useful to not-have
  • Adds GPUTexture.usage() - Too useful to not-have
  • Adds GPUTextureView.size() - Too useful to not-have
  • Adds GPUTextureView.texture() - Too useful to not-have
  • Changes GPU.request_adapter() - arguments include a canvas object
  • Changes GPU.request_adapter_async() - arguments include a canvas object
  • Hides GPUBuffer.get_mapped_range()
  • Hides GPUBuffer.map_async()
  • Hides GPUBuffer.unmap()
  • Hides GPUDevice.import_external_texture() - Specific to browsers.
  • Hides GPUDevice.pop_error_scope()
  • Hides GPUDevice.push_error_scope()
  • Hides GPUQueue.copy_external_image_to_texture() - Specific to browsers.

Each backend may also implement minor differences (usually additions) from the base API. For the rs backend check print(wgpu.backends.rs.apidiff.__doc__).


Adapter, device and canvas

The GPU represents the root namespace that contains the entrypoint to request an adapter.

The GPUAdapter represents a hardware or software device, with specific features, limits and properties. To actually start using that harware for computations or rendering, a GPUDevice object must be requisted from the adapter. This is a logical unit to control your hardware (or software). The device is the central object; most other GPU objects are created from it. Also see the convenience function wgpu.utils.get_default_device().

A device is controlled with a specific backend API. By default one is selected automatically. This can be overridden by setting the WGPU_BACKEND_TYPE environment variable to “Vulkan”, “Metal”, “D3D12”, “D3D11”, or “OpenGL”.

The device and all objects created from it inherit from GPUObjectBase - they represent something on the GPU.

In most render use-cases you want the result to be presented to a canvas on the screen. The GPUCanvasContext is the bridge between wgpu and the underlying GUI backend.

Buffers and textures

A GPUBuffer can be created from a device. It is used to hold data, that can be uploaded using it’s API. From the shader’s point of view, the buffer can be accessed as a typed array.

A GPUTexture is similar to a buffer, but has some image-specific features. A texture can be 1D, 2D or 3D, can have multiple levels of detail (i.e. lod or mipmaps). The texture itself represents the raw data, you can create one or more GPUTextureView objects for it, that can be attached to a shader.

To let a shader sample from a texture, you also need a GPUSampler that defines the filtering and sampling behavior beyond the edges.

WebGPU also defines the GPUExternalTexture, but this is not (yet?) used in wgpu-py.

Bind groups

Shaders need access to resources like buffers, texture views, and samplers. The access to these resources occurs via so called bindings. There are integer slots, which must be specifie both via the API, and in the shader.

Bindings are organized into GPUBindGroup s, which are essentially a list of GPUBinding s.

Further, in wgpu you need to specify a GPUBindGroupLayout, providing meta-information about the binding (type, texture dimension etc.).

Multiple bind groups layouts are collected in a GPUPipelineLayout, which represents a complete layout description for a pipeline.

Shaders and pipelines

The wgpu API knows three kinds of shaders: compute, vertex and fragment. Pipelines define how the shader is run, and with what resources.

Shaders are represented by a GPUShaderModule.

Compute shaders are combined with a pipelinelayout into a GPUComputePipeline. Similarly, a vertex and (optional) fragment shader are combined with a pipelinelayout into a GPURenderPipeline. Both of these inherit from GPUPipelineBase.

Command buffers and encoders

The actual rendering occurs by recording a series of commands and then submitting these commands.

The root object to generate commands with is the GPUCommandEncoder. This class inherits from GPUCommandsMixin (because it generates commands), and GPUDebugCommandsMixin (because it supports debugging).

Commands specific to compute and rendering are generated with a GPUComputePassEncoder and GPURenderPassEncoder respectively. You get these from the command encoder by the corresponding begin_x_pass() method. These pass encoders inherit from GPUBindingCommandsMixin (because you associate a pipeline) and the latter also from GPURenderCommandsMixin.

When you’re done generating commands, you call finish() and get the list of commands as an opaque object: the GPUCommandBuffer. You don’t really use this object except for submitting it to the GPUQueue.

The command buffers are one-time use. The GPURenderBundle and GPURenderBundleEncoder can be used to record commands to be used multiple times, but this is not yet implememted in wgpu-py.

Error handling

Errors are caught and logged using the wgpu logger.

Todo: document the role of these classes: GPUUncapturedErrorEvent GPUValidationError GPUOutOfMemoryError GPUDeviceLostInfo


These classes are not supported and/or documented yet. GPUCompilationMessage GPUCompilationInfo GPUQuerySet