Guide

The wgpu library presents a Pythonic API for the WebGPU spec. It is an API to control graphics hardware. Like OpenGL but modern. Or like Vulkan but higher level. GPU programming is a craft that requires knowledge of how GPU’s work.

Getting started

Creating a canvas

If you want to render to the screen, you need a canvas. Multiple GUI toolkits are supported, see the gui API. In general, it’s easiest to let wgpu select a GUI automatically:

from wgpu.gui.auto import WgpuCanvas, run

canvas = WgpuCanvas(title="a wgpu example")

Next, we can setup the render context, which we will need later on.

present_context = canvas.get_context("wgpu")
render_texture_format = present_context.get_preferred_format(device.adapter)
present_context.configure(device=device, format=render_texture_format)

Obtaining a device

The next step is to obtain an adapter, which represents an abstract render device. You can pass it the canvas that you just created, or pass None for the canvas if you have none (e.g. for compute or offscreen rendering). From the adapter, you can obtain a device.

adapter = wgpu.gpu.request_adapter_sync(power_preference="high-performance")
device = adapter.request_device_sync()

The wgpu.gpu object is the API entrypoint (wgpu.GPU). It contains just a handful of functions, including request_adapter(). The device is used to create most other GPU objects.

Creating buffers, textures shaders, etc.

Using the device, you can create buffers, textures, write shader code, and put these together into pipeline objects. How to do this depends a lot on what you want to achieve, and is therefore out of scope for this guide. Have a look at the examples or some of the tutorials that we link to below.

Setting up a draw function

Let’s now define a function that will actually draw the stuff we put together in the previous step.

def draw_frame():

    # We'll record commands that we do on a render pass object
    command_encoder = device.create_command_encoder()
    current_texture_view = present_context.get_current_texture()
    render_pass = command_encoder.begin_render_pass(
        color_attachments=[
            {
                "view": current_texture_view,
                "resolve_target": None,
                "clear_value": (1, 1, 1, 1),
                "load_op": wgpu.LoadOp.clear,
                "store_op": wgpu.StoreOp.store,
            }
        ],
    )

    # Perform commands, something like ...
    render_pass.set_pipeline(...)
    render_pass.set_index_buffer(...)
    render_pass.set_vertex_buffer(...)
    render_pass.set_bind_group(...)
    render_pass.draw_indexed(...)

    # When done, submit the commands to the device queue.
    render_pass.end()
    device.queue.submit([command_encoder.finish()])

    # If you want to draw continuously, request a new draw right now
    canvas.request_draw()

Starting the event loop

We can now pass the above render function to the canvas. The canvas will then call the function whenever it (re)draws the window. And finally, we call run() to enter the mainloop.

canvas.request_draw(draw_frame)
run()

Offscreen

If you render offscreen, or only do compute, you do not need a canvas. You also won’t need a GUI toolkit, draw function or enter the event loop. Instead, you will obtain a command encoder and submit its records to the queue directly.

Examples and external resources

Examples that show wgpu-py in action:

https://github.com/pygfx/wgpu-py/tree/main/examples

Note

The examples in the main branch of the repository may not match the pip installable version. Be sure to refer to the examples from the git tag that matches the version of wgpu you have installed.

External resources:

A brief history of WebGPU

For years, OpenGL has been the only cross-platform API to talk to the GPU. But over time OpenGL has grown into an inconsistent and complex API …

OpenGL is dying — Dzmitry Malyshau at Fosdem 2020

In recent years, modern API’s have emerged that solve many of OpenGL’s problems. You may have heard of Vulkan, Metal, and DX12. These API’s are much closer to the hardware, which makes the drivers more consistent and reliable. Unfortunately, the huge amount of “knobs to turn” also makes them quite hard to work with for developers.

Therefore, higher level API are needed, which use the same concepts, but are much easier to work with. The most notable one is the WebGPU specification. This is what future devs will be using to write GPU code for the browser. And for desktop and mobile as well.

As the WebGPU spec is being developed, a reference implementation is also build. It’s written in Rust and powers the WebGPU implementation in Firefox. This reference implementation, called wgpu, also exposes a C-api (via wgpu-native), so that it can be wrapped in Python. And this is precisely what wgpu-py does.

So in short, wgpu-py is a Python wrapper of wgpu, which is an desktop implementation of WebGPU, an API that wraps Vulkan, Metal and DX12, which talk to the GPU hardware.

Coordinate system

In wgpu, the Y-axis is up in normalized device coordinate (NDC): point(-1.0, -1.0) in NDC is located at the bottom-left corner of NDC. In addition, x and y in NDC should be between -1.0 and 1.0 inclusive, while z in NDC should be between 0.0 and 1.0 inclusive. Vertices out of this range in NDC will not introduce any errors, but they will be clipped.

Array data

The wgpu library makes no assumptions about how you store your data. In places where you provide data to the API, it can consume any data that supports the buffer protocol, which includes bytes, bytearray, memoryview, ctypes arrays, and numpy arrays.

In places where data is returned, the API returns a memoryview object. These objects provide a quite versatile view on ndarray data:

# One could, for instance read the content of a buffer
m = device.queue.read_buffer(buffer)
# Cast it to float32
m = m.cast("f")
# Index it
m[0]
# Show the content
print(m.tolist())

Chances are that you prefer Numpy. Converting the memoryview to a numpy array (without copying the data) is easy:

array = np.frombuffer(m, np.float32)

Debugging

If the default wgpu-backend causes issues, or if you want to run on a different backend for another reason, you can set the WGPU_BACKEND_TYPE environment variable to “Vulkan”, “Metal”, “D3D12”, or “OpenGL”.

The log messages produced (by Rust) in wgpu-native are captured and injected into Python’s “wgpu” logger. One can set the log level to “INFO” or even “DEBUG” to get detailed logging information.

Many GPU objects can be given a string label. This label will be used in Rust validation errors, and are also used in e.g. RenderDoc to identify objects. Additionally, you can insert debug markers at the render/compute pass object, which will then show up in RenderDoc.

Eventually, wgpu-native will fully validate API input. Until then, it may be worthwhile to enable the Vulkan validation layers. To do so, run a debug build of wgpu-native and make sure that the Lunar Vulkan SDK is installed.

You can run your application via RenderDoc, which is able to capture a frame, including all API calls, objects and the complete pipeline state, and display all of that information within a nice UI.

You can use adapter.request_device_sync() to provide a directory path where a trace of all API calls will be written. This trace can then be used to re-play your use-case elsewhere (it’s cross-platform).

Also see wgpu-core’s section on debugging: https://github.com/gfx-rs/wgpu/wiki/Debugging-wgpu-Applications

Freezing apps

In wgpu a PyInstaller-hook is provided to help simplify the freezing process (it e.g. ensures that the wgpu-native DLL is included). This hook requires PyInstaller version 4+.

Our hook also includes glfw when it is available, so code using wgpu.gui.auto should Just Work.

Note that PyInstaller needs wgpu to be installed in site-packages for the hook to work (i.e. it seems not to work with a pip -e . dev install).