dawn-python

ctypeslib based python bindings to the Dawn WebGPU engine with a focus on the compute pipeline. Besides the autogenerated bindings, a small utils interface (similar to wgpu-py) is provided to make it easier to run compute.

Chromium version used in pydawn: 133.0.6940.0

How to use

1. Install from PyPI

pip install dawn-python

2. Import utils, and webgpu

from pydawn import utils, webgpu

if __name__ == "__main__":
    # Creating an adapter
    adapter = utils.request_adapter_sync(power_preference=webgpu.WGPUPowerPreference_HighPerformance)

    # Creating a device with a list of required features
    dev = utils.request_device_sync(adapter, [webgpu.WGPUFeatureName_ShaderF16, webgpu.WGPUFeatureName_TimestampQuery])

For more details see examples/query_set.py

Generating bindings

1. Build dawn or use the pre-built lib in the repo
2. pip install ctypeslib2
3. clang2py -l libwebgpu_dawn.dylib webgpu.h --nm nm_patch.py -o webgpu.py

Examples

python3 examples/simple_compute.py

Projects using pydawn

• tinygrad - tinygrad is a deep learning framework, and in its WebGPU backend it uses the Dawn binaries from pydawn. It uses the same autogened python around the binaries as pydawn.

Experimental

The experimental folders in pydawn, pydawn/lib, and examples contain code related to experimental Dawn builds. Currently pydawn experimental is used to test the new chromium_experimental_subgroup_matrix feature.

Supported platforms

Platform	Architectures	Backend
macOS	x86_64, arm64	Metal
Ubuntu	x86_64, arm64	Vulkan
Windows	x86_64	DirectX

License

BSD 3-Clause, please see LICENSE