ONNX to WebGPU Compiler and Runtime

In order to learn more about GPU programming and Machine Learning compilers, I set out to build an ONNX to WebGPU compiler. This would materialize each ONNX operation as a GPU kernel on the fly. In my opinion, the most interesting part of this project are the register allocator which can reuse tensors during an evaluation of the model. This makes the memory used smaller than more naive strategies that were used at the time in projects like webonnx.