README

HelloPytorch is a simple application that demonstrates how to build a PyTorch wrapper around Halide ops using the internal Generator target.

The demo op takes two 4D tensors as input, a and b and outputs their pointwise sum.

The application builds float32 and float64 versions of the op for CPU and CUDA (provided that a GPU is present and the nvcc compiler is installed and in the path, see Makefile).

The build proceeds in four steps:

1. Build the Halide generator
2. Build Halide code for each each datatype and each architecture (-e static_library,h flag in the generator).
3. Build the PyTorch C++ interface code (-e pytorch_wrapper flag in the generator). Note that the PyTorch wrapper requires the user_context feature be enabled in the generator target for CUDA ops. This allows Halide's and PyTorch's GPU memory managers to communicate with each other.
4. Synthesize and compile a Python extension that links togethers the various operator libraries and exposes them to Python (see setup.py).

Building only requires Python 3 and PyTorch. Please follow these instructions to install the latest PyTorch: https://pytorch.org/

If everything is setup correctly, running make test should build the PyTorch extension and runs a simple test (test.py).

Tested with Python 3.7.4 and PyTorch 1.2.0