This is an example of a CUDA extension for PyTorch which computes the Hadamard product of two tensors.
For the Torch version of this example extension, please see: https://github.com/sniklaus/torch-extension
For a more advanced PyTorch extension that uses CuPy, please see: https://github.com/szagoruyko/pyinn
To build the extension, run bash install.bash and make sure that the CUDA_HOME environment variable is set. Should you receive an error message regarding an invalid device function when making use of the extension, configure the CUDA architecture within install.bash to something your graphics card supports.
The initial version of this extension only contained a CFFI example. Since then, HadamardProduct_cupy.py has been added which makes use of CuPy to perform the given task. This version of the extension does not need to be built in advance and the setup section does not apply to it, CuPy will take care of the compilation during runtime. This repository thus serves as a reference for implementing the Hadamard product once using CFFI and once using CuPy instead.
After successfully building the extension, run python test.py to test it. A minimal example of how the sample extension can be used is also shown below. To test the CuPy version, run python test_cupy.py instead.
import torch
from HadamardProduct import HadamardProduct
class Network(torch.nn.Module):
def __init__(self):
super(Network, self).__init__()
# end
def forward(self, input1, input2):
return HadamardProduct()(input1, input2)
# end
# end
net = Network().cuda()
input1 = torch.rand(64, 3, 128, 128).cuda()
input2 = torch.rand(64, 3, 128, 128).cuda()
input1 = torch.autograd.Variable(input1, requires_grad=True)
input2 = torch.autograd.Variable(input2, requires_grad=True)
output = net(input1, input2)
expected = torch.mul(input1, input2)
print(torch.sum(output.data - expected.data), '<-- should be 0.0')Please refer to the appropriate file within this repository.