torch-extension
This is an example of a CUDA extension for Torch which computes the Hadamard product of two tensors.
For the PyTorch version of this example extension, please see: https://github.com/sniklaus/pytorch-extension
setup
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 CMakeLists.txt to something your graphics card supports.
usage
After successfully building the extension, run th test.lua to test it. A minimal example of how the sample extension can be used is also shown below.
require 'torch'
require 'nn'
require 'cunn'
require 'nnex'
torch.setdefaulttensortype('torch.FloatTensor')
local net = nn.HadamardProduct():cuda()
local input1 = torch.rand(64, 3, 128, 128):cuda()
local input2 = torch.rand(64, 3, 128, 128):cuda()
local output = net:forward({ input1, input2 })
local expected = torch.cmul(input1, input2)
print(torch.sum(output - expected), '<-- should be 0.0')license
Please refer to the appropriate file within this repository.