Modifying segments of data in at::tensor object by reference

1

I have a tensor given as:

auto my_tensor = at::zeros({5,5}, at::kCUDA)

and another function that takes a tensor as an argument:

at::tensor my_func(at::Tensor t)
{
    // some operation changing the data in t
}

I want to pull call my_func on subsets of tensor in a loop.

Currently I’ve tried

const int stride = 5;
for (int i = 0; i < 5; i++)
{
    my_func(my_tensor.data() + i*stride)
}

But tensor.data() + 5 gives me a pointer–I need the at::tensor. I can’t just dereference the pointer though, because it’s a pointer to the raw data, not to the at::tensor object.

My goal is to update my_tensor in-place so I don’t have to do unnecessary copying.

How can I do this? It seems perhaps something along the lines of a TensorAccessor? I went to try it, though, and it told me packed_accessor() didn’t exist (Pytorch 0.4.1).

2

Please pass by reference
i.e.

at::tensor my_func(at::Tensor &t)
{
    // some operation changing the data in t
}
3

Thanks for the reply. What I’m actually trying to do is modify slices of the tensor in-place. I’m not concerned necessarily about how to pass the tensor itself as an argument, but rather just a slice of it.

Also, I don’t think passing by reference quite makes a difference in this case, as at::tensor just wraps a pointer to the data, not the data itself. In passing by reference, you’re just passing a reference to a pointer.

4

Oh! You are right.
Let me get back to you