sean-brock / muten

muten

A small tensor library. Supports HIP and CPU backend.

Structure

Mutiple types, one for each backend

Tensor:

- http://blog.ezyang.com/2019/05/pytorch-internals/
- Created on CPU or device (optionally, in constructor(device=CPU))
- to(device) to transfer to a different compute context
    - tensor_gpu = tensor_cpu.to(hip_context)
- All operations on tensors must be on the same context
- Operations:
    - tensor = muten::add(tensor, tensor)
    - Depending on device, use device specific impl

Autograd:

- Required to handle backwards passes for NN
- https://en.wikipedia.org/wiki/Automatic_differentiation