Get the underlying function calls for a function call and locate them in PyTorch code base

autograd
1

I want to see the following function calls happening for Tensor.to(device) function and how the moving of a Tensor is working from CPU to GPU or GPU to CPU. I used the following code to get an idea of function calls.

import torch

# Enable the profiler
with torch.autograd.profiler.profile() as prof:
    # Your code snippet involving Tensor.to() here
    tensor = torch.randn(3, 3).to(device="cuda")
    tensor = tensor.to(device="cpu")

# Print the profiler results
print(prof)

The output is

Screenshot (553)
Screenshot (553)1126×527 15 KB

So, for the Tensor.to(device) function, aten::to, atten::to_copy, cudaDeviceGetStreamPriorityRange, cudaMalloc, aten::copy functions are getting called.
I want to track these functions in PyTorch code base. I am new to PyTorch. So, I fail to locate and navigate the PyTorch code base. I do not know if there are any other tools to get more detailed information about classes that contain the functions. Any kind of help will be appreciated. Thanks in advance.

2

copy_ will call into copy_impl and dispatch to the stub which will eventually call into e.g. copy_device_to_device and there into the actual memcpy kernel.

3

Thanks for linking the class. Can you give some guidelines on how to navigate through the PyTorch source code and how to locate the functions in the source code?

4

I usually check the native_functions.yml first to check for dispatching into other internal methods. Once I’ve found the kernel name I grep -r for it in the source code repository and narrow down where it’s defined.