[Feature Request] tensor.data_ptr()

[hiyyg]

like that in torch, so that we can access the cuda memory address of a tensor in python.

[MARD1NO]

Can you give us an example to use this feature in PyTorch? I wonder in which situation you need to use tensor.data_ptr()

[hiyyg]

@MARD1NO For example:

https://github.com/NVlabs/DeepIM-PyTorch/blob/b46ccd2465ce69ac575bf454d4171a5dcb9c6908/ycb_render/ycb_renderer.py#L506

[lixinqi]

@MARD1NO For example:

https://github.com/NVlabs/DeepIM-PyTorch/blob/b46ccd2465ce69ac575bf454d4171a5dcb9c6908/ycb_render/ycb_renderer.py#L506

It seems like pytorch has no api called map_tensor. Did it implement by your team?

[hiyyg]

@MARD1NO For example: https://github.com/NVlabs/DeepIM-PyTorch/blob/b46ccd2465ce69ac575bf454d4171a5dcb9c6908/ycb_render/ycb_renderer.py#L506

It seems like pytorch has no api called map_tensor. Did it implement by your team?

@lixinqi map_tensor is not related to pytorch. It just need a pointer like tensor.data_ptr() in pytorch. But if oneflow can also give users such a pointer, it can be changed to oneflow.

[lixinqi]

@hiyyg Oneflow doesn't provide api tensor.data_ptr() right now for two reasons:

1. The fundemental of design is very different from pytorch's. Memory of pytorch tensor are synchronously allocated in the same thread running python code, While memory of oneflow tensor are asynchronously allocated in a dependent worker thread. It's illegal to directly accesss memory of tensor in main thread.
2. oneflow tensor can represent distributed tensor across nodes. For example t = oneflow.ones((32, 32), placement=flow.placement("cuda", {0:[0, 1]}), sbp=flow.sbp.split(0)), t is a distributed tensor (or consistent tensor) , the data of t[0:16, :] is located on device cuda 0, and the data of t[16:, :] is located on device cuda 1. Hence the behavior of t.data_ptr() is unable to be defined.

We will discuss this feature request in the near future.

[hiyyg]

OK. Really hope that oneday oneflow can TRULY replace pytorch.