[Python/C++] Support data copying between different GPU devices

[asfimport]

Currently, the data copying is supported from host to device, from device to host, from device to the same device. For multiple GPU systems, copying data from one device to another is needed.

See also
#2844 (comment)

Reporter: Pearu Peterson / @pearu
Assignee: Pearu Peterson / @pearu

Related issues:

• [C++] GPU support library development (is part of)

PRs and other links:

• GitHub Pull Request #3617

_{Note: This issue was originally created as ARROW-3653. Please see the migration documentation for further details.}

[asfimport]

Wes McKinney / @wesm:
Would you like to submit a patch for this?

[asfimport]

Pearu Peterson / @pearu:
Yes, I can submit a patch. However, I don't have access to a multi-GPU system. So, I would not be able to fully test it.

[asfimport]

Wes McKinney / @wesm:
I can create a user account for you on one of our multi-GPU machines. I think you should be able to test this without sudo?

[asfimport]

Pearu Peterson / @pearu:
Sounds good. As long as the nvidia driver is working and there is enough disk space to create a development environment, sudo access would not be needed. Let me know if you need anything from me for creating the account.

[asfimport]

Uwe Korn / @xhochy:
Issue resolved by pull request 3617
#3617