[DRAFT]: Experimental: Streaming DeviceSelect

[elstehle]

Description

Experimental branch, exploring the option of a streaming DeviceSelect that processes partitions of sizes of up to INT_MAX, making sure iterators are only advanced/offset on the device,

Checklist

• New or existing tests cover these changes.
• The documentation is up to date with these changes.

[mfbalin]

@elstehle Is it possible to do this for more algorithms? Doing it for DeviceCopy::Batched and DeviceMemcpy::Batched should be even easier. Then, the user wouldn't need manual loops like I have here:

https://github.com/dmlc/dgl/blob/b5ee45fd1a533a877de7ee787c43358828a86b21/graphbolt/src/cuda/expand_indptr.cu#L106-L113

[elstehle]

@elstehle Is it possible to do this for more algorithms? Doing it for DeviceCopy::Batched and DeviceMemcpy::Batched should be even easier. Then, the user wouldn't need manual loops like I have here:

Thank you for expressing interest in having support for larger-than-INT_MAX problems in DeviceMemcpy. This is definitely something that is on our agenda. At this point it is unclear whether the approach presented in this PR will be used for DeviceMemcpy but we will surely provide the solution that we think is best for each algorithm.

If you want to keep track of progress on algorithms other than DeviceSelect, please keep an eye on #50 that tracks the support for larger-than-INT_MAX problems in CUB algorithms.

[elstehle]

Closing this draft PR, since #2400 was merged in the meanwhile.