Renumber utility function for sampling output

[seunghwak]

Implements a utility function to renumber sampling function (e.g. uniform_neighbor_sample) outputs based on DGL and PyG requirements.

closes #3718

[seunghwak]

Define the function API first. Several questions to be resolved.

1. Does multi-GPU implementation make sense? If we are processing only one batch, sampled output will be small and it will easily fit within a single GPU. If we are processing many batches, different batches will be independently processed in different GPUs. I don't see a practical multi-GPU use case. Any thoughts?
2. Say edgelist_srcs is [0, 5, 5] and edgelist_hops is [1, 0, 2]. Then, should we associate hop # 0 to vertex 5 or hop # 2 to vertex 5? (so [0, 5] should be renumbered to [1, 0] or [0,1]?). If vertex 5 is associated with hop # 0 & 2, this API assumed that vertex 5 is associated with the minimum hop # (i.e. 0). Any concerns with this?

Besides the questions above, do you have any other concerns?

@ChuckHastings @tingyu66 @VibhuJawa @alexbarghi-nv

[alexbarghi-nv]

Define the function API first. Several questions to be resolved.

1. Does multi-GPU implementation make sense? If we are processing only one batch, sampled output will be small and it will easily fit within a single GPU. If we are processing many batches, different batches will be independently processed in different GPUs. I don't see a practical multi-GPU use case. Any thoughts?
2. Say edgelist_srcs is [0, 5, 5] and edgelist_hops is [1, 0, 2]. Then, should we associate hop # 0 to vertex 5 or hop # 2 to vertex 5? (so [0, 5] should be renumbered to [1, 0] or [0,1]?). If vertex 5 is associated with hop # 0 & 2, this API assumed that vertex 5 is associated with the minimum hop # (i.e. 0). Any concerns with this?

Besides the questions above, do you have any other concerns?

@ChuckHastings @tingyu66 @VibhuJawa @alexbarghi-nv

For (1): We are already partitioning, sending all samples for a single batch to a single GPU. So single GPU implementation is ok. There should be no comms/transfer between GPUs. Each worker should just run this function independent of all other workers.

For (2): I think we always want the minimum hop. So this behavior is fine.

[alexbarghi-nv]

@seunghwak this API looks good to me 👍

[tingyu66]

Define the function API first. Several questions to be resolved.

1. Does multi-GPU implementation make sense? If we are processing only one batch, sampled output will be small and it will easily fit within a single GPU. If we are processing many batches, different batches will be independently processed in different GPUs. I don't see a practical multi-GPU use case. Any thoughts?
2. Say edgelist_srcs is [0, 5, 5] and edgelist_hops is [1, 0, 2]. Then, should we associate hop # 0 to vertex 5 or hop # 2 to vertex 5? (so [0, 5] should be renumbered to [1, 0] or [0,1]?). If vertex 5 is associated with hop # 0 & 2, this API assumed that vertex 5 is associated with the minimum hop # (i.e. 0). Any concerns with this?

Besides the questions above, do you have any other concerns?

@ChuckHastings @tingyu66 @VibhuJawa @alexbarghi-nv

2). Yes, a src vertex should be associated with its minimum hop, so [0, 5] should be renumbered to [1, 0] in that case.

[VibhuJawa]

2. Say edgelist_srcs is [0, 5, 5] and edgelist_hops is [1, 0, 2]. Then, should we associate hop # 0 to vertex 5 or hop # 2 to vertex 5? (so [0, 5] should be renumbered to [1, 0] or [0,1]?). If vertex 5 is associated with hop # 0 & 2, this API assumed that vertex 5 is associated with the minimum hop # (i.e. 0). Any concerns with this?

Yes, Agreed it should be associated with the minimum hop.

API looks good to me.

[seunghwak]

OK, I will drop the multi_gpu flag and implement this.

[alexbarghi-nv]

This looks good, just needs a C API. I assume @ChuckHastings is working on that?

[ChuckHastings]

Just 2023

[alexbarghi-nv]

Tested, working.

[ChuckHastings]

/merge