Sparse int64 Large tensor support

[ChaiBapchya]

Description

Adds support to Sparse tensor creation

Checklist

Essentials

Please feel free to remove inapplicable items for your PR.

• Changes are complete (i.e. I finished coding on this PR)
• All changes have test coverage:
• Code is well-documented:
• To the my best knowledge, examples are either not affected by this change, or have been fixed to be compatible with this change

Changes

• src/c_api/c_api.cc
• python/mxnet/ndarray/sparse.py
• tests/nightly/test_large_array.py

Test

python -m nose tests/nightly/test_large_array.py:test_sparse_dot
.
----------------------------------------------------------------------
Ran 1 test in 3.688s

OK

[eric-haibin-lin]

Is there a check list of what to change for int64 support in general?

[ChaiBapchya]

@apeforest
Moreover, I have seen
32bit API has both examples uint32_t and int

MXNDArrayCreateEx
https://github.com/apache/incubator-mxnet/blob/6f95702613dd95aa4a1b467b938d79c4d13bee96/include/mxnet/c_api.h#L599

MXNDArrayReshape
https://github.com/apache/incubator-mxnet/blob/6f95702613dd95aa4a1b467b938d79c4d13bee96/include/mxnet/c_api.h#L845

Should I update them so that ndims for 32bit API are consistent?

[ChaiBapchya]

@eric-haibin-lin while I don't think there is uniform guideline for int64 support across all ops. But in most cases, switching the datatype of all the variables holding "indices" from 32bit dtype to 64bit dtype solves the issue. One still has to test that to ensure it doesn't break the functionality and supports int64.

@apeforest @access2rohit any inputs?

[access2rohit]

@eric-haibin-lin while I don't think there is uniform guideline for int64 support across all ops. But in most cases, switching the datatype of all the variables holding "indices" from 32bit dtype to 64bit dtype solves the issue. One still has to test that to ensure it doesn't break the functionality and supports int64.

@apeforest @access2rohit any inputs?

Plus ensure that the C API being used supports int64_t for indices as well. On language binding side int64_t values need to be passed to C++ backend. For e.g. ctype.int64 should be used for passing large values else they will get truncated at the language binding interface itself.

[apeforest]

Adding @reminisce to review the C API change for numpy operator.

[apeforest]

One last comment. Otherwise looks good to me. Thanks!

[eric-haibin-lin]

Does https://cwiki.apache.org/confluence/display/MXNET/Large+Tensor+Support correctly document the current practice/changes?

[ChaiBapchya]

Does https://cwiki.apache.org/confluence/display/MXNET/Large+Tensor+Support correctly document the current practice/changes?

Yes. It enlists the 3 steps needed to extend LTS for operators (includes changes on Front-end language binding side and Back-end C++ api side). I updated the wiki. Please verify. @eric-haibin-lin

[apeforest]

LGTM

[ChaiBapchya]

@mxnet-label-bot add [pr-awaiting-merge]