Embedding gradient performance optimization on GPU

[MoisesHer]

Description

This PR includes a specific Embedding-backward operator for GPU.
Two new CUDA kernels have been implemented for improving the performance of the operator when using GPU.
According to our measurements on Volta GPUs, the previous version was taken 2.2ms,
whereas the new implementation takes 0.3ms, i.e. more than 7x speedup.

Checklist

Essentials

• [X ] Changes are complete (i.e. I finished coding on this PR)
• All changes have test coverage:
• Unit tests already existed (tests/python/gpu/test_operator_gpu:test_embedding_with_type) and changes do not affect correctness.
• Code is well-documented:
• For new C++ functions in header files, their functionalities and arguments are documented.

Changes

• [x ] Embedding-backward operator for GPU, test: tests/python/gpu/test_operator_gpu:test_embedding_with_type

[ptrendx]

@sxjscience FYI

[ptrendx]

LGTM

[sxjscience]

Nice! LGTM. So the BinarySearch version of FindBounds has complexity O(|V| log |N|) where |V| is the vocabulary size and |N| is the number of indices. I guess our initial version (https://github.com/dmlc/mshadow/blob/bc49327a44650c3f2b427e953ff95d2c27566c04/mshadow/cuda/tensor_gpu-inl.cuh#L619-L672) has complexity O(|N|) for finding the boundaries. Thus, in some workloads (in which |N| is small), the O(N) version might be faster.

[ptrendx]

Moises did a performance comparison between the new version and both the old one and the old buggy one. The new kernel is faster than the old working version in all cases and ~same in speed as the buggy one. The biggest performance change is seen actually when changing how many different elements are in the input data (as small number of distinct elements limits parallelism in the backward pass).