[caffe2] early return for empty indices in SLS (#35498)
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/35498
As title
Test Plan:
Need to run remote predictor canary
In SKL T6,
numactl -m 0 -C 3 ./sparse_lengths_sum_benchmark.par -d float -e 100000 --embedding-dim 1 --average-len 0 --batch-size 16 -i 1000000
Before this diff
0.000302733 ms. 100%. SparseLengthsSum
After this diff
0.000214509 ms. 100%. SparseLengthsSum
Reviewed By: jianyuh, ellie-wen
Differential Revision: D20678075
fbshipit-source-id: c0c8359036b82ffcbcc8b2a89dfb62db7f0a9c14
