Optimizing the Performance of Decoder fl_asr_decode

[mtmd]

Original Issue: #707

Summary

This pull request improves the performance of the decoder (fl_asr_decode), targeting the transformer model, by offering the following optimizations:

1. Adding support for AMP in the AM.
2. Adding support for the batch processing in the AM.
3. Offering an efficient CUDA kernel for PositionalEmbedding.

Test Plan (required)

Using the following flag file, the average decoding time on V100 can be improved by more that 4X using the optimizations that this pull request offers. The custom kernel for positional embedding may also improve the training performance. However, the impact is not measured.

--am=/path/to/am_transformer_ctc_stride3_letters_300Mparams.bin
--tokens=/path/to/tokens.txt
--lexicon=/path/to/lexicon.txt
--lm=/path/to/lm_common_crawl_small_4gram_prun0-6-15_200kvocab.bin
--datadir=/path/to/lists
--test=test-other.lst
--uselexicon=true
--decodertype=wrd
--lmweight=2
--wordscore=0
--beamsize=50
--beamthreshold=100
--nthread_decoder=16
--smearing=max
--show=false
--lmtype=kenlm
--lm_vocab=/path/to/lm_common_crawl_200kvocab.txt
--batchsize=24
--fl_optim_mode=O1
--fl_amp_use_mixed_precision=true