[MXNET-938] RNN.hybrid_forward() spends a significant time on _rnn_param_concat

• Type: Improvement
• Status: To Do
• Priority: Major
• Resolution: Unresolved
• Component/s: Gluon
• Assignee: Unassigned
• Reporter: Gautham Kollu
• Created: 10/Sep/18 23:53
• Updated: 13/Nov/18 19:00

Description

Issue :
RNN.hybrid_forward() spends a significant time on _rnn_param_concat

Root Cause:
Creating a simple LSTM (256 -> 256) with gluon (gluon_lstm.py) and running inference on CPU with the profile shows the following output.  { Refer gluon_lstm.py for the code}

_rnn_param_concat() takes more time than the RNN operator.

 

-------------------	-------	--------	-----------		------------		------------		------------
operator
=================
Name	Total	Count	Time (ms)	Min	Time (ms)	Max	Time (ms)	Avg	Time (ms)
		------	---------	—	----------	—	----------	—	----------
RNN		72	72.312		0.065		33.15		1.0043
_zeros		788	1.76		0.001		0.011		0.0022
DeleteVariable		1110	0.637		0		0.002		0.0006
SwapAxis		462	2.223		0.003		0.015		0.0048
_rnn_param_concat		76	103.813		1.122		2.08		1.366

Issue Links

• links to: GitHub Pull Request #12512