Intel MKL-DNN RNN Support

[sbodenstein]

MKL-DNN recently got support for RNNs. Are there any plans for adding this as a backend to mx.sym.RNN, and what sort of timeframe? Thanks!

@zheng-da

[TaoLv]

@sbodenstein Thanks for asking that. It's already on our plan list for Q2. Since it's still an experimental interface, we will keep tracking the development progress from mkldnn team.
FYI, we are working intensively on PR #10104 and #10311 to provide fused RNN operator for mxnet before mkldnn RNN interfaces become mature. Hopefully, after mkldnn RNN is integrated into mxnet, code of #10104 and #10311 would still exsit as a reference implementation for cpu. @pengzhao-intel

[Roshrini]

@eric-haibin-lin Can you please add labels-
FeatureRequest, RNN

[azai91]

@sbodenstein we plan on implementing MKL-DNN RNN support when MKLDNN 0.15 is released which should be within the next few weeks.

[sbodenstein]

@azai91: fantastic, thanks so much!

[neverjoe]

Any good news for mkl-dnn rnn support?

[pengzhao-intel]

@neverjoe thanks for the asking.
MKL-DNN provides the RNN API now in the new latest (v0.17) and we have upgraded it to MXNet now.
We have the discussion with @szha for the further integration.

Do you mind providing more details about what's your models and feature requests for MKL-DNN?

[neverjoe]

It's happy to hear that. I am looking for the fastest gru inference implementation. Thanks. PatricZhao <[email protected]>于2018年11月13日周二 上午10:53写道：

@neverjoe <https://github.com/neverjoe> thanks for the asking. MKL-DNN provides the RNN API now in the new latest (v0.17) and we have upgraded it to MXNet now. We have the discussion with @szha <https://github.com/szha> for the further integration. Do you mind providing more details about what's your models and feature requests for MKL-DNN? — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#10542 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ABsFmIvmhMNnihdE3jxm03CgcWDeUEtwks5uujQ1gaJpZM4TTi1_> .

-- Sent by Inbox

[neverjoe]

Is there any python test demo for GruForwardInference?

[pengzhao-intel]

There's no python interface, but you can test by mkldnn binary very simple.
https://github.com/intel/mkl-dnn/tree/master/tests/benchdnn

oneapi-src/oneDNN@3ced2c7

./benchdnn --rnn --batch=./inputs/rnn/test_rnn_small

In the test file, it includes several configurations:

--reset --alg=VANILLA_GRU
--direction=left2right
--activation=TANH
--prop=FWD_D --batch=rnn_gru_small
--prop=BWD_DW --batch=rnn_gru_small

[neverjoe]

Very Thanks! PatricZhao <[email protected]>于2018年11月13日周二 下午4:07写道：

There's no python interface, but you can test by mkldnn binary very simple. https://github.com/intel/mkl-dnn/tree/master/tests/benchdnn ***@***.*** <oneapi-src/oneDNN@3ced2c7> ./benchdnn --rnn --batch=./inputs/rnn/test_rnn_small In the test file, it includes several configurations: --reset --alg=VANILLA_GRU --direction=left2right --activation=TANH --prop=FWD_D --batch=rnn_gru_small --prop=BWD_DW --batch=rnn_gru_small — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#10542 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ABsFmNGvIFo7dx4mGqcUno5hPz6KOYzYks5uun24gaJpZM4TTi1_> .

-- Sent by Inbox

[anirudh2290]

@pengzhao-intel Can you point me to the further discussion for MKLDNN RNN API Integration.

[pengzhao-intel]

@anirudh2290 yes, we have the plan but a little complicated mixed the next major release of MXNet 2.0 and MKLDNN 1.0.

We can talk for more details by email, please ping me [email protected]

[bgawrych]

@szha Can we close this issue? MXNet already have support for oneDNN RNN's

[szha]

@bgawrych thanks!