[Chromium] Implement MKL-DNN backend - Githubissues

intel / webml-polyfill

Deprecated, the Web Neural Network Polyfill project has been moved to https://github.com/webmachinelearning/webnn-polyfill

Apache License 2.0

161 stars 42 forks source link

[Chromium] Implement MKL-DNN backend #6

Closed BruceDai closed 5 years ago

huningxin commented 6 years ago

https://github.com/intel/mkl-dnn

huningxin commented 5 years ago

WebNN POC ops to mkl-dnn ops mapping

Op Type	mkl-dnn
ADD	Sum
AVERAGE_POOL_2D	Pooling with mkldnn_pooling_max
CONCATENATION	Concat
CONV_2D	Convolution
DEPTHWISE_CONV_2D	Convolution with grouped weights
MAX_POOL_2D	Pooling with mkldnn_pooling_avg
MUL	Not supported https://github.com/intel/mkl-dnn/issues/255
RESHAPE	Reorder
SOFTMAX	Softmax
FULLY_CONNECTED	Inner Product

huningxin commented 5 years ago

TODOs:

[x] refactory model and compilation impl to accomodate both clDNN and mkl-dnn
[x] refactory clDNN compilation impl
[x] add mkl-dnn compilation impl
[x] add mkl-dnn execution impl
[x] implement ops with mkl-dnn

huningxin commented 5 years ago

https://github.com/otcshare/chromium-src/pull/65 for code refactoring has been merged. The mkl-dnn support is under development. Move the milestone to January.

huningxin commented 5 years ago

Done mkl-dnn compilation and execution impls. Will implement CONV_2D in first PR and create issues to track other ops.

huningxin commented 5 years ago

Created issues:

DEPTHWISE_CONV_2D: https://github.com/intel/webml-polyfill/issues/453
ATROUS_CONV_2D and ATROUS_DEPTHWISE_CONV_2D: https://github.com/intel/webml-polyfill/issues/454
AVERAGE_POOL_2D and MAX_POOL_2D: https://github.com/intel/webml-polyfill/issues/455
ADD: https://github.com/intel/webml-polyfill/issues/456
CONCATENATION: https://github.com/intel/webml-polyfill/issues/457
SOFTMAX: https://github.com/intel/webml-polyfill/issues/458
RESHAPE: https://github.com/intel/webml-polyfill/issues/459
FULLY_CONNECTED: https://github.com/intel/webml-polyfill/issues/460
RESIZE_BILINEAR: https://github.com/intel/webml-polyfill/issues/461

huningxin commented 5 years ago

@BruceDai , MKLDNN backend merged. Please help verify the basic functionalities and CONV_2D op on both Windows and Linux. Thanks!