This repo is dedicated to improving Theano performance on CPU, especially in Intel® Xeon® and Intel® Xeon Phi™ processors.

Key Features

• New backend of Intel® MKL (version >= 2017.0 which includes neural network primitives)
• Further graph optimizations
• CPU friendly OPs
• Switch to Intel® MKL backend automatically in Intel® Architecture
• Out-of-box performance improvements and good portability

Benchmark

• Hardwares
  • Intel® Xeon® CPU E5-2699 v4 @ 2.20GHz, 128GB RAM
  • Intel® Xeon Phi™ CPU 7250F @ 1.40GHz, 98G RAM
• Softwares
  • Script: convnet-benchmarks
  • Stock Theano, commit ID: 2fa3cec
  • Intel Theano, commit ID: 00216ef, ver-1.1
• Terminologies
  • FWD, forward for inference
  • FWD+BWD, forward and backward for training
• Results

--

Performance Tips

• Add bias after convolution to archieve high performance since this sub-graph can be replaced with MKL Op

• Use group convolution OP, AbstractConvGroup

• Use New MKL OP: LRN

• Intel® Xeon Phi™ Environment Setting, example as below

  #!/bin/sh
  export KMP_BLOCKTIME=1
  export KMP_AFFINITY=verbose, granularity=core,noduplicates,compact,0,0
  export OMP_NUM_THREADS=68
  export MKL_DYNAMIC=false
  python xxx.py
  

Branch Information

• master, stable and fully tested version based on 0.9dev2 with Intel® MKL backend
• nomkl-optimized, based on 0.9.0dev1 with generic optimizations
• others, experimental codes for different applications which may be merged into master and/or deleted soon

Installation

• Environment Setting

  Set Intel MKL library path into both LD_LIBRARY_PATH and LIBRARY_PATH

• Quick Commands

  git clone https://github.com/intel/theano.git intel-theano
  cd intel-theano
  python setup.py build
  python setup.py install --user [--mkl]   # Note: using 'mkl' option will check and download MKL if it is not available
  cp intel-theano/theanorc_icc_mkl ~/.theanorc
  

• Check Intel MKL

```
python theano/theano/sandbox/mkl/tests/test_mkl.py
     WARNING (theano.gof.cmodule): WARNING: your Theano flags `gcc.cxxflags` specify an `-march=X` flags.
     It is better to let Theano/g++ find it automatically, but we don't do it now
     mkl_available: True
     mkl_version: 20170209
     .
     ----------------------------------------------------------------------
     Ran 1 test in 2.213s

     OK
```

• Run Benchmark

```
python democase/alexnet/benchmark.py
```

• Install Guide (recommend to go througth this document and set up optimized softwares) https://github.com/intel/theano/blob/master/Install_Guide.pdf

Other Optimized Software

• Self-contained MKL in here
• Optimized Numpy in here

* Other names and trademarks may be claimed as the property of others.