Overview / Usage

A new hybrid neural network architecture for extremely fast inference on the Intel Xeon Phi 7210 processor.

This work introduces two CPU optimized algorithms, xGEMM and xCONV. These are binarized forms of general matrix multiply and convolutions.
Speed up of about 30-50 times on the Intel Xeon Phi processors as compared to CMMA algorithms for general matrix multiply. For most use cases, we see an approximate 40 times speed up in xCONV as compared to full precision convolution.

xCMMA/xGEMM code:
https://github.com/akhauriyash/XNOR-Intel-ISA/blob/master/xCMMA.c

xCONV code:
https://github.com/akhauriyash/XNOR-Intel-ISA/blob/master/xCONV.c

//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////

// xCONV.c - Tests for Intel(R) Xeon Phi(TM) Processor.

// Implemented by Yash Akhauri.

// Notes:

// - Performance tests matrix multiply algorithms on a Intel Xeon Phi 7210 Processor.

// - To compile, make sure the directory of echo ~/director/xconv.out | qsub matches.

// To Compile:

// icpc -xMIC-AVX512 -qopenmp -mkl -fp-model fast=2 -fma -unroll=4 xCONV.c -o xconv.out && echo ~/xconv.out |
qsub

//

//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////

//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////

// xCMMA.c - Tests for Intel(R) Xeon Phi(TM) Processor.

// Implemented by Yash Akhauri.

// Notes:

// - Performance tests matrix multiply algorithms on a Intel Xeon Phi 7210 Processor.

// - To compile, make sure the directory of echo ~/director/xconv.out | qsub matches.

// To Compile:

// icpc -xMIC-AVX512 -qopenmp -mkl -fp-model fast=2 -fma -unroll=4 xCMMA.c -o xcmma.out && echo ~/xcmma.out
| qsub

//

//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////