Skip to content

Latest commit

 

History

History
13 lines (9 loc) · 579 Bytes

README.md

File metadata and controls

13 lines (9 loc) · 579 Bytes
Raw

Build

To build the project, please refer to original BMXNet-v2

Optimize

This codebase optimize binary gemm kernel of BMXNet-v2 using AVX instructions. The optimization are two parts:

  1. The xnor popcnt gemm. First remove openmp because this slows down the performance. Second use AVX to optimize xnor+popcnt (this gives ~2x speedup).
  2. Convert float to bits. This part cost a lot of time and is easily optimized. This gives 5x speedup.

Supplied with a Binary_IDF

In addition to optimiza binary kernel, A Binary IDF for lossless compression is supplied under example.