README.md

High Performance Vector Math

This repository shows a non-GPU way to get faster linear algebra calculations. We test some software and hardware optimizations to get speedup.

Content

1. C simple
2. C unroll
3. C - Intel MKL
4. Python simple
5. Python numpy

Running Instructions

1. Requires C, Intel MKL library, Python and numpy
2. For best performance for C programs compile with -O3 flag for compiler level optimizations

gcc program.c -O3 -o program

3. To compile C program with Intel MKL

gcc program.c -O3 -o program -lmkl_intel_lp64 -lmkl_sequential -lmkl_core -lpthread -lm

Dot Product Observations:

Algorithm	Vector Size	Runtime(sec)	Bandwidth (GB/s)	Throughput
C simple	1,000,000	0.001501 sec	5.330960	1.333 GFLOPS
C simple	300,000,000	0.457934 sec	5.240930	1.310 GFLOPS
C unroll	1,000,000	0.000465 sec	17.202948	4.301 GFLOPS
C unroll	300,000,000	0.208287 sec	11.522536	2.881 GFLOPS
C - Intel MKL	1,000,000	0.000194 sec	41.342320	10.336 GFLOPS
C - Intel MKL	300,000,000	0.093496 sec	25.669479	6.417 GFLOPS
Python simple	1,000,000	0.610351 sec	0.013107	0.00328 GFLOPS
Python simple	300,000,000	177.1355 sec	0.013549	0.00339 GFLOPS
Python numpy	1,000,000	0.001509 sec	5.301609	1.3254 GFLOPS
Python numpy	300,000,000	0.452451 sec	5.304442	1.3261 GFLOPS

License

MIT