Code optimization with single instruction, multiple data
Strong Arm Performance
© Lead Image © vectortatu, 123RF.com
Coding for the ARM NEON vector hardware can significantly improve performance and help you get the most out of low-power systems such as the Raspberry Pi.
You have just coded that new algorithm, the one that handles all the complexities of your data to return exactly the answers you need. As you launch the program and wait, you realize that you have a problem: You are still waiting. The code is slow, far too slow to be usable. You need faster processing, preferably without upgrading your hardware.
Before you reach for the other cores on your processor, it makes sense to see if you can optimize the code for a single core first. Any single-core optimizations will reduce the number of cores ultimately needed. Taking advantage of any available single instruction, multiple data (SIMD) hardware is an effective means of accelerating mathematically intensive problems. SIMD (vector) hardware uses parallel arithmetic units executing the same operation on multiple elements of data within the same clock cycle.
ARM's implementation of SIMD, called NEON, is relatively intuitive and effective. The NEON instructions operate on 128 bits (16 bytes) of data per clock, either as sixteen 8-bit characters, eight 16-bit short integers, or even four 32-bit floating-point numbers. Modern compilers, such as GCC, have the ability to vectorize code automatically during optimization [1], but you often have room for improvement if you are willing to use the NEON instructions directly.
[...]
Buy this article as PDF
(incl. VAT)
Buy Linux Magazine
US / Canada
UK / Australia