As Apple M1 launches, there comes a tendency to port codes originally run on x86 machines to ARM. For C code, porting can be easy if the code only relies on well-defined behaviors — you just need to recompile it. But things can get complicated if your code relies on undefined behaviors — they might differ on these two architectures.
Days ago I was porting otfcc, an optimized OpenType builder and inspector, to ARM platform. I was expecting all I needed to do was a simple recompilation and everything would run out-of-the-box. But I got sucked into segmentation error and malformed output. After about a day’s debugging, I found that these errors were caused by a small difference in wrapping behavior between x86 and ARM! On x86, when casting a negative floating point number into unsigned integer, it will be wrapped. …
TL;DR: Modern compilers feature very strong capability of auto-vectorization. So just write loops and let the compiler optimize them! (with an appropriate optimization level)
Why We Need to Vectorize Our Code
In many cases, we need to perform the same operations on a huge amount of data. For example, if I need to calculate a single value of the probability density function of the Cauchy distribution, I will write:
float dcauchy_single(float x, float location, float scale) {
const float fct = M_1_PI / scale;
const float frc = (x - location) / scale;
const float y = fct / (1.0f + frc * frc);
return y;
}
That works fine for a single…
About
