Vector types and debug performance
For reasons, I have found myself writing a vector math library. Like Aras some time ago I’ve found an unpleasant surprise when looking at the debug performance of some vector types.
There is no point repeating what Aras already wrote, but I have some additional data points. My concrete example uses float3 (3D), where I would still like to use vector instructions, but none of this here is actually float3 specific.
I ended up doing this investigation because I had written a little sample particle simulation and decided to improve it by adding a vector math library. The particle simulation just used scalar code, no float3 struct in sight. Adding the vector math library did improve performance of optimized builds, but completely fell off a cliff in non-optimized debug builds.
This was unexpected. Did I not avoid all of the obviously bad things? My code, for example, does not contain templates. Vector math isn’t going to change, ever, and I need just (float|int)(2|3|4). It’s way more important to have readable, sensible code and duplicate that 6 times than to have a single template. But alas, that was not enough to save me from performance woes.
