I think there's a generic C fallback, which can also serve as a baseline. But for the big (targeted) architectures, there one handwritten assembly version per arch.
Yup.
On startup, it runs cpuid and assigns each operation the most optimal function pointer for that architecture.
In addition to things like ‘supports avx’ or ‘supports sse4’ some operations even have more explicit checks like ‘is a fifth generation celeron’. The level of optimization in that case was optimizing around the cache architecture on the cpu iirc.
Source: I did some dirty things with chromes native client and ffmpeg 10 years ago.
First, we had ICON computers in my elementary school, we'd all try to spin the trackball as quickly as it would go. Not sure if we ever broke one.
The second is when I worked at BlackBerry. I was building a feature that allowed you to use your QNX BlackBerry as a Bluetooth HID device. You could connect it to any device and use the trackpad + physical keyboard to remotely control a computer. It was fantastic. You could hook your laptop up to a project and control slides from your BlackBerry.
Then some product manager with questionable decision making told me to lock it down so it would only work with Blackberry Playbooks for "business purposes", rendering it effectively useless (since Playbooks are all ewaste). I distinctly remember that meeting where Dan Dodge argued that since it's a standard, it should not be locked down.
I respect Dan Dodge for that, I don't think I'd work with that PM again.