On Fri, Jan 06, 2023 at 11:44:58AM +0100, Daniel Vetter wrote: > > > The problem is going to happen as soon as you have cross-vendor userspace. > > > Which I'm kinda hoping is at least still the aspiration. Because with > > > cross-vendor userspace you generally iterate & open all devices before you > > > select the one you're going to use. And so we do kinda need a distinction, > > > or we need that the single-user drivers also guarantee that open() is > > > cheap. > > > > FWIW we had good support in ivpu for probe open's in form of lazy context > > allocation. It was removed recently due to review feedback that this is > > unnecessary, but we can add it back. > > Yeah once you have more than 1 multi-user accel chip in the system you > need to do that. Which is really the reason why I think smashing > multi-user client accel things into render is good, it forces drivers > to suck less. > > On that topic, does your userspace still do the drmIoctl() wrapper? Yes, it still does. We released the code BTW, the wrapper can be seen here: https://github.com/intel/linux-vpu-driver/blob/b6ed73cabf87f461cbbe4427e1b9351a548d790b/umd/vpu_driver/source/os_interface/vpu_driver_api.cpp#L41 Regards Stanislaw
