Cuda Solver

[Bird]

Amazing. Maybe they're short staffed because they're having a hard time making a profit these days.

[Julio Jerez]

it worked much better, so that's good.
It is still very difficult to get accurate timing across kernel calls, since the driver can preempt the sequence of summited kernels. is if foe example you
take the time at the beginning, them you issue few kernels, them to take the time at the end.
what happen is the somewhere in the middle of issuing the kernels, the driver stop to do the rendering,
so what you get is the sampling that happened between the two calls, and that can have a whole bunch of other stuff.
but at least it is consistent, because the measurement is a actual clock, and no the nonsense give but the functions.

I guess one only way to get a accurate measurement is to measure in every stream 0 kernel call and add the deltas together.
but that's could also be a futile, because it assume no preemption in the middle of a kernel call, which I believe the the latest hardware can do now. So I will not spend time on that, to find out at the end that it may or may not be any better.

no sure why the did not wrapped that into an intrinsic function instead of wrapping a useless call liek clock 64 to give the core ticks. at least wrapping the both .
but anyway, at least that's one problem less.