Comment 137 for bug 252094

(In reply to comment #24)
> For most of the EXA tests the performance is basically unchanged when changing
> from 10x10 to 500x500. That suggests that there is good GPU acceleration
> happening. Contrast that with the XAA results where performance generally falls
> dramatically from 10x10 to 500x500 suggesting that the CPU is getting involved
> on a per-pixel basis.

I got some details wrong above.

We do expect substantial slowdown between the 10x10 and 500x500 tests. And in fact, the original numbers reported here are basically physically impossible. The expected results are about 1000 operations per second for the 500x500 case (based on available memory bandwidth) and about 125000 to 250000 operations per second for the 10x10 case (limited this time by the maximum rate of command submission).

So something went wrong in testing that you saw such impossibly large numbers.

> The big exception is the pixmap to pixmap test where not only does EXA get
> dramatically slower from 10x10 to 500x500, but it's also from 2x to 14x slower
> than XAA for this case. So there's obviously something quite broken there,
> (perhaps migration for a fallback).

It still looks like there is a problem in this case. I'm still looking to see
if I can replicate this.

-Carl