[Intel-gfx] Performance regression using Firefox between 2.14 and 2.15

Clemens Eisserer linuxhippy at gmail.com
Thu Sep 29 17:25:30 CEST 2011


> We think we've fixed all the related bugs in the kernel, so you can try
> flipping the switch and see what happens.

Manuelle re-enabling relaxedfencing brings speed back to normal, quite
a few pages that scrolled slow in firefox are now as snappy as before
2.15 :)

> However, I think your symptoms are only a side-effect of disabling relaxed
> fencing. The cause is actually more likely to be the use of an untiled
> buffer for an image with a stride of greater then 4096 bytes. With relaxed
> fencing there is less likelihood of any such buffers being present and so
> reused for textures.

Could this be worked arround, so that even with relaxed fencing
disabled, peolpe don't run into this regression?
I was quite irritated when I first observed the slowdown, and since
relaxedfencinig is disabled by default I assume a lot of users are
affected. (but probably more forgiving when it comes to slow user
interfaces ;) )

Thanks, Clemens

More information about the Intel-gfx mailing list