[Intel-gfx] [PATCH 00/11] aliasing ppgtt support v2
chris at chris-wilson.co.uk
Tue Nov 29 01:42:25 PST 2011
On Mon, 28 Nov 2011 21:35:27 +0100, Daniel Vetter <daniel.vetter at ffwll.ch> wrote:
> Hi all,
> Changes since the last submission:
> - fixed issues pointed out by Chris Wilson on irc.
> - fixed an oops on pre-snb, shame on me for that one.
> - added two new patches to only bind objects to the global gtt when required.
> - added a new patch so that userspace can find out whether ppgtt is on. This is
> required to use MI_STORE/LOAD commands correctly from userspace batchbuffers.
> Luckily no currently released userspace code depends on this, it's just prep
> Comments, test reports, reviews and flames highly welcome.
The lazy-gtt is just missing a guard to ensure the buffer is bound in
the global gtt before reading through those PTE (impacts other code to
avoid allocating mappable GTT space). The beauty is that it forced me to
grok the rest of the lazy-gtt, it's deceptively simple.
Aside from the lazy-gtt sufficiently speeding up command submission and
reopening an old wound that prevents me from usefully analysing
performance, the series is tested-by myself. I've so far reviewed
everything but the actual mechanics of the PDE, which is not saying much
Chris Wilson, Intel Open Source Technology Centre
More information about the Intel-gfx