On 17/12/15 19:00, Dave Gordon wrote:
On 17/12/15 12:27, Chris Wilson wrote:
On Thu, Dec 17, 2015 at 11:09:54AM +0000, Nick Hoath wrote:
[snip!]
If you want to make such a big change, eliminate the default_ctx from
execlists.
-Chris
No, we need the default (or global) context for idling the engines, as
well as for sending initialisation commands during startup. We can't
make the GPU stop using any given (user) context within a known bounded
time except by telling it to switch to another context. Ergo, to stop
using ANY user context, there must exist a non-user context that we can
switch to.
After writing this last night, I realised this morning that maybe you
didn't mean "eliminate the default context per se" but rather "eliminate
the use of ring->default_context", which is a very different thing and
much less problematic.
So if that's the case, then, yes, I'll be quite happy to provide a
followup patch which eliminates most uses of ring->default_context and
in particular all those in intel_lrc.c where it's compared against
another context pointer. But I'm not going to do that extra work until
and unless this is merged, as it would just be a waste of effort.
The reason we just hit this specific use of default_context first is
because this is the one that was blocking the merge of Nick's "Fix
context/engine cleanup order" patch that you'd already R-B'd. Once this
is in, we can get Nick's patches in, and /then/ clean up all the other
comparisons made against ring->default_context.
.Dave.
_______________________________________________
Intel-gfx mailing list
Intel-gfx@xxxxxxxxxxxxxxxxxxxxx
http://lists.freedesktop.org/mailman/listinfo/intel-gfx
