[Intel-gfx] [Mesa-dev] [PATCH] i965: Share the workaround bo between all contexts
chadversary at chromium.org
Thu Jan 26 23:40:47 UTC 2017
On Thu 26 Jan 2017, Chris Wilson wrote:
> On Thu, Jan 26, 2017 at 09:39:51AM -0800, Chad Versace wrote:
> > On Thu 26 Jan 2017, Chris Wilson wrote:
> > > Since the workaround bo is used strictly as a write-only buffer, we need
> > > only allocate one per screen and use the same one from all contexts.
> > >
> > > (The caveat here is during extension initialisation, where we write into
> > > and read back register values from the buffer, but that is performed only
> > > once for the first context - and baring synchronisation issues should not
> > > be a problem. Safer would be to move that also to the screen.)
> > >
> > > v2: Give the workaround bo its own init function and don't piggy back
> > > intel_bufmgr_init() since it is not that related.
> > >
> > > v3: Drop the reference count of the workaround bo for the context since
> > > the context itself is owned by the screen (and so we can rely on the bo
> > > existing for the lifetime of the context).
> > I like this idea, but I have questions and comments about the details.
> > More questions than comments, really.
> > Today, with only Mesa changes, could we effectively do the same as
> > drm_intel_gem_bo_disable_implicit_sync(screen->workaround_bo);
> > by hacking Mesa to set no read/write domain when emitting relocs for the
> > workaround_bo? (I admit I don't fully understand the kernel's domain
> > tracking). If that does work, then it just would require a small hack to
> > brw_emit_pipe_control_write().
> Yes, for anything that is totally scratch just not setting the write
> hazard is the same. For something like the seqno page where we have
> multiple engines that we do want to be preserved, not settting the write
> hazzard had the consequence that page could be lost under memory pressure
> or across resume. (As usual there are some details that this part of the
> ABI had to be relaxed because userspace didn't have this flag.)
> But that doesn't sell many bananas.
Good. That's how I thought it worked.
More information about the Intel-gfx