[Intel-gfx] [PATCH v2 3/3] drm/i915/perf: enable filtering on multiple contexts

[Chris Wilson]

Quoting Lionel Landwerlin (2020-04-06 15:30:38)
> On 06/04/2020 17:27, Chris Wilson wrote:
> > Quoting Lionel Landwerlin (2020-04-06 15:07:30)
> >> On 06/04/2020 16:59, Chris Wilson wrote:
> >>> Quoting Lionel Landwerlin (2020-04-06 14:54:38)
> >>>> On 31/03/2020 21:08, Chris Wilson wrote:
> >>>>> Quoting Lionel Landwerlin (2020-03-31 12:48:21)
> >>>>>> Add 2 new properties to the i915-perf open ioctl to specify an array
> >>>>>> of GEM context handles as well as the length of the array.
> >>>>> Hmm. The other thought is ctx->engine[] where one context may have more
> >>>>> than one logical context instance that OA could track. The extension to
> >>>>> track multiple pinned contexts should equally work for multiple engines.
> >>>>>
> >>>>>         I915_DEFINE_CONTEXT_PARAM_ENGINES(engines, 64) = {};
> >>>>>         struct drm_i915_gem_context_param p = {
> >>>>>                 .ctx_id = gem_context_create(i915),
> >>>>>                 .param = I915_CONTEXT_PARAM_ENGINES,
> >>>>>                 .value = to_user_pointer(&engines),
> >>>>>                 .size = sizeof(struct i915_context_param_engines),
> >>>>>         };
> >>>>>         gem_context_set_param(i915, &p);
> >>>>>
> >>>>> would do the trick in creating one context with 64 rcs0 that they may
> >>>>> want to track. And that also opens the door to virtual engines over top.
> >>>>> -Chris
> >>>> I rather punt this away for now :)
> >>>>
> >>>> I can't see use cases for Iris/Vulkan.
> >>> There's immediate use cases for iris, since it uses 2 contexts instead
> >>> of 2 logical instances within one GL/GEM context.
> >>
> >> I don't understand this. Are you saying Iris could use the stuff from
> >> above and still select what logical context to dispatch to?
> > Yes. And can be done so only by changing the context setup to create one
> > context with two independent rcs engines (different logical GPU state,
> > and rings, *only* sharing the same vm). And since legacy EXEC_DEFAULT[0]
> > and EXEC_RENDER[1] alias to the same engine, very little code needs to be
> > changed to support (1 ctx, 2 engines) vs (2 ctx, 1 engine).
> > -Chris
> 
> Sounds like there could a performance gain somewhere...
> 
> Will look into this.

A year old, so a bit stale,
https://gitlab.freedesktop.org/ickle/mesa/-/commits/iris-composite-context

Note it doesn't setup a single shared timeline as Kenneth wanted to keep
explicit control over the fencing between the pipelines.

* fingers crossed that's the right version.
-Chris