Quoting Chris Wilson (2019-08-01 11:57:06) > Quoting Tvrtko Ursulin (2019-08-01 09:53:15) > > We could store it in ce then. We already have well defined control > > points for when vm changes when all are updated. > > We are storing it in ce; it's not like we recompute it all that often, > and when we do it's because we have rebound the vma. > > > If done like this then it looks like assigning ctx->hw_id could also do > > the default_desc update, so that we can avoid even more work done at pin > > time. > > What ctx->hw_id? You are imagining things again :-p > > Remember that we only do this on first pin from idle, not every pin. Fwiw, I quickly looked at only doing it if the vma is rebound, but that's move branches just to save a couple. The low frequency at which we have to actually compute this (walk a few more branches inside an already branchy contxt_pin) doesn't seem to justify the extra storage for me. It's not like we are recomputing lrc_desc on every submit as it once was. -Chris _______________________________________________ Intel-gfx mailing list Intel-gfx@xxxxxxxxxxxxxxxxxxxxx https://lists.freedesktop.org/mailman/listinfo/intel-gfx
