Quoting Tvrtko Ursulin (2020-11-06 15:17:12) > > On 04/11/2020 17:09, Chris Wilson wrote: > > Use MI_MATH and MI_COND_BBE we can construct a loop that runs for a > > precise number of clock cycles, as measured by the CTX_TIMESTAMP. We use > > the CTX_TIMESTAMP (as opposed to the CS_TIMESTAMP) so that the elapsed > > time is measured local to the context, and the length of the batch is > > unaffected by preemption. Since the clock ticks at a known frequency, we > > can directly translate the batch durations into cycles and so remove the > > requirement for nop calibration, and the often excessively large nop > > batches. > > > > The downside to this is that we need to use engine local registers, and > > before gen11 there is no support in the CS for relative mmio and so this > > technique does not support transparent load balancing on a virtual > > engine before Icelake. > > > > v2: More commentary, more code removal. > > I almost acked it a few times but then since a) I don't have a local > gen11+ and b) trace.pl is broken upstream I kept getting I got cold > feet. Trace.pl becuase I wanted to check if durations now works as > advertised. Although that could be done simpliy with test workloads as > well. Yes. I used a wsim that did a single wait for 100ms and 10 repeats to satisfy myself (which was very useful for testing how the relative mmio bit actually worked). > Anyway, it looks good and gem_wsim.c is inactive enough so I could > easily revert locally it if I needed to run something on my local gen9. > No point in delaying this brilliant improvement. > > Acked-by: Tvrtko Ursulin <tvrtko.ursulin@xxxxxxxxx> > > Oh one question I had - does preemption period works as expected - the > MI_MATH instructions do not prevent setting to non-preemtpable by any > chance? No. It's reduced to a boolean as it is unconditionally checked every few us. I didn't work out a way of having 2 loops. (The problem boils down to not having a conditional jump, only a conditional return; I think the predicated MI_BATCH_BUFFER_START is rcs only.) We could use preempt_us MI_NOOP, but I was hoping it wasn't critical. -Chris _______________________________________________ Intel-gfx mailing list Intel-gfx@xxxxxxxxxxxxxxxxxxxxx https://lists.freedesktop.org/mailman/listinfo/intel-gfx
