[Mesa-dev] [PATCH 1/2] i965: Move MI_BATCHBUFFER_END handling into brw_finish_batch().

[Chris Wilson]

Quoting Kenneth Graunke (2017-09-18 21:38:09)
> On Monday, September 18, 2017 11:03:37 AM PDT Chris Wilson wrote:
> > Tangential: Do we care about intel_upload_finish on batch flush any more
> > now that async access is universal?
> 
> I'm not sure.  We could certainly enable async maps for the upload BOs,
> and avoid flushing them when finishing the batch.  Then again, since the
> uploaded data is still related to the commands in the batch, it might
> make sense to "start fresh" with every batch.
> 
> I don't think it much matters, anymore.

I was hoping it would reduce one special case; the upload buffer becomes
just an unnamed BO and it would feed through the same range analysis as
named BO (eventually). Right now, just to remove one more line of magic.

The only caveat is if it is used for readback, and is then tied to later
batches before the readback is sync'ed. That seems unlikely, but I
haven't checked. On the write, tying the upload buffer to multiple
batches doesn't seem to present any problems (and offers a minor
reduction in mem/GTT pressure, although unrealistic to expect it be a
tipping point).
-Chris