[igt-dev] [PATCH i-g-t 8/9] lib/igt_fb: Speed up format conversion for local memory

[Chris Wilson]

Quoting Imre Deak (2020-01-30 12:04:09)
> On Thu, Jan 30, 2020 at 10:24:05AM +0000, Chris Wilson wrote:
> > Quoting Imre Deak (2020-01-29 18:16:00)
> > > To speed up the conversion that needs to read from a dGFX local memory
> > > use the same trick as what's used for GTT apertures and make a copy
> > > first into system memory.
> > > 
> > > Signed-off-by: Imre Deak <imre.deak at intel.com>
> > > ---
> > >  lib/igt_fb.c | 4 ++++
> > >  1 file changed, 4 insertions(+)
> > > 
> > > diff --git a/lib/igt_fb.c b/lib/igt_fb.c
> > > index 30f3bfba..ef3fa2ed 100644
> > > --- a/lib/igt_fb.c
> > > +++ b/lib/igt_fb.c
> > > @@ -3427,6 +3427,10 @@ static void create_cairo_surface__convert(int fd, struct igt_fb *fb)
> > >         if (use_enginecopy(fb) || use_blitter(fb) ||
> > >             igt_vc4_is_tiled(fb->modifier)) {
> > >                 setup_linear_mapping(&blit->base);
> > > +
> > > +               /* speed things up by working from a copy in system memory */
> > > +               cvt.src.slow_reads =
> > > +                       is_i915_device(fd) && !gem_has_mappable_ggtt(fd);
> > 
> > Any read from WC (including from GGTT) is greatly improved by using
> > memcpy_from_wc, and it even works with X/Y-tiling through fences.
> > 
> > slow_reads is a question of the mapping, at the time of use.
> 
> Yes, setting slow_reads above will lead to a memcpy_from_wc() of the
> source to a system memory buf and the conversion will happen from the
> latter. What's slow is the random access read during conversion from
> local memory.

We have fast routines for those, if needed. We can read from tiled
surfaces faster than the HW can detile in most cases.
-Chris