Quoting Chris Wilson (2017-07-20 15:15:02)
> Quoting Kenneth Graunke (2017-07-19 21:08:23)
> > On Wednesday, July 19, 2017 3:09:10 AM PDT Chris Wilson wrote:
> > > Sometimes we want to emit a relocation to a NULL surface when the
> > > constructing the batch. If we push the NULL handling into the common
> > > brw_emit_reloc() we can make the batch construction itself more
> > > readable.
> >
> > I don't like this...
> >
> > There is no such thing as a "relocation to a NULL surface". No relocation
> > is emittted in this case. It either means the field is relative to a base
> > address, and is simply an offset, or the address is unused and we're setting
> > a NULL pointer combined with other bits packed into the same DWord.
>
> There's actually no such thing as a relocation ;) I only put it out
> because there were so many duplicated checks.
The difference is that in the always use brw_emit_reloc patch end up
with
@@ -661,18 +662,17 @@ brw_emit_buffer_surface_state(struct brw_context *brw,
out_offset);
isl_buffer_fill_state(&brw->isl_dev, dw,
- .address = (bo ? bo->offset64 : 0) + buffer_offset,
+ .address = (bo ?
+ brw_emit_reloc(&brw->batch,
+ *out_offset +
brw->isl_dev.ss.addr_offset,
+ bo, buffer_offset,
+ I915_GEM_DOMAIN_SAMPLER,
+ (rw ?
I915_GEM_DOMAIN_SAMPLER : 0)) :
+ buffer_offset),
.size = buffer_size,
.format = surface_format,
.stride = pitch,
.mocs = tex_mocs[brw->gen]);
which was a bit too much of an eyesore for me.
-Chris
_______________________________________________
mesa-dev mailing list
mesa-dev@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/mesa-dev
