[Mesa-dev] [PATCH 31/59] i965/fs: optimize pack double

[Iago Toral]

On Mon, 2016-05-02 at 15:11 -0400, Connor Abbott wrote:
> On Fri, Apr 29, 2016 at 7:29 AM, Samuel Iglesias Gonsálvez
> <siglesias at igalia.com> wrote:
> > From: Iago Toral Quiroga <itoral at igalia.com>
> >
> > When we are actually creating a double using values obtained from a
> > previous unpack operation we can bypass the unpack and source from
> > the original double value directly.
> > ---
> >  src/mesa/drivers/dri/i965/brw_fs_nir.cpp | 23 +++++++++++++++++++++++
> >  1 file changed, 23 insertions(+)
> >
> > diff --git a/src/mesa/drivers/dri/i965/brw_fs_nir.cpp b/src/mesa/drivers/dri/i965/brw_fs_nir.cpp
> > index ca1f24f..4332575 100644
> > --- a/src/mesa/drivers/dri/i965/brw_fs_nir.cpp
> > +++ b/src/mesa/drivers/dri/i965/brw_fs_nir.cpp
> > @@ -1091,6 +1091,29 @@ fs_visitor::nir_emit_alu(const fs_builder &bld, nir_alu_instr *instr)
> >        break;
> >
> >     case nir_op_pack_double_2x32_split:
> > +      /* Optimize the common case where we are re-packing a double with
> > +       * the result of a previous double unpack. In this case we can take the
> > +       * 32-bit value to use in the re-pack from the original double and bypass
> > +       * the unpack operation.
> > +       */
> > +      for (int i = 0; i < 2; i++) {
> > +         if (instr->src[i].src.is_ssa) {
> > +            nir_instr *parent_instr = instr->src[i].src.ssa->parent_instr;
> > +            if (parent_instr->type == nir_instr_type_alu) {
> > +               nir_alu_instr *alu_parent = nir_instr_as_alu(parent_instr);
> > +               if (alu_parent->op == nir_op_unpack_double_2x32_split_x ||
> > +                   alu_parent->op == nir_op_unpack_double_2x32_split_y) {
> > +                  op[i] = get_nir_src(alu_parent->src[0].src);
> 
> The source could be a register here (say, if it was part of a phi web
> that got collapsed) in which case we should bail since something else
> might've overwritten it in the meantime. Implementing Topi's
> suggestion to use early exits instead will prevent this from getting
> even more unwieldy when fixing this up.

Oh right, good catch! I'll fix that and implement's Topi's changes too.
Thanks!

> > +                  op[i] = offset(retype(op[i], BRW_REGISTER_TYPE_DF), bld,
> > +                                 alu_parent->src[0].swizzle[channel]);
> > +                  op[i] = retype(op[i], BRW_REGISTER_TYPE_UD);
> > +                  if (alu_parent->op == nir_op_unpack_double_2x32_split_y)
> > +                     op[i] = horiz_offset(op[i], 1);
> > +                  op[i] = stride(op[i], 2);
> > +               }
> > +            }
> > +         }
> > +      }
> >        bld.emit(FS_OPCODE_PACK, result, op[0], op[1]);
> >        break;
> >
> > --
> > 2.5.0
> >
> > _______________________________________________
> > mesa-dev mailing list
> > mesa-dev at lists.freedesktop.org
> > https://lists.freedesktop.org/mailman/listinfo/mesa-dev
> _______________________________________________
> mesa-dev mailing list
> mesa-dev at lists.freedesktop.org
> https://lists.freedesktop.org/mailman/listinfo/mesa-dev