[Mesa-dev] [PATCH] i965/fs: Fix saturate for nir_opcode_bcsel.

[Jason Ekstrand]

On Tue, Feb 3, 2015 at 12:27 PM, Kenneth Graunke <kenneth at whitecape.org>
wrote:

> On Tuesday, February 03, 2015 07:10:20 AM you wrote:
> > On Feb 3, 2015 2:35 AM, "Kenneth Graunke" <kenneth at whitecape.org> wrote:
> > > Caught by lit_sat.shader_test.
> > >
> > > Signed-off-by: Kenneth Graunke <kenneth at whitecape.org>
> > > ---
> > >  src/mesa/drivers/dri/i965/brw_fs_nir.cpp | 1 +
> > >  1 file changed, 1 insertion(+)
> > >
> > > diff --git a/src/mesa/drivers/dri/i965/brw_fs_nir.cpp
> > > b/src/mesa/drivers/dri/i965/brw_fs_nir.cpp
> > > index 153a1be..3c611af 100644
> > > --- a/src/mesa/drivers/dri/i965/brw_fs_nir.cpp
> > > +++ b/src/mesa/drivers/dri/i965/brw_fs_nir.cpp
> > > @@ -1084,6 +1084,7 @@ fs_visitor::nir_emit_alu(nir_alu_instr *instr)
> > >        emit(CMP(reg_null_d, op[0], fs_reg(0), BRW_CONDITIONAL_NZ));
> > >        inst = emit(SEL(result, op[1], op[2]));
> > >        inst->predicate = BRW_PREDICATE_NORMAL;
> > > +      inst->saturate = instr->dest.saturate;
> > >        break;
> > >
> > >     default:
> > > --
> > > 2.2.2
> >
> > Hrm... I thought bcsel worked on integers.  You shouldn't be able to sat
> it
> > anyway... This seems strange.
> >
> > As a side-note, this is one of the downsides to typeless that we should
> > figure out how to solve.  Not 100% sure how at the moment.
>
> For LIT's Z component, I generate different code based on whether
> drivers support native integers/prefer real booleans:
>
>    bcsel(fge(0.0f, src.x), 0.0f, pow(...))
>
> or
>
>    fcsel(sge(0.0f, src.x), 0.0f, pow(...))
>
> My thinking was that bcsel uses a real boolean condition, whereas fcsel
> has to do condition != 0.0f...and that the type of the sources being
> selected shouldn't really matter.  But I suppose it does if we're doing
> saturate.
>

Exactly.


> Incidentally, making an "fsat" ALU operation would solve that ambiguity,
> wouldn't require special handling all over the place, could be optimized
> in nir_opt_algebraic, and would probably be better for nouveau...
>

We have one.  It just gets lowered to a dest modifier before you see it.
That said, maybe it was getting lowered wrong here...


> Plus, I think we can probably just emit MOV.sat in the i965 backend, and
> Matt's saturation propagation pass should clean it up for us.
>

Yeah, it should most of the time.  However, lowering the sat in SSA has its
advantages.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.freedesktop.org/archives/mesa-dev/attachments/20150203/9a30ab0d/attachment.html>