[Mesa-dev] [PATCH 41/95] i965/vec4: make the generator set correct NibCtrl for SIMD4 DF instructions

[Iago Toral]

On Mon, 2016-08-08 at 15:26 -0700, Francisco Jerez wrote:
> Iago Toral Quiroga <itoral at igalia.com> writes:
> 
> > 
> > From the HSW PRM, Command Reference, QtrCtrl:
> > 
> >    "NibCtrl is only allowed for SIMD4 instructions with a DF
> > (Double Float)
> >     source or destination type."
> > 
> > v2 (Samuel): Assert that the type is DF.
> > 
> > Signed-off-by: Samuel Iglesias Gonsálvez <siglesias at igalia.com>
> > ---
> >  src/mesa/drivers/dri/i965/brw_vec4_generator.cpp | 9 +++++++++
> >  1 file changed, 9 insertions(+)
> > 
> > diff --git a/src/mesa/drivers/dri/i965/brw_vec4_generator.cpp
> > b/src/mesa/drivers/dri/i965/brw_vec4_generator.cpp
> > index c6e040e..dc4f6db 100644
> > --- a/src/mesa/drivers/dri/i965/brw_vec4_generator.cpp
> > +++ b/src/mesa/drivers/dri/i965/brw_vec4_generator.cpp
> > @@ -1494,6 +1494,7 @@ generate_code(struct brw_codegen *p,
> >        brw_set_default_saturate(p, inst->saturate);
> >        brw_set_default_mask_control(p, inst->force_writemask_all);
> >        brw_set_default_acc_write_control(p, inst-
> > >writes_accumulator);
> > +      brw_set_default_group(p, 0);
> > 
> >        assert(inst->base_mrf + inst->mlen <= BRW_MAX_MRF(devinfo-
> > >gen));
> >        assert(inst->mlen <= BRW_MAX_MSG_LENGTH);
> > @@ -1529,6 +1530,14 @@ generate_code(struct brw_codegen *p,
> >        } else {
> >           assert(inst->exec_size == 8 || inst->exec_size == 4);
> >           brw_set_default_exec_size(p, cvt(inst->exec_size) - 1);
> > +         if (inst->exec_size == 4 && !inst->force_writemask_all) {
> > +            assert(inst->dst.type == BRW_REGISTER_TYPE_DF ||
> > +                   inst->src[0].type == BRW_REGISTER_TYPE_DF ||
> > +                   inst->src[1].type == BRW_REGISTER_TYPE_DF ||
> > +                   inst->src[2].type == BRW_REGISTER_TYPE_DF);
> > +            assert(inst->group == 0 || inst->group == 4);
> > +            brw_inst_set_group(p->devinfo, p->current, inst-
> > >group);
> I think this should be unconditional (e.g. if somebody tries to set
> group == 4 on an instruction with exec_size == 8 it would be nice to
> get
> an assertion failure instead of the group value being silently
> ignored).  How about you replace the 'brw_set_default_group(p, 0)'
> line
> above with:
> 
> > 
> > 
> >       assert(inst->group % inst->exec_size == 0);
> >       assert(inst->group % 8 == 0 ||
> >              inst->dst.type == BRW_REGISTER_TYPE_DF ||
> >              inst->src[0].type == BRW_REGISTER_TYPE_DF ||
> >              inst->src[1].type == BRW_REGISTER_TYPE_DF ||
> >              inst->src[2].type == BRW_REGISTER_TYPE_DF);
> >       brw_set_default_group(p, inst->group);
> and then drop this hunk?

Yes, that looks like better. Thanks!

I suppose that setting the group won't have any effect  if
force_writemask_all is set as too so there is no need to void setting
it in that case, right?

> > 
> >        }
> >  
> >        switch (inst->opcode) {