[Mesa-dev] [PATCH 09/12] i965/fs: make SIMD-splitting respect the original stride/offset

[Pohjolainen, Topi]

On Mon, Aug 17, 2015 at 10:09:21AM +0300, Pohjolainen, Topi wrote:
> On Fri, Aug 14, 2015 at 03:30:18PM -0700, Connor Abbott wrote:
> > In some cases, we need to emit ALU instructions with a certain stride
> > due to a HW limitation. When splitting that instruction, we need to
> > respect the original stride when creating the temporaries we load from
> > and store into. Otherwise, we'll reintroduce the problem we were trying
> > to work around.
> > 
> > Signed-off-by: Connor Abbott <connor.w.abbott at intel.com>
> > ---
> >  src/mesa/drivers/dri/i965/brw_fs.cpp | 14 ++++++++++----
> >  1 file changed, 10 insertions(+), 4 deletions(-)
> > 
> > diff --git a/src/mesa/drivers/dri/i965/brw_fs.cpp b/src/mesa/drivers/dri/i965/brw_fs.cpp
> > index 812648f..386e9a2 100644
> > --- a/src/mesa/drivers/dri/i965/brw_fs.cpp
> > +++ b/src/mesa/drivers/dri/i965/brw_fs.cpp
> > @@ -2370,12 +2370,14 @@ fs_visitor::opt_register_renaming()
> >  
> >        if (depth == 0 &&
> >            inst->dst.file == GRF &&
> > -          alloc.sizes[inst->dst.reg] == inst->exec_size / 8 &&
> > +          alloc.sizes[inst->dst.reg] ==
> > +            inst->dst.component_size(inst->exec_size) &&
> >            !inst->is_partial_write()) {
> >           if (remap[dst] == -1) {
> >              remap[dst] = dst;
> >           } else {
> > -            remap[dst] = alloc.allocate(inst->exec_size / 8);
> > +            remap[dst] =
> > +               alloc.allocate(inst->dst.component_size(inst->exec_size));
> >              inst->dst.reg = remap[dst];
> >              progress = true;
> >           }
> > @@ -4334,6 +4336,8 @@ fs_visitor::lower_simd_width()
> >                     * temporary passed as source to the lowered instruction.
> >                     */
> >                    split_inst.src[j] = lbld.vgrf(inst->src[j].type, src_size);
> > +                  split_inst.src[j].subreg_offset = inst->src[j].subreg_offset;
> 
> The fixes for the wider component size (64-bits) and the stride are clearly
> needed for doubles. I'm wondering though why the sub-register offset hasn't
> caused us any problems before. That change is not needed just for doubles,
> is it?

I meant that it is applicable for other types as well.