[Beignet] [PATCH V2 1/3] GBE: make vload3 only read 3 elements.

[Zhigang Gong]

I found that will break this optimization patchset, so this patchset
LGTM, will push latter. Thanks.


On Mon, Mar 10, 2014 at 10:23:42AM +0800, Zhigang Gong wrote:
> What's the side effect for the original implementation?
> IMO, by default, a load on an int4 vector is better than
> 3 loads on int scalar. Right?
> 
> On Fri, Mar 07, 2014 at 01:48:46PM +0800, Ruiling Song wrote:
> > clang will align the vec3 load into vec4. we have to do it in frontend.
> > 
> > Signed-off-by: Ruiling Song <ruiling.song at intel.com>
> > ---
> >  backend/src/ocl_stdlib.tmpl.h |    4 ++--
> >  1 file changed, 2 insertions(+), 2 deletions(-)
> > 
> > diff --git a/backend/src/ocl_stdlib.tmpl.h b/backend/src/ocl_stdlib.tmpl.h
> > index 76395fa..e3ac632 100755
> > --- a/backend/src/ocl_stdlib.tmpl.h
> > +++ b/backend/src/ocl_stdlib.tmpl.h
> > @@ -3854,12 +3854,12 @@ INLINE_OVERLOADABLE void vstore3(TYPE##3 v, size_t offset, SPACE TYPE *p) {\
> >    *(p + 3 * offset + 2) = v.s2; \
> >  } \
> >  INLINE_OVERLOADABLE TYPE##3 vload3(size_t offset, const SPACE TYPE *p) { \
> > -  return *(SPACE TYPE##3 *) (p + 3 * offset); \
> > +  return (TYPE##3)(*(p + 3 * offset), *(p+ 3 * offset + 1), *(p + 3 * offset + 2));\
> >  }
> >  
> >  #define DECL_UNTYPED_RDV3_SPACE(TYPE, SPACE) \
> >  INLINE_OVERLOADABLE TYPE##3 vload3(size_t offset, const SPACE TYPE *p) { \
> > -  return *(SPACE TYPE##3 *) (p + 3 * offset); \
> > +  return (TYPE##3)(*(p + 3 * offset), *(p+ 3 * offset + 1), *(p + 3 * offset + 2));\
> >  }
> >  
> >  #define DECL_UNTYPED_RW_ALL_SPACE(TYPE, SPACE) \
> > -- 
> > 1.7.9.5
> > 
> > _______________________________________________
> > Beignet mailing list
> > Beignet at lists.freedesktop.org
> > http://lists.freedesktop.org/mailman/listinfo/beignet
> _______________________________________________
> Beignet mailing list
> Beignet at lists.freedesktop.org
> http://lists.freedesktop.org/mailman/listinfo/beignet