[Beignet] [PATCH] revise built-in function "shuffle"

Zhigang Gong zhigang.gong at linux.intel.com
Wed Jul 24 01:00:22 PDT 2013 

On Wed, Jul 24, 2013 at 07:34:07AM +0000, Xing, Homer wrote:
> Yes, 3-component argument is not required by OpenCL spec.
> 
> I have discussed this issue with Yang Rong in today's noon. We think 3-component argument can help the user. 
> 
> Anyway, please feel free to simply delete 3-component version code if you like.
I will delete it. But just want to clarify, this is not a personal like or dislike.
Add a new API which is not in the spec need a some extra efforts. You can't just
simply add it into the header file.
You need to define an extension, and give the user application a chance to check
this extension and thus the user application could use this new API when it find
the extension exist, otherwise the user application could choose the APIs in the
spec to achieve the same function.

If you don't do the above things, you just simply introduce a non-portable API
which is really useless for any serious Open CL developers.

And back to this API itself, I don't think it worths to add an extension for this
new API.

> 
> -----Original Message-----
> From: Zhigang Gong [mailto:zhigang.gong at linux.intel.com] 
> Sent: Wednesday, July 24, 2013 3:32 PM
> To: Xing, Homer
> Cc: beignet at lists.freedesktop.org
> Subject: Re: [Beignet] [PATCH] revise built-in function "shuffle"
> 
> On Wed, Jul 24, 2013 at 02:58:28PM +0800, Homer Hsing wrote:
> > Signed-off-by: Homer Hsing <homer.xing at intel.com>
> > ---
> >  backend/src/ocl_stdlib.tmpl.h | 68 
> > +++++++++++++++++++++----------------------
> >  1 file changed, 34 insertions(+), 34 deletions(-)
> > 
> > diff --git a/backend/src/ocl_stdlib.tmpl.h 
> > b/backend/src/ocl_stdlib.tmpl.h index 387e6fb..92a822a 100644
> > --- a/backend/src/ocl_stdlib.tmpl.h
> > +++ b/backend/src/ocl_stdlib.tmpl.h
> > @@ -1106,68 +1106,69 @@ DECL_UNTYPED_RW_ALL(float)  #define maxmag 
> > __gen_ocl_internal_maxmag  #define minmag __gen_ocl_internal_minmag
> >  
> > +/////////////////////////////////////////////////////////////////////
> > +////////
> >  // Miscellaneous Vector Functions (see 6.11.12 of OCL 1.1 spec)  
> > //////////////////////////////////////////////////////////////////////
> > ///////
> >  #define DEC2(TYPE, XTYPE) \
> >    INLINE_OVERLOADABLE TYPE##2 shuffle(XTYPE x, uint2 mask) { \
> >      TYPE##2 y; \
> > -    y.s0 = ((TYPE *) &x)[mask.s0 & 1]; \
> > -    y.s1 = ((TYPE *) &x)[mask.s1 & 1]; \
> > +    y.s0 = ((TYPE *) &x)[mask.s0 & (vec_step(x) - 1)]; \
> > +    y.s1 = ((TYPE *) &x)[mask.s1 & (vec_step(x) - 1)]; \
> >      return y; \
> >    }
> >  
> >  #define DEC3(TYPE, XTYPE) \
> >    INLINE_OVERLOADABLE TYPE##3 shuffle(XTYPE x, uint3 mask) { \
> >      TYPE##3 y; \
> > -    y.s0 = ((TYPE *) &x)[mask.s0 & 3]; \
> > -    y.s1 = ((TYPE *) &x)[mask.s1 & 3]; \
> > -    y.s2 = ((TYPE *) &x)[mask.s2 & 3]; \
> > +    y.s0 = ((TYPE *) &x)[mask.s0 & (vec_step(x) - 1)]; \
> > +    y.s1 = ((TYPE *) &x)[mask.s1 & (vec_step(x) - 1)]; \
> > +    y.s2 = ((TYPE *) &x)[mask.s2 & (vec_step(x) - 1)]; \
> >      return y; \
> >    }
> In misc vector function section,  gentypen (or gentypem) to indicate the built-in data types {char|uchar|...}{2|4|8|16} And there is no 3-components vector type defined here.
> >  
> >  #define DEC4(TYPE, XTYPE) \
> >    INLINE_OVERLOADABLE TYPE##4 shuffle(XTYPE x, uint4 mask) { \
> >      TYPE##4 y; \
> > -    y.s0 = ((TYPE *) &x)[mask.s0 & 3]; \
> > -    y.s1 = ((TYPE *) &x)[mask.s1 & 3]; \
> > -    y.s2 = ((TYPE *) &x)[mask.s2 & 3]; \
> > -    y.s3 = ((TYPE *) &x)[mask.s3 & 3]; \
> > +    y.s0 = ((TYPE *) &x)[mask.s0 & (vec_step(x) - 1)]; \
> > +    y.s1 = ((TYPE *) &x)[mask.s1 & (vec_step(x) - 1)]; \
> > +    y.s2 = ((TYPE *) &x)[mask.s2 & (vec_step(x) - 1)]; \
> > +    y.s3 = ((TYPE *) &x)[mask.s3 & (vec_step(x) - 1)]; \
> >      return y; \
> >    }
> >  
> >  #define DEC8(TYPE, XTYPE) \
> >    INLINE_OVERLOADABLE TYPE##8 shuffle(XTYPE x, uint8 mask) { \
> >      TYPE##8 y; \
> > -    y.s0 = ((TYPE *) &x)[mask.s0 & 7]; \
> > -    y.s1 = ((TYPE *) &x)[mask.s1 & 7]; \
> > -    y.s2 = ((TYPE *) &x)[mask.s2 & 7]; \
> > -    y.s3 = ((TYPE *) &x)[mask.s3 & 7]; \
> > -    y.s4 = ((TYPE *) &x)[mask.s4 & 7]; \
> > -    y.s5 = ((TYPE *) &x)[mask.s5 & 7]; \
> > -    y.s6 = ((TYPE *) &x)[mask.s6 & 7]; \
> > -    y.s7 = ((TYPE *) &x)[mask.s7 & 7]; \
> > +    y.s0 = ((TYPE *) &x)[mask.s0 & (vec_step(x) - 1)]; \
> > +    y.s1 = ((TYPE *) &x)[mask.s1 & (vec_step(x) - 1)]; \
> > +    y.s2 = ((TYPE *) &x)[mask.s2 & (vec_step(x) - 1)]; \
> > +    y.s3 = ((TYPE *) &x)[mask.s3 & (vec_step(x) - 1)]; \
> > +    y.s4 = ((TYPE *) &x)[mask.s4 & (vec_step(x) - 1)]; \
> > +    y.s5 = ((TYPE *) &x)[mask.s5 & (vec_step(x) - 1)]; \
> > +    y.s6 = ((TYPE *) &x)[mask.s6 & (vec_step(x) - 1)]; \
> > +    y.s7 = ((TYPE *) &x)[mask.s7 & (vec_step(x) - 1)]; \
> >      return y; \
> >    }
> >  
> >  #define DEC16(TYPE, XTYPE) \
> >    INLINE_OVERLOADABLE TYPE##16 shuffle(XTYPE x, uint16 mask) { \
> >      TYPE##16 y; \
> > -    y.s0 = ((TYPE *) &x)[mask.s0 & 15]; \
> > -    y.s1 = ((TYPE *) &x)[mask.s1 & 15]; \
> > -    y.s2 = ((TYPE *) &x)[mask.s2 & 15]; \
> > -    y.s3 = ((TYPE *) &x)[mask.s3 & 15]; \
> > -    y.s4 = ((TYPE *) &x)[mask.s4 & 15]; \
> > -    y.s5 = ((TYPE *) &x)[mask.s5 & 15]; \
> > -    y.s6 = ((TYPE *) &x)[mask.s6 & 15]; \
> > -    y.s7 = ((TYPE *) &x)[mask.s7 & 15]; \
> > -    y.s8 = ((TYPE *) &x)[mask.s8 & 15]; \
> > -    y.s9 = ((TYPE *) &x)[mask.s9 & 15]; \
> > -    y.sa = ((TYPE *) &x)[mask.sa & 15]; \
> > -    y.sb = ((TYPE *) &x)[mask.sb & 15]; \
> > -    y.sc = ((TYPE *) &x)[mask.sc & 15]; \
> > -    y.sd = ((TYPE *) &x)[mask.sd & 15]; \
> > -    y.se = ((TYPE *) &x)[mask.se & 15]; \
> > -    y.sf = ((TYPE *) &x)[mask.sf & 15]; \
> > +    y.s0 = ((TYPE *) &x)[mask.s0 & (vec_step(x) - 1)]; \
> > +    y.s1 = ((TYPE *) &x)[mask.s1 & (vec_step(x) - 1)]; \
> > +    y.s2 = ((TYPE *) &x)[mask.s2 & (vec_step(x) - 1)]; \
> > +    y.s3 = ((TYPE *) &x)[mask.s3 & (vec_step(x) - 1)]; \
> > +    y.s4 = ((TYPE *) &x)[mask.s4 & (vec_step(x) - 1)]; \
> > +    y.s5 = ((TYPE *) &x)[mask.s5 & (vec_step(x) - 1)]; \
> > +    y.s6 = ((TYPE *) &x)[mask.s6 & (vec_step(x) - 1)]; \
> > +    y.s7 = ((TYPE *) &x)[mask.s7 & (vec_step(x) - 1)]; \
> > +    y.s8 = ((TYPE *) &x)[mask.s8 & (vec_step(x) - 1)]; \
> > +    y.s9 = ((TYPE *) &x)[mask.s9 & (vec_step(x) - 1)]; \
> > +    y.sa = ((TYPE *) &x)[mask.sa & (vec_step(x) - 1)]; \
> > +    y.sb = ((TYPE *) &x)[mask.sb & (vec_step(x) - 1)]; \
> > +    y.sc = ((TYPE *) &x)[mask.sc & (vec_step(x) - 1)]; \
> > +    y.sd = ((TYPE *) &x)[mask.sd & (vec_step(x) - 1)]; \
> > +    y.se = ((TYPE *) &x)[mask.se & (vec_step(x) - 1)]; \
> > +    y.sf = ((TYPE *) &x)[mask.sf & (vec_step(x) - 1)]; \
> >      return y; \
> >    }
> >  
> > @@ -1190,7 +1191,6 @@ DEF(float)
> >  #undef DEC4
> >  #undef DEC8
> >  #undef DEC16
> > -/////////////////////////////////////////////////////////////////////
> > ////////
> >  
> >  
> > //////////////////////////////////////////////////////////////////////
> > ///////
> >  // Synchronization functions
> > --
> > 1.8.1.2
> > 
> > _______________________________________________
> > Beignet mailing list
> > Beignet at lists.freedesktop.org
> > http://lists.freedesktop.org/mailman/listinfo/beignet

More information about the Beignet mailing list