[Beignet] [PATCH] GBE: disable MulAdd pattern in instruction selection temporarily.

[Zhigang Gong]

That's ok for me. Thanks.

On Tue, Nov 05, 2013 at 08:38:02AM +0000, Song, Ruiling wrote:
> After my investigation, I find the unite test depends on mad optimization which is not quite reasonable.
> So, I change the unit test to directly call mad(). And also directly map mad() builtin to ISA mad.
> I have sent the patches, welcome any comments. Thanks!
> 
> Ruiling
> -----Original Message-----
> From: Song, Ruiling 
> Sent: Monday, November 04, 2013 4:24 PM
> To: Zhigang Gong
> Cc: beignet at lists.freedesktop.org
> Subject: RE: [Beignet] [PATCH] GBE: disable MulAdd pattern in instruction selection temporarily.
> 
> In fact, even if user set 'FP_CONTRACT ON', we can also choose not do 'mad optimization.'
> It depends on us whether or not do the contraction. "Contract on" is just a hint.
> Obviously contract on will get better speed and accuracy.
> 
> I think the compiler_menger_sponge_no_shadow fail because of accuracy. Its reference is got with mad instruction.
> I think we can use the new result to replace the reference, there should be not too much difference. If they differ too much, then it would be a bug.
> 
> 
> -----Original Message-----
> From: Zhigang Gong [mailto:zhigang.gong at linux.intel.com]
> Sent: Monday, November 04, 2013 3:41 PM
> To: Song, Ruiling
> Cc: beignet at lists.freedesktop.org
> Subject: Re: [Beignet] [PATCH] GBE: disable MulAdd pattern in instruction selection temporarily.
> 
> This patch breaks utests compiler_menger_sponge_no_shadow.
> Actually, my first feeling of this patch is that it assumes that FP_CONTRACT is always off and may be cause problem if this assumption is incorrect. The unit test case compiler_menger_sponge_no_shadow should be one of that case.
> 
> Any thoughts?
> 
> On Fri, Nov 01, 2013 at 02:16:08PM +0800, Ruiling Song wrote:
> > The story starts from 'FP_CONTRACT'. In c99 spec, it describes 
> > constract expression as:
> > "A floating expression may be contracted, that is, evaluated as though 
> > it were an atomic operation, thereby omitting rounding errors implied 
> > by the source code and the expression evaluation method."
> > 
> > But user can use 'pragma FP_CONTRACT OFF' to disable float 
> > contraction, in which condition, we should not do contraction like mad optimization.
> > In SPIR 1.2, named metadata 'opencl.enable.FP_CONTRACT' will be used to do this.
> > When Clang is ready, we need refine the backend logic.
> > 
> > Signed-off-by: Ruiling Song <ruiling.song at intel.com>
> > ---
> >  backend/src/backend/gen_insn_selection.cpp |    3 +++
> >  1 file changed, 3 insertions(+)
> > 
> > diff --git a/backend/src/backend/gen_insn_selection.cpp
> > b/backend/src/backend/gen_insn_selection.cpp
> > index 7eae7ca..1f25f0e 100644
> > --- a/backend/src/backend/gen_insn_selection.cpp
> > +++ b/backend/src/backend/gen_insn_selection.cpp
> > @@ -1856,6 +1856,9 @@ namespace gbe
> >      {
> >        using namespace ir;
> >  
> > +      // XXX TODO: we need a clean support of FP_CONTRACT to remove below line 'return false'
> > +      // if 'pragma FP_CONTRACT OFF' is used in cl kernel, we should not do mad optimization.
> > +      return false;
> >        // MAD tend to increase liveness of the sources (since there are three of
> >        // them). TODO refine this strategy. Well, we should be able at least to
> >        // evaluate per basic block register pressure and selectively 
> > enable
> > --
> > 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