[Beignet] [PATCH v2] fix llvm.trunc.float instruction bug.

[Zhigang Gong]

Right, for scalar value, magnitude is the same as absolute value.
We should use RNDZ here.
Thanks for the valuable comments.

Xionghu, no need to send a new version I will fix it and push the patch.

Thanks.


On Fri, Jan 16, 2015 at 03:21:48AM +0000, Song, Ruiling wrote:
> llvm manual page says llvm.trunc share exactly same behavior with libm trunc:
> "Semantics of llvm.trunc:
> This function returns the same values as the libm trunc functions would, and handles error conditions in the same way."
> 
> So, I think magnitude shows same meaning with absolute value here. Generally magnitude means the length of a vector, so it is non-negative.
> 
> > -----Original Message-----
> > From: Zhigang Gong [mailto:zhigang.gong at linux.intel.com]
> > Sent: Friday, January 16, 2015 10:45 AM
> > To: Song, Ruiling; Luo, Xionghu; beignet at lists.freedesktop.org
> > Cc: Luo, Xionghu
> > Subject: RE: [Beignet] [PATCH v2] fix llvm.trunc.float instruction bug.
> > 
> > Should be RNDD, please see the difference between the two definitions
> > below:
> > 
> > Libm trunk():
> >   These functions round x to the nearest integer not larger in *absolute*
> > value.
> > 
> > Llvm.trunc:
> > The ‘llvm.trunc.*‘ intrinsics returns the operand rounded to the nearest
> > integer not larger in magnitude than the operand.
> > 
> > 
> > > -----Original Message-----
> > > From: Beignet [mailto:beignet-bounces at lists.freedesktop.org] On Behalf
> > > Of Song, Ruiling
> > > Sent: Friday, January 16, 2015 10:36 AM
> > > To: Luo, Xionghu; beignet at lists.freedesktop.org
> > > Cc: Luo, Xionghu
> > > Subject: Re: [Beignet] [PATCH v2] fix llvm.trunc.float instruction bug.
> > >
> > > Should be RNDZ. The llvm.trunc is same as libm trunc(), which means
> > > rounding toward zero.
> > >
> > > > -----Original Message-----
> > > > From: Beignet [mailto:beignet-bounces at lists.freedesktop.org] On
> > > > Behalf Of xionghu.luo at intel.com
> > > > Sent: Thursday, January 15, 2015 7:05 PM
> > > > To: beignet at lists.freedesktop.org
> > > > Cc: Luo, Xionghu
> > > > Subject: [Beignet] [PATCH v2] fix llvm.trunc.float instruction bug.
> > > >
> > > > From: Luo <xionghu.luo at intel.com>
> > > >
> > > > float to float trunc should use RNDU IR instruction.
> > > >
> > > > v2: fix typo. should be RNDD instead of RNDU.
> > > > Signed-off-by: Luo <xionghu.luo at intel.com>
> > > > ---
> > > >  backend/src/llvm/llvm_gen_backend.cpp | 4 +---
> > > >  1 file changed, 1 insertion(+), 3 deletions(-)
> > > >
> > > > diff --git a/backend/src/llvm/llvm_gen_backend.cpp
> > > > b/backend/src/llvm/llvm_gen_backend.cpp
> > > > index 53dec0a..ee5b6a3 100644
> > > > --- a/backend/src/llvm/llvm_gen_backend.cpp
> > > > +++ b/backend/src/llvm/llvm_gen_backend.cpp
> > > > @@ -3310,11 +3310,9 @@ error:
> > > >              ir::Type srcType = getType(ctx, llvmSrcType);
> > > >              GBE_ASSERT(srcType == dstType);
> > > >
> > > > -            const ir::Register tmp = ctx.reg(getFamily(ir::TYPE_S32));
> > > >              const ir::Register dst = this->getRegister(&I);
> > > >              const ir::Register src =
> > this->getRegister(I.getOperand(0));
> > > > -            ctx.CVT(ir::TYPE_S32, srcType, tmp, src);
> > > > -            ctx.CVT(dstType, ir::TYPE_S32, dst, tmp);
> > > > +            ctx.RNDD(dstType, dst, src);
> > > >            }
> > > >            break;
> > > >            case Intrinsic::copysign:
> > > > --
> > > > 1.9.1
> > > >
> > > > _______________________________________________
> > > > 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
> 
> _______________________________________________
> Beignet mailing list
> Beignet at lists.freedesktop.org
> http://lists.freedesktop.org/mailman/listinfo/beignet