[Beignet] [PATCH 3/3] add llvm intrinsic call translate.

[Zhigang Gong]

On Thu, Jan 15, 2015 at 08:43:50AM +0000, Song, Ruiling wrote:
> > +          case Intrinsic::copysign:
> > +          {
> > +            ir::Type srcType = getType(ctx, I.getType());
> > +            const ir::Register dst = this->getRegister(&I);
> > +            const ir::Register src0 = this->getRegister(I.getOperand(0));
> > +            const ir::Register src1 = this->getRegister(I.getOperand(1));
> > +            const ir::Register cmp = ctx.reg(ir::FAMILY_BOOL);
> > +
> > +            const ir::Register tmp1 = ctx.reg(getFamily(srcType));
> > +            const ir::Register tmp2 = ctx.reg(getFamily(srcType));
> > +
> > +            const ir::RegisterFamily family = getFamily(srcType);
> > +            const ir::ImmediateIndex zero =
> > ctx.newFloatImmediate((float)0.0);
> > +            const ir::Register zeroReg = ctx.reg(family);
> > +            ctx.LOADI(srcType, zeroReg, zero);
> > +
> > +            ctx.GE(srcType, cmp, src1, zeroReg);
> > +            ctx.ALU1(ir::OP_ABS, srcType, tmp1, src0);
> > +            ctx.SUB(srcType, tmp2, zeroReg, tmp1);
> > +            ctx.SEL(srcType, dst, cmp, tmp1, tmp2);
> For the copysign implementation, I would still prefer you use current implementation in backend/src/libocl/tmpl/ocl_math.tmpl.cl
> Gen hardware does not support denorm number. So when the input y is denorm number, the GE compare could not return exact result.
> The current implementation in libocl is robust enough for all corner cases. What do you think?
Right, our original copysign's implementation should be used rather than this new one. 

And not only just this one, but also include all the upcoming intrinsics,
if we have to implement in the backend, we need to keep using the original
version. I suggested xionghu to do the intrinsic lowering in an earlier
pass to replace these intrinsics call to normal calls which is implemented
in beignet's bitcode library.

For those intrinsics which are very simple and could map to one or two IR instructions,
we prefer to handle it at gen writer pass. For all the other complex intrinsics, we need
to lower them in the above intrinsics lowering pass and use our current implementation
directly.

Any comments?

> > +          }
> > +          break;
> >            default: NOT_IMPLEMENTED;
> >          }
> >        } else {
> > +
> > +        LibFunc::Func Func;
> > +        if(isLibFuncFunc(&I, Func))
> > +        {
> > +          switch (Func) {
> > +            case LibFunc::copysignf:
> > +              {
> > +                ir::Type srcType = getType(ctx, I.getType());
> > +                const ir::Register dst = this->getRegister(&I);
> > +                const ir::Register src0 =
> > this->getRegister(I.getOperand(0));
> > +                const ir::Register src1 =
> > this->getRegister(I.getOperand(1));
> > +                const ir::Register cmp = ctx.reg(ir::FAMILY_BOOL);
> > +
> > +                const ir::Register tmp1 = ctx.reg(getFamily(srcType));
> > +                const ir::Register tmp2 = ctx.reg(getFamily(srcType));
> > +
> > +                const ir::RegisterFamily family = getFamily(srcType);
> > +                const ir::ImmediateIndex zero =
> > ctx.newFloatImmediate((float)0.0);
> > +                const ir::Register zeroReg = ctx.reg(family);
> > +                ctx.LOADI(srcType, zeroReg, zero);
> > +
> > +                ctx.GE(srcType, cmp, src1, zeroReg);
> > +                ctx.ALU1(ir::OP_ABS, srcType, tmp1, src0);
> > +                ctx.SUB(srcType, tmp2, zeroReg, tmp1);
> 
> 
> And also here.
> 
> > +                ctx.SEL(srcType, dst, cmp, tmp1, tmp2);
> > +              }
> > +              break;
> > +            default:
> > +              GBE_ASSERTM(false, "Unsupported libFuncs");
> > +          }
> > +          return;
> > +        }
> > + 
> 
> _______________________________________________
> Beignet mailing list
> Beignet at lists.freedesktop.org
> http://lists.freedesktop.org/mailman/listinfo/beignet