AcosReg.h

Go to the documentation of this file.

// AcosReg.h:
// -------------------
//

#ifndef __IPSDKUTIL_INSTRUCTIONSET_ARITHMETIC_DETAIL_COMMON_ACOSREG_H__
#define __IPSDKUTIL_INSTRUCTIONSET_ARITHMETIC_DETAIL_COMMON_ACOSREG_H__

#include <IPSDKUtil/IPSDKUtilExports.h>
#include <IPSDKUtil/InstructionSet/RegType.h>
#include <IPSDKUtil/InstructionSet/Arithmetic/abs.h>
#include <IPSDKUtil/InstructionSet/Arithmetic/asin.h>
#include <IPSDKUtil/InstructionSet/Arithmetic/div.h>
#include <IPSDKUtil/InstructionSet/Arithmetic/mul.h>
#include <IPSDKUtil/InstructionSet/Arithmetic/sub.h>
#include <IPSDKUtil/InstructionSet/Comparison/isGreater.h>
#include <IPSDKUtil/InstructionSet/Logical/ifElse.h>
#include <IPSDKUtil/InstructionSet/assign.h>

#include <IPSDKUtil/Tools/ForceInline.h>

namespace ipsdk {
namespace simd {
namespace detail {


template <eInstructionSet::domain IS, typename T>
struct AcosReg<
    IS,
    T,
    typename boost::enable_if<
        typename boost::mpl::and_<
            typename boost::mpl::not_equal_to<
                boost::mpl::int_<IS>,
                boost::mpl::int_<eInstructionSet::eIS_Standard>
            >::type,
            typename boost::is_float<T>::type
        >::type
   >::type
>            
{
    static IPSDK_FORCEINLINE
    typename RegType<IS, T>::Type
    act(const typename RegType<IS, T>::Type& in)
    {
        typename RegType<IS, T>::Type out;
        act(in, out);
        return out;
    }

    static IPSDK_FORCEINLINE
    void
    act(const typename RegType<IS, T>::Type& in,
        typename RegType<IS, T>::Type& out)
    {
        typedef typename RegType<IS, T>::Type Reg;
        typedef typename RegMaskType<IS, T>::Type MaskReg;

        Reg cste0P5Reg, cstePiD2Reg;
        simd::detail::AssignReg<IS, T>::act(cste0P5Reg, static_cast<T>(0.5));
        simd::detail::AssignReg<IS, T>::act(cstePiD2Reg, static_cast<T>(M_PI_2));

        // check for input data range
        Reg valReg;
        MaskReg testReg;
        simd::detail::IsGreaterReg<IS, T>::act(in, cste0P5Reg, testReg);  // if (x > 0.5)

                                                         // std::sqrt(0.5-0.5*x)
        simd::detail::MulReg<IS, T>::act(cste0P5Reg, in, valReg);
        simd::detail::SubReg<IS, T>::act(cste0P5Reg, valReg, valReg);
        simd::detail::SqrtReg<IS, T>::act(valReg, valReg);

        // select value for asin computation
        Reg asinReg;
        simd::detail::IfElseReg<IS, T>::act(testReg, valReg, in, valReg);
        simd::detail::AsinReg<IS, T>::act(valReg, asinReg);

        // compute solutions
        Reg valReg1, valReg2;
        simd::detail::AddReg<IS, T>::act(asinReg, asinReg, valReg1);      // 2*asin_std(std::sqrt(0.5-0.5*x))
        simd::detail::SubReg<IS, T>::act(cstePiD2Reg, asinReg, valReg2); // M_PI_2 - asin_std(x)
        simd::detail::IfElseReg<IS, T>::act(testReg, valReg1, valReg2, out);

        
    }
};


} // end of namespace detail
} // end of namespace simd
} // end of namespace ipsdk

#endif // __IPSDKUTIL_INSTRUCTIONSET_ARITHMETIC_DETAIL_COMMON_ACOSREG_H__