NormalRandomReg.h

Go to the documentation of this file.

// NormalRandomReg.h:
// ------------
//

#ifndef __IPSDKUTIL_INSTRUCTIONSET_DETAIL_NORMALRANDOMREG_H__
#define __IPSDKUTIL_INSTRUCTIONSET_DETAIL_NORMALRANDOMREG_H__

#include <IPSDKUtil/IPSDKUtilExports.h>
#include <IPSDKUtil/InstructionSet/detail/AssignReg.h>
#include <IPSDKUtil/InstructionSet/detail/NormalRandomReg.h>
#include <IPSDKUtil/InstructionSet/Arithmetic/detail/SignedUnsignedOpReg.h>
#include <IPSDKUtil/InstructionSet/Arithmetic/cos.h>
#include <IPSDKUtil/InstructionSet/Arithmetic/log.h>
#include <IPSDKUtil/InstructionSet/Arithmetic/mul.h>
#include <IPSDKUtil/InstructionSet/Arithmetic/sin.h>
#include <IPSDKUtil/InstructionSet/Arithmetic/sqrt.h>
#include <IPSDKUtil/InstructionSet/RegType.h>
#include <IPSDKUtil/Tools/ForceInline.h>

#include <boost/math/constants/constants.hpp>

#include <math.h>

namespace ipsdk {
namespace simd {
namespace detail {


template <eInstructionSet::domain is, eInstructionSet::domain isFma>
struct NormalRandomReg {

static
IPSDK_FORCEINLINE
void
act(const typename RegType<is, ipReal32>::Type& rMean,
    const typename RegType<is, ipReal32>::Type& rSigma,
    const typename RegType<is, ipReal32>::Type& rUnifRand1,
    const typename RegType<is, ipReal32>::Type& rUnifRand2,
    typename RegType<is, ipReal32>::Type& rOut1,
    typename RegType<is, ipReal32>::Type& rOut2)
{
    typename RegType<is, ipReal32>::Type rTwo, rSqrt2LogU1;
    AssignReg<is, ipReal32>::act(rTwo, 2.0f);
    LogReg<is, isFma, ipReal32>::act(rUnifRand1, rSqrt2LogU1);
    MulReg<is, ipReal32>::act(rSqrt2LogU1, rTwo, rSqrt2LogU1);
    UnaryMinusReg<is, ipReal32>::act(rSqrt2LogU1, rSqrt2LogU1);
    SqrtReg<is, ipReal32>::act(rSqrt2LogU1, rSqrt2LogU1);

    const ipReal32 two_pi = 2.0f*boost::math::constants::pi<ipReal32>();
    typename RegType<is, ipReal32>::Type rTwoPi, rAngle;
    AssignReg<is, ipReal32>::act(rTwoPi, two_pi);
    MulReg<is, ipReal32>::act(rTwoPi, rUnifRand2, rAngle);

    typename RegType<is, ipReal32>::Type rCos, rSin;
    CosReg<is, isFma, ipReal32>::act(rAngle, rCos);
    SinReg<is, isFma, ipReal32>::act(rAngle, rSin);

    MulReg<is, ipReal32>::act(rCos, rSqrt2LogU1, rOut1);
    MulReg<is, ipReal32>::act(rOut1, rSigma, rOut1);
    AddReg<is, ipReal32>::act(rOut1, rMean, rOut1);

    MulReg<is, ipReal32>::act(rSin, rSqrt2LogU1, rOut2);
    MulReg<is, ipReal32>::act(rOut2, rSigma, rOut2);
    AddReg<is, ipReal32>::act(rOut2, rMean, rOut2);
}
};


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

#endif // __IPSDKUTIL_INSTRUCTIONSET_DETAIL_NORMALRANDOMREG_H__