SetImpl.h

Go to the documentation of this file.

// SetImpl.h:
// ------------
//

#ifndef __IPSDKUTIL_INSTRUCTIONSET_DETAIL_SSE2_SETIMPL_H__
#define __IPSDKUTIL_INSTRUCTIONSET_DETAIL_SSE2_SETIMPL_H__

#include <IPSDKUtil/IPSDKUtilExports.h>
#include <IPSDKUtil/Tools/ForceInline.h>
#include <IPSDKUtil/InstructionSet/NbEltsPerReg.h>
#include <IPSDKUtil/InstructionSet/unload.h>

namespace ipsdk {
namespace simd {
namespace detail {


template <typename T>
struct SetImpl<eInstructionSet::eIS_Sse2, T> {
    static IPSDK_FORCEINLINE
    void act(T* buffer, T value, ipUInt64 nbElts)
    {
        const ipUInt64 nbAlignedElts =
            nbElts - (nbElts%NbEltsPerReg<ePackType::ePT_Sse, T>::Value);
        typedef typename Sse2Type<T>::Type RegT;
        RegT reg;
        AssignReg<eInstructionSet::eIS_Sse2, T>::act(reg, value);

        for(ipUInt64 i=0; i<nbAlignedElts; i+=NbEltsPerReg<ePackType::ePT_Sse, T>::Value)
            UnloadReg<eInstructionSet::eIS_Sse2, T>::act(reg, buffer + i);

        for(ipUInt64 i=nbAlignedElts; i<nbElts; ++i)
            buffer[i] = value;
    }
};


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

#endif // __IPSDKUTIL_INSTRUCTIONSET_DETAIL_SSE2_SETIMPL_H__