include/native/SSEPlus_native_SSE4a.h

Go to the documentation of this file.

//
// Copyright (c) 2006-2008 Advanced Micro Devices, Inc. All Rights Reserved.
// This software is subject to the Apache v2.0 License.
//
#ifndef __SSEPLUS_NATIVE_SSE4A_H__
#define __SSEPLUS_NATIVE_SSE4A_H__

#include "../SSEPlus_base.h"
#include SSP_INCLUDE_FILE_SSE4a

SSP_FORCEINLINE __m128i ssp_extract_si64_SSE4A( __m128i a ,__m128i b )
{
    return _mm_extract_si64( a, b );
}
SSP_FORCEINLINE __m128i ssp_extracti_si64_SSE4A( __m128i a, int len, int ndx )
{
    return _mm_extracti_si64( a, len, ndx );
}
SSP_FORCEINLINE __m128i ssp_insert_si64_SSE4A( __m128i a, __m128i b )
{
    return _mm_insert_si64( a, b );
}
SSP_FORCEINLINE __m128i ssp_inserti_si64_SSE4A( __m128i a, __m128i b, int len, int ndx )
{
    return _mm_inserti_si64( a, b, len, ndx );
}
SSP_FORCEINLINE void ssp_stream_sd_SSE4A( double *dst ,__m128d src )
{
    _mm_stream_sd( dst, src );
}
SSP_FORCEINLINE void ssp_stream_ss_SSE4A( float *dst, __m128 src )
{
    _mm_stream_ss( dst, src );
}

SSP_FORCEINLINE unsigned short ssp_lzcnt16_SSE4A( unsigned short val )
{
    return __lzcnt16( val );
}
SSP_FORCEINLINE unsigned int ssp_lzcnt_SSE4A( unsigned int val )
{
    return __lzcnt( val );
}

#ifdef SYS64

SSP_FORCEINLINE ssp_u64 ssp_lzcnt64_SSE4A( ssp_u64 val )
{
    return __lzcnt64( val );
}
#endif

SSP_FORCEINLINE unsigned short ssp_popcnt16_SSE4A( unsigned short val )
{
    return __popcnt16( val );
}
SSP_FORCEINLINE unsigned int ssp_popcnt_SSE4A( unsigned int val )
{
    return __popcnt( val );
}

#ifdef SYS64

SSP_FORCEINLINE ssp_u64 ssp_popcnt64_SSE4A( ssp_u64 val )
{
    return __popcnt64( val );
}
#endif

#endif // __SSEPLUS_NATIVE_SSE4A_H__

---