"SSEPlus": SSE4.1

SSP_FORCEINLINE __m128i ssp_blend_epi16_SSE4_1	(	__m128i	a,
		__m128i	b,
		const int	mask
	)

Native implementation of _mm_blend_epi16 [SSE4.1]. (Searches MSDN)

Definition at line 58 of file SSEPlus_native_SSE4.1.h.

00059 {
00060     switch( mask & 0xFF )
00061     {
00062         CASE_256( _mm_blend_epi16, a, b );
00063     }
00064 }

SSP_FORCEINLINE __m128d ssp_blend_pd_SSE4_1	(	__m128d	a,
		__m128d	b,
		const int	mask
	)

Native implementation of _mm_blend_pd [SSE4.1]. (Searches MSDN)

Definition at line 84 of file SSEPlus_native_SSE4.1.h.

00085 {
00086     switch(mask&0x3)
00087     {
00088         CASE_4( _mm_blend_pd, a, b );
00089     }
00090 }

SSP_FORCEINLINE __m128 ssp_blend_ps_SSE4_1	(	__m128	a,
		__m128	b,
		const int	mask
	)

Native implementation of _mm_blend_ps [SSE4.1]. (Searches MSDN)

Definition at line 71 of file SSEPlus_native_SSE4.1.h.

00072 {
00073     switch( mask & 0x0F )
00074     {
00075         CASE_16( _mm_blend_ps, a, b );
00076     }
00077 }

SSP_FORCEINLINE __m128i ssp_blendv_epi8_SSE4_1	(	__m128i	a,
		__m128i	b,
		__m128i	mask
	)

Native implementation of _mm_blendv_epi8 [SSE4.1]. (Searches MSDN)

Definition at line 66 of file SSEPlus_native_SSE4.1.h.

00067 {
00068     return _mm_blendv_epi8( a, b, mask );
00069 }

SSP_FORCEINLINE __m128d ssp_blendv_pd_SSE4_1	(	__m128d	a,
		__m128d	b,
		__m128d	mask
	)

Native implementation of _mm_blendv_pd [SSE4.1]. (Searches MSDN)

Definition at line 92 of file SSEPlus_native_SSE4.1.h.

00093 {
00094     return _mm_blendv_pd( a, b, mask);
00095 }

SSP_FORCEINLINE __m128 ssp_blendv_ps_SSE4_1	(	__m128	a,
		__m128	b,
		__m128	mask
	)

Native implementation of _mm_blendv_ps [SSE4.1]. (Searches MSDN)

Definition at line 79 of file SSEPlus_native_SSE4.1.h.

00080 {
00081     return _mm_blendv_ps( a, b, mask);
00082 }

SSP_FORCEINLINE __m128d ssp_ceil_pd_SSE4_1 ( __m128d a )

Native implementation of _mm_ceil_pd [SSE4.1]. (Searches MSDN)

Definition at line 18 of file SSEPlus_native_SSE4.1.h.

00019 {
00020     return _mm_ceil_pd( a );
00021 }

SSP_FORCEINLINE __m128 ssp_ceil_ps_SSE4_1 ( __m128 a )

Native implementation of _mm_ceil_ps [SSE4.1]. (Searches MSDN)

Definition at line 38 of file SSEPlus_native_SSE4.1.h.

00039 {
00040     return _mm_ceil_ps( a );
00041 }

SSP_FORCEINLINE __m128d ssp_ceil_sd_SSE4_1	(	__m128d	dst,
		__m128d	a
	)

Native implementation of _mm_ceil_sd [SSE4.1]. (Searches MSDN)

Definition at line 23 of file SSEPlus_native_SSE4.1.h.

00024 {
00025     return _mm_ceil_sd( dst, a );
00026 }

SSP_FORCEINLINE __m128 ssp_ceil_ss_SSE4_1	(	__m128	dst,
		__m128	a
	)

Native implementation of _mm_ceil_ss [SSE4.1]. (Searches MSDN)

Definition at line 43 of file SSEPlus_native_SSE4.1.h.

00044 {
00045     return _mm_ceil_ss( dst, a );
00046 }

SSP_FORCEINLINE __m128i ssp_cmpeq_epi64_SSE4_1	(	__m128i	a,
		__m128i	b
	)

Native implementation of _mm_cmpeq_epi64 [SSE4.1]. (Searches MSDN)

Definition at line 113 of file SSEPlus_native_SSE4.1.h.

00114 {
00115     return _mm_cmpeq_epi64( a, b);
00116 }

SSP_FORCEINLINE __m128i ssp_cvtepi16_epi32_SSE4_1 ( __m128i shortValues )

Native implementation of _mm_cvtepi16_epi32 [SSE4.1]. (Searches MSDN)

Definition at line 230 of file SSEPlus_native_SSE4.1.h.

00231 {
00232     return _mm_cvtepi16_epi32( shortValues );
00233 }

SSP_FORCEINLINE __m128i ssp_cvtepi16_epi64_SSE4_1 ( __m128i shortValues )

Native implementation of _mm_cvtepi16_epi64 [SSE4.1]. (Searches MSDN)

Definition at line 245 of file SSEPlus_native_SSE4.1.h.

00246 {
00247     return _mm_cvtepi16_epi64( shortValues );
00248 }

SSP_FORCEINLINE __m128i ssp_cvtepi32_epi64_SSE4_1 ( __m128i intValues )

Native implementation of _mm_cvtepi32_epi64 [SSE4.1]. (Searches MSDN)

Definition at line 240 of file SSEPlus_native_SSE4.1.h.

00241 {
00242     return _mm_cvtepi32_epi64( intValues );
00243 }

SSP_FORCEINLINE __m128i ssp_cvtepi8_epi16_SSE4_1 ( __m128i byteValues )

Native implementation of _mm_cvtepi8_epi16 [SSE4.1]. (Searches MSDN)

Definition at line 250 of file SSEPlus_native_SSE4.1.h.

00251 {
00252     return _mm_cvtepi8_epi16( byteValues );
00253 }

SSP_FORCEINLINE __m128i ssp_cvtepi8_epi32_SSE4_1 ( __m128i byteValues )

Native implementation of _mm_cvtepi8_epi32 [SSE4.1]. (Searches MSDN)

Definition at line 225 of file SSEPlus_native_SSE4.1.h.

00226 {
00227     return _mm_cvtepi8_epi32( byteValues );
00228 }

SSP_FORCEINLINE __m128i ssp_cvtepi8_epi64_SSE4_1 ( __m128i byteValues )

Native implementation of _mm_cvtepi8_epi64 [SSE4.1]. (Searches MSDN)

Definition at line 235 of file SSEPlus_native_SSE4.1.h.

00236 {
00237     return _mm_cvtepi8_epi64( byteValues );
00238 }

SSP_FORCEINLINE __m128i ssp_cvtepu16_epi32_SSE4_1 ( __m128i shortValues )

Native implementation of _mm_cvtepu16_epi32 [SSE4.1]. (Searches MSDN)

Definition at line 260 of file SSEPlus_native_SSE4.1.h.

00261 {
00262     return _mm_cvtepu16_epi32( shortValues );
00263 }

SSP_FORCEINLINE __m128i ssp_cvtepu16_epi64_SSE4_1 ( __m128i shortValues )

Native implementation of _mm_cvtepu16_epi64 [SSE4.1]. (Searches MSDN)

Definition at line 275 of file SSEPlus_native_SSE4.1.h.

00276 {
00277     return _mm_cvtepu16_epi64( shortValues );
00278 }

SSP_FORCEINLINE __m128i ssp_cvtepu32_epi64_SSE4_1 ( __m128i intValues )

Native implementation of _mm_cvtepu32_epi64 [SSE4.1]. (Searches MSDN)

Definition at line 270 of file SSEPlus_native_SSE4.1.h.

00271 {
00272     return _mm_cvtepu32_epi64( intValues );
00273 }

SSP_FORCEINLINE __m128i ssp_cvtepu8_epi16_SSE4_1 ( __m128i byteValues )

Native implementation of _mm_cvtepu8_epi16 [SSE4.1]. (Searches MSDN)

Definition at line 280 of file SSEPlus_native_SSE4.1.h.

00281 {
00282     return _mm_cvtepu8_epi16( byteValues );
00283 }

SSP_FORCEINLINE __m128i ssp_cvtepu8_epi32_SSE4_1 ( __m128i byteValues )

Native implementation of _mm_cvtepu8_epi32 [SSE4.1]. (Searches MSDN)

Definition at line 255 of file SSEPlus_native_SSE4.1.h.

00256 {
00257     return _mm_cvtepu8_epi32( byteValues );
00258 }

SSP_FORCEINLINE __m128i ssp_cvtepu8_epi64_SSE4_1 ( __m128i shortValues )

Native implementation of _mm_cvtepu8_epi64 [SSE4.1]. (Searches MSDN)

Definition at line 265 of file SSEPlus_native_SSE4.1.h.

00266 {
00267     return _mm_cvtepu8_epi64( shortValues );
00268 }

SSP_FORCEINLINE __m128d ssp_dp_pd_SSE4_1	(	__m128d	a,
		__m128d	b,
		const int	mask
	)

Native implementation of _mm_dp_pd [SSE4.1]. (Searches MSDN)

Definition at line 105 of file SSEPlus_native_SSE4.1.h.

00106 {
00107     switch( mask & 0x3F )
00108     {
00109         CASE_128( _mm_dp_pd, a, b );
00110     }
00111 }

SSP_FORCEINLINE __m128 ssp_dp_ps_SSE4_1	(	__m128	a,
		__m128	b,
		const int	mask
	)

Native implementation of _mm_dp_ps [SSE4.1]. (Searches MSDN)

Definition at line 97 of file SSEPlus_native_SSE4.1.h.

00098 {
00099     switch( mask & 0xFF )
00100     {
00101         CASE_256( _mm_dp_ps, a, b );
00102     }
00103 }

SSP_FORCEINLINE int ssp_extract_epi32_SSE4_1	(	__m128i	src,
		const int	ndx
	)

Native implementation of _mm_extract_epi32 [SSE4.1]. (Searches MSDN)

Definition at line 210 of file SSEPlus_native_SSE4.1.h.

00211 {
00212     switch( ndx & 0x3 )
00213     {
00214         CASE_4( _mm_extract_epi32, src );
00215     }
00216 }

SSP_FORCEINLINE int ssp_extract_epi8_SSE4_1	(	__m128i	src,
		const int	ndx
	)

Native implementation of _mm_extract_epi8 [SSE4.1]. (Searches MSDN)

Definition at line 202 of file SSEPlus_native_SSE4.1.h.

00203 {
00204     switch( ndx & 0xF )
00205     {
00206         CASE_16( _mm_extract_epi8, src );
00207     }
00208 }

SSP_FORCEINLINE int ssp_extract_ps_SSE4_1	(	__m128	src,
		const int	ndx
	)

Native implementation of _mm_extract_ps [SSE4.1]. (Searches MSDN)

Definition at line 177 of file SSEPlus_native_SSE4.1.h.

00178 {
00179     switch(ndx&0x3)
00180     {
00181         CASE_4( _mm_extract_ps, src )
00182     }
00183 }

SSP_FORCEINLINE __m128d ssp_floor_pd_SSE4_1 ( __m128d a )

Native implementation of _mm_floor_pd [SSE4.1]. (Searches MSDN)

Definition at line 28 of file SSEPlus_native_SSE4.1.h.

00029 {
00030     return _mm_floor_pd( a );
00031 }

SSP_FORCEINLINE __m128 ssp_floor_ps_SSE4_1 ( __m128 a )

Native implementation of _mm_floor_ps [SSE4.1]. (Searches MSDN)

Definition at line 48 of file SSEPlus_native_SSE4.1.h.

00049 {
00050     return _mm_floor_ps( a );
00051 }

SSP_FORCEINLINE __m128d ssp_floor_sd_SSE4_1	(	__m128d	dst,
		__m128d	a
	)

Native implementation of _mm_floor_sd [SSE4.1]. (Searches MSDN)

Definition at line 33 of file SSEPlus_native_SSE4.1.h.

00034 {
00035     return _mm_floor_sd( dst, a );
00036 }

SSP_FORCEINLINE __m128 ssp_floor_ss_SSE4_1	(	__m128	dst,
		__m128	a
	)

Native implementation of _mm_floor_ss [SSE4.1]. (Searches MSDN)

Definition at line 53 of file SSEPlus_native_SSE4.1.h.

00054 {
00055     return _mm_floor_ss( dst, a );
00056 }

SSP_FORCEINLINE __m128i ssp_insert_epi32_SSE4_1	(	__m128i	dst,
		int	s,
		const int	ndx
	)

Native implementation of _mm_insert_epi32 [SSE4.1]. (Searches MSDN)

Definition at line 193 of file SSEPlus_native_SSE4.1.h.

00194 {
00195     switch( ndx & 0x3 )
00196     {
00197         CASE_4( _mm_insert_epi32, dst, s );
00198     }
00199 }

SSP_FORCEINLINE __m128i ssp_insert_epi8_SSE4_1	(	__m128i	dst,
		int	s,
		const int	ndx
	)

Native implementation of _mm_insert_epi8 [SSE4.1]. (Searches MSDN)

Definition at line 185 of file SSEPlus_native_SSE4.1.h.

00186 {
00187     switch( ndx & 0xF )
00188     {
00189         CASE_16( _mm_insert_epi8, dst, s );
00190     }
00191 }

SSP_FORCEINLINE __m128 ssp_insert_ps_SSE4_1	(	__m128	dst,
		__m128	src,
		const int	ndx
	)

Native implementation of _mm_insert_ps [SSE4.1]. (Searches MSDN)

Definition at line 169 of file SSEPlus_native_SSE4.1.h.

00170 {
00171     switch( ndx & 0xFF )
00172     {
00173         CASE_256( _mm_insert_ps, dst, src );
00174     }
00175 }

SSP_FORCEINLINE __m128i ssp_max_epi32_SSE4_1	(	__m128i	a,
		__m128i	b
	)

Native implementation of _mm_max_epi32 [SSE4.1]. (Searches MSDN)

Definition at line 143 of file SSEPlus_native_SSE4.1.h.

00144 {
00145     return _mm_max_epi32( a, b);
00146 }

SSP_FORCEINLINE __m128i ssp_max_epi8_SSE4_1	(	__m128i	a,
		__m128i	b
	)

Native implementation of _mm_max_epi8 [SSE4.1]. (Searches MSDN)

Definition at line 123 of file SSEPlus_native_SSE4.1.h.

00124 {
00125     return _mm_max_epi8( a, b);
00126 }

SSP_FORCEINLINE __m128i ssp_max_epu16_SSE4_1	(	__m128i	a,
		__m128i	b
	)

Native implementation of _mm_max_epu16 [SSE4.1]. (Searches MSDN)

Definition at line 133 of file SSEPlus_native_SSE4.1.h.

00134 {
00135     return _mm_max_epu16( a, b);
00136 }

SSP_FORCEINLINE __m128i ssp_max_epu32_SSE4_1	(	__m128i	a,
		__m128i	b
	)

Native implementation of _mm_max_epu32 [SSE4.1]. (Searches MSDN)

Definition at line 153 of file SSEPlus_native_SSE4.1.h.

00154 {
00155     return _mm_max_epu32( a, b);
00156 }

SSP_FORCEINLINE __m128i ssp_min_epi32_SSE4_1	(	__m128i	a,
		__m128i	b
	)

Native implementation of _mm_min_epi32 [SSE4.1]. (Searches MSDN)

Definition at line 138 of file SSEPlus_native_SSE4.1.h.

00139 {
00140     return _mm_min_epi32( a, b);
00141 }

SSP_FORCEINLINE __m128i ssp_min_epi8_SSE4_1	(	__m128i	a,
		__m128i	b
	)

Native implementation of _mm_min_epi8 [SSE4.1]. (Searches MSDN)

Definition at line 118 of file SSEPlus_native_SSE4.1.h.

00119 {
00120     return _mm_min_epi8( a, b);
00121 }

SSP_FORCEINLINE __m128i ssp_min_epu16_SSE4_1	(	__m128i	a,
		__m128i	b
	)

Native implementation of _mm_min_epu16 [SSE4.1]. (Searches MSDN)

Definition at line 128 of file SSEPlus_native_SSE4.1.h.

00129 {
00130     return _mm_min_epu16( a, b);
00131 }

SSP_FORCEINLINE __m128i ssp_min_epu32_SSE4_1	(	__m128i	a,
		__m128i	b
	)

Native implementation of _mm_min_epu32 [SSE4.1]. (Searches MSDN)

Definition at line 148 of file SSEPlus_native_SSE4.1.h.

00149 {
00150     return _mm_min_epu32( a, b);
00151 }

SSP_FORCEINLINE __m128i ssp_minpos_epu16_SSE4_1 ( __m128i shortValues )

Native implementation of _mm_minpos_epu16 [SSE4.1]. (Searches MSDN)

Definition at line 219 of file SSEPlus_native_SSE4.1.h.

00220 {
00221     return _mm_minpos_epu16( shortValues );
00222 }

SSP_FORCEINLINE __m128i ssp_mpsadbw_epu8_SSE4_1	(	__m128i	a,
		__m128i	b,
		const int	msk
	)

Native implementation of _mm_mpsadbw_epu8 [SSE4.1]. (Searches MSDN)

Definition at line 290 of file SSEPlus_native_SSE4.1.h.

00291 {
00292     switch( msk & 0x7 )
00293     {
00294         CASE_8( _mm_mpsadbw_epu8, a, b );
00295     }
00296 }

SSP_FORCEINLINE __m128i ssp_mul_epi32_SSE4_1	(	__m128i	a,
		__m128i	b
	)

Native implementation of _mm_mul_epi32 [SSE4.1]. (Searches MSDN)

Definition at line 163 of file SSEPlus_native_SSE4.1.h.

00164 {
00165     return _mm_mul_epi32( a, b);
00166 }

SSP_FORCEINLINE __m128i ssp_mullo_epi32_SSE4_1	(	__m128i	a,
		__m128i	b
	)

Native implementation of _mm_mullo_epi32 [SSE4.1]. (Searches MSDN)

Definition at line 158 of file SSEPlus_native_SSE4.1.h.

00159 {
00160     return _mm_mullo_epi32( a, b);
00161 }

SSP_FORCEINLINE __m128i ssp_packus_epi32_SSE4_1	(	__m128i	a,
		__m128i	b
	)

Native implementation of _mm_packus_epi32 [SSE4.1]. (Searches MSDN)

Definition at line 285 of file SSEPlus_native_SSE4.1.h.

00286 {
00287     return _mm_packus_epi32( a, b );
00288 }

SSP_FORCEINLINE __m128d ssp_round_pd_SSE4_1	(	__m128d	a,
		int	iRoundMode
	)

Native implementation of _mm_round_pd/ roundpd [SSE4.1 and SSE5]. (SSE5 .pdf documentation here)

Definition at line 325 of file SSEPlus_native_SSE4.1.h.

00326 {
00327     switch( iRoundMode & 0xF )
00328     {
00329         CASE_16( _mm_round_pd, a );
00330     }
00331 }

SSP_FORCEINLINE __m128 ssp_round_ps_SSE4_1	(	__m128	a,
		int	iRoundMode
	)

Native implementation of _mm_round_ps/ roundps [SSE4.1 and SSE5]. (SSE5 .pdf documentation here)

Definition at line 341 of file SSEPlus_native_SSE4.1.h.

00342 {
00343     switch( iRoundMode & 0xF )
00344     {
00345         CASE_16( _mm_round_ps, a );
00346     }
00347 }

SSP_FORCEINLINE __m128d ssp_round_sd_SSE4_1	(	__m128d	dst,
		__m128d	a,
		int	iRoundMode
	)

Native implementation of _mm_round_sd/ roundsd [SSE4.1 and SSE5]. (SSE5 .pdf documentation here)

Definition at line 333 of file SSEPlus_native_SSE4.1.h.

00334 {
00335     switch( iRoundMode & 0xF )
00336     {
00337         CASE_16( _mm_round_sd, dst, a );
00338     }
00339 }

SSP_FORCEINLINE __m128 ssp_round_ss_SSE4_1	(	__m128	dst,
		__m128	a,
		int	iRoundMode
	)

Native implementation of _mm_round_ss/ roundss [SSE4.1 and SSE5]. (SSE5 .pdf documentation here)

Definition at line 349 of file SSEPlus_native_SSE4.1.h.

00350 {
00351     switch( iRoundMode & 0xF )
00352     {
00353         CASE_16( _mm_round_ss, dst, a );
00354     }
00355 }

SSP_FORCEINLINE __m128i ssp_stream_load_si128_SSE4_1 ( __m128i * a )

Native implementation of _mm_stream_load_si128 [SSE4.1]. (Searches MSDN)

Definition at line 298 of file SSEPlus_native_SSE4.1.h.

00299 {
00300     return _mm_stream_load_si128( a );
00301 }

SSP_FORCEINLINE int ssp_testc_si128_SSE4_1	(	__m128i	mask,
		__m128i	a
	)

Native implementation of _mm_testc_si128 [SSE4.1]. (Searches MSDN)

Definition at line 314 of file SSEPlus_native_SSE4.1.h.

00315 {
00316     return _mm_testc_si128( mask, a);
00317 }

SSP_FORCEINLINE int ssp_testnzc_si128_SSE4_1	(	__m128i	mask,
		__m128i	b
	)

Native implementation of _mm_testnzc_si128 [SSE4.1]. (Searches MSDN)

Definition at line 319 of file SSEPlus_native_SSE4.1.h.

00320 {
00321     return _mm_testnzc_si128( mask, b);
00322 }

SSP_FORCEINLINE int ssp_testz_si128_SSE4_1	(	__m128i	mask,
		__m128i	a
	)

Native implementation of _mm_testz_si128 [SSE4.1]. (Searches MSDN)

Definition at line 309 of file SSEPlus_native_SSE4.1.h.

00310 {
00311     return _mm_testz_si128( mask, a);
00312 }


Native SSE4.1 Operations
SSP_FORCEINLINE __m128d	ssp_ceil_pd_SSE4_1 (__m128d a)
SSP_FORCEINLINE __m128d	ssp_ceil_sd_SSE4_1 (__m128d dst, __m128d a)
SSP_FORCEINLINE __m128d	ssp_floor_pd_SSE4_1 (__m128d a)
SSP_FORCEINLINE __m128d	ssp_floor_sd_SSE4_1 (__m128d dst, __m128d a)
SSP_FORCEINLINE __m128	ssp_ceil_ps_SSE4_1 (__m128 a)
SSP_FORCEINLINE __m128	ssp_ceil_ss_SSE4_1 (__m128 dst, __m128 a)
SSP_FORCEINLINE __m128	ssp_floor_ps_SSE4_1 (__m128 a)
SSP_FORCEINLINE __m128	ssp_floor_ss_SSE4_1 (__m128 dst, __m128 a)
SSP_FORCEINLINE __m128i	ssp_blend_epi16_SSE4_1 (__m128i a, __m128i b, const int mask)
SSP_FORCEINLINE __m128i	ssp_blendv_epi8_SSE4_1 (__m128i a, __m128i b, __m128i mask)
SSP_FORCEINLINE __m128	ssp_blend_ps_SSE4_1 (__m128 a, __m128 b, const int mask)
SSP_FORCEINLINE __m128	ssp_blendv_ps_SSE4_1 (__m128 a, __m128 b, __m128 mask)
SSP_FORCEINLINE __m128d	ssp_blend_pd_SSE4_1 (__m128d a, __m128d b, const int mask)
SSP_FORCEINLINE __m128d	ssp_blendv_pd_SSE4_1 (__m128d a, __m128d b, __m128d mask)
SSP_FORCEINLINE __m128	ssp_dp_ps_SSE4_1 (__m128 a, __m128 b, const int mask)
SSP_FORCEINLINE __m128d	ssp_dp_pd_SSE4_1 (__m128d a, __m128d b, const int mask)
SSP_FORCEINLINE __m128i	ssp_cmpeq_epi64_SSE4_1 (__m128i a, __m128i b)
SSP_FORCEINLINE __m128i	ssp_min_epi8_SSE4_1 (__m128i a, __m128i b)
SSP_FORCEINLINE __m128i	ssp_max_epi8_SSE4_1 (__m128i a, __m128i b)
SSP_FORCEINLINE __m128i	ssp_min_epu16_SSE4_1 (__m128i a, __m128i b)
SSP_FORCEINLINE __m128i	ssp_max_epu16_SSE4_1 (__m128i a, __m128i b)
SSP_FORCEINLINE __m128i	ssp_min_epi32_SSE4_1 (__m128i a, __m128i b)
SSP_FORCEINLINE __m128i	ssp_max_epi32_SSE4_1 (__m128i a, __m128i b)
SSP_FORCEINLINE __m128i	ssp_min_epu32_SSE4_1 (__m128i a, __m128i b)
SSP_FORCEINLINE __m128i	ssp_max_epu32_SSE4_1 (__m128i a, __m128i b)
SSP_FORCEINLINE __m128i	ssp_mullo_epi32_SSE4_1 (__m128i a, __m128i b)
SSP_FORCEINLINE __m128i	ssp_mul_epi32_SSE4_1 (__m128i a, __m128i b)
SSP_FORCEINLINE __m128	ssp_insert_ps_SSE4_1 (__m128 dst, __m128 src, const int ndx)
SSP_FORCEINLINE int	ssp_extract_ps_SSE4_1 (__m128 src, const int ndx)
SSP_FORCEINLINE __m128i	ssp_insert_epi8_SSE4_1 (__m128i dst, int s, const int ndx)
SSP_FORCEINLINE __m128i	ssp_insert_epi32_SSE4_1 (__m128i dst, int s, const int ndx)
SSP_FORCEINLINE int	ssp_extract_epi8_SSE4_1 (__m128i src, const int ndx)
SSP_FORCEINLINE int	ssp_extract_epi32_SSE4_1 (__m128i src, const int ndx)
SSP_FORCEINLINE __m128i	ssp_minpos_epu16_SSE4_1 (__m128i shortValues)
SSP_FORCEINLINE __m128i	ssp_cvtepi8_epi32_SSE4_1 (__m128i byteValues)
SSP_FORCEINLINE __m128i	ssp_cvtepi16_epi32_SSE4_1 (__m128i shortValues)
SSP_FORCEINLINE __m128i	ssp_cvtepi8_epi64_SSE4_1 (__m128i byteValues)
SSP_FORCEINLINE __m128i	ssp_cvtepi32_epi64_SSE4_1 (__m128i intValues)
SSP_FORCEINLINE __m128i	ssp_cvtepi16_epi64_SSE4_1 (__m128i shortValues)
SSP_FORCEINLINE __m128i	ssp_cvtepi8_epi16_SSE4_1 (__m128i byteValues)
SSP_FORCEINLINE __m128i	ssp_cvtepu8_epi32_SSE4_1 (__m128i byteValues)
SSP_FORCEINLINE __m128i	ssp_cvtepu16_epi32_SSE4_1 (__m128i shortValues)
SSP_FORCEINLINE __m128i	ssp_cvtepu8_epi64_SSE4_1 (__m128i shortValues)
SSP_FORCEINLINE __m128i	ssp_cvtepu32_epi64_SSE4_1 (__m128i intValues)
SSP_FORCEINLINE __m128i	ssp_cvtepu16_epi64_SSE4_1 (__m128i shortValues)
SSP_FORCEINLINE __m128i	ssp_cvtepu8_epi16_SSE4_1 (__m128i byteValues)
SSP_FORCEINLINE __m128i	ssp_packus_epi32_SSE4_1 (__m128i a, __m128i b)
SSP_FORCEINLINE __m128i	ssp_mpsadbw_epu8_SSE4_1 (__m128i a, __m128i b, const int msk)
SSP_FORCEINLINE __m128i	ssp_stream_load_si128_SSE4_1 (__m128i *a)
SSP_FORCEINLINE int	ssp_testz_si128_SSE4_1 (__m128i mask, __m128i a)
SSP_FORCEINLINE int	ssp_testc_si128_SSE4_1 (__m128i mask, __m128i a)
SSP_FORCEINLINE int	ssp_testnzc_si128_SSE4_1 (__m128i mask, __m128i b)
SSP_FORCEINLINE __m128d	ssp_round_pd_SSE4_1 (__m128d a, int iRoundMode)
SSP_FORCEINLINE __m128d	ssp_round_sd_SSE4_1 (__m128d dst, __m128d a, int iRoundMode)
SSP_FORCEINLINE __m128	ssp_round_ps_SSE4_1 (__m128 a, int iRoundMode)
SSP_FORCEINLINE __m128	ssp_round_ss_SSE4_1 (__m128 dst, __m128 a, int iRoundMode)

SSE4.1
[Native]

Native SSE4.1 Operations

Function Documentation

SSE4.1 [Native]

Native SSE4.1 Operations

Function Documentation

SSE4.1
[Native]