38 	int16* vec2L = ( int16* )vec2A;  in bbs_dotProduct_64SSE2()  local
58 			m_XMM1 = _mm_loadl_epi64( (__m128i *)&0[vec2L] );  in bbs_dotProduct_64SSE2()
65 			m_XMM3 = _mm_loadl_epi64( (__m128i *)&4[vec2L] );  in bbs_dotProduct_64SSE2()
70 			m_XMM5 = _mm_loadl_epi64( (__m128i *)&8[vec2L] );  in bbs_dotProduct_64SSE2()
78 			m_XMM8 = _mm_loadl_epi64( (__m128i *)&12[vec2L] );  in bbs_dotProduct_64SSE2()
84 			vec2L += 16;  in bbs_dotProduct_64SSE2()
106 			resultL += ( int32 )*vec1L++ * *vec2L++;  in bbs_dotProduct_64SSE2()
108 			resultL += ( int32 )*vec1L++ * *vec2L++;  in bbs_dotProduct_64SSE2()
110 			resultL += ( int32 )*vec1L++ * *vec2L++;  in bbs_dotProduct_64SSE2()
112 			resultL += ( int32 )*vec1L++ * *vec2L++;  in bbs_dotProduct_64SSE2()
114 			resultL += ( int32 )*vec1L++ * *vec2L++;  in bbs_dotProduct_64SSE2()
116 			resultL += ( int32 )*vec1L++ * *vec2L++;  in bbs_dotProduct_64SSE2()
118 			resultL += ( int32 )*vec1L++ * *vec2L++;  in bbs_dotProduct_64SSE2()
120 			resultL += ( int32 )*vec1L++ * *vec2L++;  in bbs_dotProduct_64SSE2()
122 			resultL += ( int32 )*vec1L++ * *vec2L++;  in bbs_dotProduct_64SSE2()
124 			resultL += ( int32 )*vec1L++ * *vec2L++;  in bbs_dotProduct_64SSE2()
126 			resultL += ( int32 )*vec1L++ * *vec2L++;  in bbs_dotProduct_64SSE2()
128 			resultL += ( int32 )*vec1L++ * *vec2L++;  in bbs_dotProduct_64SSE2()
130 			resultL += ( int32 )*vec1L++ * *vec2L++;  in bbs_dotProduct_64SSE2()
132 			resultL += ( int32 )*vec1L++ * *vec2L++;  in bbs_dotProduct_64SSE2()
134 			resultL += ( int32 )*vec1L++ * *vec2L++;  in bbs_dotProduct_64SSE2()
150 	int16* vec2L = ( int16* )vec2A;  in bbs_dotProduct_128SSE2()  local
168 			m_XMM2 = _mm_load_si128( (__m128i *)&0[vec2L] );  in bbs_dotProduct_128SSE2()
176 			m_XMM3 = _mm_load_si128( (__m128i *)&8[vec2L] );  in bbs_dotProduct_128SSE2()
181 			vec2L += 16;  in bbs_dotProduct_128SSE2()
208 			resultL += ( int32 )*vec1L++ * *vec2L++;  in bbs_dotProduct_128SSE2()
210 			resultL += ( int32 )*vec1L++ * *vec2L++;  in bbs_dotProduct_128SSE2()
212 			resultL += ( int32 )*vec1L++ * *vec2L++;  in bbs_dotProduct_128SSE2()
214 			resultL += ( int32 )*vec1L++ * *vec2L++;  in bbs_dotProduct_128SSE2()
216 			resultL += ( int32 )*vec1L++ * *vec2L++;  in bbs_dotProduct_128SSE2()
218 			resultL += ( int32 )*vec1L++ * *vec2L++;  in bbs_dotProduct_128SSE2()
220 			resultL += ( int32 )*vec1L++ * *vec2L++;  in bbs_dotProduct_128SSE2()
222 			resultL += ( int32 )*vec1L++ * *vec2L++;  in bbs_dotProduct_128SSE2()
224 			resultL += ( int32 )*vec1L++ * *vec2L++;  in bbs_dotProduct_128SSE2()
226 			resultL += ( int32 )*vec1L++ * *vec2L++;  in bbs_dotProduct_128SSE2()
228 			resultL += ( int32 )*vec1L++ * *vec2L++;  in bbs_dotProduct_128SSE2()
230 			resultL += ( int32 )*vec1L++ * *vec2L++;  in bbs_dotProduct_128SSE2()
232 			resultL += ( int32 )*vec1L++ * *vec2L++;  in bbs_dotProduct_128SSE2()
234 			resultL += ( int32 )*vec1L++ * *vec2L++;  in bbs_dotProduct_128SSE2()
236 			resultL += ( int32 )*vec1L++ * *vec2L++;  in bbs_dotProduct_128SSE2()
253 	int16* vec2L = ( int16* )vec2A;  in bbs_dotProduct_u128SSE2()  local
272 			m_XMM2 = _mm_loadu_si128( (__m128i *)&0[vec2L] );  in bbs_dotProduct_u128SSE2()
280 			m_XMM3 = _mm_loadu_si128( (__m128i *)&8[vec2L] );  in bbs_dotProduct_u128SSE2()
285 			vec2L += 16;  in bbs_dotProduct_u128SSE2()
313 			resultL += ( int32 )*vec1L++ * *vec2L++;  in bbs_dotProduct_u128SSE2()
315 			resultL += ( int32 )*vec1L++ * *vec2L++;  in bbs_dotProduct_u128SSE2()
317 			resultL += ( int32 )*vec1L++ * *vec2L++;  in bbs_dotProduct_u128SSE2()
319 			resultL += ( int32 )*vec1L++ * *vec2L++;  in bbs_dotProduct_u128SSE2()
321 			resultL += ( int32 )*vec1L++ * *vec2L++;  in bbs_dotProduct_u128SSE2()
323 			resultL += ( int32 )*vec1L++ * *vec2L++;  in bbs_dotProduct_u128SSE2()
325 			resultL += ( int32 )*vec1L++ * *vec2L++;  in bbs_dotProduct_u128SSE2()
327 			resultL += ( int32 )*vec1L++ * *vec2L++;  in bbs_dotProduct_u128SSE2()
329 			resultL += ( int32 )*vec1L++ * *vec2L++;  in bbs_dotProduct_u128SSE2()
331 			resultL += ( int32 )*vec1L++ * *vec2L++;  in bbs_dotProduct_u128SSE2()
333 			resultL += ( int32 )*vec1L++ * *vec2L++;  in bbs_dotProduct_u128SSE2()
335 			resultL += ( int32 )*vec1L++ * *vec2L++;  in bbs_dotProduct_u128SSE2()
337 			resultL += ( int32 )*vec1L++ * *vec2L++;  in bbs_dotProduct_u128SSE2()
339 			resultL += ( int32 )*vec1L++ * *vec2L++;  in bbs_dotProduct_u128SSE2()
341 			resultL += ( int32 )*vec1L++ * *vec2L++;  in bbs_dotProduct_u128SSE2()