Home | History | Annotate | Download | only in x86

Lines Matching refs:_mm_add_epi32

75     const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING);
76 const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING);
77 const __m128i v4 = _mm_add_epi32(u4, k__DCT_CONST_ROUNDING);
78 const __m128i v6 = _mm_add_epi32(u6, k__DCT_CONST_ROUNDING);
185 v[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING);
186 v[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING);
187 v[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING);
188 v[3] = _mm_add_epi32(u[3], k__DCT_CONST_ROUNDING);
224 u[0] = _mm_add_epi32(v[0], v[1]);
226 u[2] = _mm_add_epi32(v[3], v[4]);
230 u[6] = _mm_add_epi32(u[3], u[5]);
232 v[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING);
233 v[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING);
234 v[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING);
235 v[3] = _mm_add_epi32(u[6], k__DCT_CONST_ROUNDING);
350 const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING);
351 const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING);
352 const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING);
353 const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING);
354 const __m128i v4 = _mm_add_epi32(u4, k__DCT_CONST_ROUNDING);
355 const __m128i v5 = _mm_add_epi32(u5, k__DCT_CONST_ROUNDING);
356 const __m128i v6 = _mm_add_epi32(u6, k__DCT_CONST_ROUNDING);
357 const __m128i v7 = _mm_add_epi32(u7, k__DCT_CONST_ROUNDING);
382 const __m128i f0 = _mm_add_epi32(e0, k__DCT_CONST_ROUNDING);
383 const __m128i f1 = _mm_add_epi32(e1, k__DCT_CONST_ROUNDING);
384 const __m128i f2 = _mm_add_epi32(e2, k__DCT_CONST_ROUNDING);
385 const __m128i f3 = _mm_add_epi32(e3, k__DCT_CONST_ROUNDING);
412 const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING);
413 const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING);
414 const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING);
415 const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING);
416 const __m128i v4 = _mm_add_epi32(u4, k__DCT_CONST_ROUNDING);
417 const __m128i v5 = _mm_add_epi32(u5, k__DCT_CONST_ROUNDING);
418 const __m128i v6 = _mm_add_epi32(u6, k__DCT_CONST_ROUNDING);
419 const __m128i v7 = _mm_add_epi32(u7, k__DCT_CONST_ROUNDING);
710 v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING);
711 v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING);
712 v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING);
713 v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING);
714 v4 = _mm_add_epi32(u4, k__DCT_CONST_ROUNDING);
715 v5 = _mm_add_epi32(u5, k__DCT_CONST_ROUNDING);
716 v6 = _mm_add_epi32(u6, k__DCT_CONST_ROUNDING);
717 v7 = _mm_add_epi32(u7, k__DCT_CONST_ROUNDING);
743 u0 = _mm_add_epi32(v0, k__DCT_CONST_ROUNDING);
744 u1 = _mm_add_epi32(v1, k__DCT_CONST_ROUNDING);
745 u2 = _mm_add_epi32(v2, k__DCT_CONST_ROUNDING);
746 u3 = _mm_add_epi32(v3, k__DCT_CONST_ROUNDING);
778 u0 = _mm_add_epi32(v0, k__DCT_CONST_ROUNDING);
779 u1 = _mm_add_epi32(v1, k__DCT_CONST_ROUNDING);
780 u2 = _mm_add_epi32(v2, k__DCT_CONST_ROUNDING);
781 u3 = _mm_add_epi32(v3, k__DCT_CONST_ROUNDING);
782 u4 = _mm_add_epi32(v4, k__DCT_CONST_ROUNDING);
783 u5 = _mm_add_epi32(v5, k__DCT_CONST_ROUNDING);
784 u6 = _mm_add_epi32(v6, k__DCT_CONST_ROUNDING);
785 u7 = _mm_add_epi32(v7, k__DCT_CONST_ROUNDING);
869 w0 = _mm_add_epi32(u0, u8);
870 w1 = _mm_add_epi32(u1, u9);
871 w2 = _mm_add_epi32(u2, u10);
872 w3 = _mm_add_epi32(u3, u11);
873 w4 = _mm_add_epi32(u4, u12);
874 w5 = _mm_add_epi32(u5, u13);
875 w6 = _mm_add_epi32(u6, u14);
876 w7 = _mm_add_epi32(u7, u15);
887 v0 = _mm_add_epi32(w0, k__DCT_CONST_ROUNDING);
888 v1 = _mm_add_epi32(w1, k__DCT_CONST_ROUNDING);
889 v2 = _mm_add_epi32(w2, k__DCT_CONST_ROUNDING);
890 v3 = _mm_add_epi32(w3, k__DCT_CONST_ROUNDING);
891 v4 = _mm_add_epi32(w4, k__DCT_CONST_ROUNDING);
892 v5 = _mm_add_epi32(w5, k__DCT_CONST_ROUNDING);
893 v6 = _mm_add_epi32(w6, k__DCT_CONST_ROUNDING);
894 v7 = _mm_add_epi32(w7, k__DCT_CONST_ROUNDING);
895 v8 = _mm_add_epi32(w8, k__DCT_CONST_ROUNDING);
896 v9 = _mm_add_epi32(w9, k__DCT_CONST_ROUNDING);
897 v10 = _mm_add_epi32(w10, k__DCT_CONST_ROUNDING);
898 v11 = _mm_add_epi32(w11, k__DCT_CONST_ROUNDING);
899 v12 = _mm_add_epi32(w12, k__DCT_CONST_ROUNDING);
900 v13 = _mm_add_epi32(w13, k__DCT_CONST_ROUNDING);
901 v14 = _mm_add_epi32(w14, k__DCT_CONST_ROUNDING);
902 v15 = _mm_add_epi32(w15, k__DCT_CONST_ROUNDING);
950 w0 = _mm_add_epi32(v0, v4);
951 w1 = _mm_add_epi32(v1, v5);
952 w2 = _mm_add_epi32(v2, v6);
953 w3 = _mm_add_epi32(v3, v7);
959 v0 = _mm_add_epi32(w0, k__DCT_CONST_ROUNDING);
960 v1 = _mm_add_epi32(w1, k__DCT_CONST_ROUNDING);
961 v2 = _mm_add_epi32(w2, k__DCT_CONST_ROUNDING);
962 v3 = _mm_add_epi32(w3, k__DCT_CONST_ROUNDING);
963 v4 = _mm_add_epi32(w4, k__DCT_CONST_ROUNDING);
964 v5 = _mm_add_epi32(w5, k__DCT_CONST_ROUNDING);
965 v6 = _mm_add_epi32(w6, k__DCT_CONST_ROUNDING);
966 v7 = _mm_add_epi32(w7, k__DCT_CONST_ROUNDING);
998 u0 = _mm_add_epi32(v0, k__DCT_CONST_ROUNDING);
999 u1 = _mm_add_epi32(v1, k__DCT_CONST_ROUNDING);
1000 u2 = _mm_add_epi32(v2, k__DCT_CONST_ROUNDING);
1001 u3 = _mm_add_epi32(v3, k__DCT_CONST_ROUNDING);
1002 u4 = _mm_add_epi32(v4, k__DCT_CONST_ROUNDING);
1003 u5 = _mm_add_epi32(v5, k__DCT_CONST_ROUNDING);
1004 u6 = _mm_add_epi32(v6, k__DCT_CONST_ROUNDING);
1005 u7 = _mm_add_epi32(v7, k__DCT_CONST_ROUNDING);
1262 const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING);
1263 const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING);
1264 const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING);
1265 const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING);
1266 const __m128i v4 = _mm_add_epi32(u4, k__DCT_CONST_ROUNDING);
1267 const __m128i v5 = _mm_add_epi32(u5, k__DCT_CONST_ROUNDING);
1268 const __m128i v6 = _mm_add_epi32(u6, k__DCT_CONST_ROUNDING);
1269 const __m128i v7 = _mm_add_epi32(u7, k__DCT_CONST_ROUNDING);
1295 const __m128i f0 = _mm_add_epi32(e0, k__DCT_CONST_ROUNDING);
1296 const __m128i f1 = _mm_add_epi32(e1, k__DCT_CONST_ROUNDING);
1297 const __m128i f2 = _mm_add_epi32(e2, k__DCT_CONST_ROUNDING);
1298 const __m128i f3 = _mm_add_epi32(e3, k__DCT_CONST_ROUNDING);
1326 const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING);
1327 const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING);
1328 const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING);
1329 const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING);
1330 const __m128i v4 = _mm_add_epi32(u4, k__DCT_CONST_ROUNDING);
1331 const __m128i v5 = _mm_add_epi32(u5, k__DCT_CONST_ROUNDING);
1332 const __m128i v6 = _mm_add_epi32(u6, k__DCT_CONST_ROUNDING);
1333 const __m128i v7 = _mm_add_epi32(u7, k__DCT_CONST_ROUNDING);
1362 const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING);
1363 const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING);
1364 const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING);
1365 const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING);
1384 const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING);
1385 const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING);
1386 const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING);
1387 const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING);
1418 const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING);
1419 const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING);
1420 const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING);
1421 const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING);
1440 const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING);
1441 const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING);
1442 const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING);
1443 const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING);
1474 const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING);
1475 const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING);
1476 const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING);
1477 const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING);
1496 const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING);
1497 const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING);
1498 const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING);
1499 const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING);
1518 const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING);
1519 const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING);
1520 const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING);
1521 const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING);
1540 const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING);
1541 const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING);
1542 const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING);
1543 const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING);
1813 v[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING);
1814 v[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING);
1815 v[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING);
1816 v[3] = _mm_add_epi32(u[3], k__DCT_CONST_ROUNDING);
1817 v[4] = _mm_add_epi32(u[4], k__DCT_CONST_ROUNDING);
1818 v[5] = _mm_add_epi32(u[5], k__DCT_CONST_ROUNDING);
1819 v[6] = _mm_add_epi32(u[6], k__DCT_CONST_ROUNDING);
1820 v[7] = _mm_add_epi32(u[7], k__DCT_CONST_ROUNDING);
1843 u[0] = _mm_add_epi32(v[0], k__DCT_CONST_ROUNDING);
1844 u[1] = _mm_add_epi32(v[1], k__DCT_CONST_ROUNDING);
1845 u[2] = _mm_add_epi32(v[2], k__DCT_CONST_ROUNDING);
1846 u[3] = _mm_add_epi32(v[3], k__DCT_CONST_ROUNDING);
1875 u[0] = _mm_add_epi32(v[0], k__DCT_CONST_ROUNDING);
1876 u[1] = _mm_add_epi32(v[1], k__DCT_CONST_ROUNDING);
1877 u[2] = _mm_add_epi32(v[2], k__DCT_CONST_ROUNDING);
1878 u[3] = _mm_add_epi32(v[3], k__DCT_CONST_ROUNDING);
1879 u[4] = _mm_add_epi32(v[4], k__DCT_CONST_ROUNDING);
1880 u[5] = _mm_add_epi32(v[5], k__DCT_CONST_ROUNDING);
1881 u[6] = _mm_add_epi32(v[6], k__DCT_CONST_ROUNDING);
1882 u[7] = _mm_add_epi32(v[7], k__DCT_CONST_ROUNDING);
1913 u[0] = _mm_add_epi32(v[0], k__DCT_CONST_ROUNDING);
1914 u[1] = _mm_add_epi32(v[1], k__DCT_CONST_ROUNDING);
1915 u[2] = _mm_add_epi32(v[2], k__DCT_CONST_ROUNDING);
1916 u[3] = _mm_add_epi32(v[3], k__DCT_CONST_ROUNDING);
1917 u[4] = _mm_add_epi32(v[4], k__DCT_CONST_ROUNDING);
1918 u[5] = _mm_add_epi32(v[5], k__DCT_CONST_ROUNDING);
1919 u[6] = _mm_add_epi32(v[6], k__DCT_CONST_ROUNDING);
1920 u[7] = _mm_add_epi32(v[7], k__DCT_CONST_ROUNDING);
1961 u[0] = _mm_add_epi32(v[0], k__DCT_CONST_ROUNDING);
1962 u[1] = _mm_add_epi32(v[1], k__DCT_CONST_ROUNDING);
1963 u[2] = _mm_add_epi32(v[2], k__DCT_CONST_ROUNDING);
1964 u[3] = _mm_add_epi32(v[3], k__DCT_CONST_ROUNDING);
1965 u[4] = _mm_add_epi32(v[4], k__DCT_CONST_ROUNDING);
1966 u[5] = _mm_add_epi32(v[5], k__DCT_CONST_ROUNDING);
1967 u[6] = _mm_add_epi32(v[6], k__DCT_CONST_ROUNDING);
1968 u[7] = _mm_add_epi32(v[7], k__DCT_CONST_ROUNDING);
2021 u[0] = _mm_add_epi32(v[0], k__DCT_CONST_ROUNDING);
2022 u[1] = _mm_add_epi32(v[1], k__DCT_CONST_ROUNDING);
2023 u[2] = _mm_add_epi32(v[2], k__DCT_CONST_ROUNDING);
2024 u[3] = _mm_add_epi32(v[3], k__DCT_CONST_ROUNDING);
2025 u[4] = _mm_add_epi32(v[4], k__DCT_CONST_ROUNDING);
2026 u[5] = _mm_add_epi32(v[5], k__DCT_CONST_ROUNDING);
2027 u[6] = _mm_add_epi32(v[6], k__DCT_CONST_ROUNDING);
2028 u[7] = _mm_add_epi32(v[7], k__DCT_CONST_ROUNDING);
2029 u[8] = _mm_add_epi32(v[8], k__DCT_CONST_ROUNDING);
2030 u[9] = _mm_add_epi32(v[9], k__DCT_CONST_ROUNDING);
2031 u[10] = _mm_add_epi32(v[10], k__DCT_CONST_ROUNDING);
2032 u[11] = _mm_add_epi32(v[11], k__DCT_CONST_ROUNDING);
2033 u[12] = _mm_add_epi32(v[12], k__DCT_CONST_ROUNDING);
2034 u[13] = _mm_add_epi32(v[13], k__DCT_CONST_ROUNDING);
2035 u[14] = _mm_add_epi32(v[14], k__DCT_CONST_ROUNDING);
2036 u[15] = _mm_add_epi32(v[15], k__DCT_CONST_ROUNDING);
2150 u[0] = _mm_add_epi32(v[0], v[16]);
2151 u[1] = _mm_add_epi32(v[1], v[17]);
2152 u[2] = _mm_add_epi32(v[2], v[18]);
2153 u[3] = _mm_add_epi32(v[3], v[19]);
2154 u[4] = _mm_add_epi32(v[4], v[20]);
2155 u[5] = _mm_add_epi32(v[5], v[21]);
2156 u[6] = _mm_add_epi32(v[6], v[22]);
2157 u[7] = _mm_add_epi32(v[7], v[23]);
2158 u[8] = _mm_add_epi32(v[8], v[24]);
2159 u[9] = _mm_add_epi32(v[9], v[25]);
2160 u[10] = _mm_add_epi32(v[10], v[26]);
2161 u[11] = _mm_add_epi32(v[11], v[27]);
2162 u[12] = _mm_add_epi32(v[12], v[28]);
2163 u[13] = _mm_add_epi32(v[13], v[29]);
2164 u[14] = _mm_add_epi32(v[14], v[30]);
2165 u[15] = _mm_add_epi32(v[15], v[31]);
2183 v[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING);
2184 v[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING);
2185 v[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING);
2186 v[3] = _mm_add_epi32(u[3], k__DCT_CONST_ROUNDING);
2187 v[4] = _mm_add_epi32(u[4], k__DCT_CONST_ROUNDING);
2188 v[5] = _mm_add_epi32(u[5], k__DCT_CONST_ROUNDING);
2189 v[6] = _mm_add_epi32(u[6], k__DCT_CONST_ROUNDING);
2190 v[7] = _mm_add_epi32(u[7], k__DCT_CONST_ROUNDING);
2191 v[8] = _mm_add_epi32(u[8], k__DCT_CONST_ROUNDING);
2192 v[9] = _mm_add_epi32(u[9], k__DCT_CONST_ROUNDING);
2193 v[10] = _mm_add_epi32(u[10], k__DCT_CONST_ROUNDING);
2194 v[11] = _mm_add_epi32(u[11], k__DCT_CONST_ROUNDING);
2195 v[12] = _mm_add_epi32(u[12], k__DCT_CONST_ROUNDING);
2196 v[13] = _mm_add_epi32(u[13], k__DCT_CONST_ROUNDING);
2197 v[14] = _mm_add_epi32(u[14], k__DCT_CONST_ROUNDING);
2198 v[15] = _mm_add_epi32(u[15], k__DCT_CONST_ROUNDING);
2199 v[16] = _mm_add_epi32(u[16], k__DCT_CONST_ROUNDING);
2200 v[17] = _mm_add_epi32
2201 v[18] = _mm_add_epi32(u[18], k__DCT_CONST_ROUNDING);
2202 v[19] = _mm_add_epi32(u[19], k__DCT_CONST_ROUNDING);
2203 v[20] = _mm_add_epi32(u[20], k__DCT_CONST_ROUNDING);
2204 v[21] = _mm_add_epi32(u[21], k__DCT_CONST_ROUNDING);
2205 v[22] = _mm_add_epi32(u[22], k__DCT_CONST_ROUNDING);
2206 v[23] = _mm_add_epi32(u[23], k__DCT_CONST_ROUNDING);
2207 v[24] = _mm_add_epi32(u[24], k__DCT_CONST_ROUNDING);
2208 v[25] = _mm_add_epi32(u[25], k__DCT_CONST_ROUNDING);
2209 v[26] = _mm_add_epi32(u[26], k__DCT_CONST_ROUNDING);
2210 v[27] = _mm_add_epi32(u[27], k__DCT_CONST_ROUNDING);
2211 v[28] = _mm_add_epi32(u[28], k__DCT_CONST_ROUNDING);
2212 v[29] = _mm_add_epi32(u[29], k__DCT_CONST_ROUNDING);
2213 v[30] = _mm_add_epi32(u[30], k__DCT_CONST_ROUNDING);
2214 v[31] = _mm_add_epi32(u[31], k__DCT_CONST_ROUNDING);
2293 u[0] = _mm_add_epi32(v[0], v[8]);
2294 u[1] = _mm_add_epi32(v[1], v[9]);
2295 u[2] = _mm_add_epi32(v[2], v[10]);
2296 u[3] = _mm_add_epi32(v[3], v[11]);
2297 u[4] = _mm_add_epi32(v[4], v[12]);
2298 u[5] = _mm_add_epi32(v[5], v[13]);
2299 u[6] = _mm_add_epi32(v[6], v[14]);
2300 u[7] = _mm_add_epi32(v[7], v[15]);
2310 v[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING);
2311 v[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING);
2312 v[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING);
2313 v[3] = _mm_add_epi32(u[3], k__DCT_CONST_ROUNDING);
2314 v[4] = _mm_add_epi32(u[4], k__DCT_CONST_ROUNDING);
2315 v[5] = _mm_add_epi32(u[5], k__DCT_CONST_ROUNDING);
2316 v[6] = _mm_add_epi32(u[6], k__DCT_CONST_ROUNDING);
2317 v[7] = _mm_add_epi32(u[7], k__DCT_CONST_ROUNDING);
2318 v[8] = _mm_add_epi32(u[8], k__DCT_CONST_ROUNDING);
2319 v[9] = _mm_add_epi32(u[9], k__DCT_CONST_ROUNDING);
2320 v[10] = _mm_add_epi32(u[10], k__DCT_CONST_ROUNDING);
2321 v[11] = _mm_add_epi32(u[11], k__DCT_CONST_ROUNDING);
2322 v[12] = _mm_add_epi32(u[12], k__DCT_CONST_ROUNDING);
2323 v[13] = _mm_add_epi32(u[13], k__DCT_CONST_ROUNDING);
2324 v[14] = _mm_add_epi32(u[14], k__DCT_CONST_ROUNDING);
2325 v[15] = _mm_add_epi32(u[15], k__DCT_CONST_ROUNDING);
2388 u[0] = _mm_add_epi32(v[0], v[4]);
2389 u[1] = _mm_add_epi32(v[1], v[5]);
2390 u[2] = _mm_add_epi32(v[2], v[6]);
2391 u[3] = _mm_add_epi32(v[3], v[7]);
2396 u[8] = _mm_add_epi32(v[8], v[12]);
2397 u[9] = _mm_add_epi32(v[9], v[13]);
2398 u[10] = _mm_add_epi32(v[10], v[14]);
2399 u[11] = _mm_add_epi32(v[11], v[15]);
2405 u[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING);
2406 u[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING);
2407 u[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING);
2408 u[3] = _mm_add_epi32(u[3], k__DCT_CONST_ROUNDING);
2409 u[4] = _mm_add_epi32(u[4], k__DCT_CONST_ROUNDING);
2410 u[5] = _mm_add_epi32(u[5], k__DCT_CONST_ROUNDING);
2411 u[6] = _mm_add_epi32(u[6], k__DCT_CONST_ROUNDING);
2412 u[7] = _mm_add_epi32(u[7], k__DCT_CONST_ROUNDING);
2413 u[8] = _mm_add_epi32(u[8], k__DCT_CONST_ROUNDING);
2414 u[9] = _mm_add_epi32(u[9], k__DCT_CONST_ROUNDING);
2415 u[10] = _mm_add_epi32(u[10], k__DCT_CONST_ROUNDING);
2416 u[11] = _mm_add_epi32(u[11], k__DCT_CONST_ROUNDING);
2417 u[12] = _mm_add_epi32(u[12], k__DCT_CONST_ROUNDING);
2418 u[13] = _mm_add_epi32(u[13], k__DCT_CONST_ROUNDING);
2419 u[14] = _mm_add_epi32(u[14], k__DCT_CONST_ROUNDING);
2420 u[15] = _mm_add_epi32(u[15], k__DCT_CONST_ROUNDING);
2483 u[0] = _mm_add_epi32(v[0], k__DCT_CONST_ROUNDING);
2484 u[1] = _mm_add_epi32(v[1], k__DCT_CONST_ROUNDING);
2485 u[2] = _mm_add_epi32(v[2], k__DCT_CONST_ROUNDING);
2486 u[3] = _mm_add_epi32(v[3], k__DCT_CONST_ROUNDING);
2487 u[4] = _mm_add_epi32(v[4], k__DCT_CONST_ROUNDING);
2488 u[5] = _mm_add_epi32(v[5], k__DCT_CONST_ROUNDING);
2489 u[6] = _mm_add_epi32(v[6], k__DCT_CONST_ROUNDING);
2490 u[7] = _mm_add_epi32(v[7], k__DCT_CONST_ROUNDING);
2491 u[8] = _mm_add_epi32(v[8], k__DCT_CONST_ROUNDING);
2492 u[9] = _mm_add_epi32(v[9], k__DCT_CONST_ROUNDING);
2493 u[10] = _mm_add_epi32(v[10], k__DCT_CONST_ROUNDING);
2494 u[11] = _mm_add_epi32(v[11], k__DCT_CONST_ROUNDING);
2495 u[12] = _mm_add_epi32(v[12], k__DCT_CONST_ROUNDING);
2496 u[13] = _mm_add_epi32(v[13], k__DCT_CONST_ROUNDING);
2497 u[14] = _mm_add_epi32(v[14], k__DCT_CONST_ROUNDING);
2498 u[15] = _mm_add_epi32(v[15], k__DCT_CONST_ROUNDING);