ih264_inter_pred_filters_ssse3.c | 241 __m128i src_r0_16x8b, src_r1_16x8b, src_r0r1_16x8b; local 261 src_r0r1_16x8b = _mm_unpacklo_epi64(src_r0_16x8b, src_r1_16x8b); //a0 a1 a1 a2 a2 a3 a3 a4 b0 b1 b1 b2 b2 b3 b3 b4 262 res_r0r1_t1_8x16b = _mm_maddubs_epi16(src_r0r1_16x8b, coeff0_1_16x8b); //a0*c0+a1*c1 a1*c0+a2*c1 a2*c0+a3*c1 a3*c0+a4*c1 268 src_r0r1_16x8b = _mm_unpacklo_epi64(src_r0_16x8b, src_r1_16x8b); //a2 a3 a3 a4 a4 a5 a5 a6 b2 b3 b3 b4 b4 b5 b5 b6 269 res_r0r1_t2_8x16b = _mm_maddubs_epi16(src_r0r1_16x8b, coeff2_3_16x8b); //a2*c2+a3*c3 a3*c2+a4*c3 a4*c2+a5*c3 a5*c2+a6*c3 275 src_r0r1_16x8b = _mm_unpacklo_epi64(src_r0_16x8b, src_r1_16x8b); //a4 a5 a5 a6 a6 a7 a7 a8 b4 b5 b5 b6 b6 b7 b7 b8 276 res_r0r1_t3_8x16b = _mm_maddubs_epi16(src_r0r1_16x8b, coeff4_5_16x8b); //a4*c4+a5*c5 a5*c4+a6*c5 a6*c4+a7*c5 a7*c4+a8*c5 492 __m128i src_r0r1_16x8b, src_r2r3_16x8b, src_r4r5_16x8b; local 538 src_r0r1_16x8b = _mm_unpacklo_epi8(src_r0_16x8b, src_r1_16x8b) 1524 __m128i src_r0r1_16x8b; local 1800 __m128i src_r0r1_16x8b, src_r2r3_16x8b, src_r4r5_16x8b; local 2151 __m128i src_r0r1_16x8b, src_r2r3_16x8b, src_r4r5_16x8b; local 2293 __m128i src_r0r1_16x8b, src_r2r3_16x8b, src_r4r5_16x8b; local 2469 __m128i src_r0r1_16x8b, src_r2r3_16x8b, src_r4r5_16x8b; local 2709 __m128i src_r0r1_16x8b, src_r2r3_16x8b, src_r4r5_16x8b; local 2870 __m128i src_r0r1_16x8b, src_r2r3_16x8b, src_r4r5_16x8b; local 3053 __m128i src_r0r1_16x8b, src_r2r3_16x8b, src_r4r5_16x8b; local [all...] |