ih264_inter_pred_filters_ssse3.c | 307 __m128i src_r0_t1_16x8b, src_r1_t1_16x8b; local 323 src_r0_t1_16x8b = _mm_unpacklo_epi8(src_r0_16x8b, src_r0_sht_16x8b); //a0 a1 a1 a2 a2 a3 a3 a4 a4 a5 a5 a6 a6 a7 a7 a8 326 res_r0_t1_8x16b = _mm_maddubs_epi16(src_r0_t1_16x8b, coeff0_1_16x8b); //a0*c0+a1*c1 a1*c0+a2*c1 a2*c0+a3*c1 a3*c0+a4*c1 337 src_r0_t1_16x8b = _mm_unpacklo_epi8(src_r0_16x8b, src_r0_sht_16x8b); //a2 a3 a3 a4 a4 a5 a5 a6 a6 a7 a7 a8 a8 a9 a9 a10 340 res_r0_t2_8x16b = _mm_maddubs_epi16(src_r0_t1_16x8b, coeff2_3_16x8b); //a2*c2+a3*c3 a3*c2+a4*c3 a4*c2+a5*c3 a5*c2+a6*c3 351 src_r0_t1_16x8b = _mm_unpacklo_epi8(src_r0_16x8b, src_r0_sht_16x8b); //a4 a5 a5 a6 a6 a7 a7 a8 a8 a9 a9 a10 a10 a11 a11 a12 354 res_r0_t3_8x16b = _mm_maddubs_epi16(src_r0_t1_16x8b, coeff4_5_16x8b); //a4*c4+a5*c5 a5*c4+a6*c5 a6*c4+a7*c5 a7*c4+a8*c5 383 __m128i src_r0_t1_16x8b, src_r1_t1_16x8b; local 400 src_r0_t1_16x8b = _mm_unpacklo_epi8(src_r0_16x8b, src_r0_sht_16x8b); //a0 a1 a1 a2 a2 a3 a3 a4 a4 a5 a5 a6 a6 a7 a7 a8 403 res_r0_t1_8x16b = _mm_maddubs_epi16(src_r0_t1_16x8b, coeff0_1_16x8b); //a0*c0+a1*c1 a1*c0+a2*c1 a2*c0+a3*c1 a3*c0+a4*c 957 __m128i src_r0_t1_16x8b, src_r1_t1_16x8b; local 1175 __m128i src_r0_t1_16x8b, src_r1_t1_16x8b; local 1593 __m128i src_r0_t1_16x8b, src_r1_t1_16x8b; local 1676 __m128i src_r0_t1_16x8b, src_r1_t1_16x8b; local 2377 __m128i src_r0_t1_16x8b, src_r1_t1_16x8b; local 2568 __m128i src_r0_t1_16x8b, src_r1_t1_16x8b; local 3460 __m128i src_r0_t1_16x8b, src_r1_t1_16x8b; local 3666 __m128i src_r0_t1_16x8b, src_r1_t1_16x8b; local [all...] |