/external/libavc/common/x86/ |
ih264_ihadamard_scaling_ssse3.c | 97 __m128i src_r0, src_r1, src_r2, src_r3; local 113 src_r2 = _mm_unpacklo_epi16(src_r2_r3, sign_reg); 128 temp2 = _mm_unpacklo_epi32(src_r2, src_r3); //c0 d0 c1 d1 130 temp3 = _mm_unpackhi_epi32(src_r2, src_r3); //c2 d2 c3 d3 133 src_r2 = _mm_unpacklo_epi64(temp1, temp3); //a2 b2 c2 d2 137 temp1 = _mm_add_epi32(src_r1, src_r2); 138 temp2 = _mm_sub_epi32(src_r1, src_r2); 143 src_r2 = _mm_sub_epi32(temp0, temp1); 157 temp2 = _mm_unpacklo_epi32(src_r2, src_r3); //a2 a3 b2 b3 159 temp3 = _mm_unpackhi_epi32(src_r2, src_r3); //c2 c3 d2 d [all...] |
ih264_ihadamard_scaling_sse42.c | 97 __m128i src_r0, src_r1, src_r2, src_r3; local 110 src_r2 = _mm_cvtepi16_epi32(src_r2_r3); 126 temp2 = _mm_unpacklo_epi32(src_r2, src_r3); //c0 d0 c1 d1 128 temp3 = _mm_unpackhi_epi32(src_r2, src_r3); //c2 d2 c3 d3 131 src_r2 = _mm_unpacklo_epi64(temp1, temp3); //a2 b2 c2 d2 135 temp1 = _mm_add_epi32(src_r1, src_r2); 136 temp2 = _mm_sub_epi32(src_r1, src_r2); 141 src_r2 = _mm_sub_epi32(temp0, temp1); 155 temp2 = _mm_unpacklo_epi32(src_r2, src_r3); //a2 a3 b2 b3 157 temp3 = _mm_unpackhi_epi32(src_r2, src_r3); //c2 c3 d2 d [all...] |
ih264_resi_trans_quant_sse42.c | 118 __m128i src_r0, src_r1, src_r2, src_r3; local 131 src_r2 = _mm_loadl_epi64((__m128i *) (&pu1_src[2 * src_strd])); //a20 a21 a22 a23 0 0 0 0 0 0 0 0 -- all 8 bits 136 src_r2 = _mm_cvtepu8_epi16(src_r2); 151 src_r2 = _mm_sub_epi16(src_r2, pred_r2); 166 temp2 = _mm_unpacklo_epi16(src_r2, src_r3); //c0 d0 c1 d1 c2 d2 c3 d3 172 src_r2 = _mm_unpacklo_epi64(temp3, zero_8x16b); //a2 b2 c2 d2 179 temp1 = _mm_add_epi16(src_r1, src_r2); 181 temp2 = _mm_sub_epi16(src_r1, src_r2); 392 __m128i src_r0, src_r1, src_r2, src_r3; local 676 __m128i src_r0, src_r1, src_r2, src_r3; local [all...] |
ih264_iquant_itrans_recon_sse42.c | 114 __m128i src_r0, src_r1, src_r2, src_r3; local 146 src_r2 = _mm_unpacklo_epi16(src_r2_r3, zero_8x16b); // a20 0 a21 0 a22 0 a23 0 -- 16 bit long 151 temp6 = _mm_madd_epi16(src_r2, temp6); 363 __m128i src_r0, src_r1, src_r2, src_r3; local 397 src_r2 = _mm_unpacklo_epi16(src_r2_r3, zero_8x16b); // a20 0 a21 0 a22 0 a23 0 -- 16 bit long 402 temp6 = _mm_madd_epi16(src_r2, temp6);
|
ih264_iquant_itrans_recon_ssse3.c | 114 __m128i src_r0, src_r1, src_r2, src_r3; local 147 src_r2 = _mm_unpacklo_epi16(src_r2_r3, zero_8x16b); // a20 0 a21 0 a22 0 a23 0 -- 16 bit long 152 temp6 = _mm_madd_epi16(src_r2, temp6); [all...] |
/external/libmpeg2/common/x86/ |
impeg2_inter_pred_sse42_intr.c | 76 __m128i src_r0, src_r1, src_r2, src_r3; local 86 src_r2 = _mm_loadu_si128((__m128i *) (src + 2 * src_wd)); 91 _mm_storeu_si128((__m128i *) (dst + 2 * dst_wd), src_r2); 99 src_r2 = _mm_loadu_si128((__m128i *) (src + 2 * src_wd)); 104 _mm_storeu_si128((__m128i *) (dst + 2 * dst_wd), src_r2); 112 src_r2 = _mm_loadu_si128((__m128i *) (src + 2 * src_wd)); 117 _mm_storeu_si128((__m128i *) (dst + 2 * dst_wd), src_r2); 125 src_r2 = _mm_loadu_si128((__m128i *) (src + 2 * src_wd)); 130 _mm_storeu_si128((__m128i *) (dst + 2 * dst_wd), src_r2); 145 src_r2 = _mm_loadl_epi64((__m128i *)(src + 2 * src_wd)) 792 __m128i src_r0, src_r1, src_r2, temp0, temp1; local 873 __m128i src_r0, src_r1, src_r2, src_r3; local [all...] |
/external/libavc/encoder/x86/ |
ime_distortion_metrics_sse42.c | 108 __m128i src_r0, src_r1, src_r2, src_r3; local 118 src_r2 = _mm_loadu_si128((__m128i *) (pu1_src + 2*src_strd)); 128 res_r2 = _mm_sad_epu8(src_r2, est_r2); 141 src_r2 = _mm_loadu_si128((__m128i *) (pu1_src + 2*src_strd)); 151 res_r2 = _mm_sad_epu8(src_r2, est_r2); 164 src_r2 = _mm_loadu_si128((__m128i *) (pu1_src + 2*src_strd)); 174 res_r2 = _mm_sad_epu8(src_r2, est_r2); 187 src_r2 = _mm_loadu_si128((__m128i *) (pu1_src + 2*src_strd)); 197 res_r2 = _mm_sad_epu8(src_r2, est_r2); 252 __m128i src_r0, src_r1, src_r2, src_r3 local 348 __m128i src_r0, src_r1, src_r2, src_r3; local 503 __m128i src_r0, src_r1, src_r2, src_r3; local 1728 __m128i src_r0, src_r1, src_r2, src_r3; local [all...] |