| /external/libmpeg2/common/x86/ |
| impeg2_inter_pred_sse42_intr.c | 76 __m128i src_r0, src_r1, src_r2, src_r3; local
84 src_r0 = _mm_loadu_si128((__m128i *) (src));
89 _mm_storeu_si128((__m128i *) dst, src_r0);
97 src_r0 = _mm_loadu_si128((__m128i *) (src));
102 _mm_storeu_si128((__m128i *) dst, src_r0);
110 src_r0 = _mm_loadu_si128((__m128i *) (src));
115 _mm_storeu_si128((__m128i *) dst, src_r0);
123 src_r0 = _mm_loadu_si128((__m128i *) (src));
128 _mm_storeu_si128((__m128i *) dst, src_r0);
143 src_r0 = _mm_loadl_epi64((__m128i *)src)
479 __m128i src_r0, src_r0_1, src_r1, src_r1_1; local
684 __m128i src_r0, src_r0_1, src_r1, src_r1_1; local
792 __m128i src_r0, src_r1, src_r2, temp0, temp1; local
873 __m128i src_r0, src_r1, src_r2, src_r3; local
[all...] |
| icv_sad_ssse3.c | 98 __m128 src_r0, src_r1; local
108 src_r0 = (__m128)_mm_loadl_epi64((__m128i *) (pu1_src));
114 src_r0 = _mm_loadh_pi (src_r0, (__m64 *) (pu1_src));
135 res_r0 = _mm_sad_epu8((__m128i)src_r0, (__m128i)ref_r0);
|
| icv_variance_ssse3.c | 90 __m128 src_r0, src_r1; local
110 src_r0 = (__m128)_mm_loadl_epi64((__m128i *) (pu1_src));
116 src_r0 = _mm_loadh_pi (src_r0, (__m64 *) (pu1_src));
124 sum_r0 = _mm_sad_epu8((__m128i)src_r0, zero);
134 ssrc_r0 = _mm_unpacklo_epi8((__m128i)src_r0, zero);
136 ssrc_r2 = _mm_unpackhi_epi8((__m128i)src_r0, zero);
|
| /external/libavc/common/x86/ |
| ih264_ihadamard_scaling_ssse3.c | 97 __m128i src_r0, src_r1, src_r2, src_r3; local
110 src_r0 = _mm_unpacklo_epi16(src_r0_r1, sign_reg);
127 temp0 = _mm_unpacklo_epi32(src_r0, src_r1); //a0 b0 a1 b1
129 temp1 = _mm_unpackhi_epi32(src_r0, src_r1); //a2 b2 a3 b3
131 src_r0 = _mm_unpacklo_epi64(temp0, temp2); //a0 b0 c0 d0
136 temp0 = _mm_add_epi32(src_r0, src_r3);
139 temp3 = _mm_sub_epi32(src_r0, src_r3);
141 src_r0 = _mm_add_epi32(temp0, temp1);
156 temp0 = _mm_unpacklo_epi32(src_r0, src_r1); //a0 a1 b0 b1
158 temp1 = _mm_unpackhi_epi32(src_r0, src_r1); //c0 c1 d0 d
[all...] |
| ih264_ihadamard_scaling_sse42.c | 97 __m128i src_r0, src_r1, src_r2, src_r3; local
106 src_r0 = _mm_cvtepi16_epi32(src_r0_r1);
125 temp0 = _mm_unpacklo_epi32(src_r0, src_r1); //a0 b0 a1 b1
127 temp1 = _mm_unpackhi_epi32(src_r0, src_r1); //a2 b2 a3 b3
129 src_r0 = _mm_unpacklo_epi64(temp0, temp2); //a0 b0 c0 d0
134 temp0 = _mm_add_epi32(src_r0, src_r3);
137 temp3 = _mm_sub_epi32(src_r0, src_r3);
139 src_r0 = _mm_add_epi32(temp0, temp1);
154 temp0 = _mm_unpacklo_epi32(src_r0, src_r1); //a0 a1 b0 b1
156 temp1 = _mm_unpackhi_epi32(src_r0, src_r1); //c0 c1 d0 d
[all...] |
| ih264_resi_trans_quant_sse42.c | 118 __m128i src_r0, src_r1, src_r2, src_r3; local
129 src_r0 = _mm_loadl_epi64((__m128i *) (&pu1_src[0])); //a00 a01 a02 a03 0 0 0 0 0 0 0 0 -- all 8 bits
134 src_r0 = _mm_cvtepu8_epi16(src_r0);
149 src_r0 = _mm_sub_epi16(src_r0, pred_r0);
165 temp0 = _mm_unpacklo_epi16(src_r0, src_r1); //a0 b0 a1 b1 a2 b2 a3 b3
170 src_r0 = _mm_unpacklo_epi64(temp1, zero_8x16b); //a0 b0 c0 d0
177 temp0 = _mm_add_epi16(src_r0, src_r3);
183 temp3 = _mm_sub_epi16(src_r0, src_r3)
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_ssse3.c | 114 __m128i src_r0, src_r1, src_r2, src_r3; local
145 src_r0 = _mm_unpacklo_epi16(src_r0_r1, zero_8x16b); // a00 0 a01 0 a02 0 a03 0 -- 16 bit long
150 temp4 = _mm_madd_epi16(src_r0, temp4); //a00*b00*q00 a10*b10*q10 a20*b20*q20 a30*b30 q30 -- 32 bits long
381 __m128i src_r0; local
412 src_r0 = _mm_loadu_si128((__m128i *) (pi2_src)); //a00 a01 a02 a03 a04 a05 a06 a07 -- the source matrix 0th row
415 src_r0_1 = _mm_unpacklo_epi16(src_r0, zero_8x16b); //a00 0 a01 0 a02 0 a03 0 -- 16 bit long
416 src_r0_2 = _mm_unpackhi_epi16(src_r0, zero_8x16b); // a04 0 a05 0 a06 0 a07 0 -- 16 bit long
435 src_r0 = _mm_loadu_si128((__m128i *) (pi2_src + 8)); //a00 a01 a02 a03 a04 a05 a06 a07 a08 -- the source matrix 1st row
438 src_r0_1 = _mm_unpacklo_epi16(src_r0, zero_8x16b); //a00 0 a01 0 a02 0 a03 0 -- 16 bit long
439 src_r0_2 = _mm_unpackhi_epi16(src_r0, zero_8x16b); // a04 0 a05 0 a06 0 a07 0 -- 16 bit lon
[all...] |
| ih264_iquant_itrans_recon_sse42.c | 114 __m128i src_r0, src_r1, src_r2, src_r3; local
144 src_r0 = _mm_unpacklo_epi16(src_r0_r1, zero_8x16b); // a00 0 a01 0 a02 0 a03 0 -- 16 bit long
149 temp4 = _mm_madd_epi16(src_r0, temp4); //a00*b00*q00 a10*b10*q10 a20*b20*q20 a30*b30 q30 -- 32 bits long
363 __m128i src_r0, src_r1, src_r2, src_r3; local
395 src_r0 = _mm_unpacklo_epi16(src_r0_r1, zero_8x16b); // a00 0 a01 0 a02 0 a03 0 -- 16 bit long
400 temp4 = _mm_madd_epi16(src_r0, temp4); //a00*b00*q00 a10*b10*q10 a20*b20*q20 a30*b30 q30 -- 32 bits long
|
| /external/libavc/encoder/x86/ |
| ime_distortion_metrics_sse42.c | 108 __m128i src_r0, src_r1, src_r2, src_r3; local
116 src_r0 = _mm_loadu_si128((__m128i *) (pu1_src));
126 res_r0 = _mm_sad_epu8(src_r0, est_r0);
139 src_r0 = _mm_loadu_si128((__m128i *) (pu1_src));
149 res_r0 = _mm_sad_epu8(src_r0, est_r0);
162 src_r0 = _mm_loadu_si128((__m128i *) (pu1_src));
172 res_r0 = _mm_sad_epu8(src_r0, est_r0);
185 src_r0 = _mm_loadu_si128((__m128i *) (pu1_src));
195 res_r0 = _mm_sad_epu8(src_r0, est_r0);
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...] |
