| /external/libavc/common/x86/ |
| ih264_ihadamard_scaling_ssse3.c | 97 __m128i src_r0, src_r1, src_r2, src_r3; local
111 src_r1 = _mm_unpackhi_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
132 src_r1 = _mm_unpackhi_epi64(temp0, temp2); //a1 b1 c1 d1
137 temp1 = _mm_add_epi32(src_r1, src_r2);
138 temp2 = _mm_sub_epi32(src_r1, src_r2);
142 src_r1 = _mm_add_epi32(temp2, temp3);
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
108 src_r1 = _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
130 src_r1 = _mm_unpackhi_epi64(temp0, temp2); //a1 b1 c1 d1
135 temp1 = _mm_add_epi32(src_r1, src_r2);
136 temp2 = _mm_sub_epi32(src_r1, src_r2);
140 src_r1 = _mm_add_epi32(temp2, temp3);
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
130 src_r1 = _mm_loadl_epi64((__m128i *) (&pu1_src[src_strd])); //a10 a11 a12 a13 0 0 0 0 0 0 0 0 -- all 8 bits
135 src_r1 = _mm_cvtepu8_epi16(src_r1);
150 src_r1 = _mm_sub_epi16(src_r1, pred_r1);
165 temp0 = _mm_unpacklo_epi16(src_r0, src_r1); //a0 b0 a1 b1 a2 b2 a3 b3
171 src_r1 = _mm_unpackhi_epi64(temp1, zero_8x16b); //a1 b1 c1 d1
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
145 src_r1 = _mm_unpackhi_epi16(src_r0_r1, zero_8x16b); // a10 0 a11 0 a12 0 a13 0 -- 16 bit long
150 temp5 = _mm_madd_epi16(src_r1, temp5);
363 __m128i src_r0, src_r1, src_r2, src_r3; local
396 src_r1 = _mm_unpackhi_epi16(src_r0_r1, zero_8x16b); // a10 0 a11 0 a12 0 a13 0 -- 16 bit long
401 temp5 = _mm_madd_epi16(src_r1, temp5);
|
| ih264_iquant_itrans_recon_ssse3.c | 114 __m128i src_r0, src_r1, src_r2, src_r3; local
146 src_r1 = _mm_unpackhi_epi16(src_r0_r1, zero_8x16b); // a10 0 a11 0 a12 0 a13 0 -- 16 bit long
151 temp5 = _mm_madd_epi16(src_r1, temp5);
[all...] |
| /external/libmpeg2/common/x86/ |
| icv_sad_ssse3.c | 98 __m128 src_r0, src_r1; local
111 src_r1 = (__m128)_mm_loadl_epi64((__m128i *) (pu1_src));
117 src_r1 = _mm_loadh_pi (src_r1, (__m64 *) (pu1_src));
136 res_r1 = _mm_sad_epu8((__m128i)src_r1, (__m128i)ref_r1);
|
| icv_variance_ssse3.c | 90 __m128 src_r0, src_r1; local
113 src_r1 = (__m128)_mm_loadl_epi64((__m128i *) (pu1_src));
119 src_r1 = _mm_loadh_pi (src_r1, (__m64 *) (pu1_src));
125 sum_r1 = _mm_sad_epu8((__m128i)src_r1, zero);
135 ssrc_r1 = _mm_unpacklo_epi8((__m128i)src_r1, zero);
137 ssrc_r3 = _mm_unpackhi_epi8((__m128i)src_r1, zero);
|
| impeg2_inter_pred_sse42_intr.c | 76 __m128i src_r0, src_r1, src_r2, src_r3; local
85 src_r1 = _mm_loadu_si128((__m128i *) (src + src_wd));
90 _mm_storeu_si128((__m128i *) (dst + dst_wd), src_r1);
98 src_r1 = _mm_loadu_si128((__m128i *) (src + src_wd));
103 _mm_storeu_si128((__m128i *) (dst + dst_wd), src_r1);
111 src_r1 = _mm_loadu_si128((__m128i *) (src + src_wd));
116 _mm_storeu_si128((__m128i *) (dst + dst_wd), src_r1);
124 src_r1 = _mm_loadu_si128((__m128i *) (src + src_wd));
129 _mm_storeu_si128((__m128i *) (dst + dst_wd), src_r1);
144 src_r1 = _mm_loadl_epi64((__m128i *)(src + src_wd))
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...] |
| /external/libavc/encoder/x86/ |
| ime_distortion_metrics_sse42.c | 108 __m128i src_r0, src_r1, src_r2, src_r3; local
117 src_r1 = _mm_loadu_si128((__m128i *) (pu1_src + src_strd));
127 res_r1 = _mm_sad_epu8(src_r1, est_r1);
140 src_r1 = _mm_loadu_si128((__m128i *) (pu1_src + src_strd));
150 res_r1 = _mm_sad_epu8(src_r1, est_r1);
163 src_r1 = _mm_loadu_si128((__m128i *) (pu1_src + src_strd));
173 res_r1 = _mm_sad_epu8(src_r1, est_r1);
186 src_r1 = _mm_loadu_si128((__m128i *) (pu1_src + src_strd));
196 res_r1 = _mm_sad_epu8(src_r1, est_r1);
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...] |
| /external/mesa3d/src/gallium/drivers/swr/rasterizer/core/ |
| utils.h | 517 __m128 src_r1 = _mm_load_ps(pfSrc + 4); local
523 __m128 dst2 = _mm_unpacklo_ps(src_r1, src_g1);
524 __m128 dst3 = _mm_unpackhi_ps(src_r1, src_g1);
[all...] |
