/external/libmpeg2/common/x86/ |
D | impeg2_inter_pred_sse42_intr.c | 76 __m128i src_r0, src_r1, src_r2, src_r3; in impeg2_copy_mb_sse42() local 84 src_r0 = _mm_loadu_si128((__m128i *) (src)); in impeg2_copy_mb_sse42() 89 _mm_storeu_si128((__m128i *) dst, src_r0); in impeg2_copy_mb_sse42() 97 src_r0 = _mm_loadu_si128((__m128i *) (src)); in impeg2_copy_mb_sse42() 102 _mm_storeu_si128((__m128i *) dst, src_r0); in impeg2_copy_mb_sse42() 110 src_r0 = _mm_loadu_si128((__m128i *) (src)); in impeg2_copy_mb_sse42() 115 _mm_storeu_si128((__m128i *) dst, src_r0); in impeg2_copy_mb_sse42() 123 src_r0 = _mm_loadu_si128((__m128i *) (src)); in impeg2_copy_mb_sse42() 128 _mm_storeu_si128((__m128i *) dst, src_r0); in impeg2_copy_mb_sse42() 143 src_r0 = _mm_loadl_epi64((__m128i *)src); in impeg2_copy_mb_sse42() [all …]
|
D | icv_variance_ssse3.c | 90 __m128 src_r0, src_r1; in icv_variance_8x4_ssse3() local 110 src_r0 = (__m128)_mm_loadl_epi64((__m128i *) (pu1_src)); in icv_variance_8x4_ssse3() 116 src_r0 = _mm_loadh_pi (src_r0, (__m64 *) (pu1_src)); in icv_variance_8x4_ssse3() 124 sum_r0 = _mm_sad_epu8((__m128i)src_r0, zero); in icv_variance_8x4_ssse3() 134 ssrc_r0 = _mm_unpacklo_epi8((__m128i)src_r0, zero); in icv_variance_8x4_ssse3() 136 ssrc_r2 = _mm_unpackhi_epi8((__m128i)src_r0, zero); in icv_variance_8x4_ssse3()
|
D | icv_sad_ssse3.c | 98 __m128 src_r0, src_r1; in icv_sad_8x4_ssse3() local 108 src_r0 = (__m128)_mm_loadl_epi64((__m128i *) (pu1_src)); in icv_sad_8x4_ssse3() 114 src_r0 = _mm_loadh_pi (src_r0, (__m64 *) (pu1_src)); in icv_sad_8x4_ssse3() 135 res_r0 = _mm_sad_epu8((__m128i)src_r0, (__m128i)ref_r0); in icv_sad_8x4_ssse3()
|
/external/libavc/common/x86/ |
D | ih264_ihadamard_scaling_ssse3.c | 97 __m128i src_r0, src_r1, src_r2, src_r3; in ih264_ihadamard_scaling_4x4_ssse3() local 110 src_r0 = _mm_unpacklo_epi16(src_r0_r1, sign_reg); in ih264_ihadamard_scaling_4x4_ssse3() 127 temp0 = _mm_unpacklo_epi32(src_r0, src_r1); //a0 b0 a1 b1 in ih264_ihadamard_scaling_4x4_ssse3() 129 temp1 = _mm_unpackhi_epi32(src_r0, src_r1); //a2 b2 a3 b3 in ih264_ihadamard_scaling_4x4_ssse3() 131 src_r0 = _mm_unpacklo_epi64(temp0, temp2); //a0 b0 c0 d0 in ih264_ihadamard_scaling_4x4_ssse3() 136 temp0 = _mm_add_epi32(src_r0, src_r3); in ih264_ihadamard_scaling_4x4_ssse3() 139 temp3 = _mm_sub_epi32(src_r0, src_r3); in ih264_ihadamard_scaling_4x4_ssse3() 141 src_r0 = _mm_add_epi32(temp0, temp1); in ih264_ihadamard_scaling_4x4_ssse3() 156 temp0 = _mm_unpacklo_epi32(src_r0, src_r1); //a0 a1 b0 b1 in ih264_ihadamard_scaling_4x4_ssse3() 158 temp1 = _mm_unpackhi_epi32(src_r0, src_r1); //c0 c1 d0 d1 in ih264_ihadamard_scaling_4x4_ssse3() [all …]
|
D | ih264_resi_trans_quant_sse42.c | 118 __m128i src_r0, src_r1, src_r2, src_r3; in ih264_resi_trans_quant_4x4_sse42() 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 bi… in ih264_resi_trans_quant_4x4_sse42() 134 src_r0 = _mm_cvtepu8_epi16(src_r0); in ih264_resi_trans_quant_4x4_sse42() 149 src_r0 = _mm_sub_epi16(src_r0, pred_r0); in ih264_resi_trans_quant_4x4_sse42() 165 temp0 = _mm_unpacklo_epi16(src_r0, src_r1); //a0 b0 a1 b1 a2 b2 a3 b3 in ih264_resi_trans_quant_4x4_sse42() 170 src_r0 = _mm_unpacklo_epi64(temp1, zero_8x16b); //a0 b0 c0 d0 in ih264_resi_trans_quant_4x4_sse42() 177 temp0 = _mm_add_epi16(src_r0, src_r3); in ih264_resi_trans_quant_4x4_sse42() 183 temp3 = _mm_sub_epi16(src_r0, src_r3); in ih264_resi_trans_quant_4x4_sse42() 186 src_r0 = _mm_add_epi16(temp0, temp1); in ih264_resi_trans_quant_4x4_sse42() 203 temp0 = _mm_unpacklo_epi16(src_r0, src_r1); //a0 a1 b0 b1 c0 c1 d0 d1 in ih264_resi_trans_quant_4x4_sse42() [all …]
|
D | ih264_ihadamard_scaling_sse42.c | 97 __m128i src_r0, src_r1, src_r2, src_r3; in ih264_ihadamard_scaling_4x4_sse42() local 106 src_r0 = _mm_cvtepi16_epi32(src_r0_r1); in ih264_ihadamard_scaling_4x4_sse42() 125 temp0 = _mm_unpacklo_epi32(src_r0, src_r1); //a0 b0 a1 b1 in ih264_ihadamard_scaling_4x4_sse42() 127 temp1 = _mm_unpackhi_epi32(src_r0, src_r1); //a2 b2 a3 b3 in ih264_ihadamard_scaling_4x4_sse42() 129 src_r0 = _mm_unpacklo_epi64(temp0, temp2); //a0 b0 c0 d0 in ih264_ihadamard_scaling_4x4_sse42() 134 temp0 = _mm_add_epi32(src_r0, src_r3); in ih264_ihadamard_scaling_4x4_sse42() 137 temp3 = _mm_sub_epi32(src_r0, src_r3); in ih264_ihadamard_scaling_4x4_sse42() 139 src_r0 = _mm_add_epi32(temp0, temp1); in ih264_ihadamard_scaling_4x4_sse42() 154 temp0 = _mm_unpacklo_epi32(src_r0, src_r1); //a0 a1 b0 b1 in ih264_ihadamard_scaling_4x4_sse42() 156 temp1 = _mm_unpackhi_epi32(src_r0, src_r1); //c0 c1 d0 d1 in ih264_ihadamard_scaling_4x4_sse42() [all …]
|
D | ih264_iquant_itrans_recon_ssse3.c | 114 __m128i src_r0, src_r1, src_r2, src_r3; in ih264_iquant_itrans_recon_4x4_ssse3() local 145 src_r0 = _mm_unpacklo_epi16(src_r0_r1, zero_8x16b); // a00 0 a01 0 a02 0 a03 0 -- 16 bit long in ih264_iquant_itrans_recon_4x4_ssse3() 150 …temp4 = _mm_madd_epi16(src_r0, temp4); //a00*b00*q00 a10*b10*q10 a20*b20*q20 a30*b30 q30 -- 32 bit… in ih264_iquant_itrans_recon_4x4_ssse3() 381 __m128i src_r0; in ih264_iquant_itrans_recon_8x8_ssse3() local 412 …src_r0 = _mm_loadu_si128((__m128i *) (pi2_src)); //a00 a01 a02 a03 a04 a05 a06 a07 -- the source m… in ih264_iquant_itrans_recon_8x8_ssse3() 415 src_r0_1 = _mm_unpacklo_epi16(src_r0, zero_8x16b); //a00 0 a01 0 a02 0 a03 0 -- 16 bit long in ih264_iquant_itrans_recon_8x8_ssse3() 416 src_r0_2 = _mm_unpackhi_epi16(src_r0, zero_8x16b); // a04 0 a05 0 a06 0 a07 0 -- 16 bit long in ih264_iquant_itrans_recon_8x8_ssse3() 435 …src_r0 = _mm_loadu_si128((__m128i *) (pi2_src + 8)); //a00 a01 a02 a03 a04 a05 a06 a07 a08 -- the … in ih264_iquant_itrans_recon_8x8_ssse3() 438 src_r0_1 = _mm_unpacklo_epi16(src_r0, zero_8x16b); //a00 0 a01 0 a02 0 a03 0 -- 16 bit long in ih264_iquant_itrans_recon_8x8_ssse3() 439 src_r0_2 = _mm_unpackhi_epi16(src_r0, zero_8x16b); // a04 0 a05 0 a06 0 a07 0 -- 16 bit long in ih264_iquant_itrans_recon_8x8_ssse3() [all …]
|
D | ih264_iquant_itrans_recon_sse42.c | 114 __m128i src_r0, src_r1, src_r2, src_r3; in ih264_iquant_itrans_recon_4x4_sse42() local 144 src_r0 = _mm_unpacklo_epi16(src_r0_r1, zero_8x16b); // a00 0 a01 0 a02 0 a03 0 -- 16 bit long in ih264_iquant_itrans_recon_4x4_sse42() 149 …temp4 = _mm_madd_epi16(src_r0, temp4); //a00*b00*q00 a10*b10*q10 a20*b20*q20 a30*b30 q30 -- 32 bit… in ih264_iquant_itrans_recon_4x4_sse42() 363 __m128i src_r0, src_r1, src_r2, src_r3; in ih264_iquant_itrans_recon_chroma_4x4_sse42() local 395 src_r0 = _mm_unpacklo_epi16(src_r0_r1, zero_8x16b); // a00 0 a01 0 a02 0 a03 0 -- 16 bit long in ih264_iquant_itrans_recon_chroma_4x4_sse42() 400 …temp4 = _mm_madd_epi16(src_r0, temp4); //a00*b00*q00 a10*b10*q10 a20*b20*q20 a30*b30 q30 -- 32 bit… in ih264_iquant_itrans_recon_chroma_4x4_sse42()
|
/external/libavc/encoder/x86/ |
D | ime_distortion_metrics_sse42.c | 108 __m128i src_r0, src_r1, src_r2, src_r3; in ime_compute_sad_16x16_sse42() local 116 src_r0 = _mm_loadu_si128((__m128i *) (pu1_src)); in ime_compute_sad_16x16_sse42() 126 res_r0 = _mm_sad_epu8(src_r0, est_r0); in ime_compute_sad_16x16_sse42() 139 src_r0 = _mm_loadu_si128((__m128i *) (pu1_src)); in ime_compute_sad_16x16_sse42() 149 res_r0 = _mm_sad_epu8(src_r0, est_r0); in ime_compute_sad_16x16_sse42() 162 src_r0 = _mm_loadu_si128((__m128i *) (pu1_src)); in ime_compute_sad_16x16_sse42() 172 res_r0 = _mm_sad_epu8(src_r0, est_r0); in ime_compute_sad_16x16_sse42() 185 src_r0 = _mm_loadu_si128((__m128i *) (pu1_src)); in ime_compute_sad_16x16_sse42() 195 res_r0 = _mm_sad_epu8(src_r0, est_r0); in ime_compute_sad_16x16_sse42() 252 __m128i src_r0, src_r1, src_r2, src_r3; in ime_compute_sad_16x8_sse42() local [all …]
|
/external/mesa3d/src/gallium/drivers/swr/rasterizer/core/ |
D | format_utils.h | 451 simd4scalar src_r0 = SIMD128::load_ps(pfSrc + 0); in Transpose() local 456 simd4scalar dst0 = SIMD128::unpacklo_ps(src_r0, src_g0); in Transpose() 457 simd4scalar dst1 = SIMD128::unpackhi_ps(src_r0, src_g0); in Transpose()
|