Home
last modified time | relevance | path

Searched refs:_mm_maddubs_epi16 (Results 1 – 25 of 60) sorted by relevance

123

/external/libvpx/libvpx/vpx_dsp/x86/
Dconvolve_ssse3.h47 const __m128i x0 = _mm_maddubs_epi16(s[0], f[0]); in convolve8_8_ssse3()
48 const __m128i x1 = _mm_maddubs_epi16(s[1], f[1]); in convolve8_8_ssse3()
49 const __m128i x2 = _mm_maddubs_epi16(s[2], f[2]); in convolve8_8_ssse3()
50 const __m128i x3 = _mm_maddubs_epi16(s[3], f[3]); in convolve8_8_ssse3()
70 const __m128i x0 = _mm_maddubs_epi16(s[0], f[0]); in convolve8_8_even_offset_ssse3()
71 const __m128i x1 = _mm_maddubs_epi16(s[1], f[1]); in convolve8_8_even_offset_ssse3()
72 const __m128i x2 = _mm_maddubs_epi16(s[2], f[2]); in convolve8_8_even_offset_ssse3()
73 const __m128i x3 = _mm_maddubs_epi16(s[3], f[3]); in convolve8_8_even_offset_ssse3()
75 const __m128i x4 = _mm_maddubs_epi16(s[1], _mm_set1_epi8(64)); in convolve8_8_even_offset_ssse3()
91 const __m128i x0 = _mm_maddubs_epi16(s[0], f[0]); in convolve8_8_odd_offset_ssse3()
[all …]
Dvpx_subpixel_8t_intrin_ssse3.c88 srcRegFilt1 = _mm_maddubs_epi16(srcRegFilt1, firstFilters); in vpx_filter_block1d4_h8_intrin_ssse3()
89 srcRegFilt2 = _mm_maddubs_epi16(srcRegFilt2, secondFilters); in vpx_filter_block1d4_h8_intrin_ssse3()
249 tmp_0 = _mm_maddubs_epi16(src_reg_shift_0, kernel_reg_23); in vpx_filter_block1d16_h4_ssse3()
250 tmp_1 = _mm_maddubs_epi16(src_reg_shift_2, kernel_reg_45); in vpx_filter_block1d16_h4_ssse3()
260 tmp_0 = _mm_maddubs_epi16(src_reg_shift_0, kernel_reg_23); in vpx_filter_block1d16_h4_ssse3()
261 tmp_1 = _mm_maddubs_epi16(src_reg_shift_2, kernel_reg_45); in vpx_filter_block1d16_h4_ssse3()
343 res_reg_m10_lo = _mm_maddubs_epi16(src_reg_m10_lo, kernel_reg_23); in vpx_filter_block1d16_v4_ssse3()
344 res_reg_01_lo = _mm_maddubs_epi16(src_reg_01_lo, kernel_reg_23); in vpx_filter_block1d16_v4_ssse3()
346 res_reg_12_lo = _mm_maddubs_epi16(src_reg_12_lo, kernel_reg_45); in vpx_filter_block1d16_v4_ssse3()
347 res_reg_23_lo = _mm_maddubs_epi16(src_reg_23_lo, kernel_reg_45); in vpx_filter_block1d16_v4_ssse3()
[all …]
Dconvolve_avx2.h80 const __m128i x0 = _mm_maddubs_epi16(_mm256_castsi256_si128(s[0]), in convolve8_8_avx2()
82 const __m128i x1 = _mm_maddubs_epi16(_mm256_castsi256_si128(s[1]), in convolve8_8_avx2()
84 const __m128i x2 = _mm_maddubs_epi16(_mm256_castsi256_si128(s[2]), in convolve8_8_avx2()
86 const __m128i x3 = _mm_maddubs_epi16(_mm256_castsi256_si128(s[3]), in convolve8_8_avx2()
/external/libhevc/common/x86/
Dihevc_inter_pred_filters_ssse3_intr.c336 … res_temp1_8x16b = _mm_maddubs_epi16(src_temp3_16x8b, coeff0_1_8x16b); /* row = 0 */ in ihevc_inter_pred_luma_horz_ssse3()
342 … res_temp2_8x16b = _mm_maddubs_epi16(src_temp4_16x8b, coeff2_3_8x16b); /* row = 0 */ in ihevc_inter_pred_luma_horz_ssse3()
348 … res_temp3_8x16b = _mm_maddubs_epi16(src_temp5_16x8b, coeff4_5_8x16b); /* row = 0 */ in ihevc_inter_pred_luma_horz_ssse3()
354 … res_temp4_8x16b = _mm_maddubs_epi16(src_temp6_16x8b, coeff6_7_8x16b); /* row = 0 */ in ihevc_inter_pred_luma_horz_ssse3()
369 … res_temp11_8x16b = _mm_maddubs_epi16(src_temp13_16x8b, coeff0_1_8x16b); /* row = 1 */ in ihevc_inter_pred_luma_horz_ssse3()
375 … res_temp12_8x16b = _mm_maddubs_epi16(src_temp14_16x8b, coeff2_3_8x16b); /* row = 1 */ in ihevc_inter_pred_luma_horz_ssse3()
381 … res_temp13_8x16b = _mm_maddubs_epi16(src_temp15_16x8b, coeff4_5_8x16b); /* row = 1 */ in ihevc_inter_pred_luma_horz_ssse3()
387 … res_temp14_8x16b = _mm_maddubs_epi16(src_temp16_16x8b, coeff6_7_8x16b); /* row = 1 */ in ihevc_inter_pred_luma_horz_ssse3()
425 … res_temp1_8x16b = _mm_maddubs_epi16(src_temp3_16x8b, coeff0_1_8x16b); /* row = 0 */ in ihevc_inter_pred_luma_horz_ssse3()
431 … res_temp2_8x16b = _mm_maddubs_epi16(src_temp4_16x8b, coeff2_3_8x16b); /* row = 0 */ in ihevc_inter_pred_luma_horz_ssse3()
[all …]
Dihevc_intra_pred_filters_ssse3_intr.c735 res_temp_8x16b = _mm_maddubs_epi16(pu1_ref_16x8b, row_16x8b); in ihevc_intra_pred_luma_planar_ssse3()
736 res_temp1_8x16b = _mm_maddubs_epi16(pu1_ref_16x8b, row1_16x8b); in ihevc_intra_pred_luma_planar_ssse3()
737 res_temp2_8x16b = _mm_maddubs_epi16(pu1_ref_16x8b, row2_16x8b); in ihevc_intra_pred_luma_planar_ssse3()
738 res_temp3_8x16b = _mm_maddubs_epi16(pu1_ref_16x8b, row3_16x8b); in ihevc_intra_pred_luma_planar_ssse3()
740 res_temp4_8x16b = _mm_maddubs_epi16(res_temp4_8x16b, nt_row_16x8b); in ihevc_intra_pred_luma_planar_ssse3()
741 res_temp5_8x16b = _mm_maddubs_epi16(res_temp5_8x16b, nt_row1_16x8b); in ihevc_intra_pred_luma_planar_ssse3()
742 res_temp6_8x16b = _mm_maddubs_epi16(res_temp6_8x16b, nt_row2_16x8b); in ihevc_intra_pred_luma_planar_ssse3()
743 res_temp7_8x16b = _mm_maddubs_epi16(res_temp7_8x16b, nt_row3_16x8b); in ihevc_intra_pred_luma_planar_ssse3()
813 res_temp_8x16b = _mm_maddubs_epi16(pu1_ref_16x8b, row_16x8b); in ihevc_intra_pred_luma_planar_ssse3()
814 res_temp1_8x16b = _mm_maddubs_epi16(pu1_ref_16x8b, row1_16x8b); in ihevc_intra_pred_luma_planar_ssse3()
[all …]
Dihevc_intra_pred_filters_sse42_intr.c1497 src_temp1_8x16b = _mm_maddubs_epi16(src_temp1_8x16b, temp1_8x16b); in ihevc_intra_pred_luma_mode_3_to_9_sse42()
1498 src_temp2_8x16b = _mm_maddubs_epi16(src_temp2_8x16b, temp2_8x16b); in ihevc_intra_pred_luma_mode_3_to_9_sse42()
1499 src_temp3_8x16b = _mm_maddubs_epi16(src_temp3_8x16b, temp3_8x16b); in ihevc_intra_pred_luma_mode_3_to_9_sse42()
1500 src_temp4_8x16b = _mm_maddubs_epi16(src_temp4_8x16b, temp4_8x16b); in ihevc_intra_pred_luma_mode_3_to_9_sse42()
1636 src_temp1_8x16b = _mm_maddubs_epi16(src_temp1_8x16b, temp1_8x16b); in ihevc_intra_pred_luma_mode_3_to_9_sse42()
1637 src_temp2_8x16b = _mm_maddubs_epi16(src_temp2_8x16b, temp2_8x16b); in ihevc_intra_pred_luma_mode_3_to_9_sse42()
1638 src_temp3_8x16b = _mm_maddubs_epi16(src_temp3_8x16b, temp3_8x16b); in ihevc_intra_pred_luma_mode_3_to_9_sse42()
1639 src_temp4_8x16b = _mm_maddubs_epi16(src_temp4_8x16b, temp4_8x16b); in ihevc_intra_pred_luma_mode_3_to_9_sse42()
1642 src_temp11_8x16b = _mm_maddubs_epi16(src_temp11_8x16b, temp11_8x16b); in ihevc_intra_pred_luma_mode_3_to_9_sse42()
1643 src_temp12_8x16b = _mm_maddubs_epi16(src_temp12_8x16b, temp12_8x16b); in ihevc_intra_pred_luma_mode_3_to_9_sse42()
[all …]
Dihevc_chroma_intra_pred_filters_ssse3_intr.c1265 src_temp1_8x16b = _mm_maddubs_epi16(src_temp1_8x16b, temp1_8x16b); in ihevc_intra_pred_chroma_mode_3_to_9_ssse3()
1266 src_temp2_8x16b = _mm_maddubs_epi16(src_temp2_8x16b, temp2_8x16b); in ihevc_intra_pred_chroma_mode_3_to_9_ssse3()
1267 src_temp3_8x16b = _mm_maddubs_epi16(src_temp3_8x16b, temp3_8x16b); in ihevc_intra_pred_chroma_mode_3_to_9_ssse3()
1268 src_temp4_8x16b = _mm_maddubs_epi16(src_temp4_8x16b, temp4_8x16b); in ihevc_intra_pred_chroma_mode_3_to_9_ssse3()
1433 src_temp1_8x16b = _mm_maddubs_epi16(src_temp1_8x16b, temp1_8x16b); in ihevc_intra_pred_chroma_mode_3_to_9_ssse3()
1434 src_temp2_8x16b = _mm_maddubs_epi16(src_temp2_8x16b, temp2_8x16b); in ihevc_intra_pred_chroma_mode_3_to_9_ssse3()
1435 src_temp3_8x16b = _mm_maddubs_epi16(src_temp3_8x16b, temp3_8x16b); in ihevc_intra_pred_chroma_mode_3_to_9_ssse3()
1436 src_temp4_8x16b = _mm_maddubs_epi16(src_temp4_8x16b, temp4_8x16b); in ihevc_intra_pred_chroma_mode_3_to_9_ssse3()
1439 src_temp11_8x16b = _mm_maddubs_epi16(src_temp11_8x16b, temp11_8x16b); in ihevc_intra_pred_chroma_mode_3_to_9_ssse3()
1440 src_temp12_8x16b = _mm_maddubs_epi16(src_temp12_8x16b, temp12_8x16b); in ihevc_intra_pred_chroma_mode_3_to_9_ssse3()
[all …]
Dihevc_deblk_ssse3_intr.c153 mask_d_result_4x32b = _mm_maddubs_epi16(src_tmp_8x16b, coef_8x16b); in ihevc_deblk_luma_vert_ssse3()
166 mask_de_result_8x16b = _mm_maddubs_epi16(mask_de_result_8x16b, temp_coef0_8x16b); in ihevc_deblk_luma_vert_ssse3()
274 temp_pq_str0_16x8b = _mm_maddubs_epi16(temp_pq_str0_16x8b, const2_8x16b); in ihevc_deblk_luma_vert_ssse3()
290 temp_pq1_str0_16x8b = _mm_maddubs_epi16(temp_pq1_str0_16x8b, const2_8x16b); in ihevc_deblk_luma_vert_ssse3()
291 temp_pq1_str1_16x8b = _mm_maddubs_epi16(temp_pq1_str1_16x8b, const2_8x16b); in ihevc_deblk_luma_vert_ssse3()
319 temp_pq2_str0_16x8b = _mm_maddubs_epi16(temp_str3_16x8b, const2_8x16b); in ihevc_deblk_luma_vert_ssse3()
411 tmp_delta3_8x16b = _mm_maddubs_epi16(tmp_delta2_8x16b, coefdelta_0_8x16b); in ihevc_deblk_luma_vert_ssse3()
433 tmp_delta3_8x16b = _mm_maddubs_epi16(tmp_delta2_8x16b, coefdelta_0_8x16b); in ihevc_deblk_luma_vert_ssse3()
613 mask_d_result_4x32b = _mm_maddubs_epi16(src_tmp_p_0_8x16b, coef_8x16b); in ihevc_deblk_luma_horz_ssse3()
626 mask_de_result_8x16b = _mm_maddubs_epi16(mask_de_result_8x16b, temp_coef0_8x16b); in ihevc_deblk_luma_horz_ssse3()
[all …]
/external/libavc/common/x86/
Dih264_inter_pred_filters_ssse3.c262 …res_r0r1_t1_8x16b = _mm_maddubs_epi16(src_r0r1_16x8b, coeff0_1_16x8b); //a0*c0+a1*c1 a1*c0+a2*c1 … in ih264_inter_pred_luma_horz_ssse3()
269 …res_r0r1_t2_8x16b = _mm_maddubs_epi16(src_r0r1_16x8b, coeff2_3_16x8b); //a2*c2+a3*c3 a3*c2+a4*c3 … in ih264_inter_pred_luma_horz_ssse3()
276 …res_r0r1_t3_8x16b = _mm_maddubs_epi16(src_r0r1_16x8b, coeff4_5_16x8b); //a4*c4+a5*c5 a5*c4+a6*c5 … in ih264_inter_pred_luma_horz_ssse3()
326 …res_r0_t1_8x16b = _mm_maddubs_epi16(src_r0_t1_16x8b, coeff0_1_16x8b); //a0*c0+a1*c1 a1*c0+a2*c1 a2… in ih264_inter_pred_luma_horz_ssse3()
328 …res_r1_t1_8x16b = _mm_maddubs_epi16(src_r1_t1_16x8b, coeff0_1_16x8b); //b0*c0+b1*c1 b1*c0+b2*c1 b2… in ih264_inter_pred_luma_horz_ssse3()
340 …res_r0_t2_8x16b = _mm_maddubs_epi16(src_r0_t1_16x8b, coeff2_3_16x8b); //a2*c2+a3*c3 a3*c2+a4*c3 a4… in ih264_inter_pred_luma_horz_ssse3()
342 …res_r1_t2_8x16b = _mm_maddubs_epi16(src_r1_t1_16x8b, coeff2_3_16x8b); //b2*c2+b3*c3 b3*c2+b4*c3 b2… in ih264_inter_pred_luma_horz_ssse3()
354 …res_r0_t3_8x16b = _mm_maddubs_epi16(src_r0_t1_16x8b, coeff4_5_16x8b); //a4*c4+a5*c5 a5*c4+a6*c5 a6… in ih264_inter_pred_luma_horz_ssse3()
356 …res_r1_t3_8x16b = _mm_maddubs_epi16(src_r1_t1_16x8b, coeff4_5_16x8b); //b4*c4+b5*c5 b5*c4+b6*c5 b6… in ih264_inter_pred_luma_horz_ssse3()
403 …res_r0_t1_8x16b = _mm_maddubs_epi16(src_r0_t1_16x8b, coeff0_1_16x8b); //a0*c0+a1*c1 a1*c0+a2*c1 a2… in ih264_inter_pred_luma_horz_ssse3()
[all …]
/external/libaom/libaom/aom_dsp/x86/
Daom_subpixel_8t_intrin_ssse3.c93 srcRegFilt32b1_1 = _mm_maddubs_epi16(srcRegFilt32b1_1, firstFilters); in aom_filter_block1d4_h4_ssse3()
164 resReglo = _mm_maddubs_epi16(srcReg2345_3456_lo, firstFilters); in aom_filter_block1d4_v4_ssse3()
165 resReghi = _mm_maddubs_epi16(srcReg2345_3456_hi, firstFilters); in aom_filter_block1d4_v4_ssse3()
233 srcRegFilt1 = _mm_maddubs_epi16(srcRegFilt1, firstFilters); in aom_filter_block1d4_h8_intrin_ssse3()
234 srcRegFilt2 = _mm_maddubs_epi16(srcRegFilt2, secondFilters); in aom_filter_block1d4_h8_intrin_ssse3()
298 srcRegFilt32b3 = _mm_maddubs_epi16(srcRegFilt32b3, secondFilters); in aom_filter_block1d8_h4_ssse3()
299 srcRegFilt32b2 = _mm_maddubs_epi16(srcRegFilt32b2, thirdFilters); in aom_filter_block1d8_h4_ssse3()
367 resReg23 = _mm_maddubs_epi16(srcReg23, secondFilters); in aom_filter_block1d8_v4_ssse3()
368 resReg34 = _mm_maddubs_epi16(srcReg34, secondFilters); in aom_filter_block1d8_v4_ssse3()
369 resReg45 = _mm_maddubs_epi16(srcReg45, thirdFilters); in aom_filter_block1d8_v4_ssse3()
[all …]
Daom_subpixel_8t_intrin_avx2.c141 _mm_maddubs_epi16(srcRegFilt1_1, _mm256_castsi256_si128(firstFilters)); in aom_filter_block1d4_h4_avx2()
238 _mm_maddubs_epi16(srcRegFilt1_1, _mm256_castsi256_si128(firstFilters)); in aom_filter_block1d4_h8_avx2()
245 _mm_maddubs_epi16(srcRegFilt2, _mm256_castsi256_si128(secondFilters)); in aom_filter_block1d4_h8_avx2()
337 _mm_maddubs_epi16(srcRegFilt2, _mm256_castsi256_si128(secondFilters)); in aom_filter_block1d8_h4_avx2()
339 _mm_maddubs_epi16(srcRegFilt3, _mm256_castsi256_si128(thirdFilters)); in aom_filter_block1d8_h4_avx2()
453 _mm_maddubs_epi16(srcRegFilt1_1, _mm256_castsi256_si128(firstFilters)); in aom_filter_block1d8_h8_avx2()
455 _mm_maddubs_epi16(srcRegFilt2, _mm256_castsi256_si128(forthFilters)); in aom_filter_block1d8_h8_avx2()
466 _mm_maddubs_epi16(srcRegFilt3, _mm256_castsi256_si128(secondFilters)); in aom_filter_block1d8_h8_avx2()
468 _mm_maddubs_epi16(srcRegFilt2, _mm256_castsi256_si128(thirdFilters)); in aom_filter_block1d8_h8_avx2()
729 _mm_maddubs_epi16(srcRegFilt1_1, _mm256_castsi256_si128(firstFilters)); in aom_filter_block1d16_h8_avx2()
[all …]
Dmasked_variance_intrin_ssse3.h42 const __m128i blendAL = _mm_maddubs_epi16(ssAL, aaAL); in comp_mask_pred_16_ssse3()
43 const __m128i blendAH = _mm_maddubs_epi16(ssAH, aaAH); in comp_mask_pred_16_ssse3()
77 const __m128i blendA = _mm_maddubs_epi16(ssA, aaA); in comp_mask_pred_8_ssse3()
78 const __m128i blendB = _mm_maddubs_epi16(ssB, aaB); in comp_mask_pred_8_ssse3()
Dblend_sse4.h64 const __m128i v_p0_w = _mm_maddubs_epi16(_mm_unpacklo_epi8(v_s0_b, v_s1_b), in blend_4_u8()
78 const __m128i v_p0_w = _mm_maddubs_epi16(_mm_unpacklo_epi8(v_s0_b, v_s1_b), in blend_8_u8()
92 const __m128i v_p0_w = _mm_maddubs_epi16(_mm_unpacklo_epi8(v_s0_b, v_s1_b), in blend_16_u8()
94 const __m128i v_p1_w = _mm_maddubs_epi16(_mm_unpackhi_epi8(v_s0_b, v_s1_b), in blend_16_u8()
Dmasked_sad4d_ssse3.c28 pred_l = _mm_maddubs_epi16(data_l, mask_l); \
33 pred_r = _mm_maddubs_epi16(data_r, mask_r); \
95 pred_l = _mm_maddubs_epi16(data_l, mask_l); \
100 pred_r = _mm_maddubs_epi16(data_r, mask_r); \
161 pred = _mm_maddubs_epi16(data, mask); \
Dvariance_impl_ssse3.c48 __m128i res = _mm_maddubs_epi16(source, filters); in aom_var_filter_block2d_bil_first_pass_ssse3()
74 __m128i res = _mm_maddubs_epi16(source_shuffle, filters); in aom_var_filter_block2d_bil_first_pass_ssse3()
Dblend_mask_sse4.h113 const __m128i m_acbd = _mm_maddubs_epi16(m_ac, one_b); in aom_lowbd_blend_a64_d16_mask_subw1_subh1_w4_sse4_1()
138 const __m128i m_acbd = _mm_maddubs_epi16(m_ac, one_b); in aom_lowbd_blend_a64_d16_mask_subw1_subh1_w8_sse4_1()
161 const __m128i m_ac = _mm_maddubs_epi16(m_i0, one_b); in aom_lowbd_blend_a64_d16_mask_subw1_subh0_w4_sse4_1()
183 const __m128i m_ac = _mm_maddubs_epi16(m_i0, one_b); in aom_lowbd_blend_a64_d16_mask_subw1_subh0_w8_sse4_1()
Dmasked_sad_intrin_ssse3.c117 __m128i pred_l = _mm_maddubs_epi16(data_l, mask_l); in masked_sad_ssse3()
122 __m128i pred_r = _mm_maddubs_epi16(data_r, mask_r); in masked_sad_ssse3()
164 __m128i pred_l = _mm_maddubs_epi16(data_l, mask_l); in aom_masked_sad8xh_ssse3()
169 __m128i pred_r = _mm_maddubs_epi16(data_r, mask_r); in aom_masked_sad8xh_ssse3()
213 __m128i pred_16bit = _mm_maddubs_epi16(data, mask); in aom_masked_sad4xh_ssse3()
/external/libavc/encoder/x86/
Dih264e_half_pel_ssse3.c141 …res_r0_t1_8x16b = _mm_maddubs_epi16(src_r0_t1_16x8b, coeff0_1_16x8b); //a0*c0+a1*c1 a1*c0+a2*c1… in ih264e_sixtapfilter_horz_ssse3()
143 …res_r1_t1_8x16b = _mm_maddubs_epi16(src_r1_t1_16x8b, coeff0_1_16x8b); //b0*c0+b1*c1 b1*c0+b2*c1… in ih264e_sixtapfilter_horz_ssse3()
155 …res_r0_t2_8x16b = _mm_maddubs_epi16(src_r0_t1_16x8b, coeff2_3_16x8b); //a2*c2+a3*c3 a3*c2+a4*c3… in ih264e_sixtapfilter_horz_ssse3()
157 …res_r1_t2_8x16b = _mm_maddubs_epi16(src_r1_t1_16x8b, coeff2_3_16x8b); //b2*c2+b3*c3 b3*c2+b4*c3… in ih264e_sixtapfilter_horz_ssse3()
169 …res_r0_t3_8x16b = _mm_maddubs_epi16(src_r0_t1_16x8b, coeff4_5_16x8b); //a4*c4+a5*c5 a5*c4+a6*c5… in ih264e_sixtapfilter_horz_ssse3()
171 …res_r1_t3_8x16b = _mm_maddubs_epi16(src_r1_t1_16x8b, coeff4_5_16x8b); //b4*c4+b5*c5 b5*c4+b6*c5… in ih264e_sixtapfilter_horz_ssse3()
312 res_t1_8x16b = _mm_maddubs_epi16(src_r0r1_16x8b, coeff0_1_16x8b); in ih264e_sixtap_filter_2dvh_vert_ssse3()
313 res_t2_8x16b = _mm_maddubs_epi16(src_r2r3_16x8b, coeff2_3_16x8b); in ih264e_sixtap_filter_2dvh_vert_ssse3()
314 res_t3_8x16b = _mm_maddubs_epi16(src_r4r5_16x8b, coeff4_5_16x8b); in ih264e_sixtap_filter_2dvh_vert_ssse3()
325 res_t1_8x16b = _mm_maddubs_epi16(src_r0r1_16x8b, coeff0_1_16x8b); in ih264e_sixtap_filter_2dvh_vert_ssse3()
[all …]
/external/libaom/libaom/av1/common/x86/
Dcfl_ssse3.c51 top = _mm_maddubs_epi16(top, twos); in cfl_luma_subsampling_420_lbd_ssse3()
53 bot = _mm_maddubs_epi16(bot, twos); in cfl_luma_subsampling_420_lbd_ssse3()
58 top = _mm_maddubs_epi16(top, twos); in cfl_luma_subsampling_420_lbd_ssse3()
60 bot = _mm_maddubs_epi16(bot, twos); in cfl_luma_subsampling_420_lbd_ssse3()
65 top = _mm_maddubs_epi16(top, twos); in cfl_luma_subsampling_420_lbd_ssse3()
67 bot = _mm_maddubs_epi16(bot, twos); in cfl_luma_subsampling_420_lbd_ssse3()
74 top_1 = _mm_maddubs_epi16(top_1, twos); in cfl_luma_subsampling_420_lbd_ssse3()
75 bot_1 = _mm_maddubs_epi16(bot_1, twos); in cfl_luma_subsampling_420_lbd_ssse3()
105 top = _mm_maddubs_epi16(top, fours); in cfl_luma_subsampling_422_lbd_ssse3()
109 top = _mm_maddubs_epi16(top, fours); in cfl_luma_subsampling_422_lbd_ssse3()
[all …]
Dfilterintra_sse4.c53 const __m128i out_01 = _mm_maddubs_epi16(in, f1f0); in av1_filter_intra_predictor_sse4_1()
54 const __m128i out_23 = _mm_maddubs_epi16(in, f3f2); in av1_filter_intra_predictor_sse4_1()
55 const __m128i out_45 = _mm_maddubs_epi16(in, f5f4); in av1_filter_intra_predictor_sse4_1()
56 const __m128i out_67 = _mm_maddubs_epi16(in, f7f6); in av1_filter_intra_predictor_sse4_1()
Dintra_edge_sse4.c58 d0 = _mm_maddubs_epi16(d0, coef0); in av1_filter_intra_edge_sse4_1()
59 d1 = _mm_maddubs_epi16(d1, coef0); in av1_filter_intra_edge_sse4_1()
91 d0 = _mm_maddubs_epi16(d0, coef0); in av1_filter_intra_edge_sse4_1()
92 d1 = _mm_maddubs_epi16(d1, coef0); in av1_filter_intra_edge_sse4_1()
93 d2 = _mm_maddubs_epi16(d2, coef0); in av1_filter_intra_edge_sse4_1()
94 d3 = _mm_maddubs_epi16(d3, coef0); in av1_filter_intra_edge_sse4_1()
243 d0 = _mm_maddubs_epi16(d0, coef0); in av1_upsample_intra_edge_sse4_1()
244 d1 = _mm_maddubs_epi16(d1, coef0); in av1_upsample_intra_edge_sse4_1()
245 d2 = _mm_maddubs_epi16(d2, coef0); in av1_upsample_intra_edge_sse4_1()
246 d3 = _mm_maddubs_epi16(d3, coef0); in av1_upsample_intra_edge_sse4_1()
/external/libgav1/libgav1/src/dsp/x86/
Dobmc_sse4.cc58 RightShiftWithRounding_U16(_mm_maddubs_epi16(terms, masks), 6); in OverlapBlendFromLeft2xH_SSE4_1()
97 RightShiftWithRounding_U16(_mm_maddubs_epi16(terms, masks), 6); in OverlapBlendFromLeft4xH_SSE4_1()
125 RightShiftWithRounding_U16(_mm_maddubs_epi16(terms, masks), 6); in OverlapBlendFromLeft8xH_SSE4_1()
174 RightShiftWithRounding_U16(_mm_maddubs_epi16(terms_lo, masks_lo), 6); in OverlapBlendFromLeft_SSE4_1()
177 RightShiftWithRounding_U16(_mm_maddubs_epi16(terms_hi, masks_hi), 6); in OverlapBlendFromLeft_SSE4_1()
218 RightShiftWithRounding_U16(_mm_maddubs_epi16(terms, masks), 6); in OverlapBlendFromTop4xH_SSE4_1()
248 RightShiftWithRounding_U16(_mm_maddubs_epi16(terms, masks), 6); in OverlapBlendFromTop8xH_SSE4_1()
291 RightShiftWithRounding_U16(_mm_maddubs_epi16(terms_lo, masks), 6); in OverlapBlendFromTop_SSE4_1()
294 RightShiftWithRounding_U16(_mm_maddubs_epi16(terms_hi, masks), 6); in OverlapBlendFromTop_SSE4_1()
Dconvolve_sse4.inc29 const __m128i v_madd_21 = _mm_maddubs_epi16(src[0], taps[0]); // k2k1
30 const __m128i v_madd_43 = _mm_maddubs_epi16(src[1], taps[1]); // k4k3
31 const __m128i v_madd_65 = _mm_maddubs_epi16(src[2], taps[2]); // k6k5
36 const __m128i v_madd_10 = _mm_maddubs_epi16(src[0], taps[0]); // k1k0
37 const __m128i v_madd_32 = _mm_maddubs_epi16(src[1], taps[1]); // k3k2
38 const __m128i v_madd_54 = _mm_maddubs_epi16(src[2], taps[2]); // k5k4
39 const __m128i v_madd_76 = _mm_maddubs_epi16(src[3], taps[3]); // k7k6
45 sum = _mm_maddubs_epi16(src[0], taps[0]); // k4k3
48 const __m128i v_madd_32 = _mm_maddubs_epi16(src[0], taps[0]); // k3k2
49 const __m128i v_madd_54 = _mm_maddubs_epi16(src[1], taps[1]); // k5k4
[all …]
Dintrapred_filter_sse4.cc71 const __m128i mul_0_01 = _mm_maddubs_epi16(pixels, taps_0_1); in Filter4x2_SSE4_1()
72 const __m128i mul_0_23 = _mm_maddubs_epi16(pixels, taps_2_3); in Filter4x2_SSE4_1()
80 const __m128i mul_1_01 = _mm_maddubs_epi16(pixels, taps_4_5); in Filter4x2_SSE4_1()
81 const __m128i mul_1_23 = _mm_maddubs_epi16(pixels, taps_6_7); in Filter4x2_SSE4_1()
/external/rust/crates/libz-sys/src/zlib-ng/arch/x86/
Dadler32_ssse3.c77 … __m128i v_short_sum1 = _mm_maddubs_epi16(vbuf, dot1v); // multiply-add, resulting in 8 shorts. in adler32_ssse3()
79 __m128i v_short_sum2 = _mm_maddubs_epi16(vbuf, dot2v); in adler32_ssse3()

123