Home
last modified time | relevance | path

Searched refs:_mm_adds_epi16 (Results 1 – 25 of 50) sorted by relevance

12

/external/libaom/libaom/av1/encoder/x86/
Dav1_fwd_txfm_sse2.c66 x1[0] = _mm_adds_epi16(input[0], input[3]); in fdct8x4_new_sse2()
68 x1[1] = _mm_adds_epi16(input[1], input[2]); in fdct8x4_new_sse2()
100 x1[0] = _mm_adds_epi16(input[0], input[7]); in fdct4x8_new_sse2()
102 x1[1] = _mm_adds_epi16(input[1], input[6]); in fdct4x8_new_sse2()
104 x1[2] = _mm_adds_epi16(input[2], input[5]); in fdct4x8_new_sse2()
106 x1[3] = _mm_adds_epi16(input[3], input[4]); in fdct4x8_new_sse2()
111 x2[0] = _mm_adds_epi16(x1[0], x1[3]); in fdct4x8_new_sse2()
113 x2[1] = _mm_adds_epi16(x1[1], x1[2]); in fdct4x8_new_sse2()
126 x3[4] = _mm_adds_epi16(x2[4], x2[5]); in fdct4x8_new_sse2()
129 x3[7] = _mm_adds_epi16(x2[7], x2[6]); in fdct4x8_new_sse2()
[all …]
Dav1_fwd_txfm_sse2.h62 output[0] = _mm_adds_epi16(input[0], input[0]); in fidentity8x8_new_sse2()
63 output[1] = _mm_adds_epi16(input[1], input[1]); in fidentity8x8_new_sse2()
64 output[2] = _mm_adds_epi16(input[2], input[2]); in fidentity8x8_new_sse2()
65 output[3] = _mm_adds_epi16(input[3], input[3]); in fidentity8x8_new_sse2()
66 output[4] = _mm_adds_epi16(input[4], input[4]); in fidentity8x8_new_sse2()
67 output[5] = _mm_adds_epi16(input[5], input[5]); in fidentity8x8_new_sse2()
68 output[6] = _mm_adds_epi16(input[6], input[6]); in fidentity8x8_new_sse2()
69 output[7] = _mm_adds_epi16(input[7], input[7]); in fidentity8x8_new_sse2()
Dav1_quantize_sse2.c97 qcoeff0 = _mm_adds_epi16(qcoeff0, *round0); in quantize()
98 qcoeff1 = _mm_adds_epi16(qcoeff1, *round1); in quantize()
/external/libavc/common/x86/
Dih264_weighted_pred_sse42.c422 y_0_8x16b = _mm_adds_epi16(round_8x16b, y_0_8x16b); in ih264_weighted_pred_luma_sse42()
423 y_2_8x16b = _mm_adds_epi16(round_8x16b, y_2_8x16b); in ih264_weighted_pred_luma_sse42()
428 y_0_8x16b = _mm_adds_epi16(ofst_8x16b, y_0_8x16b); in ih264_weighted_pred_luma_sse42()
429 y_2_8x16b = _mm_adds_epi16(ofst_8x16b, y_2_8x16b); in ih264_weighted_pred_luma_sse42()
468 y_0_8x16b = _mm_adds_epi16(round_8x16b, y_0_8x16b); in ih264_weighted_pred_luma_sse42()
469 y_1_8x16b = _mm_adds_epi16(round_8x16b, y_1_8x16b); in ih264_weighted_pred_luma_sse42()
470 y_2_8x16b = _mm_adds_epi16(round_8x16b, y_2_8x16b); in ih264_weighted_pred_luma_sse42()
471 y_3_8x16b = _mm_adds_epi16(round_8x16b, y_3_8x16b); in ih264_weighted_pred_luma_sse42()
478 y_0_8x16b = _mm_adds_epi16(ofst_8x16b, y_0_8x16b); in ih264_weighted_pred_luma_sse42()
479 y_1_8x16b = _mm_adds_epi16(ofst_8x16b, y_1_8x16b); in ih264_weighted_pred_luma_sse42()
[all …]
/external/libvpx/libvpx/vp9/common/x86/
Dvp9_idct_intrin_sse2.c88 in[0] = _mm_adds_epi16(in[0], final_rounding); in vp9_iht8x8_64_add_sse2()
89 in[1] = _mm_adds_epi16(in[1], final_rounding); in vp9_iht8x8_64_add_sse2()
90 in[2] = _mm_adds_epi16(in[2], final_rounding); in vp9_iht8x8_64_add_sse2()
91 in[3] = _mm_adds_epi16(in[3], final_rounding); in vp9_iht8x8_64_add_sse2()
92 in[4] = _mm_adds_epi16(in[4], final_rounding); in vp9_iht8x8_64_add_sse2()
93 in[5] = _mm_adds_epi16(in[5], final_rounding); in vp9_iht8x8_64_add_sse2()
94 in[6] = _mm_adds_epi16(in[6], final_rounding); in vp9_iht8x8_64_add_sse2()
95 in[7] = _mm_adds_epi16(in[7], final_rounding); in vp9_iht8x8_64_add_sse2()
141 in[0] = _mm_adds_epi16(in[0], final_rounding); in write_buffer_8x16()
142 in[1] = _mm_adds_epi16(in[1], final_rounding); in write_buffer_8x16()
[all …]
/external/libvpx/libvpx/vpx_dsp/x86/
Dconvolve_ssse3.h60 sum1 = _mm_adds_epi16(sum1, sum2); in convolve8_8_ssse3()
77 __m128i temp = _mm_adds_epi16(x0, x3); in convolve8_8_even_offset_ssse3()
78 temp = _mm_adds_epi16(temp, x1); in convolve8_8_even_offset_ssse3()
79 temp = _mm_adds_epi16(temp, x2); in convolve8_8_even_offset_ssse3()
80 temp = _mm_adds_epi16(temp, x4); in convolve8_8_even_offset_ssse3()
82 temp = _mm_adds_epi16(temp, k_64); in convolve8_8_even_offset_ssse3()
101 temp = _mm_adds_epi16(x0, x1); in convolve8_8_odd_offset_ssse3()
102 temp = _mm_adds_epi16(temp, x2); in convolve8_8_odd_offset_ssse3()
103 temp = _mm_adds_epi16(temp, x3); in convolve8_8_odd_offset_ssse3()
104 temp = _mm_adds_epi16(temp, x4); in convolve8_8_odd_offset_ssse3()
[all …]
Dhighbd_loopfilter_sse2.c159 filt = _mm_adds_epi16(filt, work_a); in vpx_highbd_lpf_horizontal_16_sse2()
160 filt = _mm_adds_epi16(filt, work_a); in vpx_highbd_lpf_horizontal_16_sse2()
161 filt = signed_char_clamp_bd_sse2(_mm_adds_epi16(filt, work_a), bd); in vpx_highbd_lpf_horizontal_16_sse2()
163 filter1 = signed_char_clamp_bd_sse2(_mm_adds_epi16(filt, t4), bd); in vpx_highbd_lpf_horizontal_16_sse2()
164 filter2 = signed_char_clamp_bd_sse2(_mm_adds_epi16(filt, t3), bd); in vpx_highbd_lpf_horizontal_16_sse2()
170 qs0 = _mm_adds_epi16( in vpx_highbd_lpf_horizontal_16_sse2()
172 ps0 = _mm_adds_epi16( in vpx_highbd_lpf_horizontal_16_sse2()
173 signed_char_clamp_bd_sse2(_mm_adds_epi16(ps0, filter2), bd), t80); in vpx_highbd_lpf_horizontal_16_sse2()
174 filt = _mm_adds_epi16(filter1, t1); in vpx_highbd_lpf_horizontal_16_sse2()
177 qs1 = _mm_adds_epi16(signed_char_clamp_bd_sse2(_mm_subs_epi16(qs1, filt), bd), in vpx_highbd_lpf_horizontal_16_sse2()
[all …]
Dvpx_subpixel_8t_intrin_ssse3.c100 srcRegFilt1 = _mm_adds_epi16(srcRegFilt1, srcRegFilt2); in vpx_filter_block1d4_h8_intrin_ssse3()
251 dst_first = _mm_adds_epi16(tmp_0, tmp_1); in vpx_filter_block1d16_h4_ssse3()
262 dst_second = _mm_adds_epi16(tmp_0, tmp_1); in vpx_filter_block1d16_h4_ssse3()
354 res_reg_m1012_lo = _mm_adds_epi16(res_reg_m10_lo, res_reg_12_lo); in vpx_filter_block1d16_v4_ssse3()
355 res_reg_0123_lo = _mm_adds_epi16(res_reg_01_lo, res_reg_23_lo); in vpx_filter_block1d16_v4_ssse3()
365 res_reg_m1012_hi = _mm_adds_epi16(res_reg_m10_hi, res_reg_12_hi); in vpx_filter_block1d16_v4_ssse3()
366 res_reg_0123_hi = _mm_adds_epi16(res_reg_01_hi, res_reg_23_hi); in vpx_filter_block1d16_v4_ssse3()
440 dst_first = _mm_adds_epi16(tmp_0, tmp_1); in vpx_filter_block1d8_h4_ssse3()
523 res_reg_m1012 = _mm_adds_epi16(res_reg_m10, res_reg_12); in vpx_filter_block1d8_v4_ssse3()
524 res_reg_0123 = _mm_adds_epi16(res_reg_01, res_reg_23); in vpx_filter_block1d8_v4_ssse3()
Dvpx_subpixel_4t_intrin_sse2.c206 res_reg_m1012_lo = _mm_adds_epi16(res_reg_m10_lo, res_reg_12_lo); in vpx_filter_block1d16_v4_sse2()
207 res_reg_0123_lo = _mm_adds_epi16(res_reg_01_lo, res_reg_23_lo); in vpx_filter_block1d16_v4_sse2()
228 res_reg_m1012_hi = _mm_adds_epi16(res_reg_m10_hi, res_reg_12_hi); in vpx_filter_block1d16_v4_sse2()
229 res_reg_0123_hi = _mm_adds_epi16(res_reg_01_hi, res_reg_23_hi); in vpx_filter_block1d16_v4_sse2()
410 res_reg_m1012_lo = _mm_adds_epi16(res_reg_m10_lo, res_reg_12_lo); in vpx_filter_block1d8_v4_sse2()
411 res_reg_0123_lo = _mm_adds_epi16(res_reg_01_lo, res_reg_23_lo); in vpx_filter_block1d8_v4_sse2()
592 res_reg_m1012_lo = _mm_adds_epi16(res_reg_m10_lo, res_reg_12_lo); in vpx_filter_block1d4_v4_sse2()
593 res_reg_0123_lo = _mm_adds_epi16(res_reg_01_lo, res_reg_23_lo); in vpx_filter_block1d4_v4_sse2()
Dconvolve_sse2.h84 const __m128i nearest_src = _mm_adds_epi16(*src, *half_depth); in mm_round_epi16_sse2()
Dquantize_sse2.h41 qcoeff = _mm_adds_epi16(*coeff, round); in calculate_qcoeff()
/external/libaom/libaom/aom_dsp/x86/
Daom_subpixel_8t_intrin_ssse3.c98 srcRegFilt32b1_1 = _mm_adds_epi16(srcRegFilt32b1_1, addFilterReg32); in aom_filter_block1d4_h4_ssse3()
171 resReglo = _mm_adds_epi16(resReglo, addFilterReg32); in aom_filter_block1d4_v4_ssse3()
172 resReghi = _mm_adds_epi16(resReghi, addFilterReg32); in aom_filter_block1d4_v4_ssse3()
243 srcRegFilt1 = _mm_adds_epi16(srcRegFilt1, srcRegFilt4); in aom_filter_block1d4_h8_intrin_ssse3()
245 srcRegFilt1 = _mm_adds_epi16(srcRegFilt1, minReg); in aom_filter_block1d4_h8_intrin_ssse3()
246 srcRegFilt1 = _mm_adds_epi16(srcRegFilt1, srcRegFilt3); in aom_filter_block1d4_h8_intrin_ssse3()
247 srcRegFilt1 = _mm_adds_epi16(srcRegFilt1, addFilterReg64); in aom_filter_block1d4_h8_intrin_ssse3()
301 srcRegFilt32b1_1 = _mm_adds_epi16(srcRegFilt32b3, srcRegFilt32b2); in aom_filter_block1d8_h4_ssse3()
304 srcRegFilt32b1_1 = _mm_adds_epi16(srcRegFilt32b1_1, addFilterReg32); in aom_filter_block1d8_h4_ssse3()
373 resReg23_45 = _mm_adds_epi16(resReg23, resReg45); in aom_filter_block1d8_v4_ssse3()
[all …]
Daom_subpixel_8t_intrin_sse2.c91 srcRegFilt32b1_1 = _mm_adds_epi16(srcRegFilt32b1_1, addFilterReg32); in aom_filter_block1d16_h4_sse2()
92 srcRegFilt32b2_1 = _mm_adds_epi16(srcRegFilt32b2_1, addFilterReg32); in aom_filter_block1d16_h4_sse2()
192 resReg23_45_lo = _mm_adds_epi16(resReg23_lo, resReg45_lo); in aom_filter_block1d16_v4_sse2()
193 resReg34_56_lo = _mm_adds_epi16(resReg34_lo, resReg56_lo); in aom_filter_block1d16_v4_sse2()
218 resReg23_45_hi = _mm_adds_epi16(resReg23_hi, resReg45_hi); in aom_filter_block1d16_v4_sse2()
219 resReg34_56_hi = _mm_adds_epi16(resReg34_hi, resReg56_hi); in aom_filter_block1d16_v4_sse2()
222 resReg23_45_lo = _mm_adds_epi16(resReg23_45_lo, addFilterReg32); in aom_filter_block1d16_v4_sse2()
223 resReg34_56_lo = _mm_adds_epi16(resReg34_56_lo, addFilterReg32); in aom_filter_block1d16_v4_sse2()
224 resReg23_45_hi = _mm_adds_epi16(resReg23_45_hi, addFilterReg32); in aom_filter_block1d16_v4_sse2()
225 resReg34_56_hi = _mm_adds_epi16(resReg34_56_hi, addFilterReg32); in aom_filter_block1d16_v4_sse2()
[all …]
Daom_subpixel_8t_intrin_avx2.c146 _mm_adds_epi16(srcRegFilt1_1, _mm256_castsi256_si128(addFilterReg32)); in aom_filter_block1d4_h4_avx2()
247 srcRegFilt1_1 = _mm_adds_epi16(srcRegFilt1_1, srcRegFilt2); in aom_filter_block1d4_h8_avx2()
251 _mm_adds_epi16(srcRegFilt1_1, _mm256_castsi256_si128(addFilterReg32)); in aom_filter_block1d4_h8_avx2()
342 srcRegFilt1_1 = _mm_adds_epi16(srcRegFilt2, srcRegFilt3); in aom_filter_block1d8_h4_avx2()
346 _mm_adds_epi16(srcRegFilt1_1, _mm256_castsi256_si128(addFilterReg32)); in aom_filter_block1d8_h4_avx2()
458 srcRegFilt1_1 = _mm_adds_epi16(srcRegFilt1_1, srcRegFilt2); in aom_filter_block1d8_h8_avx2()
472 _mm_adds_epi16(srcRegFilt1_1, _mm_adds_epi16(srcRegFilt3, srcRegFilt2)); in aom_filter_block1d8_h8_avx2()
476 _mm_adds_epi16(srcRegFilt1_1, _mm256_castsi256_si128(addFilterReg32)); in aom_filter_block1d8_h8_avx2()
734 srcRegFilt1_1 = _mm_adds_epi16(srcRegFilt1_1, srcRegFilt2); in aom_filter_block1d16_h8_avx2()
748 _mm_adds_epi16(srcRegFilt1_1, _mm_adds_epi16(srcRegFilt3, srcRegFilt2)); in aom_filter_block1d16_h8_avx2()
[all …]
Dhighbd_loopfilter_sse2.c243 filter2filter1 = _mm_adds_epi16(filt, t3t4); /* signed_short_clamp */ in highbd_filter4_sse2()
250 filt = _mm_adds_epi16(filt, one); in highbd_filter4_sse2()
258 ps1ps0_work = _mm_adds_epi16(ps1ps0_work, filter2filt); in highbd_filter4_sse2()
263 *qs1qs0 = _mm_adds_epi16(qs1qs0_work, *t80); in highbd_filter4_sse2()
264 *ps1ps0 = _mm_adds_epi16(ps1ps0_work, *t80); in highbd_filter4_sse2()
296 filter = _mm_adds_epi16(filter, x); in highbd_filter4_dual_sse2()
297 filter = _mm_adds_epi16(filter, x); in highbd_filter4_dual_sse2()
298 filter = _mm_adds_epi16(filter, x); in highbd_filter4_dual_sse2()
303 __m128i filter1 = _mm_adds_epi16(filter, t4); in highbd_filter4_dual_sse2()
304 __m128i filter2 = _mm_adds_epi16(filter, t3); in highbd_filter4_dual_sse2()
[all …]
Dquantize_x86.h43 qcoeff = _mm_adds_epi16(*coeff, round); in calculate_qcoeff()
55 qcoeff = _mm_adds_epi16(*coeff, round); in calculate_qcoeff_log_scale()
/external/libhevc/common/x86/
Dihevc_weighted_pred_ssse3_intr.c1324 lvl_shift1_8x16b = _mm_adds_epi16(lvl_shift1_8x16b, lvl_shift2_8x16b); in ihevc_weighted_pred_bi_default_ssse3()
1325 lvl_shift1_8x16b = _mm_adds_epi16(lvl_shift1_8x16b, const_temp_8x16b); in ihevc_weighted_pred_bi_default_ssse3()
1362 src_temp1_8x16b = _mm_adds_epi16(src_temp1_8x16b, src_temp2_8x16b); in ihevc_weighted_pred_bi_default_ssse3()
1363 src_temp3_8x16b = _mm_adds_epi16(src_temp3_8x16b, src_temp4_8x16b); in ihevc_weighted_pred_bi_default_ssse3()
1364 src_temp5_8x16b = _mm_adds_epi16(src_temp5_8x16b, src_temp6_8x16b); in ihevc_weighted_pred_bi_default_ssse3()
1365 src_temp7_8x16b = _mm_adds_epi16(src_temp7_8x16b, src_temp8_8x16b); in ihevc_weighted_pred_bi_default_ssse3()
1373 src_temp1_8x16b = _mm_adds_epi16(src_temp1_8x16b, lvl_shift1_8x16b); in ihevc_weighted_pred_bi_default_ssse3()
1374 src_temp3_8x16b = _mm_adds_epi16(src_temp3_8x16b, lvl_shift1_8x16b); in ihevc_weighted_pred_bi_default_ssse3()
1375 src_temp5_8x16b = _mm_adds_epi16(src_temp5_8x16b, lvl_shift1_8x16b); in ihevc_weighted_pred_bi_default_ssse3()
1376 src_temp7_8x16b = _mm_adds_epi16(src_temp7_8x16b, lvl_shift1_8x16b); in ihevc_weighted_pred_bi_default_ssse3()
[all …]
Dihevc_weighted_pred_sse42_intr.c1810 lvl_shift1_8x16b = _mm_adds_epi16(lvl_shift1_8x16b, lvl_shift2_8x16b); in ihevc_weighted_pred_bi_default_sse42()
1811 lvl_shift1_8x16b = _mm_adds_epi16(lvl_shift1_8x16b, const_temp_8x16b); in ihevc_weighted_pred_bi_default_sse42()
1848 src_temp1_8x16b = _mm_adds_epi16(src_temp1_8x16b, src_temp2_8x16b); in ihevc_weighted_pred_bi_default_sse42()
1849 src_temp3_8x16b = _mm_adds_epi16(src_temp3_8x16b, src_temp4_8x16b); in ihevc_weighted_pred_bi_default_sse42()
1850 src_temp5_8x16b = _mm_adds_epi16(src_temp5_8x16b, src_temp6_8x16b); in ihevc_weighted_pred_bi_default_sse42()
1851 src_temp7_8x16b = _mm_adds_epi16(src_temp7_8x16b, src_temp8_8x16b); in ihevc_weighted_pred_bi_default_sse42()
1859 src_temp1_8x16b = _mm_adds_epi16(src_temp1_8x16b, lvl_shift1_8x16b); in ihevc_weighted_pred_bi_default_sse42()
1860 src_temp3_8x16b = _mm_adds_epi16(src_temp3_8x16b, lvl_shift1_8x16b); in ihevc_weighted_pred_bi_default_sse42()
1861 src_temp5_8x16b = _mm_adds_epi16(src_temp5_8x16b, lvl_shift1_8x16b); in ihevc_weighted_pred_bi_default_sse42()
1862 src_temp7_8x16b = _mm_adds_epi16(src_temp7_8x16b, lvl_shift1_8x16b); in ihevc_weighted_pred_bi_default_sse42()
[all …]
/external/fec/
Dviterbi39_sse2.c147 m0 = _mm_adds_epi16(vp->old_metrics->v[i],metric); in update_viterbi39_blk_sse2()
148 m3 = _mm_adds_epi16(vp->old_metrics->v[16+i],metric); in update_viterbi39_blk_sse2()
149 m1 = _mm_adds_epi16(vp->old_metrics->v[16+i],m_metric); in update_viterbi39_blk_sse2()
150 m2 = _mm_adds_epi16(vp->old_metrics->v[i],m_metric); in update_viterbi39_blk_sse2()
Dviterbi615_sse2.c150 m0 = _mm_adds_epi16(vp->old_metrics->v[i],metric); in update_viterbi615_blk_sse2()
151 m3 = _mm_adds_epi16(vp->old_metrics->v[1024+i],metric); in update_viterbi615_blk_sse2()
152 m1 = _mm_adds_epi16(vp->old_metrics->v[1024+i],m_metric); in update_viterbi615_blk_sse2()
153 m2 = _mm_adds_epi16(vp->old_metrics->v[i],m_metric); in update_viterbi615_blk_sse2()
/external/libvpx/libvpx/vp9/encoder/x86/
Dvp9_quantize_sse2.c69 qcoeff0 = _mm_adds_epi16(qcoeff0, round); in vp9_quantize_fp_sse2()
71 qcoeff1 = _mm_adds_epi16(qcoeff1, round); in vp9_quantize_fp_sse2()
140 qcoeff0 = _mm_adds_epi16(qcoeff0, round); in vp9_quantize_fp_sse2()
141 qcoeff1 = _mm_adds_epi16(qcoeff1, round); in vp9_quantize_fp_sse2()
Dvp9_dct_ssse3.c321 qcoeff0 = _mm_adds_epi16(qcoeff0, round); in vp9_fdct8x8_quant_ssse3()
323 qcoeff1 = _mm_adds_epi16(qcoeff1, round); in vp9_fdct8x8_quant_ssse3()
393 qcoeff0 = _mm_adds_epi16(qcoeff0, round); in vp9_fdct8x8_quant_ssse3()
394 qcoeff1 = _mm_adds_epi16(qcoeff1, round); in vp9_fdct8x8_quant_ssse3()
/external/libaom/libaom/av1/common/x86/
Dav1_inv_txfm_ssse3.h41 in0 = _mm_adds_epi16(_in0, _in1); \
50 in0 = _mm_adds_epi16(_in0, _in1); \
57 out0 = _mm_adds_epi16(_in0, _in1); \
/external/mesa3d/src/intel/compiler/
Dbrw_fs_bank_conflicts.cpp121 _mm_adds_epi16(v.v[0], w.v[0]), in adds()
122 _mm_adds_epi16(v.v[1], w.v[1]) in adds()
159 const __m128i v8 = _mm_adds_epi16(v.v[0], v.v[1]); in sums()
160 const __m128i v4 = _mm_adds_epi16(v8, _mm_shuffle_epi32(v8, 0x4e)); in sums()
161 const __m128i v2 = _mm_adds_epi16(v4, _mm_shuffle_epi32(v4, 0xb1)); in sums()
162 const __m128i v1 = _mm_adds_epi16(v2, _mm_shufflelo_epi16(v2, 0xb1)); in sums()
/external/webp/src/dsp/
Ddec_sse2.c209 const __m128i B = _mm_adds_epi16(A, CD); in TransformAC3()
210 const __m128i m0 = _mm_adds_epi16(B, d4); in TransformAC3()
211 const __m128i m1 = _mm_adds_epi16(B, c4); in TransformAC3()
226 dst0 = _mm_adds_epi16(dst0, _mm_srai_epi16(m0, 3)); in TransformAC3()
227 dst1 = _mm_adds_epi16(dst1, _mm_srai_epi16(m1, 3)); in TransformAC3()
228 dst2 = _mm_adds_epi16(dst2, _mm_srai_epi16(m2, 3)); in TransformAC3()
229 dst3 = _mm_adds_epi16(dst3, _mm_srai_epi16(m3, 3)); in TransformAC3()

12