Line data Source code
1 : /*
2 : * Copyright (c) 2016, Alliance for Open Media. All rights reserved
3 : *
4 : * This source code is subject to the terms of the BSD 2 Clause License and
5 : * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
6 : * was not distributed with this source code in the LICENSE file, you can
7 : * obtain it at www.aomedia.org/license/software. If the Alliance for Open
8 : * Media Patent License 1.0 was not distributed with this source code in the
9 : * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
10 : */
11 :
12 : #include <immintrin.h> /* AVX2 */
13 :
14 : #include "./aom_dsp_rtcd.h"
15 : #include "aom_ports/mem.h"
16 :
17 0 : void aom_lpf_horizontal_edge_8_avx2(unsigned char *s, int p,
18 : const unsigned char *_blimit,
19 : const unsigned char *_limit,
20 : const unsigned char *_thresh) {
21 : __m128i mask, hev, flat, flat2;
22 0 : const __m128i zero = _mm_set1_epi16(0);
23 0 : const __m128i one = _mm_set1_epi8(1);
24 : __m128i q7p7, q6p6, q5p5, q4p4, q3p3, q2p2, q1p1, q0p0, p0q0, p1q1;
25 : __m128i abs_p1p0;
26 :
27 0 : const __m128i thresh =
28 0 : _mm_broadcastb_epi8(_mm_cvtsi32_si128((int)_thresh[0]));
29 0 : const __m128i limit = _mm_broadcastb_epi8(_mm_cvtsi32_si128((int)_limit[0]));
30 0 : const __m128i blimit =
31 0 : _mm_broadcastb_epi8(_mm_cvtsi32_si128((int)_blimit[0]));
32 :
33 0 : q4p4 = _mm_loadl_epi64((__m128i *)(s - 5 * p));
34 0 : q4p4 = _mm_castps_si128(
35 0 : _mm_loadh_pi(_mm_castsi128_ps(q4p4), (__m64 *)(s + 4 * p)));
36 0 : q3p3 = _mm_loadl_epi64((__m128i *)(s - 4 * p));
37 0 : q3p3 = _mm_castps_si128(
38 0 : _mm_loadh_pi(_mm_castsi128_ps(q3p3), (__m64 *)(s + 3 * p)));
39 0 : q2p2 = _mm_loadl_epi64((__m128i *)(s - 3 * p));
40 0 : q2p2 = _mm_castps_si128(
41 0 : _mm_loadh_pi(_mm_castsi128_ps(q2p2), (__m64 *)(s + 2 * p)));
42 0 : q1p1 = _mm_loadl_epi64((__m128i *)(s - 2 * p));
43 0 : q1p1 = _mm_castps_si128(
44 0 : _mm_loadh_pi(_mm_castsi128_ps(q1p1), (__m64 *)(s + 1 * p)));
45 0 : p1q1 = _mm_shuffle_epi32(q1p1, 78);
46 0 : q0p0 = _mm_loadl_epi64((__m128i *)(s - 1 * p));
47 0 : q0p0 = _mm_castps_si128(
48 : _mm_loadh_pi(_mm_castsi128_ps(q0p0), (__m64 *)(s - 0 * p)));
49 0 : p0q0 = _mm_shuffle_epi32(q0p0, 78);
50 :
51 : {
52 : __m128i abs_p1q1, abs_p0q0, abs_q1q0, fe, ff, work;
53 0 : abs_p1p0 =
54 0 : _mm_or_si128(_mm_subs_epu8(q1p1, q0p0), _mm_subs_epu8(q0p0, q1p1));
55 0 : abs_q1q0 = _mm_srli_si128(abs_p1p0, 8);
56 0 : fe = _mm_set1_epi8(0xfe);
57 0 : ff = _mm_cmpeq_epi8(abs_p1p0, abs_p1p0);
58 0 : abs_p0q0 =
59 0 : _mm_or_si128(_mm_subs_epu8(q0p0, p0q0), _mm_subs_epu8(p0q0, q0p0));
60 0 : abs_p1q1 =
61 0 : _mm_or_si128(_mm_subs_epu8(q1p1, p1q1), _mm_subs_epu8(p1q1, q1p1));
62 0 : flat = _mm_max_epu8(abs_p1p0, abs_q1q0);
63 0 : hev = _mm_subs_epu8(flat, thresh);
64 0 : hev = _mm_xor_si128(_mm_cmpeq_epi8(hev, zero), ff);
65 :
66 0 : abs_p0q0 = _mm_adds_epu8(abs_p0q0, abs_p0q0);
67 0 : abs_p1q1 = _mm_srli_epi16(_mm_and_si128(abs_p1q1, fe), 1);
68 0 : mask = _mm_subs_epu8(_mm_adds_epu8(abs_p0q0, abs_p1q1), blimit);
69 0 : mask = _mm_xor_si128(_mm_cmpeq_epi8(mask, zero), ff);
70 : // mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2 > blimit) * -1;
71 0 : mask = _mm_max_epu8(abs_p1p0, mask);
72 : // mask |= (abs(p1 - p0) > limit) * -1;
73 : // mask |= (abs(q1 - q0) > limit) * -1;
74 :
75 0 : work = _mm_max_epu8(
76 : _mm_or_si128(_mm_subs_epu8(q2p2, q1p1), _mm_subs_epu8(q1p1, q2p2)),
77 : _mm_or_si128(_mm_subs_epu8(q3p3, q2p2), _mm_subs_epu8(q2p2, q3p3)));
78 0 : mask = _mm_max_epu8(work, mask);
79 0 : mask = _mm_max_epu8(mask, _mm_srli_si128(mask, 8));
80 0 : mask = _mm_subs_epu8(mask, limit);
81 0 : mask = _mm_cmpeq_epi8(mask, zero);
82 : }
83 :
84 : // lp filter
85 : {
86 0 : const __m128i t4 = _mm_set1_epi8(4);
87 0 : const __m128i t3 = _mm_set1_epi8(3);
88 0 : const __m128i t80 = _mm_set1_epi8(0x80);
89 0 : const __m128i t1 = _mm_set1_epi16(0x1);
90 0 : __m128i qs1ps1 = _mm_xor_si128(q1p1, t80);
91 0 : __m128i qs0ps0 = _mm_xor_si128(q0p0, t80);
92 0 : __m128i qs0 = _mm_xor_si128(p0q0, t80);
93 0 : __m128i qs1 = _mm_xor_si128(p1q1, t80);
94 : __m128i filt;
95 : __m128i work_a;
96 : __m128i filter1, filter2;
97 : __m128i flat2_q6p6, flat2_q5p5, flat2_q4p4, flat2_q3p3, flat2_q2p2;
98 : __m128i flat2_q1p1, flat2_q0p0, flat_q2p2, flat_q1p1, flat_q0p0;
99 :
100 0 : filt = _mm_and_si128(_mm_subs_epi8(qs1ps1, qs1), hev);
101 0 : work_a = _mm_subs_epi8(qs0, qs0ps0);
102 0 : filt = _mm_adds_epi8(filt, work_a);
103 0 : filt = _mm_adds_epi8(filt, work_a);
104 0 : filt = _mm_adds_epi8(filt, work_a);
105 : /* (aom_filter + 3 * (qs0 - ps0)) & mask */
106 0 : filt = _mm_and_si128(filt, mask);
107 :
108 0 : filter1 = _mm_adds_epi8(filt, t4);
109 0 : filter2 = _mm_adds_epi8(filt, t3);
110 :
111 0 : filter1 = _mm_unpacklo_epi8(zero, filter1);
112 0 : filter1 = _mm_srai_epi16(filter1, 0xB);
113 0 : filter2 = _mm_unpacklo_epi8(zero, filter2);
114 0 : filter2 = _mm_srai_epi16(filter2, 0xB);
115 :
116 : /* Filter1 >> 3 */
117 0 : filt = _mm_packs_epi16(filter2, _mm_subs_epi16(zero, filter1));
118 0 : qs0ps0 = _mm_xor_si128(_mm_adds_epi8(qs0ps0, filt), t80);
119 :
120 : /* filt >> 1 */
121 0 : filt = _mm_adds_epi16(filter1, t1);
122 0 : filt = _mm_srai_epi16(filt, 1);
123 0 : filt = _mm_andnot_si128(_mm_srai_epi16(_mm_unpacklo_epi8(zero, hev), 0x8),
124 : filt);
125 0 : filt = _mm_packs_epi16(filt, _mm_subs_epi16(zero, filt));
126 0 : qs1ps1 = _mm_xor_si128(_mm_adds_epi8(qs1ps1, filt), t80);
127 : // loopfilter done
128 :
129 : {
130 : __m128i work;
131 0 : flat = _mm_max_epu8(
132 : _mm_or_si128(_mm_subs_epu8(q2p2, q0p0), _mm_subs_epu8(q0p0, q2p2)),
133 : _mm_or_si128(_mm_subs_epu8(q3p3, q0p0), _mm_subs_epu8(q0p0, q3p3)));
134 0 : flat = _mm_max_epu8(abs_p1p0, flat);
135 0 : flat = _mm_max_epu8(flat, _mm_srli_si128(flat, 8));
136 0 : flat = _mm_subs_epu8(flat, one);
137 0 : flat = _mm_cmpeq_epi8(flat, zero);
138 0 : flat = _mm_and_si128(flat, mask);
139 :
140 0 : q5p5 = _mm_loadl_epi64((__m128i *)(s - 6 * p));
141 0 : q5p5 = _mm_castps_si128(
142 0 : _mm_loadh_pi(_mm_castsi128_ps(q5p5), (__m64 *)(s + 5 * p)));
143 :
144 0 : q6p6 = _mm_loadl_epi64((__m128i *)(s - 7 * p));
145 0 : q6p6 = _mm_castps_si128(
146 0 : _mm_loadh_pi(_mm_castsi128_ps(q6p6), (__m64 *)(s + 6 * p)));
147 :
148 0 : flat2 = _mm_max_epu8(
149 : _mm_or_si128(_mm_subs_epu8(q4p4, q0p0), _mm_subs_epu8(q0p0, q4p4)),
150 : _mm_or_si128(_mm_subs_epu8(q5p5, q0p0), _mm_subs_epu8(q0p0, q5p5)));
151 :
152 0 : q7p7 = _mm_loadl_epi64((__m128i *)(s - 8 * p));
153 0 : q7p7 = _mm_castps_si128(
154 0 : _mm_loadh_pi(_mm_castsi128_ps(q7p7), (__m64 *)(s + 7 * p)));
155 :
156 0 : work = _mm_max_epu8(
157 : _mm_or_si128(_mm_subs_epu8(q6p6, q0p0), _mm_subs_epu8(q0p0, q6p6)),
158 : _mm_or_si128(_mm_subs_epu8(q7p7, q0p0), _mm_subs_epu8(q0p0, q7p7)));
159 :
160 0 : flat2 = _mm_max_epu8(work, flat2);
161 0 : flat2 = _mm_max_epu8(flat2, _mm_srli_si128(flat2, 8));
162 0 : flat2 = _mm_subs_epu8(flat2, one);
163 0 : flat2 = _mm_cmpeq_epi8(flat2, zero);
164 0 : flat2 = _mm_and_si128(flat2, flat); // flat2 & flat & mask
165 : }
166 :
167 : // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
168 : // flat and wide flat calculations
169 : {
170 0 : const __m128i eight = _mm_set1_epi16(8);
171 0 : const __m128i four = _mm_set1_epi16(4);
172 : __m128i p7_16, p6_16, p5_16, p4_16, p3_16, p2_16, p1_16, p0_16;
173 : __m128i q7_16, q6_16, q5_16, q4_16, q3_16, q2_16, q1_16, q0_16;
174 : __m128i pixelFilter_p, pixelFilter_q;
175 : __m128i pixetFilter_p2p1p0, pixetFilter_q2q1q0;
176 : __m128i sum_p7, sum_q7, sum_p3, sum_q3, res_p, res_q;
177 :
178 0 : p7_16 = _mm_unpacklo_epi8(q7p7, zero);
179 0 : p6_16 = _mm_unpacklo_epi8(q6p6, zero);
180 0 : p5_16 = _mm_unpacklo_epi8(q5p5, zero);
181 0 : p4_16 = _mm_unpacklo_epi8(q4p4, zero);
182 0 : p3_16 = _mm_unpacklo_epi8(q3p3, zero);
183 0 : p2_16 = _mm_unpacklo_epi8(q2p2, zero);
184 0 : p1_16 = _mm_unpacklo_epi8(q1p1, zero);
185 0 : p0_16 = _mm_unpacklo_epi8(q0p0, zero);
186 0 : q0_16 = _mm_unpackhi_epi8(q0p0, zero);
187 0 : q1_16 = _mm_unpackhi_epi8(q1p1, zero);
188 0 : q2_16 = _mm_unpackhi_epi8(q2p2, zero);
189 0 : q3_16 = _mm_unpackhi_epi8(q3p3, zero);
190 0 : q4_16 = _mm_unpackhi_epi8(q4p4, zero);
191 0 : q5_16 = _mm_unpackhi_epi8(q5p5, zero);
192 0 : q6_16 = _mm_unpackhi_epi8(q6p6, zero);
193 0 : q7_16 = _mm_unpackhi_epi8(q7p7, zero);
194 :
195 0 : pixelFilter_p = _mm_add_epi16(_mm_add_epi16(p6_16, p5_16),
196 : _mm_add_epi16(p4_16, p3_16));
197 0 : pixelFilter_q = _mm_add_epi16(_mm_add_epi16(q6_16, q5_16),
198 : _mm_add_epi16(q4_16, q3_16));
199 :
200 0 : pixetFilter_p2p1p0 = _mm_add_epi16(p0_16, _mm_add_epi16(p2_16, p1_16));
201 0 : pixelFilter_p = _mm_add_epi16(pixelFilter_p, pixetFilter_p2p1p0);
202 :
203 0 : pixetFilter_q2q1q0 = _mm_add_epi16(q0_16, _mm_add_epi16(q2_16, q1_16));
204 0 : pixelFilter_q = _mm_add_epi16(pixelFilter_q, pixetFilter_q2q1q0);
205 0 : pixelFilter_p =
206 0 : _mm_add_epi16(eight, _mm_add_epi16(pixelFilter_p, pixelFilter_q));
207 0 : pixetFilter_p2p1p0 = _mm_add_epi16(
208 : four, _mm_add_epi16(pixetFilter_p2p1p0, pixetFilter_q2q1q0));
209 0 : res_p = _mm_srli_epi16(
210 : _mm_add_epi16(pixelFilter_p, _mm_add_epi16(p7_16, p0_16)), 4);
211 0 : res_q = _mm_srli_epi16(
212 : _mm_add_epi16(pixelFilter_p, _mm_add_epi16(q7_16, q0_16)), 4);
213 0 : flat2_q0p0 = _mm_packus_epi16(res_p, res_q);
214 0 : res_p = _mm_srli_epi16(
215 : _mm_add_epi16(pixetFilter_p2p1p0, _mm_add_epi16(p3_16, p0_16)), 3);
216 0 : res_q = _mm_srli_epi16(
217 : _mm_add_epi16(pixetFilter_p2p1p0, _mm_add_epi16(q3_16, q0_16)), 3);
218 :
219 0 : flat_q0p0 = _mm_packus_epi16(res_p, res_q);
220 :
221 0 : sum_p7 = _mm_add_epi16(p7_16, p7_16);
222 0 : sum_q7 = _mm_add_epi16(q7_16, q7_16);
223 0 : sum_p3 = _mm_add_epi16(p3_16, p3_16);
224 0 : sum_q3 = _mm_add_epi16(q3_16, q3_16);
225 :
226 0 : pixelFilter_q = _mm_sub_epi16(pixelFilter_p, p6_16);
227 0 : pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q6_16);
228 0 : res_p = _mm_srli_epi16(
229 : _mm_add_epi16(pixelFilter_p, _mm_add_epi16(sum_p7, p1_16)), 4);
230 0 : res_q = _mm_srli_epi16(
231 : _mm_add_epi16(pixelFilter_q, _mm_add_epi16(sum_q7, q1_16)), 4);
232 0 : flat2_q1p1 = _mm_packus_epi16(res_p, res_q);
233 :
234 0 : pixetFilter_q2q1q0 = _mm_sub_epi16(pixetFilter_p2p1p0, p2_16);
235 0 : pixetFilter_p2p1p0 = _mm_sub_epi16(pixetFilter_p2p1p0, q2_16);
236 0 : res_p = _mm_srli_epi16(
237 : _mm_add_epi16(pixetFilter_p2p1p0, _mm_add_epi16(sum_p3, p1_16)), 3);
238 0 : res_q = _mm_srli_epi16(
239 : _mm_add_epi16(pixetFilter_q2q1q0, _mm_add_epi16(sum_q3, q1_16)), 3);
240 0 : flat_q1p1 = _mm_packus_epi16(res_p, res_q);
241 :
242 0 : sum_p7 = _mm_add_epi16(sum_p7, p7_16);
243 0 : sum_q7 = _mm_add_epi16(sum_q7, q7_16);
244 0 : sum_p3 = _mm_add_epi16(sum_p3, p3_16);
245 0 : sum_q3 = _mm_add_epi16(sum_q3, q3_16);
246 :
247 0 : pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q5_16);
248 0 : pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p5_16);
249 0 : res_p = _mm_srli_epi16(
250 : _mm_add_epi16(pixelFilter_p, _mm_add_epi16(sum_p7, p2_16)), 4);
251 0 : res_q = _mm_srli_epi16(
252 : _mm_add_epi16(pixelFilter_q, _mm_add_epi16(sum_q7, q2_16)), 4);
253 0 : flat2_q2p2 = _mm_packus_epi16(res_p, res_q);
254 :
255 0 : pixetFilter_p2p1p0 = _mm_sub_epi16(pixetFilter_p2p1p0, q1_16);
256 0 : pixetFilter_q2q1q0 = _mm_sub_epi16(pixetFilter_q2q1q0, p1_16);
257 :
258 0 : res_p = _mm_srli_epi16(
259 : _mm_add_epi16(pixetFilter_p2p1p0, _mm_add_epi16(sum_p3, p2_16)), 3);
260 0 : res_q = _mm_srli_epi16(
261 : _mm_add_epi16(pixetFilter_q2q1q0, _mm_add_epi16(sum_q3, q2_16)), 3);
262 0 : flat_q2p2 = _mm_packus_epi16(res_p, res_q);
263 :
264 0 : sum_p7 = _mm_add_epi16(sum_p7, p7_16);
265 0 : sum_q7 = _mm_add_epi16(sum_q7, q7_16);
266 0 : pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q4_16);
267 0 : pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p4_16);
268 0 : res_p = _mm_srli_epi16(
269 : _mm_add_epi16(pixelFilter_p, _mm_add_epi16(sum_p7, p3_16)), 4);
270 0 : res_q = _mm_srli_epi16(
271 : _mm_add_epi16(pixelFilter_q, _mm_add_epi16(sum_q7, q3_16)), 4);
272 0 : flat2_q3p3 = _mm_packus_epi16(res_p, res_q);
273 :
274 0 : sum_p7 = _mm_add_epi16(sum_p7, p7_16);
275 0 : sum_q7 = _mm_add_epi16(sum_q7, q7_16);
276 0 : pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q3_16);
277 0 : pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p3_16);
278 0 : res_p = _mm_srli_epi16(
279 : _mm_add_epi16(pixelFilter_p, _mm_add_epi16(sum_p7, p4_16)), 4);
280 0 : res_q = _mm_srli_epi16(
281 : _mm_add_epi16(pixelFilter_q, _mm_add_epi16(sum_q7, q4_16)), 4);
282 0 : flat2_q4p4 = _mm_packus_epi16(res_p, res_q);
283 :
284 0 : sum_p7 = _mm_add_epi16(sum_p7, p7_16);
285 0 : sum_q7 = _mm_add_epi16(sum_q7, q7_16);
286 0 : pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q2_16);
287 0 : pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p2_16);
288 0 : res_p = _mm_srli_epi16(
289 : _mm_add_epi16(pixelFilter_p, _mm_add_epi16(sum_p7, p5_16)), 4);
290 0 : res_q = _mm_srli_epi16(
291 : _mm_add_epi16(pixelFilter_q, _mm_add_epi16(sum_q7, q5_16)), 4);
292 0 : flat2_q5p5 = _mm_packus_epi16(res_p, res_q);
293 :
294 0 : sum_p7 = _mm_add_epi16(sum_p7, p7_16);
295 0 : sum_q7 = _mm_add_epi16(sum_q7, q7_16);
296 0 : pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q1_16);
297 0 : pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p1_16);
298 0 : res_p = _mm_srli_epi16(
299 : _mm_add_epi16(pixelFilter_p, _mm_add_epi16(sum_p7, p6_16)), 4);
300 0 : res_q = _mm_srli_epi16(
301 : _mm_add_epi16(pixelFilter_q, _mm_add_epi16(sum_q7, q6_16)), 4);
302 0 : flat2_q6p6 = _mm_packus_epi16(res_p, res_q);
303 : }
304 : // wide flat
305 : // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
306 :
307 0 : flat = _mm_shuffle_epi32(flat, 68);
308 0 : flat2 = _mm_shuffle_epi32(flat2, 68);
309 :
310 0 : q2p2 = _mm_andnot_si128(flat, q2p2);
311 0 : flat_q2p2 = _mm_and_si128(flat, flat_q2p2);
312 0 : q2p2 = _mm_or_si128(q2p2, flat_q2p2);
313 :
314 0 : qs1ps1 = _mm_andnot_si128(flat, qs1ps1);
315 0 : flat_q1p1 = _mm_and_si128(flat, flat_q1p1);
316 0 : q1p1 = _mm_or_si128(qs1ps1, flat_q1p1);
317 :
318 0 : qs0ps0 = _mm_andnot_si128(flat, qs0ps0);
319 0 : flat_q0p0 = _mm_and_si128(flat, flat_q0p0);
320 0 : q0p0 = _mm_or_si128(qs0ps0, flat_q0p0);
321 :
322 0 : q6p6 = _mm_andnot_si128(flat2, q6p6);
323 0 : flat2_q6p6 = _mm_and_si128(flat2, flat2_q6p6);
324 0 : q6p6 = _mm_or_si128(q6p6, flat2_q6p6);
325 0 : _mm_storel_epi64((__m128i *)(s - 7 * p), q6p6);
326 0 : _mm_storeh_pi((__m64 *)(s + 6 * p), _mm_castsi128_ps(q6p6));
327 :
328 0 : q5p5 = _mm_andnot_si128(flat2, q5p5);
329 0 : flat2_q5p5 = _mm_and_si128(flat2, flat2_q5p5);
330 0 : q5p5 = _mm_or_si128(q5p5, flat2_q5p5);
331 0 : _mm_storel_epi64((__m128i *)(s - 6 * p), q5p5);
332 0 : _mm_storeh_pi((__m64 *)(s + 5 * p), _mm_castsi128_ps(q5p5));
333 :
334 0 : q4p4 = _mm_andnot_si128(flat2, q4p4);
335 0 : flat2_q4p4 = _mm_and_si128(flat2, flat2_q4p4);
336 0 : q4p4 = _mm_or_si128(q4p4, flat2_q4p4);
337 0 : _mm_storel_epi64((__m128i *)(s - 5 * p), q4p4);
338 0 : _mm_storeh_pi((__m64 *)(s + 4 * p), _mm_castsi128_ps(q4p4));
339 :
340 0 : q3p3 = _mm_andnot_si128(flat2, q3p3);
341 0 : flat2_q3p3 = _mm_and_si128(flat2, flat2_q3p3);
342 0 : q3p3 = _mm_or_si128(q3p3, flat2_q3p3);
343 0 : _mm_storel_epi64((__m128i *)(s - 4 * p), q3p3);
344 0 : _mm_storeh_pi((__m64 *)(s + 3 * p), _mm_castsi128_ps(q3p3));
345 :
346 0 : q2p2 = _mm_andnot_si128(flat2, q2p2);
347 0 : flat2_q2p2 = _mm_and_si128(flat2, flat2_q2p2);
348 0 : q2p2 = _mm_or_si128(q2p2, flat2_q2p2);
349 0 : _mm_storel_epi64((__m128i *)(s - 3 * p), q2p2);
350 0 : _mm_storeh_pi((__m64 *)(s + 2 * p), _mm_castsi128_ps(q2p2));
351 :
352 0 : q1p1 = _mm_andnot_si128(flat2, q1p1);
353 0 : flat2_q1p1 = _mm_and_si128(flat2, flat2_q1p1);
354 0 : q1p1 = _mm_or_si128(q1p1, flat2_q1p1);
355 0 : _mm_storel_epi64((__m128i *)(s - 2 * p), q1p1);
356 0 : _mm_storeh_pi((__m64 *)(s + 1 * p), _mm_castsi128_ps(q1p1));
357 :
358 0 : q0p0 = _mm_andnot_si128(flat2, q0p0);
359 0 : flat2_q0p0 = _mm_and_si128(flat2, flat2_q0p0);
360 0 : q0p0 = _mm_or_si128(q0p0, flat2_q0p0);
361 0 : _mm_storel_epi64((__m128i *)(s - 1 * p), q0p0);
362 0 : _mm_storeh_pi((__m64 *)(s - 0 * p), _mm_castsi128_ps(q0p0));
363 : }
364 0 : }
365 :
366 : DECLARE_ALIGNED(32, static const uint8_t, filt_loopfilter_avx2[32]) = {
367 : 0, 128, 1, 128, 2, 128, 3, 128, 4, 128, 5, 128, 6, 128, 7, 128,
368 : 8, 128, 9, 128, 10, 128, 11, 128, 12, 128, 13, 128, 14, 128, 15, 128
369 : };
370 :
371 0 : void aom_lpf_horizontal_edge_16_avx2(unsigned char *s, int p,
372 : const unsigned char *_blimit,
373 : const unsigned char *_limit,
374 : const unsigned char *_thresh) {
375 : __m128i mask, hev, flat, flat2;
376 0 : const __m128i zero = _mm_set1_epi16(0);
377 0 : const __m128i one = _mm_set1_epi8(1);
378 : __m128i p7, p6, p5;
379 : __m128i p4, p3, p2, p1, p0, q0, q1, q2, q3, q4;
380 : __m128i q5, q6, q7;
381 : __m256i p256_7, q256_7, p256_6, q256_6, p256_5, q256_5, p256_4, q256_4,
382 : p256_3, q256_3, p256_2, q256_2, p256_1, q256_1, p256_0, q256_0;
383 :
384 0 : const __m128i thresh =
385 0 : _mm_broadcastb_epi8(_mm_cvtsi32_si128((int)_thresh[0]));
386 0 : const __m128i limit = _mm_broadcastb_epi8(_mm_cvtsi32_si128((int)_limit[0]));
387 0 : const __m128i blimit =
388 0 : _mm_broadcastb_epi8(_mm_cvtsi32_si128((int)_blimit[0]));
389 :
390 0 : p256_4 =
391 0 : _mm256_castpd_si256(_mm256_broadcast_pd((__m128d const *)(s - 5 * p)));
392 0 : p256_3 =
393 0 : _mm256_castpd_si256(_mm256_broadcast_pd((__m128d const *)(s - 4 * p)));
394 0 : p256_2 =
395 0 : _mm256_castpd_si256(_mm256_broadcast_pd((__m128d const *)(s - 3 * p)));
396 0 : p256_1 =
397 0 : _mm256_castpd_si256(_mm256_broadcast_pd((__m128d const *)(s - 2 * p)));
398 0 : p256_0 =
399 0 : _mm256_castpd_si256(_mm256_broadcast_pd((__m128d const *)(s - 1 * p)));
400 0 : q256_0 =
401 0 : _mm256_castpd_si256(_mm256_broadcast_pd((__m128d const *)(s - 0 * p)));
402 0 : q256_1 =
403 0 : _mm256_castpd_si256(_mm256_broadcast_pd((__m128d const *)(s + 1 * p)));
404 0 : q256_2 =
405 0 : _mm256_castpd_si256(_mm256_broadcast_pd((__m128d const *)(s + 2 * p)));
406 0 : q256_3 =
407 0 : _mm256_castpd_si256(_mm256_broadcast_pd((__m128d const *)(s + 3 * p)));
408 0 : q256_4 =
409 0 : _mm256_castpd_si256(_mm256_broadcast_pd((__m128d const *)(s + 4 * p)));
410 :
411 0 : p4 = _mm256_castsi256_si128(p256_4);
412 0 : p3 = _mm256_castsi256_si128(p256_3);
413 0 : p2 = _mm256_castsi256_si128(p256_2);
414 0 : p1 = _mm256_castsi256_si128(p256_1);
415 0 : p0 = _mm256_castsi256_si128(p256_0);
416 0 : q0 = _mm256_castsi256_si128(q256_0);
417 0 : q1 = _mm256_castsi256_si128(q256_1);
418 0 : q2 = _mm256_castsi256_si128(q256_2);
419 0 : q3 = _mm256_castsi256_si128(q256_3);
420 0 : q4 = _mm256_castsi256_si128(q256_4);
421 :
422 : {
423 0 : const __m128i abs_p1p0 =
424 0 : _mm_or_si128(_mm_subs_epu8(p1, p0), _mm_subs_epu8(p0, p1));
425 0 : const __m128i abs_q1q0 =
426 0 : _mm_or_si128(_mm_subs_epu8(q1, q0), _mm_subs_epu8(q0, q1));
427 0 : const __m128i fe = _mm_set1_epi8(0xfe);
428 0 : const __m128i ff = _mm_cmpeq_epi8(abs_p1p0, abs_p1p0);
429 0 : __m128i abs_p0q0 =
430 0 : _mm_or_si128(_mm_subs_epu8(p0, q0), _mm_subs_epu8(q0, p0));
431 0 : __m128i abs_p1q1 =
432 0 : _mm_or_si128(_mm_subs_epu8(p1, q1), _mm_subs_epu8(q1, p1));
433 : __m128i work;
434 0 : flat = _mm_max_epu8(abs_p1p0, abs_q1q0);
435 0 : hev = _mm_subs_epu8(flat, thresh);
436 0 : hev = _mm_xor_si128(_mm_cmpeq_epi8(hev, zero), ff);
437 :
438 0 : abs_p0q0 = _mm_adds_epu8(abs_p0q0, abs_p0q0);
439 0 : abs_p1q1 = _mm_srli_epi16(_mm_and_si128(abs_p1q1, fe), 1);
440 0 : mask = _mm_subs_epu8(_mm_adds_epu8(abs_p0q0, abs_p1q1), blimit);
441 0 : mask = _mm_xor_si128(_mm_cmpeq_epi8(mask, zero), ff);
442 : // mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2 > blimit) * -1;
443 0 : mask = _mm_max_epu8(flat, mask);
444 : // mask |= (abs(p1 - p0) > limit) * -1;
445 : // mask |= (abs(q1 - q0) > limit) * -1;
446 0 : work = _mm_max_epu8(
447 : _mm_or_si128(_mm_subs_epu8(p2, p1), _mm_subs_epu8(p1, p2)),
448 : _mm_or_si128(_mm_subs_epu8(p3, p2), _mm_subs_epu8(p2, p3)));
449 0 : mask = _mm_max_epu8(work, mask);
450 0 : work = _mm_max_epu8(
451 : _mm_or_si128(_mm_subs_epu8(q2, q1), _mm_subs_epu8(q1, q2)),
452 : _mm_or_si128(_mm_subs_epu8(q3, q2), _mm_subs_epu8(q2, q3)));
453 0 : mask = _mm_max_epu8(work, mask);
454 0 : mask = _mm_subs_epu8(mask, limit);
455 0 : mask = _mm_cmpeq_epi8(mask, zero);
456 : }
457 :
458 : // lp filter
459 : {
460 0 : const __m128i t4 = _mm_set1_epi8(4);
461 0 : const __m128i t3 = _mm_set1_epi8(3);
462 0 : const __m128i t80 = _mm_set1_epi8(0x80);
463 0 : const __m128i te0 = _mm_set1_epi8(0xe0);
464 0 : const __m128i t1f = _mm_set1_epi8(0x1f);
465 0 : const __m128i t1 = _mm_set1_epi8(0x1);
466 0 : const __m128i t7f = _mm_set1_epi8(0x7f);
467 :
468 0 : __m128i ps1 = _mm_xor_si128(p1, t80);
469 0 : __m128i ps0 = _mm_xor_si128(p0, t80);
470 0 : __m128i qs0 = _mm_xor_si128(q0, t80);
471 0 : __m128i qs1 = _mm_xor_si128(q1, t80);
472 : __m128i filt;
473 : __m128i work_a;
474 : __m128i filter1, filter2;
475 : __m128i flat2_p6, flat2_p5, flat2_p4, flat2_p3, flat2_p2, flat2_p1,
476 : flat2_p0, flat2_q0, flat2_q1, flat2_q2, flat2_q3, flat2_q4, flat2_q5,
477 : flat2_q6, flat_p2, flat_p1, flat_p0, flat_q0, flat_q1, flat_q2;
478 :
479 0 : filt = _mm_and_si128(_mm_subs_epi8(ps1, qs1), hev);
480 0 : work_a = _mm_subs_epi8(qs0, ps0);
481 0 : filt = _mm_adds_epi8(filt, work_a);
482 0 : filt = _mm_adds_epi8(filt, work_a);
483 0 : filt = _mm_adds_epi8(filt, work_a);
484 : /* (aom_filter + 3 * (qs0 - ps0)) & mask */
485 0 : filt = _mm_and_si128(filt, mask);
486 :
487 0 : filter1 = _mm_adds_epi8(filt, t4);
488 0 : filter2 = _mm_adds_epi8(filt, t3);
489 :
490 : /* Filter1 >> 3 */
491 0 : work_a = _mm_cmpgt_epi8(zero, filter1);
492 0 : filter1 = _mm_srli_epi16(filter1, 3);
493 0 : work_a = _mm_and_si128(work_a, te0);
494 0 : filter1 = _mm_and_si128(filter1, t1f);
495 0 : filter1 = _mm_or_si128(filter1, work_a);
496 0 : qs0 = _mm_xor_si128(_mm_subs_epi8(qs0, filter1), t80);
497 :
498 : /* Filter2 >> 3 */
499 0 : work_a = _mm_cmpgt_epi8(zero, filter2);
500 0 : filter2 = _mm_srli_epi16(filter2, 3);
501 0 : work_a = _mm_and_si128(work_a, te0);
502 0 : filter2 = _mm_and_si128(filter2, t1f);
503 0 : filter2 = _mm_or_si128(filter2, work_a);
504 0 : ps0 = _mm_xor_si128(_mm_adds_epi8(ps0, filter2), t80);
505 :
506 : /* filt >> 1 */
507 0 : filt = _mm_adds_epi8(filter1, t1);
508 0 : work_a = _mm_cmpgt_epi8(zero, filt);
509 0 : filt = _mm_srli_epi16(filt, 1);
510 0 : work_a = _mm_and_si128(work_a, t80);
511 0 : filt = _mm_and_si128(filt, t7f);
512 0 : filt = _mm_or_si128(filt, work_a);
513 0 : filt = _mm_andnot_si128(hev, filt);
514 0 : ps1 = _mm_xor_si128(_mm_adds_epi8(ps1, filt), t80);
515 0 : qs1 = _mm_xor_si128(_mm_subs_epi8(qs1, filt), t80);
516 : // loopfilter done
517 :
518 : {
519 : __m128i work;
520 0 : work = _mm_max_epu8(
521 : _mm_or_si128(_mm_subs_epu8(p2, p0), _mm_subs_epu8(p0, p2)),
522 : _mm_or_si128(_mm_subs_epu8(q2, q0), _mm_subs_epu8(q0, q2)));
523 0 : flat = _mm_max_epu8(work, flat);
524 0 : work = _mm_max_epu8(
525 : _mm_or_si128(_mm_subs_epu8(p3, p0), _mm_subs_epu8(p0, p3)),
526 : _mm_or_si128(_mm_subs_epu8(q3, q0), _mm_subs_epu8(q0, q3)));
527 0 : flat = _mm_max_epu8(work, flat);
528 0 : work = _mm_max_epu8(
529 : _mm_or_si128(_mm_subs_epu8(p4, p0), _mm_subs_epu8(p0, p4)),
530 : _mm_or_si128(_mm_subs_epu8(q4, q0), _mm_subs_epu8(q0, q4)));
531 0 : flat = _mm_subs_epu8(flat, one);
532 0 : flat = _mm_cmpeq_epi8(flat, zero);
533 0 : flat = _mm_and_si128(flat, mask);
534 :
535 0 : p256_5 = _mm256_castpd_si256(
536 0 : _mm256_broadcast_pd((__m128d const *)(s - 6 * p)));
537 0 : q256_5 = _mm256_castpd_si256(
538 0 : _mm256_broadcast_pd((__m128d const *)(s + 5 * p)));
539 0 : p5 = _mm256_castsi256_si128(p256_5);
540 0 : q5 = _mm256_castsi256_si128(q256_5);
541 0 : flat2 = _mm_max_epu8(
542 : _mm_or_si128(_mm_subs_epu8(p5, p0), _mm_subs_epu8(p0, p5)),
543 : _mm_or_si128(_mm_subs_epu8(q5, q0), _mm_subs_epu8(q0, q5)));
544 :
545 0 : flat2 = _mm_max_epu8(work, flat2);
546 0 : p256_6 = _mm256_castpd_si256(
547 0 : _mm256_broadcast_pd((__m128d const *)(s - 7 * p)));
548 0 : q256_6 = _mm256_castpd_si256(
549 0 : _mm256_broadcast_pd((__m128d const *)(s + 6 * p)));
550 0 : p6 = _mm256_castsi256_si128(p256_6);
551 0 : q6 = _mm256_castsi256_si128(q256_6);
552 0 : work = _mm_max_epu8(
553 : _mm_or_si128(_mm_subs_epu8(p6, p0), _mm_subs_epu8(p0, p6)),
554 : _mm_or_si128(_mm_subs_epu8(q6, q0), _mm_subs_epu8(q0, q6)));
555 :
556 0 : flat2 = _mm_max_epu8(work, flat2);
557 :
558 0 : p256_7 = _mm256_castpd_si256(
559 0 : _mm256_broadcast_pd((__m128d const *)(s - 8 * p)));
560 0 : q256_7 = _mm256_castpd_si256(
561 0 : _mm256_broadcast_pd((__m128d const *)(s + 7 * p)));
562 0 : p7 = _mm256_castsi256_si128(p256_7);
563 0 : q7 = _mm256_castsi256_si128(q256_7);
564 0 : work = _mm_max_epu8(
565 : _mm_or_si128(_mm_subs_epu8(p7, p0), _mm_subs_epu8(p0, p7)),
566 : _mm_or_si128(_mm_subs_epu8(q7, q0), _mm_subs_epu8(q0, q7)));
567 :
568 0 : flat2 = _mm_max_epu8(work, flat2);
569 0 : flat2 = _mm_subs_epu8(flat2, one);
570 0 : flat2 = _mm_cmpeq_epi8(flat2, zero);
571 0 : flat2 = _mm_and_si128(flat2, flat); // flat2 & flat & mask
572 : }
573 :
574 : // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
575 : // flat and wide flat calculations
576 : {
577 0 : const __m256i eight = _mm256_set1_epi16(8);
578 0 : const __m256i four = _mm256_set1_epi16(4);
579 : __m256i pixelFilter_p, pixelFilter_q, pixetFilter_p2p1p0,
580 : pixetFilter_q2q1q0, sum_p7, sum_q7, sum_p3, sum_q3, res_p, res_q;
581 :
582 0 : const __m256i filter =
583 : _mm256_load_si256((__m256i const *)filt_loopfilter_avx2);
584 0 : p256_7 = _mm256_shuffle_epi8(p256_7, filter);
585 0 : p256_6 = _mm256_shuffle_epi8(p256_6, filter);
586 0 : p256_5 = _mm256_shuffle_epi8(p256_5, filter);
587 0 : p256_4 = _mm256_shuffle_epi8(p256_4, filter);
588 0 : p256_3 = _mm256_shuffle_epi8(p256_3, filter);
589 0 : p256_2 = _mm256_shuffle_epi8(p256_2, filter);
590 0 : p256_1 = _mm256_shuffle_epi8(p256_1, filter);
591 0 : p256_0 = _mm256_shuffle_epi8(p256_0, filter);
592 0 : q256_0 = _mm256_shuffle_epi8(q256_0, filter);
593 0 : q256_1 = _mm256_shuffle_epi8(q256_1, filter);
594 0 : q256_2 = _mm256_shuffle_epi8(q256_2, filter);
595 0 : q256_3 = _mm256_shuffle_epi8(q256_3, filter);
596 0 : q256_4 = _mm256_shuffle_epi8(q256_4, filter);
597 0 : q256_5 = _mm256_shuffle_epi8(q256_5, filter);
598 0 : q256_6 = _mm256_shuffle_epi8(q256_6, filter);
599 0 : q256_7 = _mm256_shuffle_epi8(q256_7, filter);
600 :
601 0 : pixelFilter_p = _mm256_add_epi16(_mm256_add_epi16(p256_6, p256_5),
602 : _mm256_add_epi16(p256_4, p256_3));
603 0 : pixelFilter_q = _mm256_add_epi16(_mm256_add_epi16(q256_6, q256_5),
604 : _mm256_add_epi16(q256_4, q256_3));
605 :
606 0 : pixetFilter_p2p1p0 =
607 0 : _mm256_add_epi16(p256_0, _mm256_add_epi16(p256_2, p256_1));
608 0 : pixelFilter_p = _mm256_add_epi16(pixelFilter_p, pixetFilter_p2p1p0);
609 :
610 0 : pixetFilter_q2q1q0 =
611 0 : _mm256_add_epi16(q256_0, _mm256_add_epi16(q256_2, q256_1));
612 0 : pixelFilter_q = _mm256_add_epi16(pixelFilter_q, pixetFilter_q2q1q0);
613 :
614 0 : pixelFilter_p = _mm256_add_epi16(
615 : eight, _mm256_add_epi16(pixelFilter_p, pixelFilter_q));
616 :
617 0 : pixetFilter_p2p1p0 = _mm256_add_epi16(
618 : four, _mm256_add_epi16(pixetFilter_p2p1p0, pixetFilter_q2q1q0));
619 :
620 0 : res_p = _mm256_srli_epi16(
621 : _mm256_add_epi16(pixelFilter_p, _mm256_add_epi16(p256_7, p256_0)), 4);
622 :
623 0 : flat2_p0 = _mm256_castsi256_si128(
624 0 : _mm256_permute4x64_epi64(_mm256_packus_epi16(res_p, res_p), 168));
625 :
626 0 : res_q = _mm256_srli_epi16(
627 : _mm256_add_epi16(pixelFilter_p, _mm256_add_epi16(q256_7, q256_0)), 4);
628 :
629 0 : flat2_q0 = _mm256_castsi256_si128(
630 0 : _mm256_permute4x64_epi64(_mm256_packus_epi16(res_q, res_q), 168));
631 :
632 0 : res_p =
633 0 : _mm256_srli_epi16(_mm256_add_epi16(pixetFilter_p2p1p0,
634 : _mm256_add_epi16(p256_3, p256_0)),
635 : 3);
636 :
637 0 : flat_p0 = _mm256_castsi256_si128(
638 0 : _mm256_permute4x64_epi64(_mm256_packus_epi16(res_p, res_p), 168));
639 :
640 0 : res_q =
641 0 : _mm256_srli_epi16(_mm256_add_epi16(pixetFilter_p2p1p0,
642 : _mm256_add_epi16(q256_3, q256_0)),
643 : 3);
644 :
645 0 : flat_q0 = _mm256_castsi256_si128(
646 0 : _mm256_permute4x64_epi64(_mm256_packus_epi16(res_q, res_q), 168));
647 :
648 0 : sum_p7 = _mm256_add_epi16(p256_7, p256_7);
649 :
650 0 : sum_q7 = _mm256_add_epi16(q256_7, q256_7);
651 :
652 0 : sum_p3 = _mm256_add_epi16(p256_3, p256_3);
653 :
654 0 : sum_q3 = _mm256_add_epi16(q256_3, q256_3);
655 :
656 0 : pixelFilter_q = _mm256_sub_epi16(pixelFilter_p, p256_6);
657 :
658 0 : pixelFilter_p = _mm256_sub_epi16(pixelFilter_p, q256_6);
659 :
660 0 : res_p = _mm256_srli_epi16(
661 : _mm256_add_epi16(pixelFilter_p, _mm256_add_epi16(sum_p7, p256_1)), 4);
662 :
663 0 : flat2_p1 = _mm256_castsi256_si128(
664 0 : _mm256_permute4x64_epi64(_mm256_packus_epi16(res_p, res_p), 168));
665 :
666 0 : res_q = _mm256_srli_epi16(
667 : _mm256_add_epi16(pixelFilter_q, _mm256_add_epi16(sum_q7, q256_1)), 4);
668 :
669 0 : flat2_q1 = _mm256_castsi256_si128(
670 0 : _mm256_permute4x64_epi64(_mm256_packus_epi16(res_q, res_q), 168));
671 :
672 0 : pixetFilter_q2q1q0 = _mm256_sub_epi16(pixetFilter_p2p1p0, p256_2);
673 :
674 0 : pixetFilter_p2p1p0 = _mm256_sub_epi16(pixetFilter_p2p1p0, q256_2);
675 :
676 0 : res_p =
677 0 : _mm256_srli_epi16(_mm256_add_epi16(pixetFilter_p2p1p0,
678 : _mm256_add_epi16(sum_p3, p256_1)),
679 : 3);
680 :
681 0 : flat_p1 = _mm256_castsi256_si128(
682 0 : _mm256_permute4x64_epi64(_mm256_packus_epi16(res_p, res_p), 168));
683 :
684 0 : res_q =
685 0 : _mm256_srli_epi16(_mm256_add_epi16(pixetFilter_q2q1q0,
686 : _mm256_add_epi16(sum_q3, q256_1)),
687 : 3);
688 :
689 0 : flat_q1 = _mm256_castsi256_si128(
690 0 : _mm256_permute4x64_epi64(_mm256_packus_epi16(res_q, res_q), 168));
691 :
692 0 : sum_p7 = _mm256_add_epi16(sum_p7, p256_7);
693 :
694 0 : sum_q7 = _mm256_add_epi16(sum_q7, q256_7);
695 :
696 0 : sum_p3 = _mm256_add_epi16(sum_p3, p256_3);
697 :
698 0 : sum_q3 = _mm256_add_epi16(sum_q3, q256_3);
699 :
700 0 : pixelFilter_p = _mm256_sub_epi16(pixelFilter_p, q256_5);
701 :
702 0 : pixelFilter_q = _mm256_sub_epi16(pixelFilter_q, p256_5);
703 :
704 0 : res_p = _mm256_srli_epi16(
705 : _mm256_add_epi16(pixelFilter_p, _mm256_add_epi16(sum_p7, p256_2)), 4);
706 :
707 0 : flat2_p2 = _mm256_castsi256_si128(
708 0 : _mm256_permute4x64_epi64(_mm256_packus_epi16(res_p, res_p), 168));
709 :
710 0 : res_q = _mm256_srli_epi16(
711 : _mm256_add_epi16(pixelFilter_q, _mm256_add_epi16(sum_q7, q256_2)), 4);
712 :
713 0 : flat2_q2 = _mm256_castsi256_si128(
714 0 : _mm256_permute4x64_epi64(_mm256_packus_epi16(res_q, res_q), 168));
715 :
716 0 : pixetFilter_p2p1p0 = _mm256_sub_epi16(pixetFilter_p2p1p0, q256_1);
717 :
718 0 : pixetFilter_q2q1q0 = _mm256_sub_epi16(pixetFilter_q2q1q0, p256_1);
719 :
720 0 : res_p =
721 0 : _mm256_srli_epi16(_mm256_add_epi16(pixetFilter_p2p1p0,
722 : _mm256_add_epi16(sum_p3, p256_2)),
723 : 3);
724 :
725 0 : flat_p2 = _mm256_castsi256_si128(
726 0 : _mm256_permute4x64_epi64(_mm256_packus_epi16(res_p, res_p), 168));
727 :
728 0 : res_q =
729 0 : _mm256_srli_epi16(_mm256_add_epi16(pixetFilter_q2q1q0,
730 : _mm256_add_epi16(sum_q3, q256_2)),
731 : 3);
732 :
733 0 : flat_q2 = _mm256_castsi256_si128(
734 0 : _mm256_permute4x64_epi64(_mm256_packus_epi16(res_q, res_q), 168));
735 :
736 0 : sum_p7 = _mm256_add_epi16(sum_p7, p256_7);
737 :
738 0 : sum_q7 = _mm256_add_epi16(sum_q7, q256_7);
739 :
740 0 : pixelFilter_p = _mm256_sub_epi16(pixelFilter_p, q256_4);
741 :
742 0 : pixelFilter_q = _mm256_sub_epi16(pixelFilter_q, p256_4);
743 :
744 0 : res_p = _mm256_srli_epi16(
745 : _mm256_add_epi16(pixelFilter_p, _mm256_add_epi16(sum_p7, p256_3)), 4);
746 :
747 0 : flat2_p3 = _mm256_castsi256_si128(
748 0 : _mm256_permute4x64_epi64(_mm256_packus_epi16(res_p, res_p), 168));
749 :
750 0 : res_q = _mm256_srli_epi16(
751 : _mm256_add_epi16(pixelFilter_q, _mm256_add_epi16(sum_q7, q256_3)), 4);
752 :
753 0 : flat2_q3 = _mm256_castsi256_si128(
754 0 : _mm256_permute4x64_epi64(_mm256_packus_epi16(res_q, res_q), 168));
755 :
756 0 : sum_p7 = _mm256_add_epi16(sum_p7, p256_7);
757 :
758 0 : sum_q7 = _mm256_add_epi16(sum_q7, q256_7);
759 :
760 0 : pixelFilter_p = _mm256_sub_epi16(pixelFilter_p, q256_3);
761 :
762 0 : pixelFilter_q = _mm256_sub_epi16(pixelFilter_q, p256_3);
763 :
764 0 : res_p = _mm256_srli_epi16(
765 : _mm256_add_epi16(pixelFilter_p, _mm256_add_epi16(sum_p7, p256_4)), 4);
766 :
767 0 : flat2_p4 = _mm256_castsi256_si128(
768 0 : _mm256_permute4x64_epi64(_mm256_packus_epi16(res_p, res_p), 168));
769 :
770 0 : res_q = _mm256_srli_epi16(
771 : _mm256_add_epi16(pixelFilter_q, _mm256_add_epi16(sum_q7, q256_4)), 4);
772 :
773 0 : flat2_q4 = _mm256_castsi256_si128(
774 0 : _mm256_permute4x64_epi64(_mm256_packus_epi16(res_q, res_q), 168));
775 :
776 0 : sum_p7 = _mm256_add_epi16(sum_p7, p256_7);
777 :
778 0 : sum_q7 = _mm256_add_epi16(sum_q7, q256_7);
779 :
780 0 : pixelFilter_p = _mm256_sub_epi16(pixelFilter_p, q256_2);
781 :
782 0 : pixelFilter_q = _mm256_sub_epi16(pixelFilter_q, p256_2);
783 :
784 0 : res_p = _mm256_srli_epi16(
785 : _mm256_add_epi16(pixelFilter_p, _mm256_add_epi16(sum_p7, p256_5)), 4);
786 :
787 0 : flat2_p5 = _mm256_castsi256_si128(
788 0 : _mm256_permute4x64_epi64(_mm256_packus_epi16(res_p, res_p), 168));
789 :
790 0 : res_q = _mm256_srli_epi16(
791 : _mm256_add_epi16(pixelFilter_q, _mm256_add_epi16(sum_q7, q256_5)), 4);
792 :
793 0 : flat2_q5 = _mm256_castsi256_si128(
794 0 : _mm256_permute4x64_epi64(_mm256_packus_epi16(res_q, res_q), 168));
795 :
796 0 : sum_p7 = _mm256_add_epi16(sum_p7, p256_7);
797 :
798 0 : sum_q7 = _mm256_add_epi16(sum_q7, q256_7);
799 :
800 0 : pixelFilter_p = _mm256_sub_epi16(pixelFilter_p, q256_1);
801 :
802 0 : pixelFilter_q = _mm256_sub_epi16(pixelFilter_q, p256_1);
803 :
804 0 : res_p = _mm256_srli_epi16(
805 : _mm256_add_epi16(pixelFilter_p, _mm256_add_epi16(sum_p7, p256_6)), 4);
806 :
807 0 : flat2_p6 = _mm256_castsi256_si128(
808 0 : _mm256_permute4x64_epi64(_mm256_packus_epi16(res_p, res_p), 168));
809 :
810 0 : res_q = _mm256_srli_epi16(
811 : _mm256_add_epi16(pixelFilter_q, _mm256_add_epi16(sum_q7, q256_6)), 4);
812 :
813 0 : flat2_q6 = _mm256_castsi256_si128(
814 0 : _mm256_permute4x64_epi64(_mm256_packus_epi16(res_q, res_q), 168));
815 : }
816 :
817 : // wide flat
818 : // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
819 :
820 0 : p2 = _mm_andnot_si128(flat, p2);
821 0 : flat_p2 = _mm_and_si128(flat, flat_p2);
822 0 : p2 = _mm_or_si128(flat_p2, p2);
823 :
824 0 : p1 = _mm_andnot_si128(flat, ps1);
825 0 : flat_p1 = _mm_and_si128(flat, flat_p1);
826 0 : p1 = _mm_or_si128(flat_p1, p1);
827 :
828 0 : p0 = _mm_andnot_si128(flat, ps0);
829 0 : flat_p0 = _mm_and_si128(flat, flat_p0);
830 0 : p0 = _mm_or_si128(flat_p0, p0);
831 :
832 0 : q0 = _mm_andnot_si128(flat, qs0);
833 0 : flat_q0 = _mm_and_si128(flat, flat_q0);
834 0 : q0 = _mm_or_si128(flat_q0, q0);
835 :
836 0 : q1 = _mm_andnot_si128(flat, qs1);
837 0 : flat_q1 = _mm_and_si128(flat, flat_q1);
838 0 : q1 = _mm_or_si128(flat_q1, q1);
839 :
840 0 : q2 = _mm_andnot_si128(flat, q2);
841 0 : flat_q2 = _mm_and_si128(flat, flat_q2);
842 0 : q2 = _mm_or_si128(flat_q2, q2);
843 :
844 0 : p6 = _mm_andnot_si128(flat2, p6);
845 0 : flat2_p6 = _mm_and_si128(flat2, flat2_p6);
846 0 : p6 = _mm_or_si128(flat2_p6, p6);
847 0 : _mm_storeu_si128((__m128i *)(s - 7 * p), p6);
848 :
849 0 : p5 = _mm_andnot_si128(flat2, p5);
850 0 : flat2_p5 = _mm_and_si128(flat2, flat2_p5);
851 0 : p5 = _mm_or_si128(flat2_p5, p5);
852 0 : _mm_storeu_si128((__m128i *)(s - 6 * p), p5);
853 :
854 0 : p4 = _mm_andnot_si128(flat2, p4);
855 0 : flat2_p4 = _mm_and_si128(flat2, flat2_p4);
856 0 : p4 = _mm_or_si128(flat2_p4, p4);
857 0 : _mm_storeu_si128((__m128i *)(s - 5 * p), p4);
858 :
859 0 : p3 = _mm_andnot_si128(flat2, p3);
860 0 : flat2_p3 = _mm_and_si128(flat2, flat2_p3);
861 0 : p3 = _mm_or_si128(flat2_p3, p3);
862 0 : _mm_storeu_si128((__m128i *)(s - 4 * p), p3);
863 :
864 0 : p2 = _mm_andnot_si128(flat2, p2);
865 0 : flat2_p2 = _mm_and_si128(flat2, flat2_p2);
866 0 : p2 = _mm_or_si128(flat2_p2, p2);
867 0 : _mm_storeu_si128((__m128i *)(s - 3 * p), p2);
868 :
869 0 : p1 = _mm_andnot_si128(flat2, p1);
870 0 : flat2_p1 = _mm_and_si128(flat2, flat2_p1);
871 0 : p1 = _mm_or_si128(flat2_p1, p1);
872 0 : _mm_storeu_si128((__m128i *)(s - 2 * p), p1);
873 :
874 0 : p0 = _mm_andnot_si128(flat2, p0);
875 0 : flat2_p0 = _mm_and_si128(flat2, flat2_p0);
876 0 : p0 = _mm_or_si128(flat2_p0, p0);
877 0 : _mm_storeu_si128((__m128i *)(s - 1 * p), p0);
878 :
879 0 : q0 = _mm_andnot_si128(flat2, q0);
880 0 : flat2_q0 = _mm_and_si128(flat2, flat2_q0);
881 0 : q0 = _mm_or_si128(flat2_q0, q0);
882 : _mm_storeu_si128((__m128i *)(s - 0 * p), q0);
883 :
884 0 : q1 = _mm_andnot_si128(flat2, q1);
885 0 : flat2_q1 = _mm_and_si128(flat2, flat2_q1);
886 0 : q1 = _mm_or_si128(flat2_q1, q1);
887 0 : _mm_storeu_si128((__m128i *)(s + 1 * p), q1);
888 :
889 0 : q2 = _mm_andnot_si128(flat2, q2);
890 0 : flat2_q2 = _mm_and_si128(flat2, flat2_q2);
891 0 : q2 = _mm_or_si128(flat2_q2, q2);
892 0 : _mm_storeu_si128((__m128i *)(s + 2 * p), q2);
893 :
894 0 : q3 = _mm_andnot_si128(flat2, q3);
895 0 : flat2_q3 = _mm_and_si128(flat2, flat2_q3);
896 0 : q3 = _mm_or_si128(flat2_q3, q3);
897 0 : _mm_storeu_si128((__m128i *)(s + 3 * p), q3);
898 :
899 0 : q4 = _mm_andnot_si128(flat2, q4);
900 0 : flat2_q4 = _mm_and_si128(flat2, flat2_q4);
901 0 : q4 = _mm_or_si128(flat2_q4, q4);
902 0 : _mm_storeu_si128((__m128i *)(s + 4 * p), q4);
903 :
904 0 : q5 = _mm_andnot_si128(flat2, q5);
905 0 : flat2_q5 = _mm_and_si128(flat2, flat2_q5);
906 0 : q5 = _mm_or_si128(flat2_q5, q5);
907 0 : _mm_storeu_si128((__m128i *)(s + 5 * p), q5);
908 :
909 0 : q6 = _mm_andnot_si128(flat2, q6);
910 0 : flat2_q6 = _mm_and_si128(flat2, flat2_q6);
911 0 : q6 = _mm_or_si128(flat2_q6, q6);
912 0 : _mm_storeu_si128((__m128i *)(s + 6 * p), q6);
913 : }
914 : _mm256_zeroupper();
915 0 : }
|
