1 /*
2 * Copyright (c) 2010 The WebM project authors. All Rights Reserved.
3 *
4 * Use of this source code is governed by a BSD-style license
5 * that can be found in the LICENSE file in the root of the source
6 * tree. An additional intellectual property rights grant can be found
7 * in the file PATENTS. All contributing project authors may
8 * be found in the AUTHORS file in the root of the source tree.
9 */
10
11 #include "./vpx_config.h"
12 #include "vp9/common/vp9_common.h"
13 #include "vp9/common/vp9_loopfilter.h"
14 #include "vp9/common/vp9_onyxc_int.h"
15
signed_char_clamp(int t)16 static INLINE int8_t signed_char_clamp(int t) {
17 return (int8_t)clamp(t, -128, 127);
18 }
19
20 // should we apply any filter at all: 11111111 yes, 00000000 no
filter_mask(uint8_t limit,uint8_t blimit,uint8_t p3,uint8_t p2,uint8_t p1,uint8_t p0,uint8_t q0,uint8_t q1,uint8_t q2,uint8_t q3)21 static INLINE int8_t filter_mask(uint8_t limit, uint8_t blimit,
22 uint8_t p3, uint8_t p2,
23 uint8_t p1, uint8_t p0,
24 uint8_t q0, uint8_t q1,
25 uint8_t q2, uint8_t q3) {
26 int8_t mask = 0;
27 mask |= (abs(p3 - p2) > limit) * -1;
28 mask |= (abs(p2 - p1) > limit) * -1;
29 mask |= (abs(p1 - p0) > limit) * -1;
30 mask |= (abs(q1 - q0) > limit) * -1;
31 mask |= (abs(q2 - q1) > limit) * -1;
32 mask |= (abs(q3 - q2) > limit) * -1;
33 mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2 > blimit) * -1;
34 return ~mask;
35 }
36
flat_mask4(uint8_t thresh,uint8_t p3,uint8_t p2,uint8_t p1,uint8_t p0,uint8_t q0,uint8_t q1,uint8_t q2,uint8_t q3)37 static INLINE int8_t flat_mask4(uint8_t thresh,
38 uint8_t p3, uint8_t p2,
39 uint8_t p1, uint8_t p0,
40 uint8_t q0, uint8_t q1,
41 uint8_t q2, uint8_t q3) {
42 int8_t mask = 0;
43 mask |= (abs(p1 - p0) > thresh) * -1;
44 mask |= (abs(q1 - q0) > thresh) * -1;
45 mask |= (abs(p2 - p0) > thresh) * -1;
46 mask |= (abs(q2 - q0) > thresh) * -1;
47 mask |= (abs(p3 - p0) > thresh) * -1;
48 mask |= (abs(q3 - q0) > thresh) * -1;
49 return ~mask;
50 }
51
flat_mask5(uint8_t thresh,uint8_t p4,uint8_t p3,uint8_t p2,uint8_t p1,uint8_t p0,uint8_t q0,uint8_t q1,uint8_t q2,uint8_t q3,uint8_t q4)52 static INLINE int8_t flat_mask5(uint8_t thresh,
53 uint8_t p4, uint8_t p3,
54 uint8_t p2, uint8_t p1,
55 uint8_t p0, uint8_t q0,
56 uint8_t q1, uint8_t q2,
57 uint8_t q3, uint8_t q4) {
58 int8_t mask = ~flat_mask4(thresh, p3, p2, p1, p0, q0, q1, q2, q3);
59 mask |= (abs(p4 - p0) > thresh) * -1;
60 mask |= (abs(q4 - q0) > thresh) * -1;
61 return ~mask;
62 }
63
64 // is there high edge variance internal edge: 11111111 yes, 00000000 no
hev_mask(uint8_t thresh,uint8_t p1,uint8_t p0,uint8_t q0,uint8_t q1)65 static INLINE int8_t hev_mask(uint8_t thresh, uint8_t p1, uint8_t p0,
66 uint8_t q0, uint8_t q1) {
67 int8_t hev = 0;
68 hev |= (abs(p1 - p0) > thresh) * -1;
69 hev |= (abs(q1 - q0) > thresh) * -1;
70 return hev;
71 }
72
filter4(int8_t mask,uint8_t thresh,uint8_t * op1,uint8_t * op0,uint8_t * oq0,uint8_t * oq1)73 static INLINE void filter4(int8_t mask, uint8_t thresh, uint8_t *op1,
74 uint8_t *op0, uint8_t *oq0, uint8_t *oq1) {
75 int8_t filter1, filter2;
76
77 const int8_t ps1 = (int8_t) *op1 ^ 0x80;
78 const int8_t ps0 = (int8_t) *op0 ^ 0x80;
79 const int8_t qs0 = (int8_t) *oq0 ^ 0x80;
80 const int8_t qs1 = (int8_t) *oq1 ^ 0x80;
81 const uint8_t hev = hev_mask(thresh, *op1, *op0, *oq0, *oq1);
82
83 // add outer taps if we have high edge variance
84 int8_t filter = signed_char_clamp(ps1 - qs1) & hev;
85
86 // inner taps
87 filter = signed_char_clamp(filter + 3 * (qs0 - ps0)) & mask;
88
89 // save bottom 3 bits so that we round one side +4 and the other +3
90 // if it equals 4 we'll set to adjust by -1 to account for the fact
91 // we'd round 3 the other way
92 filter1 = signed_char_clamp(filter + 4) >> 3;
93 filter2 = signed_char_clamp(filter + 3) >> 3;
94
95 *oq0 = signed_char_clamp(qs0 - filter1) ^ 0x80;
96 *op0 = signed_char_clamp(ps0 + filter2) ^ 0x80;
97
98 // outer tap adjustments
99 filter = ROUND_POWER_OF_TWO(filter1, 1) & ~hev;
100
101 *oq1 = signed_char_clamp(qs1 - filter) ^ 0x80;
102 *op1 = signed_char_clamp(ps1 + filter) ^ 0x80;
103 }
104
vp9_lpf_horizontal_4_c(uint8_t * s,int p,const uint8_t * blimit,const uint8_t * limit,const uint8_t * thresh,int count)105 void vp9_lpf_horizontal_4_c(uint8_t *s, int p /* pitch */,
106 const uint8_t *blimit, const uint8_t *limit,
107 const uint8_t *thresh, int count) {
108 int i;
109
110 // loop filter designed to work using chars so that we can make maximum use
111 // of 8 bit simd instructions.
112 for (i = 0; i < 8 * count; ++i) {
113 const uint8_t p3 = s[-4 * p], p2 = s[-3 * p], p1 = s[-2 * p], p0 = s[-p];
114 const uint8_t q0 = s[0 * p], q1 = s[1 * p], q2 = s[2 * p], q3 = s[3 * p];
115 const int8_t mask = filter_mask(*limit, *blimit,
116 p3, p2, p1, p0, q0, q1, q2, q3);
117 filter4(mask, *thresh, s - 2 * p, s - 1 * p, s, s + 1 * p);
118 ++s;
119 }
120 }
121
vp9_lpf_horizontal_4_dual_c(uint8_t * s,int p,const uint8_t * blimit0,const uint8_t * limit0,const uint8_t * thresh0,const uint8_t * blimit1,const uint8_t * limit1,const uint8_t * thresh1)122 void vp9_lpf_horizontal_4_dual_c(uint8_t *s, int p, const uint8_t *blimit0,
123 const uint8_t *limit0, const uint8_t *thresh0,
124 const uint8_t *blimit1, const uint8_t *limit1,
125 const uint8_t *thresh1) {
126 vp9_lpf_horizontal_4_c(s, p, blimit0, limit0, thresh0, 1);
127 vp9_lpf_horizontal_4_c(s + 8, p, blimit1, limit1, thresh1, 1);
128 }
129
vp9_lpf_vertical_4_c(uint8_t * s,int pitch,const uint8_t * blimit,const uint8_t * limit,const uint8_t * thresh,int count)130 void vp9_lpf_vertical_4_c(uint8_t *s, int pitch, const uint8_t *blimit,
131 const uint8_t *limit, const uint8_t *thresh,
132 int count) {
133 int i;
134
135 // loop filter designed to work using chars so that we can make maximum use
136 // of 8 bit simd instructions.
137 for (i = 0; i < 8 * count; ++i) {
138 const uint8_t p3 = s[-4], p2 = s[-3], p1 = s[-2], p0 = s[-1];
139 const uint8_t q0 = s[0], q1 = s[1], q2 = s[2], q3 = s[3];
140 const int8_t mask = filter_mask(*limit, *blimit,
141 p3, p2, p1, p0, q0, q1, q2, q3);
142 filter4(mask, *thresh, s - 2, s - 1, s, s + 1);
143 s += pitch;
144 }
145 }
146
vp9_lpf_vertical_4_dual_c(uint8_t * s,int pitch,const uint8_t * blimit0,const uint8_t * limit0,const uint8_t * thresh0,const uint8_t * blimit1,const uint8_t * limit1,const uint8_t * thresh1)147 void vp9_lpf_vertical_4_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0,
148 const uint8_t *limit0, const uint8_t *thresh0,
149 const uint8_t *blimit1, const uint8_t *limit1,
150 const uint8_t *thresh1) {
151 vp9_lpf_vertical_4_c(s, pitch, blimit0, limit0, thresh0, 1);
152 vp9_lpf_vertical_4_c(s + 8 * pitch, pitch, blimit1, limit1,
153 thresh1, 1);
154 }
155
filter8(int8_t mask,uint8_t thresh,uint8_t flat,uint8_t * op3,uint8_t * op2,uint8_t * op1,uint8_t * op0,uint8_t * oq0,uint8_t * oq1,uint8_t * oq2,uint8_t * oq3)156 static INLINE void filter8(int8_t mask, uint8_t thresh, uint8_t flat,
157 uint8_t *op3, uint8_t *op2,
158 uint8_t *op1, uint8_t *op0,
159 uint8_t *oq0, uint8_t *oq1,
160 uint8_t *oq2, uint8_t *oq3) {
161 if (flat && mask) {
162 const uint8_t p3 = *op3, p2 = *op2, p1 = *op1, p0 = *op0;
163 const uint8_t q0 = *oq0, q1 = *oq1, q2 = *oq2, q3 = *oq3;
164
165 // 7-tap filter [1, 1, 1, 2, 1, 1, 1]
166 *op2 = ROUND_POWER_OF_TWO(p3 + p3 + p3 + 2 * p2 + p1 + p0 + q0, 3);
167 *op1 = ROUND_POWER_OF_TWO(p3 + p3 + p2 + 2 * p1 + p0 + q0 + q1, 3);
168 *op0 = ROUND_POWER_OF_TWO(p3 + p2 + p1 + 2 * p0 + q0 + q1 + q2, 3);
169 *oq0 = ROUND_POWER_OF_TWO(p2 + p1 + p0 + 2 * q0 + q1 + q2 + q3, 3);
170 *oq1 = ROUND_POWER_OF_TWO(p1 + p0 + q0 + 2 * q1 + q2 + q3 + q3, 3);
171 *oq2 = ROUND_POWER_OF_TWO(p0 + q0 + q1 + 2 * q2 + q3 + q3 + q3, 3);
172 } else {
173 filter4(mask, thresh, op1, op0, oq0, oq1);
174 }
175 }
176
vp9_lpf_horizontal_8_c(uint8_t * s,int p,const uint8_t * blimit,const uint8_t * limit,const uint8_t * thresh,int count)177 void vp9_lpf_horizontal_8_c(uint8_t *s, int p, const uint8_t *blimit,
178 const uint8_t *limit, const uint8_t *thresh,
179 int count) {
180 int i;
181
182 // loop filter designed to work using chars so that we can make maximum use
183 // of 8 bit simd instructions.
184 for (i = 0; i < 8 * count; ++i) {
185 const uint8_t p3 = s[-4 * p], p2 = s[-3 * p], p1 = s[-2 * p], p0 = s[-p];
186 const uint8_t q0 = s[0 * p], q1 = s[1 * p], q2 = s[2 * p], q3 = s[3 * p];
187
188 const int8_t mask = filter_mask(*limit, *blimit,
189 p3, p2, p1, p0, q0, q1, q2, q3);
190 const int8_t flat = flat_mask4(1, p3, p2, p1, p0, q0, q1, q2, q3);
191 filter8(mask, *thresh, flat, s - 4 * p, s - 3 * p, s - 2 * p, s - 1 * p,
192 s, s + 1 * p, s + 2 * p, s + 3 * p);
193 ++s;
194 }
195 }
196
vp9_lpf_horizontal_8_dual_c(uint8_t * s,int p,const uint8_t * blimit0,const uint8_t * limit0,const uint8_t * thresh0,const uint8_t * blimit1,const uint8_t * limit1,const uint8_t * thresh1)197 void vp9_lpf_horizontal_8_dual_c(uint8_t *s, int p, const uint8_t *blimit0,
198 const uint8_t *limit0, const uint8_t *thresh0,
199 const uint8_t *blimit1, const uint8_t *limit1,
200 const uint8_t *thresh1) {
201 vp9_lpf_horizontal_8_c(s, p, blimit0, limit0, thresh0, 1);
202 vp9_lpf_horizontal_8_c(s + 8, p, blimit1, limit1, thresh1, 1);
203 }
204
vp9_lpf_vertical_8_c(uint8_t * s,int pitch,const uint8_t * blimit,const uint8_t * limit,const uint8_t * thresh,int count)205 void vp9_lpf_vertical_8_c(uint8_t *s, int pitch, const uint8_t *blimit,
206 const uint8_t *limit, const uint8_t *thresh,
207 int count) {
208 int i;
209
210 for (i = 0; i < 8 * count; ++i) {
211 const uint8_t p3 = s[-4], p2 = s[-3], p1 = s[-2], p0 = s[-1];
212 const uint8_t q0 = s[0], q1 = s[1], q2 = s[2], q3 = s[3];
213 const int8_t mask = filter_mask(*limit, *blimit,
214 p3, p2, p1, p0, q0, q1, q2, q3);
215 const int8_t flat = flat_mask4(1, p3, p2, p1, p0, q0, q1, q2, q3);
216 filter8(mask, *thresh, flat, s - 4, s - 3, s - 2, s - 1,
217 s, s + 1, s + 2, s + 3);
218 s += pitch;
219 }
220 }
221
vp9_lpf_vertical_8_dual_c(uint8_t * s,int pitch,const uint8_t * blimit0,const uint8_t * limit0,const uint8_t * thresh0,const uint8_t * blimit1,const uint8_t * limit1,const uint8_t * thresh1)222 void vp9_lpf_vertical_8_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0,
223 const uint8_t *limit0, const uint8_t *thresh0,
224 const uint8_t *blimit1, const uint8_t *limit1,
225 const uint8_t *thresh1) {
226 vp9_lpf_vertical_8_c(s, pitch, blimit0, limit0, thresh0, 1);
227 vp9_lpf_vertical_8_c(s + 8 * pitch, pitch, blimit1, limit1,
228 thresh1, 1);
229 }
230
filter16(int8_t mask,uint8_t thresh,uint8_t flat,uint8_t flat2,uint8_t * op7,uint8_t * op6,uint8_t * op5,uint8_t * op4,uint8_t * op3,uint8_t * op2,uint8_t * op1,uint8_t * op0,uint8_t * oq0,uint8_t * oq1,uint8_t * oq2,uint8_t * oq3,uint8_t * oq4,uint8_t * oq5,uint8_t * oq6,uint8_t * oq7)231 static INLINE void filter16(int8_t mask, uint8_t thresh,
232 uint8_t flat, uint8_t flat2,
233 uint8_t *op7, uint8_t *op6,
234 uint8_t *op5, uint8_t *op4,
235 uint8_t *op3, uint8_t *op2,
236 uint8_t *op1, uint8_t *op0,
237 uint8_t *oq0, uint8_t *oq1,
238 uint8_t *oq2, uint8_t *oq3,
239 uint8_t *oq4, uint8_t *oq5,
240 uint8_t *oq6, uint8_t *oq7) {
241 if (flat2 && flat && mask) {
242 const uint8_t p7 = *op7, p6 = *op6, p5 = *op5, p4 = *op4,
243 p3 = *op3, p2 = *op2, p1 = *op1, p0 = *op0;
244
245 const uint8_t q0 = *oq0, q1 = *oq1, q2 = *oq2, q3 = *oq3,
246 q4 = *oq4, q5 = *oq5, q6 = *oq6, q7 = *oq7;
247
248 // 15-tap filter [1, 1, 1, 1, 1, 1, 1, 2, 1, 1, 1, 1, 1, 1, 1]
249 *op6 = ROUND_POWER_OF_TWO(p7 * 7 + p6 * 2 + p5 + p4 + p3 + p2 + p1 + p0 +
250 q0, 4);
251 *op5 = ROUND_POWER_OF_TWO(p7 * 6 + p6 + p5 * 2 + p4 + p3 + p2 + p1 + p0 +
252 q0 + q1, 4);
253 *op4 = ROUND_POWER_OF_TWO(p7 * 5 + p6 + p5 + p4 * 2 + p3 + p2 + p1 + p0 +
254 q0 + q1 + q2, 4);
255 *op3 = ROUND_POWER_OF_TWO(p7 * 4 + p6 + p5 + p4 + p3 * 2 + p2 + p1 + p0 +
256 q0 + q1 + q2 + q3, 4);
257 *op2 = ROUND_POWER_OF_TWO(p7 * 3 + p6 + p5 + p4 + p3 + p2 * 2 + p1 + p0 +
258 q0 + q1 + q2 + q3 + q4, 4);
259 *op1 = ROUND_POWER_OF_TWO(p7 * 2 + p6 + p5 + p4 + p3 + p2 + p1 * 2 + p0 +
260 q0 + q1 + q2 + q3 + q4 + q5, 4);
261 *op0 = ROUND_POWER_OF_TWO(p7 + p6 + p5 + p4 + p3 + p2 + p1 + p0 * 2 +
262 q0 + q1 + q2 + q3 + q4 + q5 + q6, 4);
263 *oq0 = ROUND_POWER_OF_TWO(p6 + p5 + p4 + p3 + p2 + p1 + p0 +
264 q0 * 2 + q1 + q2 + q3 + q4 + q5 + q6 + q7, 4);
265 *oq1 = ROUND_POWER_OF_TWO(p5 + p4 + p3 + p2 + p1 + p0 +
266 q0 + q1 * 2 + q2 + q3 + q4 + q5 + q6 + q7 * 2, 4);
267 *oq2 = ROUND_POWER_OF_TWO(p4 + p3 + p2 + p1 + p0 +
268 q0 + q1 + q2 * 2 + q3 + q4 + q5 + q6 + q7 * 3, 4);
269 *oq3 = ROUND_POWER_OF_TWO(p3 + p2 + p1 + p0 +
270 q0 + q1 + q2 + q3 * 2 + q4 + q5 + q6 + q7 * 4, 4);
271 *oq4 = ROUND_POWER_OF_TWO(p2 + p1 + p0 +
272 q0 + q1 + q2 + q3 + q4 * 2 + q5 + q6 + q7 * 5, 4);
273 *oq5 = ROUND_POWER_OF_TWO(p1 + p0 +
274 q0 + q1 + q2 + q3 + q4 + q5 * 2 + q6 + q7 * 6, 4);
275 *oq6 = ROUND_POWER_OF_TWO(p0 +
276 q0 + q1 + q2 + q3 + q4 + q5 + q6 * 2 + q7 * 7, 4);
277 } else {
278 filter8(mask, thresh, flat, op3, op2, op1, op0, oq0, oq1, oq2, oq3);
279 }
280 }
281
vp9_lpf_horizontal_16_c(uint8_t * s,int p,const uint8_t * blimit,const uint8_t * limit,const uint8_t * thresh,int count)282 void vp9_lpf_horizontal_16_c(uint8_t *s, int p, const uint8_t *blimit,
283 const uint8_t *limit, const uint8_t *thresh,
284 int count) {
285 int i;
286
287 // loop filter designed to work using chars so that we can make maximum use
288 // of 8 bit simd instructions.
289 for (i = 0; i < 8 * count; ++i) {
290 const uint8_t p3 = s[-4 * p], p2 = s[-3 * p], p1 = s[-2 * p], p0 = s[-p];
291 const uint8_t q0 = s[0 * p], q1 = s[1 * p], q2 = s[2 * p], q3 = s[3 * p];
292 const int8_t mask = filter_mask(*limit, *blimit,
293 p3, p2, p1, p0, q0, q1, q2, q3);
294 const int8_t flat = flat_mask4(1, p3, p2, p1, p0, q0, q1, q2, q3);
295 const int8_t flat2 = flat_mask5(1,
296 s[-8 * p], s[-7 * p], s[-6 * p], s[-5 * p], p0,
297 q0, s[4 * p], s[5 * p], s[6 * p], s[7 * p]);
298
299 filter16(mask, *thresh, flat, flat2,
300 s - 8 * p, s - 7 * p, s - 6 * p, s - 5 * p,
301 s - 4 * p, s - 3 * p, s - 2 * p, s - 1 * p,
302 s, s + 1 * p, s + 2 * p, s + 3 * p,
303 s + 4 * p, s + 5 * p, s + 6 * p, s + 7 * p);
304 ++s;
305 }
306 }
307
mb_lpf_vertical_edge_w(uint8_t * s,int p,const uint8_t * blimit,const uint8_t * limit,const uint8_t * thresh,int count)308 static void mb_lpf_vertical_edge_w(uint8_t *s, int p,
309 const uint8_t *blimit,
310 const uint8_t *limit,
311 const uint8_t *thresh,
312 int count) {
313 int i;
314
315 for (i = 0; i < count; ++i) {
316 const uint8_t p3 = s[-4], p2 = s[-3], p1 = s[-2], p0 = s[-1];
317 const uint8_t q0 = s[0], q1 = s[1], q2 = s[2], q3 = s[3];
318 const int8_t mask = filter_mask(*limit, *blimit,
319 p3, p2, p1, p0, q0, q1, q2, q3);
320 const int8_t flat = flat_mask4(1, p3, p2, p1, p0, q0, q1, q2, q3);
321 const int8_t flat2 = flat_mask5(1, s[-8], s[-7], s[-6], s[-5], p0,
322 q0, s[4], s[5], s[6], s[7]);
323
324 filter16(mask, *thresh, flat, flat2,
325 s - 8, s - 7, s - 6, s - 5, s - 4, s - 3, s - 2, s - 1,
326 s, s + 1, s + 2, s + 3, s + 4, s + 5, s + 6, s + 7);
327 s += p;
328 }
329 }
330
vp9_lpf_vertical_16_c(uint8_t * s,int p,const uint8_t * blimit,const uint8_t * limit,const uint8_t * thresh)331 void vp9_lpf_vertical_16_c(uint8_t *s, int p, const uint8_t *blimit,
332 const uint8_t *limit, const uint8_t *thresh) {
333 mb_lpf_vertical_edge_w(s, p, blimit, limit, thresh, 8);
334 }
335
vp9_lpf_vertical_16_dual_c(uint8_t * s,int p,const uint8_t * blimit,const uint8_t * limit,const uint8_t * thresh)336 void vp9_lpf_vertical_16_dual_c(uint8_t *s, int p, const uint8_t *blimit,
337 const uint8_t *limit, const uint8_t *thresh) {
338 mb_lpf_vertical_edge_w(s, p, blimit, limit, thresh, 16);
339 }
340