2
* Copyright (c) 2010 The WebM project authors. All Rights Reserved.
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.
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"
16
static INLINE int8_t signed_char_clamp(int t) {
17
return (int8_t)clamp(t, -128, 127);
20
// should we apply any filter at all: 11111111 yes, 00000000 no
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) {
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;
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) {
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;
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;
64
// is there high edge variance internal edge: 11111111 yes, 00000000 no
65
static INLINE int8_t hev_mask(uint8_t thresh, uint8_t p1, uint8_t p0,
66
uint8_t q0, uint8_t q1) {
68
hev |= (abs(p1 - p0) > thresh) * -1;
69
hev |= (abs(q1 - q0) > thresh) * -1;
73
static INLINE void filter4(int8_t mask, uint8_t hev, uint8_t *op1,
74
uint8_t *op0, uint8_t *oq0, uint8_t *oq1) {
75
int8_t filter1, filter2;
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;
82
// add outer taps if we have high edge variance
83
int8_t filter = signed_char_clamp(ps1 - qs1) & hev;
86
filter = signed_char_clamp(filter + 3 * (qs0 - ps0)) & mask;
88
// save bottom 3 bits so that we round one side +4 and the other +3
89
// if it equals 4 we'll set to adjust by -1 to account for the fact
90
// we'd round 3 the other way
91
filter1 = signed_char_clamp(filter + 4) >> 3;
92
filter2 = signed_char_clamp(filter + 3) >> 3;
94
*oq0 = signed_char_clamp(qs0 - filter1) ^ 0x80;
95
*op0 = signed_char_clamp(ps0 + filter2) ^ 0x80;
97
// outer tap adjustments
98
filter = ROUND_POWER_OF_TWO(filter1, 1) & ~hev;
100
*oq1 = signed_char_clamp(qs1 - filter) ^ 0x80;
101
*op1 = signed_char_clamp(ps1 + filter) ^ 0x80;
104
void vp9_loop_filter_horizontal_edge_c(uint8_t *s, int p /* pitch */,
105
const uint8_t *blimit,
106
const uint8_t *limit,
107
const uint8_t *thresh,
111
// loop filter designed to work using chars so that we can make maximum use
112
// of 8 bit simd instructions.
113
for (i = 0; i < 8 * count; ++i) {
114
const uint8_t p3 = s[-4 * p], p2 = s[-3 * p], p1 = s[-2 * p], p0 = s[-p];
115
const uint8_t q0 = s[0 * p], q1 = s[1 * p], q2 = s[2 * p], q3 = s[3 * p];
116
const int8_t mask = filter_mask(*limit, *blimit,
117
p3, p2, p1, p0, q0, q1, q2, q3);
118
const int8_t hev = hev_mask(*thresh, p1, p0, q0, q1);
119
filter4(mask, hev, s - 2 * p, s - 1 * p, s, s + 1 * p);
124
void vp9_loop_filter_vertical_edge_c(uint8_t *s, int pitch,
125
const uint8_t *blimit,
126
const uint8_t *limit,
127
const uint8_t *thresh,
131
// loop filter designed to work using chars so that we can make maximum use
132
// of 8 bit simd instructions.
133
for (i = 0; i < 8 * count; ++i) {
134
const uint8_t p3 = s[-4], p2 = s[-3], p1 = s[-2], p0 = s[-1];
135
const uint8_t q0 = s[0], q1 = s[1], q2 = s[2], q3 = s[3];
136
const int8_t mask = filter_mask(*limit, *blimit,
137
p3, p2, p1, p0, q0, q1, q2, q3);
138
const int8_t hev = hev_mask(*thresh, p1, p0, q0, q1);
139
filter4(mask, hev, s - 2, s - 1, s, s + 1);
144
static INLINE void filter8(int8_t mask, uint8_t hev, uint8_t flat,
145
uint8_t *op3, uint8_t *op2,
146
uint8_t *op1, uint8_t *op0,
147
uint8_t *oq0, uint8_t *oq1,
148
uint8_t *oq2, uint8_t *oq3) {
150
const uint8_t p3 = *op3, p2 = *op2, p1 = *op1, p0 = *op0;
151
const uint8_t q0 = *oq0, q1 = *oq1, q2 = *oq2, q3 = *oq3;
153
// 7-tap filter [1, 1, 1, 2, 1, 1, 1]
154
*op2 = ROUND_POWER_OF_TWO(p3 + p3 + p3 + 2 * p2 + p1 + p0 + q0, 3);
155
*op1 = ROUND_POWER_OF_TWO(p3 + p3 + p2 + 2 * p1 + p0 + q0 + q1, 3);
156
*op0 = ROUND_POWER_OF_TWO(p3 + p2 + p1 + 2 * p0 + q0 + q1 + q2, 3);
157
*oq0 = ROUND_POWER_OF_TWO(p2 + p1 + p0 + 2 * q0 + q1 + q2 + q3, 3);
158
*oq1 = ROUND_POWER_OF_TWO(p1 + p0 + q0 + 2 * q1 + q2 + q3 + q3, 3);
159
*oq2 = ROUND_POWER_OF_TWO(p0 + q0 + q1 + 2 * q2 + q3 + q3 + q3, 3);
161
filter4(mask, hev, op1, op0, oq0, oq1);
165
void vp9_mbloop_filter_horizontal_edge_c(uint8_t *s, int p,
166
const uint8_t *blimit,
167
const uint8_t *limit,
168
const uint8_t *thresh,
172
// loop filter designed to work using chars so that we can make maximum use
173
// of 8 bit simd instructions.
174
for (i = 0; i < 8 * count; ++i) {
175
const uint8_t p3 = s[-4 * p], p2 = s[-3 * p], p1 = s[-2 * p], p0 = s[-p];
176
const uint8_t q0 = s[0 * p], q1 = s[1 * p], q2 = s[2 * p], q3 = s[3 * p];
178
const int8_t mask = filter_mask(*limit, *blimit,
179
p3, p2, p1, p0, q0, q1, q2, q3);
180
const int8_t hev = hev_mask(*thresh, p1, p0, q0, q1);
181
const int8_t flat = flat_mask4(1, p3, p2, p1, p0, q0, q1, q2, q3);
182
filter8(mask, hev, flat, s - 4 * p, s - 3 * p, s - 2 * p, s - 1 * p,
183
s, s + 1 * p, s + 2 * p, s + 3 * p);
188
void vp9_mbloop_filter_vertical_edge_c(uint8_t *s, int pitch,
189
const uint8_t *blimit,
190
const uint8_t *limit,
191
const uint8_t *thresh,
195
for (i = 0; i < 8 * count; ++i) {
196
const uint8_t p3 = s[-4], p2 = s[-3], p1 = s[-2], p0 = s[-1];
197
const uint8_t q0 = s[0], q1 = s[1], q2 = s[2], q3 = s[3];
198
const int8_t mask = filter_mask(*limit, *blimit,
199
p3, p2, p1, p0, q0, q1, q2, q3);
200
const int8_t hev = hev_mask(thresh[0], p1, p0, q0, q1);
201
const int8_t flat = flat_mask4(1, p3, p2, p1, p0, q0, q1, q2, q3);
202
filter8(mask, hev, flat, s - 4, s - 3, s - 2, s - 1,
203
s, s + 1, s + 2, s + 3);
208
static INLINE void filter16(int8_t mask, uint8_t hev,
209
uint8_t flat, uint8_t flat2,
210
uint8_t *op7, uint8_t *op6,
211
uint8_t *op5, uint8_t *op4,
212
uint8_t *op3, uint8_t *op2,
213
uint8_t *op1, uint8_t *op0,
214
uint8_t *oq0, uint8_t *oq1,
215
uint8_t *oq2, uint8_t *oq3,
216
uint8_t *oq4, uint8_t *oq5,
217
uint8_t *oq6, uint8_t *oq7) {
218
if (flat2 && flat && mask) {
219
const uint8_t p7 = *op7, p6 = *op6, p5 = *op5, p4 = *op4,
220
p3 = *op3, p2 = *op2, p1 = *op1, p0 = *op0;
222
const uint8_t q0 = *oq0, q1 = *oq1, q2 = *oq2, q3 = *oq3,
223
q4 = *oq4, q5 = *oq5, q6 = *oq6, q7 = *oq7;
225
// 15-tap filter [1, 1, 1, 1, 1, 1, 1, 2, 1, 1, 1, 1, 1, 1, 1]
226
*op6 = ROUND_POWER_OF_TWO(p7 * 7 + p6 * 2 + p5 + p4 + p3 + p2 + p1 + p0 +
228
*op5 = ROUND_POWER_OF_TWO(p7 * 6 + p6 + p5 * 2 + p4 + p3 + p2 + p1 + p0 +
230
*op4 = ROUND_POWER_OF_TWO(p7 * 5 + p6 + p5 + p4 * 2 + p3 + p2 + p1 + p0 +
232
*op3 = ROUND_POWER_OF_TWO(p7 * 4 + p6 + p5 + p4 + p3 * 2 + p2 + p1 + p0 +
233
q0 + q1 + q2 + q3, 4);
234
*op2 = ROUND_POWER_OF_TWO(p7 * 3 + p6 + p5 + p4 + p3 + p2 * 2 + p1 + p0 +
235
q0 + q1 + q2 + q3 + q4, 4);
236
*op1 = ROUND_POWER_OF_TWO(p7 * 2 + p6 + p5 + p4 + p3 + p2 + p1 * 2 + p0 +
237
q0 + q1 + q2 + q3 + q4 + q5, 4);
238
*op0 = ROUND_POWER_OF_TWO(p7 + p6 + p5 + p4 + p3 + p2 + p1 + p0 * 2 +
239
q0 + q1 + q2 + q3 + q4 + q5 + q6, 4);
240
*oq0 = ROUND_POWER_OF_TWO(p6 + p5 + p4 + p3 + p2 + p1 + p0 +
241
q0 * 2 + q1 + q2 + q3 + q4 + q5 + q6 + q7, 4);
242
*oq1 = ROUND_POWER_OF_TWO(p5 + p4 + p3 + p2 + p1 + p0 +
243
q0 + q1 * 2 + q2 + q3 + q4 + q5 + q6 + q7 * 2, 4);
244
*oq2 = ROUND_POWER_OF_TWO(p4 + p3 + p2 + p1 + p0 +
245
q0 + q1 + q2 * 2 + q3 + q4 + q5 + q6 + q7 * 3, 4);
246
*oq3 = ROUND_POWER_OF_TWO(p3 + p2 + p1 + p0 +
247
q0 + q1 + q2 + q3 * 2 + q4 + q5 + q6 + q7 * 4, 4);
248
*oq4 = ROUND_POWER_OF_TWO(p2 + p1 + p0 +
249
q0 + q1 + q2 + q3 + q4 * 2 + q5 + q6 + q7 * 5, 4);
250
*oq5 = ROUND_POWER_OF_TWO(p1 + p0 +
251
q0 + q1 + q2 + q3 + q4 + q5 * 2 + q6 + q7 * 6, 4);
252
*oq6 = ROUND_POWER_OF_TWO(p0 +
253
q0 + q1 + q2 + q3 + q4 + q5 + q6 * 2 + q7 * 7, 4);
255
filter8(mask, hev, flat, op3, op2, op1, op0, oq0, oq1, oq2, oq3);
259
void vp9_mb_lpf_horizontal_edge_w_c(uint8_t *s, int p,
260
const uint8_t *blimit,
261
const uint8_t *limit,
262
const uint8_t *thresh,
266
// loop filter designed to work using chars so that we can make maximum use
267
// of 8 bit simd instructions.
268
for (i = 0; i < 8 * count; ++i) {
269
const uint8_t p3 = s[-4 * p], p2 = s[-3 * p], p1 = s[-2 * p], p0 = s[-p];
270
const uint8_t q0 = s[0 * p], q1 = s[1 * p], q2 = s[2 * p], q3 = s[3 * p];
271
const int8_t mask = filter_mask(*limit, *blimit,
272
p3, p2, p1, p0, q0, q1, q2, q3);
273
const int8_t hev = hev_mask(*thresh, p1, p0, q0, q1);
274
const int8_t flat = flat_mask4(1, p3, p2, p1, p0, q0, q1, q2, q3);
275
const int8_t flat2 = flat_mask5(1,
276
s[-8 * p], s[-7 * p], s[-6 * p], s[-5 * p], p0,
277
q0, s[4 * p], s[5 * p], s[6 * p], s[7 * p]);
279
filter16(mask, hev, flat, flat2,
280
s - 8 * p, s - 7 * p, s - 6 * p, s - 5 * p,
281
s - 4 * p, s - 3 * p, s - 2 * p, s - 1 * p,
282
s, s + 1 * p, s + 2 * p, s + 3 * p,
283
s + 4 * p, s + 5 * p, s + 6 * p, s + 7 * p);
288
void vp9_mb_lpf_vertical_edge_w_c(uint8_t *s, int p,
289
const uint8_t *blimit,
290
const uint8_t *limit,
291
const uint8_t *thresh) {
294
for (i = 0; i < 8; ++i) {
295
const uint8_t p3 = s[-4], p2 = s[-3], p1 = s[-2], p0 = s[-1];
296
const uint8_t q0 = s[0], q1 = s[1], q2 = s[2], q3 = s[3];
297
const int8_t mask = filter_mask(*limit, *blimit,
298
p3, p2, p1, p0, q0, q1, q2, q3);
299
const int8_t hev = hev_mask(*thresh, p1, p0, q0, q1);
300
const int8_t flat = flat_mask4(1, p3, p2, p1, p0, q0, q1, q2, q3);
301
const int8_t flat2 = flat_mask5(1, s[-8], s[-7], s[-6], s[-5], p0,
302
q0, s[4], s[5], s[6], s[7]);
304
filter16(mask, hev, flat, flat2,
305
s - 8, s - 7, s - 6, s - 5, s - 4, s - 3, s - 2, s - 1,
306
s, s + 1, s + 2, s + 3, s + 4, s + 5, s + 6, s + 7);