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 <math.h>
13 #include <stdlib.h>
14 
15 #include "config/aom_dsp_rtcd.h"
16 #include "config/av1_rtcd.h"
17 
18 #include "av1/common/cdef.h"
19 
20 /* Generated from gen_filter_tables.c. */
21 DECLARE_ALIGNED(16, const int, cdef_directions[8][2]) = {
22   { -1 * CDEF_BSTRIDE + 1, -2 * CDEF_BSTRIDE + 2 },
23   { 0 * CDEF_BSTRIDE + 1, -1 * CDEF_BSTRIDE + 2 },
24   { 0 * CDEF_BSTRIDE + 1, 0 * CDEF_BSTRIDE + 2 },
25   { 0 * CDEF_BSTRIDE + 1, 1 * CDEF_BSTRIDE + 2 },
26   { 1 * CDEF_BSTRIDE + 1, 2 * CDEF_BSTRIDE + 2 },
27   { 1 * CDEF_BSTRIDE + 0, 2 * CDEF_BSTRIDE + 1 },
28   { 1 * CDEF_BSTRIDE + 0, 2 * CDEF_BSTRIDE + 0 },
29   { 1 * CDEF_BSTRIDE + 0, 2 * CDEF_BSTRIDE - 1 }
30 };
31 
32 /* Detect direction. 0 means 45-degree up-right, 2 is horizontal, and so on.
33    The search minimizes the weighted variance along all the lines in a
34    particular direction, i.e. the squared error between the input and a
35    "predicted" block where each pixel is replaced by the average along a line
36    in a particular direction. Since each direction have the same sum(x^2) term,
37    that term is never computed. See Section 2, step 2, of:
38    http://jmvalin.ca/notes/intra_paint.pdf */
cdef_find_dir_c(const uint16_t * img,int stride,int32_t * var,int coeff_shift)39 int cdef_find_dir_c(const uint16_t *img, int stride, int32_t *var,
40                     int coeff_shift) {
41   int i;
42   int32_t cost[8] = { 0 };
43   int partial[8][15] = { { 0 } };
44   int32_t best_cost = 0;
45   int best_dir = 0;
46   /* Instead of dividing by n between 2 and 8, we multiply by 3*5*7*8/n.
47      The output is then 840 times larger, but we don't care for finding
48      the max. */
49   static const int div_table[] = { 0, 840, 420, 280, 210, 168, 140, 120, 105 };
50   for (i = 0; i < 8; i++) {
51     int j;
52     for (j = 0; j < 8; j++) {
53       int x;
54       /* We subtract 128 here to reduce the maximum range of the squared
55          partial sums. */
56       x = (img[i * stride + j] >> coeff_shift) - 128;
57       partial[0][i + j] += x;
58       partial[1][i + j / 2] += x;
59       partial[2][i] += x;
60       partial[3][3 + i - j / 2] += x;
61       partial[4][7 + i - j] += x;
62       partial[5][3 - i / 2 + j] += x;
63       partial[6][j] += x;
64       partial[7][i / 2 + j] += x;
65     }
66   }
67   for (i = 0; i < 8; i++) {
68     cost[2] += partial[2][i] * partial[2][i];
69     cost[6] += partial[6][i] * partial[6][i];
70   }
71   cost[2] *= div_table[8];
72   cost[6] *= div_table[8];
73   for (i = 0; i < 7; i++) {
74     cost[0] += (partial[0][i] * partial[0][i] +
75                 partial[0][14 - i] * partial[0][14 - i]) *
76                div_table[i + 1];
77     cost[4] += (partial[4][i] * partial[4][i] +
78                 partial[4][14 - i] * partial[4][14 - i]) *
79                div_table[i + 1];
80   }
81   cost[0] += partial[0][7] * partial[0][7] * div_table[8];
82   cost[4] += partial[4][7] * partial[4][7] * div_table[8];
83   for (i = 1; i < 8; i += 2) {
84     int j;
85     for (j = 0; j < 4 + 1; j++) {
86       cost[i] += partial[i][3 + j] * partial[i][3 + j];
87     }
88     cost[i] *= div_table[8];
89     for (j = 0; j < 4 - 1; j++) {
90       cost[i] += (partial[i][j] * partial[i][j] +
91                   partial[i][10 - j] * partial[i][10 - j]) *
92                  div_table[2 * j + 2];
93     }
94   }
95   for (i = 0; i < 8; i++) {
96     if (cost[i] > best_cost) {
97       best_cost = cost[i];
98       best_dir = i;
99     }
100   }
101   /* Difference between the optimal variance and the variance along the
102      orthogonal direction. Again, the sum(x^2) terms cancel out. */
103   *var = best_cost - cost[(best_dir + 4) & 7];
104   /* We'd normally divide by 840, but dividing by 1024 is close enough
105      for what we're going to do with this. */
106   *var >>= 10;
107   return best_dir;
108 }
109 
110 const int cdef_pri_taps[2][2] = { { 4, 2 }, { 3, 3 } };
111 const int cdef_sec_taps[2] = { 2, 1 };
112 
113 /* Smooth in the direction detected. */
cdef_filter_block_c(uint8_t * dst8,uint16_t * dst16,int dstride,const uint16_t * in,int pri_strength,int sec_strength,int dir,int pri_damping,int sec_damping,int bsize,int coeff_shift)114 void cdef_filter_block_c(uint8_t *dst8, uint16_t *dst16, int dstride,
115                          const uint16_t *in, int pri_strength, int sec_strength,
116                          int dir, int pri_damping, int sec_damping, int bsize,
117                          int coeff_shift) {
118   int i, j, k;
119   const int s = CDEF_BSTRIDE;
120   const int *pri_taps = cdef_pri_taps[(pri_strength >> coeff_shift) & 1];
121   const int *sec_taps = cdef_sec_taps;
122   for (i = 0; i < 4 << (bsize == BLOCK_8X8 || bsize == BLOCK_4X8); i++) {
123     for (j = 0; j < 4 << (bsize == BLOCK_8X8 || bsize == BLOCK_8X4); j++) {
124       int16_t sum = 0;
125       int16_t y;
126       int16_t x = in[i * s + j];
127       int max = x;
128       int min = x;
129       for (k = 0; k < 2; k++) {
130         int16_t p0 = in[i * s + j + cdef_directions[dir][k]];
131         int16_t p1 = in[i * s + j - cdef_directions[dir][k]];
132         sum += pri_taps[k] * constrain(p0 - x, pri_strength, pri_damping);
133         sum += pri_taps[k] * constrain(p1 - x, pri_strength, pri_damping);
134         if (p0 != CDEF_VERY_LARGE) max = AOMMAX(p0, max);
135         if (p1 != CDEF_VERY_LARGE) max = AOMMAX(p1, max);
136         min = AOMMIN(p0, min);
137         min = AOMMIN(p1, min);
138         int16_t s0 = in[i * s + j + cdef_directions[(dir + 2) & 7][k]];
139         int16_t s1 = in[i * s + j - cdef_directions[(dir + 2) & 7][k]];
140         int16_t s2 = in[i * s + j + cdef_directions[(dir + 6) & 7][k]];
141         int16_t s3 = in[i * s + j - cdef_directions[(dir + 6) & 7][k]];
142         if (s0 != CDEF_VERY_LARGE) max = AOMMAX(s0, max);
143         if (s1 != CDEF_VERY_LARGE) max = AOMMAX(s1, max);
144         if (s2 != CDEF_VERY_LARGE) max = AOMMAX(s2, max);
145         if (s3 != CDEF_VERY_LARGE) max = AOMMAX(s3, max);
146         min = AOMMIN(s0, min);
147         min = AOMMIN(s1, min);
148         min = AOMMIN(s2, min);
149         min = AOMMIN(s3, min);
150         sum += sec_taps[k] * constrain(s0 - x, sec_strength, sec_damping);
151         sum += sec_taps[k] * constrain(s1 - x, sec_strength, sec_damping);
152         sum += sec_taps[k] * constrain(s2 - x, sec_strength, sec_damping);
153         sum += sec_taps[k] * constrain(s3 - x, sec_strength, sec_damping);
154       }
155       y = clamp((int16_t)x + ((8 + sum - (sum < 0)) >> 4), min, max);
156       if (dst8)
157         dst8[i * dstride + j] = (uint8_t)y;
158       else
159         dst16[i * dstride + j] = (uint16_t)y;
160     }
161   }
162 }
163 
164 /* Compute the primary filter strength for an 8x8 block based on the
165    directional variance difference. A high variance difference means
166    that we have a highly directional pattern (e.g. a high contrast
167    edge), so we can apply more deringing. A low variance means that we
168    either have a low contrast edge, or a non-directional texture, so
169    we want to be careful not to blur. */
adjust_strength(int strength,int32_t var)170 static INLINE int adjust_strength(int strength, int32_t var) {
171   const int i = var >> 6 ? AOMMIN(get_msb(var >> 6), 12) : 0;
172   /* We use the variance of 8x8 blocks to adjust the strength. */
173   return var ? (strength * (4 + i) + 8) >> 4 : 0;
174 }
175 
av1_cdef_filter_fb(uint8_t * dst8,uint16_t * dst16,int dstride,uint16_t * in,int xdec,int ydec,int dir[CDEF_NBLOCKS][CDEF_NBLOCKS],int * dirinit,int var[CDEF_NBLOCKS][CDEF_NBLOCKS],int pli,cdef_list * dlist,int cdef_count,int level,int sec_strength,int damping,int coeff_shift)176 void av1_cdef_filter_fb(uint8_t *dst8, uint16_t *dst16, int dstride,
177                         uint16_t *in, int xdec, int ydec,
178                         int dir[CDEF_NBLOCKS][CDEF_NBLOCKS], int *dirinit,
179                         int var[CDEF_NBLOCKS][CDEF_NBLOCKS], int pli,
180                         cdef_list *dlist, int cdef_count, int level,
181                         int sec_strength, int damping, int coeff_shift) {
182   int bi;
183   int bx;
184   int by;
185   const int pri_strength = level << coeff_shift;
186   sec_strength <<= coeff_shift;
187   damping += coeff_shift - (pli != AOM_PLANE_Y);
188   const int bw_log2 = 3 - xdec;
189   const int bh_log2 = 3 - ydec;
190   if (dirinit && pri_strength == 0 && sec_strength == 0) {
191     // If we're here, both primary and secondary strengths are 0, and
192     // we still haven't written anything to y[] yet, so we just copy
193     // the input to y[]. This is necessary only for av1_cdef_search()
194     // and only av1_cdef_search() sets dirinit.
195     for (bi = 0; bi < cdef_count; bi++) {
196       by = dlist[bi].by;
197       bx = dlist[bi].bx;
198       // TODO(stemidts/jmvalin): SIMD optimisations
199       for (int iy = 0; iy < 1 << bh_log2; iy++) {
200         memcpy(&dst16[(bi << (bw_log2 + bh_log2)) + (iy << bw_log2)],
201                &in[((by << bh_log2) + iy) * CDEF_BSTRIDE + (bx << bw_log2)],
202                ((size_t)1 << bw_log2) * sizeof(*dst16));
203       }
204     }
205     return;
206   }
207 
208   if (pli == 0) {
209     if (!dirinit || !*dirinit) {
210       for (bi = 0; bi < cdef_count; bi++) {
211         by = dlist[bi].by;
212         bx = dlist[bi].bx;
213         dir[by][bx] = cdef_find_dir(&in[8 * by * CDEF_BSTRIDE + 8 * bx],
214                                     CDEF_BSTRIDE, &var[by][bx], coeff_shift);
215       }
216       if (dirinit) *dirinit = 1;
217     }
218   }
219   if (pli == 1 && xdec != ydec) {
220     for (bi = 0; bi < cdef_count; bi++) {
221       static const int conv422[8] = { 7, 0, 2, 4, 5, 6, 6, 6 };
222       static const int conv440[8] = { 1, 2, 2, 2, 3, 4, 6, 0 };
223       by = dlist[bi].by;
224       bx = dlist[bi].bx;
225       dir[by][bx] = (xdec ? conv422 : conv440)[dir[by][bx]];
226     }
227   }
228 
229   const int bsize =
230       ydec ? (xdec ? BLOCK_4X4 : BLOCK_8X4) : (xdec ? BLOCK_4X8 : BLOCK_8X8);
231   const int t = pri_strength;
232   const int s = sec_strength;
233   for (bi = 0; bi < cdef_count; bi++) {
234     by = dlist[bi].by;
235     bx = dlist[bi].bx;
236     if (dst8) {
237       cdef_filter_block(
238           &dst8[(by << bh_log2) * dstride + (bx << bw_log2)], NULL, dstride,
239           &in[(by * CDEF_BSTRIDE << bh_log2) + (bx << bw_log2)],
240           (pli ? t : adjust_strength(t, var[by][bx])), s, t ? dir[by][bx] : 0,
241           damping, damping, bsize, coeff_shift);
242     } else {
243       cdef_filter_block(
244           NULL,
245           &dst16[dirinit ? bi << (bw_log2 + bh_log2)
246                          : (by << bh_log2) * dstride + (bx << bw_log2)],
247           dirinit ? 1 << bw_log2 : dstride,
248           &in[(by * CDEF_BSTRIDE << bh_log2) + (bx << bw_log2)],
249           (pli ? t : adjust_strength(t, var[by][bx])), s, t ? dir[by][bx] : 0,
250           damping, damping, bsize, coeff_shift);
251     }
252   }
253 }
254