FFmpeg
 All Data Structures Namespaces Files Functions Variables Typedefs Enumerations Enumerator Macros Groups Pages
vf_dctdnoiz.c
Go to the documentation of this file.
1 /*
2  * Copyright (c) 2013-2014 Clément Bœsch
3  *
4  * This file is part of FFmpeg.
5  *
6  * FFmpeg is free software; you can redistribute it and/or
7  * modify it under the terms of the GNU Lesser General Public
8  * License as published by the Free Software Foundation; either
9  * version 2.1 of the License, or (at your option) any later version.
10  *
11  * FFmpeg is distributed in the hope that it will be useful,
12  * but WITHOUT ANY WARRANTY; without even the implied warranty of
13  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
14  * Lesser General Public License for more details.
15  *
16  * You should have received a copy of the GNU Lesser General Public
17  * License along with FFmpeg; if not, write to the Free Software
18  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
19  */
20 
21 /**
22  * A simple, relatively efficient and slow DCT image denoiser.
23  *
24  * @see http://www.ipol.im/pub/art/2011/ys-dct/
25  *
26  * The DCT factorization used is based on "Fast and numerically stable
27  * algorithms for discrete cosine transforms" from Gerlind Plonkaa & Manfred
28  * Tasche (DOI: 10.1016/j.laa.2004.07.015).
29  */
30 
31 #include "libavutil/avassert.h"
32 #include "libavutil/eval.h"
33 #include "libavutil/opt.h"
34 #include "internal.h"
35 
36 static const char *const var_names[] = { "c", NULL };
37 enum { VAR_C, VAR_VARS_NB };
38 
39 #define MAX_THREADS 8
40 
41 typedef struct DCTdnoizContext {
42  const AVClass *class;
43 
44  /* coefficient factor expression */
45  char *expr_str;
48 
50  int pr_width, pr_height; // width and height to process
51  float sigma; // used when no expression are st
52  float th; // threshold (3*sigma)
53  float *cbuf[2][3]; // two planar rgb color buffers
54  float *slices[MAX_THREADS]; // slices buffers (1 slice buffer per thread)
55  float *weights; // dct coeff are cumulated with overlapping; these values are used for averaging
56  int p_linesize; // line sizes for color and weights
57  int overlap; // number of block overlapping pixels
58  int step; // block step increment (blocksize - overlap)
59  int n; // 1<<n is the block size
60  int bsize; // block size, 1<<n
62  const float *src, int src_linesize,
63  float *dst, int dst_linesize,
64  int thread_id);
65  void (*color_decorrelation)(float **dst, int dst_linesize,
66  const uint8_t *src, int src_linesize,
67  int w, int h);
68  void (*color_correlation)(uint8_t *dst, int dst_linesize,
69  float **src, int src_linesize,
70  int w, int h);
72 
73 #define MIN_NBITS 3 /* blocksize = 1<<3 = 8 */
74 #define MAX_NBITS 4 /* blocksize = 1<<4 = 16 */
75 #define DEFAULT_NBITS 3
76 
77 #define OFFSET(x) offsetof(DCTdnoizContext, x)
78 #define FLAGS AV_OPT_FLAG_FILTERING_PARAM|AV_OPT_FLAG_VIDEO_PARAM
79 static const AVOption dctdnoiz_options[] = {
80  { "sigma", "set noise sigma constant", OFFSET(sigma), AV_OPT_TYPE_FLOAT, {.dbl=0}, 0, 999, .flags = FLAGS },
81  { "s", "set noise sigma constant", OFFSET(sigma), AV_OPT_TYPE_FLOAT, {.dbl=0}, 0, 999, .flags = FLAGS },
82  { "overlap", "set number of block overlapping pixels", OFFSET(overlap), AV_OPT_TYPE_INT, {.i64=-1}, -1, (1<<MAX_NBITS)-1, .flags = FLAGS },
83  { "expr", "set coefficient factor expression", OFFSET(expr_str), AV_OPT_TYPE_STRING, {.str=NULL}, .flags = FLAGS },
84  { "e", "set coefficient factor expression", OFFSET(expr_str), AV_OPT_TYPE_STRING, {.str=NULL}, .flags = FLAGS },
85  { "n", "set the block size, expressed in bits", OFFSET(n), AV_OPT_TYPE_INT, {.i64=DEFAULT_NBITS}, MIN_NBITS, MAX_NBITS, .flags = FLAGS },
86  { NULL }
87 };
88 
89 AVFILTER_DEFINE_CLASS(dctdnoiz);
90 
91 static void av_always_inline fdct8_1d(float *dst, const float *src,
92  int dst_stridea, int dst_strideb,
93  int src_stridea, int src_strideb)
94 {
95  int i;
96 
97  for (i = 0; i < 8; i++) {
98  const float x00 = src[0*src_stridea] + src[7*src_stridea];
99  const float x01 = src[1*src_stridea] + src[6*src_stridea];
100  const float x02 = src[2*src_stridea] + src[5*src_stridea];
101  const float x03 = src[3*src_stridea] + src[4*src_stridea];
102  const float x04 = src[0*src_stridea] - src[7*src_stridea];
103  const float x05 = src[1*src_stridea] - src[6*src_stridea];
104  const float x06 = src[2*src_stridea] - src[5*src_stridea];
105  const float x07 = src[3*src_stridea] - src[4*src_stridea];
106  const float x08 = x00 + x03;
107  const float x09 = x01 + x02;
108  const float x0a = x00 - x03;
109  const float x0b = x01 - x02;
110  const float x0c = 1.38703984532215f*x04 + 0.275899379282943f*x07;
111  const float x0d = 1.17587560241936f*x05 + 0.785694958387102f*x06;
112  const float x0e = -0.785694958387102f*x05 + 1.17587560241936f*x06;
113  const float x0f = 0.275899379282943f*x04 - 1.38703984532215f*x07;
114  const float x10 = 0.353553390593274f * (x0c - x0d);
115  const float x11 = 0.353553390593274f * (x0e - x0f);
116  dst[0*dst_stridea] = 0.353553390593274f * (x08 + x09);
117  dst[1*dst_stridea] = 0.353553390593274f * (x0c + x0d);
118  dst[2*dst_stridea] = 0.461939766255643f*x0a + 0.191341716182545f*x0b;
119  dst[3*dst_stridea] = 0.707106781186547f * (x10 - x11);
120  dst[4*dst_stridea] = 0.353553390593274f * (x08 - x09);
121  dst[5*dst_stridea] = 0.707106781186547f * (x10 + x11);
122  dst[6*dst_stridea] = 0.191341716182545f*x0a - 0.461939766255643f*x0b;
123  dst[7*dst_stridea] = 0.353553390593274f * (x0e + x0f);
124  dst += dst_strideb;
125  src += src_strideb;
126  }
127 }
128 
129 static void av_always_inline idct8_1d(float *dst, const float *src,
130  int dst_stridea, int dst_strideb,
131  int src_stridea, int src_strideb,
132  int add)
133 {
134  int i;
135 
136  for (i = 0; i < 8; i++) {
137  const float x00 = 1.4142135623731f *src[0*src_stridea];
138  const float x01 = 1.38703984532215f *src[1*src_stridea] + 0.275899379282943f*src[7*src_stridea];
139  const float x02 = 1.30656296487638f *src[2*src_stridea] + 0.541196100146197f*src[6*src_stridea];
140  const float x03 = 1.17587560241936f *src[3*src_stridea] + 0.785694958387102f*src[5*src_stridea];
141  const float x04 = 1.4142135623731f *src[4*src_stridea];
142  const float x05 = -0.785694958387102f*src[3*src_stridea] + 1.17587560241936f*src[5*src_stridea];
143  const float x06 = 0.541196100146197f*src[2*src_stridea] - 1.30656296487638f*src[6*src_stridea];
144  const float x07 = -0.275899379282943f*src[1*src_stridea] + 1.38703984532215f*src[7*src_stridea];
145  const float x09 = x00 + x04;
146  const float x0a = x01 + x03;
147  const float x0b = 1.4142135623731f*x02;
148  const float x0c = x00 - x04;
149  const float x0d = x01 - x03;
150  const float x0e = 0.353553390593274f * (x09 - x0b);
151  const float x0f = 0.353553390593274f * (x0c + x0d);
152  const float x10 = 0.353553390593274f * (x0c - x0d);
153  const float x11 = 1.4142135623731f*x06;
154  const float x12 = x05 + x07;
155  const float x13 = x05 - x07;
156  const float x14 = 0.353553390593274f * (x11 + x12);
157  const float x15 = 0.353553390593274f * (x11 - x12);
158  const float x16 = 0.5f*x13;
159  dst[0*dst_stridea] = (add ? dst[ 0*dst_stridea] : 0) + 0.25f * (x09 + x0b) + 0.353553390593274f*x0a;
160  dst[1*dst_stridea] = (add ? dst[ 1*dst_stridea] : 0) + 0.707106781186547f * (x0f + x15);
161  dst[2*dst_stridea] = (add ? dst[ 2*dst_stridea] : 0) + 0.707106781186547f * (x0f - x15);
162  dst[3*dst_stridea] = (add ? dst[ 3*dst_stridea] : 0) + 0.707106781186547f * (x0e + x16);
163  dst[4*dst_stridea] = (add ? dst[ 4*dst_stridea] : 0) + 0.707106781186547f * (x0e - x16);
164  dst[5*dst_stridea] = (add ? dst[ 5*dst_stridea] : 0) + 0.707106781186547f * (x10 - x14);
165  dst[6*dst_stridea] = (add ? dst[ 6*dst_stridea] : 0) + 0.707106781186547f * (x10 + x14);
166  dst[7*dst_stridea] = (add ? dst[ 7*dst_stridea] : 0) + 0.25f * (x09 + x0b) - 0.353553390593274f*x0a;
167  dst += dst_strideb;
168  src += src_strideb;
169  }
170 }
171 
172 
173 static void av_always_inline fdct16_1d(float *dst, const float *src,
174  int dst_stridea, int dst_strideb,
175  int src_stridea, int src_strideb)
176 {
177  int i;
178 
179  for (i = 0; i < 16; i++) {
180  const float x00 = src[ 0*src_stridea] + src[15*src_stridea];
181  const float x01 = src[ 1*src_stridea] + src[14*src_stridea];
182  const float x02 = src[ 2*src_stridea] + src[13*src_stridea];
183  const float x03 = src[ 3*src_stridea] + src[12*src_stridea];
184  const float x04 = src[ 4*src_stridea] + src[11*src_stridea];
185  const float x05 = src[ 5*src_stridea] + src[10*src_stridea];
186  const float x06 = src[ 6*src_stridea] + src[ 9*src_stridea];
187  const float x07 = src[ 7*src_stridea] + src[ 8*src_stridea];
188  const float x08 = src[ 0*src_stridea] - src[15*src_stridea];
189  const float x09 = src[ 1*src_stridea] - src[14*src_stridea];
190  const float x0a = src[ 2*src_stridea] - src[13*src_stridea];
191  const float x0b = src[ 3*src_stridea] - src[12*src_stridea];
192  const float x0c = src[ 4*src_stridea] - src[11*src_stridea];
193  const float x0d = src[ 5*src_stridea] - src[10*src_stridea];
194  const float x0e = src[ 6*src_stridea] - src[ 9*src_stridea];
195  const float x0f = src[ 7*src_stridea] - src[ 8*src_stridea];
196  const float x10 = x00 + x07;
197  const float x11 = x01 + x06;
198  const float x12 = x02 + x05;
199  const float x13 = x03 + x04;
200  const float x14 = x00 - x07;
201  const float x15 = x01 - x06;
202  const float x16 = x02 - x05;
203  const float x17 = x03 - x04;
204  const float x18 = x10 + x13;
205  const float x19 = x11 + x12;
206  const float x1a = x10 - x13;
207  const float x1b = x11 - x12;
208  const float x1c = 1.38703984532215f*x14 + 0.275899379282943f*x17;
209  const float x1d = 1.17587560241936f*x15 + 0.785694958387102f*x16;
210  const float x1e = -0.785694958387102f*x15 + 1.17587560241936f *x16;
211  const float x1f = 0.275899379282943f*x14 - 1.38703984532215f *x17;
212  const float x20 = 0.25f * (x1c - x1d);
213  const float x21 = 0.25f * (x1e - x1f);
214  const float x22 = 1.40740373752638f *x08 + 0.138617169199091f*x0f;
215  const float x23 = 1.35331800117435f *x09 + 0.410524527522357f*x0e;
216  const float x24 = 1.24722501298667f *x0a + 0.666655658477747f*x0d;
217  const float x25 = 1.09320186700176f *x0b + 0.897167586342636f*x0c;
218  const float x26 = -0.897167586342636f*x0b + 1.09320186700176f *x0c;
219  const float x27 = 0.666655658477747f*x0a - 1.24722501298667f *x0d;
220  const float x28 = -0.410524527522357f*x09 + 1.35331800117435f *x0e;
221  const float x29 = 0.138617169199091f*x08 - 1.40740373752638f *x0f;
222  const float x2a = x22 + x25;
223  const float x2b = x23 + x24;
224  const float x2c = x22 - x25;
225  const float x2d = x23 - x24;
226  const float x2e = 0.25f * (x2a - x2b);
227  const float x2f = 0.326640741219094f*x2c + 0.135299025036549f*x2d;
228  const float x30 = 0.135299025036549f*x2c - 0.326640741219094f*x2d;
229  const float x31 = x26 + x29;
230  const float x32 = x27 + x28;
231  const float x33 = x26 - x29;
232  const float x34 = x27 - x28;
233  const float x35 = 0.25f * (x31 - x32);
234  const float x36 = 0.326640741219094f*x33 + 0.135299025036549f*x34;
235  const float x37 = 0.135299025036549f*x33 - 0.326640741219094f*x34;
236  dst[ 0*dst_stridea] = 0.25f * (x18 + x19);
237  dst[ 1*dst_stridea] = 0.25f * (x2a + x2b);
238  dst[ 2*dst_stridea] = 0.25f * (x1c + x1d);
239  dst[ 3*dst_stridea] = 0.707106781186547f * (x2f - x37);
240  dst[ 4*dst_stridea] = 0.326640741219094f*x1a + 0.135299025036549f*x1b;
241  dst[ 5*dst_stridea] = 0.707106781186547f * (x2f + x37);
242  dst[ 6*dst_stridea] = 0.707106781186547f * (x20 - x21);
243  dst[ 7*dst_stridea] = 0.707106781186547f * (x2e + x35);
244  dst[ 8*dst_stridea] = 0.25f * (x18 - x19);
245  dst[ 9*dst_stridea] = 0.707106781186547f * (x2e - x35);
246  dst[10*dst_stridea] = 0.707106781186547f * (x20 + x21);
247  dst[11*dst_stridea] = 0.707106781186547f * (x30 - x36);
248  dst[12*dst_stridea] = 0.135299025036549f*x1a - 0.326640741219094f*x1b;
249  dst[13*dst_stridea] = 0.707106781186547f * (x30 + x36);
250  dst[14*dst_stridea] = 0.25f * (x1e + x1f);
251  dst[15*dst_stridea] = 0.25f * (x31 + x32);
252  dst += dst_strideb;
253  src += src_strideb;
254  }
255 }
256 
257 static void av_always_inline idct16_1d(float *dst, const float *src,
258  int dst_stridea, int dst_strideb,
259  int src_stridea, int src_strideb,
260  int add)
261 {
262  int i;
263 
264  for (i = 0; i < 16; i++) {
265  const float x00 = 1.4142135623731f *src[ 0*src_stridea];
266  const float x01 = 1.40740373752638f *src[ 1*src_stridea] + 0.138617169199091f*src[15*src_stridea];
267  const float x02 = 1.38703984532215f *src[ 2*src_stridea] + 0.275899379282943f*src[14*src_stridea];
268  const float x03 = 1.35331800117435f *src[ 3*src_stridea] + 0.410524527522357f*src[13*src_stridea];
269  const float x04 = 1.30656296487638f *src[ 4*src_stridea] + 0.541196100146197f*src[12*src_stridea];
270  const float x05 = 1.24722501298667f *src[ 5*src_stridea] + 0.666655658477747f*src[11*src_stridea];
271  const float x06 = 1.17587560241936f *src[ 6*src_stridea] + 0.785694958387102f*src[10*src_stridea];
272  const float x07 = 1.09320186700176f *src[ 7*src_stridea] + 0.897167586342636f*src[ 9*src_stridea];
273  const float x08 = 1.4142135623731f *src[ 8*src_stridea];
274  const float x09 = -0.897167586342636f*src[ 7*src_stridea] + 1.09320186700176f*src[ 9*src_stridea];
275  const float x0a = 0.785694958387102f*src[ 6*src_stridea] - 1.17587560241936f*src[10*src_stridea];
276  const float x0b = -0.666655658477747f*src[ 5*src_stridea] + 1.24722501298667f*src[11*src_stridea];
277  const float x0c = 0.541196100146197f*src[ 4*src_stridea] - 1.30656296487638f*src[12*src_stridea];
278  const float x0d = -0.410524527522357f*src[ 3*src_stridea] + 1.35331800117435f*src[13*src_stridea];
279  const float x0e = 0.275899379282943f*src[ 2*src_stridea] - 1.38703984532215f*src[14*src_stridea];
280  const float x0f = -0.138617169199091f*src[ 1*src_stridea] + 1.40740373752638f*src[15*src_stridea];
281  const float x12 = x00 + x08;
282  const float x13 = x01 + x07;
283  const float x14 = x02 + x06;
284  const float x15 = x03 + x05;
285  const float x16 = 1.4142135623731f*x04;
286  const float x17 = x00 - x08;
287  const float x18 = x01 - x07;
288  const float x19 = x02 - x06;
289  const float x1a = x03 - x05;
290  const float x1d = x12 + x16;
291  const float x1e = x13 + x15;
292  const float x1f = 1.4142135623731f*x14;
293  const float x20 = x12 - x16;
294  const float x21 = x13 - x15;
295  const float x22 = 0.25f * (x1d - x1f);
296  const float x23 = 0.25f * (x20 + x21);
297  const float x24 = 0.25f * (x20 - x21);
298  const float x25 = 1.4142135623731f*x17;
299  const float x26 = 1.30656296487638f*x18 + 0.541196100146197f*x1a;
300  const float x27 = 1.4142135623731f*x19;
301  const float x28 = -0.541196100146197f*x18 + 1.30656296487638f*x1a;
302  const float x29 = 0.176776695296637f * (x25 + x27) + 0.25f*x26;
303  const float x2a = 0.25f * (x25 - x27);
304  const float x2b = 0.176776695296637f * (x25 + x27) - 0.25f*x26;
305  const float x2c = 0.353553390593274f*x28;
306  const float x1b = 0.707106781186547f * (x2a - x2c);
307  const float x1c = 0.707106781186547f * (x2a + x2c);
308  const float x2d = 1.4142135623731f*x0c;
309  const float x2e = x0b + x0d;
310  const float x2f = x0a + x0e;
311  const float x30 = x09 + x0f;
312  const float x31 = x09 - x0f;
313  const float x32 = x0a - x0e;
314  const float x33 = x0b - x0d;
315  const float x37 = 1.4142135623731f*x2d;
316  const float x38 = 1.30656296487638f*x2e + 0.541196100146197f*x30;
317  const float x39 = 1.4142135623731f*x2f;
318  const float x3a = -0.541196100146197f*x2e + 1.30656296487638f*x30;
319  const float x3b = 0.176776695296637f * (x37 + x39) + 0.25f*x38;
320  const float x3c = 0.25f * (x37 - x39);
321  const float x3d = 0.176776695296637f * (x37 + x39) - 0.25f*x38;
322  const float x3e = 0.353553390593274f*x3a;
323  const float x34 = 0.707106781186547f * (x3c - x3e);
324  const float x35 = 0.707106781186547f * (x3c + x3e);
325  const float x3f = 1.4142135623731f*x32;
326  const float x40 = x31 + x33;
327  const float x41 = x31 - x33;
328  const float x42 = 0.25f * (x3f + x40);
329  const float x43 = 0.25f * (x3f - x40);
330  const float x44 = 0.353553390593274f*x41;
331  dst[ 0*dst_stridea] = (add ? dst[ 0*dst_stridea] : 0) + 0.176776695296637f * (x1d + x1f) + 0.25f*x1e;
332  dst[ 1*dst_stridea] = (add ? dst[ 1*dst_stridea] : 0) + 0.707106781186547f * (x29 + x3d);
333  dst[ 2*dst_stridea] = (add ? dst[ 2*dst_stridea] : 0) + 0.707106781186547f * (x29 - x3d);
334  dst[ 3*dst_stridea] = (add ? dst[ 3*dst_stridea] : 0) + 0.707106781186547f * (x23 - x43);
335  dst[ 4*dst_stridea] = (add ? dst[ 4*dst_stridea] : 0) + 0.707106781186547f * (x23 + x43);
336  dst[ 5*dst_stridea] = (add ? dst[ 5*dst_stridea] : 0) + 0.707106781186547f * (x1b - x35);
337  dst[ 6*dst_stridea] = (add ? dst[ 6*dst_stridea] : 0) + 0.707106781186547f * (x1b + x35);
338  dst[ 7*dst_stridea] = (add ? dst[ 7*dst_stridea] : 0) + 0.707106781186547f * (x22 + x44);
339  dst[ 8*dst_stridea] = (add ? dst[ 8*dst_stridea] : 0) + 0.707106781186547f * (x22 - x44);
340  dst[ 9*dst_stridea] = (add ? dst[ 9*dst_stridea] : 0) + 0.707106781186547f * (x1c + x34);
341  dst[10*dst_stridea] = (add ? dst[10*dst_stridea] : 0) + 0.707106781186547f * (x1c - x34);
342  dst[11*dst_stridea] = (add ? dst[11*dst_stridea] : 0) + 0.707106781186547f * (x24 + x42);
343  dst[12*dst_stridea] = (add ? dst[12*dst_stridea] : 0) + 0.707106781186547f * (x24 - x42);
344  dst[13*dst_stridea] = (add ? dst[13*dst_stridea] : 0) + 0.707106781186547f * (x2b - x3b);
345  dst[14*dst_stridea] = (add ? dst[14*dst_stridea] : 0) + 0.707106781186547f * (x2b + x3b);
346  dst[15*dst_stridea] = (add ? dst[15*dst_stridea] : 0) + 0.176776695296637f * (x1d + x1f) - 0.25f*x1e;
347  dst += dst_strideb;
348  src += src_strideb;
349  }
350 }
351 
352 #define DEF_FILTER_FREQ_FUNCS(bsize) \
353 static av_always_inline void filter_freq_##bsize(const float *src, int src_linesize, \
354  float *dst, int dst_linesize, \
355  AVExpr *expr, double *var_values, \
356  int sigma_th) \
357 { \
358  unsigned i; \
359  DECLARE_ALIGNED(32, float, tmp_block1)[bsize * bsize]; \
360  DECLARE_ALIGNED(32, float, tmp_block2)[bsize * bsize]; \
361  \
362  /* forward DCT */ \
363  fdct##bsize##_1d(tmp_block1, src, 1, bsize, 1, src_linesize); \
364  fdct##bsize##_1d(tmp_block2, tmp_block1, bsize, 1, bsize, 1); \
365  \
366  for (i = 0; i < bsize*bsize; i++) { \
367  float *b = &tmp_block2[i]; \
368  /* frequency filtering */ \
369  if (expr) { \
370  var_values[VAR_C] = FFABS(*b); \
371  *b *= av_expr_eval(expr, var_values, NULL); \
372  } else { \
373  if (FFABS(*b) < sigma_th) \
374  *b = 0; \
375  } \
376  } \
377  \
378  /* inverse DCT */ \
379  idct##bsize##_1d(tmp_block1, tmp_block2, 1, bsize, 1, bsize, 0); \
380  idct##bsize##_1d(dst, tmp_block1, dst_linesize, 1, bsize, 1, 1); \
381 } \
382  \
383 static void filter_freq_sigma_##bsize(DCTdnoizContext *s, \
384  const float *src, int src_linesize, \
385  float *dst, int dst_linesize, int thread_id) \
386 { \
387  filter_freq_##bsize(src, src_linesize, dst, dst_linesize, NULL, NULL, s->th); \
388 } \
389  \
390 static void filter_freq_expr_##bsize(DCTdnoizContext *s, \
391  const float *src, int src_linesize, \
392  float *dst, int dst_linesize, int thread_id) \
393 { \
394  filter_freq_##bsize(src, src_linesize, dst, dst_linesize, \
395  s->expr[thread_id], s->var_values[thread_id], 0); \
396 }
397 
400 
401 #define DCT3X3_0_0 0.5773502691896258f /* 1/sqrt(3) */
402 #define DCT3X3_0_1 0.5773502691896258f /* 1/sqrt(3) */
403 #define DCT3X3_0_2 0.5773502691896258f /* 1/sqrt(3) */
404 #define DCT3X3_1_0 0.7071067811865475f /* 1/sqrt(2) */
405 #define DCT3X3_1_2 -0.7071067811865475f /* -1/sqrt(2) */
406 #define DCT3X3_2_0 0.4082482904638631f /* 1/sqrt(6) */
407 #define DCT3X3_2_1 -0.8164965809277261f /* -2/sqrt(6) */
408 #define DCT3X3_2_2 0.4082482904638631f /* 1/sqrt(6) */
409 
410 static av_always_inline void color_decorrelation(float **dst, int dst_linesize,
411  const uint8_t *src, int src_linesize,
412  int w, int h,
413  int r, int g, int b)
414 {
415  int x, y;
416  float *dstp_r = dst[0];
417  float *dstp_g = dst[1];
418  float *dstp_b = dst[2];
419 
420  for (y = 0; y < h; y++) {
421  const uint8_t *srcp = src;
422 
423  for (x = 0; x < w; x++) {
424  dstp_r[x] = srcp[r] * DCT3X3_0_0 + srcp[g] * DCT3X3_0_1 + srcp[b] * DCT3X3_0_2;
425  dstp_g[x] = srcp[r] * DCT3X3_1_0 + srcp[b] * DCT3X3_1_2;
426  dstp_b[x] = srcp[r] * DCT3X3_2_0 + srcp[g] * DCT3X3_2_1 + srcp[b] * DCT3X3_2_2;
427  srcp += 3;
428  }
429  src += src_linesize;
430  dstp_r += dst_linesize;
431  dstp_g += dst_linesize;
432  dstp_b += dst_linesize;
433  }
434 }
435 
436 static av_always_inline void color_correlation(uint8_t *dst, int dst_linesize,
437  float **src, int src_linesize,
438  int w, int h,
439  int r, int g, int b)
440 {
441  int x, y;
442  const float *src_r = src[0];
443  const float *src_g = src[1];
444  const float *src_b = src[2];
445 
446  for (y = 0; y < h; y++) {
447  uint8_t *dstp = dst;
448 
449  for (x = 0; x < w; x++) {
450  dstp[r] = av_clip_uint8(src_r[x] * DCT3X3_0_0 + src_g[x] * DCT3X3_1_0 + src_b[x] * DCT3X3_2_0);
451  dstp[g] = av_clip_uint8(src_r[x] * DCT3X3_0_1 + src_b[x] * DCT3X3_2_1);
452  dstp[b] = av_clip_uint8(src_r[x] * DCT3X3_0_2 + src_g[x] * DCT3X3_1_2 + src_b[x] * DCT3X3_2_2);
453  dstp += 3;
454  }
455  dst += dst_linesize;
456  src_r += src_linesize;
457  src_g += src_linesize;
458  src_b += src_linesize;
459  }
460 }
461 
462 #define DECLARE_COLOR_FUNCS(name, r, g, b) \
463 static void color_decorrelation_##name(float **dst, int dst_linesize, \
464  const uint8_t *src, int src_linesize, \
465  int w, int h) \
466 { \
467  color_decorrelation(dst, dst_linesize, src, src_linesize, w, h, r, g, b); \
468 } \
469  \
470 static void color_correlation_##name(uint8_t *dst, int dst_linesize, \
471  float **src, int src_linesize, \
472  int w, int h) \
473 { \
474  color_correlation(dst, dst_linesize, src, src_linesize, w, h, r, g, b); \
475 }
476 
477 DECLARE_COLOR_FUNCS(rgb, 0, 1, 2)
478 DECLARE_COLOR_FUNCS(bgr, 2, 1, 0)
479 
480 static int config_input(AVFilterLink *inlink)
481 {
482  AVFilterContext *ctx = inlink->dst;
483  DCTdnoizContext *s = ctx->priv;
484  int i, x, y, bx, by, linesize, *iweights, max_slice_h, slice_h;
485  const int bsize = 1 << s->n;
486 
487  switch (inlink->format) {
488  case AV_PIX_FMT_BGR24:
489  s->color_decorrelation = color_decorrelation_bgr;
490  s->color_correlation = color_correlation_bgr;
491  break;
492  case AV_PIX_FMT_RGB24:
493  s->color_decorrelation = color_decorrelation_rgb;
494  s->color_correlation = color_correlation_rgb;
495  break;
496  default:
497  av_assert0(0);
498  }
499 
500  s->pr_width = inlink->w - (inlink->w - bsize) % s->step;
501  s->pr_height = inlink->h - (inlink->h - bsize) % s->step;
502  if (s->pr_width != inlink->w)
503  av_log(ctx, AV_LOG_WARNING, "The last %d horizontal pixels won't be denoised\n",
504  inlink->w - s->pr_width);
505  if (s->pr_height != inlink->h)
506  av_log(ctx, AV_LOG_WARNING, "The last %d vertical pixels won't be denoised\n",
507  inlink->h - s->pr_height);
508 
509  max_slice_h = s->pr_height / ((s->bsize - 1) * 2);
510  s->nb_threads = FFMIN3(MAX_THREADS, ctx->graph->nb_threads, max_slice_h);
511  av_log(ctx, AV_LOG_DEBUG, "threads: [max=%d hmax=%d user=%d] => %d\n",
512  MAX_THREADS, max_slice_h, ctx->graph->nb_threads, s->nb_threads);
513 
514  s->p_linesize = linesize = FFALIGN(s->pr_width, 32);
515  for (i = 0; i < 2; i++) {
516  s->cbuf[i][0] = av_malloc(linesize * s->pr_height * sizeof(*s->cbuf[i][0]));
517  s->cbuf[i][1] = av_malloc(linesize * s->pr_height * sizeof(*s->cbuf[i][1]));
518  s->cbuf[i][2] = av_malloc(linesize * s->pr_height * sizeof(*s->cbuf[i][2]));
519  if (!s->cbuf[i][0] || !s->cbuf[i][1] || !s->cbuf[i][2])
520  return AVERROR(ENOMEM);
521  }
522 
523  /* eval expressions are probably not thread safe when the eval internal
524  * state can be changed (typically through load & store operations) */
525  if (s->expr_str) {
526  for (i = 0; i < s->nb_threads; i++) {
527  int ret = av_expr_parse(&s->expr[i], s->expr_str, var_names,
528  NULL, NULL, NULL, NULL, 0, ctx);
529  if (ret < 0)
530  return ret;
531  }
532  }
533 
534  /* each slice will need to (pre & re)process the top and bottom block of
535  * the previous one in in addition to its processing area. This is because
536  * each pixel is averaged by all the surrounding blocks */
537  slice_h = (int)ceilf(s->pr_height / s->nb_threads) + (s->bsize - 1) * 2;
538  for (i = 0; i < s->nb_threads; i++) {
539  s->slices[i] = av_malloc_array(linesize, slice_h * sizeof(*s->slices[i]));
540  if (!s->slices[i])
541  return AVERROR(ENOMEM);
542  }
543 
544  s->weights = av_malloc(s->pr_height * linesize * sizeof(*s->weights));
545  if (!s->weights)
546  return AVERROR(ENOMEM);
547  iweights = av_calloc(s->pr_height, linesize * sizeof(*iweights));
548  if (!iweights)
549  return AVERROR(ENOMEM);
550  for (y = 0; y < s->pr_height - bsize + 1; y += s->step)
551  for (x = 0; x < s->pr_width - bsize + 1; x += s->step)
552  for (by = 0; by < bsize; by++)
553  for (bx = 0; bx < bsize; bx++)
554  iweights[(y + by)*linesize + x + bx]++;
555  for (y = 0; y < s->pr_height; y++)
556  for (x = 0; x < s->pr_width; x++)
557  s->weights[y*linesize + x] = 1. / iweights[y*linesize + x];
558  av_free(iweights);
559 
560  return 0;
561 }
562 
563 static av_cold int init(AVFilterContext *ctx)
564 {
565  DCTdnoizContext *s = ctx->priv;
566 
567  s->bsize = 1 << s->n;
568  if (s->overlap == -1)
569  s->overlap = s->bsize - 1;
570 
571  if (s->overlap > s->bsize - 1) {
572  av_log(s, AV_LOG_ERROR, "Overlap value can not except %d "
573  "with a block size of %dx%d\n",
574  s->bsize - 1, s->bsize, s->bsize);
575  return AVERROR(EINVAL);
576  }
577 
578  if (s->expr_str) {
579  switch (s->n) {
580  case 3: s->filter_freq_func = filter_freq_expr_8; break;
581  case 4: s->filter_freq_func = filter_freq_expr_16; break;
582  default: av_assert0(0);
583  }
584  } else {
585  switch (s->n) {
586  case 3: s->filter_freq_func = filter_freq_sigma_8; break;
587  case 4: s->filter_freq_func = filter_freq_sigma_16; break;
588  default: av_assert0(0);
589  }
590  }
591 
592  s->th = s->sigma * 3.;
593  s->step = s->bsize - s->overlap;
594  return 0;
595 }
596 
598 {
599  static const enum AVPixelFormat pix_fmts[] = {
602  };
604  return 0;
605 }
606 
607 typedef struct ThreadData {
608  float *src, *dst;
609 } ThreadData;
610 
612  void *arg, int jobnr, int nb_jobs)
613 {
614  int x, y;
615  DCTdnoizContext *s = ctx->priv;
616  const ThreadData *td = arg;
617  const int w = s->pr_width;
618  const int h = s->pr_height;
619  const int slice_start = (h * jobnr ) / nb_jobs;
620  const int slice_end = (h * (jobnr+1)) / nb_jobs;
621  const int slice_start_ctx = FFMAX(slice_start - s->bsize + 1, 0);
622  const int slice_end_ctx = FFMIN(slice_end, h - s->bsize + 1);
623  const int slice_h = slice_end_ctx - slice_start_ctx;
624  const int src_linesize = s->p_linesize;
625  const int dst_linesize = s->p_linesize;
626  const int slice_linesize = s->p_linesize;
627  float *dst;
628  const float *src = td->src + slice_start_ctx * src_linesize;
629  const float *weights = s->weights + slice_start * dst_linesize;
630  float *slice = s->slices[jobnr];
631 
632  // reset block sums
633  memset(slice, 0, (slice_h + s->bsize - 1) * dst_linesize * sizeof(*slice));
634 
635  // block dct sums
636  for (y = 0; y < slice_h; y += s->step) {
637  for (x = 0; x < w - s->bsize + 1; x += s->step)
638  s->filter_freq_func(s, src + x, src_linesize,
639  slice + x, slice_linesize,
640  jobnr);
641  src += s->step * src_linesize;
642  slice += s->step * slice_linesize;
643  }
644 
645  // average blocks
646  slice = s->slices[jobnr] + (slice_start - slice_start_ctx) * slice_linesize;
647  dst = td->dst + slice_start * dst_linesize;
648  for (y = slice_start; y < slice_end; y++) {
649  for (x = 0; x < w; x++)
650  dst[x] = slice[x] * weights[x];
651  slice += slice_linesize;
652  dst += dst_linesize;
653  weights += dst_linesize;
654  }
655 
656  return 0;
657 }
658 
659 static int filter_frame(AVFilterLink *inlink, AVFrame *in)
660 {
661  AVFilterContext *ctx = inlink->dst;
662  DCTdnoizContext *s = ctx->priv;
663  AVFilterLink *outlink = inlink->dst->outputs[0];
664  int direct, plane;
665  AVFrame *out;
666 
667  if (av_frame_is_writable(in)) {
668  direct = 1;
669  out = in;
670  } else {
671  direct = 0;
672  out = ff_get_video_buffer(outlink, outlink->w, outlink->h);
673  if (!out) {
674  av_frame_free(&in);
675  return AVERROR(ENOMEM);
676  }
677  av_frame_copy_props(out, in);
678  }
679 
680  s->color_decorrelation(s->cbuf[0], s->p_linesize,
681  in->data[0], in->linesize[0],
682  s->pr_width, s->pr_height);
683  for (plane = 0; plane < 3; plane++) {
684  ThreadData td = {
685  .src = s->cbuf[0][plane],
686  .dst = s->cbuf[1][plane],
687  };
688  ctx->internal->execute(ctx, filter_slice, &td, NULL, s->nb_threads);
689  }
690  s->color_correlation(out->data[0], out->linesize[0],
691  s->cbuf[1], s->p_linesize,
692  s->pr_width, s->pr_height);
693 
694  if (!direct) {
695  int y;
696  uint8_t *dst = out->data[0];
697  const uint8_t *src = in->data[0];
698  const int dst_linesize = out->linesize[0];
699  const int src_linesize = in->linesize[0];
700  const int hpad = (inlink->w - s->pr_width) * 3;
701  const int vpad = (inlink->h - s->pr_height);
702 
703  if (hpad) {
704  uint8_t *dstp = dst + s->pr_width * 3;
705  const uint8_t *srcp = src + s->pr_width * 3;
706 
707  for (y = 0; y < s->pr_height; y++) {
708  memcpy(dstp, srcp, hpad);
709  dstp += dst_linesize;
710  srcp += src_linesize;
711  }
712  }
713  if (vpad) {
714  uint8_t *dstp = dst + s->pr_height * dst_linesize;
715  const uint8_t *srcp = src + s->pr_height * src_linesize;
716 
717  for (y = 0; y < vpad; y++) {
718  memcpy(dstp, srcp, inlink->w * 3);
719  dstp += dst_linesize;
720  srcp += src_linesize;
721  }
722  }
723 
724  av_frame_free(&in);
725  }
726 
727  return ff_filter_frame(outlink, out);
728 }
729 
730 static av_cold void uninit(AVFilterContext *ctx)
731 {
732  int i;
733  DCTdnoizContext *s = ctx->priv;
734 
735  av_free(s->weights);
736  for (i = 0; i < 2; i++) {
737  av_free(s->cbuf[i][0]);
738  av_free(s->cbuf[i][1]);
739  av_free(s->cbuf[i][2]);
740  }
741  for (i = 0; i < s->nb_threads; i++) {
742  av_free(s->slices[i]);
743  av_expr_free(s->expr[i]);
744  }
745 }
746 
747 static const AVFilterPad dctdnoiz_inputs[] = {
748  {
749  .name = "default",
750  .type = AVMEDIA_TYPE_VIDEO,
751  .filter_frame = filter_frame,
752  .config_props = config_input,
753  },
754  { NULL }
755 };
756 
757 static const AVFilterPad dctdnoiz_outputs[] = {
758  {
759  .name = "default",
760  .type = AVMEDIA_TYPE_VIDEO,
761  },
762  { NULL }
763 };
764 
766  .name = "dctdnoiz",
767  .description = NULL_IF_CONFIG_SMALL("Denoise frames using 2D DCT."),
768  .priv_size = sizeof(DCTdnoizContext),
769  .init = init,
770  .uninit = uninit,
772  .inputs = dctdnoiz_inputs,
773  .outputs = dctdnoiz_outputs,
774  .priv_class = &dctdnoiz_class,
776 };