FFmpeg
 All Data Structures Namespaces Files Functions Variables Typedefs Enumerations Enumerator Macros Groups Pages
vf_spp.c
Go to the documentation of this file.
1 /*
2  * Copyright (c) 2003 Michael Niedermayer <michaelni@gmx.at>
3  * Copyright (c) 2013 Clément Bœsch <u pkh me>
4  *
5  * This file is part of FFmpeg.
6  *
7  * FFmpeg is free software; you can redistribute it and/or modify
8  * it under the terms of the GNU General Public License as published by
9  * the Free Software Foundation; either version 2 of the License, or
10  * (at your option) any later version.
11  *
12  * FFmpeg is distributed in the hope that it will be useful,
13  * but WITHOUT ANY WARRANTY; without even the implied warranty of
14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
15  * GNU General Public License for more details.
16  *
17  * You should have received a copy of the GNU General Public License along
18  * with FFmpeg; if not, write to the Free Software Foundation, Inc.,
19  * 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
20  */
21 
22 /**
23  * @file
24  * Simple post processing filter
25  *
26  * This implementation is based on an algorithm described in
27  * "Aria Nosratinia Embedded Post-Processing for
28  * Enhancement of Compressed Images (1999)"
29  *
30  * Originally written by Michael Niedermayer for the MPlayer project, and
31  * ported by Clément Bœsch for FFmpeg.
32  */
33 
34 #include "libavutil/avassert.h"
35 #include "libavutil/imgutils.h"
36 #include "libavutil/opt.h"
37 #include "libavutil/pixdesc.h"
38 #include "internal.h"
39 #include "vf_spp.h"
40 
41 enum mode {
45 };
46 
47 static const AVClass *child_class_next(const AVClass *prev)
48 {
49  return prev ? NULL : avcodec_dct_get_class();
50 }
51 
52 static void *child_next(void *obj, void *prev)
53 {
54  SPPContext *s = obj;
55  return prev ? NULL : s->dct;
56 }
57 
58 #define OFFSET(x) offsetof(SPPContext, x)
59 #define FLAGS AV_OPT_FLAG_FILTERING_PARAM|AV_OPT_FLAG_VIDEO_PARAM
60 static const AVOption spp_options[] = {
61  { "quality", "set quality", OFFSET(log2_count), AV_OPT_TYPE_INT, {.i64 = 3}, 0, MAX_LEVEL, FLAGS },
62  { "qp", "force a constant quantizer parameter", OFFSET(qp), AV_OPT_TYPE_INT, {.i64 = 0}, 0, 63, FLAGS },
63  { "mode", "set thresholding mode", OFFSET(mode), AV_OPT_TYPE_INT, {.i64 = MODE_HARD}, 0, NB_MODES - 1, FLAGS, "mode" },
64  { "hard", "hard thresholding", 0, AV_OPT_TYPE_CONST, {.i64 = MODE_HARD}, INT_MIN, INT_MAX, FLAGS, "mode" },
65  { "soft", "soft thresholding", 0, AV_OPT_TYPE_CONST, {.i64 = MODE_SOFT}, INT_MIN, INT_MAX, FLAGS, "mode" },
66  { "use_bframe_qp", "use B-frames' QP", OFFSET(use_bframe_qp), AV_OPT_TYPE_INT, {.i64 = 0}, 0, 1, FLAGS },
67  { NULL }
68 };
69 
70 static const AVClass spp_class = {
71  .class_name = "spp",
72  .item_name = av_default_item_name,
73  .option = spp_options,
74  .version = LIBAVUTIL_VERSION_INT,
75  .category = AV_CLASS_CATEGORY_FILTER,
76  .child_class_next = child_class_next,
78 };
79 
80 // XXX: share between filters?
81 DECLARE_ALIGNED(8, static const uint8_t, ldither)[8][8] = {
82  { 0, 48, 12, 60, 3, 51, 15, 63 },
83  { 32, 16, 44, 28, 35, 19, 47, 31 },
84  { 8, 56, 4, 52, 11, 59, 7, 55 },
85  { 40, 24, 36, 20, 43, 27, 39, 23 },
86  { 2, 50, 14, 62, 1, 49, 13, 61 },
87  { 34, 18, 46, 30, 33, 17, 45, 29 },
88  { 10, 58, 6, 54, 9, 57, 5, 53 },
89  { 42, 26, 38, 22, 41, 25, 37, 21 },
90 };
91 
92 static const uint8_t offset[127][2] = {
93  {0,0},
94  {0,0}, {4,4}, // quality = 1
95  {0,0}, {2,2}, {6,4}, {4,6}, // quality = 2
96  {0,0}, {5,1}, {2,2}, {7,3}, {4,4}, {1,5}, {6,6}, {3,7}, // quality = 3
97 
98  {0,0}, {4,0}, {1,1}, {5,1}, {3,2}, {7,2}, {2,3}, {6,3}, // quality = 4
99  {0,4}, {4,4}, {1,5}, {5,5}, {3,6}, {7,6}, {2,7}, {6,7},
100 
101  {0,0}, {0,2}, {0,4}, {0,6}, {1,1}, {1,3}, {1,5}, {1,7}, // quality = 5
102  {2,0}, {2,2}, {2,4}, {2,6}, {3,1}, {3,3}, {3,5}, {3,7},
103  {4,0}, {4,2}, {4,4}, {4,6}, {5,1}, {5,3}, {5,5}, {5,7},
104  {6,0}, {6,2}, {6,4}, {6,6}, {7,1}, {7,3}, {7,5}, {7,7},
105 
106  {0,0}, {4,4}, {0,4}, {4,0}, {2,2}, {6,6}, {2,6}, {6,2}, // quality = 6
107  {0,2}, {4,6}, {0,6}, {4,2}, {2,0}, {6,4}, {2,4}, {6,0},
108  {1,1}, {5,5}, {1,5}, {5,1}, {3,3}, {7,7}, {3,7}, {7,3},
109  {1,3}, {5,7}, {1,7}, {5,3}, {3,1}, {7,5}, {3,5}, {7,1},
110  {0,1}, {4,5}, {0,5}, {4,1}, {2,3}, {6,7}, {2,7}, {6,3},
111  {0,3}, {4,7}, {0,7}, {4,3}, {2,1}, {6,5}, {2,5}, {6,1},
112  {1,0}, {5,4}, {1,4}, {5,0}, {3,2}, {7,6}, {3,6}, {7,2},
113  {1,2}, {5,6}, {1,6}, {5,2}, {3,0}, {7,4}, {3,4}, {7,0},
114 };
115 
116 static void hardthresh_c(int16_t dst[64], const int16_t src[64],
117  int qp, const uint8_t *permutation)
118 {
119  int i;
120  int bias = 0; // FIXME
121 
122  unsigned threshold1 = qp * ((1<<4) - bias) - 1;
123  unsigned threshold2 = threshold1 << 1;
124 
125  memset(dst, 0, 64 * sizeof(dst[0]));
126  dst[0] = (src[0] + 4) >> 3;
127 
128  for (i = 1; i < 64; i++) {
129  int level = src[i];
130  if (((unsigned)(level + threshold1)) > threshold2) {
131  const int j = permutation[i];
132  dst[j] = (level + 4) >> 3;
133  }
134  }
135 }
136 
137 static void softthresh_c(int16_t dst[64], const int16_t src[64],
138  int qp, const uint8_t *permutation)
139 {
140  int i;
141  int bias = 0; //FIXME
142 
143  unsigned threshold1 = qp * ((1<<4) - bias) - 1;
144  unsigned threshold2 = threshold1 << 1;
145 
146  memset(dst, 0, 64 * sizeof(dst[0]));
147  dst[0] = (src[0] + 4) >> 3;
148 
149  for (i = 1; i < 64; i++) {
150  int level = src[i];
151  if (((unsigned)(level + threshold1)) > threshold2) {
152  const int j = permutation[i];
153  if (level > 0) dst[j] = (level - threshold1 + 4) >> 3;
154  else dst[j] = (level + threshold1 + 4) >> 3;
155  }
156  }
157 }
158 
159 static void store_slice_c(uint8_t *dst, const int16_t *src,
160  int dst_linesize, int src_linesize,
161  int width, int height, int log2_scale,
162  const uint8_t dither[8][8])
163 {
164  int y, x;
165 
166 #define STORE(pos) do { \
167  temp = ((src[x + y*src_linesize + pos] << log2_scale) + d[pos]) >> 6; \
168  if (temp & 0x100) \
169  temp = ~(temp >> 31); \
170  dst[x + y*dst_linesize + pos] = temp; \
171 } while (0)
172 
173  for (y = 0; y < height; y++) {
174  const uint8_t *d = dither[y];
175  for (x = 0; x < width; x += 8) {
176  int temp;
177  STORE(0);
178  STORE(1);
179  STORE(2);
180  STORE(3);
181  STORE(4);
182  STORE(5);
183  STORE(6);
184  STORE(7);
185  }
186  }
187 }
188 
189 static void store_slice16_c(uint16_t *dst, const int16_t *src,
190  int dst_linesize, int src_linesize,
191  int width, int height, int log2_scale,
192  const uint8_t dither[8][8], int depth)
193 {
194  int y, x;
195  unsigned int mask = -1<<depth;
196 
197 #define STORE16(pos) do { \
198  temp = ((src[x + y*src_linesize + pos] << log2_scale) + (d[pos]>>1)) >> 5; \
199  if (temp & mask ) \
200  temp = ~(temp >> 31); \
201  dst[x + y*dst_linesize + pos] = temp; \
202 } while (0)
203 
204  for (y = 0; y < height; y++) {
205  const uint8_t *d = dither[y];
206  for (x = 0; x < width; x += 8) {
207  int temp;
208  STORE16(0);
209  STORE16(1);
210  STORE16(2);
211  STORE16(3);
212  STORE16(4);
213  STORE16(5);
214  STORE16(6);
215  STORE16(7);
216  }
217  }
218 }
219 
220 static inline void add_block(uint16_t *dst, int linesize, const int16_t block[64])
221 {
222  int y;
223 
224  for (y = 0; y < 8; y++) {
225  *(uint32_t *)&dst[0 + y*linesize] += *(uint32_t *)&block[0 + y*8];
226  *(uint32_t *)&dst[2 + y*linesize] += *(uint32_t *)&block[2 + y*8];
227  *(uint32_t *)&dst[4 + y*linesize] += *(uint32_t *)&block[4 + y*8];
228  *(uint32_t *)&dst[6 + y*linesize] += *(uint32_t *)&block[6 + y*8];
229  }
230 }
231 
232 static void filter(SPPContext *p, uint8_t *dst, uint8_t *src,
233  int dst_linesize, int src_linesize, int width, int height,
234  const uint8_t *qp_table, int qp_stride, int is_luma, int depth)
235 {
236  int x, y, i;
237  const int count = 1 << p->log2_count;
238  const int linesize = is_luma ? p->temp_linesize : FFALIGN(width+16, 16);
239  DECLARE_ALIGNED(16, uint64_t, block_align)[32];
240  int16_t *block = (int16_t *)block_align;
241  int16_t *block2 = (int16_t *)(block_align + 16);
242  uint16_t *psrc16 = (uint16_t*)p->src;
243  const int sample_bytes = (depth+7) / 8;
244 
245  for (y = 0; y < height; y++) {
246  int index = 8 + 8*linesize + y*linesize;
247  memcpy(p->src + index*sample_bytes, src + y*src_linesize, width*sample_bytes);
248  if (sample_bytes == 1) {
249  for (x = 0; x < 8; x++) {
250  p->src[index - x - 1] = p->src[index + x ];
251  p->src[index + width + x ] = p->src[index + width - x - 1];
252  }
253  } else {
254  for (x = 0; x < 8; x++) {
255  psrc16[index - x - 1] = psrc16[index + x ];
256  psrc16[index + width + x ] = psrc16[index + width - x - 1];
257  }
258  }
259  }
260  for (y = 0; y < 8; y++) {
261  memcpy(p->src + ( 7-y)*linesize * sample_bytes, p->src + ( y+8)*linesize * sample_bytes, linesize * sample_bytes);
262  memcpy(p->src + (height+8+y)*linesize * sample_bytes, p->src + (height-y+7)*linesize * sample_bytes, linesize * sample_bytes);
263  }
264 
265  for (y = 0; y < height + 8; y += 8) {
266  memset(p->temp + (8 + y) * linesize, 0, 8 * linesize * sizeof(*p->temp));
267  for (x = 0; x < width + 8; x += 8) {
268  int qp;
269 
270  if (p->qp) {
271  qp = p->qp;
272  } else{
273  const int qps = 3 + is_luma;
274  qp = qp_table[(FFMIN(x, width - 1) >> qps) + (FFMIN(y, height - 1) >> qps) * qp_stride];
275  qp = FFMAX(1, ff_norm_qscale(qp, p->qscale_type));
276  }
277  for (i = 0; i < count; i++) {
278  const int x1 = x + offset[i + count - 1][0];
279  const int y1 = y + offset[i + count - 1][1];
280  const int index = x1 + y1*linesize;
281  p->dct->get_pixels(block, p->src + sample_bytes*index, sample_bytes*linesize);
282  p->dct->fdct(block);
283  p->requantize(block2, block, qp, p->dct->idct_permutation);
284  p->dct->idct(block2);
285  add_block(p->temp + index, linesize, block2);
286  }
287  }
288  if (y) {
289  if (sample_bytes == 1) {
290  p->store_slice(dst + (y - 8) * dst_linesize, p->temp + 8 + y*linesize,
291  dst_linesize, linesize, width,
292  FFMIN(8, height + 8 - y), MAX_LEVEL - p->log2_count,
293  ldither);
294  } else {
295  store_slice16_c((uint16_t*)(dst + (y - 8) * dst_linesize), p->temp + 8 + y*linesize,
296  dst_linesize/2, linesize, width,
297  FFMIN(8, height + 8 - y), MAX_LEVEL - p->log2_count,
298  ldither, depth);
299  }
300  }
301  }
302 }
303 
305 {
306  static const enum PixelFormat pix_fmts[] = {
321  };
323  return 0;
324 }
325 
326 static int config_input(AVFilterLink *inlink)
327 {
328  SPPContext *spp = inlink->dst->priv;
329  const int h = FFALIGN(inlink->h + 16, 16);
330  const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(inlink->format);
331  const int bps = desc->comp[0].depth_minus1 + 1;
332 
333  av_opt_set_int(spp->dct, "bits_per_sample", bps, 0);
334  avcodec_dct_init(spp->dct);
335 
336  if (ARCH_X86)
337  ff_spp_init_x86(spp);
338 
339  spp->hsub = desc->log2_chroma_w;
340  spp->vsub = desc->log2_chroma_h;
341  spp->temp_linesize = FFALIGN(inlink->w + 16, 16);
342  spp->temp = av_malloc_array(spp->temp_linesize, h * sizeof(*spp->temp));
343  spp->src = av_malloc_array(spp->temp_linesize, h * sizeof(*spp->src) * 2);
344 
345  if (!spp->temp || !spp->src)
346  return AVERROR(ENOMEM);
347  return 0;
348 }
349 
350 static int filter_frame(AVFilterLink *inlink, AVFrame *in)
351 {
352  AVFilterContext *ctx = inlink->dst;
353  SPPContext *spp = ctx->priv;
354  AVFilterLink *outlink = ctx->outputs[0];
355  AVFrame *out = in;
356  int qp_stride = 0;
357  const int8_t *qp_table = NULL;
358  const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(inlink->format);
359  const int depth = desc->comp[0].depth_minus1 + 1;
360 
361  /* if we are not in a constant user quantizer mode and we don't want to use
362  * the quantizers from the B-frames (B-frames often have a higher QP), we
363  * need to save the qp table from the last non B-frame; this is what the
364  * following code block does */
365  if (!spp->qp) {
366  qp_table = av_frame_get_qp_table(in, &qp_stride, &spp->qscale_type);
367 
368  if (qp_table && !spp->use_bframe_qp && in->pict_type != AV_PICTURE_TYPE_B) {
369  int w, h;
370 
371  /* if the qp stride is not set, it means the QP are only defined on
372  * a line basis */
373  if (!qp_stride) {
374  w = FF_CEIL_RSHIFT(inlink->w, 4);
375  h = 1;
376  } else {
377  w = qp_stride;
378  h = FF_CEIL_RSHIFT(inlink->h, 4);
379  }
380 
381  if (w * h > spp->non_b_qp_alloc_size) {
382  int ret = av_reallocp_array(&spp->non_b_qp_table, w, h);
383  if (ret < 0) {
384  spp->non_b_qp_alloc_size = 0;
385  return ret;
386  }
387  spp->non_b_qp_alloc_size = w * h;
388  }
389 
390  av_assert0(w * h <= spp->non_b_qp_alloc_size);
391  memcpy(spp->non_b_qp_table, qp_table, w * h);
392  }
393  }
394 
395  if (spp->log2_count && !ctx->is_disabled) {
396  if (!spp->use_bframe_qp && spp->non_b_qp_table)
397  qp_table = spp->non_b_qp_table;
398 
399  if (qp_table || spp->qp) {
400  const int cw = FF_CEIL_RSHIFT(inlink->w, spp->hsub);
401  const int ch = FF_CEIL_RSHIFT(inlink->h, spp->vsub);
402 
403  /* get a new frame if in-place is not possible or if the dimensions
404  * are not multiple of 8 */
405  if (!av_frame_is_writable(in) || (inlink->w & 7) || (inlink->h & 7)) {
406  const int aligned_w = FFALIGN(inlink->w, 8);
407  const int aligned_h = FFALIGN(inlink->h, 8);
408 
409  out = ff_get_video_buffer(outlink, aligned_w, aligned_h);
410  if (!out) {
411  av_frame_free(&in);
412  return AVERROR(ENOMEM);
413  }
414  av_frame_copy_props(out, in);
415  out->width = in->width;
416  out->height = in->height;
417  }
418 
419  filter(spp, out->data[0], in->data[0], out->linesize[0], in->linesize[0], inlink->w, inlink->h, qp_table, qp_stride, 1, depth);
420 
421  if (out->data[2]) {
422  filter(spp, out->data[1], in->data[1], out->linesize[1], in->linesize[1], cw, ch, qp_table, qp_stride, 0, depth);
423  filter(spp, out->data[2], in->data[2], out->linesize[2], in->linesize[2], cw, ch, qp_table, qp_stride, 0, depth);
424  }
425  emms_c();
426  }
427  }
428 
429  if (in != out) {
430  if (in->data[3])
431  av_image_copy_plane(out->data[3], out->linesize[3],
432  in ->data[3], in ->linesize[3],
433  inlink->w, inlink->h);
434  av_frame_free(&in);
435  }
436  return ff_filter_frame(outlink, out);
437 }
438 
439 static int process_command(AVFilterContext *ctx, const char *cmd, const char *args,
440  char *res, int res_len, int flags)
441 {
442  SPPContext *spp = ctx->priv;
443 
444  if (!strcmp(cmd, "level")) {
445  if (!strcmp(args, "max"))
446  spp->log2_count = MAX_LEVEL;
447  else
448  spp->log2_count = av_clip(strtol(args, NULL, 10), 0, MAX_LEVEL);
449  return 0;
450  }
451  return AVERROR(ENOSYS);
452 }
453 
455 {
456  SPPContext *spp = ctx->priv;
457  int ret;
458 
459  spp->avctx = avcodec_alloc_context3(NULL);
460  spp->dct = avcodec_dct_alloc();
461  if (!spp->avctx || !spp->dct)
462  return AVERROR(ENOMEM);
463 
464  if (opts) {
465  AVDictionaryEntry *e = NULL;
466 
467  while ((e = av_dict_get(*opts, "", e, AV_DICT_IGNORE_SUFFIX))) {
468  if ((ret = av_opt_set(spp->dct, e->key, e->value, 0)) < 0)
469  return ret;
470  }
471  av_dict_free(opts);
472  }
473 
474  spp->store_slice = store_slice_c;
475  switch (spp->mode) {
476  case MODE_HARD: spp->requantize = hardthresh_c; break;
477  case MODE_SOFT: spp->requantize = softthresh_c; break;
478  }
479  return 0;
480 }
481 
482 static av_cold void uninit(AVFilterContext *ctx)
483 {
484  SPPContext *spp = ctx->priv;
485 
486  av_freep(&spp->temp);
487  av_freep(&spp->src);
488  if (spp->avctx) {
489  avcodec_close(spp->avctx);
490  av_freep(&spp->avctx);
491  }
492  av_freep(&spp->dct);
493  av_freep(&spp->non_b_qp_table);
494 }
495 
496 static const AVFilterPad spp_inputs[] = {
497  {
498  .name = "default",
499  .type = AVMEDIA_TYPE_VIDEO,
500  .config_props = config_input,
501  .filter_frame = filter_frame,
502  },
503  { NULL }
504 };
505 
506 static const AVFilterPad spp_outputs[] = {
507  {
508  .name = "default",
509  .type = AVMEDIA_TYPE_VIDEO,
510  },
511  { NULL }
512 };
513 
515  .name = "spp",
516  .description = NULL_IF_CONFIG_SMALL("Apply a simple post processing filter."),
517  .priv_size = sizeof(SPPContext),
518  .init_dict = init_dict,
519  .uninit = uninit,
521  .inputs = spp_inputs,
522  .outputs = spp_outputs,
524  .priv_class = &spp_class,
526 };