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 "libavcodec/dsputil.h"
35 #include "libavutil/avassert.h"
36 #include "libavutil/imgutils.h"
37 #include "libavutil/opt.h"
38 #include "libavutil/pixdesc.h"
39 #include "internal.h"
40 #include "vf_spp.h"
41 
42 enum mode {
46 };
47 
48 #define OFFSET(x) offsetof(SPPContext, x)
49 #define FLAGS AV_OPT_FLAG_FILTERING_PARAM|AV_OPT_FLAG_VIDEO_PARAM
50 static const AVOption spp_options[] = {
51  { "quality", "set quality", OFFSET(log2_count), AV_OPT_TYPE_INT, {.i64 = 3}, 0, MAX_LEVEL, FLAGS },
52  { "qp", "force a constant quantizer parameter", OFFSET(qp), AV_OPT_TYPE_INT, {.i64 = 0}, 0, 63, FLAGS },
53  { "mode", "set thresholding mode", OFFSET(mode), AV_OPT_TYPE_INT, {.i64 = MODE_HARD}, 0, NB_MODES - 1, FLAGS, "mode" },
54  { "hard", "hard thresholding", 0, AV_OPT_TYPE_CONST, {.i64 = MODE_HARD}, INT_MIN, INT_MAX, FLAGS, "mode" },
55  { "soft", "soft thresholding", 0, AV_OPT_TYPE_CONST, {.i64 = MODE_SOFT}, INT_MIN, INT_MAX, FLAGS, "mode" },
56  { "use_bframe_qp", "use B-frames' QP", OFFSET(use_bframe_qp), AV_OPT_TYPE_INT, {.i64 = 0}, 0, 1, FLAGS },
57  { NULL }
58 };
59 
61 
62 // XXX: share between filters?
63 DECLARE_ALIGNED(8, static const uint8_t, ldither)[8][8] = {
64  { 0, 48, 12, 60, 3, 51, 15, 63 },
65  { 32, 16, 44, 28, 35, 19, 47, 31 },
66  { 8, 56, 4, 52, 11, 59, 7, 55 },
67  { 40, 24, 36, 20, 43, 27, 39, 23 },
68  { 2, 50, 14, 62, 1, 49, 13, 61 },
69  { 34, 18, 46, 30, 33, 17, 45, 29 },
70  { 10, 58, 6, 54, 9, 57, 5, 53 },
71  { 42, 26, 38, 22, 41, 25, 37, 21 },
72 };
73 
74 static const uint8_t offset[127][2] = {
75  {0,0},
76  {0,0}, {4,4}, // quality = 1
77  {0,0}, {2,2}, {6,4}, {4,6}, // quality = 2
78  {0,0}, {5,1}, {2,2}, {7,3}, {4,4}, {1,5}, {6,6}, {3,7}, // quality = 3
79 
80  {0,0}, {4,0}, {1,1}, {5,1}, {3,2}, {7,2}, {2,3}, {6,3}, // quality = 4
81  {0,4}, {4,4}, {1,5}, {5,5}, {3,6}, {7,6}, {2,7}, {6,7},
82 
83  {0,0}, {0,2}, {0,4}, {0,6}, {1,1}, {1,3}, {1,5}, {1,7}, // quality = 5
84  {2,0}, {2,2}, {2,4}, {2,6}, {3,1}, {3,3}, {3,5}, {3,7},
85  {4,0}, {4,2}, {4,4}, {4,6}, {5,1}, {5,3}, {5,5}, {5,7},
86  {6,0}, {6,2}, {6,4}, {6,6}, {7,1}, {7,3}, {7,5}, {7,7},
87 
88  {0,0}, {4,4}, {0,4}, {4,0}, {2,2}, {6,6}, {2,6}, {6,2}, // quality = 6
89  {0,2}, {4,6}, {0,6}, {4,2}, {2,0}, {6,4}, {2,4}, {6,0},
90  {1,1}, {5,5}, {1,5}, {5,1}, {3,3}, {7,7}, {3,7}, {7,3},
91  {1,3}, {5,7}, {1,7}, {5,3}, {3,1}, {7,5}, {3,5}, {7,1},
92  {0,1}, {4,5}, {0,5}, {4,1}, {2,3}, {6,7}, {2,7}, {6,3},
93  {0,3}, {4,7}, {0,7}, {4,3}, {2,1}, {6,5}, {2,5}, {6,1},
94  {1,0}, {5,4}, {1,4}, {5,0}, {3,2}, {7,6}, {3,6}, {7,2},
95  {1,2}, {5,6}, {1,6}, {5,2}, {3,0}, {7,4}, {3,4}, {7,0},
96 };
97 
98 static void hardthresh_c(int16_t dst[64], const int16_t src[64],
99  int qp, const uint8_t *permutation)
100 {
101  int i;
102  int bias = 0; // FIXME
103 
104  unsigned threshold1 = qp * ((1<<4) - bias) - 1;
105  unsigned threshold2 = threshold1 << 1;
106 
107  memset(dst, 0, 64 * sizeof(dst[0]));
108  dst[0] = (src[0] + 4) >> 3;
109 
110  for (i = 1; i < 64; i++) {
111  int level = src[i];
112  if (((unsigned)(level + threshold1)) > threshold2) {
113  const int j = permutation[i];
114  dst[j] = (level + 4) >> 3;
115  }
116  }
117 }
118 
119 static void softthresh_c(int16_t dst[64], const int16_t src[64],
120  int qp, const uint8_t *permutation)
121 {
122  int i;
123  int bias = 0; //FIXME
124 
125  unsigned threshold1 = qp * ((1<<4) - bias) - 1;
126  unsigned threshold2 = threshold1 << 1;
127 
128  memset(dst, 0, 64 * sizeof(dst[0]));
129  dst[0] = (src[0] + 4) >> 3;
130 
131  for (i = 1; i < 64; i++) {
132  int level = src[i];
133  if (((unsigned)(level + threshold1)) > threshold2) {
134  const int j = permutation[i];
135  if (level > 0) dst[j] = (level - threshold1 + 4) >> 3;
136  else dst[j] = (level + threshold1 + 4) >> 3;
137  }
138  }
139 }
140 
141 static void store_slice_c(uint8_t *dst, const int16_t *src,
142  int dst_linesize, int src_linesize,
143  int width, int height, int log2_scale,
144  const uint8_t dither[8][8])
145 {
146  int y, x;
147 
148 #define STORE(pos) do { \
149  temp = ((src[x + y*src_linesize + pos] << log2_scale) + d[pos]) >> 6; \
150  if (temp & 0x100) \
151  temp = ~(temp >> 31); \
152  dst[x + y*dst_linesize + pos] = temp; \
153 } while (0)
154 
155  for (y = 0; y < height; y++) {
156  const uint8_t *d = dither[y];
157  for (x = 0; x < width; x += 8) {
158  int temp;
159  STORE(0);
160  STORE(1);
161  STORE(2);
162  STORE(3);
163  STORE(4);
164  STORE(5);
165  STORE(6);
166  STORE(7);
167  }
168  }
169 }
170 
171 static inline void add_block(int16_t *dst, int linesize, const int16_t block[64])
172 {
173  int y;
174 
175  for (y = 0; y < 8; y++) {
176  *(uint32_t *)&dst[0 + y*linesize] += *(uint32_t *)&block[0 + y*8];
177  *(uint32_t *)&dst[2 + y*linesize] += *(uint32_t *)&block[2 + y*8];
178  *(uint32_t *)&dst[4 + y*linesize] += *(uint32_t *)&block[4 + y*8];
179  *(uint32_t *)&dst[6 + y*linesize] += *(uint32_t *)&block[6 + y*8];
180  }
181 }
182 
183 // XXX: export the function?
184 static inline int norm_qscale(int qscale, int type)
185 {
186  switch (type) {
187  case FF_QSCALE_TYPE_MPEG1: return qscale;
188  case FF_QSCALE_TYPE_MPEG2: return qscale >> 1;
189  case FF_QSCALE_TYPE_H264: return qscale >> 2;
190  case FF_QSCALE_TYPE_VP56: return (63 - qscale + 2) >> 2;
191  }
192  return qscale;
193 }
194 
195 static void filter(SPPContext *p, uint8_t *dst, uint8_t *src,
196  int dst_linesize, int src_linesize, int width, int height,
197  const uint8_t *qp_table, int qp_stride, int is_luma)
198 {
199  int x, y, i;
200  const int count = 1 << p->log2_count;
201  const int linesize = is_luma ? p->temp_linesize : FFALIGN(width+16, 16);
202  DECLARE_ALIGNED(16, uint64_t, block_align)[32];
203  int16_t *block = (int16_t *)block_align;
204  int16_t *block2 = (int16_t *)(block_align + 16);
205 
206  for (y = 0; y < height; y++) {
207  int index = 8 + 8*linesize + y*linesize;
208  memcpy(p->src + index, src + y*src_linesize, width);
209  for (x = 0; x < 8; x++) {
210  p->src[index - x - 1] = p->src[index + x ];
211  p->src[index + width + x ] = p->src[index + width - x - 1];
212  }
213  }
214  for (y = 0; y < 8; y++) {
215  memcpy(p->src + ( 7-y)*linesize, p->src + ( y+8)*linesize, linesize);
216  memcpy(p->src + (height+8+y)*linesize, p->src + (height-y+7)*linesize, linesize);
217  }
218 
219  for (y = 0; y < height + 8; y += 8) {
220  memset(p->temp + (8 + y) * linesize, 0, 8 * linesize * sizeof(*p->temp));
221  for (x = 0; x < width + 8; x += 8) {
222  int qp;
223 
224  if (p->qp) {
225  qp = p->qp;
226  } else{
227  const int qps = 3 + is_luma;
228  qp = qp_table[(FFMIN(x, width - 1) >> qps) + (FFMIN(y, height - 1) >> qps) * qp_stride];
229  qp = FFMAX(1, norm_qscale(qp, p->qscale_type));
230  }
231  for (i = 0; i < count; i++) {
232  const int x1 = x + offset[i + count - 1][0];
233  const int y1 = y + offset[i + count - 1][1];
234  const int index = x1 + y1*linesize;
235  p->pdsp.get_pixels(block, p->src + index, linesize);
236  p->fdsp.fdct(block);
237  p->requantize(block2, block, qp, p->idsp.idct_permutation);
238  p->idsp.idct(block2);
239  add_block(p->temp + index, linesize, block2);
240  }
241  }
242  if (y)
243  p->store_slice(dst + (y - 8) * dst_linesize, p->temp + 8 + y*linesize,
244  dst_linesize, linesize, width,
245  FFMIN(8, height + 8 - y), MAX_LEVEL - p->log2_count,
246  ldither);
247  }
248 }
249 
251 {
252  static const enum PixelFormat pix_fmts[] = {
259  };
261  return 0;
262 }
263 
264 static int config_input(AVFilterLink *inlink)
265 {
266  SPPContext *spp = inlink->dst->priv;
267  const int h = FFALIGN(inlink->h + 16, 16);
268  const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(inlink->format);
269 
270  spp->hsub = desc->log2_chroma_w;
271  spp->vsub = desc->log2_chroma_h;
272  spp->temp_linesize = FFALIGN(inlink->w + 16, 16);
273  spp->temp = av_malloc_array(spp->temp_linesize, h * sizeof(*spp->temp));
274  spp->src = av_malloc_array(spp->temp_linesize, h * sizeof(*spp->src));
275  if (!spp->use_bframe_qp) {
276  /* we are assuming here the qp blocks will not be smaller that 16x16 */
277  spp->non_b_qp_alloc_size = FF_CEIL_RSHIFT(inlink->w, 4) * FF_CEIL_RSHIFT(inlink->h, 4);
278  spp->non_b_qp_table = av_calloc(spp->non_b_qp_alloc_size, sizeof(*spp->non_b_qp_table));
279  if (!spp->non_b_qp_table)
280  return AVERROR(ENOMEM);
281  }
282  if (!spp->temp || !spp->src)
283  return AVERROR(ENOMEM);
284  return 0;
285 }
286 
287 static int filter_frame(AVFilterLink *inlink, AVFrame *in)
288 {
289  AVFilterContext *ctx = inlink->dst;
290  SPPContext *spp = ctx->priv;
291  AVFilterLink *outlink = ctx->outputs[0];
292  AVFrame *out = in;
293  int qp_stride = 0;
294  const int8_t *qp_table = NULL;
295 
296  /* if we are not in a constant user quantizer mode and we don't want to use
297  * the quantizers from the B-frames (B-frames often have a higher QP), we
298  * need to save the qp table from the last non B-frame; this is what the
299  * following code block does */
300  if (!spp->qp) {
301  qp_table = av_frame_get_qp_table(in, &qp_stride, &spp->qscale_type);
302 
303  if (qp_table && !spp->use_bframe_qp && in->pict_type != AV_PICTURE_TYPE_B) {
304  int w, h;
305 
306  /* if the qp stride is not set, it means the QP are only defined on
307  * a line basis */
308  if (!qp_stride) {
309  w = FF_CEIL_RSHIFT(inlink->w, 4);
310  h = 1;
311  } else {
312  w = FF_CEIL_RSHIFT(qp_stride, 4);
313  h = FF_CEIL_RSHIFT(inlink->h, 4);
314  }
315  av_assert0(w * h <= spp->non_b_qp_alloc_size);
316  memcpy(spp->non_b_qp_table, qp_table, w * h);
317  }
318  }
319 
320  if (spp->log2_count && !ctx->is_disabled) {
321  if (!spp->use_bframe_qp && spp->non_b_qp_table)
322  qp_table = spp->non_b_qp_table;
323 
324  if (qp_table || spp->qp) {
325  const int cw = FF_CEIL_RSHIFT(inlink->w, spp->hsub);
326  const int ch = FF_CEIL_RSHIFT(inlink->h, spp->vsub);
327 
328  /* get a new frame if in-place is not possible or if the dimensions
329  * are not multiple of 8 */
330  if (!av_frame_is_writable(in) || (inlink->w & 7) || (inlink->h & 7)) {
331  const int aligned_w = FFALIGN(inlink->w, 8);
332  const int aligned_h = FFALIGN(inlink->h, 8);
333 
334  out = ff_get_video_buffer(outlink, aligned_w, aligned_h);
335  if (!out) {
336  av_frame_free(&in);
337  return AVERROR(ENOMEM);
338  }
339  av_frame_copy_props(out, in);
340  out->width = in->width;
341  out->height = in->height;
342  }
343 
344  filter(spp, out->data[0], in->data[0], out->linesize[0], in->linesize[0], inlink->w, inlink->h, qp_table, qp_stride, 1);
345  filter(spp, out->data[1], in->data[1], out->linesize[1], in->linesize[1], cw, ch, qp_table, qp_stride, 0);
346  filter(spp, out->data[2], in->data[2], out->linesize[2], in->linesize[2], cw, ch, qp_table, qp_stride, 0);
347  emms_c();
348  }
349  }
350 
351  if (in != out) {
352  if (in->data[3])
353  av_image_copy_plane(out->data[3], out->linesize[3],
354  in ->data[3], in ->linesize[3],
355  inlink->w, inlink->h);
356  av_frame_free(&in);
357  }
358  return ff_filter_frame(outlink, out);
359 }
360 
361 static int process_command(AVFilterContext *ctx, const char *cmd, const char *args,
362  char *res, int res_len, int flags)
363 {
364  SPPContext *spp = ctx->priv;
365 
366  if (!strcmp(cmd, "level")) {
367  if (!strcmp(args, "max"))
368  spp->log2_count = MAX_LEVEL;
369  else
370  spp->log2_count = av_clip(strtol(args, NULL, 10), 0, MAX_LEVEL);
371  return 0;
372  }
373  return AVERROR(ENOSYS);
374 }
375 
376 static av_cold int init(AVFilterContext *ctx)
377 {
378  SPPContext *spp = ctx->priv;
379 
380  spp->avctx = avcodec_alloc_context3(NULL);
381  if (!spp->avctx)
382  return AVERROR(ENOMEM);
383  ff_idctdsp_init(&spp->idsp, spp->avctx);
384  ff_fdctdsp_init(&spp->fdsp, spp->avctx);
385  ff_pixblockdsp_init(&spp->pdsp, spp->avctx);
386  spp->store_slice = store_slice_c;
387  switch (spp->mode) {
388  case MODE_HARD: spp->requantize = hardthresh_c; break;
389  case MODE_SOFT: spp->requantize = softthresh_c; break;
390  }
391  if (ARCH_X86)
392  ff_spp_init_x86(spp);
393  return 0;
394 }
395 
396 static av_cold void uninit(AVFilterContext *ctx)
397 {
398  SPPContext *spp = ctx->priv;
399 
400  av_freep(&spp->temp);
401  av_freep(&spp->src);
402  if (spp->avctx) {
403  avcodec_close(spp->avctx);
404  av_freep(&spp->avctx);
405  }
406  av_freep(&spp->non_b_qp_table);
407 }
408 
409 static const AVFilterPad spp_inputs[] = {
410  {
411  .name = "default",
412  .type = AVMEDIA_TYPE_VIDEO,
413  .config_props = config_input,
414  .filter_frame = filter_frame,
415  },
416  { NULL }
417 };
418 
419 static const AVFilterPad spp_outputs[] = {
420  {
421  .name = "default",
422  .type = AVMEDIA_TYPE_VIDEO,
423  },
424  { NULL }
425 };
426 
428  .name = "spp",
429  .description = NULL_IF_CONFIG_SMALL("Apply a simple post processing filter."),
430  .priv_size = sizeof(SPPContext),
431  .init = init,
432  .uninit = uninit,
434  .inputs = spp_inputs,
435  .outputs = spp_outputs,
437  .priv_class = &spp_class,
439 };