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 {
42 MODE_HARD,
43 MODE_SOFT,
44 NB_MODES
45 };
46
child_class_next(const AVClass * prev)47 static const AVClass *child_class_next(const AVClass *prev)
48 {
49 return prev ? NULL : avcodec_dct_get_class();
50 }
51
child_next(void * obj,void * prev)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 #define TFLAGS AV_OPT_FLAG_FILTERING_PARAM|AV_OPT_FLAG_VIDEO_PARAM|AV_OPT_FLAG_RUNTIME_PARAM
61 static const AVOption spp_options[] = {
62 { "quality", "set quality", OFFSET(log2_count), AV_OPT_TYPE_INT, {.i64 = 3}, 0, MAX_LEVEL, TFLAGS },
63 { "qp", "force a constant quantizer parameter", OFFSET(qp), AV_OPT_TYPE_INT, {.i64 = 0}, 0, 63, FLAGS },
64 { "mode", "set thresholding mode", OFFSET(mode), AV_OPT_TYPE_INT, {.i64 = MODE_HARD}, 0, NB_MODES - 1, FLAGS, "mode" },
65 { "hard", "hard thresholding", 0, AV_OPT_TYPE_CONST, {.i64 = MODE_HARD}, INT_MIN, INT_MAX, FLAGS, "mode" },
66 { "soft", "soft thresholding", 0, AV_OPT_TYPE_CONST, {.i64 = MODE_SOFT}, INT_MIN, INT_MAX, FLAGS, "mode" },
67 { "use_bframe_qp", "use B-frames' QP", OFFSET(use_bframe_qp), AV_OPT_TYPE_BOOL, {.i64 = 0}, 0, 1, FLAGS },
68 { NULL }
69 };
70
71 static const AVClass spp_class = {
72 .class_name = "spp",
73 .item_name = av_default_item_name,
74 .option = spp_options,
75 .version = LIBAVUTIL_VERSION_INT,
76 .category = AV_CLASS_CATEGORY_FILTER,
77 .child_class_next = child_class_next,
78 .child_next = child_next,
79 };
80
81 // XXX: share between filters?
82 DECLARE_ALIGNED(8, static const uint8_t, ldither)[8][8] = {
83 { 0, 48, 12, 60, 3, 51, 15, 63 },
84 { 32, 16, 44, 28, 35, 19, 47, 31 },
85 { 8, 56, 4, 52, 11, 59, 7, 55 },
86 { 40, 24, 36, 20, 43, 27, 39, 23 },
87 { 2, 50, 14, 62, 1, 49, 13, 61 },
88 { 34, 18, 46, 30, 33, 17, 45, 29 },
89 { 10, 58, 6, 54, 9, 57, 5, 53 },
90 { 42, 26, 38, 22, 41, 25, 37, 21 },
91 };
92
93 static const uint8_t offset[127][2] = {
94 {0,0},
95 {0,0}, {4,4}, // quality = 1
96 {0,0}, {2,2}, {6,4}, {4,6}, // quality = 2
97 {0,0}, {5,1}, {2,2}, {7,3}, {4,4}, {1,5}, {6,6}, {3,7}, // quality = 3
98
99 {0,0}, {4,0}, {1,1}, {5,1}, {3,2}, {7,2}, {2,3}, {6,3}, // quality = 4
100 {0,4}, {4,4}, {1,5}, {5,5}, {3,6}, {7,6}, {2,7}, {6,7},
101
102 {0,0}, {0,2}, {0,4}, {0,6}, {1,1}, {1,3}, {1,5}, {1,7}, // quality = 5
103 {2,0}, {2,2}, {2,4}, {2,6}, {3,1}, {3,3}, {3,5}, {3,7},
104 {4,0}, {4,2}, {4,4}, {4,6}, {5,1}, {5,3}, {5,5}, {5,7},
105 {6,0}, {6,2}, {6,4}, {6,6}, {7,1}, {7,3}, {7,5}, {7,7},
106
107 {0,0}, {4,4}, {0,4}, {4,0}, {2,2}, {6,6}, {2,6}, {6,2}, // quality = 6
108 {0,2}, {4,6}, {0,6}, {4,2}, {2,0}, {6,4}, {2,4}, {6,0},
109 {1,1}, {5,5}, {1,5}, {5,1}, {3,3}, {7,7}, {3,7}, {7,3},
110 {1,3}, {5,7}, {1,7}, {5,3}, {3,1}, {7,5}, {3,5}, {7,1},
111 {0,1}, {4,5}, {0,5}, {4,1}, {2,3}, {6,7}, {2,7}, {6,3},
112 {0,3}, {4,7}, {0,7}, {4,3}, {2,1}, {6,5}, {2,5}, {6,1},
113 {1,0}, {5,4}, {1,4}, {5,0}, {3,2}, {7,6}, {3,6}, {7,2},
114 {1,2}, {5,6}, {1,6}, {5,2}, {3,0}, {7,4}, {3,4}, {7,0},
115 };
116
hardthresh_c(int16_t dst[64],const int16_t src[64],int qp,const uint8_t * permutation)117 static void hardthresh_c(int16_t dst[64], const int16_t src[64],
118 int qp, const uint8_t *permutation)
119 {
120 int i;
121 int bias = 0; // FIXME
122
123 unsigned threshold1 = qp * ((1<<4) - bias) - 1;
124 unsigned threshold2 = threshold1 << 1;
125
126 memset(dst, 0, 64 * sizeof(dst[0]));
127 dst[0] = (src[0] + 4) >> 3;
128
129 for (i = 1; i < 64; i++) {
130 int level = src[i];
131 if (((unsigned)(level + threshold1)) > threshold2) {
132 const int j = permutation[i];
133 dst[j] = (level + 4) >> 3;
134 }
135 }
136 }
137
softthresh_c(int16_t dst[64],const int16_t src[64],int qp,const uint8_t * permutation)138 static void softthresh_c(int16_t dst[64], const int16_t src[64],
139 int qp, const uint8_t *permutation)
140 {
141 int i;
142 int bias = 0; //FIXME
143
144 unsigned threshold1 = qp * ((1<<4) - bias) - 1;
145 unsigned threshold2 = threshold1 << 1;
146
147 memset(dst, 0, 64 * sizeof(dst[0]));
148 dst[0] = (src[0] + 4) >> 3;
149
150 for (i = 1; i < 64; i++) {
151 int level = src[i];
152 if (((unsigned)(level + threshold1)) > threshold2) {
153 const int j = permutation[i];
154 if (level > 0) dst[j] = (level - threshold1 + 4) >> 3;
155 else dst[j] = (level + threshold1 + 4) >> 3;
156 }
157 }
158 }
159
store_slice_c(uint8_t * dst,const int16_t * src,int dst_linesize,int src_linesize,int width,int height,int log2_scale,const uint8_t dither[8][8])160 static void store_slice_c(uint8_t *dst, const int16_t *src,
161 int dst_linesize, int src_linesize,
162 int width, int height, int log2_scale,
163 const uint8_t dither[8][8])
164 {
165 int y, x;
166
167 #define STORE(pos) do { \
168 temp = ((src[x + y*src_linesize + pos] << log2_scale) + d[pos]) >> 6; \
169 if (temp & 0x100) \
170 temp = ~(temp >> 31); \
171 dst[x + y*dst_linesize + pos] = temp; \
172 } while (0)
173
174 for (y = 0; y < height; y++) {
175 const uint8_t *d = dither[y];
176 for (x = 0; x < width; x += 8) {
177 int temp;
178 STORE(0);
179 STORE(1);
180 STORE(2);
181 STORE(3);
182 STORE(4);
183 STORE(5);
184 STORE(6);
185 STORE(7);
186 }
187 }
188 }
189
store_slice16_c(uint16_t * dst,const int16_t * src,int dst_linesize,int src_linesize,int width,int height,int log2_scale,const uint8_t dither[8][8],int depth)190 static void store_slice16_c(uint16_t *dst, const int16_t *src,
191 int dst_linesize, int src_linesize,
192 int width, int height, int log2_scale,
193 const uint8_t dither[8][8], int depth)
194 {
195 int y, x;
196 unsigned int mask = -1<<depth;
197
198 #define STORE16(pos) do { \
199 temp = ((src[x + y*src_linesize + pos] << log2_scale) + (d[pos]>>1)) >> 5; \
200 if (temp & mask ) \
201 temp = ~(temp >> 31); \
202 dst[x + y*dst_linesize + pos] = temp; \
203 } while (0)
204
205 for (y = 0; y < height; y++) {
206 const uint8_t *d = dither[y];
207 for (x = 0; x < width; x += 8) {
208 int temp;
209 STORE16(0);
210 STORE16(1);
211 STORE16(2);
212 STORE16(3);
213 STORE16(4);
214 STORE16(5);
215 STORE16(6);
216 STORE16(7);
217 }
218 }
219 }
220
add_block(uint16_t * dst,int linesize,const int16_t block[64])221 static inline void add_block(uint16_t *dst, int linesize, const int16_t block[64])
222 {
223 int y;
224
225 for (y = 0; y < 8; y++) {
226 dst[0 + y*linesize] += block[0 + y*8];
227 dst[1 + y*linesize] += block[1 + y*8];
228 dst[2 + y*linesize] += block[2 + y*8];
229 dst[3 + y*linesize] += block[3 + y*8];
230 dst[4 + y*linesize] += block[4 + y*8];
231 dst[5 + y*linesize] += block[5 + y*8];
232 dst[6 + y*linesize] += block[6 + y*8];
233 dst[7 + y*linesize] += block[7 + y*8];
234 }
235 }
236
filter(SPPContext * p,uint8_t * dst,uint8_t * src,int dst_linesize,int src_linesize,int width,int height,const uint8_t * qp_table,int qp_stride,int is_luma,int depth)237 static void filter(SPPContext *p, uint8_t *dst, uint8_t *src,
238 int dst_linesize, int src_linesize, int width, int height,
239 const uint8_t *qp_table, int qp_stride, int is_luma, int depth)
240 {
241 int x, y, i;
242 const int count = 1 << p->log2_count;
243 const int linesize = is_luma ? p->temp_linesize : FFALIGN(width+16, 16);
244 DECLARE_ALIGNED(16, uint64_t, block_align)[32];
245 int16_t *block = (int16_t *)block_align;
246 int16_t *block2 = (int16_t *)(block_align + 16);
247 uint16_t *psrc16 = (uint16_t*)p->src;
248 const int sample_bytes = (depth+7) / 8;
249
250 for (y = 0; y < height; y++) {
251 int index = 8 + 8*linesize + y*linesize;
252 memcpy(p->src + index*sample_bytes, src + y*src_linesize, width*sample_bytes);
253 if (sample_bytes == 1) {
254 for (x = 0; x < 8; x++) {
255 p->src[index - x - 1] = p->src[index + x ];
256 p->src[index + width + x ] = p->src[index + width - x - 1];
257 }
258 } else {
259 for (x = 0; x < 8; x++) {
260 psrc16[index - x - 1] = psrc16[index + x ];
261 psrc16[index + width + x ] = psrc16[index + width - x - 1];
262 }
263 }
264 }
265 for (y = 0; y < 8; y++) {
266 memcpy(p->src + ( 7-y)*linesize * sample_bytes, p->src + ( y+8)*linesize * sample_bytes, linesize * sample_bytes);
267 memcpy(p->src + (height+8+y)*linesize * sample_bytes, p->src + (height-y+7)*linesize * sample_bytes, linesize * sample_bytes);
268 }
269
270 for (y = 0; y < height + 8; y += 8) {
271 memset(p->temp + (8 + y) * linesize, 0, 8 * linesize * sizeof(*p->temp));
272 for (x = 0; x < width + 8; x += 8) {
273 int qp;
274
275 if (p->qp) {
276 qp = p->qp;
277 } else{
278 const int qps = 3 + is_luma;
279 qp = qp_table[(FFMIN(x, width - 1) >> qps) + (FFMIN(y, height - 1) >> qps) * qp_stride];
280 qp = FFMAX(1, ff_norm_qscale(qp, p->qscale_type));
281 }
282 for (i = 0; i < count; i++) {
283 const int x1 = x + offset[i + count - 1][0];
284 const int y1 = y + offset[i + count - 1][1];
285 const int index = x1 + y1*linesize;
286 p->dct->get_pixels_unaligned(block, p->src + sample_bytes*index, sample_bytes*linesize);
287 p->dct->fdct(block);
288 p->requantize(block2, block, qp, p->dct->idct_permutation);
289 p->dct->idct(block2);
290 add_block(p->temp + index, linesize, block2);
291 }
292 }
293 if (y) {
294 if (sample_bytes == 1) {
295 p->store_slice(dst + (y - 8) * dst_linesize, p->temp + 8 + y*linesize,
296 dst_linesize, linesize, width,
297 FFMIN(8, height + 8 - y), MAX_LEVEL - p->log2_count,
298 ldither);
299 } else {
300 store_slice16_c((uint16_t*)(dst + (y - 8) * dst_linesize), p->temp + 8 + y*linesize,
301 dst_linesize/2, linesize, width,
302 FFMIN(8, height + 8 - y), MAX_LEVEL - p->log2_count,
303 ldither, depth);
304 }
305 }
306 }
307 }
308
query_formats(AVFilterContext * ctx)309 static int query_formats(AVFilterContext *ctx)
310 {
311 static const enum AVPixelFormat pix_fmts[] = {
312 AV_PIX_FMT_YUV444P, AV_PIX_FMT_YUV422P,
313 AV_PIX_FMT_YUV420P, AV_PIX_FMT_YUV411P,
314 AV_PIX_FMT_YUV410P, AV_PIX_FMT_YUV440P,
315 AV_PIX_FMT_YUVJ444P, AV_PIX_FMT_YUVJ422P,
316 AV_PIX_FMT_YUVJ420P, AV_PIX_FMT_YUVJ440P,
317 AV_PIX_FMT_YUV444P10, AV_PIX_FMT_YUV422P10,
318 AV_PIX_FMT_YUV420P10,
319 AV_PIX_FMT_YUV444P9, AV_PIX_FMT_YUV422P9,
320 AV_PIX_FMT_YUV420P9,
321 AV_PIX_FMT_GRAY8,
322 AV_PIX_FMT_GBRP,
323 AV_PIX_FMT_GBRP9,
324 AV_PIX_FMT_GBRP10,
325 AV_PIX_FMT_NONE
326 };
327
328 AVFilterFormats *fmts_list = ff_make_format_list(pix_fmts);
329 if (!fmts_list)
330 return AVERROR(ENOMEM);
331 return ff_set_common_formats(ctx, fmts_list);
332 }
333
config_input(AVFilterLink * inlink)334 static int config_input(AVFilterLink *inlink)
335 {
336 SPPContext *s = inlink->dst->priv;
337 const int h = FFALIGN(inlink->h + 16, 16);
338 const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(inlink->format);
339 const int bps = desc->comp[0].depth;
340
341 av_opt_set_int(s->dct, "bits_per_sample", bps, 0);
342 avcodec_dct_init(s->dct);
343
344 if (ARCH_X86)
345 ff_spp_init_x86(s);
346
347 s->hsub = desc->log2_chroma_w;
348 s->vsub = desc->log2_chroma_h;
349 s->temp_linesize = FFALIGN(inlink->w + 16, 16);
350 s->temp = av_malloc_array(s->temp_linesize, h * sizeof(*s->temp));
351 s->src = av_malloc_array(s->temp_linesize, h * sizeof(*s->src) * 2);
352
353 if (!s->temp || !s->src)
354 return AVERROR(ENOMEM);
355 return 0;
356 }
357
filter_frame(AVFilterLink * inlink,AVFrame * in)358 static int filter_frame(AVFilterLink *inlink, AVFrame *in)
359 {
360 AVFilterContext *ctx = inlink->dst;
361 SPPContext *s = ctx->priv;
362 AVFilterLink *outlink = ctx->outputs[0];
363 AVFrame *out = in;
364 int qp_stride = 0;
365 const int8_t *qp_table = NULL;
366 const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(inlink->format);
367 const int depth = desc->comp[0].depth;
368
369 /* if we are not in a constant user quantizer mode and we don't want to use
370 * the quantizers from the B-frames (B-frames often have a higher QP), we
371 * need to save the qp table from the last non B-frame; this is what the
372 * following code block does */
373 if (!s->qp) {
374 qp_table = av_frame_get_qp_table(in, &qp_stride, &s->qscale_type);
375
376 if (qp_table && !s->use_bframe_qp && in->pict_type != AV_PICTURE_TYPE_B) {
377 int w, h;
378
379 /* if the qp stride is not set, it means the QP are only defined on
380 * a line basis */
381 if (!qp_stride) {
382 w = AV_CEIL_RSHIFT(inlink->w, 4);
383 h = 1;
384 } else {
385 w = qp_stride;
386 h = AV_CEIL_RSHIFT(inlink->h, 4);
387 }
388
389 if (w * h > s->non_b_qp_alloc_size) {
390 int ret = av_reallocp_array(&s->non_b_qp_table, w, h);
391 if (ret < 0) {
392 s->non_b_qp_alloc_size = 0;
393 return ret;
394 }
395 s->non_b_qp_alloc_size = w * h;
396 }
397
398 av_assert0(w * h <= s->non_b_qp_alloc_size);
399 memcpy(s->non_b_qp_table, qp_table, w * h);
400 }
401 }
402
403 if (s->log2_count && !ctx->is_disabled) {
404 if (!s->use_bframe_qp && s->non_b_qp_table)
405 qp_table = s->non_b_qp_table;
406
407 if (qp_table || s->qp) {
408 const int cw = AV_CEIL_RSHIFT(inlink->w, s->hsub);
409 const int ch = AV_CEIL_RSHIFT(inlink->h, s->vsub);
410
411 /* get a new frame if in-place is not possible or if the dimensions
412 * are not multiple of 8 */
413 if (!av_frame_is_writable(in) || (inlink->w & 7) || (inlink->h & 7)) {
414 const int aligned_w = FFALIGN(inlink->w, 8);
415 const int aligned_h = FFALIGN(inlink->h, 8);
416
417 out = ff_get_video_buffer(outlink, aligned_w, aligned_h);
418 if (!out) {
419 av_frame_free(&in);
420 return AVERROR(ENOMEM);
421 }
422 av_frame_copy_props(out, in);
423 out->width = in->width;
424 out->height = in->height;
425 }
426
427 filter(s, out->data[0], in->data[0], out->linesize[0], in->linesize[0], inlink->w, inlink->h, qp_table, qp_stride, 1, depth);
428
429 if (out->data[2]) {
430 filter(s, out->data[1], in->data[1], out->linesize[1], in->linesize[1], cw, ch, qp_table, qp_stride, 0, depth);
431 filter(s, out->data[2], in->data[2], out->linesize[2], in->linesize[2], cw, ch, qp_table, qp_stride, 0, depth);
432 }
433 emms_c();
434 }
435 }
436
437 if (in != out) {
438 if (in->data[3])
439 av_image_copy_plane(out->data[3], out->linesize[3],
440 in ->data[3], in ->linesize[3],
441 inlink->w, inlink->h);
442 av_frame_free(&in);
443 }
444 return ff_filter_frame(outlink, out);
445 }
446
process_command(AVFilterContext * ctx,const char * cmd,const char * args,char * res,int res_len,int flags)447 static int process_command(AVFilterContext *ctx, const char *cmd, const char *args,
448 char *res, int res_len, int flags)
449 {
450 SPPContext *s = ctx->priv;
451
452 if (!strcmp(cmd, "level") || !strcmp(cmd, "quality")) {
453 if (!strcmp(args, "max"))
454 s->log2_count = MAX_LEVEL;
455 else
456 s->log2_count = av_clip(strtol(args, NULL, 10), 0, MAX_LEVEL);
457 return 0;
458 }
459 return AVERROR(ENOSYS);
460 }
461
init_dict(AVFilterContext * ctx,AVDictionary ** opts)462 static av_cold int init_dict(AVFilterContext *ctx, AVDictionary **opts)
463 {
464 SPPContext *s = ctx->priv;
465 int ret;
466
467 s->dct = avcodec_dct_alloc();
468 if (!s->dct)
469 return AVERROR(ENOMEM);
470
471 if (opts) {
472 AVDictionaryEntry *e = NULL;
473
474 while ((e = av_dict_get(*opts, "", e, AV_DICT_IGNORE_SUFFIX))) {
475 if ((ret = av_opt_set(s->dct, e->key, e->value, 0)) < 0)
476 return ret;
477 }
478 av_dict_free(opts);
479 }
480
481 s->store_slice = store_slice_c;
482 switch (s->mode) {
483 case MODE_HARD: s->requantize = hardthresh_c; break;
484 case MODE_SOFT: s->requantize = softthresh_c; break;
485 }
486 return 0;
487 }
488
uninit(AVFilterContext * ctx)489 static av_cold void uninit(AVFilterContext *ctx)
490 {
491 SPPContext *s = ctx->priv;
492
493 av_freep(&s->temp);
494 av_freep(&s->src);
495 av_freep(&s->dct);
496 av_freep(&s->non_b_qp_table);
497 }
498
499 static const AVFilterPad spp_inputs[] = {
500 {
501 .name = "default",
502 .type = AVMEDIA_TYPE_VIDEO,
503 .config_props = config_input,
504 .filter_frame = filter_frame,
505 },
506 { NULL }
507 };
508
509 static const AVFilterPad spp_outputs[] = {
510 {
511 .name = "default",
512 .type = AVMEDIA_TYPE_VIDEO,
513 },
514 { NULL }
515 };
516
517 AVFilter ff_vf_spp = {
518 .name = "spp",
519 .description = NULL_IF_CONFIG_SMALL("Apply a simple post processing filter."),
520 .priv_size = sizeof(SPPContext),
521 .init_dict = init_dict,
522 .uninit = uninit,
523 .query_formats = query_formats,
524 .inputs = spp_inputs,
525 .outputs = spp_outputs,
526 .process_command = process_command,
527 .priv_class = &spp_class,
528 .flags = AVFILTER_FLAG_SUPPORT_TIMELINE_INTERNAL,
529 };
530