libavfilter/vf_hqdn3d.c

   1 /*
   2  * Copyright (c) 2003 Daniel Moreno <comac AT comac DOT darktech DOT org>
   3  * Copyright (c) 2010 Baptiste Coudurier
   4  * Copyright (c) 2012 Loren Merritt
   5  *
   6  * This file is part of FFmpeg, ported from MPlayer.
   7  *
   8  * FFmpeg is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License as published by
  10  * the Free Software Foundation; either version 2 of the License, or
  11  * (at your option) any later version.
  12  *
  13  * FFmpeg is distributed in the hope that it will be useful,
  14  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  15  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  16  * GNU General Public License for more details.
  17  *
  18  * You should have received a copy of the GNU General Public License along
  19  * with FFmpeg; if not, write to the Free Software Foundation, Inc.,
  20  * 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
  21  */
  22
  23 /**
  24  * @file
  25  * high quality 3d video denoiser, ported from MPlayer
  26  * libmpcodecs/vf_hqdn3d.c.
  27  */
  28
  29 #include <float.h>
  30
  31 #include "config.h"
  32 #include "libavutil/attributes.h"
  33 #include "libavutil/common.h"
  34 #include "libavutil/emms.h"
  35 #include "libavutil/mem.h"
  36 #include "libavutil/pixdesc.h"
  37 #include "libavutil/intreadwrite.h"
  38 #include "libavutil/opt.h"
  39
  40 #include "avfilter.h"
  41 #include "filters.h"
  42 #include "video.h"
  43 #include "vf_hqdn3d.h"
  44
  45 #define LUT_BITS (depth==16 ? 8 : 4)
  46 #define LOAD(x) (((depth == 8 ? src[x] : AV_RN16A(src + (x) * 2)) << (16 - depth))\
  47                  + (((1 << (16 - depth)) - 1) >> 1))
  48 #define STORE(x,val) (depth == 8 ? dst[x] = (val) >> (16 - depth) : \
  49                                    AV_WN16A(dst + (x) * 2, (val) >> (16 - depth)))
  50
  51 av_always_inline
  52 static uint32_t lowpass(int prev, int cur, int16_t *coef, int depth)
  53 {
  54     int d = (prev - cur) >> (8 - LUT_BITS);
  55     return cur + coef[d];
  56 }
  57
  58 av_always_inline
  59 static void denoise_temporal(uint8_t *src, uint8_t *dst,
  60                              uint16_t *frame_ant,
  61                              int w, int h, int sstride, int dstride,
  62                              int16_t *temporal, int depth)
  63 {
  64     long x, y;
  65     uint32_t tmp;
  66
  67     temporal += 256 << LUT_BITS;
  68
  69     for (y = 0; y < h; y++) {
  70         for (x = 0; x < w; x++) {
  71             frame_ant[x] = tmp = lowpass(frame_ant[x], LOAD(x), temporal, depth);
  72             STORE(x, tmp);
  73         }
  74         src += sstride;
  75         dst += dstride;
  76         frame_ant += w;
  77     }
  78 }
  79
  80 av_always_inline
  81 static void denoise_spatial(HQDN3DContext *s,
  82                             uint8_t *src, uint8_t *dst,
  83                             uint16_t *line_ant, uint16_t *frame_ant,
  84                             int w, int h, int sstride, int dstride,
  85                             int16_t *spatial, int16_t *temporal, int depth)
  86 {
  87     long x, y;
  88     uint32_t pixel_ant;
  89     uint32_t tmp;
  90
  91     spatial  += 256 << LUT_BITS;
  92     temporal += 256 << LUT_BITS;
  93
  94     /* First line has no top neighbor. Only left one for each tmp and
  95      * last frame */
  96     pixel_ant = LOAD(0);
  97     for (x = 0; x < w; x++) {
  98         line_ant[x] = tmp = pixel_ant = lowpass(pixel_ant, LOAD(x), spatial, depth);
  99         frame_ant[x] = tmp = lowpass(frame_ant[x], tmp, temporal, depth);
 100         STORE(x, tmp);
 101     }
 102
 103     for (y = 1; y < h; y++) {
 104         src += sstride;
 105         dst += dstride;
 106         frame_ant += w;
 107         if (s->denoise_row[depth]) {
 108             s->denoise_row[depth](src, dst, line_ant, frame_ant, w, spatial, temporal);
 109             continue;
 110         }
 111         pixel_ant = LOAD(0);
 112         for (x = 0; x < w-1; x++) {
 113             line_ant[x] = tmp = lowpass(line_ant[x], pixel_ant, spatial, depth);
 114             pixel_ant = lowpass(pixel_ant, LOAD(x+1), spatial, depth);
 115             frame_ant[x] = tmp = lowpass(frame_ant[x], tmp, temporal, depth);
 116             STORE(x, tmp);
 117         }
 118         line_ant[x] = tmp = lowpass(line_ant[x], pixel_ant, spatial, depth);
 119         frame_ant[x] = tmp = lowpass(frame_ant[x], tmp, temporal, depth);
 120         STORE(x, tmp);
 121     }
 122 }
 123
 124 av_always_inline
 125 static int denoise_depth(HQDN3DContext *s,
 126                          uint8_t *src, uint8_t *dst,
 127                          uint16_t *line_ant, uint16_t **frame_ant_ptr,
 128                          int w, int h, int sstride, int dstride,
 129                          int16_t *spatial, int16_t *temporal, int depth)
 130 {
 131     // FIXME: For 16-bit depth, frame_ant could be a pointer to the previous
 132     // filtered frame rather than a separate buffer.
 133     long x, y;
 134     uint16_t *frame_ant = *frame_ant_ptr;
 135     if (!frame_ant) {
 136         uint8_t *frame_src = src;
 137         *frame_ant_ptr = frame_ant = av_malloc_array(w, h*sizeof(uint16_t));
 138         if (!frame_ant)
 139             return AVERROR(ENOMEM);
 140         for (y = 0; y < h; y++, src += sstride, frame_ant += w)
 141             for (x = 0; x < w; x++)
 142                 frame_ant[x] = LOAD(x);
 143         src = frame_src;
 144         frame_ant = *frame_ant_ptr;
 145     }
 146
 147     if (spatial[0])
 148         denoise_spatial(s, src, dst, line_ant, frame_ant,
 149                         w, h, sstride, dstride, spatial, temporal, depth);
 150     else
 151         denoise_temporal(src, dst, frame_ant,
 152                          w, h, sstride, dstride, temporal, depth);
 153     emms_c();
 154     return 0;
 155 }
 156
 157 #define denoise(...)                                                          \
 158     do {                                                                      \
 159         int ret = AVERROR_BUG;                                                \
 160         switch (s->depth) {                                                   \
 161             case  8: ret = denoise_depth(__VA_ARGS__,  8); break;             \
 162             case  9: ret = denoise_depth(__VA_ARGS__,  9); break;             \
 163             case 10: ret = denoise_depth(__VA_ARGS__, 10); break;             \
 164             case 12: ret = denoise_depth(__VA_ARGS__, 12); break;             \
 165             case 14: ret = denoise_depth(__VA_ARGS__, 14); break;             \
 166             case 16: ret = denoise_depth(__VA_ARGS__, 16); break;             \
 167         }                                                                     \
 168         if (ret < 0) {                                                        \
 169             av_frame_free(&out);                                              \
 170             if (!direct)                                                      \
 171                 av_frame_free(&in);                                           \
 172             return ret;                                                       \
 173         }                                                                     \
 174     } while (0)
 175
 176 static void precalc_coefs(double dist25, int depth, int16_t *ct)
 177 {
 178     int i;
 179     double gamma, simil, C;
 180
 181     gamma = log(0.25) / log(1.0 - FFMIN(dist25,252.0)/255.0 - 0.00001);
 182
 183     for (i = -(256<<LUT_BITS); i < 256<<LUT_BITS; i++) {
 184         double f = (i * (1 << (9-LUT_BITS)) + (1<<(8-LUT_BITS)) - 1) / 512.0; // midpoint of the bin
 185         simil = FFMAX(0, 1.0 - fabs(f) / 255.0);
 186         C = pow(simil, gamma) * 256.0 * f;
 187         ct[(256<<LUT_BITS)+i] = lrint(C);
 188     }
 189
 190     ct[0] = !!dist25;
 191 }
 192
 193 #define PARAM1_DEFAULT 4.0
 194 #define PARAM2_DEFAULT 3.0
 195 #define PARAM3_DEFAULT 6.0
 196
 197 static av_cold int init(AVFilterContext *ctx)
 198 {
 199     HQDN3DContext *s = ctx->priv;
 200
 201     if (!s->strength[LUMA_SPATIAL])
 202         s->strength[LUMA_SPATIAL] = PARAM1_DEFAULT;
 203     if (!s->strength[CHROMA_SPATIAL])
 204         s->strength[CHROMA_SPATIAL] = PARAM2_DEFAULT * s->strength[LUMA_SPATIAL] / PARAM1_DEFAULT;
 205     if (!s->strength[LUMA_TMP])
 206         s->strength[LUMA_TMP]   = PARAM3_DEFAULT * s->strength[LUMA_SPATIAL] / PARAM1_DEFAULT;
 207     if (!s->strength[CHROMA_TMP])
 208         s->strength[CHROMA_TMP] = s->strength[LUMA_TMP] * s->strength[CHROMA_SPATIAL] / s->strength[LUMA_SPATIAL];
 209
 210     av_log(ctx, AV_LOG_VERBOSE, "ls:%f cs:%f lt:%f ct:%f\n",
 211            s->strength[LUMA_SPATIAL], s->strength[CHROMA_SPATIAL],
 212            s->strength[LUMA_TMP], s->strength[CHROMA_TMP]);
 213
 214     return 0;
 215 }
 216
 217 static av_cold void uninit(AVFilterContext *ctx)
 218 {
 219     HQDN3DContext *s = ctx->priv;
 220
 221     av_freep(&s->coefs[0]);
 222     av_freep(&s->coefs[1]);
 223     av_freep(&s->coefs[2]);
 224     av_freep(&s->coefs[3]);
 225     av_freep(&s->line[0]);
 226     av_freep(&s->line[1]);
 227     av_freep(&s->line[2]);
 228     av_freep(&s->frame_prev[0]);
 229     av_freep(&s->frame_prev[1]);
 230     av_freep(&s->frame_prev[2]);
 231 }
 232
 233 static const enum AVPixelFormat pix_fmts[] = {
 234     AV_PIX_FMT_YUV420P, AV_PIX_FMT_YUV422P, AV_PIX_FMT_YUV444P,
 235     AV_PIX_FMT_YUV410P, AV_PIX_FMT_YUV411P, AV_PIX_FMT_YUV440P,
 236     AV_PIX_FMT_YUVJ420P, AV_PIX_FMT_YUVJ422P, AV_PIX_FMT_YUVJ444P, AV_PIX_FMT_YUVJ440P,
 237     AV_PIX_FMT_YUV420P9, AV_PIX_FMT_YUV422P9, AV_PIX_FMT_YUV444P9,
 238     AV_PIX_FMT_YUV420P10, AV_PIX_FMT_YUV422P10, AV_PIX_FMT_YUV444P10,
 239     AV_PIX_FMT_YUV440P10,
 240     AV_PIX_FMT_YUV444P12, AV_PIX_FMT_YUV422P12, AV_PIX_FMT_YUV420P12,
 241     AV_PIX_FMT_YUV440P12,
 242     AV_PIX_FMT_YUV444P14, AV_PIX_FMT_YUV422P14, AV_PIX_FMT_YUV420P14,
 243     AV_PIX_FMT_YUV420P16, AV_PIX_FMT_YUV422P16, AV_PIX_FMT_YUV444P16,
 244     AV_PIX_FMT_NONE
 245 };
 246
 247 static void calc_coefs(AVFilterContext *ctx)
 248 {
 249     HQDN3DContext *s = ctx->priv;
 250
 251     for (int i = 0; i < 4; i++)
 252         precalc_coefs(s->strength[i], s->depth, s->coefs[i]);
 253 }
 254
 255 static int config_input(AVFilterLink *inlink)
 256 {
 257     AVFilterContext *ctx = inlink->dst;
 258     HQDN3DContext *s = inlink->dst->priv;
 259     const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(inlink->format);
 260     int i, depth;
 261
 262     uninit(inlink->dst);
 263
 264     s->hsub  = desc->log2_chroma_w;
 265     s->vsub  = desc->log2_chroma_h;
 266     s->depth = depth = desc->comp[0].depth;
 267
 268     for (i = 0; i < 3; i++) {
 269         s->line[i] = av_malloc_array(inlink->w, sizeof(*s->line[i]));
 270         if (!s->line[i])
 271             return AVERROR(ENOMEM);
 272     }
 273
 274     for (i = 0; i < 4; i++) {
 275         s->coefs[i] = av_malloc((512<<LUT_BITS) * sizeof(int16_t));
 276         if (!s->coefs[i])
 277             return AVERROR(ENOMEM);
 278     }
 279
 280     calc_coefs(ctx);
 281
 282 #if ARCH_X86
 283     ff_hqdn3d_init_x86(s);
 284 #endif
 285
 286     return 0;
 287 }
 288
 289 typedef struct ThreadData {
 290     AVFrame *in, *out;
 291     int direct;
 292 } ThreadData;
 293
 294 static int do_denoise(AVFilterContext *ctx, void *data, int job_nr, int n_jobs)
 295 {
 296     HQDN3DContext *s = ctx->priv;
 297     const ThreadData *td = data;
 298     AVFrame *out = td->out;
 299     AVFrame *in = td->in;
 300     int direct = td->direct;
 301
 302     denoise(s, in->data[job_nr], out->data[job_nr],
 303                 s->line[job_nr], &s->frame_prev[job_nr],
 304                 AV_CEIL_RSHIFT(in->width,  (!!job_nr * s->hsub)),
 305                 AV_CEIL_RSHIFT(in->height, (!!job_nr * s->vsub)),
 306                 in->linesize[job_nr], out->linesize[job_nr],
 307                 s->coefs[job_nr ? CHROMA_SPATIAL : LUMA_SPATIAL],
 308                 s->coefs[job_nr ? CHROMA_TMP     : LUMA_TMP]);
 309
 310     return 0;
 311 }
 312
 313 static int filter_frame(AVFilterLink *inlink, AVFrame *in)
 314 {
 315     AVFilterContext *ctx  = inlink->dst;
 316     AVFilterLink *outlink = ctx->outputs[0];
 317
 318     AVFrame *out;
 319     int direct = av_frame_is_writable(in) && !ctx->is_disabled;
 320     ThreadData td;
 321
 322     if (direct) {
 323         out = in;
 324     } else {
 325         out = ff_get_video_buffer(outlink, outlink->w, outlink->h);
 326         if (!out) {
 327             av_frame_free(&in);
 328             return AVERROR(ENOMEM);
 329         }
 330
 331         av_frame_copy_props(out, in);
 332     }
 333
 334     td.in = in;
 335     td.out = out;
 336     td.direct = direct;
 337     /* one thread per plane */
 338     ff_filter_execute(ctx, do_denoise, &td, NULL, 3);
 339
 340     if (ctx->is_disabled) {
 341         av_frame_free(&out);
 342         return ff_filter_frame(outlink, in);
 343     }
 344
 345     if (!direct)
 346         av_frame_free(&in);
 347
 348     return ff_filter_frame(outlink, out);
 349 }
 350
 351 static int process_command(AVFilterContext *ctx, const char *cmd, const char *args,
 352                            char *res, int res_len, int flags)
 353 {
 354     int ret;
 355
 356     ret = ff_filter_process_command(ctx, cmd, args, res, res_len, flags);
 357     if (ret < 0)
 358         return ret;
 359
 360     calc_coefs(ctx);
 361
 362     return 0;
 363 }
 364
 365 #define OFFSET(x) offsetof(HQDN3DContext, x)
 366 #define FLAGS AV_OPT_FLAG_VIDEO_PARAM | AV_OPT_FLAG_FILTERING_PARAM | AV_OPT_FLAG_RUNTIME_PARAM
 367 static const AVOption hqdn3d_options[] = {
 368     { "luma_spatial",   "spatial luma strength",    OFFSET(strength[LUMA_SPATIAL]),   AV_OPT_TYPE_DOUBLE, { .dbl = 0.0 }, 0, DBL_MAX, FLAGS },
 369     { "chroma_spatial", "spatial chroma strength",  OFFSET(strength[CHROMA_SPATIAL]), AV_OPT_TYPE_DOUBLE, { .dbl = 0.0 }, 0, DBL_MAX, FLAGS },
 370     { "luma_tmp",       "temporal luma strength",   OFFSET(strength[LUMA_TMP]),       AV_OPT_TYPE_DOUBLE, { .dbl = 0.0 }, 0, DBL_MAX, FLAGS },
 371     { "chroma_tmp",     "temporal chroma strength", OFFSET(strength[CHROMA_TMP]),     AV_OPT_TYPE_DOUBLE, { .dbl = 0.0 }, 0, DBL_MAX, FLAGS },
 372     { NULL }
 373 };
 374
 375 AVFILTER_DEFINE_CLASS(hqdn3d);
 376
 377 static const AVFilterPad avfilter_vf_hqdn3d_inputs[] = {
 378     {
 379         .name         = "default",
 380         .type         = AVMEDIA_TYPE_VIDEO,
 381         .config_props = config_input,
 382         .filter_frame = filter_frame,
 383     },
 384 };
 385
 386
 387 const FFFilter ff_vf_hqdn3d = {
 388     .p.name        = "hqdn3d",
 389     .p.description = NULL_IF_CONFIG_SMALL("Apply a High Quality 3D Denoiser."),
 390     .p.priv_class  = &hqdn3d_class,
 391     .p.flags       = AVFILTER_FLAG_SUPPORT_TIMELINE_INTERNAL | AVFILTER_FLAG_SLICE_THREADS,
 392     .priv_size     = sizeof(HQDN3DContext),
 393     .init          = init,
 394     .uninit        = uninit,
 395     FILTER_INPUTS(avfilter_vf_hqdn3d_inputs),
 396     FILTER_OUTPUTS(ff_video_default_filterpad),
 397     FILTER_PIXFMTS_ARRAY(pix_fmts),
 398     .process_command = process_command,
 399 };