vf_hqdn3d.c 12.5 KB
Newer Older
1 2 3
/*
 * Copyright (c) 2003 Daniel Moreno <comac AT comac DOT darktech DOT org>
 * Copyright (c) 2010 Baptiste Coudurier
4
 * Copyright (c) 2012 Loren Merritt
5
 *
6
 * This file is part of Libav, ported from MPlayer.
7
 *
8
 * Libav is free software; you can redistribute it and/or modify
9 10 11 12
 * it under the terms of the GNU General Public License as published by
 * the Free Software Foundation; either version 2 of the License, or
 * (at your option) any later version.
 *
13
 * Libav is distributed in the hope that it will be useful,
14 15 16 17 18
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
 * GNU General Public License for more details.
 *
 * You should have received a copy of the GNU General Public License along
19
 * with Libav; if not, write to the Free Software Foundation, Inc.,
20 21 22 23 24 25 26 27 28
 * 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
 */

/**
 * @file
 * high quality 3d video denoiser, ported from MPlayer
 * libmpcodecs/vf_hqdn3d.c.
 */

29
#include "libavutil/common.h"
30
#include "libavutil/pixdesc.h"
31
#include "libavutil/intreadwrite.h"
32
#include "avfilter.h"
33
#include "formats.h"
34
#include "internal.h"
35
#include "video.h"
36 37

typedef struct {
38
    int16_t *coefs[4];
39
    uint16_t *line;
Loren Merritt's avatar
Loren Merritt committed
40
    uint16_t *frame_prev[3];
41
    double strength[4];
42
    int hsub, vsub;
43
    int depth;
Loren Merritt's avatar
Loren Merritt committed
44
    void (*denoise_row[17])(uint8_t *src, uint8_t *dst, uint16_t *line_ant, uint16_t *frame_ant, ptrdiff_t w, int16_t *spatial, int16_t *temporal);
45 46
} HQDN3DContext;

Loren Merritt's avatar
Loren Merritt committed
47 48 49 50 51
void ff_hqdn3d_row_8_x86(uint8_t *src, uint8_t *dst, uint16_t *line_ant, uint16_t *frame_ant, ptrdiff_t w, int16_t *spatial, int16_t *temporal);
void ff_hqdn3d_row_9_x86(uint8_t *src, uint8_t *dst, uint16_t *line_ant, uint16_t *frame_ant, ptrdiff_t w, int16_t *spatial, int16_t *temporal);
void ff_hqdn3d_row_10_x86(uint8_t *src, uint8_t *dst, uint16_t *line_ant, uint16_t *frame_ant, ptrdiff_t w, int16_t *spatial, int16_t *temporal);
void ff_hqdn3d_row_16_x86(uint8_t *src, uint8_t *dst, uint16_t *line_ant, uint16_t *frame_ant, ptrdiff_t w, int16_t *spatial, int16_t *temporal);

52
#define LUT_BITS (depth==16 ? 8 : 4)
53 54 55 56 57
#define RIGHTSHIFT(a,b) (((a)+(((1<<(b))-1)>>1))>>(b))
#define LOAD(x) ((depth==8 ? src[x] : AV_RN16A(src+(x)*2)) << (16-depth))
#define STORE(x,val) (depth==8 ? dst[x] = RIGHTSHIFT(val, 16-depth)\
                    : AV_WN16A(dst+(x)*2, RIGHTSHIFT(val, 16-depth)))

58
av_always_inline
59
static uint32_t lowpass(int prev, int cur, int16_t *coef, int depth)
60
{
61
    int d = (prev - cur) >> (8 - LUT_BITS);
Loren Merritt's avatar
Loren Merritt committed
62
    return cur + coef[d];
63 64
}

65
av_always_inline
Loren Merritt's avatar
Loren Merritt committed
66 67 68
static void denoise_temporal(uint8_t *src, uint8_t *dst,
                             uint16_t *frame_ant,
                             int w, int h, int sstride, int dstride,
69
                             int16_t *temporal, int depth)
70
{
Loren Merritt's avatar
Loren Merritt committed
71
    long x, y;
72
    uint32_t tmp;
Loren Merritt's avatar
Loren Merritt committed
73

74
    temporal += 256 << LUT_BITS;
75

Loren Merritt's avatar
Loren Merritt committed
76 77
    for (y = 0; y < h; y++) {
        for (x = 0; x < w; x++) {
78
            frame_ant[x] = tmp = lowpass(frame_ant[x], LOAD(x), temporal, depth);
79
            STORE(x, tmp);
80
        }
Loren Merritt's avatar
Loren Merritt committed
81 82 83
        src += sstride;
        dst += dstride;
        frame_ant += w;
84 85 86
    }
}

87
av_always_inline
Loren Merritt's avatar
Loren Merritt committed
88 89
static void denoise_spatial(HQDN3DContext *hqdn3d,
                            uint8_t *src, uint8_t *dst,
90
                            uint16_t *line_ant, uint16_t *frame_ant,
Loren Merritt's avatar
Loren Merritt committed
91
                            int w, int h, int sstride, int dstride,
92
                            int16_t *spatial, int16_t *temporal, int depth)
93
{
Loren Merritt's avatar
Loren Merritt committed
94 95
    long x, y;
    uint32_t pixel_ant;
96
    uint32_t tmp;
97

98 99
    spatial  += 256 << LUT_BITS;
    temporal += 256 << LUT_BITS;
100

101 102
    /* First line has no top neighbor. Only left one for each tmp and
     * last frame */
103
    pixel_ant = LOAD(0);
104
    for (x = 0; x < w; x++) {
105 106
        line_ant[x] = tmp = pixel_ant = lowpass(pixel_ant, LOAD(x), spatial, depth);
        frame_ant[x] = tmp = lowpass(frame_ant[x], tmp, temporal, depth);
107
        STORE(x, tmp);
108 109
    }

Loren Merritt's avatar
Loren Merritt committed
110
    for (y = 1; y < h; y++) {
111 112 113
        src += sstride;
        dst += dstride;
        frame_ant += w;
Loren Merritt's avatar
Loren Merritt committed
114 115 116 117
        if (hqdn3d->denoise_row[depth]) {
            hqdn3d->denoise_row[depth](src, dst, line_ant, frame_ant, w, spatial, temporal);
            continue;
        }
118
        pixel_ant = LOAD(0);
119
        for (x = 0; x < w-1; x++) {
120 121 122
            line_ant[x] = tmp = lowpass(line_ant[x], pixel_ant, spatial, depth);
            pixel_ant = lowpass(pixel_ant, LOAD(x+1), spatial, depth);
            frame_ant[x] = tmp = lowpass(frame_ant[x], tmp, temporal, depth);
123
            STORE(x, tmp);
124
        }
125 126
        line_ant[x] = tmp = lowpass(line_ant[x], pixel_ant, spatial, depth);
        frame_ant[x] = tmp = lowpass(frame_ant[x], tmp, temporal, depth);
127
        STORE(x, tmp);
128 129 130
    }
}

131
av_always_inline
Loren Merritt's avatar
Loren Merritt committed
132 133
static void denoise_depth(HQDN3DContext *hqdn3d,
                          uint8_t *src, uint8_t *dst,
134 135 136
                          uint16_t *line_ant, uint16_t **frame_ant_ptr,
                          int w, int h, int sstride, int dstride,
                          int16_t *spatial, int16_t *temporal, int depth)
137
{
138 139
    // FIXME: For 16bit depth, frame_ant could be a pointer to the previous
    // filtered frame rather than a separate buffer.
Loren Merritt's avatar
Loren Merritt committed
140 141 142
    long x, y;
    uint16_t *frame_ant = *frame_ant_ptr;
    if (!frame_ant) {
143
        uint8_t *frame_src = src;
Loren Merritt's avatar
Loren Merritt committed
144
        *frame_ant_ptr = frame_ant = av_malloc(w*h*sizeof(uint16_t));
145
        for (y = 0; y < h; y++, src += sstride, frame_ant += w)
Loren Merritt's avatar
Loren Merritt committed
146
            for (x = 0; x < w; x++)
147
                frame_ant[x] = LOAD(x);
148 149
        src = frame_src;
        frame_ant = *frame_ant_ptr;
150 151
    }

152
    if (spatial[0])
Loren Merritt's avatar
Loren Merritt committed
153
        denoise_spatial(hqdn3d, src, dst, line_ant, frame_ant,
154
                        w, h, sstride, dstride, spatial, temporal, depth);
155
    else
Loren Merritt's avatar
Loren Merritt committed
156
        denoise_temporal(src, dst, frame_ant,
157
                         w, h, sstride, dstride, temporal, depth);
158 159
}

160 161 162 163 164
#define denoise(...) \
    switch (hqdn3d->depth) {\
        case  8: denoise_depth(__VA_ARGS__,  8); break;\
        case  9: denoise_depth(__VA_ARGS__,  9); break;\
        case 10: denoise_depth(__VA_ARGS__, 10); break;\
165
        case 16: denoise_depth(__VA_ARGS__, 16); break;\
166 167
    }

168
static int16_t *precalc_coefs(double dist25, int depth)
169 170
{
    int i;
Loren Merritt's avatar
Loren Merritt committed
171
    double gamma, simil, C;
172 173 174
    int16_t *ct = av_malloc((512<<LUT_BITS)*sizeof(int16_t));
    if (!ct)
        return NULL;
175

176
    gamma = log(0.25) / log(1.0 - FFMIN(dist25,252.0)/255.0 - 0.00001);
177

178 179
    for (i = -255<<LUT_BITS; i <= 255<<LUT_BITS; i++) {
        double f = ((i<<(9-LUT_BITS)) + (1<<(8-LUT_BITS)) - 1) / 512.0; // midpoint of the bin
180 181
        simil = 1.0 - FFABS(f) / 255.0;
        C = pow(simil, gamma) * 256.0 * f;
182
        ct[(256<<LUT_BITS)+i] = lrint(C);
183 184
    }

Loren Merritt's avatar
Loren Merritt committed
185
    ct[0] = !!dist25;
186
    return ct;
187 188 189 190 191 192
}

#define PARAM1_DEFAULT 4.0
#define PARAM2_DEFAULT 3.0
#define PARAM3_DEFAULT 6.0

193
static int init(AVFilterContext *ctx, const char *args)
194 195
{
    HQDN3DContext *hqdn3d = ctx->priv;
Loren Merritt's avatar
Loren Merritt committed
196 197
    double lum_spac, lum_tmp, chrom_spac, chrom_tmp;
    double param1, param2, param3, param4;
198

Loren Merritt's avatar
Loren Merritt committed
199 200 201 202
    lum_spac   = PARAM1_DEFAULT;
    chrom_spac = PARAM2_DEFAULT;
    lum_tmp    = PARAM3_DEFAULT;
    chrom_tmp  = lum_tmp * chrom_spac / lum_spac;
203 204 205

    if (args) {
        switch (sscanf(args, "%lf:%lf:%lf:%lf",
Loren Merritt's avatar
Loren Merritt committed
206
                       &param1, &param2, &param3, &param4)) {
207
        case 1:
Loren Merritt's avatar
Loren Merritt committed
208 209 210 211
            lum_spac   = param1;
            chrom_spac = PARAM2_DEFAULT * param1 / PARAM1_DEFAULT;
            lum_tmp    = PARAM3_DEFAULT * param1 / PARAM1_DEFAULT;
            chrom_tmp  = lum_tmp * chrom_spac / lum_spac;
212 213
            break;
        case 2:
Loren Merritt's avatar
Loren Merritt committed
214 215 216 217
            lum_spac   = param1;
            chrom_spac = param2;
            lum_tmp    = PARAM3_DEFAULT * param1 / PARAM1_DEFAULT;
            chrom_tmp  = lum_tmp * chrom_spac / lum_spac;
218 219
            break;
        case 3:
Loren Merritt's avatar
Loren Merritt committed
220 221 222 223
            lum_spac   = param1;
            chrom_spac = param2;
            lum_tmp    = param3;
            chrom_tmp  = lum_tmp * chrom_spac / lum_spac;
224 225
            break;
        case 4:
Loren Merritt's avatar
Loren Merritt committed
226 227 228 229
            lum_spac   = param1;
            chrom_spac = param2;
            lum_tmp    = param3;
            chrom_tmp  = param4;
230 231 232 233
            break;
        }
    }

234 235 236 237 238
    hqdn3d->strength[0] = lum_spac;
    hqdn3d->strength[1] = lum_tmp;
    hqdn3d->strength[2] = chrom_spac;
    hqdn3d->strength[3] = chrom_tmp;

239
    av_log(ctx, AV_LOG_VERBOSE, "ls:%lf cs:%lf lt:%lf ct:%lf\n",
Loren Merritt's avatar
Loren Merritt committed
240 241
           lum_spac, chrom_spac, lum_tmp, chrom_tmp);
    if (lum_spac < 0 || chrom_spac < 0 || isnan(chrom_tmp)) {
242 243 244 245 246 247 248 249 250 251 252 253 254
        av_log(ctx, AV_LOG_ERROR,
               "Invalid negative value for luma or chroma spatial strength, "
               "or resulting value for chroma temporal strength is nan.\n");
        return AVERROR(EINVAL);
    }

    return 0;
}

static void uninit(AVFilterContext *ctx)
{
    HQDN3DContext *hqdn3d = ctx->priv;

255 256 257 258
    av_freep(&hqdn3d->coefs[0]);
    av_freep(&hqdn3d->coefs[1]);
    av_freep(&hqdn3d->coefs[2]);
    av_freep(&hqdn3d->coefs[3]);
Loren Merritt's avatar
Loren Merritt committed
259 260 261 262
    av_freep(&hqdn3d->line);
    av_freep(&hqdn3d->frame_prev[0]);
    av_freep(&hqdn3d->frame_prev[1]);
    av_freep(&hqdn3d->frame_prev[2]);
263 264 265 266
}

static int query_formats(AVFilterContext *ctx)
{
267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287
    static const enum AVPixelFormat pix_fmts[] = {
        AV_PIX_FMT_YUV420P,
        AV_PIX_FMT_YUV422P,
        AV_PIX_FMT_YUV444P,
        AV_PIX_FMT_YUV410P,
        AV_PIX_FMT_YUV411P,
        AV_PIX_FMT_YUV440P,
        AV_PIX_FMT_YUVJ420P,
        AV_PIX_FMT_YUVJ422P,
        AV_PIX_FMT_YUVJ444P,
        AV_PIX_FMT_YUVJ440P,
        AV_NE( AV_PIX_FMT_YUV420P9BE, AV_PIX_FMT_YUV420P9LE ),
        AV_NE( AV_PIX_FMT_YUV422P9BE, AV_PIX_FMT_YUV422P9LE ),
        AV_NE( AV_PIX_FMT_YUV444P9BE, AV_PIX_FMT_YUV444P9LE ),
        AV_NE( AV_PIX_FMT_YUV420P10BE, AV_PIX_FMT_YUV420P10LE ),
        AV_NE( AV_PIX_FMT_YUV422P10BE, AV_PIX_FMT_YUV422P10LE ),
        AV_NE( AV_PIX_FMT_YUV444P10BE, AV_PIX_FMT_YUV444P10LE ),
        AV_NE( AV_PIX_FMT_YUV420P16BE, AV_PIX_FMT_YUV420P16LE ),
        AV_NE( AV_PIX_FMT_YUV422P16BE, AV_PIX_FMT_YUV422P16LE ),
        AV_NE( AV_PIX_FMT_YUV444P16BE, AV_PIX_FMT_YUV444P16LE ),
        AV_PIX_FMT_NONE
288 289
    };

290
    ff_set_common_formats(ctx, ff_make_format_list(pix_fmts));
291 292 293 294 295 296 297

    return 0;
}

static int config_input(AVFilterLink *inlink)
{
    HQDN3DContext *hqdn3d = inlink->dst->priv;
298
    const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(inlink->format);
299
    int i;
300

301 302 303
    hqdn3d->hsub  = desc->log2_chroma_w;
    hqdn3d->vsub  = desc->log2_chroma_h;
    hqdn3d->depth = desc->comp[0].depth_minus1+1;
304

Loren Merritt's avatar
Loren Merritt committed
305 306
    hqdn3d->line = av_malloc(inlink->w * sizeof(*hqdn3d->line));
    if (!hqdn3d->line)
307 308
        return AVERROR(ENOMEM);

309
    for (i = 0; i < 4; i++) {
310 311 312 313 314
        hqdn3d->coefs[i] = precalc_coefs(hqdn3d->strength[i], hqdn3d->depth);
        if (!hqdn3d->coefs[i])
            return AVERROR(ENOMEM);
    }

Loren Merritt's avatar
Loren Merritt committed
315 316 317 318 319 320 321
#if HAVE_YASM
    hqdn3d->denoise_row[ 8] = ff_hqdn3d_row_8_x86;
    hqdn3d->denoise_row[ 9] = ff_hqdn3d_row_9_x86;
    hqdn3d->denoise_row[10] = ff_hqdn3d_row_10_x86;
    hqdn3d->denoise_row[16] = ff_hqdn3d_row_16_x86;
#endif

322 323 324
    return 0;
}

325 326 327 328
static int null_draw_slice(AVFilterLink *link, int y, int h, int slice_dir)
{
    return 0;
}
329

330
static int end_frame(AVFilterLink *inlink)
331 332 333 334 335
{
    HQDN3DContext *hqdn3d = inlink->dst->priv;
    AVFilterLink *outlink = inlink->dst->outputs[0];
    AVFilterBufferRef *inpic  = inlink ->cur_buf;
    AVFilterBufferRef *outpic = outlink->out_buf;
336 337 338
    int ret, c;

    for (c = 0; c < 3; c++) {
Loren Merritt's avatar
Loren Merritt committed
339
        denoise(hqdn3d, inpic->data[c], outpic->data[c],
340 341 342 343 344 345
                hqdn3d->line, &hqdn3d->frame_prev[c],
                inpic->video->w >> (!!c * hqdn3d->hsub),
                inpic->video->h >> (!!c * hqdn3d->vsub),
                inpic->linesize[c], outpic->linesize[c],
                hqdn3d->coefs[c?2:0], hqdn3d->coefs[c?3:1]);
    }
346

347 348 349 350
    if ((ret = ff_draw_slice(outlink, 0, inpic->video->h, 1)) < 0 ||
        (ret = ff_end_frame(outlink)) < 0)
        return ret;
    return 0;
351 352
}

353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372
static const AVFilterPad avfilter_vf_hqdn3d_inputs[] = {
    {
        .name         = "default",
        .type         = AVMEDIA_TYPE_VIDEO,
        .start_frame  = ff_inplace_start_frame,
        .draw_slice   = null_draw_slice,
        .config_props = config_input,
        .end_frame    = end_frame
    },
    { NULL }
};

static const AVFilterPad avfilter_vf_hqdn3d_outputs[] = {
    {
        .name = "default",
        .type = AVMEDIA_TYPE_VIDEO
    },
    { NULL }
};

373 374 375 376 377 378 379 380 381
AVFilter avfilter_vf_hqdn3d = {
    .name          = "hqdn3d",
    .description   = NULL_IF_CONFIG_SMALL("Apply a High Quality 3D Denoiser."),

    .priv_size     = sizeof(HQDN3DContext),
    .init          = init,
    .uninit        = uninit,
    .query_formats = query_formats,

382 383 384
    .inputs    = avfilter_vf_hqdn3d_inputs,

    .outputs   = avfilter_vf_hqdn3d_outputs,
385
};