/*
* Copyright (c) 2003 Daniel Moreno <comac AT comac DOT darktech DOT org>
* Copyright (c) 2010 Baptiste Coudurier
+ * Copyright (c) 2012 Loren Merritt
*
* This file is part of Libav, ported from MPlayer.
*
* libmpcodecs/vf_hqdn3d.c.
*/
+#include <float.h>
+
+#include "config.h"
+#include "libavutil/attributes.h"
+#include "libavutil/common.h"
#include "libavutil/pixdesc.h"
+#include "libavutil/intreadwrite.h"
+#include "libavutil/opt.h"
+
#include "avfilter.h"
#include "formats.h"
#include "internal.h"
#include "video.h"
+#include "vf_hqdn3d.h"
-typedef struct {
- int coefs[4][512*16];
- uint32_t *line;
- uint16_t *frame_prev[3];
- int hsub, vsub;
-} HQDN3DContext;
+#define LUT_BITS (depth==16 ? 8 : 4)
+#define LOAD(x) (((depth == 8 ? src[x] : AV_RN16A(src + (x) * 2)) << (16 - depth))\
+ + (((1 << (16 - depth)) - 1) >> 1))
+#define STORE(x,val) (depth == 8 ? dst[x] = (val) >> (16 - depth) : \
+ AV_WN16A(dst + (x) * 2, (val) >> (16 - depth)))
-static inline uint32_t lowpass(unsigned int prev, unsigned int cur, int *coef)
+av_always_inline
+static uint32_t lowpass(int prev, int cur, int16_t *coef, int depth)
{
- int dmul = prev-cur;
- unsigned int d = (dmul+0x10007FF)>>12; // 0x1000 to convert to unsigned, 7FF for rounding
+ int d = (prev - cur) >> (8 - LUT_BITS);
return cur + coef[d];
}
+av_always_inline
static void denoise_temporal(uint8_t *src, uint8_t *dst,
uint16_t *frame_ant,
int w, int h, int sstride, int dstride,
- int *temporal)
+ int16_t *temporal, int depth)
{
long x, y;
- uint32_t pixel;
+ uint32_t tmp;
+
+ temporal += 256 << LUT_BITS;
for (y = 0; y < h; y++) {
for (x = 0; x < w; x++) {
- pixel = lowpass(frame_ant[x]<<8, src[x]<<16, temporal);
- frame_ant[x] = ((pixel+0x1000007F)>>8);
- dst[x]= ((pixel+0x10007FFF)>>16);
+ frame_ant[x] = tmp = lowpass(frame_ant[x], LOAD(x), temporal, depth);
+ STORE(x, tmp);
}
src += sstride;
dst += dstride;
}
}
-static void denoise_spatial(uint8_t *src, uint8_t *dst,
- uint32_t *line_ant,
+av_always_inline
+static void denoise_spatial(HQDN3DContext *s,
+ uint8_t *src, uint8_t *dst,
+ uint16_t *line_ant, uint16_t *frame_ant,
int w, int h, int sstride, int dstride,
- int *horizontal, int *vertical)
+ int16_t *spatial, int16_t *temporal, int depth)
{
long x, y;
- long sline_offs = 0, dline_offs = 0;
uint32_t pixel_ant;
- uint32_t pixel;
+ uint32_t tmp;
- /* First pixel has no left nor top neighbor. */
- pixel = line_ant[0] = pixel_ant = src[0]<<16;
- dst[0]= ((pixel+0x10007FFF)>>16);
+ spatial += 256 << LUT_BITS;
+ temporal += 256 << LUT_BITS;
- /* First line has no top neighbor, only left. */
- for (x = 1; x < w; x++) {
- pixel = line_ant[x] = lowpass(pixel_ant, src[x]<<16, horizontal);
- dst[x]= ((pixel+0x10007FFF)>>16);
+ /* First line has no top neighbor. Only left one for each tmp and
+ * last frame */
+ pixel_ant = LOAD(0);
+ for (x = 0; x < w; x++) {
+ line_ant[x] = tmp = pixel_ant = lowpass(pixel_ant, LOAD(x), spatial, depth);
+ frame_ant[x] = tmp = lowpass(frame_ant[x], tmp, temporal, depth);
+ STORE(x, tmp);
}
for (y = 1; y < h; y++) {
- uint32_t pixel_ant;
- sline_offs += sstride, dline_offs += dstride;
- /* First pixel on each line doesn't have previous pixel */
- pixel_ant = src[sline_offs]<<16;
- pixel = line_ant[0] = lowpass(line_ant[0], pixel_ant, vertical);
- dst[dline_offs]= ((pixel+0x10007FFF)>>16);
-
- for (x = 1; x < w; x++) {
- uint32_t pixel;
- /* The rest are normal */
- pixel_ant = lowpass(pixel_ant, src[sline_offs+x]<<16, horizontal);
- pixel = line_ant[x] = lowpass(line_ant[x], pixel_ant, vertical);
- dst[dline_offs+x]= ((pixel+0x10007FFF)>>16);
+ src += sstride;
+ dst += dstride;
+ frame_ant += w;
+ if (s->denoise_row[depth]) {
+ s->denoise_row[depth](src, dst, line_ant, frame_ant, w, spatial, temporal);
+ continue;
}
+ pixel_ant = LOAD(0);
+ for (x = 0; x < w-1; x++) {
+ 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);
+ STORE(x, tmp);
+ }
+ line_ant[x] = tmp = lowpass(line_ant[x], pixel_ant, spatial, depth);
+ frame_ant[x] = tmp = lowpass(frame_ant[x], tmp, temporal, depth);
+ STORE(x, tmp);
}
}
-static void denoise(uint8_t *src, uint8_t *dst,
- uint32_t *line_ant, uint16_t **frame_ant_ptr,
- int w, int h, int sstride, int dstride,
- int *horizontal, int *vertical, int *temporal)
+av_always_inline
+static int denoise_depth(HQDN3DContext *s,
+ uint8_t *src, uint8_t *dst,
+ 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)
{
+ // FIXME: For 16bit depth, frame_ant could be a pointer to the previous
+ // filtered frame rather than a separate buffer.
long x, y;
- long sline_offs = 0, dline_offs = 0;
- uint32_t pixel_ant;
- uint32_t pixel;
uint16_t *frame_ant = *frame_ant_ptr;
-
if (!frame_ant) {
+ uint8_t *frame_src = src;
*frame_ant_ptr = frame_ant = av_malloc(w*h*sizeof(uint16_t));
- for (y = 0; y < h; y++) {
- uint16_t *frame_dst = frame_ant+y*w;
- uint8_t *frame_src = src+y*sstride;
+ if (!frame_ant)
+ return AVERROR(ENOMEM);
+ for (y = 0; y < h; y++, src += sstride, frame_ant += w)
for (x = 0; x < w; x++)
- frame_dst[x] = frame_src[x]<<8;
- }
+ frame_ant[x] = LOAD(x);
+ src = frame_src;
+ frame_ant = *frame_ant_ptr;
}
- if (!horizontal[0] && !vertical[0]) {
+ if (spatial[0])
+ denoise_spatial(s, src, dst, line_ant, frame_ant,
+ w, h, sstride, dstride, spatial, temporal, depth);
+ else
denoise_temporal(src, dst, frame_ant,
- w, h, sstride, dstride, temporal);
- return;
- }
- if (!temporal[0]) {
- denoise_spatial(src, dst, line_ant,
- w, h, sstride, dstride, horizontal, vertical);
- return;
- }
-
- /* First pixel has no left nor top neighbor. Only previous frame */
- line_ant[0] = pixel_ant = src[0]<<16;
- pixel = lowpass(frame_ant[0]<<8, pixel_ant, temporal);
- frame_ant[0] = ((pixel+0x1000007F)>>8);
- dst[0]= ((pixel+0x10007FFF)>>16);
-
- /* First line has no top neighbor. Only left one for each pixel and
- * last frame */
- for (x = 1; x < w; x++) {
- line_ant[x] = pixel_ant = lowpass(pixel_ant, src[x]<<16, horizontal);
- pixel = lowpass(frame_ant[x]<<8, pixel_ant, temporal);
- frame_ant[x] = ((pixel+0x1000007F)>>8);
- dst[x]= ((pixel+0x10007FFF)>>16);
- }
-
- for (y = 1; y < h; y++) {
- uint32_t pixel_ant;
- uint16_t *line_prev = frame_ant+y*w;
- sline_offs += sstride, dline_offs += dstride;
- /* First pixel on each line doesn't have previous pixel */
- pixel_ant = src[sline_offs]<<16;
- line_ant[0] = lowpass(line_ant[0], pixel_ant, vertical);
- pixel = lowpass(line_prev[0]<<8, line_ant[0], temporal);
- line_prev[0] = ((pixel+0x1000007F)>>8);
- dst[dline_offs]= ((pixel+0x10007FFF)>>16);
-
- for (x = 1; x < w; x++) {
- uint32_t pixel;
- /* The rest are normal */
- pixel_ant = lowpass(pixel_ant, src[sline_offs+x]<<16, horizontal);
- line_ant[x] = lowpass(line_ant[x], pixel_ant, vertical);
- pixel = lowpass(line_prev[x]<<8, line_ant[x], temporal);
- line_prev[x] = ((pixel+0x1000007F)>>8);
- dst[dline_offs+x]= ((pixel+0x10007FFF)>>16);
- }
- }
+ w, h, sstride, dstride, temporal, depth);
+ emms_c();
+ return 0;
}
-static void precalc_coefs(int *ct, double dist25)
+#define denoise(...) \
+ do { \
+ int ret = AVERROR_INVALIDDATA; \
+ switch (s->depth) { \
+ case 8: ret = denoise_depth(__VA_ARGS__, 8); break; \
+ case 9: ret = denoise_depth(__VA_ARGS__, 9); break; \
+ case 10: ret = denoise_depth(__VA_ARGS__, 10); break; \
+ case 16: ret = denoise_depth(__VA_ARGS__, 16); break; \
+ } \
+ if (ret < 0) { \
+ av_frame_free(&out); \
+ if (!direct) \
+ av_frame_free(&in); \
+ return ret; \
+ } \
+ } while (0)
+
+static int16_t *precalc_coefs(double dist25, int depth)
{
int i;
double gamma, simil, C;
+ int16_t *ct = av_malloc((512<<LUT_BITS)*sizeof(int16_t));
+ if (!ct)
+ return NULL;
- gamma = log(0.25) / log(1.0 - dist25/255.0 - 0.00001);
+ gamma = log(0.25) / log(1.0 - FFMIN(dist25,252.0)/255.0 - 0.00001);
- for (i = -255*16; i <= 255*16; i++) {
- simil = 1.0 - FFABS(i) / (16*255.0);
- C = pow(simil, gamma) * 65536.0 * i / 16.0;
- ct[16*256+i] = lrint(C);
+ 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
+ simil = 1.0 - FFABS(f) / 255.0;
+ C = pow(simil, gamma) * 256.0 * f;
+ ct[(256<<LUT_BITS)+i] = lrint(C);
}
ct[0] = !!dist25;
+ return ct;
}
#define PARAM1_DEFAULT 4.0
#define PARAM2_DEFAULT 3.0
#define PARAM3_DEFAULT 6.0
-static int init(AVFilterContext *ctx, const char *args)
+static av_cold int init(AVFilterContext *ctx)
{
- HQDN3DContext *hqdn3d = ctx->priv;
- double lum_spac, lum_tmp, chrom_spac, chrom_tmp;
- double param1, param2, param3, param4;
-
- lum_spac = PARAM1_DEFAULT;
- chrom_spac = PARAM2_DEFAULT;
- lum_tmp = PARAM3_DEFAULT;
- chrom_tmp = lum_tmp * chrom_spac / lum_spac;
-
- if (args) {
- switch (sscanf(args, "%lf:%lf:%lf:%lf",
- ¶m1, ¶m2, ¶m3, ¶m4)) {
- case 1:
- 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;
- break;
- case 2:
- lum_spac = param1;
- chrom_spac = param2;
- lum_tmp = PARAM3_DEFAULT * param1 / PARAM1_DEFAULT;
- chrom_tmp = lum_tmp * chrom_spac / lum_spac;
- break;
- case 3:
- lum_spac = param1;
- chrom_spac = param2;
- lum_tmp = param3;
- chrom_tmp = lum_tmp * chrom_spac / lum_spac;
- break;
- case 4:
- lum_spac = param1;
- chrom_spac = param2;
- lum_tmp = param3;
- chrom_tmp = param4;
- break;
- }
- }
+ HQDN3DContext *s = ctx->priv;
- av_log(ctx, AV_LOG_VERBOSE, "ls:%lf cs:%lf lt:%lf ct:%lf\n",
- lum_spac, chrom_spac, lum_tmp, chrom_tmp);
- if (lum_spac < 0 || chrom_spac < 0 || isnan(chrom_tmp)) {
- 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);
- }
+ if (!s->strength[LUMA_SPATIAL])
+ s->strength[LUMA_SPATIAL] = PARAM1_DEFAULT;
+ if (!s->strength[CHROMA_SPATIAL])
+ s->strength[CHROMA_SPATIAL] = PARAM2_DEFAULT * s->strength[LUMA_SPATIAL] / PARAM1_DEFAULT;
+ if (!s->strength[LUMA_TMP])
+ s->strength[LUMA_TMP] = PARAM3_DEFAULT * s->strength[LUMA_SPATIAL] / PARAM1_DEFAULT;
+ if (!s->strength[CHROMA_TMP])
+ s->strength[CHROMA_TMP] = s->strength[LUMA_TMP] * s->strength[CHROMA_SPATIAL] / s->strength[LUMA_SPATIAL];
- precalc_coefs(hqdn3d->coefs[0], lum_spac);
- precalc_coefs(hqdn3d->coefs[1], lum_tmp);
- precalc_coefs(hqdn3d->coefs[2], chrom_spac);
- precalc_coefs(hqdn3d->coefs[3], chrom_tmp);
+ av_log(ctx, AV_LOG_VERBOSE, "ls:%f cs:%f lt:%f ct:%f\n",
+ s->strength[LUMA_SPATIAL], s->strength[CHROMA_SPATIAL],
+ s->strength[LUMA_TMP], s->strength[CHROMA_TMP]);
return 0;
}
-static void uninit(AVFilterContext *ctx)
+static av_cold void uninit(AVFilterContext *ctx)
{
- HQDN3DContext *hqdn3d = ctx->priv;
-
- av_freep(&hqdn3d->line);
- av_freep(&hqdn3d->frame_prev[0]);
- av_freep(&hqdn3d->frame_prev[1]);
- av_freep(&hqdn3d->frame_prev[2]);
+ HQDN3DContext *s = ctx->priv;
+
+ av_freep(&s->coefs[0]);
+ av_freep(&s->coefs[1]);
+ av_freep(&s->coefs[2]);
+ av_freep(&s->coefs[3]);
+ av_freep(&s->line);
+ av_freep(&s->frame_prev[0]);
+ av_freep(&s->frame_prev[1]);
+ av_freep(&s->frame_prev[2]);
}
static int query_formats(AVFilterContext *ctx)
{
- static const enum PixelFormat pix_fmts[] = {
- PIX_FMT_YUV420P, PIX_FMT_YUV422P, PIX_FMT_YUV411P, PIX_FMT_NONE
+ 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
};
ff_set_common_formats(ctx, ff_make_format_list(pix_fmts));
static int config_input(AVFilterLink *inlink)
{
- HQDN3DContext *hqdn3d = inlink->dst->priv;
+ HQDN3DContext *s = inlink->dst->priv;
+ const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(inlink->format);
+ int i;
+
+ uninit(inlink->dst);
- hqdn3d->hsub = av_pix_fmt_descriptors[inlink->format].log2_chroma_w;
- hqdn3d->vsub = av_pix_fmt_descriptors[inlink->format].log2_chroma_h;
+ s->hsub = desc->log2_chroma_w;
+ s->vsub = desc->log2_chroma_h;
+ s->depth = desc->comp[0].depth;
- hqdn3d->line = av_malloc(inlink->w * sizeof(*hqdn3d->line));
- if (!hqdn3d->line)
+ s->line = av_malloc(inlink->w * sizeof(*s->line));
+ if (!s->line)
return AVERROR(ENOMEM);
- return 0;
-}
+ for (i = 0; i < 4; i++) {
+ s->coefs[i] = precalc_coefs(s->strength[i], s->depth);
+ if (!s->coefs[i])
+ return AVERROR(ENOMEM);
+ }
+
+ if (ARCH_X86)
+ ff_hqdn3d_init_x86(s);
-static int null_draw_slice(AVFilterLink *link, int y, int h, int slice_dir)
-{
return 0;
}
-static int end_frame(AVFilterLink *inlink)
+static int filter_frame(AVFilterLink *inlink, AVFrame *in)
{
- HQDN3DContext *hqdn3d = inlink->dst->priv;
+ HQDN3DContext *s = inlink->dst->priv;
AVFilterLink *outlink = inlink->dst->outputs[0];
- AVFilterBufferRef *inpic = inlink ->cur_buf;
- AVFilterBufferRef *outpic = outlink->out_buf;
- int cw = inpic->video->w >> hqdn3d->hsub;
- int ch = inpic->video->h >> hqdn3d->vsub;
- int ret;
-
- denoise(inpic->data[0], outpic->data[0],
- hqdn3d->line, &hqdn3d->frame_prev[0], inpic->video->w, inpic->video->h,
- inpic->linesize[0], outpic->linesize[0],
- hqdn3d->coefs[0],
- hqdn3d->coefs[0],
- hqdn3d->coefs[1]);
- denoise(inpic->data[1], outpic->data[1],
- hqdn3d->line, &hqdn3d->frame_prev[1], cw, ch,
- inpic->linesize[1], outpic->linesize[1],
- hqdn3d->coefs[2],
- hqdn3d->coefs[2],
- hqdn3d->coefs[3]);
- denoise(inpic->data[2], outpic->data[2],
- hqdn3d->line, &hqdn3d->frame_prev[2], cw, ch,
- inpic->linesize[2], outpic->linesize[2],
- hqdn3d->coefs[2],
- hqdn3d->coefs[2],
- hqdn3d->coefs[3]);
-
- if ((ret = ff_draw_slice(outlink, 0, inpic->video->h, 1)) < 0 ||
- (ret = ff_end_frame(outlink)) < 0)
- return ret;
- return 0;
+ AVFrame *out;
+ int c, direct = av_frame_is_writable(in);
+
+ if (direct) {
+ out = in;
+ } else {
+ out = ff_get_video_buffer(outlink, outlink->w, outlink->h);
+ if (!out) {
+ av_frame_free(&in);
+ return AVERROR(ENOMEM);
+ }
+
+ av_frame_copy_props(out, in);
+ out->width = outlink->w;
+ out->height = outlink->h;
+ }
+
+ for (c = 0; c < 3; c++) {
+ denoise(s, in->data[c], out->data[c],
+ s->line, &s->frame_prev[c],
+ in->width >> (!!c * s->hsub),
+ in->height >> (!!c * s->vsub),
+ in->linesize[c], out->linesize[c],
+ s->coefs[c?2:0], s->coefs[c?3:1]);
+ }
+
+ if (!direct)
+ av_frame_free(&in);
+
+ return ff_filter_frame(outlink, out);
}
-AVFilter avfilter_vf_hqdn3d = {
+#define OFFSET(x) offsetof(HQDN3DContext, x)
+#define FLAGS AV_OPT_FLAG_VIDEO_PARAM
+static const AVOption options[] = {
+ { "luma_spatial", "spatial luma strength", OFFSET(strength[LUMA_SPATIAL]), AV_OPT_TYPE_DOUBLE, { .dbl = 0.0 }, 0, DBL_MAX, FLAGS },
+ { "chroma_spatial", "spatial chroma strength", OFFSET(strength[CHROMA_SPATIAL]), AV_OPT_TYPE_DOUBLE, { .dbl = 0.0 }, 0, DBL_MAX, FLAGS },
+ { "luma_tmp", "temporal luma strength", OFFSET(strength[LUMA_TMP]), AV_OPT_TYPE_DOUBLE, { .dbl = 0.0 }, 0, DBL_MAX, FLAGS },
+ { "chroma_tmp", "temporal chroma strength", OFFSET(strength[CHROMA_TMP]), AV_OPT_TYPE_DOUBLE, { .dbl = 0.0 }, 0, DBL_MAX, FLAGS },
+ { NULL },
+};
+
+static const AVClass hqdn3d_class = {
+ .class_name = "hqdn3d",
+ .item_name = av_default_item_name,
+ .option = options,
+ .version = LIBAVUTIL_VERSION_INT,
+};
+
+static const AVFilterPad avfilter_vf_hqdn3d_inputs[] = {
+ {
+ .name = "default",
+ .type = AVMEDIA_TYPE_VIDEO,
+ .config_props = config_input,
+ .filter_frame = filter_frame,
+ },
+ { NULL }
+};
+
+static const AVFilterPad avfilter_vf_hqdn3d_outputs[] = {
+ {
+ .name = "default",
+ .type = AVMEDIA_TYPE_VIDEO
+ },
+ { NULL }
+};
+
+AVFilter ff_vf_hqdn3d = {
.name = "hqdn3d",
.description = NULL_IF_CONFIG_SMALL("Apply a High Quality 3D Denoiser."),
.priv_size = sizeof(HQDN3DContext),
+ .priv_class = &hqdn3d_class,
.init = init,
.uninit = uninit,
.query_formats = query_formats,
- .inputs = (const AVFilterPad[]) {{ .name = "default",
- .type = AVMEDIA_TYPE_VIDEO,
- .draw_slice = null_draw_slice,
- .config_props = config_input,
- .end_frame = end_frame },
- { .name = NULL}},
+ .inputs = avfilter_vf_hqdn3d_inputs,
- .outputs = (const AVFilterPad[]) {{ .name = "default",
- .type = AVMEDIA_TYPE_VIDEO },
- { .name = NULL}},
+ .outputs = avfilter_vf_hqdn3d_outputs,
};