[FFmpeg-devel] [PATCH] vf_colorspace: add floyd-steinberg dithering option to full conversion.
Ronald S. Bultje
rsbultje at gmail.com
Tue May 3 19:49:35 CEST 2016
---
doc/filters.texi | 13 ++++
libavfilter/colorspacedsp.c | 12 ++++
libavfilter/colorspacedsp.h | 6 ++
libavfilter/colorspacedsp_template.c | 128 +++++++++++++++++++++++++++++++++++
libavfilter/vf_colorspace.c | 58 +++++++++++++++-
5 files changed, 214 insertions(+), 3 deletions(-)
diff --git a/doc/filters.texi b/doc/filters.texi
index b17b115..98a002b 100644
--- a/doc/filters.texi
+++ b/doc/filters.texi
@@ -5104,6 +5104,19 @@ YUV 4:4:4 planar 12-bits
Do a fast conversion, which skips gamma/primary correction. This will take
significantly less CPU, but will be mathematically incorrect. To get output
compatible with that produced by the colormatrix filter, use fast=1.
+
+ at item dither
+Specify dithering mode.
+
+The accepted values are:
+ at table @samp
+ at item none
+No dithering
+
+ at item fsb
+Floyd-Steinberg dithering
+ at end table
+
@end table
The filter converts the transfer characteristics, color space and color
diff --git a/libavfilter/colorspacedsp.c b/libavfilter/colorspacedsp.c
index d4c43c3..f95805b 100644
--- a/libavfilter/colorspacedsp.c
+++ b/libavfilter/colorspacedsp.c
@@ -20,6 +20,9 @@
#include "colorspacedsp.h"
+/*
+ * SS_W/H are synonyms for AVPixFmtDescriptor->log2_chroma_w/h.
+ */
#define SS_W 0
#define SS_H 0
@@ -114,6 +117,15 @@ void ff_colorspacedsp_init(ColorSpaceDSPContext *dsp)
init_rgb2yuv_fn(1, 10);
init_rgb2yuv_fn(2, 12);
+#define init_rgb2yuv_fsb_fn(idx, bit) \
+ dsp->rgb2yuv_fsb[idx][0] = rgb2yuv_fsb_444p##bit##_c; \
+ dsp->rgb2yuv_fsb[idx][1] = rgb2yuv_fsb_422p##bit##_c; \
+ dsp->rgb2yuv_fsb[idx][2] = rgb2yuv_fsb_420p##bit##_c
+
+ init_rgb2yuv_fsb_fn(0, 8);
+ init_rgb2yuv_fsb_fn(1, 10);
+ init_rgb2yuv_fsb_fn(2, 12);
+
#define init_yuv2yuv_fn(idx1, idx2, bit1, bit2) \
dsp->yuv2yuv[idx1][idx2][0] = yuv2yuv_444p##bit1##to##bit2##_c; \
dsp->yuv2yuv[idx1][idx2][1] = yuv2yuv_422p##bit1##to##bit2##_c; \
diff --git a/libavfilter/colorspacedsp.h b/libavfilter/colorspacedsp.h
index 4e70c6c..2ca7b19 100644
--- a/libavfilter/colorspacedsp.h
+++ b/libavfilter/colorspacedsp.h
@@ -32,6 +32,11 @@ typedef void (*rgb2yuv_fn)(uint8_t *yuv[3], ptrdiff_t yuv_stride[3],
int16_t *rgb[3], ptrdiff_t rgb_stride,
int w, int h, const int16_t rgb2yuv_coeffs[3][3][8],
const int16_t yuv_offset[8]);
+typedef void (*rgb2yuv_fsb_fn)(uint8_t *yuv[3], ptrdiff_t yuv_stride[3],
+ int16_t *rgb[3], ptrdiff_t rgb_stride,
+ int w, int h, const int16_t rgb2yuv_coeffs[3][3][8],
+ const int16_t yuv_offset[8],
+ int *rnd[3][2]);
typedef void (*yuv2yuv_fn)(uint8_t *yuv_out[3], ptrdiff_t yuv_out_stride[3],
uint8_t *yuv_in[3], ptrdiff_t yuv_in_stride[3],
int w, int h, const int16_t yuv2yuv_coeffs[3][3][8],
@@ -40,6 +45,7 @@ typedef void (*yuv2yuv_fn)(uint8_t *yuv_out[3], ptrdiff_t yuv_out_stride[3],
typedef struct ColorSpaceDSPContext {
yuv2rgb_fn yuv2rgb[3 /* 0: 8bit, 1: 10bit, 2: 12bit */][3 /* 0: 444, 1: 422, 2: 420 */];
rgb2yuv_fn rgb2yuv[3 /* 0: 8bit, 1: 10bit, 2: 12bit */][3 /* 0: 444, 1: 422, 2: 420 */];
+ rgb2yuv_fsb_fn rgb2yuv_fsb[3 /* 0: 8bit, 1: 10bit, 2: 12bit */][3 /* 0: 444, 1: 422, 2: 420 */];
yuv2yuv_fn yuv2yuv[3 /* in_depth */][3 /* out_depth */][3 /* 0: 444, 1: 422, 2: 420 */];
void (*multiply3x3)(int16_t *data[3], ptrdiff_t stride,
diff --git a/libavfilter/colorspacedsp_template.c b/libavfilter/colorspacedsp_template.c
index f225391..db4a8d2 100644
--- a/libavfilter/colorspacedsp_template.c
+++ b/libavfilter/colorspacedsp_template.c
@@ -199,6 +199,134 @@ static void fn(rgb2yuv)(uint8_t *_yuv[3], ptrdiff_t yuv_stride[3],
}
}
+/* floyd-steinberg dithering - for any mid-top pixel A in a 3x2 block of pixels:
+ * 1 A 2
+ * 3 4 5
+ * the rounding error is distributed over the neighbouring pixels:
+ * 2: 7/16th, 3: 3/16th, 4: 5/16th and 5: 1/16th
+ */
+static void fn(rgb2yuv_fsb)(uint8_t *_yuv[3], ptrdiff_t yuv_stride[3],
+ int16_t *rgb[3], ptrdiff_t s,
+ int w, int h, const int16_t rgb2yuv_coeffs[3][3][8],
+ const int16_t yuv_offset[8],
+ int *rnd_scratch[3][2])
+{
+ pixel **yuv = (pixel **) _yuv;
+ pixel *yuv0 = yuv[0], *yuv1 = yuv[1], *yuv2 = yuv[2];
+ const int16_t *rgb0 = rgb[0], *rgb1 = rgb[1], *rgb2 = rgb[2];
+ int y, x;
+ const int sh = 29 - BIT_DEPTH;
+ const int rnd = 1 << (sh - 1);
+ int cry = rgb2yuv_coeffs[0][0][0];
+ int cgy = rgb2yuv_coeffs[0][1][0];
+ int cby = rgb2yuv_coeffs[0][2][0];
+ int cru = rgb2yuv_coeffs[1][0][0];
+ int cgu = rgb2yuv_coeffs[1][1][0];
+ int cburv = rgb2yuv_coeffs[1][2][0];
+ int cgv = rgb2yuv_coeffs[2][1][0];
+ int cbv = rgb2yuv_coeffs[2][2][0];
+ ptrdiff_t s0 = yuv_stride[0] / sizeof(pixel);
+ const int uv_offset = 128 << (BIT_DEPTH - 8);
+ unsigned mask = (1 << sh) - 1;
+
+ for (x = 0; x < w; x++) {
+ rnd_scratch[0][0][x] =
+ rnd_scratch[0][1][x] = rnd;
+ }
+ av_assert2(rgb2yuv_coeffs[1][2][0] == rgb2yuv_coeffs[2][0][0]);
+ w = AV_CEIL_RSHIFT(w, SS_W);
+ h = AV_CEIL_RSHIFT(h, SS_H);
+ for (x = 0; x < w; x++) {
+ rnd_scratch[1][0][x] =
+ rnd_scratch[1][1][x] =
+ rnd_scratch[2][0][x] =
+ rnd_scratch[2][1][x] = rnd;
+ }
+ for (y = 0; y < h; y++) {
+ for (x = 0; x < w; x++) {
+ int r00 = rgb0[x << SS_W], g00 = rgb1[x << SS_W], b00 = rgb2[x << SS_W];
+ int y00;
+#if SS_W == 1
+ int r01 = rgb0[x * 2 + 1], g01 = rgb1[x * 2 + 1], b01 = rgb2[x * 2 + 1];
+ int y01;
+#if SS_H == 1
+ int r10 = rgb0[x * 2 + 0 + s], g10 = rgb1[x * 2 + 0 + s], b10 = rgb2[x * 2 + 0 + s];
+ int r11 = rgb0[x * 2 + 1 + s], g11 = rgb1[x * 2 + 1 + s], b11 = rgb2[x * 2 + 1 + s];
+ int y10, y11;
+#endif
+#endif
+ int u, v, diff;
+
+ y00 = r00 * cry + g00 * cgy + b00 * cby + rnd_scratch[0][y & !SS_H][x << SS_W];
+ diff = (y00 & mask) - rnd;
+ yuv0[x << SS_W] = av_clip_pixel(yuv_offset[0] + (y00 >> sh));
+ rnd_scratch[0][ (y & !SS_H)][(x << SS_W) + 1] += (diff * 7 + 8) >> 4;
+ rnd_scratch[0][!(y & !SS_H)][(x << SS_W) - 1] += (diff * 3 + 8) >> 4;
+ rnd_scratch[0][!(y & !SS_H)][(x << SS_W) + 0] += (diff * 5 + 8) >> 4;
+ rnd_scratch[0][!(y & !SS_H)][(x << SS_W) + 1] += (diff * 1 + 8) >> 4;
+ rnd_scratch[0][ (y & !SS_H)][(x << SS_W) + 0] = rnd;
+#if SS_W == 1
+ y01 = r01 * cry + g01 * cgy + b01 * cby + rnd_scratch[0][y & !SS_H][x * 2 + 1];
+ diff = (y01 & mask) - rnd;
+ yuv0[x * 2 + 1] = av_clip_pixel(yuv_offset[0] + (y01 >> sh));
+ rnd_scratch[0][ (y & !SS_H)][x * 2 + 2] += (diff * 7 + 8) >> 4;
+ rnd_scratch[0][!(y & !SS_H)][x * 2 + 0] += (diff * 3 + 8) >> 4;
+ rnd_scratch[0][!(y & !SS_H)][x * 2 + 1] += (diff * 5 + 8) >> 4;
+ rnd_scratch[0][!(y & !SS_H)][x * 2 + 2] += (diff * 1 + 8) >> 4;
+ rnd_scratch[0][ (y & !SS_H)][x * 2 + 1] = rnd;
+#if SS_H == 1
+ y10 = r10 * cry + g10 * cgy + b10 * cby + rnd_scratch[0][1][x * 2 + 0];
+ diff = (y10 & mask) - rnd;
+ yuv0[x * 2 + 0 + s0] = av_clip_pixel(yuv_offset[0] + (y10 >> sh));
+ rnd_scratch[0][1][x * 2 + 1] += (diff * 7 + 8) >> 4;
+ rnd_scratch[0][0][x * 2 - 1] += (diff * 3 + 8) >> 4;
+ rnd_scratch[0][0][x * 2 + 0] += (diff * 5 + 8) >> 4;
+ rnd_scratch[0][0][x * 2 + 1] += (diff * 1 + 8) >> 4;
+ rnd_scratch[0][1][x * 2 + 0] = rnd;
+
+ y11 = r11 * cry + g11 * cgy + b11 * cby + rnd_scratch[0][1][x * 2 + 1];
+ diff = (y11 & mask) - rnd;
+ yuv0[x * 2 + 1 + s0] = av_clip_pixel(yuv_offset[0] + (y11 >> sh));
+ rnd_scratch[0][1][x * 2 + 2] += (diff * 7 + 8) >> 4;
+ rnd_scratch[0][0][x * 2 + 0] += (diff * 3 + 8) >> 4;
+ rnd_scratch[0][0][x * 2 + 1] += (diff * 5 + 8) >> 4;
+ rnd_scratch[0][0][x * 2 + 2] += (diff * 1 + 8) >> 4;
+ rnd_scratch[0][1][x * 2 + 1] = rnd;
+#endif
+#endif
+
+ u = avg(r00, r01, r10, r11) * cru +
+ avg(g00, g01, g10, g11) * cgu +
+ avg(b00, b01, b10, b11) * cburv + rnd_scratch[1][y & 1][x];
+ diff = (u & mask) - rnd;
+ yuv1[x] = av_clip_pixel(uv_offset + (u >> sh));
+ rnd_scratch[1][ (y & 1)][x + 1] += (diff * 7 + 8) >> 4;
+ rnd_scratch[1][!(y & 1)][x - 1] += (diff * 3 + 8) >> 4;
+ rnd_scratch[1][!(y & 1)][x + 0] += (diff * 5 + 8) >> 4;
+ rnd_scratch[1][!(y & 1)][x + 1] += (diff * 1 + 8) >> 4;
+ rnd_scratch[1][ (y & 1)][x + 0] = rnd;
+
+ v = avg(r00, r01, r10, r11) * cburv +
+ avg(g00, g01, g10, g11) * cgv +
+ avg(b00, b01, b10, b11) * cbv + rnd_scratch[2][y & 1][x];
+ diff = (v & mask) - rnd;
+ yuv2[x] = av_clip_pixel(uv_offset + (v >> sh));
+ rnd_scratch[2][ (y & 1)][x + 1] += (diff * 7 + 8) >> 4;
+ rnd_scratch[2][!(y & 1)][x - 1] += (diff * 3 + 8) >> 4;
+ rnd_scratch[2][!(y & 1)][x + 0] += (diff * 5 + 8) >> 4;
+ rnd_scratch[2][!(y & 1)][x + 1] += (diff * 1 + 8) >> 4;
+ rnd_scratch[2][ (y & 1)][x + 0] = rnd;
+ }
+
+ yuv0 += s0 * (1 << SS_H);
+ yuv1 += yuv_stride[1] / sizeof(pixel);
+ yuv2 += yuv_stride[2] / sizeof(pixel);
+ rgb0 += s * (1 << SS_H);
+ rgb1 += s * (1 << SS_H);
+ rgb2 += s * (1 << SS_H);
+ }
+}
+
#undef IN_BIT_DEPTH
#undef OUT_BIT_DEPTH
#define OUT_BIT_DEPTH BIT_DEPTH
diff --git a/libavfilter/vf_colorspace.c b/libavfilter/vf_colorspace.c
index 4003f70..408d0e0 100644
--- a/libavfilter/vf_colorspace.c
+++ b/libavfilter/vf_colorspace.c
@@ -34,6 +34,12 @@
#include "internal.h"
#include "video.h"
+enum DitherMode {
+ DITHER_NONE,
+ DITHER_FSB,
+ DITHER_NB,
+};
+
enum Colorspace {
CS_UNSPECIFIED,
CS_BT470M,
@@ -121,10 +127,12 @@ typedef struct ColorSpaceContext {
enum AVColorPrimaries in_prm, out_prm, user_prm, iprm;
enum AVPixelFormat in_format, user_format;
int fast_mode;
+ enum DitherMode dither;
int16_t *rgb[3];
ptrdiff_t rgb_stride;
unsigned rgb_sz;
+ int *dither_scratch[3][2], *dither_scratch_base[3][2];
const struct ColorPrimaries *in_primaries, *out_primaries;
int lrgb2lrgb_passthrough;
@@ -142,6 +150,7 @@ typedef struct ColorSpaceContext {
DECLARE_ALIGNED(16, int16_t, yuv_offset)[2 /* in, out */][8];
yuv2rgb_fn yuv2rgb;
rgb2yuv_fn rgb2yuv;
+ rgb2yuv_fsb_fn rgb2yuv_fsb;
yuv2yuv_fn yuv2yuv;
double yuv2rgb_dbl_coeffs[3][3], rgb2yuv_dbl_coeffs[3][3];
int in_y_rng, in_uv_rng, out_y_rng, out_uv_rng;
@@ -481,8 +490,13 @@ static int convert(AVFilterContext *ctx, void *data, int job_nr, int n_jobs)
s->dsp.multiply3x3(rgb, s->rgb_stride, w, h, s->lrgb2lrgb_coeffs);
apply_lut(rgb, s->rgb_stride, w, h, s->delin_lut);
}
- s->rgb2yuv(out_data, td->out_linesize, rgb, s->rgb_stride, w, h,
- s->rgb2yuv_coeffs, s->yuv_offset[1]);
+ if (s->dither == DITHER_FSB) {
+ s->rgb2yuv_fsb(out_data, td->out_linesize, rgb, s->rgb_stride, w, h,
+ s->rgb2yuv_coeffs, s->yuv_offset[1], s->dither_scratch);
+ } else {
+ s->rgb2yuv(out_data, td->out_linesize, rgb, s->rgb_stride, w, h,
+ s->rgb2yuv_coeffs, s->yuv_offset[1]);
+ }
}
return 0;
@@ -784,6 +798,8 @@ static int create_filtergraph(AVFilterContext *ctx,
av_assert2(s->rgb2yuv_coeffs[1][2][0] == s->rgb2yuv_coeffs[2][0][0]);
s->rgb2yuv = s->dsp.rgb2yuv[(out_desc->comp[0].depth - 8) >> 1]
[out_desc->log2_chroma_h + out_desc->log2_chroma_w];
+ s->rgb2yuv_fsb = s->dsp.rgb2yuv_fsb[(out_desc->comp[0].depth - 8) >> 1]
+ [out_desc->log2_chroma_h + out_desc->log2_chroma_w];
emms = 1;
}
@@ -834,6 +850,12 @@ static void uninit(AVFilterContext *ctx)
av_freep(&s->rgb[1]);
av_freep(&s->rgb[2]);
s->rgb_sz = 0;
+ av_freep(&s->dither_scratch_base[0][0]);
+ av_freep(&s->dither_scratch_base[0][1]);
+ av_freep(&s->dither_scratch_base[1][0]);
+ av_freep(&s->dither_scratch_base[1][1]);
+ av_freep(&s->dither_scratch_base[2][0]);
+ av_freep(&s->dither_scratch_base[2][1]);
av_freep(&s->lin_lut);
}
@@ -888,15 +910,39 @@ static int filter_frame(AVFilterLink *link, AVFrame *in)
else if (s->iall != CS_UNSPECIFIED)
in->color_trc = default_trc[s->iall];
if (rgb_sz != s->rgb_sz) {
+ const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(out->format);
+ int uvw = in->width >> desc->log2_chroma_w;
+
av_freep(&s->rgb[0]);
av_freep(&s->rgb[1]);
av_freep(&s->rgb[2]);
s->rgb_sz = 0;
+ av_freep(&s->dither_scratch_base[0][0]);
+ av_freep(&s->dither_scratch_base[0][1]);
+ av_freep(&s->dither_scratch_base[1][0]);
+ av_freep(&s->dither_scratch_base[1][1]);
+ av_freep(&s->dither_scratch_base[2][0]);
+ av_freep(&s->dither_scratch_base[2][1]);
s->rgb[0] = av_malloc(rgb_sz);
s->rgb[1] = av_malloc(rgb_sz);
s->rgb[2] = av_malloc(rgb_sz);
- if (!s->rgb[0] || !s->rgb[1] || !s->rgb[2]) {
+ s->dither_scratch_base[0][0] = av_malloc(sizeof(int) * (in->width + 4));
+ s->dither_scratch_base[0][1] = av_malloc(sizeof(int) * (in->width + 4));
+ s->dither_scratch_base[1][0] = av_malloc(sizeof(int) * (uvw + 4));
+ s->dither_scratch_base[1][1] = av_malloc(sizeof(int) * (uvw + 4));
+ s->dither_scratch_base[2][0] = av_malloc(sizeof(int) * (uvw + 4));
+ s->dither_scratch_base[2][1] = av_malloc(sizeof(int) * (uvw + 4));
+ s->dither_scratch[0][0] = &s->dither_scratch_base[0][0][1];
+ s->dither_scratch[0][1] = &s->dither_scratch_base[0][1][1];
+ s->dither_scratch[1][0] = &s->dither_scratch_base[1][0][1];
+ s->dither_scratch[1][1] = &s->dither_scratch_base[1][1][1];
+ s->dither_scratch[2][0] = &s->dither_scratch_base[2][0][1];
+ s->dither_scratch[2][1] = &s->dither_scratch_base[2][1][1];
+ if (!s->rgb[0] || !s->rgb[1] || !s->rgb[2] ||
+ !s->dither_scratch[0][0] || !s->dither_scratch[0][1] ||
+ !s->dither_scratch[1][0] || !s->dither_scratch[1][1] ||
+ !s->dither_scratch[2][0] || !s->dither_scratch[2][1]) {
uninit(ctx);
return AVERROR(ENOMEM);
}
@@ -1046,6 +1092,12 @@ static const AVOption colorspace_options[] = {
{ "fast", "Ignore primary chromaticity and gamma correction",
OFFSET(fast_mode), AV_OPT_TYPE_BOOL, { .i64 = 0 },
0, 1, FLAGS },
+ { "dither", "Dithering mode",
+ OFFSET(dither), AV_OPT_TYPE_INT, { .i64 = DITHER_NONE },
+ DITHER_NONE, DITHER_NB - 1, FLAGS, "dither" },
+ ENUM("none", DITHER_NONE, "dither"),
+ ENUM("fsb", DITHER_FSB, "dither"),
+
{ NULL }
};
--
2.8.1
More information about the ffmpeg-devel
mailing list