mpv/video/out/filter_kernels.c

410 lines
13 KiB
C

/*
* Some of the filter code was taken from Glumpy:
* # Copyright (c) 2009-2016 Nicolas P. Rougier. All rights reserved.
* # Distributed under the (new) BSD License.
* (https://github.com/glumpy/glumpy/blob/master/glumpy/library/build-spatial-filters.py)
*
* Also see:
* - http://vector-agg.cvs.sourceforge.net/viewvc/vector-agg/agg-2.5/include/agg_image_filters.h
* - Vapoursynth plugin fmtconv (WTFPL Licensed), which is based on
* dither plugin for avisynth from the same author:
* https://github.com/vapoursynth/fmtconv/tree/master/src/fmtc
* - Paul Heckbert's "zoom"
* - XBMC: ConvolutionKernels.cpp etc.
*
* This file is part of mpv.
*
* This file can be distributed under the 3-clause license ("New BSD License").
*
* You can alternatively redistribute the non-Glumpy parts of this file and/or
* modify it under the terms of the GNU Lesser General Public
* License as published by the Free Software Foundation; either
* version 2.1 of the License, or (at your option) any later version.
*/
#include <stddef.h>
#include <string.h>
#include <math.h>
#include <assert.h>
#include "filter_kernels.h"
#include "common/common.h"
// NOTE: all filters are designed for discrete convolution
const struct filter_window *mp_find_filter_window(const char *name)
{
if (!name)
return NULL;
for (const struct filter_window *w = mp_filter_windows; w->name; w++) {
if (strcmp(w->name, name) == 0)
return w;
}
return NULL;
}
const struct filter_kernel *mp_find_filter_kernel(const char *name)
{
if (!name)
return NULL;
for (const struct filter_kernel *k = mp_filter_kernels; k->f.name; k++) {
if (strcmp(k->f.name, name) == 0)
return k;
}
return NULL;
}
// sizes = sorted list of available filter sizes, terminated with size 0
// inv_scale = source_size / dest_size
bool mp_init_filter(struct filter_kernel *filter, const int *sizes,
double inv_scale)
{
assert(filter->f.radius > 0);
// Only downscaling requires widening the filter
filter->filter_scale = MPMAX(1.0, inv_scale);
double src_radius = filter->f.radius * filter->filter_scale;
// Polar filters are dependent solely on the radius
if (filter->polar) {
filter->size = 1; // Not meaningful for EWA/polar scalers.
// Safety precaution to avoid generating a gigantic shader
if (src_radius > 16.0) {
src_radius = 16.0;
filter->filter_scale = src_radius / filter->f.radius;
return false;
}
return true;
}
int size = ceil(2.0 * src_radius);
// round up to smallest available size that's still large enough
if (size < sizes[0])
size = sizes[0];
const int *cursize = sizes;
while (size > *cursize && *cursize)
cursize++;
if (*cursize) {
filter->size = *cursize;
return true;
} else {
// The filter doesn't fit - instead of failing completely, use the
// largest filter available. This is incorrect, but better than refusing
// to do anything.
filter->size = cursize[-1];
filter->filter_scale = (filter->size/2.0) / filter->f.radius;
return false;
}
}
// Sample from a blurred and tapered window
static double sample_window(struct filter_window *kernel, double x)
{
if (!kernel->weight)
return 1.0;
// All windows are symmetric, this makes life easier
x = fabs(x);
// Stretch and taper the window size as needed
x = kernel->blur > 0.0 ? x / kernel->blur : x;
x = x <= kernel->taper ? 0.0 : (x - kernel->taper) / (1 - kernel->taper);
if (x < kernel->radius)
return kernel->weight(kernel, x);
return 0.0;
}
// Evaluate a filter's kernel and window at a given absolute position
static double sample_filter(struct filter_kernel *filter, double x)
{
// The window is always stretched to the entire kernel
double w = sample_window(&filter->w, x / filter->f.radius * filter->w.radius);
double k = w * sample_window(&filter->f, x);
return k < 0 ? (1 - filter->clamp) * k : k;
}
// Calculate the 1D filtering kernel for N sample points.
// N = number of samples, which is filter->size
// The weights will be stored in out_w[0] to out_w[N - 1]
// f = x0 - abs(x0), subpixel position in the range [0,1) or [0,1].
static void mp_compute_weights(struct filter_kernel *filter, double f,
float *out_w)
{
assert(filter->size > 0);
double sum = 0;
for (int n = 0; n < filter->size; n++) {
double x = f - (n - filter->size / 2 + 1);
double w = sample_filter(filter, x / filter->filter_scale);
out_w[n] = w;
sum += w;
}
// Normalize to preserve energy
for (int n = 0; n < filter->size; n++)
out_w[n] /= sum;
}
// Fill the given array with weights for the range [0.0, 1.0]. The array is
// interpreted as rectangular array of count * filter->size items, with a
// stride of `stride` floats in between each array element. (For polar filters,
// the `count` indicates the row size and filter->size/stride are ignored)
//
// There will be slight sampling error if these weights are used in a OpenGL
// texture as LUT directly. The sampling point of a texel is located at its
// center, so out_array[0] will end up at 0.5 / count instead of 0.0.
// Correct lookup requires a linear coordinate mapping from [0.0, 1.0] to
// [0.5 / count, 1.0 - 0.5 / count].
void mp_compute_lut(struct filter_kernel *filter, int count, int stride,
float *out_array)
{
if (filter->polar) {
filter->radius_cutoff = 0.0;
// Compute a 1D array indexed by radius
for (int x = 0; x < count; x++) {
double r = x * filter->f.radius / (count - 1);
out_array[x] = sample_filter(filter, r);
if (fabs(out_array[x]) > filter->value_cutoff)
filter->radius_cutoff = r;
}
} else {
// Compute a 2D array indexed by subpixel position
for (int n = 0; n < count; n++) {
mp_compute_weights(filter, n / (double)(count - 1),
out_array + stride * n);
}
}
}
typedef struct filter_window params;
static double box(params *p, double x)
{
// This is mathematically 1.0 everywhere, the clipping is done implicitly
// based on the radius.
return 1.0;
}
static double triangle(params *p, double x)
{
return fmax(0.0, 1.0 - fabs(x / p->radius));
}
static double cosine(params *p, double x)
{
return cos(x);
}
static double hanning(params *p, double x)
{
return 0.5 + 0.5 * cos(M_PI * x);
}
static double hamming(params *p, double x)
{
return 0.54 + 0.46 * cos(M_PI * x);
}
static double quadric(params *p, double x)
{
if (x < 0.5) {
return 0.75 - x * x;
} else if (x < 1.5) {
double t = x - 1.5;
return 0.5 * t * t;
}
return 0.0;
}
#define POW3(x) ((x) <= 0 ? 0 : (x) * (x) * (x))
static double bicubic(params *p, double x)
{
return (1.0/6.0) * ( POW3(x + 2)
- 4 * POW3(x + 1)
+ 6 * POW3(x)
- 4 * POW3(x - 1));
}
static double bessel_i0(double x)
{
double s = 1.0;
double y = x * x / 4.0;
double t = y;
int i = 2;
while (t > 1e-12) {
s += t;
t *= y / (i * i);
i += 1;
}
return s;
}
static double kaiser(params *p, double x)
{
if (x > 1)
return 0;
double i0a = 1.0 / bessel_i0(p->params[1]);
return bessel_i0(p->params[0] * sqrt(1.0 - x * x)) * i0a;
}
static double blackman(params *p, double x)
{
double a = p->params[0];
double a0 = (1-a)/2.0, a1 = 1/2.0, a2 = a/2.0;
double pix = M_PI * x;
return a0 + a1*cos(pix) + a2*cos(2 * pix);
}
static double welch(params *p, double x)
{
return 1.0 - x*x;
}
// Family of cubic B/C splines
static double cubic_bc(params *p, double x)
{
double b = p->params[0],
c = p->params[1];
double p0 = (6.0 - 2.0 * b) / 6.0,
p2 = (-18.0 + 12.0 * b + 6.0 * c) / 6.0,
p3 = (12.0 - 9.0 * b - 6.0 * c) / 6.0,
q0 = (8.0 * b + 24.0 * c) / 6.0,
q1 = (-12.0 * b - 48.0 * c) / 6.0,
q2 = (6.0 * b + 30.0 * c) / 6.0,
q3 = (-b - 6.0 * c) / 6.0;
if (x < 1.0) {
return p0 + x * x * (p2 + x * p3);
} else if (x < 2.0) {
return q0 + x * (q1 + x * (q2 + x * q3));
}
return 0.0;
}
static double spline16(params *p, double x)
{
if (x < 1.0) {
return ((x - 9.0/5.0 ) * x - 1.0/5.0 ) * x + 1.0;
} else {
return ((-1.0/3.0 * (x-1) + 4.0/5.0) * (x-1) - 7.0/15.0 ) * (x-1);
}
}
static double spline36(params *p, double x)
{
if (x < 1.0) {
return ((13.0/11.0 * x - 453.0/209.0) * x - 3.0/209.0) * x + 1.0;
} else if (x < 2.0) {
return ((-6.0/11.0 * (x-1) + 270.0/209.0) * (x-1) - 156.0/ 209.0) * (x-1);
} else {
return ((1.0/11.0 * (x-2) - 45.0/209.0) * (x-2) + 26.0/209.0) * (x-2);
}
}
static double spline64(params *p, double x)
{
if (x < 1.0) {
return ((49.0/41.0 * x - 6387.0/2911.0) * x - 3.0/2911.0) * x + 1.0;
} else if (x < 2.0) {
return ((-24.0/41.0 * (x-1) + 4032.0/2911.0) * (x-1) - 2328.0/2911.0) * (x-1);
} else if (x < 3.0) {
return ((6.0/41.0 * (x-2) - 1008.0/2911.0) * (x-2) + 582.0/2911.0) * (x-2);
} else {
return ((-1.0/41.0 * (x-3) + 168.0/2911.0) * (x-3) - 97.0/2911.0) * (x-3);
}
}
static double gaussian(params *p, double x)
{
return exp(-2.0 * x * x / p->params[0]);
}
static double sinc(params *p, double x)
{
if (fabs(x) < 1e-8)
return 1.0;
x *= M_PI;
return sin(x) / x;
}
static double jinc(params *p, double x)
{
if (fabs(x) < 1e-8)
return 1.0;
x *= M_PI;
return 2.0 * j1(x) / x;
}
static double sphinx(params *p, double x)
{
if (fabs(x) < 1e-8)
return 1.0;
x *= M_PI;
return 3.0 * (sin(x) - x * cos(x)) / (x * x * x);
}
const struct filter_window mp_filter_windows[] = {
{"box", 1, box},
{"triangle", 1, triangle},
{"bartlett", 1, triangle},
{"cosine", M_PI_2, cosine},
{"hanning", 1, hanning},
{"tukey", 1, hanning, .taper = 0.5},
{"hamming", 1, hamming},
{"quadric", 1.5, quadric},
{"welch", 1, welch},
{"kaiser", 1, kaiser, .params = {6.33, NAN} },
{"blackman", 1, blackman, .params = {0.16, NAN} },
{"gaussian", 2, gaussian, .params = {1.00, NAN} },
{"sinc", 1, sinc},
{"jinc", 1.2196698912665045, jinc},
{"sphinx", 1.4302966531242027, sphinx},
{0}
};
const struct filter_kernel mp_filter_kernels[] = {
// Spline filters
{{"spline16", 2, spline16}},
{{"spline36", 3, spline36}},
{{"spline64", 4, spline64}},
// Sinc filters
{{"sinc", 2, sinc, .resizable = true}},
{{"lanczos", 3, sinc, .resizable = true}, .window = "sinc"},
{{"ginseng", 3, sinc, .resizable = true}, .window = "jinc"},
// Jinc filters
{{"jinc", 3, jinc, .resizable = true}, .polar = true},
{{"ewa_lanczos", 3, jinc, .resizable = true}, .polar = true, .window = "jinc"},
{{"ewa_hanning", 3, jinc, .resizable = true}, .polar = true, .window = "hanning" },
{{"ewa_ginseng", 3, jinc, .resizable = true}, .polar = true, .window = "sinc"},
// Radius is based on the true jinc radius, slightly sharpened as per
// calculations by Nicolas Robidoux. Source: Imagemagick's magick/resize.c
{{"ewa_lanczossharp", 3.2383154841662362, jinc, .blur = 0.9812505644269356,
.resizable = true}, .polar = true, .window = "jinc"},
// Similar to the above, but softened instead. This one makes hash patterns
// disappear completely. Blur determined by trial and error.
{{"ewa_lanczossoft", 3.2383154841662362, jinc, .blur = 1.015,
.resizable = true}, .polar = true, .window = "jinc"},
// Very soft (blurred) hanning-windowed jinc; removes almost all aliasing.
// Blur parameter picked to match orthogonal and diagonal contributions
{{"haasnsoft", 3.2383154841662362, jinc, .blur = 1.11, .resizable = true},
.polar = true, .window = "hanning"},
// Cubic filters
{{"bicubic", 2, bicubic}},
{{"bcspline", 2, cubic_bc, .params = {0.5, 0.5} }},
{{"catmull_rom", 2, cubic_bc, .params = {0.0, 0.5} }},
{{"mitchell", 2, cubic_bc, .params = {1.0/3.0, 1.0/3.0} }},
{{"robidoux", 2, cubic_bc, .params = {12 / (19 + 9 * M_SQRT2),
113 / (58 + 216 * M_SQRT2)} }},
{{"robidouxsharp", 2, cubic_bc, .params = {6 / (13 + 7 * M_SQRT2),
7 / (2 + 12 * M_SQRT2)} }},
{{"ewa_robidoux", 2, cubic_bc, .params = {12 / (19 + 9 * M_SQRT2),
113 / (58 + 216 * M_SQRT2)}},
.polar = true},
{{"ewa_robidouxsharp", 2,cubic_bc, .params = {6 / (13 + 7 * M_SQRT2),
7 / (2 + 12 * M_SQRT2)}},
.polar = true},
// Miscellaneous filters
{{"box", 1, box, .resizable = true}},
{{"nearest", 0.5, box}},
{{"triangle", 1, triangle, .resizable = true}},
{{"gaussian", 2, gaussian, .params = {1.0, NAN}, .resizable = true}},
{{0}}
};