mirror of https://github.com/GNOME/gimp.git
Bug 780859 - Brush hardness blur is slow
Add a specialized convolution algorithm for the hardness blur. It uses the same kernel as before, but performs the convolution in (amortized) O(1)-per-pixel time, instead of O(n^2), where n is the size of the kernel (or the blur radius). Note that the new code performs the convolution in the input color space, instead of always using a linear space. Since our brush pixmaps (but the not masks) are currently perceptual, the result is a bit different.
This commit is contained in:
Ell
parent
c585c99e80
commit
da30b86ffc
|
|
@ -19,6 +19,8 @@
|
|||
|
||||
#include "config.h"
|
||||
|
||||
#include <string.h>
|
||||
|
||||
#include <gdk-pixbuf/gdk-pixbuf.h>
|
||||
#include <gegl.h>
|
||||
|
||||
|
|
@ -45,10 +47,8 @@ static void gimp_brush_transform_bounding_box (GimpBrush *brush,
|
|||
gint *width,
|
||||
gint *height);
|
||||
|
||||
static gdouble gimp_brush_transform_array_sum (gfloat *arr,
|
||||
gint len);
|
||||
static void gimp_brush_transform_fill_blur_kernel (gfloat *arr,
|
||||
gint len);
|
||||
static void gimp_brush_transform_blur (GimpTempBuf *buf,
|
||||
gint r);
|
||||
static gint gimp_brush_transform_blur_kernel_size (gint height,
|
||||
gint width,
|
||||
gdouble hardness);
|
||||
|
|
@ -347,36 +347,7 @@ gimp_brush_real_transform_mask (GimpBrush *brush,
|
|||
|
||||
if (hardness < 1.0)
|
||||
{
|
||||
GimpTempBuf *blur_src;
|
||||
GeglBuffer *src_buffer;
|
||||
GeglBuffer *dest_buffer;
|
||||
gint kernel_len = kernel_size * kernel_size;
|
||||
gfloat blur_kernel[kernel_len];
|
||||
|
||||
gimp_brush_transform_fill_blur_kernel (blur_kernel, kernel_len);
|
||||
|
||||
blur_src = gimp_temp_buf_copy (result);
|
||||
|
||||
src_buffer = gimp_temp_buf_create_buffer (blur_src);
|
||||
dest_buffer = gimp_temp_buf_create_buffer (result);
|
||||
|
||||
gimp_temp_buf_unref (blur_src);
|
||||
|
||||
gimp_gegl_convolve (src_buffer,
|
||||
GEGL_RECTANGLE (0, 0,
|
||||
gimp_temp_buf_get_width (blur_src),
|
||||
gimp_temp_buf_get_height (blur_src)),
|
||||
dest_buffer,
|
||||
GEGL_RECTANGLE (0, 0,
|
||||
gimp_temp_buf_get_width (result),
|
||||
gimp_temp_buf_get_height (result)),
|
||||
blur_kernel, kernel_size,
|
||||
gimp_brush_transform_array_sum (blur_kernel,
|
||||
kernel_len),
|
||||
GIMP_NORMAL_CONVOL, FALSE);
|
||||
|
||||
g_object_unref (src_buffer);
|
||||
g_object_unref (dest_buffer);
|
||||
gimp_brush_transform_blur (result, kernel_size / 2);
|
||||
}
|
||||
|
||||
return result;
|
||||
|
|
@ -658,36 +629,7 @@ gimp_brush_real_transform_pixmap (GimpBrush *brush,
|
|||
|
||||
if (hardness < 1.0)
|
||||
{
|
||||
GimpTempBuf *blur_src;
|
||||
GeglBuffer *src_buffer;
|
||||
GeglBuffer *dest_buffer;
|
||||
gint kernel_len = kernel_size * kernel_size;
|
||||
gfloat blur_kernel[kernel_len];
|
||||
|
||||
gimp_brush_transform_fill_blur_kernel (blur_kernel, kernel_len);
|
||||
|
||||
blur_src = gimp_temp_buf_copy (result);
|
||||
|
||||
src_buffer = gimp_temp_buf_create_buffer (blur_src);
|
||||
dest_buffer = gimp_temp_buf_create_buffer (result);
|
||||
|
||||
gimp_temp_buf_unref (blur_src);
|
||||
|
||||
gimp_gegl_convolve (src_buffer,
|
||||
GEGL_RECTANGLE (0, 0,
|
||||
gimp_temp_buf_get_width (blur_src),
|
||||
gimp_temp_buf_get_height (blur_src)),
|
||||
dest_buffer,
|
||||
GEGL_RECTANGLE (0, 0,
|
||||
gimp_temp_buf_get_width (result),
|
||||
gimp_temp_buf_get_height (result)),
|
||||
blur_kernel, kernel_size,
|
||||
gimp_brush_transform_array_sum (blur_kernel,
|
||||
kernel_len),
|
||||
GIMP_NORMAL_CONVOL, FALSE);
|
||||
|
||||
g_object_unref (src_buffer);
|
||||
g_object_unref (dest_buffer);
|
||||
gimp_brush_transform_blur (result, kernel_size / 2);
|
||||
}
|
||||
|
||||
return result;
|
||||
|
|
@ -775,35 +717,178 @@ gimp_brush_transform_bounding_box (GimpBrush *brush,
|
|||
*height = MAX (1, *height);
|
||||
}
|
||||
|
||||
static gdouble
|
||||
gimp_brush_transform_array_sum (gfloat *arr,
|
||||
gint len)
|
||||
{
|
||||
gfloat total = 0;
|
||||
gint i;
|
||||
|
||||
for (i = 0; i < len; i++)
|
||||
{
|
||||
total += arr [i];
|
||||
}
|
||||
|
||||
return total;
|
||||
}
|
||||
|
||||
/* Blurs the brush mask/pixmap using a convolution of the form:
|
||||
*
|
||||
* 12 11 10 9 8
|
||||
* 7 6 5 4 3
|
||||
* 2 1 0 1 2
|
||||
* 3 4 5 6 7
|
||||
* 8 9 10 11 12
|
||||
*
|
||||
* (i.e., an array, wrapped into a matrix, whose i-th element is
|
||||
* `abs (i - a / 2)`, where `a` is the length of the array.) `r` specifies the
|
||||
* convolution kernel's radius.
|
||||
*/
|
||||
static void
|
||||
gimp_brush_transform_fill_blur_kernel (gfloat *arr,
|
||||
gint len)
|
||||
gimp_brush_transform_blur (GimpTempBuf *buf,
|
||||
gint r)
|
||||
{
|
||||
gint half_point = ((gint) len / 2) + 1;
|
||||
gint i;
|
||||
typedef struct
|
||||
{
|
||||
gint sum;
|
||||
gint weighted_sum;
|
||||
gint middle_sum;
|
||||
} Sums;
|
||||
|
||||
for (i = 0; i < len; i++)
|
||||
const Babl *format = gimp_temp_buf_get_format (buf);
|
||||
gint components = babl_format_get_n_components (format);
|
||||
gint width = gimp_temp_buf_get_width (buf);
|
||||
gint height = gimp_temp_buf_get_height (buf);
|
||||
gint stride = components * width;
|
||||
guchar *data = gimp_temp_buf_get_data (buf);
|
||||
gint rw = MIN (r, width - 1);
|
||||
gint rh = MIN (r, height - 1);
|
||||
gint n = 2 * r + 1;
|
||||
gint weight = (n * n / 2) * (n * n / 2 + 1);
|
||||
gint x;
|
||||
gint y;
|
||||
gint c;
|
||||
Sums *sums;
|
||||
guchar *d;
|
||||
Sums *s;
|
||||
|
||||
if (rw <= 0 || rh <= 0)
|
||||
return;
|
||||
|
||||
sums = g_new (Sums, width * height * components);
|
||||
|
||||
d = data;
|
||||
s = sums;
|
||||
|
||||
for (y = 0; y < height; y++)
|
||||
{
|
||||
if (i < half_point)
|
||||
arr [i] = half_point - i;
|
||||
else
|
||||
arr [i] = i - half_point;
|
||||
struct
|
||||
{
|
||||
gint sum;
|
||||
gint weighted_sum;
|
||||
gint leading_sum;
|
||||
gint leading_weighted_sum;
|
||||
} acc[components];
|
||||
|
||||
memset (acc, 0, sizeof (acc));
|
||||
|
||||
for (x = 0; x <= rw; x++)
|
||||
{
|
||||
for (c = 0; c < components; c++)
|
||||
{
|
||||
acc[c].sum += d[components * x + c];
|
||||
acc[c].weighted_sum += -x * d[components * x + c];
|
||||
}
|
||||
}
|
||||
|
||||
for (x = 0; x < width; x++)
|
||||
{
|
||||
for (c = 0; c < components; c++)
|
||||
{
|
||||
if (x > 0)
|
||||
{
|
||||
acc[c].weighted_sum += acc[c].sum;
|
||||
acc[c].leading_weighted_sum += acc[c].leading_sum;
|
||||
|
||||
if (x < width - r)
|
||||
{
|
||||
acc[c].sum += d[components * r];
|
||||
acc[c].weighted_sum += -r * d[components * r];
|
||||
}
|
||||
}
|
||||
|
||||
acc[c].leading_sum += d[0];
|
||||
|
||||
s->sum = acc[c].sum;
|
||||
s->weighted_sum = acc[c].weighted_sum;
|
||||
s->middle_sum = 2 * acc[c].leading_weighted_sum -
|
||||
acc[c].weighted_sum;
|
||||
|
||||
if (x >= r)
|
||||
{
|
||||
acc[c].sum -= d[components * -r];
|
||||
acc[c].weighted_sum -= r * d[components * -r];
|
||||
acc[c].leading_sum -= d[components * -r];
|
||||
acc[c].leading_weighted_sum -= r * d[components * -r];
|
||||
}
|
||||
|
||||
d++;
|
||||
s++;
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
for (x = 0; x < width; x++)
|
||||
{
|
||||
struct
|
||||
{
|
||||
gint weighted_sum;
|
||||
gint leading_sum;
|
||||
gint trailing_sum;
|
||||
} acc[components];
|
||||
|
||||
memset (acc, 0, sizeof (acc));
|
||||
|
||||
d = data + components * x;
|
||||
s = sums + components * x;
|
||||
|
||||
for (y = 1; y <= rh; y++)
|
||||
{
|
||||
for (c = 0; c < components; c++)
|
||||
{
|
||||
acc[c].weighted_sum += n * y * s[stride * y + c].sum -
|
||||
s[stride * y + c].weighted_sum;
|
||||
acc[c].trailing_sum += s[stride * y + c].sum;
|
||||
}
|
||||
}
|
||||
|
||||
for (y = 0; y < height; y++)
|
||||
{
|
||||
for (c = 0; c < components; c++)
|
||||
{
|
||||
if (y > 0)
|
||||
{
|
||||
acc[c].weighted_sum += s->weighted_sum +
|
||||
n * (acc[c].leading_sum -
|
||||
acc[c].trailing_sum);
|
||||
acc[c].trailing_sum -= s->sum;
|
||||
|
||||
if (y < height - r)
|
||||
{
|
||||
acc[c].weighted_sum += n * r * s[stride * r].sum -
|
||||
s[stride * r].weighted_sum;
|
||||
acc[c].trailing_sum += s[stride * r].sum;
|
||||
}
|
||||
}
|
||||
|
||||
acc[c].leading_sum += s->sum;
|
||||
|
||||
*d = (acc[c].weighted_sum + s->middle_sum + weight / 2) / weight;
|
||||
|
||||
acc[c].weighted_sum += s->weighted_sum;
|
||||
|
||||
if (y >= r)
|
||||
{
|
||||
acc[c].weighted_sum -= n * r * s[stride * -r].sum +
|
||||
s[stride * -r].weighted_sum;
|
||||
acc[c].leading_sum -= s[stride * -r].sum;
|
||||
}
|
||||
|
||||
d++;
|
||||
s++;
|
||||
}
|
||||
|
||||
d += stride - components;
|
||||
s += stride - components;
|
||||
}
|
||||
}
|
||||
|
||||
g_free (sums);
|
||||
}
|
||||
|
||||
static gint
|
||||
|
|
|
|||
Loading…
Reference in New Issue