[gegl] sampler.h: enlarge samplers' buffer tiles to 256x256. may need to be tweaked further
- From: Nicolas Robidoux <nrobidoux src gnome org>
- To: commits-list gnome org
- Cc:
- Subject: [gegl] sampler.h: enlarge samplers' buffer tiles to 256x256. may need to be tweaked further
- Date: Wed, 19 Dec 2012 03:38:04 +0000 (UTC)
commit 9db8bfed10d5821d3bc617db1ae256c8735ace3a
Author: Nicolas Robidoux <nrobidoux git gnome org>
Date: Tue Dec 18 22:37:58 2012 -0500
sampler.h: enlarge samplers' buffer tiles to 256x256. may need to be tweaked further
gegl/buffer/gegl-sampler-lohalo.c | 166 ++++++++++++++++++---------------
gegl/buffer/gegl-sampler.c | 16 ++-
gegl/buffer/gegl-sampler.h | 2 +-
operations/transform/transform-core.c | 139 ++++------------------------
4 files changed, 122 insertions(+), 201 deletions(-)
---
diff --git a/gegl/buffer/gegl-sampler-lohalo.c b/gegl/buffer/gegl-sampler-lohalo.c
index f0dadc4..989bde3 100644
--- a/gegl/buffer/gegl-sampler-lohalo.c
+++ b/gegl/buffer/gegl-sampler-lohalo.c
@@ -411,7 +411,7 @@ gegl_sampler_lohalo_class_init (GeglSamplerLohaloClass *klass)
/*
* IMPORTANT: LOHALO_OFFSET_0 SHOULD BE AN INTEGER >= 2.
*/
-#define LOHALO_OFFSET_0 (14)
+#define LOHALO_OFFSET_0 (16)
#define LOHALO_SIZE_0 ( 1 + 2 * LOHALO_OFFSET_0 )
/*
@@ -424,7 +424,7 @@ gegl_sampler_lohalo_class_init (GeglSamplerLohaloClass *klass)
* mipmap level's offset should almost never be smaller than half the
* previous level's offset.
*/
-#define LOHALO_OFFSET_MIPMAP (14)
+#define LOHALO_OFFSET_MIPMAP (16)
#define LOHALO_SIZE_MIPMAP ( 1 + 2 * LOHALO_OFFSET_MIPMAP )
#define LOHALO_OFFSET_1 LOHALO_OFFSET_MIPMAP
@@ -2250,36 +2250,95 @@ gegl_sampler_lohalo_get ( GeglSampler* restrict self,
const gfloat c_minor_y = minor_unit_y / minor_mag;
/*
- * Accumulator for the EWA weights:
+ * Remainder of the ellipse geometry computation:
*/
- gdouble total_weight = (gdouble) 0.0;
/*
- * Storage for the EWA contribution:
+ * Major and minor axis direction vectors:
*/
- gfloat ewa_newval[channels];
- ewa_newval[0] = (gfloat) 0.0;
- ewa_newval[1] = (gfloat) 0.0;
- ewa_newval[2] = (gfloat) 0.0;
- ewa_newval[3] = (gfloat) 0.0;
+ const gdouble major_x = major_mag * major_unit_x;
+ const gdouble major_y = major_mag * major_unit_y;
+ const gdouble minor_x = minor_mag * minor_unit_x;
+ const gdouble minor_y = minor_mag * minor_unit_y;
+
+ /*
+ * Ellipse coefficients:
+ */
+ const gdouble ellipse_a =
+ major_y * major_y + minor_y * minor_y;
+ const gdouble folded_ellipse_b =
+ major_x * major_y + minor_x * minor_y;
+ const gdouble ellipse_c =
+ major_x * major_x + minor_x * minor_x;
+ const gdouble ellipse_f = major_mag * minor_mag;
+
+ /*
+ * Bounding box of the ellipse:
+ */
+ const gdouble bounding_box_factor =
+ ellipse_f * ellipse_f /
+ ( ellipse_c * ellipse_a - folded_ellipse_b * folded_ellipse_b );
+ const gfloat bounding_box_half_width =
+ sqrtf( (gfloat) (ellipse_c * bounding_box_factor) );
+ const gfloat bounding_box_half_height =
+ sqrtf( (gfloat) (ellipse_a * bounding_box_factor) );
/*
- * Grab the pixel values located within the context_rect of
- * "pure" LBB-Nohalo. Farther ones will be accessed through
+ * Relative weight of the contribution of LBB-Nohalo:
+ */
+ const gfloat theta = (gfloat) ( (gdouble) 1. / ellipse_f );
+
+ /*
+ * Grab the pixel values located within the level 0
+ * context_rect. Farther ones will be accessed through
* higher mipmap levels.
*/
+ const gint out_left_0 =
+ LOHALO_MAX
+ (
+ (gint) ceil ( (double) ( x_0 - bounding_box_half_width ) )
+ ,
+ -LOHALO_OFFSET_0
+ );
+ const gint out_rite_0 =
+ LOHALO_MIN
+ (
+ (gint) floorf ( (double) ( x_0 + bounding_box_half_width ) )
+ ,
+ LOHALO_OFFSET_0
+ );
+ const gint out_top_0 =
+ LOHALO_MAX
+ (
+ (gint) ceilf ( (double) ( y_0 - bounding_box_half_height ) )
+ ,
+ -LOHALO_OFFSET_0
+ );
+ const gint out_bot_0 =
+ LOHALO_MIN
+ (
+ (gint) floorf ( (double) ( y_0 + bounding_box_half_height ) )
+ ,
+ LOHALO_OFFSET_0
+ );
+
+ /*
+ * Accumulator for the EWA weights:
+ */
+ gdouble total_weight = (gdouble) 0.0;
/*
- * TODO: Given that we now use a somewhat large level 0
- * context_rect, it makes sense to restrict the EWA
- * computation to the bounding box even at level 0. With
- * small context_rect, this is a waste. But right now, when
- * barely downsmapling, there are lots of multiplications by
- * zero weights.
+ * Storage for the EWA contribution:
*/
+ gfloat ewa_newval[channels];
+ ewa_newval[0] = (gfloat) 0;
+ ewa_newval[1] = (gfloat) 0;
+ ewa_newval[2] = (gfloat) 0;
+ ewa_newval[3] = (gfloat) 0;
+
{
- gint i = -LOHALO_OFFSET_0;
+ gint i = out_top_0;
do
{
- gint j = -LOHALO_OFFSET_0;
+ gint j = out_left_0;
do
{
ewa_update (j,
@@ -2295,8 +2354,8 @@ gegl_sampler_lohalo_get ( GeglSampler* restrict self,
input_bptr,
&total_weight,
ewa_newval);
- } while ( ++j <= LOHALO_OFFSET_0 );
- } while ( ++i <= LOHALO_OFFSET_0 );
+ } while ( ++j <= out_rite_0 );
+ } while ( ++i <= out_bot_0 );
}
{
@@ -2307,7 +2366,7 @@ gegl_sampler_lohalo_get ( GeglSampler* restrict self,
* determine the alignment of the level 1 mipmap level
* w.r.t. the current level 0.
*
- * At level 0, we can access pixels which are
+ * At level 0, we can access pixels which are at most
* LOHALO_OFFSET_0 away from the anchor pixel location in
* box distance.
*/
@@ -2318,47 +2377,11 @@ gegl_sampler_lohalo_get ( GeglSampler* restrict self,
const gint odd_ix_0 = ix_0 % 2;
const gint odd_iy_0 = iy_0 % 2;
/*
- * Find the closest locations, on all four sides, of level
- * 1 pixels which involve data not found in the level 0
+ * Find the closest locations, on all four sides, of level 1
+ * pixels which involve data not found in the level 0
* LOHALO_SIZE_0xLOHALO_SIZE_0.
*/
LOHALO_FIND_CLOSEST_LOCATIONS(0,1)
-
- /*
- * Remainder of the ellipse geometry computation:
- */
- /*
- * Major and minor axis direction vectors:
- */
- const gdouble major_x = major_mag * major_unit_x;
- const gdouble major_y = major_mag * major_unit_y;
- const gdouble minor_x = minor_mag * minor_unit_x;
- const gdouble minor_y = minor_mag * minor_unit_y;
-
- /*
- * Ellipse coefficients:
- */
- const gdouble ellipse_a =
- major_y * major_y + minor_y * minor_y;
- const gdouble folded_ellipse_b =
- major_x * major_y + minor_x * minor_y;
- const gdouble ellipse_c =
- major_x * major_x + minor_x * minor_x;
- const gdouble ellipse_f = major_mag * minor_mag;
-
- /*
- * Bounding box of the ellipse:
- */
- const gdouble bounding_box_factor =
- ellipse_f * ellipse_f
- /
- (
- ellipse_c * ellipse_a - folded_ellipse_b * folded_ellipse_b
- );
- const gfloat bounding_box_half_width =
- sqrtf( (gfloat) (ellipse_c * bounding_box_factor) );
- const gfloat bounding_box_half_height =
- sqrtf( (gfloat) (ellipse_a * bounding_box_factor) );
/*
* Bounding box shrunk a smidgen given that a location
* very close to the edge of the bounding box will get a
@@ -2369,11 +2392,6 @@ gegl_sampler_lohalo_get ( GeglSampler* restrict self,
const gfloat fudged_bounding_box_half_height =
bounding_box_half_height - LOHALO_FUDGEF;
- /*
- * Relative weight of the contribution of LBB-Nohalo:
- */
- const gfloat theta = (gfloat) ( (gdouble) 1. / ellipse_f );
-
if (( x_0 - fudged_bounding_box_half_width < closest_left_1 )
||
( x_0 + fudged_bounding_box_half_width > closest_rite_1 )
@@ -2429,7 +2447,7 @@ gegl_sampler_lohalo_get ( GeglSampler* restrict self,
* The "in" indices are the closest relative mipmap 1
* indices of needed mipmap values:
*/
- LOHALO_FIND_CLOSEST_INDICES(0,1)
+ LOHALO_FIND_CLOSEST_INDICES(0,1)
/*
* The "out" indices are the farthest relative mipmap
* 1 indices we use at this level:
@@ -2440,7 +2458,7 @@ gegl_sampler_lohalo_get ( GeglSampler* restrict self,
* Update using mipmap level 1 values.
*/
LOHALO_MIPMAP_EWA_UPDATE(1)
- {
+ {
/*
* In order to know whether we use even higher
* mipmap level values, we need to check whether
@@ -2454,7 +2472,7 @@ gegl_sampler_lohalo_get ( GeglSampler* restrict self,
* access pixels which are LOHALO_OFFSET_MIPMAP away
* from the level 1 anchor pixel location in box
* distance.
- */
+ */
/*
* Determine whether the anchor level_1 pixel
* locations are odd (VS even):
@@ -2488,7 +2506,7 @@ gegl_sampler_lohalo_get ( GeglSampler* restrict self,
* pixel values will get 0 coefficients--but we
* used a quick and dirty bounding box test
* which lets through false positives.)
- */
+ */
/*
* Nearest mipmap anchor pixel location:
@@ -2514,12 +2532,12 @@ gegl_sampler_lohalo_get ( GeglSampler* restrict self,
* the center of a level 1 pixel is at a box
* distance of 1 from the center of the closest
* level 2 pixel.
- */
+ */
const gfloat x_2 =
x_1 + (gfloat) ( 2 * ( ix_1 - 2 * ix_2 ) - 1 );
const gfloat y_2 =
y_1 + (gfloat) ( 2 * ( iy_1 - 2 * iy_2 ) - 1 );
-
+
/*
* Key index ranges:
*/
@@ -2527,7 +2545,7 @@ gegl_sampler_lohalo_get ( GeglSampler* restrict self,
* The "in" indices are the closest relative
* mipmap 2 indices of needed mipmap values:
*/
- LOHALO_FIND_CLOSEST_INDICES(1,2)
+ LOHALO_FIND_CLOSEST_INDICES(1,2)
/*
* The "out" indices are the farthest relative
* mipmap 1 indices we use at this level:
@@ -2572,7 +2590,7 @@ gegl_sampler_lohalo_get ( GeglSampler* restrict self,
x_2 + (gfloat) ( 2 * ( ix_2 - 2 * ix_3 ) - 1 );
const gfloat y_3 =
y_2 + (gfloat) ( 2 * ( iy_2 - 2 * iy_3 ) - 1 );
- LOHALO_FIND_CLOSEST_INDICES(2,3)
+ LOHALO_FIND_CLOSEST_INDICES(2,3)
LOHALO_FIND_FARTHEST_INDICES(3)
LOHALO_MIPMAP_EWA_UPDATE(3)
}
diff --git a/gegl/buffer/gegl-sampler.c b/gegl/buffer/gegl-sampler.c
index 5f80788..f2bc629 100644
--- a/gegl/buffer/gegl-sampler.c
+++ b/gegl/buffer/gegl-sampler.c
@@ -267,10 +267,12 @@ gegl_sampler_get_ptr (GeglSampler *const sampler,
*/
fetch_rectangle.x =
x + sampler->context_rect[0].x -
- (maximum_width_and_height - sampler->context_rect[0].width ) / (gint) 8;
+ (maximum_width_and_height - sampler->context_rect[0].width ) /
+ (gint) 8;
fetch_rectangle.y =
y + sampler->context_rect[0].y -
- (maximum_width_and_height - sampler->context_rect[0].height) / (gint) 8;
+ (maximum_width_and_height - sampler->context_rect[0].height) /
+ (gint) 8;
fetch_rectangle.width = maximum_width_and_height;
fetch_rectangle.height = maximum_width_and_height;
@@ -347,9 +349,11 @@ gegl_sampler_get_from_buffer (GeglSampler *const sampler,
GeglRectangle fetch_rectangle;
fetch_rectangle.x = x -
- (maximum_width_and_height - sampler->context_rect[0].width ) / (gint) 8;
+ (maximum_width_and_height - sampler->context_rect[0].width ) /
+ (gint) 8;
fetch_rectangle.y = y -
- (maximum_width_and_height - sampler->context_rect[0].height) / (gint) 8;
+ (maximum_width_and_height - sampler->context_rect[0].height) /
+ (gint) 8;
fetch_rectangle.width = maximum_width_and_height;
fetch_rectangle.height = maximum_width_and_height;
@@ -438,11 +442,11 @@ gegl_sampler_get_from_mipmap (GeglSampler *const sampler,
fetch_rectangle.x =
x + sampler->context_rect[level].x -
(maximum_width_and_height - sampler->context_rect[level].width ) /
- (gint) 8;
+ (gint) 8;
fetch_rectangle.y =
y + sampler->context_rect[level].y -
(maximum_width_and_height - sampler->context_rect[level].height) /
- (gint) 8;
+ (gint) 8;
fetch_rectangle.width = maximum_width_and_height;
fetch_rectangle.height = maximum_width_and_height;
diff --git a/gegl/buffer/gegl-sampler.h b/gegl/buffer/gegl-sampler.h
index 7b01ef3..3d73ccd 100644
--- a/gegl/buffer/gegl-sampler.h
+++ b/gegl/buffer/gegl-sampler.h
@@ -41,7 +41,7 @@ G_BEGIN_DECLS
* The way the samplers use mipmap levels, square buffers are
* preferable to rectangular ones.
*/
-#define GEGL_SAMPLER_MAXIMUM_WIDTH_AND_HEIGHT 64
+#define GEGL_SAMPLER_MAXIMUM_WIDTH_AND_HEIGHT 256
typedef struct _GeglSamplerClass GeglSamplerClass;
diff --git a/operations/transform/transform-core.c b/operations/transform/transform-core.c
index 9965ea0..1723726 100644
--- a/operations/transform/transform-core.c
+++ b/operations/transform/transform-core.c
@@ -797,20 +797,14 @@ transform_affine (GeglBuffer *dest,
const GeglRectangle *dest_extent;
gint x,
y;
- gfloat * restrict dest_buf,
- *dest_ptr;
GeglMatrix3 inverse;
GeglMatrix2 inverse_jacobian;
- gdouble u_start,
- v_start,
- u_float,
- v_float,
- base_u,
+ gdouble base_u,
base_v;
- const Babl *format = babl_format ("RaGaBaA float");
gint dest_pixels,
flip_x = 0,
flip_y = 0;
+ const Babl *format = babl_format ("RaGaBaA float");
/*
* XXX: fast paths as existing in files in the same dir as
@@ -966,24 +960,26 @@ transform_affine (GeglBuffer *dest,
while (gegl_buffer_iterator_next (i))
{
GeglRectangle *roi = &i->roi[0];
+ gfloat * restrict dest_buf = (gfloat *)i->data[0];
- dest_buf = (gfloat *)i->data[0];
-
- u_start = base_u +
- inverse.coeff [0][0] * ( roi->x + flip_x * roi->width ) +
- inverse.coeff [0][1] * ( roi->y + flip_y * roi->height );
- v_start = base_v +
- inverse.coeff [1][0] * ( roi->x + flip_x * roi->width ) +
- inverse.coeff [1][1] * ( roi->y + flip_y * roi->height );
-
- dest_ptr = dest_buf +
- (gint) 4 * flip_x * (roi->width - (gint) 1) +
- (gint) 4 * flip_y * (roi->height - (gint) 1) * roi->width;
+ gfloat * restrict dest_ptr =
+ dest_buf +
+ (gint) 4 * ( flip_x * (roi->width - (gint) 1) +
+ flip_y * (roi->height - (gint) 1) * roi->width );
+
+ gdouble u_start =
+ base_u +
+ inverse.coeff [0][0] * ( roi->x + flip_x * roi->width ) +
+ inverse.coeff [0][1] * ( roi->y + flip_y * roi->height );
+ gdouble v_start =
+ base_v +
+ inverse.coeff [1][0] * ( roi->x + flip_x * roi->width ) +
+ inverse.coeff [1][1] * ( roi->y + flip_y * roi->height );
for (y = roi->height; y--;)
{
- u_float = u_start;
- v_float = v_start;
+ gdouble u_float = u_start;
+ gdouble v_float = v_start;
for (x = roi->width; x--;)
{
@@ -993,7 +989,6 @@ transform_affine (GeglBuffer *dest,
&inverse_jacobian,
dest_ptr,
GEGL_ABYSS_NONE);
-
dest_ptr += (gint) 4 - (gint) 8 * flip_x;
u_float += inverse_jacobian.coeff [0][0];
@@ -1008,102 +1003,6 @@ transform_affine (GeglBuffer *dest,
}
}
-#if 0
-static void
-transform_generic (GeglBuffer *dest,
- GeglBuffer *src,
- GeglMatrix3 *matrix,
- GeglSampler *sampler,
- gint level)
-{
- GeglBufferIterator *i;
- const GeglRectangle *dest_extent;
- gint x,
- y;
- gfloat * restrict dest_buf,
- *dest_ptr;
- GeglMatrix3 inverse;
- gdouble u_start,
- v_start,
- w_start,
- u_float,
- v_float,
- w_float;
- const Babl *format = babl_format ("RaGaBaA float");
- gint dest_pixels;
-
- g_object_get (dest, "pixels", &dest_pixels, NULL);
- dest_extent = gegl_buffer_get_extent (dest);
-
- i = gegl_buffer_iterator_new (dest,
- dest_extent,
- level,
- format,
- GEGL_BUFFER_WRITE,
- GEGL_ABYSS_NONE);
-
- while (gegl_buffer_iterator_next (i))
- {
- GeglRectangle *roi = &i->roi[0];
-
- dest_buf = (gfloat *)i->data[0];
-
- gegl_matrix3_copy_into (&inverse, matrix);
- gegl_matrix3_invert (&inverse);
-
- u_start = inverse.coeff [0][0] * (roi->x + (gdouble) 0.5) +
- inverse.coeff [0][1] * (roi->y + (gdouble) 0.5) +
- inverse.coeff [0][2];
- v_start = inverse.coeff [1][0] * (roi->x + (gdouble) 0.5) +
- inverse.coeff [1][1] * (roi->y + (gdouble) 0.5) +
- inverse.coeff [1][2];
- w_start = inverse.coeff [2][0] * (roi->x + (gdouble) 0.5) +
- inverse.coeff [2][1] * (roi->y + (gdouble) 0.5) +
- inverse.coeff [2][2];
-
- for (dest_ptr = dest_buf, y = roi->height; y--;)
- {
- u_float = u_start;
- v_float = v_start;
- w_float = w_start;
-
- for (x = roi->width; x--;)
- {
- GeglMatrix2 inverse_jacobian;
- gdouble w_recip = 1.0 / w_float;
- gdouble u = u_float * w_recip;
- gdouble v = v_float * w_recip;
-
- inverse_jacobian.coeff [0][0] =
- (inverse.coeff [0][0] - inverse.coeff [2][0] * u) * w_recip;
- inverse_jacobian.coeff [0][1] =
- (inverse.coeff [0][1] - inverse.coeff [2][1] * u) * w_recip;
- inverse_jacobian.coeff [1][0] =
- (inverse.coeff [1][0] - inverse.coeff [2][0] * v) * w_recip;
- inverse_jacobian.coeff [1][1] =
- (inverse.coeff [1][1] - inverse.coeff [2][1] * v) * w_recip;
-
- gegl_sampler_get (sampler,
- u,
- v,
- &inverse_jacobian,
- dest_ptr,
- GEGL_ABYSS_NONE);
- dest_ptr+=4;
-
- u_float += inverse.coeff [0][0];
- v_float += inverse.coeff [1][0];
- w_float += inverse.coeff [2][0];
- }
-
- u_start += inverse.coeff [0][1];
- v_start += inverse.coeff [1][1];
- w_start += inverse.coeff [2][1];
- }
- }
-}
-#endif
-
static void
transform_generic (GeglBuffer *dest,
GeglBuffer *src,
[
Date Prev][
Date Next] [
Thread Prev][
Thread Next]
[
Thread Index]
[
Date Index]
[
Author Index]
