[gtk+/gtk-3-16] gtkcairoblur: Unroll inner loop for common radius values
- From: Matthias Clasen <matthiasc src gnome org>
- To: commits-list gnome org
- Cc:
- Subject: [gtk+/gtk-3-16] gtkcairoblur: Unroll inner loop for common radius values
- Date: Mon, 23 Mar 2015 03:13:46 +0000 (UTC)
commit c96df5a3507a24cb270f26b4e48bd9838211ed1d
Author: Alexander Larsson <alexl redhat com>
Date: Thu Mar 19 16:29:32 2015 +0100
gtkcairoblur: Unroll inner loop for common radius values
This unrolls the inner blur loop for radius 1-10, allowing
the compiler to use a divide-by-constant operation instead
of a generic division.
Here is the blur-performance output before:
Radius 1: 124.95 msec, 32.01 kpixels/msec:
Radius 2: 117.27 msec, 34.11 kpixels/msec:
Radius 3: 123.57 msec, 32.37 kpixels/msec:
Radius 4: 118.17 msec, 33.85 kpixels/msec:
Radius 5: 119.32 msec, 33.52 kpixels/msec:
Radius 6: 124.17 msec, 32.21 kpixels/msec:
Radius 7: 121.04 msec, 33.05 kpixels/msec:
Radius 8: 130.64 msec, 30.62 kpixels/msec:
Radius 9: 119.47 msec, 33.48 kpixels/msec:
Radius 10: 117.95 msec, 33.91 kpixels/msec:
Radius 11: 122.38 msec, 32.68 kpixels/msec:
Radius 12: 121.92 msec, 32.81 kpixels/msec:
Radius 13: 125.45 msec, 31.89 kpixels/msec:
Radius 14: 121.63 msec, 32.89 kpixels/msec:
Radius 15: 120.18 msec, 33.28 kpixels/msec:
And after:
Radius 1: 42.26 msec, 94.65 kpixels/msec:
Radius 2: 59.15 msec, 67.62 kpixels/msec:
Radius 3: 60.29 msec, 66.35 kpixels/msec:
Radius 4: 64.53 msec, 61.99 kpixels/msec:
Radius 5: 60.07 msec, 66.59 kpixels/msec:
Radius 6: 62.43 msec, 64.07 kpixels/msec:
Radius 7: 60.36 msec, 66.27 kpixels/msec:
Radius 8: 59.59 msec, 67.13 kpixels/msec:
Radius 9: 76.17 msec, 52.51 kpixels/msec:
Radius 10: 79.41 msec, 50.37 kpixels/msec:
Radius 11: 118.92 msec, 33.64 kpixels/msec:
Radius 12: 121.31 msec, 32.97 kpixels/msec:
Radius 13: 118.30 msec, 33.81 kpixels/msec:
Radius 14: 116.82 msec, 34.24 kpixels/msec:
Radius 15: 116.99 msec, 34.19 kpixels/msec:
I.e. almost double performance for the unrolled radius values.
https://bugzilla.gnome.org/show_bug.cgi?id=746468
gtk/gtkcairoblur.c | 39 +++++++++++++++++++++++++++++----------
1 files changed, 29 insertions(+), 10 deletions(-)
---
diff --git a/gtk/gtkcairoblur.c b/gtk/gtkcairoblur.c
index 82f5d33..170f460 100644
--- a/gtk/gtkcairoblur.c
+++ b/gtk/gtkcairoblur.c
@@ -73,18 +73,37 @@ blur_xspan (guchar *row,
* only divide down after all three passes. (SSE parallel implementation
* of the divide step is possible.)
*/
- for (i = -d + offset; i < row_width + offset; i++)
- {
- if (i >= 0 && i < row_width)
- sum += row[i];
-
- if (i >= offset)
- {
- if (i >= d)
- sum -= row[i - d];
- tmp_buffer[i - offset] = (sum + d / 2) / d;
- }
+#define BLUR_ROW_KERNEL(D) \
+ for (i = -(D) + offset; i < row_width + offset; i++) \
+ { \
+ if (i >= 0 && i < row_width) \
+ sum += row[i]; \
+ \
+ if (i >= offset) \
+ { \
+ if (i >= (D)) \
+ sum -= row[i - (D)]; \
+ \
+ tmp_buffer[i - offset] = (sum + (D) / 2) / (D); \
+ } \
+ } \
+ break;
+
+ /* We unroll the values for d for radius 2-10 to avoid a generic
+ * divide operation (not radius 1, because its a no-op) */
+ switch (d)
+ {
+ case get_box_filter_size (2): BLUR_ROW_KERNEL (get_box_filter_size (2));
+ case get_box_filter_size (3): BLUR_ROW_KERNEL (get_box_filter_size (3));
+ case get_box_filter_size (4): BLUR_ROW_KERNEL (get_box_filter_size (4));
+ case get_box_filter_size (5): BLUR_ROW_KERNEL (get_box_filter_size (5));
+ case get_box_filter_size (6): BLUR_ROW_KERNEL (get_box_filter_size (6));
+ case get_box_filter_size (7): BLUR_ROW_KERNEL (get_box_filter_size (7));
+ case get_box_filter_size (8): BLUR_ROW_KERNEL (get_box_filter_size (8));
+ case get_box_filter_size (9): BLUR_ROW_KERNEL (get_box_filter_size (9));
+ case get_box_filter_size (10): BLUR_ROW_KERNEL (get_box_filter_size (10));
+ default: BLUR_ROW_KERNEL (d);
}
memcpy (row, tmp_buffer, row_width);
[
Date Prev][
Date Next] [
Thread Prev][
Thread Next]
[
Thread Index]
[
Date Index]
[
Author Index]