[FFmpeg-cvslog] lavfi/gblur: doing several columns at the same time

Ruiling Song git at videolan.org
Wed May 8 04:43:04 EEST 2019


ffmpeg | branch: master | Ruiling Song <ruiling.song at intel.com> | Mon May  6 11:48:31 2019 +0800| [06ba4783a0ba331153b71993bc1908958d483c2d] | committer: Ruiling Song

lavfi/gblur: doing several columns at the same time

Instead of doing each column one by one, doing several columns
together gives about 30% better performance.

Reviewed-by: Paul B Mahol <onemda at gmail.com>
Signed-off-by: Ruiling Song <ruiling.song at intel.com>

> http://git.videolan.org/gitweb.cgi/ffmpeg.git/?a=commit;h=06ba4783a0ba331153b71993bc1908958d483c2d
---

 libavfilter/vf_gblur.c | 62 ++++++++++++++++++++++++++++++++++----------------
 1 file changed, 42 insertions(+), 20 deletions(-)

diff --git a/libavfilter/vf_gblur.c b/libavfilter/vf_gblur.c
index f77a3fffc3..b91a8c074a 100644
--- a/libavfilter/vf_gblur.c
+++ b/libavfilter/vf_gblur.c
@@ -108,6 +108,40 @@ static int filter_horizontally(AVFilterContext *ctx, void *arg, int jobnr, int n
     return 0;
 }
 
+static void do_vertical_columns(float *buffer, int width, int height,
+                                int column_begin, int column_end, int steps,
+                                float nu, float boundaryscale, int column_step)
+{
+    const int numpixels = width * height;
+    int i, x, k, step;
+    float *ptr;
+    for (x = column_begin; x < column_end;) {
+        for (step = 0; step < steps; step++) {
+            ptr = buffer + x;
+            for (k = 0; k < column_step; k++) {
+                ptr[k] *= boundaryscale;
+            }
+            /* Filter downwards */
+            for (i = width; i < numpixels; i += width) {
+                for (k = 0; k < column_step; k++) {
+                    ptr[i + k] += nu * ptr[i - width + k];
+                }
+            }
+            i = numpixels - width;
+
+            for (k = 0; k < column_step; k++)
+                ptr[i + k] *= boundaryscale;
+
+            /* Filter upwards */
+            for (; i > 0; i -= width) {
+                for (k = 0; k < column_step; k++)
+                    ptr[i - width + k] += nu * ptr[i + k];
+            }
+        }
+        x += column_step;
+    }
+}
+
 static int filter_vertically(AVFilterContext *ctx, void *arg, int jobnr, int nb_jobs)
 {
     GBlurContext *s = ctx->priv;
@@ -117,31 +151,19 @@ static int filter_vertically(AVFilterContext *ctx, void *arg, int jobnr, int nb_
     const int slice_start = (width *  jobnr   ) / nb_jobs;
     const int slice_end   = (width * (jobnr+1)) / nb_jobs;
     const float boundaryscale = s->boundaryscaleV;
-    const int numpixels = width * height;
     const int steps = s->steps;
     const float nu = s->nuV;
     float *buffer = s->buffer;
-    int i, x, step;
-    float *ptr;
-
-    /* Filter vertically along each column */
-    for (x = slice_start; x < slice_end; x++) {
-        for (step = 0; step < steps; step++) {
-            ptr = buffer + x;
-            ptr[0] *= boundaryscale;
-
-            /* Filter downwards */
-            for (i = width; i < numpixels; i += width)
-                ptr[i] += nu * ptr[i - width];
-
-            ptr[i = numpixels - width] *= boundaryscale;
+    int aligned_end;
 
-            /* Filter upwards */
-            for (; i > 0; i -= width)
-                ptr[i - width] += nu * ptr[i];
-        }
-    }
+    aligned_end = slice_start + (((slice_end - slice_start) >> 3) << 3);
+    /* Filter vertically along columns (process 8 columns in each step) */
+    do_vertical_columns(buffer, width, height, slice_start, aligned_end,
+                        steps, nu, boundaryscale, 8);
 
+    /* Filter un-aligned columns one by one */
+    do_vertical_columns(buffer, width, height, aligned_end, slice_end,
+                        steps, nu, boundaryscale, 1);
     return 0;
 }
 



More information about the ffmpeg-cvslog mailing list