[FFmpeg-devel] [PATCH] avfilter/lut: reduce dereference in the inner loop

[Yayoi]

For rgb, with a 1080p source, 69 to 74fps on core i5(2 core, 1.8GHz),
and 136 to 160 fps on an core i7(4770R, 3.2Ghz)
Changed the yuv code for consistency, even though the performance
increase is not as obvious as rgb
---
 libavfilter/vf_lut.c | 29 ++++++++++++++++++-----------
 1 file changed, 18 insertions(+), 11 deletions(-)

diff --git a/libavfilter/vf_lut.c b/libavfilter/vf_lut.c
index 0b7a2ca..e262c6e 100644
--- a/libavfilter/vf_lut.c
+++ b/libavfilter/vf_lut.c
@@ -299,26 +299,31 @@ static int filter_frame(AVFilterLink *inlink, AVFrame *in)
 
     if (s->is_rgb) {
         /* packed */
+        const int w = inlink->w;
+        const int h = in->height;
+        const uint8_t (*tab)[256] = (const uint8_t (*)[256])s->lut;
+        const int in_linesize  =  in->linesize[0];
+        const int out_linesize = out->linesize[0];
+        const int step = s->step;
+
         inrow0  = in ->data[0];
         outrow0 = out->data[0];
 
-        for (i = 0; i < in->height; i ++) {
-            int w = inlink->w;
-            const uint8_t (*tab)[256] = (const uint8_t (*)[256])s->lut;
+        for (i = 0; i < h; i ++) {
             inrow  = inrow0;
             outrow = outrow0;
             for (j = 0; j < w; j++) {
-                switch (s->step) {
+                switch (step) {
                 case 4:  outrow[3] = tab[3][inrow[3]]; // Fall-through
                 case 3:  outrow[2] = tab[2][inrow[2]]; // Fall-through
                 case 2:  outrow[1] = tab[1][inrow[1]]; // Fall-through
                 default: outrow[0] = tab[0][inrow[0]];
                 }
-                outrow += s->step;
-                inrow  += s->step;
+                outrow += step;
+                inrow  += step;
             }
-            inrow0  += in ->linesize[0];
-            outrow0 += out->linesize[0];
+            inrow0  += in_linesize;
+            outrow0 += out_linesize;
         }
     } else {
         /* planar */
@@ -327,16 +332,18 @@ static int filter_frame(AVFilterLink *inlink, AVFrame *in)
             int hsub = plane == 1 || plane == 2 ? s->hsub : 0;
             int h = FF_CEIL_RSHIFT(inlink->h, vsub);
             int w = FF_CEIL_RSHIFT(inlink->w, hsub);
+            const uint8_t *tab = s->lut[plane];
+            const int in_linesize  =  in->linesize[plane];
+            const int out_linesize = out->linesize[plane];
 
             inrow  = in ->data[plane];
             outrow = out->data[plane];
 
             for (i = 0; i < h; i++) {
-                const uint8_t *tab = s->lut[plane];
                 for (j = 0; j < w; j++)
                     outrow[j] = tab[inrow[j]];
-                inrow  += in ->linesize[plane];
-                outrow += out->linesize[plane];
+                inrow  += in_linesize;
+                outrow += out_linesize;
             }
         }
     }
-- 
1.8.3.4 (Apple Git-47)