[FFmpeg-cvslog] avcodec/dfa: Optimize output reshuffle loop

[Michael Niedermayer]

ffmpeg | branch: master | Michael Niedermayer <michael at niedermayer.cc> | Fri Mar 11 20:57:04 2022 +0100| [18bc612f2fd33b6ac943bf1a0fdaa55b1f4c9d25] | committer: Michael Niedermayer

avcodec/dfa: Optimize output reshuffle loop

18035 -> 4018 dezicycles (Tested with LOGOS.DFA, gcc 7, 3950X)

Signed-off-by: Michael Niedermayer <michael at niedermayer.cc>

> http://git.videolan.org/gitweb.cgi/ffmpeg.git/?a=commit;h=18bc612f2fd33b6ac943bf1a0fdaa55b1f4c9d25
---

 libavcodec/dfa.c | 14 +++++++++++---
 1 file changed, 11 insertions(+), 3 deletions(-)

diff --git a/libavcodec/dfa.c b/libavcodec/dfa.c
index 0cf3641a38..ab78d66763 100644
--- a/libavcodec/dfa.c
+++ b/libavcodec/dfa.c
@@ -388,9 +388,17 @@ static int dfa_decode_frame(AVCodecContext *avctx,
     for (i = 0; i < avctx->height; i++) {
         if(version == 0x100) {
             int j;
-            for(j = 0; j < avctx->width; j++) {
-                dst[j] = buf[ (i&3)*(avctx->width /4) + (j/4) +
-                             ((j&3)*(avctx->height/4) + (i/4))*avctx->width];
+            const uint8_t *buf1 = buf + (i&3)*(avctx->width/4) + (i/4)*avctx->width;
+            int stride = (avctx->height/4)*avctx->width;
+            for(j = 0; j < avctx->width/4; j++) {
+                dst[4*j+0] = buf1[j + 0*stride];
+                dst[4*j+1] = buf1[j + 1*stride];
+                dst[4*j+2] = buf1[j + 2*stride];
+                dst[4*j+3] = buf1[j + 3*stride];
+            }
+            j *= 4;
+            for(; j < avctx->width; j++) {
+                dst[j] = buf1[(j/4) + (j&3)*stride];
             }
         } else {
             memcpy(dst, buf, avctx->width);