[FFmpeg-devel] avcodec/pngdec: Optimize has_trns code

Submitted by Michael Niedermayer on Aug. 18, 2019, 8:29 a.m.

Details

Message ID 20190818082902.2425-1-michael@niedermayer.cc
State Accepted
Commit 361b3c873ee011598e16397db699bbfe5c045d7c
Headers show

Commit Message

Michael Niedermayer Aug. 18, 2019, 8:29 a.m.
add inner loop specialisations for 2 bpp and 4 bpp
These are all cases for which i found testsamples.

30M cycles -> 5M cycles

Testcase: fate-rgbapng-4816
Testcase: 16097/clusterfuzz-testcase-minimized-ffmpeg_AV_CODEC_ID_APNG_fuzzer-5664690889293824

Found-by: continuous fuzzing process https://github.com/google/oss-fuzz/tree/master/projects/ffmpeg
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
---
 libavcodec/pngdec.c | 37 +++++++++++++++++++++++++++++--------
 1 file changed, 29 insertions(+), 8 deletions(-)

Comments

Michael Niedermayer Sept. 2, 2019, 4:45 p.m.
On Sun, Aug 18, 2019 at 10:29:02AM +0200, Michael Niedermayer wrote:
> add inner loop specialisations for 2 bpp and 4 bpp
> These are all cases for which i found testsamples.
> 
> 30M cycles -> 5M cycles
> 
> Testcase: fate-rgbapng-4816
> Testcase: 16097/clusterfuzz-testcase-minimized-ffmpeg_AV_CODEC_ID_APNG_fuzzer-5664690889293824
> 
> Found-by: continuous fuzzing process https://github.com/google/oss-fuzz/tree/master/projects/ffmpeg
> Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
> ---
>  libavcodec/pngdec.c | 37 +++++++++++++++++++++++++++++--------
>  1 file changed, 29 insertions(+), 8 deletions(-)

will apply

[...]

Patch hide | download patch | download mbox

diff --git a/libavcodec/pngdec.c b/libavcodec/pngdec.c
index cad5796545..2d6c1b218e 100644
--- a/libavcodec/pngdec.c
+++ b/libavcodec/pngdec.c
@@ -24,6 +24,7 @@ 
 #include "libavutil/avassert.h"
 #include "libavutil/bprint.h"
 #include "libavutil/imgutils.h"
+#include "libavutil/intreadwrite.h"
 #include "libavutil/stereo3d.h"
 #include "libavutil/mastering_display_metadata.h"
 
@@ -1367,15 +1368,35 @@  exit_loop:
         for (y = 0; y < s->height; ++y) {
             uint8_t *row = &s->image_buf[s->image_linesize * y];
 
-            /* since we're updating in-place, we have to go from right to left */
-            for (x = s->width; x > 0; --x) {
-                uint8_t *pixel = &row[s->bpp * (x - 1)];
-                memmove(pixel, &row[raw_bpp * (x - 1)], raw_bpp);
+            if (s->bpp == 2 && byte_depth == 1) {
+                uint8_t *pixel = &row[2 * s->width - 1];
+                uint8_t *rowp  = &row[1 * s->width - 1];
+                int tcolor = s->transparent_color_be[0];
+                for (x = s->width; x > 0; --x) {
+                    *pixel-- = *rowp == tcolor ? 0 : 0xff;
+                    *pixel-- = *rowp--;
+                }
+            } else if (s->bpp == 4 && byte_depth == 1) {
+                uint8_t *pixel = &row[4 * s->width - 1];
+                uint8_t *rowp  = &row[3 * s->width - 1];
+                int tcolor = AV_RL24(s->transparent_color_be);
+                for (x = s->width; x > 0; --x) {
+                    *pixel-- = AV_RL24(rowp-2) == tcolor ? 0 : 0xff;
+                    *pixel-- = *rowp--;
+                    *pixel-- = *rowp--;
+                    *pixel-- = *rowp--;
+                }
+            } else {
+                /* since we're updating in-place, we have to go from right to left */
+                for (x = s->width; x > 0; --x) {
+                    uint8_t *pixel = &row[s->bpp * (x - 1)];
+                    memmove(pixel, &row[raw_bpp * (x - 1)], raw_bpp);
 
-                if (!memcmp(pixel, s->transparent_color_be, raw_bpp)) {
-                    memset(&pixel[raw_bpp], 0, byte_depth);
-                } else {
-                    memset(&pixel[raw_bpp], 0xff, byte_depth);
+                    if (!memcmp(pixel, s->transparent_color_be, raw_bpp)) {
+                        memset(&pixel[raw_bpp], 0, byte_depth);
+                    } else {
+                        memset(&pixel[raw_bpp], 0xff, byte_depth);
+                    }
                 }
             }
         }