Closed GoogleCodeExporter closed 9 years ago
Also for this formula, it would be use to use a constant for 32768.
y = (dy >> 1) - 32768;
Original comment by fbarch...@google.com
on 23 May 2013 at 1:33
r645 was ok
svn up -r 645 source/scale.cc
out\Debug\convert.exe faces_352x288_ARGB.raw faces_352x288_P420.yuv
out\Debug\convert.exe -f 0 faces_352x288_P420.yuv faces_699x531_P420.yuv
out\Debug\convert.exe -f 0 faces_699x531_P420.yuv face2_352x288_P420.yuv
out\Debug\psnr.exe -s 352 288 -v faces_352x288_P420.yuv face2_352x288_P420.yuv
Frame PSNR-Y PSNR-U PSNR-V PSNR-All Frame
0 41.197008 46.917811 49.997085 42.544299 min
Original comment by fbarch...@google.com
on 28 May 2013 at 11:05
This is showing the old code for point sampling was the same as ffmpeg for r645
upsampling, but different in r705.
out\Debug\convert.exe -f 0 faces_352x288_P420.yuv faces_640x480_P420.yuv
ffmpeg -y -f rawvideo -pix_fmt yuv420p -s 352x288 -i faces_352x288_P420.yuv
-sws_flags neighbor -vf scale=640:480 -f rawvideo facef_640x480_P420.yuv
out\Debug\psnr.exe faces_640x480_P420.yuv facef_640x480_P420.yuv
Was r645
Frame PSNR-Y PSNR-U PSNR-V PSNR-All Frame
Global: 128.000000 128.000000 128.000000 128.000000 1
Avg: 128.000000 128.000000 128.000000 128.000000 1
Min: 128.000000 128.000000 128.000000 128.000000 0
Now r705
d:\src\libyuv\trunk>out\Debug\psnr.exe faces_640x480_P420.yuv
facef_640x480_P420.yuv
Frame PSNR-Y PSNR-U PSNR-V PSNR-All Frame
Global: 32.353306 40.023221 43.114150 33.845889 1
Avg: 32.353306 40.023221 43.114150 33.845889 1
Min: 32.353306 40.023221 43.114150 33.845889 0
Original comment by fbarch...@google.com
on 29 May 2013 at 4:43
r706 fixes point upsampling. exactly matches ffmpeg
r707 fixes point downsampling. exactly matches ffmpeg and up and then down is
lossless.
For point sampler, the remaining task is specialized scalers should match
general purpose one.
Original comment by fbarch...@google.com
on 29 May 2013 at 5:42
r708 fixes point down sample to 1/2 size. exactly matches ffmpeg and general
purpose scaling.
Performance improved:
Was ScaleDownBy2_None (6547 ms)
Now ScaleDownBy2_None (4471 ms)
Original comment by fbarch...@google.com
on 29 May 2013 at 7:31
r709 fixes 1/4 and 1/8 size for point.
Original issue of up sample and then down sample is lossless
out\Debug\convert.exe faces_1280x720_ARGB.raw faces_352x288_P420.yuv
out\Debug\convert.exe -f 0 faces_352x288_P420.yuv faces_699x531_P420.yuv
out\Debug\convert.exe -f 0 faces_699x531_P420.yuv face2_352x288_P420.yuv
out\Debug\psnr.exe -s 352 288 -v faces_352x288_P420.yuv face2_352x288_P420.yuv
Size: 352x288
Frame PSNR-Y PSNR-U PSNR-V PSNR-All Frame
0 128.000000 128.000000 128.000000 128.000000
Global: 128.000000 128.000000 128.000000 128.000000 1
Avg: 128.000000 128.000000 128.000000 128.000000 1
Min: 128.000000 128.000000 128.000000 128.000000 0
Bilinear is close, but not identical.
out\Debug\convert.exe -f 1 faces_352x288_P420.yuv faces_699x531_P420.yuv
out\Debug\convert.exe -f 1 faces_699x531_P420.yuv face2_352x288_P420.yuv
out\Debug\psnr.exe -s 352 288 -v faces_352x288_P420.yuv face2_352x288_P420.yuv
Size: 352x288
Frame PSNR-Y PSNR-U PSNR-V PSNR-All Frame
0 40.454419 46.045956 47.859297 41.744798 min
Global: 40.454419 46.045956 47.859297 41.744798 1
Avg: 40.454419 46.045956 47.859297 41.744798 1
Min: 40.454419 46.045956 47.859297 41.744798 0
Original comment by fbarch...@google.com
on 30 May 2013 at 5:40
Original comment by fbarch...@google.com
on 30 May 2013 at 7:17
Original issue reported on code.google.com by
fbarch...@chromium.org
on 20 May 2013 at 9:49