Denis Oliver Kropp wrote: > Phil Endecott wrote: >> Dear All, >> >> I have been playing around with the djpeg program (a trivial wrapper >> around libjpeg for decoding) and there is a performance difference of >> about 25% between the defaults and the fastest settings, with no >> noticeable difference in the output (to my eye - no doubt there are >> cases where it can be seen). > > What does it do? 4:4:4 to 4:2:0 reduction while converting to RGB?
The main thing is to do with rounding in the IDCT. There is also a small gain from using a rectangular upsampling filter during the colour conversion. I believe that it's possible to make it even faster using MMX (on x86), but the code that's available seems to have bit-rotted. I'll investigate this further. >> I've also noticed that djpeg runs about 15% faster if compiled with -Os >> rather than -O4. > > On x86? Yes; it's a VIA C3. > On embedded architectures, -O2 is often better than -O3, but if you have > a very small instruction cache, -Os could be best. I use -Os for everything on my ARM systems, though I think it could benefit from some tweaking. Regards, Phil. _______________________________________________ directfb-dev mailing list [email protected] http://mail.directfb.org/cgi-bin/mailman/listinfo/directfb-dev
