Hi, 2016-04-27 14:52 GMT+02:00 Alexandra Hájková <[email protected]>: > HEVC - no speed loss > H264 - no speed loss > OPUS - 0.7 % speed up > AAC - 2.6 % speed up > DCA - 0.5 % speed up
I have applied all patches except 5 (which maybe could be move to after 8), 9 and 11+, and it seems that for h264, aac and dca, the same code is generated, at least using size on object files. For the sake of thoroughness, and for dcadec.o, I run strip and got the same md5sum for before and after the patch series. Am I missing something? btw, how long are the sequences? Because sequences that take less than 2s to decode will certainly have a lot of noise in the measures. For dnxhddec, I got: From: 6980 256 0 7236 1c44 libavcodec/dnxhddec.o To : 12676 256 0 12932 3284 libavcodec/dnxhddec.o and benchmarking the dnxhd_decode_dct_block function: From: 5953 UNITS in dnxhd, 16774226 runs, 2990 skips To : 5566 UNITS in dnxhd, 16773790 runs, 3426 skips So there's an improvement that probably warrants the code size increase, but I'm somewhat baffled by the results: - I haven't checked the generated code but that looks like a lot more inlining goes there; - I would have thought that, as dnxhd uses larger codewords, the benefit of the "likely" branch wouldn't be there. -- Christophe _______________________________________________ libav-devel mailing list [email protected] https://lists.libav.org/mailman/listinfo/libav-devel
