2014-07-10 14:53 GMT+02:00 Michael Niedermayer <michae...@gmx.at>: > On Thu, Jul 10, 2014 at 06:47:54AM +0000, Christophe Gisquet wrote: >> When dealing with MVs, both components may be processed at a time. > > mixing memory access sizes can lead to speed penalties on some cpus > when such mixed accesses are close together > did you benchmark this ?
I only have a x86 platform to test it on. No, I haven't benchmarked, nor analyzed the disassembled code. Indeed I would better, as things sometimes work the opposite of what seemed obvious. (also replying to Ronald) Maybe adding some DECLARE_ALIGNED(4, ...) where applicable would fix the alignment issue? Second point, can someone more used to those macros than me document them? I guessed that in AV_[RW][BLN]: - R/W means read/write - Big/Little/Natural endianness Other aspect of this documentation: indicate that aligned addresses may provide better performance or may be required to prevent crashes (e.g. the 128 ones). Some of these might be obvious to most for most platforms, but still. -- Christophe _______________________________________________ ffmpeg-devel mailing list ffmpeg-devel@ffmpeg.org http://ffmpeg.org/mailman/listinfo/ffmpeg-devel
