The chief difference between round and lrint/llrint lies in somewhat nuanced semantics: 1. (minor) round inherently returns a double precision value, so this is then implicitly cast to an integer afterwards. Nothing inherently wrong with this, but the implicit cast may be easily missed. 2. l(l)rint uses the superior "Dutch/Gaussian" rounding, i.e it rounds half-way points to even numbers instead of biasing away from zero, which is actually forced by round. Of course, this does not eliminate all sources of such bias (e.g y-0.5 always being even), but it is observed to be far better in practice.
On x86-64, they map to cvtsd2si and cvttsd2si (extra t for truncation), which turn out to have identical cycle times. Same goes for the vector variants. This patchset accordingly removes all needless round instructions. In a few cases, speedups are actually obtained, i.e changes are not purely cosmetic/ theoretical for the halfway cases. When observed, they have been noted in respective commit messages. Patchset tested with FATE. Ganesh Ajjanagadde (12): lavfi/vf_idet: replace round and cast by lrint lavfi/vf_colorchannelmixer: replace round by lrint lavfi/vf_colorlevels: replace round by lrint lavfi/vf_drawtext: replace round by llrint lavfi/vsrc_mandelbrot: replace round by lrint lavf/hlsenc: replace round by lrint lavc/libvpxenc: replace round by lrint lavfi/vf_crop: replace round by lrint lavfi/af_dynaudnorm: replace round by lrint lavfi/vf_cropdetect: replace round by lrint lavfi/vf_histogram: replace round by lrint lavd/pulse_audio_enc: replace lround by lrint libavcodec/libvpxenc.c | 4 ++-- libavdevice/pulse_audio_enc.c | 2 +- libavfilter/af_dynaudnorm.c | 2 +- libavfilter/vf_colorchannelmixer.c | 38 +++++++++++++++++++------------------- libavfilter/vf_colorlevels.c | 16 ++++++++-------- libavfilter/vf_crop.c | 2 +- libavfilter/vf_cropdetect.c | 2 +- libavfilter/vf_drawtext.c | 2 +- libavfilter/vf_histogram.c | 2 +- libavfilter/vf_idet.c | 2 +- libavfilter/vsrc_mandelbrot.c | 2 +- libavformat/hlsenc.c | 2 +- 12 files changed, 38 insertions(+), 38 deletions(-) -- 2.6.4 _______________________________________________ ffmpeg-devel mailing list ffmpeg-devel@ffmpeg.org http://ffmpeg.org/mailman/listinfo/ffmpeg-devel
