[FFmpeg-devel] [PATCH] E-AC-3 spectral extension (bump)

Christophe Gisquet christophe.gisquet
Thu Mar 25 15:52:31 CET 2010


2010/3/24 Michael Niedermayer <michaelni at gmx.at>:
> would we overall have fewer unaligned accesses if we would place the array
> so that not element 0 but element 1 is aligned?

I've just now see that this might be troublesome later on, as
int32_to_float_fmul_scalar and ff_imdct_half are used on those buffers
and both expect 16-bytes aligned data (movaps used).

>> num_copy_sections can definitely be >1. ?That is the whole point of it.
>
> if num_copy_sections is only rarely >1 than this optimization is maybe
> not worth it

At least the samples on mplayerhq and the French TNT-HD broadcasts are
in this situation.

Last but not least, I've used START/STOP_TIMER and oprofile to cycle
counts and % of execution time. It appears that the CSI samples have
CRC whose computation eats most of the runtime. Here are those figures
as food for thought:

      | (1) | (2) |  (3) |  (4)  |  (5)  |  (6) |
rambo | 1.0 |  64 | 3870 | 14000 |  11.2 |  0.0 | 11.2
CSI   | 5.1 | 256 | 5020 | 83000 |   0.9 | 81.4 |  5.1
CSI   | 2.0 | 128 | 1920 | 12700 |   0.3 | 81.1 |  1.6

(1) Channel configuration
(2) Bitrate (Kb/s)
(3) dezicycles for the RMS computation loop
(4) dezicycles for ff_eac3_apply_spectral_extension
(5) oprofile % for ff_eac3_apply_spectral_extension
(6) oprofile % for av_crc
(7) oprofile pseudo % for ff_eac3_apply_spectral_extension if no av_crc

>> spx_src_start_freq, spx_dst_start_freq, spx_dst_end_freq ?
>
> sound good

Locally changed, though I'm wondering if Justin expects me to be the
one submitting the patch :)

Best regards,
Christophe



More information about the ffmpeg-devel mailing list