[FFmpeg-devel] [PATCH 2/2] configure: Use intel math.h header.

[Matt Oliver]

>
> Sounds reasonable. Just curious, with this patch, how much faster is ICL
> compared to MSVC?
>

General msvc/icl benchmarks show a significant speedup but as for FFmpeg
specifically I havnt done a full performance check yet. Also at this point
im not sure what the best way to benchmark FFmpeg for this is. If there are
some specific benchmarks you want me to run let me know and ill compare the
difference.


> How about Linux? Intel Math Library is available for Linux too.
>

In theory this would work perfectly fine with icc but as I dont have an icc
install to test it on I didnt want to submit a untested patch. Some one
with icc should have a look at that one and confirm (also whether having
the header in the msvcrt folder is a little too confusing if also being
used for icc). There should be absolutely no problem to doing this with icc
and I can extend the patch to cover that aswell if wanted (if so feel free
to suggest a better location for the header file).

What is this ../include/math.h? Also from the configure settings this
> header is not included by MSVC so what is the point of this?
>

../include/math.h is there because msvc doesnt support include_next so this
is the equivalent to pass on to the internal math.h file. This code is
there as fallback to prevent issues arising with msvc picking up the
compat/msvcrt path (which can happen if compiling any other file in that
directory). Also I left it there so that it can be used to add some cleaner
msvc compat in a future patch i.e. add in the missing defines not found in
older versions of msvc so it is seamlessly C99 compat without changing any
other FFmpeg code (currently mathematics.h tries to do this but anyone
writing code with gcc will normally just include math.h and not realise
there is a problem with msvc - this would fix that).