[FFmpeg-devel] [PATCH 2/5] fate: avoid framemd5, use framecrc its faster

[Nicolas George]

L'octidi 28 floréal, an CCXXI, Michael Niedermayer a écrit :
> in that case comparing plain C of both against each other would be
> interresting too
> if openssl is faster in plain C too then we should consider completely
> replacing our code by it otherwise it would be better to keep our C
> code with openssls asm or something like that

I built openssl with:

./Configure no-asm linux-x86_64 shared --prefix=/tmp/ossl

and compared. Here are the result of the benchmark, doctored to have both
OpenSSL asm and OpenSSL no-asm together:

lavu          MD5      size:  1048576  runs:     1024  time:    8.212 +- 0.121
openssl-noasm MD5      size:  1048576  runs:     1024  time:    5.424 +- 0.353
openssl-asm   MD5      size:  1048576  runs:     1024  time:    5.085 +- 0.320

lavu          SHA-1    size:  1048576  runs:     1024  time:    7.034 +- 0.227
openssl-noasm SHA-1    size:  1048576  runs:     1024  time:    8.194 +- 0.223
openssl-asm   SHA-1    size:  1048576  runs:     1024  time:    4.694 +- 0.182

lavu          SHA-224  size:  1048576  runs:     1024  time:   14.966 +- 0.431
openssl-noasm SHA-224  size:  1048576  runs:     1024  time:   17.409 +- 0.999
openssl-asm   SHA-224  size:  1048576  runs:     1024  time:   12.882 +- 0.407

lavu          SHA-256  size:  1048576  runs:     1024  time:   15.018 +- 0.866
openssl-noasm SHA-256  size:  1048576  runs:     1024  time:   17.391 +- 1.019
openssl-asm   SHA-256  size:  1048576  runs:     1024  time:   12.921 +- 0.821

Clearly, their C MD5 is faster and there are tricks that may be borrowed.

OTOH, their C SHA are worse than lavu's, but there would be benefit in
importing / reimplementing an assembly version.

Regards,

-- 
  Nicolas George