[Zlib-devel] AMD, minigzip and 64 bits benchmark

[Gilles Vollant]

Hello,

The document http://www.amd.com/us-en/assets/content_ty...s/30579_low.pdf at
page 57 talk about an 64 bis optimized minigzip for AMD64:
"Mini-GZIP
The Mini-GZIP benchmark contains a 64-bit benchmark that measures the
performance of an AMD64-optimized version of ZLIB. It also contains a 32-bit
benchmark implemented with the standard (unmodified) ZLIB. The source code
for this benchmark is included on the CD."

Somes websites used the "performance analysis" CD from AMD, to tell that 64
bits minigzip is twice faster than 32 bits
http://www.aceshardware.com/read.jsp?id=60000257
http://www.overclockers.com/articles837/index05.asp
http://www.3dvelocity.com/reviews/AMD%20Athlon%2064%20FX-51/a64fx51_11.htm
http://www.xbitlabs.com/articles/cpu/print/athlon64-fx51.html
http://www.ukgamer.com/article.php4?id=152&page=15

http://www.x86-secret.com/articles/cpu/32vs64/32vs64-5.htm (in french)

The x86-secret.com webmaster send me both source and binaries from the AMD
CD :
The 32 bits versions is a zLib 1.14 DLL compiler with zLib 1.14 with
Microsoft Visual C++ using the nt\Makefile.nt from this distribution, which
did NOT activate the optimisation.
The 64 bits versions used a longest_match function rewritten in AMD64
assembly code.



I tried compile zLib 1.14 with Visual C++ 6.0 with optimisation (just add
"-Ox" in nt/makefile.nt), and I obtain a 32 bits version only a bit slower
than 64 bits version. If I added the Brian Raiter asm 686 code (to compare
with AMD64 assembly code from AMD), this version is a bit faster than AMD 64
bits executable.

And, to terminate, my gvmat64.asm I wrote these day is faster than 64 bits
code from AMD. This is the only way to have 64 bits code faster than Brian
Raiter 32 bits assembly code :-)

Several french site talk about this story
http://www.x86-secret.com/?option=newsd&nid=837
http://www.clubic.com/actualite-18469-amd-montre-du-doigt-pour-ses-benchs-64
-bits-.html
http://www.hardware.fr/news/imprimer/7276/