zlib1.dll

Mark Adler wrote:

> The last version of zlib before 1.2.1 is available here:
>
>     http://www.alumni.caltech.edu/~madler/zlib-1.2.0.7.tar.gz

Given that the new release is at the door, I prepared a DLL
distribution:

  http://www.cs.toronto.edu/~cosmin/pngtech/zlib/zlib-1207-dll.zip

Gerard advised that a DLL distribution should always be available, to
reduce (eliminate?) the need for private builds.

In the zlib FAQ it is written that the DLL can be downloaded from the
zlib site. Therefore, each time there is a new zlib release, somebody
has to synchronize with Mark and make the DLL release at the same time,
or shortly after.

I am volunteering for this task. As a test, I prepared zlib-1207-dll,
built from the 1.2.0.7 release candidate.

I must tell you, however, that I only have Visual C++ 6.0 and MinGW-3.0
(gcc-3.2.3). Gilles Vollant pointed out that Visual C++ 7.0 and 7.1
optimize the code for PIII, but I don't have VC7, and I don't plan to
buy it any time soon. If you have it, and are willing to provide the
zlib1.dll builds yourself, write me. You don't have to write docs,
because I did it already (they're in the zlib-1207-dll distribution).

You'll only have to:
 1. Build the DLL from the unaltered sources, as provided by Mark;
 2. Update the zlib version number in README.txt
 3. Make an archive and send it to Mark, to be put in the zlib www site.

**

I will write some performance considerations below, based on my testing
on an AMD Duron 900MHz and Win2k.

- I noticed no performance difference between zlib1.dll as compiled with
Microsoft C vs. MinGW (no ASM). When using gcc, I disabled the frame
pointers (-fomit-frame-pointer)

- Inlining did *not* prove to be beneficial. I noticed very, very minor
performance decreases (1-2%) when I built zlib with inlining, both with
Microsoft C (with /Ob2) and with gcc (-O3 vs. -O2). Maybe this
difference was caused by the OS overhead, but even in this case, the
lesson is that zlib should be compiled without inlining.

- The ASM deflation code (ASMV) provided little benefit, around 2-3%.

- The ASM inflation code (ASMINF) provided a considerable benefit, over
20%.

The tests were run many times, and the outcomes were averaged (except
the first one, which was discarded to eliminate the impact of loading
the data into the cache).


Best regards,
Cosmin

--
Send the message body "help" to [email protected]