| 1 |
This is a patched version of zlib modified to use |
| 2 |
Pentium-optimized assembly code in the deflation algorithm. The files |
| 3 |
changed/added by this patch are: |
| 4 |
|
| 5 |
README.586 |
| 6 |
match.S |
| 7 |
|
| 8 |
The effectiveness of these modifications is a bit marginal, as the the |
| 9 |
program's bottleneck seems to be mostly L1-cache contention, for which |
| 10 |
there is no real way to work around without rewriting the basic |
| 11 |
algorithm. The speedup on average is around 5-10% (which is generally |
| 12 |
less than the amount of variance between subsequent executions). |
| 13 |
However, when used at level 9 compression, the cache contention can |
| 14 |
drop enough for the assembly version to achieve 10-20% speedup (and |
| 15 |
sometimes more, depending on the amount of overall redundancy in the |
| 16 |
files). Even here, though, cache contention can still be the limiting |
| 17 |
factor, depending on the nature of the program using the zlib library. |
| 18 |
This may also mean that better improvements will be seen on a Pentium |
| 19 |
with MMX, which suffers much less from L1-cache contention, but I have |
| 20 |
not yet verified this. |
| 21 |
|
| 22 |
Note that this code has been tailored for the Pentium in particular, |
| 23 |
and will not perform well on the Pentium Pro (due to the use of a |
| 24 |
partial register in the inner loop). |
| 25 |
|
| 26 |
If you are using an assembler other than GNU as, you will have to |
| 27 |
translate match.S to use your assembler's syntax. (Have fun.) |
| 28 |
|
| 29 |
Brian Raiter |
| 30 |
[email protected] |
| 31 |
April, 1998 |
| 32 |
|
| 33 |
|
| 34 |
Added for zlib 1.1.3: |
| 35 |
|
| 36 |
The patches come from |
| 37 |
http://www.muppetlabs.com/~breadbox/software/assembly.html |
| 38 |
|
| 39 |
To compile zlib with this asm file, copy match.S to the zlib directory |
| 40 |
then do: |
| 41 |
|
| 42 |
CFLAGS="-O3 -DASMV" ./configure |
| 43 |
make OBJA=match.o |