[Gdal-dev] MrSid Driver performance

[Andrey Kiselev]

On Thu, Nov 18, 2004 at 02:45:23PM +0000, Gao, Peng wrote:
> > First, Andrey thanks a lot for implementing block caching in the
> > driver for DSDK 4.xx, now decompress an RGB MrSID file is very
> > faster ! 
> > 
> > I've made some tests on performances with the 4.xx driver on a big
> > 10000x10000 MrSID RGB file. The file has the version 1.0.1 (with
> > mrsidinfo) and I suppose this is an MG2 file ; I haven't got any big
> > MrSID MG3 file for testing...
> > 
> > I found that the driver can be improved a lot simply by changing the
> > blocks size to 256x256 instead strips calculated with the
> > poImageReader->getStripHeight() DSDK method. With this changes my
> > tests seem to be between 2 and 3 times more faster !
> > 
> > My tests are (on Win32 and Solaris) : - gdal_translate execution
> > time - OpenEV visualisation (it's faster and more reactive) - a sort
> > of gdal_translate in our software - visualisation time in our
> > software
> 
> I fisrt posted this as a followup, but it was buried in many old
> posts.
> 
> Is there any progress made on this? Using a strip may be good for
> format translation,  but not so for viewing where the rectanglar
> display is much smaller than the size of the image. I did some GDAL
> performance test, it performs well for JP2 and Imagine, but much
> slower for MrSID than another package I am using, even after I changed
> the block size to the similar rectangler size. Is there any plan for a
> performance tuning?

I have made several tests, but doesn't found any perfomance hit (on my
data). What is really important here is the band oriented RasterIO
method which is not implemented for 4.x DSDK. That mathod should be very
helpful for applications like OpenEV and Mapserver. I have a bunch of
MrSID related issues in my TODO list and that is one of them.

						Andrey

-- 
Andrey V. Kiselev
Home phone:  +7 812 5970603  ICQ# 26871517