| Re: C++ in IDL [message #44286 is a reply to message #29423] |
Wed, 01 June 2005 10:00   |
JD Smith
Messages: 850
Registered: December 1999
|
Senior Member |
|
|
On Wed, 01 Jun 2005 10:10:34 -0600, David Fanning wrote:
> Michael Wallace writes:
>
>> There's that much of a difference? I thought sure that IDL would have
>> been designed for number crunching considering the market they are
>> trying to serve. Of course, if the IDL code wasn't optimized...
>
> I've seen this kind of comparison many times. On further
> investigation the IDL code is usually, uh, not well written. :-)
> I would take any comparisons not performed by J.D. with a
> grain of salt. :-)
I've been known to push on IDL to get as much performance as possible,
and still, for many types of problems, a direct C-coded approach,
compiled with optimization, will out-perform the best-optimized IDL
version by a factor of 10 or more. Here's a good example from your
site in which we collected and refined the best-optimized IDL
algorithms from the experts (Craig, Wayne, etc.), and still I found a
simple compiled C approach is 20x faster:
http://www.dfanning.com/code_tips/drizzling.html
IDL truly excels at basic large array operations (adding 1 to 1
million integers, etc.); for these types of problems you'll find
comparable performance to compiled code (ballpark anyway). For more
complex algorithms, you can make huge gains in performance with
careful vectorization and other methods in IDL, but, in the end, may
still (but not always) miss compiled performance by a non-negligible
margin.
This is not actually surprising. It's part of the deal you make when
you code in IDL at a high-level, vs., e.g., C or FORTRAN at a low
level. Did you think you would get all of those many conveniences
which IDL provides, like loose typing, dynamic variable content,
optional and keyword arguments, variable type checking, etc., for
free? A much fairer comparison would be between, e.g., Matlab's
array-crunching performance and IDL's. Among the so-called 4G
languages (including scripting extensions like Perl's PDL, etc.), IDL
delivers among the best performance for basic array operations.
I might suggest prototyping everything in IDL, profiling to find the
algorithm or sub-algorithm which is limiting performance, and then
re-coding that in C as a DLM. Sometimes, as in the case referenced
above, the C version can be quite simple and compact, especially if
all of the setup, display, analysis and data juggling is handled in
IDL. Simple DLMs are actually not difficult to produce, and it's
worth your while to learn how to make use of them. Their one real
drawback is cross-platform compatibility, but with MAKE_DLL and auto
glue, you can actually recover this to some degree.
JD
|
|
|
|
comp.lang.idl-pvwave archive
Members
Search
Help
Login
Home
