And the beat goes on... (compiling the CWI code)

[R.D. Silverman]

Originally Posted by R.D. Silverman

I have a working version of the matrix build code, compiled from the new code. It seems
to work correctly and produces a matrix identical with the old code.

The new CWI code has two versions for LA: gauss.c and lanczos.c

I managed to get a 64-bit version of the former built. But when I run it
it gives the same matrix weight mismatch error that I got before.

I assume that CWI is no longer supporting this, but I'd like to ask someone
there (who is still there?) about this problem.

I could work around the problem, telling the code not to exit, but I don't want to
waste a week running LA code only to find out that I get the wrong answer....

Compiling lanczos.c is still ongoing.

Perhaps gauss.c and buildmatrix.c are not intended to work together.......

Or it could be an array size limitation somewhere......

or......

[R.D. Silverman]

Originally Posted by R.D. Silverman

Perhaps gauss.c and buildmatrix.c are not intended to work together.......

Or it could be an array size limitation somewhere......

or......

I am clearly brain dead. I've been pounding away all day and getting nowhere.

I have scrutinized the code that adds up the column weights while the matrix is being built
and written to disk. I have scrutinized the code in gauss.c that reads the matrix back in.
I'll be damned if I know where the problem is coming from.

The next step is a column by column comparison of the column weights when they are
computed by the matrix build and when they are read by the LA code. This will have to wait.
I have other work to do.

[firejuggler]

take the night/day/week off, migth see the problem better in a few days.

[R.D. Silverman]

Originally Posted by firejuggler

take the night/day/week off, migth see the problem better in a few days.

Good advice.

But I have a matrix to solve and I am blankety blank frustrated.

Trying to debug someone else's code is a <substitute your favorite expression here>

[R.D. Silverman]

Originally Posted by R.D. Silverman

Good advice.

But I have a matrix to solve and I am blankety blank frustrated.

Trying to debug someone else's code is a <substitute your favorite expression here>

And I just ran into another problem. The LA source calls a function
ConstructTmpName that is nowhere in the distribution.

There is a construct_tmp_name with different arguments.

I found ConstructTmpName in an old (1996) CWI distribution in an
include file with the same name as the file in the new distribution that has
construct_tmp_name

I've been running (up until now) 32 bit post processing code built from the '96 code.

I replaced the new version with the old file and got everything to compile and link.

Miracle!

The LA code, compiled and linked with gcc seems to run fine.
(I will know in a week, when it finishes)

The same code, compiled and linked with Visual Studio fails.

I now have to work on getting the multi-threaded LA code compiled and linked.
And I need to put together 64 bit filter code, although the 32-bit code works OK.
Neither of these tasks is urgent.

[ldesnogu]

Originally Posted by R.D. Silverman

And I just ran into another problem. The LA source calls a function
ConstructTmpName that is nowhere in the distribution.

There is a construct_tmp_name with different arguments.

I found ConstructTmpName in an old (1996) CWI distribution in an
include file with the same name as the file in the new distribution that has
construct_tmp_name

I *guess* you can safely pick the ConstructTmpName from statistics.c and just remove the second parameter (100) in the call found in gauss.c.