Tela Installation Frequencly Asked Questions For 1.22

First and foremost you need a working C++ compiler.
For example, g++ or Cfront can be used.

LAPACK and BLAS libraries, GNU Readline library, HDF library
plus include files. PlotMTV1.4.1t or later for graphics.
In addition you may find it convenient to have bison++ and
flex++ parser generators installed. A Fortran compiler is
not strictly necessary if you have precomplied LAPACK/BLAS
libraries, but the necessary Fortran runtime libraries must
be there.

 Q1.2: Where to obtain more information?
Read the file INSTALL in the main source directory.


COMPILATION

 Q2.1: The compilation stops in files.ct in HDF include files with many errors.
The HDF library include files expect one of the HDF-specific machine
names to be defined. For a list see ls MAKEFILE* in the HDF source
directory. If your machine is not directly supported, you can usually
do rather well by using an alternative. For example on HPUX HDF library
compiles with -DSUN or -DIRIS. Just be aware that byte ordering is
significant, therefore on Linux you should use SUN386. Of course you
want to use the same -D macro when compiling libdf.a and Tela.

 Q2.2: I have problems with the Makefile on IRIS system?
Some IRIX versions have a buggy make. At least 5.2 has this problem.
Use GNU make to avoid problems (this is a good hint for all other
software packages too).

 Q2.3: It doesn't find some include files.
Check variables containing INCLUDE in makeinc and Makefile.
Do grep INCLUDE makeinc Makefile.

 Q2.4: It complains about multiply defined outputters e.g. in codegen.C.
There are two possibilities. 1) BROKEN_UCHAR_OUTPUTTER has been defined
wrong in makeinc. 2) your C++ compiler is buggy and is unwilling to
compile it for you. In the latter case the only fix is to manually insert
(char*) type casts in source code in appropriate places.

 Q2.5: The compiler gets stuck in objarithm.C.
On some systems, for example IRIX6.x in 64-bit mode, it takes very long
time to compile objarithm.C using the -O3 optimization. Can't help it,
the compiler just spends that time optimizating the thing. Just wait.

 Q2.6: The compiler runs out of memory.
This may happen under AIX if you use -O3. Buy more memory, try to reconfigure
your virtual memory (hope you can find someone who knows how), or reduce
optimization level to -O2.

 Q2.7: The compiler crashes in fftpack.C.
This may happen under AIX if you use -O3. Use -O2 for that file.

 Q2.8: flex++ complains about bad options!
Your flex++ probably is not the flex++ Tela wants. Sometimes the
ordinary GNU flex installer does ln -s flex flex++ because flex
has some C++ features. Get the real flex++ from
ftp.funet.fi:/pub/sci/math/tela/needed/, or define FLEX++ = true
in makeinc.


LINKING

 Q3.1: All LAPACK functions seem to be undefined in linking phase.
The Fortran compiler encodes the external names for the linker.
Among the used practices for Fortran subroutine foo are:

    foo         (no encoding) HPUX, AIX
    foo_        Most other Unix systems and f2c
    FOO         Cray UNICOS

(On AIX this can be changed by the user.)
Tela C++ code must conform to the practice used by the Fortran compiler.
The file affected is lapack.H. A mismatch in NO_UNDERSCORE_IN_FORTRAN_NAMES
shows up in linking phase as an army of undefined symbols.

 Q3.2: tputs or something similar seems to be undefined.
You probably forgot to insert -ltermcap to READLINELIB.

 Q3.3: alloca is undefined.
Your C++ system may not support alloca, or has a buggy support
that configure has rejected. This is the case if C_ALLOCA is defined
in makeinc. Compile alloca.c by cc -c alloca.c by hand and ensure
that alloca.o is among the linked-in files.

 Q3.4: Some Fortran-specific stuff seems to be undefined.
The Fortran runtime support are found in many libraries on different
systems. Try -lF77, -lU77, -lI77, -lisam. If you compiled using F-90
compiler the library names are probably different (probably not -lF90
though, that would evidently make life too dull). A useful trick
in looking for undefined symbols is to write (in C-shell)

cd /usr/lib      # or whatever
foreach i (lib*.a)
echo $i
nm $i | grep UNDEFINED_SYMBOL
end


INSTALLATION

 Q4.1: install.sh seems to install only part of the files.
Yes. I know. install.sh actually does not work properly.
I should write a better one. Meanwhile, get and install (!) GNU install.


RUNTIME ERRORS

 Q5.1: The linear algebra functions seem to produce garbage results.
Check that the floating point precision of your LAPACK/BLAS routines
matches Treal and Tint in Tela (in def.H). By default Treal is defined
to be double and Tint is defined to be int.

 Q5.2: The FFT routines produce garbage.
This should no longer occur if you are using Tela version 1.21 or later.
On earlier versions this is probably because you failed to compile FFTPACK
with -r8 or equivalent switch. FFTPACK is formally single precision (REAL)
while Tela expects double precision (C type double), therefore the Fortran
compiler must be forced to promote REALs to DOUBLE PRECISION using a command
line switch. On Tela>=1.21 this no longer happens because fftpack has been
translated in C++.

 Q5.3: Sometimes I see weird hex numbers where I expect to see a string in output?
Some C++ systems handle iostream unsigned char* outputting incorrectly,
i.e. they output not a string but the pointer in hex. The configure
script tries to recognize this but the check has been somehow fooled
in your case. Set BROKEN_UCHAR_OUTPUTTER to 1 in makeinc.


PROBLEMS ON SPECIFIC MACHINES

 Q6.1: Problems on Cray UNICOS 8.0.3 using scc4.0 and C++1.0?
- The compiler sometimes crashes on the VECTORIZE macro.
  By replacing every occurrence of VECTORIZE by the corresponding
  pragma should help. I used the following small filter program
  to do the job:

#include <stdio.h>
#include <string.h>

#define MAXLINE 1024

int main()
{
        char s[MAXLINE+1];
        char *pos;
        int i;
        while (1) {
                fgets(s,MAXLINE,stdin);
                if (feof(stdin)) break;
                pos = strstr(s,"VECTORIZED");
                if (pos) {
                        for (i=0; s[i] && (&s[i]) != pos; i++) putchar(s[i]);
                        printf("\n#pragma _CRI ivdep\n");
                        i+= strlen("VECTORIZED");
                        for (; s[i]; i++) putchar(s[i]);
                } else {
                        fputs(s,stdout);
                }
        }
        return 0;
}

I filtered gatscat.C and object.C through this.

- I could not get the readline library working, thus
  noreadline.c was used instead.

- The bison++ parser generator compiles out of box but
  produces incorrect results! The resulting parser 
  would give syntax error at every ';', for example.

- Some of the elementary functions give floating point exceptions
  on some bad input arguments because the IEEE arithmetic is not
  in use on Cray.

- The -haggress option produces "out of memory during compilation"
  style messages on objarithm.C. It compiled, nevertheless, but
  some of the loops are probably not vectorized.

- The WHEN_COMPILED string must be made simpler in Makefile

tela-bugs@fmi.fi

Back to Tela home page