FORTRAN Programming

Chapter Overview
================

This chapter briefly describes the Fortran compiler options available
on the C3880. Note that by default the compiler does not vectorize
your code. 

Invoking the Fortran Compiler
=============================

Use the fc command to invoke the Fortran compiler. The simplest
way to use the compiler on your Fortran code myprog.f is: 

    % fc myprog.f


Using the compiler in this manner: 

* does not optimize or vectorize your code 
* sends error and warning messages to your terminal 
* provides no output listing file 
* produces an executable file called a.out if there are no errors


In order to optimize, vectorize, or parallelize your code with
the compiler, you must supply one of the -On options, where n
represents 1, 2, or 3. You should compile your code without any
optimization and test it with a small test case. If the answers
are correct, apply options O1, O2, and O3. The O3 should be used
only if you want to parallelize. Check your answers after each
step. At a minimum, every code running on the CONVEX C3880 should
be vectorized using -O2. 

    % fc -O2 -LST -or all myprog.f > myprog.1


This vectorizes the code and produces an output listing named
myprog.l and an executable file named a.out. The output listing
includes an optimization report (-or). Following is a list of
other commonly used command line options. 
Option Description 
-c makes compiler produce an object file (objectfilename.o) only
    
-cxdb enables debug information to be generated for cxdb 
-l name searches specified library file when linking object 
-o name specifies alternate name of executable file rather than
    a.out 
-il allows subprogram inlining (debugging may be difficult with
    this option on) 
-re allows subprograms to be called recursively 
-LST lists source of the code 
-xr lists cross-reference of the code 
-ds allows dynamic selection of multiple loop copies 
-ur causes compiles to automatically select and unroll loops 
-pa instruments code for CONVEX performance analyzer (cxpa). Also
    -pab for block-level and -par for routine-level instrumenting
    
-p instrument for prof analyzer 
-bp instrument for bprof analyzer 
-pg instrument for gprof analyzer

After reaching optimization levels -O2 and -O3, you may want to
incorporate some extra options on your compile command. Add the
-ur option to automatically unroll loops the compiler chooses.
Also try the -ds option to produce multiple copies of some loops
that the CONVEX can choose between dynamically at run-time, depending
on the loop ``trip count.'' CONVEX debugger CXdb and performance
analyzer CXpa are available on out system and are documented by
CONVEX. For more information on optimization, refer to the CONVEX
manual CONVEX FORTRAN Optimization Guide. 

Other command line options are available for the fc compiler.
A full listing can be viewed online by entering the command man
fc. 

Software Information
====================

Information about recent releases is available in the /usr/doc
directory in the following files: 
Fortranv8.0fc.relnotes
Application Compilerv1.1apc.relnotes
X-based debuggerv2.0cxdb.relnotes
support librariesv2.0all.relnotes

Fortran

* CONVEX Fortran version 7.0 introduced Fortran 90 support. This
    support has been extended to include masked array assignments,
    using the WHERE statement and construct as well as two new
    array manipulation intrinsics (EOSHIFT and CSHIFT). 
* Version 8.0 of CONVEX Fortran supports two new compiler directives
    that allow you to increase optimization opportunities by declaring
    variables and arrays private to loops and tasks. The new compiler
    directives are C$DIR DO_PRIVATE and C$DIR TASK_PRIVATE. 
* The -cfcwpa option causes the compiler to compute word addresses
    when performing Cray pointer arithmetic on explicitly-declared
    Cray pointers. 
* The CONVEX Fortran cross-reference utility, fcxref, has been
    enhanced to generate module-interface reports.


Application Compiler

* An interactive browser for the program database named pdbview
    is included.   This X-based tool allows you to filter error
    messages, construct call graphs, and examine the different
    analyses performed by the compiler. 
* Procedure inlining and cloning have been enhanced. 
* The Application Compiler can increase the effectiveness of its
    call frequency analysis by using profile information you supply.
    
* To reduce bank conflicts in vector code, the compiler looks
    for multi-dimensional arrays whose dimensions can cause bank
    conflicts during vector loads and stores of the array. 


CXdb

CXdb, an X-based debugger, includes a significant number of enhancements
and bug fixes. Consult the /usr/doc/v2.0cxdb.relnotes file for
more detailed information. 
.