|
Most Research Computing Linux computers have three different kinds of
compilers: GNU compilers, Portland Group compilers, and Intel compilers.
Below is a table that shows the different compilers and their commands:
|
Language
|
File Suffix
|
Command
|
Type of Compiler
|
|
Fortran 77
|
.f
|
pgf77
|
Portland Group
|
|
|
.f
|
ifort
|
Intel
|
|
|
.f .F .for
|
g77
|
GNU
|
|
Fortran 90
|
.f .f90
|
pgf90
|
Portland Group
|
|
|
.f .f90
|
ifort
|
Intel
|
|
C
|
.c
|
pgcc
|
Portland Group
|
|
|
.c
|
icc
|
Intel
|
|
|
.c
|
gcc
|
GNU
|
|
C++
|
.c .C
|
pgCC
|
Portland
|
|
|
.c
|
icc
|
Intel
|
|
|
.c
|
g++
|
GNU
|
No one compiler works best for all kinds of code. If you are interested in
speeding up your application, you will want to test various compiler options
and compilers to see which ones work the best for your application. There is a
man page for each compiler that lists the compiler options.
In general, gcc/g++ is the most flexible since it runs on a variety of
different computer platforms. Many people use gcc/g++ so it will be easy to
port their code to other platforms. g77 is the Fortran77 GNU compiler and is
not as robust as the Portland Group and Intel Fortran compilers.
The Portland Group compilers will compile f77, f90, C, and C++ codes. Many
Fortran codes use extensions to the language, and the Portland Group compilers
support those extensions. Portland Group recommends you use the pgf90 for both
Fortran 90 and Fortran 77 codes to gain additional code speed up.
The Intel compilers are more restrictive than the Portland Group and GNU
compilers and don't always allow language extensions. However, they do a good
job of compiling code to make use of the hyper-threading/SSE enhancements on
the Pentium 4/Xeon processors.
All of the compiler commands listed above are in the standard user path. In
addition, the compilers have been linked as provided below.
Useful Compiler Options
Following are some useful compiler optimization options to know about for
each of the three kinds of compilers. All three kinds of compilers generate
code that is optimized for the platform where they are compiled. The comments
below apply to both C/C++ and Fortran77/90 codes. These optimizations will not
necessarily speed up all kinds of code and should be tested individually.
Portland Group: pgcc, pgCC, pgf77, pgf90
If you want to use a number of more aggressive optimizations that make use
of the features of the Pentium4/Xeon/AMD processors to speed up your code and
use the options -fastsse -O4. In addition, following are other optimization
options to try.
|
PG Compiler Option
|
Description of Option
|
|
-Munroll
|
Unroll loops.
|
|
-Mvect
|
Vectorize code.
|
|
-Mvect=sse
|
Make use of the sse/see2 vectorization.
|
|
-Mconcur
|
Auto-parallelization using OpenMP.
|
| -mp |
parallelize using user-inserted OpenMP directives |
|
-Mipa
|
Inter- procedural analysis.
|
Intel Compilers
Use the option -O3 to obtain more aggressive optimization. You can use the
-parallel option to have the compiler insert multi-threaded code in loops that
can be safely parallelized and make use of the hyper-threading capabilities of
the Pentium 4/Xeon processors. Following are some other optimization
options.
|
Intel Compiler Option
|
Description of Option
|
|
-ipo
|
Multi-file interprocedural analysis.
|
|
-unroll[n]
|
Unroll loops.
|
| -parallel |
Auto-parallelization using OpenMP |
|
-openmp
|
Generate OpenMP code in loops that can be parallelized.
|
GNU gcc, g++
Use the option -O3 for the most aggressive optimization. Following are some
additional optimization options.
|
gcc/g++ Compiler Option
|
Description of Option
|
|
-march=machinetype (i.e., pentium4)
|
Generate instruction of the machine type (i.e., pentium4).
|
|
-ffast-math
|
Use optimized math functions (this should be used with caution).
|
|
-mfpmath=sse
|
Make use of sse vectorization for math calculations.
|
|
-funroll-loops
|
Unroll loops.
|
Send questions, corrections, or additional information about these compilers
to Susan A. Schwarz
|
