Preface

This guide is part of a set of manuals that describe how to use the PGI Fortran, C, and C++ compilers and program development tools. These compilers and tools include the PGF77,PGFORTRAN, PGC++, andPGCC compilers, the PGI profiler, and the PGI debugger. They work in conjunction with an x86-64 assembler and linker. You can use the PGI compilers and tools to compile, debug, optimize, and profile serial and parallel applications for x86 processor-based systems.

The PGI Compiler User’s Guide provides operating instructions for the PGI command-level development environment. The PGI Compiler Reference Manual contains details concerning the PGI compilers' interpretation of the Fortran language, implementation of Fortran language extensions, and command-level compilation. Users are expected to have previous experience with or knowledge of the Fortran programming language. Neither guide teaches the Fortran programming language.

Audience Description

This manual is intended for scientists and engineers using the PGI compilers. To use these compilers, you should be aware of the role of high-level languages, such as Fortran, C, and C⁠+⁠+, as well as assembly-language in the software development process; and you should have some level of understanding of programming. The PGI compilers are available on a variety of x86-64/x64 hardware platforms and operating systems. You need to be familiar with the basic commands available on your system.

Compatibility and Conformance to Standards

Your system needs to be running a properly installed and configured version of this PGI product. For information on installing PGI compilers and tools, refer to the Release Notes and Installation Guide included with your software.

For further information, refer to the following:

  • American National Standard Programming Language FORTRAN, ANSI X3. -1978 (1978).
  • ISO/IEC 1539-1 : 1991, Information technology – Programming Languages – Fortran, Geneva, 1991 (Fortran 90).
  • ISO/IEC 1539-1 : 1997, Information technology – Programming Languages – Fortran, Geneva, 1997 (Fortran 95).
  • ISO/IEC 1539-1 : 2004, Information technology – Programming Languages – Fortran, Geneva, 2004 (Fortran 2003).
  • ISO/IEC 1539-1 : 2010, Information technology – Programming Languages – Fortran, Geneva, 2010 (Fortran 2008).
  • Fortran 95 Handbook Complete ISO/ANSI Reference, Adams et al, The MIT Press, Cambridge, Mass, 1997.
  • The Fortran 2003 Handbook, Adams et al, Springer, 2009.
  • OpenMP Application Program Interface, Version 3.1, July 2011, http://www.openmp.org.
  • Programming in VAX Fortran, Version 4.0, Digital Equipment Corporation (September, 1984).
  • IBM VS Fortran, IBM Corporation, Rev. GC26-4119.
  • Military Standard, Fortran, DOD Supplement to American National Standard Programming Language Fortran, ANSI x.3-1978, MIL-STD-1753 (November 9, 1978).
  • American National Standard Programming Language C, ANSI X3.159-1989.
  • ISO/IEC 9899:1999, Information technology – Programming Languages – C, Geneva, 1999 (C99).
  • ISO/IEC 9899:2011, Information Technology – Programming Languages – C, Geneva, 2011 (C11).
  • ISO/IEC 14882:2011, Information Technology – Programming Languages – C++, Geneva, 2011 (C++11).

Organization

Users typically begin by wanting to know how to use a product and often then find that they need more information and facts about specific areas of the product. Knowing how as well as why you might use certain options or perform certain tasks is key to using the PGI compilers and tools effectively and efficiently. However, once you have this knowledge and understanding, you very likely might find yourself wanting to know much more about specific areas or specific topics.

This guide contains the essential information on how to use the compiler and is divided into these sections:

Getting Started provides an introduction to the PGI compilers and describes their use and overall features.

Use Command-line Options provides an overview of the command-line options as well as task-related lists of options.

Optimizing and Parallelizing describes standard optimization techniques that, with little effort, allow users to significantly improve the performance of programs.

Using Function Inlining describes how to use function inlining and shows how to create an inline library.

Using OpenMP provides a description of the OpenMP Fortran parallelization directives and of the OpenMP C and C++ parallelization pragmas, and shows examples of their use.

Using MPI describes how to use MPI with PGI products.

Using an Accelerator describes how to use the PGI Accelerator compilers.

Eclipse describes how to use the PGI C/C++ compilers from within Eclipse, a free, open-source, integrated software development environment.

Using Directives and Pragmas provides a description of each Fortran optimization directive and C/C++ optimization pragma, and shows examples of their use.

Creating and Using Libraries discusses PGI support libraries, shared object files, and environment variables that affect the behavior of the PGI compilers.

Using Environment Variables describes the environment variables that affect the behavior of the PGI compilers.

Distributing Files – Deployment describes the deployment of your files once you have built, debugged and compiled them successfully.

Inter-language Calling provides examples showing how to place C language calls in a Fortran program and Fortran language calls in a C program.

Programming Considerations for 64-Bit Environments discusses issues of which programmers should be aware when targeting 64-bit processors.

C/C++ Inline Assembly and Intrinsics describes how to use inline assembly code in C and C⁠+⁠+ programs, as well as how to use intrinsic functions that map directly to x86-64 machine instructions.

Hardware and Software Constraints

This guide describes versions of the PGI compilers that produce assembly code forx64 processor-based systems. Details concerning environment-specific values and defaults and system-specific features or limitations are presented in the release notes delivered with the PGI compilers.

Conventions

This guide uses the following conventions:

italic
is used for emphasis.
Constant Width
is used for filenames, directories, arguments, options, examples, and for language statements in the text, including assembly language statements.
Bold
is used for commands.
[ item1 ]
in general, square brackets indicate optional items. In this case item1 is optional. In the context of p/t-sets, square brackets are required to specify a p/t-set.
{ item2 | item 3 }
braces indicate that a selection is required. In this case, you must select either item2 or item3.
filename ...
ellipsis indicate a repetition. Zero or more of the preceding item may occur. In this example, multiple filenames are allowed.
FORTRAN
Fortran language statements are shown in the text of this guide using a reduced fixed point size.
C/C++
C/C++ language statements are shown in the test of this guide using a reduced fixed point size.

The PGI compilers and tools are supported on a wide variety of Linux, macOS and Windows operating systems running on 64-bit x86-compatible processors, and on Linux running on OpenPOWER processors. (Currently, the PGI debugger is supported on x86-64/x64 only.) See the Compatibility and Installation section on the PGI website for a comprehensive listing of supported platforms.

Note: Support for 32-bit development was deprecated in PGI 2016 and is no longer available as of the PGI 2017 release. PGI 2017 is only available for 64-bit operating systems and does not include the ability to compile 32-bit applications for execution on either 32- or 64-bit operating systems.

Terms

A number of terms related to systems, processors, compilers and tools are used throughout this guide. For example:

accelerator FMA -mcmodel=medium shared library
AVX host -mcmodel=small SIMD
CUDA hyperthreading (HT) MPI SSE
device large arrays MPICH static linking
DLL license keys multicore Win64
driver linux86-64 NUMA x64
DWARF LLVM OpenPOWER s86
dynamic library manycore osx86-64 x87

For a complete definition of these terms and other terms in this guide with which you may be unfamiliar, please refer to the PGI online glossary.

The following table lists the PGI compilers and tools and their corresponding commands:

Table 1. PGI Compilers and Commands
Compiler or Tool Language or Function Command
PGF77 ANSI FORTRAN 77 pgf77
PGFORTRAN ISO/ANSI Fortran 2003 pgfortran
PGCC ISO/ANSI C11 and K&R C pgcc
PGC++ ISO/ANSI C++14 with GNU compatibility pgc++ on Linux and macOS
PGI Debugger Source code debugger pgdbg
PGI Profiler Performance profiler pgprof

In general, the designation PGI Fortran is used to refer to the PGI Fortran 2003 compiler, and pgfortran is used to refer to the command that invokes the compiler. A similar convention is used for each of the PGI compilers and tools.

For simplicity, examples of command-line invocation of the compilers generally reference the pgfortran command, and most source code examples are written in Fortran. Usage of the PGF77 compiler, whose features are a subset of PGFORTRAN, is similar.Usage of PGC⁠+⁠+ and PGCC is consistent with PGFORTRAN and PGF77, though there are command-line options and features of these compilers that do not apply to PGFORTRAN and PGF77, and vice versa.

There are a wide variety of 64-bit x86-compatible processors in use. All are supported by the PGI compilers and tools. Most of these processors are forward-compatible, but not backward-compatible, meaning that code compiled to target a given processor will not necessarily execute correctly on a previous-generation processor.

A table listing the processor options that PGI supports is available in the Release Notes. The table also includes the features utilized by the PGI compilers that distinguish them from a compatibility standpoint.

In this manual, the convention is to use "x86" to specify the group of processors that are "32-bit" but not "64-bit". The convention is to use "x64" to specify the group of processors that are both "32-bit" and "64-bit". x86 processor-based systems can run only 32-bit operating systems. x64 processor-based systems can run either 32-bit or 64-bit operating systems, and can execute all 32-bit x86 binaries in either case. x64 processors have additional registers and 64-bit addressing capabilities that are utilized by the PGI compilers and tools when running on a 64-bit operating system. The prefetch, SSE1, SSE2, SSE3, and AVX processor features further distinguish the various processors. Where such distinctions are important with respect to a given compiler option or feature, it is explicitly noted in this manual.

Note: The default for performing scalar floating-point arithmetic is to use SSE instructions on targets that support SSE1 and SSE2.
Note: Support for 32-bit development was deprecated in PGI 2016 and is no longer available as of the PGI 2017 release. PGI 2017 is only available for 64-bit operating systems and does not include the ability to compile 32-bit applications for execution on either 32-bit or 64-bit operating systems.

1. Getting Started

This section describes how to use the PGI compilers.

1.1. Overview

The command used to invoke a compiler, such as the pgfortran command, is called a compiler driver. The compiler driver controls the following phases of compilation: preprocessing, compiling, assembling, and linking. Once a file is compiled and an executable file is produced, you can execute, debug, or profile the program on your system. Executables produced by the PGI compilers are unconstrained, meaning they can be executed on any compatible x86-64 processor-based system, regardless of whether the PGI compilers are installed on that system.

In general, using a PGI compiler involves three steps:

  1. Produce program source code in a file containing a .f extension or another appropriate extension, as described in Input Files. This program may be one that you have written or one that you are modifying.
  2. Compile the program using the appropriate compiler command.
  3. Execute, debug, or profile the executable file on your system.

You might also want to deploy your application, though this is not a required step.

The PGI compilers allow many variations on these general program development steps. These variations include the following:

  • Stop the compilation after preprocessing, compiling or assembling to save and examine intermediate results.
  • Provide options to the driver that control compiler optimization or that specify various features or limitations.
  • Include as input intermediate files such as preprocessor output, compiler output, or assembler output.

1.2. Creating an Example

Let's look at a simple example of using the PGI compiler to create, compile, and execute a program that prints:

hello
  1. Create your program. For this example, suppose you enter the following simple Fortran program in the file hello.f:
    print *, "hello"
    end 
  2. Compile the program. When you created your program, you called it hello.f. In this example, we compile it from a shell command prompt using the default pgfortran driver option. Use the following syntax:
    $ pgfortran hello.f 

    By default, the executable output is placed in the file a.out, or, on Windows platforms, in a filename based on the name of the first source or object file on the command line. However, you can specify an output file name by using the -⁠o option.

    To place the executable output in the file hello, use this command:

    $ pgfortran -o hello hello.f 
  3. Execute the program. To execute the resulting hello program, simply type the filename at the command prompt and press the Return or Enter key on your keyboard:
    $ hello
    Below is the expected output:
    hello

1.3. Invoking the Command-level PGI Compilers

To translate and link a Fortran, C, or C⁠+⁠+ program, the pgf77, pgf95,pgfortran, pgcc and pgc⁠+⁠+ commands do the following:

  1. Preprocess the source text file.
  2. Check the syntax of the source text.
  3. Generate an assembly language file.
  4. Pass control to the subsequent assembly and linking steps.

1.3.1. Command-line Syntax

The compiler command-line syntax, using pgfortran as an example, is:

pgfortran [options] [path]filename [...]
Where:
options
is one or more command-line options, all of which are described in detail in Use Command-line Options.
path
is the pathname to the directory containing the file named by filename. If you do not specify the path for a filename, the compiler uses the current directory. You must specify the path separately for each filename not in the current directory.
filename
is the name of a source file, preprocessed source file, assembly-language file, object file, or library to be processed by the compilation system. You can specify more than one [path]filename.

1.3.2. Command-line Options

The command-line options control various aspects of the compilation process. For a complete alphabetical listing and a description of all the command-line options, refer to Use Command-Line Options.

The following list provides important information about proper use of command-line options.
  • Command-line options and their arguments are case sensitive.
  • The compiler drivers recognize characters preceded by a hyphen (-⁠) as command-line options. For example, the -⁠Mlist option specifies that the compiler creates a listing file.

    Note: The convention for the text of this manual is to show command-line options using a dash instead of a hyphen; for example, you see -⁠Mlist.
  • The order of options and the filename is flexible. That is, you can place options before and after the filename argument on the command line. However, the placement of some options is significant, such as the -⁠l option, in which the order of the filenames determines the search order.

    Note: If two or more options contradict each other, the last one in the command line takes precedence.

1.3.3. Fortran Directives and C/C++ Pragmas

You can insert Fortran directives and C/C++ pragmas in program source code to alter the effects of certain command-line options and to control various aspects of the compilation process for a specific routine or a specific program loop. For more information on Fortran directives and C/C⁠+⁠+ pragmas, refer to Using OpenMP and Using Directives and Pragmas.

1.4. Filename Conventions

The PGI compilers use the filenames that you specify on the command line to find and to create input and output files. This section describes the input and output filename conventions for the phases of the compilation process.

1.4.1. Input Files

You can specify assembly-language files, preprocessed source files, Fortran/C/C++ source files, object files, and libraries as inputs on the command line. The compiler driver determines the type of each input file by examining the filename extensions.

Note: For systems with a case-insensitive file system, use the -⁠Mpreprocess option, described in ‘Command-Line Options Reference’ section of the PGI Compiler Reference Manual, under the commands for Fortran preprocessing.

The drivers use the following conventions:

filename.f
indicates a Fortran source file.
filename.F
indicates a Fortran source file that can contain macros and preprocessor directives (to be preprocessed).
filename.FOR
indicates a Fortran source file that can contain macros and preprocessor directives (to be preprocessed).
filename.F90
indicates a Fortran 90/95 source file that can contain macros and preprocessor directives (to be preprocessed).
filename.F95
indicates a Fortran 90/95 source file that can contain macros and preprocessor directives (to be preprocessed).
filename.f90
indicates a Fortran 90/95 source file that is in freeform format.
filename.f95
indicates a Fortran 90/95 source file that is in freeform format.
filename.cuf
indicates a Fortran 90/95 source file in free format with CUDA Fortran extensions.
filename.CUF
indicates a Fortran 90/95 source file in free format with CUDA Fortran extensions and that can contain macros and preprocessor directives (to be preprocessed).
filename.c
indicates a C source file that can contain macros and preprocessor directives (to be preprocessed).
filename.C
indicates a C++ source file that can contain macros and preprocessor directives (to be preprocessed).
filename.i
indicates a preprocessed C or C++ source file.
filename.cc
indicates a C++ source file that can contain macros and preprocessor directives (to be preprocessed).
filename.cpp
indicates a C++ source file that can contain macros and preprocessor directives (to be preprocessed).
filename.s
indicates an assembly-language file.
filename.o
(Linux and macOS) indicates an object file.
filename.obj
(Windows systems only) indicates an object file.
filename.a
(Linux and macOS) indicates a library of object files.
filename.lib
(Windows systems only) indicates a statically-linked library of object files or an import library.
filename.so
(Linux only) indicates a library of shared object files.
filename.dll
(Windows systems only) indicates a dynamically-linked library.
filename.dylib
(macOS systems only) indicates a dynamically-linked library.

The driver passes files with .s extensions to the assembler and files with .o, .obj, .so, .dll, .a, and .lib extensions to the linker. Input files with unrecognized extensions, or no extension, are also passed to the linker.

Files with a .F (Capital F) or .FOR suffix are first preprocessed by the Fortran compilers and the output is passed to the compilation phase. The Fortran preprocessor functions like cpp for C programs, but is built in to the Fortran compilers rather than implemented through an invocation of cpp. This design ensures consistency in the preprocessing step regardless of the type or revision of operating system under which you are compiling.

Any input files not needed for a particular phase of processing are not processed. For example, if on the command line you specify an assembly-language file (filename.s) and the -⁠S option to stop before the assembly phase, the compiler takes no action on the assembly language file. Processing stops after compilation and the assembler does not run. In this scenario, the compilation must have been completed in a previous pass which created the .s file. For a complete description of the -⁠S option, refer to Output Files.

In addition to specifying primary input files on the command line, code within other files can be compiled as part of include files using the INCLUDE statement in a Fortran source file or the preprocessor #include directive in Fortran source files that use a .F extension or C and C++ source files.

When linking a program with a library, the linker extracts only those library components that the program needs. The compiler drivers link in several libraries by default. For more information about libraries, refer to Create and Use Libraries.

1.4.2. Output Files

By default, an executable output file produced by one of the PGI compilers is placed in the file a.out, or, on Windows, in a filename based on the name of the first source or object file on the command line. As the Hello example shows, you can use the -⁠o option to specify the output file name.

If you use option -⁠F (Fortran only), -⁠P (C/C++ only), -⁠S or -⁠c, the compiler produces a file containing the output of the last completed phase for each input file, as specified by the option supplied.

The output file is a preprocessed source file, an assembly-language file, or an unlinked object file respectively. Similarly, the -⁠E option does not produce a file, but displays the preprocessed source file on the standard output. Using any of these options, the -⁠o option is valid only if you specify a single input file. If no errors occur during processing, you can use the files created by these options as input to a future invocation of any of the PGI compiler drivers.

The following table lists the stop-after options and the output files that the compilers create when you use these options. It also indicates the accepted input files.

Table 2. Option Descriptions
Option Stop After Input Output
-E preprocessing Source files preprocessed file to standard out
-F preprocessing Source files. This option is not valid for pgcc or pgc⁠+⁠+. preprocessed file (.f)
-P preprocessing Source files. This option is not valid for pgf77, pgf95, orpgfortran. preprocessed file (.i)
-S compilation Source files or preprocessed files assembly-language file (.s)
-c assembly Source files, or preprocessed files, or assembly-language files unlinked object file (.o or .obj)
none linking Source files, or preprocessed files, assembly-language files, object files, or libraries executable file (a.out or .exe)

If you specify multiple input files or do not specify an object filename, the compiler uses the input filenames to derive corresponding default output filenames of the following form, where filename is the input filename without its extension:

filename.f
indicates a preprocessed file, if you compiled a Fortran file using the -⁠F option.
filename.i
indicates a preprocessed file, if you compiled using the -⁠P option.
filename.lst
indicates a listing file from the -⁠Mlist option.
filename.o or filename.obj
indicates a object file from the -⁠c option.
filename.s
indicates an assembly-language file from the -⁠S option.
Note: Unless you specify otherwise, the destination directory for any output file is the current working directory. If the file exists in the destination directory, the compiler overwrites it.

The following example demonstrates the use of output filename extensions.

$ pgfortran -c proto.f proto1.F 

This produces the output files proto.o and proto1.o, or, on Windows, proto.obj and proto1.obj, all of which are binary object files. Prior to compilation, the file proto1.F is preprocessed because it has a .F filename extension.

1.5. Fortran, C, and C++ Data Types

The PGI Fortran, C, and C++ compilers recognize scalar and aggregate data types. A scalar data type holds a single value, such as the integer value 42 or the real value 112.6. An aggregate data type consists of one or more scalar data type objects, such as an array of integer values.

For information about the format and alignment of each data type in memory, and the range of values each type can have on x64 processor-based systems, refer to ‘Fortran, C, and C++ Data Types’ section of the PGI Compiler Reference Manual.

For more information on x86-64-specific data representation, refer to the System V Application Binary Interface Processor Supplement by AT&T UNIX System Laboratories, Inc. (Prentice Hall, Inc.) listed in the ‘Related Publications’ section in the Preface.

For more information on x64 processor-based systems and the application binary interface (ABI) for those systems, see http://www.x86-64.org/documentation/abi.pdf.

1.6. Parallel Programming Using the PGI Compilers

The PGI compilers support many styles of parallel programming:

  • Automatic shared-memory parallel programs compiled using the -⁠Mconcur option to pgf77, pgf95,pgfortran, pgcc or pgc⁠+⁠+. Parallel programs of this variety can be run on shared-memory parallel (SMP) systems such as dual-core or multi-processor workstations.
  • OpenMP shared-memory parallel programs compiled using the -⁠mp option to pgf77, pgf95,pgfortran, pgcc or pgc⁠+⁠+. Parallel programs of this variety can be run on SMP systems. Carefully coded user-directed parallel programs using OpenMP directives can often achieve significant speed-ups on dual-core workstations or large numbers of processors on SMP server systems. Using OpenMP contains complete descriptions of user-directed parallel programming.
  • Distributed computing using an MPI message-passing library for communication between distributed processes.
  • Accelerated computing using either a low-level model such as CUDA Fortran or a high-level model such as the PGI Accelerator model or OpenACC to target a many-core GPU or other attached accelerator.

The first two types of parallel programs are collectively referred to as SMP parallel programs.

On a single silicon die, today's CPUs incorporate two or more complete processor cores – functional units, registers, level 1 cache, level 2 cache, and so on. These CPUs are known as multicore processors. For purposes of threads or OpenMP parallelism, these cores function as two or more distinct processors. However, the processing cores are on a single chip occupying a single socket on a system motherboard. For purposes of PGI software licensing, a multicore processor is treated as a single CPU.

1.6.1. Run SMP Parallel Programs

When you execute an SMP parallel program, by default it uses only one processor. To run on more than one processor, set the NCPUS environment variable to the desired number of processors. For information on how to set environment variables, refer to Setting Environment Variables.

Note: If you set NCPUS to a number larger than the number of physical processors, your program may execute very slowly.

1.7. Platform-specific considerations

The following list are the platforms supported by the PGI compilers and tools:

  • 64-bit Linux – supported on 64-bit Linux operating systems running on an x86-64 compatible processor.
  • 64-bit Windows – supported on 64-bit Windows operating systems running an x64-compatible processor.
  • 64-bit macOS – supported on 64-bit macOS operating systems running on a 64-bit Intel-based Mac system.

The following sections describe the specific considerations required to use the PGI compilers on the various platforms: Linux, Windows, and macOS.

1.7.1. Using the PGI Compilers on Linux

Linux Header Files

The Linux system header files contain many GNU gcc extensions. PGI supports many of these extensions, thus allowing the PGI C and C++ compilers to compile most programs that the GNU compilers can compile. A few header files not interoperable with the PGI compilers have been rewritten. These files are included in $PGI/linux86-64/include and in $PGI/linux86-64/include --⁠⁠gcc*, such as sigset.h, asm/byteorder.h, stddef.h, asm/posix_types.h and others. Also, PGI’s version of stdarg.h supports changes in newer versions of Linux.

If you are using the PGI C or C++ compilers, please make sure that the supplied versions of these include files are found before the system versions. This hierarchy happens by default unless you explicitly add a -⁠I option that references one of the system include directories.

Running Parallel Programs on Linux

You may encounter difficulties running auto-parallel or OpenMP programs on Linux systems when the per-thread stack size is set to the default (2MB). If you have unexplained failures, please try setting the environment variable OMP_STACKSIZE to a larger value, such as 8MB. For information on setting environment variables, refer to Setting Environment Variables.

If your program is still failing, you may be encountering the hard 8 MB limit on main process stack sizes in Linux. You can work around the problem by issuing the following command:

In csh:

	% limit stacksize unlimited

In bash, sh, zsh, or ksh, use:

	$ ulimit -s unlimited

1.7.2. Using the PGI Compilers on Windows

PGI on the Windows Start Menu

PGI provides a Start menu entry that provides access to different versions of PGI command shells as well as easy access to the PGI Debugger, the PGI Profiler, documentation, and licensing. The following sections provide a quick overview of the menu selections.

To access the main PGI menu, from the Start menu, select Start | All Programs | PGI.

Command Shell Submenus

From the PGI menu, you have access to PGI command shells for each version of PGI installed on your system. For example, if you have both PGI 14.1 and PGI 13.9 installed, then you have a submenu for each of these versions.

The PGI submenus for each version include the following:

  • PGI Bash (64)—Select this option to launch a Cygwin bash shell in which the environment is pre-initialized to use the 64-bit PGI compilers and tools. The default environment variables are already set and available. (Available only on x64 systems with Cygwin installed.)
  • PGI Cmd (64)—Select this option to launch a Microsoft command shell in which the environment is pre-initialized to use the 64-bit PGI compilers and tools. The default environment variables are already set and available. (Available only on x64 systems.)

The command window launched by PGI menu command or desktop icon can be customized using the "Properties" selection on the menu accessible by right-clicking the window's title bar.

Debugger & Profiler Submenu

From the Debugger & Profiler menu, you have access to the PGI debugging and profiling tools. The PGI debugger is a symbolic debugger for Fortran, C and assembly language programs. It provides debugger features, such as execution control using breakpoints, single-stepping, and examination and modification of application variables, memory locations, and registers.

  • PGI Debugger—Select this option to launch the PGI debugger.
  • PGI Profiler—Select this option to launch the PGI performance profiler. The profiler provides a way to visualize and diagnose the performance of the components of your program, and provides features for helping you to understand why certain parts of your program have high execution times.

Documentation Submenu

All PGI documentation is available online. The Documentation submenu contains a link to the online location.

Licensing Submenu

From the Licensing menu, you have access to the PGI License Agreement and an automated license generating tool:

  • Generate License—Select this option to display the PGI License Setup dialog that walks you through the steps required to download and install a PGI license. To complete this process you need an Internet connection.
  • License Agreement—Select this option to display the license agreement that is associated with use of PGI software.

1.7.3. PGI on the Windows Desktop

By default, a PGI installation creates a shortcut on the Windows desktop. This shortcut launches a Cygwin bash shell if Cygwin is installed; otherwise it launches a Microsoft command shell. The environment for this shell is pre-configured to use PGI compilers and tools.

BASH Shell Environment (Cygwin)

A UNIX-like shell environment, Cygwin, is bundled with PGI compilers and tools for Windows to provide a familiar development environment for Linux or UNIX users.

After installation you'll have a PGI icon on your Windows desktop. Double-left-click on this icon to launch an instance of the Cygwin bash command shell window. Working within BASH is very much like working within the sh or ksh shells on a Linux system; yet BASH has a command history feature similar to csh and several other unique features. Shell programming is fully supported.

The BASH shell window is pre-initialized for usage of the PGI compilers and tools, so there is no need to set environment variables or modify your command path when the command window comes up. In addition to the PGI compiler commands, within BASH you have access to over 100 common commands and utilities, including but not limited to the following:

vi gzip / gunzip ftp
tar / untar grep / egrep / fgrep awk
sed cksum cp
cat diff du
date kill ls
find mv printenv / env
more / less touch wc
rm / rmdir make  

If you are familiar with program development in a Linux environment, editing, compiling, and executing programs within bash will be very comfortable. If you have not previously used such an environment, you might want to familiarize yourself with vi or other editors and with makefiles. The Web has an extensive online tutorial available for the vi editor as well as a number of thorough introductions to the construction and use of makefiles.

ar or ranlib

For library compatibility, PGI provides versions of ar and ranlib that are compatible with native Windows object-file formats. For more information on these commands, refer to Creating and Using Static Libraries on Windows.

1.7.4. Using the PGI Compilers on macOS

PGI products for macOS support most of the features of the 32-and 64-bit versions for Linux environments. Typically the PGI compilers and tools on macOS function identically to their Linux counterparts.

macOS Header Files

The macOS header files contain numerous non-standard extensions. PGI supports many of these extensions, thus allowing the PGI C and C++ compilers to compile most programs that the GNU compilers can compile. A few header files not interoperable with the PGI compilers have been rewritten. These files are included in $PGI/osx86/17.7/include or $PGI/osx86-64/17.7/include. These files are: stdarg.h, stddef.h, and others.

If you are using the PGI C or C++ compilers, please make sure that the supplied versions of these include files are found before the system versions. This will happen by default unless you explicitly add a -⁠I option that references one of the system include directories.

Apple macOS Debugging Requirements

Both the -⁠g and -⁠Mkeepobj switches play important roles when compiling a program on macOS for debugging.

  • To debug a program with symbol information on macOS, files must be compiled with the -⁠g switch to keep the program's object files, the files with a ".o" extension. Further, these object files must remain in the same directory in which they were created.
  • If a program is built with separate compile and link steps, by compiling with the -⁠c switch which generates the ".o" object files, then using the -⁠g switch guarantees the required object files are available for debugging.

Use the following command sequence to compile and then link your code.

To compile the programs, use these commands:

pgcc -c -g main.c
pgcc -c -g foo.c
pgcc -c -g bar.c

To link, use this command:

pgcc -g main.o foo.o bar.o

Linking on macOS

On macOS, the PGI compilers do not support static linking of user binaries. For compatibility with future Apple updates, the compilers support dynamic linking of user binaries. For more information on dynamic linking, refer to Creating and Using Dynamic Libraries on Apple macOX.

Running Parallel Programs on macOS

You may encounter difficulties running auto-parallel or OpenMP programs on macOS systems when the per-thread stack size is set to the default (8MB). If you have unexplained failures, please try setting the environment variable OMP_STACKSIZE to a larger value, such as 16MB. For information on how to set environment variables, refer to Setting Environment Variables.

1.8. Site-Specific Customization of the Compilers

If you are using the PGI compilers and want all your users to have access to specific libraries or other files, there are special files that allow you to customize the compilers for your site.

1.8.1. Use siterc Files

The PGI compiler drivers utilize a file named siterc to enable site-specific customization of the behavior of the PGI compilers. The siterc file is located in the bin subdirectory of the PGI installation directory. Using siterc, you can control how the compiler drivers invoke the various components in the compilation tool chain.

1.8.2. Using User rc Files

In addition to the siterc file, user rc files can reside in a given user's home directory, as specified by the user's HOME environment variable. You can use these files to control the respective PGI compilers. All of these files are optional.

On Linux and macOS, these files are named .mypgf77rc,.mypgf90rc, .mypgccrc, and .mypgc++rc.

On Windows, these files are named mypgf77rc, mypgf90rc, mypgf95rc, mypgfortranrc and mypgccrc.

The following examples show how you can use these rc files to tailor a given installation for a particular purpose.

Table 3. Examples of Usine siterc and User rc Files
To do this... Add the line shown to the indicated file(s)
Make available to all linux86-64 compilations the libraries found in /⁠opt/newlibs/64 set SITELIB=/opt/newlibs/64;
to /opt/pgi/linux86-64/17.7/bin/siterc
Add to all linux86-64 compilations a new library path: /opt/local/fast append SITELIB=/opt/local/fast;
to /opt/pgi/linux86-64/17.7/bin/siterc
With linux86-64 compilations, change -⁠Mmpi to link in /opt/mympi/64/libmpix.a set MPILIBDIR=/opt/mympi/64;
set MPILIBNAME=mpix;
to /opt/pgi/linux86-64/17.7/bin/siterc
With linux86-64 compilations, always add -⁠DIS64BIT -⁠DAMD set SITEDEF=IS64BIT AMD;
to /opt/pgi/linux86-64/17.7/bin/siterc
Build an F90 or F95 executable for linux86-64 or linux86 that resolves PGI shared objects in the relative directory ./REDIST set set RPATH=./REDIST;
to ~/.mypgfortranrc
Note.This only affects the behavior of PGFORTRAN for the given user.

1.9. Common Development Tasks

Now that you have a brief introduction to the compiler, let's look at some common development tasks that you might wish to perform.

  • When you compile code you can specify a number of options on the command line that define specific characteristics related to how the program is compiled and linked, typically enhancing or overriding the default behavior of the compiler. For a list of the most common command line options and information on all the command line options, refer to Use Command-line Options.
  • Code optimization and parallelization allows the compiler to organize your code for efficient execution. While possibly increasing compilation time and making the code more difficult to debug, these techniques typically produce code that runs significantly faster than code that does not use them. For more information on optimization and parallelization, refer to Optimizing and Parallelizing.
  • Function inlining, a special type of optimization, replaces a call to a function or a subroutine with the body of the function or subroutine. This process can speed up execution by eliminating parameter passing and the function or subroutine call and return overhead. In addition, function inlining allows the compiler to optimize the function with the rest of the code. However, function inlining may also result in much larger code size with no increase in execution speed. For more information on function inlining, refer to Using Function Inlining.
  • Directives and pragmas allow users to place hints in the source code to help the compiler generate better assembly code. You typically use directives and pragmas to control the actions of the compiler in a particular portion of a program without affecting the program as a whole. You place them in your source code where you want them to take effect. A directive or pragma typically stays in effect from the point where included until the end of the compilation unit or until another directive or pragma changes its status. For more information on directives and pragmas, refer to Using OpenMP and Using Directives and Pragmas.
  • A library is a collection of functions or subprograms used to develop software. Libraries contain "helper" code and data, which provide services to independent programs, allowing code and data to be shared and changed in a modular fashion. The functions and programs in a library are grouped for ease of use and linking. When creating your programs, it is often useful to incorporate standard libraries or proprietary ones. For more information on this topic, refer to Creating and Using Libraries.
  • Environment variables define a set of dynamic values that can affect the way running processes behave on a computer. It is often useful to use these variables to set and pass information that alters the default behavior of the PGI compilers and the executables which they generate. For more information on these variables, refer to Using Environment Variables.
  • Deployment, though possibly an infrequent task, can present some unique issues related to concerns of porting the code to other systems. Deployment, in this context, involves distribution of a specific file or set of files that are already compiled and configured. The distribution must occur in such a way that the application executes accurately on another system which may not be configured exactly the same as the system on which the code was created. For more information on what you might need to know to successfully deploy your code, refer to Distributing Files – Deployment.
  • An intrinsic is a function available in a given language whose implementation is handled specially by the compiler. Intrinsics make using processor-specific enhancements easier because they provide a C/C++ language interface to assembly instructions. In doing so, the compiler manages details that the user would normally have to be concerned with, such as register names, register allocations, and memory locations of data. For C/C++ programs, PGI provides support for MMX SSE, SSE2, SSE3, SSSE3, SSE4A, ABM, and AVX intrinsics. For more information on these intrinsics, refer to the C/C++ MMX/SSE Inline Intrinsics section of the PGI Compiler Reference Manual.

2. Use Command-line Options

A command line option allows you to control specific behavior when a program is compiled and linked. This section describes the syntax for properly using command-line options and provides a brief overview of a few of the more common options.

Note: For a complete list of command-line options, their descriptions and use, refer to the ‘Command-Line Options Reference’ section of the PGI Compiler Reference Manual.

2.1. Command-line Option Overview

Before looking at all the command-line options, first become familiar with the syntax for these options. There are a large number of options available to you, yet most users only use a few of them. So, start simple and progress into using the more advanced options.

By default, the PGI compilers generate code that is optimized for the type of processor on which compilation is performed, the compilation host. Before adding options to your command-line, review Help with Command-line Options and Frequently-used Options.

2.1.1. Command-line Options Syntax

On a command-line, options need to be preceded by a hyphen (-⁠). If the compiler does not recognize an option, you get an unknown switch error. The error can be downgraded to a warning by adding the -noswitcherror option.

This document uses the following notation when describing options:

[item]

Square brackets indicate that the enclosed item is optional.

{item | item}

Braces indicate that you must select one and only one of the enclosed items. A vertical bar (|) separates the choices.

...

Horizontal ellipses indicate that zero or more instances of the preceding item are valid.

Note: Some options do not allow a space between the option and its argument or within an argument. When applicable, the syntax section of the option description in the ‘Command-Line Options Reference’ section of the PGI Compiler Reference Manual contains this information.

2.1.2. Command-line Suboptions

Some options accept several suboptions. You can specify these suboptions either by using the full option statement multiple times or by using a comma-separated list for the suboptions.

The following two command lines are equivalent:

pgfortran -Mvect=simd -Mvect=noaltcode
pgfortran -Mvect=simd,noaltcode

2.1.3. Command-line Conflicting Options

Some options have an opposite or negated counterpart. For example, both -⁠Mvect and -⁠Mnovect are available. -⁠Mvect enables vectorization and -⁠Mnovect disables it. If you used both of these commands on a command line, they would conflict.

Note:Rule: When you use conflicting options on a command line, the last encountered option takes precedence over any previous one.

The conflicting options rule is important for a number of reasons.

  • Some options, such as -⁠fast, include other options. Therefore, it is possible for you to be unaware that you have conflicting options.
  • You can use this rule to create makefiles that apply specific flags to a set of files, as shown in the following example.

Example: Makefiles with Options

In this makefile fragment, CCFLAGS uses vectorization. CCNOVECTFLAGS uses the flags defined for CCFLAGS but disables vectorization.

CCFLAGS=c -Mvect=simd
CCNOVECTFLAGS=$(CCFLAGS) -Mnovect

2.2. Help with Command-line Options

If you are just getting started with the PGI compilers and tools, it is helpful to know which options are available, when to use them, and which options most users find effective.

Using -help

The -⁠help option is useful because it provides information about all options supported by a given compiler.

You can use -⁠help in one of three ways:

  • Use -⁠help with no parameters to obtain a list of all the available options with a brief one-line description of each.
  • Add a parameter to -⁠help to restrict the output to information about a specific option. The syntax for this usage is:

    -help <command line option>

    Suppose you use the following command to restrict the output to information about the -⁠fast option:

    $ pgfortran -help -fast

    The output you see is similar to:

    -fast Common optimizations; includes -O2 -Munroll=c:1 -Mnoframe -Mlre

    In the following example, we add the -⁠help parameter to restrict the output to information about the help command. The usage information for -⁠help shows how groups of options can be listed or examined according to function.

    $ pgfortran -help -help
    	-help[=groups|asm|debug|language|linker|opt|other|overall|phase|prepro|
             suffix|switch|target|variable]
  • Add a parameter to -⁠help to restrict the output to a specific set of options or to a building process. The syntax for this usage is this:

    -help=<subgroup>

    By using the command pgfortran -⁠help -⁠help, as previously shown, we can see output that shows the available subgroups. You can use the following command to restrict the output on the -⁠help command to information about only the options related to only one group, such as debug information generation.

    $ pgfortran -help=debug

    The output you see is similar to this:

    Debugging switches:
    -M[no]bounds Generate code to check array bounds
    -Mchkstk Check for sufficient stack space upon subprogram entry
    -Mcoff Generate COFF format object
    -Mdwarf1 Generate DWARF1 debug information with -g
    -Mdwarf2 Generate DWARF2 debug information with -g
    -Mdwarf3 Generate DWARF3 debug information with -g
    -Melf Generate ELF format object
    -g Generate information for debugger
    -gopt Generate information for debugger without disabling
    optimizations 

    For a complete description of subgroups, refer to the -⁠help description in the Command-line Options Reference section of the PGI Compiler Reference Manual.

2.3. Getting Started with Performance

One of the top priorities of most users is performance and optimization. This section provides a quick overview of a few of the command-line options that are useful in improving performance.

2.3.1. Using -fast

PGI compilers implement a wide range of options that allow users a fine degree of control on each optimization phase. When it comes to optimization of code, the quickest way to start is to use the option -⁠fast. These options create a generally optimal set of flags. They incorporate optimization options to enable use of vector streaming SIMD instructions for 64-bit targets. They enable vectorization with SIMD instructions, cache alignment, and flush to zero mode.

Note: The contents of the -⁠fast option are host-dependent. Further, you should use these options on both compile and link command lines.

The following table shows the typical -⁠fast options.

Table 4. Typical -⁠fast Options
Use this option... To do this...
-O2 Specifies a code optimization level of 2.
-Munroll=c:1 Unrolls loops, executing multiple instances of the original loop during each iteration.
-Mnoframe Indicates to not generate code to set up a stack frame.
Note. With this option, a stack trace does not work.
-Mlre Indicates loop-carried redundancy elimination.
-Mpre Indicates partial redundancy elimination

-⁠fast typically includes the options shown in this table:

Table 5. Additional -⁠fast Options
Use this option... To do this...
-Mvect=simd Generates packed SIMD instructions.
-Mcache_align Aligns long objects on cache-line boundaries.
-Mflushz Sets flush-to-zero mode.
-M[no]vect Controls automatic vector pipelining.
Note: For best performance on processors that support SIMD instructions, use the PGFORTRAN compiler, even for FORTRAN 77 code, and the -⁠fast option.

To see the specific behavior of -⁠fast for your target, use the following command:

$ pgfortran -help -fast

2.3.2. Other Performance-Related Options

While -⁠fast is designed to be the quickest route to best performance, it is limited to routine boundaries. Depending on the nature and writing style of the source code, the compiler often can perform further optimization by knowing the global context of usage of a given routine. For instance, determining the possible value range of actual parameters of a routine could enable a loop to be vectorized; similarly, determining static occurrence of calls helps to decide which routine is beneficial to inline.

These types of global optimizations are under control of Interprocedural Analysis (IPA) in PGI compilers. Option -⁠Mipa enables Interprocedural Analysis. -⁠Mipa=fast is the recommended option to get best performances for global optimization. You can also add the suboption inline to enable automatic global inlining across files. You might consider using -⁠Mipa=fast,inline. This option for interprocedural analysis and global optimization can improve performance.

Note: As of the PGI 16.3 release, -⁠Mipa has been disabled on Windows.

You may also obtain further performance improvements by experimenting with the -⁠M<pgflag> options described in the section ‘-⁠M Options by Category’ section of the PGI Compiler Reference Manual . These options include, but are not limited to, -⁠Mvect, -⁠Munroll, -⁠Minline, -⁠Mconcur, -⁠Mpfi and -⁠Mpfo. However, performance improvements using these options are typically application- and system-dependent. It is important to time your application carefully when using these options to ensure no performance degradations occur.

For more information on optimization, refer to Optimizing and Parallelizing. For specific information about these options, refer to the ‘Optimization Controls’ section of the PGI Compiler Reference Manual .

2.4. Targeting Multiple Systems—Using the -⁠tp Option

The -⁠tp option allows you to set the target architecture. By default, the PGI compiler uses all supported instructions wherever possible when compiling on a given system. As a result, executables created on a given system may not be usable on previous generation systems. For example, executables created on a Pentium 4 may fail to execute on a Pentium III or Pentium II.

Processor-specific optimizations can be specified or limited explicitly by using the -⁠tp option. Thus, it is possible to create executables that are usable on previous generation systems. Using a -⁠tp flag option of k8 or p7 produces an executable that runs on most x86-64 hardware in use today.

For more information about the -⁠tp option, refer to the -⁠tp <target> [,target...] description in the ‘Command-Line Options Reference’ section of the PGI Compiler Reference Manual.

2.5. Frequently-used Options

In addition to overall performance, there are a number of other options that many users find useful when getting started. The following table provides a brief summary of these options.

For more information on these options, refer to the complete description of each option available in the ‘Command-Line Options Reference’ section of the PGI Compiler Reference Manual. Also, there are a number of suboptions available with each of the -⁠M options listed. For more information on those options, refer to the specific section on ‘M Options by Category’.

Table 6. Commonly Used Command-Line Options
Use this option... To do this...
-⁠fast
-⁠fastsse
These options create a generally optimal set of flags for targets that support SIMD capability. They incorporate optimization options to enable use of vector streaming SIMD instructions (64-bit targets) and enable vectorization with SIMD instructions, cache aligned and flushz.
-⁠g Instructs the compiler to include symbolic debugging information in the object module; sets the optimization level to zero unless a -⁠O option is present on the command line.
--⁠gopt Instructs the compiler to include symbolic debugging information in the object file, and to generate optimized code identical to that generated when -⁠g is not specified.
-⁠help Provides information about available options.
-⁠mcmodel=medium Enables medium=model core generation for 64-bit targets, which is useful when the data space of the program exceeds 4GB.
-⁠Mconcur Instructs the compiler to enable auto-concurrentization of loops. If specified, the compiler uses multiple processors to execute loops that it determines to be parallelizable; thus, loop iterations are split to execute optimally in a multithreaded execution context.
-⁠Minfo Instructs the compiler to produce information on standard error.
-⁠Minline Enables function inlining.
-⁠Mipa=fast,inline Enables interprocedural analysis and optimization. Also enables automatic procedure inlining.
-⁠Mpfi or -⁠Mpfo Enable profile feedback driven optimizations
-⁠Mkeepasm Keeps the generated assembly files.
-⁠Munroll Invokes the loop unroller to unroll loops, executing multiple instances of the loop during each iteration. This also sets the optimization level to 2 if the level is set to less than 2, or if no -⁠O or -⁠g options are supplied.
-⁠M[no]vect Enables/Disables the code vectorizer.
--⁠[no_]exceptions Removes exception handling from user code. For C++, declares that the functions in this file generate no C++ exceptions, allowing more optimal code generation.
-⁠o Names the output file.
-⁠O <level> Specifies code optimization level where <level> is 0, 1, 2, 3, or 4.
-⁠tp <target> [,target...] Specify the target processor(s); for the 64-bit compilers, more than one target is allowed, and enables generation of PGI Unified Binary executables.
-⁠Wl, <option> Compiler driver passes the specified options to the linker.

3. Optimizing and Parallelizing

Source code that is readable, maintainable, and produces correct results is not always organized for efficient execution. Normally, the first step in the program development process involves producing code that executes and produces the correct results. This first step usually involves compiling without much worry about optimization. After code is compiled and debugged, code optimization and parallelization become an issue.

Invoking one of the PGI compiler commands with certain options instructs the compiler to generate optimized code. Optimization is not always performed since it increases compilation time and may make debugging difficult. However, optimization produces more efficient code that usually runs significantly faster than code that is not optimized.

The compilers optimize code according to the specified optimization level. You can use a number of options to specify the optimization levels, including -⁠O, -⁠Mvect, -⁠Mipa, and -⁠Mconcur. In addition, you can use several of the -⁠M<pgflag> switches to control specific types of optimization and parallelization.

Note: As of the PGI 16.3 release, -⁠Mipa has been disabled on Windows.

This chapter describes these optimization options:

-fast -Minline -O -Munroll
-Mconcur -Mipa=fast -Mpfi -Mvect
-Minfo -Mneginfo -Mpfo -Msafeptr
-Mipa=fast,inline      
-fast -Minline -O -Munroll
-Mconcur -Mpfi -Mvect -Minfo
-Mneginfo -Mpfo    

This chapter also describes how to choose optimization options to use with the PGI compilers. This overview is helpful if you are just getting started with one of the PGI compilers, or wish to experiment with individual optimizations.

Complete specifications of each of these options is available in the Command-Line Options Reference section of the PGI Compiler Reference Manual .

3.1. Overview of Optimization

In general, optimization involves using transformations and replacements that generate more efficient code. This is done by the compiler and involves replacements that are independent of the particular target processor's architecture as well as replacements that take advantage of the x86-64 architecture, instruction set and registers.

For discussion purposes, we categorize optimization:

3.1.1. Local Optimization

A basic block is a sequence of statements in which the flow of control enters at the beginning and leaves at the end without the possibility of branching, except at the end. Local optimization is performed on a block-by-block basis within a program’s basic blocks.

The PGI compilers perform many types of local optimization including: algebraic identity removal, constant folding, common sub-expression elimination, redundant load and store elimination, scheduling, strength reduction, and peephole optimizations.

3.1.2. Global Optimization

This optimization is performed on a subprogram/function over all its basic blocks. The optimizer performs control-flow and data-flow analysis for an entire program unit. All loops, including those formed by ad hoc branches such as IFs or GOTOs, are detected and optimized.

Global optimization includes: constant propagation, copy propagation, dead store elimination, global register allocation, invariant code motion, and induction variable elimination.

3.1.3. Loop Optimization: Unrolling, Vectorization and Parallelization

The performance of certain classes of loops may be improved through vectorization or unrolling options. Vectorization transforms loops to improve memory access performance and make use of packed SSE instructions which perform the same operation on multiple data items concurrently. Unrolling replicates the body of loops to reduce loop branching overhead and provide better opportunities for local optimization, vectorization and scheduling of instructions. Performance for loops on systems with multiple processors may also improve using the parallelization features of the PGI compilers.

3.1.4. Interprocedural Analysis (IPA) and Optimization

Interprocedural analysis (IPA) allows use of information across function call boundaries to perform optimizations that would otherwise be unavailable. For example, if the actual argument to a function is in fact a constant in the caller, it may be possible to propagate that constant into the callee and perform optimizations that are not valid if the dummy argument is treated as a variable. A wide range of optimizations are enabled or improved by using IPA, including but not limited to data alignment optimizations, argument removal, constant propagation, pointer disambiguation, pure function detection, F90/F95 array shape propagation, data placement, empty function removal, automatic function inlining, inlining of functions from pre-compiled libraries, and interprocedural optimization of functions from pre-compiled libraries.

3.1.5. Function Inlining

This optimization allows a call to a function to be replaced by a copy of the body of that function. This optimization will sometimes speed up execution by eliminating the function call and return overhead. Function inlining may also create opportunities for other types of optimization. Function inlining is not always beneficial. When used improperly it may increase code size and generate less efficient code.

3.1.6. Profile-Feedback Optimization (PFO)

Profile-feedback optimization (PFO) makes use of information from a trace file produced by specially instrumented executables which capture and save information on branch frequency, function and subroutine call frequency, semi-invariant values, loop index ranges, and other input data dependent information that can only be collected dynamically during execution of a program.

By definition, use of profile-feedback optimization is a two-phase process: compilation and execution of a specially-instrumented executable, followed by a subsequent compilation which reads a trace file generated during the first phase and uses the information in that trace file to guide compiler optimizations.

3.2. Getting Started with Optimization

The first concern should be getting the program to execute and produce correct results. To get the program running, start by compiling and linking without optimization. Add -⁠O0 to the compile line to select no optimization; or add -⁠g to debug the program easily and isolate any coding errors exposed during porting to x86-64 platforms. For more information on debugging, refer to the Debugger User's Guide.

To get started quickly with optimization, a good set of options to use with any of the PGI compilers is -⁠fast . For example:

$ pgfortran -fast -Mipa=fast,inline prog.f
Note: As of the PGI 16.3 release,
-⁠Mipa
has been disabled on Windows.

For all of the PGI Fortran, C, and C++ compilers, the -⁠fast -⁠Mipa=fast,inline options generally produce code that is well-optimized without the possibility of significant slowdowns due to pathological cases.

  • The-⁠fast option is an aggregate option that includes a number of individual PGI compiler options; which PGI compiler options are included depends on the target for which compilation is performed.
  • The -⁠Mipa=fast,inline option invokes interprocedural analysis (IPA), including several IPA suboptions. The inline suboption enables automatic inlining with IPA. If you do not wish to use automatic inlining, you can compile with -⁠Mipa=fast and use several IPA suboptions without inlining.

By experimenting with individual compiler options on a file-by-file basis, further significant performance gains can sometimes be realized. However, depending on the coding style, individual optimizations can sometimes cause slowdowns, and must be used carefully to ensure performance improvements.

There are other useful command line options related to optimization and parallelization, such as -⁠help, -⁠Minfo, -⁠Mneginfo, -⁠dryrun, and -⁠v.

3.2.1. -help

As described in Help with Command-Line Options, you can see a specification of any command-line option by invoking any of the PGI compilers with -⁠help in combination with the option in question, without specifying any input files.

For example, you might want information on -⁠O:

$ pgfortran  -help -O

The resulting output is similar to this:

-O Set opt level. All -O1 optimizations plus traditional scheduling and
  global scalar optimizations performed

Or you can see the full functionality of -⁠help itself, which can return information on either an individual option or groups of options:

$ pgfortran  -help -help

The resulting output is similar to this:

-help[=groups|asm|debug|language|linker|opt|other|overall|
phase|prepro|suffix|switch|target|variable]
Show compiler switches

3.2.2. -Minfo

You can use the -⁠Minfo option to display compile-time optimization listings. When this option is used, the PGI compilers issue informational messages to standard error (stderr) as compilation proceeds. From these messages, you can determine which loops are optimized using unrolling, SSE instructions , vectorization, parallelization, interprocedural optimizations and various miscellaneous optimizations. You can also see where and whether functions are inlined.

For more information on -⁠Minfo, refer to ‘Optimization Controls’ section of the PGI Compiler Reference Manual.

3.2.3. -Mneginfo

You can use the -⁠Mneginfo option to display informational messages to standard error (stderr) that explain why certain optimizations are inhibited.

For more information on -⁠Mneginfo, refer to ‘Optimization Controls’ section of the PGI Compiler Reference Manual.

3.2.4. -dryrun

The -⁠dryrun option can be useful as a diagnostic tool if you need to see the steps used by the compiler driver to preprocess, compile, assemble and link in the presence of a given set of command line inputs. When you specify the -⁠dryrun option, these steps are printed to standard error (stderr) but are not actually performed. For example, you can use this option to inspect the default and user-specified libraries that are searched during the link phase, and the order in which they are searched by the linker.

3.2.5. -v

The -⁠v option is similar to -⁠dryrun, except each compilation step is performed and not simply printed.

3.2.6. PGI Profiler

The PGI profiler is a profiling tool that provides a way to visualize the performance of the components of your program. Using tables and graphs, the profiler associates execution time and resource utilization data with the source code and instructions of your program. This association allows you to see where a program’s execution time is spent. Through resource utilization data and compiler analysis information, the profiler helps you to understand why certain parts of your program have high execution times. This information may help you with selecting which optimization options to use with your program.

The profiler also allows you to correlate the messages produced by -⁠Minfo and -⁠Mneginfo, described above, to your program’s source code. This feature is known as the Common Compiler Feedback Format.

For more information on the profiler, refer to the Profiler User's Guide.

3.3. Common Compiler Feedback Format (CCFF)

Using the Common Compiler Feedback Format (CCFF), PGI compilers save information about how your program was optimized, or why a particular optimization was not made, in the executable file. To append this information to the object file, use the compiler option -⁠Minfo=ccff.

If you choose to use the PGI profiler to aid with your optimization, it can extract this information and associate it with source code and other performance data, allowing you to view all of this information simultaneously in one of the available profiler panels.

3.4. Local and Global Optimization

This section describes local and global optimization.

3.4.1. -Msafeptr

The -⁠Msafeptr option can significantly improve performance of C/C++ programs in which there is known to be no pointer aliasing. For obvious reasons, this command-line option must be used carefully. There are a number of suboptions for -⁠Msafeptr:

  • -Msafeptr=all – All pointers are safe. Equivalent to the default setting: -⁠Msafeptr.
  • -Msafeptr=arg – Function formal argument pointers are safe. Equivalent to -⁠Msafeptr=dummy.
  • -Msafeptr=global – Global pointers are safe.
  • -Msafeptr=local – Local pointers are safe. Equivalent to -⁠Msafeptr=auto.
  • -Msafeptr=static – Static local pointers are safe.

If your C/C++ program has pointer aliasing and you also want automating inlining, then compiling with -⁠Mipa=fast or -⁠Mipa=fast,inline includes pointer aliasing optimizations. IPA may be able to optimize some of the alias references in your program and leave intact those that cannot be safely optimizied.

3.4.2. -O

Using the PGI compiler commands with the -⁠O<level> option (the capital O is for Optimize), you can specify any integer level from 0 to 4.

-O0

Level zero specifies no optimization. A basic block is generated for each language statement. At this level, the compiler generates a basic block for each statement.

Performance will almost always be slowest using this optimization level. This level is useful for the initial execution of a program. It is also useful for debugging, since there is a direct correlation between the program text and the code generated. To enable debugging, include -⁠g on your compile line.

-O1

Level one specifies local optimization. Scheduling of basic blocks is performed. Register allocation is performed.

Local optimization is a good choice when the code is very irregular, such as code that contains many short statements containing IF statements and does not contain loops (DO or DO WHILE statements ). Although this case rarely occurs, for certain types of code, this optimization level may perform better than level-two (-⁠O2).

-O

When no level is specified, level two global optimizations are performed, including traditional scalar optimizations, induction recognition, and loop invariant motion. No SIMD vectorization is enabled.

-O2

Level two specifies global optimization. This level performs all level-one local optimization as well as level two global optimization described in -⁠O. In addition, more advanced optimizations such as SIMD code generation, cache alignment, and partial redundancy elimination are enabled.

-O3

Level three specifies aggressive global optimization. This level performs all level-one and level-two optimizations and enables more aggressive hoisting and scalar replacement optimizations that may or may not be profitable.

-O4

Level four performs all level-one, level-two, and level-three optimizations and enables hoisting of guarded invariant floating point expressions.

Types of Optimizations

The PGI compilers perform many different types of local optimizations, including but not limited to:

  • Algebraic identity removal
  • Constant folding
  • Common subexpression elimination
  • Local register optimization
  • Peephole optimizations
  • Redundant load and store elimination
  • Strength reductions

Level-two optimization (-⁠O2 or -⁠O) specifies global optimization. The -⁠fast option generally specifies global optimization; however, the -⁠fast switch varies from release to release, depending on a reasonable selection of switches for any one particular release. The -⁠O or -⁠O2 level performs all level-one local optimizations as well as global optimizations. Control flow analysis is applied and global registers are allocated for all functions and subroutines. Loop regions are given special consideration. This optimization level is a good choice when the program contains loops, the loops are short, and the structure of the code is regular.

The PGI compilers perform many different types of global optimizations, including but not limited to:

  • Branch to branch elimination
  • Constant propagation
  • Copy propagation
  • Dead store elimination
  • Global register allocation
  • Induction variable elimination
  • Invariant code motion

You can explicitly select the optimization level on the command line. For example, the following command line specifies level-two optimization which results in global optimization:

$ pgfortran  -O2 prog.f

The default optimization level changes depending on which options you select on the command line. For example, when you select the -⁠g debugging option, the default optimization level is set to level-zero (-⁠O0). However, if you need to debug optimized code, you can use the -⁠gopt option to generate debug information without perturbing optimization. For a description of the default levels, refer to Default Optimization Levels.

The -⁠fast option includes -⁠O2 on all targets. If you want to override the default for -⁠fast with -⁠O3 while maintaining all other elements of -⁠fast, simply compile as follows:

$ pgfortran  -fast -O3 prog.f

3.5. Loop Unrolling using -Munroll

This optimization unrolls loops, which reduces branch overhead, and can improve execution speed by creating better opportunities for instruction scheduling. A loop with a constant count may be completely unrolled or partially unrolled. A loop with a non-constant count may also be unrolled. A candidate loop must be an innermost loop containing one to four blocks of code.

The following example shows the use of the -⁠Munroll option:

$ pgfortran -Munroll prog.f

The -⁠Munroll option is included as part of -⁠fast on all targets. The loop unroller expands the contents of a loop and reduces the number of times a loop is executed. Branching overhead is reduced when a loop is unrolled two or more times, since each iteration of the unrolled loop corresponds to two or more iterations of the original loop; the number of branch instructions executed is proportionately reduced. When a loop is unrolled completely, the loop’s branch overhead is eliminated altogether.

Loop unrolling may be beneficial for the instruction scheduler. When a loop is completely unrolled or unrolled two or more times, opportunities for improved scheduling may be presented. The code generator can take advantage of more possibilities for instruction grouping or filling instruction delays found within the loop.

Examples Showing Effect of Unrolling

The following side-by-side examples show the effect of code unrolling on a segment that computes a dot product.

Note: This example is only meant to represent how the compiler can transform the loop; it is not meant to imply that the programmer needs to manually change code. In fact, manually unrolling your code can sometimes inhibit the compiler’s analysis and optimization.
Table 7. Example of Effect of Code Unrolling
Dot Product Code Unrolled Dot Product Code
 REAL*4 A(100), B(100), Z
  INTEGER I
  DO I=1, 100
    Z = Z + A(i) * B(i)
  END DO
 END
 REAL*4 A(100), B(100), Z
  INTEGER I
  DO I=1, 100, 2
    Z = Z + A(i) * B(i)
    Z = Z + A(i+1) * B(i+1)
  END DO
 END

Using the -⁠Minfo option, the compiler informs you when a loop is being unrolled. For example, a message similar to the following, indicating the line number, and the number of times the code is unrolled, displays when a loop is unrolled:

dot:
  5, Loop unrolled 5 times

Using the c:<m> and n:<m> sub-options to -⁠Munroll, or using -⁠Mnounroll, you can control whether and how loops are unrolled on a file-by-file basis. Using directives or pragmas, you can precisely control whether and how a given loop is unrolled. For more information on -⁠Munroll, refer to Use Command-line Options.

3.6. Vectorization using -Mvect

The -⁠Mvect option is included as part of -⁠fast on all targets. If your program contains computationally-intensive loops, the -⁠Mvect option may be helpful. If in addition you specify -⁠Minfo, and your code contains loops that can be vectorized, the compiler reports relevant information on the optimizations applied.

When a PGI compiler command is invoked with the -⁠Mvect option, the vectorizer scans code searching for loops that are candidates for high-⁠level transformations such as loop distribution, loop interchange, cache tiling, and idiom recognition (replacement of a recognizable code sequence, such as a reduction loop, with optimized code sequences or function calls). When the vectorizer finds vectorization opportunities, it internally rearranges or replaces sections of loops (the vectorizer changes the code generated; your source code’s loops are not altered). In addition to performing these loop transformations, the vectorizer produces extensive data dependence information for use by other phases of compilation and detects opportunities to use vector or packed Streaming SIMD Extensions (SSE) instructions on x86 processors where these are supported.

The -⁠Mvect option can speed up code which contains well-behaved countable loops which operate on large floating point arrays in Fortran and their C/C++ counterparts. However, it is possible that some codes will show a decrease in performance when compiled with the -⁠Mvect option due to the generation of conditionally executed code segments, inability to determine data alignment, and other code generation factors. For this reason, it is recommended that you check carefully whether particular program units or loops show improved performance when compiled with this option enabled.

3.6.1. Vectorization Sub-options

The vectorizer performs high-level loop transformations on countable loops. A loop is countable if the number of iterations is set only before loop execution and cannot be modified during loop execution. Some of the vectorizer transformations can be controlled by arguments to the -⁠Mvect command line option. The following sections describe the arguments that affect the operation of the vectorizer. In addition, some of these vectorizer operations can be controlled from within code using directives and pragmas. For details on the use of directives and pragmas, refer to Using Directives and Pragmas.

The vectorizer performs the following operations:

  • Loop interchange
  • Loop splitting
  • Loop fusion
  • Memory-hierarchy (cache tiling) optimizations
  • Generation of SSE instructions on processors where these are supported
  • Generation of prefetch instructions on processors where these are supported
  • Loop iteration peeling to maximize vector alignment
  • Alternate code generation

By default, -⁠Mvect without any sub-options is equivalent to:

-Mvect=assoc,cachesize=c

where c is the actual cache size of the machine.

This enables the options for nested loop transformation and various other vectorizer options. These defaults may vary depending on the target system. The following table lists and briefly describes some of the -⁠Mvect suboptions.

Table 8. -Mvect Suboptions
Use this option ... To instruct the vectorizer to do this ...
-Mvect=altcode Generate appropriate code for vectorized loops.
-Mvect=[no]assoc Perform[disable] associativity conversions that can change the results of a computation due to a round-off error. For example, a typical optimization is to change one arithmetic operation to another arithmetic operation that is mathematically correct, but can be computationally different and generate faster code. This option is provided to enable or disable this transformation, since a round-off error for such associativity conversions may produce unacceptable results.
-Mvect=cachesize:n Tiles nested loop operations, assuming a data cache size of n bytes. By default, the vectorizer attempts to tile nested loop operations, such as matrix multiply, using multi-dimensional strip-mining techniques to maximize re-use of items in the data cache.
-Mvect=fuse Enable loop fusion.
-Mvect=gather Enable vectorization of indirect array references.
-Mvect=idiom Enable idiom recognition.
-Mvect=levels:<n> Set the maximum next level of loops to optimize.
-Mvect=nocond Disable vectorization of loops with conditions.
-Mvect=partial Enable partial loop vectorization via inner loop distribution.
-Mvect=prefetch Automatically generate prefetch instructions when vectorizable loops are encountered, even in cases where SSE instructions are not generated.
-Mvect=short Enable short vector operations.
-Mvect=simd Automatically generate packed SSE (Streaming SIMD Extensions), and prefetch instructions when vectorizable loops are encountered. SIMD instructions, first introduced on Pentium III and AthlonXP processors, operate on single-precision floating-point data.
-Mvect=sizelimit:n Limit the size of vectorized loops.
-Mvect=sse Equivalent to -⁠Mvect=simd.
-Mvect=tile Enable loop tiling.
-Mvect=uniform Perform consistent optimizations in both vectorized and residual loops. Be aware that this may affect the performance of the residual loop.
Note: Inserting no in front of the option disables the option. For example, to disable the generation of SSE (or SIMD) instructions on x86, compile with -⁠Mvect=nosimd.

3.6.2. Vectorization Example Using SIMD Instructions

One of the most important vectorization options is -Mvect=simd. When you use this option, the compiler automatically generates SSE instructions, where possible, when targeting x86 processors on which these instructions are supported. This process can improve performance by several factors compared with the equivalent scalar code. All of the PGI Fortran, C and C++ compilers support this capability. The PGI Release Notes show which x86 and x64 processors PGI supports.

In the program in Vector operation using SIMD instructions, the vectorizer recognizes the vector operation in subroutine 'loop' when either the compiler switch -Mvect=simd or -fast is used. This example shows the compilation, informational messages, and runtime results using the SSE instructions on a x86 4 Core Intel Sandybridge 2.5 GHz system, along with issues that affect SSE performance.

Loops vectorized using x86SSE instructions operate much more efficiently when processing vectors that are aligned to a cache-line boundary. You can cause unconstrained data objects of size 16 bytes or greater to be cache-aligned by compiling with the -⁠Mcache_align switch. An unconstrained data object is a data object that is not a common block member and not a member of an aggregate data structure.

Note: For stack-based local variables to be properly aligned, the main program or function must be compiled with -⁠Mcache_align.

The -⁠Mcache_align switch has no effect on the alignment of Fortran allocatable or automatic arrays. If you have arrays that are constrained, such as vectors that are members of Fortran common blocks, you must specifically pad your data structures to ensure proper cache alignment. You can use -⁠Mcache_align for only the beginning address of each common block to be cache-aligned.

The following examples show the results of compiling the sample code in Vector operation using SIMD instructions both with and without the option -⁠Mvect=simd.

Vector operation using SIMD instructions

program vector_op
 parameter (N = 9999)
 real*4 x(N), y(N), z(N), W(N)
 do i = 1, n
   y(i) = i
   z(i) = 2*i
   w(i) = 4*i
 enddo
 do j = 1, 200000
   call loop(x,y,z,w,1.0e0,N)
 enddo
 print *, x(1),x(771),x(3618),x(6498),x(9999)
end
subroutine loop(a,b,c,d,s,n)
 integer i, n
 real*4 a(n), b(n), c(n), d(n),s
 do i = 1, n
   a(i) = b(i) + c(i) - s * d(i)
 enddo
end

Assume the preceding program is compiled as follows, where -Mvect=nosimd disables x86 SSE vectorization:

% pgfortran -fast -Mvect=nosimd -Minfo vadd.f -Mfree -o vadd
 vector_op:
 4, Loop unrolled 16 times
    Generates 1 prefetches in scalar loop
 9, Loop not vectorized/parallelized: contains call
 loop:
 18, Loop unrolled 4 times

The following output shows a sample result if the generated executable is run and timed on a x86 4 Core Intel Sandybridge 2.5 GHz system:

% /bin/time vadd
 -1.000000 -771.000 -3618.000 -6498.00 -9999.00
 1.35user 0.00system 0:01.35elapsed 99%CPU (0avgtext+0avgdata 3936maxresident)k
0inputs+0outputs (0major+290minor)pagefaults 0swaps

Now, recompile with x86 SSE vectorization enabled, and you see results similar to these:

% pgfortran -fast -Minfo vadd.f -Mfree -o vadd
vector_op:
      4, Loop not vectorized: may not be beneficial
         Unrolled inner loop 8 times
         Residual loop unrolled 7 times (completely unrolled)
         Generated 1 prefetches in scalar loop
      9, Loop not vectorized/parallelized: contains call
loop:
     17, Generated 4 alternate versions of the loop
         Generated vector sse code for the loop
         Generated 3 prefetch instructions for the loop

Notice the informational message for the loop at line 17.

  • The first two lines of the message indicate that the loop was vectorized, x86 SSE instructions were generated, and four alternate versions of the loop were also generated. The loop count and alignments of the arrays determine which of these versions is executed.
  • The last line of the informational message indicates that prefetch instructions have been generated for three loads to minimize latency of data transfers from main memory.

Executing again, you should see results similar to the following:

 % /bin/time vadd
 -1.000000 -771.000 -3618.00 -6498.00 -9999.0
 0.60user 0.00system 0:00.61elapsed 99%CPU (0avgtext+0avgdata 3920maxresident)k
0inputs+0outputs (0major+289minor)pagefaults 0swaps

The SIMD result is 2.25 times faster than the equivalent non-SIMD version of the program.

Speed-up realized by a given loop or program can vary widely based on a number of factors:

  • When the vectors of data are resident in the data cache, performance improvement using vector x86 SSE or SSE2 instructions is most effective.
  • If data is aligned properly, performance will be better in general than when using vector x86 SSE operations on unaligned data.
  • If the compiler can guarantee that data is aligned properly, even more efficient sequences of x86 SSE instructions can be generated.
  • The efficiency of loops that operate on single-precision data can be higher. x86 SSE2 vector instructions can operate on four single-precision elements concurrently, but only two double-precision elements.
Note: Compiling with -⁠Mvect=simd can result in numerical differences from the executables generated with less optimization. Certain vectorizable operations, for example dot products, are sensitive to order of operations and the associative transformations necessary to enable vectorization (or parallelization).

3.7. Auto-Parallelization using -Mconcur

With the -⁠Mconcur option the compiler scans code searching for loops that are candidates for auto-parallelization. -⁠Mconcur must be used at both compile-time and link-time. When the parallelizer finds opportunities for auto-parallelization, it parallelizes loops and you are informed of the line or loop being parallelized if the -⁠Minfo option is present on the compile line. For a complete specification of -⁠Mconcur, refer to the ‘Optimization Controls’ section of the PGI Compiler Reference Manual.

A loop is considered parallelizable if it doesn't contain any cross-iteration data dependencies. Cross-iteration dependencies from reductions and expandable scalars are excluded from consideration, enabling more loops to be parallelizable. In general, loops with calls are not parallelized due to unknown side effects. Also, loops with low trip counts are not parallelized since the overhead in setting up and starting a parallel loop will likely outweigh the potential benefits. In addition, the default is not to parallelize innermost loops, since these often by definition are vectorizable and it is seldom profitable to both vectorize and parallelize the same loop, especially on multicore processors. Compiler switches and directives are available to let you override most of these restrictions on auto-parallelization.

3.7.1. Auto-Parallelization Sub-options

The parallelizer performs various operations that can be controlled by arguments to the -⁠Mconcur command line option. The following sections describe these arguments that affect the operation of the parallelizer. In addition, these parallelizer operations can be controlled from within code using directives and pragmas. For details on the use of directives and pragmas, refer to Using Directives and Pragmas.

By default, -⁠Mconcur without any sub-options is equivalent to:

-Mconcur=dist:block

This enables parallelization of loops with blocked iteration allocation across the available threads of execution. These defaults may vary depending on the target system. The following table lists and briefly describes some of the -⁠Mconcur suboptions.

Table 9. -Mconcur Suboptions
Use this option ... To instruct the parallelizer to do this...
-Mconcur=allcores Use all available cores. Specify this option at link time.
-Mconcur=[no]altcode

Generate [do not generate] alternate serial code for parallelized loops. If altcode is specified without arguments, the parallelizer determines an appropriate cutoff length and generates serial code to be executed whenever the loop count is less than or equal to that length.

If altcode:n is specified, the serial altcode is executed whenever the loop count is less than or equal to n. Specifying noaltcode disables this option and no alternate serial code is generated.

-Mconcur=[no]assoc Enable [disable] parallelization of loops with associative reductions.
-Mconcur=bind Bind threads to cores. Specify this option at link time.
-Mconcur=cncall Specifies that it is safe to parallelize loops that contain subroutine or function calls. By default, such loops are excluded from consideration for auto-parallelization. Also, no minimum loop count threshold must be satisfied before parallelization occurs, and last values of scalars are assumed to be safe.
-Mconcur=dist:{block|cyclic} Specifies whether to assign loop iterations to the available threads in blocks or in a cyclic (round-robin) fashion. Block distribution is the default. If cyclic is specified, iterations are allocated to processors cyclically. That is, processor 0 performs iterations 0, 3, 6, etc.; processor 1 performs iterations 1, 4, 7, etc.; and processor 2 performs iterations 2, 5, 8, etc.
-Mconcur=innermost Enable parallelization of innermost loops.
-Mconcur=levels:<n> Parallelize loops nested at most n levels deep.
-Mconcur=[no]numa Use thread/processors affinity when running on a NUMA architecture. Specifying -⁠Mconcur=nonuma disables this option.

The environment variable NCPUS is checked at runtime for a parallel program. If NCPUS is set to 1, a parallel program runs serially, but will use the parallel routines generated during compilation. If NCPUS is set to a value greater than 1, the specified number of processors are used to execute the program. Setting NCPUS to a value exceeding the number of physical processors can produce inefficient execution. Executing a program on multiple processors in an environment where some of the processors are being time-shared with another executing job can also result in inefficient execution.

As with the vectorizer, the -⁠Mconcur option can speed up code if it contains well-behaved countable loops and/or computationally intensive nested loops that operate on arrays. However, it is possible that some codes show a decrease in performance on multi-processor systems when compiled with -⁠Mconcur due to parallelization overheads, memory bandwidth limitations in the target system, false-sharing of cache lines, or other architectural or code-generation factors. For this reason, it is recommended that you check carefully whether particular program units or loops show improved performance when compiled using this option.

If the compiler is not able to successfully auto-parallelize your application, you should refer to Using OpenMP. It is possible that insertion of explicit parallelization directives or pragmas, and use of the compiler option -⁠mp might enable the application to run in parallel.

3.7.2. Loops That Fail to Parallelize

In spite of the sophisticated analysis and transformations performed by the compiler, programmers may notice loops that are seemingly parallel, but are not parallelized. In this subsection, we look at some examples of common situations where parallelization does not occur.

Innermost Loops

As noted earlier in this section, the PGI compilers will not parallelize innermost loops by default, because it is usually not profitable. However, you can override this default using the -⁠Mconcur=innermost command-line option.

Timing Loops

Often, loops occur in programs that are similar to timing loops. The outer loop in the following example is one such loop.

do j = 1, 2
  do i = 1, n
1    a(i) = b(i) + c(i)
 enddo
enddo

The outer loop in the preceding example is not parallelized because the compiler detects a cross-iteration dependence in the assignment to a(i). Suppose the outer loop were parallelized. Then both processors would simultaneously attempt to make assignments into a(1:n). Now in general the values computed by each processor for a(1:n) will differ, so that simultaneous assignment into a(1:n) will produce values different from sequential execution of the loops.

In this example, values computed for a(1:n) don’t depend on j, so that simultaneous assignment by both processors does not yield incorrect results. However, it is beyond the scope of the compilers’ dependence analysis to determine that values computed in one iteration of a loop don’t differ from values computed in another iteration. So the worst case is assumed, and different iterations of the outer loop are assumed to compute different values for a(1:n). Is this assumption too pessimistic? If j doesn’t occur anywhere within a loop, the loop exists only to cause some delay, most probably to improve timing resolution. It is not usually valid to parallelize timing loops; to do so would distort the timing information for the inner loops.

Scalars

Quite often, scalars will inhibit parallelization of non-innermost loops. There are two separate cases that present problems. In the first case, scalars appear to be expandable, but appear in non-innermost loops, as in the following example.

do j = 1, n 
 x = b(j) 
 do i = 1, n 
  a(i,j) = x + c(i,j) 
 enddo
enddo

There are a number of technical problems to be resolved in order to recognize expandable scalars in non-innermost loops. Until this generalization occurs, scalars like x in the preceding code segment inhibit parallelization of loops in which they are assigned. In the following example, scalar k is not expandable, and it is not an accumulator for a reduction.

    k = 1 
     do i = 1, n 
       do j = 1, n 
1       a(j,i) = b(k) * x 
       enddo
       k = i 
2      if (i .gt. n/2) k = n - (i - n/2) 
     enddo

If the outer loop is parallelized, conflicting values are stored into k by the various processors. The variable k cannot be made local to each processor because its value must remain coherent among the processors. It is possible the loop could be parallelized if all assignments to k are placed in critical sections. However, it is not clear where critical sections should be introduced because in general the value for k could depend on another scalar (or on k itself), and code to obtain the value of other scalars must reside in the same critical section.

In the previous example, the assignment to k within a conditional at label 2 prevents k from being recognized as an induction variable. If the conditional statement at label 2 is removed, k would be an induction variable whose value varies linearly with j, and the loop could be parallelized.

Scalar Last Values

During parallelization, scalars within loops often need to be privatized; that is, each execution thread has its own independent copy of the scalar. Problems can arise if a privatized scalar is accessed outside the loop. For example, consider the following loops in C/C++ and Fortran:

/* C/C++ version */
for (i = 1; i<N; i++){
  if( x[i] > 5.0 ) 
    t = x[i];
}
v = t;
  
f(v);
! Fortran version
do I = 1,N
  if (x(I) > 5.0 ) then
    t = x(I)
  endif
enddo
v = t

call f(v)

The value of t may not be computed on the last iteration of the loop. Normally, if a scalar is assigned within a loop and used following the loop, the PGI compilers save the last value of the scalar. However, if the loop is parallelized and the scalar is not assigned on every iteration, it may be difficult, without resorting to costly critical sections, to determine on what iteration t is last assigned. Analysis allows the compiler to determine that a scalar is assigned on each iteration and hence that the loop is safe to parallelize if the scalar is used later, as illustrated in the following C/C++ and Fortran examples.

/* C/C++ version */
for (i=1;i<n;i++) {
 if (x[i]>0.0) {
  t=2.0;
 }
 else {
  t=3.0;
  y[i]=t;
 }
}
v=t;
! Fortran version
do I = 1,N
 if (x(I)>0.0) then
  t=2.0
 else
  t=3.0
  y(i)=t
 endif
enddo
v=t

Notice that t is assigned on every iteration of the loop. However, there are cases where a scalar may be privatizable, but if it is used after the loop, it is unsafe to parallelize. Examine the following loops in which each use of t within the loop is reached by a definition from the same iteration.

/* C/C++ Version */
for (i=1;i<N;i++){
 if(x[i]>0.0){
  t=x[i];
  y[i]=t;
 }
}
v=t;

f(v);
! Fortran Version
do I = 1,N
 if (x(I)>0.0) then
  t=x(I)
  y(i)=t
 endif
enddo
v=t

call f(v)

Here t is privatizable, but the use of t outside the loop may yield incorrect results, since the compiler may not be able to detect on which iteration of the parallelized loop t is last assigned. The compiler detects the previous cases. When a scalar is used after the loop but is not defined on every iteration of the loop, parallelization does not occur.

When the programmer knows that the scalar is assigned on the last iteration of the loop, the programmer may use a directive or pragma to let the compiler know the loop is safe to parallelize. The directive or pragma safe_lastval informs the compiler that, for a given loop, all scalars are assigned in the last iteration of the loop; thus, it is safe to parallelize the loop. We could add the following line to any of our previous examples.

!pgi$l safe_lastval    ! Fortran Version
#pragma loop safe_lastval     /* C/C++ Version */

The resulting code looks similar to this:

/* C/C++ Version */
#pragma loop safe_lastval
...
for (i=1;i<N;i++){
 if(x[i]>5.0 ) t=x[i];
}
	v = t;
! Fortran Version
!pgi$l safe_lastv
...
do I = 1,N
 if (x(I) > 5.0 ) then
  t = x(I)
 endif
enddo
v = t

In addition, a command-line option -⁠Msafe_lastval provides this information for all loops within the routines being compiled, which essentially provides global scope.

3.8. Processor-Specific Optimization and the Unified Binary

Every brand of processor has differences, some subtle, in hardware features such as instruction sets and cache size. The compilers make architecture-specific decisions about things such as instruction selection, instruction scheduling, and vectorization. By default, the PGI compilers produce code specifically targeted to the type of processor on which the compilation is performed. That is, the default is to use all supported instructions wherever possible when compiling on a given system. As a result, executables created on a given system may not be usable on previous generation systems. For example, executables created on a Pentium 4 may fail to execute on a Pentium III or Pentium II.

All PGI compilers have the capability of generating unified binaries, which provide a low-overhead means for generating a single executable that is compatible with and has good performance on more than one hardware platform.

You can use the -⁠tp option to control compilation behavior by specifying the processor or processors with which the generated code is compatible. The compilers generate and combine into one executable multiple binary code streams, each optimized for a specific platform. At runtime, the executable senses the environment and dynamically selects the appropriate code stream. For specific information on the -⁠tp option, refer to the PGI Compiler Reference Manual.

Executable size is automatically controlled via unified binary culling. Only those functions and subroutines where the target affects the generated code have unique binary images, resulting in a code-size savings of 10% to 90% compared to generating full copies of code for each target.

Programs can use the PGI Unified Binary even if all of the object files and libraries are not compiled as unified binaries. Like any other object file, you can use PGI Unified Binary object files to create programs or libraries. No special start up code is needed; support is linked in from the PGI libraries.

The -⁠Mipa option disables generation of PGI Unified Binary. Instead, the default target auto-detect rules for the host are used to select the target processor.

3.9. Interprocedural Analysis and Optimization using -Mipa

The PGI Fortran, C and C++ compilers use interprocedural analysis (IPA) that results in minimal changes to makefiles and the standard edit-build-run application development cycle. Other than adding -⁠Mipa to the command line, no other changes are required. For reference and background, the process of building a program without IPA is described later in this section, followed by the minor modifications required to use IPA with the PGI compilers. While the PGCC compiler is used here to show how IPA works, similar capabilities apply to each of the PGI Fortran, C and C++ compilers.

Note: The examples use Linux file naming conventions. As of the PGI 16.3 release, -⁠Mipa has been disabled on Windows.

3.9.1. Building a Program Without IPA – Single Step

Using the pgcc command-level compiler driver, multiple source files can be compiled and linked into a single executable with one command. The following example compiles and links three source files:

% pgcc -o a.out file1.c file2.c file3.c

In actuality, the pgcc driver executes several steps to produce the assembly code and object files corresponding to each source file, and subsequently to link the object files together into a single executable file. This command is roughly equivalent to the following commands performed individually:

% pgcc -S -o file1.s file1.c
% as -o file1.o file1.s
% pgcc -S -o file2.s file2.c
% as -o file2.o file2.s
% pgcc -S -o file3.s file3.c
% as -o file3.o file3.s
% pgcc -o a.out file1.o file2.o file3.o

If any of the three source files is edited, the executable can be rebuilt with the same command line:

% pgcc -o a.out file1.c file2.c file3.c
Note: This always works as intended, but has the side-effect of recompiling all of the source files, even if only one has changed. For applications with a large number of source files, this can be time-consuming and inefficient.

3.9.2. Building a Program Without IPA – Several Steps

It is also possible to use individual pgcc commands to compile each source file into a corresponding object file, and one to link the resulting object files into an executable:

% pgcc -c file1.c
% pgcc -c file2.c
% pgcc -c file3.c
% pgcc -o a.out file1.o file2.o file3.o

The pgcc driver invokes the compiler and assembler as required to process each source file, and invokes the linker for the final link command. If you modify one of the source files, the executable can be rebuilt by compiling just that file and then relinking:

% pgcc -c file1.c
% pgcc -o a.out file1.o file2.o file3.o

3.9.3. Building a Program Without IPA Using Make

The program compilation and linking process can be simplified greatly using the make utility on systems where it is supported. Suppose you create a makefile containing the following lines:

a.out: file1.o file2.o file3.o
 pgcc $(OPT) -o a.out file1.o file2.o file3.o 
file1.o: file1.c
 pgcc $(OPT) -c file1.c
file2.o: file2.c
 pgcc $(OPT) -c file2.c
file3.o: file3.c
pgcc $(OPT) -c file3.c

It is then possible to type a single make command:

% make

The make utility determines which object files are out of date with respect to their corresponding source files, and invokes the compiler to recompile only those source files and to relink the executable. If you subsequently edit one or more source files, the executable can be rebuilt with the minimum number of recompilations using the same single make command.

3.9.4. Building a Program with IPA

Interprocedural analysis and optimization (IPA) by the PGI compilers alters the standard and make utility command-level interfaces as little as possible. IPA occurs in three phases:

  • Collection: Create a summary of each function or procedure, collecting the useful information for interprocedural optimizations. This is done during the compile step if the -⁠Mipa switch is present on the command line; summary information is collected and stored in the object file.
  • Propagation: Process all the object files to propagate the interprocedural summary information across function and file boundaries. This is done during the link step, when all the object files are combined, if the -⁠Mipa switch is present on the link command line.
  • Recompile/Optimization: Recompile each of the object files with the propagated interprocedural information, producing a specialized object file. This process is also performed during the link step when the -⁠Mipa switch is present on the link command line.

When linking with -⁠Mipa, the PGI compilers automatically regenerate IPA-optimized versions of each object file, essentially recompiling each file. If there are IPA-optimized objects from a previous build, the compilers will minimize the recompile time by reusing those objects if they are still valid. They will still be valid if the IPA-optimized object is newer than the original object file, and the propagated IPA information for that file has not changed since it was optimized.

After each object file has been recompiled, the regular linker is invoked to build the application with the IPA-optimized object files. The IPA-optimized object files are saved in the same directory as the original object files, for use in subsequent program builds.

3.9.5. Building a Program with IPA – Single Step

By adding the -⁠Mipa command line switch, several source files can be compiled and linked with interprocedural optimizations with one command:

% pgcc -Mipa=fast -o a.out file1.c file2.c file3.c

Just like compiling without -⁠Mipa, the driver executes several steps to produce the assembly and object files to create the executable:

% pgcc -Mipa=fast -S -o file1.s file1.c
% as -o file1.o file1.s
% pgcc -Mipa=fast -S -o file2.s file2.c
% as -o file2.o file2.s
% pgcc -Mipa=fast -S -o file3.s file3.c
% as -o file3.o file3.s
% pgcc -Mipa=fast -o a.out file1.o file2.o file3.o

In the last step, an IPA linker is invoked to read all the IPA summary information and perform the interprocedural propagation. The IPA linker reinvokes the compiler on each of the object files to recompile them with interprocedural information. This creates three new objects with mangled names:

file1_ipa5_a.out.oo.o, file2_ipa5_a.out.oo.o, file3_ipa5_a.out.oo.o

The system linker is then invoked to link these IPA-optimized objects into the final executable. Later, if one of the three source files is edited, the executable can be rebuilt with the same command line:

% pgcc -Mipa=fast -o a.out file1.c file2.c file3.c

This works, but again has the side-effect of compiling each source file, and recompiling each object file at link time.

3.9.6. Building a Program with IPA – Several Steps

Just by adding the -⁠Mipa command-line switch, it is possible to use individual pgcc commands to compile each source file, followed by a command to link the resulting object files into an executable:

% pgcc -Mipa=fast -c file1.c
% pgcc -Mipa=fast -c file2.c
% pgcc -Mipa=fast -c file3.c
% pgcc -Mipa=fast -o a.out file1.o file2.o file3.o

The pgcc driver invokes the compiler and assembler as required to process each source file, and invokes the IPA linker for the final link command. If you modify one of the source files, the executable can be rebuilt by compiling just that file and then relinking:

% pgcc -Mipa=fast -c file1.c
% pgcc -Mipa=fast -o a.out file1.o file2.o file3.o

When the IPA linker is invoked, it will determine that the IPA-optimized object for file1.o (file1_ipa5_a.out.oo.o) is stale, since it is older than the object file1.o; and hence it needs to be rebuilt, and reinvokes the compiler to generate it. In addition, depending on the nature of the changes to the source file file1.c, the interprocedural optimizations previously performed for file2 and file3 may now be inaccurate. For instance, IPA may have propagated a constant argument value in a call from a function in file1.c to a function in file2.c; if the value of the argument has changed, any optimizations based on that constant value are invalid. The IPA linker determines which, if any, of the previously created IPA-optimized objects need to be regenerated; and, as appropriate, reinvokes the compiler to regenerate them. Only those objects that are stale or which have new or different IPA information are regenerated. This approach saves compile time.

3.9.7. Building a Program with IPA Using Make

As shown earlier, programs can be built with IPA using the make utility. Just add the command-line switch -⁠Mipa, as shown here:

OPT=-Mipa=fast 
a.out: file1.o file2.o file3.o 
 pgcc $(OPT) -o a.out file1.o file2.o file3.o 
file1.o: file1.c 
 pgcc $(OPT) -c file1.c 
file2.o: file2.c 
 pgcc $(OPT) -c file2.c 
file3.o: file3.c 
 pgcc $(OPT) -c file3.c

Using the single make command invokes the compiler to generate any of the object files that are out-of-date, then invokes pgcc to link the objects into the executable. At link time, pgcc calls the IPA linker to regenerate any stale or invalid IPA-optimized objects.

% make

3.9.8. Questions about IPA

Question: Why is the object file so large?

Answer: An object file created with -⁠Mipa contains several additional sections. One is the summary information used to drive the interprocedural analysis. In addition, the object file contains the compiler internal representation of the source file, so the file can be recompiled at link time with interprocedural optimizations. There may be additional information when inlining is enabled. The total size of the object file may be 5-10 times its original size. The extra sections are not added to the final executable.

Question: What if I compile with -⁠Mipa and link without -⁠Mipa?

Answer: The PGI compilers generate a legal object file, even when the source file is compiled with -⁠Mipa. If you compile with -⁠Mipa and link without -⁠Mipa, the linker is invoked on the original object files. A legal executable is generated. While this executable does not have the benefit of interprocedural optimizations, any other optimizations do apply.

Question: What if I compile without -⁠Mipa and link with -⁠Mipa?

Answer: At link time, the IPA linker must have summary information about all the functions or routines used in the program. This information is created only when a file is compiled with -⁠Mipa. If you compile a file without -⁠Mipa and then try to get interprocedural optimizations by linking with -⁠Mipa, the IPA linker will issue a message that some routines have no IPA summary information, and will proceed to run the system linker using the original object files. If some files were compiled with -⁠Mipa and others were not, it will determine the safest approximation of the IPA summary information for those files not compiled with -⁠Mipa, and use that to recompile the other files using interprocedural optimizations.

Question: Can I build multiple applications in the same directory with -⁠Mipa?

Answer: Yes. Suppose you have three source files: main1.c, main2.c, and sub.c, where sub.c is shared between the two applications. Suppose you build the first application with -⁠Mipa, using this command:

% pgcc -Mipa=fast -o app1 main1.c sub.c

The IPA linker creates two IPA-optimized object files and uses them to build the first application.

main1_ipa4_app1.oo sub_ipa4_app1.oo

Now suppose you build the second application using this command:

% pgcc -Mipa=fast -o app2 main2.c sub.c

The IPA linker creates two more IPA-optimized object files:

main2_ipa4_app2.oo sub_ipa4_app2.oo
Note: There are now three object files for sub.c: the original sub.o, and two IPA-optimized objects, one for each application in which it appears.

Question: How is the mangled name for the IPA-optimized object files generated?

Answer: The mangled name has ‘_ipa’ appended, followed by the decimal number of the length of the executable file name, followed by an underscore and the executable file name itself. The suffix is changed to .oo so that linking *.o does not pull in the IPA-optimized objects. If the IPA linker determines that the file would not benefit from any interprocedural optimizations, it does not have to recompile the file at link time, and uses the original object.

Question: Can I use parallel make environments (e.g., pmake) with IPA?

Answer: No. IPA is not compatible with parallel make environments.

3.10. Profile-Feedback Optimization using -Mpfi/-Mpfo

The PGI compilers support many common profile-feedback optimizations, including semi-invariant value optimizations and block placement. These are performed under control of the -⁠Mpfi/-⁠Mpfo command-line options.

When invoked with the -⁠Mpfi option, the PGI compilers instrument the generated executable for collection of profile and data feedback information. This information can be used in subsequent compilations that include the -⁠Mpfo optimization option. -⁠Mpfi must be used at both compile-time and link-time. Programs compiled with -⁠Mpfi include extra code to collect runtime statistics and write them out to a trace file. When the resulting program is executed, a profile feedback trace file pgfi.out is generated in the current working directory.

Note: Programs compiled and linked with -⁠Mpfi execute more slowly due to the instrumentation and data collection overhead. You should use executables compiled with -⁠Mpfi only for execution of training runs.

When invoked with the -⁠Mpfo option, the PGI compilers use data from a pgfi.out profile feedback tracefile to enable or enhance certain performance optimizations. Use of this option requires the presence of a pgfi.out trace file in the current working directory.

3.11. Default Optimization Levels

The following table shows the interaction between the -⁠O<level>, -⁠g, and -⁠M<opt> options. In the table, level can be 0, 1, 2, 3 or 4, and <opt> can be vect, concur, unroll or ipa. The default optimization level is dependent upon these command-line options.

Table 10. Optimization and -⁠O, -⁠g and -⁠M<opt> Options
Optimize Option Debug Option -M<opt> Option Optimization Level
none none none 1
none none -M<opt> 2
none -g none 0
-O none or -g none 2
-O<level> none or -g none level
-O<level> <= 2 none or -g -M<opt> 2

Code that is not optimized yet compiled using the option -⁠O0 can be significantly slower than code generated at other optimization levels. The -⁠M<opt> option, where <opt> is vect, concur, unroll or ipa, sets the optimization level to 2 if no -⁠O options are supplied. The -⁠fast option sets the optimization level to a target-dependent optimization level if no -⁠O options are supplied.

3.12. Local Optimization Using Directives and Pragmas

Command-line options let you specify optimizations for an entire source file. Directives supplied within a Fortran source file and pragmas supplied within a C or C++ source file provide information to the compiler and alter the effects of certain command-line options or the default behavior of the compiler. (Many directives have a corresponding command-line option.)

While a command line option affects the entire source file that is being compiled, directives and pragmas let you do the following:

  • Apply, or disable, the effects of a particular command-line option to selected subprograms or to selected loops in the source file (for example, an optimization).
  • Globally override command-line options.
  • Tune selected routines or loops based on your knowledge or on information obtained through profiling.

Using Directives and Pragmas provides details on how to add directives and pragmas to your source files.

3.13. Execution Timing and Instruction Counting

As this chapter describes, once you have a program that compiles, executes and gives correct results, you may optimize your code for execution efficiency.

Selecting the correct optimization level requires some thought and may require that you compare several optimization levels before arriving at the best solution. To compare optimization levels, you need to measure the execution time for your program. There are several approaches you can take for timing execution.

  • You can use shell commands that provide execution time statistics.
  • You can include function calls in your code that provide timing information.
  • You can profile sections of code.

Timing functions available with the PGI compilers include these:

  • 3F timing routines.
  • The SECNDS pre-declared function in PGF77, PGF95, or PGFORTRAN.
  • The SYSTEM_CLOCK or CPU_CLOCK intrinsics in PGF95 or PGFORTRAN.

In general, when timing a program, you should try to eliminate or reduce the amount of system level activities such as I/O, program loading, and task switching.

The following example shows a fragment that indicates how to use SYSTEM_CLOCK effectively within a Fortran program unit.

Using SYSTEM_CLOCK code fragment

integer :: nprocs, hz, clock0, clock1
 real :: time
call system_clock (count_rate=hz)
 call system_clock(count=clock0)
 < do work>
 call system_clock(count=clock1)
 t = (clock1 - clock0)
 time = real (t) / real(hz)

Or you can use the F90 cpu_time subroutine:

real :: t1, t2, time
call cpu_time(t1)
 < do work>
call cpu_time(t2)
time = t2 - t1

3.14. Portability of Multi-Threaded Programs on Linux

PGI created the library libnuma to handle the variations between various implementations of Linux.

Some older versions of Linux are lacking certain features that support multi-processor and multicore systems; in particular, the system call 'sched_setaffinity' and the numa library libnuma. The PGI runtime library uses these features to implement some -⁠Mconcur and -⁠mp operations.

These variations led to the creation of the PGI library: libnuma, which is used on all 64-bit Linux systems, but is not needed on Windows or macOS.

When a program is linked with the system libnuma library, the program depends on that library to run. On systems without a libnuma library, the PGI version of libnuma provides the required stubs so that the program links and executes properly. If the program is linked with libnuma, the differences between systems is masked by the different versions of libnuma.

When a program is deployed to the target system, the proper set of libraries, real or stub, should be deployed with the program.

This facility requires that the program be dynamically linked with libnuma.

3.14.1. libnuma

Not all systems have libnuma. Typically, only numa systems have this library. PGI supplies a stub version of libnuma which satisfies the calls from the PGI runtime to libnuma. libnuma is a shared library that is linked dynamically at runtime.

The reason to have a numa library on all systems is to allow multi-threaded programs, such as programs compiled with -⁠Mconcur or -⁠mp, to be compiled, linked, and executed without regard to whether the host or target systems has a numa library. When the numa library is not available, a multi-threaded program still runs because the calls to the numa library are satisfied by the PGI stub library.

During installation, the installation procedure checks for the existence of a real libnuma among the system libraries. If the real library is not found, the PGI stub version is substituted.

4. Using Function Inlining

Function inlining replaces a call to a function or a subroutine with the body of the function or subroutine. This can speed up execution by eliminating parameter passing and function/subroutine call and return overhead. It also allows the compiler to optimize the function with the rest of the code. Note that using function inlining indiscriminately can result in much larger code size and no increase in execution speed.

The PGI compilers provide two categories of inlining:

  • Automatic function inlining – In C/C++, you can inline static functions with the inline keyword by using the -⁠Mautoinline option, which is included with -⁠fast.
  • Function inlining – You can inline functions which were extracted to the inline libraries in C/Fortran/C++. There are two ways of enabling function inlining: with and without the lib suboption. For the latter, you create inline libraries, for example using the pgfortran compiler driver and the -⁠o and -⁠Mextract options.

There are important restrictions on inlining. Inlining only applies to certain types of functions. Refer to Restrictions on Inlining for more details on function inlining limitations.

This section describes how to use the following options related to function inlining:

  -Mautoinline
  -Mextract
  -Minline
  -Mnoinline
  -Mrecursive

4.1. Automatic function inlining in C/C++

To enable automatic function inlining in C/C++ for static functions with the inline keyword, use the -⁠Mautoinline option (included in -⁠fast). Use -⁠Mnoautoinline to disable it.

Several -⁠Mautoinline suboptions let you determine the selection criteria. These suboptions are:

maxsize:n
Automatically inline functions size n and less
totalsize:n
Limit automatic inlining to total size of n

4.2. Invoking Function Inlining

To invoke the function inliner, use the -⁠Minline option. If you do not specify an inline library, the compiler performs a special prepass on all source files named on the compiler command line before it compiles any of them. This pass extracts functions that meet the requirements for inlining and puts them in a temporary inline library for use by the compilation pass.

Several -⁠Minline suboptions let you determine the selection criteria for functions to be inlined. These suboptions include:

except:func
Inlines all eligible functions except func, a function in the source text. You can use a comma-separated list to specify multiple functions.
[name:]func
Inlines all functions in the source text whose name matches func. You can use a comma-separated list to specify multiple functions.
[maxsize:]number
A numeric option is assumed to be a size. Functions of size number or less are inlined. If both number and function are specified, then functions matching the given name(s) or meeting the size requirements are inlined.
reshape
Fortran subprograms with array arguments are not inlined by default if the array shape does not match the shape in the caller. Use this option to override the default.
smallsize:number
Always inline functions of size smaller than number regardless of other size limits.
totalsize:number
Stop inlining in a function when the function's total inlined size reaches the number specified.
[lib:]file.ext
Instructs the inliner to inline the functions within the library file file.ext. If no inline library is specified, functions are extracted from a temporary library created during an extract prepass.
Tip: Create the library file using the -⁠Mextract option.

If you specify both a function name and a maxsize n, the compiler inlines functions that match the function name or have n or fewer statements.

If a name is used without a keyword, then a name with a period is assumed to be an inline library and a name without a period is assumed to be a function name. If a number is used without a keyword, the number is assumed to be a size.

Inlining can be disabled with -⁠Mnoinline.

In the following example, the compiler inlines functions with fewer than approximately 100 statements in the source file myprog.f and writes the executable code in the default output file a.out.

   $ pgfortran -Minline=maxsize:100 myprog.f 

For more information on the -⁠Minline options, refer to ‘-⁠M Options by Category’ section of the PGI Compiler Reference Manual.

4.3. Using an Inline Library

If you specify one or more inline libraries on the command line with the -Minline option, the compiler does not perform an initial extract pass. The compiler selects functions to inline from the specified inline library. If you also specify a size or function name, all functions in the inline library meeting the selection criteria are selected for inline expansion at points in the source text where they are called.

If you do not specify a function name or a size limitation for the -Minline option, the compiler inlines every function in the inline library that matches a function in the source text.

In the following example, the compiler inlines the function proc from the inline library lib.il and writes the executable code in the default output file a.out.

	$ pgfortran -Minline=name:proc,lib:lib.il myprog.f

The following command line is equivalent to the preceding line, with the exception that in the following example does not use the keywords name: and lib:. You typically use keywords to avoid name conflicts when you use an inline library name that does not contain a period. Otherwise, without the keywords, a period informs the compiler that the file on the command line is an inline library.

$ pgfortran -Minline=proc,lib.il myprog.f

4.4. Creating an Inline Library

You can create or update an inline library using the -⁠Mextract command-line option. If you do not specify selection criteria with the -⁠Mextract option, the compiler attempts to extract all subprograms.

Several -⁠Mextract options let you determine the selection criteria for creating or updating an inline library. These selection criteria include:

func
Extracts the function func. you can use a comma-separated list to specify multiple functions.
[name:]func
Extracts the functions whose name matches func, a function in the source text.
[size:]n
Limits the size of the extracted functions to functions with a statement count less than or equal to n, the specified size.
Note: The size n may not exactly equal the number of statements in a selected function; the size parameter is merely a rough gauge.
[lib:]ext.lib
Stores the extracted information in the library directory ext.lib.

If no inline library is specified, functions are extracted to a temporary library created during an extract prepass for use during the compilation stage.

When you use the -⁠Mextract option, only the extract phase is performed; the compile and link phases are not performed. The output of an extract pass is a library of functions available for inlining. This output is placed in the inline library file specified on the command line with the -⁠o filename specification. If the library file exists, new information is appended to it. If the file does not exist, it is created. You can use a command similar to the following:

$ pgfortran -Mextract=lib:lib.il myfunc.f

You can use the -⁠Minline option with the -⁠Mextract option. In this case, the extracted library of functions can have other functions inlined into the library. Using both options enables you to obtain more than one level of inlining. In this situation, if you do not specify a library with the -⁠Minline option, the inline process consists of two extract passes. The first pass is a hidden pass implied by the -⁠Minline option, during which the compiler extracts functions and places them into a temporary library. The second pass uses the results of the first pass but puts its results into the library that you specify with the -⁠o option.

4.4.1. Working with Inline Libraries

An inline library is implemented as a directory with each inline function in the library stored as a file using an encoded form of the inlinable function.

A special file named TOC in the inline library directory serves as a table of contents for the inline library. This is a printable, ASCII file which you can examine to locate information about the library contents, such as names and sizes of functions, the source file from which they were extracted, the version number of the extractor which created the entry, and so on.

Libraries and their elements can be manipulated using ordinary system commands.

  • Inline libraries can be copied or renamed.
  • Elements of libraries can be deleted or copied from one library to another.
  • The ls or dir command can be used to determine the last-change date of a library entry.

4.4.2. Dependencies

When a library is created or updated using one of the PGI compilers, the last-change date of the library directory is updated. This allows a library to be listed as a dependence in a makefile and ensures that the necessary compilations are performed when a library is changed.

4.4.3. Updating Inline Libraries – Makefiles

If you use inline libraries you must be certain that they remain up-to-date with the source files into which they are inlined. One way to assure inline libraries are updated is to include them in a makefile.

The makefile fragment in the following example assumes the file utils.f contains a number of small functions used in the files parser.f and alloc.f.

This portion of the makefile:

  • Maintains the inline library utils.il.
  • Updates the library whenever you change utils.f or one of the include files it uses.
  • Compiles parser.f and alloc.f whenever you update the library.

Sample Makefile

SRC = mydir
FC = pgfortran
FFLAGS = -O2
main.o: $(SRC)/main.f $(SRC)/global.h
 $(FC) $(FFLAGS) -c $(SRC)/main.f
utils.o: $(SRC)/utils.f $(SRC)/global.h $(SRC)/utils.h
 $(FC) $(FFLAGS) -c $(SRC)/utils.f
utils.il: $(SRC)/utils.f $(SRC)/global.h $(SRC)/utils.h
 $(FC) $(FFLAGS) -Mextract=15 -o utils.il $(SRC)/utils.f
parser.o: $(SRC)/parser.f $(SRC)/global.h utils.il
 $(FC) $(FFLAGS) -Minline=utils.il -c $(SRC)/parser.f
alloc.o: $(SRC)/alloc.f $(SRC)/global.h utils.il
 $(FC) $(FFLAGS) -Minline=utils.il -c $(SRC)/alloc.f 
myprog: main.o utils.o parser.o alloc.o
$(FC) -o myprog main.o utils.o parser.o alloc.o

4.5. Error Detection during Inlining

You can specify the -⁠Minfo=inline option to request inlining information from the compiler when you invoke the inliner. For example:

$ pgfortran -Minline=mylib.il -Minfo=inline myext.f

4.6. Examples

Assume the program dhry consists of a single source file dhry.f. The following command line builds an executable file for dhry in which proc7 is inlined wherever it is called:

$ pgfortran dhry.f -Minline=proc7

The following command lines build an executable file for dhry in which proc7 plus any functions of approximately 10 or fewer statements are inlined (one level only).

Note: The specified functions are inlined only if they are previously placed in the inline library, temp.il, during the extract phase.
$ pgfortran dhry.f -Mextract=lib:temp.il 
$ pgfortran dhry.f -Minline=10,proc7,temp.il

Using the same source file dhry.f, the following example builds an executable for dhry in which all functions of roughly ten or fewer statements are inlined. Two levels of inlining are performed. This means that if function A calls function B, and B calls C, and both B and C are inlinable, then the version of B which is inlined into A will have had C inlined into it.

$ pgfortran dhry.f -Minline=maxsize:10

4.7. Restrictions on Inlining

The following Fortran subprograms cannot be extracted:

  • Main or BLOCK DATA programs.
  • Subprograms containing alternate return, assigned GO TO, DATA, SAVE, or EQUIVALENCE statements.
  • Subprograms containing FORMAT statements.
  • Subprograms containing multiple entries.

A Fortran subprogram is not inlined if any of the following applies:

  • It is referenced in a statement function.
  • A common block mismatch exists; in other words, the caller must contain all common blocks specified in the callee, and elements of the common blocks must agree in name, order, and type (except that the caller's common block can have additional members appended to the end of the common block).
  • An argument mismatch exists; in other words, the number and type (size) of actual and formal parameters must be equal.
  • A name clash exists, such as a call to subroutine xyz in the extracted subprogram and a variable named xyz in the caller.

The following types of C and C++ functions cannot be inlined:

  • Functions containing switch statements
  • Functions which reference a static variable whose definition is nested within the function
  • Functions which accept a variable number of arguments

Certain C/C++ functions can only be inlined into the file that contains their definition:

  • Static functions
  • Functions which call a static function
  • Functions which reference a static variable

5. Using OpenMP

The PGF77 and PGFORTRAN Fortran compilers support the OpenMP Fortran Application Program Interface. The PGCC and PGC⁠+⁠+ compilers support the OpenMP C/C⁠+⁠+ Application Program Interface.

OpenMP is a specification for a set of compiler directives, an applications programming interface (API), and a set of environment variables that can be used to specify shared memory parallelism in FORTRAN and C/C⁠+⁠+ programs. OpenMP may be used to obtain most of the parallel performance you can expect from your code, or it may serve as a stepping stone to parallelizing an entire application with MPI.

This section provides information on OpenMP as it is supported by PGI compilers. Currently, all PGI compilers support the version 3.1 OpenMP specification.

Use the -mp compiler switch to enable processing of the OpenMP pragmas listed in this section. As of the PGI 2011 Release, the OpenMP runtime library is linked by default. Note that GNU pthreads are not completely interoperable with OpenMP threads.

Note: When using pgc++ on Linux or macOS, the GNU STL is thread-safe to the extent listed in the GNU documentation as required by the C++11 standard. If an STL thread-safe issue is suspected, the suspect code can be run sequentially inside of an OpenMP region using #pragma omp critical sections.

This section describes how to use the following option supporting OpenMP: -⁠mp

5.1. OpenMP Overview

Let’s look at the OpenMP shared-memory parallel programming model and some common OpenMP terminology.

5.1.1. OpenMP Shared-Memory Parallel Programming Model

The OpenMP shared-memory programming model is a collection of compiler directives or pragmas, library routines, and environment variables that can be used to specify shared-memory parallelism in Fortran and in C/C++ programs.

Fortran directives and C/C++ pragmas
Allow users to mark sections of code that can be executed in parallel when the code is compiled using the -⁠mp switch.

When this switch is not present, the compiler ignores these directives and pragmas.

Fixed-form Fortran OpenMP directives begin with !$OMP, C$OMP, or *$OMP, beginning in column 1. Free-form Fortran OpenMP pragmas begin with !$OMP. OpenMP pragmas for C/C++ begin with #pragma omp. This format allows the user to have a single source code file for use with or without the -⁠mp switch, as these lines are then merely viewed as comments when -⁠mp is not present.

These directives and pragmas allow the user to create task, loop, and parallel section work-sharing constructs and synchronization constructs. They also allow the user to define how data is shared or copied between parallel threads of execution.

Note: The data environment is controlled either by using clauses on the directives or pragmas, or with additional directives or pragmas.
Runtime library routines
Are available to query the parallel runtime environment, for example to determine how many threads are participating in execution of a parallel region.
Environment variables
Are available to control the execution behavior of parallel programs. For more information, see the OpenMP website.
Macro substitution
C and C++ pragmas are subject to macro replacement after #pragma omp.

5.1.2. Terminology

For OpenMP 3.1 there are a number of terms for which it is useful to have common definitions.

Thread
An execution entity with a stack and associated static memory, called threadprivate memory.
  • An OpenMP thread is a thread that is managed by the OpenMP runtime system.
  • A thread-safe routine is a routine that performs the intended function even when executed concurrently, that is, by more than one thread.
Region
All code encountered during a specific instance of the execution of a given construct or of an OpenMP library routine. A region includes any code in called routines as well as any implicit code introduced by the OpenMP implementation.

Regions are nested if one region is (dynamically) enclosed by another region, that is, a region is encountered during the execution of another region. PGI supports both lexically and non-lexically nested parallel regions.

Parallel region
In OpenMP 3.1 there is a distinction between a parallel region and an active parallel region. A parallel region can be either inactive or active.
  • An inactive parallel region is executed by a single thread.
  • An active parallel region is a parallel region that is executed by a team consisting of more than one thread.
    Note: The definition of an active parallel region changed between OpenMP 2.5 and OpenMP 3.1. In OpenMP 2.5, the definition was a parallel region whose IF clause evaluates to true. To examine the significance of this change, look at the following example:
         program test
         logical omp_in_parallel
    
    !$omp parallel
         print *, omp_in_parallel()
    !$omp end parallel
    
         stop
         end

    Suppose we run this program with OMP_NUM_THREADS set to one. In OpenMP 2.5, this program yields T while in OpenMP 3.1, the program yields F. In OpenMP 3.1, execution is not occurring by more than one thread. Therefore, change in this definition may mean previous programs require modification.

Task
A specific instance of executable code and its data environment, generated when a thread encounters a task construct or a parallel construct.

5.1.3. OpenMP Example

Look at the following simple OpenMP example involving loops.

OpenMP Loop Example

      PROGRAM MAIN
      INTEGER I, N, OMP_GET_THREAD_NUM
      REAL*8 V(1000), GSUM, LSUM
      GSUM = 0.0D0
      N = 1000
      DO I = 1, N
         V(I) = DBLE(I)
      ENDDO
!$OMP PARALLEL PRIVATE(I,LSUM) SHARED(V,GSUM,N)
      LSUM = 0.0D0
!$OMP DO
      DO I = 1, N
         LSUM = LSUM + V(I)
      ENDDO
!$OMP END DO
!$OMP CRITICAL
      print *, "Thread ",OMP_GET_THREAD_NUM()," local sum: ",LSUM
      GSUM = GSUM + LSUM
!$OMP END CRITICAL
!$OMP END PARALLEL

      PRINT *, "Global Sum: ",GSUM
      STOP
      END

If you execute this example with the environment variable OMP_NUM_THREADS set to 4, then the output looks similar to this:

Thread            0 local sum:    31375.00000000000    
Thread            1 local sum:    93875.00000000000    
Thread            2 local sum:    156375.0000000000    
Thread            3 local sum:    218875.0000000000    
Global Sum:    500500.0000000000    
FORTRAN STOP 

5.2. Task Overview

Every part of an OpenMP program is part of a task. A task, whose execution can be performed immediately or delayed, has these characteristics:

  • Code to execute
  • A data environment – that is, it owns its data
  • An assigned thread that executes the code and uses the data.

There are two activities associated with tasks: packaging and execution.

  • Packaging: Each encountering thread packages a new instance of a task – code and data.
  • Execution: Some thread in the team executes the task at some later time.

In the following sections, we use this terminology:

Task
The package of code and instructions for allocating data created when a thread encounters a task construct. A task can be implicit or explicit.
  • An explicit task is a task generated when a task construct is encountered during execution.
  • An implicit task is a task generated by the implicit parallel region or generated when a parallel construct is encountered during execution.
Task construct
A task directive or pragma plus a structured block.
Task region
The dynamic sequence of instructions produced by the execution of a task by a thread.

5.3. Fortran Parallelization Directives

Parallelization directives are comments in a program that are interpreted by the PGI Fortran compilers when the option -⁠mp is specified on the command line. The form of a parallelization directive is:

sentinel directive_name    [clauses]

With the exception of the SGI-compatible DOACROSS directive, the sentinel must comply with these rules:

  • Be one of these: !$OMP, C$OMP, or *$OMP.
  • Must start in column 1 (one) for free-form code.
  • Must appear as a single word without embedded white space.
  • The sentinel marking a DOACROSS directive is C$.

The directive_name can be any of the directives listed in Directive and Pragma Summary Table. The valid clauses depend on the directive. Directive and Pragma Clauses provides a list of clauses, the directives and pragmas to which they apply, and their functionality.

In addition to the sentinel rules, the directive must also comply with these rules:

  • Standard Fortran syntax restrictions, such as line length, case insensitivity, and so on, apply to the directive line.
  • Initial directive lines must have a space or zero in column six.
  • Continuation directive lines must have a character other than a space or a zero in column six. Continuation lines for C$DOACROSS directives are specified using the C$& sentinel.
  • Directives which are presented in pairs must be used in pairs.

Clauses associated with directives have these characteristics:

  • The order in which clauses appear in the parallelization directives is not significant.
  • Commas separate clauses within the directives, but commas are not allowed between the directive name and the first clause.
  • Clauses on directives may be repeated as needed, subject to the restrictions listed in the description of each clause.

5.4. C/C++ Parallelization Pragmas

Parallelization pragmas are #pragma statements in a C or C++ program that are interpreted by the PGI C and C++ compilers when the option -⁠mp is specified on the command line. The form of a parallelization pragma is:

#pragma omp    pragma_name    [clauses]

The format for pragmas include these rules:

  • The pragmas follow the conventions of the C and C++ rules.
  • Whitespace can appear before and after the #.
  • Preprocessing tokens following the #pragma omp are subject to macro replacement.
  • The order in which clauses appear in the parallelization pragmas is not significant.
  • Spaces separate clauses within the pragmas.
  • Clauses on pragmas may be repeated as needed subject to the restrictions listed in the description of each clause.

For the purposes of the OpenMP pragmas, a C/C++ structured block is defined to be a statement or compound statement (a sequence of statements beginning with { and ending with }) that has a single entry and a single exit. No statement or compound statement is a C/C++ structured block if there is a jump into or out of that statement.

5.5. Directive and Pragma Recognition

The compiler option -⁠mp enables recognition of the parallelization directives and pragmas.

The use of this option also implies:

-⁠Miomutex
For directives, critical sections are generated around Fortran I/O statements.

For pragmas, calls to I/O library functions are system-dependent and are not necessarily guaranteed to be thread-safe. I/O library calls within parallel regions should be protected by critical regions to ensure they function correctly on all systems.

5.6. Directive and Pragma Summary Table

The following table provides a brief summary of the directives and pragmas that PGI supports.

Note: In the table, the values in uppercase letters are Fortran directives while the names in lowercase letters are C/C++ pragmas.

5.6.1. Directive and Pragma Summary Table

Table 11. Directive and Pragma Summary Table
Fortran Directive and C++ Pragma Description
ATOMIC [TYPE] ... END ATOMIC and atomic Semantically equivalent to enclosing a single statement in the CRITCIAL...END CRITICAL directive or critical pragma.

TYPE may be empty or one of the following: UPDATE, READ, WRITE, or CAPTURE. The END ATOMIC directive is only allowed when ending ATOMIC CAPTURE regions.

Note: Only certain statements are allowed.
BARRIER and barrier Synchronizes all threads at a specific point in a program so that all threads complete work to that point before any thread continues.
CRITICAL ... END CRITICAL and critical Defines a subsection of code within a parallel region, a critical section, which is executed one thread at a time.
DO...END DO and for Provides a mechanism for distribution of loop iterations across the available threads in a parallel region.
C$DOACROSS Specifies that the compiler should parallelize the loop to which it applies, even though that loop is not contained within a parallel region.
FLUSH and flush When this appears, all processor-visible data items, or, when a list is present (FLUSH [list]), only those specified in the list, are written to memory, thus ensuring that all the threads in a team have a consistent view of certain objects in memory.
MASTER ... END MASTER and master Designates code that executes on the master thread and that is skipped by the other threads.
ORDERED and ordered Defines a code block that is executed by only one thread at a time, and in the order of the loop iterations; this makes the ordered code block sequential, while allowing parallel execution of statements outside the code block.
PARALLEL DO and parallel for Enables you to specify which loops the compiler should parallelize.
PARALLEL ... END PARALLEL and parallel Supports a fork/join execution model in which a single thread executes all statements until a parallel region is encountered.
PARALLEL SECTIONS and parallel sections Defines a non-iterative work-sharing construct without the need to define an enclosing parallel region.
PARALLEL WORKSHARE ... END PARALLEL WORKSHARE Provides a short form method for including a WORKSHARE directive inside a PARALLEL construct.
SECTIONS ... END SECTIONS and sections Defines a non-iterative work-sharing construct within a parallel region.
SINGLE ... END SINGLE and single Designates code that executes on a single thread and that is skipped by the other threads.
TASK and task Defines an explicit task.
TASKYIELD and taskyield Specifies a scheduling point for a task where the currently executing task may be yielded, and a different deferred task may be executed.
TASKWAIT and taskwait Specifies a wait on the completion of child tasks generated since the beginning of the current task.
THREADPRIVATE and threadprivate When a common block or variable that is initialized appears in this directive or pragma, each thread’s copy is initialized once prior to its first use.
WORKSHARE ... END WORKSHARE Provides a mechanism to effect parallel execution of non-iterative but implicitly data parallel constructs.

5.7. Directive and Pragma Clauses

Some directives and pragmas accept clauses that further allow a user to control the scope attributes of variables for the duration of the directive or pragma. Not all clauses are allowed on all directives, so the clauses that are valid are included with the description of the directive and pragma.

The following table provides a brief summary of the clauses associated with OPENMP directives and pragmas that PGI supports.

For complete information on these clauses, refer to the OpenMP documentation available on the World Wide Web.

Table 12. Directive and Pragma Summary Table
This clause... Applies to this directive Applies to this pragma Has this functionality
CAPTURE ATOMIC atomic Specifies that the atomic action is reading and updating, or writing and updating a value, capturing the intermediate state.
COLLAPSE (n) DO...END DO
PARALLEL DO
PARALLEL WORKSHARE
parallel for Specifies how many loops are associated with the loop construct.
COPYIN (list) PARALLEL
PARALLEL DO
PARALLEL SECTIONS
PARALLEL WORKSHARE
parallel
parallel for
Allows threads to access the master thread's value, for a threadprivate variable. You assign the same value to threadprivate variables for each thread in the team executing the parallel region. Then, for each variable specified, the value of the variable in the master thread of the team is copied to the threadprivate copies at the beginning of the parallel region.
COPYPRIVATE(list) SINGLE single Specifies that one or more variables should be shared among all threads. This clause provides a mechanism to use a private variable to broadcast a value from one member of a team to the other members.
DEFAULT PARALLEL
PARALLEL DO
PARALLEL SECTIONS
PARALLEL WORKSHARE
parallel
parallel for
Specifies the behavior of unscoped variables in a parallel region, such as the data-sharing attributes of variables.
FINAL TASK task Specifies that all subtasks of this task will be run immediately.
FIRSTPRIVATE(list) DO
PARALLEL
PARALLEL DO
PARALLEL SECTIONS
PARALLEL WORKSHARE
SECTIONS
SINGLE
for
parallel
parallel for
sections
single
Specifies that each thread should have its own instance of a variable, and that each variable in the list should be initialized with the value of the original variable, because it exists before the parallel construct.
IF() PARALLEL ... END PARALLEL
PARALLEL DO ...
END PARALLEL DO
PARALLEL SECTIONS ...
END PARALLEL SECTIONS
PARALLEL WORKSHARE
parallel
parallel for
parallel sections
Specifies whether a loop should be executed in parallel or in serial.
LASTPRIVATE(list) DO
PARALLEL DO ...
END PARALLEL DO
PARALLEL SECTIONS ...
END PARALLEL SECTIONS
SECTIONS
parallel
parallel for
parallel sections
sections
Specifies that the enclosing context's version of the variable is set equal to the private version of whichever thread executes the final iteration of a loop construct or last section of an OpenMP section.
MERGEABLE TASK task Specifies that this task will run with the same data environment, including OpenMP internal control variables, as when it is encountered.
NOWAIT DO ... END DO
SECTIONS
SINGLE
WORKSHARE ...
END WORKSHARE
for
sections
single
Eliminates the barrier implicit at the end of a parallel region.
NUM_THREADS PARALLEL
PARALLEL DO ...
END PARALLEL DO
PARALLEL SECTIONS ...
END PARALLEL SECTIONS
PARALLEL WORKSHARE
parallel
parallel for
parallel sections
Sets the number of threads in a thread team.
ORDERED DO...END DO
PARALLEL DO...
END PARALLEL DO
parallel for Specifies that this block within the parallel DO or FOR region needs to be execute serially in the same order indicated by the enclosing loop.
PRIVATE DO
PARALLEL
PARALLEL DO ...
END PARALLEL DO
PARALLEL SECTIONS ...
END PARALLEL SECTIONS
PARALLEL WORKSHARE
SECTIONS
SINGLE
for
parallel
parallel for
parallel sections
sections
single
Specifies that each thread should have its own instance of a variable.
READ ATOMIC atomic Specifies that the atomic action is reading a value.

REDUCTION
({operator
| intrinsic } :
list)

DO
PARALLEL
PARALLEL DO ...
END PARALLEL DO
PARALLEL SECTIONS ...
END PARALLEL SECTIONS
PARALLEL WORKSHARE
SECTIONS
for
parallel
parallel for
parallel sections
sections
Specifies that one or more variables in list that are private to each thread are the subject of a reduction operation at the end of the parallel region.

SCHEDULE
(type[
,chunk])

DO ... END DO
PARALLEL DO...
END PARALLEL DO
for
parallel for
Applies to the looping directive, allowing the user to specify the chunking method for parallelization. Work is assigned to threads in different manners depending on the scheduling type or chunk size used.
SHARED PARALLEL
PARALLEL DO ...
END PARALLEL DO
PARALLEL SECTIONS ...
END PARALLEL SECTIONS
PARALLEL WORKSHARE
parallel
parallel for
parallel sections
Specifies that one or more variables should be shared among all threads. All threads within a team access the same storage area for shared variables.
UNTIED TASK
TASKWAIT
task
taskwait
Specifies that any thread in the team can resume the task region after a suspension.
UPDATE ATOMIC atomic Specifies that the atomic action is updating a value.
WRITE ATOMIC atomic Specifies that the atomic action is writing a value.

5.8. Runtime Library Routines

User-callable functions are available to the programmer to query and alter the parallel execution environment.

Any C/C++ program unit that invokes these functions should include the statement #include <omp.h>. The omp.h header file contains definitions for each of the C/C++ library routines and the required type definitions. For example, to use the omp_get_num_threads function, use this syntax:

#include <omp.h>
int omp_get_num_threads(void);
Note: Unlimited OpenMP thread counts are available in all PGI configurations. The number of threads is unlicensed in the OpenMP runtime libraries – up to the hard limit of 256 threads. The OpenPOWER compiler relies on the llvm OpenMP runtime, which has a maximum of 231 threads.

The following table summarizes the runtime library calls.

Note: The Fortran call is shown first followed by the equivalent C/C++ call.
Table 13. Runtime Library Routines Summary
Runtime Library Routines with Examples
omp_get_num_threads
Returns the number of threads in the team executing the parallel region from which it is called. When called from a serial region, this function returns 1. A nested parallel region is the same as a single parallel region. By default, the value returned by this function is equal to the value of the environment variable OMP_NUM_THREADS or to the value set by the last previous call to omp_set_num_threads().
Fortran
integer function omp_get_num_threads()
C/C++
int omp_get_num_threads(void); 
omp_set_num_threads
Sets the number of threads to use for the next parallel region.

This subroutine or function can only be called from a serial region of code. If it is called from within a parallel region, or from within a subroutine or function that is called from within a parallel region, the results are undefined. Further, this subroutine or function has precedence over the OMP_NUM_THREADS environment variable.

Fortran
subroutine omp_set_num_threads(scalar_integer_exp)
C/C++
void omp_set_num_threads(int num_threads);
omp_get_thread_num

Returns the thread number within the team. The thread number lies between 0 and omp_get_num_threads()-1. When called from a serial region, this function returns 0. A nested parallel region is the same as a single parallel region.

Fortran
integer function omp_get_thread_num()
C/C++
int omp_get_thread_num(void); 
omp_get_ancestor_thread_num

Returns, for a given nested level of the current thread, the thread number of the ancestor.

Fortran
integer function omp_get_ancestor_thread_num(level)
    integer level
C/C++
int omp_get_ancestor_thread_num(int level); 
omp_get_active_level
Returns the number of enclosing active parallel regions enclosing the task that contains the call.
Fortran
integer function omp_get_active_level()
C/C++
int omp_get_active_level(void);
omp_get_level
Returns the number of parallel regions enclosing the task that contains the call.
Fortran
integer function omp_get_level()
C/C++
int omp_get_level(void);
omp_get_max_threads
Returns the maximum value that can be returned by calls to omp_get_num_threads().

If omp_set_num_threads() is used to change the number of processors, subsequent calls to omp_get_max_threads() return the new value. Further, this function returns the maximum value whether executing from a parallel or serial region of code.

Fortran
integer function omp_get_max_threads()
C/C++
int omp_get_max_threads(void);
omp_get_num_procs
Returns the number of processors that are available to the program
Fortran
integer function omp_get_num_procs()
C/C++
int omp_get_num_procs(void);
omp_get_stack_size
Returns the value of the OpenMP internal control variable that specifies the size that is used to create a stack for a newly created thread.

This value may not be the size of the stack of the current thread.

Fortran
!omp_get_stack_size interface
 function omp_get_stack_size ()
 use omp_lib_kinds
 integer ( kind=OMP_STACK_SIZE_KIND ) 
 :: omp_get_stack_size
 end function omp_get_stack_size
 end interface
C/C++
size_t omp_get_stack_size(void);
omp_set_stack_size
Changes the value of the OpenMP internal control variable that specifies the size to be used to create a stack for a newly created thread.

The integer argument specifies the stack size in kilobytes. The size of the stack of the current thread cannot be changed. In the PGI implementation, all OpenMP or auto-parallelization threads are created just prior to the first parallel region; therefore, only calls to omp_set_stack_size() that occur prior to the first region have an effect.

Fortran
subroutine omp_set_stack_size(integer(KIND=OMP_STACK_SIZE_KIND))
C/C++
void omp_set_stack_size(size_t stack_size);
omp_get_team_size
Returns, for a given nested level of the current thread, the size of the thread team to which the ancestor belongs.
Fortran
integer function omp_get_team_size (level)
integer level
C/C++
int omp_get_team_size(int level);
omp_in_final
Returns whether or not the call is within a final task.

Returns .TRUE. for directives and non-zero for pragmas if called from within a final task region.

Fortran
integer function omp_in_final()
C/C++
int omp_in_final(void);
omp_in_parallel
Returns whether or not the call is within a parallel region.

Returns .TRUE. for directives and non-zero for pragmas if called from within a parallel region and .FALSE. for directives and zero for pragmas if called outside of a parallel region. When called from within a parallel region that is serialized, for example in the presence of an IF clause evaluating .FALSE. for directives and zero for pragmas, the function returns .FALSE. for directives and zero for pragmas.

Fortran
logical function omp_in_parallel()
C/C++
int omp_in_parallel(void);
omp_set_dynamic
Allows automatic dynamic adjustment of the number of threads used for execution of parallel regions.

This function is recognized, but currently has no effect.

Fortran
subroutine omp_set_dynamic(scalar_logical_exp)
C/C++
void omp_set_dynamic(int dynamic_threads);
omp_get_dynamic
Allows the user to query whether automatic dynamic adjustment of the number of threads used for execution of parallel regions is enabled.

This function is recognized, but currently always returns .FALSE. for directives and zero for pragmas.

Fortran
logical function omp_get_dynamic()
C/C++
void omp_get_dynamic(void);
omp_set_nested
Allows enabling/disabling of nested parallel regions.
Fortran
subroutine omp_set_nested(nested)
logical nested
C/C++
void omp_set_nested(int nested);
omp_get_nested
Allows the user to query whether dynamic adjustment of the number of threads available for execution of parallel regions is enabled.
Fortran
logical function omp_get_nested()
C/C++
int omp_get_nested(void);
omp_set_schedule
Set the value of the run_sched_var.
Fortran
subroutine omp_set_schedule(kind, modifier)
   include 'omp_lib_kinds.h'
   integer (kind=omp_sched_kind) kind
   integer modifier
C/C++
void omp_set_schedule(omp_sched_t kind, int chunk_size);
omp_get_schedule
Retrieve the value of the run_sched_var.
Fortran
subroutine omp_get_schedule(kind, modifier)
   include 'omp_lib_kinds.h'
   integer (kind=omp_sched_kind) kind
   integer modifier
C/C++
void omp_get_schedule(omp_sched_t *kind, int *chunk_size);
omp_get_wtime
Returns the elapsed wall clock time, in seconds, as a DOUBLE PRECISION value for directives and as a floating-point double value for pragmas.

Times returned are per-thread times, and are not necessarily globally consistent across all threads.

Fortran
double precision function omp_get_wtime()
C/C++
double omp_get_wtime(void);
omp_get_wtick
Returns the resolution of omp_get_wtime(), in seconds, as a DOUBLE PRECISION value for Fortran directives and as a floating-point double value for C/C++ pragmas.
Fortran
double precision function omp_get_wtick()
C/C++
double omp_get_wtick();
omp_init_lock
Initializes a lock associated with the variable lock for use in subsequent calls to lock routines.

The initial state of the lock is unlocked. If the variable is already associated with a lock, it is illegal to make a call to this routine.

Fortran
subroutine omp_init_lock(lock)
   include 'omp_lib_kinds.h'
   integer(kind=omp_lock_kind) lock
C/C++
void omp_init_lock(omp_lock_t *lock);
void omp_init_nest_lock(omp_nest_lock_t *lock);
omp_destroy_lock
Disassociates a lock associated with the variable.
Fortran
subroutine omp_destroy_lock(lock)
   include 'omp_lib_kinds.h'
   integer(kind=omp_lock_kind) lock
C/C++
void omp_destroy_lock(omp_lock_t *lock);
void omp_destroy_nest_lock(omp_nest_lock_t *lock);
omp_set_lock
Causes the calling thread to wait until the specified lock is available.

The thread gains ownership of the lock when it is available. If the variable is not already associated with a lock, it is illegal to make a call to this routine.

Fortran
subroutine omp_set_lock(lock)
   include 'omp_lib_kinds.h'
   integer(kind=omp_lock_kind) lock
C/C++
void omp_set_lock(omp_lock_t *lock);
void omp_set_nest_lock(omp_nest_lock_t *lock);
omp_unset_lock
Causes the calling thread to release ownership of the lock associated with integer_var.

If the variable is not already associated with a lock, it is illegal to make a call to this routine.

Fortran
subroutine omp_unset_lock(lock)
   include 'omp_lib_kinds.h'
   integer(kind=omp_lock_kind) lock
C/C++
void omp_unset_lock(omp_lock_t *lock);
void omp_unset_nest_lock(omp_nest_lock_t *lock);
omp_test_lock
Causes the calling thread to try to gain ownership of the lock associated with the variable.

The function returns .TRUE. for directives and non-zero for pragmas if the thread gains ownership of the lock; otherwise, it returns .FALSE. for directives and zero for pragmas.

If the variable is not already associated with a lock, it is illegal to make a call to this routine.

Fortran
logical function omp_test_lock(lock)
   include 'omp_lib_kinds.h'
   integer(kind=omp_lock_kind) lock
C/C++
int omp_test_lock(omp_lock_t *lock);
int omp_test_nest_lock(omp_nest_lock_t *lock);

5.9. Environment Variables

You can use OpenMP environment variables to control the behavior of OpenMP programs. These environment variables allow you to set and pass information that can alter the behavior of directives and pragmas.

The following summary table is a quick reference for the OpenMP environment variables that PGI uses.

Table 14. OpenMP-related Environment Variable Summary Table
Environment Variable Default Description
OMP_DYNAMIC FALSE Currently has no effect.
Typically enables (TRUE) or disables (FALSE) the dynamic adjustment of the number of threads.
OMP_MAX_ACTIVE_LEVELS 1 Specifies the maximum number of nested parallel regions.
OMP_NESTED FALSE Enables (TRUE) or disables (FALSE) nested parallelism.
OMP_NUM_THREADS 1 Specifies the number of threads to use during execution of parallel regions at the corresponding nested level. For example, OMP_NUM_THREADS=4,2 uses 4 threads at the first nested parallel level, and 2 at the next nested parallel level.
OMP_SCHEDULE STATIC with chunk size of 1 Specifies the type of iteration scheduling and optionally the chunk size to use for omp for and omp parallel for loops that include the runtime schedule clause. The supported schedule types, which can be specified in upper- or lower-case are static, dynamic, guided, and auto.
OMP_PROC_BIND FALSE Specifies whether executing threads should be bound to a core during execution. Allowable values are "true" and "false".
OMP_STACKSIZE   Overrides the default stack size for a newly created thread.
OMP_THREAD_LIMIT 64 Specifies the absolute maximum number of threads that can be used in a program.
OMP_WAIT_POLICY ACTIVE Sets the behavior of idle threads, defining whether they spin or sleep when idle. The values are ACTIVE and PASSIVE.

6. Using MPI

Message Passing Interface (MPI) is an industry-standard application programming interface designed for rapid data exchange between processors in a cluster application. MPI is computer software used in computer clusters that allows the processes of a parallel application to communicate with one another.

PGI provides MPI support with PGI compilers and tools. PGI compilers provide explicit support to build MPI applications on Windows using Microsoft’s implementation of MPI, MS-MPI, on macOS using MPICH, and on Linux using MPICH, MVAPICH2, and Open MPI. Of course, you may always build using an arbitrary version of MPI; to do this, use the -I, -L, and -l option.

PGI products for Linux include Open MPI, PGI products for macOS includes MPICH, and PGI products for Windows includes MS-MPI. PGI users with permanent licenses can download PGI-built versions of MVAPICH2 and MPICH from pgroup.com. This section describes how to use the MPI capabilities of PGI compilers and how to configure PGI compilers so these capabilities can be used with custom MPI installations.

The debugger is enabled to support MPI applications running locally with a limited number of processes. The PGI Debugger Guide describes the MPI-enabled tools in detail.

6.1. MPI Overview

This section contains general information applicable to various MPI distributions. For distribution-specific information, refer to the sections later in this section.

MPI is a set of function calls and libraries that are used to send messages between multiple processes. These processes can be located on the same system or on a collection of distributed servers. Unlike OpenMP, the distributed nature of MPI allows it to work in almost any parallel environment.

6.3. Debugging MPI Applications

The PGI debugger provides support for symbolic debugging of MPI applications. The number and location of processes that can be debugged is limited by your license. PGI node-locked licenses limit processes to a single system whereas PGI network floating licenses support general development on clusters.

For all distributions of MPI except MPICH1, support for which has been deprecated, and SGI MPI, you can initiate an MPI debugging session from either the command line or from within the debugger. For MPICH1 and SGI MPI, debugging must be initiated at the command line. For specific information on how to initiate a debugging session for a particular version of MPI, refer to the PGI Debugger Guide.

The debugger can display the contents of message queues for instances of MPI that have been configured to support that feature. The version of MPICH1 provided with PGI products for Linux and macOS is configured with this support. At this time, MS-MPI does not support displaying message queue contents.

For more information on MPI and displaying message queues, refer to the documentation for your specific distribution of MPI.

6.4. Using MPICH on Linux and Apple macOS

PGI products for Linux and macOS include MPI header files, libraries and tools required to compile, execute and debug MPI programs. PGI compilers can be installed on a single system, and that system can be treated as if it is a small cluster by using this version of MPI.

Example

MPI Hello World Example

The following MPI example program uses Open MPI.

$ cd my_example_dir
$ cp -r $PGI/linux86-64/17.7/EXAMPLES/MPI/mpihello .
$ cd mpihello
$ export PATH=$PGI/linux86-64/2017/mpi/openmpi/bin:$PATH
$ mpif90 mpihello.f -o mpihello
$ mpiexec mpihello
Hello world! I'm node 0
$ mpiexec -np 4 mpihello
Hello world! I'm node 0
Hello world! I'm node 2
Hello world! I'm node 1
Hello world! I'm node 3

6.5. Using MPICH1, MPICH2, and MVAPICH1 on Linux

The -Mmpi sub-options mpich1, mpich2 and mvapich1 have been deprecated in the PGI 2017 release and will be disabled in a future release. To use these deprecated options, you must first set the environment variable MPIDIR to the root of the MPI installation directory you intend to use, that is, the directory that contains bin, include, lib and so on. If you do not set MPIDIR, you will see an error like the following:

$ pgfortran -Mmpi=mpich2 pi.f90 -o fpi
-Mmpi=mpich2 has been deprecated and will be removed in a future release
pgfortran-Error-The environment variable $MPIDIR must be set to use MPICH2

Once MPIDIR is set, you can use -⁠Mmpi with the mpich1, mpich2 and mvapich1 sub-options as you have in previous releases.

6.6. Using MVAPICH2 on Linux

PGI users with permanent licenses can download a PGI-built version of MVAPICH2 for Linux that includes everything required to compile, execute and debug MPI programs using MVAPICH2.

To build an application using MVAPICH2, use the MVAPICH2 compiler wrappers: mpicc, mpic++, mpif77, and mpif90. These wrappers autonmatically set up the compiler commands with the correct include file search paths, library directories, and link libraries.

To build an application using MVAPICH2 for debugging, add -⁠g to the compiler wrapper command line arguments.

6.7. Using MS-MPI on Windows

PGI products on Windows include a version of Microsoft’s MPI. You can compile, run, and debug locally on your system using this instance of MS-MPI.

To compile the application, use the -Mmpi=msmpi option. This option automatically sets up the appropriate include and library paths to use the MS-MPI headers and libraries. To compile for debugging, add -⁠g.

6.8. Using Open MPI on Linux

PGI products for Linux ship with a PGI-built version of Open MPI that includes everything required to compile, execute and debug MPI programs using Open MPI.

To build an application using Open MPI, use the Open MPI compiler wrappers: mpicc, mpic⁠+⁠+, mpif77, and mpif90. These wrappers automatically set up the compiler commands with the correct include file search paths, library directories, and link libraries.

To build an application using Open MPI for debugging, add -⁠g to the compiler wrapper command line arguments.

6.9. Using SGI MPI on Linux

PGI compilers and tools support SGI’s version of MPI. If you want to build your MPI application using an instance of SGI MPI, you must first set either MPIDIR or MPI_ROOT to the root of the SGI MPI installation directory that you want to use, that is, the directory that contains bin, include, lib, and so on. Then you can compile with the -Mmpi=sgimpi option. Add -g for debugging. To run programs built with SGI MPI, you must include the SGI MPI lib directory in the LD_LIBRARY_PATH environment variable.

6.10. Using MPI Compiler Wrappers

When you use MPI compiler wrappers to build with the -fpic or -mcmodel=medium options, then you must specify -shlib to link with the correct libraries. Here are a few examples:

For a static link to the MPI libraries, use this command:

% mpicc hello.f

For a dynamic link to the MPI libraries, use this command:

% mpicc hello.f -shlib

To compile with -fpic, which, by default, invokes dynamic linking, use this command:

% mpicc -fpic -shlib hello.f

To compile with -mcmodel=medium, use this command:

% mpicc -mcmodel=medium -shlib hello.f

6.11. Limitations

The Open Source Cluster utilities, in particular the MPICH and ScaLAPACK libraries, are provided with support necessary to build and define their proper use. However, use of these libraries on linux86-64 systems is subject to the following limitations:

  • MPI libraries are limited to Messages of length < 2GB, and integer arguments are INTEGER*4 in FORTRAN, and int in C.
  • Integer arguments for ScaLAPACK libraries are INTEGER*4 in FORTRAN, and int in C.
  • Arrays passed must be < 2GB in size.

6.12. Testing and Benchmarking

The Examples directory contains various benchmarks and tests. Copy this directory into a local working directory by issuing the following command:

% cp -r $PGI/linux86/17.7/EXAMPLES/MPI .

NAS Parallel Benchmarks

The NPB2.3 subdirectory contains version 2.3 of the NAS Parallel Benchmarks in MPI. Issue the following commands to run the BT benchmark on four nodes of your cluster:

% cd MPI/NPB2.3/BT
% make BT NPROCS=4 CLASS=W
% cd ../bin
% mpirun -np 4 bt.W.4

There are several other NAS parallel benchmarks available in this directory. Similar commands are used to build and run each of them. If you want to run a larger problem, try building the Class A version of BT by substituting "A" for "W" in the previous commands.

ScaLAPACK

The ScaLAPACK test times execution of the 3D PBLAS (parallel BLAS) on your cluster. To run this test, execute the following commands:

% cd scalapack
% make
% mpirun -np 4 pdbla3tim

7. Using an Accelerator

An accelerator is a special-purpose co-processor attached to a CPU and to which the CPU can offload data and executable kernels to perform compute-intensive calculations. This section describes a collection of compiler directives used to specify regions of code in Fortran and C programs that can be offloaded from a host CPU to an attached accelerator.

7.1. Overview

The programming model and directives described in this section allow programmers to create high-level host+accelerator programs without the need to explicitly initialize the accelerator, manage data or program transfers between the host and accelerator, or initiate accelerator startup and shutdown. Rather, all of these details are implicit in the programming model and are managed by the PGI Fortran , C, and C++ accelerator compilers.

The method described provides a model for accelerator programming that is portable across operating systems and various host CPUs and accelerators. The directives allow a programmer to migrate applications incrementally to accelerator targets using standards-compliant Fortran or C.

This programming model allows the programmer to augment information available to the compilers, including specification of data local to an accelerator region, guidance on mapping of loops onto an accelerator, and similar performance-related details.

7.1.1. Components

The PGI Accelerator compiler technology includes the following components:

  • PGFORTRAN™ native OpenMP and OpenACC Fortran 2003 compiler.
  • PGCC® native OpenMP and OpenACC ANSI C11 and K&R C compiler.
  • PGC++® native OpenMP and OpenACC ANSI C++14 compiler.
  • PGI profiler for OpenACC, CUDA, OpenMP, and multi-thread programs.
  • NVIDIA CUDA Toolkit components
  • A simple command-line tool to detect whether the system has an appropriate GPU or accelerator card

No accelerator-enabled debugger is included with this release

7.1.2. Availability

The PGI 17.7 Fortran & C Accelerator compilers are available on x86-64 processor-based workstations and servers with an attached NVIDIA CUDA-enabled GPU or Tesla card. These compilers target all platforms that PGI supports. All examples included in this section are developed and presented on such a platform. For a list of supported GPUs, refer to the ‘Accelerator Installation and Supported Platforms’ list in the latest PGI Release Notes.

7.1.3. User-directed Accelerator Programming

In user-directed accelerator programming the user specifies the regions of a host program to be targeted for offloading to an accelerator device. The bulk of a user’s program, as well as regions containing constructs that are not supported on the targeted accelerator, are executed on the host. This section concentrates on specification of loops and regions of code to be offloaded to an accelerator.

7.1.4. Features Not Covered or Implemented

This section does not describe features or limitations of the host programming environment as a whole. Further, it does not cover automatic detection and offloading of regions of code to an accelerator by a compiler or other tool. While future versions of the PGI compilers may allow for automatic offloading, this feature is not currently supported.

7.2. Terminology

Clear and consistent terminology is important in describing any programming model. This section provides definitions of the terms required for you to effectively use this section and the associated programming model.

Accelerator
a special-purpose co-processor attached to a CPU and to which the CPU can offload data and executable kernels to perform compute-intensive calculations.
Compute intensity
for a given loop, region, or program unit, the ratio of the number of arithmetic operations performed on computed data divided by the number of memory transfers required to move that data between two levels of a memory hierarchy.
Compute region
a structure block defined by an OpenACC compute construct. A compute construct is a structured block containing loops which are compiled for the accelerator. A compute region may require device memory to be allocated and data to be copied from host to device upon region entry, and data to be copied from device to host memory and device memory deallocated upon exit. The dynamic range of a compute construct, including any code in procedures called from within the construct, is the compute region. In this release, compute regions may not contain other compute regions or data regions.
Construct
a structured block identified by the programmer or implicitly defined by the language. Certain actions may occur when program execution reaches the start and end of a construct, such as device memory allocation or data movement between the host and device memory. Loops in a compute construct are targeted for execution on the accelerator. The dynamic range of a construct including any code in procedures called from within the construct, is called a region.
CUDA
stands for Compute Unified Device Architecture; the CUDA environment from NVIDIA is a C-like programming environment used to explicitly control and program an NVIDIA GPU.
Data region
a region defined by an OpenACC data construct, or an implicit data region for a function or subroutine containing OpenACC directives. Data regions typically require device memory to be allocated and data to be copied from host to device memory upon entry, and data to be copied from device to host memory and device memory deallocated upon exit. Data regions may contain other data regions and compute regions.
Device
a general reference to any type of accelerator.
Device memory
memory attached to an accelerator which is physically separate from the host memory.
Directive
in C, a #pragma, or in Fortran, a specially formatted comment statement that is interpreted by a compiler to augment information about or specify the behavior of the program.
DMA
Direct Memory Access, a method to move data between physically separate memories; this is typically performed by a DMA engine, separate from the host CPU, that can access the host physical memory as well as an IO device or GPU physical memory.
GPU
a Graphics Processing Unit; one type of accelerator device.
GPGPU
General Purpose computation on Graphics Processing Units.
Host
the main CPU that in this context has an attached accelerator device. The host CPU controls the program regions and data loaded into and executed on the device.
Loop trip count
the number of times a particular loop executes.
OpenACC
a parallel programming standard describing a set of compiler directives which can be applied to standard C, C++, and Fortran to specify regions of code for offloading from a host CPU to an attached accelerator.
Private data
with respect to an iterative loop, data which is used only during a particular loop iteration. With respect to a more general region of code, data which is used within the region but is not initialized prior to the region and is re-initialized prior to any use after the region.
Region
the dynamic range of a construct, including any procedures invoked from within the construct.
Structured block
in C, an executable statement, possibly compound, with a single entry at the top and a single exit at the bottom. In Fortran, a block of executable statements with a single entry at the top and a single exit at the bottom.
Vector operation
a single operation or sequence of operations applied uniformly to each element of an array.
Visible device copy
a copy of a variable, array, or subarray allocated in device memory, that is visible to the program unit being compiled.

7.3. System Requirements

For NVIDIA GPUs

To use the PGI Accelerator compiler features on NVIDIA GPUs, you must install the NVIDIA drivers. You may download these components from the NVIDIA website at http://www.nvidia.com/cuda

These are not PGI products. They are licensed and supported by NVIDIA.

Note: You must be using an operating system that is supported by both the current PGI release and by the CUDA software and drivers.

7.4. Supported Processors and GPUs

This PGI Accelerator compiler release supports all AMD64 and Intel 64 host processors. Use the -⁠tp=<target> flag as documented in the release to specify the target processor.

Use the -⁠acc flag to enable OpenACC directives and the -⁠ta=<target> flag to target NVIDIA GPUs. You can then use the generated code on any supported system with CUDA installed that has a CUDA-enabled GeForce, Quadro, or Tesla card.

For more information on these flags as they relate to accelerator technology, refer to Applicable Command-line Options.

For a complete list of supported CUDA GPUs, refer to the NVIDIA website at: http://www.nvidia.com/object/cuda_learn_products.html

You can detect whether the system has CUDA properly installed and has an attached GPU by running the pgaccelinfo command, which is delivered as part of the PGI compilers software package.

7.5. Installation and Licensing

Note: The PGI Accelerator compilers have a different license key than the -x64 only version of the PGI products.

7.5.1. Required Files

Note: If you are installing on Windows, the required files are built for you.

The default NVIDIA Compute Capability for generated code in this release is cc2+ or fermi+, enabling code generation for NVIDIA Fermi and Kepler GPUs. You can use the -⁠ta flag to specify other compute capabilities.

You can also change the default to one or more of the supported compute capabilities by adding a line similar to the following one to the sitenvrc file. This example sets the compute capability to enable code generation for all of the supported compute capabilities. Notice that the compute capabilities are separated by a space.

set COMPUTECAP=20 30 35 50;

Place the sitenvrc file in the following directory, where $PGI is the PGI installation directory, which is typically /opt/pgi or /usr/pgi.

$PGI/linux86-64/17.7/bin/ 

7.5.2. Command-line Flag

After acquiring the PGI Accelerator compilers license key, you can use the -⁠acc or -⁠ta option with the pgfortran, pgcc or pgc++ commands.

For more information on the -⁠ta flag and the suboptions that relate to the target accelerators, refer to Applicable Command-line Options.

The compiler automatically invokes the necessary CUDA software tools to create the kernel code and embeds the kernels in the Linux object file.

Note: To access the accelerator libraries, you must link an accelerator program with the -⁠ta flag as well.

7.6. Execution Model

The execution model targeted by the PGI Accelerator compilers is host-directed execution with an attached accelerator device, such as a GPU. The bulk of a user application executes on the host. Compute intensive regions are offloaded to the accelerator device under control of the host. The accelerator device executes kernels, which may be as simple as a tightly-nested loop, or as complex as a subroutine, depending on the accelerator hardware.

7.6.1. Host Functions

Even in accelerator-targeted regions, the host must orchestrate the execution; it

  • allocates memory on the accelerator device
  • initiates data transfer
  • sends the kernel code to the accelerator
  • passes kernel arguments
  • queues the kernel
  • waits for completion
  • transfers results back to the host
  • deallocates memory
Note: In most cases, the host can queue a sequence of kernels to be executed on the device, one after the other.

7.6.2. Levels of Parallelism

Most current GPUs support two levels of parallelism:

  • an outer doall (fully parallel) loop level
  • an inner synchronous (SIMD or vector) loop level

Each level can be multidimensional with 2 or 3 dimensions, but the domain must be strictly rectangular. The synchronous level may not be fully implemented with SIMD or vector operations, so explicit synchronization is supported and required across this level. No synchronization is supported between parallel threads across the doall level.

The execution model on the device side exposes these two levels of parallelism and the programmer is required to understand the difference between, for example, a fully parallel loop and a loop that is vectorizable but requires synchronization across iterations. All fully parallel loops can be scheduled for either doall or synchronous parallel execution, but by definition SIMD vector loops that require synchronization can only be scheduled for synchronous parallel execution.

7.7. Memory Model

The most significant difference between a host-only program and a host+accelerator program is that the memory on the accelerator can be completely separate from host memory, which is the case on most current GPUs. For example:

  • The host cannot read or write accelerator memory by reference because it is not mapped into the virtual memory space of the host.
  • All data movement between host memory and accelerator memory must be performed by the host through runtime library calls that explicitly move data between the separate memories.
  • It is not valid to assume the accelerator can read or write host memory, though this may be supported by accelerators in the future.

7.7.1. Separate Host and Accelerator Memory Considerations

The programmer must be aware of the potentially separate memories for many reasons, including but not limited to:

  • Memory bandwidth between host memory and accelerator memory determines the compute intensity required to effectively accelerate a given region of code.
  • Limited size of accelerator memory may prohibit offloading of regions of code that operate on very large amounts of data.

7.7.2. Accelerator Memory

On the accelerator side, current GPUs implement a weak memory model. In particular, they do not support memory coherence between threads unless those threads are parallel only at the synchronous level and the memory operations are separated by an explicit barrier. Otherwise, if one thread updates a memory location and another reads the same location, or two threads store a value to the same location, the hardware does not guarantee the results. While the results of running such a program might be inconsistent, it is not accurate to say that the results are incorrect. By definition, such programs are defined as being in error. While a compiler can detect some potential errors of this nature, it is nonetheless possible to write an accelerator region that produces inconsistent numerical results.

7.7.3. Cache Management

Some current GPUs have a software-managed cache, some have hardware-managed caches, and most have hardware caches that can be used only in certain situations and are limited to read-only data. In low-level programming models such as CUDA, it is up to the programmer to manage these caches. However, in the PGI Accelerator programming model, the compiler manages these caches using hints from the programmer in the form of directives.

7.8. Running an Accelerator Program

Running a program that has accelerator directives and was compiled and linked with the -ta flag is the same as running the program compiled without the -ta flag.

  • When running programs on NVIDIA GPUs, the program looks for and dynamically loads the CUDA libraries. If the libraries are not available, or if they are in a different directory than they were when the program was compiled, you may need to append the appropriate library directory to your LD_LIBRARY_PATH environment variable on Linux or to the PATH environment variable on Windows.
  • On Linux, if you have no server running on your NVIDIA GPU, when your program reaches its first accelerator region, there may be a 0.5 to 1.5 second pause to warm up the GPU from a power-off audience. You can avoid this delay by running the pgcudainit program in the background, which keeps the GPU powered on.
  • If you compile a program for a particular accelerator type, then run the program on a system without that accelerator, or on a system where the target libraries are not in a directory where the runtime library can find them, the program may fail at runtime with an error message.
  • If you set the environment variable PGI_ACC_NOTIFY to a nonzero integer value, the runtime library prints a line to standard error every time it launches a kernel on the accelerator.

7.9. Accelerator Directives

This section provides an overview of the Fortran and C directives used to delineate accelerator regions and to augment information available to the compiler for scheduling of loops and classification of data.

7.9.1. Enable Accelerator Directives

PGI Accelerator compilers enable accelerator directives with the -⁠acc and -⁠ta command line option. For more information on this option as it relates to the Accelerator, refer to Applicable Command-line Options.

Note: The syntax used to define directives allows compilers to ignore accelerator directives if support is disabled or not provided.

_ACCEL macro

The _ACCEL macro name is defined to have a value yyyymm where yyyy is the year and mm is the month designation of the version of the Accelerator directives supported by the implementation. For example, the version for May, 2009 is 200905. The PGI compilers define this macro when accelerator directives are enabled.

_OPENACC macro

The _OPENACC macro name is defined to have a value yyyymm where yyyy is the year and mm is the month designation of the version of the OpenACC directives supported by the implementation. For example, the version for June, 2013 is 201306. All OpenACC compilers define this macro when OpenACC directives are enabled.

7.9.2. Format

The specific format of the directive depends on the language and the format or form of the source.

Directives include a name and clauses, and the format of the directive depends on the type:

  • C directives, described in ‘C Directives’
  • Free-form Fortran directives, described in ‘Free-Form Fortran Directives’
  • Fixed-form Fortran directives, described in ‘Fixed-Form Fortran Directives’
Note: This document uses free form for all PGI Accelerator compiler Fortran directive examples.

Rules

The following rules apply to all OpenACC compiler directives:

  • Only one directive-name can be specified per directive.
  • The order in which clauses appear is not significant.
  • Clauses may be repeated unless otherwise specified.
  • For clauses that have a list argument, a list is a comma-separated list of variable names, array names, or, in some cases, subarrays with subscript ranges.

7.9.3. C Directives

In C, OpenACC compiler directives are specified using #pragma

Syntax

The syntax of an OpenACC compiler directive is:

  #pragma acc directive-name [clause [,clause]...] new-line

Rules

In addition to the general directive rules, the following rules apply to OpenACC C directives:

  • Each directive starts with #pragma acc.
  • The remainder of the directive follows the C conventions for pragmas.
  • White space may be used before and after the #; white space may be required to separate words in a directive.
  • Preprocessing tokens following the #pragma acc are subject to macro replacement.
  • C directives are case sensitive.
  • An Accelerator directive applies to the immediately following structured block or loop.

7.9.4. Free-Form Fortran Directives

OpenACC Fortran directives can be either Free-Form or Fixed-Form directives. Free-Form Accelerator directives are specified with the !$acc mechanism.

Syntax

The syntax of directives in free-form source files is:

  !$acc directive-name [clause [,clause]...]

Rules

In addition to the general directive rules, the following rules apply to OpenACC Free-Form Fortran directives:

  • The comment prefix (!) may appear in any column, but may only be preceded by white space (spaces and tabs).
  • The sentinel (!$acc) must appear as a single word, with no intervening white space.
  • Line length, white space, and continuation rules apply to the directive line.
  • Initial directive lines must have a space after the sentinel.
  • Continued directive lines must have an ampersand (&) as the last nonblank character on the line, prior to any comment placed in the directive.
  • Comments may appear on the same line as the directive, starting with an exclamation point and extending to the end of the line.
  • If the first nonblank character after the sentinel is an exclamation point, the line is ignored.
  • Directives are case-insensitive.
  • Directives cannot be embedded within continued statements.
  • Statements must not be embedded within continued directives.

7.9.5. Fixed-Form Fortran Directives

Fixed-Form Accelerator directives are specified using one of three formats.

Syntax

The syntax of directives in fixed-form source files is one these three formats:

  !$acc directive-name [clause [,clause]...]
  c$acc directive-name [clause [,clause]...]
  *$acc directive-name [clause [,clause]...]

Rules

In addition to the general directive rules, the following rules apply to Accelerator Fixed-Form Fortran directives:

  • The sentinel (!$acc, c$acc, or *$acc) must occupy columns 1-5.
  • Fixed form line length, white space, continuation, and column rules apply to the directive line.
  • Initial directive lines must have a space or zero in column 6, and continuation directive lines must have a character other than a space or zero in column 6.
  • Comments may appear on the same line as a directive, starting with an exclamation point on or after column 7 and continuing to the end of the line.
  • Directives are case-insensitive.
  • Directives cannot be embedded within continued statements.
  • Statements must not be embedded within continued directives.

7.9.6. OpenACC Directive Summary

PGI currently supports these types of accelerator directives:

  • Parallel Directive
  • Kernels Directive
  • Loop Directive
  • Combined Directive
  • Data Directive
  • Enter Data and Exit Data Directives
  • Host_Data Directive
  • Cache Directive
  • Declare Directive
  • Update Directive
  • Routine Directive
  • Wait Directive

Table 16 lists and briefly describes each of the accelerator directives that PGI currently supports. For a complete description of each directive, refer to ‘PGI Accelerator Directives’ in the PGI Compiler Reference Manual.

Table 16. PGI Accelerator Directive Summary Table
This directive... Accepts these clauses... Has this functionality...
Parallel Directive
  • async [( int-expr )]
  • wait [( int-expr-list )]
  • num_gangs ( int-expr )
  • num_workers( int-expr )
  • vector_length( int-expr )
  • if( condition )
  • reduction( operator : list )
  • copy ( list )
  • copyin( list )
  • copyout( list )
  • create( list )
  • present( list )
  • deviceptr( list )
  • private( list )
  • firstprivate( list )
Defines the region of the program that should be compiled for parallel execution on the accelerator device.
C Syntax
#pragma acc parallel [clause [, clause]...] new-line
    structured block

Fortran Syntax

!$acc parallel [clause [, clause]...]
    structured block
!$acc end parallel
Kernels Directive
  • async [( int-expr )]
  • wait [( int-expr-list )]
  • if( condition )
  • copy ( list )
  • copyin( list )
  • copyout( list )
  • create( list )
  • present( list )
  • deviceptr( list )
Defines the region of the program that should be compiled into a sequence of kernels for execution on the accelerator device.
C Syntax
#pragma acc kernels [clause [, clause]...] new-line
    structured block

Fortran Syntax

!$acc kernels [clause [, clause]...]
    structured block
!$acc end kernels
Data Directive
  • if( condition )
  • copy ( list )
  • copyin( list )
  • copyout( list )
  • create( list )
  • deviceptr( list )
Defines data, typically arrays, that should be allocated in the device memory for the duration of the data region, whether data should be copied from the host to the device memory upon region entry, and copied from the device to host memory upon region exit.
C Syntax
#pragma acc data [clause [, clause]...] new-line
    structured block
Fortran Syntax
!$acc data [clause [, clause]...]
    structured block
!$acc end data
Enter Data Directive
  • if( condition )
  • async [( int-expr )]
  • wait [( int-expr-list )]
  • copyin( list )
  • create( list )
Defines data, typically arrays, that should be allocated in the device memory for the duration of the program or until an exit data regdirective that deallocates the data, and whether data should be copied from the host to the device memory at the enter data directive.
C Syntax
#pragma acc enter data [clause [, clause]...] new-line

FortranSyntax

!$acc enter data [clause [, clause]...]
Exit Data Directive
  • if( condition )
  • async [( int-expr )]
  • wait [( int-expr-list )]
  • copyout( list )
  • delete( list )
Defines data, typically arrays, that should be allocated in the device memory, and whether data should be copied from the device to the host memory.
C Syntax
#pragma acc exit data [clause [, clause]...] new-line

Fortran Syntax

!$acc exit data [clause [, clause]...]
Host_Data Directive
  • use_device( list )
Makes the address of the device data available on the host.
C Syntax
#pragma acc host_data [clause [, clause]...] new-line
    structured block

Fortran Syntax

!$acc host_data [clause [, clause]...]
    structured block
!$acc end host_data
Loop Directive
  • collapse( n )
  • gang [( gang-arg-list )]
  • worker [( [num:] int-expr )]
  • vector [( [length:] int-expr )]
  • seq
  • independent
  • private( list )
  • reduction( operator : list )
  • shortloop
Describes what type of parallelism to use to execute the loop and declare loop-private variables and arrays and reduction operations. Applies to a loop which must appear on the following line.
C Syntax
#pragma acc loop [clause [,clause]...] new-line
for loop

Fortran Syntax

!$acc loop [clause [,clause]...]
do loop
Combined Parallel Loop Directive Any clause that is allowed on a parallel directive or a loop directive is allowed on a combined parallel loop directive. Is a shortcut for specifying a loop directive nested immediately inside an accelerator parallel directive. The meaning is identical to explicitly specifying a parallel construct containing a loop directive.
C Syntax
#pragma acc parallel loop [clause [, clause]...] new-line
for loop

Fortran Syntax

!$acc parallel loop [clause [, clause]...]
do loop
Combined Kernels Loop Directive Any clause that is allowed on a kernels directive or a loop directive is allowed on a combined kernels loop directive. Is a shortcut for specifying a loop directive nested immediately inside an accelerator kernels directive. The meaning is identical to explicitly specifying a kernels construct containing a loop directive.
C Syntax
#pragma acc kernels loop [clause [, clause]...] new-line
for loop

Fortran Syntax

!$acc kernels loop [clause [, clause]...]
do loop
Cache Directive   Specifies array elements or subarrays that should be fetched into the highest level of the cache for the body of a loop. Must appear at the top of (inside of) the loop.
C Syntax
#pragma acc cache ( list )... new-line

Fortran Syntax

!$acc cache ( list )
Declare Directive
  • copy ( list )
  • copyin( list )
  • copyout( list )
  • create( list )
  • deviceptr( list )
  • device_resident( list )

Specifies that an array or arrays are to be allocated in the device memory for the duration of the implicit data region of a function, subroutine, or program.

Specifies whether the data values are to be transferred from the host to the device memory upon entry to the implicit data region, and from the device to the host memory upon exit from the implicit data region.

Creates a visible device copy of the variable or array.

C Syntax
#pragma acc declare clause [,clause]... new-line

FortranSyntax

!$acc declare clause [,clause]...
Update Directive
  • async [( int-expr )]
  • wait [( int-expr )]
  • device-type ( devtype-list )
  • if ( condition )
  • self ( list )
  • host ( list )
  • device( list )
Used during the lifetime of accelerator data to update all or part of a host memory array with values from the corresponding array in device memory, or to update all or part of a device memory array with values from the corresponding array in host memory.
C Syntax
#pragma acc update clause [,clause]... new-line

Fortran Syntax

!$acc update clause [,clause]...
Routine Directive
  • gang
  • worker
  • vector
  • seq
Used to tell the compiler to compile a given procedure for an accelerator as well as the host. In a file or routine with a procedure call the routine directive tells the implementation the attributes of the procedure when called on the accelerator.
C Syntax
#pragma acc routine clause [,clause]... new-line
or
#pragma acc routine ( name ) clause [,clause]... new-line

Fortran Syntax

!$acc routine clause [,clause]...

or

!$acc routine( name ) clause [,clause]...
Wait Directive
  • if( condition )
  • device( list )
Specifies to wait until all operations on a specific device async queue or all async queues are complete.
C Syntax
#pragma acc wait [(int-expr-list)] [clause [,clause]...] new-line

Fortran Syntax

!$acc wait [(int-expr-list)] [clause [,clause]...]

7.10. Accelerator Directive Clauses

Table 17 provides an alphabetical listing and brief description of each clause that is applicable for the various Accelerator directives. The table also indicates for which directives the clause is applicable.

Table 17. Directive Clauses Summary
Use this clause... In these directives... To do this...
async [(int-expr)]
  • Parallel
  • Kernels
  • Data
  • Enter Data
  • Exit Data
  • Update
The parallel or kernels region or data operations may be processed asynchronously while the local thread continues with the code following the construct or directive. (The async clause on the data construct is a PGI extension)
collapse (n)
  • Loop
Specifies how many tightly nested loops are associated with the loop construct.
copy (list)
  • Parallel
  • Kernels
  • Data Region
  • Declare
At execution time, the implementation will test whether each of the items in the list are already present in the device memory. If not, corresponding device memory will be allocated and the data will be copied to the device at entry to the region, and data will be copied back to the host and the memory deallocated at exit from the region. For compatibility with OpenACC 2.0, this may also be spelled present_or_copy or pcopy.
copyin (list)
  • Parallel
  • Kernels
  • Data
  • Declare
  • Enter Data
At execution time, the implementation will test whether each of the items in the list are already present in the device memory. If not, corresponding device memory will be allocated and the data will be copied to the device at entry to the region, and the memory deallocated at exit from the region. For compatibility with OpenACC 2.0, this may also be spelled present_or_copyin or pcopyin.
copyout (list)
  • Parallel
  • Kernels
  • Data
  • Declare
  • Exit Data
At execution time, the implementation will test whether each of the items in the list are already present in the device memory. If not, corresponding device memory will be allocated at entry to the region, and data will be copied back to the host and the memory deallocated at exit from the region. For compatibility with OpenACC 2.0, this may also be spelled present_or_copyout or pcopyout.
create (list)
  • Parallel
  • Kernels
  • Data
  • Declare
  • Enter Data
At execution time, the implementation will test whether each of the items in the list are already present in the device memory. If not, corresponding device memory will be allocated at entry to the region, and the memory deallocated at exit from the region. For compatibility with OpenACC 2.0, this may also be spelled present_or_create or pcreate.
delete (list)
  • Exit Data
At execution time, the implementation will deallocate the corresponding device memory.
device (list)
  • Update
Copies the variables, arrays, or subarrays in the list argument from host memory to the visible device copy of the variables, arrays, or subarrays in device memory.
device_resident (list)
  • Declare
Specifies that the memory of the named variables should be allocated in the accelerator device memory and not in the host memory.
deviceptr (list)
  • Parallel
  • Kernels
  • Data
  • Declare
Declares that the items in the list are device pointers, so the data need not be allocated or moved between the host and device..
firstprivate (list)
  • Parallel
Declares that a copy of each item on the list will be created for each parallel gang, and that the copy will be initialized with the value of that item on the host when the parallel construct is encountered.
gang [(gang-arg-list)]
  • Loop
  • Routine
Specifies that the iterations of the associate loop or loops are to be executed in parallel by distributing the iterations among the gangs created by the parallel construct.
host (list)
  • Update
Copies the visible device copies of the variables, arrays, or subarrays in the list argument to the associated host memory locations. The copy occurs after completion of the compute or data region.
if (condition)
  • Parallel
  • Kernels
When present, tells the compiler to generate two copies of the region – one for the accelerator, one for the host – and to generate code to decide which copy to execute.
if (condition)
  • Data Region
  • Enter Data
  • Exit Data
  • Update
Conditionally allocate memory on, and move data to and/or from the device.
independent
  • Loop
Tells the compiler that the iterations of this loop are data-independent of each other, thus allowing the compiler to generate code to execute the iterations in parallel, without synchronization.
num_gangs (int-expr)
  • Parallel
Defines the number of parallel gangs that will execute the region.
num_workers (int-expr)
  • Parallel
Defines the number of workers within each gang that will be active after a gang transitions from worker-single mode to worker-partitioned mode.
present (list)
  • Parallel
  • Kernels
  • Data
  • Declare
Tells the implementation that the items in the list are already present in device memory.
private (list)
  • Loop
Specifies that a copy of each item in the list will be created for each thread that executes one or more iterations of the associated loop or loops.
private (list)
  • Parallel
Declares that a copy of each item on the list will be created for each parallel gang.
reduction (operator: list)
  • Loop
For each variable in the list, a private copy is created for each thread that executes iterations of the associated loop or loops and initialized for the operator. At the end of the loop, the values for each thread are combined using the reduction operator, and the result combined with the value of the original variable and stored in the original variable.
reduction (operator: list)
  • Parallel
For each variable in the list, a private copy is created for each parallel gang and initialized for the operator. At the end of the region, the values for each gang are combined using the reduction operator, and the result combined with the value of the original variable and stored in the original variable.
self (list)
  • Update
Specifies that the items in the list are to be copied from the accelerator device memory to the local memory. The self clause is a synonym for the host clause.
seq
  • Loop
Tells the compiler to execute this loop sequentially on the accelerator. There is no maximum number of iterations for a seq schedule.
shortloop
  • Loop
If the loop has the vector clause, this tells the compiler that the trip count for the loop is less than or equal to the number of vector lanes created for this kernel, as specified by the vector_length on the parallel construct or vector clause in a kernels construct. (PGI extension)
use_device (list)
  • Host_Data
Tells the compiler to use the device address of any item in the list in code within the construct.
vector [( [length:] int-expr)]
  • Loop
Tells the compiler to execute this loop in vector or SIMD mode on the accelerator.
vector_length (int-expr)
  • Parallel
Defines the number of vector lanes that will be active after a worker transitions from vector-single mode to vector-partitioned mode.
wait [( int-expr-list )]
  • Parallel
  • Kernels
  • Enter Data
  • Exit Data
  • Update
The compute, data or update operation may not be launched or executed until all operations enqueued up to this point by this thread on the associated asynchronous device activity queues have completed.
worker [( [num:] int-expr)]
  • Loop
Specifies that the iterations of the associated loop or loops are to be executed in parallel by distributing the iterations among the multiple workers within a single gang.

7.11. OpenAcc Runtime Libraries

This section provides an overview of the user-callable functions and library routines that are available for use by programmers to query the accelerator features and to control behavior of accelerator-enabled programs at runtime.

Note: In Fortran, none of the OpenACC runtime library routines may be called from a PURE or ELEMENTAL procedure.

7.11.1. Runtime Library Definitions

There are separate runtime library files for C and for Fortran.

C Runtime Library Files

In C, prototypes for the runtime library routines are available in a header file named accel.h. All the library routines are extern functions with ‘C’ linkage. This file defines:

  • The prototypes of all routines in this section.
  • Any data types used in those prototypes, including an enumeration type to describe types of accelerators.

Fortran Runtime Library Files

In Fortran, interface declarations are provided in a Fortran include file named accel_lib.h and in a Fortran module named accel_lib. These files define:

  • Interfaces for all routines in this section.
  • Integer parameters to define integer kinds for arguments to those routines.
  • Integer parameters to describe types of accelerators.
  • The integer parameter accel_version with a value yyyymm where yyyy and mm are the year and month designations of the version of the Accelerator programming model supported. This value matches the value of the preprocessor variable _ACCEL.

7.11.2. Runtime Library Routines

Table 18 lists and briefly describes the runtime library routines supported by PGI in addition to the standard OpenACC runtine API routines.

Table 18. Accelerator Runtime Library Routines
This Runtime Library Routine... Does this...
acc_allocs Returns the number of arrays allocated in data or compute regions.
acc_bytesalloc Returns the total bytes allocated by data or compute regions.
acc_bytesin Returns the total bytes copied in to the accelerator by data or compute regions.
acc_bytesout Returns the total bytes copied out from the accelerator by data or compute regions.
acc_copyins Returns the number of arrays copied in to the accelerator by data or compute regions.
acc_copyouts Returns the number of arrays copied out from the accelerator by data or compute regions.
acc_disable_time Tells the runtime to stop profiling accelerator regions and kernels.
acc_enable_time Tells the runtime to start profiling accelerator regions and kernels, if it is not already doing so.
acc_exec_time Returns the number of microseconds spent on the accelerator executing kernels.
acc_frees Returns the number of arrays freed or deallocated in data or compute regions.
acc_get_device Returns the type of accelerator device used to run the next accelerator region, if one is selected.
acc_get_device_num Returns the number of the device being used to execute an accelerator region.
acc_get_free_memory Returns the total available free memory on the attached accelerator device.
acc_get_memory Returns the total memory on the attached accelerator device.
acc_get_num_devices Returns the number of accelerator devices of the given type attached to the host.
acc_kernels Returns the number of accelerator kernels launched since the start of the program.
acc_present_dump Summarizes all data present on the current device.
acc_present_dump_all Summarizes all data present on all devices.
acc_regions Returns the number of accelerator regions entered since the start of the program.
acc_total_time Returns the number of microseconds spent in accelerator compute regions and in moving data for accelerator data regions.

7.12. Environment Variables

PGI supports environment variables that modify the behavior of accelerator regions. This section defines the user-setable environment variables used to control behavior of accelerator-enabled programs at execution. These environment variables must comply with these rules:

  • The names of the environment variables must be upper case.
  • The values assigned environment variables are case insensitive and may have leading and trailing white space.
  • The behavior is implementation-defined if the values of the environment variables change after the program has started, even if the program itself modifies the values.

Table 19 lists and briefly describes the Accelerator environment variables that PGI supports.

Table 19. Accelerator Environment Variables
This environment variable... Does this...
ACC_DEVICE_TYPE Controls which accelerator device to use when executing accelerator regions, if the program has been compiled to use more than one different type of device. The value of this environment variable is implementation-defined, and currently may be the string NVIDIA, or HOST.
ACC_DEVICE_NUM Controls the default device number to use when executing accelerator regions. The value of this environment variable must be a nonnegative integer between zero and the number of devices attached to the host.
PGI_ACC_BUFFERSIZE For NVIDIA CUDA devices, this defines the size of the pinned buffer used to transfer data between host and device.
PGI_ACC_CUDA_PROFSTOP Set to 1 (or any positive value) to tell the PGI runtime environment to insert an 'atexit(cuProfilerStop)' call upon exit. This behavior may be desired in the case where a profile is incomplete or where a message is issued to call cudaProfilerStop().
PGI_ACC_DEBUG Set to 1 to instruct the PGI runtime to generate information about device memory allocation, data movement, kernel launches, and more. PGI_ACC_DEBUG is designed mostly for use in debugging the runtime itself, but it may be helpful in understanding how the program interacts with the device. Expect copious amounts of output.
PGI_ACC_GANGLIMIT When set to any value greater than zero, limit the number of gangs (CUDA thread blocks) that will be launched by a kernel to that value. Set to zero to allow an unlimited number of gangs.
PGI_ACC_NOTIFY When set to an integer value, the value is used as a bit mask to print information about kernel launches (value 1), data transfers (value 2), wait operations or synchronizations with the device (value 4), region entry/exit (value 8), and data allocate/free (value 16).
PGI_ACC_TIME Enables a lightweight profiler to measure data movement and accelerator kernel execution time and print a summary at the end of program execution.

7.13. Applicable Command-line Options

The following command line options are applicable specifically when working with accelerators.

-ta
Use this option to enable recognition of the !$ACC directives in Fortran, and #pragma acc directives in C.
-tp
Use this option to specify the target host processor architecture.
-acc
Use this option to enable OpenACC directives. You can use the -⁠acc suboptions to specify loop autoparallelization, how the compiler reports compute regions failures to accelerate, and whether to issue a warning or an error for non-OpenACC accelerator directives.
-Minfo or -⁠Minfo=accel
Use this option to see messages about the success or failure of the compiler in translating the accelerator region into GPU kernels.

The -ta flag has the following accelerator-related suboptions:

tesla
Select NVIDIA accelerator target. This option has a number of suboptions:
cc20, cc30, cc35, cc50, cc60 Generate code for compute capability 2.0, 3.0, 3.5, 5.0 or 6.0 respectively; multiple selections are valid.
cuda7.5 or 7.5 Specify the CUDA 7.5 version of the toolkit. This is the default.
cuda8.0 or 8.0 Specify the CUDA 8.0 version of the toolkit.
deepcopy Enable full deep copy of aggregate data structions in OpenACC; Fortran only.
fastmath Use routines from the fast math library.
[no]flushz Control flush-to-zero mode for floating point computations in the GPU code.
[no]fma Do [not] generate fused multiply-add instructions.
keep Keep the kernel files.
kepler Generate code for Kepler Architecture equivalent to NVIDIA compute capability 3.x.
maxregcount:n
  • Specify the maximum number of registers to use on the GPU.
  • Leaving this blank indicates no limit.
[no]fma Control generation of fused multiply-add instructions.
noL1 Prevent the use of L1 hardware data cache to cache global variables.
[no]rdc Control generation of relocatable device code (default on).
host
Generate host code for accelerator regions.

The compiler automatically invokes the necessary software tools to create the kernel code and embeds the kernels in the object file.

Note: To access accelerator libraries, you must link an accelerator program with the -⁠ta flag.

7.14. Profiling Accelerator Kernels

This release supports the environment variable PGI_ACC_TIME for 64-bits. Setting this environment variable to a nonzero value enables collection and printing of simple timing information about the accelerator regions and generated kernels.

Note: Turn off all CUDA Profilers (NVIDIA's Visual Profiler, NVPROF, CUDA_PROFILE, etc) when enabling PGI_ACC_TIME, they use the same library to gather performance data and cannot be used concurently.
Note: Windows Users: To ensure that all the performance information is collected we recommend that 'acc_shutdown' is called before your application is finished and 'main' exits.

Accelerator Kernel Timing Data

bb04.f90
  s1
   15: region entered 1 times
	 time(us): total=1490738 
                 init=1489138 region=1600
                 kernels=155 data=1445
	 w/o init: total=1600 max=1600 
                 min=1600 avg=1600
   18: kernel launched 1 times
   time(us): total=155 max=155 min=155 avg=155

In this example, a number of things are occurring:

  • For each accelerator region, the file name bb04.f90 and subroutine or function name s1 is printed, with the line number of the accelerator region, which in the example is 15.
  • The library counts how many times the region is entered (1 in the example) and the microseconds spent in the region (in this example 1490738), which is split into initialization time (in this example 1489138) and execution time (in this example 1600).
  • The execution time is then divided into kernel execution time and data transfer time between the host and GPU.
  • For each kernel, the line number is given, (18 in the example), along with a count of kernel launches, and the total, maximum, minimum, and average time spent in the kernel, all of which are 155 in this example.

7.16. Supported Intrinsics

An intrinsic is a function available in a given language whose implementation is handled specifically by the compiler. Typically, an intrinsic substitutes a sequence of automatically-generated instructions for the original function call. Since the compiler has an intimate knowledge of the intrinsic function, it can better integrate it and optimize it for the situation.

Intrinsics make the use of processor-specific enhancements easier because they provide a language interface to assembly instructions. In doing so, the compiler manages things that the user would normally have to be concerned with, such as register names, register allocations, and memory locations of data.

This section contains an overview of the Fortran and C intrinsics that the accelerator supports.

7.16.1. Supported Fortran Intrinsics Summary Table

Table 20 is an alphabetical summary of the supported Fortran intrinsics that the accelerator supports. These functions are specific to Fortran 90/95 unless otherwise specified.

Note: For complete descriptions of these intrinsics, refer to ‘Fortran Intrinsics’ of the PGI Fortran Reference Manual.

In most cases PGI provides support for all the data types for which the intrinsic is valid. When support is available for only certain data types, the middle column of the table specifies which ones, using the following codes:

I for integer S for single precision real C for single precision complex
  D for double precision real Z for double precision complex
Table 20. Supported Fortran Intrinsics
This intrinsic Returns this value ...
ABS I,S,D absolute value of the supplied argument.
ACOS   arccosine of the specified value.
AINT   truncation of the supplied value to a whole number.
ANINT   nearest whole number to the supplied argument.
ASIN   arcsine of the specified value.
ATAN   arctangent of the specified value.
ATAN2   arctangent of the specified value.
COS S,D cosine of the specified value.
COSH   hyperbolic cosine of the specified value.
DBLE S,D conversion of the value to double precision real.
DPROD   double precision real product.
EXP S,D exponential value of the argument.
IAND   result of a bit-by-bit logical AND on the arguments.
IEOR   result of a bit-by-bit logical exclusive OR on the arguments.
INT I,S,D conversion of the value to integer type.
IOR   result of a bit-by-bit logical OR on the arguments.
LOG S,D natural logarithm of the specified value.
LOG10   base-10 logarithm of the specified value.
MAX   maximum value of the supplied arguments.
MIN   minimum value of the supplied arguments.
MOD I remainder of the division.
NINT   nearest integer to the real argument.
NOT   result of a bit-by-bit logical complement on the argument.
REAL I,S,D conversion of the argument to real.
SIGN   absolute value of A times the sign of B.
SIN S,D value of the sine of the argument.
SINH   hyperbolic sine of the argument.
SQRT S,D square root of the argument.
TAN   tangent of the specified value.
TANH   hyperbolic tangent of the specified value.

7.16.2. Supported C Intrinsics Summary Table

This section contains two alphabetical summaries – one for double functions and a second for float functions. These lists contain only those C intrinsics that the accelerator supports.

Table 21. Supported C Intrinsic Double Functions
This intrinsic Returns this value ...
acos arccosine of the specified value.
asin arcsine of the specified value.
atan arctangent of the specified value.
atan2 arctangent of y/x, where y is the first argument, x the second.
cos cosine of the specified value.
cosh hyperbolic cosine of the specified value.
exp exponential value of the argument.
fabs absolute value of the argument.
fmax maximum value of the two supplied arguments
fmin minimum value of the two supplied arguments
log natural logarithm of the specified value.
log10 base-10 logarithm of the specified value.
pow value of the first argument raised to the power of the second argument.
sin value of the sine of the argument.
sinh hyperbolic sine of the argument.
sqrt square root of the argument.
tan tangent of the specified value.
tanh hyperbolic tangent of the specified value.
Table 22. Supported C Intrinsic Float Functions
This intrinsic Returns this value ...
acosf arccosine of the specified value.
asinf arcsine of the specified value.
atanf arctangent of the specified value.
atan2f arctangent of y/x, where y is the first argument, x the second.
cosf cosine of the specified value.
coshf hyperbolic cosine of the specified value.
expf exponential value of the floating-point argument.
fabsf absolute value of the floating-point argument.
logf natural logarithm of the specified value.
log10f base-10 logarithm of the specified value.
powf value of the first argument raised to the power of the second argument.
sinf value of the sine of the argument.
sinhf hyperbolic sine of the argument.
sqrtf square root of the argument.
tanf tangent of the specified value.
tanhf hyperbolic tangent of the specified value.

7.17. References related to Accelerators

  • ISO/IEC 1539-1:1997, Information Technology – Programming Languages – Fortran, Geneva, 1997 (Fortran 95).
  • American National Standard Programming Language C, ANSI X3.159-1989 (ANSI C).
  • ISO/IEC 9899:1999, Information Technology – Programming Languages – C, Geneva, 1999 (C99).
  • ISO/IEC 9899:2011, Information Technology – Programming Languages – C, Geneva, 2011 (C11).
  • ISO/IEC 14882:2011, Information Technology – Programming Languages – C++, Geneva, 2011 (C++11).
  • Debugger User's Guide.
  • Profiler User's Guide.
  • PGI Fortran Reference Manual.

8. Eclipse

This document explains how to install and use the PGI plugin for Eclipse CDT (C/C++ development tool). PGI Eclipse integration is only available on Linux.

8.1. Install Eclipse CDT

To install the Eclipse plugin for the PGI C and C++ compilers:

  1. Download the Eclipse plugin from the PGI download webpage.
  2. Untar and install the PGI Eclipse plugin componennts.
    $ tar zxf pgieclipse-2017.tar.gz
    $ ./insall_compiilers

    Note you will be asked to provide the location of your current PGI installation. The default is /opt/pgi.

  3. Before you install the actual PGI Eclipse plugin, from within Eclipse, check your CDT version.
    1. Go to Help -> About Eclipse
    2. Click the Eclipse CDT button.

      You might need to hover the mouse pointer on the button to see the hint.

    3. Select Eclipse C/C++ Development Tools.

      The first number in the feature version specifies which plugin version is selected.

  4. Go to Help -> Install New software.
  5. Click the Add button to add a new software repository.
  6. In the Add Repository dialog box:
    1. Click Local.
    2. Select your PGI installation directory, such as /opt/pgi.
    3. Browse inside 2017/eclipse and select the directory matching your CDT version.
    4. Click OK.

      The Add Repository dialog should show the path to the local directory containing the plugin for your CDT version. For example, if PGI compilers are installed in /opt/pgi, then the CDT 7 plugin is located in /opt/pgi/<os-version>/2017/eclipse/cdt7; the CDT 8 plugin is in /opt/pgi/<os-version>/2017/eclipse/cdt8, and so on.

    5. Click OK in the Add Repository dialog.

    The install form now shows “PGI C/C++ Compiler Plugin” as an option to install.

  7. Check the box next to PGI option and select Next to get to the Install Details view.
  8. Click Next again.
  9. Review and accept the End-User License agreement.
  10. Click Finish.

You are prompted to restart. Select Restart to complete installation of the plugin.

8.2. Use Eclipse CDT

To use the Eclipse plugin for the PGI C and C++ compilers, the directory containing PGI compilers and tools should be included in your PATH prior to starting the Eclipse IDE. For details on how to include this directory in your PATH environment variable, refer to Using Environment Variables, and specifically to PATH.

This plugin does not currently support the Code Analysis feature of Eclipse CDT. This feature is not disabled by default for PGI projects and is the cause of spurious syntax errors during pre-compilation. This feature can be disabled manually at either the Project or Workspace level; any actual compilation or link errors are reported at build time.

  • To disable Code Analysis at the Workspace level, select the menu item Window | Preferences, then select C/C++, then Code Analysis; uncheck all category items, which should deselect all sub-items.
  • To disable Code Analysis at the Project level, select the menu item Project | Properties, then select C/C++ General, then Code Analysis; uncheck the top-level categories to deselect everything.

The PGI plugin follows the same rules for creating, building, and running a project as any other compiler supported by Eclipse. For more information, refer to Eclipse documentation and tutorials at: http://www.eclipse.org/documentation/.

9. Using Directives and Pragmas

It is often useful to be able to alter the effects of certain command line options or default behavior of the compiler. Fortran directives and C/C++ pragmas provide pragmatic information that control the actions of the compiler in a particular portion of a program without affecting the program as a whole. That is, while a command line option affects the entire source file that is being compiled, directives and pragmas apply, or disable, the effects of a command line option to selected subprograms or to selected loops in the source file, for example, to optimize a specific area of code. Use directives and pragmas to tune selected routines or loops.

9.1. PGI Proprietary Fortran Directives

PGI Fortran compilers support proprietary directives that may have any of the following forms:

!pgi$g directive
!pgi$r directive
!pgi$l directive
!pgi$ directive
Note: If the input is in fixed format, the comment character must begin in column 1 and either * or C is allowed in place of !.

The scope indicator controls the scope of the directive. This indicator occurs after the $. Some directives ignore the scope indicator.

The valid scopes, shown in the previous forms of the directive, are these:

g
(global) indicates the directive applies to the end of the source file.
r
(routine) indicates the directive applies to the next subprogram.
l
(loop) indicates the directive applies to the next loop, but not to any loop contained within the loop body. Loop-scoped directives are only applied to DO loops.
blank
indicates that the default scope for the directive is applied.

The body of the directive may immediately follow the scope indicator. Alternatively, any number of blanks may precede the name of the directive. Any names in the body of the directive, including the directive name, may not contain embedded blanks. Blanks may surround any special characters, such as a comma or an equal sign.

The directive name, including the directive prefix, may contain upper or lower case letters, and the case is not significant. Case is significant for any variable names that appear in the body of the directive if the command line option -⁠Mupcase is selected. For compatibility with other vendors’ directives, the prefix cpgi$ may be substituted with cdir$ or cvd$.

9.2. PGI Proprietary C and C++ Pragmas

Pragmas may be supplied in a C/C++ source file to provide information to the compiler. Many pragmas have a corresponding command-line option. Pragmas may also toggle an option, selectively enabling and disabling the option.

The general syntax of a pragma is:

#pragma [ scope ] pragma-body 

The optional scope field is an indicator for the scope of the pragma; some pragmas ignore the scope indicator.

The valid scopes are:

global
indicates the pragma applies to the entire source file.
routine
indicates the pragma applies to the next function.
loop
indicates the pragma applies to the next loop (but not to any loop contained within the loop body). Loop-scoped pragmas are only applied to for and while loops.

If a scope indicator is not present, the default scope, if any, is applied. Whitespace must appear after the pragma keyword and between the scope indicator and the body of the pragma. Whitespace may also surround any special characters, such as a comma or an equal sign. Case is significant for the names of the pragmas and any variable names that appear in the body of the pragma.

9.3. PGI Proprietary Optimization Directive and Pragma Summary

The following table summarizes the supported Fortran directives and C/C++ pragmas. The following terms are useful in understanding the table.

  • Functionality is a brief summary of the way to use the directive or pragma. For a complete description, refer to the ‘Directives and Pragmas Reference’ section of the PGI Compiler Reference Manual.
  • Many of the directives and pragmas can be preceded by NO. The default entry indicates the default for the directive or pragma. N/A appears if a default does not apply.
  • The scope entry indicates the allowed scope indicators for each directive or pragma, with L for loop, R for routine, and G for global. The default scope is surrounded by parentheses and N/A appears if the directive or pragma is not available in the given language.
    Note: The "*" in the scope indicates this:
    For routine-scoped directive
    The scope includes the code following the directive or pragma until the end of the routine.
    For globally-scoped directive
    The scope includes the code following the directive or pragma until the end of the file rather than for the entire file.
Note:

The name of a directive or pragma may also be prefixed with -⁠M.

For example, you can use the directive -⁠Mbounds, which is equivalent to the directive bounds and you can use -⁠Mopt, which is equivalent to opt. For pragmas, you can use the directive -⁠Mnoassoc, which is equivalent to the pragma noassoc, and -⁠Mvintr, which is equivalent to vintr.

Table 23. Proprietary Optimization-Related Fortran Directive and C/C++ Pragma Summary
Directive or pragma Functionality Default Fortran Scope C/C++ Scope
altcode (noaltcode) Do/don’t generate alternate code for vectorized and parallelized loops. altcode (L)RG (L)RG
assoc (noassoc) Do/don’t perform associative transformations. assoc (L)RG (L)RG
bounds (nobounds) Do/don’t perform array bounds checking. nobounds (R)G* (R)G
cncall (nocncall) Loops are considered for parallelization, even if they contain calls to user-defined subroutines or functions, or if their loop counts do not exceed usual thresholds. nocncall (L)RG (L)RG
concur (noconcur) Do/don’t enable auto-concurrentization of loops. concur (L)RG (L)RG
depchk (nodepchk) Do/don’t ignore potential data dependencies. depchk (L)RG (L)RG
eqvchk (noeqvchk) Do/don’t check EQUIVALENCE for data dependencies. eqvchk (L)RG N/A
fcon (nofcon) Do/don’t assume unsuffixed real constants are single precision. nofcon N/A (R)G
invarif (noinvarif) Do/don’t remove invariant if constructs from loops. invarif (L)RG (L)RG
ivdep Ignore potential data dependencies. ivdep (L)RG N/A
lstval (nolstval) Do/don’t compute last values. lstval (L)RG (L)RG
prefetch Control how prefetch instructions are emitted      
opt Select optimization level. N/A (R)G (R)G
safe (nosafe) Do/don’t treat pointer arguments as safe. safe N/A (R)G
safe_lastval Parallelize when loop contains a scalar used outside of loop. not enabled (L) (L)
safeptr (nosafeptr) Do/don’t ignore potential data dependencies to pointers. nosafeptr N/A L(R)G
single (nosingle) Do/don’t convert float parameters to double. nosingle N/A (R)G*
tp Generate PGI Unified Binary code optimized for specified targets. N/A (R)G (R)G
unroll (nounroll) Do/don’t unroll loops. nounroll (L)RG (L)RG
vector (novector) Do/don't perform vectorizations. vector (L)RG* (L)RG
vintr (novintr) Do/don’t recognize vector intrinsics. vintr (L)RG (L)RG

9.4. Scope of Fortran Directives and Command-Line Options

During compilation the effect of a directive may be to either turn an option on, or turn an option off. Directives apply to the section of code following the directive, corresponding to the specified scope, which may include the following loop, the following routine, or the rest of the program. This section presents several examples that show the effect of directives as well as their scope. Consider the following Fortran code:

 integer maxtime, time
 parameter (n = 1000, maxtime = 10)
 double precision a(n,n), b(n,n), c(n,n)
   do time = 1, maxtime
     do i = 1, n
       do j = 1, n
         c(i,j) = a(i,j) + b(i,j)
       enddo
     enddo
   enddo
 end

When compiled with -⁠Mvect, both interior loops are interchanged with the outer loop.

$ pgfortran -Mvect dirvect1.f

Directives alter this behavior either globally or on a routine or loop by loop basis. To assure that vectorization is not applied, use the novector directive with global scope.

!pgi$g novector
    integer maxtime, time
    parameter (n = 1000, maxtime = 10)
    double precision a(n,n), b(n,n), c(n,n)
    do time = 1, maxtime
      do i = 1, n
	do j = 1, n
	  c(i,j) = a(i,j) + b(i,j)
        enddo
      enddo
   enddo
 end

In this version, the compiler disables vectorization for the entire source file. Another use of the directive scoping mechanism turns an option on or off locally, either for a specific procedure or for a specific loop:

 integer maxtime, time
 parameter (n = 1000, maxtime = 10)
 double precision a(n,n), b(n,n), c(n,n)
!pgi$l novector
   do time = 1, maxtime
     do i = 1, n
       do j = 1, n
         c(i,j) = a(i,j) + b(i,j)
       enddo
     enddo
   enddo
 end

Loop level scoping does not apply to nested loops. That is, the directive only applies to the following loop. In this example, the directive turns off vector transformations for the top-level loop. If the outer loop were a timing loop, this would be a practical use for a loop-scoped directive.

9.5. Scope of C/C++ Pragmas and Command-Line Options

During compilation a pragma either turns an option on or turns an option off. Pragmas apply to the section of code corresponding to the specified scope – either the entire file, the following loop, or the following or current routine. This section presents several examples showing the effect of pragmas and the use of the pragma scope indicators.

Note: In all cases, pragmas override a corresponding command-line option.

For pragmas that have only routine and global scope, there are two rules for determining the scope of the pragma. We cover these special scope rules at the end of this section.

Consider the following program:

main() {
   float a[100][100], b[100][100], c[100][100];
   int time, maxtime, n, i, j;
   maxtime=10;
   n=100;
   for (time=0; time<maxtime;time++)
     for (j=0; j<n;j++)
       for (i=0; i<n;i++)
         c[i][j] = a[i][j] + b[i][j];
}

When this is compiled using the -⁠Mvect command-line option, both interior loops are interchanged with the outer loop. Pragmas alter this behavior either globally or on a routine or loop by loop basis. To ensure that vectorization is not applied, use the novector pragma with global scope.

main() {
#pragma global novector
   float a[100][100], b[100][100],c[100][100];
   int time, maxtime, n, i, j;
   maxtime=10;
   n=100;
   for (time=0; time<maxtime;time++)
     for (j=0; j<n;j++)
       for (i=0; i<n;i++)
         c[i][j] = a[i][j] + b[i][j];
}

In this version, the compiler does not perform vectorization for the entire source file. Another use of the pragma scoping mechanism turns an option on or off locally either for a specific procedure or for a specific loop. The following example shows the use of a loop-scoped pragma.

main() {
   float a[100][100], b[100][100],c[100][100];
   int time, maxtime, n, i, j;
   maxtime=10;
   n=100;
#pragma loop novector
   for (time=0; time<maxtime;time++)
     for (j=0; j<n;j++)
       for (i=0; i<n;i++)
         c[i][j] = a[i][j] + b[i][j];
}

Loop level scoping does not apply to nested loops. That is, the pragma only applies to the following loop. In this example, the pragma turns off vector transformations for the top-level loop. If the outer loop were a timing loop, this would be a practical use for a loop-scoped pragma. The following example shows routine pragma scope:

#include "math.h"
func1() {
#pragma routine novector
 float a[100][100], b[100][100];
 float c[100][100], d[100][100];
 int i,j;
 for (i=0;i<100;i++)
   for (j=0;j<100;j++)
     a[i][j] = a[i][j] + b[i][j] * c[i][j];
     c[i][j] = c[i][j] + b[i][j] * d[i][j];
}
func2() {
 float a[200][200], b[200][200];
 float c[200][200], d[200][200];
 int i,j;
 for (i=0;i<200;i++)
   for (j=0;j<200;j++)
     a[i][j] = a[i][j] + b[i][j] * c[i][j];
     c[i][j] = c[i][j] + b[i][j] * d[i][j];
}

When this source is compiled using the -⁠Mvect command-line option, func2 is vectorized but func1 is not vectorized. In the following example, the global novector pragma turns off vectorization for the entire file.

#include "math.h"
func1() {
#pragma global novector
 float a[100][100], b[100][100];
 float c[100][100], d[100][100];
 int i,j;
 for (i=0;i<100;i++)
   for (j=0;j<100;j++)
     a[i][j] = a[i][j] + b[i][j] * c[i][j];
     c[i][j] = c[i][j] + b[i][j] * d[i][j];
}
func2() {
 float a[200][200], b[200][200];
 float c[200][200], d[200][200];
 int i,j;
 for (i=0;i<200;i++)
   for (j=0;j<200;j++)
     a[i][j] = a[i][j] + b[i][j] * c[i][j];
     c[i][j] = c[i][j] + b[i][j] * d[i][j];
}

Special Scope Rules

Special rules apply for a pragma with loop, routine, and global scope. When the pragma is placed within a routine, it applies to the routine from its point in the routine to the end of the routine. The same rule applies for one of these pragmas with global scope.

However, there are several pragmas for which only routine and global scope applies and which affect code immediately following the pragma:

  • bounds and fcon – The bounds and fcon pragmas behave in a similar manner to pragmas with loop scope. That is, they apply to the code following the pragma.
  • opt and safe – When the opt or safe pragmas are placed within a routine, they apply to the entire routine as if they had been placed at the beginning of the routine.

9.6. Prefetch Directives and Pragmas

Today’s processors are so fast that it is difficult to bring data into them quickly enough to keep them busy. Prefetch instructions can increase the speed of an application substantially by bringing data into cache so that it is available when the processor needs it.

When vectorization is enabled using the -⁠Mvect or -⁠Mprefetch compiler options, or an aggregate option such as -⁠fast that incorporates -⁠Mvect, the PGI compilers selectively emit instructions to explicitly prefetch data into the data cache prior to first use. You can control how these prefetch instructions are emitted by using prefetch directives and pragmas.

For a list of processors that support prefetch instructions refer to the PGI Release Notes.

9.6.1. Prefetch Directive Syntax in Fortran

The syntax of a prefetch directive is as follows:

  !$mem prefetch <var1>[,<var2>[,...]]

where <varn> is any valid variable, member, or array element reference.

9.6.2. Prefetch Directive Format Requirements

Note: The sentinel for prefetch directives is !$mem, which is distinct from the !pgi$ sentinel used for optimization directives. Any prefetch directives that use the !pgi$ sentinel are ignored by the PGI compilers.
  • The "c" must be in column 1 for fixed format.
  • Either * or ! is allowed in place of c for fixed format.
  • The scope indicators g, r and l used with the !pgi$ sentinel are not supported.
  • The directive name, including the directive prefix, may contain upper or lower case letters and is case insensitive (case is not significant).
  • If the command line option -⁠Mupcase is used, any variable names that appear in the body of the directive are case sensitive.

9.6.3. Sample Usage of Prefetch Directive

Prefetch Directive Use

This example uses prefetch directives to prefetch data in a matrix multiplication inner loop where a row of one source matrix has been gathered into a contiguous vector.

 real*8 a(m,n), b(n,p), c(m,p), arow(n)
...
 do j = 1, p
!$mem prefetch arow(1),b(1,j)
!$mem prefetch arow(5),b(5,j)
!$mem prefetch arow(9),b(9,j)
 do k = 1, n, 4
!$mem prefetch arow(k+12),b(k+12,j)
 c(i,j) = c(i,j) + arow(k) * b(k,j)
 c(i,j) = c(i,j) + arow(k+1) * b(k+1,j)
 c(i,j) = c(i,j) + arow(k+2) * b(k+2,j)
 c(i,j) = c(i,j) + arow(k+3) * b(k+3,j)
 enddo
 enddo

This pattern of prefetch directives the compiler emits prefetch instructions whereby elements of arow and b are fetched into the data cache starting four iterations prior to first use. By varying the prefetch distance in this way, it is sometimes possible to reduce the effects of main memory latency and improve performance.

9.6.4. Prefetch Pragma Syntax in C/C++

The syntax of a prefetch pragma is as follows:

#pragma mem prefetch <var1>[,<var2>[,...]]

where <varn> is any valid variable, member, or array element reference.

9.6.5. Sample Usage of Prefetch Pragma

Prefetch Pragma in C

This example uses the prefetch pragma to prefetch data from the source vector x for eight iterations beyond the current iteration.

 for (i=0; i<n; i++)  {
   #pragma mem prefetch x[i+8]
   y[i] = y[i] + a*x[i];
}

9.7. !$PRAGMA C

When programs are compiled using one of the PGI Fortran compilers on Linux, Win64, and macOS systems, an underscore is appended to Fortran global names, including names of functions, subroutines, and common blocks. This mechanism distinguishes Fortran name space from C/C++ name space.

You can use !$PRAGMA C in the Fortran program to call a C/C++ function from Fortran. The statement would look similar to this:

!$PRAGMA C(name[,name]...)
Note: This statement directs the compiler to recognize the routine 'name' as a C function, thus preventing the Fortran compiler from appending an underscore to the routine name.

9.8. IGNORE_TKR Directive

This directive indicates to the compiler to ignore the type, kind, and/or rank (/TKR/) of the specified dummy arguments in an interface of a procedure. The compiler also ignores the type, kind, and/or rank of the actual arguments when checking all the specifics in a generic call for ambiguities.

9.8.1. IGNORE_TKR Directive Syntax

The syntax for the IGNORE_TKR directive is this:

!DIR$ IGNORE_TKR [ [(<letter>) <dummy_arg>] ... ]
<letter>
is one or any combination of the following:
T – type K – kind R – rank

For example, KR indicates to ignore both kind and rank rules and TKR indicates to ignore the type, kind, and rank arguments.

<dummy_arg>
if specified, indicates the dummy argument for which TKR rules should be ignored. If not specified, TKR rules are ignored for all dummy arguments in the procedure that contains the directive.

9.8.2. IGNORE_TKR Directive Format Requirements

The following rules apply to this directive:

  • IGNORE_TKR must not specify dummy arguments that are allocatable, Fortran 90 pointers, or assumed-shape arrays.
  • IGNORE_TKR may appear in the body of an interface block or in the body of a module procedure, and may specify dummy argument names only.
  • IGNORE_TKR may appear before or after the declarations of the dummy arguments it specifies.
  • If dummy argument names are specified, IGNORE_TKR applies only to those particular dummy arguments.
  • If no dummy argument names are specified, IGNORE_TKR applies to all dummy arguments except those that are allocatable objects, Fortran 90 pointers, or assumed-shape arrays.

9.8.3. Sample Usage of IGNORE_TKR Directive

Consider this subroutine fragment:

subroutine example(A,B,C,D)
!DIR$ IGNORE_TKR A, (R) B, (TK) C, (K) D

Table 24 indicates which rules are ignored for which dummy arguments in the preceding sample subroutine fragment:

Table 24. IGNORE_TKR Example
Dummy Argument Ignored Rules
A Type, Kind and Rank
B Only rank
C Type and Kind
D Only Kind

Notice that no letters were specified for A, so all type, kind, and rank rules are ignored.

9.9. !DEC$ Directives

PGI Fortran compilers for Microsoft Windows support several de-facto standard Fortran directives that help with inter-language calling and importing and exporting routines to and from DLLs.

9.9.1. !DEC$ Directive Syntax

These directives all take the form:

!DEC$ directive

9.9.2. Format Requirements

You must follow the following format requirements for the directive to be recognized in your program:

  • The directive must begin in column 1 when the file is fixed format or compiled with -⁠Mfixed.
  • The directive prefix !DEC$ requires a space between the prefix and the directive keyword, such as ATTRIBUTES.
  • The ! must begin the prefix when compiling Fortran 90/95 free-form format.
  • The characters C or * can be used in place of ! in either form of the prefix when compiling F77-style fixed-form format.
  • The directives are completely case insensitive.

9.9.3. Summary Table

The following table summarizes the supported !DEC$ directives. For a complete description of each directive, refer to the ‘!DEC$ Directives’ section of the ‘Directives and Pragmas Reference’ section in the PGI Compiler Reference Manual.

Table 25. !DEC$ Directives Summary Table
Directive Functionality
ALIAS Specifies an alternative name with which to resolve a routine.
ATTRIBUTES Lets you specify properties for data objects and procedures.
DECORATE Specifies that the name specified in the ALIAS directive should have the prefix and postfix decorations performed on it that are associated with the calling conventions that are in effect. This directive has no effect if ALIAS is not specified.
DISTRIBUTE Tells the compiler at what point within a loop to split into two loops.

10. Creating and Using Libraries

A library is a collection of functions or subprograms that are grouped for reference and ease of linking. This section discusses issues related to PGI-supplied compiler libraries. Specifically, it addresses the use of C/C++ builtin functions in place of the corresponding libc routines, creation of dynamically linked libraries, known as shared objects or shared libraries, and math libraries.

Note: This section does not duplicate material related to using libraries for inlining, described in Creating an Inline Library or information related to runtime library routines available to OpenMP programmers, described in Runtime Library Routines.

PGI provides libraries that export C interfaces by using Fortran modules. On Windows, PGI also provides additions to the supported library functionality for runtime functions included in DFLIB.

This section has examples that include the following options related to creating and using libraries.

-Bdynamic -def<file> -implib <file> -Mmakeimplib
-Bstatic -dynamiclib -l -o
-c -fpic -Mmakedll -shared

10.1. Using builtin Math Functions in C/C++

The name of the math header file is math.h. Include the math header file in all of your source files that use a math library routine as in the following example, which calculates the inverse cosine of 3.5.

#include <math.h>
#include<stdio.h>
#define PI 3.1415926535
void main()
{
  double x, y;
  x = PI/3.0;
  y = acos(0.5);
  printf(‘%f %f\n’,x,y);
}

Including math.h causes PGI C and C++ to use builtin functions, which are much more efficient than library calls. In particular, if you include math.h, the following intrinsics calls are processed using builtins:

abs acosf asinf atan atan2 atan2f
atanf cos cosf exp expf fabs
fabsf fmax fmaxf fmin fminf log
log10 log10f logf pow powf sin
sinf sqrt sqrtf tan tanf  

10.2. Using System Library Routines

Release 17.7 of the PGI runtime libraries makes use of Linux system libraries to implement, for example, OpenMP and Fortran I/O. The PGI runtime libraries make use of several additional system library routines.

On 64-bit Linux systems, the system library routines that PGI supports include these:

aio_error aio_write pthread_mutex_init sleep
aio_read calloc pthread_mutex_lock  
aio_return getrlimit pthread_mutex_unlock  
aio_suspend pthread_attr_init setrlimit  

10.3. Creating and Using Shared Object Files on Linux

All of the PGI Fortran, C, and C++ compilers support creation of shared object files. Unlike statically-linked object and library files, shared object files link and resolve references with an executable at runtime via a dynamic linker supplied with your operating system. The PGI compilers must generate position independent code to support creation of shared objects by the linker. However, this is not the default. You must create object files with position independent code and shared object files that will include them.

10.3.1. Procedure to create a use a shared object file

The following steps describe how to create and use a shared object file.

  1. Create an object file with position independent code. To do this, compile your code with the appropriate PGI compiler using the -⁠fpic option, or one of the equivalent options, such as -⁠fPIC, -⁠Kpic, and -⁠KPIC, which are supported for compatibility with other systems. For example, use the following command to create an object file with position independent code using pgfortran:
    % pgfortran -c -fpic tobeshared.f
  2. Produce a shared object file.

    To do this, use the appropriate PGI compiler to invoke the linker supplied with your system. It is customary to name such files using a .so filename extension. On Linux, you do this by passing the -⁠shared option to the linker:

    % pgfortran -shared -o tobeshared.so tobeshared.o
    Note: Compilation and generation of the shared object can be performed in one step using both the -⁠fpic option and the appropriate option for generation of a shared object file.
  3. Use a shared object file.

    To do this, use the appropriate PGI compiler to compile and link the program which will reference functions or subroutines in the shared object file, and list the shared object on the link line, as shown here:

    % pgfortran -o myprog myprog.f tobeshared.so
  4. Make the executable available.

    You now have an executable myprog which does not include any code from functions or subroutines in tobeshared.so, but which can be executed and dynamically linked to that code. By default, when the program is linked to produce myprog, no assumptions are made on the location of tobeshared.so. Therefore, for myprog to execute correctly, you must initialize the environment variable LD_LIBRARY_PATH to include the directory containing tobeshared.so. If LD_LIBRARY_PATH is already initialized, it is important not to overwrite its contents. If you have placed tobeshared.so in directory /home/myusername/bin, you can initialize LD_LIBRARY_PATH to include that directory and preserve its existing contents, as shown in the following:

    % setenv LD_LIBRARY_PATH "$LD_LIBRARY_PATH":/home/myusername/bin

    If you know that tobeshared.so always resides in a specific directory, you can create the executable myprog in a form that assumes this directory by using the -⁠R link-time option. For example, you can link as follows:

    % pgfortran -o myprog myprof.f tobeshared.so -R/home/myusername/bin
    Note: As with the -⁠L option, there is no space between -⁠R and the directory name. If the -⁠R option is used, it is not necessary to initialize LD_LIBRARY_PATH.

    In the previous example, the dynamic linker always looks in /home/myusername/bin to resolve references to tobeshared.so. By default, if the LD_LIBRARY_PATH environment variable is not set, the linker only searches /usr/lib and /lib for shared objects.

10.3.2. ldd Command

The ldd command is a useful tool when working with shared object files and executables that reference them. When applied to an executable, as shown in the following example, ldd lists all shared object files referenced in the executable along with the pathname of the directory from which they will be extracted.

% ldd myprog 

If the pathname is not hard-coded using the-⁠R option, and if LD_LIBRARY_PATH is not initialized, the pathname is listed as "not found". For more information on ldd, its options and usage, see the online man page for ldd.

10.4. Creating and Using Dynamic Libraries on Apple macOX

Note: PGI compilers for macOS do not support static linking on user executables. Apple only ships dynamic versions of its system libraries – not static versions. You can create static libraries; however, you cannot create 100% static executables.

10.5. PGI Runtime Libraries on Windows

Both statically- and dynamically-linked library (DLL) versions are available with the PGI runtime libraries on Windows. The static libraries are used by default.

  • You can use the dynamically-linked version of the runtime by specifying -⁠Bdynamic at both compile and link time.
  • You can explicitly specify static linking, the default, by using -⁠Bstatic at compile and link time.

For details on why you might choose one type of linking over another type, refer to Creating and Using Dynamic-Link Libraries on Windows.

10.6. Creating and Using Static Libraries on Windows

The Microsoft Library Manager (LIB.EXE) is the tool that is typically used to create and manage a static library of object files on Windows. LIB is provided with the PGI compilers as part of the Microsoft Open Tools. Refer to http://www.msdn2.com for a complete LIB reference – search for LIB.EXE. For a list of available options, invoke LIB with the /? switch.

For compatibility with legacy makefiles, PGI provides a wrapper for LIB and LINK called ar. This version of ar is compatible with Windows and object-file formats.

PGI also provides ranlib as a placeholder for legacy makefile support.

10.6.1. ar command

The ar command is a legacy archive wrapper that interprets legacy ar command line options and translates these to LINK/LIB options. You can use it to create libraries of object files.

Syntax

The syntax for the ar command is this:

ar [options] [archive] [object file].

Where:

  • The first argument must be a command line switch, and the leading dash on the first option is optional.
  • The single character options, such as -⁠d and -⁠v, may be combined into one option, such as -⁠dv.

    Thus, ar dv, ar -dv, and ar -d -v all mean the same thing.

  • The first non-switch argument must be the library name.
  • Exactly one of -⁠d, -⁠r, -⁠t, or -⁠x must appear on the command line.

Options

The options available for the ar command are these:

-c
This switch is for compatibility; it is ignored.
-d
Deletes the named object files from the library.
-r
Replaces in or adds the named object files to the library.
-t
Writes a table of contents of the library to standard out.
-v
Writes a verbose file-by-file description of the making of the new library to standard out.
-x
Extracts the named files by copying them into the current directory.

10.6.2. ranlib command

The ranlib command is a wrapper that allows use of legacy scripts and makefiles that use the ranlib command. The command actually does nothing; it merely exists for compatibility.

Syntax

The syntax for the ranlib command is this:

ranlib [options] [archive]

Options

The options available for the ar command are these:

-help
Short help information is printed out.
-V
Version information is printed out.

10.7. Creating and Using Dynamic-Link Libraries on Windows

There are several differences between static- and dynamic-link libraries on Windows. Libraries of either type are used when resolving external references for linking an executable, but the process differs for each type of library. When linking with a static library, the code needed from the library is incorporated into the executable. When linking with a DLL, external references are resolved using the DLL's import library, not the DLL itself. The code in the DLL associated with the external references does not become a part of the executable. The DLL is loaded when the executable that needs it is run. For the DLL to be loaded in this manner, the DLL must be in your path.

Static libraries and DLLs also handle global data differently. Global data in static libraries is automatically accessible to other objects linked into an executable. Global data in a DLL can only be accessed from outside the DLL if the DLL exports the data and the image that uses the data imports it.

To access global data, the C compilers support the Microsoft storage class extensions: __declspec(dllimport) and __declspec(dllexport). These extensions may appear as storage class modifiers and enable functions and data to be imported and exported:
extern int __declspec(dllimport) intfunc();
	float __declspec(dllexport) fdata;

The PGI Fortran compilers support the DEC$ ATTRIBUTES extensions DLLIMPORT and DLLEXPORT:

cDEC$ ATTRIBUTES DLLEXPORT :: object [,object] ...
cDEC$ ATTRIBUTES DLLIMPORT :: object [,object] ...

Here c is one of C, c, !, or *. object is the name of the subprogram or common block that is exported or imported. Further, common block names are enclosed within slashes (/), as shown here:

cDEC$ ATTRIBUTES DLLIMPORT :: intfunc
!DEC$ ATTRIBUTES DLLEXPORT :: /fdata/

For more information on these extensions, refer to !DEC$ Directives.

The examples in this section further illustrate the use of these extensions.

To create a DLL from the command line, use the -⁠Mmakedll option.

The following switches apply to making and using DLLs with the PGI compilers:

-Bdynamic
Compile for and link to the DLL version of the PGI runtime libraries. This flag is required when linking with any DLL built by the PGI compilers. This flag corresponds to the /MD flag used by Microsoft’s cl compilers.

When you use the PGI compiler flag -⁠Bdynamic to create an executable that links to the DLL form of the runtime, the executable built is smaller than one built without -⁠Bdynamic. The PGI runtime DLLs, however, must be available on the system where the executable is run. You must use the -⁠Bdynamic flag when linking an executable against a DLL built by the PGI compilers.

-Bstatic
Compile for and link to the static version of the PGI runtime libraries. This flag corresponds to the /MT flag used by Microsoft’s cl compilers.

On Windows, you must use-⁠Bstatic for both compiling and linking.

-Mmakedll
Generate a dynamic-link library or DLL. Implies -⁠Bdynamic.
-Mmakeimplib
Generate an import library without generating a DLL. Use this flag when you want to generate an import library for a DLL but are not yet ready to build the DLL itself. This situation might arise, for example, when building DLLs with mutual imports, as shown in Build DLLs Containing Mutual Imports: Fortran.
-o <file>
Passed to the linker. Name the DLL or import library <file>.
-def <file>
When used with -⁠Mmakedll, this flag is passed to the linker and a .def file named <file> is generated for the DLL. The .def file contains the symbols exported by the DLL. Generating a .def file is not required when building a DLL but can be a useful debugging tool if the DLL does not contain the symbols that you expect it to contain.

When used with -⁠Mmakeimplib, this flag is passed to lib which requires a .def file to create an import library. The .def file can be empty if the list of symbols to export are passed to lib on the command line or explicitly marked as DLLEXPORT in the source code.

-implib <file>
Passed to the colinker. Generate an import library named <file> for the DLL. A DLL’s import library is the interface used when linking an executable that depends on routines in a DLL.

To use the PGI compilers to create an executable that links to the DLL form of the runtime, use the compiler flag -⁠Bdynamic. The executable built will be smaller than one built without -⁠Bdynamic; the PGI runtime DLLs, however, must be available on the system where the executable is run. The -⁠Bdynamic flag must be used when an executable is linked against a DLL built by the PGI compilers.

The following examples outline how to use -⁠Bdynamic, -⁠Mmakedll and -⁠Mmakeimplib to build and use DLLs with the PGI compilers.

10.7.1. Build a DLL: Fortran

This example builds a DLL from a single source file, object1.f, which exports data and a subroutine using DLLEXPORT. The source file, prog1.f, uses DLLIMPORT to import the data and subroutine from the DLL.

object1.f

 subroutine sub1(i)
!DEC$ ATTRIBUTES DLLEXPORT :: sub1
 integer i
 common /acommon/ adata
 integer adata
!DEC$ ATTRIBUTES DLLEXPORT :: /acommon/
 print *, "sub1 adata", adata
 print *, "sub1 i ", i
 adata = i
 end

prog1.f

 program prog1
 common /acommon/ adata
 integer adata
 external sub1
!DEC$ ATTRIBUTES DLLIMPORT:: sub1, /acommon/
 adata = 11
 call sub1(12)
 print *, "main adata", adata
 end
  1. Create the DLL obj1.dll and its import library obj1.lib using the following series of commands:
    % pgfortran -Bdynamic -c object1.f
    % pgfortran -Mmakedll object1.obj -o obj1.dll
            
  2. Compile the main program:
    % pgfortran -Bdynamic -o prog1 prog1.f -defaultlib:obj1

    The -⁠Bdynamic and -⁠Mmakedll switches cause the compiler to link against the PGI runtime DLLs instead of the PGI runtime static libraries. The -⁠Bdynamic switch is required when linking against any PGI-compiled DLL, such as obj1.dll. The -defaultlib: switch specifies that obj1.lib, the DLL’s import library, should be used to resolve imports.

  3. Ensure that obj1.dll is in your path, then run the executable prog1 to determine if the DLL was successfully created and linked:
    % prog1
    sub1 adata 11
    sub1 i 12
    main adata 12

    Should you wish to change obj1.dll without changing the subroutine or function interfaces, no rebuilding of prog1 is necessary. Just recreate obj1.dll and the new obj1.dll is loaded at runtime.

10.7.2. Build a DLL: C

In this example, we build a DLL out of a single source file, object2.c, which exports data and a subroutine using __declspec(dllexport). The main source file, prog2.c, uses __declspec(dllimport) to import the data and subroutine from the DLL.

object2.c

int __declspec(dllexport) data;
void __declspec(dllexport)
func2(int i)
{
  printf("func2: data == %d\n", data);
  printf("func2: i == %d\n", i);
  data = i;
}

prog2.c

int __declspec(dllimport) data;
void __declspec(dllimport) func2(int);
int
main()
{
  data = 11;
  func2(12);
  printf("main: data == %d\n",data);
  return 0;
  }
  1. Step 1: Create the DLL obj2.dll and its import library obj2.lib using the following series of commands:
    % pgcc -Bdynamic -c object2.c
    % pgcc -Mmakedll object2.obj -o obj2.dll
  2. Step 2: Compile the main program:
    % pgcc -Bdynamic -o prog2 prog2.c -defaultlib:obj2

    The -⁠Bdynamic switch causes the compiler to link against the PGI runtime DLLs instead of the PGI runtime static libraries. The -⁠Bdynamic switch is required when linking against any PGI-compiled DLL such as obj2.dll. The -⁠defaultlib: switch specifies that obj2.lib, the DLL’s import library, should be used to resolve the imported data and subroutine in prog2.c.

  3. Step 3: Ensure that obj2.dll is in your path, then run the executable prog2 to determine if the DLL was successfully created and linked:
    % prog2
    func2: data == 11
    func2: i == 12
    main: data == 12

    Should you wish to change obj2.dll without changing the subroutine or function interfaces, no rebuilding of prog2 is necessary. Just recreate obj2.dll and the new obj2.dll is loaded at runtime.

10.7.3. Build DLLs Containing Circular Mutual Imports: C

In this example we build two DLLs, obj3.dll and obj4.dll, each of which imports a routine that is exported by the other. To link the first DLL, the import library for the second DLL must be available. Usually an import library is created when a DLL is linked. In this case, however, the second DLL cannot be linked without the import library for the first DLL. When such circular imports exist, an import library for one of the DLLs must be created in a separate step without creating the DLL. The PGI drivers call the Microsoft lib tool to create import libraries in this situation. Once the DLLs are built, we can use them to build the main program.

/* object3.c */

 void __declspec(dllimport) func_4b(void);
 void __declspec(dllexport)
 func_3a(void)
 {
   printf("func_3a, calling a routine in obj4.dll\n");
   func_4b();
 }
 void __declspec(dllexport)
 func_3b(void)
 {
   printf("func_3b\n");
 }
/* object4.c */
 void __declspec(dllimport) func_3b(void);
 void __declspec(dllexport)
 func_4a(void)
 {
   printf("func_4a, calling a routine in obj3.dll\n");
   func_3b();
 }
 void __declspec(dllexport)
 func_4b(void)
 {
   printf("func_4b\n");
 }
/* prog3.c */
 void __declspec(dllimport) func_3a(void);
 void __declspec(dllimport) func_4a(void);
 int
 main()
 {
   func_3a();
   func_4a();
   return 0;
 }
  1. Step 1: Use -⁠Mmakeimplib with the PGI compilers to build an import library for the first DLL without building the DLL itself.
    % pgcc -Bdynamic -c object3.c
    % pgcc -Mmakeimplib -o obj3.lib object3.obj
    Tip: The -⁠def=<deffile> option can also be used with -⁠Mmakeimplib. Use a .def file when you need to export additional symbols from the DLL. A .def file is not needed in this example because all symbols are exported using __declspec(dllexport).
  2. Step 2: Use the import library, obj3.lib, created in Step 1, to link the second DLL.
    % pgcc -Bdynamic -c object4.c
    % pgcc -Mmakedll -o obj4.dll object4.obj -defaultlib:obj3
  3. Step 3: Use the import library, obj4.lib, created in Step 2, to link the first DLL.
    % pgcc -Mmakedll -o obj3.dll object3.obj -defaultlib:obj4
  4. Step 4: Compile the main program and link against the import libraries for the two DLLs
    % pgcc -Bdynamic prog3.c -o prog3 -defaultlib:obj3 -defaultlib:obj4
  5. Step 5: Execute prog3.exe to ensure that the DLLs were create properly.
    % prog3
    func_3a, calling a routine in obj4.dll
    func_4b
    func_4a, calling a routine in obj3.dll
    func_3b

10.7.4. Build DLLs Containing Mutual Imports: Fortran

In this example we build two DLLs when each DLL is dependent on the other, and use them to build the main program.

In the following source files, object2.f95 makes calls to routines defined in object3.f95, and vice versa. This situation of mutual imports requires two steps to build each DLL.

To link the first DLL, the import library for the second DLL must be available. Usually an import library is created when a DLL is linked. In this case, however, the second DLL cannot be linked without the import library for the first DLL. When such circular imports exist, an import library for one of the DLLs must be created in a separate step without creating the DLL. The PGI drivers call the Microsoft lib tool to create import libraries in this situation.

Once the DLLs are built, we can use them to build the main program.

object2.f95

 subroutine func_2a
	external func_3b 
!DEC$ ATTRIBUTES DLLEXPORT :: func_2a 
!DEC$ ATTRIBUTES DLLIMPORT :: func_3b 
    print*,"func_2a, calling a routine in obj3.dll"
	call func_3b()  end subroutine
 subroutine func_2b 
!DEC$ ATTRIBUTES DLLEXPORT :: func_2b
  print*,"func_2b"  
end subroutine

object3.f95

 subroutine func_3a
   external func_2b
   !DEC$ ATTRIBUTES DLLEXPORT :: func_3a 
   !DEC$ ATTRIBUTES DLLIMPORT :: func_2b  
   print*,"func_3a, calling a routine in obj2.dll"  
   call func_2b()  end subroutine
 subroutine func_3b 
 !DEC$ ATTRIBUTES DLLEXPORT :: func_3b  
    print*,"func_3b"  
 end subroutine

prog2.f95

 program prog2
     external func_2a
     external func_3a
 !DEC$ ATTRIBUTES DLLIMPORT :: func_2a 
 !DEC$ ATTRIBUTES DLLIMPORT :: func_3a  
     call func_2a()  
     call func_3a()  
 end program 
  1. Use -⁠Mmakeimplib with the PGI compilers to build an import library for the first DLL without building the DLL itself.
    % pgfortran -Bdynamic -c object2.f95
    % pgfortran -Mmakeimplib -o obj2.lib object2.obj
    Tip: The -def=<deffile> option can also be used with -⁠Mmakeimplib. Use a .def file when you need to export additional symbols from the DLL. A .def file is not needed in this example because all symbols are exported using DLLEXPORT.
  2. Use the import library, obj2.lib, created in Step 1, to link the second DLL.
    % pgfortran -Bdynamic -c object3.f95
    % pgfortran -Mmakedll -o obj3.dll object3.obj -defaultlib:obj2
  3. Use the import library, obj3.lib, created in Step 2, to link the first DLL.
    % pgfortran -Mmakedll -o obj2.dll object2.obj -defaultlib:obj3
  4. Compile the main program and link against the import libraries for the two DLLs.
    % pgfortran -Bdynamic prog2.f95 -o prog2 -defaultlib:obj2 -defaultlib:obj3
  5. Execute prog2 to ensure that the DLLs were create properly.
    % prog2
    func_2a, calling a routine in obj3.dll
    func_3b
    func_3a, calling a routine in obj2.dll
    func_2b

10.7.5. Import a Fortran module from a DLL

In this example we import a Fortran module from a DLL. We use the source file defmod.f90 to create a DLL containing a Fortran module. We then use the source file use_mod.f90 to build a program that imports and uses the Fortran module from defmod.f90.

defmod.f90

module testm
 type a_type
 integer :: an_int
 end type a_type
 type(a_type) :: a, b
!DEC$ ATTRIBUTES DLLEXPORT :: a,b
 contains
 subroutine print_a
!DEC$ ATTRIBUTES DLLEXPORT :: print_a
 write(*,*) a%an_int
 end subroutine
 subroutine print_b
!DEC$ ATTRIBUTES DLLEXPORT :: print_b
 write(*,*) b%an_int
 end subroutine
end module

usemod.f90

use testm
 a%an_int = 1
 b%an_int = 2
 call print_a
 call print_b
end
  1. Create the DLL.
    % pgf90 -Mmakedll -o defmod.dll defmod.f90
    Creating library defmod.lib and object defmod.exp
  2. Create the exe and link against the import library for the imported DLL.
    % pgf90 -Bdynamic -o usemod usemod.f90 -defaultlib:defmod.lib
  3. Run the exe to ensure that the module was imported from the DLL properly.
    % usemod
    1
    2

10.8. Using LIB3F

The PGI Fortran compilers include complete support for the de facto standard LIB3F library routines on both Linux and Windows operating systems. See the PGI Fortran Language Reference manual for a complete list of available routines in the PGI implementation of LIB3F.

10.9. LAPACK, BLAS and FFTs

All PGI products now include a BLAS and LAPACK library based on the customized OpenBLAS project source and built with PGI compilers. The LAPACK library is called liblapack.a or on Windows, liblapack.lib. The BLAS library is called libblas.a or on Windows, libblas.lib. These libraries are installed to $PGI/<target>/lib, where <target> is replaced with the appropriate target name (linux86-64, osx86-64, or win64).

To use these libraries, simply link them in using the -⁠l option when linking your main program:

% pgfortran myprog.f -llapack -lblas

10.11. The C++ Standard Template Library

On Linux and macOS, the GNU-compatible pgc++ compiler uses the GNU g++ header files and Standard Template Library (STL) directly. The versions used are dependent on the version of the GNU compilers installed on your system, or specified when makelocalrc was run during installation of the PGI compilers.

11. Using Environment Variables

Environment variables allow you to set and pass information that can alter the default behavior of the PGI compilers and the executables which they generate. This section includes explanations of the environment variables specific to PGI compilers. Other environment variables are referenced and documented in other sections of this User’s Guide, the accompanying PGI Compiler Reference Manual, the Debugger User's Guide and the Profiler User's Guide.

  • You use OpenMP environment variables to control the behavior of OpenMP programs. For consistency related to the OpenMP environment, the details of the OpenMP-related environment variables are included in OpenMP section: Environment Variables.
  • You can use environment variables to control the behavior of the PGI debugger or PGI profiler. For a description of environment variables that affect these tools, refer to the Debugger User's Guide andProfiler User's Guide, respectively.

11.1. Setting Environment Variables

Before we look at the environment variables that you might use with the PGI compilers and tools, let’s take a look at how to set environment variables. To illustrate how to set these variables in various environments, let’s look at how a user might initialize the shell environment prior to using the PGI compilers and tools.

11.1.1. Setting Environment Variables on Linux

Let’s assume that you want access to the PGI products when you log in. Let’s further assume that you installed the PGI compilers in /opt/pgi and that the license file is in /opt/pgi/license.dat. For access at startup, you can add the following lines to your startup file.

In csh, use these commands:

 % setenv PGI /opt/pgi
 % setenv MANPATH "$MANPATH":$PGI/linux86-64/17.7/man
 % setenv LM_LICENSE_FILE $PGI/license.dat
 % set path = ($PGI/linux86-64/17.7/bin $path)

In bash, sh, zsh, or ksh, use these commands:

 $ PGI=/opt/pgi; export PGI 
 $ MANPATH=$MANPATH:$PGI/linux86-64/17.7/man; export MANPATH 
 $ LM_LICENSE_FILE=$PGI/license.dat; export LM_LICENSE_FILE 
 $ PATH=$PGI/linux86-64/17.7/bin:$PATH; export PATH 

11.1.2. Setting Environment Variables on Windows

In Windows, when you access PGI 17.7 (for example, using Start | ALL Programs | PGI | Command Shells 17.7), you have options that PGI provides for setting your environment variables – either the DOS command environment or the Cygwin Bash environment.

When you open either of these shells available to you, the default environment variables are already set and available to you.

You may want to use other environment variables, such as the OpenMP ones. This section explains how to do that.

Suppose that your home directory is C:\tmp. The following examples show how you might set the temporary directory to your home directory, and then verify that it is set.

Command prompt:

Once you have launched a command shell, enter the following:

DOS> set TMPDIR=C:\tmp
DOS> echo %TMPDIR%
C:\tmp
DOS>

Cygwin Bash prompt:

From the PGI 17.7 menu item, select PGI option and at the Cygwin Bash prompt, enter the following:

 PGI$ export TMPDIR=C:\tmp
 PGI$ echo $TMPDIR
 C:\tmp
PGI$

11.1.3. Setting Environment Variables on Apple macOS

Let’s assume that you want access to the PGI products when you log in. Let’s further assume that you installed the PGI compilers in /opt/pgi and that the license file is in /opt/pgi/license.dat. For access at startup, you can add the following lines to your startup file.

For x86-64 osx86-64 in a csh:

% set path = (/opt/pgi/osx86-64/17.7/bin $path)

For x86-64 osx86-64 in a bash, sh, zsh, or ksh:

$ PATH=/opt/pgi/osx86-64/17.7/bin:$PATH; export PATH 

11.3. PGI Environment Variables

You use the environment variables listed in Table 26 to alter the default behavior of the PGI compilers and the executables which they generate. This section provides more detailed descriptions about the variables in this table that are not OpenMP environment variables.

11.3.1. FLEXLM_BATCH

By default, on Windows the license server creates interactive pop-up messages to issue warning and errors. You can use the environment variable FLEXLM_BATCH to prevent interactive pop-up windows. To do this, set the environment variable FLEXLM_BATCH to 1.

The following csh example prevents interactive pop-up messages for licensing warnings and errors:

% set FLEXLM_BATCH = 1;

11.3.2. FORTRANOPT

FORTRANOPT allows the user to adjust the behavior of the PGI Fortran compilers.

  • If FORTRANOPT exists and contains the value vaxio, the record length in the open statement is in units of 4-byte words, and the $ edit descriptor only has an effect for lines beginning with a space or a plus sign (+).
  • If FORTRANOPT exists and contains the value format_relaxed, an I/O item corresponding to a numerical edit descriptor (such as F, E, I, and so on) is not required to be a type implied by the descriptor.
  • In a non-Windows environment, if FORTRANOPT exists and contains the value crif, a sequential formatted or list-directed record is allowed to be terminated with the character sequence \r\n (carriage return, newline). This approach is useful when reading records from a file produced on a Window’s system.

The following example causes the PGI Fortran compilers to use VAX I/O conventions:

% setenv FORTRANOPT vaxio 

11.3.3. GMON_OUT_PREFIX

GMON_OUT_PREFIX specifies the name of the output file for programs that are compiled and linked with the -pg option. The default name is gmon.out.

If GMON_OUT_PREFIX is set, the name of the output file has GMON_OUT_PREFIX as a prefix. Further, the suffix is the pid of the running process. The prefix and suffix are separated by a dot. For example, if the output file is mygmon, then the full filename may look something similar to this: mygmon.0012348567.

The following example causes the PGI Fortran compilers to use pgout as the output file for programs compiled and linked with the -pg option.

% setenv GMON_OUT_PREFIX pgout 

11.3.4. LD_LIBRARY_PATH

The LD_LIBRARY_PATH variable is a colon-separated set of directories specifying where libraries should first be searched, prior to searching the standard set of directories. This variable is useful when debugging a new library or using a nonstandard library for special purposes.

The following csh example adds the current directory to your LD_LIBRARY_PATH variable.

% setenv LD_LIBRARY_PATH "$LD_LIBRARY_PATH":"./"

11.3.5. LM_LICENSE_FILE

The LM_LICENSE_FILE variable specifies the full path of the license file that is required for running the PGI software.

For example, once the license file is in place, you can execute the following csh commands to make the products you have purchased accessible and to initialize your environment for use of FLEXlm. These commands assume that you use the default installation directory: /opt/pgi
% setenv PGI /opt/pgi
% setenv LM_LICENSE_FILE "$LM_LICENSE_FILE":/opt/pgi/license.dat

To set the environment variable LM_LICENSE_FILE to the full path of the license key file, do this:

  1. Open the System Properties dialog: Start | Control Panel | System.
  2. Select the Advanced tab.
  3. Click the Environment Variables button.
    • If LM_LICENSE_FILE is not already an environment variable, create a new system variable for it. Set its value to the full path, including the name of the license key file, license.dat.
    • If LM_LICENSE_FILE already exists as an environment variable, append the path to the license file to the variable’s current value using a semi-colon to separate entries.

11.3.6. MANPATH

The MANPATH variable sets the directories that are searched for manual pages associated with the commands that the user types. When using PGI products, it is important that you set your PATH to include the location of the PGI products and then set the MANPATH variable to include the man pages associated with the products. MANPATH is not required for macOS.

The following csh example targets the linux86-64 version of the compilers and tool s and allows the user access to the manual pages associated with them.

% set path = (/opt/pgi/linux86-64/17.7/bin $path
% setenv MANPATH "$MANPATH":/opt/pgi/linux86-64/17.7/man

11.3.7. MPSTKZ

MPSTKZ increases the size of the stacks used by threads executing in parallel regions. You typically use this variable with programs that utilize large amounts of thread-local storage in the form of private variables or local variables in functions or subroutines called within parallel regions. The value should be an integer <n> concatenated with M or m to specify stack sizes of n megabytes.

For example, the following setting specifies a stack size of 8 megabytes.

% setenv MPSTKZ 8M

11.3.8. MP_BIND

You can set MP_BIND to yes or y to bind processes or threads executing in a parallel region to physical processor. Set it to no or n to disable such binding. The default is to not bind processes to processors. This variable is an execution-time environment variable interpreted by the PGI runtime support libraries. It does not affect the behavior of the PGI compilers in any way.

Note: The MP_BIND environment variable is not supported on all platforms.
% setenv MP_BIND y 

11.3.9. MP_BLIST

MP_BLIST allows you to specifically define the thread-CPU relationship.

Note: This variable is only in effect when MP_BIND is yes.

While the MP_BIND variable binds processors or threads to a physical processor, MP_BLIST allows you to specifically define which thread is associated with which processor. The list defines the processor-thread relationship order, beginning with thread 0. This list overrides the default binding.

For example, the following setting for MP_BLIST maps CPUs 3, 2, 1 and 0 to threads 0, 1, 2 and 3 respectively.

% setenv MP_BLIST=3,2,1,0 

11.3.10. MP_SPIN

When a thread executing in a parallel region enters a barrier, it spins on a semaphore. You can use MP_SPIN to specify the number of times it checks the semaphore before calling sched_yield() (on Linux or macOS) or_sleep()(on Windows). These calls cause the