These PGI compilers generate host CPU code for 64-bit little-endian OpenPOWER CPUs, and GPU device code for NVIDIA Kepler and Pascal GPUs.

These compilers include all GPU OpenACC features available in the PGI C/C⁠+⁠+/Fortran compilers for x86-64.

Documentation includes the pgcc, pgc⁠+⁠+ and pgfortran man pages and the -⁠help option. In addition, you can find both HTML and PDF versions of these installation and release notes, the PGI Compiler User’s Guide and PGI Compiler Reference Manual for OpenPOWER on the PGI website.

Release 2019 includes the following components:

• PGFORTRAN™ native OpenMP and OpenACC Fortran 2003 compiler.
• PGCC^® native OpenMP and OpenACC ISO C11 and K&R C compiler.
• PGC++^® native OpenMP and OpenACC ISO C++17 compiler.
• PGI Profiler^® OpenACC, CUDA, OpenMP, and multi-thread profiler.
• Open MPI version 3.1.3 including support for NVIDIA GPUDirect. As NVIDIA GPUDirect depends on InfiniBand support, Open MPI is also configured to use InfiniBand hardware if it is available on the system. InfiniBand support requires OFED 3.18 or later.
• ScaLAPACK 2.0.2 linear algebra math library for distributed-memory systems for use with Open MPI and the PGI compilers.
• BLAS and LAPACK library based on the customized OpenBLAS project source.
• Documentation in man page format and online in both HTML and PDF formats.

The PGI compilers for OpenPOWER are command-line compatible with the corresponding PGI products on Linux x86-64, meaning target-independent compiler options should behave consistently across the two platforms. The intent and expectation is that makefiles and build scripts used to drive PGI compilers on Linux x86-64 should work with little or no modification on OpenPOWER. The -⁠help compiler option lists all compiler options by default, or can be used to check the functionality of a specific option. For example:

% pgcc -help -fast
Reading rcfile /opt/pgi/linuxpower/19.4/bin/.pgccrc
-fast               Common optimizations; includes -O2 -Munroll=c:1 -Mlre -Mautoinline
                    == -Mvect=simd -Mflushz
-M[no]vect[=[no]simd|[no]assoc|[no]fuse]
                    Control automatic vector pipelining
    [no]simd        Generate [don't generate] SIMD instructions
    [no]assoc       Allow [disallow] reassociation
    [no]fuse        Enable [disable] loop fusion
-M[no]flushz        Set SSE to flush-to-zero mode
%

The PGI compilers use NVIDIA's CUDA Toolkit when building programs for execution on an NVIDIA GPU. Every PGI installation package puts the required CUDA Toolkit components into a PGI installation directory called 2019/cuda.

An NVIDIA CUDA driver must be installed on a system with a GPU before you can run a program compiled for the GPU on that system. PGI products do not contain CUDA Drivers. You must download and install the appropriate CUDA Driver from NVIDIA. The CUDA Driver version must be at least as new as the version of the CUDA Toolkit with which you compiled your code.

The PGI tool pgaccelinfo prints the driver version as its first line of output. You can use it to find out which version of the CUDA Driver is installed on your system.

You can let the compiler pick which version of the CUDA Toolkit to use or you can instruct it to use a particular version. The rest of this section describes all of your options.

If you do not specify a version of the CUDA Toolkit, the compiler uses the version of the CUDA Driver installed on the system on which you are compiling to determine which CUDA Toolkit to use. This auto-detect feature was introduced in the PGI 18.7 release; auto-detect is especially convenient when you are compiling and running your application on the same system. In the absence of any other information, the compiler will look for a CUDA Toolkit version in the PGI 2019/cuda directory that matches the version of the CUDA Driver installed on the system. If a match is not found, the compiler searches for the newest CUDA Toolkit version that is not newer than the CUDA Driver version. If there is no CUDA Driver installed, the PGI 19.4 compilers fall back to the default of CUDA 9.2.

When you specify a CUDA Toolkit version, you can additionally instruct the compiler to use a CUDA Toolkit installation different from the defaults bundled with the current PGI compilers. While most users do not need to use any other CUDA Toolkit installation than those provided with PGI, situations do arise where this capability is needed. Developers working with pre-release CUDA software may occasionally need to test with a CUDA Toolkit version not included in a PGI release. Conversely, some developers might find a need to compile with a CUDA Toolkit older than the oldest CUDA Toolkit installed with a PGI release. For these users, PGI compilers can interoperate with components from a CUDA Toolkit installed outside of the PGI installation directories.

PGI tests extensively using the co-installed versions of the CUDA Toolkits and fully supports their use. Use of CUDA Toolkit components not included with a PGI install is done with your understanding that functionality differences may exist.

To use a CUDA toolkit that is not installed with a PGI release, such as CUDA 9.1 with PGI 19.4, there are three options:

• Use the rcfile variable DEFAULT_CUDA_HOME to override the base default

  set DEFAULT_CUDA_HOME = /opt/cuda-9.1;

• Set the environment variable CUDA_HOME

  export CUDA_HOME=/opt/cuda-9.1

• Use the compiler compilation line assignment CUDA_HOME=

  pgfortran CUDA_HOME=/opt/cuda-9.1

The PGI compilers use the following order of precedence when determining which version of the CUDA Toolkit to use.

1. If you do not tell the compiler which CUDA Toolkit version to use, the compiler picks the CUDA Toolkit from the PGI installation directory 2019/cuda that matches the version of the CUDA Driver installed on your system. If the PGI installation directory does not contain a direct match, the newest version in that directory which is not newer than the CUDA driver version is used. If there is no CUDA driver installed on your system, the compiler falls back on an internal default; in PGI 19.4, this default is CUDA 9.2.
2. The rcfile variable DEFAULT_CUDA_HOME will override the base default.
3. The environment variable CUDA_HOME will override all of the above defaults.
4. The environment variable PGI_CUDA_HOME overrides all of the above; it is available for advanced users in case they need to override an already-defined CUDA_HOME.
5. A user-specified cudaX.Y sub-option to -⁠Mcuda and -⁠ta=tesla will override all of the above defaults and the CUDA Toolkit located in the PGI installation directory 2019/cuda will be used.
6. The compiler compilation line assignment CUDA_HOME= will override all of the above defaults (including the cudaX.Y sub-option).

The compilers can generate code for NVIDIA GPU compute capabilities 3.0 through 7.5. The compilers construct a default list of compute capabilities that matches the compute capabilities supported by the GPUs found on the system used in compilation. If there are no GPUs detected, the compilers select cc35, cc60, and cc70.

You can override the default by specifying one or more compute capabilities using either command-line options or an rcfile.

To change the default with a command-line option, provide a comma-separated list of compute capabilities to -⁠ta=tesla: for OpenACC or -⁠Mcuda= for CUDA Fortran.

To change the default with an rcfile, set the DEFCOMPUTECAP value to a blank-separated list of compute capabilities in the siterc file located in your installation's bin directory:

set DEFCOMPUTECAP=60 70;

Alternatively, if you don't have permissions to change the siterc file, you can add the DEFCOMPUTECAP definition to a separate .mypgirc file in your home directory.

The generation of device code can be time consuming, so you may notice an increase in compile time as the number of compute capabilities increases.

Many open-source software packages have been ported for use with PGI compilers on Linux for OpenPOWER.

The following PGI-compiled open-source software packages are included in the PGI OpenPOWER download package:

• OpenBLAS 0.3.3 – customized BLAS and LAPACK libraries based on the OpenBLAS project source.
• Open MPI 3.1.3 – open-source MPI implementation.
• ScaLAPACK 2.0.2 – a library of high-performance linear algebra routines for parallel distributed memory machines. ScaLAPACK uses Open MPI 3.1.3.

The following list of open-source software packages have been precompiled for execution on OpenPOWER targets using the PGI compilers and are available to download from the PGI website.

• ESMF 7.1.0r for Open MPI 3.1.3 – The Earth System Modeling Framework for building climate, numerical weather prediction, data assimilation, and other Earth science software applications.
• NetCDF 4.6.2 – A set of software libraries and self-describing, machine-independent data formats that support the creation, access, and sharing of array-oriented scientific data, written in C. Included in this package are the following components:
  • NetCDF-C++ 4.3.0 – C++ interfaces to NetCDF libraries.
  • NetCDF-Fortran 4.4.4 – Fortran interfaces to NetCDF libraries.
  • Parallel NetCDF 1.11.0 – for Open MPI 3.1.3.
  • HDF5 1.10.4 – data model, library, and file format for storing and managing data.
  • SZIP 2.1.1 – extended-Rice lossless compression algorithm.
  • ZLIB 1.2.11 – file compression library.

In addition, these software packages have also been ported to PGI on OpenPOWER but due to licensing restrictions, they are not available in binary format directly from PGI. You can find instructions for building them in the Porting & Tuning Guides section of the PGI website.

• FFTW 2.1.5 – version 2 of the Fast Fourier Transform library, includes MPI bindings built with Open MPI 3.1.3.
• FFTW 3.3.8 – version 3 of the Fast Fourier Transform library, includes MPI bindings built with Open MPI 3.1.3.

For additional information about building these and other packages, please see the Porting & Tuning Guides section of the PGI website.

The following PGI features are limited or are not implemented in the 19.4 release for OpenPOWER+Tesla:

• -Mipa is not enabled (no PGI inter-procedural analysis/optimization); the command-line option is accepted and silently ignored.
• -Mpfi/-Mpfo are not enabled (no profile-feedback optimization); the command-line options are accepted and silently ignored.
• Passing -M[no]llvm to MPI wrappers (mpicc, mpifort, etc.) is not supported. Doing so will cause unresolved symbol errors and segmentation faults when compiling.
• Using -V<version> with the PGI 2019 compilers to target PGI 2018 or earlier releases is not supported. This is a known limitation that is the result of the switch to using LLVM compilers as the default code generator for the PGI 2019 compilers.

Some specific issues related to the PGI Profiler:

• Debugging information is not always available on OpenPower which can cause the profiler to crash when attempting to unwind the call-stack. We recommend you use the --cpu-profiling-unwind-stack off option to disable call-stack tracing if you encounter any problem profiling on OpenPower.
• The Profiler relies on being able to directly call 'dlsym'. If this system call is intercepted by the program being profiled or by some other library the profiler may hang at startup. We have encountered this specific problem with some implementations of MPI. We recommend you disable any features that may be intercepting the 'dlsym' system call or disable CPU profiling with the --cpu-profiling off option.
  • To disable 'dlsym' interception when using IBM's spectrum MPI set the environment variable: PAMI_DISABLE_CUDA_HOOK=1, omit the following option: -gpu and add the options: -x PAMI_DISABLE_CUDA_HOOK and -disable_gpu_hooks.