Compilers

Global

Programming Models Support

The current compiler support for the existing programming models for GPU offload is the following:

Programming Model GCC IBM XL C/C++ IBM XL Fortran PGI Accelerator
CUDA Supported Supported Supported
CUDA Fortran Supported Supported Supported
OpenACC Experimental Not Supported Supported
OpenMP >= Version 7 >= Versions 13.1.6 / 15.1.6 Not Supported

Invocation

The invocation commands of the compilers are the following:

Compiler C C++ Fortran 77 Fortran 90 Fortran 03 Fortran 08
GCC gcc g++ gfortran gfortran gfortran gfortran
IBM XL xlc_r xlC_r xlf_r xlf90_r xlf2003_r xlf2008_r
PGI pgcc pgCC pgf90 pgf90 pgf90 pgf90

For IBM XL C/C++ & Fortran, the commands specified above constitute the « thread-safe » variants, to be used when dealing with multithreaded applications. It is recommended to exclusively use these variants, even in the case of non-multithreaded applications, though.

Compiler Options Overview

The table below provides the main equivalences between the options of the different compilers:

Purpose GCC IBM XL C/C++ IBM XL Fortran PGI Accelerator
Architecture -mcpu=power8 -qarch=pwr8 -m64
Disabled Optimization -O0 -O0 -qnoopt -O0
Recommended Optimization -O3 -mcpu=power8 -funroll-loops -O3 [-qipa] [-qhot] -fast -Mipa=fast,inline [-Msmartalloc]
Profile-Guided Optimization (PGO) -fprofile-generate/-fprofile-use -qpdf1/-qpdf2 -Mpfi/-Mpfo
Inter-Procedural Optimization -flto -qipa -Mipa
Vectorization -ftree-vectorize -qsimd=auto -Mvect
Automatic Parallelization -floop-parallelize-all -qsmp=auto -Mconcur
OpenMP Support -fopenmp -qsmp=omp -mp
OpenACC Offload N/A N/A -acc -Minfo=accel -ta:tesla:cc70
OpenMP Offload N/A -qoffload -qtgtarch=sm_60 N/A
Loop Optimization -fpeel-loops -funroll-loops -qhot
Debugging Symbols -g -g -g
Verbose -V

GPU Device Target Specification

The target device can be explicitely specified through the following options:

Target GCC IBM XL C/C++ IBM XL Fortran PGI Accelerator
Tesla K80 -qtgtarch=sm_37 -ta:cc37
Tesla P100 -qtgtarch=sm_60 -ta:cc60
Tesla V100 -qtgtarch=sm_70 -ta:cc70

For IBM XL C/C++ & IBM XL Fortran, the former -qxflag option has been replaced by the -qtgtarch option.

Compiler-Specific

IBM XL C/C++ & IBM XL Fortran

Optimization Levels

The IBM XL C/C++ and IBM XL Fortran compilers allow the following optimization levels:

Options Purpose
-O0 -qnohot -qnosmp -qstrict Disable any optimization
-O2 Limited optimization (safe)
-O3 -qstrict Medium optimization without modification of the code semantics
-O3 Medium optimization with potential modifications of the code semantics
-O4 Agressive optimization + Light Inter-Procedural Analysis (IPA)
-O5 Equivalent to -O4 level with deeper IPA

Increasing the optimization level tends to lead to a potentially significant increase in compilation time. It is therefore recommended to keep limited optimization levels while porting an application onto the platform.

For a given optimization level, the IBM XL compilers automatically introduce additional options by default. These options can be in turn complemented with other additional options. The table below details the main possible combinations:

Optimization Level Additional Options Introduced By Default Additional Options Other Options with Potential Benefits
-O0 -qsimd=auto -qarch=pwr9
-O2 -qmaxmem=8192 -qarch=pwr9 -qmaxmem=-1
-qsimd=auto -qtune=pwr9 -qhot=level=0
-qstrict=all
-O3 -qhot=noarraypad:level=0:novector:fastmath -qarch=pwr9
-qnostrict -qtune=pwr9
-qmaxmem=-1
-qsimd=auto
-O4 -qarch=auto -qarch=pwr9 -qsmp=auto
-qcache=auto -qcache
-qhot=noarraypad:level=1:vector:fastmath -qtune=pwr9
-qipa
-qmaxmem=-1
-qnostrict
-qsimd=auto
-qtune=auto
-O5 -qarch=auto -qarch=pwr9 -qsmp=auto
-qcache=auto -qcache
-qhot=noarraypad:level=1:vector:fastmath -qtune=pwr9
-qipa=level=2
-qmaxmem=-1
-qnostrict
-qsimd=auto
-qtune=auto

Other Compiler Options

  • Floating-Point and Integer Control
Option Purpose
-qautodbl={dbl4,dbl8,dbl} Convert single-precision floating-point calculations to double-precision and double-precision calculations to extended-precision
-qdpc Promote basic real constants as double-precision constants
-qintsize={2,4*,8} Set default size of INTEGER / LOGICAL data entities
-qrealsize={4*,8} Set default size of REAL / DOUBLE PRECISION / (DOUBLE) COMPLEX values
  • Object Code Control
Option Purpose
-qsaveopt Save compilation-related information to the corresponding object file

Note: Compilation-related information stored thanks to the -qsaveopt option can be read through the strings --all command:

$ strings -a <binary> | grep '^@(#)opt'
@(#)opt c /opt/ibm/xlC/13.1.6/bin/.orig/xlc_r -F/opt/ibm/xlC/13.1.6/etc/xlc.cfg.rhel.7.3.gcc.4.8.5.cuda.9.0 -DHAVE_CONFIG_H -I. -I/gpfs/shared/fftw -q64 -qfullpath -qsaveopt -O3 -qarch=pwr8 -qtune=pwr8 fftwf77.lo -MD -MP -MF .deps/fftwf77.Tpo -c fftwf77.c

PGI Accelerator

Default Compiler Options

A user-specific set of default compiler options can be defined through the following procedure:

  • Create Following File inside Home Directory:

    ~/.mypgirc

  • Populate File with Acceptable Directives:

    • Default Compute Capability:

      set DEFCOMPUTECAP=60;

    • Default CUDA Version:

      set DEFCUDAVERSION=10.1;

Note: These options would override any system-wide options defined by system administrator.

-fast Compilation Flag configuration

The -fast compilation flag is a shortcut for a consistent set of predefined optimization options.

Its current signification can be made explicit through the following command:

$ pgcc -fast -help
Reading rcfile /opt/pgi/linuxpower/20.02/bin/.pgccrc
-fast               Common optimizations; includes -O2 -Munroll=c:1 -Mlre -Mautoinline
                    == -Mvect=simd
-M[no]vect[=[no]simd|[no]assoc]
                    Control automatic vector pipelining
    [no]simd        Generate [don't generate] SIMD instructions
    [no]assoc       Allow [disallow] reassociation

results matching ""

    No results matching ""