papi_wrap

[jg piccinali]

• https://github.com/bcumming/papi-wrap

cirra/gnu (F90)

setup

cd /users/piccinal/papi-wrap.git/
module use \
/users/piccinal/SAVESPACE.EB/cirra/eb26/easybuild/apps/modules/all
mll PAPI/5.4.1

compile

make -f makefile.gnu CXX=g++ F90=gfortran \
PAPI_PATH=$EBROOTPAPI \
PAPI_LIB_PATH=$EBROOTPAPI/lib  clean all

• cd test/

make -f makefile.gnu F90=gfortran \
PAPI_PATH=$EBROOTPAPI \
PAPI_LIB_PATH=$EBROOTPAPI/lib clean saxpy

run

• export CSCSPERF_EVENTS="PAPI_TOT_CYC|PAPI_TOT_INS"
• export OMP_NUM_THREADS=1
• ./saxpy

----------------------------------------------
   Collector init :: wall time 0.346686 seconds
 [      TOTAL ]
 [  595390457 ] PAPI_TOT_CYC
 [  800007227 ] PAPI_TOT_INS
-----------------------------------------------
   Collector saxpy :: wall time 0.125008 seconds
 [      TOTAL ]
 [  448084488 ] PAPI_TOT_CYC
 [  800002134 ] PAPI_TOT_INS
------------------------------------------------
 handles with values           0 and           1
 the saxpy loop took   
200000000 floating point operations   
100000000.50000000   

• grep pw_ saxpy.f90

    call pw_new_collector('init', handle1)
    call pw_new_collector('saxpy', handle2)
    call pw_start_collector(handle1)
    call pw_stop_collector(handle1)
    call pw_start_collector(handle2)
    call pw_stop_collector(handle2)
    call pw_print()

CPI

cpi = PAPI_TOT_CYC/PAPI_TOT_INS
= 448084488 / 800002134 
= 0.56

cirra/gnu (C)

setup

cd /users/piccinal/papi-wrap.git/
module use \
/users/piccinal/SAVESPACE.EB/cirra/eb26/easybuild/apps/modules/all
mll PAPI/5.4.1

compile

• cd test/

make -f makefile.gnu CC=gcc \
PAPI_PATH=$EBROOTPAPI \
PAPI_LIB_PATH=$EBROOTPAPI/lib clean dot

run

• export CSCSPERF_EVENTS="PAPI_TOT_CYC|PAPI_TOT_INS"
• export OMP_NUM_THREADS=1
• ./dot

two collectors with handles 0 and 1
the sum is 3.33333e+11
------------------------------
   Collector initialize :: wall time 0.00921297 seconds
 [      TOTAL ]
 [    8266664 ] PAPI_TOT_CYC
 [    8003202 ] PAPI_TOT_INS
------------------------------
   Collector dot product :: wall time 0.00216699 seconds
 [      TOTAL ]
 [    3010844 ] PAPI_TOT_CYC
 [    6001928 ] PAPI_TOT_INS
------------------------------

• grep pw_ dot.c

    handle1 =  pw_new_collector("initialize");
    handle2 =  pw_new_collector("dot product");
    pw_start_collector(handle1);
    pw_stop_collector(handle1);
    pw_start_collector(handle2);
    pw_stop_collector(handle2);
    pw_print();

CPI

cpi = PAPI_TOT_CYC/PAPI_TOT_INS
= 3010844 / 6001928
= 0.501

[jg piccinali]

Cray/XC

src:

• ~/reframe.git/cscs-checks/tools/profiling_and_debugging/src/F90/

build

module load papi-wrap/1.0-CrayGNU-18.07
export CRAYPE_LINK_TYPE=dynamic
make FC="ftn -D_PAPIWRAP -fopenmp -std=c++11" \
  TOOL="-L$EBROOTPAPIMINWRAP/lib -L$PAT_BUILD_PAPI_BASEDIR/lib -lpapi -lpapi_wrap -lstdc++"

run

• OMP_NUM_THREADS=1 srun -Cgpu -n1 ./jacobi

MPI-3.1#1
 Jacobi           1 MPI process(es) with           1 OpenMP-      201307  thread(s)/process
--------------------------------------------------------------------------------
   Collector init :: wall time 0.063961 seconds
 [      TOTAL ]
 [  135147929 ] PAPI_TOT_CYC
 [  348038802 ] PAPI_TOT_INS
--------------------------------------------------------------------------------
   Collector jacobi :: wall time 0.531846 seconds
 [      TOTAL ]
 [ 1782792727 ] PAPI_TOT_CYC
 [ 4012856105 ] PAPI_TOT_INS

_main.F90

program MAIN
#ifdef _PAPIWRAP
    use m_papi_wrap
#endif
    use VariableDef
    use JacobiMod
    implicit none
    include 'mpif.h'

    TYPE(JacobiData) :: myData

#ifdef _PAPIWRAP
    integer :: papiw_handle1, papiw_handle2
#endif
!   sets default values or reads from stdin
!    * inits MPI and OpenMP if needed
!    * distribute MPI data, calculate MPI bounds
!    */
    call Init(mydata)
#ifdef _PAPIWRAP
    call pw_new_collector('init', papiw_handle1)
    call pw_new_collector('jacobi', papiw_handle2)
#endif

    if ( allocated(myData%afU) .and. allocated(myData%afF) ) then
!        /* matrix init */
#ifdef _PAPIWRAP
        call pw_start_collector(papiw_handle1)
#endif
        call InitializeMatrix(myData)
#ifdef _PAPIWRAP
        call pw_stop_collector(papiw_handle1)
#endif

!        /* starting timer */
        mydata%fTimeStart = MPI_Wtime()

!        /* running calculations */
#ifdef _PAPIWRAP
        call pw_start_collector(papiw_handle2)
#endif
        call Jacobi(myData)
#ifdef _PAPIWRAP
        call pw_stop_collector(papiw_handle2)
#endif

!        /* stopping timer */
        mydata%fTimeStop = MPI_Wtime()

!        /* error checking */
        call CheckError(myData)

!        /* print result summary */
        call PrintResults(myData)
    else
        write (*,*) " Memory allocation failed ...\n"
    end if

!    /* cleanup */
#ifdef _PAPIWRAP
    call pw_print()
#endif
    call Finish(myData)

end program MAIN