int ipm_control(const int ctl, char *cmd, void *data)
{
char *reg;
switch( ctl )
{
case 1: /* enter region, IPM1 compatible */
CHECK_REGION(reg, cmd);
if(reg)
ipm_region(1, reg);
break;
case -1: /* exit region, IPM1 compatible */
if(reg)
CHECK_REGION(reg, cmd);
ipm_region(-1, reg);
break;
/* general case */
case 0:
default:
if( !strncmp(cmd, "enter", 5) ) {
CHECK_REGION(reg, data);
if(reg)
ipm_region(1, reg);
}
if( !strncmp(cmd, "exit", 4) ) {
CHECK_REGION(reg, data);
if(reg)
ipm_region(-1, reg);
}
#ifdef HAVE_POSIXIO_TRACE
if( !strncmp(cmd, "traceoff", 8) )
task.tracestate = 0;
if( !strncmp(cmd, "traceon", 7) )
task.tracestate = 1;
#endif
}
return 0;
}
void foobar() {
ipm_region(1, "foobar");
foo();
bar();
ipm_region(-1, "foobar");
}
void bar() {
ipm_region(1, "bar");
foo();
ipm_region(-1, "bar");
}
void foo() {
ipm_region(1, "foo");
usleep(10000);
ipm_region(-1, "foo");
}
int main(int argc, char *argv[]) {
int i,j,k,ii,ireg,three_reg=0;
double gflops;
volatile double a = 0.5, b = 2.2;
double c = 0.11;
long int nflop=128*1024*1024;
long int nfma = 0;
task.flags = 0;
ipm_init(task.flags);
/*
IPM_CLEAR_TASKDATA(task);
task.mpi_rank = 0;
task.mpi_size = 1;
task.hpm_eventset = 0;
*/
task.flags |= IPM_HPM_ACTIVE;
while(--argc && argv++) {
if(!strcmp("-v",*argv)) {
task.flags |= VERBOSE;
} else if(!strcmp("-x",*argv)) {
task.hpm_eventset = -1;
task.flags &= ~IPM_HPM_ACTIVE;
} else if(!strcmp("-r",*argv)) {
three_reg = 1;
} else if(!strcmp("-s",*argv)) {
--argc; argv++;
task.hpm_eventset = (int)atol(*argv);
} else if(!strcmp("-n",*argv)) {
--argc; argv++;
nflop = atol(*argv);
/* printf("nflop = 2 * %ld\n", nflop); */
} else if(!strcmp("-fma",*argv)) {
--argc; argv++;
nfma = atol(*argv);
}
}
ipm_hpm_start();
if(0) {
printf("IPM: %d hpm_test eventset=%d events :", task.mpi_rank, task.hpm_eventset);
for(i=0;i<MAXSIZE_HPMCOUNTERS;i++) {
ii = ipm_hpm_eorder[task.hpm_eventset][i];
printf("%s ", ipm_hpm_ename[task.hpm_eventset][ii]);
}
printf("\n");
for(ireg=0;ireg<MAXSIZE_REGION;ireg++) {
printf("IPM: %d hpm_test init region(%d) %s count",
task.mpi_rank,ireg, task.region_label[ireg]);
for(i=0;i<MAXSIZE_HPMCOUNTERS;i++) {
ii = ipm_hpm_eorder[task.hpm_eventset][i];
printf(" %lld", task.hpm_count[ireg][task.hpm_eventset][ii]);
}
printf("\n");
}
}
if(!three_reg) {
ipm_region(1,"ipm_noregion");
ipm_region(-1,"ipm_noregion");
do_flops(nflop);
do_fmas(nfma);
ipm_region(-1,"ipm_noregion");
} else {
ipm_region(1,"ipm_noregion");
do_flops(nflop);
ipm_region(-1,"ipm_noregion");
ipm_region(1,"a");
do_flops(nflop*2);
ipm_region(-1,"a");
ipm_region(1,"b");
do_flops(nflop*3);
ipm_region(-1,"b");
}
ipm_hpm_stop();
IPM_PRINT_TASKDATA(task);
return 0;
}
int ipm_finalize(int flags)
{
int rv, i;
if(ipm_state!=STATE_ACTIVE &&
ipm_state!=STATE_NOTACTIVE ) {
IPMERR("ipm_finalize() called with ipm_state=%d\n", ipm_state);
return IPM_EOTHER;
}
ipm_state = STATE_IN_FINALIZE;
/* this should be close to user code */
ipm_region_end(&ipm_app);
ipm_region(-1, "ipm_main");
/* update global timing statistics */
gettimeofday( &(task.t_stop), 0);
task.wtime = ipm_wtime() -task.wtime;
task.utime = ipm_utime() -task.utime;
task.stime = ipm_stime() -task.stime;
task.mtime = ipm_mtime() -task.mtime;
task.iotime = ipm_iotime() -task.iotime;
task.omptime = ipm_omptime()-task.omptime;
ipm_get_procmem(&(task.procmem));
task.procmem /= (1024.0*1024.0*1024.0);
#ifdef HAVE_SELFMONITOR
ipm_selfmon.t_finalize = ipm_wtime()-ipm_selfmon.t_finalize;
#endif
#ifdef HAVE_PAPI
//rstack_adjust_ctrs();
#endif
/* write out banner report */
if( !(task.flags&FLAG_REPORT_NONE) ) {
fflush(stdout);
ipm_banner(stdout);
}
#if defined(HAVE_MPI) && defined(HAVE_CALLPATH)
ipm_unify_callsite_ids();
#endif
/* call output routine for each module */
for( i=0; i<MAXNUM_MODULES; i++ ) {
if( i==IPM_MODULE_SELFMONITOR ||
i==IPM_MODULE_MPI)
continue;
if( modules[i].output!=0 ) {
IPMDBG("calling output() for module %d\n", i);
rv = modules[i].output(&(modules[i]), flags);
}
}
ipm_write_profile_log();
rstack_cleanup(ipm_rstack);
#ifdef HAVE_SELFMONITOR
mod_selfmonitor_output(&(modules[IPM_MODULE_SELFMONITOR]), flags);
#endif
/* call finalization routine for each module */
for( i=0; i<MAXNUM_MODULES; i++ ) {
if( modules[i].finalize!=0 ) {
IPMDBG("calling finalize() for module %d\n", i);
rv = modules[i].finalize(&(modules[i]), flags);
}
}
/* TODO: check errors in modules */
ipm_state=STATE_FINALIZED;
return IPM_OK;
}
int ipm_init(int flags)
{
int i, state, rv;
double t_init;
char cmdline[MAXSIZE_CMDLINE];
char realpath[MAXSIZE_CMDLINE];
char *target;
state=ipm_state;
ipm_state=STATE_IN_INIT;
/* check if IPM_TARGET is set and if it is,
only monitor matching processes */
target=getenv("IPM_TARGET");
ipm_get_exec_cmdline(cmdline,realpath);
/* IPM_TARGET specifies string that has to appear in cmdline */
if( target && target[0]!='!' && !strstr(cmdline,target) ) {
ipm_state=STATE_NOTACTIVE;
return IPM_OK;
}
/* IPM_TARGET specifies string that must not appear in cmdline */
if( target && target[0]=='!' && strstr(cmdline,target) ) {
ipm_state=STATE_NOTACTIVE;
return IPM_OK;
}
ipm_time_init(flags);
rstack_init(flags);
t_init = ipm_wtime();
taskdata_init(&task);
htable_init(ipm_htable);
/* need to get env variables before modules init */
ipm_get_env();
/* init local data structures */
for( i=0; i<MAXNUM_MODULES; i++ ) {
ipm_module_init( &(modules[i]) );
}
/* --- initialize modules --- */
#ifdef HAVE_MPI
modules[IPM_MODULE_MPI].init=mod_mpi_init;
#endif
#ifdef HAVE_POSIXIO
modules[IPM_MODULE_POSIXIO].init=mod_posixio_init;
#endif
#ifdef HAVE_MPIIO
modules[IPM_MODULE_MPIIO].init=mod_mpiio_init;
#endif
#ifdef HAVE_CALLPATH
modules[IPM_MODULE_CALLPATH].init=mod_callpath_init;
#endif
#ifdef HAVE_KEYHIST
modules[IPM_MODULE_KEYHIST].init=mod_keyhist_init;
#endif
#ifdef HAVE_PAPI
modules[IPM_MODULE_PAPI].init=mod_papi_init;
#endif
#ifdef HAVE_SELFMONITOR
modules[IPM_MODULE_SELFMONITOR].init=mod_selfmonitor_init;
#endif
#ifdef HAVE_PROCCTRL
modules[IPM_MODULE_PROCCTRL].init=mod_procctrl_init;
#endif
#ifdef HAVE_CLUSTERING
modules[IPM_MODULE_CLUSTERING].init=mod_clustering_init;
#endif
#ifdef HAVE_OMPTRACEPOINTS
modules[IPM_MODULE_OMPTRACEPOINTS].init=mod_omptracepoints_init;
#endif
#ifdef HAVE_CUDA
modules[IPM_MODULE_CUDA].init=mod_cuda_init;
#endif
#ifdef HAVE_CUFFT
modules[IPM_MODULE_CUFFT].init=mod_cufft_init;
#endif
#ifdef HAVE_CUBLAS
modules[IPM_MODULE_CUBLAS].init=mod_cublas_init;
#endif
/* TODO: handle errors in module initialization, set
ipm_state to STATE_ERROR */
for( i=0; i<MAXNUM_MODULES; i++ ) {
if( modules[i].init ) { /* call init function if it is set */
rv=modules[i].init(&(modules[i]), flags);
if(rv!=IPM_OK) {
IPMERR("Error initializing module %d (%s), error %d\n",
i, modules[i].name?modules[i].name:"", rv);
}
#ifdef HAVE_MPI
if( i==IPM_MODULE_POSIXIO )
modules[i].state=STATE_NOTACTIVE;
#endif
}
}
/* --- done initializing modules --- */
#ifdef DELAYED_MPI_FINALIZE
rv = atexit(ipm_atexit_handler);
if(!rv) {
task.flags|=FLAG_USING_ATEXIT;
} else {
IPMERR("Error installing atexit() handler\n");
task.flags&=~FLAG_USING_ATEXIT;
}
#endif
signal(SIGXCPU, ipm_sig_handler);
signal(SIGTERM, ipm_sig_handler);
signal(SIGABRT, ipm_sig_handler);
#ifdef HAVE_SELFMONITOR
ipm_selfmon.t_init = ipm_wtime()-t_init;
#endif
/* this should be close to user code */
ipm_region(1, "ipm_main");
ipm_region_begin(&ipm_app);
ipm_state=STATE_ACTIVE;
return IPM_OK;
}
