void FATR util_mic_set_affinity_() {
char affinity[BUFSZ];
char num_threads[BUFSZ];
int pos;
int micdev;
int nprocs;
int ranks_per_dev;
int rank_on_dev;
int nthreads;
int ppn;
int ranks_per_device=util_getenv_nwc_ranks_per_device_();
if (ranks_per_device == 0) {
return ;
} else if (ranks_per_device < 0){
ranks_per_device = RANKS_PER_DEVICE;
}
pos=snprintf(affinity, BUFSZ, "KMP_PLACE_THREADS=");
micdev=util_mic_get_device_();
ppn=util_cgetppn();
#pragma offload target(mic:micdev) out(nprocs)
{
/* do one offload to query the coprocessor for the number of cores */
nprocs = ((int) sysconf(_SC_NPROCESSORS_ONLN) / 4) - 1;
}
rank_on_dev = util_my_smp_index() % util_nwc_ranks_per_device_();
nthreads = nprocs / ranks_per_device * DEFAULT_OFFLOAD_THREAD_MULTIPLIER;
pos+=snprintf(affinity+pos, BUFSZ-pos, "%dc,%dt,%do",
nprocs / ranks_per_device, DEFAULT_OFFLOAD_THREAD_MULTIPLIER,
rank_on_dev * (nprocs / ranks_per_device));
snprintf(num_threads, BUFSZ, "OMP_NUM_THREADS=%d", nthreads);
printf("%02d: micdev=%d nprocs=%d rank_on_dev=%d ranks_per_device=%d affinity='%s' pos=%d\n",
GA_Nodeid(), micdev, nprocs, rank_on_dev, ranks_per_device, affinity, pos);
fflush(stdout);
#pragma offload target(mic:micdev) in(affinity) in(num_threads)
{
/* set the affinity masks and the number of offloaded OpenMP threads */
kmp_set_defaults("KMP_AFFINITY=compact");
kmp_set_defaults(affinity);
kmp_set_defaults(num_threads);
}
}
void calculate_production_rate(REACTION_DATA_ver2 &reaction_data, vector<double> &Rf, vector<double> &Rb, vector<double> &S_chem)
{
kmp_set_defaults("KMP_AFFINITY = scatter");
if (S_chem.size() != reaction_data.NS) S_chem.resize(reaction_data.NS);
#pragma omp parallel for
for (int s = 0; s <= reaction_data.NS-1; s++) S_chem[s] = 0.0;
#pragma omp parallel for
for (int s = 0; s <= reaction_data.NS-1; s++)
{
for (int k = 0; k <= reaction_data.NR-1; k++)
{
double Rf_m_Rb = Rf[k] - Rb[k];
S_chem[s] += (reaction_data.reaction_k[k].Product_coeff[s] - reaction_data.reaction_k[k].Reactant_coeff[s]) * Rf_m_Rb;
}
}
#pragma omp parallel for
for (int s = 0; s <= reaction_data.NS-1; s++)
S_chem[s] = S_chem[s] * reaction_data.species_data[s].basic_data.M;
}
int main(int argc, char* argv[]) {
try {
cmdline::parser cl;
cl.add<int>("threads", 't', "number of threads", false);
cl.add<int>("iter", 'i', "millions of iterations", false, 10);
cl.parse_check(argc, argv);
std::cout << "Initializing" << std::endl;
if (cl.exist("threads")) {
omp_set_num_threads(cl.get<int>("threads"));
}
#if 0
kmp_set_defaults("KMP_AFFINITY=compact");
#endif
#pragma omp parallel for
for (auto i = 0; i < FLOPS_ARRAY_SIZE; i++) {
fa[i] = (float)i + 0.1f;
fb[i] = (float)i + 0.2f;
}
int max_threads = omp_get_max_threads();
std::cout << "Starting compute on " << max_threads << " threads"
<< std::endl;
Timer timer;
float a = 1.1f;
int iters = 1000000 * cl.get<int>("iter");
#pragma omp parallel for
for (int i = 0; i < max_threads; ++i) {
// each thread will work its own array section
int offset = i * LOOP_COUNT;
// loop many times to get lots of calculations
for (int j = 0; j < iters; ++j) {
// scale 1st array and add in the 2nd array
for (int k = 0; k < LOOP_COUNT; ++k) {
fa[k + offset] = a * fa[k + offset] + fb[k + offset];
}
}
}
double gflops =
(double)(1e-9 * max_threads * LOOP_COUNT * iters * FLOPSPERCALC);
// elasped time
double sec = timer();
std::cout << "Gflops = " << gflops << std::endl;
std::cout << "secs = " << sec << std::endl;
std::cout << "Gflops/s = " << (gflops / sec) << std::endl;
}
catch (std::string& ex) {
std::cerr << "error: " << ex << std::endl;
}
catch (const char* ex) {
std::cerr << "error: " << ex << std::endl;
}
}
int test_kmp_set_defaults_lock_bug()
{
/* checks that omp_get_num_threads is equal to the number of
threads */
int nthreads_lib;
int nthreads = 0;
nthreads_lib = -1;
#pragma omp parallel
{
omp_set_lock(&lock);
nthreads++;
omp_unset_lock(&lock);
#pragma omp single
{
nthreads_lib = omp_get_num_threads ();
} /* end of single */
} /* end of parallel */
kmp_set_defaults("OMP_NUM_THREADS");
#pragma omp parallel
{
omp_set_lock(&lock);
nthreads++;
omp_unset_lock(&lock);
} /* end of parallel */
return (nthreads == 2*nthreads_lib);
}
void initomp (int nthreads, int verbose)
{
char schedule[1024];
if (verbose == 1)
{
sprintf (schedule,
"KMP_AFFINITY=granularity=fine,compact,verbose");
}
else
{
sprintf (schedule,
"KMP_AFFINITY=granularity=fine,compact");
}
kmp_set_defaults (schedule);
omp_set_num_threads (nthreads);
}
/*
* FN-01: Prepararization OP2A
* @author Minkwan Kim
* @version 1.0 25/5/2015
*/
void ApplicationOP2A::preparation(int argc, char *argv[], string modulename)
{
/*
* ==============================================
* STEP 1: Initialize Parallel Communication:
* - Development Status : Done
* - Last modified on: July 23, 2014
* by: Minkwan Kim
* =============================================
* */
time_running.initStartTime();
#ifdef MPI
MPI_Init(&argc, &argv); // INITIALIZE MPI
MPI_Comm_size(MPI_COMM_WORLD, &NP); // FINDOUT HOW MANY PROCESSORS IN THERE
MPI_Comm_rank(MPI_COMM_WORLD, &P); // FINDOUT WHICH PROCESSOR I AM
t0 = MPI_Wtime();
#endif
if (NP > OP2A_MAX_N_TASK)
throw Common::ExceptionNPExceed (FromHere(), "Number of processors exceeds MAX_PROCESSOR. Need to adjust value of MAX_N_TASKS");
kmp_set_defaults("KMP_AFFINITY = scatter");
/*
* =================================================
* STEP 2: Show Version information:
* - Development Status : Done
* - Last modified on: July 23, 2014
* - by: Minkwan Kim
* =================================================
*/
Version versionOP2A(OP2A_VERSION_MAIN, OP2A_VERSION_SUB, m_now->tm_year + 1900, m_now->tm_mon + 1, m_now->tm_mday, modulename.c_str());
if (P == 0)
{
versionOP2A.info(); // Show the version information
cout << " --> t = " << time_running.getDeltaT() << "[sec]" << endl << endl;
}
}
void calculate_reaction_rate(REACTION_DATA_ver2 &reaction_data, vector<double> &rhos, vector<double> &T, vector<double> &kf, vector<double> &kb, vector<double> &Rf, vector<double> &Rb)
{
kmp_set_defaults("KMP_AFFINITY = scatter");
// 1. Initialize Rf and Rb
Rf.reserve(reaction_data.NR);
Rb.reserve(reaction_data.NR);
// 2. Calculate 10^-3 * rho_s / Ms
vector<double> rhos_Ms(reaction_data.NS, 0.0);
#pragma omp parallel for
for (int s = 0; s <= reaction_data.NS-1; s++) rhos_Ms[s] = 0.001 * rhos[s]/reaction_data.species_data[s].basic_data.M;
double n_mix = 0.0;
for (int s = 0; s <= reaction_data.NS-1; s++) n_mix = n_mix + rhos[s]/reaction_data.species_data[s].basic_data.m;
n_mix = n_mix * 1.0e-6;
//3. Calculate Forward/backward reaction rate
for (int k = 0; k <= reaction_data.NR-1; k++)
{
double Tf, Tb;
calculate_reaction_temperature(reaction_data.reaction_k[k], T, Tf, Tb);
kf[k] = cal_kf(reaction_data.reaction_k[k], Tf);
kb[k] = cal_kb(reaction_data.reaction_k[k], n_mix, Tb);
double Rf_temp = 1.0;
for (int j = 0; j <= reaction_data.NS-1; j++) Rf_temp *= pow(rhos_Ms[j], reaction_data.reaction_k[k].Reactant_coeff[j]);
Rf[k] = 1000.0*kf[k] * Rf_temp;
double Rb_temp = 1.0;
for (int j = 0; j <= reaction_data.NS-1; j++) Rf_temp *= pow(rhos_Ms[j], reaction_data.reaction_k[k].Product_coeff[j]);
Rb[k] = 1000.0*kb[k] * Rb_temp;
}
}
/***************************************************************************//**
*
* @ingroup Auxiliary
*
* PLASMA_Init_Affinity - Initialize PLASMA.
*
*******************************************************************************
*
* @param[in] cores
* Number of cores to use (threads to launch).
* If cores = 0, cores = PLASMA_NUM_THREADS if it is set, the
* system number of core otherwise.
*
* @param[in] coresbind
* Array to specify where to bind each thread.
* Each thread i is binded to coresbind[hwloc(i)] if hwloc is
* provided, or to coresbind[i] otherwise.
* If coresbind = NULL, coresbind = PLASMA_AFF_THREADS if it
* is set, the identity function otherwise.
*
*******************************************************************************
*
* @return
* \retval PLASMA_SUCCESS successful exit
*
******************************************************************************/
int PLASMA_Init_Affinity(int cores, int *coresbind)
{
plasma_context_t *plasma;
int status;
int core;
/* Create context and insert in the context map */
plasma = plasma_context_create();
if (plasma == NULL) {
plasma_fatal_error("PLASMA_Init", "plasma_context_create() failed");
return PLASMA_ERR_OUT_OF_RESOURCES;
}
status = plasma_context_insert(plasma, pthread_self());
if (status != PLASMA_SUCCESS) {
plasma_fatal_error("PLASMA_Init", "plasma_context_insert() failed");
return PLASMA_ERR_OUT_OF_RESOURCES;
}
/* Init number of cores and topology */
plasma_topology_init();
/* Set number of cores */
if ( cores < 1 ) {
plasma->world_size = plasma_get_numthreads();
if ( plasma->world_size == -1 ) {
plasma->world_size = 1;
plasma_warning("PLASMA_Init", "Could not find the number of cores: the thread number is set to 1");
}
}
else
plasma->world_size = cores;
if (plasma->world_size <= 0) {
plasma_fatal_error("PLASMA_Init", "failed to get system size");
return PLASMA_ERR_NOT_FOUND;
}
/* Check if not more cores than the hard limit */
if (plasma->world_size > CONTEXT_THREADS_MAX) {
plasma_fatal_error("PLASMA_Init", "not supporting so many cores");
return PLASMA_ERR_INTERNAL_LIMIT;
}
/* Get the size of each NUMA node */
plasma->group_size = plasma_get_numthreads_numa();
while ( ((plasma->world_size)%(plasma->group_size)) != 0 )
(plasma->group_size)--;
/* Initialize barriers */
plasma_barrier_init(plasma);
plasma_barrier_bw_init(plasma);
/* Initialize default thread attributes */
status = pthread_attr_init(&plasma->thread_attr);
if (status != 0) {
plasma_fatal_error("PLASMA_Init", "pthread_attr_init() failed");
return status;
}
/* Set scope to system */
status = pthread_attr_setscope(&plasma->thread_attr, PTHREAD_SCOPE_SYSTEM);
if (status != 0) {
plasma_fatal_error("PLASMA_Init", "pthread_attr_setscope() failed");
return status;
}
/* Set concurrency */
status = pthread_setconcurrency(plasma->world_size);
if (status != 0) {
plasma_fatal_error("PLASMA_Init", "pthread_setconcurrency() failed");
return status;
}
/* Launch threads */
memset(plasma->thread_id, 0, CONTEXT_THREADS_MAX*sizeof(pthread_t));
if (coresbind != NULL) {
memcpy(plasma->thread_bind, coresbind, plasma->world_size*sizeof(int));
}
else {
plasma_get_affthreads(plasma->thread_bind);
}
/* Assign rank and thread ID for the master */
plasma->thread_rank[0] = 0;
plasma->thread_id[0] = pthread_self();
for (core = 1; core < plasma->world_size; core++) {
plasma->thread_rank[core] = core;
pthread_create(
&plasma->thread_id[core],
&plasma->thread_attr,
plasma_parallel_section,
(void*)plasma);
}
/* Ensure BLAS are sequential and set thread affinity for the master */
#if defined(PLASMA_WITH_MKL)
#if defined(__ICC) || defined(__INTEL_COMPILER)
kmp_set_defaults("KMP_AFFINITY=disabled");
#endif
#endif
/* Initialize the dynamic scheduler */
plasma->quark = QUARK_Setup(plasma->world_size);
plasma_barrier(plasma);
plasma_setlapack_sequential(plasma);
return PLASMA_SUCCESS;
}
//
// Main program - pedal to the metal...calculate using tons o'flops!
//
int main(int argc, char *argv[] )
{
int i,j,k;
int numthreads;
double tstart, tstop, ttime;
double gflops = 0.0;
float a=1.1;
//
// initialize the compute arrays
//
//
omp_set_num_threads(2);
kmp_set_defaults("KMP_AFFINITY=compact");
#pragma omp parallel
#pragma omp master
numthreads = omp_get_num_threads();
printf("Initializing\r\n");
#pragma omp parallel for
for(i=0; i<FLOPS_ARRAY_SIZE; i++)
{
fa[i] = (float)i + 0.1;
fb[i] = (float)i + 0.2;
}
printf("Starting Compute on %d threads\r\n",numthreads);
tstart = dtime();
// scale the calculation across threads requested
// need to set environment variables OMP_NUM_THREADS and KMP_AFFINITY
#pragma omp parallel for private(j,k)
for (i=0; i<numthreads; i++)
{
// each thread will work it's own array section
// calc offset into the right section
int offset = i*LOOP_COUNT;
// loop many times to get lots of calculations
for(j=0; j<MAXFLOPS_ITERS; j++)
{
// scale 1st array and add in the 2nd array
for(k=0; k<LOOP_COUNT; k++)
{
fa[k+offset] = a * fa[k+offset] + fb[k+offset];
}
}
}
tstop = dtime();
// # of gigaflops we just calculated
gflops = (double)( 1.0e-9*numthreads*LOOP_COUNT*
MAXFLOPS_ITERS*FLOPSPERCALC);
//elasped time
ttime = tstop - tstart;
//
// Print the results
//
if ((ttime) > 0.0)
{
printf("GFlops = %10.3lf, Secs = %10.3lf, GFlops per sec = %10.3lf\r\n", gflops, ttime, gflops/ttime);
}
return( 0 );
}
int main(int argc, char *argv[])
{
#ifdef _OPENMP
#ifndef KMP_AFFINITY
kmp_set_defaults("KMP_AFFINITY=compact, granularity=fine");
// kmp_set_defaults("KMP_AFFINITY=scatter, granularity=fine");//this gives much slower performance...
#endif
#pragma omp parallel
#pragma omp master
//#ifndef OMP_NUM_THREADS
omp_set_num_threads(240);
//#endif
printf("\nComputing 7-point stencil on Intel Xeon Phi in %d threads.\n", omp_get_num_threads());
#endif
const int nx = problem_dim;
const int ny = problem_dim;
const int nz = problem_dim;
const int problem_size = sizeof(float)*nx*ny*nz;
float *f1 = (float *)_mm_malloc(problem_size, 64);
float *f2 = (float *)_mm_malloc(problem_size, 64);
assert(f1 != MAP_FAILED);
assert(f2 != MAP_FAILED);
float *answer = (float *)_mm_malloc(problem_size, 64);
float *f_final = NULL;
int count = 0;
float c0, c1;
float l = 1.0;
float kappa = 0.1;
float dx = l / nx;
float dy = l / ny;
float dz = l / nz;
float dt = 0.1*dx*dx / kappa;
float scale = 0.1;
count = scale / dt;
f_final = (count % 2)? f2 : f1;
create_field<float>(f1, nx, ny, nz, dx, dy, dz, kappa, 0.0);
c1 = kappa*dt/(dx*dx);
c0 = 1.0 - 6*c1;
printf("Running heat kernel %d times\n", count);
fflush(stdout);
float *f1_t = f1;
float *f2_t = f2;
struct timeval time_begin, time_end;
gettimeofday(&time_begin, NULL);
for (int i = 0; i < count; ++i) {
compute_7pt_stencile_mm512_ps(f2_t, f1_t, nx, ny, nz, c0, c1);
float *t = f1_t;
f1_t = f2_t;
f2_t = t;
}
gettimeofday(&time_end, NULL);
float time = count * dt;
create_field<float>(answer, nx, ny, nz, dx, dy, dz, kappa, time);
float err = accuracy<float>(answer,f_final, nx*ny*nz);
double elapsed_time = (time_end.tv_sec - time_begin.tv_sec)
+ (time_end.tv_usec - time_begin.tv_usec)*1.0e-6;
float Gflops = (nx*ny*nz)*8.0*count/elapsed_time * 1.0e-09;
float Gstens = (nx*ny*nz)*1.0*count/elapsed_time * 1.0e-06;
double thput = (nx * ny * nz) * sizeof(float) * 3.0 * count
/ elapsed_time * 1.0e-09;
fprintf(stderr, "Elapsed time : %.3f (s)\n", elapsed_time);
fprintf(stderr, "FLOPS : %.3f (GFlops)\n", Gflops);
fprintf(stderr, "Updates : %.3f (Mupdates/sec)\n", Gstens);
fprintf(stderr, "Throughput : %.3f (GB/s)\n", thput);
fprintf(stderr, "Accuracy : %e\n", err);
_mm_free(f1);
_mm_free(f2);
_mm_free(answer);
return 0;
}
int main(int argc, char* argv[])
{
double sum_delta = 0.0, sum_ref = 0.0, L1norm = 0.0;
unsigned int seed = 123;
int verbose = 0;
if (argc > 2)
{
printf("usage: Black-Scholes <verbose> verbose = 1 for validtating result, the default is 0. \n");
exit(1);
}
if (argc == 1)
verbose = 0;
else if (argc == 2)
verbose = atoi(argv[1]);
kmp_set_defaults("KMP_AFFINITY=compact,granularity=fine");
#ifdef _OPENMP
int ThreadNum = omp_get_max_threads();
omp_set_num_threads(ThreadNum);
#else
int ThreadNum = 1;
#endif
int OptPerThread = OPT_N / ThreadNum;
int mem_size = sizeof(double) * OptPerThread;
setlocale(LC_ALL,"");
printf("Black-Scholes Formula Double Precision.\n");
printf("Compiler Version = %d\n", __INTEL_COMPILER/100);
printf("Release Update = %d\n", __INTEL_COMPILER_UPDATE);
printf("Build Time = %s %s\n", __DATE__, __TIME__);
printf("Input Dataset = %d\n", OPT_N);
printf("Repetitions = %d\n", NUM_ITERATIONS);
printf("Chunk Size = %d\n", CHUNKSIZE);
printf("Worker Threads = %d\n\n", ThreadNum);
if (verbose)
printf("Allocate and initialize memory on %d boundary,\n", SIMDALIGN);
#pragma omp parallel reduction(+ : sum_delta) reduction(+ : sum_ref)
{
#ifdef _OPENMP
int threadID = omp_get_thread_num();
#else
int threadID = 0;
#endif
double *CallResult = (double *)_mm_malloc(mem_size, SIMDALIGN);
double *PutResult = (double *)_mm_malloc(mem_size, SIMDALIGN);
double *StockPrice = (double *)_mm_malloc(mem_size, SIMDALIGN);
double *OptionStrike = (double *)_mm_malloc(mem_size, SIMDALIGN);
double *OptionYears = (double *)_mm_malloc(mem_size, SIMDALIGN);
seed += threadID;
for(int i = OptPerThread-1; i > -1 ; i--)
{
CallResult[i] = 0.0;
PutResult[i] = -1.0;
StockPrice[i] = RandDouble(5.0, 30.0, &seed);
OptionStrike[i] = RandDouble(1.0, 100.0, &seed);
OptionYears[i] = RandDouble(0.25, 10.0, &seed);
}
#pragma omp barrier
if (threadID == 0) {
start_cyc = _rdtsc();
}
for(int i = 0; i < NUM_ITERATIONS; i++)
for (int chunkBase = 0; chunkBase < OptPerThread; chunkBase += CHUNKSIZE)
{
#pragma simd vectorlength(CHUNKSIZE)
#pragma simd
#pragma vector aligned
for(int opt = chunkBase; opt < (chunkBase+CHUNKSIZE); opt++)
{
double CNDD1, CNDD2;
double T = OptionYears[opt];
double X = OptionStrike[opt];
double XexpRT = X*exp2(RLOG2E * T);
double S = StockPrice[opt];
double sqrtT = sqrt(T);
double d1 = log2(S / X) / (VLOG2E * sqrtT) + RVV * sqrtT;
CNDF_C (&CNDD1, &d1 );
double d2 = d1 - VOLATILITY * sqrtT;
CNDF_C (&CNDD2, &d2 );
double CallVal = S * CNDD1 - XexpRT * CNDD2;
double PutVal = CallVal + XexpRT - S;
CallResult[opt] = CallVal ;
PutResult[opt] = PutVal ;
}
}
#pragma omp barrier
if (threadID == 0) {
end_cyc = _rdtsc();
}
if (verbose)
{
double delta = 0.0, ref = 0.0, L1norm = 0.0;
int max_index = 0;
for(int i = 0; i < OptPerThread; i++)
{
double callReference, putReference;
BlackScholesReference(
callReference,
putReference,
StockPrice[i], //Stock price
OptionStrike[i], //Option strike
OptionYears[i], //Option years
RISKFREE, //Riskless rate
VOLATILITY //Volatility rate
);
ref = callReference;
delta = fabs(callReference - CallResult[i]);
sum_delta += delta;
sum_ref += fabs(ref);
}
}
_mm_free(CallResult);
_mm_free(PutResult);
_mm_free(StockPrice);
_mm_free(OptionStrike);
_mm_free(OptionYears);
} //parallel section
const unsigned long long cyc = end_cyc - start_cyc;
double sec = cyc/(FREQ*1e9);
printf("=============================================\n");
printf("Total Cycles = %lld\n", cyc);
printf("Cycles/OptionPair at thread 0 = %5.2f\n", cyc/(1.0f*NUM_ITERATIONS*OptPerThread));
printf("Time Elapsed = %5.2f\n", sec);
printf("Options/sec = %5.2f\n", (2.0f*NUM_ITERATIONS*OPT_N)/(1e9*sec));
printf("=============================================\n");
if (verbose)
{
L1norm = sum_delta / sum_ref;
printf("L1 norm: %E\n", L1norm);
printf((L1norm < 9e-5) ? "TEST PASSED\n" : "TEST FAILED\n");
}
}