int main(int argc, char *argv[])
{
if (argc < 2) {
printf("Missing size of array!\n");
return EXIT_FAILURE;
}
int size_array = atoi(argv[1]);
int *array = (int *) malloc(size_array * sizeof(uint32_t));
for (int i = 0; i < size_array; i++) {
array[i] = getrand(0, 100000);
// printf ("array[%d]= %d ",i,array[i]);
}
double time = wtime();
for (int i = 0; i < size_array - 1; i++) {
for (int j = 0; j < size_array - i - 1; j++) {
if (array[j] > array[j + 1]) {
int tmp = array[j];
array[j] = array[j + 1];
array[j + 1] = tmp;
}
}
}
/* for (int i = 0; i < size_array; i++) {
printf ("array[%d]= %d ",i,array[i]);
}*/
time = wtime() - time;
FILE *tb;
tb = fopen("bubblesort.dat", "a");
fprintf(tb, "%d %.6f\n", size_array, time);
free(array);
return EXIT_SUCCESS;
}
int main(int argc, char **argv){
int i, me, target;
unsigned int size;
double t;
MPI_Status status;
MPI_Init(&argc, &argv);
MPI_Comm_rank(MPI_COMM_WORLD, &me);
target = 1 - me;
init_buf(send_buf, me);
init_buf(recv_buf, target);
if(me==0) print_items();
for(size=1;size<MAX_SIZE+1;size*=2){
MPI_Barrier(MPI_COMM_WORLD);
for(i=0;i<LOOP+WARMUP;i++){
if(WARMUP == i)
t = wtime();
if(me == 0){
MPI_Send(send_buf, size, MPI_CHAR, target, 9, MPI_COMM_WORLD);
MPI_Recv(recv_buf, size, MPI_CHAR, target, 5, MPI_COMM_WORLD, &status);
}
else {
MPI_Recv(recv_buf, size, MPI_CHAR, target, 9, MPI_COMM_WORLD, &status);
MPI_Send(send_buf, size, MPI_CHAR, target, 5, MPI_COMM_WORLD);
}
}
MPI_Barrier(MPI_COMM_WORLD);
t = wtime() - t;
if(me == 0)
print_results(size, t);
}
MPI_Finalize();
return 0;
}
// startrun: startup hierarchical N-body code.
// ___________________________________________
// This runs once.
local void startrun(void) {
printf("startrun\n");
startrun_time_0 = wtime();
bodyptr p1, p2, p;
stream gravstr;
define_body(sizeof(body), Precision, NDIM); // setup phat body struct
define_body_offset(PosTag, BodyOffset(Pos));
define_body_offset(VelTag, BodyOffset(Vel));
define_body_offset(MassTag, BodyOffset(Mass));
define_body_offset(PhiTag, BodyOffset(Phi));
define_body_offset(AccTag, BodyOffset(Acc));
infile = getparam("in"); // set I/O file names
outfile = getparam("out");
savefile = getparam("save");
if (strnull(getparam("restore"))) // starting a new run?
newrun();
else // else resume old run
oldrun();
if (ABS(nstatic) > nbody) // check nstatic is OK
error("%s: absurd value for nstatic\n", getargv0());
p1 = bodytab + MAX(nstatic, 0); // set dynamic body range
p2 = bodytab + nbody + MIN(nstatic, 0);
testcalc = TRUE; // determine type of calc:
for (p = p1; p < p2; p++)
testcalc = testcalc && (Mass(p) == 0); // look for dynamic masses
strfile = getparam("stream");
logfile = getparam("log");
#if defined(EXTGRAV)
if (! strnull(getparam("gravgsp"))) { // was GSP file given?
gravstr = stropen(getparam("gravgsp"), "r");
get_history(gravstr);
gravgsp = get_gsprof(gravstr); // read external field GSP
strclose(gravstr);
}
#endif
startrun_time_1 = wtime();
}
int main(int argc, char **argv)
{
int n;
int repeat;
double dot;
long start_time, end_time;
if ((argc != 3)) {
printf("Uso: %s <tamanho dos vetores> <repeticoes>\n", argv[0]);
exit(EXIT_FAILURE);
}
n = atoi(argv[1]); // tamanho dos vetores
repeat = atoi(argv[2]); // numero de repeticoes (variar carga)
// Cria vetores
double *a = (double *) malloc(sizeof(double) * n);
double *b = (double *) malloc(sizeof(double) * n);
if (a == NULL || b == NULL) {
printf("Erro de alocacao de memoria\n");
exit(EXIT_FAILURE);
}
init_vectors(a, b, n);
start_time = wtime();
dot = dot_product(a, b, n, repeat);
end_time = wtime();
printf("Produto escalar = %f\n", dot);
printf("Tempo de calculo = %ld usec\n", (long) (end_time - start_time));
free((void *) a);
free((void *) b);
return EXIT_SUCCESS;
}
int main(int argc, char **argv)
{
pthread_t thread1;
int ret = -1;
int i = 0;
double time1, time2;
ret = pthread_create(&thread1, NULL, thread1_fn, NULL);
assert(ret == 0);
time1 = wtime();
for(i=0; i<ITERATIONS; i++)
{
wakeywakey();
}
time2 = wtime();
printf("time for %d iterations: %f seconds.\n", ITERATIONS, (time2-time1));
printf("per iteration: %f\n", (time2-time1)/(double)ITERATIONS);
return(0);
}
void init_synchronization(void)
{
current_synchronization = form_of_synchronization;
max_counter = First_max_counter;
interval = First_interval;
first_measurement_run = True;
logging(DBG_SYNC, "starting with max_counter = %d interval = %9.1f\n",
max_counter, interval*1.0e6);
if( current_synchronization == SYNC_REAL) {
if( ! mpi_wtime_is_global ) determine_time_differences();
if( lrootproc() ) start_batch = wtime();
logging(DBG_SYNC, "---- new start_batch ----------------\n");
MPI_Bcast(&start_batch, 1, MPI_DOUBLE, 0, get_measurement_comm());
}
}
depth_t bc::bfs_sssp(
index_t root
)
{
sa[root] = 0;
sp_count[root] = 1;
depth_t level = 0;
dist[root]=0;
while(true)
{
double ltm= wtime();
index_t front_count = 0;
for(vertex_t vert_id = 0; vert_id<g->vert_count; vert_id++)
{
if(sa[vert_id] == level)
{
index_t my_beg = g->beg_pos[vert_id];
index_t my_end = g->beg_pos[vert_id + 1];
for(; my_beg<my_end; my_beg++)
{
vertex_t nebr=g->csr[my_beg];
path_t weit=g->weight[my_beg];
if(dist[nebr]>dist[vert_id]+weit)
{
dist[nebr]=dist[vert_id]+weit;
sp_count[nebr]=0; //prior parent is wrong
sa[nebr]=level+1;
front_count++;
}
if(dist[nebr]==dist[vert_id]+weit)
sp_count[nebr]+=sp_count[vert_id];
}
}
}
// std::cout<<"Level "<<(int) level<<": "<<front_count<<" "
// <<wtime() - ltm<<"\n";
if(front_count == 0) break;
level ++;
}
return level+1;
}
double stop_synchronization(void)
{
stop_batch = stop_sync = wtime();
if( current_synchronization == SYNC_REAL ) {
if( stop_sync - start_sync > interval )
invalid[counter] = INVALID_TOOK_TOO_LONG;
logging(DBG_SYNC, "stop_sync = %9.1f ", normalize_time(stop_sync));
switch( invalid[counter] ) {
case INVALID_TOOK_TOO_LONG: logging(DBG_SYNC, "invalid_too_long\n"); break;
case INVALID_STARTED_LATE: logging(DBG_SYNC, "invalid_started_late\n"); break;
default: logging(DBG_SYNC, "\n");
}
}
return stop_sync;
}
int main(){
double t;
int i, me, target;
unsigned int size;
me = xmp_node_num();
target = 3 - me;
init_buf(local_buf, me);
init_buf(target_buf, me);
if(me==1) print_items();
for(size=4;size<MAX_SIZE+1;size*=2){ // size must be more than 4 when using Fujitsu RDMA
xmp_sync_all(NULL);
for(i=0;i<LOOP+WARMUP;i++){
if(WARMUP == i)
t = wtime();
if(me == 1){
local_buf[0:size] = target_buf[0:size]:[target];
xmp_sync_memory(NULL);
#ifdef DEBUG
if(local_buf[0] != '2' && local_buf[size-1] != '2') fprintf(stderr, "Error !\n");
local_buf[0] = '1'; local_buf[size-1] = '1';
#endif
xmp_sync_all(NULL);
}
else{
xmp_sync_all(NULL);
local_buf[0:size] = target_buf[0:size]:[target];
#ifdef DEBUG
if(local_buf[0] != '1' && local_buf[size-1] != '1') fprintf(stderr, "Error !\n");
local_buf[0] = '2'; local_buf[size-1] = '2';
#endif
}
xmp_sync_all(NULL);
}
int main()
{
Init();
double start = wtime();
double start_linked_list = wtime();
RunThoughLinkedList();
double end_linked_list = wtime();
double start_explicit = wtime();
RunExplicit();
double end_explicit = wtime();
double end = wtime();
printf("Time through Linked List %7.2f\n"
"Time through explicit %7.2f\n"
"Total Time taken %7.2f\n",
end_linked_list-start_linked_list,
end_explicit-start_explicit,
end-start
);
}
int main(int argc, char **argv)
{
long int j, iter; /* dummies */
double scalar; /* constant used in Triad operation */
int iterations; /* number of times vector loop gets repeated */
long int length, /* vector length per processor */
total_length, /* total vector length */
offset; /* offset between vectors a and b, and b and c */
double bytes; /* memory IO size */
size_t space; /* memory used for a single vector */
double nstream_time, /* timing parameters */
avgtime = 0.0,
maxtime = 0.0,
mintime = 366.0*8760.0*3600.0; /* set the minimum time to a
large value; one leap year should be enough */
int Num_procs, /* process parameters */
my_ID, /* rank of calling process */
root=0; /* ID of master process */
int error=0; /* error flag for individual process */
/**********************************************************************************
* process and test input parameters
***********************************************************************************/
MPI_Init(&argc,&argv);
MPI_Comm_size(MPI_COMM_WORLD,&Num_procs);
MPI_Comm_rank(MPI_COMM_WORLD,&my_ID);
if (my_ID == root) {
printf("MPI stream triad: A = B + scalar*C\n");
if (argc != 4) {
printf("Usage: %s <# iterations> <vector length> <offset>\n", *argv);
error = 1;
goto ENDOFTESTS;
}
iterations = atoi(*++argv);
if (iterations < 1) {
printf("ERROR: Invalid number of iterations: %d\n", iterations);
error = 1;
goto ENDOFTESTS;
}
total_length = atol(*++argv);
if (total_length < Num_procs) {
printf("ERROR: Invalid vector length: %ld\n", total_length);
error = 1;
goto ENDOFTESTS;
}
else length = total_length/Num_procs;
offset = atol(*++argv);
if (offset < 0) {
printf("ERROR: Invalid array offset: %ld\n", offset);
error = 1;
goto ENDOFTESTS;
}
#ifdef STATIC_ALLOCATION
if ((3*length + 2*offset) > N) {
printf("ERROR: vector length/offset %ld/%ld too ", total_length, offset);
printf("large; increase MAXLENGTH in Makefile or decrease vector length\n");
error = 1;
goto ENDOFTESTS;
}
#endif
ENDOFTESTS:
;
}
bail_out(error);
/* broadcast initialization data */
MPI_Bcast(&length,1, MPI_LONG, root, MPI_COMM_WORLD);
MPI_Bcast(&offset,1, MPI_LONG, root, MPI_COMM_WORLD);
MPI_Bcast(&iterations,1, MPI_INT, root, MPI_COMM_WORLD);
#ifndef STATIC_ALLOCATION
space = (3*length + 2*offset)*sizeof(double);
a = (double *) malloc(space);
if (!a && my_ID == root) {
printf("ERROR: Could not allocate %ld bytes for vectors\n", (long int)space);
error = 1;
}
bail_out(error);
#endif
b = a + length + offset;
c = b + length + offset;
bytes = 3.0 * sizeof(double) * length * Num_procs;
if (my_ID == root) {
printf("Number of processes = %d\n", Num_procs);
printf("Vector length = %ld\n", total_length);
printf("Offset = %ld\n", offset);
printf("Number of iterations = %d\n", iterations);
}
#pragma vector always
for (j=0; j<length; j++) {
a[j] = 0.0;
b[j] = 2.0;
c[j] = 2.0;
}
/* --- MAIN LOOP --- repeat Triad iterations times --- */
scalar = SCALAR;
for (iter=0; iter<iterations; iter++) {
MPI_Barrier(MPI_COMM_WORLD);
if (my_ID == root) {
nstream_time = wtime();
}
#pragma vector always
for (j=0; j<length; j++) a[j] = b[j]+scalar*c[j];
if (my_ID == root) {
if (iter>0 || iterations==1) { /* skip the first iteration */
nstream_time = wtime() - nstream_time;
avgtime = avgtime + nstream_time;
mintime = MIN(mintime, nstream_time);
maxtime = MAX(maxtime, nstream_time);
}
}
/* insert a dependency between iterations to avoid dead-code elimination */
#pragma vector always
for (j=0; j<length; j++) b[j] = a[j];
}
/*********************************************************************
** Analyze and output results.
*********************************************************************/
if (my_ID == root) {
if (checkTRIADresults(iterations, length)) {
avgtime = avgtime/(double)(MAX(iterations-1,1));
printf("Rate (MB/s): %lf, Avg time (s): %lf, Min time (s): %lf",
1.0E-06 * bytes/mintime, avgtime, mintime);
printf(", Max time (s): %lf\n", maxtime);
}
else error = 1;
}
bail_out(error);
MPI_Finalize();
}
int main( int argc, char *argv[] )
{
unsigned iter;
FILE *infile, *resfile;
char *resfilename;
// algorithmic parameters
algoparam_t param;
int np;
double runtime, flop;
double residual=0.0;
// check arguments
if( argc < 2 )
{
usage( argv[0] );
return 1;
}
// check input file
if( !(infile=fopen(argv[1], "r")) )
{
fprintf(stderr,
"\nError: Cannot open \"%s\" for reading.\n\n", argv[1]);
usage(argv[0]);
return 1;
}
// check result file
resfilename= (argc>=3) ? argv[2]:"heat.ppm";
if( !(resfile=fopen(resfilename, "w")) )
{
fprintf(stderr,
"\nError: Cannot open \"%s\" for writing.\n\n",
resfilename);
usage(argv[0]);
return 1;
}
// check input
if( !read_input(infile, ¶m) )
{
fprintf(stderr, "\nError: Error parsing input file.\n\n");
usage(argv[0]);
return 1;
}
print_params(¶m);
if( !initialize(¶m) )
{
fprintf(stderr, "Error in Solver initialization.\n\n");
usage(argv[0]);
return 1;
}
// full size (param.resolution are only the inner points)
np = param.resolution + 2;
#if _EXTRAE_
Extrae_init();
#endif
// starting time
runtime = wtime();
iter = 0;
while(1) {
switch( param.algorithm ) {
case 0: // JACOBI
residual = relax_jacobi(param.u, param.uhelp, np, np);
// Copy uhelp into u
copy_mat(param.uhelp, param.u, np, np);
break;
case 1: // GAUSS
residual = relax_gauss(param.u, np, np);
break;
}
iter++;
// solution good enough ?
if (residual < 0.00005) break;
// max. iteration reached ? (no limit with maxiter=0)
if (param.maxiter>0 && iter>=param.maxiter) break;
}
// Flop count after iter iterations
flop = iter * 11.0 * param.resolution * param.resolution;
// stopping time
runtime = wtime() - runtime;
#if _EXTRAE_
Extrae_fini();
#endif
fprintf(stdout, "Time: %04.3f \n", runtime);
fprintf(stdout, "Flops and Flops per second: (%3.3f GFlop => %6.2f MFlop/s)\n",
flop/1000000000.0,
flop/runtime/1000000);
fprintf(stdout, "Convergence to residual=%f: %d iterations\n", residual, iter);
// for plot...
coarsen( param.u, np, np,
param.uvis, param.visres+2, param.visres+2 );
write_image( resfile, param.uvis,
param.visres+2,
param.visres+2 );
finalize( ¶m );
return 0;
}
int main(int argc, char* argv[])
{
double t1, t2, t3, t4, t5;
double sum1, sum2, sum3, sum4;
int arg = 1, len = 0, iters = 0, verb = 0, run = 1;
int do_vcopy = 1, do_vadd = 1, do_vjacobi = 1;
while(argc>arg) {
if (strcmp(argv[arg],"-v")==0) verb++;
else if (strcmp(argv[arg],"-vv")==0) verb+=2;
else if (strcmp(argv[arg],"-n")==0) run = 0;
else if (strcmp(argv[arg],"-c")==0) do_vadd = 0, do_vjacobi = 0;
else if (strcmp(argv[arg],"-a")==0) do_vcopy = 0, do_vjacobi = 0;
else if (strcmp(argv[arg],"-j")==0) do_vcopy = 0, do_vadd = 0;
else
break;
arg++;
}
if (argc>arg) { len = atoi(argv[arg]); arg++; }
if (argc>arg) { iters = atoi(argv[arg]); arg++; }
if (len == 0) len = 10000;
if (iters == 0) iters = 20;
len = len * 1000;
printf("Alloc/init 3 double arrays of length %d ...\n", len);
double* a = (double*) malloc(len * sizeof(double));
double* b = (double*) malloc(len * sizeof(double));
double* c = (double*) malloc(len * sizeof(double));
for(int i = 0; i<len; i++) {
a[i] = 1.0;
b[i] = (double) (i % 20);
c[i] = 3.0;
}
// Generate vectorized variants & run against naive/original
#if __AVX__
bool do32 = true;
#else
bool do32 = false;
#endif
// vcopy
if (do_vcopy) {
vcopy_t vcopy16, vcopy32;
Rewriter* rc16 = dbrew_new();
if (verb>1) dbrew_verbose(rc16, true, true, true);
dbrew_set_function(rc16, (uint64_t) vcopy);
dbrew_config_parcount(rc16, 3);
dbrew_config_force_unknown(rc16, 0);
dbrew_set_vectorsize(rc16, 16);
vcopy16 = (vcopy_t) dbrew_rewrite(rc16, a, b, len);
if (verb) decode_func(rc16, "vcopy16");
if (do32) {
Rewriter* rc32 = dbrew_new();
if (verb>1) dbrew_verbose(rc32, true, true, true);
dbrew_set_function(rc32, (uint64_t) vcopy);
dbrew_config_parcount(rc32, 3);
dbrew_config_force_unknown(rc32, 0);
dbrew_set_vectorsize(rc32, 32);
vcopy32 = (vcopy_t) dbrew_rewrite(rc32, a, b, len);
if (verb) decode_func(rc32, "vcopy32");
}
printf("Running %d iterations of vcopy ...\n", iters);
t1 = wtime();
for(int iter = 0; iter < iters; iter++)
naive_vcopy(a, b, len);
t2 = wtime();
for(int iter = 0; iter < iters; iter++)
vcopy(a, b, len);
t3 = wtime();
if (run)
for(int iter = 0; iter < iters; iter++)
vcopy16(a, b, len);
t4 = wtime();
if (do32 && run)
for(int iter = 0; iter < iters; iter++)
vcopy32(a, b, len);
t5 = wtime();
printf(" naive: %.3f s, un-rewritten: %.3f s, rewritten-16: %.3f s",
t2-t1, t3-t2, t4-t3);
if (do32)
printf(", rewritten-32: %.3f s", t5-t4);
printf("\n");
}
// vadd
if (do_vadd) {
vadd_t vadd16, vadd32;
Rewriter* ra16 = dbrew_new();
if (verb>1) dbrew_verbose(ra16, true, true, true);
dbrew_set_function(ra16, (uint64_t) vadd);
dbrew_config_parcount(ra16, 4);
dbrew_config_force_unknown(ra16, 0);
dbrew_set_vectorsize(ra16, 16);
vadd16 = (vadd_t) dbrew_rewrite(ra16, a, b, c, len);
if (verb) decode_func(ra16, "vadd16");
if (do32) {
Rewriter* ra32 = dbrew_new();
if (verb>1) dbrew_verbose(ra32, true, true, true);
dbrew_set_function(ra32, (uint64_t) vadd);
dbrew_config_parcount(ra32, 4);
dbrew_config_force_unknown(ra32, 0);
dbrew_set_vectorsize(ra32, 32);
vadd32 = (vadd_t) dbrew_rewrite(ra32, a, b, c, len);
if (verb) decode_func(ra32, "vadd32");
}
sum1 = 0.0, sum2 = 0.0, sum3 = 0.0, sum4 = 0.0;
printf("Running %d iterations of vadd ...\n", iters);
t1 = wtime();
for(int iter = 0; iter < iters; iter++)
naive_vadd(a, b, c, len);
for(int i = 0; i < len; i++) sum1 += a[i];
t2 = wtime();
for(int iter = 0; iter < iters; iter++)
vadd(a, b, c, len);
for(int i = 0; i < len; i++) sum2 += a[i];
t3 = wtime();
if (run)
for(int iter = 0; iter < iters; iter++)
vadd16(a, b, c, len);
for(int i = 0; i < len; i++) sum3 += a[i];
t4 = wtime();
if (do32 && run)
for(int iter = 0; iter < iters; iter++)
vadd32(a, b, c, len);
for(int i = 0; i < len; i++) sum4 += a[i];
t5 = wtime();
printf(" naive: %.3f s, un-rewritten: %.3f s, rewritten-16: %.3f s",
t2-t1, t3-t2, t4-t3);
if (do32)
printf(", rewritten-32: %.3f s", t5-t4);
printf("\n");
printf(" sum naive: %f, sum rewritten-16: %f, sum rewritten-16: %f\n",
sum1, sum3, sum4);
}
// vjacobi_1d
if (do_vjacobi) {
vcopy_t vjacobi_1d16, vjacobi_1d32;
Rewriter* rj16 = dbrew_new();
if (verb>1) dbrew_verbose(rj16, true, true, true);
dbrew_set_function(rj16, (uint64_t) vjacobi_1d);
dbrew_config_parcount(rj16, 3);
dbrew_config_force_unknown(rj16, 0);
dbrew_set_vectorsize(rj16, 16);
vjacobi_1d16 = (vcopy_t) dbrew_rewrite(rj16, a, b, len);
if (verb) decode_func(rj16, "vjacobi_1d16");
if (do32) {
Rewriter* rj32 = dbrew_new();
if (verb>1) dbrew_verbose(rj32, true, true, true);
dbrew_set_function(rj32, (uint64_t) vjacobi_1d);
dbrew_config_parcount(rj32, 3);
dbrew_config_force_unknown(rj32, 0);
dbrew_set_vectorsize(rj32, 32);
vjacobi_1d32 = (vcopy_t) dbrew_rewrite(rj32, a, b, len);
if (verb) decode_func(rj32, "vjacobi_1d32");
}
sum1 = 0.0, sum2 = 0.0, sum3 = 0.0, sum4 = 0.0;
printf("Running %d iterations of vjacobi_1d ...\n", iters);
t1 = wtime();
for(int iter = 0; iter < iters; iter++)
naive_vjacobi_1d(a+1, b+1, len-2);
for(int i = 0; i < len; i++) sum1 += a[i];
t2 = wtime();
for(int iter = 0; iter < iters; iter++)
vjacobi_1d(a+1, b+1, len-2);
for(int i = 0; i < len; i++) sum2 += a[i];
t3 = wtime();
if (run)
for(int iter = 0; iter < iters; iter++)
vjacobi_1d16(a+1, b+1, len-2);
for(int i = 0; i < len; i++) sum3 += a[i];
t4 = wtime();
if (do32 && run)
for(int iter = 0; iter < iters; iter++)
vjacobi_1d32(a+1, b+1, len-2);
for(int i = 0; i < len; i++) sum4 += a[i];
t5 = wtime();
printf(" naive: %.3f s, un-rewritten: %.3f s, rewritten-16: %.3f s",
t2-t1, t3-t2, t4-t3);
if (do32)
printf(", rewritten-32: %.3f s", t5-t4);
printf("\n");
printf(" sum naive: %f, sum rewritten-16: %f, sum rewritten-16: %f\n",
sum1, sum3, sum4);
}
}
int main(int argc, char ** argv)
{
int Block_order;
size_t Block_size;
size_t Colblock_size;
int Tile_order=32;
int tiling;
int Num_procs; /* Number of ranks */
int order; /* overall matrix order */
int send_to, recv_from; /* communicating ranks */
size_t bytes; /* total amount of data to be moved */
int my_ID; /* rank */
int root=0; /* root rank of a communicator */
int iterations; /* number of times to run the pipeline algorithm */
int i, j, it, jt, ID;/* dummies */
int iter; /* index of iteration */
int phase; /* phase in the staged communication */
size_t colstart; /* sequence number of first column owned by calling rank */
int error=0; /* error flag */
double *A_p; /* original matrix column block */
double *B_p; /* transposed matrix column block */
double *Work_in_p; /* workspace for the transpose function */
double *Work_out_p;/* workspace for the transpose function */
double abserr, abserr_tot; /* computed error */
double epsilon = 1.e-8; /* error tolerance */
double local_trans_time, /* timing parameters */
trans_time,
avgtime;
MPI_Status status; /* completion status of message */
MPI_Win shm_win_A; /* Shared Memory window object */
MPI_Win shm_win_B; /* Shared Memory window object */
MPI_Win shm_win_Work_in; /* Shared Memory window object */
MPI_Win shm_win_Work_out; /* Shared Memory window object */
MPI_Info rma_winfo;/* info for window */
MPI_Comm shm_comm_prep;/* Shared Memory prep Communicator */
MPI_Comm shm_comm; /* Shared Memory Communicator */
int shm_procs; /* # of ranks in shared domain */
int shm_ID; /* MPI rank within coherence domain */
int group_size; /* number of ranks per shared memory group */
int Num_groups; /* number of shared memory group */
int group_ID; /* sequence number of shared memory group */
int size_mul; /* size multiplier; 0 for non-root ranks in coherence domain*/
int istart;
MPI_Request send_req, recv_req;
/*********************************************************************************
** Initialize the MPI environment
**********************************************************************************/
MPI_Init(&argc,&argv);
MPI_Comm_rank(MPI_COMM_WORLD, &my_ID);
MPI_Comm_size(MPI_COMM_WORLD, &Num_procs);
root = 0;
/*********************************************************************
** process, test and broadcast input parameter
*********************************************************************/
if (my_ID == root){
if (argc != 4 && argc !=5){
printf("Usage: %s <#ranks per coherence domain> <# iterations> <matrix order> [tile size]\n",
*argv);
error = 1;
goto ENDOFTESTS;
}
group_size = atoi(*++argv);
if (group_size < 1) {
printf("ERROR: # ranks per coherence domain must be >= 1 : %d \n",group_size);
error = 1;
goto ENDOFTESTS;
}
if (Num_procs%group_size) {
printf("ERROR: toal # %d ranks not divisible by ranks per coherence domain %d\n",
Num_procs, group_size);
error = 1;
goto ENDOFTESTS;
}
iterations = atoi(*++argv);
if (iterations < 1){
printf("ERROR: iterations must be >= 1 : %d \n",iterations);
error = 1;
goto ENDOFTESTS;
}
order = atoi(*++argv);
if (order < Num_procs) {
printf("ERROR: matrix order %d should at least # procs %d\n",
order, Num_procs);
error = 1; goto ENDOFTESTS;
}
if (order%Num_procs) {
printf("ERROR: matrix order %d should be divisible by # procs %d\n",
order, Num_procs);
error = 1; goto ENDOFTESTS;
}
if (argc == 5) Tile_order = atoi(*++argv);
ENDOFTESTS:;
}
bail_out(error);
/* Broadcast input data to all ranks */
MPI_Bcast(&order, 1, MPI_INT, root, MPI_COMM_WORLD);
MPI_Bcast(&iterations, 1, MPI_INT, root, MPI_COMM_WORLD);
MPI_Bcast(&Tile_order, 1, MPI_INT, root, MPI_COMM_WORLD);
MPI_Bcast(&group_size, 1, MPI_INT, root, MPI_COMM_WORLD);
if (my_ID == root) {
printf("Parallel Research Kernels version %s\n", PRKVERSION);
printf("MPI+SHM Matrix transpose: B = A^T\n");
printf("Number of ranks = %d\n", Num_procs);
printf("Rank group size = %d\n", group_size);
printf("Matrix order = %d\n", order);
printf("Number of iterations = %d\n", iterations);
if ((Tile_order > 0) && (Tile_order < order))
printf("Tile size = %d\n", Tile_order);
else printf("Untiled\n");
#ifndef SYNCHRONOUS
printf("Non-");
#endif
printf("Blocking messages\n");
}
/* Setup for Shared memory regions */
/* first divide WORLD in groups of size group_size */
MPI_Comm_split(MPI_COMM_WORLD, my_ID/group_size, my_ID%group_size, &shm_comm_prep);
/* derive from that a SHM communicator */
MPI_Comm_split_type(shm_comm_prep, MPI_COMM_TYPE_SHARED, 0, MPI_INFO_NULL, &shm_comm);
MPI_Comm_rank(shm_comm, &shm_ID);
MPI_Comm_size(shm_comm, &shm_procs);
/* do sanity check, making sure groups did not shrink in second comm split */
if (shm_procs != group_size) MPI_Abort(MPI_COMM_WORLD, 666);
/* a non-positive tile size means no tiling of the local transpose */
tiling = (Tile_order > 0) && (Tile_order < order);
bytes = 2 * sizeof(double) * order * order;
/*********************************************************************
** The matrix is broken up into column blocks that are mapped one to a
** rank. Each column block is made up of Num_procs smaller square
** blocks of order block_order.
*********************************************************************/
Num_groups = Num_procs/group_size;
Block_order = order/Num_groups;
group_ID = my_ID/group_size;
colstart = Block_order * group_ID;
Colblock_size = order * Block_order;
Block_size = Block_order * Block_order;
/*********************************************************************
** Create the column block of the test matrix, the column block of the
** transposed matrix, and workspace (workspace only if #procs>1)
*********************************************************************/
/* RMA win info */
MPI_Info_create(&rma_winfo);
/* This key indicates that passive target RMA will not be used.
* It is the one info key that MPICH actually uses for optimization. */
MPI_Info_set(rma_winfo, "no_locks", "true");
/* only the root of each SHM domain specifies window of nonzero size */
size_mul = (shm_ID==0);
int offset = 32;
MPI_Aint size= (Colblock_size+offset)*sizeof(double)*size_mul; int disp_unit;
MPI_Win_allocate_shared(size, sizeof(double), rma_winfo, shm_comm,
(void *) &A_p, &shm_win_A);
MPI_Win_lock_all(MPI_MODE_NOCHECK,shm_win_A);
MPI_Win_shared_query(shm_win_A, MPI_PROC_NULL, &size, &disp_unit, (void *)&A_p);
if (A_p == NULL){
printf(" Error allocating space for original matrix on node %d\n",my_ID);
error = 1;
}
bail_out(error);
A_p += offset;
/* recompute memory size (overwritten by prior query */
size= (Colblock_size+offset)*sizeof(double)*size_mul;
MPI_Win_allocate_shared(size, sizeof(double), rma_winfo, shm_comm,
(void *) &B_p, &shm_win_B);
MPI_Win_lock_all(MPI_MODE_NOCHECK,shm_win_B);
MPI_Win_shared_query(shm_win_B, MPI_PROC_NULL, &size, &disp_unit, (void *)&B_p);
if (B_p == NULL){
printf(" Error allocating space for transposed matrix by group %d\n",group_ID);
error = 1;
}
bail_out(error);
B_p += offset;
if (Num_groups>1) {
size = Block_size*sizeof(double)*size_mul;
MPI_Win_allocate_shared(size, sizeof(double),rma_winfo, shm_comm,
(void *) &Work_in_p, &shm_win_Work_in);
MPI_Win_lock_all(MPI_MODE_NOCHECK,shm_win_Work_in);
MPI_Win_shared_query(shm_win_Work_in, MPI_PROC_NULL, &size, &disp_unit,
(void *)&Work_in_p);
if (Work_in_p == NULL){
printf(" Error allocating space for in block by group %d\n",group_ID);
error = 1;
}
bail_out(error);
/* recompute memory size (overwritten by prior query */
size = Block_size*sizeof(double)*size_mul;
MPI_Win_allocate_shared(size, sizeof(double), rma_winfo,
shm_comm, (void *) &Work_out_p, &shm_win_Work_out);
MPI_Win_lock_all(MPI_MODE_NOCHECK,shm_win_Work_out);
MPI_Win_shared_query(shm_win_Work_out, MPI_PROC_NULL, &size, &disp_unit,
(void *)&Work_out_p);
if (Work_out_p == NULL){
printf(" Error allocating space for out block by group %d\n",group_ID);
error = 1;
}
bail_out(error);
}
/* Fill the original column matrix */
istart = 0;
int chunk_size = Block_order/group_size;
if (tiling) {
for (j=shm_ID*chunk_size;j<(shm_ID+1)*chunk_size;j+=Tile_order) {
for (i=0;i<order; i+=Tile_order)
for (jt=j; jt<MIN((shm_ID+1)*chunk_size,j+Tile_order); jt++)
for (it=i; it<MIN(order,i+Tile_order); it++) {
A(it,jt) = (double) ((double)order*(jt+colstart) + it);
B(it,jt) = -1.0;
}
}
}
else {
for (j=shm_ID*chunk_size;j<(shm_ID+1)*chunk_size;j++)
for (i=0;i<order; i++) {
A(i,j) = (double)((double)order*(j+colstart) + i);
B(i,j) = -1.0;
}
}
/* NEED A STORE FENCE HERE */
MPI_Win_sync(shm_win_A);
MPI_Win_sync(shm_win_B);
MPI_Barrier(shm_comm);
for (iter=0; iter<=iterations; iter++) {
/* start timer after a warmup iteration */
if (iter == 1) {
MPI_Barrier(MPI_COMM_WORLD);
local_trans_time = wtime();
}
/* do the local transpose */
istart = colstart;
if (!tiling) {
for (i=shm_ID*chunk_size; i<(shm_ID+1)*chunk_size; i++) {
for (j=0; j<Block_order; j++)
B(j,i) = A(i,j);
}
}
else {
for (i=shm_ID*chunk_size; i<(shm_ID+1)*chunk_size; i+=Tile_order) {
for (j=0; j<Block_order; j+=Tile_order)
for (it=i; it<MIN(Block_order,i+Tile_order); it++)
for (jt=j; jt<MIN(Block_order,j+Tile_order);jt++) {
B(jt,it) = A(it,jt);
}
}
}
for (phase=1; phase<Num_groups; phase++){
recv_from = ((group_ID + phase )%Num_groups);
send_to = ((group_ID - phase + Num_groups)%Num_groups);
istart = send_to*Block_order;
if (!tiling) {
for (i=shm_ID*chunk_size; i<(shm_ID+1)*chunk_size; i++)
for (j=0; j<Block_order; j++){
Work_out(j,i) = A(i,j);
}
}
else {
for (i=shm_ID*chunk_size; i<(shm_ID+1)*chunk_size; i+=Tile_order)
for (j=0; j<Block_order; j+=Tile_order)
for (it=i; it<MIN(Block_order,i+Tile_order); it++)
for (jt=j; jt<MIN(Block_order,j+Tile_order);jt++) {
Work_out(jt,it) = A(it,jt);
}
}
/* NEED A LOAD/STORE FENCE HERE */
MPI_Win_sync(shm_win_Work_in);
MPI_Win_sync(shm_win_Work_out);
MPI_Barrier(shm_comm);
if (shm_ID==0) {
#ifndef SYNCHRONOUS
/* if we place the Irecv outside this block, it would not be
protected by a local barrier, which creates a race */
MPI_Irecv(Work_in_p, Block_size, MPI_DOUBLE,
recv_from*group_size, phase, MPI_COMM_WORLD, &recv_req);
MPI_Isend(Work_out_p, Block_size, MPI_DOUBLE, send_to*group_size,
phase, MPI_COMM_WORLD, &send_req);
MPI_Wait(&recv_req, &status);
MPI_Wait(&send_req, &status);
#else
MPI_Sendrecv(Work_out_p, Block_size, MPI_DOUBLE, send_to*group_size,
phase, Work_in_p, Block_size, MPI_DOUBLE,
recv_from*group_size, phase, MPI_COMM_WORLD, &status);
#endif
}
/* NEED A LOAD FENCE HERE */
MPI_Win_sync(shm_win_Work_in);
MPI_Win_sync(shm_win_Work_out);
MPI_Barrier(shm_comm);
istart = recv_from*Block_order;
/* scatter received block to transposed matrix; no need to tile */
for (j=shm_ID*chunk_size; j<(shm_ID+1)*chunk_size; j++)
for (i=0; i<Block_order; i++)
B(i,j) = Work_in(i,j);
} /* end of phase loop */
} /* end of iterations */
local_trans_time = wtime() - local_trans_time;
MPI_Reduce(&local_trans_time, &trans_time, 1, MPI_DOUBLE, MPI_MAX, root,
MPI_COMM_WORLD);
abserr = 0.0;
istart = 0;
/* for (j=shm_ID;j<Block_order;j+=group_size) for (i=0;i<order; i++) { */
for (j=shm_ID*chunk_size; j<(shm_ID+1)*chunk_size; j++)
for (i=0;i<order; i++) {
abserr += ABS(B(i,j) - (double)((double)order*i + j+colstart));
}
MPI_Reduce(&abserr, &abserr_tot, 1, MPI_DOUBLE, MPI_SUM, root, MPI_COMM_WORLD);
if (my_ID == root) {
if (abserr_tot < epsilon) {
printf("Solution validates\n");
avgtime = trans_time/(double)iterations;
printf("Rate (MB/s): %lf Avg time (s): %lf\n",1.0E-06*bytes/avgtime, avgtime);
#ifdef VERBOSE
printf("Summed errors: %f \n", abserr_tot);
#endif
}
else {
printf("ERROR: Aggregate squared error %e exceeds threshold %e\n", abserr_tot, epsilon);
error = 1;
}
}
bail_out(error);
MPI_Win_unlock_all(shm_win_A);
MPI_Win_unlock_all(shm_win_B);
MPI_Win_free(&shm_win_A);
MPI_Win_free(&shm_win_B);
if (Num_groups>1) {
MPI_Win_unlock_all(shm_win_Work_in);
MPI_Win_unlock_all(shm_win_Work_out);
MPI_Win_free(&shm_win_Work_in);
MPI_Win_free(&shm_win_Work_out);
}
MPI_Info_free(&rma_winfo);
MPI_Finalize();
exit(EXIT_SUCCESS);
} /* end of main */
/*---------------------------------------------------------------------------
*
* Compute matrix product using BLAS routine DGEMM.
*
* Input
* int argc - length of argv[] array
* char* argv[] - pointer to command line parameter array
* int verbosity - program verification: verbosity > 0 gives more output
*
* Output
* double - elapsed time for product computation
*/
double multiply_by_blas( int argc, char* argv[], int verbosity )
{
int rows, cols, mids;
double **a, **b, **c;
double t1, t2;
double sec;
double gflop_count;
/*
* process command line arguments
*/
rows = atoi( argv[0] );
mids = atoi( argv[1] );
cols = atoi( argv[2] );
gflop_count = 2.0 * rows * mids * cols / 1.0e9;
if ( verbosity > 0 )
{
printf( "BLAS: rows = %d, mids = %d, columns = %d\n",
rows, mids, cols );
}
/*
* allocate and initialize matrices
*/
a = (double**) allocateMatrix( rows, mids );
b = (double**) allocateMatrix( mids, cols );
c = (double**) allocateMatrix( rows, cols );
initialize_matrices( a, b, c, rows, cols, mids, verbosity );
/*
* compute product: There is an implicit matrix transpose when
* passing from Fortran to C and vice-versa. To compute C :=
* alpha * A * B + beta * C we use dgemm() to compute C' := alpha
* * B' * A' + beta * C'. The first two arguments to dgemm() are
* 'N' indicating we don't want a transpose in addition to the
* implicit one. The matrices A and B are passed in reverse order
* so dgemm() receives (after the implicit transpose) B' and A'.
* Arguments 3 and 4 are the dimensions of C' and argument 5 is
* the column dimension of B' (and the row dimension of A').
*/
t1 = wtime();
dgemm( 'N', 'N', cols, rows, mids, 1.0, &b[0][0], cols, &a[0][0], mids,
0.0, &c[0][0], cols );
t2 = wtime();
sec = t2 - t1;
if ( verbosity > 1 )
printf( "checksum = %f\n", checksum( c, rows, cols ) );
printf( "BLAS: %6.3f secs %6.3f gflops ( %5d x %5d x %5d )\n",
sec, gflop_count / sec, rows, mids, cols );
/*
* clean up
*/
deallocateMatrix( a );
deallocateMatrix( b );
deallocateMatrix( c );
return t2 - t1;
}
int main(int argc, char ** argv)
{
int vector_length; /* length of vectors to be aggregated */
int total_length; /* bytes needed to store reduction vectors */
double reduce_time, /* timing parameters */
avgtime = 0.0,
maxtime = 0.0,
mintime = 366.0*24.0*3600.0; /* set the minimum time to a large
value; one leap year should be enough */
double epsilon=1.e-8; /* error tolerance */
int i, iter; /* dummies */
double element_value; /* reference element value for final vector */
int iterations; /* number of times the reduction is carried out */
static double /* use static so it goes on the heap, not stack */
RESTRICT vector[MEMWORDS];/* we would like to allocate "vector" dynamically,
but need to be able to flush the thing in some
versions of the reduction algorithm -> static */
/*****************************************************************************
** process and test input parameters
******************************************************************************/
if (argc != 3){
printf("Usage: %s <# iterations> <vector length>\n", *argv);
return(EXIT_FAILURE);
}
iterations = atoi(*++argv);
if (iterations < 1){
printf("ERROR: Iterations must be positive : %d \n", iterations);
exit(EXIT_FAILURE);
}
vector_length = atoi(*++argv);
if (vector_length < 1){
printf("ERROR: vector length must be >= 1 : %d \n",vector_length);
exit(EXIT_FAILURE);
}
/* make sure we stay within the memory allocated for vector */
total_length = 2*vector_length;
if (total_length/2 != vector_length || total_length > MEMWORDS) {
printf("Vector length of %d too large; ", vector_length);
printf("increase MEMWORDS in Makefile or reduce vector length\n");
exit(EXIT_FAILURE);
}
printf("Serial Vector Reduction\n");
printf("Vector length = %d\n", vector_length);
printf("Number of iterations = %d\n", iterations);
for (iter=0; iter<iterations; iter++) {
/* initialize the arrays, assuming first-touch memory placement */
for (i=0; i<vector_length; i++) {
VEC0(i) = (double)(1);
VEC1(i) = (double)(2);
}
reduce_time = wtime();
/* do actual reduction */
/* first do the "local" part, which is the same for all algorithms */
for (i=0; i<vector_length; i++) {
VEC0(i) += VEC1(i);
}
reduce_time = wtime() - reduce_time;
#ifdef VERBOSE
printf("\nFinished with reduction, using %lf seconds \n", reduce_time);
#endif
if (iter>0 || iterations==1) { /* skip the first iteration */
avgtime = avgtime + reduce_time;
mintime = MIN(mintime, reduce_time);
maxtime = MAX(maxtime, reduce_time);
}
} /* end of iter loop */
/* verify correctness */
element_value = (2.0+1.0);
for (i=0; i<vector_length; i++) {
if (ABS(VEC0(i) - element_value) >= epsilon) {
printf("First error at i=%d; value: %lf; reference value: %lf\n",
i, VEC0(i), element_value);
exit(EXIT_FAILURE);
}
}
printf("Solution validates\n");
#ifdef VERBOSE
printf("Element verification value: %lf\n", element_value);
#endif
avgtime = avgtime/(double)(MAX(iterations-1,1));
printf("Rate (MFlops/s): %lf, Avg time (s): %lf, Min time (s): %lf",
1.0E-06 * (2.0-1.0)*vector_length/mintime, avgtime, mintime);
printf(", Max time (s): %lf\n", maxtime);
exit(EXIT_SUCCESS);
}
if(me == 1){
local_buf[0:size] = target_buf[0:size]:[target];
xmp_sync_memory(NULL);
#ifdef DEBUG
if(local_buf[0] != '2' && local_buf[size-1] != '2') fprintf(stderr, "Error !\n");
local_buf[0] = '1'; local_buf[size-1] = '1';
#endif
xmp_sync_all(NULL);
}
else{
xmp_sync_all(NULL);
local_buf[0:size] = target_buf[0:size]:[target];
#ifdef DEBUG
if(local_buf[0] != '1' && local_buf[size-1] != '1') fprintf(stderr, "Error !\n");
local_buf[0] = '2'; local_buf[size-1] = '2';
#endif
}
xmp_sync_all(NULL);
}
xmp_sync_all(NULL);
t = wtime() - t;
if(me == 1)
print_results(size, t);
}
return 0;
}
static double
wtime()
{
static struct timeval tv0 = {.tv_sec = 0};
struct timeval tv;
int cc;
cc = gettimeofday(&tv, 0);
assert(cc == 0);
if (tv0.tv_sec == 0) {
tv0 = tv;
assert(tv0.tv_sec != 0);
}
double dt = ((double)(tv.tv_sec - tv0.tv_sec)
+ ((double)(tv.tv_usec - tv0.tv_usec) * 1e-6));
return dt;
}
/* Puts 200 key-value pairs to output KVO. It is a map-function. It
runs only on rank0. Inputs (KV0 and KVS0) are dummy. */
static int
addkeysfn(const struct kmr_kv_box kv0,
const KMR_KVS *kvs0, KMR_KVS *kvo, void *p, const long ind)
{
assert(kvs0 == 0 && kv0.klen == 0 && kv0.vlen == 0 && kvo != 0);
char k[80];
char v[80];
int cc;
for (int i = 0; i < 200; i++) {
snprintf(k, 80, "key%d", i);
snprintf(v, 80, "value%d", i);
struct kmr_kv_box kv = {
.klen = (int)(strlen(k) + 1),
.vlen = (int)(strlen(v) + 1),
.k.p = k,
.v.p = v
};
cc = kmr_add_kv(kvo, kv);
assert(cc == MPI_SUCCESS);
}
return MPI_SUCCESS;
}
static int
replacevaluefn(const struct kmr_kv_box kv0,
const KMR_KVS *kvs0, KMR_KVS *kvo, void *p,
const long i)
{
assert(kvs0 != 0 && kvo != 0);
int cc, x;
char gomi;
cc = sscanf((&((char *)kv0.k.p)[3]), "%d%c", &x, &gomi);
assert(cc == 1);
char v[80];
snprintf(v, 10, "newvalue%d", x);
struct kmr_kv_box kv = {.klen = kv0.klen,
.vlen = (int)(strlen(v) + 1),
.k.p = kv0.k.p,
.v.p = v
};
cc = kmr_add_kv(kvo, kv);
assert(cc == MPI_SUCCESS);
return MPI_SUCCESS;
}
static int
emptyreducefn(const struct kmr_kv_box kv[], const long n,
const KMR_KVS *kvs, KMR_KVS *kvo, void *p)
{
return MPI_SUCCESS;
}
/* Do KMR operations many times. */
static void
simple0(int nprocs, int rank)
{
int cc;
KMR *mr = kmr_create_context(MPI_COMM_WORLD, MPI_INFO_NULL, 0);
double t0, t1;
t0 = wtime();
for (int i = 0; i < 10000; i++) {
/* Check timeout. */
t1 = wtime();
KMR_KVS *to0 = kmr_create_kvs(mr, KMR_KV_INTEGER, KMR_KV_INTEGER);
if (rank == 0) {
struct kmr_kv_box kv = {
.klen = (int)sizeof(long),
.vlen = (int)sizeof(long),
.k.i = 0,
.v.i = ((t1 - t0) > 20.0)
};
cc = kmr_add_kv(to0, kv);
assert(cc == MPI_SUCCESS);
}
cc = kmr_add_kv_done(to0);
assert(cc == MPI_SUCCESS);
KMR_KVS *to1 = kmr_create_kvs(mr, KMR_KV_INTEGER, KMR_KV_INTEGER);
cc = kmr_replicate(to0, to1, kmr_noopt);
assert(cc == MPI_SUCCESS);
struct kmr_kv_box tok = {.klen = (int)sizeof(long), .k.p = 0,
.vlen = 0, .v.p = 0};
struct kmr_kv_box tov;
cc = kmr_find_key(to1, tok, &tov);
assert(cc == MPI_SUCCESS);
cc = kmr_free_kvs(to1);
assert(cc == MPI_SUCCESS);
if (tov.v.i) {
if (rank == 0) {
printf("loops %d\n", i);
}
break;
}
/* Put some pairs. */
KMR_KVS *kvs0 = kmr_create_kvs(mr, KMR_KV_OPAQUE, KMR_KV_OPAQUE);
cc = kmr_map_on_rank_zero(kvs0, 0, kmr_noopt, addkeysfn);
assert(cc == MPI_SUCCESS);
/* Replicate pairs to all ranks. */
KMR_KVS *kvs1 = kmr_create_kvs(mr, KMR_KV_OPAQUE, KMR_KV_OPAQUE);
cc = kmr_replicate(kvs0, kvs1, kmr_noopt);
assert(cc == MPI_SUCCESS);
/* Map pairs. */
KMR_KVS *kvs2 = kmr_create_kvs(mr, KMR_KV_OPAQUE, KMR_KV_OPAQUE);
cc = kmr_map(kvs1, kvs2, 0, kmr_noopt, replacevaluefn);
assert(cc == MPI_SUCCESS);
/* Collect pairs by theirs keys. */
KMR_KVS *kvs3 = kmr_create_kvs(mr, KMR_KV_OPAQUE, KMR_KV_OPAQUE);
cc = kmr_shuffle(kvs2, kvs3, kmr_noopt);
assert(cc == MPI_SUCCESS);
/* Reduce collected pairs. */
KMR_KVS *kvs4 = kmr_create_kvs(mr, KMR_KV_OPAQUE, KMR_KV_OPAQUE);
cc = kmr_reduce(kvs3, kvs4, 0, kmr_noopt, emptyreducefn);
assert(cc == MPI_SUCCESS);
cc = kmr_free_kvs(kvs4);
assert(cc == MPI_SUCCESS);
}
cc = kmr_free_context(mr);
assert(cc == MPI_SUCCESS);
}
int
main(int argc, char *argv[])
{
char cmd[256];
int pid = getpid();
int N = 8;
int nprocs, rank, thlv;
/*MPI_Init(&argc, &argv);*/
MPI_Init_thread(&argc, &argv, MPI_THREAD_SERIALIZED, &thlv);
MPI_Comm_size(MPI_COMM_WORLD, &nprocs);
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
kmr_init();
MPI_Barrier(MPI_COMM_WORLD);
if (rank == 0) {printf("Check leakage by observing heap size.\n");}
if (rank == 0) {printf("Watch VSZ changes (loops %d times)...\n", N);}
if (rank == 0) {printf("(Each loop will take approx. 20 sec).\n");}
fflush(0);
usleep(50 * 1000);
MPI_Barrier(MPI_COMM_WORLD);
for (int i = 0; i < N; i++) {
simple0(nprocs, rank);
MPI_Barrier(MPI_COMM_WORLD);
if (rank == 0) {
snprintf(cmd, sizeof(cmd), "ps l %d", pid);
system(cmd);
}
fflush(0);
}
MPI_Barrier(MPI_COMM_WORLD);
if (rank == 0) printf("OK\n");
fflush(0);
kmr_fin();
MPI_Finalize();
return 0;
}
int main(int argc, char ** argv)
{
int my_ID; /* Thread ID */
long vector_length; /* length of vectors to be aggregated */
long total_length; /* bytes needed to store reduction vectors */
double reduce_time, /* timing parameters */
avgtime;
double epsilon=1.e-8; /* error tolerance */
int group_size, /* size of aggregating half of thread pool */
old_size, /* group size in previous binary tree iteration */
i, id, iter, stage; /* dummies */
double element_value; /* reference element value for final vector */
char *algorithm; /* reduction algorithm selector */
int intalgorithm; /* integer encoding of algorithm selector */
int iterations; /* number of times the reduction is carried out */
int flag[MAX_THREADS*LINEWORDS]; /* used for pairwise synchronizations */
int start[MAX_THREADS],
end[MAX_THREADS];/* segments of vectors for bucket algorithm */
long segment_size;
int my_donor, my_segment;
int nthread_input, /* thread parameters */
nthread;
double RESTRICT *vector;/* vector pair to be reduced */
int num_error=0; /* flag that signals that requested and obtained
numbers of threads are the same */
/*****************************************************************************
** process and test input parameters
******************************************************************************/
printf("Parallel Research Kernels version %s\n", PRKVERSION);
printf("OpenMP Vector Reduction\n");
if (argc != 4 && argc != 5){
printf("Usage: %s <# threads> <# iterations> <vector length> ", *argv);
printf("[<alghorithm>]\n");
printf("Algorithm: linear, binary-barrier, binary-p2p, or long-optimal\n");
return(EXIT_FAILURE);
}
/* Take number of threads to request from command line */
nthread_input = atoi(*++argv);
if ((nthread_input < 1) || (nthread_input > MAX_THREADS)) {
printf("ERROR: Invalid number of threads: %d\n", nthread_input);
exit(EXIT_FAILURE);
}
omp_set_num_threads(nthread_input);
iterations = atoi(*++argv);
if (iterations < 1){
printf("ERROR: Iterations must be positive : %d \n", iterations);
exit(EXIT_FAILURE);
}
vector_length = atol(*++argv);
if (vector_length < 1){
printf("ERROR: vector length must be >= 1 : %ld \n",vector_length);
exit(EXIT_FAILURE);
}
total_length = vector_length*2*nthread_input*sizeof(double);
vector = (double *) prk_malloc(total_length);
if (!vector) {
printf("ERROR: Could not allocate space for vectors: %ld\n", total_length);
exit(EXIT_FAILURE);
}
algorithm = "binary-p2p";
if (argc == 5) algorithm = *++argv;
intalgorithm = NONE;
if (!strcmp(algorithm,"linear" )) intalgorithm = LINEAR;
if (!strcmp(algorithm,"binary-barrier")) intalgorithm = BINARY_BARRIER;
if (!strcmp(algorithm,"binary-p2p" )) intalgorithm = BINARY_P2P;
if (!strcmp(algorithm,"long-optimal" )) intalgorithm = LONG_OPTIMAL;
if (intalgorithm == NONE) {
printf("Wrong algorithm: %s; choose linear, binary-barrier, ", algorithm);
printf("binary-p2p, or long-optimal\n");
exit(EXIT_FAILURE);
}
else {
if (nthread_input == 1) intalgorithm = LOCAL;
}
#pragma omp parallel private(i, old_size, group_size, my_ID, iter, start, end, \
segment_size, stage, id, my_donor, my_segment)
{
my_ID = omp_get_thread_num();
#pragma omp master
{
nthread = omp_get_num_threads();
if (nthread != nthread_input) {
num_error = 1;
printf("ERROR: number of requested threads %d does not equal ",
nthread_input);
printf("number of spawned threads %d\n", nthread);
}
else {
printf("Number of threads = %d\n",nthread_input);
printf("Vector length = %ld\n", vector_length);
printf("Reduction algorithm = %s\n", algorithm);
printf("Number of iterations = %d\n", iterations);
}
}
bail_out(num_error);
for (iter=0; iter<=iterations; iter++) {
/* start timer after a warmup iteration */
if (iter == 1) {
#pragma omp barrier
#pragma omp master
{
reduce_time = wtime();
}
}
/* in case of the long-optimal algorithm we need a barrier before the
reinitialization to make sure that we don't overwrite parts of the
vector before other threads are done with those parts */
if (intalgorithm == LONG_OPTIMAL) {
#pragma omp barrier
}
/* initialize the arrays, assuming first-touch memory placement */
for (i=0; i<vector_length; i++) {
VEC0(my_ID,i) = (double)(my_ID+1);
VEC1(my_ID,i) = (double)(my_ID+1+nthread);
}
if (intalgorithm == BINARY_P2P) {
/* we need a barrier before setting all flags to zero, to avoid
zeroing some that are still in use in a previous iteration */
#pragma omp barrier
flag(my_ID) = 0;
/* we also need a barrier after setting the flags, to make each is
visible to all threads, and to synchronize before the timer starts */
#pragma omp barrier
}
/* do actual reduction */
/* first do the "local" part, which is the same for all algorithms */
for (i=0; i<vector_length; i++) {
VEC0(my_ID,i) += VEC1(my_ID,i);
}
/* now do the "non-local" part */
switch (intalgorithm) {
case LOCAL:
break;
case LINEAR:
{
#pragma omp barrier
#pragma omp master
{
for (id=1; id<nthread; id++) {
for (i=0; i<vector_length; i++) {
VEC0(0,i) += VEC0(id,i);
}
}
}
}
break;
case BINARY_BARRIER:
group_size = nthread;
while (group_size >1) {
/* barrier to make sure threads have completed their updates before
the results are being read */
#pragma omp barrier
old_size = group_size;
group_size = (group_size+1)/2;
/* Threads in "first half" of group aggregate data from threads in
second half; must make sure the counterpart is within old group.
If group size is odd, the last thread in the group does not have
a counterpart. */
if (my_ID < group_size && my_ID+group_size<old_size) {
for (i=0; i<vector_length; i++) {
VEC0(my_ID,i) += VEC0(my_ID+group_size,i);
}
}
}
break;
case BINARY_P2P:
group_size = nthread;
while (group_size >1) {
old_size = group_size;
group_size = (group_size+1)/2;
/* synchronize between each pair of threads that collaborate to
aggregate a new subresult, to make sure the donor of the pair has
updated its vector in the previous round before it is being read */
if (my_ID < group_size && my_ID+group_size<old_size) {
while (flag(my_ID+group_size) == 0) {
#pragma omp flush
}
/* make sure I read the latest version of vector from memory */
#pragma omp flush
for (i=0; i<vector_length; i++) {
VEC0(my_ID,i) += VEC0(my_ID+group_size,i);
}
}
else {
if (my_ID < old_size) {
/* I am a producer of data in this iteration; make sure my
updated version can be seen by all threads */
flag(my_ID) = 1;
#pragma omp flush
}
}
}
break;
case LONG_OPTIMAL:
/* compute starts and ends of subvectors to be passed among threads */
segment_size = (vector_length+nthread-1)/nthread;
for (id=0; id<nthread; id++) {
start[id] = segment_size*id;
end[id] = MIN(vector_length,segment_size*(id+1));
}
/* first do the Bucket Reduce Scatter in nthread-1 stages */
my_donor = (my_ID-1+nthread)%nthread;
for (stage=1; stage<nthread; stage++) {
#pragma omp barrier
my_segment = (my_ID-stage+nthread)%nthread;
for (i=start[my_segment]; i<end[my_segment]; i++) {
VEC0(my_ID,i) += VEC0(my_donor,i);
}
}
/* next, each thread pushes its contribution into the master thread
vector; no need to synchronize, because of the push model */
my_segment = (my_ID+1)%nthread;
if (my_ID != 0)
for (i=start[my_segment]; i<end[my_segment]; i++) {
VEC0(0,i) = VEC0(my_ID,i);
}
break;
} /* end of algorithm switch statement */
} /* end of iter loop */
#pragma omp barrier
#pragma omp master
{
reduce_time = wtime() - reduce_time;
}
} /* end of OpenMP parallel region */
/* verify correctness */
element_value = (double)nthread*(2.0*(double)nthread+1.0);
for (i=0; i<vector_length; i++) {
if (ABS(VEC0(0,i) - element_value) >= epsilon) {
printf("First error at i=%d; value: %lf; reference value: %lf\n",
i, VEC0(0,i), element_value);
exit(EXIT_FAILURE);
}
}
printf("Solution validates\n");
#ifdef VERBOSE
printf("Element verification value: %lf\n", element_value);
#endif
avgtime = reduce_time/iterations;
printf("Rate (MFlops/s): %lf Avg time (s): %lf\n",
1.0E-06 * (2.0*nthread-1.0)*vector_length/avgtime, avgtime);
exit(EXIT_SUCCESS);
}
int main(int argc, char ** argv) {
int Num_procs; /* number of ranks */
int Num_procsx, Num_procsy; /* number of ranks in each coord direction */
int my_ID; /* MPI rank */
int my_IDx, my_IDy; /* coordinates of rank in rank grid */
int right_nbr; /* global rank of right neighboring tile */
int left_nbr; /* global rank of left neighboring tile */
int top_nbr; /* global rank of top neighboring tile */
int bottom_nbr; /* global rank of bottom neighboring tile */
DTYPE *top_buf_out; /* communication buffer */
DTYPE *top_buf_in; /* " " */
DTYPE *bottom_buf_out; /* " " */
DTYPE *bottom_buf_in; /* " " */
DTYPE *right_buf_out; /* " " */
DTYPE *right_buf_in; /* " " */
DTYPE *left_buf_out; /* " " */
DTYPE *left_buf_in; /* " " */
int root = 0;
int n, width, height;/* linear global and local grid dimension */
long nsquare; /* total number of grid points */
int i, j, ii, jj, kk, it, jt, iter, leftover; /* dummies */
int istart, iend; /* bounds of grid tile assigned to calling rank */
int jstart, jend; /* bounds of grid tile assigned to calling rank */
DTYPE norm, /* L1 norm of solution */
local_norm, /* contribution of calling rank to L1 norm */
reference_norm;
DTYPE f_active_points; /* interior of grid with respect to stencil */
DTYPE flops; /* floating point ops per iteration */
int iterations; /* number of times to run the algorithm */
double local_stencil_time,/* timing parameters */
stencil_time,
avgtime;
int stencil_size; /* number of points in stencil */
int nthread_input, /* thread parameters */
nthread;
DTYPE * RESTRICT in; /* input grid values */
DTYPE * RESTRICT out; /* output grid values */
long total_length_in; /* total required length to store input array */
long total_length_out;/* total required length to store output array */
int error=0; /* error flag */
DTYPE weight[2*RADIUS+1][2*RADIUS+1]; /* weights of points in the stencil */
MPI_Request request[8];
/*******************************************************************************
** Initialize the MPI environment
********************************************************************************/
MPI_Init(&argc,&argv);
MPI_Comm_rank(MPI_COMM_WORLD, &my_ID);
MPI_Comm_size(MPI_COMM_WORLD, &Num_procs);
/*******************************************************************************
** process, test, and broadcast input parameters
********************************************************************************/
if (my_ID == root) {
printf("Parallel Research Kernels version %s\n", PRKVERSION);
printf("MPI+OPENMP stencil execution on 2D grid\n");
#ifndef STAR
printf("ERROR: Compact stencil not supported\n");
error = 1;
goto ENDOFTESTS;
#endif
if (argc != 4){
printf("Usage: %s <#threads><#iterations> <array dimension> \n",
*argv);
error = 1;
goto ENDOFTESTS;
}
/* Take number of threads to request from command line */
nthread_input = atoi(*++argv);
if ((nthread_input < 1) || (nthread_input > MAX_THREADS)) {
printf("ERROR: Invalid number of threads: %d\n", nthread_input);
error = 1;
goto ENDOFTESTS;
}
iterations = atoi(*++argv);
if (iterations < 1){
printf("ERROR: iterations must be >= 1 : %d \n",iterations);
error = 1;
goto ENDOFTESTS;
}
n = atoi(*++argv);
nsquare = (long) n * (long) n;
if (nsquare < Num_procs){
printf("ERROR: grid size %ld must be at least # ranks: %d\n",
nsquare, Num_procs);
error = 1;
goto ENDOFTESTS;
}
if (RADIUS < 0) {
printf("ERROR: Stencil radius %d should be non-negative\n", RADIUS);
error = 1;
goto ENDOFTESTS;
}
if (2*RADIUS +1 > n) {
printf("ERROR: Stencil radius %d exceeds grid size %d\n", RADIUS, n);
error = 1;
goto ENDOFTESTS;
}
ENDOFTESTS:;
}
bail_out(error);
/* determine best way to create a 2D grid of ranks (closest to square, for
best surface/volume ratio); we do this brute force for now
*/
for (Num_procsx=(int) (sqrt(Num_procs+1)); Num_procsx>0; Num_procsx--) {
if (!(Num_procs%Num_procsx)) {
Num_procsy = Num_procs/Num_procsx;
break;
}
}
my_IDx = my_ID%Num_procsx;
my_IDy = my_ID/Num_procsx;
/* compute neighbors; don't worry about dropping off the edges of the grid */
right_nbr = my_ID+1;
left_nbr = my_ID-1;
top_nbr = my_ID+Num_procsx;
bottom_nbr = my_ID-Num_procsx;
MPI_Bcast(&n, 1, MPI_INT, root, MPI_COMM_WORLD);
MPI_Bcast(&iterations, 1, MPI_INT, root, MPI_COMM_WORLD);
MPI_Bcast(&nthread_input, 1, MPI_INT, root, MPI_COMM_WORLD);
omp_set_num_threads(nthread_input);
if (my_ID == root) {
printf("Number of ranks = %d\n", Num_procs);
printf("Number of threads = %d\n", omp_get_max_threads());
printf("Grid size = %d\n", n);
printf("Radius of stencil = %d\n", RADIUS);
printf("Tiles in x/y-direction = %d/%d\n", Num_procsx, Num_procsy);
printf("Type of stencil = star\n");
#if DOUBLE
printf("Data type = double precision\n");
#else
printf("Data type = single precision\n");
#endif
#if LOOPGEN
printf("Script used to expand stencil loop body\n");
#else
printf("Compact representation of stencil loop body\n");
#endif
printf("Number of iterations = %d\n", iterations);
}
/* compute amount of space required for input and solution arrays */
width = n/Num_procsx;
leftover = n%Num_procsx;
if (my_IDx<leftover) {
istart = (width+1) * my_IDx;
iend = istart + width;
}
else {
istart = (width+1) * leftover + width * (my_IDx-leftover);
iend = istart + width - 1;
}
width = iend - istart + 1;
if (width == 0) {
printf("ERROR: rank %d has no work to do\n", my_ID);
error = 1;
}
bail_out(error);
height = n/Num_procsy;
leftover = n%Num_procsy;
if (my_IDy<leftover) {
jstart = (height+1) * my_IDy;
jend = jstart + height;
}
else {
jstart = (height+1) * leftover + height * (my_IDy-leftover);
jend = jstart + height - 1;
}
height = jend - jstart + 1;
if (height == 0) {
printf("ERROR: rank %d has no work to do\n", my_ID);
error = 1;
}
bail_out(error);
if (width < RADIUS || height < RADIUS) {
printf("ERROR: rank %d has work tile smaller then stencil radius\n",
my_ID);
error = 1;
}
bail_out(error);
total_length_in = (width+2*RADIUS)*(height+2*RADIUS)*sizeof(DTYPE);
if (total_length_in/(height+2*RADIUS) != (width+2*RADIUS)*sizeof(DTYPE)) {
printf("ERROR: Space for %d x %d input array cannot be represented\n",
width+2*RADIUS, height+2*RADIUS);
error = 1;
}
bail_out(error);
total_length_out = width*height*sizeof(DTYPE);
in = (DTYPE *) prk_malloc(total_length_in);
out = (DTYPE *) prk_malloc(total_length_out);
if (!in || !out) {
printf("ERROR: rank %d could not allocate space for input/output array\n",
my_ID);
error = 1;
}
bail_out(error);
/* fill the stencil weights to reflect a discrete divergence operator */
for (jj=-RADIUS; jj<=RADIUS; jj++) for (ii=-RADIUS; ii<=RADIUS; ii++)
WEIGHT(ii,jj) = (DTYPE) 0.0;
stencil_size = 4*RADIUS+1;
for (ii=1; ii<=RADIUS; ii++) {
WEIGHT(0, ii) = WEIGHT( ii,0) = (DTYPE) (1.0/(2.0*ii*RADIUS));
WEIGHT(0,-ii) = WEIGHT(-ii,0) = -(DTYPE) (1.0/(2.0*ii*RADIUS));
}
norm = (DTYPE) 0.0;
f_active_points = (DTYPE) (n-2*RADIUS)*(DTYPE) (n-2*RADIUS);
/* intialize the input and output arrays */
#pragma omp parallel for private (i)
for (j=jstart; j<=jend; j++) for (i=istart; i<=iend; i++) {
IN(i,j) = COEFX*i+COEFY*j;
OUT(i,j) = (DTYPE)0.0;
}
/* allocate communication buffers for halo values */
top_buf_out = (DTYPE *) prk_malloc(4*sizeof(DTYPE)*RADIUS*width);
if (!top_buf_out) {
printf("ERROR: Rank %d could not allocated comm buffers for y-direction\n", my_ID);
error = 1;
}
bail_out(error);
top_buf_in = top_buf_out + RADIUS*width;
bottom_buf_out = top_buf_out + 2*RADIUS*width;
bottom_buf_in = top_buf_out + 3*RADIUS*width;
right_buf_out = (DTYPE *) prk_malloc(4*sizeof(DTYPE)*RADIUS*height);
if (!right_buf_out) {
printf("ERROR: Rank %d could not allocated comm buffers for x-direction\n", my_ID);
error = 1;
}
bail_out(error);
right_buf_in = right_buf_out + RADIUS*height;
left_buf_out = right_buf_out + 2*RADIUS*height;
left_buf_in = right_buf_out + 3*RADIUS*height;
for (iter = 0; iter<=iterations; iter++){
/* start timer after a warmup iteration */
if (iter == 1) {
MPI_Barrier(MPI_COMM_WORLD);
local_stencil_time = wtime();
}
/* need to fetch ghost point data from neighbors in y-direction */
if (my_IDy < Num_procsy-1) {
MPI_Irecv(top_buf_in, RADIUS*width, MPI_DTYPE, top_nbr, 101,
MPI_COMM_WORLD, &(request[1]));
for (kk=0,j=jend-RADIUS+1; j<=jend; j++) for (i=istart; i<=iend; i++) {
top_buf_out[kk++]= IN(i,j);
}
MPI_Isend(top_buf_out, RADIUS*width,MPI_DTYPE, top_nbr, 99,
MPI_COMM_WORLD, &(request[0]));
}
if (my_IDy > 0) {
MPI_Irecv(bottom_buf_in,RADIUS*width, MPI_DTYPE, bottom_nbr, 99,
MPI_COMM_WORLD, &(request[3]));
for (kk=0,j=jstart; j<=jstart+RADIUS-1; j++) for (i=istart; i<=iend; i++) {
bottom_buf_out[kk++]= IN(i,j);
}
MPI_Isend(bottom_buf_out, RADIUS*width,MPI_DTYPE, bottom_nbr, 101,
MPI_COMM_WORLD, &(request[2]));
}
if (my_IDy < Num_procsy-1) {
MPI_Wait(&(request[0]), MPI_STATUS_IGNORE);
MPI_Wait(&(request[1]), MPI_STATUS_IGNORE);
for (kk=0,j=jend+1; j<=jend+RADIUS; j++) for (i=istart; i<=iend; i++) {
IN(i,j) = top_buf_in[kk++];
}
}
if (my_IDy > 0) {
MPI_Wait(&(request[2]), MPI_STATUS_IGNORE);
MPI_Wait(&(request[3]), MPI_STATUS_IGNORE);
for (kk=0,j=jstart-RADIUS; j<=jstart-1; j++) for (i=istart; i<=iend; i++) {
IN(i,j) = bottom_buf_in[kk++];
}
}
/* need to fetch ghost point data from neighbors in x-direction */
if (my_IDx < Num_procsx-1) {
MPI_Irecv(right_buf_in, RADIUS*height, MPI_DTYPE, right_nbr, 1010,
MPI_COMM_WORLD, &(request[1+4]));
for (kk=0,j=jstart; j<=jend; j++) for (i=iend-RADIUS+1; i<=iend; i++) {
right_buf_out[kk++]= IN(i,j);
}
MPI_Isend(right_buf_out, RADIUS*height, MPI_DTYPE, right_nbr, 990,
MPI_COMM_WORLD, &(request[0+4]));
}
if (my_IDx > 0) {
MPI_Irecv(left_buf_in, RADIUS*height, MPI_DTYPE, left_nbr, 990,
MPI_COMM_WORLD, &(request[3+4]));
for (kk=0,j=jstart; j<=jend; j++) for (i=istart; i<=istart+RADIUS-1; i++) {
left_buf_out[kk++]= IN(i,j);
}
MPI_Isend(left_buf_out, RADIUS*height, MPI_DTYPE, left_nbr, 1010,
MPI_COMM_WORLD, &(request[2+4]));
}
if (my_IDx < Num_procsx-1) {
MPI_Wait(&(request[0+4]), MPI_STATUS_IGNORE);
MPI_Wait(&(request[1+4]), MPI_STATUS_IGNORE);
for (kk=0,j=jstart; j<=jend; j++) for (i=iend+1; i<=iend+RADIUS; i++) {
IN(i,j) = right_buf_in[kk++];
}
}
if (my_IDx > 0) {
MPI_Wait(&(request[2+4]), MPI_STATUS_IGNORE);
MPI_Wait(&(request[3+4]), MPI_STATUS_IGNORE);
for (kk=0,j=jstart; j<=jend; j++) for (i=istart-RADIUS; i<=istart-1; i++) {
IN(i,j) = left_buf_in[kk++];
}
}
/* Apply the stencil operator */
#pragma omp parallel for private (i, j, ii, jj)
for (j=MAX(jstart,RADIUS); j<=MIN(n-RADIUS-1,jend); j++) {
for (i=MAX(istart,RADIUS); i<=MIN(n-RADIUS-1,iend); i++) {
#if LOOPGEN
#include "loop_body_star.incl"
#else
for (jj=-RADIUS; jj<=RADIUS; jj++) OUT(i,j) += WEIGHT(0,jj)*IN(i,j+jj);
for (ii=-RADIUS; ii<0; ii++) OUT(i,j) += WEIGHT(ii,0)*IN(i+ii,j);
for (ii=1; ii<=RADIUS; ii++) OUT(i,j) += WEIGHT(ii,0)*IN(i+ii,j);
#endif
}
}
#pragma omp parallel for private (i)
/* add constant to solution to force refresh of neighbor data, if any */
for (j=jstart; j<=jend; j++) for (i=istart; i<=iend; i++) IN(i,j)+= 1.0;
}
local_stencil_time = wtime() - local_stencil_time;
MPI_Reduce(&local_stencil_time, &stencil_time, 1, MPI_DOUBLE, MPI_MAX, root,
MPI_COMM_WORLD);
/* compute L1 norm in parallel */
local_norm = (DTYPE) 0.0;
#pragma omp parallel for reduction(+:local_norm) private (i)
for (j=MAX(jstart,RADIUS); j<=MIN(n-RADIUS-1,jend); j++) {
for (i=MAX(istart,RADIUS); i<=MIN(n-RADIUS-1,iend); i++) {
local_norm += (DTYPE)ABS(OUT(i,j));
}
}
MPI_Reduce(&local_norm, &norm, 1, MPI_DTYPE, MPI_SUM, root, MPI_COMM_WORLD);
/*******************************************************************************
** Analyze and output results.
********************************************************************************/
/* verify correctness */
if (my_ID == root) {
norm /= f_active_points;
if (RADIUS > 0) {
reference_norm = (DTYPE) (iterations+1) * (COEFX + COEFY);
}
else {
reference_norm = (DTYPE) 0.0;
}
if (ABS(norm-reference_norm) > EPSILON) {
printf("ERROR: L1 norm = "FSTR", Reference L1 norm = "FSTR"\n",
norm, reference_norm);
error = 1;
}
else {
printf("Solution validates\n");
#if VERBOSE
printf("Reference L1 norm = "FSTR", L1 norm = "FSTR"\n",
reference_norm, norm);
#endif
}
}
bail_out(error);
if (my_ID == root) {
/* flops/stencil: 2 flops (fma) for each point in the stencil,
plus one flop for the update of the input of the array */
flops = (DTYPE) (2*stencil_size+1) * f_active_points;
avgtime = stencil_time/iterations;
printf("Rate (MFlops/s): "FSTR" Avg time (s): %lf\n",
1.0E-06 * flops/avgtime, avgtime);
}
MPI_Finalize();
exit(EXIT_SUCCESS);
}
/*---------------------------------------------------------------------------
*
* Compute matrix product using tiling. The loop order used for the tile
* products is specified in string variable "mode".
*
* Input
* int argc - length of argv[] array
* char* argv[] - pointer to command line parameter array
* int verbosity - program verification: verbosity > 0 gives more output
* char* order - string indicating loop order, e.g., "ijk" or "jki"
*
* Output
* double - elapsed time for product computation
*/
double multiply_by_tiles( int argc, char* argv[], int verbosity, char* order )
{
int rows, cols, mids;
int rows_per_tile, cols_per_tile, mids_per_tile;
int row_start, row_end;
int col_start, col_end;
int mid_start, mid_end;
double **a, **b, **c;
double t1, t2;
double sec;
double gflop_count;
/*
* process command line arguments
*/
rows = atoi( argv[0] );
mids = atoi( argv[1] );
cols = atoi( argv[2] );
rows_per_tile = atoi( argv[3] );
mids_per_tile = atoi( argv[4] );
cols_per_tile = atoi( argv[5] );
gflop_count = 2.0 * rows * mids * cols / 1.0e9;
if ( verbosity > 0 )
{
printf( "Tiles(%3s): rows = %d, mids = %d, columns = %d\n",
order, rows, mids, cols );
printf( "block rows = %d, mids = %d, columns = %d\n",
rows_per_tile, mids_per_tile, cols_per_tile );
}
/*
* allocate and initialize matrices
*/
a = (double**) allocateMatrix( rows, mids );
b = (double**) allocateMatrix( mids, cols );
c = (double**) allocateMatrix( rows, cols );
initialize_matrices( a, b, c, rows, cols, mids, verbosity );
/*
* compute product
*/
t1 = wtime();
for ( row_start = 0; row_start < rows; row_start += rows_per_tile )
{
row_end = row_start + rows_per_tile - 1;
if ( row_end >= rows ) row_end = rows - 1;
for ( col_start = 0; col_start < cols; col_start += cols_per_tile )
{
col_end = col_start + cols_per_tile - 1;
if ( col_end >= cols ) col_end = cols - 1;
for ( mid_start = 0; mid_start < mids; mid_start += mids_per_tile )
{
mid_end = mid_start + mids_per_tile - 1;
if ( mid_end >= mids ) mid_end = mids - 1;
do_product( a, b, c, row_start, row_end, col_start,
col_end, mid_start, mid_end );
}
}
}
t2 = wtime();
sec = t2 - t1;
if ( verbosity > 1 )
printf( "checksum = %f\n", checksum( c, rows, cols ) );
printf( "tiles(%3s): %6.3f secs %6.3f gflops ",
order, sec, gflop_count / sec );
printf( "( %5d x %5d x %5d ) ( %4d x %4d x %4d )\n",
rows, mids, cols, rows_per_tile, mids_per_tile,
cols_per_tile );
/*
* clean up
*/
deallocateMatrix( a );
deallocateMatrix( b );
deallocateMatrix( c );
return t2 - t1;
}
int main(int argc, char ** argv) {
int Num_procs; /* number of ranks */
int Num_procsx,
Num_procsy; /* number of ranks in each coord direction */
int Num_groupsx,
Num_groupsy; /* number of blocks in each coord direction */
int my_group; /* sequence number of shared memory block */
int my_group_IDx,
my_group_IDy; /* coordinates of block within block grid */
int group_size; /* number of ranks in shared memory group */
int group_sizex,
group_sizey; /* number of ranks in block in each coord direction */
int my_ID; /* MPI rank */
int my_global_IDx,
my_global_IDy; /* coordinates of rank in overall rank grid */
int my_local_IDx,
my_local_IDy; /* coordinates of rank within shared memory block */
int right_nbr; /* global rank of right neighboring tile */
int left_nbr; /* global rank of left neighboring tile */
int top_nbr; /* global rank of top neighboring tile */
int bottom_nbr; /* global rank of bottom neighboring tile */
int local_nbr[4]; /* list of synchronizing local neighbors */
int num_local_nbrs; /* number of synchronizing local neighbors */
int dummy;
DTYPE *top_buf_out; /* communication buffer */
DTYPE *top_buf_in; /* " " */
DTYPE *bottom_buf_out; /* " " */
DTYPE *bottom_buf_in; /* " " */
DTYPE *right_buf_out; /* " " */
DTYPE *right_buf_in; /* " " */
DTYPE *left_buf_out; /* " " */
DTYPE *left_buf_in; /* " " */
int root = 0;
long n, width, height;/* linear global and block grid dimension */
int width_rank,
height_rank; /* linear local dimension */
int iter, leftover; /* dummies */
int istart_rank,
iend_rank; /* bounds of grid tile assigned to calling rank */
int jstart_rank,
jend_rank; /* bounds of grid tile assigned to calling rank */
int istart, iend; /* bounds of grid block containing tile */
int jstart, jend; /* bounds of grid block containing tile */
DTYPE norm, /* L1 norm of solution */
local_norm, /* contribution of calling rank to L1 norm */
reference_norm; /* value to be matched by computed norm */
DTYPE f_active_points; /* interior of grid with respect to stencil */
DTYPE flops; /* floating point ops per iteration */
int iterations; /* number of times to run the algorithm */
double local_stencil_time,/* timing parameters */
stencil_time,
avgtime;
int stencil_size; /* number of points in stencil */
DTYPE * RESTRICT in; /* input grid values */
DTYPE * RESTRICT out; /* output grid values */
long total_length_in; /* total required length to store input array */
long total_length_out;/* total required length to store output array */
int error=0; /* error flag */
DTYPE weight[2*RADIUS+1][2*RADIUS+1]; /* weights of points in the stencil */
MPI_Request request[8]; /* requests for sends & receives in 4 coord directions */
MPI_Win shm_win_in; /* shared memory window object for IN array */
MPI_Win shm_win_out; /* shared memory window object for OUT array */
MPI_Comm shm_comm_prep; /* preparatory shared memory communicator */
MPI_Comm shm_comm; /* Shared Memory Communicator */
int shm_procs; /* # of rankes in shared domain */
int shm_ID; /* MPI rank in shared memory domain */
MPI_Aint size_in; /* size of the IN array in shared memory window */
MPI_Aint size_out; /* size of the OUT array in shared memory window */
int size_mul; /* one for shm_comm root, zero for the other ranks */
int disp_unit; /* ignored */
/*******************************************************************************
** Initialize the MPI environment
********************************************************************************/
MPI_Init(&argc,&argv);
MPI_Comm_rank(MPI_COMM_WORLD, &my_ID);
MPI_Comm_size(MPI_COMM_WORLD, &Num_procs);
/*******************************************************************************
** process, test, and broadcast input parameters
********************************************************************************/
if (my_ID == root) {
printf("Parallel Research Kernels version %s\n", PRKVERSION);
printf("MPI+SHM stencil execution on 2D grid\n");
#if !STAR
printf("ERROR: Compact stencil not supported\n");
error = 1;
goto ENDOFTESTS;
#endif
if (argc != 4){
printf("Usage: %s <#ranks per coherence domain><# iterations> <array dimension> \n",
*argv);
error = 1;
goto ENDOFTESTS;
}
group_size = atoi(*++argv);
if (group_size < 1) {
printf("ERROR: # ranks per coherence domain must be >= 1 : %d \n",group_size);
error = 1;
goto ENDOFTESTS;
}
if (Num_procs%group_size) {
printf("ERROR: total # %d ranks not divisible by ranks per coherence domain %d\n",
Num_procs, group_size);
error = 1;
goto ENDOFTESTS;
}
iterations = atoi(*++argv);
if (iterations < 0){
printf("ERROR: iterations must be >= 0 : %d \n",iterations);
error = 1;
goto ENDOFTESTS;
}
n = atol(*++argv);
long nsquare = n * n;
if (nsquare < Num_procs){
printf("ERROR: grid size must be at least # ranks: %ld\n", nsquare);
error = 1;
goto ENDOFTESTS;
}
if (RADIUS < 0) {
printf("ERROR: Stencil radius %d should be non-negative\n", RADIUS);
error = 1;
goto ENDOFTESTS;
}
if (2*RADIUS +1 > n) {
printf("ERROR: Stencil radius %d exceeds grid size %ld\n", RADIUS, n);
error = 1;
goto ENDOFTESTS;
}
ENDOFTESTS:;
}
bail_out(error);
MPI_Bcast(&n, 1, MPI_LONG, root, MPI_COMM_WORLD);
MPI_Bcast(&iterations, 1, MPI_INT, root, MPI_COMM_WORLD);
MPI_Bcast(&group_size, 1, MPI_INT, root, MPI_COMM_WORLD);
/* determine best way to create a 2D grid of ranks (closest to square, for
best surface/volume ratio); we do this brute force for now. The
decomposition needs to be such that shared memory groups can evenly
tessellate the rank grid
*/
for (Num_procsx=(int) (sqrt(Num_procs+1)); Num_procsx>0; Num_procsx--) {
if (!(Num_procs%Num_procsx)) {
Num_procsy = Num_procs/Num_procsx;
for (group_sizex=(int)(sqrt(group_size+1)); group_sizex>0; group_sizex--) {
if (!(group_size%group_sizex) && !(Num_procsx%group_sizex)) {
group_sizey=group_size/group_sizex;
break;
}
}
if (!(Num_procsy%group_sizey)) break;
}
}
if (my_ID == root) {
printf("Number of ranks = %d\n", Num_procs);
printf("Grid size = %ld\n", n);
printf("Radius of stencil = %d\n", RADIUS);
printf("Tiles in x/y-direction = %d/%d\n", Num_procsx, Num_procsy);
printf("Tiles per shared memory domain = %d\n", group_size);
printf("Tiles in x/y-direction in group = %d/%d\n", group_sizex, group_sizey);
printf("Type of stencil = star\n");
#if LOCAL_BARRIER_SYNCH
printf("Local synchronization = barrier\n");
#else
printf("Local synchronization = point to point\n");
#endif
#if DOUBLE
printf("Data type = double precision\n");
#else
printf("Data type = single precision\n");
#endif
#if LOOPGEN
printf("Script used to expand stencil loop body\n");
#else
printf("Compact representation of stencil loop body\n");
#endif
printf("Number of iterations = %d\n", iterations);
}
/* Setup for Shared memory regions */
/* first divide WORLD in groups of size group_size */
MPI_Comm_split(MPI_COMM_WORLD, my_ID/group_size, my_ID%group_size, &shm_comm_prep);
/* derive from that an SHM communicator */
MPI_Comm_split_type(shm_comm_prep, MPI_COMM_TYPE_SHARED, 0, MPI_INFO_NULL, &shm_comm);
MPI_Comm_rank(shm_comm, &shm_ID);
MPI_Comm_size(shm_comm, &shm_procs);
/* do sanity check, making sure groups did not shrink in second comm split */
if (shm_procs != group_size) MPI_Abort(MPI_COMM_WORLD, 666);
Num_groupsx = Num_procsx/group_sizex;
Num_groupsy = Num_procsy/group_sizey;
my_group = my_ID/group_size;
my_group_IDx = my_group%Num_groupsx;
my_group_IDy = my_group/Num_groupsx;
my_local_IDx = my_ID%group_sizex;
my_local_IDy = (my_ID%group_size)/group_sizex;
my_global_IDx = my_group_IDx*group_sizex+my_local_IDx;
my_global_IDy = my_group_IDy*group_sizey+my_local_IDy;
/* set all neighboring ranks to -1 (no communication with those ranks) */
left_nbr = right_nbr = top_nbr = bottom_nbr = -1;
/* keep track of local neighbors for local synchronization */
num_local_nbrs = 0;
if (my_local_IDx == group_sizex-1 && my_group_IDx != (Num_groupsx-1)) {
right_nbr = (my_group+1)*group_size+shm_ID-group_sizex+1;
}
if (my_local_IDx != group_sizex-1) {
local_nbr[num_local_nbrs++] = shm_ID + 1;
}
if (my_local_IDx == 0 && my_group_IDx != 0) {
left_nbr = (my_group-1)*group_size+shm_ID+group_sizex-1;
}
if (my_local_IDx != 0) {
local_nbr[num_local_nbrs++] = shm_ID - 1;
}
if (my_local_IDy == group_sizey-1 && my_group_IDy != (Num_groupsy-1)) {
top_nbr = (my_group+Num_groupsx)*group_size + my_local_IDx;
}
if (my_local_IDy != group_sizey-1) {
local_nbr[num_local_nbrs++] = shm_ID + group_sizex;
}
if (my_local_IDy == 0 && my_group_IDy != 0) {
bottom_nbr = (my_group-Num_groupsx)*group_size + group_sizex*(group_sizey-1)+my_local_IDx;
}
if (my_local_IDy != 0) {
local_nbr[num_local_nbrs++] = shm_ID - group_sizex;
}
/* compute amount of space required for input and solution arrays for the block,
and also compute index sets */
width = n/Num_groupsx;
leftover = n%Num_groupsx;
if (my_group_IDx<leftover) {
istart = (width+1) * my_group_IDx;
iend = istart + width;
}
else {
istart = (width+1) * leftover + width * (my_group_IDx-leftover);
iend = istart + width - 1;
}
width = iend - istart + 1;
if (width == 0) {
printf("ERROR: rank %d has no work to do\n", my_ID);
error = 1;
}
bail_out(error);
height = n/Num_groupsy;
leftover = n%Num_groupsy;
if (my_group_IDy<leftover) {
jstart = (height+1) * my_group_IDy;
jend = jstart + height;
}
else {
jstart = (height+1) * leftover + height * (my_group_IDy-leftover);
jend = jstart + height - 1;
}
height = jend - jstart + 1;
if (height == 0) {
printf("ERROR: rank %d has no work to do\n", my_ID);
error = 1;
}
bail_out(error);
if (width < RADIUS || height < RADIUS) {
printf("ERROR: rank %d has work tile smaller then stencil radius; w=%ld,h=%ld\n",
my_ID, width, height);
error = 1;
}
bail_out(error);
total_length_in = (width+2*RADIUS)*(height+2*RADIUS)*sizeof(DTYPE);
total_length_out = width*height*sizeof(DTYPE);
/* only the root of each SHM domain specifies window of nonzero size */
size_mul = (shm_ID==0);
size_in= total_length_in*size_mul;
MPI_Win_allocate_shared(size_in, sizeof(double), MPI_INFO_NULL, shm_comm,
(void *) &in, &shm_win_in);
MPI_Win_lock_all(MPI_MODE_NOCHECK, shm_win_in);
MPI_Win_shared_query(shm_win_in, MPI_PROC_NULL, &size_in, &disp_unit, (void *)&in);
if (in == NULL){
printf("Error allocating space for input array by group %d\n",my_group);
error = 1;
}
bail_out(error);
size_out= total_length_out*size_mul;
MPI_Win_allocate_shared(size_out, sizeof(double), MPI_INFO_NULL, shm_comm,
(void *) &out, &shm_win_out);
MPI_Win_lock_all(MPI_MODE_NOCHECK, shm_win_out);
MPI_Win_shared_query(shm_win_out, MPI_PROC_NULL, &size_out, &disp_unit, (void *)&out);
if (out == NULL){
printf("Error allocating space for output array by group %d\n", my_group);
error = 1;
}
bail_out(error);
/* determine index set assigned to each rank */
width_rank = width/group_sizex;
leftover = width%group_sizex;
if (my_local_IDx<leftover) {
istart_rank = (width_rank+1) * my_local_IDx;
iend_rank = istart_rank + width_rank;
}
else {
istart_rank = (width_rank+1) * leftover + width_rank * (my_local_IDx-leftover);
iend_rank = istart_rank + width_rank - 1;
}
istart_rank += istart;
iend_rank += istart;
width_rank = iend_rank - istart_rank + 1;
height_rank = height/group_sizey;
leftover = height%group_sizey;
if (my_local_IDy<leftover) {
jstart_rank = (height_rank+1) * my_local_IDy;
jend_rank = jstart_rank + height_rank;
}
else {
jstart_rank = (height_rank+1) * leftover + height_rank * (my_local_IDy-leftover);
jend_rank = jstart_rank + height_rank - 1;
}
jstart_rank+=jstart;
jend_rank+=jstart;
height_rank = jend_rank - jstart_rank + 1;
if (height_rank*width_rank==0) {
error = 1;
printf("Rank %d has no work to do\n", my_ID);
}
bail_out(error);
/* allocate communication buffers for halo values */
top_buf_out = (DTYPE *) prk_malloc(4*sizeof(DTYPE)*RADIUS*width_rank);
if (!top_buf_out) {
printf("ERROR: Rank %d could not allocated comm buffers for y-direction\n", my_ID);
error = 1;
}
bail_out(error);
top_buf_in = top_buf_out + RADIUS*width_rank;
bottom_buf_out = top_buf_out + 2*RADIUS*width_rank;
bottom_buf_in = top_buf_out + 3*RADIUS*width_rank;
right_buf_out = (DTYPE *) prk_malloc(4*sizeof(DTYPE)*RADIUS*height_rank);
if (!right_buf_out) {
printf("ERROR: Rank %d could not allocated comm buffers for x-direction\n", my_ID);
error = 1;
}
bail_out(error);
right_buf_in = right_buf_out + RADIUS*height_rank;
left_buf_out = right_buf_out + 2*RADIUS*height_rank;
left_buf_in = right_buf_out + 3*RADIUS*height_rank;
/* fill the stencil weights to reflect a discrete divergence operator */
for (int jj=-RADIUS; jj<=RADIUS; jj++) for (int ii=-RADIUS; ii<=RADIUS; ii++)
WEIGHT(ii,jj) = (DTYPE) 0.0;
stencil_size = 4*RADIUS+1;
for (int ii=1; ii<=RADIUS; ii++) {
WEIGHT(0, ii) = WEIGHT( ii,0) = (DTYPE) (1.0/(2.0*ii*RADIUS));
WEIGHT(0,-ii) = WEIGHT(-ii,0) = -(DTYPE) (1.0/(2.0*ii*RADIUS));
}
norm = (DTYPE) 0.0;
f_active_points = (DTYPE) (n-2*RADIUS)*(DTYPE) (n-2*RADIUS);
/* intialize the input and output arrays */
for (int j=jstart_rank; j<=jend_rank; j++) for (int i=istart_rank; i<=iend_rank; i++) {
IN(i,j) = COEFX*i+COEFY*j;
OUT(i,j) = (DTYPE)0.0;
}
/* LOAD/STORE FENCE */
MPI_Win_sync(shm_win_in);
MPI_Win_sync(shm_win_out);
MPI_Barrier(shm_comm);
for (iter = 0; iter<=iterations; iter++){
/* start timer after a warmup iteration */
if (iter == 1) {
MPI_Barrier(MPI_COMM_WORLD);
local_stencil_time = wtime();
}
/* need to fetch ghost point data from neighbors in y-direction */
if (top_nbr != -1) {
MPI_Irecv(top_buf_in, RADIUS*width_rank, MPI_DTYPE, top_nbr, 101,
MPI_COMM_WORLD, &(request[1]));
for (int kk=0,j=jend_rank-RADIUS+1; j<=jend_rank; j++)
for (int i=istart_rank; i<=iend_rank; i++) {
top_buf_out[kk++]= IN(i,j);
}
MPI_Isend(top_buf_out, RADIUS*width_rank,MPI_DTYPE, top_nbr, 99,
MPI_COMM_WORLD, &(request[0]));
}
if (bottom_nbr != -1) {
MPI_Irecv(bottom_buf_in,RADIUS*width_rank, MPI_DTYPE, bottom_nbr, 99,
MPI_COMM_WORLD, &(request[3]));
for (int kk=0,j=jstart_rank; j<=jstart_rank+RADIUS-1; j++)
for (int i=istart_rank; i<=iend_rank; i++) {
bottom_buf_out[kk++]= IN(i,j);
}
MPI_Isend(bottom_buf_out, RADIUS*width_rank,MPI_DTYPE, bottom_nbr, 101,
MPI_COMM_WORLD, &(request[2]));
}
if (top_nbr != -1) {
MPI_Wait(&(request[0]), MPI_STATUS_IGNORE);
MPI_Wait(&(request[1]), MPI_STATUS_IGNORE);
for (int kk=0,j=jend_rank+1; j<=jend_rank+RADIUS; j++)
for (int i=istart_rank; i<=iend_rank; i++) {
IN(i,j) = top_buf_in[kk++];
}
}
if (bottom_nbr != -1) {
MPI_Wait(&(request[2]), MPI_STATUS_IGNORE);
MPI_Wait(&(request[3]), MPI_STATUS_IGNORE);
for (int kk=0,j=jstart_rank-RADIUS; j<=jstart_rank-1; j++)
for (int i=istart_rank; i<=iend_rank; i++) {
IN(i,j) = bottom_buf_in[kk++];
}
}
/* LOAD/STORE FENCE */
MPI_Win_sync(shm_win_in);
/* need to fetch ghost point data from neighbors in x-direction */
if (right_nbr != -1) {
MPI_Irecv(right_buf_in, RADIUS*height_rank, MPI_DTYPE, right_nbr, 1010,
MPI_COMM_WORLD, &(request[1+4]));
for (int kk=0,j=jstart_rank; j<=jend_rank; j++)
for (int i=iend_rank-RADIUS+1; i<=iend_rank; i++) {
right_buf_out[kk++]= IN(i,j);
}
MPI_Isend(right_buf_out, RADIUS*height_rank, MPI_DTYPE, right_nbr, 990,
MPI_COMM_WORLD, &(request[0+4]));
}
if (left_nbr != -1) {
MPI_Irecv(left_buf_in, RADIUS*height_rank, MPI_DTYPE, left_nbr, 990,
MPI_COMM_WORLD, &(request[3+4]));
for (int kk=0,j=jstart_rank; j<=jend_rank; j++)
for (int i=istart_rank; i<=istart_rank+RADIUS-1; i++) {
left_buf_out[kk++]= IN(i,j);
}
MPI_Isend(left_buf_out, RADIUS*height_rank, MPI_DTYPE, left_nbr, 1010,
MPI_COMM_WORLD, &(request[2+4]));
}
if (right_nbr != -1) {
MPI_Wait(&(request[0+4]), MPI_STATUS_IGNORE);
MPI_Wait(&(request[1+4]), MPI_STATUS_IGNORE);
for (int kk=0,j=jstart_rank; j<=jend_rank; j++)
for (int i=iend_rank+1; i<=iend_rank+RADIUS; i++) {
IN(i,j) = right_buf_in[kk++];
}
}
if (left_nbr != -1) {
MPI_Wait(&(request[2+4]), MPI_STATUS_IGNORE);
MPI_Wait(&(request[3+4]), MPI_STATUS_IGNORE);
for (int kk=0,j=jstart_rank; j<=jend_rank; j++)
for (int i=istart_rank-RADIUS; i<=istart_rank-1; i++) {
IN(i,j) = left_buf_in[kk++];
}
}
/* LOAD/STORE FENCE */
MPI_Win_sync(shm_win_in);
/* Apply the stencil operator */
for (int j=MAX(jstart_rank,RADIUS); j<=MIN(n-RADIUS-1,jend_rank); j++) {
for (int i=MAX(istart_rank,RADIUS); i<=MIN(n-RADIUS-1,iend_rank); i++) {
#if LOOPGEN
#include "loop_body_star.incl"
#else
for (int jj=-RADIUS; jj<=RADIUS; jj++) OUT(i,j) += WEIGHT(0,jj)*IN(i,j+jj);
for (int ii=-RADIUS; ii<0; ii++) OUT(i,j) += WEIGHT(ii,0)*IN(i+ii,j);
for (int ii=1; ii<=RADIUS; ii++) OUT(i,j) += WEIGHT(ii,0)*IN(i+ii,j);
#endif
}
}
/* LOAD/STORE FENCE */
MPI_Win_sync(shm_win_out);
#if LOCAL_BARRIER_SYNCH
MPI_Barrier(shm_comm); // needed to avoid writing IN while other ranks are reading it
#else
for (int i=0; i<num_local_nbrs; i++) {
MPI_Irecv(&dummy, 0, MPI_INT, local_nbr[i], 666, shm_comm, &(request[i]));
MPI_Send(&dummy, 0, MPI_INT, local_nbr[i], 666, shm_comm);
}
MPI_Waitall(num_local_nbrs, request, MPI_STATUSES_IGNORE);
#endif
/* add constant to solution to force refresh of neighbor data, if any */
for (int j=jstart_rank; j<=jend_rank; j++)
for (int i=istart_rank; i<=iend_rank; i++) IN(i,j)+= 1.0;
/* LOAD/STORE FENCE */
MPI_Win_sync(shm_win_in);
#if LOCAL_BARRIER_SYNCH
MPI_Barrier(shm_comm); // needed to avoid reading IN while other ranks are writing it
#else
for (int i=0; i<num_local_nbrs; i++) {
MPI_Irecv(&dummy, 0, MPI_INT, local_nbr[i], 666, shm_comm, &(request[i]));
MPI_Send(&dummy, 0, MPI_INT, local_nbr[i], 666, shm_comm);
}
MPI_Waitall(num_local_nbrs, request, MPI_STATUSES_IGNORE);
#endif
} /* end of iterations */
local_stencil_time = wtime() - local_stencil_time;
MPI_Reduce(&local_stencil_time, &stencil_time, 1, MPI_DOUBLE, MPI_MAX, root,
MPI_COMM_WORLD);
/* compute L1 norm in parallel */
local_norm = (DTYPE) 0.0;
for (int j=MAX(jstart_rank,RADIUS); j<=MIN(n-RADIUS-1,jend_rank); j++) {
for (int i=MAX(istart_rank,RADIUS); i<=MIN(n-RADIUS-1,iend_rank); i++) {
local_norm += (DTYPE)ABS(OUT(i,j));
}
}
MPI_Reduce(&local_norm, &norm, 1, MPI_DTYPE, MPI_SUM, root, MPI_COMM_WORLD);
/*******************************************************************************
** Analyze and output results.
********************************************************************************/
/* verify correctness */
if (my_ID == root) {
norm /= f_active_points;
if (RADIUS > 0) {
reference_norm = (DTYPE) (iterations+1) * (COEFX + COEFY);
}
else {
reference_norm = (DTYPE) 0.0;
}
if (ABS(norm-reference_norm) > EPSILON) {
printf("ERROR: L1 norm = "FSTR", Reference L1 norm = "FSTR"\n",
norm, reference_norm);
error = 1;
}
else {
printf("Solution validates\n");
#if VERBOSE
printf("Reference L1 norm = "FSTR", L1 norm = "FSTR"\n",
reference_norm, norm);
#endif
}
}
bail_out(error);
MPI_Win_unlock_all(shm_win_in);
MPI_Win_unlock_all(shm_win_out);
MPI_Win_free(&shm_win_in);
MPI_Win_free(&shm_win_out);
if (my_ID == root) {
/* flops/stencil: 2 flops (fma) for each point in the stencil,
plus one flop for the update of the input of the array */
flops = (DTYPE) (2*stencil_size+1) * f_active_points;
avgtime = stencil_time/iterations;
printf("Rate (MFlops/s): "FSTR" Avg time (s): %lf\n",
1.0E-06 * flops/avgtime, avgtime);
}
MPI_Finalize();
exit(EXIT_SUCCESS);
}
int main(int argc, char ** argv) {
int args_used = 1; // keeps track of # consumed arguments
uint64_t L; // dimension of grid in cells
uint64_t iterations; // total number of simulation steps
uint64_t n; // total number of particles in the simulation
char *init_mode; // particle initialization mode (char)
uint64_t particle_mode; // particle initialization mode (int)
double rho; // attenuation factor for geometric particle distribution
int64_t k, m; // determine initial horizontal and vertical velocity of
// particles-- (2*k)+1 cells per time step
double alpha, beta; // slope and offset values for linear particle distribution
bbox_t grid_patch, // whole grid
init_patch; // subset of grid used for localized initialization
int correctness = 1; // determines whether simulation was correct
double *Qgrid; // field of fixed charges
particle_t *particles, *p; // the particles array
uint64_t iter, i; // dummies
double fx, fy, ax, ay; // forces and accelerations
#if UNUSED
int particles_per_cell;// number of particles per cell to be injected
int error=0; // used for graceful exit after error
#endif
double avg_time, pic_time;// timing parameters
int nthread_input, // thread parameters
nthread;
int num_error=0; // flag that signals that requested and obtained
// numbers of threads are the same
random_draw_t dice;
printf("Parallel Research Kernels Version %s\n", PRKVERSION);
printf("OpenMP Particle-in-Cell execution on 2D grid\n");
/*******************************************************************************
** process and test input parameters
********************************************************************************/
if (argc<7) {
printf("Usage: %s <#threads> <#simulation steps> <grid size> <#particles> <k (particle charge semi-increment)> ", argv[0]);
printf("<m (vertical particle velocity)>\n");
printf(" <init mode> <init parameters>]\n");
printf(" init mode \"GEOMETRIC\" parameters: <attenuation factor>\n");
printf(" \"SINUSOIDAL\" parameters: none\n");
printf(" \"LINEAR\" parameters: <negative slope> <constant offset>\n");
printf(" \"PATCH\" parameters: <xleft> <xright> <ybottom> <ytop>\n");
exit(SUCCESS);
}
/* Take number of threads to request from command line */
nthread_input = atoi(*++argv);
if ((nthread_input < 1) || (nthread_input > MAX_THREADS)) {
printf("ERROR: Invalid number of threads: %d\n", nthread_input);
exit(EXIT_FAILURE);
}
omp_set_num_threads(nthread_input);
iterations = atol(*++argv); args_used++;
if (iterations<1) {
printf("ERROR: Number of time steps must be positive: %" PRIu64 "\n", iterations);
exit(FAILURE);
}
L = atol(*++argv); args_used++;
if (L<1 || L%2) {
printf("ERROR: Number of grid cells must be positive and even: %" PRIu64 "\n", L);
exit(FAILURE);
}
grid_patch = (bbox_t){0, L+1, 0, L+1};
n = atol(*++argv); args_used++;
if (n<1) {
printf("ERROR: Number of particles must be positive: %" PRIu64 "\n", n);
exit(FAILURE);
}
particle_mode = UNDEFINED;
k = atoi(*++argv); args_used++;
if (k<0) {
printf("ERROR: Particle semi-charge must be non-negative: %" PRIu64 "\n", k);
exit(FAILURE);
}
m = atoi(*++argv); args_used++;
init_mode = *++argv; args_used++;
/* Initialize particles with geometric distribution */
if (strcmp(init_mode, "GEOMETRIC") == 0) {
if (argc<args_used+1) {
printf("ERROR: Not enough arguments for GEOMETRIC\n");
exit(FAILURE);
}
particle_mode = GEOMETRIC;
rho = atof(*++argv); args_used++;
}
/* Initialize with a sinusoidal particle distribution (single period) */
if (strcmp(init_mode, "SINUSOIDAL") == 0) {
particle_mode = SINUSOIDAL;
}
/* Initialize particles with linear distribution */
/* The linear function is f(x) = -alpha * x + beta , x in [0,1]*/
if (strcmp(init_mode, "LINEAR") == 0) {
if (argc<args_used+2) {
printf("ERROR: Not enough arguments for LINEAR initialization\n");
exit(EXIT_FAILURE);
}
particle_mode = LINEAR;
alpha = atof(*++argv); args_used++;
beta = atof(*++argv); args_used++;
if (beta <0 || beta<alpha) {
printf("ERROR: linear profile gives negative particle density\n");
exit(EXIT_FAILURE);
}
}
/* Initialize particles uniformly within a "patch" */
if (strcmp(init_mode, "PATCH") == 0) {
if (argc<args_used+4) {
printf("ERROR: Not enough arguments for PATCH initialization\n");
exit(FAILURE);
}
particle_mode = PATCH;
init_patch.left = atoi(*++argv); args_used++;
init_patch.right = atoi(*++argv); args_used++;
init_patch.bottom = atoi(*++argv); args_used++;
init_patch.top = atoi(*++argv); args_used++;
if (bad_patch(&init_patch, &grid_patch)) {
printf("ERROR: inconsistent initial patch\n");
exit(FAILURE);
}
}
#pragma omp parallel
{
#pragma omp master
{
nthread = omp_get_num_threads();
if (nthread != nthread_input) {
num_error = 1;
printf("ERROR: number of requested threads %d does not equal ",
nthread_input);
printf("number of spawned threads %d\n", nthread);
}
else {
printf("Number of threads = %d\n",nthread_input);
printf("Grid size = %lld\n", L);
printf("Number of particles requested = %lld\n", n);
printf("Number of time steps = %lld\n", iterations);
printf("Initialization mode = %s\n", init_mode);
switch(particle_mode) {
case GEOMETRIC: printf(" Attenuation factor = %lf\n", rho); break;
case SINUSOIDAL: break;
case LINEAR: printf(" Negative slope = %lf\n", alpha);
printf(" Offset = %lf\n", beta); break;
case PATCH: printf(" Bounding box = %" PRIu64 "%" PRIu64 "%" PRIu64 "%" PRIu64 "\n",
init_patch.left, init_patch.right,
init_patch.bottom, init_patch.top); break;
default: printf("ERROR: Unsupported particle initializating mode\n");
exit(FAILURE);
}
printf("Particle charge semi-increment = %"PRIu64"\n", k);
printf("Vertical velocity = %"PRIu64"\n", m);
/* Initialize grid of charges and particles */
Qgrid = initializeGrid(L);
LCG_init(&dice);
switch(particle_mode) {
case GEOMETRIC: particles = initializeGeometric(n, L, rho, k, m, &n, &dice); break;
case SINUSOIDAL: particles = initializeSinusoidal(n, L, k, m, &n, &dice); break;
case LINEAR: particles = initializeLinear(n, L, alpha, beta, k, m, &n, &dice); break;
case PATCH: particles = initializePatch(n, L, init_patch, k, m, &n, &dice); break;
default: printf("ERROR: Unsupported particle distribution\n"); exit(FAILURE);
}
printf("Number of particles placed = %lld\n", n);
}
}
bail_out(num_error);
}
for (iter=0; iter<=iterations; iter++) {
/* start the timer after one warm-up time step */
if (iter==1) {
pic_time = wtime();
}
/* Calculate forces on particles and update positions */
#pragma omp parallel for private(i, p, fx, fy, ax, ay)
for (i=0; i<n; i++) {
p = particles;
fx = 0.0;
fy = 0.0;
computeTotalForce(p[i], L, Qgrid, &fx, &fy);
ax = fx * MASS_INV;
ay = fy * MASS_INV;
/* Update particle positions, taking into account periodic boundaries */
p[i].x = fmod(p[i].x + p[i].v_x*DT + 0.5*ax*DT*DT + L, L);
p[i].y = fmod(p[i].y + p[i].v_y*DT + 0.5*ay*DT*DT + L, L);
/* Update velocities */
p[i].v_x += ax * DT;
p[i].v_y += ay * DT;
}
}
pic_time = wtime() - pic_time;
/* Run the verification test */
for (i=0; i<n; i++) {
correctness *= verifyParticle(particles[i], iterations, Qgrid, L);
}
if (correctness) {
printf("Solution validates\n");
#ifdef VERBOSE
printf("Simulation time is %lf seconds\n", pic_time);
#endif
avg_time = n*iterations/pic_time;
printf("Rate (Mparticles_moved/s): %lf\n", 1.0e-6*avg_time);
} else {
printf("Solution does not validate\n");
}
return(EXIT_SUCCESS);
}
int main(int argc, char ** argv) {
int N;
int tile_size=32; /* default tile size for tiling of local transpose */
int num_iterations;/* number of times to do the transpose */
int tiling; /* boolean: true if tiling is used */
double total_bytes; /* combined size of matrices */
double start_time, /* timing parameters */
end_time, avgtime;
/*********************************************************************
** read and test input parameters
*********************************************************************/
if(argc != 3 && argc != 4){
if(MYTHREAD == 0)
printf("Usage: %s <# iterations> <matrix order> [tile size]\n", *argv);
upc_global_exit(EXIT_FAILURE);
}
num_iterations = atoi(*++argv);
if(num_iterations < 1){
if(MYTHREAD == 0)
printf("ERROR: iterations must be >= 1 : %d \n", num_iterations);
upc_global_exit(EXIT_FAILURE);
}
N = atoi(*++argv);
if(N < 0){
if(MYTHREAD == 0)
printf("ERROR: Matrix Order must be greater than 0 : %d \n", N);
upc_global_exit(EXIT_FAILURE);
}
if (argc == 4)
tile_size = atoi(*++argv);
/*a non-positive tile size means no tiling of the local transpose */
tiling = (tile_size > 0) && (tile_size < N);
if(!tiling)
tile_size = N;
int sizex = N / THREADS;
if(N % THREADS != 0) {
if(MYTHREAD == 0)
printf("N %% THREADS != 0\n");
upc_global_exit(EXIT_FAILURE);
}
int sizey = N;
if(MYTHREAD == 0) {
printf("Parallel Research Kernels version %s\n", PRKVERSION);
printf("UPC matrix transpose: B = A^T\n");
printf("Number of threads = %d\n", THREADS);
printf("Matrix order = %d\n", N);
printf("Number of iterations = %d\n", num_iterations);
if (tiling)
printf("Tile size = %d\n", tile_size);
else printf("Untiled\n");
}
/*********************************************************************
** Allocate memory for input and output matrices
*********************************************************************/
total_bytes = 2.0 * sizeof(double) * N * N;
int myoffsetx = MYTHREAD * sizex;
int myoffsety = 0;
upc_barrier;
debug("Allocating arrays (%d, %d), offset (%d, %d)", sizex, sizey, myoffsetx, myoffsety);
local_shared_block_ptrs in_array = shared_2d_array_alloc(sizex, sizey, myoffsetx, myoffsety);
local_shared_block_ptrs out_array = shared_2d_array_alloc(sizex, sizey, myoffsetx, myoffsety);
local_shared_block_ptrs buf_array = shared_2d_array_alloc(sizex, sizey, myoffsetx, myoffsety);
in_arrays[MYTHREAD] = in_array;
out_arrays[MYTHREAD] = out_array;
buf_arrays[MYTHREAD] = buf_array;
double **in_array_private = shared_2d_array_to_private(in_array, sizex, sizey, myoffsetx, myoffsety);
double **out_array_private = shared_2d_array_to_private(out_array, sizex, sizey, myoffsetx, myoffsety);
double **buf_array_private = shared_2d_array_to_private(buf_array, sizex, sizey, myoffsetx, myoffsety);
upc_barrier;
/*********************************************************************
** Initialize the matrices
*********************************************************************/
for(int y=myoffsety; y<myoffsety + sizey; y++){
for(int x=myoffsetx; x<myoffsetx + sizex; x++){
in_array_private[y][x] = (double) (x+N*y);
out_array[y][x] = -1.0;
}
}
upc_barrier;
for(int y=myoffsety; y<myoffsety + sizey; y++){
for(int x=myoffsetx; x<myoffsetx + sizex; x++){
if(in_array_private[y][x] !=(double) (x+N*y))
die("x=%d y=%d in_array=%f != %f", x, y, in_array[y][x], (x+N*y));
if(out_array_private[y][x] != -1.0)
die("out_array_private error");
}
}
/*********************************************************************
** Transpose
*********************************************************************/
int transfer_size = sizex * sizex * sizeof(double);
if(MYTHREAD == 0)
debug("transfer size = %d", transfer_size);
for(int iter=0; iter<=num_iterations; iter++){
/* start timer after a warmup iteration */
if(iter == 1){
upc_barrier;
start_time = wtime();
}
for(int i=0; i<THREADS; i++){
int local_blk_id = (MYTHREAD + i) % THREADS;
int remote_blk_id = MYTHREAD;
int remote_thread = local_blk_id;
upc_memget(&buf_array_private[local_blk_id * sizex][myoffsetx],
&in_arrays[remote_thread][remote_blk_id * sizex][remote_thread * sizex], transfer_size);
#define OUT_ARRAY(x,y) out_array_private[local_blk_id * sizex + x][myoffsetx + y]
#define BUF_ARRAY(x,y) buf_array_private[local_blk_id * sizex + x][myoffsetx + y]
if(!tiling){
for(int x=0; x<sizex; x++){
for(int y=0; y<sizex; y++){
OUT_ARRAY(x,y) = BUF_ARRAY(y,x);
}
}
}
else{
for(int x=0; x<sizex; x+=tile_size){
for(int y=0; y<sizex; y+=tile_size){
for(int bx=x; bx<MIN(sizex, x+tile_size); bx++){
for(int by=y; by<MIN(sizex, y+tile_size); by++){
OUT_ARRAY(bx,by) = BUF_ARRAY(by,bx);
}
}
}
}
}
}
upc_barrier;
}
upc_barrier;
end_time = wtime();
/*********************************************************************
** Analyze and output results.
*********************************************************************/
for(int y=myoffsety; y<myoffsety + sizey; y++){
for(int x=myoffsetx; x<myoffsetx + sizex; x++){
if(in_array_private[y][x] != (double)(x+ N*y))
die("Error in input: x=%d y=%d", x, y);
if(out_array_private[y][x] != (double)(y + N*x))
die("x=%d y=%d in_array=%f != %f %d %d", x, y, out_array[y][x], (double)(y + N*x), (int)(out_array[y][x]) % N, (int)(out_array[y][x]) / N);
}
}
if(MYTHREAD == 0){
printf("Solution validates\n");
double transfer_size = 2 * N * N * sizeof(double);
avgtime = (end_time - start_time) / num_iterations;
double rate = transfer_size / avgtime * 1.0E-06;
printf("Rate (MB/s): %lf Avg time (s): %lf\n",rate, avgtime);
}
}
/***********************************************************************
* Read the input file.
***********************************************************************/
int read_input ( FILE *fp_in, FILE *fp_out, input_data *input_vars,
para_data *para_vars, time_data *time_vars )
{
/***********************************************************************
* Local variables.
***********************************************************************/
double t1, t2;
int ierr = 0;
char *error = NULL;
char *line = NULL;
size_t len = 0;
ssize_t read;
char *tmpData = NULL;
int tmpStrLen, i;
/***********************************************************************
* Read the input file. Echo to output file. Call for an input variable
* check. Only root reads, echoes, checks input.
***********************************************************************/
t1 = wtime ();
if ( IPROC == ROOT )
{
if ( !fp_in )
{
tmpStrLen = strlen (" ***ERROR: READ_INPUT:"
" Problem reading input file.\n");
ALLOC_STR ( error, tmpStrLen + 1, &ierr );
snprintf ( error, tmpStrLen + 1,
" ***ERROR: READ_INPUT:"
" Problem reading input file.\n" );
print_error ( fp_out, error, IPROC, ROOT );
FREE ( error );
ierr = 1;
}
else
{
while ( (read = getline(&line, &len, fp_in)) != -1 )
{
i = 0;
while ( isspace(line[i]) )
{
i++;
}
// Parallel processing inputs
// npey: number of process elements in y-dir
if ( strncmp(&line[i], "npey=", strlen("npey=")) == 0 )
{
get_input_value ( &line[i], "npey=", &tmpData );
NPEY = atoi ( tmpData );
}
// npez: input number of process elements in z-dir
else if ( strncmp(&line[i], "npez=", strlen("npez=")) == 0 )
{
get_input_value ( &line[i], "npez=", &tmpData );
NPEZ = atoi ( tmpData );
}
// ichunk:
else if ( strncmp(&line[i], "ichunk=", strlen("ichunk=")) == 0 )
{
get_input_value ( &line[i], "ichunk=", &tmpData );
ICHUNK = atoi ( tmpData );
}
// nthreads: input number of threads
else if ( strncmp(&line[i], "nthreads=", strlen("nthreads=")) == 0 )
{
get_input_value ( &line[i], "nthreads=", &tmpData );
NTHREADS = atoi ( tmpData );
}
// nnested:
else if ( strncmp(&line[i], "nnested=", strlen("nnested=")) == 0 )
{
get_input_value ( &line[i], "nnested=", &tmpData );
NNESTED = atoi ( tmpData );
}
// Geometry inputs
// ndimen:
else if ( strncmp(&line[i], "ndimen=", strlen("ndimen=")) == 0 )
{
get_input_value ( &line[i], "ndimen=", &tmpData );
NDIMEN = atoi ( tmpData );
}
// nx:
else if ( strncmp(&line[i], "nx=", strlen("nx=")) == 0 )
{
get_input_value ( &line[i], "nx=", &tmpData );
NX = atoi ( tmpData );
}
// ny:
else if ( strncmp(&line[i], "ny=", strlen("ny=")) == 0 )
{
get_input_value ( &line[i], "ny=", &tmpData );
NY = atoi ( tmpData );
}
// nz:
else if ( strncmp(&line[i], "nz=", strlen("nz=")) == 0 )
{
get_input_value ( &line[i], "nz=", &tmpData );
NZ = atoi ( tmpData );
}
// lx:
else if ( strncmp(&line[i], "lx=", strlen("lx=")) == 0 )
{
get_input_value ( &line[i], "lx=", &tmpData );
LX = atof ( tmpData );
}
// ly:
else if ( strncmp(&line[i], "ly=", strlen("ly=")) == 0 )
{
get_input_value ( &line[i], "ly=", &tmpData );
LY = atof ( tmpData );
}
// lz:
else if ( strncmp(&line[i], "lz=", strlen("lz=")) == 0 )
{
get_input_value ( &line[i], "lz=", &tmpData );
LZ = atof ( tmpData );
}
// Sn inputs
// nmom:
else if ( strncmp(&line[i], "nmom=", strlen("nmom=")) == 0 )
{
get_input_value ( &line[i], "nmom=", &tmpData );
NMOM = atoi ( tmpData );
}
// nang:
else if ( strncmp(&line[i], "nang=", strlen("nang=")) == 0 )
{
get_input_value ( &line[i], "nang=", &tmpData );
NANG = atoi ( tmpData );
}
// Data inputs
// ng:
else if ( strncmp(&line[i], "ng=", strlen("ng=")) == 0 )
{
get_input_value ( &line[i], "ng=", &tmpData );
NG = atoi ( tmpData );
}
// mat_opt:
else if ( strncmp(&line[i], "mat_opt=", strlen("mat_opt=")) == 0 )
{
get_input_value ( &line[i], "mat_opt=", &tmpData );
MAT_OPT = atoi ( tmpData );
}
// src_opt:
else if ( strncmp(&line[i], "src_opt=", strlen("src_opt=")) == 0 )
{
get_input_value ( &line[i], "src_opt=", &tmpData );
SRC_OPT = atoi ( tmpData );
}
// scatp:
else if ( strncmp(&line[i], "scatp=", strlen("scatp=")) == 0 )
{
get_input_value ( &line[i], "scatp=", &tmpData );
SCATP = atoi ( tmpData );
}
// Control inputs
// epsi:
else if ( strncmp(&line[i], "epsi=", strlen("epsi=")) == 0 )
{
get_input_value ( &line[i], "epsi=", &tmpData );
EPSI = atof ( tmpData );
}
// tf:
else if ( strncmp(&line[i], "tf=", strlen("tf=")) == 0 )
{
get_input_value ( &line[i], "tf=", &tmpData );
TF = atof ( tmpData );
}
// iitm:
else if ( strncmp(&line[i], "iitm=", strlen("iitm=")) == 0 )
{
get_input_value ( &line[i], "iitm=", &tmpData );
IITM = atoi ( tmpData );
}
// oitm:
else if ( strncmp(&line[i], "oitm=", strlen("oitm=")) == 0 )
{
get_input_value ( &line[i], "oitm=", &tmpData );
OITM = atoi ( tmpData );
}
// timedep:
else if ( strncmp(&line[i], "timedep=", strlen("timedep=")) == 0 )
{
get_input_value ( &line[i], "timedep=", &tmpData );
TIMEDEP = atoi ( tmpData );
}
// nsteps:
else if ( strncmp(&line[i], "nsteps=", strlen("nsteps=")) == 0 )
{
get_input_value ( &line[i], "nsteps=", &tmpData );
NSTEPS = atoi ( tmpData );
}
// it_det:
else if ( strncmp(&line[i], "it_det=", strlen("it_det=")) == 0 )
{
get_input_value ( &line[i], "it_det=", &tmpData );
IT_DET = atoi ( tmpData );
}
// fluxp:
else if ( strncmp(&line[i], "fluxp=", strlen("fluxp=")) == 0 )
{
get_input_value ( &line[i], "fluxp=", &tmpData );
FLUXP = atoi ( tmpData );
}
// fixup:
else if ( strncmp(&line[i], "fixup=", strlen("fixup=")) == 0 )
{
get_input_value ( &line[i], "fixup=", &tmpData );
FIXUP = atoi ( tmpData );
}
}
// Free temp data from memory
FREE ( tmpData );
}
}
bcast_i_scalar ( &ierr, COMM_SNAP, ROOT, NPROC );
if ( ierr != 0 )
{
return ierr;
}
if ( IPROC == ROOT )
{
input_echo ( input_vars, fp_out );
ierr = input_check ( fp_out, input_vars, para_vars );
}
/***********************************************************************
* Broadcast the data to all processes.
***********************************************************************/
ierr = var_bcast ( input_vars, para_vars );
t2 = wtime();
time_vars->tinp = t2 - t1;
return ierr;
}
int main(int argc, char *argv[])
{
int my_ID, myrow, mycol, /* my index and row and column index */
root=0, /* ID of root rank */
Num_procs, /* number of ranks */
nprow, npcol, /* row, column dimensions of rank grid */
order, /* matrix order */
mynrows, myfrow, /* my number of rows and index of first row*/
mylrow, /* and last row */
myncols, myfcol, /* my number of cols and index of first row*/
mylcol, /* and last row */
*mm, /* arrays that hold m_i's and n_j's */
*nn,
nb, /* block factor for SUMMA */
inner_block_flag, /* flag to select local DGEMM blocking */
error=0, /* error flag */
*ranks, /* work array for row and column ranks */
lda, ldb, ldc, /* leading array dimensions of a, b, and c */
i, j, ii, jj, /* dummy variables */
iter, iterations;
double *a, *b, *c, /* arrays that hold local a, b, c */
*work1, *work2, /* work arrays to pass to dpmmmult */
local_dgemm_time, /* timing parameters */
dgemm_time,
avgtime;
double
forder, nflops, /* float matrix order + total flops */
checksum, /* array checksum for verification test */
checksum_local=0.0,
ref_checksum; /* reference checkcum for verification */
MPI_Group world_group,
temp_group;
MPI_Comm comm_row, /* communicators for row and column ranks */
comm_col; /* of rank grid */
/* initialize */
MPI_Init(&argc,&argv);
MPI_Comm_rank( MPI_COMM_WORLD, &my_ID );
MPI_Comm_size( MPI_COMM_WORLD, &Num_procs );
/*********************************************************************
** process, test and broadcast input parameters
*********************************************************************/
if (my_ID == root) {
if (argc != 5) {
printf("Usage: %s <# iterations> <matrix order> <outer block size> ",
*argv);
printf("<local block flag (non-zero=yes, zero=no)>\n");
error = 1;
goto ENDOFTESTS;
}
iterations = atoi(*++argv);
if(iterations < 1) {
printf("ERROR: iterations must be positive: %d \n",iterations);
error = 1;
goto ENDOFTESTS;
}
order = atoi(*++argv);
if (order < Num_procs) {
printf("ERROR: matrix order too small: %d\n", order);
error = 1;
goto ENDOFTESTS;
}
nb = atoi(*++argv);
/* a non-positive tile size means no outer level tiling */
inner_block_flag = atoi(*++argv);
ENDOFTESTS:
;
}
bail_out(error);
MPI_Bcast(&order, 1, MPI_INT, root, MPI_COMM_WORLD);
MPI_Bcast(&iterations, 1, MPI_INT, root, MPI_COMM_WORLD);
MPI_Bcast(&nb, 1, MPI_INT, root, MPI_COMM_WORLD);
MPI_Bcast(&inner_block_flag, 1, MPI_INT, root, MPI_COMM_WORLD);
/* compute rank grid to most closely match a square; to do so,
compute largest divisor of Num_procs, using hare-brained method.
The small term epsilon is used to guard against roundoff errors
in case Num_procs is a perfect square */
nprow = (int) (sqrt((double) Num_procs + epsilon));
while (Num_procs%nprow) nprow--;
npcol = Num_procs/nprow;
if (my_ID == root) {
printf("Parallel Research Kernels version %s\n", PRKVERSION);
printf("MPI Dense matrix-matrix multiplication: C = A x B\n");
printf("Number of ranks = %d\n", Num_procs);
printf("Rank grid = %d rows x %d columns\n", nprow, npcol);
printf("Matrix order = %d\n", order);
printf("Outer block size = %d\n", nb);
printf("Number of iterations = %d\n", iterations);
if (inner_block_flag)
printf("Using local dgemm blocking\n");
else
printf("No local dgemm blocking\n");
}
/* set up row and column communicators */
ranks = (int *) malloc (3*Num_procs*sizeof(int));
if (!ranks) {
printf("ERROR: Proc %d could not allocate rank work arrays\n",
my_ID);
error = 1;
}
bail_out(error);
mm = ranks + Num_procs;
nn = mm + Num_procs;
/* 1. extract group of ranks that make up WORLD */
MPI_Comm_group( MPI_COMM_WORLD, &world_group );
/* 2. create list of all ranks in same row of rank grid */
ranks[0] = my_ID/npcol * npcol;
for (i=1; i<npcol; i++) ranks[i] = ranks[i-1] + 1;
/* create row group and communicator */
MPI_Group_incl( world_group, npcol, ranks, &temp_group );
MPI_Comm_create( MPI_COMM_WORLD, temp_group, &comm_row );
/* 3. create list of all ranks in same column of rank grid */
ranks[0] = my_ID%npcol;
for (i=1; i<nprow; i++) ranks[i] = ranks[i-1] + npcol;
/* create column group and communicator */
MPI_Group_incl( world_group, nprow, ranks, &temp_group );
MPI_Comm_create( MPI_COMM_WORLD, temp_group, &comm_col );
/* extract this node's row and column index */
MPI_Comm_rank( comm_row, &mycol );
MPI_Comm_rank( comm_col, &myrow );
/* mynrows = number of rows assigned to me; distribute excess
rows evenly if nprow does not divide matrix order evenly */
if (myrow < order%nprow) mynrows = (order/nprow)+1;
else mynrows = (order/nprow);
/* make sure lda is a multiple of the block size nb */
if (mynrows%nb==0 || mynrows<nb) lda = mynrows;
else lda = (mynrows/nb+1)*nb;
/* myncols = number of colums assigned to me; distribute excess
columns evenly if npcol does not divide order evenly */
if (mycol < order%npcol) myncols = (order/npcol)+1;
else myncols = (order/npcol);
/* get space for local blocks of A, B, C */
a = (double *) malloc( lda*myncols*sizeof(double) );
b = (double *) malloc( lda*myncols*sizeof(double) );
c = (double *) malloc( lda*myncols*sizeof(double) );
if ( a == NULL || b == NULL || c == NULL ) {
error = 1;
printf("ERROR: Proc %d could not allocate a, b, and/or c\n",my_ID);
}
bail_out(error);
/* get space for two work arrays for dgemm */
work1 = (double *) malloc( nb*lda*sizeof(double) );
work2 = (double *) malloc( nb*myncols*sizeof(double) );
if ( !work1 || !work2 ) {
printf("ERROR: Proc %d could not allocate work buffers\n", my_ID);
error = 1;
}
bail_out(error);
/* collect array that holds mynrows from all nodes in my row
of the rank grid (array of all m_i) */
MPI_Allgather( &mynrows, 1, MPI_INT, mm, 1, MPI_INT, comm_col );
/* myfrow = first row on my node */
for (myfrow=1,i=0; i<myrow; i++) myfrow += mm[i];
mylrow = myfrow+mynrows-1;
/* collect array that holds myncols from all nodes in my column
of the rank grid (array of all n_j) */
MPI_Allgather( &myncols, 1, MPI_INT, nn, 1, MPI_INT, comm_row );
/* myfcol = first col on my node */
for (myfcol=1,i=0; i<mycol; i++) myfcol += nn[i];
mylcol = myfcol+myncols-1;
/* initialize matrices A, B, and C */
ldc = ldb = lda;
for (jj=0, j=myfcol; j<=mylcol; j++,jj++ )
for (ii=0, i=myfrow; i<=mylrow; i++, ii++ ) {
A(ii,jj) = (double) (j-1);
B(ii,jj) = (double) (j-1);
C(ii,jj) = 0.0;
}
for (iter=0; iter<=iterations; iter++) {
/* start timer after a warmup iteration */
if (iter == 1) {
MPI_Barrier(MPI_COMM_WORLD);
local_dgemm_time = wtime();
}
/* actual matrix-vector multiply */
dgemm(order, nb, inner_block_flag, a, lda, b, lda, c, lda,
mm, nn, comm_row, comm_col, work1, work2 );
} /* end of iterations */
local_dgemm_time = wtime() - local_dgemm_time;
MPI_Reduce(&local_dgemm_time, &dgemm_time, 1, MPI_DOUBLE, MPI_MAX, root,
MPI_COMM_WORLD);
/* verification test */
for (jj=0, j=myfcol; j<=mylcol; j++, jj++)
for (ii=0, i=myfrow; i<=mylrow; i++, ii++)
checksum_local += C(ii,jj);
MPI_Reduce(&checksum_local, &checksum, 1, MPI_DOUBLE, MPI_SUM,
root, MPI_COMM_WORLD);
forder = (double) order;
ref_checksum = (0.25*forder*forder*forder*(forder-1.0)*(forder-1.0));
ref_checksum *= (iterations+1);
if (my_ID == root) {
if (ABS((checksum - ref_checksum)/ref_checksum) > epsilon) {
printf("ERROR: Checksum = %lf, Reference checksum = %lf\n",
checksum, ref_checksum);
error = 1;
}
else {
printf("Solution validates\n");
#ifdef VERBOSE
printf("Reference checksum = %lf, checksum = %lf\n",
ref_checksum, checksum);
#endif
}
}
bail_out(error);
/* report elapsed time */
nflops = 2.0*forder*forder*forder;
if ( my_ID == root ) {
avgtime = dgemm_time/iterations;
printf("Rate (MFlops/s): %lf Avg time (s): %lf\n",
1.0E-06 * nflops/avgtime, avgtime);
}
MPI_Finalize();
}
int main(int argc, char ** argv) {
int Num_procs; /* number of ranks */
int Num_procsx, Num_procsy; /* number of ranks in each coord direction */
int my_ID; /* SHMEM rank */
int my_IDx, my_IDy; /* coordinates of rank in rank grid */
int right_nbr; /* global rank of right neighboring tile */
int left_nbr; /* global rank of left neighboring tile */
int top_nbr; /* global rank of top neighboring tile */
int bottom_nbr; /* global rank of bottom neighboring tile */
DTYPE *top_buf_out; /* communication buffer */
DTYPE *top_buf_in[2]; /* " " */
DTYPE *bottom_buf_out; /* " " */
DTYPE *bottom_buf_in[2];/* " " */
DTYPE *right_buf_out; /* " " */
DTYPE *right_buf_in[2]; /* " " */
DTYPE *left_buf_out; /* " " */
DTYPE *left_buf_in[2]; /* " " */
int root = 0;
int n, width, height;/* linear global and local grid dimension */
int i, j, ii, jj, kk, it, jt, iter, leftover; /* dummies */
int istart, iend; /* bounds of grid tile assigned to calling rank */
int jstart, jend; /* bounds of grid tile assigned to calling rank */
DTYPE reference_norm;
DTYPE f_active_points; /* interior of grid with respect to stencil */
int stencil_size; /* number of points in the stencil */
DTYPE flops; /* floating point ops per iteration */
int iterations; /* number of times to run the algorithm */
double avgtime, /* timing parameters */
*local_stencil_time, *stencil_time;
DTYPE * RESTRICT in; /* input grid values */
DTYPE * RESTRICT out; /* output grid values */
long total_length_in; /* total required length to store input array */
long total_length_out;/* total required length to store output array */
int error=0; /* error flag */
DTYPE weight[2*RADIUS+1][2*RADIUS+1]; /* weights of points in the stencil */
int *arguments; /* command line parameters */
int count_case=4; /* number of neighbors of a rank */
long *pSync_bcast; /* work space for collectives */
long *pSync_reduce; /* work space for collectives */
double *pWrk_time; /* work space for collectives */
DTYPE *pWrk_norm; /* work space for collectives */
int *iterflag; /* synchronization flags */
int sw; /* double buffering switch */
DTYPE *local_norm, *norm; /* local and global error norms */
/*******************************************************************************
** Initialize the SHMEM environment
********************************************************************************/
prk_shmem_init();
my_ID=prk_shmem_my_pe();
Num_procs=prk_shmem_n_pes();
pSync_bcast = (long *) prk_shmem_malloc(PRK_SHMEM_BCAST_SYNC_SIZE*sizeof(long));
pSync_reduce = (long *) prk_shmem_malloc(PRK_SHMEM_REDUCE_SYNC_SIZE*sizeof(long));
pWrk_time = (double *) prk_shmem_malloc(PRK_SHMEM_REDUCE_MIN_WRKDATA_SIZE*sizeof(double));
pWrk_norm = (DTYPE *) prk_shmem_malloc(PRK_SHMEM_REDUCE_MIN_WRKDATA_SIZE*sizeof(DTYPE));
local_stencil_time = (double *) prk_shmem_malloc(sizeof(double));
stencil_time = (double *) prk_shmem_malloc(sizeof(double));
local_norm = (DTYPE *) prk_shmem_malloc(sizeof(DTYPE));
norm = (DTYPE *) prk_shmem_malloc(sizeof(DTYPE));
iterflag = (int *) prk_shmem_malloc(2*sizeof(int));
if (!(pSync_bcast && pSync_reduce && pWrk_time && pWrk_norm && iterflag &&
local_stencil_time && stencil_time && local_norm && norm))
{
printf("Could not allocate scalar variables on rank %d\n", my_ID);
error = 1;
}
bail_out(error);
for(i=0;i<PRK_SHMEM_BCAST_SYNC_SIZE;i++)
pSync_bcast[i]=PRK_SHMEM_SYNC_VALUE;
for(i=0;i<PRK_SHMEM_REDUCE_SYNC_SIZE;i++)
pSync_reduce[i]=PRK_SHMEM_SYNC_VALUE;
arguments=(int*)prk_shmem_malloc(2*sizeof(int));
/*******************************************************************************
** process, test, and broadcast input parameters
********************************************************************************/
if (my_ID == root) {
#ifndef STAR
printf("ERROR: Compact stencil not supported\n");
error = 1;
goto ENDOFTESTS;
#endif
if (argc != 3){
printf("Usage: %s <# iterations> <array dimension> \n",
*argv);
error = 1;
goto ENDOFTESTS;
}
iterations = atoi(*++argv);
arguments[0]=iterations;
if (iterations < 1){
printf("ERROR: iterations must be >= 1 : %d \n",iterations);
error = 1;
goto ENDOFTESTS;
}
n = atoi(*++argv);
arguments[1]=n;
long nsquare = (long)n * (long)n;
if (nsquare < Num_procs){
printf("ERROR: grid size must be at least # ranks: %ld\n", nsquare);
error = 1;
goto ENDOFTESTS;
}
if (RADIUS < 0) {
printf("ERROR: Stencil radius %d should be non-negative\n", RADIUS);
error = 1;
goto ENDOFTESTS;
}
if (2*RADIUS +1 > n) {
printf("ERROR: Stencil radius %d exceeds grid size %d\n", RADIUS, n);
error = 1;
goto ENDOFTESTS;
}
ENDOFTESTS:;
}
bail_out(error);
/* determine best way to create a 2D grid of ranks (closest to square, for
best surface/volume ratio); we do this brute force for now
*/
for (Num_procsx=(int) (sqrt(Num_procs+1)); Num_procsx>0; Num_procsx--) {
if (!(Num_procs%Num_procsx)) {
Num_procsy = Num_procs/Num_procsx;
break;
}
}
my_IDx = my_ID%Num_procsx;
my_IDy = my_ID/Num_procsx;
/* compute neighbors; don't worry about dropping off the edges of the grid */
right_nbr = my_ID+1;
left_nbr = my_ID-1;
top_nbr = my_ID+Num_procsx;
bottom_nbr = my_ID-Num_procsx;
iterflag[0] = iterflag[1] = 0;
if(my_IDx==0) count_case--;
if(my_IDx==Num_procsx-1) count_case--;
if(my_IDy==0) count_case--;
if(my_IDy==Num_procsy-1) count_case--;
if (my_ID == root) {
printf("Parallel Research Kernels version %s\n", PRKVERSION);
printf("SHMEM stencil execution on 2D grid\n");
printf("Number of ranks = %d\n", Num_procs);
printf("Grid size = %d\n", n);
printf("Radius of stencil = %d\n", RADIUS);
printf("Tiles in x/y-direction = %d/%d\n", Num_procsx, Num_procsy);
printf("Type of stencil = star\n");
#ifdef DOUBLE
printf("Data type = double precision\n");
#else
printf("Data type = single precision\n");
#endif
#if LOOPGEN
printf("Script used to expand stencil loop body\n");
#else
printf("Compact representation of stencil loop body\n");
#endif
#if SPLITFENCE
printf("Split fence = ON\n");
#else
printf("Split fence = OFF\n");
#endif
printf("Number of iterations = %d\n", iterations);
}
shmem_barrier_all();
shmem_broadcast32(&arguments[0], &arguments[0], 2, root, 0, 0, Num_procs, pSync_bcast);
iterations=arguments[0];
n=arguments[1];
shmem_barrier_all();
prk_shmem_free(arguments);
/* compute amount of space required for input and solution arrays */
width = n/Num_procsx;
leftover = n%Num_procsx;
if (my_IDx<leftover) {
istart = (width+1) * my_IDx;
iend = istart + width + 1;
}
else {
istart = (width+1) * leftover + width * (my_IDx-leftover);
iend = istart + width;
}
width = iend - istart + 1;
if (width == 0) {
printf("ERROR: rank %d has no work to do\n", my_ID);
error = 1;
}
bail_out(error);
height = n/Num_procsy;
leftover = n%Num_procsy;
if (my_IDy<leftover) {
jstart = (height+1) * my_IDy;
jend = jstart + height + 1;
}
else {
jstart = (height+1) * leftover + height * (my_IDy-leftover);
jend = jstart + height;
}
height = jend - jstart + 1;
if (height == 0) {
printf("ERROR: rank %d has no work to do\n", my_ID);
error = 1;
}
bail_out(error);
if (width < RADIUS || height < RADIUS) {
printf("ERROR: rank %d has work tile smaller then stencil radius\n",
my_ID);
error = 1;
}
bail_out(error);
total_length_in = (width+2*RADIUS);
total_length_in *= (height+2*RADIUS);
total_length_in *= sizeof(DTYPE);
total_length_out = width;
total_length_out *= height;
total_length_out *= sizeof(DTYPE);
in = (DTYPE *) malloc(total_length_in);
out = (DTYPE *) malloc(total_length_out);
if (!in || !out) {
printf("ERROR: rank %d could not allocate space for input/output array\n",
my_ID);
error = 1;
}
bail_out(error);
/* fill the stencil weights to reflect a discrete divergence operator */
for (jj=-RADIUS; jj<=RADIUS; jj++) for (ii=-RADIUS; ii<=RADIUS; ii++)
WEIGHT(ii,jj) = (DTYPE) 0.0;
stencil_size = 4*RADIUS+1;
for (ii=1; ii<=RADIUS; ii++) {
WEIGHT(0, ii) = WEIGHT( ii,0) = (DTYPE) (1.0/(2.0*ii*RADIUS));
WEIGHT(0,-ii) = WEIGHT(-ii,0) = -(DTYPE) (1.0/(2.0*ii*RADIUS));
}
norm[0] = (DTYPE) 0.0;
f_active_points = (DTYPE) (n-2*RADIUS)*(DTYPE) (n-2*RADIUS);
/* intialize the input and output arrays */
for (j=jstart; j<jend; j++) for (i=istart; i<iend; i++) {
IN(i,j) = COEFX*i+COEFY*j;
OUT(i,j) = (DTYPE)0.0;
}
/* allocate communication buffers for halo values */
top_buf_out=(DTYPE*)malloc(2*sizeof(DTYPE)*RADIUS*width);
if (!top_buf_out) {
printf("ERROR: Rank %d could not allocate output comm buffers for y-direction\n", my_ID);
error = 1;
}
bail_out(error);
bottom_buf_out = top_buf_out+RADIUS*width;
top_buf_in[0]=(DTYPE*)prk_shmem_malloc(4*sizeof(DTYPE)*RADIUS*width);
if(!top_buf_in)
{
printf("ERROR: Rank %d could not allocate input comm buffers for y-direction\n", my_ID);
error=1;
}
bail_out(error);
top_buf_in[1] = top_buf_in[0] + RADIUS*width;
bottom_buf_in[0] = top_buf_in[1] + RADIUS*width;
bottom_buf_in[1] = bottom_buf_in[0] + RADIUS*width;
right_buf_out=(DTYPE*)malloc(2*sizeof(DTYPE)*RADIUS*height);
if (!right_buf_out) {
printf("ERROR: Rank %d could not allocate output comm buffers for x-direction\n", my_ID);
error = 1;
}
bail_out(error);
left_buf_out=right_buf_out+RADIUS*height;
right_buf_in[0]=(DTYPE*)prk_shmem_malloc(4*sizeof(DTYPE)*RADIUS*height);
if(!right_buf_in)
{
printf("ERROR: Rank %d could not allocate input comm buffers for x-dimension\n", my_ID);
error=1;
}
bail_out(error);
right_buf_in[1] = right_buf_in[0] + RADIUS*height;
left_buf_in[0] = right_buf_in[1] + RADIUS*height;
left_buf_in[1] = left_buf_in[0] + RADIUS*height;
/* make sure all symmetric heaps are allocated before being used */
shmem_barrier_all();
for (iter = 0; iter<=iterations; iter++){
/* start timer after a warmup iteration */
if (iter == 1) {
shmem_barrier_all();
local_stencil_time[0] = wtime();
}
/* sw determines which incoming buffer to select */
sw = iter%2;
/* need to fetch ghost point data from neighbors */
if (my_IDy < Num_procsy-1) {
for (kk=0,j=jend-RADIUS; j<=jend-1; j++) for (i=istart; i<=iend; i++) {
top_buf_out[kk++]= IN(i,j);
}
shmem_putmem(bottom_buf_in[sw], top_buf_out, RADIUS*width*sizeof(DTYPE), top_nbr);
#if SPLITFENCE
shmem_fence();
shmem_int_inc(&iterflag[sw], top_nbr);
#endif
}
if (my_IDy > 0) {
for (kk=0,j=jstart; j<=jstart+RADIUS-1; j++) for (i=istart; i<=iend; i++) {
bottom_buf_out[kk++]= IN(i,j);
}
shmem_putmem(top_buf_in[sw], bottom_buf_out, RADIUS*width*sizeof(DTYPE), bottom_nbr);
#if SPLITFENCE
shmem_fence();
shmem_int_inc(&iterflag[sw], bottom_nbr);
#endif
}
if(my_IDx < Num_procsx-1) {
for(kk=0,j=jstart;j<=jend;j++) for(i=iend-RADIUS;i<=iend-1;i++) {
right_buf_out[kk++]=IN(i,j);
}
shmem_putmem(left_buf_in[sw], right_buf_out, RADIUS*height*sizeof(DTYPE), right_nbr);
#if SPLITFENCE
shmem_fence();
shmem_int_inc(&iterflag[sw], right_nbr);
#endif
}
if(my_IDx>0) {
for(kk=0,j=jstart;j<=jend;j++) for(i=istart;i<=istart+RADIUS-1;i++) {
left_buf_out[kk++]=IN(i,j);
}
shmem_putmem(right_buf_in[sw], left_buf_out, RADIUS*height*sizeof(DTYPE), left_nbr);
#if SPLITFENCE
shmem_fence();
shmem_int_inc(&iterflag[sw], left_nbr);
#endif
}
#if SPLITFENCE == 0
shmem_fence();
if(my_IDy<Num_procsy-1) shmem_int_inc(&iterflag[sw], top_nbr);
if(my_IDy>0) shmem_int_inc(&iterflag[sw], bottom_nbr);
if(my_IDx<Num_procsx-1) shmem_int_inc(&iterflag[sw], right_nbr);
if(my_IDx>0) shmem_int_inc(&iterflag[sw], left_nbr);
#endif
shmem_int_wait_until(&iterflag[sw], SHMEM_CMP_EQ, count_case*(iter/2+1));
if (my_IDy < Num_procsy-1) {
for (kk=0,j=jend; j<=jend+RADIUS-1; j++) for (i=istart; i<=iend; i++) {
IN(i,j) = top_buf_in[sw][kk++];
}
}
if (my_IDy > 0) {
for (kk=0,j=jstart-RADIUS; j<=jstart-1; j++) for (i=istart; i<=iend; i++) {
IN(i,j) = bottom_buf_in[sw][kk++];
}
}
if (my_IDx < Num_procsx-1) {
for (kk=0,j=jstart; j<=jend; j++) for (i=iend; i<=iend+RADIUS-1; i++) {
IN(i,j) = right_buf_in[sw][kk++];
}
}
if (my_IDx > 0) {
for (kk=0,j=jstart; j<=jend; j++) for (i=istart-RADIUS; i<=istart-1; i++) {
IN(i,j) = left_buf_in[sw][kk++];
}
}
/* Apply the stencil operator */
for (j=MAX(jstart,RADIUS); j<=MIN(n-RADIUS-1,jend); j++) {
for (i=MAX(istart,RADIUS); i<=MIN(n-RADIUS-1,iend); i++) {
#if LOOPGEN
#include "loop_body_star.incl"
#else
for (jj=-RADIUS; jj<=RADIUS; jj++) OUT(i,j) += WEIGHT(0,jj)*IN(i,j+jj);
for (ii=-RADIUS; ii<0; ii++) OUT(i,j) += WEIGHT(ii,0)*IN(i+ii,j);
for (ii=1; ii<=RADIUS; ii++) OUT(i,j) += WEIGHT(ii,0)*IN(i+ii,j);
#endif
}
}
/* add constant to solution to force refresh of neighbor data, if any */
for (j=jstart; j<jend; j++) for (i=istart; i<iend; i++) IN(i,j)+= 1.0;
}
local_stencil_time[0] = wtime() - local_stencil_time[0];
shmem_barrier_all();
shmem_double_max_to_all(&stencil_time[0], &local_stencil_time[0], 1, 0, 0,
Num_procs, pWrk_time, pSync_reduce);
/* compute L1 norm in parallel */
local_norm[0] = (DTYPE) 0.0;
for (j=MAX(jstart,RADIUS); j<MIN(n-RADIUS,jend); j++) {
for (i=MAX(istart,RADIUS); i<MIN(n-RADIUS,iend); i++) {
local_norm[0] += (DTYPE)ABS(OUT(i,j));
}
}
shmem_barrier_all();
#ifdef DOUBLE
shmem_double_sum_to_all(&norm[0], &local_norm[0], 1, 0, 0, Num_procs, pWrk_norm, pSync_reduce);
#else
shmem_float_sum_to_all(&norm[0], &local_norm[0], 1, 0, 0, Num_procs, pWrk_norm, pSync_reduce);
#endif
/*******************************************************************************
** Analyze and output results.
********************************************************************************/
/* verify correctness */
if (my_ID == root) {
norm[0] /= f_active_points;
if (RADIUS > 0) {
reference_norm = (DTYPE) (iterations+1) * (COEFX + COEFY);
}
else {
reference_norm = (DTYPE) 0.0;
}
if (ABS(norm[0]-reference_norm) > EPSILON) {
printf("ERROR: L1 norm = "FSTR", Reference L1 norm = "FSTR"\n",
norm[0], reference_norm);
error = 1;
}
else {
printf("Solution validates\n");
#ifdef VERBOSE
printf("Reference L1 norm = "FSTR", L1 norm = "FSTR"\n",
reference_norm, norm[0]);
#endif
}
}
bail_out(error);
if (my_ID == root) {
/* flops/stencil: 2 flops (fma) for each point in the stencil,
plus one flop for the update of the input of the array */
flops = (DTYPE) (2*stencil_size+1) * f_active_points;
avgtime = stencil_time[0]/iterations;
printf("Rate (MFlops/s): "FSTR" Avg time (s): %lf\n",
1.0E-06 * flops/avgtime, avgtime);
}
prk_shmem_free(top_buf_in);
prk_shmem_free(right_buf_in);
free(top_buf_out);
free(right_buf_out);
prk_shmem_free(pSync_bcast);
prk_shmem_free(pSync_reduce);
prk_shmem_free(pWrk_time);
prk_shmem_free(pWrk_norm);
prk_shmem_finalize();
exit(EXIT_SUCCESS);
}
int main(int argc, char **argv){
int iter, r; /* dummies */
int lsize; /* logarithmic linear size of grid */
int lsize2; /* logarithmic size of grid */
int size; /* linear size of grid */
s64Int size2; /* matrix order (=total # points in grid) */
int radius, /* stencil parameters */
stencil_size;
s64Int row, col, first, last; /* dummies */
s64Int i, j; /* dummies */
int iterations; /* number of times the multiplication is done */
s64Int elm; /* sequence number of matrix nonzero */
s64Int nent; /* number of nonzero entries */
double sparsity; /* fraction of non-zeroes in matrix */
double sparse_time,/* timing parameters */
avgtime = 0.0,
maxtime = 0.0,
mintime = 366.0*24.0*3600.0; /* set the minimum time to
a large value; one leap year should be enough */
double * RESTRICT matrix; /* sparse matrix entries */
double * RESTRICT vector; /* vector multiplying the sparse matrix */
double * RESTRICT result; /* computed matrix-vector product */
double temp; /* temporary scalar storing reduction data */
double vector_sum; /* checksum of result */
double reference_sum; /* checksum of "rhs" */
double epsilon = 1.e-8; /* error tolerance */
s64Int * RESTRICT colIndex; /* column indices of sparse matrix entries */
int nthread_input, /* thread parameters */
nthread;
int num_error=0; /* flag that signals that requested and
obtained numbers of threads are the same */
size_t vector_space, /* variables used to hold malloc sizes */
matrix_space,
index_space;
if (argc != 5) {
printf("Usage: %s <# threads> <# iterations> <2log grid size> <stencil radius>\n",*argv);
exit(EXIT_FAILURE);
}
/* Take number of threads to request from command line */
nthread_input = atoi(*++argv);
if ((nthread_input < 1) || (nthread_input > MAX_THREADS)) {
printf("ERROR: Invalid number of threads: %d\n", nthread_input);
exit(EXIT_FAILURE);
}
omp_set_num_threads(nthread_input);
iterations = atoi(*++argv);
if (iterations < 1){
printf("ERROR: Iterations must be positive : %d \n", iterations);
exit(EXIT_FAILURE);
}
lsize = atoi(*++argv);
lsize2 = 2*lsize;
size = 1<<lsize;
if (lsize <0) {
printf("ERROR: Log of grid size must be greater than or equal to zero: %d\n",
(int) lsize);
exit(EXIT_FAILURE);
}
/* compute number of points in the grid */
size2 = size*size;
radius = atoi(*++argv);
if (radius <0) {
printf("ERROR: Stencil radius must be non-negative: %d\n", (int) size);
exit(EXIT_FAILURE);
}
/* emit error if (periodic) stencil overlaps with itself */
if (size <2*radius+1) {
printf("ERROR: Grid extent %d smaller than stencil diameter 2*%d+1= %d\n",
size, radius, radius*2+1);
exit(EXIT_FAILURE);
}
/* compute total size of star stencil in 2D */
stencil_size = 4*radius+1;
/* sparsity follows from number of non-zeroes per row */
sparsity = (double)(4*radius+1)/(double)size2;
/* compute total number of non-zeroes */
nent = size2*stencil_size;
matrix_space = nent*sizeof(double);
if (matrix_space/sizeof(double) != nent) {
printf("ERROR: Cannot represent space for matrix: %ld\n", matrix_space);
exit(EXIT_FAILURE);
}
matrix = (double *) malloc(matrix_space);
if (!matrix) {
printf("ERROR: Could not allocate space for sparse matrix: "FSTR64U"\n", nent);
exit(EXIT_FAILURE);
}
vector_space = 2*size2*sizeof(double);
if (vector_space/sizeof(double) != 2*size2) {
printf("ERROR: Cannot represent space for vectors: %ld\n", vector_space);
exit(EXIT_FAILURE);
}
vector = (double *) malloc(vector_space);
if (!vector) {
printf("ERROR: Could not allocate space for vectors: %d\n", (int)(2*size2));
exit(EXIT_FAILURE);
}
result = vector + size2;
index_space = nent*sizeof(s64Int);
if (index_space/sizeof(s64Int) != nent) {
printf("ERROR: Cannot represent space for column indices: %ld\n", index_space);
exit(EXIT_FAILURE);
}
colIndex = (s64Int *) malloc(index_space);
if (!colIndex) {
printf("ERROR: Could not allocate space for column indices: "FSTR64U"\n",
nent*sizeof(s64Int));
exit(EXIT_FAILURE);
}
#pragma omp parallel private (row, col, elm, first, last, iter)
{
#pragma omp master
{
nthread = omp_get_num_threads();
printf("OpenMP Sparse matrix-vector multiplication\n");
if (nthread != nthread_input) {
num_error = 1;
printf("ERROR: number of requested threads %d does not equal ",
nthread_input);
printf("number of spawned threads %d\n", nthread);
}
else {
printf("Number of threads = %16d\n",nthread_input);
printf("Matrix order = "FSTR64U"\n", size2);
printf("Stencil diameter = %16d\n", 2*radius+1);
printf("Sparsity = %16.10lf\n", sparsity);
#ifdef SCRAMBLE
printf("Using scrambled indexing\n");
#else
printf("Using canonical indexing\n");
#endif
printf("Number of iterations = %16d\n", iterations);
}
}
bail_out(num_error);
/* initialize the input and result vectors */
#pragma omp for
for (row=0; row<size2; row++) result[row] = vector[row] = 0.0;
/* fill matrix with nonzeroes corresponding to difference stencil. We use the
scrambling for reordering the points in the grid. */
#pragma omp for private (i,j,r)
for (row=0; row<size2; row++) {
j = row/size; i=row%size;
elm = row*stencil_size;
colIndex[elm] = REVERSE(LIN(i,j),lsize2);
for (r=1; r<=radius; r++, elm+=4) {
colIndex[elm+1] = REVERSE(LIN((i+r)%size,j),lsize2);
colIndex[elm+2] = REVERSE(LIN((i-r+size)%size,j),lsize2);
colIndex[elm+3] = REVERSE(LIN(i,(j+r)%size),lsize2);
colIndex[elm+4] = REVERSE(LIN(i,(j-r+size)%size),lsize2);
}
// sort colIndex to make sure the compressed row accesses
// vector elements in increasing order
qsort(&(colIndex[row*stencil_size]), stencil_size, sizeof(s64Int), compare);
for (elm=row*stencil_size; elm<(row+1)*stencil_size; elm++)
matrix[elm] = 1.0/(double)(colIndex[elm]+1);
}
for (iter=0; iter<iterations; iter++) {
#pragma omp barrier
#pragma omp master
{
sparse_time = wtime();
}
/* fill vector */
#pragma omp for
for (row=0; row<size2; row++) vector[row] += (double) (row+1);
/* do the actual matrix-vector multiplication */
#pragma omp for
for (row=0; row<size2; row++) {
temp = 0.0;
first = stencil_size*row; last = first+stencil_size-1;
#pragma simd reduction(+:temp)
for (col=first; col<=last; col++) {
temp += matrix[col]*vector[colIndex[col]];
}
result[row] += temp;
}
#pragma omp master
{
sparse_time = wtime() - sparse_time;
if (iter>0 || iterations==1) { /* skip the first iteration */
avgtime = avgtime + sparse_time;
mintime = MIN(mintime, sparse_time);
maxtime = MAX(maxtime, sparse_time);
}
}
}
} /* end of parallel region */
/* verification test */
reference_sum = 0.5 * (double) nent * (double) iterations *
(double) (iterations +1);
vector_sum = 0.0;
for (row=0; row<size2; row++) vector_sum += result[row];
if (ABS(vector_sum-reference_sum) > epsilon) {
printf("ERROR: Vector sum = %lf, Reference vector sum = %lf\n",
vector_sum, reference_sum);
exit(EXIT_FAILURE);
}
else {
printf("Solution validates\n");
#ifdef VERBOSE
printf("Reference sum = %lf, vector sum = %lf\n",
reference_sum, vector_sum);
#endif
}
avgtime = avgtime/(double)(MAX(iterations-1,1));
printf("Rate (MFlops/s): %lf, Avg time (s): %lf, Min time (s): %lf",
1.0E-06 * (2.0*nent)/mintime, avgtime, mintime);
printf(", Max time (s): %lf\n", maxtime);
exit(EXIT_SUCCESS);
}
int main(int argc, char ** argv) {
long order; /* order of a the matrix */
int Tile_order=32; /* default tile size for tiling of local transpose */
int iterations; /* number of times to do the transpose */
int tiling; /* boolean: true if tiling is used */
int i, j, it, jt, iter; /* dummies */
double bytes; /* combined size of matrices */
double * RESTRICT A; /* buffer to hold original matrix */
double * RESTRICT B; /* buffer to hold transposed matrix */
double abserr; /* absolute error */
double epsilon=1.e-8; /* error tolerance */
double transpose_time,/* timing parameters */
avgtime;
int nthread_input,
nthread;
int num_error=0; /* flag that signals that requested and
obtained numbers of threads are the same */
/*********************************************************************
** read and test input parameters
*********************************************************************/
printf("Parallel Research Kernels version %s\n", PRKVERSION);
printf("OpenMP Matrix transpose: B = A^T\n");
if (argc != 4 && argc != 5){
printf("Usage: %s <# threads> <# iterations> <matrix order> [tile size]\n",
*argv);
exit(EXIT_FAILURE);
}
/* Take number of threads to request from command line */
nthread_input = atoi(*++argv);
if ((nthread_input < 1) || (nthread_input > MAX_THREADS)) {
printf("ERROR: Invalid number of threads: %d\n", nthread_input);
exit(EXIT_FAILURE);
}
omp_set_num_threads(nthread_input);
iterations = atoi(*++argv);
if (iterations < 1){
printf("ERROR: iterations must be >= 1 : %d \n",iterations);
exit(EXIT_FAILURE);
}
order = atoi(*++argv);
if (order < 0){
printf("ERROR: Matrix Order must be greater than 0 : %d \n", order);
exit(EXIT_FAILURE);
}
if (argc == 5) Tile_order = atoi(*++argv);
/* a non-positive tile size means no tiling of the local transpose */
tiling = (Tile_order > 0) && (Tile_order < order);
if (!tiling) Tile_order = order;
/*********************************************************************
** Allocate space for the input and transpose matrix
*********************************************************************/
A = (double *)malloc(order*order*sizeof(double));
if (A == NULL){
printf(" ERROR: cannot allocate space for input matrix: %ld\n",
order*order*sizeof(double));
exit(EXIT_FAILURE);
}
B = (double *)malloc(order*order*sizeof(double));
if (B == NULL){
printf(" ERROR: cannot allocate space for output matrix: %ld\n",
order*order*sizeof(double));
exit(EXIT_FAILURE);
}
bytes = 2.0 * sizeof(double) * order * order;
#pragma omp parallel private (iter)
{
#pragma omp master
{
nthread = omp_get_num_threads();
if (nthread != nthread_input) {
num_error = 1;
printf("ERROR: number of requested threads %d does not equal ",
nthread_input);
printf("number of spawned threads %d\n", nthread);
}
else {
printf("Number of threads = %i;\n",nthread_input);
printf("Matrix order = %ld\n", order);
printf("Number of iterations = %d\n", iterations);
if (tiling) {
printf("Tile size = %d\n", Tile_order);
#ifdef COLLAPSE
printf("Using loop collapse\n");
#endif
}
else
printf("Untiled\n");
}
}
bail_out(num_error);
/* Fill the original matrix, set transpose to known garbage value. */
if (tiling) {
#ifdef COLLAPSE
#pragma omp for private (i,it,jt) collapse(2)
#else
#pragma omp for private (i,it,jt)
#endif
for (j=0; j<order; j+=Tile_order)
for (i=0; i<order; i+=Tile_order)
for (jt=j; jt<MIN(order,j+Tile_order);jt++)
for (it=i; it<MIN(order,i+Tile_order); it++){
A(it,jt) = (double) (order*jt + it);
B(it,jt) = 0.0;
}
}
else {
#pragma omp for private (i)
for (j=0;j<order;j++)
for (i=0;i<order; i++) {
A(i,j) = (double) (order*j + i);
B(i,j) = 0.0;
}
}
for (iter = 0; iter<=iterations; iter++){
/* start timer after a warmup iteration */
if (iter == 1) {
#pragma omp barrier
#pragma omp master
{
transpose_time = wtime();
}
}
/* Transpose the matrix */
if (!tiling) {
#pragma omp for private (j)
for (i=0;i<order; i++)
for (j=0;j<order;j++) {
B(j,i) += A(i,j);
A(i,j) += 1.0;
}
}
else {
#ifdef COLLAPSE
#pragma omp for private (j,it,jt) collapse(2)
#else
#pragma omp for private (j,it,jt)
#endif
for (i=0; i<order; i+=Tile_order)
for (j=0; j<order; j+=Tile_order)
for (it=i; it<MIN(order,i+Tile_order); it++)
for (jt=j; jt<MIN(order,j+Tile_order);jt++) {
B(jt,it) += A(it,jt);
A(it,jt) += 1.0;
}
}
} /* end of iter loop */
#pragma omp barrier
#pragma omp master
{
transpose_time = wtime() - transpose_time;
}
} /* end of OpenMP parallel region */
abserr = test_results (order, B, iterations);
/*********************************************************************
** Analyze and output results.
*********************************************************************/
if (abserr < epsilon) {
printf("Solution validates\n");
avgtime = transpose_time/iterations;
printf("Rate (MB/s): %lf Avg time (s): %lf\n",
1.0E-06 * bytes/avgtime, avgtime);
#ifdef VERBOSE
printf("Squared errors: %f \n", abserr);
#endif
exit(EXIT_SUCCESS);
}
else {
printf("ERROR: Aggregate squared error %lf exceeds threshold %e\n",
abserr, epsilon);
exit(EXIT_FAILURE);
}
} /* end of main */
int main(int argc, char ** argv)
{
int my_ID; /* Thread ID */
int vector_length; /* length of vector loop containing the branch */
int nfunc; /* number of functions used in INS_HEAVY option */
int rank; /* matrix rank used in INS_HEAVY option */
double branch_time, /* timing parameters */
no_branch_time;
double ops; /* double precision representation of integer ops */
int iterations; /* number of times the branch loop is carried out */
int i, iter, aux; /* dummies */
char *branch_type; /* string defining branching type */
int btype; /* integer encoding branching type */
int total=0,
total_ref; /* computed and stored verification values */
int nthread_input; /* thread parameters */
int nthread;
int num_error=0; /* flag that signals that requested and obtained
numbers of threads are the same */
/**********************************************************************************
** process and test input parameters
**********************************************************************************/
if (argc != 5){
printf("Usage: %s <# threads> <# iterations> <vector length>", *argv);
printf("<branching type>\n");
printf("branching type: vector_go, vector_stop, no_vector, ins_heavy\n");
exit(EXIT_FAILURE);
}
nthread_input = atoi(*++argv);
if ((nthread_input < 1) || (nthread_input > MAX_THREADS)) {
printf("ERROR: Invalid number of threads: %d\n", nthread_input);
exit(EXIT_FAILURE);
}
omp_set_num_threads(nthread_input);
iterations = atoi(*++argv);
if (iterations < 1 || iterations%2==1){
printf("ERROR: Iterations must be positive and even : %d \n", iterations);
exit(EXIT_FAILURE);
}
vector_length = atoi(*++argv);
if (vector_length < 1){
printf("ERROR: loop length must be >= 1 : %d \n",vector_length);
exit(EXIT_FAILURE);
}
branch_type = *++argv;
if (!strcmp(branch_type,"vector_stop")) btype = VECTOR_STOP;
else if (!strcmp(branch_type,"vector_go" )) btype = VECTOR_GO;
else if (!strcmp(branch_type,"no_vector" )) btype = NO_VECTOR;
else if (!strcmp(branch_type,"ins_heavy" )) btype = INS_HEAVY;
else {
printf("Wrong branch type: %s; choose vector_stop, vector_go, ", branch_type);
printf("no_vector, or ins_heavy\n");
exit(EXIT_FAILURE);
}
#pragma omp parallel private(i, my_ID, iter, aux, nfunc, rank) reduction(+:total)
{
int * RESTRICT vector; int * RESTRICT index;
int factor = -1;
#pragma omp master
{
nthread = omp_get_num_threads();
printf("Parallel Research Kernels version %s\n", PRKVERSION);
printf("OpenMP Branching Bonanza\n");
if (nthread != nthread_input) {
num_error = 1;
printf("ERROR: number of requested threads %d does not equal ",
nthread_input);
printf("number of spawned threads %d\n", nthread);
}
else {
printf("Number of threads = %d\n", nthread_input);
printf("Vector length = %d\n", vector_length);
printf("Number of iterations = %d\n", iterations);
printf("Branching type = %s\n", branch_type);
}
}
bail_out(num_error);
my_ID = omp_get_thread_num();
vector = malloc(vector_length*2*sizeof(int));
if (!vector) {
printf("ERROR: Thread %d failed to allocate space for vector\n", my_ID);
num_error = 1;
}
bail_out(num_error);
/* grab the second half of vector to store index array */
index = vector + vector_length;
/* initialize the array with entries with varying signs; array "index" is only
used to obfuscate the compiler (i.e. it won't vectorize a loop containing
indirect referencing). It functions as the identity operator. */
for (i=0; i<vector_length; i++) {
vector[i] = 3 - (i&7);
index[i] = i;
}
#pragma omp barrier
#pragma omp master
{
branch_time = wtime();
}
/* do actual branching */
switch (btype) {
case VECTOR_STOP:
/* condition vector[index[i]]>0 inhibits vectorization */
for (iter=0; iter<iterations; iter+=2) {
#pragma vector always
for (i=0; i<vector_length; i++) {
aux = -(3 - (i&7));
if (vector[index[i]]>0) vector[i] -= 2*vector[i];
else vector[i] -= 2*aux;
}
#pragma vector always
for (i=0; i<vector_length; i++) {
aux = (3 - (i&7));
if (vector[index[i]]>0) vector[i] -= 2*vector[i];
else vector[i] -= 2*aux;
}
}
break;
case VECTOR_GO:
/* condition aux>0 allows vectorization */
for (iter=0; iter<iterations; iter+=2) {
#pragma vector always
for (i=0; i<vector_length; i++) {
aux = -(3 - (i&7));
if (aux>0) vector[i] -= 2*vector[i];
else vector[i] -= 2*aux;
}
#pragma vector always
for (i=0; i<vector_length; i++) {
aux = (3 - (i&7));
if (aux>0) vector[i] -= 2*vector[i];
else vector[i] -= 2*aux;
}
}
break;
case NO_VECTOR:
/* condition aux>0 allows vectorization, but indirect indexing inbibits it */
for (iter=0; iter<iterations; iter+=2) {
#pragma vector always
for (i=0; i<vector_length; i++) {
aux = -(3 - (i&7));
if (aux>0) vector[i] -= 2*vector[index[i]];
else vector[i] -= 2*aux;
}
#pragma vector always
for (i=0; i<vector_length; i++) {
aux = (3 - (i&7));
if (aux>0) vector[i] -= 2*vector[index[i]];
else vector[i] -= 2*aux;
}
}
break;
case INS_HEAVY:
fill_vec(vector, vector_length, iterations, WITH_BRANCHES, &nfunc, &rank);
}
#pragma omp master
{
branch_time = wtime() - branch_time;
if (btype == INS_HEAVY) {
printf("Number of matrix functions = %d\n", nfunc);
printf("Matrix order = %d\n", rank);
}
}
/* do the whole thing once more, but now without branches */
#pragma omp barrier
#pragma omp master
{
no_branch_time = wtime();
}
/* do actual branching */
switch (btype) {
case VECTOR_STOP:
case VECTOR_GO:
for (iter=0; iter<iterations; iter+=2) {
#pragma vector always
for (i=0; i<vector_length; i++) {
aux = -(3-(i&7));
vector[i] -= (vector[i] + aux);
}
for (i=0; i<vector_length; i++) {
aux = (3-(i&7));
vector[i] -= (vector[i] + aux);
}
}
break;
case NO_VECTOR:
for (iter=0; iter<iterations; iter+=2) {
#pragma vector always
for (i=0; i<vector_length; i++) {
aux = -(3-(i&7));
vector[i] -= (vector[index[i]]+aux);
}
#pragma vector always
for (i=0; i<vector_length; i++) {
aux = (3-(i&7));
vector[i] -= (vector[index[i]]+aux);
}
}
break;
case INS_HEAVY:
fill_vec(vector, vector_length, iterations, WITHOUT_BRANCHES, &nfunc, &rank);
}
#pragma omp master
{
no_branch_time = wtime() - no_branch_time;
ops = (double)vector_length * (double)iterations * (double)nthread;
if (btype == INS_HEAVY) ops *= rank*(rank*19 + 6);
else ops *= 4;
}
for (total = 0, i=0; i<vector_length; i++) total += vector[i];
} /* end of OPENMP parallel region */
/* compute verification values */
total_ref = ((vector_length%8)*(vector_length%8-8) + vector_length)/2*nthread;
if (total == total_ref) {
printf("Solution validates\n");
printf("Rate (Mops/s) with branches: %lf time (s): %lf\n",
ops/(branch_time*1.e6), branch_time);
printf("Rate (Mops/s) without branches: %lf time (s): %lf\n",
ops/(no_branch_time*1.e6), no_branch_time);
#ifdef VERBOSE
printf("Array sum = %d, reference value = %d\n", total, total_ref);
#endif
}
else {
printf("ERROR: array sum = %d, reference value = %d\n", total, total_ref);
}
exit(EXIT_SUCCESS);
}