static inline void init_comp_semaphores() {
// allocate completion semaphore
comp[MYTHREAD] = bupc_sem_alloc(0);
completion = malloc(THREADS * sizeof(bupc_sem_t*));
// wait for all threads to have allocated their semaphore
upc_barrier;
// copy all semaphore pointers to local memory
int i; for (i = 0; i < THREADS; i++) {
completion[i] = comp[i];
}
}
int main(int argc, char ** argv) {
long m, n; /* grid dimensions */
int i, j, iter; /* dummies */
int iterations; /* number of times to run the pipeline algorithm */
double pipeline_time, /* timing parameters */
avgtime, max_time;
double epsilon = 1.e-8; /* error tolerance */
double corner_val; /* verification value at top right corner of grid */
double *vector;/* array holding grid values */
long total_length; /* total required length to store grid values */
/*******************************************************************************
** process and test input parameters
********************************************************************************/
if(MYTHREAD == THREADS-1){
printf("Parallel Research Kernels version %s\n", PRKVERSION);
printf("UPC pipeline execution on 2D grid\n");
}
if (argc != 4){
if(MYTHREAD == THREADS-1){
printf("Usage: %s <# iterations> <first array dimension> ", *argv);
printf("<second array dimension>\n");
}
upc_global_exit(EXIT_FAILURE);
}
iterations = atoi(*++argv);
if (iterations < 1){
if(MYTHREAD == THREADS-1)
printf("ERROR: iterations must be >= 1 : %d \n",iterations);
upc_global_exit(EXIT_FAILURE);
}
m = atol(*++argv);
n = atol(*++argv);
if (m < 1 || n < 1){
if(MYTHREAD == THREADS-1)
printf("ERROR: grid dimensions must be positive: %d, %d \n", m, n);
upc_global_exit(EXIT_FAILURE);
}
if(MYTHREAD == THREADS-1){
printf("Number of threads = %d\n", THREADS);
printf("Grid sizes = %ld, %ld\n", m, n);
printf("Number of iterations = %d\n", iterations);
#if USE_BUPC_EXT
printf("Using Berkeley UPC extensions\n");
#endif
}
/*********************************************************************
** Allocate memory for input and output matrices
*********************************************************************/
#if USE_BUPC_EXT
bupc_sem_t *myflag = bupc_sem_alloc(BUPC_SEM_INTEGER | BUPC_SEM_MPRODUCER);
upc_barrier;
allflags[MYTHREAD] = myflag;
upc_barrier;
bupc_sem_t *mypeer = allflags[(MYTHREAD+1) % THREADS];
#endif
long segment_size = m / THREADS;
int leftover = m % THREADS;
int myoffsetx, sizex;
if(MYTHREAD < leftover){
myoffsetx = (segment_size + 1) * MYTHREAD;
sizex = segment_size + 1;
}else{
myoffsetx = (segment_size + 1) * leftover + segment_size * (MYTHREAD - leftover);
sizex = segment_size;
}
#if USE_BUPC_EXT
if(MYTHREAD != 0){
myoffsetx -= 1;
sizex += 1;
}
#endif
int sizey = n;
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);
in_arrays[MYTHREAD] = in_array;
double **in_array_private = shared_2d_array_to_private(in_array, sizex, sizey, myoffsetx, myoffsety);
if(MYTHREAD == 0)
current_max_line[MYTHREAD] = sizey;
else
current_max_line[MYTHREAD] = 0;
upc_barrier;
/*********************************************************************
** Initialize the matrices
*********************************************************************/
/* clear the array */
for (j=0; j<n; j++)
for (i=myoffsetx; i<myoffsetx + sizex; i++)
ARRAY(i, j) = 0.0;
/* set boundary values (bottom and left side of grid */
if(MYTHREAD == 0)
for (j=0; j<n; j++)
ARRAY(0, j) = (double) j;
for (i=myoffsetx; i<myoffsetx + sizex; i++)
ARRAY(i, 0) = (double) i;
upc_barrier;
for (iter = 0; iter<=iterations; iter++){
/* start timer after a warmup iteration */
if (iter == 1)
pipeline_time = wtime();
if(MYTHREAD == 0)
debug("start it %d, %f", iter, ARRAY(0, 0));
if(MYTHREAD != THREADS - 1) // Send the element in line 0
in_arrays[MYTHREAD + 1][0][myoffsetx + sizex -1] = ARRAY(myoffsetx + sizex - 1, 0);
for (j=1; j<n; j++) {
#if USE_BUPC_EXT
if(MYTHREAD > 0){
bupc_sem_wait(myflag);
}
for (i=myoffsetx+1; i<myoffsetx + sizex; i++)
ARRAY(i, j) = ARRAY(i-1, j) + ARRAY(i, j-1) - ARRAY(i-1, j-1);
if(MYTHREAD != THREADS - 1){
in_arrays[MYTHREAD + 1][j][myoffsetx + sizex -1] = ARRAY(myoffsetx + sizex - 1, j);
bupc_sem_post(mypeer);
}
#else
while(j > current_max_line[MYTHREAD]) // Normally not necessary: bupc_poll();
;
if(MYTHREAD > 0)
ARRAY(myoffsetx, j) = in_arrays[MYTHREAD - 1][j][myoffsetx-1] + ARRAY(myoffsetx, j-1) - in_arrays[MYTHREAD-1][j-1][myoffsetx-1];
for (i=myoffsetx+1; i<myoffsetx + sizex; i++)
ARRAY(i, j) = ARRAY(i-1, j) + ARRAY(i, j-1) - ARRAY(i-1, j-1);
if(MYTHREAD < THREADS - 1)
current_max_line[MYTHREAD+1] = j;
#endif
}
/* copy top right corner value to bottom left corner to create dependency; we
need a barrier to make sure the latest value is used. This also guarantees
that the flags for the next iteration (if any) are not getting clobbered */
if(MYTHREAD == 0)
current_max_line[MYTHREAD] = sizey;
else
current_max_line[MYTHREAD] = 0;
if(MYTHREAD == THREADS - 1){
in_arrays[0][0][0] = -ARRAY(m-1, n-1);
}
upc_barrier;
}
pipeline_time = wtime() - pipeline_time;
times[MYTHREAD] = pipeline_time;
upc_barrier;
// Compute max_time
if(MYTHREAD == THREADS - 1){
max_time = times[MYTHREAD];
for(i=1; i<THREADS; i++){
if(max_time < times[i])
max_time = times[i];
}
}
/*******************************************************************************
** Analyze and output results.
********************************************************************************/
/* verify correctness, using top right value; */
if( MYTHREAD == THREADS - 1){
corner_val = (double)((iterations+1)*(n+m-2));
if (fabs(ARRAY(m-1,n-1)-corner_val)/corner_val > epsilon) {
printf("ERROR: checksum %lf does not match verification value %lf\n",
ARRAY(m-1, n-1), corner_val);
exit(EXIT_FAILURE);
}
#if VERBOSE
printf("checksum %lf verification value %lf\n",
ARRAY(m-1, n-1), corner_val);
printf("Solution validates; verification value = %lf\n", corner_val);
#else
printf("Solution validates\n");
#endif
avgtime = max_time/iterations;
printf("Rate (MFlops/s): %lf Avg time (s): %lf\n",
1.0E-06 * 2 * ((double)(m-1)*(double)(n-1))/avgtime, avgtime);
exit(EXIT_SUCCESS);
}
}