uint64_t accumulate (size_t sndlen,
int mytask,
int origin,
int target)
{
TRACE_ERR((stderr, "(%u) Do test ... sndlen = %zu\n", mytask, sndlen));
unsigned i;
uint64_t t1 = GetTimeBase();
double scale = 1.0;
if (mytask == origin)
{
for (i = 0; i < ITERATIONS; i++)
{
//fprintf(stderr, "(%u) Starting Iteration %d of size %zu dstaddr %p\n", mytask, i, sndlen, rcvbuf);
A1_Get ( target,
srcbuf,
dstbuf,
sndlen * sizeof(double) );
A1_Flush (target);
}
}
A1_AllFence();
MPI_Barrier(MPI_COMM_WORLD);
uint64_t t2 = GetTimeBase();
return ((t2 - t1) / ITERATIONS);
}
double time_get(double *src_buf, double *dst_buf, int chunk, int loop,
int proc, int levels)
{
int i, bal = 0;
int stride[2];
int count[2];
int stride_levels = levels;
double *tmp_buf, *tmp_buf_ptr;
double start_time, stop_time, total_time = 0;
stride[0] = SIZE * sizeof(double);
count[0] = chunk * sizeof(double); count[1] = chunk;
if(CHECK_RESULT) {
tmp_buf = (double *)malloc(SIZE * SIZE * sizeof(double));
assert(tmp_buf != NULL);
fill_array(tmp_buf, SIZE*SIZE, proc);
tmp_buf_ptr = tmp_buf;
}
start_time = TIMER();
for(i=0; i<loop; i++) {
if(levels)
A1_GetS(proc, stride_levels, count, src_buf, stride, dst_buf, stride);
else
A1_Get(proc, src_buf, dst_buf, count[0]);
if(CHECK_RESULT) {
sprintf(check_type, "A1_GetS:");
check_result(tmp_buf_ptr, dst_buf, stride, count, stride_levels);
}
/* prepare next src and dst ptrs: avoid cache locality */
if(bal == 0) {
src_buf += 128;
dst_buf += 128;
if(CHECK_RESULT) tmp_buf_ptr += 128;
bal = 1;
} else {
src_buf -= 128;
dst_buf -= 128;
if(CHECK_RESULT) tmp_buf_ptr -= 128;
bal = 0;
}
}
stop_time = TIMER();
total_time = (stop_time - start_time);
if(CHECK_RESULT) free(tmp_buf);
if(total_time == 0.0){
total_time=0.000001; /* workaround for inaccurate timers */
warn_accuracy++;
}
return(total_time/loop);
}