def srumma(g_a, g_b, g_c, chunk_size, multiplier, g_counter):
# statically partition the task list among nprocs
task_list = get_task_list(chunk_size, multiplier)
task_id = ga.read_inc(g_counter, 0)
# the srumma algorithm, more or less
task_prev = task_list[task_id]
a_prev, a_nb_prev = ga.nbget(g_a, task_prev.alo, task_prev.ahi)
b_prev, b_nb_prev = ga.nbget(g_b, task_prev.blo, task_prev.bhi)
task_id = ga.read_inc(g_counter, 0)
while task_id < multiplier**3:
task_next = task_list[task_id]
a_next, a_nb_next = ga.nbget(g_a, task_next.alo, task_next.ahi)
b_next, b_nb_next = ga.nbget(g_b, task_next.blo, task_next.bhi)
ga.nbwait(a_nb_prev)
ga.nbwait(b_nb_prev)
result = np.dot(a_prev, b_prev)
ga.acc(g_c, result, task_prev.clo, task_prev.chi)
task_prev = task_next
a_prev, a_nb_prev = a_next, a_nb_next
b_prev, b_nb_prev = b_next, b_nb_next
task_id = ga.read_inc(g_counter, 0)
ga.nbwait(a_nb_prev)
ga.nbwait(b_nb_prev)
result = np.dot(a_prev, b_prev)
ga.acc(g_c, result, task_prev.clo, task_prev.chi)
ga.sync()
def srumma(g_a, g_b, g_c, chunk_size, multiplier):
# statically partition the task list among nprocs
task_list = get_task_list(chunk_size, multiplier)
ntasks = multiplier**3 // nproc
start = me * ntasks
stop = (me + 1) * ntasks
if me + 1 == nproc:
stop += multiplier**3 % nproc
# the srumma algorithm, more or less
task_prev = task_list[start]
a_prev, a_nb_prev = ga.nbget(g_a, task_prev.alo, task_prev.ahi)
b_prev, b_nb_prev = ga.nbget(g_b, task_prev.blo, task_prev.bhi)
for i in range(start + 1, stop):
task_next = task_list[i]
a_next, a_nb_next = ga.nbget(g_a, task_next.alo, task_next.ahi)
b_next, b_nb_next = ga.nbget(g_b, task_next.blo, task_next.bhi)
ga.nbwait(a_nb_prev)
ga.nbwait(b_nb_prev)
result = np.dot(a_prev, b_prev)
ga.acc(g_c, result, task_prev.clo, task_prev.chi)
task_prev = task_next
a_prev, a_nb_prev = a_next, a_nb_next
b_prev, b_nb_prev = b_next, b_nb_next
ga.nbwait(a_nb_prev)
ga.nbwait(b_nb_prev)
result = np.dot(a_prev, b_prev)
ga.acc(g_c, result, task_prev.clo, task_prev.chi)
ga.sync()
def srumma(g_a, g_b, g_c, chunk_size, multiplier, g_counter):
task_list = get_task_list(chunk_size, multiplier)
### get first integer from g_counter and assign to 'task_id'
# the srumma algorithm, more or less
task_prev = task_list[task_id]
a_prev,a_nb_prev = ga.nbget(g_a, task_prev.alo, task_prev.ahi)
b_prev,b_nb_prev = ga.nbget(g_b, task_prev.blo, task_prev.bhi)
### get next integer from g_counter and assign to 'task_id'
while task_id < multiplier**3:
task_next = task_list[task_id]
a_next,a_nb_next = ga.nbget(g_a, task_next.alo, task_next.ahi)
b_next,b_nb_next = ga.nbget(g_b, task_next.blo, task_next.bhi)
ga.nbwait(a_nb_prev)
ga.nbwait(b_nb_prev)
result = np.dot(a_prev,b_prev)
ga.acc(g_c, result, task_prev.clo, task_prev.chi)
task_prev = task_next
a_prev,a_nb_prev = a_next,a_nb_next
b_prev,b_nb_prev = b_next,b_nb_next
### get next integer from g_counter and assign to 'task_id'
ga.nbwait(a_nb_prev)
ga.nbwait(b_nb_prev)
result = np.dot(a_prev,b_prev)
ga.acc(g_c, result, task_prev.clo, task_prev.chi)
ga.sync()
def srumma(g_a, g_b, g_c, chunk_size, multiplier, g_counter):
# statically partition the task list among nprocs
task_list = get_task_list(chunk_size, multiplier)
task_id = ga.read_inc(g_counter, 0)
# the srumma algorithm, more or less
task_prev = task_list[task_id]
a_prev,a_nb_prev = ga.nbget(g_a, task_prev.alo, task_prev.ahi)
b_prev,b_nb_prev = ga.nbget(g_b, task_prev.blo, task_prev.bhi)
task_id = ga.read_inc(g_counter, 0)
while task_id < multiplier**3:
task_next = task_list[task_id]
a_next,a_nb_next = ga.nbget(g_a, task_next.alo, task_next.ahi)
b_next,b_nb_next = ga.nbget(g_b, task_next.blo, task_next.bhi)
ga.nbwait(a_nb_prev)
ga.nbwait(b_nb_prev)
result = np.dot(a_prev,b_prev)
ga.acc(g_c, result, task_prev.clo, task_prev.chi)
task_prev = task_next
a_prev,a_nb_prev = a_next,a_nb_next
b_prev,b_nb_prev = b_next,b_nb_next
task_id = ga.read_inc(g_counter, 0)
ga.nbwait(a_nb_prev)
ga.nbwait(b_nb_prev)
result = np.dot(a_prev,b_prev)
ga.acc(g_c, result, task_prev.clo, task_prev.chi)
ga.sync()
def srumma(g_a, g_b, g_c, chunk_size, multiplier):
# statically partition the task list among nprocs
task_list = get_task_list(chunk_size, multiplier)
ntasks = multiplier**3 // nproc
start = me*ntasks
stop = (me+1)*ntasks
if me+1 == nproc:
stop += multiplier**3 % nproc
# the srumma algorithm, more or less
task_prev = task_list[start]
a_prev,a_nb_prev = ga.nbget(g_a, task_prev.alo, task_prev.ahi)
b_prev,b_nb_prev = ga.nbget(g_b, task_prev.blo, task_prev.bhi)
for i in range(start+1,stop):
task_next = task_list[i]
a_next,a_nb_next = ga.nbget(g_a, task_next.alo, task_next.ahi)
b_next,b_nb_next = ga.nbget(g_b, task_next.blo, task_next.bhi)
ga.nbwait(a_nb_prev)
ga.nbwait(b_nb_prev)
result = np.dot(a_prev,b_prev)
ga.acc(g_c, result, task_prev.clo, task_prev.chi)
task_prev = task_next
a_prev,a_nb_prev = a_next,a_nb_next
b_prev,b_nb_prev = b_next,b_nb_next
ga.nbwait(a_nb_prev)
ga.nbwait(b_nb_prev)
result = np.dot(a_prev,b_prev)
ga.acc(g_c, result, task_prev.clo, task_prev.chi)
ga.sync()