def dispatch_kernel(self):
start_time, done_events = self.kernel.dispatch(
0,
0,
self.ctxs,
self.cmd_qs,
C_cpu=self.thread_coarsening_cpu,
C_gpu=self.thread_coarsening_gpu)
fw.host_synchronize(self.cmd_qs, done_events)
end_time = datetime.datetime.now()
seconds = (end_time - start_time).total_seconds()
return seconds
kernel.get_data_types_and_shapes()
logging.debug("Dispatching Kernel...")
lw_callback = False
start_time, done_events = kernel.dispatch(0,
0,
ctxs,
cmd_qs,
C_cpu=1,
C_gpu=1)
logging.debug("Waiting for events... \n")
fw.host_synchronize(cmd_qs, done_events)
sched_end_time = datetime.datetime.now()
seconds = (sched_end_time - sched_start_time).total_seconds()
# print("%s with partition %d and dataset %d ran %fs" % (info['name'], partition, dataset, seconds))
if args.profile:
dump = os.path.join("profiling", "dumps")
if not os.path.exists(dump):
os.makedirs(dump)
with open(os.path.join(dump, name), "w") as f:
f.write(str(seconds))
if not lw_callback:
dump_dev = fw.dump_device_history()
def select_main(kernels, select=baseline_select):
cmd_qs, ctxs, gpus, cpus = fw.host_initialize(4, 2)
aCPU, aGPU = fw.nCPU, fw.nGPU
cpu_q, gpu_q, mixed_q, mixedg_q, mixedc_q = [], [], [], [], []
num_dispatched = 0
num_kernels = len(kernels)
rCPU, rGPU = 0, 0
num_global_work_items_max = 0
for i in range(len(kernels)):
kernels[i].build_kernel(gpus, cpus, ctxs)
kernels[i].random_data()
p = kernels[i].partition
num_global_work_items = kernels[i].get_num_global_work_items()
if num_global_work_items > num_global_work_items_max:
num_global_work_items_max = num_global_work_items
if p == 0:
heapq.heappush(cpu_q, (p, -num_global_work_items, i))
rCPU += 1
elif p == 10:
heapq.heappush(gpu_q, (-p, -num_global_work_items, i))
rGPU += 1
elif p >= 5:
heapq.heappush(mixedg_q, (abs(p-5), -num_global_work_items, i))
heapq.heappush(mixed_q, (abs(p-5), -num_global_work_items, i))
rCPU += 1
rGPU += 1
else:
heapq.heappush(mixedc_q, (abs(p-5), -num_global_work_items, i))
heapq.heappush(mixed_q, (abs(p-5), -num_global_work_items, i))
rCPU +=1
rGPU +=1
logging.debug( "Task Queue Stats")
logging.debug(rCPU)
logging.debug(rGPU)
logging.debug( "CPU " + str(len(cpu_q)))
logging.debug( "GPU " + str(len(gpu_q)))
logging.debug( "Mixed " + str(len(mixed_q)))
logging.debug( "Mixed CPU " +str(len(mixedc_q)))
logging.debug( "Mixed GPU " +str(len(mixedg_q)))
events = []
now, soon = None, None
while (len(cpu_q) > 0 or len(gpu_q) > 0 or len(mixed_q) > 0 or num_dispatched != num_kernels) and (len(cpu_q) > 0 or len(gpu_q) > 0 or
len(mixedc_q) > 0 or len(mixedg_q) > 0 or num_dispatched != num_kernels):
logging.debug( "READY DEVICES")
logging.debug( "GPU " + str(len(fw.ready_queue['gpu'])))
logging.debug( "CPU " + str(len(fw.ready_queue['cpu'])))
logging.debug( "Number of tasks left")
logging.debug( "Mixed Queue " + str(len(mixed_q)))
logging.debug( "CPU Queue " + str(len(cpu_q)))
logging.debug( "GPU Queue " + str(len(gpu_q)))
logging.debug( "Mixed CPU " + str(len(mixedc_q)))
logging.debug( "Mixed GPU " + str(len(mixedg_q)))
logging.debug( "Number of available devices (CPU and GPU) " + str(fw.nCPU) + " " + str(fw.nGPU))
if fw.nCPU > 0 or fw.nGPU > 0:
logging.debug( "Entering selection phase")
if soon == None:
if select is baseline_select or select is look_ahead_select:
now, soon = select(kernels, M=mixed_q, G=gpu_q, C=cpu_q)
else:
now, soon = select(kernels, M1=mixedc_q, M2=mixedg_q, G=gpu_q, C=cpu_q, feat_max=num_global_work_items_max, threshold=0.4)
else:
now, soon = soon, None
if now == None:
logging.debug( "Searching for available devices")
continue
i, p = now
if fw.nCPU > rCPU and fw.nGPU > rGPU:
logging.debug( "DISPATCH MULTIPLE")
g_factor = fw.nGPU if rGPU == 0 else fw.nGPU / rGPU
c_factor = fw.nCPU if rCPU == 0 else fw.nCPU / rCPU
free_gpus, free_cpus = [], []
# try:
while fw.test_and_set(0,1):
pass
if p == 0:
for j in range(c_factor):
free_cpus.append(fw.ready_queue['cpu'].popleft())
elif p == 10:
for j in range(g_factor):
free_gpus.append(fw.ready_queue['gpu'].popleft())
else:
for j in range(c_factor):
free_cpus.append(fw.ready_queue['cpu'].popleft())
for j in range(g_factor):
free_gpus.append(fw.ready_queue['gpu'].popleft())
fw.rqlock[0] = 0
# except:
# logging.debug( free_cpus, free_gpus, framework.ready_queue, framework.nCPU, framework.nGPU, c_factor, g_factor
if kernels[i].partition == 0:
rCPU -= 1
elif kernels[i].partition == 10:
rGPU -= 1
else:
rCPU -= 1
rGPU -= 1
kernels[i].partition = p
# kernels[i].build_kernel(gpus, cpus, ctxs)
# kernels[i].random_data()
logging.debug( "Dispatching Multiple " + str(kernels[i].name))
start_time, done_events = kernels[i].dispatch_multiple(free_gpus, free_cpus, ctxs, cmd_qs)
events.extend(done_events)
num_dispatched += 1
# if False:
# pass
else:
logging.debug( "DISPATCH")
cpu, gpu = -1, -1
if p == 0:
while fw.test_and_set(0, 1):
pass
cpu = fw.ready_queue['cpu'].popleft()
fw.rqlock[0] = 0
elif p == 10:
while fw.test_and_set(0, 1):
pass
gpu = fw.ready_queue['gpu'].popleft()
fw.rqlock[0] = 0
else:
while fw.test_and_set(0, 1):
pass
cpu = fw.ready_queue['cpu'].popleft()
gpu = fw.ready_queue['gpu'].popleft()
fw.rqlock[0] = 0
if kernels[i].partition == 0:
rCPU -= 1
elif kernels[i].partition == 10:
rGPU -= 1
else:
rCPU -=1
rGPU -=1
kernels[i].partition = p
logging.debug( "Dispatching " + str(kernels[i].name) + " with partition class " + str(kernels[i].partition))
start_time, done_events = kernels[i].dispatch(gpu, cpu, ctxs, cmd_qs)
events.extend(done_events)
num_dispatched +=1
else:
logging.debug( "Devices unavailable")
fw.host_synchronize(cmd_qs, events)
return fw.dump_device_history()