def main():
ctx = cl.create_some_context()
prof_overhead, latency = perf.get_profiling_overhead(ctx)
print("command latency: %g s" % latency)
print("profiling overhead: %g s -> %.1f %%" %
(prof_overhead, 100 * prof_overhead / latency))
queue = cl.CommandQueue(
ctx, properties=cl.command_queue_properties.PROFILING_ENABLE)
print("empty kernel: %g s" % perf.get_empty_kernel_time(queue))
print("float32 add: %g GOps/s" % (perf.get_add_rate(queue) / 1e9))
for tx_type in [
perf.HostToDeviceTransfer, perf.DeviceToHostTransfer,
perf.DeviceToDeviceTransfer
]:
print("----------------------------------------")
print(tx_type.__name__)
print("----------------------------------------")
print("latency: %g s" % perf.transfer_latency(queue, tx_type))
for i in range(6, 28, 2):
bs = 1 << i
print("bandwidth @ %d bytes: %g GB/s" %
(bs, perf.transfer_bandwidth(queue, tx_type, bs) / 1e9))
def main():
ctx = cl.create_some_context()
prof_overhead, latency = perf.get_profiling_overhead(ctx)
print("command latency: %g s" % latency)
print("profiling overhead: %g s -> %.1f %%" % (
prof_overhead, 100*prof_overhead/latency))
queue = cl.CommandQueue(ctx, properties=cl.command_queue_properties.PROFILING_ENABLE)
print("empty kernel: %g s" % perf.get_empty_kernel_time(queue))
print("float32 add: %g GOps/s" % (perf.get_add_rate(queue)/1e9))
for tx_type in [
perf.HostToDeviceTransfer,
perf.DeviceToHostTransfer,
perf.DeviceToDeviceTransfer]:
print("----------------------------------------")
print(tx_type.__name__)
print("----------------------------------------")
print("latency: %g s" % perf.transfer_latency(queue, tx_type))
for i in range(6, 28, 2):
bs = 1<<i
print("bandwidth @ %d bytes: %g GB/s" % (
bs, perf.transfer_bandwidth(queue, tx_type, bs)/1e9))
def main():
ctx = cl.create_some_context()
prof_overhead, latency = perf.get_profiling_overhead(ctx)
print("command latency: %g s" % latency)
print("profiling overhead: {:g} s -> {:.1f} %".format(
prof_overhead, 100 * prof_overhead / latency))
queue = cl.CommandQueue(
ctx, properties=cl.command_queue_properties.PROFILING_ENABLE)
print("empty kernel: %g s" % perf.get_empty_kernel_time(queue))
print("float32 add: %g GOps/s" % (perf.get_add_rate(queue) / 1e9))
for tx_type in [
perf.HostToDeviceTransfer, perf.DeviceToHostTransfer,
perf.DeviceToDeviceTransfer
]:
print("----------------------------------------")
print(tx_type.__name__)
print("----------------------------------------")
print("latency: %g s" % perf.transfer_latency(queue, tx_type))
for i in range(6, 31, 2):
bs = 1 << i
try:
result = "%g GB/s" % (
perf.transfer_bandwidth(queue, tx_type, bs) / 1e9)
except Exception as e:
result = "exception: %s" % e.__class__.__name__
print("bandwidth @ %d bytes: %s" % (bs, result))
def profile_opencl_device(platform_id=0, device_id=0, verbose=False):
interact = Interactor(__file__.split(os.sep)[-1])
interact.set_verbosity(verbose)
platform = cl.get_platforms()[platform_id]
device = platform.get_devices()[device_id]
context = cl.Context([device])
queue = cl.CommandQueue(context, properties=cl.command_queue_properties.PROFILING_ENABLE)
interact('Collecting profiling info for the following device:')
interact('Platform:\t' + platform.name)
interact('Device:\t' + device.name)
interact('Version:\t' + device.version.strip())
interact('Please wait, this may take a while...')
prof_overhead, latency = clperf.get_profiling_overhead(context)
h2d_latency = clperf.transfer_latency(queue, clperf.HostToDeviceTransfer) * 1000
d2h_latency = clperf.transfer_latency(queue, clperf.DeviceToHostTransfer) * 1000
d2d_latency = clperf.transfer_latency(queue, clperf.DeviceToDeviceTransfer) * 1000
device_profile = {
'profiling overhead (time)': prof_overhead * 1000,
'profiling overhead (percentage)': f'{(100 * prof_overhead / latency):.2f}%',
'command latency': latency * 1000,
'host-to-device transfer latency': h2d_latency,
'device-to-host transfer latency': d2h_latency,
'device-to-device transfer latency': d2d_latency
}
for tx_type, tx_type_name in zip(
[clperf.HostToDeviceTransfer, clperf.DeviceToHostTransfer, clperf.DeviceToDeviceTransfer],
['host-device', 'device-host', 'device-device']
):
tx_type_bw = tx_type_name + ' bandwidth'
device_profile[tx_type_bw] = {}
for i in range(6, 31, 2):
bs = 1 << i
try:
bw = str(clperf.transfer_bandwidth(queue, tx_type, bs)/1e9) + ' GB/s'
except Exception as e:
bw = 'exception: ' + e.__class__.__name__
device_profile[tx_type_bw][f'{bs} bytes'] = bw
return device_profile
def run_perf_tests(self) -> None:
"""[summary]
Raises:
RuntimeError: [description]
"""
if self.queue is not None:
prof_overhead, latency = perf.get_profiling_overhead(self.ctx)
logging.debug("command latency: %g s" % latency)
logging.debug("profiling overhead: %g s -> %.1f %%" %
(prof_overhead, 100 * prof_overhead / latency))
logging.debug("empty kernel: %g s" %
perf.get_empty_kernel_time(self.queue))
logging.debug("float32 add: %g GOps/s" %
(perf.get_add_rate(self.queue) / 1e9))
for tx_type in [
perf.HostToDeviceTransfer, perf.DeviceToHostTransfer,
perf.DeviceToDeviceTransfer
]:
logging.debug("----------------------------------------")
logging.debug(tx_type.__name__)
logging.debug("----------------------------------------")
logging.debug("latency: %g s" %
perf.transfer_latency(self.queue, tx_type))
for i in range(6, 30, 2):
bs = 1 << i
try:
result = "%g GB/s" % (perf.transfer_bandwidth(
self.queue, tx_type, bs) / 1e9)
except Exception as e:
result = "exception: %s" % e.__class__.__name__
logging.debug("bandwidth @ %d bytes: %s" % (bs, result))
else:
raise RuntimeError("perf tests cannot be executed without a queue")
import pyopencl as cl
import pyopencl.characterize.performance as performance
context = cl.create_some_context()
queue = cl.CommandQueue(
context, properties=cl.command_queue_properties.PROFILING_ENABLE)
overhead, latency = performance.get_profiling_overhead(context)
print("\n\nCommand Latency: {} s".format(latency))
print("Profiling Overhead: {} s -> {}".format(overhead,
100 * overhead / latency))
# XXX Both these lines break the program on a Mac XXX
print("Empty Kernel: {} s".format(performance.get_empty_kernel_time(queue)))
print("Float32 Add: {} GOps/s\n".format(performance.get_add_rate(queue) / 1e9))
for transfer_type in [
performance.HostToDeviceTransfer, performance.DeviceToHostTransfer,
performance.DeviceToDeviceTransfer
]:
print("\n" + transfer_type.__name__)
print(" Latency: {0} s".format(
performance.transfer_latency(queue, transfer_type)))
for exponent in range(6, 28, 2):
bytes = 1 << exponent # This bit shift << operation does 'two to the exponent' (2^exponent)
print(" Bandwidth at {0} Bytes: {1} GB/s".format(
bytes,
performance.transfer_bandwidth(queue, transfer_type, bytes) / 1e9))
# XXX Find out more about pyopencl.characterize - why does this exist? XXX
context = cl.create_some_context()
queue = cl.CommandQueue(context, properties=cl.command_queue_properties.PROFILING_ENABLE)
overhead, latency = performance.get_profiling_overhead(context)
print("\n\nCommand Latency: {} s".format(latency))
print("Profiling Overhead: {} s -> {}".format(overhead, 100 * overhead / latency))
# XXX Both these lines break the program on a Mac XXX
print("Empty Kernel: {} s".format(performance.get_empty_kernel_time(queue)))
print("Float32 Add: {} GOps/s\n".format(performance.get_add_rate(queue) / 1e9))
for transfer_type in [
performance.HostToDeviceTransfer,
performance.DeviceToHostTransfer,
performance.DeviceToDeviceTransfer,
]:
print("\n" + transfer_type.__name__)
print(" Latency: {0} s".format(performance.transfer_latency(queue, transfer_type)))
for exponent in range(6, 28, 2):
bytes = 1 << exponent # This bit shift << operation does 'two to the exponent' (2^exponent)
print(
" Bandwidth at {0} Bytes: {1} GB/s".format(
bytes, performance.transfer_bandwidth(queue, transfer_type, bytes) / 1e9
)
)
