def run_benchmark(name, arguments_list, precision, num_runs, platform, device):
binary = "./clblast_client_x" + name
# Loops over sub-benchmarks per benchmark
results = []
for arguments in arguments_list:
# Sets the arguments
constant_arguments = ["-warm_up", "-q", "-no_abbrv", "-cblas 0", "-cublas 0"]
common_arguments = ["-precision %d" % precision, "-runs %d" % num_runs]
opencl_arguments = ["-platform %d" % platform, "-device %d" % device]
all_arguments = opencl_arguments + common_arguments + constant_arguments
for name, value in arguments.items():
all_arguments.append("-" + name + " " + str(value))
# Calls the binary and parses the results
benchmark_output = utils.run_binary(binary, all_arguments)
result = utils.parse_results(benchmark_output)
# For half-precision: also runs single-precision for comparison
if precision == 16:
all_arguments = [arg if arg != "-precision 16" else "-precision 32" for arg in all_arguments]
benchmark_output = utils.run_binary(binary, all_arguments)
result_extra = utils.parse_results(benchmark_output)
for index in range(len(min(result, result_extra))):
result[index]["GBs_1_FP32"] = result_extra[index]["GBs_1"]
result[index]["GBs_2"] = result_extra[index]["GBs_2"]
result[index]["GFLOPS_1_FP32"] = result_extra[index]["GFLOPS_1"]
result[index]["GFLOPS_2"] = result_extra[index]["GFLOPS_2"]
results.extend(result)
return results
return BPF(text=text, cflags=flags)
if __name__ == "__main__":
signal.signal(signal.SIGINT, lambda x, y: sys.exit())
signal.signal(signal.SIGTERM, lambda x, y: sys.exit())
parser = argparse.ArgumentParser()
# parser.add_argument(
# "-p", "--pid", required=1, type=int, help="The PID of webserver.py"
# )
args = parser.parse_args()
# Check for root
if os.geteuid() != 0:
parser.error("Need superuser privileges to run")
bpfbox = BPFBoxRules()
try:
pid = run_binary(f"python3 {WEBSERVER_PATH}")
b = bpfbox.load_program(pid=pid)
os.kill(pid, signal.SIGUSR1)
except:
os.kill(pid, signal.SIGKILL)
sys.exit(-1)
while 1:
b.trace_print()
time.sleep(0.1)