ampere_research/pytorch/run.py

#! /bin/python3

import perf_stat

import argparse
import os, sys
import subprocess, signal
import json

parser = argparse.ArgumentParser()
parser.add_argument('arch')
parser.add_argument('matrix_file')
parser.add_argument('iterations', type=int)
parser.add_argument('baseline_time_s', type=int)
parser.add_argument('baseline_delay_s', type=int)
parser.add_argument('--perf', action='store_true')
parser.add_argument('--power', action='store_true')
parser.add_argument('-d', '--debug', action='store_true')
args = parser.parse_args()

program_altra = [
        'apptainer', 'run', 'pytorch-altra.sif', '-c',
        'numactl --cpunodebind=0 --membind=0 '
        + f'python spmv.py {args.matrix_file} {args.iterations}']
perf = ['perf', 'stat']
perf_altra = [['-d', '-d'],
        ['-M', 'branch_misprediction_ratio'],
        ['-M', 'dtlb_walk_ratio,itlb_walk_ratio'],
        ['-M', 'l1d_cache_miss_ratio,l1i_cache_miss_ratio'],
        ['-M', 'l2_cache_miss_ratio,l2_tlb_miss_ratio,ll_cache_read_miss_ratio']]

def baseline_power(baseline_time_s: int) -> list:
    power_process = subprocess.Popen(['./power.py', '-s', str(baseline_time_s)],
            stdout=subprocess.PIPE, stderr=subprocess.DEVNULL, text=True)
    return [float(x) for x in power_process.communicate()[0].strip().split('\n')]

def run_program(program: list[str]) -> tuple[dict, str]:
    process = subprocess.run(program,
            stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True)
    #print(json.loads(process.stdout))
    #print(process.stderr)
    return (json.loads(process.stdout), process.stderr)

result = dict()
result['architecture'] = args.arch
result['iterations'] = args.iterations
result['baseline_time_s'] = args.baseline_time_s
result['baseline_delay_s'] = args.baseline_delay_s

if args.power is True:
    result['power_before'] = baseline_power(args.baseline_time_s)
    if args.debug:
        print(result)

    run_program(program_altra) # Warmup
    power_process = subprocess.Popen(['./power.py'],
            stdout=subprocess.PIPE, stderr=subprocess.DEVNULL, text=True)

    if args.arch == 'altra':
        result = result | run_program(program_altra)[0]
    elif args.arch == 'x86':
        print("Arch not implemented yet!")
        exit(1)
    else:
        print("Unrecognized arch!")
        exit(1)

    power_process.send_signal(signal.SIGINT)
    if args.debug:
        print(result)

    result['power'] = [float(x) for x in power_process.communicate()[0].strip().split('\n')]
    if args.debug:
        print(result)
        #print(len(result['power']))
        #print(sum(result['power']) / len(result['power']))

    import time
    time.sleep(args.baseline_delay_s)
    result['power_after'] = baseline_power(args.baseline_time_s)
    if args.debug:
        print(result)

if args.perf is True:
    if args.arch == 'altra':
        for perf_args in perf_altra:
            output = run_program(perf + perf_args + program_altra)[1]
            print(output, file=sys.stderr)
            result = result | perf_stat.parse_output(output, perf_stat.CPU.ALTRA)
            if args.debug:
                print(result)
    elif args.arch == 'x86':
        print("no implement")
        exit(1)
    else:
        print("Unrecognized arch!")
        exit(1)

    result = result | perf_stat.derive_stats(result)

    if args.debug:
        print(result)

print(json.dumps(result))

#arch = subprocess.run(['uname', '-m'], stdout=subprocess.PIPE, text=True).stdout.strip()
#baseline = subprocess.run(
#        ['./power.sh', args.baseline_time_s],
#        stdout=subprocess.PIPE, stderr=subprocess.STDOUT, text=True)
#print(baseline)
#for line in baseline.stdout.split('\n'):
#    print("line")
#    print(line)

#os.path.basename(args.matrix_file)

#subprocess.run(
#    ['apptainer', 'run', 'pytorch-altra.sif', '-c',
#        f'"numactl --cpunodebind=0 --membind=0 python spmv.py {args.matrix_file} {args.iterations}"'
#    ],
#    stdout=subprocess.PIPE, stderr=subprocess.STDOUT, text=True)