gs_run.py

import _problems
import qmctoolscl

import pyopencl as cl
import numpy as np
import pandas as pd
import argparse 

parser = argparse.ArgumentParser(prog='Global Size Experiment')
parser.add_argument('-q','--qrproblem',type=str,nargs='?',default="lattice",help='quasi-random problem type')
parser.add_argument('-t','--tag',type=str,nargs='?',default="debug",help='experiment tag')
parser.add_argument('-m','--np2',type=int,nargs='?',default=1,help='problems from n = 2^0,...,2^m')
parser.add_argument('-k','--dp2',type=int,nargs='?',default=2,help='problems from d = 2^0,...,2^k')
parser.add_argument('-v','--gsnp2',type=int,nargs='?',default=1,help='global size n from 2^0,...,2^v')
parser.add_argument('-w','--gsdp2',type=int,nargs='?',default=2,help='global size d from 2^0,...,2^w')
parser.add_argument('-a','--gsminnp2',type=int,nargs='?',default=1,help='minimum global size n from 2^a,...,2^v')
parser.add_argument('-b','--gsmindp2',type=int,nargs='?',default=2,help='global size d from 2^b,...,2^w')
parser.add_argument('-s','--skip',type=int,nargs='?',default=0,help='initial runs skipped for each problem, default 0')
parser.add_argument('-r','--runs',type=int,nargs='?',default=1,help='runs after skipped runs which count for each problem, default 1')
parser.add_argument('-p','--platform',type=int,nargs='?',default=0,help='platform index, default 0')
parser.add_argument('-d','--device',type=int,nargs='?',default=0,help='device index within platform, default 0')
parser.add_argument('-f','--force',type=bool,nargs='?',default=False,help='if True, overwrite previous tag results directory')
args = parser.parse_args()

experiment_dir = "gs_%s.%s"%(args.qrproblem,args.tag)
run_problem = getattr(_problems,"run_%s"%args.qrproblem)
n = 2**args.np2 
d = 2**args.dp2 
gs_n_pows2 = np.arange(args.gsminnp2,args.gsnp2+1) 
gs_d_pows2 = np.arange(args.gsmindp2,args.gsdp2+1)
gs_ns = 2**gs_n_pows2
gs_ds = 2**gs_d_pows2
num_gs_ns = len(gs_ns)
num_gs_ds = len(gs_ds)

_problems.remake_dir(experiment_dir,force=args.force)

kwargs_c = {"backend": "c"}

df_c_perf,df_c_process = np.zeros((num_gs_ns,num_gs_ds),dtype=np.float64),np.zeros((num_gs_ns,num_gs_ds),dtype=np.float64) 

print("logging C program %d trials: "%(args.skip+args.runs),end='',flush=True)
for t in range(args.skip+args.runs):
    print("%d, "%(t+1),end='',flush=True)
    c_perf,c_process = run_problem(n=n,d=d,kwargs=kwargs_c)
    if t>=args.skip:
        df_c_perf[:] += c_perf 
        df_c_process[:] += c_process
print("\n")
df_c_perf /= args.runs
df_c_process /= args.runs

pd.DataFrame(df_c_perf,index=gs_ns,columns=gs_ds).to_csv("%s/gs_%s.c_perf.csv"%(experiment_dir,args.qrproblem))
pd.DataFrame(df_c_process,index=gs_ns,columns=gs_ds).to_csv("%s/gs_%s.c_process.csv"%(experiment_dir,args.qrproblem))

df_cl_perf,df_cl_process = np.zeros((num_gs_ns,num_gs_ds),dtype=np.float64),np.zeros((num_gs_ns,num_gs_ds),dtype=np.float64)

platform = cl.get_platforms()[args.platform]
device = platform.get_devices()[args.device]
print("\nPlatform: %s\nDevice: %s\n"%(platform.name,device.name))

kwargs_cl = {
    "backend": "cl", 
    "wait": True}

print("logging gs_n_pows2 from %d up to %d and gs_d_pows2 from %d up to %d"%(args.gsminnp2,args.gsnp2,args.gsmindp2,args.gsdp2))
for i in range(num_gs_ns):
    print("gs_n_pow2 = %d: gs_d_pow2 = "%gs_n_pows2[i],end='',flush=True)
    gs_n = gs_ns[i]
    for j in range(num_gs_ds):
        print("%d, "%gs_d_pows2[j],end='',flush=True)
        gs_d = gs_ds[j]
        kwargs_cl["global_size"] = (1,gs_n,gs_d)
        for t in range(args.skip+args.runs):
            try:
                cl_perf,cl_process = run_problem(n=n,d=d,kwargs=kwargs_cl)
            except cl._cl.LogicError:
                cl_perf,cl_process = np.nan,np.nan
            if t>=args.skip:
                df_cl_perf[i,j] += cl_perf 
                df_cl_process[i,j] += cl_process
    print()
print()
df_cl_perf /= args.runs 
df_cl_process /= args.runs

pd.DataFrame(df_cl_perf,index=gs_ns,columns=gs_ds).to_csv("%s/gs_%s.cl_perf.csv"%(experiment_dir,args.qrproblem))
pd.DataFrame(df_cl_process,index=gs_ns,columns=gs_ds).to_csv("%s/gs_%s.cl_process.csv"%(experiment_dir,args.qrproblem))