xref: /XiangShan/scripts/top-down/top_down.py (revision effccb7d09c4946a549cc139dc1b99ed9e726752)

from multiprocessing import Process, Manager
import threading
import os.path as osp
import os
import resource
import json
import argparse
import psutil
import numpy as np
import pandas as pd
import utils as u
import configs as cf
from draw import draw


def batch():
    paths = u.glob_stats(cf.stats_dir, fname='simulator_err.txt')

    manager = Manager()
    all_bmk_dict = manager.dict()

    semaphore = threading.Semaphore(psutil.cpu_count())

    # for workload, path in paths:
    def extract_and_post_process(gloabl_dict, workload, path):
        with semaphore:
            flag_file = osp.join(osp.dirname(path), 'simulator_out.txt')
            with open(flag_file, encoding='utf-8') as f:
                contents = f.read()
                if 'EXCEEDING CYCLE/INSTR LIMIT' not in contents and 'HIT GOOD TRAP' not in contents:
                    print('Skip unfinished job:', workload)
                    return

            print('Process finished job:', workload)

            d = u.xs_get_stats(path, cf.targets)
            if len(d):

                # add bmk and point after topdown processing
                segments = workload.split('_')
                if len(segments):
                    d['point'] = segments[-1]
                    d['workload'] = '_'.join(segments[:-1])
                    d['bmk'] = segments[0]

            gloabl_dict[workload] = d
        return

    jobs = [Process(target=extract_and_post_process, args=(
        all_bmk_dict, workload, path)) for workload, path in paths]
    _ = [p.start() for p in jobs]
    _ = [p.join() for p in jobs]

    df = pd.DataFrame.from_dict(all_bmk_dict, orient='index')
    df = df.sort_index()
    df = df.reindex(sorted(df.columns), axis=1)

    df = df.fillna(0)

    df.to_csv(cf.CSV_PATH, index=True)


def proc_input(wl_df: pd.DataFrame, js: dict, workload: str):
    # we implement the weighted metrics computation with the following formula:
    # weight = vec_weight matmul matrix_perf
    # (N, 1) = (1, W) matmul (W, N)
    # To make sure the matrix_perf is in the same order as the vec_weight,
    # we sort the matrix_perf by point
    assert isinstance(wl_df['point'][0], np.int64)
    wl_df = wl_df.sort_values(by=['point'])
    # We also sort the vec_weight by point
    wl_js = dict(js[workload])
    wl_df['cpi'] = 1.0 / wl_df['ipc']
    vec_weight = pd.DataFrame.from_dict(wl_js['points'], orient='index')

    # convert string index into int64
    vec_weight.index = vec_weight.index.astype(np.int64)
    # select only existing points
    vec_weight = vec_weight.loc[wl_df['point']]
    # make their sum equals 1.0
    vec_weight.columns = ['weight']

    vec_weight['weight'] = vec_weight['weight'].astype(np.float64)
    coverage = np.sum(vec_weight.values)
    vec_weight = vec_weight / coverage

    # Drop these auxiliary fields
    to_drop = {'bmk', 'point', 'workload', 'ipc'}
    to_drop = to_drop.intersection(set(wl_df.columns.to_list()))
    wl_df = wl_df.drop(to_drop, axis=1)

    weight_metrics = np.matmul(vec_weight.values.reshape(1, -1), wl_df.values)
    weight_metrics_df = pd.DataFrame(weight_metrics, columns=wl_df.columns)
    # We have to process coverage here to avoid apply weight on top of weight
    weight_metrics_df['coverage'] = coverage
    return weight_metrics_df.values, weight_metrics_df.columns


def proc_bmk(bmk_df: pd.DataFrame, js: dict):
    # Similar to per-input proc, we view the instruction count as the weight
    # and compute weighted metrics with matrix multiplication
    workloads = bmk_df['workload'].unique()
    metric_list = []
    for wl in workloads:
        metrics, cols = proc_input(bmk_df[bmk_df['workload'] == wl], js, wl)
        metric_list.append(metrics)
    metrics = np.concatenate(metric_list, axis=0)
    metrics = pd.DataFrame(metrics, columns=cols)

    input_dict = {}
    for workload in workloads:
        if workload.startswith(workload):
            input_dict[workload] = int(js[workload]['insts'])
    input_insts = pd.DataFrame.from_dict(
        input_dict, orient='index', columns=['insts'])
    # make their sum equals 1.0
    vec_weight = input_insts / np.sum(input_insts.values)
    weight_metric = np.matmul(vec_weight.values.reshape(1, -1), metrics.values)
    return weight_metric, metrics.columns


def compute_weighted_metrics():
    df = pd.read_csv(cf.CSV_PATH, index_col=0)
    bmks = df['bmk'].unique()
    with open(cf.JSON_FILE, 'r', encoding='utf-8') as f:
        js = json.load(f)
    weighted = {}
    for bmk in bmks:
        if bmk not in cf.spec_bmks['06']['int'] and cf.INT_ONLY:
            continue
        if bmk not in cf.spec_bmks['06']['float'] and cf.FP_ONLY:
            continue
        df_bmk = df[df['bmk'] == bmk]
        workloads = df_bmk['workload'].unique()
        n_wl = len(workloads)
        if n_wl == 1:
            metrics, cols = proc_input(df_bmk, js, workloads[0])
        else:
            metrics, cols = proc_bmk(df_bmk, js)
        weighted[bmk] = metrics[0]
    weighted_df = pd.DataFrame.from_dict(
        weighted, orient='index', columns=cols)
    if 'cpi' in weighted_df.columns:
        weighted_df = weighted_df.sort_values(by='cpi', ascending=False)
    else:
        weighted_df = weighted_df.sort_index()
    weighted_df.to_csv(cf.OUT_CSV)


if __name__ == '__main__':
    parser = argparse.ArgumentParser(usage='generate top-down results')
    parser.add_argument('-s', '--stat-dir', action='store', required=True,
                        help='stat output directory')
    parser.add_argument('-j', '--json', action='store', required=True,
                        help='specify json file', default='resources/spec06_rv64gcb_o2_20m.json')
    opt = parser.parse_args()
    cf.stats_dir = opt.stat_dir
    cf.JSON_FILE = opt.json
    if not osp.exists('results'):
        os.makedirs('results')
    if resource.getrlimit(resource.RLIMIT_NOFILE)[0] <= 8192:
        resource.setrlimit(resource.RLIMIT_NOFILE, (8192, 8192))

    batch()
    compute_weighted_metrics()
    draw()