#!/usr/bin/python2

"""
python-libpfm4 provides python bindings to the libpfm4
library and the perf_event kernel subsystem. This
script builds on them to provide a *stat like interface
to CPU performance counters.

Run as: ./cpistat -c cpulist -e eventlist

Depends on libpfm4: http://perfmon2.sf.net/

git://perfmon2.git.sourceforge.net/gitroot/perfmon2/libpfm4
"""

import sys, os, optparse, time, struct, perfmon

if __name__ == '__main__':
    parser = optparse.OptionParser()
    parser.add_option('-e', '--events', help='Events to use',
                       action='store', dest='events')
    parser.add_option('-c', '--cpulist', help='CPUs to monitor',
                       action='store', dest='cpulist')
    parser.set_defaults(events='PERF_COUNT_HW_CPU_CYCLES,' +
                               'PERF_COUNT_HW_INSTRUCTIONS')
    (options, args) = parser.parse_args()

    show_per_cpu = False
    if not options.cpulist:
        ncpus = os.sysconf('SC_NPROCESSORS_ONLN')
        cpus = range(0, ncpus)
    else:
        cpus = options.cpulist.split(',')
        cpus = [ int(c) for c in cpus ]
        show_per_cpu = True

    if options.events:
        events = options.events.split(',')
    else:
        raise ValueError('You need to specify events to monitor')

    s = perfmon.SystemWideSession(cpus, events)

    s.start()
    # Measuring loop
    interval = 1
    iters = -1
    infinite = True
    if len(args) == 2:
        interval = int(args[0])
        iters = int(args[1])
        infinite = False

    delta = {}
    last = {}
    sum = {}
    for e in events:
        delta[e] = {}
        last[e] = {}
        sum[e] = {}
        for c in cpus:
            delta[e][c] = 0
            last[e][c] = 0

    while infinite or iters:
        for i in range(0, len(events)):
          e = events[i]
          sum[e] = 0
          for c in cpus:
              count = struct.unpack('L', s.read(c, i))[0]
              delta[e][c] = count - last[e][c]
              last[e][c] = count
              if show_per_cpu:
                  print '''CPU%d: %s\t%lu''' % (c, e, delta[e][c])
              sum[e] += delta[e][c]

        cycles = sum['PERF_COUNT_HW_CPU_CYCLES']
        instructions = sum['PERF_COUNT_HW_INSTRUCTIONS']
        CPI = cycles * 1.0/instructions
        print ('cycles: %12lu, instructions: %12lu, CPI: %2.4f'
               % (cycles, instructions, CPI))
        sys.stdout.flush()
        time.sleep(interval)
        iters = iters - 1