Home | History | Annotate | Download | only in hardware_PerfCounterVerification 
      1 # Copyright (c) 2014 The Chromium OS Authors. All rights reserved.
      2 # Use of this source code is governed by a BSD-style license that can be
      3 # found in the LICENSE file.
      4 
      5 import os
      6 
      7 from autotest_lib.client.bin import test
      8 from autotest_lib.client.bin import utils
      9 from autotest_lib.client.common_lib import error
     10 
     11 import numpy
     12 
     13 import perf_lbr_verification
     14 import perf_verification
     15 import stats_utils
     16 
     17 
     18 INTEL_LBR_UARCHS = (
     19     # 'Broadwell',  # Waiting on kernel support.
     20     'Haswell',
     21     'IvyBridge',
     22     'SandyBridge')
     23 
     24 
     25 class hardware_PerfCounterVerification(test.test):
     26     """Verify perf counters count what we think they count.
     27 
     28     For cycles and instructions, we expect a strong correlation between
     29     the number of iterations of a "noploop" program and the number of
     30     cycles and instructions. For TLB misses, we expect a strong correlation
     31     between number of misses and number of iterations of a matching benchmark
     32     Each loop iteration should retire a constant number of additional
     33     instructions, and should take a nearly constant number of additional
     34     cycles or misses.
     35     """
     36 
     37     version = 1
     38     preserve_srcdir = True
     39 
     40     def initialize(self, perf_cmd='stat', events=('cycles', 'instructions')):
     41         self.job.require_gcc()
     42         self.perf_cmd = perf_cmd
     43         self.events = events
     44 
     45     def setup(self):
     46         os.chdir(self.srcdir)
     47         utils.make('clean')
     48         utils.make()
     49 
     50     def warmup(self):
     51         if self.perf_cmd == 'record -b':
     52             uarch = utils.get_intel_cpu_uarch()
     53             if uarch not in INTEL_LBR_UARCHS:
     54                 raise error.TestNAError('Unsupported microarchitecture.')
     55         unsupported_boards = ['gizmo']
     56         board = utils.get_board()
     57         if board in unsupported_boards:
     58             raise error.TestNAError('Unsupported board')
     59 
     60     def run_once(self, program, multiplier, **kwargs):
     61         program = os.path.join(self.srcdir, program)
     62         if self.perf_cmd == 'stat':
     63             self.facts = perf_verification.GatherPerfStats(
     64                     program, ','.join(self.events), multiplier)
     65         elif self.perf_cmd == 'record -b':
     66             branch = perf_lbr_verification.ReadBranchAddressesFile(
     67                     os.path.join(self.srcdir, 'noploop_branch.txt'))
     68             self.facts = perf_lbr_verification.GatherPerfBranchSamples(
     69                     program, branch, ','.join(self.events),
     70                     10000)
     71         else:
     72             raise error.TestError('Unrecognized perf_cmd')
     73 
     74 
     75     def postprocess_iteration(self):
     76         if self.perf_cmd == 'stat':
     77             dt = numpy.dtype([('loops', numpy.int)] +
     78                              [(e, numpy.int) for e in self.events])
     79         elif self.perf_cmd == 'record -b':
     80             dt = numpy.dtype([('loops', numpy.int),
     81                               ('branch_count', numpy.int)])
     82         arr = stats_utils.FactsToNumpyArray(self.facts, dt)
     83         results = {}
     84         is_tlb_benchmark = ('iTLB-misses' in dt.names or
     85                             'dTLB-misses' in dt.names)
     86         for y_var in dt.names:
     87             if y_var == 'loops': continue
     88             if y_var == 'cycles' and is_tlb_benchmark: continue
     89             (slope, intercept), r2 = stats_utils.LinearRegression(
     90                     arr['loops'], arr[y_var])
     91             prefix = y_var + '_'
     92             results[prefix+'slope'] = slope
     93             results[prefix+'intercept'] = intercept
     94             results[prefix+'r_squared'] = r2
     95             if y_var in ('dTLB-misses', 'iTLB-misses'):
     96                 misses_per_milion_cycles = [x[y_var] * 1.0e6 / x['cycles']
     97                                             for x in self.facts]
     98                 rvar = prefix+'misses_per_milion_cycles'
     99                 results[rvar] = numpy.max(misses_per_milion_cycles)
    100 
    101         # Output the standard Autotest way:
    102         self.write_perf_keyval(results)
    103         # ... And the CrOS-specific way:
    104         for k, v in results.iteritems():
    105           self.output_perf_value(k, v)
    106 
    107         if ('cycles' in self.events and not is_tlb_benchmark and
    108             results['cycles_r_squared'] < 0.996):
    109             raise error.TestFail('Poor correlation for cycles ~ loops')
    110         if ('instructions' in self.events and
    111             results['instructions_r_squared'] < 0.999):
    112             raise error.TestFail('Poor correlation for instructions ~ loops')
    113         if ('iTLB-misses' in self.events and
    114             results['iTLB-misses_r_squared'] < 0.999):
    115             raise error.TestFail('Poor correlation for iTLB-misses ~ loops')
    116         if ('dTLB-misses' in self.events and
    117             results['dTLB-misses_r_squared'] < 0.999):
    118             raise error.TestFail('Poor correlation for dTLB-misses ~ loops')
    119         if (self.perf_cmd == 'record -b' and
    120             results['branch_count_r_squared'] < 0.9999999):
    121             raise error.TestFail('Poor correlation for branch_count ~ loops')
    122