import collections import contextlib import cProfile import inspect import gc import multiprocessing import os import random import sys import time import unittest import warnings import zlib from functools import lru_cache from io import StringIO from unittest import result, runner, signals, suite, loader, case from .loader import TestLoader from numba.core import config try: from multiprocessing import TimeoutError except ImportError: from Queue import Empty as TimeoutError def make_tag_decorator(known_tags): """ Create a decorator allowing tests to be tagged with the *known_tags*. """ def tag(*tags): """ Tag a test method with the given tags. Can be used in conjunction with the --tags command-line argument for runtests.py. """ for t in tags: if t not in known_tags: raise ValueError("unknown tag: %r" % (t,)) def decorate(func): if (not callable(func) or isinstance(func, type) or not func.__name__.startswith('test_')): raise TypeError("@tag(...) should be used on test methods") try: s = func.tags except AttributeError: s = func.tags = set() s.update(tags) return func return decorate return tag # Chances are the next queried class is the same as the previous, locally 128 # entries seems to be fastest. # Current number of test classes can be found with: # $ ./runtests.py -l|sed -e 's/\(.*\)\..*/\1/'|grep ^numba|sort|uniq|wc -l # as of writing it's 658. @lru_cache(maxsize=128) def _get_mtime(cls): """ Gets the mtime of the file in which a test class is defined. """ return str(os.path.getmtime(inspect.getfile(cls))) def cuda_sensitive_mtime(x): """ Return a key for sorting tests bases on mtime and test name. For CUDA tests, interleaving tests from different classes is dangerous as the CUDA context might get reset unexpectedly between methods of a class, so for CUDA tests the key prioritises the test module and class ahead of the mtime. """ cls = x.__class__ key = _get_mtime(cls) + str(x) from numba.cuda.testing import CUDATestCase if CUDATestCase in cls.mro(): key = "%s.%s %s" % (str(cls.__module__), str(cls.__name__), key) return key def parse_slice(useslice): """Parses the argument string "useslice" as a shard index and number and returns a function that filters on those arguments. i.e. input useslice="1:3" leads to output something like `lambda x: zlib.crc32(x) % 3 == 1`. """ if callable(useslice): return useslice if not useslice: return lambda x: True try: (index, count) = useslice.split(":") index = int(index) count = int(count) except Exception: msg = ( "Expected arguments shard index and count to follow " "option `-j i:t`, where i is the shard number and t " "is the total number of shards, found '%s'" % useslice) raise ValueError(msg) if count == 0: return lambda x: True elif count < 0 or index < 0 or index >= count: raise ValueError("Sharding out of range") else: def decide(test): func = getattr(test, test._testMethodName) if "always_test" in getattr(func, 'tags', {}): return True return abs(zlib.crc32(test.id().encode('utf-8'))) % count == index return decide class TestLister(object): """Simply list available tests rather than running them.""" def __init__(self, useslice): self.useslice = parse_slice(useslice) def run(self, test): result = runner.TextTestResult(sys.stderr, descriptions=True, verbosity=1) self._test_list = _flatten_suite(test) masked_list = list(filter(self.useslice, self._test_list)) self._test_list.sort(key=cuda_sensitive_mtime) for t in masked_list: print(t.id()) print('%d tests found. %s selected' % (len(self._test_list), len(masked_list))) return result class SerialSuite(unittest.TestSuite): """A simple marker to make sure tests in this suite are run serially. Note: As the suite is going through internals of unittest, it may get unpacked and stuffed into a plain TestSuite. We need to set an attribute on the TestCase objects to remember they should not be run in parallel. """ def addTest(self, test): if not isinstance(test, unittest.TestCase): # It's a sub-suite, recurse for t in test: self.addTest(t) else: # It's a test case, mark it serial test._numba_parallel_test_ = False super(SerialSuite, self).addTest(test) class BasicTestRunner(runner.TextTestRunner): def __init__(self, useslice, **kwargs): runner.TextTestRunner.__init__(self, **kwargs) self.useslice = parse_slice(useslice) def run(self, test): run = list(filter(self.useslice, _flatten_suite(test))) run.sort(key=cuda_sensitive_mtime) wrapped = unittest.TestSuite(run) return super(BasicTestRunner, self).run(wrapped) # "unittest.main" is really the TestProgram class! # (defined in a module named itself "unittest.main"...) class NumbaTestProgram(unittest.main): """ A TestProgram subclass adding the following options: * a -R option to enable reference leak detection * a --profile option to enable profiling of the test run * a -m option for parallel execution * a -l option to (only) list tests Currently the options are only added in 3.4+. """ refleak = False profile = False multiprocess = False useslice = None list = False tags = None exclude_tags = None random_select = None random_seed = 42 def __init__(self, *args, **kwargs): topleveldir = kwargs.pop('topleveldir', None) kwargs['testLoader'] = TestLoader(topleveldir) # HACK to force unittest not to change warning display options # (so that NumbaWarnings don't appear all over the place) sys.warnoptions.append(':x') self.nomultiproc = kwargs.pop('nomultiproc', False) super(NumbaTestProgram, self).__init__(*args, **kwargs) def _getParentArgParser(self): # NOTE: this hook only exists on Python 3.4+. The options won't be # added in earlier versions (which use optparse - 3.3 - or getopt() # - 2.x). parser = super(NumbaTestProgram, self)._getParentArgParser() if self.testRunner is None: parser.add_argument('-R', '--refleak', dest='refleak', action='store_true', help='Detect reference / memory leaks') parser.add_argument('-m', '--multiprocess', dest='multiprocess', nargs='?', type=int, const=multiprocessing.cpu_count(), help='Parallelize tests') parser.add_argument('-l', '--list', dest='list', action='store_true', help='List tests without running them') parser.add_argument('--tags', dest='tags', type=str, help='Comma-separated list of tags to select ' 'a subset of the test suite') parser.add_argument('--exclude-tags', dest='exclude_tags', type=str, help='Comma-separated list of tags to de-select ' 'a subset of the test suite') parser.add_argument('--random', dest='random_select', type=float, help='Random proportion of tests to select') parser.add_argument('--profile', dest='profile', action='store_true', help='Profile the test run') parser.add_argument('-j', '--slice', dest='useslice', nargs='?', type=str, const="None", help='Shard the test sequence') def git_diff_str(x): if x != 'ancestor': raise ValueError("invalid option for --gitdiff") return x parser.add_argument('-g', '--gitdiff', dest='gitdiff', type=git_diff_str, default=False, nargs='?', help=('Run tests from changes made against ' 'origin/release0.59 ' 'as identified by `git diff`. ' 'If set to "ancestor", the diff compares ' 'against the common ancestor.')) return parser def _handle_tags(self, argv, tagstr): found = None for x in argv: if tagstr in x: if found is None: found = x else: raise ValueError("argument %s supplied repeatedly" % tagstr) if found is not None: posn = argv.index(found) try: if found == tagstr: # --tagstr tag_args = argv[posn + 1].strip() argv.remove(tag_args) else: # --tagstr= if '=' in found: tag_args = found.split('=')[1].strip() else: raise AssertionError('unreachable') except IndexError: # at end of arg list, raise msg = "%s requires at least one tag to be specified" raise ValueError(msg % tagstr) # see if next arg is "end options" or some other flag if tag_args.startswith('-'): raise ValueError("tag starts with '-', probably a syntax error") # see if tag is something like "=" which is likely a syntax # error of form `--tags =`, note the space prior to `=`. if '=' in tag_args: msg = "%s argument contains '=', probably a syntax error" raise ValueError(msg % tagstr) attr = tagstr[2:].replace('-', '_') setattr(self, attr, tag_args) argv.remove(found) def parseArgs(self, argv): if '-l' in argv: argv.remove('-l') self.list = True super(NumbaTestProgram, self).parseArgs(argv) # If at this point self.test doesn't exist, it is because # no test ID was given in argv. Use the default instead. if not hasattr(self, 'test') or not self.test.countTestCases(): self.testNames = (self.defaultTest,) self.createTests() if self.tags: tags = [s.strip() for s in self.tags.split(',')] self.test = _choose_tagged_tests(self.test, tags, mode='include') if self.exclude_tags: tags = [s.strip() for s in self.exclude_tags.split(',')] self.test = _choose_tagged_tests(self.test, tags, mode='exclude') if self.random_select: self.test = _choose_random_tests(self.test, self.random_select, self.random_seed) if self.gitdiff is not False: self.test = _choose_gitdiff_tests( self.test, use_common_ancestor=(self.gitdiff == 'ancestor'), ) if self.verbosity <= 0: # We aren't interested in informational messages / warnings when # running with '-q'. self.buffer = True def _do_discovery(self, argv, Loader=None): # Disable unittest's implicit test discovery when parsing # CLI arguments, as it can select other tests than Numba's # (e.g. some test_xxx module that may happen to be directly # reachable from sys.path) return def runTests(self): if self.refleak: self.testRunner = RefleakTestRunner if not hasattr(sys, "gettotalrefcount"): warnings.warn("detecting reference leaks requires a debug build " "of Python, only memory leaks will be detected") elif self.list: self.testRunner = TestLister(self.useslice) elif self.testRunner is None: self.testRunner = BasicTestRunner(self.useslice, verbosity=self.verbosity, failfast=self.failfast, buffer=self.buffer) if self.multiprocess and not self.nomultiproc: if self.multiprocess < 1: msg = ("Value specified for the number of processes to use in " "running the suite must be > 0") raise ValueError(msg) self.testRunner = ParallelTestRunner(runner.TextTestRunner, self.multiprocess, self.useslice, verbosity=self.verbosity, failfast=self.failfast, buffer=self.buffer) def run_tests_real(): super(NumbaTestProgram, self).runTests() if self.profile: filename = os.path.splitext( os.path.basename(sys.modules['__main__'].__file__) )[0] + '.prof' p = cProfile.Profile(timer=time.perf_counter) # 3.3+ p.enable() try: p.runcall(run_tests_real) finally: p.disable() print("Writing test profile data into %r" % (filename,)) p.dump_stats(filename) else: run_tests_real() # These are tests which are generated and injected into the test suite, what # gets injected depends on features of the test environment, e.g. TBB presence # it's important for doing the CI "slice tests" that these are run at the end # See notes in `_flatten_suite` for why. Simple substring matching is used to # determine a match. _GENERATED = ( "numba.cuda.tests.cudapy.test_libdevice.TestLibdeviceCompilation", "numba.tests.test_num_threads", "numba.tests.test_parallel_backend", "numba.tests.test_svml", "numba.tests.test_ufuncs", ) def _flatten_suite_inner(test): """ Workhorse for _flatten_suite """ tests = [] if isinstance(test, (unittest.TestSuite, list, tuple)): for x in test: tests.extend(_flatten_suite_inner(x)) else: tests.append(test) return tests def _flatten_suite(test): """ Expand nested suite into list of test cases. """ tests = _flatten_suite_inner(test) # Strip out generated tests and stick them at the end, this is to make sure # that tests appear in a consistent order regardless of features available. # This is so that a slice through the test suite e.g. (1::N) would likely be # consistent up to the point of the generated tests, which rely on specific # features. generated = set() for t in tests: for g in _GENERATED: if g in str(t): generated.add(t) normal = set(tests) - generated def key(x): return x.__module__, type(x).__name__, x._testMethodName tests = sorted(normal, key=key) tests.extend(sorted(list(generated), key=key)) return tests def _choose_gitdiff_tests(tests, *, use_common_ancestor=False): try: from git import Repo except ImportError: raise ValueError("gitpython needed for git functionality") repo = Repo('.') path = os.path.join('numba', 'tests') if use_common_ancestor: print(f"Git diff by common ancestor") target = 'origin/release0.59...HEAD' else: target = 'origin/release0.59..HEAD' gdiff_paths = repo.git.diff(target, path, name_only=True).split() # normalise the paths as they are unix style from repo.git.diff gdiff_paths = [os.path.normpath(x) for x in gdiff_paths] selected = [] gdiff_paths = [os.path.join(repo.working_dir, x) for x in gdiff_paths] for test in _flatten_suite(tests): assert isinstance(test, unittest.TestCase) fname = inspect.getsourcefile(test.__class__) if fname in gdiff_paths: selected.append(test) print("Git diff identified %s tests" % len(selected)) return unittest.TestSuite(selected) def _choose_tagged_tests(tests, tags, mode='include'): """ Select tests that are tagged/not tagged with at least one of the given tags. Set mode to 'include' to include the tests with tags, or 'exclude' to exclude the tests with the tags. """ selected = [] tags = set(tags) for test in _flatten_suite(tests): assert isinstance(test, unittest.TestCase) func = getattr(test, test._testMethodName) try: # Look up the method's underlying function (Python 2) func = func.im_func except AttributeError: pass found_tags = getattr(func, 'tags', None) # only include the test if the tags *are* present if mode == 'include': if found_tags is not None and found_tags & tags: selected.append(test) elif mode == 'exclude': # only include the test if the tags *are not* present if found_tags is None or not (found_tags & tags): selected.append(test) else: raise ValueError("Invalid 'mode' supplied: %s." % mode) return unittest.TestSuite(selected) def _choose_random_tests(tests, ratio, seed): """ Choose a given proportion of tests at random. """ rnd = random.Random() rnd.seed(seed) if isinstance(tests, unittest.TestSuite): tests = _flatten_suite(tests) tests = rnd.sample(tests, int(len(tests) * ratio)) tests = sorted(tests, key=lambda case: case.id()) return unittest.TestSuite(tests) # The reference leak detection code is liberally taken and adapted from # Python's own Lib/test/regrtest.py. def _refleak_cleanup(): # Collect cyclic trash and read memory statistics immediately after. func1 = sys.getallocatedblocks try: func2 = sys.gettotalrefcount except AttributeError: func2 = lambda: 42 # Flush standard output, so that buffered data is sent to the OS and # associated Python objects are reclaimed. for stream in (sys.stdout, sys.stderr, sys.__stdout__, sys.__stderr__): if stream is not None: stream.flush() sys._clear_type_cache() # This also clears the various internal CPython freelists. gc.collect() return func1(), func2() class ReferenceLeakError(RuntimeError): pass class IntPool(collections.defaultdict): def __missing__(self, key): return key class RefleakTestResult(runner.TextTestResult): warmup = 3 repetitions = 6 def _huntLeaks(self, test): self.stream.flush() repcount = self.repetitions nwarmup = self.warmup rc_deltas = [0] * (repcount - nwarmup) alloc_deltas = [0] * (repcount - nwarmup) # Preallocate ints likely to be stored in rc_deltas and alloc_deltas, # to make sys.getallocatedblocks() less flaky. _int_pool = IntPool() for i in range(-200, 200): _int_pool[i] for i in range(repcount): # Use a pristine, silent result object to avoid recursion res = result.TestResult() test.run(res) # Poorly-written tests may fail when run several times. # In this case, abort the refleak run and report the failure. if not res.wasSuccessful(): self.failures.extend(res.failures) self.errors.extend(res.errors) raise AssertionError del res alloc_after, rc_after = _refleak_cleanup() if i >= nwarmup: rc_deltas[i - nwarmup] = _int_pool[rc_after - rc_before] alloc_deltas[i - nwarmup] = _int_pool[alloc_after - alloc_before] alloc_before, rc_before = alloc_after, rc_after return rc_deltas, alloc_deltas def addSuccess(self, test): try: rc_deltas, alloc_deltas = self._huntLeaks(test) except AssertionError: # Test failed when repeated assert not self.wasSuccessful() return # These checkers return False on success, True on failure def check_rc_deltas(deltas): return any(deltas) def check_alloc_deltas(deltas): # At least 1/3rd of 0s if 3 * deltas.count(0) < len(deltas): return True # Nothing else than 1s, 0s and -1s if not set(deltas) <= set((1, 0, -1)): return True return False failed = False for deltas, item_name, checker in [ (rc_deltas, 'references', check_rc_deltas), (alloc_deltas, 'memory blocks', check_alloc_deltas)]: if checker(deltas): msg = '%s leaked %s %s, sum=%s' % ( test, deltas, item_name, sum(deltas)) failed = True try: raise ReferenceLeakError(msg) except Exception: exc_info = sys.exc_info() if self.showAll: self.stream.write("%s = %r " % (item_name, deltas)) self.addFailure(test, exc_info) if not failed: super(RefleakTestResult, self).addSuccess(test) class RefleakTestRunner(runner.TextTestRunner): resultclass = RefleakTestResult class ParallelTestResult(runner.TextTestResult): """ A TestResult able to inject results from other results. """ def add_results(self, result): """ Add the results from the other *result* to this result. """ self.stream.write(result.stream.getvalue()) self.stream.flush() self.testsRun += result.testsRun self.failures.extend(result.failures) self.errors.extend(result.errors) self.skipped.extend(result.skipped) self.expectedFailures.extend(result.expectedFailures) self.unexpectedSuccesses.extend(result.unexpectedSuccesses) class _MinimalResult(object): """ A minimal, picklable TestResult-alike object. """ __slots__ = ( 'failures', 'errors', 'skipped', 'expectedFailures', 'unexpectedSuccesses', 'stream', 'shouldStop', 'testsRun', 'test_id') def fixup_case(self, case): """ Remove any unpicklable attributes from TestCase instance *case*. """ # Python 3.3 doesn't reset this one. case._outcomeForDoCleanups = None def __init__(self, original_result, test_id=None): for attr in self.__slots__: setattr(self, attr, getattr(original_result, attr, None)) for case, _ in self.expectedFailures: self.fixup_case(case) for case, _ in self.errors: self.fixup_case(case) for case, _ in self.failures: self.fixup_case(case) self.test_id = test_id class _FakeStringIO(object): """ A trivial picklable StringIO-alike for Python 2. """ def __init__(self, value): self._value = value def getvalue(self): return self._value class _MinimalRunner(object): """ A minimal picklable object able to instantiate a runner in a child process and run a test case with it. """ def __init__(self, runner_cls, runner_args): self.runner_cls = runner_cls self.runner_args = runner_args # Python 2 doesn't know how to pickle instance methods, so we use __call__ # instead. def __call__(self, test): # Executed in child process kwargs = self.runner_args # Force recording of output in a buffer (it will be printed out # by the parent). kwargs['stream'] = StringIO() runner = self.runner_cls(**kwargs) result = runner._makeResult() # Avoid child tracebacks when Ctrl-C is pressed. signals.installHandler() signals.registerResult(result) result.failfast = runner.failfast result.buffer = runner.buffer with self.cleanup_object(test): test(result) # HACK as cStringIO.StringIO isn't picklable in 2.x result.stream = _FakeStringIO(result.stream.getvalue()) return _MinimalResult(result, test.id()) @contextlib.contextmanager def cleanup_object(self, test): """ A context manager which cleans up unwanted attributes on a test case (or any other object). """ vanilla_attrs = set(test.__dict__) try: yield test finally: spurious_attrs = set(test.__dict__) - vanilla_attrs for name in spurious_attrs: del test.__dict__[name] def _split_nonparallel_tests(test, sliced): """ Split test suite into parallel and serial tests. """ ptests = [] stests = [] flat = [*filter(sliced, _flatten_suite(test))] def is_parallelizable_test_case(test): # Guard for the fake test case created by unittest when test # discovery fails, as it isn't picklable (e.g. "LoadTestsFailure") method_name = test._testMethodName method = getattr(test, method_name) if method.__name__ != method_name and method.__name__ == "testFailure": return False # Was parallel execution explicitly disabled? return getattr(test, "_numba_parallel_test_", True) for t in flat: if is_parallelizable_test_case(t): ptests.append(t) else: stests.append(t) return ptests, stests # A test can't run longer than 10 minutes _TIMEOUT = 600 class ParallelTestRunner(runner.TextTestRunner): """ A test runner which delegates the actual running to a pool of child processes. """ resultclass = ParallelTestResult timeout = _TIMEOUT def __init__(self, runner_cls, nprocs, useslice, **kwargs): runner.TextTestRunner.__init__(self, **kwargs) self.runner_cls = runner_cls self.nprocs = nprocs self.useslice = parse_slice(useslice) self.runner_args = kwargs def _run_inner(self, result): # We hijack TextTestRunner.run()'s inner logic by passing this # method as if it were a test case. child_runner = _MinimalRunner(self.runner_cls, self.runner_args) # Split the tests and recycle the worker process to tame memory usage. chunk_size = 100 splitted_tests = [self._ptests[i:i + chunk_size] for i in range(0, len(self._ptests), chunk_size)] for tests in splitted_tests: pool = multiprocessing.Pool(self.nprocs) try: self._run_parallel_tests(result, pool, child_runner, tests) except: # On exception, kill still active workers immediately pool.terminate() # Make sure exception is reported and not ignored raise else: # Close the pool cleanly unless asked to early out if result.shouldStop: pool.terminate() break else: pool.close() finally: # Always join the pool (this is necessary for coverage.py) pool.join() if not result.shouldStop: stests = SerialSuite(self._stests) stests.run(result) return result def _run_parallel_tests(self, result, pool, child_runner, tests): remaining_ids = set(t.id() for t in tests) tests.sort(key=cuda_sensitive_mtime) it = pool.imap_unordered(child_runner, tests) while True: try: child_result = it.__next__(self.timeout) except StopIteration: return except TimeoutError as e: # Diagnose the names of unfinished tests msg = ("Tests didn't finish before timeout (or crashed):\n%s" % "".join("- %r\n" % tid for tid in sorted(remaining_ids)) ) e.args = (msg,) + e.args[1:] raise e else: result.add_results(child_result) remaining_ids.discard(child_result.test_id) if child_result.shouldStop: result.shouldStop = True return def run(self, test): self._ptests, self._stests = _split_nonparallel_tests(test, self.useslice) print("Parallel: %s. Serial: %s" % (len(self._ptests), len(self._stests))) # This will call self._run_inner() on the created result object, # and print out the detailed test results at the end. return super(ParallelTestRunner, self).run(self._run_inner)