diff 2.00/testcases.py @ 21:ec6f1a132109

A pretty usable version Test groups and testconfs in non-ZIP archives or ZIP archives with comments are not yet supported.
author Oleg Oshmyan <chortos@inbox.lv>
date Fri, 06 Aug 2010 15:39:29 +0000
parents f2279b7602d3
children f07b7a431ea6
line wrap: on
line diff
--- a/2.00/testcases.py	Mon Jun 14 21:02:06 2010 +0000
+++ b/2.00/testcases.py	Fri Aug 06 15:39:29 2010 +0000
@@ -1,35 +1,380 @@
-#!/usr/bin/python
+#! /usr/bin/env python
 # Copyright (c) 2010 Chortos-2 <chortos@inbox.lv>
 
+from __future__ import division, with_statement
+
+try:
+	from compat import *
+	import files, problem, config
+except ImportError:
+	import __main__
+	__main__.import_error(sys.exc_info()[1])
+else:
+	from __main__ import clock, options
+
+import glob, re, sys, tempfile, time
+from subprocess import Popen, PIPE, STDOUT
+
+import os
+devnull = open(os.path.devnull, 'w+')
+
+try:
+	from signal import SIGTERM, SIGKILL
+except ImportError:
+	SIGTERM = 15
+	SIGKILL = 9
+
 try:
-	import files as _files, problem as _problem
-except ImportError as e:
-	import __main__
-	__main__.import_error(e)
+	from _subprocess import TerminateProcess
+except ImportError:
+	# CPython 2.5 does define _subprocess.TerminateProcess even though it is
+	# not used in the subprocess module, but maybe something else does not
+	try:
+		import ctypes
+		TerminateProcess = ctypes.windll.kernel32.TerminateProcess
+	except (ImportError, AttributeError):
+		TerminateProcess = None
+
+__all__ = ('TestCase', 'load_problem', 'TestCaseNotPassed',
+           'TimeLimitExceeded', 'WrongAnswer', 'NonZeroExitCode',
+           'CannotStartTestee', 'CannotStartValidator',
+           'CannotReadOutputFile')
+
+
+
+# Exceptions
+
+class TestCaseNotPassed(Exception): __slots__ = ()
+class TimeLimitExceeded(TestCaseNotPassed): __slots__ = ()
+
+class WrongAnswer(TestCaseNotPassed):
+	__slots__ = 'comment'
+	def __init__(self, comment=''):
+		self.comment = comment
+
+class NonZeroExitCode(TestCaseNotPassed):
+	__slots__ = 'exitcode'
+	def __init__(self, exitcode):
+		self.exitcode = exitcode
+
+class ExceptionWrapper(TestCaseNotPassed):
+	__slots__ = 'upstream'
+	def __init__(self, upstream):
+		self.upstream = upstream
+
+class CannotStartTestee(ExceptionWrapper): __slots__ = ()
+class CannotStartValidator(ExceptionWrapper): __slots__ = ()
+class CannotReadOutputFile(ExceptionWrapper): __slots__ = ()
+class CannotReadInputFile(ExceptionWrapper): __slots__ = ()
+class CannotReadAnswerFile(ExceptionWrapper): __slots__ = ()
+
+
+
+# Test case types
 
 class TestCase(object):
-	__slots__ = 'problem', 'infile', 'outfile'
+	__slots__ = ('problem', 'id', 'isdummy', 'infile', 'outfile', 'points',
+	             'process', 'time_started', 'time_stopped', 'time_limit_string',
+	             'realinname', 'realoutname', 'maxtime', 'maxmemory')
+	
+	if ABCMeta:
+		__metaclass__ = ABCMeta
 	
-	def __init__(case, prob, infile, outfile):
+	def __init__(case, prob, id, isdummy, points):
 		case.problem = prob
-		case.infile = infile
-		case.outfile = outfile
+		case.id = id
+		case.isdummy = isdummy
+		case.points = points
+		case.maxtime = case.problem.config.maxtime
+		case.maxmemory = case.problem.config.maxmemory
+		if case.maxtime:
+			case.time_limit_string = '/%.3f' % case.maxtime
+		else:
+			case.time_limit_string = ''
+		if not isdummy:
+			case.realinname = case.problem.config.testcaseinname
+			case.realoutname = case.problem.config.testcaseoutname
+		else:
+			case.realinname = case.problem.config.dummyinname
+			case.realoutname = case.problem.config.dummyoutname
+	
+	@abstractmethod
+	def test(case): raise NotImplementedError
 	
 	def __call__(case):
-		os.copy()
+		try:
+			return case.test()
+		finally:
+			case.cleanup()
+	
+	def cleanup(case):
+		if not getattr(case, 'time_started', None):
+			case.time_started = case.time_stopped = clock()
+		elif not getattr(case, 'time_stopped', None):
+			case.time_stopped = clock()
+		#if getattr(case, 'infile', None):
+		#	case.infile.close()
+		#if getattr(case, 'outfile', None):
+		#	case.outfile.close()
+		if getattr(case, 'process', None):
+			# Try killing after three unsuccessful TERM attempts in a row
+			# (except on Windows, where TERMing is killing)
+			for i in range(3):
+				try:
+					try:
+						case.process.terminate()
+					except AttributeError:
+						# Python 2.5
+						if TerminateProcess and hasattr(proc, '_handle'):
+							# Windows API
+							TerminateProcess(proc._handle, 1)
+						else:
+							# POSIX
+							os.kill(proc.pid, SIGTERM)
+				except Exception:
+					time.sleep(0)
+					case.process.poll()
+				else:
+					break
+			else:
+				# If killing the process is unsuccessful three times in a row,
+				# just silently stop trying
+				for i in range(3):
+					try:
+						try:
+							case.process.kill()
+						except AttributeError:
+							# Python 2.5
+							if TerminateProcess and hasattr(proc, '_handle'):
+								# Windows API
+								TerminateProcess(proc._handle, 1)
+							else:
+								# POSIX
+								os.kill(proc.pid, SIGKILL)
+					except Exception:
+						time.sleep(0)
+						case.process.poll()
+					else:
+						break
+	
+	def open_infile(case):
+		try:
+			case.infile = files.File('/'.join((case.problem.name, case.realinname.replace('$', case.id))))
+		except IOError:
+			e = sys.exc_info()[1]
+			raise CannotReadInputFile(e)
+	
+	def open_outfile(case):
+		try:
+			case.outfile = files.File('/'.join((case.problem.name, case.realoutname.replace('$', case.id))))
+		except IOError:
+			e = sys.exc_info()[1]
+			raise CannotReadAnswerFile(e)
+
 
-def load_problem(prob):
+class ValidatedTestCase(TestCase):
+	__slots__ = 'validator'
+	
+	def __init__(case, *args):
+		TestCase.__init__(case, *args)
+		if not case.problem.config.tester:
+			case.validator = None
+		else:
+			case.validator = case.problem.config.tester
+	
+	# TODO
+	def validate(case, output):
+		if not case.validator:
+			# Compare the output with the reference output
+			case.open_outfile()
+			with case.outfile.open() as refoutput:
+				for line, refline in zip(output, refoutput):
+					if not isinstance(refline, basestring):
+						line = bytes(line, sys.getdefaultencoding())
+					if line != refline:
+						raise WrongAnswer()
+				try:
+					try:
+						next(output)
+					except NameError:
+						output.next()
+				except StopIteration:
+					pass
+				else:
+					raise WrongAnswer()
+				try:
+					try:
+						next(refoutput)
+					except NameError:
+						refoutput.next()
+				except StopIteration:
+					pass
+				else:
+					raise WrongAnswer()
+			return case.points
+		elif callable(case.validator):
+			return case.validator(output)
+		else:                 
+			# Call the validator program
+			output.close()
+			case.open_outfile()
+			if case.problem.config.ansname:
+				case.outfile.copy(case.problem.config.ansname)
+			case.process = Popen(case.validator, stdin=devnull, stdout=PIPE, stderr=STDOUT, universal_newlines=True, bufsize=-1)
+			comment = case.process.communicate()[0].strip()
+			lower = comment.lower()
+			match = re.match(r'(ok|correct|wrong(?:(?:\s|_)*answer)?)(?:$|\s+|[.,!:]+\s*)', lower)
+			if match:
+				comment = comment[match.end():]
+			if not case.problem.config.maxexitcode:
+				if case.process.returncode:
+					raise WrongAnswer(comment)
+				else:
+					return case.points, comment
+			else:
+				return case.points * case.process.returncode / case.problem.config.maxexitcode, comment
+
+
+class BatchTestCase(ValidatedTestCase):
+	__slots__ = ()
+	
+	def test(case):
+		if sys.platform == 'win32' or not case.maxmemory:
+			preexec_fn = None
+		else:
+			def preexec_fn():
+				try:
+					import resource
+					maxmemory = int(case.maxmemory * 1048576)
+					resource.setrlimit(resource.RLIMIT_AS, (maxmemory, maxmemory))
+					# I would also set a CPU time limit but I do not want the time
+					# that passes between the calls to fork and exec to be counted in
+				except MemoryError:
+					# We do not have enough memory for ourselves;
+					# let the parent know about this
+					raise
+				except Exception:
+					# Well, at least we tried
+					pass
+		case.open_infile()
+		case.time_started = None
+		if case.problem.config.stdio:
+			if options.erase and not case.validator:
+				# FIXME: 2.5 lacks the delete parameter
+				with tempfile.NamedTemporaryFile(delete=False) as f:
+					inputdatafname = f.name 
+			else:
+				inputdatafname = case.problem.config.inname
+			case.infile.copy(inputdatafname)
+			# FIXME: inputdatafname should be deleted on __exit__
+			with open(inputdatafname, 'rU') as infile:
+				with tempfile.TemporaryFile('w+') if options.erase and not case.validator else open(case.problem.config.outname, 'w+') as outfile:
+					try:
+						try:
+							case.process = Popen(case.problem.config.path, stdin=infile, stdout=outfile, stderr=devnull, universal_newlines=True, bufsize=-1, preexec_fn=preexec_fn)
+						except MemoryError:
+							# If there is not enough memory for the forked test.py,
+							# opt for silent dropping of the limit
+							case.process = Popen(case.problem.config.path, stdin=infile, stdout=outfile, stderr=devnull, universal_newlines=True, bufsize=-1)
+					except OSError:
+						raise CannotStartTestee(sys.exc_info()[1])
+					case.time_started = clock()
+					# If we use a temporary file, it may not be a true file object,
+					# and if so, Popen will relay the standard output through pipes
+					if not case.maxtime:
+						case.process.communicate()
+						case.time_stopped = clock()
+					else:
+						time_end = case.time_started + case.maxtime
+						# FIXME: emulate communicate()
+						while True:
+							exitcode = case.process.poll()
+							now = clock()
+							if exitcode is not None:
+								case.time_stopped = now
+								break
+							elif now >= time_end:
+								raise TimeLimitExceeded()
+					if config.globalconf.force_zero_exitcode and case.process.returncode:
+						raise NonZeroExitCode(case.process.returncode)
+					outfile.seek(0)
+					return case.validate(outfile)
+		else:
+			if case.problem.config.inname:
+				case.infile.copy(case.problem.config.inname)
+			try:
+				try:
+					case.process = Popen(case.problem.config.path, stdin=devnull, stdout=devnull, stderr=STDOUT, preexec_fn=preexec_fn)
+				except MemoryError:
+					# If there is not enough memory for the forked test.py,
+					# opt for silent dropping of the limit
+					case.process = Popen(case.problem.config.path, stdin=devnull, stdout=devnull, stderr=STDOUT)
+			except OSError:
+				raise CannotStartTestee(sys.exc_info()[1])
+			case.time_started = clock()
+			if not case.maxtime:
+				case.process.wait()
+				case.time_stopped = clock()
+			else:
+				time_end = case.time_started + case.maxtime
+				while True:
+					exitcode = case.process.poll()
+					now = clock()
+					if exitcode is not None:
+						case.time_stopped = now
+						break
+					elif now >= time_end:
+						raise TimeLimitExceeded()
+			if config.globalconf.force_zero_exitcode and case.process.returncode:
+				raise NonZeroExitCode(case.process.returncode)
+			with open(case.problem.config.outname, 'rU') as output:
+				return case.validate(output)
+
+
+# This is the only test case type not executing any programs to be tested
+class OutputOnlyTestCase(ValidatedTestCase):
+	__slots__ = ()
+	def cleanup(case): pass
+
+class BestOutputTestCase(ValidatedTestCase):
+	__slots__ = ()
+
+# This is the only test case type executing two programs simultaneously
+class ReactiveTestCase(TestCase):
+	__slots__ = ()
+	# The basic idea is to launch the program to be tested and the grader
+	# and to pipe their standard I/O from and to each other,
+	# and then to capture the grader's exit code and use it
+	# like the exit code of a test validator is used.
+
+
+def load_problem(prob, _types={'batch'   : BatchTestCase,
+                               'outonly' : OutputOnlyTestCase,
+                               'bestout' : BestOutputTestCase,
+                               'reactive': ReactiveTestCase}):
 	if prob.config.usegroups:
 		pass
 	else:
+		# We will need to iterate over these configuration variables twice
+		try:
+			len(prob.config.dummies)
+		except Exception:
+			prob.config.dummies = tuple(prob.config.dummies)
+		try:
+			len(prob.config.tests)
+		except Exception:
+			prob.config.dummies = tuple(prob.config.tests)
+		# First get prob.cache.padoutput right
+		for i in prob.config.dummies:
+			s = 'sample ' + str(i).zfill(prob.config.paddummies)
+			prob.cache.padoutput = max(prob.cache.padoutput, len(s))
 		for i in prob.config.tests:
-			s = str(i).zfill(prob.config.padwithzeroestolength)
-			prob.cache.padoutputtolength = max(prob.cache.padoutputtolength, len(s))
-			infile = _files.TestCaseFile(prob, prob.config.testcaseinname.replace('$', s))
-			if infile:
-				if prob.config.kind != _problem.BATCH:
-					yield TestCase(prob, infile, None)
-				else:
-					outfile = _files.TestCaseFile(prob, prob.config.testcaseoutname.replace('$', s))
-					if outfile:
-						yield TestCase(prob, infile, outfile)
\ No newline at end of file
+			s = str(i).zfill(prob.config.padtests)
+			prob.cache.padoutput = max(prob.cache.padoutput, len(s))
+		# Now yield the actual test cases
+		for i in prob.config.dummies:
+			s = str(i).zfill(prob.config.paddummies)
+			yield _types[prob.config.kind](prob, s, True, 0)
+		for i in prob.config.tests:
+			s = str(i).zfill(prob.config.padtests)
+			yield _types[prob.config.kind](prob, s, False, prob.config.pointmap.get(i, prob.config.pointmap.get(None, prob.config.maxexitcode if prob.config.maxexitcode else 1)))
\ No newline at end of file