Matt300209's picture
Add files using upload-large-folder tool
948620a verified
import gc
import hashlib
import inspect
import os
import re
import sys
import time
from distutils.core import Distribution, Extension
from distutils.command.build_ext import build_ext
import Cython
from ..Compiler.Main import Context
from ..Compiler.Options import (default_options, CompilationOptions,
get_directive_defaults)
from ..Compiler.Visitor import CythonTransform, EnvTransform
from ..Compiler.ParseTreeTransforms import SkipDeclarations
from ..Compiler.TreeFragment import parse_from_strings
from .Dependencies import strip_string_literals, cythonize, cached_function
from .Cache import get_cython_cache_dir
from ..Compiler import Pipeline
import cython as cython_module
import importlib.util
from importlib.machinery import ExtensionFileLoader
def load_dynamic(name, path):
spec = importlib.util.spec_from_file_location(name, loader=ExtensionFileLoader(name, path))
module = importlib.util.module_from_spec(spec)
spec.loader.exec_module(module)
return module
class UnboundSymbols(EnvTransform, SkipDeclarations):
def __init__(self):
super(EnvTransform, self).__init__(context=None)
self.unbound = set()
def visit_NameNode(self, node):
if not self.current_env().lookup(node.name):
self.unbound.add(node.name)
return node
def __call__(self, node):
super().__call__(node)
return self.unbound
@cached_function
def unbound_symbols(code, context=None):
if context is None:
context = Context([], get_directive_defaults(),
options=CompilationOptions(default_options))
from ..Compiler.ParseTreeTransforms import AnalyseDeclarationsTransform
tree = parse_from_strings('(tree fragment)', code)
for phase in Pipeline.create_pipeline(context, 'pyx'):
if phase is None:
continue
tree = phase(tree)
if isinstance(phase, AnalyseDeclarationsTransform):
break
import builtins
return tuple(UnboundSymbols()(tree) - set(dir(builtins)))
def unsafe_type(arg, context=None):
py_type = type(arg)
if py_type is int:
return 'long'
else:
return safe_type(arg, context)
def safe_type(arg, context=None):
py_type = type(arg)
if py_type in (list, tuple, dict, str):
return py_type.__name__
elif py_type is complex:
return 'double complex'
elif py_type is float:
return 'double'
elif py_type is bool:
return 'bint'
elif 'numpy' in sys.modules and isinstance(arg, sys.modules['numpy'].ndarray):
return 'numpy.ndarray[numpy.%s_t, ndim=%s]' % (arg.dtype.name, arg.ndim)
else:
for base_type in py_type.__mro__:
if base_type.__module__ in ('__builtin__', 'builtins'):
return 'object'
module = context.find_module(base_type.__module__, need_pxd=False)
if module:
entry = module.lookup(base_type.__name__)
if entry.is_type:
return '%s.%s' % (base_type.__module__, base_type.__name__)
return 'object'
def _get_build_extension():
dist = Distribution()
# Ensure the build respects distutils configuration by parsing
# the configuration files
config_files = dist.find_config_files()
dist.parse_config_files(config_files)
build_extension = build_ext(dist)
build_extension.finalize_options()
return build_extension
@cached_function
def _create_context(cython_include_dirs):
return Context(
list(cython_include_dirs),
get_directive_defaults(),
options=CompilationOptions(default_options)
)
_cython_inline_cache = {}
_cython_inline_default_context = _create_context(('.',))
def _populate_unbound(kwds, unbound_symbols, locals=None, globals=None):
for symbol in unbound_symbols:
if symbol not in kwds:
if locals is None or globals is None:
calling_frame = inspect.currentframe().f_back.f_back.f_back
if locals is None:
locals = calling_frame.f_locals
if globals is None:
globals = calling_frame.f_globals
if not isinstance(locals, dict):
# FrameLocalsProxy is stricter than dict on how it looks up keys
# and this means our "EncodedStrings" don't match the keys in locals.
# Therefore copy to a dict.
locals = dict(locals)
if symbol in locals:
kwds[symbol] = locals[symbol]
elif symbol in globals:
kwds[symbol] = globals[symbol]
else:
print("Couldn't find %r" % symbol)
def _inline_key(orig_code, arg_sigs, language_level):
key = orig_code, arg_sigs, sys.version_info, sys.executable, language_level, Cython.__version__
return hashlib.sha256(str(key).encode('utf-8')).hexdigest()
def cython_inline(code, get_type=unsafe_type,
lib_dir=os.path.join(get_cython_cache_dir(), 'inline'),
cython_include_dirs=None, cython_compiler_directives=None,
force=False, quiet=False, locals=None, globals=None, language_level=None, **kwds):
if get_type is None:
get_type = lambda x: 'object'
ctx = _create_context(tuple(cython_include_dirs)) if cython_include_dirs else _cython_inline_default_context
cython_compiler_directives = dict(cython_compiler_directives) if cython_compiler_directives else {}
if language_level is None and 'language_level' not in cython_compiler_directives:
language_level = '3'
if language_level is not None:
cython_compiler_directives['language_level'] = language_level
key_hash = None
# Fast path if this has been called in this session.
_unbound_symbols = _cython_inline_cache.get(code)
if _unbound_symbols is not None:
_populate_unbound(kwds, _unbound_symbols, locals, globals)
args = sorted(kwds.items())
arg_sigs = tuple([(get_type(value, ctx), arg) for arg, value in args])
key_hash = _inline_key(code, arg_sigs, language_level)
invoke = _cython_inline_cache.get((code, arg_sigs, key_hash))
if invoke is not None:
arg_list = [arg[1] for arg in args]
return invoke(*arg_list)
orig_code = code
code, literals = strip_string_literals(code)
code = strip_common_indent(code)
if locals is None:
locals = inspect.currentframe().f_back.f_back.f_locals
if globals is None:
globals = inspect.currentframe().f_back.f_back.f_globals
try:
_cython_inline_cache[orig_code] = _unbound_symbols = unbound_symbols(code)
_populate_unbound(kwds, _unbound_symbols, locals, globals)
except AssertionError:
if not quiet:
# Parsing from strings not fully supported (e.g. cimports).
print("Could not parse code as a string (to extract unbound symbols).")
cimports = []
for name, arg in list(kwds.items()):
if arg is cython_module:
cimports.append('\ncimport cython as %s' % name)
del kwds[name]
arg_names = sorted(kwds)
arg_sigs = tuple([(get_type(kwds[arg], ctx), arg) for arg in arg_names])
if key_hash is None:
key_hash = _inline_key(orig_code, arg_sigs, language_level)
module_name = "_cython_inline_" + key_hash
if module_name in sys.modules:
module = sys.modules[module_name]
else:
build_extension = None
if cython_inline.so_ext is None:
# Figure out and cache current extension suffix
build_extension = _get_build_extension()
cython_inline.so_ext = build_extension.get_ext_filename('')
lib_dir = os.path.abspath(lib_dir)
module_path = os.path.join(lib_dir, module_name + cython_inline.so_ext)
if not os.path.exists(lib_dir):
os.makedirs(lib_dir)
if force or not os.path.isfile(module_path):
cflags = []
define_macros = []
c_include_dirs = []
qualified = re.compile(r'([.\w]+)[.]')
for type, _ in arg_sigs:
m = qualified.match(type)
if m:
cimports.append('\ncimport %s' % m.groups()[0])
# one special case
if m.groups()[0] == 'numpy':
import numpy
c_include_dirs.append(numpy.get_include())
define_macros.append(("NPY_NO_DEPRECATED_API", "NPY_1_7_API_VERSION"))
# cflags.append('-Wno-unused')
module_body, func_body = extract_func_code(code)
params = ', '.join(['%s %s' % a for a in arg_sigs])
module_code = """
%(module_body)s
%(cimports)s
def __invoke(%(params)s):
%(func_body)s
return locals()
""" % {'cimports': '\n'.join(cimports),
'module_body': module_body,
'params': params,
'func_body': func_body }
for key, value in literals.items():
module_code = module_code.replace(key, value)
pyx_file = os.path.join(lib_dir, module_name + '.pyx')
fh = open(pyx_file, 'w')
try:
fh.write(module_code)
finally:
fh.close()
extension = Extension(
name=module_name,
sources=[pyx_file],
include_dirs=c_include_dirs or None,
extra_compile_args=cflags or None,
define_macros=define_macros or None,
)
if build_extension is None:
build_extension = _get_build_extension()
build_extension.extensions = cythonize(
[extension],
include_path=cython_include_dirs or ['.'],
compiler_directives=cython_compiler_directives,
quiet=quiet)
build_extension.build_temp = os.path.dirname(pyx_file)
build_extension.build_lib = lib_dir
build_extension.run()
if sys.platform == 'win32' and sys.version_info >= (3, 8):
with os.add_dll_directory(os.path.abspath(lib_dir)):
module = load_dynamic(module_name, module_path)
else:
module = load_dynamic(module_name, module_path)
_cython_inline_cache[orig_code, arg_sigs, key_hash] = module.__invoke
arg_list = [kwds[arg] for arg in arg_names]
return module.__invoke(*arg_list)
# The code template used for cymeit benchmark runs.
# We keep the benchmark repetition separate from the benchmarked code
# to prevent the C compiler from doing unhelpful loop optimisations.
_CYMEIT_TEMPLATE = """
def __PYX_repeat_benchmark(benchmark, timer, size_t number):
cdef size_t i
t0 = timer()
for i in range(number):
benchmark()
t1 = timer()
return t1 - t0
def __PYX_make_benchmark():
{setup_code}
def __PYX_run_benchmark():
{benchmark_code}
return __PYX_run_benchmark
"""
def cymeit(code, setup_code=None, import_module=None, directives=None, timer=time.perf_counter, repeat=9):
"""Benchmark a Cython code string similar to 'timeit'.
'setup_code': string of setup code that will be run before taking the timings.
'import_module': a module namespace to run the benchmark in
(usually a compiled Cython module).
'directives': Cython directives to use when compiling the benchmark code.
'timer': The timer function. Defaults to 'time.perf_counter', returning float seconds.
Nanosecond timers are detected (and can only be used) if they return integers.
'repeat': The number of timings to take and return.
Returns a tuple: (list of single-loop timings, number of loops run for each)
"""
import textwrap
# Compile the benchmark code as an inline closure function.
setup_code = strip_common_indent(setup_code) if setup_code else ''
code = strip_common_indent(code) if code.strip() else 'pass'
module_namespace = __import__(import_module).__dict__ if import_module else None
cymeit_code = _CYMEIT_TEMPLATE.format(
setup_code=textwrap.indent(setup_code, ' '*4).strip(),
benchmark_code=textwrap.indent(code, ' '*8).strip(),
)
namespace = cython_inline(
cymeit_code,
cython_compiler_directives=directives,
locals=module_namespace,
)
make_benchmark = namespace['__PYX_make_benchmark']
repeat_benchmark = namespace['__PYX_repeat_benchmark']
# Based on 'timeit' in CPython 3.13.
def timeit(number):
benchmark = make_benchmark()
gcold = gc.isenabled()
gc.disable()
try:
timing = repeat_benchmark(benchmark, timer, number)
finally:
if gcold:
gc.enable()
return timing
# Find a sufficiently large number of loops, warm up the system.
timer_returns_nanoseconds = isinstance(timer(), int)
one_second = 1_000_000_000 if timer_returns_nanoseconds else 1.0
# Run for at least 0.2 seconds, either as integer nanoseconds or floating point seconds.
min_runtime = one_second // 5 if timer_returns_nanoseconds else one_second / 5
def autorange():
i = 1
while True:
for j in 1, 2, 5:
number = i * j
time_taken = timeit(number)
assert isinstance(time_taken, int if timer_returns_nanoseconds else float)
if time_taken >= min_runtime:
return number
elif timer_returns_nanoseconds and (time_taken < 10 and number >= 10):
# Arbitrary sanity check to prevent endless loops for non-ns timers.
raise RuntimeError(f"Timer seems to return non-ns timings: {timer}")
i *= 10
autorange() # warmup
number = autorange()
# Run and repeat the benchmark.
timings = [
timeit(number)
for _ in range(repeat)
]
half = number // 2 # for integer rounding
timings = [
(timing + half) // number if timer_returns_nanoseconds else timing / number
for timing in timings
]
return (timings, number)
# Cached suffix used by cython_inline above. None should get
# overridden with actual value upon the first cython_inline invocation
cython_inline.so_ext = None
_find_non_space = re.compile(r'\S').search
def strip_common_indent(code):
min_indent = None
lines = code.splitlines()
for line in lines:
match = _find_non_space(line)
if not match:
continue # blank
indent = match.start()
if line[indent] == '#':
continue # comment
if min_indent is None or min_indent > indent:
min_indent = indent
for ix, line in enumerate(lines):
match = _find_non_space(line)
if not match or not line or line[indent:indent+1] == '#':
continue
lines[ix] = line[min_indent:]
return '\n'.join(lines)
module_statement = re.compile(r'^((cdef +(extern|class))|cimport|(from .+ cimport)|(from .+ import +[*]))')
def extract_func_code(code):
module = []
function = []
current = function
code = code.replace('\t', ' ')
lines = code.split('\n')
for line in lines:
if not line.startswith(' '):
if module_statement.match(line):
current = module
else:
current = function
current.append(line)
return '\n'.join(module), ' ' + '\n '.join(function)
def get_body(source):
ix = source.index(':')
if source[:5] == 'lambda':
return "return %s" % source[ix+1:]
else:
return source[ix+1:]
# Lots to be done here... It would be especially cool if compiled functions
# could invoke each other quickly.
class RuntimeCompiledFunction:
def __init__(self, f):
self._f = f
self._body = get_body(inspect.getsource(f))
def __call__(self, *args, **kwds):
all = inspect.getcallargs(self._f, *args, **kwds)
return cython_inline(self._body, locals=self._f.__globals__, globals=self._f.__globals__, **all)