# Copyright 2013-2019 The Meson development team # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # http://www.apache.org/licenses/LICENSE-2.0 # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import glob import re import os import typing as T from pathlib import Path from .. import mlog from .. import mesonlib from ..environment import detect_cpu_family from .base import (DependencyException, ExternalDependency) if T.TYPE_CHECKING: from ..environment import Environment from ..compilers import Compiler TV_ResultTuple = T.Tuple[T.Optional[str], T.Optional[str], bool] class CudaDependency(ExternalDependency): supported_languages = ['cuda', 'cpp', 'c'] # see also _default_language def __init__(self, environment: 'Environment', kwargs: T.Dict[str, T.Any]) -> None: compilers = environment.coredata.compilers[self.get_for_machine_from_kwargs(kwargs)] language = self._detect_language(compilers) if language not in self.supported_languages: raise DependencyException(f'Language \'{language}\' is not supported by the CUDA Toolkit. Supported languages are {self.supported_languages}.') super().__init__('cuda', environment, kwargs, language=language) self.lib_modules: T.Dict[str, T.List[str]] = {} self.requested_modules = self.get_requested(kwargs) if 'cudart' not in self.requested_modules: self.requested_modules = ['cudart'] + self.requested_modules (self.cuda_path, self.version, self.is_found) = self._detect_cuda_path_and_version() if not self.is_found: return if not os.path.isabs(self.cuda_path): raise DependencyException(f'CUDA Toolkit path must be absolute, got \'{self.cuda_path}\'.') # nvcc already knows where to find the CUDA Toolkit, but if we're compiling # a mixed C/C++/CUDA project, we still need to make the include dir searchable if self.language != 'cuda' or len(compilers) > 1: self.incdir = os.path.join(self.cuda_path, 'include') self.compile_args += [f'-I{self.incdir}'] if self.language != 'cuda': arch_libdir = self._detect_arch_libdir() self.libdir = os.path.join(self.cuda_path, arch_libdir) mlog.debug('CUDA library directory is', mlog.bold(self.libdir)) else: self.libdir = None self.is_found = self._find_requested_libraries() @classmethod def _detect_language(cls, compilers: T.Dict[str, 'Compiler']) -> str: for lang in cls.supported_languages: if lang in compilers: return lang return list(compilers.keys())[0] def _detect_cuda_path_and_version(self) -> TV_ResultTuple: self.env_var = self._default_path_env_var() mlog.debug('Default path env var:', mlog.bold(self.env_var)) version_reqs = self.version_reqs if self.language == 'cuda': nvcc_version = self._strip_patch_version(self.get_compiler().version) mlog.debug('nvcc version:', mlog.bold(nvcc_version)) if version_reqs: # make sure nvcc version satisfies specified version requirements (found_some, not_found, found) = mesonlib.version_compare_many(nvcc_version, version_reqs) if not_found: msg = f'The current nvcc version {nvcc_version} does not satisfy the specified CUDA Toolkit version requirements {version_reqs}.' return self._report_dependency_error(msg, (None, None, False)) # use nvcc version to find a matching CUDA Toolkit version_reqs = [f'={nvcc_version}'] else: nvcc_version = None paths = [(path, self._cuda_toolkit_version(path), default) for (path, default) in self._cuda_paths()] if version_reqs: return self._find_matching_toolkit(paths, version_reqs, nvcc_version) defaults = [(path, version) for (path, version, default) in paths if default] if defaults: return (defaults[0][0], defaults[0][1], True) platform_msg = 'set the CUDA_PATH environment variable' if self._is_windows() \ else 'set the CUDA_PATH environment variable/create the \'/usr/local/cuda\' symbolic link' msg = f'Please specify the desired CUDA Toolkit version (e.g. dependency(\'cuda\', version : \'>=10.1\')) or {platform_msg} to point to the location of your desired version.' return self._report_dependency_error(msg, (None, None, False)) def _find_matching_toolkit(self, paths: T.List[TV_ResultTuple], version_reqs: T.List[str], nvcc_version: T.Optional[str]) -> TV_ResultTuple: # keep the default paths order intact, sort the rest in the descending order # according to the toolkit version part_func: T.Callable[[TV_ResultTuple], bool] = lambda t: not t[2] defaults_it, rest_it = mesonlib.partition(part_func, paths) defaults = list(defaults_it) paths = defaults + sorted(rest_it, key=lambda t: mesonlib.Version(t[1]), reverse=True) mlog.debug(f'Search paths: {paths}') if nvcc_version and defaults: default_src = f"the {self.env_var} environment variable" if self.env_var else "the \'/usr/local/cuda\' symbolic link" nvcc_warning = 'The default CUDA Toolkit as designated by {} ({}) doesn\'t match the current nvcc version {} and will be ignored.'.format(default_src, os.path.realpath(defaults[0][0]), nvcc_version) else: nvcc_warning = None for (path, version, default) in paths: (found_some, not_found, found) = mesonlib.version_compare_many(version, version_reqs) if not not_found: if not default and nvcc_warning: mlog.warning(nvcc_warning) return (path, version, True) if nvcc_warning: mlog.warning(nvcc_warning) return (None, None, False) def _default_path_env_var(self) -> T.Optional[str]: env_vars = ['CUDA_PATH'] if self._is_windows() else ['CUDA_PATH', 'CUDA_HOME', 'CUDA_ROOT'] env_vars = [var for var in env_vars if var in os.environ] user_defaults = {os.environ[var] for var in env_vars} if len(user_defaults) > 1: mlog.warning('Environment variables {} point to conflicting toolkit locations ({}). Toolkit selection might produce unexpected results.'.format(', '.join(env_vars), ', '.join(user_defaults))) return env_vars[0] if env_vars else None def _cuda_paths(self) -> T.List[T.Tuple[str, bool]]: return ([(os.environ[self.env_var], True)] if self.env_var else []) \ + (self._cuda_paths_win() if self._is_windows() else self._cuda_paths_nix()) def _cuda_paths_win(self) -> T.List[T.Tuple[str, bool]]: env_vars = os.environ.keys() return [(os.environ[var], False) for var in env_vars if var.startswith('CUDA_PATH_')] def _cuda_paths_nix(self) -> T.List[T.Tuple[str, bool]]: # include /usr/local/cuda default only if no env_var was found pattern = '/usr/local/cuda-*' if self.env_var else '/usr/local/cuda*' return [(path, os.path.basename(path) == 'cuda') for path in glob.iglob(pattern)] toolkit_version_regex = re.compile(r'^CUDA Version\s+(.*)$') path_version_win_regex = re.compile(r'^v(.*)$') path_version_nix_regex = re.compile(r'^cuda-(.*)$') cudart_version_regex = re.compile(r'#define\s+CUDART_VERSION\s+([0-9]+)') def _cuda_toolkit_version(self, path: str) -> str: version = self._read_toolkit_version_txt(path) if version: return version version = self._read_cuda_runtime_api_version(path) if version: return version mlog.debug('Falling back to extracting version from path') path_version_regex = self.path_version_win_regex if self._is_windows() else self.path_version_nix_regex try: m = path_version_regex.match(os.path.basename(path)) if m: return m.group(1) else: mlog.warning(f'Could not detect CUDA Toolkit version for {path}') except Exception as e: mlog.warning('Could not detect CUDA Toolkit version for {}: {}'.format(path, str(e))) return '0.0' def _read_cuda_runtime_api_version(self, path_str: str) -> T.Optional[str]: path = Path(path_str) for i in path.rglob('cuda_runtime_api.h'): raw = i.read_text() m = self.cudart_version_regex.search(raw) if not m: continue try: vers_int = int(m.group(1)) except ValueError: continue # use // for floor instead of / which produces a float major = vers_int // 1000 # type: int minor = (vers_int - major * 1000) // 10 # type: int return f'{major}.{minor}' return None def _read_toolkit_version_txt(self, path: str) -> T.Optional[str]: # Read 'version.txt' at the root of the CUDA Toolkit directory to determine the tookit version version_file_path = os.path.join(path, 'version.txt') try: with open(version_file_path) as version_file: version_str = version_file.readline() # e.g. 'CUDA Version 10.1.168' m = self.toolkit_version_regex.match(version_str) if m: return self._strip_patch_version(m.group(1)) except Exception as e: mlog.debug('Could not read CUDA Toolkit\'s version file {}: {}'.format(version_file_path, str(e))) return None @classmethod def _strip_patch_version(cls, version: str) -> str: return '.'.join(version.split('.')[:2]) def _detect_arch_libdir(self) -> str: arch = detect_cpu_family(self.env.coredata.compilers.host) machine = self.env.machines[self.for_machine] msg = '{} architecture is not supported in {} version of the CUDA Toolkit.' if machine.is_windows(): libdirs = {'x86': 'Win32', 'x86_64': 'x64'} if arch not in libdirs: raise DependencyException(msg.format(arch, 'Windows')) return os.path.join('lib', libdirs[arch]) elif machine.is_linux(): libdirs = {'x86_64': 'lib64', 'ppc64': 'lib', 'aarch64': 'lib64', 'loongarch64': 'lib64'} if arch not in libdirs: raise DependencyException(msg.format(arch, 'Linux')) return libdirs[arch] elif machine.is_darwin(): libdirs = {'x86_64': 'lib64'} if arch not in libdirs: raise DependencyException(msg.format(arch, 'macOS')) return libdirs[arch] else: raise DependencyException('CUDA Toolkit: unsupported platform.') def _find_requested_libraries(self) -> bool: all_found = True for module in self.requested_modules: args = self.clib_compiler.find_library(module, self.env, [self.libdir] if self.libdir else []) if args is None: self._report_dependency_error(f'Couldn\'t find requested CUDA module \'{module}\'') all_found = False else: mlog.debug(f'Link args for CUDA module \'{module}\' are {args}') self.lib_modules[module] = args return all_found def _is_windows(self) -> bool: return self.env.machines[self.for_machine].is_windows() @T.overload def _report_dependency_error(self, msg: str) -> None: ... @T.overload def _report_dependency_error(self, msg: str, ret_val: TV_ResultTuple) -> TV_ResultTuple: ... def _report_dependency_error(self, msg: str, ret_val: T.Optional[TV_ResultTuple] = None) -> T.Optional[TV_ResultTuple]: if self.required: raise DependencyException(msg) mlog.debug(msg) return ret_val def log_details(self) -> str: module_str = ', '.join(self.requested_modules) return 'modules: ' + module_str def log_info(self) -> str: return self.cuda_path if self.cuda_path else '' def get_requested(self, kwargs: T.Dict[str, T.Any]) -> T.List[str]: candidates = mesonlib.extract_as_list(kwargs, 'modules') for c in candidates: if not isinstance(c, str): raise DependencyException('CUDA module argument is not a string.') return candidates def get_link_args(self, language: T.Optional[str] = None, raw: bool = False) -> T.List[str]: args = [] if self.libdir: args += self.clib_compiler.get_linker_search_args(self.libdir) for lib in self.requested_modules: args += self.lib_modules[lib] return args