xref: /aosp_15_r20/external/tensorflow/third_party/gpus/crosstool/windows/msvc_wrapper_for_nvcc.py.tpl (revision b6fb3261f9314811a0f4371741dbb8839866f948)
1#!/usr/bin/env python
2# Copyright 2015 The TensorFlow Authors. All Rights Reserved.
3#
4# Licensed under the Apache License, Version 2.0 (the "License");
5# you may not use this file except in compliance with the License.
6# You may obtain a copy of the License at
7#
8#     http://www.apache.org/licenses/LICENSE-2.0
9#
10# Unless required by applicable law or agreed to in writing, software
11# distributed under the License is distributed on an "AS IS" BASIS,
12# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13# See the License for the specific language governing permissions and
14# limitations under the License.
15# ==============================================================================
16
17"""Crosstool wrapper for compiling CUDA programs with nvcc on Windows.
18
19DESCRIPTION:
20  This script is the Windows version of //third_party/gpus/crosstool/crosstool_wrapper_is_not_gcc
21"""
22
23from argparse import ArgumentParser
24import os
25import subprocess
26import re
27import sys
28import tempfile
29
30# Template values set by cuda_autoconf.
31CPU_COMPILER = ('%{cpu_compiler}')
32GCC_HOST_COMPILER_PATH = ('%{gcc_host_compiler_path}')
33
34NVCC_PATH = '%{nvcc_path}'
35NVCC_VERSION = '%{cuda_version}'
36NVCC_TEMP_DIR = "%{nvcc_tmp_dir}"
37
38def Log(s):
39  print('gpus/crosstool: {0}'.format(s))
40
41
42def GetOptionValue(argv, option):
43  """Extract the list of values for option from options.
44
45  Args:
46    option: The option whose value to extract.
47
48  Returns:
49    1. A list of values, either directly following the option,
50    (eg., /opt val1 val2) or values collected from multiple occurrences of
51    the option (eg., /opt val1 /opt val2).
52    2. The leftover options.
53  """
54
55  parser = ArgumentParser(prefix_chars='-/')
56  parser.add_argument(option, nargs='*', action='append')
57  option = option.lstrip('-/').replace('-', '_')
58  args, leftover = parser.parse_known_args(argv)
59  if args and vars(args)[option]:
60    return (sum(vars(args)[option], []), leftover)
61  return ([], leftover)
62
63def _update_options(nvcc_options):
64  if NVCC_VERSION in ("7.0",):
65    return nvcc_options
66
67  update_options = { "relaxed-constexpr" : "expt-relaxed-constexpr" }
68  return [ update_options[opt] if opt in update_options else opt
69                    for opt in nvcc_options ]
70
71def GetNvccOptions(argv):
72  """Collect the -nvcc_options values from argv.
73
74  Args:
75    argv: A list of strings, possibly the argv passed to main().
76
77  Returns:
78    1. The string that can be passed directly to nvcc.
79    2. The leftover options.
80  """
81
82  parser = ArgumentParser()
83  parser.add_argument('-nvcc_options', nargs='*', action='append')
84
85  args, leftover = parser.parse_known_args(argv)
86
87  if args.nvcc_options:
88    options = _update_options(sum(args.nvcc_options, []))
89    return (['--' + a for a in options], leftover)
90  return ([], leftover)
91
92
93def InvokeNvcc(argv, log=False):
94  """Call nvcc with arguments assembled from argv.
95
96  Args:
97    argv: A list of strings, possibly the argv passed to main().
98    log: True if logging is requested.
99
100  Returns:
101    The return value of calling os.system('nvcc ' + args)
102  """
103
104  src_files = [f for f in argv if
105               re.search('\.cpp$|\.cc$|\.c$|\.cxx$|\.C$', f)]
106  if len(src_files) == 0:
107    raise Error('No source files found for cuda compilation.')
108
109  out_file = [ f for f in argv if f.startswith('/Fo') ]
110  if len(out_file) != 1:
111    raise Error('Please specify exactly one output file for cuda compilation.')
112  out = ['-o', out_file[0][len('/Fo'):]]
113
114  nvcc_compiler_options, argv = GetNvccOptions(argv)
115
116  opt_option, argv = GetOptionValue(argv, '/O')
117  opt = ['-g']
118  if (len(opt_option) > 0 and opt_option[0] != 'd'):
119    opt = ['-O2']
120
121  include_options, argv = GetOptionValue(argv, '/I')
122  includes = ["-I " + include for include in include_options]
123
124  defines, argv = GetOptionValue(argv, '/D')
125  defines = ['-D' + define for define in defines]
126
127  undefines, argv = GetOptionValue(argv, '/U')
128  undefines = ['-U' + define for define in undefines]
129
130  fatbin_options, argv = GetOptionValue(argv, '-Xcuda-fatbinary')
131  fatbin_options = ['--fatbin-options=' + option for option in fatbin_options]
132
133  # The rest of the unrecognized options should be passed to host compiler
134  host_compiler_options = [option for option in argv if option not in (src_files + out_file)]
135
136  m_options = ["-m64"]
137
138  nvccopts = ['-D_FORCE_INLINES']
139  compute_capabilities, argv = GetOptionValue(argv, "--cuda-gpu-arch")
140  for capability in compute_capabilities:
141    capability = capability[len('sm_'):]
142    nvccopts += [
143        r'-gencode=arch=compute_%s,"code=sm_%s"' % (capability, capability)
144    ]
145  compute_capabilities, argv = GetOptionValue(argv, '--cuda-include-ptx')
146  for capability in compute_capabilities:
147    capability = capability[len('sm_'):]
148    nvccopts += [
149        r'-gencode=arch=compute_%s,"code=compute_%s"' % (capability, capability)
150    ]
151  _, argv = GetOptionValue(argv, '--no-cuda-include-ptx')
152
153  # nvcc doesn't respect the INCLUDE and LIB env vars from MSVC,
154  # so we explicity specify the system include paths and library search paths.
155  if 'INCLUDE' in os.environ:
156    nvccopts += [('--system-include="%s"' % p) for p in os.environ['INCLUDE'].split(";")]
157  if 'LIB' in os.environ:
158    nvccopts += [('--library-path="%s"' % p) for p in os.environ['LIB'].split(";")]
159
160  nvccopts += nvcc_compiler_options
161  nvccopts += undefines
162  nvccopts += defines
163  nvccopts += m_options
164  nvccopts += fatbin_options
165  nvccopts += ['--compiler-options=' + ",".join(host_compiler_options)]
166  nvccopts += ['-x', 'cu'] + opt + includes + out + ['-c'] + src_files
167  # Specify a unique temp directory for nvcc to generate intermediate files,
168  # then Bazel can ignore files under NVCC_TEMP_DIR during dependency check
169  # http://docs.nvidia.com/cuda/cuda-compiler-driver-nvcc/index.html#options-for-guiding-compiler-driver
170  # Different actions are sharing NVCC_TEMP_DIR, so we cannot remove it if the directory already exists.
171  if os.path.isfile(NVCC_TEMP_DIR):
172    os.remove(NVCC_TEMP_DIR)
173  if not os.path.exists(NVCC_TEMP_DIR):
174    os.makedirs(NVCC_TEMP_DIR)
175  # Provide a unique dir for each compiling action to avoid conflicts.
176  tempdir = tempfile.mkdtemp(dir = NVCC_TEMP_DIR)
177  nvccopts += ['--keep', '--keep-dir', tempdir]
178  # Force C++17 dialect (note, everything in just one string!)
179  nvccopts += ['--std c++17']
180  if log:
181    Log([NVCC_PATH] + nvccopts)
182
183  # Store command line options in a file to avoid hitting the character limit.
184  optsfile = tempfile.NamedTemporaryFile(mode='w', dir=tempdir, delete=False)
185  optsfile.write("\n".join(nvccopts))
186  optsfile.close()
187
188  proc = subprocess.Popen([NVCC_PATH, "--options-file", optsfile.name],
189                          stdout=sys.stdout,
190                          stderr=sys.stderr,
191                          env=os.environ.copy(),
192                          shell=True)
193  proc.wait()
194  return proc.returncode
195
196def main():
197  parser = ArgumentParser()
198  parser.add_argument('-x', nargs=1)
199  parser.add_argument('--cuda_log', action='store_true')
200  args, leftover = parser.parse_known_args(sys.argv[1:])
201
202  if args.x and args.x[0] == 'cuda':
203    if args.cuda_log: Log('-x cuda')
204    if args.cuda_log: Log('using nvcc')
205    return InvokeNvcc(leftover, log=args.cuda_log)
206
207  # Strip our flags before passing through to the CPU compiler for files which
208  # are not -x cuda. We can't just pass 'leftover' because it also strips -x.
209  # We not only want to pass -x to the CPU compiler, but also keep it in its
210  # relative location in the argv list (the compiler is actually sensitive to
211  # this).
212  cpu_compiler_flags = [flag for flag in sys.argv[1:]
213                             if not flag.startswith(('--cuda_log'))
214                             and not flag.startswith(('-nvcc_options'))]
215
216  return subprocess.call([CPU_COMPILER] + cpu_compiler_flags)
217
218if __name__ == '__main__':
219  sys.exit(main())
220