1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
|
import itertools
import os
import os.path
import sys
import json
import subprocess
import optparse
import textwrap
# Explicitly enable local imports
# Don't forget to add imported scripts to inputs of the calling command!
sys.path.append(os.path.dirname(os.path.abspath(__file__)))
import process_command_files as pcf
import thinlto_cache
from process_whole_archive_option import ProcessWholeArchiveOption
CUDA_LIBRARIES = {
'-lcublas_static': '-lcublas',
'-lcublasLt_static': '-lcublasLt',
'-lcudart_static': '-lcudart',
'-lcudnn_static': '-lcudnn',
'-lcudnn_adv_infer_static': '-lcudnn',
'-lcudnn_adv_train_static': '-lcudnn',
'-lcudnn_cnn_infer_static': '-lcudnn',
'-lcudnn_cnn_train_static': '-lcudnn',
'-lcudnn_ops_infer_static': '-lcudnn',
'-lcudnn_ops_train_static': '-lcudnn',
'-lcufft_static_nocallback': '-lcufft',
'-lcupti_static': '-lcupti',
'-lcurand_static': '-lcurand',
'-lcusolver_static': '-lcusolver',
'-lcusparse_static': '-lcusparse',
'-lmyelin_compiler_static': '-lmyelin',
'-lmyelin_executor_static': '-lnvcaffe_parser',
'-lmyelin_pattern_library_static': '',
'-lmyelin_pattern_runtime_static': '',
'-lnvinfer_static': '-lnvinfer',
'-lnvinfer_plugin_static': '-lnvinfer_plugin',
'-lnvonnxparser_static': '-lnvonnxparser',
'-lnvparsers_static': '-lnvparsers',
'-lnvrtc_static': '-lnvrtc',
'-lnvrtc-builtins_static': '-lnvrtc-builtins',
'-lnvptxcompiler_static': '',
'-lnppc_static': '-lnppc',
'-lnppial_static': '-lnppial',
'-lnppicc_static': '-lnppicc',
'-lnppicom_static': '-lnppicom',
'-lnppidei_static': '-lnppidei',
'-lnppif_static': '-lnppif',
'-lnppig_static': '-lnppig',
'-lnppim_static': '-lnppim',
'-lnppist_static': '-lnppist',
'-lnppisu_static': '-lnppisu',
'-lnppitc_static': '-lnppitc',
'-lnpps_static': '-lnpps',
}
class CUDAManager:
def __init__(self, known_arches, nvprune_exe):
self.fatbin_libs = self._known_fatbin_libs(set(CUDA_LIBRARIES))
self.prune_args = []
if known_arches:
for arch in known_arches.split(':'):
self.prune_args.append('-gencode')
self.prune_args.append(self._arch_flag(arch))
self.nvprune_exe = nvprune_exe
def has_cuda_fatbins(self, cmd):
return bool(set(cmd) & self.fatbin_libs)
@property
def can_prune_libs(self):
return self.prune_args and self.nvprune_exe
def _known_fatbin_libs(self, libs):
libs_wo_device_code = {
'-lcudart_static',
'-lcupti_static',
'-lnppc_static',
}
return set(libs) - libs_wo_device_code
def _arch_flag(self, arch):
_, ver = arch.split('_', 1)
return 'arch=compute_{},code={}'.format(ver, arch)
def prune_lib(self, inp_fname, out_fname):
if self.prune_args:
prune_command = [self.nvprune_exe] + self.prune_args + ['--output-file', out_fname, inp_fname]
subprocess.check_call(prune_command)
def write_linker_script(self, f):
# This script simply says:
# * Place all `.nv_fatbin` input sections from all input files into one `.nv_fatbin` output section of output file
# * Place it after `.bss` section
#
# Motivation can be found here: https://maskray.me/blog/2021-07-04-sections-and-overwrite-sections#insert-before-and-insert-after
# TL;DR - we put section with a lot of GPU code directly after the last meaningful section in the binary
# (which turns out to be .bss)
# In that case, we decrease chances of relocation overflows from .text to .bss,
# because now these sections are close to each other
script = textwrap.dedent("""
SECTIONS {
.nv_fatbin : { *(.nv_fatbin) }
} INSERT AFTER .bss
""").strip()
f.write(script)
def tmpdir_generator(base_path, prefix):
for idx in itertools.count():
path = os.path.abspath(os.path.join(base_path, prefix + '_' + str(idx)))
os.makedirs(path)
yield path
def process_cuda_library_by_external_tool(cmd, build_root, tool_name, callable_tool_executor, allowed_cuda_libs):
tmpdir_gen = tmpdir_generator(build_root, 'cuda_' + tool_name + '_libs')
new_flags = []
cuda_deps = set()
# Because each directory flag only affects flags that follow it,
# for correct pruning we need to process that in reversed order
for flag in reversed(cmd):
if flag in allowed_cuda_libs:
cuda_deps.add('lib' + flag[2:] + '.a')
flag += '_' + tool_name
elif flag.startswith('-L') and os.path.exists(flag[2:]) and os.path.isdir(flag[2:]) and any(f in cuda_deps for f in os.listdir(flag[2:])):
from_dirpath = flag[2:]
from_deps = list(cuda_deps & set(os.listdir(from_dirpath)))
if from_deps:
to_dirpath = next(tmpdir_gen)
for f in from_deps:
from_path = os.path.join(from_dirpath, f)
to_path = os.path.join(to_dirpath, f[:-2] + '_' + tool_name +'.a')
callable_tool_executor(from_path, to_path)
cuda_deps.remove(f)
# do not remove current directory
# because it can contain other libraries we want link to
# instead we just add new directory with processed by tool libs
new_flags.append('-L' + to_dirpath)
new_flags.append(flag)
assert not cuda_deps, ('Unresolved CUDA deps: ' + ','.join(cuda_deps))
return reversed(new_flags)
def process_cuda_libraries_by_objcopy(cmd, build_root, objcopy_exe):
if not objcopy_exe:
return cmd
def run_objcopy(from_path, to_path):
rename_section_command = [objcopy_exe, "--rename-section", ".ctors=.init_array", from_path, to_path]
subprocess.check_call(rename_section_command)
possible_libraries = set(CUDA_LIBRARIES.keys())
possible_libraries.update([
'-lcudadevrt',
'-lcufilt',
'-lculibos',
])
possible_libraries.update([
lib_name + "_pruner" for lib_name in possible_libraries
])
return process_cuda_library_by_external_tool(list(cmd), build_root, 'objcopy', run_objcopy, possible_libraries)
def process_cuda_libraries_by_nvprune(cmd, cuda_manager, build_root):
if not cuda_manager.has_cuda_fatbins(cmd):
return cmd
# add custom linker script
to_dirpath = next(tmpdir_generator(build_root, 'cuda_linker_script'))
script_path = os.path.join(to_dirpath, 'script')
with open(script_path, 'w') as f:
cuda_manager.write_linker_script(f)
flags_with_linker = list(cmd) + ['-Wl,--script={}'.format(script_path)]
if not cuda_manager.can_prune_libs:
return flags_with_linker
return process_cuda_library_by_external_tool(flags_with_linker, build_root, 'pruner', cuda_manager.prune_lib, cuda_manager.fatbin_libs)
def remove_excessive_flags(cmd):
flags = []
for flag in cmd:
if not flag.endswith('.ios.interface') and not flag.endswith('.pkg.fake'):
flags.append(flag)
return flags
def fix_cmd_for_dynamic_cuda(cmd):
flags = []
for flag in cmd:
if flag in CUDA_LIBRARIES:
flags.append(CUDA_LIBRARIES[flag])
else:
flags.append(flag)
return flags
def remove_libs(cmd, libs):
excluded_flags = ['-l{}'.format(lib) for lib in libs]
flags = []
for flag in cmd:
if flag in excluded_flags:
continue
flags.append(flag)
return flags
def parse_args(args):
parser = optparse.OptionParser()
parser.disable_interspersed_args()
parser.add_option('--custom-step')
parser.add_option('--python')
parser.add_option('--source-root')
parser.add_option('--build-root')
parser.add_option('--clang-ver')
parser.add_option('--dynamic-cuda', action='store_true')
parser.add_option('--cuda-architectures',
help='List of supported CUDA architectures, separated by ":" (e.g. "sm_52:compute_70:lto_90a"')
parser.add_option('--nvprune-exe')
parser.add_option('--objcopy-exe')
parser.add_option('--arch')
parser.add_option('--linker-output')
parser.add_option('--whole-archive-peers', action='append')
parser.add_option('--whole-archive-libs', action='append')
parser.add_option('--exclude-libs', action='append')
thinlto_cache.add_options(parser)
return parser.parse_args(args)
if __name__ == '__main__':
args = sys.argv[1:]
plugins = []
if '--start-plugins' in args:
ib = args.index('--start-plugins')
ie = args.index('--end-plugins')
plugins = args[ib + 1:ie]
args = args[:ib] + args[ie + 1:]
for p in plugins:
res = subprocess.check_output([sys.executable, p, sys.argv[0]] + args).decode().strip()
if res:
args = json.loads(res)[1:]
opts, args = parse_args(args)
args = pcf.skip_markers(args)
cmd = args
cmd = remove_excessive_flags(cmd)
if opts.dynamic_cuda:
cmd = fix_cmd_for_dynamic_cuda(cmd)
else:
cuda_manager = CUDAManager(opts.cuda_architectures, opts.nvprune_exe)
cmd = process_cuda_libraries_by_nvprune(cmd, cuda_manager, opts.build_root)
cmd = process_cuda_libraries_by_objcopy(cmd, opts.build_root, opts.objcopy_exe)
if opts.exclude_libs:
cmd = remove_libs(cmd, opts.exclude_libs)
cmd = ProcessWholeArchiveOption(opts.arch, opts.whole_archive_peers, opts.whole_archive_libs).construct_cmd(cmd)
if opts.custom_step:
assert opts.python
subprocess.check_call([opts.python] + [opts.custom_step] + args)
if opts.linker_output:
stdout = open(opts.linker_output, 'w')
else:
stdout = sys.stdout
thinlto_cache.preprocess(opts, cmd)
rc = subprocess.call(cmd, shell=False, stderr=sys.stderr, stdout=stdout)
thinlto_cache.postprocess(opts)
sys.exit(rc)
|
