Skip to content

P
pytensor
提交 9a7c799b authored 作者: Frédéric Bastien's avatar Frédéric Bastien 提交者： GitHub
浏览文件
操作

Merge pull request #6146 from slefrancois/pr_win

add windows pr tests
上级 8136f204 b7d23606
隐藏空白字符变更
内嵌 并排
正在显示 包含 224 行增加 和 89 行删除
.jenkins/jenkins_buildbot_mac.sh
浏览文件 @ 9a7c799b
...@@ -17,6 +17,9 @@ export PATH=/usr/local/cuda/bin:${PATH} ...@@ -17,6 +17,9 @@ export PATH=/usr/local/cuda/bin:${PATH}
export DYLD_LIBRARY_PATH=/usr/local/cuda/lib:${DYLD_LIBRARY_PATH} export DYLD_LIBRARY_PATH=/usr/local/cuda/lib:${DYLD_LIBRARY_PATH}
export CPLUS_INCLUDE_PATH=/usr/local/cuda/include:${HOME}/cuda/include:${CPLUS_INCLUDE_PATH} export CPLUS_INCLUDE_PATH=/usr/local/cuda/include:${HOME}/cuda/include:${CPLUS_INCLUDE_PATH}
# CUDNN
export CUDNNPATH=${HOME}/cuda
# Build libgpuarray # Build libgpuarray
GPUARRAY_CONFIG="Release" GPUARRAY_CONFIG="Release"
DEVICE=cuda DEVICE=cuda
...@@ -61,4 +64,4 @@ set -x ...@@ -61,4 +64,4 @@ set -x
# Fast run and float32 # Fast run and float32
FILE=${BUILDBOT_DIR}/theano_python_fastrun_f32_tests.xml FILE=${BUILDBOT_DIR}/theano_python_fastrun_f32_tests.xml
NAME=mac_fastrun_f32 NAME=mac_fastrun_f32
THEANO_FLAGS=$THEANO_FLAGS,base_compiledir=$BASECOMPILEDIR,mode=FAST_RUN,warn.ignore_bug_before=all,on_opt_error=raise,on_shape_error=raise,floatX=float32,dnn.library_path=${HOME}/cuda/lib,gcc.cxxflags="-L${LIBDIR}/lib" python bin/theano-nose ${THEANO_PARAM} ${XUNIT}${FILE} ${SUITE}${NAME} THEANO_FLAGS=$THEANO_FLAGS,base_compiledir=$BASECOMPILEDIR,mode=FAST_RUN,warn.ignore_bug_before=all,on_opt_error=raise,on_shape_error=raise,floatX=float32,dnn.base_path=${CUDNNPATH},gcc.cxxflags="-L${LIBDIR}/lib" python bin/theano-nose ${THEANO_PARAM} ${XUNIT}${FILE} ${SUITE}${NAME}
\ No newline at end of file \ No newline at end of file
.jenkins/jenkins_buildbot_windows.bat
浏览文件 @ 9a7c799b
REM CUDNN PATH
set CUDNNPATH=C:\lib\cuda
REM Set conda python, cuda, cmake path
set PATH=%PATH%;C:\ProgramData\Miniconda2;%CUDNNPATH%\bin;C:\Program Files\CMake\bin
set BUILDBOT_DIR=%WORKSPACE%\nightly_build set BUILDBOT_DIR=%WORKSPACE%\nightly_build
set COMPILEDIR=C:\\Jenkins\\theano_cache\\buildbot_windows set COMPILEDIR=C:\Jenkins\theano_cache\buildbot_windows
REM Set test reports using nosetests xunit REM Set test reports using nosetests xunit
set XUNIT=--with-xunit --xunit-file= set XUNIT=--with-xunit --xunit-file=
...@@ -13,6 +19,7 @@ REM Build libgpuarray ...@@ -13,6 +19,7 @@ REM Build libgpuarray
set GPUARRAY_CONFIG="Release" set GPUARRAY_CONFIG="Release"
set DEVICE=cuda set DEVICE=cuda
set LIBDIR=%WORKSPACE%\local set LIBDIR=%WORKSPACE%\local
set PATH=%PATH%;%LIBDIR%\bin
REM Make fresh clones of libgpuarray (with no history since we dont need it) REM Make fresh clones of libgpuarray (with no history since we dont need it)
rmdir libgpuarray /s/q rmdir libgpuarray /s/q
...@@ -23,18 +30,14 @@ rmdir %LIBDIR% /s/q ...@@ -23,18 +30,14 @@ rmdir %LIBDIR% /s/q
mkdir %LIBDIR% mkdir %LIBDIR%
REM Build libgpuarray REM Build libgpuarray
set PATH=%PATH%;C:\Program Files\CMake\bin
mkdir libgpuarray\build mkdir libgpuarray\build
cd libgpuarray\build cd libgpuarray\build
cmake .. -DCMAKE_BUILD_TYPE=%GPUARRAY_CONFIG% -G "NMake Makefiles" cmake .. -DCMAKE_BUILD_TYPE=%GPUARRAY_CONFIG% -G "NMake Makefiles" -DCMAKE_INSTALL_PREFIX=%LIBDIR%
nmake nmake
cmake --build . --target install
cd ..\.. cd ..\..
REM Copy lib and export paths REM Set conda gcc path
C:\Windows\System32\robocopy /E libgpuarray C:\lib\libgpuarray > nul
set PATH=%PATH%;C:\libgpuarray\lib;C:\lib\cuda\bin
REM Set conda python path
set PATH=%PATH%;C:\ProgramData\Miniconda2;C:\ProgramData\Miniconda2\Library\mingw-w64\bin;C:\ProgramData\Miniconda2\Library\usr\bin;C:\ProgramData\Miniconda2\Library\bin;C:\ProgramData\Miniconda2\Scripts set PATH=%PATH%;C:\ProgramData\Miniconda2;C:\ProgramData\Miniconda2\Library\mingw-w64\bin;C:\ProgramData\Miniconda2\Library\usr\bin;C:\ProgramData\Miniconda2\Library\bin;C:\ProgramData\Miniconda2\Scripts
REM Build the pygpu modules REM Build the pygpu modules
...@@ -52,5 +55,5 @@ echo "Directory of stdout/stderr %BUILDBOT_DIR%" ...@@ -52,5 +55,5 @@ echo "Directory of stdout/stderr %BUILDBOT_DIR%"
REM Fast run and float32 REM Fast run and float32
set FILE=%BUILDBOT_DIR%\theano_python2_fastrun_f32_tests.xml set FILE=%BUILDBOT_DIR%\theano_python2_fastrun_f32_tests.xml
set NAME=win_fastrun_f32 set NAME=win_fastrun_f32
set THEANO_FLAGS=%THEANO_FLAGS%,compiledir=%COMPILEDIR%,mode=FAST_RUN,warn.ignore_bug_before=all,on_opt_error=raise,on_shape_error=raise,floatX=float32,dnn.include_path=C:\\lib\\cuda\\include,dnn.library_path=C:\\lib\\cuda\\lib\\x64,gcc.cxxflags='-I"C:\\Program Files\\NVIDIA GPU Computing Toolkit\\CUDA\\v8.0\\include" -I"C:\\lib\\libgpuarray\\src" -L"C:\\Program Files\\NVIDIA GPU Computing Toolkit\CUDA\\v8.0\\lib\\x64" -LC:\\lib\\libgpuarray\\lib' set THEANO_FLAGS=%THEANO_FLAGS%,compiledir=%COMPILEDIR:\=\\%,mode=FAST_RUN,warn.ignore_bug_before=all,on_opt_error=raise,on_shape_error=raise,floatX=float32,dnn.base_path="%CUDNNPATH%",gcc.cxxflags='-I%LIBDIR:\=\\%\\include -L%LIBDIR:\=\\%\\lib'
python bin\theano-nose %THEANO_PARAM% %XUNIT%%FILE% %SUITE%%NAME% python bin\theano-nose %THEANO_PARAM% %XUNIT%%FILE% %SUITE%%NAME%
.jenkins/jenkins_pr_mac.sh
浏览文件 @ 9a7c799b
...@@ -18,6 +18,9 @@ export PATH=/usr/local/cuda/bin:${PATH} ...@@ -18,6 +18,9 @@ export PATH=/usr/local/cuda/bin:${PATH}
export DYLD_LIBRARY_PATH=/usr/local/cuda/lib:${DYLD_LIBRARY_PATH} export DYLD_LIBRARY_PATH=/usr/local/cuda/lib:${DYLD_LIBRARY_PATH}
export CPLUS_INCLUDE_PATH=/usr/local/cuda/include:${HOME}/cuda/include:${CPLUS_INCLUDE_PATH} export CPLUS_INCLUDE_PATH=/usr/local/cuda/include:${HOME}/cuda/include:${CPLUS_INCLUDE_PATH}
# CUDNN
export CUDNNPATH=${HOME}/cuda
# Build libgpuarray # Build libgpuarray
GPUARRAY_CONFIG="Release" GPUARRAY_CONFIG="Release"
DEVICE=cuda DEVICE=cuda
...@@ -53,5 +56,5 @@ python -c 'import pygpu; print(pygpu.__file__)' ...@@ -53,5 +56,5 @@ python -c 'import pygpu; print(pygpu.__file__)'
# Testing theano # Testing theano
THEANO_PARAM="theano --with-timer --timer-top-n 10 --with-xunit --xunit-file=theano_mac_pr_tests.xml" THEANO_PARAM="theano --with-timer --timer-top-n 10 --with-xunit --xunit-file=theano_mac_pr_tests.xml"
FLAGS=init_gpu_device=$DEVICE,gpuarray.preallocate=1000,mode=FAST_RUN,on_opt_error=raise,on_shape_error=raise,cmodule.age_thresh_use=604800,base_compiledir=$BASECOMPILEDIR,dnn.library_path=$HOME/cuda/lib,gcc.cxxflags="-L${LIBDIR}/lib" FLAGS=init_gpu_device=$DEVICE,gpuarray.preallocate=1000,mode=FAST_RUN,on_opt_error=raise,on_shape_error=raise,cmodule.age_thresh_use=604800,base_compiledir=$BASECOMPILEDIR,dnn.base_path=${CUDNNPATH},gcc.cxxflags="-L${LIBDIR}/lib"
THEANO_FLAGS=${FLAGS} python bin/theano-nose ${THEANO_PARAM} THEANO_FLAGS=${FLAGS} python bin/theano-nose ${THEANO_PARAM}
.jenkins/jenkins_pr_win.bat 0 → 100644
浏览文件 @ 9a7c799b
REM CUDNN PATH
set CUDNNPATH=C:\lib\cuda
REM Set conda python, cuda, cmake path
set PATH=%PATH%;C:\ProgramData\Miniconda2;%CUDNNPATH%\bin;C:\Program Files\CMake\bin
REM Set cache dir and copy from master
set COMPILEDIR=%WORKSPACE%\cache
C:\Windows\System32\robocopy /E /purge C:\Jenkins\theano_cache\buildbot_windows %COMPILEDIR% > nul
set THEANO_FLAGS=init_gpu_device=cuda
REM Build libgpuarray
set GPUARRAY_CONFIG="Release"
set DEVICE=cuda
set LIBDIR=%WORKSPACE%\local
set PATH=%PATH%;%LIBDIR%\bin
REM Make fresh clones of libgpuarray (with no history since we dont need it)
rmdir libgpuarray /s/q
git clone --depth 1 "https://github.com/Theano/libgpuarray.git"
REM Clean up previous installs (to make sure no old files are left)
rmdir %LIBDIR% /s/q
mkdir %LIBDIR%
REM Build libgpuarray
rmdir libgpuarray\build /s/q
mkdir libgpuarray\build
cd libgpuarray\build
cmake .. -DCMAKE_BUILD_TYPE=%GPUARRAY_CONFIG% -G "NMake Makefiles" -DCMAKE_INSTALL_PREFIX=%LIBDIR%
nmake
cmake --build . --target install
cd ..\..
REM Add conda gcc toolchain path
set PATH=%PATH%;C:\ProgramData\Miniconda2\Library\mingw-w64\bin;C:\ProgramData\Miniconda2\Library\usr\bin;C:\ProgramData\Miniconda2\Library\bin;C:\ProgramData\Miniconda2\Scripts
REM Build the pygpu modules
cd libgpuarray
python setup.py build_ext --inplace
mkdir %LIBDIR%\lib\python
set PYTHONPATH=%PYTHONPATH%;%LIBDIR%\lib\python
REM Then install
python setup.py install --home=%LIBDIR%
cd ..
set THEANO_PARAM=theano --with-timer --timer-top-n 10 --with-xunit --xunit-file=theano_win_pr_tests.xml
set NAME=pr_win
set THEANO_FLAGS=%THEANO_FLAGS%,mode=FAST_RUN,floatX=float32,on_opt_error=raise,on_shape_error=raise,cmodule.age_thresh_use=604800,compiledir=%COMPILEDIR:\=\\%,dnn.base_path="%CUDNNPATH%",gcc.cxxflags='-I%LIBDIR:\=\\%\\include -L%LIBDIR:\=\\%\\lib'
python bin\theano-nose %THEANO_PARAM% --xunit-testsuite-name=%NAME%
theano/configdefaults.py
浏览文件 @ 9a7c799b
...@@ -232,6 +232,41 @@ AddConfigVar('gpuarray.single_stream', ...@@ -232,6 +232,41 @@ AddConfigVar('gpuarray.single_stream',
in_c_key=False) in_c_key=False)
def get_cuda_root():
# We look for the cuda path since we need headers from there
v = os.getenv('CUDA_ROOT', "")
if v:
return v
v = os.getenv('CUDA_PATH', "")
if v:
return v
s = os.getenv("PATH")
if not s:
return ''
for dir in s.split(os.path.pathsep):
if os.path.exists(os.path.join(dir, "nvcc")):
return os.path.dirname(os.path.abspath(dir))
return ''
AddConfigVar('cuda.root',
"Location of the cuda installation",
StrParam(get_cuda_root),
in_c_key=False)
def default_cuda_include():
if theano.config.cuda.root:
return os.path.join(theano.config.cuda.root, 'include')
return ''
AddConfigVar('cuda.include_path',
"Location of the cuda includes",
StrParam(default_cuda_include),
in_c_key=False)
def safe_no_dnn_workmem(workmem): def safe_no_dnn_workmem(workmem):
""" """
Make sure the user is not attempting to use dnn.conv.workmem`. Make sure the user is not attempting to use dnn.conv.workmem`.
...@@ -325,39 +360,68 @@ AddConfigVar('dnn.conv.precision', ...@@ -325,39 +360,68 @@ AddConfigVar('dnn.conv.precision',
in_c_key=False) in_c_key=False)
def get_cuda_root(): # We want to default to the cuda root if cudnn is installed there
v = os.getenv('CUDA_ROOT', "") def default_dnn_base_path():
if v: root = get_cuda_root()
return v # The include doesn't change location between OS.
s = os.getenv("PATH") if root and os.path.exists(os.path.join(root, 'include', 'cudnn.h')):
if not s: return root
return ''
for dir in s.split(os.path.pathsep):
if os.path.exists(os.path.join(dir, "nvcc")):
return os.path.dirname(os.path.abspath(dir))
return '' return ''
def default_dnn_include_path(): AddConfigVar('dnn.base_path',
cuda_root = get_cuda_root() "Install location of cuDNN.",
if cuda_root == '': StrParam(default_dnn_base_path),
return '' in_c_key=False)
return os.path.join(cuda_root, 'include')
def default_dnn_inc_path():
if theano.config.dnn.base_path != '':
return os.path.join(theano.config.dnn.base_path, 'include')
return ''
AddConfigVar('dnn.include_path', AddConfigVar('dnn.include_path',
"Location of the cudnn header (defaults to the cuda root)", "Location of the cudnn header",
# We keep the default here since cudnn needs headers from cuda StrParam(default_dnn_inc_path),
StrParam(default_dnn_include_path()),
# Added elsewhere in the c key only when needed.
in_c_key=False) in_c_key=False)
def default_dnn_lib_path():
if theano.config.dnn.base_path != '':
path = os.path.join(theano.config.dnn.base_path, 'lib')
if sys.platform == 'win32':
path = os.path.join(path, 'x64')
return path
return ''
AddConfigVar('dnn.library_path', AddConfigVar('dnn.library_path',
"Location of the cudnn library (defaults to the cuda root)", "Location of the cudnn link library.",
StrParam(''), StrParam(default_dnn_lib_path),
# Added elsewhere in the c key only when needed. in_c_key=False)
def default_dnn_bin_path():
if type(theano.config.dnn).base_path.is_default:
return ''
else:
if theano.config.dnn.base_path != '':
if sys.platform == 'win32':
return os.path.join(theano.config.dnn.base_path, 'bin')
else:
return theano.config.dnn.library_path
return ''
AddConfigVar('dnn.bin_path',
"Location of the cuDNN load library "
"(on non-windows platforms, "
"this is the same as dnn.library_path)",
StrParam(default_dnn_bin_path),
in_c_key=False) in_c_key=False)
AddConfigVar('dnn.enabled', AddConfigVar('dnn.enabled',
"'auto', use cuDNN if available, but silently fall back" "'auto', use cuDNN if available, but silently fall back"
" to not using it if not present." " to not using it if not present."
......
theano/gof/cc.py
浏览文件 @ 9a7c799b
...@@ -1065,7 +1065,8 @@ class CLinker(link.Linker): ...@@ -1065,7 +1065,8 @@ class CLinker(link.Linker):
ret += x.c_header_dirs() ret += x.c_header_dirs()
except utils.MethodNotDefined: except utils.MethodNotDefined:
pass pass
return utils.uniq(ret) # filter out empty strings/None
return [r for r in utils.uniq(ret) if r]
def libraries(self): def libraries(self):
""" """
...@@ -1107,7 +1108,8 @@ class CLinker(link.Linker): ...@@ -1107,7 +1108,8 @@ class CLinker(link.Linker):
ret += x.c_lib_dirs() ret += x.c_lib_dirs()
except utils.MethodNotDefined: except utils.MethodNotDefined:
pass pass
return utils.uniq(ret) # filter out empty strings/None
return [r for r in utils.uniq(ret) if r]
def __compile__(self, input_storage=None, output_storage=None, def __compile__(self, input_storage=None, output_storage=None,
storage_map=None, keep_lock=False): storage_map=None, keep_lock=False):
......
theano/gof/type.py
浏览文件 @ 9a7c799b
...@@ -658,28 +658,21 @@ class CDataType(Type): ...@@ -658,28 +658,21 @@ class CDataType(Type):
""" """
__props__ = ('ctype', 'freefunc', 'headers', 'header_dirs', __props__ = ('ctype', 'freefunc', 'headers', 'header_dirs',
'libraries', 'lib_dirs', 'extra_support_code', 'libraries', 'lib_dirs', 'extra_support_code',
'version') 'compile_args', 'version')
def __init__(self, ctype, freefunc=None, headers=None, header_dirs=None, def __init__(self, ctype, freefunc=None, headers=(), header_dirs=(),
libraries=None, lib_dirs=None, extra_support_code="", libraries=(), lib_dirs=(), compile_args=(),
version=None): extra_support_code="", version=None):
assert isinstance(ctype, string_types) assert isinstance(ctype, string_types)
self.ctype = ctype self.ctype = ctype
if freefunc is not None: if freefunc is not None:
assert isinstance(freefunc, string_types) assert isinstance(freefunc, string_types)
self.freefunc = freefunc self.freefunc = freefunc
if headers is None:
headers = ()
self.headers = tuple(headers) self.headers = tuple(headers)
if header_dirs is None:
header_dirs = ()
self.header_dirs = tuple(header_dirs) self.header_dirs = tuple(header_dirs)
if libraries is None:
libraries = ()
self.libraries = tuple(libraries) self.libraries = tuple(libraries)
if lib_dirs is None:
lib_dirs = ()
self.lib_dirs = tuple(lib_dirs) self.lib_dirs = tuple(lib_dirs)
self.compile_args = tuple(compile_args)
self.extra_support_code = extra_support_code self.extra_support_code = extra_support_code
self._fn = None self._fn = None
self.version = None self.version = None
...@@ -787,6 +780,9 @@ if (py_%(name)s == NULL) { %(freefunc)s(%(name)s); } ...@@ -787,6 +780,9 @@ if (py_%(name)s == NULL) { %(freefunc)s(%(name)s); }
def c_lib_dirs(self): def c_lib_dirs(self):
return self.lib_dirs return self.lib_dirs
def c_compile_args(self):
return self.compile_args
def c_code_cache_version(self): def c_code_cache_version(self):
v = (3, ) v = (3, )
if self.version is not None: if self.version is not None:
......
theano/gpuarray/dnn.py
浏览文件 @ 9a7c799b
...@@ -72,16 +72,16 @@ def _dnn_lib(): ...@@ -72,16 +72,16 @@ def _dnn_lib():
if _dnn_lib.handle is None: if _dnn_lib.handle is None:
import ctypes.util import ctypes.util
if config.dnn.library_path != "": if config.dnn.bin_path != "":
if sys.platform == 'darwin': if sys.platform == 'darwin':
dnn_handle = _load_lib(os.path.join(config.dnn.library_path, 'libcudnn.dylib')) dnn_handle = _load_lib(os.path.join(config.dnn.bin_path, 'libcudnn.dylib'))
elif sys.platform == 'win32': elif sys.platform == 'win32':
for name in WIN32_CUDNN_NAMES: for name in WIN32_CUDNN_NAMES:
dnn_handle = _load_lib(os.path.join(config.dnn.library_path, name)) dnn_handle = _load_lib(os.path.join(config.dnn.bin_path, name))
if dnn_handle is not None: if dnn_handle is not None:
break break
else: else:
dnn_handle = _load_lib(os.path.join(config.dnn.library_path, 'libcudnn.so')) dnn_handle = _load_lib(os.path.join(config.dnn.bin_path, 'libcudnn.so'))
else: else:
lib_name = ctypes.util.find_library('cudnn') lib_name = ctypes.util.find_library('cudnn')
if lib_name is None and sys.platform == 'win32': if lib_name is None and sys.platform == 'win32':
...@@ -135,12 +135,13 @@ if ((err = cudnnCreate(&_handle)) != CUDNN_STATUS_SUCCESS) { ...@@ -135,12 +135,13 @@ if ((err = cudnnCreate(&_handle)) != CUDNN_STATUS_SUCCESS) {
} }
""" """
params = ["-l", "cudnn", "-I" + os.path.dirname(__file__)]
path_wrapper = "\"" if os.name == 'nt' else "" path_wrapper = "\"" if os.name == 'nt' else ""
params = ["-l", "cudnn"] params = ["-l", "cudnn"]
params.extend(['-I%s%s%s' % (path_wrapper, os.path.dirname(__file__), path_wrapper)]) params.extend(['-I%s%s%s' % (path_wrapper, os.path.dirname(__file__), path_wrapper)])
if config.dnn.include_path: if config.dnn.include_path:
params.extend(['-I%s%s%s' % (path_wrapper, config.dnn.include_path, path_wrapper)]) params.extend(['-I%s%s%s' % (path_wrapper, config.dnn.include_path, path_wrapper)])
if config.cuda.include_path:
params.extend(['-I%s%s%s' % (path_wrapper, config.cuda.include_path, path_wrapper)])
if config.dnn.library_path: if config.dnn.library_path:
params.extend(['-L%s%s%s' % (path_wrapper, config.dnn.library_path, path_wrapper)]) params.extend(['-L%s%s%s' % (path_wrapper, config.dnn.library_path, path_wrapper)])
# Do not run here the test program. It would run on the # Do not run here the test program. It would run on the
...@@ -223,6 +224,24 @@ def dnn_available(context_name): ...@@ -223,6 +224,24 @@ def dnn_available(context_name):
dnn_available.msg = None dnn_available.msg = None
def CUDNNDataType(name, freefunc=None):
cargs = []
if config.dnn.bin_path:
if sys.platform == 'darwin':
cargs.append('-Wl,-rpath,' + config.dnn.bin_path)
else:
cargs.append('-Wl,-rpath,"' + config.dnn.bin_path + '"')
return CDataType(name, freefunc,
headers=['cudnn.h'],
header_dirs=[config.dnn.include_path,
config.cuda.include_path],
libraries=['cudnn'],
lib_dirs=[config.dnn.library_path],
compile_args=cargs,
version=version(raises=False))
class DnnVersion(Op): class DnnVersion(Op):
__props__ = () __props__ = ()
...@@ -230,22 +249,20 @@ class DnnVersion(Op): ...@@ -230,22 +249,20 @@ class DnnVersion(Op):
return ['cudnn.h'] return ['cudnn.h']
def c_header_dirs(self): def c_header_dirs(self):
return [config.dnn.include_path] if config.dnn.include_path else [] return [config.dnn.include_path, config.cuda.include_path]
def c_libraries(self): def c_libraries(self):
return ['cudnn'] return ['cudnn']
def c_lib_dirs(self): def c_lib_dirs(self):
if config.dnn.library_path: return [config.dnn.library_path]
return [config.dnn.library_path]
return []
def c_compile_args(self): def c_compile_args(self):
if config.dnn.library_path: if config.dnn.bin_path:
if sys.platform == 'darwin': if sys.platform == 'darwin':
return ['-Wl,-rpath,' + config.dnn.library_path] return ['-Wl,-rpath,' + config.dnn.bin_path]
else: else:
return ['-Wl,-rpath,"' + config.dnn.library_path + '"'] return ['-Wl,-rpath,"' + config.dnn.bin_path + '"']
return [] return []
def c_support_code(self): def c_support_code(self):
...@@ -305,12 +322,7 @@ def version(raises=True): ...@@ -305,12 +322,7 @@ def version(raises=True):
return version.v return version.v
version.v = None version.v = None
handle_type = CDataType('cudnnHandle_t', 'cudnnDestroy', handle_type = CUDNNDataType('cudnnHandle_t', 'cudnnDestroy')
headers=['cudnn.h'],
header_dirs=[config.dnn.include_path],
libraries=['cudnn'],
lib_dirs=[config.dnn.library_path],
version=version(raises=False))
# Get cuDNN definitions to be used. # Get cuDNN definitions to be used.
cudnn = cudnn_defs.get_definitions(version(raises=False)) cudnn = cudnn_defs.get_definitions(version(raises=False))
...@@ -368,25 +380,21 @@ class DnnBase(COp): ...@@ -368,25 +380,21 @@ class DnnBase(COp):
'gpuarray_helper.h'] 'gpuarray_helper.h']
def c_header_dirs(self): def c_header_dirs(self):
dirs = [os.path.dirname(__file__), pygpu.get_include()] return [os.path.dirname(__file__), pygpu.get_include(),
if config.dnn.include_path: config.dnn.include_path, config.cuda.include_path]
dirs.append(config.dnn.include_path)
return dirs
def c_libraries(self): def c_libraries(self):
return ['cudnn', 'gpuarray'] return ['cudnn', 'gpuarray']
def c_lib_dirs(self): def c_lib_dirs(self):
if config.dnn.library_path: return [config.dnn.library_path]
return [config.dnn.library_path]
return []
def c_compile_args(self): def c_compile_args(self):
if config.dnn.library_path: if config.dnn.bin_path:
if sys.platform == 'darwin': if sys.platform == 'darwin':
return ['-Wl,-rpath,' + config.dnn.library_path] return ['-Wl,-rpath,' + config.dnn.bin_path]
else: else:
return ['-Wl,-rpath,"' + config.dnn.library_path + '"'] return ['-Wl,-rpath,"' + config.dnn.bin_path + '"']
return [] return []
def c_code_cache_version(self): def c_code_cache_version(self):
...@@ -418,7 +426,8 @@ class GpuDnnConvDesc(COp): ...@@ -418,7 +426,8 @@ class GpuDnnConvDesc(COp):
return ['cudnn.h', 'cudnn_helper.h'] return ['cudnn.h', 'cudnn_helper.h']
def c_header_dirs(self): def c_header_dirs(self):
return [os.path.dirname(__file__), config.dnn.include_path] return [os.path.dirname(__file__), config.dnn.include_path,
config.cuda.include_path]
def c_libraries(self): def c_libraries(self):
return ['cudnn'] return ['cudnn']
...@@ -426,6 +435,14 @@ class GpuDnnConvDesc(COp): ...@@ -426,6 +435,14 @@ class GpuDnnConvDesc(COp):
def c_lib_dirs(self): def c_lib_dirs(self):
return [config.dnn.library_path] return [config.dnn.library_path]
def c_compile_args(self):
if config.dnn.bin_path:
if sys.platform == 'darwin':
return ['-Wl,-rpath,' + config.dnn.bin_path]
else:
return ['-Wl,-rpath,"' + config.dnn.bin_path + '"']
return []
def do_constant_folding(self, node): def do_constant_folding(self, node):
return False return False
...@@ -466,9 +483,8 @@ class GpuDnnConvDesc(COp): ...@@ -466,9 +483,8 @@ class GpuDnnConvDesc(COp):
kern_shape = theano.tensor.basic.cast(kern_shape, 'int64') kern_shape = theano.tensor.basic.cast(kern_shape, 'int64')
node = Apply(self, [kern_shape], node = Apply(self, [kern_shape],
[CDataType("cudnnConvolutionDescriptor_t", [CUDNNDataType("cudnnConvolutionDescriptor_t",
freefunc="cudnnDestroyConvolutionDescriptor", freefunc="cudnnDestroyConvolutionDescriptor")()])
version=version(raises=False))()])
# DebugMode cannot compare the values of CDataType variables, so by # DebugMode cannot compare the values of CDataType variables, so by
# default it returns False all the time. To prevent DebugMode from # default it returns False all the time. To prevent DebugMode from
# complaining because of the MergeOptimizer, we make this variable # complaining because of the MergeOptimizer, we make this variable
...@@ -1278,9 +1294,8 @@ class GpuDnnPoolDesc(Op): ...@@ -1278,9 +1294,8 @@ class GpuDnnPoolDesc(Op):
def make_node(self): def make_node(self):
node = Apply(self, [], node = Apply(self, [],
[CDataType("cudnnPoolingDescriptor_t", [CUDNNDataType("cudnnPoolingDescriptor_t",
freefunc="cudnnDestroyPoolingDescriptor", freefunc="cudnnDestroyPoolingDescriptor")()])
version=version(raises=False))()])
# DebugMode cannot compare the values of CDataType variables, so by # DebugMode cannot compare the values of CDataType variables, so by
# default it returns False all the time. To prevent DebugMode from # default it returns False all the time. To prevent DebugMode from
# complaining because of the MergeOptimizer, we make this variable # complaining because of the MergeOptimizer, we make this variable
...@@ -1938,9 +1953,8 @@ class GpuDnnBatchNormGrad(DnnBase): ...@@ -1938,9 +1953,8 @@ class GpuDnnBatchNormGrad(DnnBase):
return [shape[0], shape[2], shape[2]] return [shape[0], shape[2], shape[2]]
gpudata_type = CDataType('gpudata *', 'gpudata_release') gpudata_type = CDataType('gpudata *', 'gpudata_release')
dropoutdesc_type = CDataType('cudnnDropoutDescriptor_t', dropoutdesc_type = CUDNNDataType('cudnnDropoutDescriptor_t',
'cudnnDestroyDropoutDescriptor', 'cudnnDestroyDropoutDescriptor')
version=version(raises=False))
class GpuDnnDropoutOp(DnnBase): class GpuDnnDropoutOp(DnnBase):
...@@ -2008,9 +2022,8 @@ def dropout(x, dropout=0.0, seed=4242): ...@@ -2008,9 +2022,8 @@ def dropout(x, dropout=0.0, seed=4242):
y, odesc = GpuDnnDropoutOp()(x, desc) y, odesc = GpuDnnDropoutOp()(x, desc)
return y, desc, odesc, states return y, desc, odesc, states
rnndesc_type = CDataType('cudnnRNNDescriptor_t', rnndesc_type = CUDNNDataType('cudnnRNNDescriptor_t',
'cudnnDestroyRNNDescriptor', 'cudnnDestroyRNNDescriptor')
version=version(raises=False))
def as_i32(v): def as_i32(v):
......
Markdown 格式
0% 或
您添加了 0 人 到此讨论。请谨慎行事。
请先完成此评论的编辑！
请 注册 或者 后发表评论