def get_argument(argument, option):
if requires_checked_cast(argument):
return CHECKED_USE.get(argument['type'],
'{}_').format(argument['name'])
elif argument['type'] == 'bool' and 'if_true' in argument:
if bool_option_is_string(argument):
tpl = '({}) ? "{}" : "{}"'
else:
tpl = '({}) ? {} : {}'
return tpl.format(argument['name'], argument['if_true'],
argument['if_false'])
elif argument['type'] == "CONSTANT":
if bool_option_is_string(
argument): # this is a bool that is actually a string...
return '"{}"'.format(argument['name'])
v = str(argument['name'])
for pattern, replacement in CONSTANT_REPLACEMENTS:
v = re.sub(pattern, replacement, v)
return CodeTemplate(v).substitute(backend_type_env)
# e.g. argument 0, i.e. repeat the 0th argument in this position...
elif argument['type'] == 'argument':
index = int(argument['name'])
return get_argument(option['arguments'][index], option)
else:
return argument['name']
def get_argument(argument, option):
if replace_with_null(argument):
return 'NULL'
elif requires_checked_cast(argument):
checked_use = CHECKED_USE.get(argument['type'],
'{}_').format(argument['name'])
if nullable_argument(argument):
checked_use = CHECKED_USE_NULLABLE.substitute(
env={}, arg_name=argument['name'], usage=checked_use)
return checked_use
elif argument['type'] == 'bool' and 'if_true' in argument:
if bool_option_is_string(argument):
tpl = '({}) ? "{}" : "{}"'
else:
tpl = '({}) ? {} : {}'
return tpl.format(argument['name'], argument['if_true'],
argument['if_false'])
elif argument['type'] == 'CONSTANT':
# this is a bool that is actually a string...
if bool_option_is_string(argument):
return '"{}"'.format(argument['name'])
v = str(argument.get('default', argument['name']))
for pattern, replacement in CONSTANT_REPLACEMENTS:
v = re.sub(pattern, replacement, v)
return CodeTemplate(v).substitute(backend_type_env)
# e.g. argument 0, i.e. repeat the 0th argument in this position...
elif argument['type'] == 'argument':
index = int(argument['name'])
return get_argument(option['arguments'][index], option)
else:
return argument['name']
def allocate_arg(env, arg, output_count):
name = arg['name']
allocation = CodeTemplate(ALLOC_WRAP[arg['type']]).substitute(env)
if arg.get('mask', False):
allocation = 'output_mask[{}] ? {} : nullptr'.format(output_count, allocation)
return [
'auto {}_ = {};'.format(name, allocation),
'auto {} = Tensor({}_,false);'.format(name, name),
]
def allocate_arg(env, arg, output_count):
name = arg['name']
allocation = CodeTemplate(ALLOC_WRAP[arg['type']]).substitute(env)
tensor_arg = '{}_'.format(name)
if arg.get('mask', False):
allocation = 'output_mask[{}] ? {} : nullptr'.format(output_count, allocation)
tensor_arg = ('{}_ == nullptr ? (TensorImpl*)UndefinedTensor::singleton() : (TensorImpl*){}_'
.format(name, name))
return [
'auto {}_ = {};'.format(name, allocation),
'auto {} = Tensor({}, false);'.format(name, tensor_arg),
]
def notebook_generator(conn, dataset, comp):
# initialization
nb = nbf.v4.new_notebook()
instance = CodeTemplate(conn, dataset, comp)
nb['cells'].append(nbf.v4.new_code_cell(instance.dependency()))
nb['cells'].append(nbf.v4.new_code_cell(instance.modifiers()))
ctx, conn_docs = instance.connection()
nb['cells'].append(nbf.v4.new_markdown_cell(conn_docs))
nb['cells'].append(nbf.v4.new_code_cell(ctx))
nb['cells'].append(nbf.v4.new_code_cell(instance.dataset()))
for docs, analysis, plot in instance.analysis():
nb['cells'].append(nbf.v4.new_markdown_cell(docs))
nb['cells'].append(nbf.v4.new_code_cell(analysis))
nb['cells'].append(nbf.v4.new_code_cell(plot))
fname = 'analysis_notebook.ipynb'
with open(fname, 'w') as f:
nbf.write(nb, f)
else:
self.filenames.remove(filename)
def check_all_files_written(self):
if len(self.undeclared_files) > 0:
raise Exception(
"trying to write files {} which are not ".format(self.undeclared_files) +
"in the list of outputs this script produces. " +
"use will_write to add them.")
if len(self.filenames) > 0:
raise Exception("Outputs declared with 'will_write' were " +
"never written: {}".format(self.filenames))
TEMPLATE_PATH = options.source_path + "/templates"
GENERATOR_DERIVED = CodeTemplate.from_file(
TEMPLATE_PATH + "/GeneratorDerived.h")
STORAGE_DERIVED_CPP = CodeTemplate.from_file(
TEMPLATE_PATH + "/StorageDerived.cpp")
STORAGE_DERIVED_H = CodeTemplate.from_file(TEMPLATE_PATH + "/StorageDerived.h")
TYPE_DERIVED_CPP = CodeTemplate.from_file(TEMPLATE_PATH + "/TypeDerived.cpp")
TYPE_DERIVED_H = CodeTemplate.from_file(TEMPLATE_PATH + "/TypeDerived.h")
TYPE_H = CodeTemplate.from_file(TEMPLATE_PATH + "/Type.h")
TYPE_CPP = CodeTemplate.from_file(TEMPLATE_PATH + "/Type.cpp")
TENSOR_DERIVED_CPP = CodeTemplate.from_file(
TEMPLATE_PATH + "/TensorDerived.cpp")
TENSOR_SPARSE_CPP = CodeTemplate.from_file(
TEMPLATE_PATH + "/TensorSparse.cpp")
TENSOR_DENSE_CPP = CodeTemplate.from_file(
TEMPLATE_PATH + "/TensorDense.cpp")
import re
import yaml
from code_template import CodeTemplate
# temporary things we cannot handle
EXCLUDE_PATTERN = "bernoulli.*|normal.*|exponential.*|random.*|arange.*"
# what has to be done to add a Operation ...
# 1. add virtual dispatch declaration to Type.h and default impl to Type.cpp
TYPE_METHOD_DECLARATION = CodeTemplate("""\
virtual ${return_type} ${method_prefix}${api_name}(${formals}) ;
""")
TYPE_METHOD_DEFINITION = CodeTemplate("""\
${return_type} Type::${method_prefix}${api_name}(${formals}) {
throw std::runtime_error(std::string("${api_name} is not implemented for type ") + toString());
}
""")
# 2. add virtual override to TypeDerived.h
TYPE_DERIVED_DECLARATION = CodeTemplate("""\
virtual ${return_type} ${method_prefix}${api_name}(${formals}) override;
""")
# 3. add override definition to TypeDerived.cpp
TYPE_DERIVED_DEFINITION = CodeTemplate("""\
${return_type} ${Type}::${method_prefix}${api_name}(${formals}) {
${type_definition_body}
}
""")
# 4. add non-virtual declaration to Tensor.h
TENSOR_METHOD_DECLARATION = CodeTemplate("""\
${return_type} ${api_name}(${method_formals})${const_mark};
""")
# 5. add non-virtual declaration to Tensor.cpp
else:
self.filenames.remove(filename)
def check_all_files_written(self):
if len(self.undeclared_files) > 0:
raise Exception(
"trying to write files {} which are not ".format(self.undeclared_files) +
"in the list of outputs this script produces. " +
"use will_write to add them.")
if len(self.filenames) > 0:
raise Exception("Outputs declared with 'will_write' were " +
"never written: {}".format(self.filenames))
TEMPLATE_PATH = options.source_path + "/templates"
GENERATOR_DERIVED = CodeTemplate.from_file(
TEMPLATE_PATH + "/GeneratorDerived.h")
TYPE_DERIVED_CPP = CodeTemplate.from_file(TEMPLATE_PATH + "/TypeDerived.cpp")
SPARSE_TYPE_DERIVED_CPP = CodeTemplate.from_file(TEMPLATE_PATH + "/SparseTypeDerived.cpp")
TYPE_DERIVED_H = CodeTemplate.from_file(TEMPLATE_PATH + "/TypeDerived.h")
TYPE_H = CodeTemplate.from_file(TEMPLATE_PATH + "/Type.h")
TYPE_EXTENDED_INTERFACE_H = CodeTemplate.from_file(TEMPLATE_PATH + "/TypeExtendedInterface.h")
TYPE_DEFAULT_H = CodeTemplate.from_file(TEMPLATE_PATH + "/TypeDefault.h")
TYPE_DEFAULT_CPP = CodeTemplate.from_file(TEMPLATE_PATH + "/TypeDefault.cpp")
TYPE_EXTENSION_H = CodeTemplate.from_file(TEMPLATE_PATH + "/TypeExtension.h")
TYPE_EXTENSION_CPP = CodeTemplate.from_file(TEMPLATE_PATH + "/TypeExtension.cpp")
TYPE_EXTENSION_DERIVED_H = CodeTemplate.from_file(TEMPLATE_PATH + "/TypeExtensionDerived.h")
TYPE_EXTENSION_DERIVED_CPP = CodeTemplate.from_file(TEMPLATE_PATH + "/TypeExtensionDerived.cpp")
LEGACY_TH_DISPATCHER_H = CodeTemplate.from_file(TEMPLATE_PATH + "/LegacyTHDispatcher.h")
LEGACY_TH_DISPATCHER_CPP = CodeTemplate.from_file(TEMPLATE_PATH + "/LegacyTHDispatcher.cpp")
LEGACY_TH_DISPATCHER_DERIVED_CPP = CodeTemplate.from_file(TEMPLATE_PATH + "/LegacyTHDispatcherDerived.cpp")
from code_template import CodeTemplate
FILE = CodeTemplate("""\
#include "TH/TH.h"
#ifdef AT_CUDA_ENABLED
#undef THNN_
#include "THC/THC.h"
#endif
#include "ATen/Utils.h"
${copy_includes}
namespace at {
${copy_functions}
}
""")
CASE = CodeTemplate("""\
case ${src_id}:
${THTensor}_copy${cuda}${src_scalar_name}(${state,}dst_->tensor,static_cast<${src_tensor}*>(src.pImpl)->tensor);
break;
""")
FUNCTION = CodeTemplate("""\
void ${Type}::s_copy(const Tensor & src, Tensor & dst) const {
// code generated by function_wrapper
auto dst_ = checked_cast_tensor<${Tensor}>(dst.pImpl,"dst",0,false);
(void) dst_; //silence unused warning
switch(src.type().ID()) {
${copy_body}
else:
self.filenames.remove(filename)
def check_all_files_written(self):
if len(self.undeclared_files) > 0:
raise Exception("trying to write files {} which are not ".format(
self.undeclared_files) +
"in the list of outputs this script produces. " +
"use will_write to add them.")
if len(self.filenames) > 0:
raise Exception("Outputs declared with 'will_write' were " +
"never written: {}".format(self.filenames))
TEMPLATE_PATH = options.source_path + "/templates"
TYPE_DERIVED_CPP = CodeTemplate.from_file(TEMPLATE_PATH + "/TypeDerived.cpp")
SPARSE_TYPE_DERIVED_CPP = CodeTemplate.from_file(TEMPLATE_PATH +
"/SparseTypeDerived.cpp")
TYPE_DERIVED_H = CodeTemplate.from_file(TEMPLATE_PATH + "/TypeDerived.h")
TYPE_DEFAULT_H = CodeTemplate.from_file(TEMPLATE_PATH + "/TypeDefault.h")
TYPE_DEFAULT_CPP = CodeTemplate.from_file(TEMPLATE_PATH + "/TypeDefault.cpp")
OPS_ALREADY_MOVED_TO_C10_CPP = CodeTemplate.from_file(
TEMPLATE_PATH + "/OpsAlreadyMovedToC10.cpp")
BACKEND_SELECT_REGISTER_CPP = CodeTemplate.from_file(
TEMPLATE_PATH + "/BackendSelectRegister.cpp")
TENSOR_H = CodeTemplate.from_file(TEMPLATE_PATH + "/TensorBody.h")
TENSOR_METHODS_H = CodeTemplate.from_file(TEMPLATE_PATH + "/TensorMethods.h")
FUNCTIONS_H = CodeTemplate.from_file(TEMPLATE_PATH + "/Functions.h")
LEGACY_TH_FUNCTIONS_H = CodeTemplate.from_file(TEMPLATE_PATH +
import re
from code_template import CodeTemplate
import sys
if sys.version_info[0] == 3:
string_type = str
else:
string_type = basestring
# temporary things we cannot handle
EXCLUDE_PATTERN = "bernoulli.*|normal.*|exponential.*|random.*|arange.*"
# what has to be done to add a Operation ...
# 1. if broadcasting or without the full list of arguments, add a non-virtual
# declaration under Type.h
TYPE_METHOD_DECLARATION_NON_VIRTUAL = CodeTemplate("""\
${return_type} ${method_prefix}${api_name}(${formals}) ;
""")
# 2. broadcasting functions are implemented in Type.cpp
TYPE_METHOD_DEFINITION_BROADCAST = CodeTemplate("""\
${return_type} Type::${method_prefix}${api_name}(${formals}) {
Tensor ${broadcast_returns};
std::tie(${broadcast_returns}) = ${broadcast_function}(${broadcast_actuals});
return ${method_prefix_derived}${api_name}(${broadcast_modified_actuals});
}
""")
# 3. functions without the full list of arguments are implemented in TypeMethods.h
TYPE_METHOD_INLINE = CodeTemplate("""\
inline ${return_type} Type::${method_prefix}${api_name}(${formals}) {
${return_call}${method_prefix}${api_name}(${actuals_with_constants});
}
""")
from code_template import CodeTemplate
FILE = CodeTemplate("""\
#include "TH/TH.h"
#ifdef AT_CUDA_ENABLED
#undef THNN_
#include "THC/THC.h"
#endif
#include "ATen/Utils.h"
${copy_includes}
namespace at {
${copy_functions}
}
""")
CASE = CodeTemplate("""\
case ${src_id}:
${THTensor}_copy${cuda}${src_scalar_name}(${state,}dst_->tensor,static_cast<${src_tensor}*>(src.pImpl)->tensor);
break;
""")
FUNCTION = CodeTemplate("""\
void ${Type}::copy(const Tensor & src, Tensor & dst) {
// code generated by function_wrapper
auto dst_ = checked_cast<${Tensor}>(dst.pImpl,"dst",0);
switch(src.type().ID()) {
${copy_body}
default:
else:
self.filenames.remove(filename)
def check_all_files_written(self):
if len(self.undeclared_files) > 0:
raise Exception(
"trying to write files {} which are not ".format(self.undeclared_files) +
"in the list of outputs this script produces. " +
"use will_write to add them.")
if len(self.filenames) > 0:
raise Exception("Outputs declared with 'will_write' were " +
"never written: {}".format(self.filenames))
TEMPLATE_PATH = options.source_path + "/templates"
TYPE_DERIVED_CPP = CodeTemplate.from_file(TEMPLATE_PATH + "/TypeDerived.cpp")
SPARSE_TYPE_DERIVED_CPP = CodeTemplate.from_file(TEMPLATE_PATH + "/SparseTypeDerived.cpp")
TYPE_DERIVED_H = CodeTemplate.from_file(TEMPLATE_PATH + "/TypeDerived.h")
TYPE_DEFAULT_H = CodeTemplate.from_file(TEMPLATE_PATH + "/TypeDefault.h")
TYPE_DEFAULT_CPP = CodeTemplate.from_file(TEMPLATE_PATH + "/TypeDefault.cpp")
TYPE_EXTENSION_H = CodeTemplate.from_file(TEMPLATE_PATH + "/TypeExtension.h")
TYPE_EXTENSION_CPP = CodeTemplate.from_file(TEMPLATE_PATH + "/TypeExtension.cpp")
REGISTRATION_DECLARATIONS_H = CodeTemplate.from_file(TEMPLATE_PATH + "/RegistrationDeclarations.h")
TENSOR_H = CodeTemplate.from_file(TEMPLATE_PATH + "/Tensor.h")
TENSOR_METHODS_H = CodeTemplate.from_file(TEMPLATE_PATH + "/TensorMethods.h")
FUNCTIONS_H = CodeTemplate.from_file(TEMPLATE_PATH + "/Functions.h")
LEGACY_TH_FUNCTIONS_H = CodeTemplate.from_file(TEMPLATE_PATH + "/LegacyTHFunctions.h")
LEGACY_TH_FUNCTIONS_CPP = CodeTemplate.from_file(TEMPLATE_PATH + "/LegacyTHFunctions.cpp")
import re
from code_template import CodeTemplate
# temporary things we cannot handle
EXCLUDE_PATTERN = "bernoulli.*|normal.*|exponential.*|random.*|arange.*"
# what has to be done to add a Operation ...
# 1. add virtual dispatch declaration to Type.h and default impl to Type.cpp
TYPE_METHOD_DECLARATION = CodeTemplate("""\
virtual ${return_type} ${method_prefix}${api_name}(${formals}) ;
""")
TYPE_METHOD_DEFINITION = CodeTemplate("""\
${return_type} Type::${method_prefix}${api_name}(${formals}) {
throw std::runtime_error(std::string("${api_name} is not implemented for type ") + toString());
}
""")
# 2. add virtual override to TypeDerived.h
TYPE_DERIVED_DECLARATION = CodeTemplate("""\
virtual ${return_type} ${method_prefix}${api_name}(${formals}) override;
""")
# 3. add override definition to TypeDerived.cpp
TYPE_DERIVED_DEFINITION = CodeTemplate("""\
${return_type} ${Type}::${method_prefix}${api_name}(${formals}) {
${type_definition_body}
}
""")
# 4. add non-virtual declaration to Tensor.h
TENSOR_METHOD_DECLARATION = CodeTemplate("""\
${return_type} ${api_name}(${method_formals})${const_mark};
""")
# 5. add non-virtual declaration to Tensor.cpp
TENSOR_METHOD_DEFINITION = CodeTemplate("""\
parser.add_argument("--output_prefix", default="", help="")
parser.add_argument(
"--install_dir", default=".", help="where to put generated file")
parser.add_argument("--aten_root", default="", help="root directory of aten")
args, _ = parser.parse_known_args()
if args.aten_root:
if not os.path.exists(args.aten_root):
raise ValueError('aten_root ({}) does not exist'.format(
args.aten_root))
sys.path.append(os.path.join(args.aten_root, 'src', 'ATen'))
from code_template import CodeTemplate as CT
else:
from src.ATen.code_template import CodeTemplate as CT
OP_TEMPLATE = CT.from_file(
os.path.join(args.template_dir, 'aten_op_template.h'))
try:
# use faster C loader if available
from yaml import CLoader as Loader
except ImportError:
from yaml import Loader
def write(filename, s):
with open(filename, "w") as f:
f.write(s)
def read(filename):
# 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 sys
import yaml
from copy import deepcopy
project_root = sys.argv[1]
sys.path.append(project_root + "/third_party/aten/src/ATen")
from code_template import CodeTemplate as CT
OP_TEMPLATE = CT.from_file(project_root +
'/caffe2/contrib/aten/aten_op_template.h')
try:
# use faster C loader if available
from yaml import CLoader as Loader
except ImportError:
from yaml import Loader
def write(filename, s):
with open(filename, "w") as f:
f.write(s)
def read(filename):
with open(filename, "r") as f:
else:
self.filenames.remove(filename)
def check_all_files_written(self):
if len(self.undeclared_files) > 0:
raise Exception(
"trying to write files {} which are not ".format(self.undeclared_files) +
"in the list of outputs this script produces. " +
"use will_write to add them.")
if len(self.filenames) > 0:
raise Exception("Outputs declared with 'will_write' were " +
"never written: {}".format(self.filenames))
TEMPLATE_PATH = options.source_path + "/templates"
GENERATOR_DERIVED = CodeTemplate.from_file(
TEMPLATE_PATH + "/GeneratorDerived.h")
TYPE_DERIVED_CPP = CodeTemplate.from_file(TEMPLATE_PATH + "/TypeDerived.cpp")
SPARSE_TYPE_DERIVED_CPP = CodeTemplate.from_file(TEMPLATE_PATH + "/SparseTypeDerived.cpp")
TYPE_DERIVED_H = CodeTemplate.from_file(TEMPLATE_PATH + "/TypeDerived.h")
TYPE_H = CodeTemplate.from_file(TEMPLATE_PATH + "/Type.h")
TYPE_EXTENDED_INTERFACE_H = CodeTemplate.from_file(TEMPLATE_PATH + "/TypeExtendedInterface.h")
TYPE_DEFAULT_H = CodeTemplate.from_file(TEMPLATE_PATH + "/TypeDefault.h")
TYPE_DEFAULT_CPP = CodeTemplate.from_file(TEMPLATE_PATH + "/TypeDefault.cpp")
TYPE_EXTENSION_H = CodeTemplate.from_file(TEMPLATE_PATH + "/TypeExtension.h")
TYPE_EXTENSION_CPP = CodeTemplate.from_file(TEMPLATE_PATH + "/TypeExtension.cpp")
LEGACY_TH_DISPATCHER_H = CodeTemplate.from_file(TEMPLATE_PATH + "/LegacyTHDispatcher.h")
LEGACY_TH_DISPATCHER_CPP = CodeTemplate.from_file(TEMPLATE_PATH + "/LegacyTHDispatcher.cpp")
LEGACY_TH_DISPATCHER_DERIVED_CPP = CodeTemplate.from_file(TEMPLATE_PATH + "/LegacyTHDispatcherDerived.cpp")
LEGACY_TH_DISPATCHER_DERIVED_H = CodeTemplate.from_file(TEMPLATE_PATH + "/LegacyTHDispatcherDerived.h")
else:
self.filenames.remove(filename)
def check_all_files_written(self):
if len(self.undeclared_files) > 0:
raise Exception("trying to write files {} which are not ".format(
self.undeclared_files) +
"in the list of outputs this script produces. " +
"use will_write to add them.")
if len(self.filenames) > 0:
raise Exception("Outputs declared with 'will_write' were " +
"never written: {}".format(self.filenames))
TEMPLATE_PATH = options.source_path + "/templates"
GENERATOR_DERIVED = CodeTemplate.from_file(TEMPLATE_PATH +
"/GeneratorDerived.h")
TYPE_DERIVED_CPP = CodeTemplate.from_file(TEMPLATE_PATH + "/TypeDerived.cpp")
SPARSE_TYPE_DERIVED_CPP = CodeTemplate.from_file(TEMPLATE_PATH +
"/SparseTypeDerived.cpp")
TYPE_DERIVED_H = CodeTemplate.from_file(TEMPLATE_PATH + "/TypeDerived.h")
TYPE_H = CodeTemplate.from_file(TEMPLATE_PATH + "/Type.h")
TYPE_EXTENDED_INTERFACE_H = CodeTemplate.from_file(TEMPLATE_PATH +
"/TypeExtendedInterface.h")
TYPE_DEFAULT_H = CodeTemplate.from_file(TEMPLATE_PATH + "/TypeDefault.h")
TYPE_DEFAULT_CPP = CodeTemplate.from_file(TEMPLATE_PATH + "/TypeDefault.cpp")
REGISTER_CPU_H = CodeTemplate.from_file(TEMPLATE_PATH + "/RegisterCPU.h")
REGISTER_CPU_CPP = CodeTemplate.from_file(TEMPLATE_PATH + "/RegisterCPU.cpp")
REGISTER_CUDA_H = CodeTemplate.from_file(TEMPLATE_PATH + "/RegisterCUDA.h")
REGISTER_CUDA_CPP = CodeTemplate.from_file(TEMPLATE_PATH + "/RegisterCUDA.cpp")
def emit_body(env, option):
body = []
body += handle_sparse(env, option)
body += handle_zero_dim(env, option)
# arguments are potentially duplicated because of one argument
# referencing another
seen_names = set()
count = 0
is_cuda = 'CUDA' in backend_type_env['Backend']
# scalar_check is the heuristic conditions when a result may be a scalar_check
# if there is a THSize* argument, then its dimensions are used to determine scalar.
# otherwise, it is true if all the input tensors are scalars,
scalar_check_is_from_size = False
scalar_check = None
for arg in option['arguments']:
if is_real_argument_to_wrapper(arg):
count += 1
if arg['type'] == 'THSize*':
scalar_check_is_from_size = True
scalar_check = '{}.size() == 0'.format(arg['name'])
wrap_dim_arg = arg.get('wrap_dim', None)
if wrap_dim_arg is not None:
# wrap_dim specification can have (add) expressions, e.g. self+1
wrap_dim_params = wrap_dim_arg.split("+")
wrap_dim_params[0] = wrap_dim_params[0] + "_"
wrap_dim_target = wrap_dim_params[0]
wrap_dim_params[0] = "{}->dim()".format(wrap_dim_target)
wrap_dim_expr = "+".join(wrap_dim_params)
body.append("{} = maybe_wrap_dim({}, {});".format(
arg['name'], arg['name'], wrap_dim_expr))
# only generated checked casts the first time we see it
if not arg['name'] in seen_names and requires_checked_cast(arg):
seen_names.add(arg['name'])
# make a new allocation of TensorImpl, then wrap a Tensor around it.
if arg.get('allocate', False):
allocation = CodeTemplate(
ALLOC_WRAP[arg['type']]).substitute(env)
body.append('auto {}_ = {};'.format(
arg['name'], allocation))
body.append('auto {} = Tensor({}_,false);'.format(
arg['name'], arg['name']))
# extract the TensorImpl from an existing tensor (or Storage, etc.)
else:
# special case where we allow undefined Tensors, and thus
# the checked cast succeeds even if the Tensor is not
# defined
null_okay = 'true' if nullable_argument(arg) else 'false'
check_cast = CHECKED_CAST[arg['type']].substitute(
env,
arg_name=arg['name'],
arg_pos=count,
null_okay=null_okay)
body.append("auto {}_ = {};".format(
arg['name'], check_cast))
if drop_argument(arg, option) or replace_with_null(arg):
body.append("(void) {}_; //silence unused warning".format(
arg['name']))
# resize tensors for special ops that require it
if 'resize' in arg:
resize = arg['resize']
if isinstance(resize, str):
body.append("{}.resize_({}.sizes());".format(
arg['name'], resize))
else:
dims = [
'{}.size({})'.format(name, dim)
for name, dim in resize
]
body.append("{}.resize_({{ {} }});".format(
arg['name'], ','.join(dims)))
# also special handling where we zero some outputs.
if arg.get('cpu_zero', False) and not is_cuda:
body.append("{}.zero_();".format(arg['name']))
# isScalar() for all input tensors is and'd to form
# the test for whether the output is also a scalar
if (not arg.get('output') and 'Tensor' in arg['type']
and 'TensorList' not in arg['type']
and 'THS' not in arg['type']
and not scalar_check_is_from_size):
check = '{}->isScalar()'.format(arg['name'] + '_')
scalar_check = (check if scalar_check is None else
scalar_check + ' && ' + check)
option['derived_actuals'] = get_arguments(option)
is_nn = option['mode'] == 'NN'
if is_cuda or is_nn:
option['derived_actuals'] = ['context->thc_state'
] + option['derived_actuals']
if is_nn:
prefix = 'THNN_{}'.format(env['THType'])
elif option.get('sparse', False):
if is_cuda:
prefix = 'THCS' + env['ScalarName'] + "Tensor_"
else:
prefix = env['THTensor'].replace('TH', 'THS') + '_'
else:
prefix = env['THTensor'] + '_'
call = prefix + \
CodeTemplate("${cname}(${derived_actuals})").substitute(env)
ret = option['return']
if ret['kind'] == 'arguments':
if 'aten_custom_call' in option:
# all aten_custom_call bodies handle settings on their own.
scalar_check = None
body.append(
CodeTemplate(option['aten_custom_call']).substitute(env))
else:
body.append(call + ";")
arguments_indices = ret['arguments']
arguments = [
option['arguments'][argi] for argi in arguments_indices
]
if scalar_check is not None:
if len(arguments) > 1:
body.append("bool maybe_scalar = {};".format(scalar_check))
scalar_check = 'maybe_scalar'
for arg in arguments:
body.append("{}_->maybeScalar({});".format(
arg['name'], scalar_check))
if len(arguments_indices) == 1:
arg = arguments[0]
body.append("return {};".format(arg['name']))
else:
types = [
to_return_type(arg, option)['type'] for arg in arguments
]
# TODO: check for move semantics...
names = [arg['name'] for arg in arguments]
body.append(
CodeTemplate(
"return std::tuple<${types}>(${names});").substitute(
types=types, names=names))
elif ret['kind'] == 'type':
if ret['type'] == 'THTensor*':
maybe_scalar = "->maybeScalar({})".format(scalar_check) \
if scalar_check is not None \
else ""
return_tensor = "return Tensor((new ${Tensor}(context,${arg_name}))${maybe_scalar},false);"
body.append(
CodeTemplate(return_tensor).substitute(
env, arg_name=call, maybe_scalar=maybe_scalar))
else:
# we using int64_t for long in the API, so correct it here...
if is_actual_return_long(ret):
call = "static_cast<int64_t>({})".format(call)
body.append("return {};".format(call))
else:
raise Exception("NYI - return handling")
return body
# limitations under the License.
##############################################################################
import sys
import yaml
project_root = sys.argv[1]
sys.path.append(project_root + "/third_party/aten/src/ATen")
from code_template import CodeTemplate as CT
try:
# use faster C loader if available
from yaml import CLoader as Loader
except ImportError:
from yaml import Loader
OP_TEMPLATE = CT.from_file(project_root+'/caffe2/contrib/aten/aten_op_template.h')
def write(filename, s):
with open(filename, "w") as f:
f.write(s)
def read(filename):
with open(filename, "r") as f:
return f.read()
decls = yaml.load(read('aten/src/ATen/ATen/Declarations.yaml'), Loader=Loader)
top_env = {
from code_template import CodeTemplate
from function_wrapper import nested_dict
FILE = CodeTemplate("""\
#include "ATen/Config.h"
#include "TH/TH.h"
${cuda_includes}
#include "ATen/Utils.h"
${copy_includes}
namespace at {
${copy_functions}
}
""")
CUDA_INCLUDES = """\
#undef THNN_
#include "THC/THC.h"
"""
# NB: The copy templates static_cast both dst and src, even though
# technically we also perform a checked_cast_tensor in the prologue
# of the copy (meaning that hypothetically, an already casted tensor
# is available. However, in s_copy, the casted tensor is dst, while
# in _s_copy_from, the casted tensor is src. So we can reuse the logic
# in both cases, we unconditionally cast both tensors (and rely
# on the surrounding code to establish the necessary invariants.)
# Rather than teaching the dispatcher how to extract dispatch keys from types besides Tensor, we
# register an extra kernel for each factory op, under the `BackendSelect` dispatch key. This key
# has higher precedence than dispatch keys for actual backends, so a BackendSelect kernel will
# front-run other kernels registered for the same op.
#
# It's the responsibility of the BackendSelect factory kernels to extract the "real" dispatch
# key from non-Tensor arguments, and redispatch using this key. Here, we generate implementations
# that obtain the key from the TensorOptions argument that's passed to all Tensor factory ops.
#
# BackendSelectRegister.cpp will contain both the BackendSelect kernels and registrations for
# all factory functions that have 'backend_select' flag in its native_functions.yaml definition.
from code_template import CodeTemplate
from function_wrapper import gen_dispatch_key_init
GENERATED_COMMENT = CodeTemplate("@" + "generated from ${filename}")
FUNCTION_REGISTRATION = CodeTemplate("""\
m.impl_UNBOXED("aten::${op_name_with_overload_name}", ${function_name});
""")
FUNCTION_DEFINITION = CodeTemplate("""\
// ${schema_string}
Tensor ${function_name}(${method_formals}) {
static OperatorHandle OP = c10::Dispatcher::singleton().findSchemaOrThrow("aten::${name}", "${overload_name}");
${dispatch_key_init}
return OP.callWithDispatchKey<${formals_types}>(_dk, ${actuals});
}
""")
else:
self.filenames.remove(filename)
def check_all_files_written(self):
if len(self.undeclared_files) > 0:
raise Exception("trying to write files {} which are not ".format(
self.undeclared_files) +
"in the list of outputs this script produces. " +
"use will_write to add them.")
if len(self.filenames) > 0:
raise Exception("Outputs declared with 'will_write' were " +
"never written: {}".format(self.filenames))
TEMPLATE_PATH = options.source_path + "/templates"
TYPE_DERIVED_CPP = CodeTemplate.from_file(TEMPLATE_PATH + "/TypeDerived.cpp")
SPARSE_TYPE_DERIVED_CPP = CodeTemplate.from_file(TEMPLATE_PATH +
"/SparseTypeDerived.cpp")
TYPE_DERIVED_H = CodeTemplate.from_file(TEMPLATE_PATH + "/TypeDerived.h")
TYPE_DEFAULT_H = CodeTemplate.from_file(TEMPLATE_PATH + "/TypeDefault.h")
TYPE_DEFAULT_CPP = CodeTemplate.from_file(TEMPLATE_PATH + "/TypeDefault.cpp")
REGISTRATION_DECLARATIONS_H = CodeTemplate.from_file(
TEMPLATE_PATH + "/RegistrationDeclarations.h")
TENSOR_H = CodeTemplate.from_file(TEMPLATE_PATH + "/TensorBody.h")
TENSOR_METHODS_H = CodeTemplate.from_file(TEMPLATE_PATH + "/TensorMethods.h")
FUNCTIONS_H = CodeTemplate.from_file(TEMPLATE_PATH + "/Functions.h")
LEGACY_TH_FUNCTIONS_H = CodeTemplate.from_file(TEMPLATE_PATH +
"/LegacyTHFunctions.h")
else:
self.filenames.remove(filename)
def check_all_files_written(self):
if len(self.undeclared_files) > 0:
raise Exception("trying to write files {} which are not ".format(
self.undeclared_files) +
"in the list of outputs this script produces. " +
"use will_write to add them.")
if len(self.filenames) > 0:
raise Exception("Outputs declared with 'will_write' were " +
"never written: {}".format(self.filenames))
TEMPLATE_PATH = options.source_path + "/templates"
TYPE_DERIVED_CPP = CodeTemplate.from_file(TEMPLATE_PATH + "/TypeDerived.cpp")
SPARSE_TYPE_DERIVED_CPP = CodeTemplate.from_file(TEMPLATE_PATH +
"/SparseTypeDerived.cpp")
TYPE_DERIVED_H = CodeTemplate.from_file(TEMPLATE_PATH + "/TypeDerived.h")
TYPE_DEFAULT_H = CodeTemplate.from_file(TEMPLATE_PATH + "/TypeDefault.h")
TYPE_DEFAULT_CPP = CodeTemplate.from_file(TEMPLATE_PATH + "/TypeDefault.cpp")
OPS_ALREADY_MOVED_TO_C10_CPP = CodeTemplate.from_file(
TEMPLATE_PATH + "/OpsAlreadyMovedToC10.cpp")
TENSOR_H = CodeTemplate.from_file(TEMPLATE_PATH + "/TensorBody.h")
TENSOR_METHODS_H = CodeTemplate.from_file(TEMPLATE_PATH + "/TensorMethods.h")
FUNCTIONS_H = CodeTemplate.from_file(TEMPLATE_PATH + "/Functions.h")
LEGACY_TH_FUNCTIONS_H = CodeTemplate.from_file(TEMPLATE_PATH +
"/LegacyTHFunctions.h")
from code_template import CodeTemplate
from function_wrapper import nested_dict
FILE = CodeTemplate("""\
// ${generated_comment}
#include "ATen/Config.h"
#include "TH/TH.h"
${cuda_includes}
#include "ATen/Utils.h"
${copy_includes}
namespace at {
${copy_functions}
}
""")
CUDA_INCLUDES = """\
#undef THNN_
#include "THC/THC.h"
"""
# NB: The copy templates static_cast both dst and src, even though
# technically we also perform a checked_cast_tensor in the prologue
# of the copy (meaning that hypothetically, an already casted tensor
# is available. However, in s_copy, the casted tensor is dst, while
# in _s_copy_from, the casted tensor is src. So we can reuse the logic
# in both cases, we unconditionally cast both tensors (and rely
# This script generates BackendSelectRegister.cpp which is being used for dispatching purposes.
# We process only those factory functions that have 'backend_select' flag in its native_functions.yaml definition.
from code_template import CodeTemplate
GENERATED_COMMENT = CodeTemplate("@" + "generated from ${filename}")
FUNCTION_REGISTRATION = CodeTemplate("""\
.op(torch::RegisterOperators::options()
.schema("${schema_string}")
.impl_unboxedOnlyKernel<decltype(${function_name}), &${function_name}>(DispatchKey::BackendSelect)
.aliasAnalysis(AliasAnalysisKind::FROM_SCHEMA))
""")
FUNCTION_DEFINITION = CodeTemplate("""\
Tensor ${function_name}(${method_formals}) {
DispatchKey key = options.computeDispatchKey();
static auto op = c10::Dispatcher::singleton().findSchemaOrThrow("aten::${name}", "${overload_name}");
return op.callUnboxedWithDispatchKey<${formals_types}>(key, ${type_method_actuals});
}
""")
def needs_backend_select(declaration_option):
# We register an op under the BackendSelect dispatch key
# if a TensorOptions argument has been gathered from its declared args
# We skip all the 'new_*' and '*_like' ops as they are special cased and avoid dispatching.
# See TypeDefault.cpp
if declaration_option['name'].endswith(
'_like') or declaration_option['name'].startswith('new_'):
return False
'--source-path',
help='path to source director for tensorlib',
action='store',
default='.')
parser.add_option('-o',
'--output-dependencies',
help='only output a list of dependencies',
action='store')
parser.add_option('-n', '--no-cuda', action='store_true')
options, files = parser.parse_args()
if options.output_dependencies is not None:
output_dependencies_file = open(options.output_dependencies, 'w')
TEMPLATE_PATH = options.source_path + "/templates"
GENERATOR_DERIVED = CodeTemplate.from_file(TEMPLATE_PATH +
"/GeneratorDerived.h")
STORAGE_DERIVED_CPP = CodeTemplate.from_file(TEMPLATE_PATH +
"/StorageDerived.cpp")
STORAGE_DERIVED_H = CodeTemplate.from_file(TEMPLATE_PATH + "/StorageDerived.h")
TYPE_DERIVED_CPP = CodeTemplate.from_file(TEMPLATE_PATH + "/TypeDerived.cpp")
TYPE_DERIVED_H = CodeTemplate.from_file(TEMPLATE_PATH + "/TypeDerived.h")
TYPE_H = CodeTemplate.from_file(TEMPLATE_PATH + "/Type.h")
TYPE_CPP = CodeTemplate.from_file(TEMPLATE_PATH + "/Type.cpp")
TENSOR_DERIVED_CPP = CodeTemplate.from_file(TEMPLATE_PATH +
"/TensorDerived.cpp")
TENSOR_SPARSE_CPP = CodeTemplate.from_file(TEMPLATE_PATH + "/TensorSparse.cpp")
TENSOR_DENSE_CPP = CodeTemplate.from_file(TEMPLATE_PATH + "/TensorDense.cpp")
TENSOR_DERIVED_H = CodeTemplate.from_file(TEMPLATE_PATH + "/TensorDerived.h")
parser.add_argument("--install_dir",
default=".",
help="where to put generated file")
parser.add_argument("--aten_root", default="", help="root directory of aten")
args, _ = parser.parse_known_args()
if args.aten_root:
if not os.path.exists(args.aten_root):
raise ValueError('aten_root ({}) does not exist'.format(
args.aten_root))
sys.path.append(os.path.join(args.aten_root, 'src', 'ATen'))
from code_template import CodeTemplate as CT
else:
from src.ATen.code_template import CodeTemplate as CT
OP_TEMPLATE = CT.from_file(
os.path.join(args.template_dir, 'aten_op_template.h'))
try:
# use faster C loader if available
from yaml import CLoader as Loader
except ImportError:
from yaml import Loader
def write(filename, s):
with open(filename, "w") as f:
f.write(s)
def read(filename):
with open(filename, "r") as f:
def emit_body(env, option):
body = []
# arguments are potentially duplicated because of one argument
# referencing another
seen_names = set()
# only generated checked casts the first time we see it
count = 0
for arg in option['arguments']:
if is_real_argument_to_wrapper(arg):
count += 1
if not arg['name'] in seen_names and requires_checked_cast(arg):
seen_names.add(arg['name'])
if arg.get('allocate', False):
allocation = CodeTemplate(
ALLOC_WRAP[arg['type']]).substitute(env)
body.append('auto {}_ = {};'.format(
arg['name'], allocation))
body.append('auto {} = Tensor({}_,false);'.format(
arg['name'], arg['name']))
else:
check_cast = CHECKED_CAST[arg['type']].substitute(
env, arg_name=arg['name'], arg_pos=count)
body.append("auto {}_ = {};".format(
arg['name'], check_cast))
if 'resize' in arg:
resize = arg['resize']
if type(resize) == str:
body.append("{}.resize_as_({});".format(
arg['name'], resize))
else:
dims = ['{}.size({})'.format(name, dim)
for name, dim in resize]
body.append("{}.resize_({{ {} }});".format(
arg['name'], ','.join(dims)))
if arg.get('cpu_zero', False):
body.append("{}.zero_();".format(arg['name']))
option['actuals'] = get_arguments(option)
is_cuda = backend_type_env['Backend'] == 'CUDA'
is_nn = option['mode'] == 'NN'
if is_cuda or is_nn:
option['actuals'] = ['context->thc_state'] + option['actuals']
if is_nn:
prefix = 'THNN_{}'.format(env['THType'])
else:
prefix = env['THTensor'] + '_'
call = prefix + CodeTemplate("${cname}(${actuals})").substitute(env)
ret = option['return']
if ret['kind'] == 'arguments':
body.append(call + ";")
arguments_indices = ret['arguments']
if len(arguments_indices) == 1:
arg = option['arguments'][arguments_indices[0]]
body.append("return {};".format(arg['name']))
else:
arguments = [option['arguments'][argi]
for argi in arguments_indices]
types = [to_return_type(arg, option) for arg in arguments]
# TODO: check for move semantics...
names = [arg['name'] for arg in arguments]
body.append(CodeTemplate("return std::tuple<${types}>(${names});").substitute(
types=types, names=names))
elif ret['kind'] == 'type':
if ret['type'] == 'THTensor*':
body.append(CodeTemplate(
"return Tensor(new ${Tensor}(context,${arg_name}),false);").substitute(env, arg_name=call))
else:
# we using int64_t for long in the API, so correct it here...
if is_actual_return_long(ret):
call = "static_cast<int64_t>({})".format(call)
body.append("return {};".format(call))
else:
raise Exception("NYI - return handling")
return body
# Rather than teaching the dispatcher how to extract dispatch keys from types besides Tensor, we
# register an extra kernel for each factory op, under the `BackendSelect` dispatch key. This key
# has higher precedence than dispatch keys for actual backends, so a BackendSelect kernel will
# front-run other kernels registered for the same op.
#
# It's the responsibility of the BackendSelect factory kernels to extract the "real" dispatch
# key from non-Tensor arguments, and redispatch using this key. Here, we generate implementations
# that obtain the key from the TensorOptions argument that's passed to all Tensor factory ops.
#
# BackendSelectRegister.cpp will contain both the BackendSelect kernels and registrations for
# all factory functions that have 'backend_select' flag in its native_functions.yaml definition.
from code_template import CodeTemplate
from function_wrapper import gen_dispatch_key_init
GENERATED_COMMENT = CodeTemplate("@" + "generated from ${filename}")
FUNCTION_REGISTRATION = CodeTemplate("""\
.op(torch::RegisterOperators::options()
.schema("${schema_string}")
.impl_unboxedOnlyKernel<decltype(${function_name}), &${function_name}>(DispatchKey::BackendSelect)
.aliasAnalysis(AliasAnalysisKind::FROM_SCHEMA))
""")
FUNCTION_DEFINITION = CodeTemplate("""\
// ${schema_string}
Tensor ${function_name}(${method_formals}) {
static OperatorHandle OP = c10::Dispatcher::singleton().findSchemaOrThrow("aten::${name}", "${overload_name}");
${dispatch_key_init}
globalLegacyTypeDispatch().initForDispatchKey(_dk);
return OP.callUnboxedWithDispatchKey<${formals_types}>(_dk, ${type_method_actuals});
from collections import OrderedDict
from code_template import CodeTemplate
import sys
if sys.version_info[0] == 3:
string_type = str
else:
string_type = basestring
# temporary things we cannot handle
EXCLUDE_PATTERN = "bernoulli.*"
# what has to be done to add a Operation ...
# 1. if broadcasting or without the full list of arguments, add a non-virtual
# declaration under Type.h
TYPE_METHOD_DECLARATION_NON_VIRTUAL = CodeTemplate("""\
${return_type} ${api_name}(${formals_with_defaults}) const;
""")
# 2. broadcasting functions are implemented in Type.cpp
TYPE_METHOD_DEFINITION_BROADCAST = CodeTemplate("""\
${return_type} Type::${api_name}(${formals}) const {
Tensor ${broadcast_returns};
std::tie(${broadcast_returns}) = ${broadcast_function}(${broadcast_actuals}, "${api_name}");
return ${method_prefix_derived}${api_name}(${broadcast_modified_actuals});
}
""")
# 3. add virtual dispatch declaration to Type.h and impl to Type.cpp (this is usually
# a default impl because actual implementations are in the derived Types); method_prefix_derived
# is present for providing a base-class definition for a derived-type method with a prefix.
TYPE_METHOD_DECLARATION = CodeTemplate("""\
virtual ${return_type} ${method_prefix_derived}${api_name}(${formals_with_defaults}) const;
""")
# Rather than teaching the dispatcher how to extract dispatch keys from types besides Tensor, we
# register an extra kernel for each factory op, under the `BackendSelect` dispatch key. This key
# has higher precedence than dispatch keys for actual backends, so a BackendSelect kernel will
# front-run other kernels registered for the same op.
#
# It's the responsibility of the BackendSelect factory kernels to extract the "real" dispatch
# key from non-Tensor arguments, and redispatch using this key. Here, we generate implementations
# that obtain the key from the TensorOptions argument that's passed to all Tensor factory ops.
#
# BackendSelectRegister.cpp will contain both the BackendSelect kernels and registrations for
# all factory functions that have 'backend_select' flag in its native_functions.yaml definition.
from code_template import CodeTemplate
from function_wrapper import gen_dispatch_key_init
GENERATED_COMMENT = CodeTemplate("@" + "generated from ${filename}")
UNBOXEDONLY_FUNCTION_REGISTRATION = CodeTemplate("""\
m.impl_UNBOXED("aten::${op_name_with_overload_name}", ${function_name});
""")
FUNCTION_REGISTRATION = CodeTemplate("""\
m.impl("aten::${op_name_with_overload_name}", c10::impl::hacky_wrapper_for_legacy_signatures(TORCH_FN(${function_name})));
""")
FUNCTION_DEFINITION = CodeTemplate("""\
// ${schema_string}
Tensor ${function_name}(${method_formals}) {
static auto op = c10::Dispatcher::singleton()
.findSchemaOrThrow("aten::${name}", "${overload_name}")
.typed<${function_cpp_signature}>();
def emit_body(env, option):
body = []
body += handle_sparse(env, option)
body += handle_zero_dim(env, option)
only_zero_dim_check = handle_only_zero_dim(env, option)
if only_zero_dim_check is not None:
# code below only_zero_dim_check is unreachable so we do not need to generate the rest.
body += only_zero_dim_check
return body
body += handle_buffers(env, option)
# arguments are potentially duplicated because of one argument
# referencing another
seen_names = set()
seen_tensorlists = set()
count = 0
output_count = 0
# scalar_check is the heuristic conditions when a result may be a scalar_check
# if there is a THSize* argument, then its dimensions are used to determine scalar.
# otherwise, it is true if all the input tensors are scalars,
scalar_check_is_from_size = False
scalar_check_is_from_option = False
scalar_check = None
scalar_check_opt = option.get('scalar_check')
if scalar_check_opt is not None:
if isinstance(scalar_check_opt, bool):
scalar_check = str(scalar_check_opt).lower()
else:
scalar_check = scalar_check_opt
scalar_check_is_from_option = True
for arg in option['arguments']:
if is_real_argument_to_wrapper(arg):
count += 1
if arg['type'] == 'THSize*' and not scalar_check_is_from_option:
scalar_check_is_from_size = True
scalar_check = '{}.size() == 0'.format(arg['name'])
if arg['type'] == 'TensorList':
seen_tensorlists.add(arg['name'])
wrap_dim_target = arg.get('wrap_dim', None)
if wrap_dim_target is not None:
# for Tensors, "name_" is the TensorImpl, but for TensorLists, it is an
# std::vector of TH*s. Since TH*s have different dimension rules, we used
# "name" instead, but keep "name_" for tensor to avoid an extra function call.
if wrap_dim_target not in seen_tensorlists:
wrap_dim_target = wrap_dim_target + "_"
body.append("{} = maybe_wrap_dim({}, {});".format(
arg['name'], arg['name'], wrap_dim_target))
# only generated checked casts the first time we see it
if arg['name'] not in seen_names and requires_checked_cast(arg):
seen_names.add(arg['name'])
# make a new allocation of TensorImpl, then wrap a Tensor around it.
if arg.get('allocate', False):
body += allocate_arg(env, arg, output_count)
output_count += 1
# extract the TensorImpl from an existing tensor (or Storage, etc.)
else:
# special case where we allow undefined Tensors, and thus
# the checked cast succeeds even if the Tensor is not
# defined
null_okay = 'true' if nullable_argument(arg) else 'false'
default_init = []
if 'default_init' in arg:
default_init.append(arg['default_init'])
noelem_to_empty = 'is_noelem_tensor_size(size)' if 'size' in seen_names else 'false'
check_cast = CHECKED_CAST[arg['type']].substitute(
env,
arg_name=arg['name'],
arg_pos=count,
null_okay=null_okay,
default_init=default_init,
size=arg.get('size'),
noelem_to_empty=noelem_to_empty)
body.append("auto {}_ = {};".format(
arg['name'], check_cast))
if drop_argument(arg, option) or replace_with_null(arg):
body.append("(void) {}_; //silence unused warning".format(
arg['name']))
initializers = []
# resize tensors for special ops that require it
if 'resize' in arg:
initializers.append(resize_arg(arg))
# also special handling where we zero some outputs.
if arg.get('zero', False) or (arg.get('cpu_zero', False)
and not is_cuda):
initializers.append("{}.zero_();".format(arg['name']))
# only initialize non-null arguments
if nullable_argument(arg) and len(initializers) > 0:
body.append(
CONDITIONAL_INITIALIZER.substitute({
'name':
arg['name'],
'initializer':
initializers
}))
else:
body += initializers
# for out-of-place: isScalar() for all input tensors is and'd to form
# the test for whether the output is also a scalar
# for in-place: isScalar() shouldn't change as a result of the operation
if (not arg.get('output') and 'Tensor' in arg['type']
and 'TensorList' not in arg['type']
and 'THS' not in arg['type']
and not scalar_check_is_from_size
and not scalar_check_is_from_option
and not option['inplace']):
check = '{}->isScalar()'.format(arg['name'] + '_')
if nullable_argument(arg):
check = '(!{} || {})'.format(arg['name'] + '_', check)
scalar_check = (check if scalar_check is None else
scalar_check + ' && ' + check)
# cimpls, if it exists, contains the underlying C function names and
# arguments. Otherwise use option
cimpls = option.get('cimpls', [option])
calls = [handle_call(env, option, cimpl) for cimpl in cimpls]
ret = option['return']
if ret['kind'] == 'arguments':
if 'aten_custom_call' in option:
# all aten_custom_call bodies handle settings on their own.
scalar_check = None
body.append(
CodeTemplate(option['aten_custom_call']).substitute(env))
else:
body.extend([call + ';' for call in calls])
arguments_indices = ret['arguments']
arguments = [
option['arguments'][argi] for argi in arguments_indices
]
if scalar_check is not None:
if len(arguments) > 1:
body.append("bool maybe_scalar = {};".format(scalar_check))
scalar_check = 'maybe_scalar'
for arg in arguments:
stmt = "{}_->maybeScalar({});".format(
arg['name'], scalar_check)
if nullable_argument(arg):
stmt = "if ({}_) {}".format(arg['name'], stmt)
body.append(stmt)
if len(arguments_indices) == 1:
arg = arguments[0]
body.append("return {};".format(arg['name']))
else:
types = [
to_return_type(arg, option)['type'] for arg in arguments
]
# TODO: check for move semantics...
names = [arg['name'] for arg in arguments]
body.append(
CodeTemplate(
"return std::tuple<${types}>(${names});").substitute(
types=types, names=names))
elif ret['kind'] == 'type':
assert len(calls) == 1
call = calls[0]
if ret['type'] == 'THTensor*':
maybe_scalar = "->maybeScalar({})".format(scalar_check) \
if scalar_check is not None \
else ""
return_tensor = "return Tensor((new ${Tensor}(context,${arg_name}))${maybe_scalar},false);"
body.append(
CodeTemplate(return_tensor).substitute(
env, arg_name=call, maybe_scalar=maybe_scalar))
# return the same underlying Tensor type for both real and accreal; this ensures
# e.g. x.sum(0) and x.sum() return the same type.
elif ret['type'] == 'accreal' or ret['type'] == 'real':
body.append('return scalarTensor({});'.format(call))
else:
# we using int64_t for long in the API, so correct it here...
if is_actual_return_long(ret):
call = "static_cast<int64_t>({})".format(call)
body.append("return {};".format(call))
else:
raise Exception("NYI - return handling")
return body
else:
self.filenames.remove(filename)
def check_all_files_written(self):
if len(self.undeclared_files) > 0:
raise Exception("trying to write files {} which are not ".format(
self.undeclared_files) +
"in the list of outputs this script produces. " +
"use will_write to add them.")
if len(self.filenames) > 0:
raise Exception("Outputs declared with 'will_write' were " +
"never written: {}".format(self.filenames))
TEMPLATE_PATH = options.source_path + "/templates"
GENERATOR_DERIVED = CodeTemplate.from_file(TEMPLATE_PATH +
"/GeneratorDerived.h")
STORAGE_DERIVED_CPP = CodeTemplate.from_file(TEMPLATE_PATH +
"/StorageDerived.cpp")
STORAGE_DERIVED_H = CodeTemplate.from_file(TEMPLATE_PATH + "/StorageDerived.h")
TYPE_DERIVED_CPP = CodeTemplate.from_file(TEMPLATE_PATH + "/TypeDerived.cpp")
SPARSE_TYPE_DERIVED_CPP = CodeTemplate.from_file(TEMPLATE_PATH +
"/SparseTypeDerived.cpp")
TYPE_DERIVED_H = CodeTemplate.from_file(TEMPLATE_PATH + "/TypeDerived.h")
TYPE_H = CodeTemplate.from_file(TEMPLATE_PATH + "/Type.h")
TYPE_CPP = CodeTemplate.from_file(TEMPLATE_PATH + "/Type.cpp")
TENSOR_DERIVED_CPP = CodeTemplate.from_file(TEMPLATE_PATH +
"/TensorDerived.cpp")
TENSOR_DENSE_CPP = CodeTemplate.from_file(TEMPLATE_PATH + "/TensorDense.cpp")