# Copyright 2020-2023 Huawei Technologies Co., Ltd
#
# 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.
# ============================================================================
"""Operators info register."""
from __future__ import absolute_import
from __future__ import division
import inspect
import json
import os
import functools
import platform
import hashlib
import shutil
from mindspore._c_expression import Oplib
from mindspore import _checkparam as validator
from mindspore import log as logger
if platform.system() == "Linux":
import fcntl
# path of built-in op info register.
BUILT_IN_OPS_REGISTER_PATH = "mindspore/ops/_op_impl"
BUILT_IN_CUSTOM_OPS_REGISTER_PATH = "mindspore/ops/_op_impl/_custom_op"
KEY_NAME = "name"
ASCEND_CUSTOM_OPP_PATH = "ASCEND_CUSTOM_OPP_PATH"
def _get_reg_info_attr(op_info, attr_name, default_value=None):
"""get attr value"""
for _, item in enumerate(op_info.get("attr", [])):
if item.get(KEY_NAME) == attr_name:
return item.get("defaultValue")
return default_value
class _CustomInstaller:
"""save custom op registration information to a json file which will be used by GE"""
reg_info_hash = [] # used to avoid writing the same reg info to file multiple times
copied_paths = [] # used to avoid copying the same file multiple times
def __init__(self, op_info, func=None):
self.op_info = op_info
self.func = func
self.op_type = op_info.get("op_name") if not func else func.__name__
vendor_name = "ms"
custom_dir = os.path.join(os.path.realpath("./"), "vendors", vendor_name)
self._set_env(custom_dir)
op_impl_dir = os.path.join(custom_dir, "op_impl")
self.ai_core_config_dir = os.path.join(op_impl_dir, "ai_core", "tbe", "config")
self.ai_core_impl_dir = os.path.join(op_impl_dir, "ai_core", "tbe", vendor_name + "_impl")
self.ai_cpu_config_dir = os.path.join(op_impl_dir, "cpu", "config")
self.ai_cpu_impl_dir = os.path.join(op_impl_dir, "cpu", "aicpu_kernel", "impl")
@staticmethod
def _set_env(custom_opp_path):
"""set custom file path to env"""
if not os.environ.get(ASCEND_CUSTOM_OPP_PATH):
os.environ[ASCEND_CUSTOM_OPP_PATH] = custom_opp_path
else:
paths = os.environ[ASCEND_CUSTOM_OPP_PATH].split(':')
if custom_opp_path not in paths:
os.environ[ASCEND_CUSTOM_OPP_PATH] = custom_opp_path + ':' + os.environ[ASCEND_CUSTOM_OPP_PATH]
@staticmethod
def _create_dir(*dir_names):
"""create directory"""
for dir_name in dir_names:
if not os.path.isdir(dir_name):
try:
os.makedirs(dir_name, exist_ok=True)
except OSError as err:
if err.errno == 17: # File exists
pass
else:
raise err
@staticmethod
def _copy_file(src_path, dst_dir):
"""copy file"""
if not os.path.exists(src_path) or src_path in _CustomInstaller.copied_paths:
return
_CustomInstaller.copied_paths.append(src_path)
if os.path.isfile(src_path):
lock_file = os.path.join(dst_dir, "file.lock")
with os.fdopen(os.open(lock_file, os.O_WRONLY | os.O_CREAT | os.O_TRUNC, 0o600), 'w') as f:
fcntl.flock(f.fileno(), fcntl.LOCK_EX)
shutil.copy(src_path, dst_dir)
def check(self):
"""check if the reg info need written"""
if platform.system() != "Linux":
return False
if not os.environ.get("MS_DEV_CUSTOM_OPP_PATH"):
# only process the first time import the mindspore module
return False
if self.op_info.get("target") in ["GPU", "CPU"]:
return False
sha256 = hashlib.sha256()
value = json.dumps(self.op_info, sort_keys=True).encode()
sha256.update(value)
hash_value = sha256.hexdigest()
if hash_value in _CustomInstaller.reg_info_hash:
return False
_CustomInstaller.reg_info_hash.append(hash_value)
return True
def _find_ai_cpu_so_path(self, so_file):
"""find the absolute path of so"""
current_path = os.path.dirname(os.path.abspath(__file__))
search_paths = [current_path + "/../lib", current_path + "/../lib/plugin/ascend"]
for path in search_paths:
so_path = os.path.join(path, so_file)
if os.path.exists(so_path):
return so_path
logger.warning("For Custom op '{}', can not find the aicpu so file '{}' in the following directories:\n{}"
.format(self.op_type, so_file, "\n".join(search_paths)))
return ""
def _gen_ai_core_reg_info(self, imply_path, func_name):
"""generate reg info"""
def _get_dtype_format(idx):
data_type = []
data_format = []
for _, dtype_format in enumerate(self.op_info.get("dtype_format", [])):
if not dtype_format[idx][0]:
data_type = None
else:
data_type.append(dtype_format[idx][0])
if not dtype_format[idx][1]:
data_format = None
else:
if dtype_format[idx][1] == "DefaultFormat":
data_format.append("ND")
else:
data_format.append(dtype_format[idx][1])
return data_type, data_format
op_info = {"opFile": {"value": os.path.splitext(os.path.basename(imply_path))[0]},
"opInterface": {"value": func_name}}
# attr
attrs_name = []
for _, item in enumerate(self.op_info.get("attr", [])):
attr_name = item.get(KEY_NAME)
attrs_name.append(attr_name)
key = "attr_" + attr_name
op_info[key] = {}
for k, v in item.items():
if k != KEY_NAME:
op_info[key][k] = v
if attrs_name:
op_info["attr"] = {"list": ",".join(attrs_name)}
# input and output
inputs = self.op_info.get("inputs", [])
outputs = self.op_info.get("outputs", [])
input_num = len(inputs)
output_num = len(outputs)
for i in range(input_num + output_num):
item = inputs[i] if i < input_num else outputs[i - input_num]
key = "input" if i < input_num else "output"
key += str(item.get("index"))
op_info[key] = {KEY_NAME: item.get(KEY_NAME),
"paramType": item.get("paramType", "required"),
"shape": item.get("shape", "all")}
dtype, formats = _get_dtype_format(i)
if dtype:
op_info[key]["dtype"] = ",".join(dtype)
if formats:
op_info[key]["format"] = ",".join(formats)
return op_info
@staticmethod
def _gen_ai_cpu_reg_info(so_file):
"""generate reg info"""
op_info = {"opInfo": {"computeCost": "100",
"engine": "DNN_VM_AICPU",
"flagAsync": "False",
"flagPartial": "False",
"functionName": "RunCpuKernel",
"kernelSo": so_file,
"opKernelLib": "CUSTAICPUKernel",
"userDefined": "True"}}
return op_info
def _save_op_info(self, dst_dir, file_name, op_info):
"""save op info file"""
repo = {}
save_path = os.path.join(dst_dir, file_name)
lock_file = os.path.join(dst_dir, "file.lock")
with os.fdopen(os.open(lock_file, os.O_WRONLY | os.O_CREAT | os.O_TRUNC, 0o600), 'w') as f:
fcntl.flock(f.fileno(), fcntl.LOCK_EX)
if os.path.isfile(save_path):
with open(save_path, 'r') as fr:
json_str = fr.read()
json_str = "{}" if json_str == "" else json_str
repo = json.loads(json_str)
repo.update({self.op_type: op_info})
with os.fdopen(os.open(save_path, os.O_WRONLY | os.O_CREAT | os.O_TRUNC, 0o600), 'w') as fw:
json.dump(repo, fw, sort_keys=True, indent=4, separators=(',', ':'))
def run(self):
"""save reg info to file"""
if not self.check():
return
so_name = _get_reg_info_attr(self.op_info, "cust_aicpu")
if so_name:
_CustomInstaller._create_dir(self.ai_cpu_config_dir, self.ai_cpu_impl_dir)
# copy so file
so_file = "lib" + so_name + ".so"
imply_path = self._find_ai_cpu_so_path(so_file)
self._copy_file(imply_path, self.ai_cpu_impl_dir)
# generate and copy reg info file
op_info = self._gen_ai_cpu_reg_info(so_file)
self._save_op_info(self.ai_cpu_config_dir, "cust_aicpu_kernel.json", op_info)
else:
_CustomInstaller._create_dir(self.ai_core_config_dir, self.ai_core_impl_dir)
# copy dsl file
imply_path = os.path.realpath(inspect.getfile(self.func))
self._copy_file(imply_path, self.ai_core_impl_dir)
# generate and copy reg info file
op_info = self._gen_ai_core_reg_info(imply_path, self.func.__name__)
self._copy_file(imply_path, self.ai_core_impl_dir)
for arc_name in ["ascend910", "ascend910b", "ascend910c", "ascend310p"]:
arc_dir = os.path.join(self.ai_core_config_dir, arc_name)
_CustomInstaller._create_dir(arc_dir)
self._save_op_info(arc_dir, "aic-{}-ops-info.json".format(arc_name), op_info)
def op_info_register(op_info):
r"""
A decorator which is used to register an operator.
Note:
'op_info' should represent the operator information by string with json format.
The 'op_info' will be added into oplib.
Args:
op_info (Union[str, dict]): operator information in json format.
Examples:
>>> from mindspore.ops import op_info_register, TBERegOp, DataType
>>> abs_op_info = TBERegOp("Abs") \
... .fusion_type("ELEMWISE") \
... .async_flag(False) \
... .binfile_name("abs.so") \
... .compute_cost(10) \
... .kernel_name("abs") \
... .partial_flag(True) \
... .op_pattern("formatAgnostic") \
... .input(0, "x", None, "required", None) \
... .output(0, "y", True, "required", "all") \
... .dtype_format(DataType.F16_None, DataType.F16_None) \
... .dtype_format(DataType.F32_None, DataType.F32_None) \
... .dtype_format(DataType.I32_None, DataType.I32_None) \
... .get_op_info()
>>>
>>> @op_info_register(abs_op_info)
... def _abs_tbe():
... return
...
Returns:
Function, returns a decorator for op info register.
"""
def register_decorator(func):
if isinstance(op_info, dict):
op_info_real = json.dumps(op_info)
else:
op_info_real = op_info
validator.check_value_type("op_info", op_info_real, [str])
op_lib = Oplib()
file_path = os.path.realpath(inspect.getfile(func))
# keep the path custom ops implementation.
if BUILT_IN_CUSTOM_OPS_REGISTER_PATH in file_path:
imply_path = file_path
else:
imply_path = "" if BUILT_IN_OPS_REGISTER_PATH in file_path else file_path
if not op_lib.reg_op(op_info_real, imply_path):
raise ValueError('Invalid op info {}:\n{}\n'.format(file_path, op_info_real))
def wrapped_function(*args, **kwargs):
return func(*args, **kwargs)
return wrapped_function
return register_decorator
[文档]def custom_info_register(*reg_info):
r"""
A decorator which is used to bind the registration information to the `func` parameter of
:class:`mindspore.ops.Custom`.
Note:
The 'reg_info' will be added into oplib.
Args:
reg_info (tuple[str, dict]): Each item represents registration information in json format.
Returns:
Function, returns a decorator for op info register.
Raises:
TypeError: If `reg_info` is not a tuple.
Examples:
>>> from mindspore.ops import custom_info_register, CustomRegOp, DataType
>>> custom_func_ascend_info = CustomRegOp() \
... .input(0, "x", "dynamic") \
... .output(0, "y") \
... .dtype_format(DataType.F16_Default, DataType.F16_Default) \
... .dtype_format(DataType.F32_Default, DataType.F32_Default) \
... .target("Ascend") \
... .get_op_info()
>>>
>>> @custom_info_register(custom_func_ascend_info)
... def custom_func(x):
... pass
"""
def decorator(func):
setattr(func, "reg_info", reg_info)
if reg_info:
used_reg_info = reg_info[0]
if isinstance(used_reg_info, dict):
# ai_cpu should be parsed inside CustomRegOp, skip it here
if not _get_reg_info_attr(used_reg_info, "cust_aicpu"):
_CustomInstaller(used_reg_info, func).run()
@functools.wraps(func)
def wrapper(*args, **kwargs):
return func(*args, **kwargs)
return wrapper
return decorator
class RegOp:
"""
Base class for op info register.
Args:
op_name (str): Name of operator.
"""
def __init__(self, op_name=""):
if not isinstance(op_name, str):
raise ValueError("op name value must be string")
if not op_name.strip():
raise ValueError("op name is empty")
self.op_name = op_name
self.inputs = []
self.outputs = []
self.attr_ = []
self.fusion_type_ = ''
self.dtype_format_ = []
def _is_string(self, value):
"""
Check if the value is a str type.
Args:
value: Parameter to be checked.
Raises:
TypeError: If the type of value is not a str.
"""
if not isinstance(value, str):
raise TypeError("%s value must be str" % str(value))
return True
def _is_int(self, value):
"""
Check if the value is an int.
Args:
value: Parameter to be checked.
Raises:
TypeError: If the type of value is not an int.
"""
if not isinstance(value, int):
raise TypeError("%s value must be int" % str(value))
return True
def _is_bool(self, value):
"""
Check if the value is a bool.
Args:
value: Parameter to be checked.
Raises:
TypeError: If the type of value is not a bool.
"""
if not isinstance(value, bool):
raise TypeError("%s value must be bool" % str(value))
return True
@staticmethod
def _is_list(value):
"""
Check if the value is a list.
Args:
value: Parameter to be checked.
Raises:
TypeError: If the type of value is not a list.
"""
if not isinstance(value, list):
raise TypeError("%s value must be list" % str(value))
return True
def _check_param(self, param_list, key_list, fn_list, kwargs):
"""
Check if the parameter type is correct.
Args:
param_list (list): Parameter list to be checked.
key_list (list): The keys of output dict.
fn_list (list): Function used for parameter checking. If the function list has only one element,
all parameters will use the same function.
kwargs (dict): Other parameter information.
Raises:
TypeError: If the type of value is not list.
ValueError: If the size of param list is not equal to the size of key list, or
the size of param list is not equal to the size of function list.
"""
for i in [param_list, key_list, fn_list]:
if not isinstance(i, list):
raise TypeError("%s value must be list type" % str(i))
if len(param_list) != len(key_list) or (len(fn_list) != 1 and len(param_list) != len(fn_list)):
raise ValueError("param_list size {}, key_list size {}, must be equal.And fn_list size {}.".
format(len(param_list), len(key_list), len(fn_list)))
out_dict = {}
for idx, element in enumerate(param_list):
if element is not None:
if len(fn_list) == 1:
fn_list[0](element)
else:
fn_list[idx](element)
out_dict[key_list[idx]] = element
if kwargs:
out_dict = dict(out_dict, **kwargs)
return out_dict
def fusion_type(self, fusion_type):
"""
Fusion type of the operator.
Args:
fusion_type (str): Value of fusion type.
"""
self._is_string(fusion_type)
self.fusion_type_ = fusion_type
return self
def dtype_format(self, *args):
"""
A dtype and format supported by the operator.
Args:
args (tuple): Value of dtype and format.
Raises:
ValueError: If the size of args not equal to input size add output size.
TypeError: If the type of args is not tuple.
"""
if len(self.inputs) + len(self.outputs) != len(args):
raise ValueError("input size add output size must be equal to dtype format size")
dtype_format = []
for arg in args:
if not isinstance(arg, tuple) or (len(arg) != 2 and len(arg) != 3):
raise ValueError("dtype and format value must be tuple of two or three elements")
self._is_string(arg[0])
self._is_string(arg[1])
if len(arg) == 3:
if self._is_string(arg[2]):
dtype_format.append(arg)
else:
dtype_format.append(arg)
self.dtype_format_.append(tuple(dtype_format))
return self
def get_op_info(self):
"""
Return all registration information for this instance.
The '_' character ending the key is removed here for compatibility with previous version.
Key will be unified into an underlined form later.
"""
op_info = {}
for key, value in self.__dict__.items():
if isinstance(key, str) and key.endswith('_'):
key = key.rstrip('_')
key_dic = {"dynamic_shape_support": "dynamicShapeSupport",
"dynamic_rank_support": "dynamicRankSupport",
"dynamic_compile_static": "dynamicCompileStatic",
"need_check_support": "needCheckSupport",
"dynamic_format": "dynamicFormat"
}
key = key_dic.get(key, key)
op_info[key] = value
return op_info
class CpuRegOp(RegOp):
"""Class for Cpu op info register"""
def __init__(self, op_name):
super(CpuRegOp, self).__init__(op_name)
self.imply_type = "CPU"
def input(self, index=None, name=None, param_type=None, **kwargs):
"""
Register Cpu op input information.
Args:
index (int): Order of the input. Default: ``None`` .
name (str): Name of the input. Default: ``None`` .
param_type (str): Param type of the input. Default: ``None`` .
kwargs (dict): Other information of the input.
"""
param_list = [index, name, param_type]
key_list = ["index", "name", "paramType"]
fn_list = [self._is_int, self._is_string, self._is_string]
input_dict = self._check_param(param_list, key_list, fn_list, kwargs)
self.inputs.append(input_dict)
return self
def output(self, index=None, name=None, param_type=None, **kwargs):
"""
Register AiCPU op output information.
Args:
index (int): Order of the output. Default: ``None`` .
name (str): Name of the output. Default: ``None`` .
param_type (str): Param type of the output. Default: ``None`` .
kwargs (dict): Other information of the output.
"""
param_list = [index, name, param_type]
key_list = ["index", "name", "paramType"]
fn_list = [self._is_int, self._is_string, self._is_string]
output_dict = self._check_param(param_list, key_list, fn_list, kwargs)
self.outputs.append(output_dict)
return self
def attr(self, name=None, value_type=None, value=None, **kwargs):
"""
Register AiCPU op attribute information.
Args:
name (str): Name of the attribute. Default: ``None`` .
value_type (str): Value type of the attribute. Default: ``None`` .
value (str): Value of the attribute. Default: ``None`` .
kwargs (dict): Other information of the attribute.
"""
param_list = [name, value_type, value]
key_list = ["name", "type", "value"]
fn_list = [self._is_string]
attr_dict = self._check_param(param_list, key_list, fn_list, kwargs)
self.attr_.append(attr_dict)
return self
class AkgRegOp(RegOp):
"""Class for Akg op info register."""
def __init__(self, op_name, processor):
super(AkgRegOp, self).__init__(op_name)
self.imply_type = "AKG"
self.processor = processor
def input(self, index=None, name=None, param_type=None, **kwargs):
"""
Register Akg op input information.
Args:
index (int): Order of the input. Default: ``None`` .
name (str): Name of the input. Default: ``None`` .
param_type (str): Param type of the input. Default: ``None`` .
kwargs (dict): Other information of the input.
"""
param_list = [index, name, param_type]
key_list = ["index", "name", "paramType"]
fn_list = [self._is_int, self._is_string, self._is_string]
input_dict = self._check_param(param_list, key_list, fn_list, kwargs)
self.inputs.append(input_dict)
return self
def output(self, index=None, name=None, **kwargs):
"""
Register Akg op output information.
Args:
index (int): Order of the output. Default: ``None`` .
name (str): Name of the output. Default: ``None`` .
kwargs (dict): Other information of the output.
"""
param_list = [index, name]
key_list = ["index", "name"]
fn_list = [self._is_int, self._is_string]
output_dict = self._check_param(param_list, key_list, fn_list, kwargs)
self.outputs.append(output_dict)
return self
def attr(self, name=None, param_type=None, value_type=None, **kwargs):
"""
Register Akg op attribute information.
Args:
name (str): Name of the attribute. Default: ``None`` .
param_type (str): Param type of the attribute. Default: ``None`` .
value_type (str): Value type of the attribute. Default: ``None`` .
kwargs (dict): Other information of the attribute.
"""
param_list = [name, param_type, value_type]
key_list = ["name", "paramType", "type"]
fn_list = [self._is_string]
attr_dict = self._check_param(param_list, key_list, fn_list, kwargs)
self.attr_.append(attr_dict)
return self
class AkgGpuRegOp(AkgRegOp):
"""Class for AkgGpu op info register"""
def __init__(self, op_name):
super(AkgGpuRegOp, self).__init__(op_name, "CUDA")
class AkgAscendRegOp(AkgRegOp):
"""Class for AkgAscend op info register"""
def __init__(self, op_name):
super(AkgAscendRegOp, self).__init__(op_name, "AiCore")
class AkgCpuRegOp(AkgRegOp):
"""Class for AkgCpu op info register"""
def __init__(self, op_name):
super(AkgCpuRegOp, self).__init__(op_name, "CPU")
[文档]class AiCPURegOp(CpuRegOp):
r"""
Class for AiCPU operator information registration.
Args:
op_name (str): Name of operator.
Examples:
>>> from mindspore.ops import AiCPURegOp, DataType
>>> stack_op_info = AiCPURegOp("Stack") \
... .fusion_type("OPAQUE") \
... .attr("axis", "int") \
... .input(0, "x", "dynamic") \
... .output(0, "y", "required") \
... .dtype_format(DataType.I8_Default, DataType.I8_Default) \
... .dtype_format(DataType.I16_Default, DataType.I16_Default) \
... .dtype_format(DataType.I32_Default, DataType.I32_Default) \
... .dtype_format(DataType.I64_Default, DataType.I64_Default) \
... .dtype_format(DataType.U8_Default, DataType.U8_Default) \
... .dtype_format(DataType.U16_Default, DataType.U16_Default) \
... .dtype_format(DataType.U32_Default, DataType.U32_Default) \
... .dtype_format(DataType.U64_Default, DataType.U64_Default) \
... .dtype_format(DataType.F16_Default, DataType.F16_Default) \
... .dtype_format(DataType.F32_Default, DataType.F32_Default) \
... .dtype_format(DataType.F64_Default, DataType.F64_Default) \
... .dtype_format(DataType.BOOL_Default, DataType.BOOL_Default) \
... .get_op_info()
>>>
"""
def __init__(self, op_name):
super(AiCPURegOp, self).__init__(op_name)
self.imply_type = "AiCPU"
[文档]class TBERegOp(RegOp):
r"""
Class for TBE operator information registration. TBE (Tensor Boost Engine) is the Ascend operator development
tool, which is extended on the basis of the TVM framework to develop custom operators.
Args:
op_name (str): Name of operator.
Examples:
>>> from mindspore.ops import TBERegOp, DataType
>>> op_name_op_info = TBERegOp("OpName") \
... .fusion_type("ELEMWISE") \
... .async_flag(False) \
... .binfile_name("op_name.so") \
... .compute_cost(10) \
... .kernel_name("op_name") \
... .partial_flag(True) \
... .op_pattern("formatAgnostic") \
... .need_check_supported(True) \
... .dynamic_shape(True) \
... .dynamic_rank_support(True) \
... .dynamic_compile_static(True) \
... .attr("format", "required", "str", "all") \
... .input(0, "x1", None, "required", None) \
... .input(0, "x2", None, "required", None) \
... .input(1, "axis", None, "required", None) \
... .output(0, "y", True, "required", "all") \
... .real_input_index([1, 0]) \
... .input_to_attr_index([2]) \
... .unknown_shape_formats(["ND", "ND", "ND", "ND"]) \
... .reshape_type("NC") \
... .is_dynamic_format(True) \
... .dtype_format(DataType.F16_None, DataType.F16_None, DataType.F16_None, DataType.F16_None) \
... .dtype_format(DataType.F32_None, DataType.F32_None, DataType.F32_None, DataType.F32_None) \
... .dtype_format(DataType.I32_None, DataType.I32_None, DataType.I32_None, DataType.I32_None) \
... .get_op_info()
>>>
"""
def __init__(self, op_name):
super(TBERegOp, self).__init__(op_name)
self.imply_type = "TBE"
self.async_flag_ = False
self.binfile_ = ''
self.compute_cost_ = 10
self.kernel_ = ''
self.partial_flag_ = False
self.reshape_type_ = ''
self.dynamic_rank_support_ = False
self.dynamic_shape_support_ = False
self.dynamic_compile_static_ = False
self.need_check_support_ = False
self.dynamic_format_ = False
self.op_pattern_ = ""
self.real_input_index_ = []
self.input_to_attr_index_ = []
self.unknown_shape_formats_ = []
[文档] def dynamic_rank_support(self, dynamic_rank_support):
"""
Description whether the operator supports dynamic rank (dynamic dimension).
Args:
dynamic_rank_support (bool): Description whether the operator supports dynamic rank (dynamic dimension).
True: indicates that dynamic rank is supported, and the operator supports
shape (- 2), which is used to determine whether dynamic is performed.
False: indicates that the operator does not support dynamic rank.
Default: ``False`` .
"""
if self._is_bool(dynamic_rank_support):
self.dynamic_rank_support_ = dynamic_rank_support
return self
[文档] def async_flag(self, async_flag=False):
"""
Define the calculation efficiency of the operator, whether the asynchronous calculation is supported.
Args:
async_flag (bool): Value of async flag. Default: ``False`` .
"""
self._is_bool(async_flag)
self.async_flag_ = async_flag
return self
[文档] def binfile_name(self, binfile_name):
"""
Set the binary file name of the operator, it is optional.
Args:
binfile_name (str): The binary file name of the operator.
"""
self._is_string(binfile_name)
self.binfile_ = binfile_name
return self
[文档] def compute_cost(self, compute_cost=10):
"""
Define the calculation efficiency of operator, which refers to the value of the cost model
in the tiling module.
Args:
compute_cost (int): Value of compute cost. Default: ``10`` .
"""
self._is_int(compute_cost)
self.compute_cost_ = compute_cost
return self
[文档] def kernel_name(self, kernel_name):
"""
The name of operator kernel.
Args:
kernel_name (str): Name of operator kernel.
"""
self._is_string(kernel_name)
self.kernel_ = kernel_name
return self
[文档] def partial_flag(self, partial_flag=True):
"""
Define the calculation efficiency of operator, whether the partial calculation is supported.
Args:
partial_flag (bool): Value of partial flag. Default: ``True`` .
"""
self._is_bool(partial_flag)
self.partial_flag_ = partial_flag
return self
[文档] def reshape_type(self, reshape_type):
"""
Reshape type of operator.
Args:
reshape_type (str): Value of reshape type. For example, if the input shape is :math:`(2, 3)`
and `reshape_type` is set to "CH", then the new shape is :math:`(1, 2, 3, 1)`.
"CH" means the C and H dimensions are kept and
new dimensions are added for N and W dimension.
"""
self._is_string(reshape_type)
self.reshape_type_ = reshape_type
return self
[文档] def dynamic_shape(self, dynamic_shape=False):
"""
Whether the operator supports dynamic shape.
Args:
dynamic_shape (bool): Value of dynamic shape. Default: ``False`` .
"""
self._is_bool(dynamic_shape)
self.dynamic_shape_support_ = dynamic_shape
return self
[文档] def dynamic_compile_static(self, dynamic_compile_static=False):
"""
Whether the operator supports dynamic compile static.
Args:
dynamic_compile_static (bool): Value of dynamic compile static. Default: ``False`` .
"""
if self._is_bool(dynamic_compile_static):
self.dynamic_compile_static_ = dynamic_compile_static
return self
[文档] def need_check_supported(self, need_check_supported=False):
"""
Whether the operator needs check supports.
Args:
need_check_supported (bool): Value of need_check_supported. Default: ``False`` .
"""
if self._is_bool(need_check_supported):
self.need_check_support_ = need_check_supported
return self
[文档] def op_pattern(self, pattern=None):
"""
The behavior type of operator, such as broadcast, reduce and so on.
Args:
pattern (str): Value of op pattern, e.g. "broadcast", "reduce". Default: ``None`` .
"""
if pattern is not None and self._is_string(pattern):
self.op_pattern_ = pattern
return self
[文档] def attr(self, name=None, param_type=None, value_type=None, value=None, default_value=None, **kwargs):
"""
Register TBE op attribute information.
Args:
name (str): Name of the attribute. Default: ``None`` .
param_type (str): Param type of the attribute. Default: ``None`` .
value_type (str): Type of the attribute. Default: ``None`` .
value (str): Value of the attribute. Default: ``None`` .
default_value (str): Default value of attribute. Default: ``None`` .
kwargs (dict): Other information of the attribute.
"""
param_list = [name, param_type, value_type, value, default_value]
key_list = ["name", "paramType", "type", "value", "defaultValue"]
fn_list = [self._is_string]
attr_dict = self._check_param(param_list, key_list, fn_list, kwargs)
self.attr_.append(attr_dict)
return self
[文档] def output(self, index=None, name=None, need_compile=None, param_type=None, shape=None, **kwargs):
"""
Register TBE op output information.
Args:
index (int): Order of the output. Default: ``None`` .
name (str): Name of the output. Default: ``None`` .
need_compile (bool): Whether the output needs to be compiled or not. Default: ``None`` .
param_type (str): Type of the output. Default: ``None`` .
shape (str): Shape of the output. Default: ``None`` .
kwargs (dict): Other information of the output.
"""
param_list = [index, name, need_compile, param_type, shape]
key_list = ["index", "name", "need_compile", "paramType", "shape"]
fn_list = [self._is_int, self._is_string, self._is_bool, self._is_string, self._is_string]
output_dict = self._check_param(param_list, key_list, fn_list, kwargs)
self.outputs.append(output_dict)
return self
[文档]class CustomRegOp(RegOp):
r"""
Class used for generating the registration information for the `func` parameter of :class:`mindspore.ops.Custom`.
The registration information mainly specifies the supported data types and formats of input and output tensors,
attributes and target of `func`.
Args:
op_name (str): kernel name. The name will be record in the reg_op_name attr of the kernel node.
Besides, the operator will generate a unique name automatically to identify the reg info.
Default: ``"Custom"`` .
Examples:
>>> from mindspore.ops import CustomRegOp, DataType
>>> custom_op_ascend_info = CustomRegOp() \
... .input(0, "x", "dynamic") \
... .output(0, "y") \
... .dtype_format(DataType.F16_Default, DataType.F16_Default) \
... .dtype_format(DataType.F32_Default, DataType.F32_Default) \
... .target("Ascend") \
... .get_op_info()
"""
def __init__(self, op_name="Custom"):
super(CustomRegOp, self).__init__(op_name)
self.target_ = "UnKnown"
[文档] def output(self, index=None, name=None, param_type="required", **kwargs):
"""
Specifies the output tensor information for the `func` parameter of :class:`mindspore.ops.Custom`. Each
invocation of this function will generate one output tensor information, which means, if `func` has two output
tensors, then this function should be invoked two times continuously. The output tensor information will be
generated as a dict: {"index": `index`, "name": `name`, "param_type": `param_type`}.
Args:
index (int): Index of the output, starts from 0. 0 means the first output tensor, 1 means the second output
tensor and so on. If ``None`` , key "index" will not appear in the output tensor information dict.
Default: ``None`` .
name (str): Name of the `index` 'th output. If ``None`` , key "name" will not appear in the output tensor
information dict. Default: ``None`` .
param_type (str): Parameter type of the `index` 'th output, can be one of
[ ``"required"`` , ``"dynamic"`` , ``"optional"`` ]. If ``None`` , key "param_type" will not appear in
the output tensor information dict. Default: ``"required"`` .
- ``"required"``: means the `index` 'th output exist and can only be a single tensor.
- ``"dynamic"``: means the `index` 'th output exist and may be multiple tensors.
- ``"optional"``: means the `index` 'th output may exist and be a single tensor or may not exist.
kwargs (dict): Other information of the output, used for extension.
Raises:
TypeError: If `index` is neither int nor None.
TypeError: If `name` is neither str nor None.
TypeError: If `param_type` is neither str nor None.
"""
param_list = [index, name, param_type]
key_list = ["index", "name", "paramType"]
fn_list = [self._is_int, self._is_string, self._is_string]
output_dict = self._check_param(param_list, key_list, fn_list, kwargs)
self.outputs.append(output_dict)
return self
[文档] def attr(self, name=None, param_type=None, value_type=None, default_value=None, **kwargs):
"""
Specifies the attributes information for the `func` parameter of :class:`mindspore.ops.Custom`. Each
invocation of this function will generate one attribute information, that means, if `func` has two attributes,
then this function should be invoked two times continuously. The attributes information will be
generated as a dict: {"name": `name`, "param_type": `param_type`, "value_type": `value_type`, "default_value":
`default_value`}.
Args:
name (str): Name of the attribute. If ``None`` , key "name" will not appear in the attributes tensor
information dict. Default: ``None`` .
param_type (str): Parameter type of the attribute, can be one of ["required", "optional"]. If ``None`` ,
key "param_type" will not appear in the attributes tensor information dict. Default: ``None`` .
- "required": means must provide a value for this attribute either by setting a default value in the
registration information or providing an input value when calling the Custom operator.
- "optional": means does not have to provide a value for this attribute.
value_type (str): Value type of the attribute, can be one of ["int", "str", "bool", "float", "listInt",
"listStr", "listBool", "listFloat"]. If ``None`` , key "value_type" will not appear in the attributes
tensor information dict. Default: ``None`` .
- "int": string representation of Python type int.
- "str": string representation of Python type str.
- "bool": string representation of Python type bool.
- "float": string representation of Python type float.
- "listInt": string representation of Python type list of int.
- "listStr": string representation of Python type list of str.
- "listBool": string representation of Python type list of bool.
- "listFloat": string representation of Python type list of float.
default_value (str): Default value of the attribute. `default_value` and `value_type` are used together.
If the real default value of the attribute is float type with value 1.0, then the `value_type` should be
"float" and `default_value` should be "1.0". If the real default value of the attribute is a list of int
with value [1, 2, 3], then the `value_type` should be "listInt" and `default_value` should be "1,2,3",
each item should split by ','. If ``None`` , means the attribute has no default value and key
"default_value" will not appear in the attributes tensor information dict. It is used for "akg",
"aicpu" and "tbe" Custom operators currently. Default: ``None`` .
kwargs (dict): Other information of the attribute, used for extension.
Raises:
TypeError: If `name` is neither str nor None.
TypeError: If `param_type` is neither str nor None.
TypeError: If `value_type` is neither str nor None.
TypeError: If `default_value` is neither str nor None.
"""
param_list = [name, param_type, value_type, default_value]
key_list = ["name", "paramType", "type", "defaultValue"]
fn_list = [self._is_string]
attr_dict = self._check_param(param_list, key_list, fn_list, kwargs)
self.attr_.append(attr_dict)
return self
[文档] def target(self, target=None):
"""
Specifies the target that this registration information is used for.
Args:
target (str): Device target for current operator information, should be one of ["Ascend", "GPU", "CPU"].
For the same `func` of :class:`mindspore.ops.Custom`, it may support different data types and formats
on different targets, use `target` to specify which target that this registration information is used
for. If ``None`` , it will be inferred automatically inside :class:`mindspore.ops.Custom`.
Default: ``None`` .
Raises:
TypeError: If `target` is neither str nor None.
"""
if target is not None:
self._is_string(target)
self.target_ = target
return self
[文档] def get_op_info(self):
"""
Return the generated registration information as a dict. This function should be invoked at last on the
`CustomRegOp` instance as shown in the above example.
"""
op_info = {}
for k, v in self.__dict__.items():
if isinstance(k, str) and k.endswith('_'):
k = k.rstrip('_')
op_info[k] = v
if _get_reg_info_attr(op_info, "cust_aicpu"):
_CustomInstaller(op_info).run()
return op_info
[文档]class DataType:
r"""
Various combinations of dtype and format of Ascend ops.
current support:
.. code-block::
None_None = ("", "")
None_Default = ("", "DefaultFormat")
BOOL_None = ("bool", "")
BOOL_Default = ("bool", "DefaultFormat")
BOOL_5HD = ("bool", "NC1HWC0")
BOOL_FracZ = ("bool", "FRACTAL_Z")
BOOL_FracNZ = ("bool", "FRACTAL_NZ")
BOOL_C1HWNCoC0 = ("bool", "C1HWNCoC0")
BOOL_NCHW = ("bool", "NCHW")
BOOL_NHWC = ("bool", "NHWC")
BOOL_HWCN = ("bool", "HWCN")
BOOL_NDHWC = ("bool", "NDHWC")
BOOL_ChannelLast = ("bool", "ChannelLast")
I8_None = ("int8", "")
I8_Default = ("int8", "DefaultFormat")
I8_5HD = ("int8", "NC1HWC0")
I8_FracZ = ("int8", "FRACTAL_Z")
I8_FracNZ = ("int8", "FRACTAL_NZ")
I8_C1HWNCoC0 = ("int8", "C1HWNCoC0")
I8_NCHW = ("int8", "NCHW")
I8_NHWC = ("int8", "NHWC")
I8_HWCN = ("int8", "HWCN")
I8_NDHWC = ("int8", "NDHWC")
I8_ChannelLast = ("int8", "ChannelLast")
I8_NDC1HWC0 = ("int8", "NDC1HWC0")
U8_None = ("uint8", "")
U8_Default = ("uint8", "DefaultFormat")
U8_5HD = ("uint8", "NC1HWC0")
U8_FracZ = ("uint8", "FRACTAL_Z")
U8_FracNZ = ("uint8", "FRACTAL_NZ")
U8_C1HWNCoC0 = ("uint8", "C1HWNCoC0")
U8_NCHW = ("uint8", "NCHW")
U8_NHWC = ("uint8", "NHWC")
U8_HWCN = ("uint8", "HWCN")
U8_NDHWC = ("uint8", "NDHWC")
U8_ChannelLast = ("uint8", "ChannelLast")
U8_NDC1HWC0 = ("uint8", "NDC1HWC0")
I16_None = ("int16", "")
I16_Default = ("int16", "DefaultFormat")
I16_5HD = ("int16", "NC1HWC0")
I16_FracZ = ("int16", "FRACTAL_Z")
I16_FracNZ = ("int16", "FRACTAL_NZ")
I16_C1HWNCoC0 = ("int16", "C1HWNCoC0")
I16_NCHW = ("int16", "NCHW")
I16_NHWC = ("int16", "NHWC")
I16_HWCN = ("int16", "HWCN")
I16_NDHWC = ("int16", "NDHWC")
I16_ChannelLast = ("int16", "ChannelLast")
U16_None = ("uint16", "")
U16_Default = ("uint16", "DefaultFormat")
U16_5HD = ("uint16", "NC1HWC0")
U16_FracZ = ("uint16", "FRACTAL_Z")
U16_FracNZ = ("uint16", "FRACTAL_NZ")
U16_C1HWNCoC0 = ("uint16", "C1HWNCoC0")
U16_NCHW = ("uint16", "NCHW")
U16_NHWC = ("uint16", "NHWC")
U16_HWCN = ("uint16", "HWCN")
U16_NDHWC = ("uint16", "NDHWC")
U16_ChannelLast = ("uint16", "ChannelLast")
I32_None = ("int32", "")
I32_Default = ("int32", "DefaultFormat")
I32_5HD = ("int32", "NC1HWC0")
I32_FracZ = ("int32", "FRACTAL_Z")
I32_FracNZ = ("int32", "FRACTAL_NZ")
I32_C1HWNCoC0 = ("int32", "C1HWNCoC0")
I32_NCHW = ("int32", "NCHW")
I32_NHWC = ("int32", "NHWC")
I32_HWCN = ("int32", "HWCN")
I32_NDHWC = ("int32", "NDHWC")
I32_ChannelLast = ("int32", "ChannelLast")
U32_None = ("uint32", "")
U32_Default = ("uint32", "DefaultFormat")
U32_5HD = ("uint32", "NC1HWC0")
U32_FracZ = ("uint32", "FRACTAL_Z")
U32_FracNZ = ("uint32", "FRACTAL_NZ")
U32_C1HWNCoC0 = ("uint32", "C1HWNCoC0")
U32_NCHW = ("uint32", "NCHW")
U32_NHWC = ("uint32", "NHWC")
U32_HWCN = ("uint32", "HWCN")
U32_NDHWC = ("uint32", "NDHWC")
U32_ChannelLast = ("uint32", "ChannelLast")
I64_None = ("int64", "")
I64_Default = ("int64", "DefaultFormat")
I64_5HD = ("int64", "NC1HWC0")
I64_FracZ = ("int64", "FRACTAL_Z")
I64_FracNZ = ("int64", "FRACTAL_NZ")
I64_C1HWNCoC0 = ("int64", "C1HWNCoC0")
I64_NCHW = ("int64", "NCHW")
I64_NHWC = ("int64", "NHWC")
I64_HWCN = ("int64", "HWCN")
I64_NDHWC = ("int64", "NDHWC")
I64_ChannelLast = ("int64", "ChannelLast")
U64_None = ("uint64", "")
U64_Default = ("uint64", "DefaultFormat")
U64_5HD = ("uint64", "NC1HWC0")
U64_FracZ = ("uint64", "FRACTAL_Z")
U64_FracNZ = ("uint64", "FRACTAL_NZ")
U64_C1HWNCoC0 = ("uint64", "C1HWNCoC0")
U64_NCHW = ("uint64", "NCHW")
U64_NHWC = ("uint64", "NHWC")
U64_HWCN = ("uint64", "HWCN")
U64_NDHWC = ("uint64", "NDHWC")
U64_ChannelLast = ("uint64", "ChannelLast")
F16_None = ("float16", "")
F16_Default = ("float16", "DefaultFormat")
F16_5HD = ("float16", "NC1HWC0")
F16_FracZ = ("float16", "FRACTAL_Z")
F16_FracNZ = ("float16", "FRACTAL_NZ")
F16_C1HWNCoC0 = ("float16", "C1HWNCoC0")
F16_NCHW = ("float16", "NCHW")
F16_NHWC = ("float16", "NHWC")
F16_HWCN = ("float16", "HWCN")
F16_NDHWC = ("float16", "NDHWC")
F16_NCDHW = ("float16", "NCDHW")
F16_DHWCN = ("float16", "DHWCN")
F16_NDC1HWC0 = ("float16", "NDC1HWC0")
F16_FRACTAL_Z_3D = ("float16", "FRACTAL_Z_3D")
F16_FracZNLSTM = ("float16", "FRACTAL_ZN_LSTM")
F16_FracZNRNN = ("float16", "FRACTAL_ZN_RNN")
F16_ND_RNNBIAS = ("float16", "ND_RNN_BIAS")
F16_ChannelLast = ("float16", "ChannelLast")
F32_None = ("float32", "")
F32_Default = ("float32", "DefaultFormat")
F32_5HD = ("float32", "NC1HWC0")
F32_FracZ = ("float32", "FRACTAL_Z")
F32_FracNZ = ("float32", "FRACTAL_NZ")
F32_C1HWNCoC0 = ("float32", "C1HWNCoC0")
F32_NCHW = ("float32", "NCHW")
F32_NHWC = ("float32", "NHWC")
F32_HWCN = ("float32", "HWCN")
F32_NDHWC = ("float32", "NDHWC")
F32_NCDHW = ("float32", "NCDHW")
F32_DHWCN = ("float32", "DHWCN")
F32_NDC1HWC0 = ("float32", "NDC1HWC0")
F32_FRACTAL_Z_3D = ("float32", "FRACTAL_Z_3D")
F32_FracZNLSTM = ("float32", "FRACTAL_ZN_LSTM")
F32_FracZNRNN = ("float32", "FRACTAL_ZN_RNN")
F32_ND_RNNBIAS = ("float32", "ND_RNN_BIAS")
F32_ChannelLast = ("float32", "ChannelLast")
F64_None = ("float64", "")
F64_Default = ("float64", "DefaultFormat")
F64_5HD = ("float64", "NC1HWC0")
F64_FracZ = ("float64", "FRACTAL_Z")
F64_FracNZ = ("float64", "FRACTAL_NZ")
F64_C1HWNCoC0 = ("float64", "C1HWNCoC0")
F64_NCHW = ("float64", "NCHW")
F64_NHWC = ("float64", "NHWC")
F64_HWCN = ("float64", "HWCN")
F64_NDHWC = ("float64", "NDHWC")
F64_ChannelLast = ("float64", "ChannelLast")
C64_Default = ("complex64", "DefaultFormat")
C128_Default = ("complex128", "DefaultFormat")
"""
None_None = ("", "")
None_Default = ("", "DefaultFormat")
BOOL_None = ("bool", "")
BOOL_Default = ("bool", "DefaultFormat")
BOOL_5HD = ("bool", "NC1HWC0")
BOOL_FracZ = ("bool", "FRACTAL_Z")
BOOL_FracNZ = ("bool", "FRACTAL_NZ")
BOOL_C1HWNCoC0 = ("bool", "C1HWNCoC0")
BOOL_NCHW = ("bool", "NCHW")
BOOL_NHWC = ("bool", "NHWC")
BOOL_HWCN = ("bool", "HWCN")
BOOL_NDHWC = ("bool", "NDHWC")
BOOL_ChannelLast = ("bool", "ChannelLast")
BOOL_Default_Tuple = ("bool", "DefaultFormat", "tuple")
BOOL_Default_List = ("bool", "DefaultFormat", "list")
I8_None = ("int8", "")
I8_Default = ("int8", "DefaultFormat")
I8_5HD = ("int8", "NC1HWC0")
I8_FracZ = ("int8", "FRACTAL_Z")
I8_FracNZ = ("int8", "FRACTAL_NZ")
I8_C1HWNCoC0 = ("int8", "C1HWNCoC0")
I8_NCHW = ("int8", "NCHW")
I8_NHWC = ("int8", "NHWC")
I8_HWCN = ("int8", "HWCN")
I8_NDHWC = ("int8", "NDHWC")
I8_NCDHW = ("int8", "NCDHW")
I8_ChannelLast = ("int8", "ChannelLast")
I8_NDC1HWC0 = ("int8", "NDC1HWC0")
I8_NC1HWC0 = ("int8", "NC1HWC0")
I8_Default_Tuple = ("int8", "DefaultFormat", "tuple")
I8_Default_List = ("int8", "DefaultFormat", "list")
U8_None = ("uint8", "")
U8_Default = ("uint8", "DefaultFormat")
U8_5HD = ("uint8", "NC1HWC0")
U8_FracZ = ("uint8", "FRACTAL_Z")
U8_FracNZ = ("uint8", "FRACTAL_NZ")
U8_C1HWNCoC0 = ("uint8", "C1HWNCoC0")
U8_NCHW = ("uint8", "NCHW")
U8_NHWC = ("uint8", "NHWC")
U8_HWCN = ("uint8", "HWCN")
U8_NDHWC = ("uint8", "NDHWC")
U8_NCDHW = ("uint8", "NCDHW")
U8_ChannelLast = ("uint8", "ChannelLast")
U8_NDC1HWC0 = ("uint8", "NDC1HWC0")
U8_NC1HWC0 = ("uint8", "NC1HWC0")
U8_Default_Tuple = ("uint8", "DefaultFormat", "tuple")
U8_Default_List = ("uint8", "DefaultFormat", "list")
I16_None = ("int16", "")
I16_Default = ("int16", "DefaultFormat")
I16_5HD = ("int16", "NC1HWC0")
I16_FracZ = ("int16", "FRACTAL_Z")
I16_FracNZ = ("int16", "FRACTAL_NZ")
I16_C1HWNCoC0 = ("int16", "C1HWNCoC0")
I16_NCHW = ("int16", "NCHW")
I16_NHWC = ("int16", "NHWC")
I16_HWCN = ("int16", "HWCN")
I16_NDHWC = ("int16", "NDHWC")
I16_ChannelLast = ("int16", "ChannelLast")
I16_Default_Tuple = ("int16", "DefaultFormat", "tuple")
I16_Default_List = ("int16", "DefaultFormat", "list")
U16_None = ("uint16", "")
U16_Default = ("uint16", "DefaultFormat")
U16_5HD = ("uint16", "NC1HWC0")
U16_FracZ = ("uint16", "FRACTAL_Z")
U16_FracNZ = ("uint16", "FRACTAL_NZ")
U16_C1HWNCoC0 = ("uint16", "C1HWNCoC0")
U16_NCHW = ("uint16", "NCHW")
U16_NHWC = ("uint16", "NHWC")
U16_HWCN = ("uint16", "HWCN")
U16_NDHWC = ("uint16", "NDHWC")
U16_ChannelLast = ("uint16", "ChannelLast")
U16_Default_Tuple = ("uint16", "DefaultFormat", "tuple")
U16_Default_List = ("uint16", "DefaultFormat", "list")
I32_None = ("int32", "")
I32_Default = ("int32", "DefaultFormat")
I32_5HD = ("int32", "NC1HWC0")
I32_FracZ = ("int32", "FRACTAL_Z")
I32_FracNZ = ("int32", "FRACTAL_NZ")
I32_C1HWNCoC0 = ("int32", "C1HWNCoC0")
I32_NCHW = ("int32", "NCHW")
I32_NHWC = ("int32", "NHWC")
I32_HWCN = ("int32", "HWCN")
I32_NDHWC = ("int32", "NDHWC")
I32_NDC1HWC0 = ("int32", "NDC1HWC0")
I32_NCDHW = ("int32", "NCDHW")
I32_ChannelLast = ("int32", "ChannelLast")
I32_Default_Tuple = ("int32", "DefaultFormat", "tuple")
I32_Default_List = ("int32", "DefaultFormat", "list")
U32_None = ("uint32", "")
U32_Default = ("uint32", "DefaultFormat")
U32_5HD = ("uint32", "NC1HWC0")
U32_FracZ = ("uint32", "FRACTAL_Z")
U32_FracNZ = ("uint32", "FRACTAL_NZ")
U32_C1HWNCoC0 = ("uint32", "C1HWNCoC0")
U32_NCHW = ("uint32", "NCHW")
U32_NHWC = ("uint32", "NHWC")
U32_HWCN = ("uint32", "HWCN")
U32_NDHWC = ("uint32", "NDHWC")
U32_ChannelLast = ("uint32", "ChannelLast")
U32_Default_Tuple = ("uint32", "DefaultFormat", "tuple")
U32_Default_List = ("uint32", "DefaultFormat", "list")
I64_None = ("int64", "")
I64_Default = ("int64", "DefaultFormat")
I64_5HD = ("int64", "NC1HWC0")
I64_FracZ = ("int64", "FRACTAL_Z")
I64_FracNZ = ("int64", "FRACTAL_NZ")
I64_C1HWNCoC0 = ("int64", "C1HWNCoC0")
I64_NCHW = ("int64", "NCHW")
I64_NHWC = ("int64", "NHWC")
I64_HWCN = ("int64", "HWCN")
I64_NDHWC = ("int64", "NDHWC")
I64_ChannelLast = ("int64", "ChannelLast")
I64_Default_Tuple = ("int64", "DefaultFormat", "tuple")
I64_Default_List = ("int64", "DefaultFormat", "list")
U64_None = ("uint64", "")
U64_Default = ("uint64", "DefaultFormat")
U64_5HD = ("uint64", "NC1HWC0")
U64_FracZ = ("uint64", "FRACTAL_Z")
U64_FracNZ = ("uint64", "FRACTAL_NZ")
U64_C1HWNCoC0 = ("uint64", "C1HWNCoC0")
U64_NCHW = ("uint64", "NCHW")
U64_NHWC = ("uint64", "NHWC")
U64_HWCN = ("uint64", "HWCN")
U64_NDHWC = ("uint64", "NDHWC")
U64_ChannelLast = ("uint64", "ChannelLast")
U64_Default_Tuple = ("uint64", "DefaultFormat", "tuple")
U64_Default_List = ("uint64", "DefaultFormat", "list")
F16_None = ("float16", "")
F16_Default = ("float16", "DefaultFormat")
F16_5HD = ("float16", "NC1HWC0")
F16_FracZ = ("float16", "FRACTAL_Z")
F16_FracNZ = ("float16", "FRACTAL_NZ")
F16_C1HWNCoC0 = ("float16", "C1HWNCoC0")
F16_NCHW = ("float16", "NCHW")
F16_NHWC = ("float16", "NHWC")
F16_HWCN = ("float16", "HWCN")
F16_NDHWC = ("float16", "NDHWC")
F16_NCDHW = ("float16", "NCDHW")
F16_DHWCN = ("float16", "DHWCN")
F16_NDC1HWC0 = ("float16", "NDC1HWC0")
F16_FRACTAL_Z_3D = ("float16", "FRACTAL_Z_3D")
F16_FracZNLSTM = ("float16", "FRACTAL_ZN_LSTM")
F16_FracZNRNN = ("float16", "FRACTAL_ZN_RNN")
F16_ND_RNNBIAS = ("float16", "ND_RNN_BIAS")
F16_ChannelLast = ("float16", "ChannelLast")
F16_Default_Tuple = ("float16", "DefaultFormat", "tuple")
F16_Default_List = ("float16", "DefaultFormat", "list")
F32_None = ("float32", "")
F32_Default = ("float32", "DefaultFormat")
F32_5HD = ("float32", "NC1HWC0")
F32_FracZ = ("float32", "FRACTAL_Z")
F32_FracNZ = ("float32", "FRACTAL_NZ")
F32_C1HWNCoC0 = ("float32", "C1HWNCoC0")
F32_NCHW = ("float32", "NCHW")
F32_NHWC = ("float32", "NHWC")
F32_HWCN = ("float32", "HWCN")
F32_NDHWC = ("float32", "NDHWC")
F32_NCDHW = ("float32", "NCDHW")
F32_DHWCN = ("float32", "DHWCN")
F32_NDC1HWC0 = ("float32", "NDC1HWC0")
F32_FRACTAL_Z_3D = ("float32", "FRACTAL_Z_3D")
F32_FracZNLSTM = ("float32", "FRACTAL_ZN_LSTM")
F32_FracZNRNN = ("float32", "FRACTAL_ZN_RNN")
F32_ND_RNNBIAS = ("float32", "ND_RNN_BIAS")
F32_ChannelLast = ("float32", "ChannelLast")
F32_Default_Tuple = ("float32", "DefaultFormat", "tuple")
F32_Default_List = ("float32", "DefaultFormat", "list")
F64_None = ("float64", "")
F64_Default = ("float64", "DefaultFormat")
F64_5HD = ("float64", "NC1HWC0")
F64_FracZ = ("float64", "FRACTAL_Z")
F64_FracNZ = ("float64", "FRACTAL_NZ")
F64_C1HWNCoC0 = ("float64", "C1HWNCoC0")
F64_NCHW = ("float64", "NCHW")
F64_NHWC = ("float64", "NHWC")
F64_HWCN = ("float64", "HWCN")
F64_NDHWC = ("float64", "NDHWC")
F64_ChannelLast = ("float64", "ChannelLast")
F64_Default_Tuple = ("float64", "DefaultFormat", "tuple")
F64_Default_List = ("float64", "DefaultFormat", "list")
C64_Default = ("complex64", "DefaultFormat")
C128_Default = ("complex128", "DefaultFormat")
C64_Default_Tuple = ("complex64", "DefaultFormat", "tuple")
C128_Default_Tuple = ("complex128", "DefaultFormat", "tuple")