Description File Parameters

compile_flag

INT32

Deprecated.

Compilation type.

• 0: precise build. In precise build mode, the system performs build based on the user-defined shape without any escape. However, this flag is invalid for AI CPU operators.
• 1: fuzzy build. In fuzzy build mode, the system generalizes the runtime dimensions of dynamic-shape operators before build. Use this feature if you want to run multiple inferences based on one build while the operator's shape range is unknown.

Defaults to 0.

Only operators of the Transformer network are supported currently.

No

op

String

Operator type.

Yes

name

String

Name of a single-operator model file.

If name is not set, the default naming rule of the model file is SN_opType_Input description (dataType_format_shape)_Output description (dataType_format_shape), for example, 0_Add_3_2_3_3_3_2_3_3_3_2_3_3.om.

You can check the enum values of dataType and format from the ${INSTALL_DIR}/include/graph/types.h file. The enum values are sorted in ascending order. Replace ${INSTALL_DIR} with the actual CANN component directory. If the Ascend-CANN-Toolkit package is installed as the root user, the CANN component directory is /usr/local/Ascend/ascend-toolkit/latest.

No

input_desc

TensorDesc array

Operator input description.

Yes

output_desc

TensorDesc array

Operator output description.

Yes

attr

Attr array

Operator attribute.

No

dynamic_input

String

(Optional) Dynamic input. Must be the same as the input name field defined in the operator information library.

Assume that an operator has two groups of dynamic inputs, which are defined as follows:

.DYNAMIC_INPUT(x,...)

.DYNAMIC_INPUT(y,...)

where two groups of dynamic inputs are defined (x and y). The number of inputs in each group is determined by the number of dynamic_input fields. For details, see the description of the name parameter in the TensorDesc array.

• If the constructed single-operator description file contains the name field, this parameter is optional.
• If the constructed single-operator description file does not contain the name field, this parameter is required.
• dynamic_input takes precedence over name.

format

String

(Required) Tensor format supported by the original operator framework.

The mapping between the possible formats and enum values is as follows:

• NCHW: 0
• NHWC: 1
• ND: 2, any format.
• NC1HWC0: 3, a 5D format. C0 is closely related to the micro-architecture, and the value is equal to the Cube Unit size, for example, 16. C1 is obtained by dividing the C dimension by C0, that is, C1 = C/C0. When the division is not exact, the result is rounded up.
• FRACTAL_Z: 4, a format to define convolution weights, converted from the Filter Matrix. It is transferred to the Cube Unit in 4D format of "C1HW,N1,N0,C0". The data is tiled into two layers, as shown in the following figure.

  

  The data of input layer, related to the cube size, is contiguously stored in column-major order (n format). The data of the second layer, related to the matrix size, is contiguously stored in row-major order (Z format).

• FRACTAL_NZ: 29, the fractal format. The format of the output matrix is NW1H1H0W0 during Cube Unit computation. A matrix is divided into (H1 x W1) fractals in column-major order, which looks like an N-shaped layout. Each fractal consists of (H0 x W0) elements in row-major order, resembling a Z-shaped layout. Thus the NW1H1H0W0 format is referred to as the Nz format. (H0 x W0) indicates the size of a fractal, as shown in the following figure.

  

• RESERVED: 40. For any unspecified optional input, set format to RESERVED and set type to UNDEFINED; for any specified optional input, set format and type to the actual values.

After your model is converted, you will see the supported format represented as the corresponding enum value in the file name of the generated offline model (.om file). For example, 1 maps to format NHWC.

origin_format

String

(Optional) Original format of the tensor.

If this field is not included, the implemented format of a tensor is the same as the original format by default.

name

String

(Optional) Tensor name. Required when the operator has dynamic inputs.

This parameter sets the input names in each group of dynamic inputs. Each input name consists of the input name defined in the operator prototype plus an ID, which is indexed starting at 0.

• If the constructed single-operator description file contains the dynamic_input field, this parameter is optional.
• If the constructed single-operator description file does not contain the dynamic_input field, this parameter is required.
• dynamic_input takes precedence over name.

shape

Array of ints

(Required) Tensor shape, for example, [1, 224, 224, 3]. The actual shape product cannot be greater than the maximum int32 value (2147483647).

• For a static shape:

  The dimensions and shape are fixed. In this scenario, shape_range does not need to be configured.

• For a constant shape:

  To specify the input and output shapes of an operator as scalars, set this parameter to "[]" format, for example, "shape": []. In this scenario, shape_range does not need to be configured.

• For a dynamic shape, the shape value can be like the following:
  • If the shape is determined, its unknown dimensions are set to -1, for example, [16, -1, 20, -1]. In this scenario, this parameter must be used together with shape_range. See the following example:

            "shape": [-1,16],
            "shape_range": [[0,32]],

  • If the shape is unknown, its value is -2, for example, "shape": [-2]. In this scenario, shape_range is not supported in the current version and does not need to be configured.

  In the dynamic shape operator execution scenario, the version of the operator library package in the operator execution environment must be the same as that in the operator model compilation environment. Otherwise, an error is reported when the operator is loaded. For details, see Failure in Loading a Single Operator Due to an OPP Version Issue.

origin_shape

String

(Optional) Original shape of the tensor.

If this field is not included, the implemented shape of a tensor is the same as the original shape by default.

type

String

(Required) Tensor data type. The mapping between the possible data types and enum values is as follows.

• bool: 12
• int8: 2
• uint8: 4
• int16: 6
• uint16: 7
• int32: 3
• uint32: 8
• int64: 9
• uint64: 10
• float16/fp16/half: 1
• float/float32: 0
• double: 11
• complex32: 33
• complex64: 16
• complex128: 17
• uint1: 30
• bfloat16: 27
• int4: 29
• UNDEFINED: 28. For any unspecified optional input, set type to UNDEFINED and set format to RESERVED; for any specified optional input, set format and type to the actual values.

After your model is converted, you will see the supported data type represented as the corresponding enum value in the file name of the generated offline model. For example, 2 maps to type int8.

shape_range

int[2] array

(Optional) Value range of an unknown/dynamic shape (excluding shape = -2).

For example, for shape [16, -1, 20, -1], -1 indicates an unknown dimension.

Assume that shape_range is [1, 128], [1, -1], where [1, 128] indicates the value range from 1 to 128, corresponding to the first -1 dimension in the shape value, and [1, -1] indicates the value range from 1 to infinity, corresponding to the second -1 dimension accordingly.

is_const

bool

(Optional) Whether the input is a constant.

• true: constant
• false (default): non-constant.

const_value

list

(Optional) Constant value.

Currently, const_value supports only a one-dimensional list. The number of lists to configure is determined by the value of shape. For example, if the value of shape is 2, there are two lists in const_value.

The data type of const_value is determined by type. If type is set to float16, single-operator build automatically converts const_value to the float16 format.

name

String

(Required) Attribute name

type

String

(Required) Attribute data type:

• bool
• int
• float
• string
• list_bool
• list_int
• list_float
• list_string
• list_list_int
• data_type

value

Determined by type

(Required) Attribute value. Varies according to type.

• bool: true/false
• int: 10
• float: 1.0
• string: "NCHW"
• list_bool: [false, true]
• list_int: [1, 224, 224, 3]
• list_float: [1.0, 0.0]
• list_string: ["str1","str2"]
• list_list_int: [[1, 3, 5, 7], [2, 4, 6, 8]]
• data_type: "DT_FLOAT" or the number corresponding to the enum value, for example, 0.

  For details about other values, see the enum values of DataType or the numbers corresponding to the enum values in ${INSTALL_DIR}/include/graph/types.h. Replace ${INSTALL_DIR} with the actual CANN component directory. If the Ascend-CANN-Toolkit package is installed as the root user, the CANN component directory is /usr/local/Ascend/ascend-toolkit/latest.