TransdataOperation

Description

Converts data formats. Converts the ACL_FORMAT_ND format to the ACL_FORMAT_FRACTAL_NZ format, or converts the ACL_FORMAT_FRACTAL_NZ format to the ACL_FORMAT_ND format.

Equivalents: ND and ACL_FORMAT_ND
Equivalents: NZ, ACL_FORMAT_FRACTAL_NZ, FRACTAL_NZ

ACL_FORMAT_ND
- Description in aclFormat
  ACL_FORMAT_ND = 2: any format, applicable to operators that take singular inputs, such as Square and Tanh.
- Detailed description
  Format is the physical layout of data and defines the dimensions for interpreting data. The ND format indicates 1D, 2D, 3D, and 4D. For example, in 4D format, N indicates the number of batches, H indicates the height of the feature map, W indicates the width of the feature map, and C indicates the channel of the feature map.
  
  In 2D format, data is stored in row-first order.
  
  Figure 1 Example of the ACL_FORMAT_ND format
ACL_FORMAT_FRACTAL_NZ
- Description in aclFormat
  ACL_FORMAT_FRACTAL_NZ = 29: a format for internal use only.
- Detailed description
  FRACTAL_NZ is a fractal format (NW1H1H0W0). The entire matrix is divided into (H1*W1) fractals, and each fractal has (H0*W0) elements for each N layers.
  
  Figure 2 Example of the ACL_FORMAT_FRACTAL_NZ format
  
  N=1, W1=4, H1=4.

Definition

struct TransdataParam {
     enum TransdataType : int { 
         UNDEFINED = 0, 
         FRACTAL_NZ_TO_ND, 
         ND_TO_FRACTAL_NZ 
     };
     TransdataType transdataType = UNDEFINED;
     SVector<int64_t> outCrops = {0, 0};
     uint8_t rsv[8] = {0};
};

Parameters

Member	Type	Default Value	Description
transdataType	TransdataType	UNDEFINED	Data format conversion type. Conversion between FRACTAL_NZ and ND is supported. UNDEFINED: default format FRACTAL_NZ_TO_ND: from FRACTAL_NZ to ND ND_TO_FRACTAL_NZ: from ND to FRACTAL_NZ
outCrops	SVector<int64_t>	{0,0}	Last two dimensions of the shape in the original ND data format. This parameter is used only when FRACTAL_NZ is converted to ND
rsv[8]	uint8_t	{0}	Reserved

Member

Type

Default Value

Description

transdataType

TransdataType

UNDEFINED

Data format conversion type. Conversion between FRACTAL_NZ and ND is supported.

UNDEFINED: default format
FRACTAL_NZ_TO_ND: from FRACTAL_NZ to ND
ND_TO_FRACTAL_NZ: from ND to FRACTAL_NZ

outCrops

SVector<int64_t>

{0,0}

Last two dimensions of the shape in the original ND data format. This parameter is used only when FRACTAL_NZ is converted to ND

rsv[8]

uint8_t

{0}

Reserved

Input and Output of ND-to-NZ Conversion

Parameter	Dimension	Data Type	Format	Description
x	[batch, m, n] [m, n]	float16/int8	ND	Input tensor
y	[batch, n1, m1m0, n0]	float16/int8	NZ	An output tensor of the same data type as the input tensor. When the input tensor is of type float16, n0 and m0 are 16. When the input tensor is of type int8, n0 is 32 and m0 is 16. When the input tensor is 2-dimensional, the batch of the output tensor is 1.

Parameter

Dimension

Data Type

Format

Description

[batch, m, n]
[m, n]

float16/int8

Input tensor

[batch, n1, m1m0, n0]

float16/int8

An output tensor of the same data type as the input tensor.

When the input tensor is of type float16, n0 and m0 are 16.
When the input tensor is of type int8, n0 is 32 and m0 is 16.
When the input tensor is 2-dimensional, the batch of the output tensor is 1.

Input and Output of NZ-to-ND Conversion

Parameter	Dimension	Data Type	Format	Description
x	[batch, n1, m1m0, n0]	float16	NZ	Input tensor
y	[batch, m, n]	float16	ND	An output tensor of the same data type as the input tensor

Functions

Converts a tensor in the ACL_FORMAT_ND data format to the ACL_FORMAT_FRACTAL_NZ data format.
- The data type is float16.
  - 2D tensor input
    Input a tensor with shape [m,n], m0 = n0 = 16. The shape of the output tensor is [1,⌈n/n0⌉,⌈m/m0⌉∗m0,n0].
  - 3D tensor input
    Input a tensor with shape [batch,m,n], m0 = n0 = 16. The shape of the output tensor is [batch,⌈n/n0⌉,⌈m/m0⌉∗m0,n0].
- The data type is int8.
  - 2D tensor input
    Input a tensor with shape [m,n], m0 = 16, n0 = 32. The shape of the output tensor is [1,⌈n/n0⌉,⌈m/m0⌉∗m0,n0].
  - 3D tensor input
    Input a tensor with shape [batch,m,n], m0 = 16, n0 = 32. The shape of the output tensor is [batch,⌈n/n0⌉,⌈m/m0⌉∗m0,n0].
Converts a tensor from the ACL_FORMAT_FRACTAL_NZ data format to the ACL_FORMAT_ND data format.
- The data type is float16.
  Input a tensor with shape [b, n1, m1m0, n0], outCrops = {m, n}. The shape of the output tensor is [b, m, n]. The value range of m is (, ], and the value range of n is (, ]).

Restrictions

The NZ dimensions are expressed in the format of {b, n1, m1m0, n0}. The corresponding ND dimensions are {b, m, n}. b indicates the batch. If the batch size is 1, the dimension value is 1 and cannot be omitted. If there are multiple batches, this dimension is the result of merging all batch dimensions.
m0 and n0 indicate the alignment bits. When float16 is used, n0 and m0 are both 16. When int8 is used, n0 is 32 and m0 is 16. m1m0 indicates that the original ND dimension m is aligned upward through the alignment bits. n1 indicates the quotient of dividing the original ND dimension n by n0 after the upward alignment through the alignment bits. For example, if the original ND dimensions are {8, 100, 30}, the corresponding NZ dimensions are {8, 2, 112, 16}.
The length of outCrops must be 2. The value must meet the following requirements:
- If m0m1 falls within (k₁ x 16, (k₁ + 1) x 16] (where k₁ is a positive integer), this interval is the value range of outCrops[0].
- If n0 x n1 falls within (k₂ x 16, (k₂ + 1) x 16] (where k₂ is a positive integer), this interval is the value range of outCrops[1].
In-place write is not supported.

Parent topic: atb/infer_op_params.h