文档

Tensor

Function

Onboard tensor abstraction, in which the memory location, data type, size, and format of tensors can be specified as the data dependency identifiers of instructions.

Prototype

1
class Tensor(mem_type, dtype=None, size=None, format=None, is_inited=False)

Parameters

Parameter

Input Type

Description

mem_type

String

Location of the memory space where the abstracted tensor is located, such as GM, UB, L1, L0A, L0B, L0C, FB, and BT.

dtype

String

Data type, such as BOOL, UINT1, UINT2, UINT8, UINT16, UINT32, BF16, UINT64, INT4, INT8, INT16, INT32, INT64, FP16, and FP32.

size

list

Shape of a tensor.

format

String

Data format. For details, see Format.

is_inited

bool

Switch that indicates whether the tensor class is ready. Once enabled, instructions that utilize the tensors as the input can be initiated.

Constraints

You need to create a tensor whose shape is [1] and is_inited is True for scalar creation.

Example

1
2
3
4
5
6
7
8
9
from mskpp import Tensor, Core
gm_tmp= Tensor("GM", "FP16", [48, 16], format="ND")
with Core("AIV0") as aiv:  # Computing logic on AIV0.
    ...
    gm_temp.load(result, set_value=0)
with Core("AIC0") as aic:
    in_x = Tensor("GM", "FP16", [48, 16], format="ND")
    in_x.load(gm_tmp, expect_value=0) # Computing logic on AIC0.
    ...