dequantize

Classes

DequantizeOperator

DequantizeOperator(op: AirOperator, model: AirModel, platform: SocPlatform, prefix: str = 'aot', attributes: dict[str, str] = {})

DEQUANTIZE operator.

This operator converts a quantized tensor to a float tensor.

Supported inputs: INT8, INT16, FLOAT16 (output must be FLOAT32).

Parameters:

• op

  (AirOperator) –

  The AIR operator to wrap.

• model

  (AirModel) –

  The AIR model.

• platform

  (SocPlatform) –

  The target platform for code generation.

• prefix

  (str, default: 'aot' ) –

  Prefix for generated code files. Defaults to "aot".

• attributes

  (dict[str, str], default: {} ) –

  Attributes for template values. Defaults to {}.

Functions

validate

validate()

Validate the operator configuration.

cmsis_float_dependencies

cmsis_float_dependencies() -> set[type]

DEQUANTIZE's float dependency is limited to the native FP16 cast path.

DEQUANTIZE never calls a gated float kernel: arm_dequantize_s8_f32 / arm_dequantize_s16_f32 live in the ungated arm_nnfunctions.h and emit into a plain float buffer. The only path that touches the float-only ns-cmsis-nn API is the native FP16 input cast (available only on Armv8.1-M FP16 cores), which binds the input as float16_t — a type introduced by arm_nnfunctions_flt.h behind ARM_NN_ENABLE_F16. The portable FP16->FP32 widening path used on non-FP16 cores works purely on uint16_t bit patterns and therefore declares no dependency.

Returns:

• set[type] –

  set[type]: {np.float16} when the operator lowers to the native

• set[type] –

  FP16 cast; otherwise an empty set.

compute_values

compute_values() -> dict[str, str]

Compute the values for the operator.

Returns:

• dict[str, str] –

  dict[str, str]: Computed values for the operator.

emit

emit(save_path: Path)

Generate the source code for the operator.

Parameters:

• save_path

  (Path) –

  Path to save the generated code.

Functions