HexagonTpu.TensorInfo (hexagon_tpu v0.1.0)

Copy Markdown

Metadata for one graph input/output tensor, extracted from the QNN context binary.

type is the Nx type of the raw on-device data. For quantized tensors (quant not nil) the logical values are floats recovered with float = (q - zero_point) * scale (see HexagonTpu.Quant).

Summary

Types

quant()
t()

Functions

qnn_type(code)

Maps a raw Qnn_DataType_t integer to {qnn_type_atom, nx_type}.

Types

quant()

@type quant() ::
  nil
  | %{scale: float(), zero_point: integer()}
  | %{axis: non_neg_integer(), scales: [float()], zero_points: [integer()]}

t()

@type t() :: %HexagonTpu.TensorInfo{
  byte_size: non_neg_integer(),
  index: non_neg_integer(),
  name: String.t(),
  qnn_type: atom(),
  quant: quant(),
  shape: tuple(),
  type: Nx.Type.t()
}

Functions

qnn_type(code)

Maps a raw Qnn_DataType_t integer to {qnn_type_atom, nx_type}.