def pack_biases(
biases: np.ndarray,
ifm_scale: float,
ifm_dtype: np.dtype,
weight_scales: np.ndarray,
ofm_scale: float,
is_activation_tanh_or_sigmoid: bool = False,
) -> np.ndarray:
"""
The NPU requires the each bias value to be packed with
output scale parameters in a 80-bit format (that is returned
via npu_encode_bias API). This function will pack such values
to a binary artifact that the NPU will use in the execution.
Parameters
----------
biases : numpy.ndarray
The values of biases
ifm_scale : float
The quantization scale parameter of input feature map
ifm_dtype : numpy.dtype
The data type of input feature map data.
weight_scales : numpy.ndarray
The quantization scale parameter of weight feature map
This could be a tuple if per-channel quantization is present.
ofm_scale : float
The quantization scale parameter of output feature map.
is_activation_tanh_or_sigmoid : bool
Indicates whether the fused activation function is tanh or sigmoid.
Returns
-------
scale_bias : numpy.ndarray
Packed scales/biases as the hardware requires them.
"""
# The BYOC infra should not partition anything else.
supported_ifm_dtypes = (np.uint8, np.int8, np.int16)
assert ifm_dtype in supported_ifm_dtypes
if weight_scales.size == 1:
weight_scales = [weight_scales] * biases.size
hw_bias_scales = _calculate_hw_bias_scales(ifm_scale, weight_scales,
ofm_scale, ifm_dtype,
is_activation_tanh_or_sigmoid)
assert len(hw_bias_scales) == biases.size
biases = biases.astype("int64")
packed_biases = bytearray()
for idx, scale in enumerate(hw_bias_scales):
packed_biases.extend(vapi.npu_encode_bias(biases[idx], *scale))
scale_bias = np.frombuffer(packed_biases, dtype=np.uint8)
scale_bias = np.reshape(scale_bias, (-1, 10))
return scale_bias
def test_encode_bias():
bias_lower_limit = -(1 << (40 - 1))
bias_upper_limit = (1 << (40 - 1)) - 1
scale_lower_limit = 0
scale_upper_limit = (1 << 32) - 1
shift_lower_limit = 0
shift_upper_limit = (1 << 6) - 1
for _ in range(30):
bias = np.int64(random.randint(bias_lower_limit, bias_upper_limit))
scale = int(random.randint(scale_lower_limit, scale_upper_limit))
shift = int(random.randint(shift_lower_limit, shift_upper_limit))
biases_enc = npu_encode_bias(bias, scale, shift)
assert isinstance(biases_enc, bytearray)
assert len(biases_enc) == 10