Fp16Overflow

Fp16Overflow()

The input magnitude below which a packed FP16 convolution fed from a model input cannot overflow a lane block.

operator_id: str = Field(..., description='Plan operator id of the convolution')

entry: str = Field(..., description='The ns-cmsis-nn entry; its packed FP16 lane sums 32 taps from the bias')

input: str = Field(..., description='The model input tensor the layer reads')

path: list[str] = Field(..., description='Operator types between the model input and the layer, in order')

threshold: float = Field(..., ge=0, description="Input magnitude below which no float16 lane block can reach 65504, allowing for float16 rounding: (65504 / (1 + g) - |bias|) / sum|w| over the first 32 taps and 65504 / (1 + g) / sum|w| over each later block, where g = n*u / (1 - n*u) bounds n round-to-nearest operations (the block's taps, plus one for the bias in block 0) at u = 2^-11; the minimum over output channels and blocks. The layer's float16 output can still overflow when its exact value exceeds the float16 range")

channel: int = Field(..., ge=0, description='Output channel attaining the threshold')

block: int = Field(..., ge=0, description='Block of 32 taps attaining it; block 0 starts from the bias')

weight_sum: float = Field(..., ge=0, description='sum|w| over that block')

bias: float = Field(..., ge=0, description='|bias| of that channel when the block is 0, else 0')