Normalization Module

The norm module is a wrapper which acts as the single interface for all normalization hardware modules. It current supports the instantiation of these normalization layers:

• BatchNorm2d

• LayerNorm

• GroupNorm

• InstanceNorm

• RMSNorm

Note that it implements all normalization layers with the assumption that affine=false.

Parameter Overview

The module has the following parameters, following the hardware metadata standard (see here). Besides PRECISION_DIM_* parameters, which dictate the numerical precision, and TENSOR_SIZE_DIM_*, which is directly inferred from Pytorch tensor shapes, the following parameters can be adjusted to affect hardware performance.

Parameter	Description
NORM_TYPE	Selects the type of normalization layer. Set to "BATCH_NORM", "LAYER_NORM", "INSTANCE_NORM", "GROUP_NORM", or "RMS_NORM".
ISQRT_LUT_MEMFILE	Sets to path to the initialization LUT for inverse square root unit.
SCALE_LUT_MEMFILE	(Only used in BATCH_NORM) Sets scale LUT mem file.
SHIFT_LUT_MEMFILE	(Only used in BATCH_NORM) Sets shift LUT mem file.
DATA_IN_0_PARALLELISM_DIM_0	Hardware compute parallelism across 1st spatial dimension.
DATA_IN_0_PARALLELISM_DIM_1	Hardware compute parallelism across 2nd spatial dimension.
DATA_IN_0_PARALLELISM_DIM_2	Number of channels to normalize across. For LAYER_NORM, set to the number of total channels. For GROUP_NORM, set to total channels divided by number of groups.
DATA_IN_0_PRECISION_0	Input data width.
DATA_IN_0_PRECISION_1	Input fractional data width, must be <= input data width.
DATA_OUT_0_PRECISION_0	Output data width.
DATA_OUT_0_PRECISION_1	Output fractional data width, must be <= output data width.