ttnn.Conv2dSliceConfig

class ttnn.Conv2dSliceConfig

Bases: pybind11_object

Conv2dSliceConfig is a structure that is used to configure how the input & output tensors of Conv2D are sliced when they are placed in DRAM. 
Conv2D only supports inputs in L1. If the input tensor or output tensor are too large to fit into L1, then the Conv2d_DRAM version can be used. 
It slices the input & output into slices and applies the Conv2D op on each slice. 
Conv2dSliceConfig determines how this slicing happens.

class SliceTypeEnum

Bases: pybind11_object

Members:

SliceHeight

SliceWidth

SliceHeight = <SliceTypeEnum.SliceHeight: 0>

SliceWidth = <SliceTypeEnum.SliceWidth: 1>

property name

property value

property num_slices: The number of slices that the input & output tensors are divided into.

The output tensor is divided into num_slices slices along the slice_type dimension.

The corresponding input tensor needed to calculate that output is determined and sliced.

If the size of the slice dimension is not divisible by num_slices, then the last slice will be smaller than the rest.

property slice_type: The type of slice to be used. Can be either SliceHeight or SliceWidth. When the tensor is in [N, H, W, C] format, then it can slice either along the height or width dimension.

Slicing along the width is preferable as it reduces the size of the output of the Halo operation.

Use SliceHeight only when the height dimension is much larger than the width dimension.