Mojo function

pack_filter

pack_filter(filter: LayoutTensor[dtype, layout, origin, address_space=address_space, element_layout=element_layout, layout_int_type=layout_int_type, linear_idx_type=linear_idx_type, masked=masked, alignment=alignment], packed_filter: LayoutTensor[dtype, layout, origin, address_space=address_space, element_layout=element_layout, layout_int_type=layout_int_type, linear_idx_type=linear_idx_type, masked=masked, alignment=alignment], num_groups: Int)

This packs the filter form RSCF to FRSCf. Use the default micro kernel size for dynamic shapes.

pack_filter[simd_size: Int, micro_kernel_f_size: Int](filter: LayoutTensor[dtype, layout, origin, address_space=address_space, element_layout=element_layout, layout_int_type=layout_int_type, linear_idx_type=linear_idx_type, masked=masked, alignment=alignment], packed_filter: LayoutTensor[dtype, layout, origin, address_space=address_space, element_layout=element_layout, layout_int_type=layout_int_type, linear_idx_type=linear_idx_type, masked=masked, alignment=alignment], num_groups: Int)

This packs the filter form RSCF to FRSCf.

F is first broken down to segments of size micro_kernel_f_size, then the remainder is further divided by simd_size. The last residual elements if any is padded with zero to fill simd_size.

Parameters:

simd_size (Int): Can differ from the simd size of the input type.
micro_kernel_f_size (Int): The size of the last dimension in FRSCf, which is equals the size of the micro kernel's F dimension.

Args:

filter (LayoutTensor): Filter in RSCF layout (if 2D).
packed_filter (LayoutTensor): Packed filter in FRSCf layout (if 2D). F - the index of continuous segments in micro kernel. R, S, C - original R, S, C. f - the index within a continuous segments.
num_groups (Int): The number of groups in the convolution.