128B swizzle on a 16×16 matrix — tiling into 16×8 groups eliminates bank conflicts
Tile
Row Range
Column
Original 16×16 Matrix
Tiled for 128B Swizzle (two 16×8 groups)
Bank Sector Activity
dtype
Hardware computes swizzle_col based on physical offset, as if each row has 8 elements. 16 elements per row means each original row will be treated as 2 rows in swizzle computation.