Constant LOG_SPLIT_BLOCK 

pub const LOG_SPLIT_BLOCK: usize = _; // 7usize

Base-2 log of the low-factor length the tensor split targets.

The row fold consumes a chunk of as many rows as a row has columns. So the low factor spans exactly that many rows. For 128 columns that is 16 tables of 256 field elements, or 64 KiB. That footprint stays resident in a performance core’s L1 data cache across every hi-block. A wider factor needs proportionally more tables and starts missing that cache.