Each outU[k] port carries a unique row partition of U and must be consumed independently. The buffer holds TP_DIM_ROWS / TP_CASC_LEN rows; the column stride is padded beyond TP_DIM_COLS for alignment. Only the first TP_DIM_COLS columns of each row contain valid data.
All outS ports carry identical data and the buffer holds TP_DIM_COLS valid singular values followed by padding. Only one port needs to be consumed, but all must be connected.
All outV ports carry identical data. The valid V matrix occupies the first TP_DIM_COLS x TP_DIM_COLS elements of the buffer; any remaining elements should be discarded. Only one port needs to be consumed, but all must be connected.