cumsum_tile

template<CumsumMode mode = CumsumMode::Default>
void ckernel::cumsum_tile_init()

Please refer to documentation for any_init.

template<CumsumMode mode = CumsumMode::Default>
void ckernel::cumsum_tile(uint32_t idst, bool first = true)

Calculates the columnwise (top to bottom) cumulative sum. For multi tile cumulative sum, tiles must come in NWH order (for example using reader_unary_transpose_wh_8bank) and first must be false for all tiles where H != 0. Tiles are also output in NWH order so writer_unary_transpose_wh can be used to store them back in NHW order. The running column totals are carried between the tiles of one column in SFPU registers, so no other SFPU operation may run between the cumsum_tile calls of the same column. After any other SFPU operation, call cumsum_tile_init again before the next cumsum_tile, with the same mode. CumsumMode::Compensated runs a compensated (Kahan) sum: the rounding error of each add is kept next to the running totals and fed back into the next add, so the error does not grow with the length of the scan. CumsumMode::Compensated is only available on Wormhole and Blackhole. Not supported on Quasar (compile error). The DST register buffer must be in acquired state via acquire_dst call. This call is blocking and is only available on the compute engine.

Return value: None

Argument Description

idst The index of the tile in DST register buffer to perform the computation on

first Set true for tiles in the first row