kernel_cost.bend source

kernel_cost.bend on the hub · documented module

# Prediction parameters, not machine bounds or correctness certificates. The# worker count is not a profile value: callers pass the runtime's pool size.import Basetype Configuration is Data:  Configuration{mac_ns: F32,copy_byte_ns: F32,operation_ns: F32,grain_ns: F32,wake_round_ns: F32}# One element's cost. Bytes price loading, storing and the loop that moves# them, as the copies that measured copy_byte_ns do; operations are the other# scalar work, counted from source.def element_ns(+bytes: F32,+operations: F32,config: Configuration) -> F32:  Configuration{mac_ns,byte_ns,operation_ns,grain,wake} = config  (bytes * byte_ns + operations * operation_ns : F32)# The official runtime grows fork tasks and queues fork-free tasks for its# work round (monk_step, row_grow, cube_run). No fixed number of leaves can# be subtracted as already completed by growth. Price complete waves of at# most workers leaves, using the largest path (including clones) per wave.# This is a conservative scheduling estimate, not a bound on the runtime's# static ring distribution. Callers charge the two wake rounds separately.def region_ns(+workers: U32,+leaves: U32,+span: F32) -> F32:  +pool = U32.max(workers,1)  +waves = (leaves / pool + Bool.pick(U32,U32.is_zero((leaves % pool : U32)),0,1) : U32)  (span * U32.to_f32(waves) : F32)