def store_tail

stelliferous@0.0.2.0 · kernel_gemm.bend:78

import Base
import stelliferous@0.0.2.0/kernel_tile.bend as Tile

78def store_tail(+count: U32,output: Array<F32>,+offset: U32,+bias: F32,+accumulator: Tile.Tile8) -> Array<F32>:
79  output = store_lane(U32.is_lt(0,count),output,offset,(Tile.lane0(accumulator) + bias : F32))
80  output = store_lane(U32.is_lt(1,count),output,(offset + 1 : U32),(Tile.lane1(accumulator) + bias : F32))
81  output = store_lane(U32.is_lt(2,count),output,(offset + 2 : U32),(Tile.lane2(accumulator) + bias : F32))
82  output = store_lane(U32.is_lt(3,count),output,(offset + 3 : U32),(Tile.lane3(accumulator) + bias : F32))
83  output = store_lane(U32.is_lt(4,count),output,(offset + 4 : U32),(Tile.lane4(accumulator) + bias : F32))
84  output = store_lane(U32.is_lt(5,count),output,(offset + 5 : U32),(Tile.lane5(accumulator) + bias : F32))
85  output = store_lane(U32.is_lt(6,count),output,(offset + 6 : U32),(Tile.lane6(accumulator) + bias : F32))
86  store_lane(U32.is_lt(7,count),output,(offset + 7 : U32),(Tile.lane7(accumulator) + bias : F32))