For the complete documentation index, see llms.txt. Markdown versions of all pages are available by appending .md to any URL (e.g. /get-started.md).
Mojo function
sum
def sum[axis: Int](inp: TileTensor[Engine=inp.Engine, address_space=inp.address_space, linear_idx_type=inp.linear_idx_type], outp: TileTensor[Engine=outp.Engine, address_space=outp.address_space, linear_idx_type=outp.linear_idx_type])
Computes lane-wise sums along one outer dimension.
Constraints:
Shapes must be static. The surviving dimension and SIMD element width must match between input and output.
Parameters:
- axis (
Int): The outer dimension to reduce, either 0 or 1.
Args:
- inp (
TileTensor[Engine=inp.Engine, address_space=inp.address_space, linear_idx_type=inp.linear_idx_type]): The rank-2 input tensor, including nested outer dimensions. - outp (
TileTensor[Engine=outp.Engine, address_space=outp.address_space, linear_idx_type=outp.linear_idx_type]): The rank-1 output tensor. Values are cast to its dtype before accumulation.
def sum[axis: Int](inp: TileTensor[Engine=inp.Engine, address_space=inp.address_space, linear_idx_type=inp.linear_idx_type]) -> TileTensor[inp.dtype, Layout[TypeList[ComptimeInt[(Int((add (mul Coord[inp.LayoutType.__shape_types[SIMDLength((Int(1) - axis))]].static_product, inp.Engine.element_size), inp.Engine.element_size, -1)) // inp.Engine.element_size)]](), TypeList[ComptimeInt[inp.Engine.element_size]]()], MutUntrackedOrigin, Engine=DefaultEngine[element_width=Coord[ComptimeInt[inp.Engine.element_size]].static_product], address_space=inp.address_space, linear_idx_type=_get_index_type[Layout[TypeList[ComptimeInt[(Coord[inp.LayoutType.__shape_types[SIMDLength((Int(1) - axis))]].static_product * inp.Engine.element_size)]](), TypeList[ComptimeInt[Int(1)]]()]](inp.address_space)]
Computes lane-wise sums into a new contiguous tensor.
Examples:
from std.collections import Array
from layout import TileTensor, row_major
from layout.math import sum
var data: Array[Int32, 6] = [0, 1, 2, 3, 4, 5]
var tensor = TileTensor(data, row_major[2, 3]())
var result = sum[0](tensor)
print(result[0], result[1], result[2]) # 3 5 7Parameters:
- axis (
Int): The outer dimension to reduce, either 0 or 1.
Args:
- inp (
TileTensor[Engine=inp.Engine, address_space=inp.address_space, linear_idx_type=inp.linear_idx_type]): A statically shaped rank-2 tensor.
Returns:
TileTensor[inp.dtype, Layout[TypeList[ComptimeInt[(Int((add (mul Coord[inp.LayoutType.__shape_types[SIMDLength((Int(1) - axis))]].static_product, inp.Engine.element_size), inp.Engine.element_size, -1)) // inp.Engine.element_size)]](), TypeList[ComptimeInt[inp.Engine.element_size]]()], MutUntrackedOrigin, Engine=DefaultEngine[element_width=Coord[ComptimeInt[inp.Engine.element_size]].static_product], address_space=inp.address_space, linear_idx_type=_get_index_type[Layout[TypeList[ComptimeInt[(Coord[inp.LayoutType.__shape_types[SIMDLength((Int(1) - axis))]].static_product * inp.Engine.element_size)]](), TypeList[ComptimeInt[Int(1)]]()]](inp.address_space)]: A rank-1 tensor with the input dtype and SIMD element width. Adjacent
elements have disjoint storage.