Mojo module

warp

This module includes intrinsics for NVIDIA GPUs shuffle instructions.

Aliases

broadcast: Broadcasts a SIMD value across the warp.
lane_group_max: Reduces a SIMD value to its maximum within a lane group.
lane_group_max_and_broadcast: Reduces and broadcasts max within a lane group.
lane_group_min: Reduces a SIMD value to its minimum within a lane group.
lane_group_reduce: Takes in an input function to computes warp shuffle based reduction operation.
lane_group_sum: Computes the sum within a lane group.
lane_group_sum_and_broadcast: Computes sum and broadcasts within a lane group.
max: Computes the maximum value across the warp.
min: Computes the minimum value across the warp.
reduce: Takes in an input function to computes warp shuffle based reduction operation.
shuffle_down: Copies values from other lanes in the warp.
shuffle_idx: Copies a value from a source lane to other lanes in a warp.
shuffle_up: Copies values from other lanes in the warp.
shuffle_xor: Copies values from between lanes (butterfly pattern).
sum: Computes the sum across the warp.