LATTICE: Constraint-Directed Scheduling, Memory Planning, and Pipeline Refinement for NPUs
Overview
LATTICE schedules commands on an AI accelerator so that more data fits in on-chip memory, then refines the pipeline around the chosen memory layout. The order of commands determines which buffers must coexist. On a neural processing unit (NPU), a poor order can force intermediate data out to external memory. LATTICE jointly considers buffer lifetimes, placement and address reuse, and evaluates spill traffic and modeled execution time on command traces.
LATTICE 研究 AI 加速芯片的编译调度。它先调整 NPU 指令顺序以缓解片上内存压力,再根据实际内存布局优化流水线。指令执行顺序决定哪些缓冲区必须同时驻留。片上空间不足会导致中间数据被写回外存,再重新载入。LATTICE 编译器根据数据生命周期安排指令,在分配地址后利用复用约束调整执行时序,并以指令轨迹评估外存流量和建模执行时间。
Preprint, 2026. See the linked paper for the full method, experiments, and latest version.
Research by Minnan Pei (裴旻楠) and collaborators · Email