LOMA: Fast Auto-Scheduling on DNN Accelerators through Loop-Order-based Memory Allocation

Symons, Arne; Mei, Linyan; Verhelst, Marian

doi:10.1109/aicas51828.2021.9458493

Cited by 12 publications

(7 citation statements)

References 10 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…To reduce the large mapping search space, LOMA [1] proposed a bottom-up memory allocation strategy independent of the loop ordering. This is possible due to the fact that for a single loop ordering o, the optimal memory allocation m can be inferred with a one-to-one relationship in a bottom-up fashion.…”

Section: Memory Allocationmentioning

confidence: 99%

“…Others, like Interstellar [8] and ZigZag [10] prune some part of the search space during the search through heuristics. LOMA [1] combines an exhaustive search with optional user-provided constraints, providing both unconstrained and constrained search. Timeloop [2] embeds a random search engine in an unconstrained space, failing to consistently provide near-optimum schedules in fast search time.…”

Section: Related Workmentioning

confidence: 99%

“…As a result, many design space exploration (DSE) schedulers [2], [8], [9], [1], have been proposed to automatically find the optimal schedule given a DNN workload and an accelerator HW architecture. However, the above-mentioned schedulers This work is funded in part by the Convolve project evaluated by the EU Horizon Europe research and innovation program under grant agreement No.…”

Section: Introductionmentioning

confidence: 99%

“…Using a dualengine strategy, SALSA consistently reaches near-optimal schedules with an average error margin of 0.007%. 2) To prove its superiority, we extensively compare SALSA with 2 SotA schedulers, LOMA [1] and Timeloop [2]. SALSA always finds schedules with higher or equal quality than Timeloop and LOMA while consequently reducing the search time.…”

Section: Introductionmentioning

confidence: 99%

“…We introduce a new strategy, combining exhaustive search with simulated annealing to address the dynamic nature of the loop ordering design space size across layers. SALSA is extensively benchmarked against two SotA schedulers, LOMA [1] and Timeloop [2] on 5 different DNNs, on average SALSA finds schedules with 11.9% and 7.6% lower energy while speeding-up the search by 1.7× and 24× compared to LOMA and Timeloop, respectively.…”

mentioning

confidence: 99%

See 4 more Smart Citations

SALSA: Simulated Annealing based Loop-Ordering Scheduler for DNN Accelerators

Symons

Mei

et al. 2023

2023 IEEE 5th International Conference on Artificial Intelligence Circuits and Systems (AICAS)

Self Cite

View full text Add to dashboard Cite

To meet the growing need for computational power for DNNs, multiple specialized hardware architectures have been proposed. Each DNN layer should be mapped onto the hardware with the most efficient schedule, however, SotA schedulers struggle to consistently provide optimum schedules in a reasonable time across all DNN-HW combinations.This paper proposes SALSA, a fast dual-engine scheduler to generate optimal execution schedules for both even and uneven mapping. We introduce a new strategy, combining exhaustive search with simulated annealing to address the dynamic nature of the loop ordering design space size across layers. SALSA is extensively benchmarked against two SotA schedulers, LOMA [1] and Timeloop [2] on 5 different DNNs, on average SALSA finds schedules with 11.9% and 7.6% lower energy while speeding-up the search by 1.7× and 24× compared to LOMA and Timeloop, respectively.

show abstract

Section: Memory Allocationmentioning

confidence: 99%