Automatic memory-efficient scheduling of CNNs

Luc Waeijen, Savvas Sioutas, Yifan He, Maurice Peemen, Henk Corporaal

Research output: Chapter in Book/Report/Conference proceedingConference contributionAcademicpeer-review

8 Downloads (Pure)

Abstract

Accessing large external DRAM is costly, and poses a challenge to efficiently evaluate data-intensive convolutional neural networks (CNNs) on embedded devices. These external memory accesses can be minimized by exploiting data reuse in on-chip memory. Selecting the combination of code transformations that minimize the external DRAM accesses is however an extremely complex task. In this work a mathematical model is presented to quickly and very precisely evaluate combinations of code transformations on CNNs. An accompanying open source tool is developed which leverages this model to perform automated design space exploration and code generation for CNNs. The correctness of the developed model is demonstrated by measurement of seven neural networks. Results show the transformations selected by the tool can reduce external memory accesses by over an order of magnitude.

Original languageEnglish
Title of host publicationEmbedded Computer Systems
Subtitle of host publicationArchitectures, Modeling, and Simulation - 19th International Conference, SAMOS 2019, Proceedings
EditorsMaxime Pelcat, Matthias Jung, Dionisios N. Pnevmatikatos
Place of PublicationCham
PublisherSpringer
Pages387-400
Number of pages14
ISBN (Electronic)978-3-030-27562-4
ISBN (Print)978-3-030-27561-7
DOIs
Publication statusPublished - 1 Jan 2019
Event19th International Conference on Embedded Computer Systems: Architectures, Modeling, and Simulation, SAMOS 2019 - Samos, Greece
Duration: 7 Jul 201911 Jul 2019

Publication series

NameLecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
Volume11733 LNCS
ISSN (Print)0302-9743
ISSN (Electronic)1611-3349

Conference

Conference19th International Conference on Embedded Computer Systems: Architectures, Modeling, and Simulation, SAMOS 2019
CountryGreece
CitySamos
Period7/07/1911/07/19

Fingerprint

Scheduling
Neural Networks
Neural networks
Data storage equipment
External Memory
Dynamic random access storage
Data Reuse
Design Space Exploration
Code Generation
Evaluate
Leverage
Open Source
Correctness
Chip
Mathematical Model
Mathematical models
Minimise
Model

Keywords

  • CNN
  • Memory efficient
  • Reuse
  • Scheduling

Cite this

Waeijen, L., Sioutas, S., He, Y., Peemen, M., & Corporaal, H. (2019). Automatic memory-efficient scheduling of CNNs. In M. Pelcat, M. Jung, & D. N. Pnevmatikatos (Eds.), Embedded Computer Systems: Architectures, Modeling, and Simulation - 19th International Conference, SAMOS 2019, Proceedings (pp. 387-400). (Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics); Vol. 11733 LNCS). Cham: Springer. https://doi.org/10.1007/978-3-030-27562-4_28
Waeijen, Luc ; Sioutas, Savvas ; He, Yifan ; Peemen, Maurice ; Corporaal, Henk. / Automatic memory-efficient scheduling of CNNs. Embedded Computer Systems: Architectures, Modeling, and Simulation - 19th International Conference, SAMOS 2019, Proceedings. editor / Maxime Pelcat ; Matthias Jung ; Dionisios N. Pnevmatikatos. Cham : Springer, 2019. pp. 387-400 (Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)).
@inproceedings{39aefe7fd879435aac3debeb9daa557a,
title = "Automatic memory-efficient scheduling of CNNs",
abstract = "Accessing large external DRAM is costly, and poses a challenge to efficiently evaluate data-intensive convolutional neural networks (CNNs) on embedded devices. These external memory accesses can be minimized by exploiting data reuse in on-chip memory. Selecting the combination of code transformations that minimize the external DRAM accesses is however an extremely complex task. In this work a mathematical model is presented to quickly and very precisely evaluate combinations of code transformations on CNNs. An accompanying open source tool is developed which leverages this model to perform automated design space exploration and code generation for CNNs. The correctness of the developed model is demonstrated by measurement of seven neural networks. Results show the transformations selected by the tool can reduce external memory accesses by over an order of magnitude.",
keywords = "CNN, Memory efficient, Reuse, Scheduling",
author = "Luc Waeijen and Savvas Sioutas and Yifan He and Maurice Peemen and Henk Corporaal",
year = "2019",
month = "1",
day = "1",
doi = "10.1007/978-3-030-27562-4_28",
language = "English",
isbn = "978-3-030-27561-7",
series = "Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)",
publisher = "Springer",
pages = "387--400",
editor = "Maxime Pelcat and Matthias Jung and Pnevmatikatos, {Dionisios N.}",
booktitle = "Embedded Computer Systems",
address = "Germany",

}

Waeijen, L, Sioutas, S, He, Y, Peemen, M & Corporaal, H 2019, Automatic memory-efficient scheduling of CNNs. in M Pelcat, M Jung & DN Pnevmatikatos (eds), Embedded Computer Systems: Architectures, Modeling, and Simulation - 19th International Conference, SAMOS 2019, Proceedings. Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), vol. 11733 LNCS, Springer, Cham, pp. 387-400, 19th International Conference on Embedded Computer Systems: Architectures, Modeling, and Simulation, SAMOS 2019, Samos, Greece, 7/07/19. https://doi.org/10.1007/978-3-030-27562-4_28

Automatic memory-efficient scheduling of CNNs. / Waeijen, Luc; Sioutas, Savvas; He, Yifan; Peemen, Maurice; Corporaal, Henk.

Embedded Computer Systems: Architectures, Modeling, and Simulation - 19th International Conference, SAMOS 2019, Proceedings. ed. / Maxime Pelcat; Matthias Jung; Dionisios N. Pnevmatikatos. Cham : Springer, 2019. p. 387-400 (Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics); Vol. 11733 LNCS).

Research output: Chapter in Book/Report/Conference proceedingConference contributionAcademicpeer-review

TY - GEN

T1 - Automatic memory-efficient scheduling of CNNs

AU - Waeijen, Luc

AU - Sioutas, Savvas

AU - He, Yifan

AU - Peemen, Maurice

AU - Corporaal, Henk

PY - 2019/1/1

Y1 - 2019/1/1

N2 - Accessing large external DRAM is costly, and poses a challenge to efficiently evaluate data-intensive convolutional neural networks (CNNs) on embedded devices. These external memory accesses can be minimized by exploiting data reuse in on-chip memory. Selecting the combination of code transformations that minimize the external DRAM accesses is however an extremely complex task. In this work a mathematical model is presented to quickly and very precisely evaluate combinations of code transformations on CNNs. An accompanying open source tool is developed which leverages this model to perform automated design space exploration and code generation for CNNs. The correctness of the developed model is demonstrated by measurement of seven neural networks. Results show the transformations selected by the tool can reduce external memory accesses by over an order of magnitude.

AB - Accessing large external DRAM is costly, and poses a challenge to efficiently evaluate data-intensive convolutional neural networks (CNNs) on embedded devices. These external memory accesses can be minimized by exploiting data reuse in on-chip memory. Selecting the combination of code transformations that minimize the external DRAM accesses is however an extremely complex task. In this work a mathematical model is presented to quickly and very precisely evaluate combinations of code transformations on CNNs. An accompanying open source tool is developed which leverages this model to perform automated design space exploration and code generation for CNNs. The correctness of the developed model is demonstrated by measurement of seven neural networks. Results show the transformations selected by the tool can reduce external memory accesses by over an order of magnitude.

KW - CNN

KW - Memory efficient

KW - Reuse

KW - Scheduling

UR - http://www.scopus.com/inward/record.url?scp=85071467951&partnerID=8YFLogxK

U2 - 10.1007/978-3-030-27562-4_28

DO - 10.1007/978-3-030-27562-4_28

M3 - Conference contribution

AN - SCOPUS:85071467951

SN - 978-3-030-27561-7

T3 - Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)

SP - 387

EP - 400

BT - Embedded Computer Systems

A2 - Pelcat, Maxime

A2 - Jung, Matthias

A2 - Pnevmatikatos, Dionisios N.

PB - Springer

CY - Cham

ER -

Waeijen L, Sioutas S, He Y, Peemen M, Corporaal H. Automatic memory-efficient scheduling of CNNs. In Pelcat M, Jung M, Pnevmatikatos DN, editors, Embedded Computer Systems: Architectures, Modeling, and Simulation - 19th International Conference, SAMOS 2019, Proceedings. Cham: Springer. 2019. p. 387-400. (Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)). https://doi.org/10.1007/978-3-030-27562-4_28