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 language | English |
|---|---|
| Title of host publication | Embedded Computer Systems |
| Subtitle of host publication | Architectures, Modeling, and Simulation - 19th International Conference, SAMOS 2019, Proceedings |
| Editors | Maxime Pelcat, Matthias Jung, Dionisios N. Pnevmatikatos |
| Place of Publication | Cham |
| Publisher | Springer |
| Pages | 387-400 |
| Number of pages | 14 |
| ISBN (Electronic) | 978-3-030-27562-4 |
| ISBN (Print) | 978-3-030-27561-7 |
| DOIs | |
| Publication status | Published - 1 Jan 2019 |
| Event | 19th International Conference on Embedded Computer Systems: Architectures, Modeling, and Simulation, SAMOS 2019 - Samos, Greece Duration: 7 Jul 2019 → 11 Jul 2019 |
Publication series
| Name | Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) |
|---|---|
| Volume | 11733 LNCS |
| ISSN (Print) | 0302-9743 |
| ISSN (Electronic) | 1611-3349 |
Conference
| Conference | 19th International Conference on Embedded Computer Systems: Architectures, Modeling, and Simulation, SAMOS 2019 |
|---|---|
| Country | Greece |
| City | Samos |
| Period | 7/07/19 → 11/07/19 |
Fingerprint
Keywords
- CNN
- Memory efficient
- Reuse
- Scheduling
Cite this
}
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 proceeding › Conference contribution › Academic › peer-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 -