摘要
Conventional CGRA multi-task resource allocation predominantly employs spatial partition, resulting in low temporal utilization of tiles. Our study reveals that fusing multiple tasks to share the whole CGRA can improve temporal utilization of tiles and introduce superior task throughput. However, this fuse execution mode brings two challenges: as nodes from multiple tasks' DFGs multiplex the same tiles, load misses from any task will block the execution of other tasks. Besides, existing CGRA mappers fail to constrain the scheduling when fuse mapping multiple tasks' DFGs collaboratively, which results in an imbalanced distribution of nodes along different modulo cycles. Both challenges hinder the full exploitation of the utilization and throughput advantages inherent to the fuse execution mode. To address these challenges, we propose FexMo, a novel framework supporting fused execution mode for multi-task CGRA. Architecturally, FexMo introduces checkpoints with clock-gating mechanisms. The contexts of any tasks that encounter load misses are checkpointed to preserve progress without blocking other tasks, and will be resumed for retrying every time at the same modulo cycle. During load misses, tasks execute NOPs on the clock-gated tiles, preventing errors and saving power. For fuse mapping, FexMo proposes inter-DFGs co-scheduling strategy, which evenly distributes nodes from all tasks across modulo cycles, maximizing temporal utilization of tiles and fully leveraging the throughput benefits of fuse execution mode. Extensive experiments demonstrate that FexMo achieves an average utilization gain of 36.9% and a 1.99 × throughput improvement over the conventional partition execution mode. FexMo-Mapper completes mapping within one minute, delivering solutions with 2.72 × higher quality than conventional mappers on average. Implementation using the ASAP7 PDK flow shows FexMo achieves 434MHz clock frequency with negligible area overhead. Despite inducing additional power consumption, FexMo maintains 1.22 × higher energy efficiency than the baseline on average.
| 源语言 | 英语 |
|---|---|
| 主期刊名 | MICRO 2025 - 58th IEEE/ACM International Symposium on Microarchitecture |
| 出版商 | IEEE Computer Society |
| 页 | 1236-1249 |
| 页数 | 14 |
| ISBN(电子版) | 9798400715730 |
| DOI | |
| 出版状态 | 已出版 - 17 10月 2025 |
| 活动 | 58th IEEE/ACM International Symposium on Microarchitecture , MICRO 2025 - Seoul, 韩国 期限: 18 10月 2025 → 22 10月 2025 |
丛书
| 姓名 | Proceedings of the Annual International Symposium on Microarchitecture, MICRO |
|---|---|
| 卷 | Part of 213862 |
| ISSN(印刷版) | 1072-4451 |
会议
| 会议 | 58th IEEE/ACM International Symposium on Microarchitecture , MICRO 2025 |
|---|---|
| 国家/地区 | 韩国 |
| 市 | Seoul |
| 时期 | 18/10/25 → 22/10/25 |
联合国可持续发展目标
此成果有助于实现下列可持续发展目标:
-
可持续发展目标 7 经济适用的清洁能源
学术指纹
探究 'FexMo: Enabling Fuse Execution Mode for Multi-task CGRAs' 的科研主题。它们共同构成独一无二的学术指纹。引用此
- APA
- Author
- BIBTEX
- Harvard
- Standard
- RIS
- Vancouver