Ruiqi Tang
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Ruiqi Tang has a total of 18 co-authors. Here are the top co-authors by number of shared publications:
- Shichen Peng
School of Microelectronics, State Key Laboratory of Integrated Chips and Systems, Fudan University, Shanghai, China
6 shared publications - Qiuyu Li
State Key Laboratory of Integrated Chips and Systems, Fudan University, Shanghai, China
4 shared publications - Xiaoyang Zeng
Fudan University
4 shared publications - Yibo Fan
Fudan University
4 shared publications - Yupeng Gui
State Key Laboratory of Integrated Chips and Systems, Fudan University, Shanghai, China
4 shared publications - Zhengyang Cao
Fudan University
4 shared publications
Latest Publications
- 11 November 2026
- Published by Elsevier in Integration
- Vol. 112
Bypass-bounded logarithmic response-ratio arbitration for SoC burst interfaces
- 11 November 2026
- Published by Elsevier in Integration
- Vol. 112
Bypass-bounded logarithmic response-ratio arbitration for SoC burst interfaces
- Open Access
- 19 August 2026
- Published by Elsevier in SSRN Electronic Journal
LineJoin-CMO: Safe Clean-Line Invalidation without Draining Unrelated Transient State
A line-scoped cache-block operation (CBO) can be delayed by transient activity that cannot change its target line. Replacing this global drain is not merely an address-filtering problem: an existing target owner must retain a path to completion, new target work must stop entering when ownership clears, and the single-ported directory must remain exclusive through check and update. This paper presents LineJoin-CMO, an effect-scoped quiescence protocol implemented at register-transfer level (RTL) in the synthesizable CVA6/HPDcache CBO.INVAL path. LineJoin-CMO classifies transient state by its remaining effect on the target cache line, drains already-admitted owners, closes target-state admission at the owner-clear boundary, and acquires the directory only after a stable phase transition. Directed tests and a reduced formal model expose distinct bounded non-completion and late-admission failures when the progress and closure boundaries are removed. Across controlled 32-epoch descriptor-ring workloads with 1, 4, and 32 measured overlap opportunities, the first-CBO-to-final-consumer reduction grows from 1.56% to 5.68% and 22.21%. At 32 opportunities, direct fence blocking falls from 2,869 to 158 cycles; endpoint decomposition shows that 1,715 cycles of downstream contention offset part of the local cache-maintenance service saving. Across 24 matched mappings in FreePDK45 and TSMC28, LineJoin-CMO adds one sequential cell and at most 0.287% mapped standard-cell area. At the TSMC28 slow-slow corner and a 3-ns target, the critical delay increases by 2.11% while retaining 0.445 ns slack. These results establish effect-scoped quiescence as a practical, low-cost optimization for clean-line invalidation under overlapping transient traffic.
- 19 August 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessLineJoin-CMO: Safe Clean-Line Invalidation without Draining Unrelated Transient State
A line-scoped cache-block operation (CBO) can be delayed by transient activity that cannot change its target line. Replacing this global drain is not merely an address-filtering problem: an existing target owner must retain a path to completion, new target work must stop entering when ownership clears, and the single-ported directory must remain exclusive through check and update. This paper presents LineJoin-CMO, an effect-scoped quiescence protocol implemented at register-transfer level (RTL) in the synthesizable CVA6/HPDcache CBO.INVAL path. LineJoin-CMO classifies transient state by its remaining effect on the target cache line, drains already-admitted owners, closes target-state admission at the owner-clear boundary, and acquires the directory only after a stable phase transition. Directed tests and a reduced formal model expose distinct bounded non-completion and late-admission failures when the progress and closure boundaries are removed. Across controlled 32-epoch descriptor-ring workloads with 1, 4, and 32 measured overlap opportunities, the first-CBO-to-final-consumer reduction grows from 1.56% to 5.68% and 22.21%. At 32 opportunities, direct fence blocking falls from 2,869 to 158 cycles; endpoint decomposition shows that 1,715 cycles of downstream contention offset part of the local cache-maintenance service saving. Across 24 matched mappings in FreePDK45 and TSMC28, LineJoin-CMO adds one sequential cell and at most 0.287% mapped standard-cell area. At the TSMC28 slow-slow corner and a 3-ns target, the critical delay increases by 2.11% while retaining 0.445 ns slack. These results establish effect-scoped quiescence as a practical, low-cost optimization for clean-line invalidation under overlapping transient traffic.
- Open Access
- 25 July 2026
- Published by Elsevier in SSRN Electronic Journal
Bypass-Bounded Logarithmic Response-Ratio Arbitration for SoC Burst Interfaces
At a non-preemptive system-on-chip (SoC) burst interface, cache-line and direct memory access (DMA) requests have unequal lengths, so grant order affects response. Round robin bounds intervening grants but ignores length, shortest-job-first can repeatedly defer a long head, and exact highest response ratio next (HRRN) requires cross-product comparison or division. We present bypass-bounded logarithmic response-ratio arbitration (BBLR). Its ordinary path ranks heads by the difference between the leading-one positions of W + 1 and declared length, where W is accumulated accepted-service wait. At bypass threshold B, forced-set round robin selects among saturated heads. With N sources, one visible head per source, and eventual burst completion, a continuously eligible head encounters at most B + N - 1 competing grants. We evaluate BBLR using 12.93 million post-L2 events from six public x86 applications and single-outstanding 16/64-beat DMA traffic in 18 environments. At B = 16, BBLR increases mean wall-clock response by 24.4% relative to bounded-SJF under the same forced-service rule. It reduces 99th-percentile and maximum response by 26.5% and 35.4% and improves Jain fairness by 14.2%. Eight-port controllers were routed in ASAP7 and a TSMC28 OpenROAD compatibility flow. Relative to exact HRRN, flat BBLR reduces area by 67.1–72.5% and critical-path delay by 64.8–72.2%. A two-stage pipeline further lowers route-aware delay by 16.0–20.3%. Differential replay and formal checks pass. A serialized PULP Advanced eXtensible Interface (AXI) write instance passes under independent stalls. Changing B shifts the mean–tail–fairness tradeoff. The competing-grant bound remains independent of request length.
- 25 July 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessBypass-Bounded Logarithmic Response-Ratio Arbitration for SoC Burst Interfaces
At a non-preemptive system-on-chip (SoC) burst interface, cache-line and direct memory access (DMA) requests have unequal lengths, so grant order affects response. Round robin bounds intervening grants but ignores length, shortest-job-first can repeatedly defer a long head, and exact highest response ratio next (HRRN) requires cross-product comparison or division. We present bypass-bounded logarithmic response-ratio arbitration (BBLR). Its ordinary path ranks heads by the difference between the leading-one positions of W + 1 and declared length, where W is accumulated accepted-service wait. At bypass threshold B, forced-set round robin selects among saturated heads. With N sources, one visible head per source, and eventual burst completion, a continuously eligible head encounters at most B + N - 1 competing grants. We evaluate BBLR using 12.93 million post-L2 events from six public x86 applications and single-outstanding 16/64-beat DMA traffic in 18 environments. At B = 16, BBLR increases mean wall-clock response by 24.4% relative to bounded-SJF under the same forced-service rule. It reduces 99th-percentile and maximum response by 26.5% and 35.4% and improves Jain fairness by 14.2%. Eight-port controllers were routed in ASAP7 and a TSMC28 OpenROAD compatibility flow. Relative to exact HRRN, flat BBLR reduces area by 67.1–72.5% and critical-path delay by 64.8–72.2%. A two-stage pipeline further lowers route-aware delay by 16.0–20.3%. Differential replay and formal checks pass. A serialized PULP Advanced eXtensible Interface (AXI) write instance passes under independent stalls. Changing B shifts the mean–tail–fairness tradeoff. The competing-grant bound remains independent of request length.
- Open Access
- 22 June 2026
- Published by Association for Computing Machinery (ACM)
- p. 411-418
STP: Semantic-Triggered Prefetching for Event-Driven Workloads
- 22 June 2026
- Published by Association for Computing Machinery (ACM)
- p. 411-418
Open AccessSTP: Semantic-Triggered Prefetching for Event-Driven Workloads