Yashuai Lü
Research Subjects
Yashuai Lü has a total of 19 co-authors. Here are the top co-authors by number of shared publications:
- Hui Guo
National University of Defense Technology
2 shared publications - Lei Wang
National Innovation Institute of Defense Technology, Academy of Military Science, Beijing, China
2 shared publications - Libo Huang
National University of Defense Technology
2 shared publications - Nong Xiao
Sun Yat-sen University
2 shared publications - Run Yan
National University of Defense Technology
2 shared publications - Yongwen Wang
National University of Defense Technology
2 shared publications
Latest Publications
- Open Access
- 5 September 2026
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Architecture and Code Optimization
- Vol. 23 (4) , 1-26
NeuroComp: A Cross-Platform Compiler for Efficient SNN Deployment on Neuromorphic Hardware
Spiking Neural Networks (SNNs) offer substantial energy efficiency advantages, yet deploying them across diverse neuromorphic hardware remains challenging. Existing toolchains exhibit a critical tradeoff: platform-specific frameworks lack portability, while low-level programming interfaces demand extensive hardware expertise and produce non-portable code. Furthermore, conventional deep learning compilers fail to handle the unique temporal dynamics of SNNs and the idiosyncratic constraints of neuromorphic architectures. To bridge this gap, we present NeuroComp , a cross-platform compiler that automatically generates functionally correct and high-performance code for heterogeneous neuromorphic backends. Our core insight is that the diverse execution behaviors of neuromorphic hardware can be universally modeled as parallelizable scalar computations expressible via nested loops. Building on this, NeuroComp introduces a unified multi-level loop abstraction to systematically align SNN computational semantics with hardware capacities. To efficiently navigate the vast implementation space, we propose a decoupled two-stage scheduling strategy. The first stage performs batch partitioning to guarantee functional correctness and strict architectural compliance, while the second stage optimizes inter-batch execution orders for maximum data reuse, guided by lightweight, hardware-specific cost models and heuristic pruning. We evaluate NeuroComp on three representative architectural paradigms: an instruction-extended CPU, a dedicated accelerator, and a many-core hardware. Experimental results demonstrate that NeuroComp consistently achieves competitive performance compared to meticulously hand-optimized implementations and significantly outperforms existing generic automated compilers, effectively minimizing development effort while unlocking cross-platform portability.
- 5 September 2026
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Architecture and Code Optimization
- Vol. 23 (4) , 1-26
Open AccessNeuroComp: A Cross-Platform Compiler for Efficient SNN Deployment on Neuromorphic Hardware
Spiking Neural Networks (SNNs) offer substantial energy efficiency advantages, yet deploying them across diverse neuromorphic hardware remains challenging. Existing toolchains exhibit a critical tradeoff: platform-specific frameworks lack portability, while low-level programming interfaces demand extensive hardware expertise and produce non-portable code. Furthermore, conventional deep learning compilers fail to handle the unique temporal dynamics of SNNs and the idiosyncratic constraints of neuromorphic architectures. To bridge this gap, we present NeuroComp , a cross-platform compiler that automatically generates functionally correct and high-performance code for heterogeneous neuromorphic backends. Our core insight is that the diverse execution behaviors of neuromorphic hardware can be universally modeled as parallelizable scalar computations expressible via nested loops. Building on this, NeuroComp introduces a unified multi-level loop abstraction to systematically align SNN computational semantics with hardware capacities. To efficiently navigate the vast implementation space, we propose a decoupled two-stage scheduling strategy. The first stage performs batch partitioning to guarantee functional correctness and strict architectural compliance, while the second stage optimizes inter-batch execution orders for maximum data reuse, guided by lightweight, hardware-specific cost models and heuristic pruning. We evaluate NeuroComp on three representative architectural paradigms: an instruction-extended CPU, a dedicated accelerator, and a many-core hardware. Experimental results demonstrate that NeuroComp consistently achieves competitive performance compared to meticulously hand-optimized implementations and significantly outperforms existing generic automated compilers, effectively minimizing development effort while unlocking cross-platform portability.
- Open Access
- 1 July 2026
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Architecture and Code Optimization
- Vol. 23 (3) , 1-25
Comperity: A Stateless Spiking Neural Network Accelerator with Differential Spike Encoding
Spiking Neural Networks (SNNs) leverage discrete, event-driven processing to achieve exceptional energy efficiency through intrinsic bit-sparsity. While current SNN accelerators effectively reduce latency by skipping zero-valued spikes, they often overlook the structural redundancy present in spatially adjacent activation patterns, or rely on expensive global search mechanisms to exploit it. We observe that neighboring spike rows frequently exhibit high similarity due to spatial locality. To exploit this efficiently, we introduce Differential Spike Encoding (DSE) , a logic-driven paradigm that decomposes adjacent rows into a shared Base Vector and sparse Differential Vectors . We present Comperity , a specialized hardware accelerator that implements a stateless, index-driven dataflow . By replacing complex pattern matching with lightweight AND/XOR logic arrays, Comperity significantly reduces arithmetic redundancy with negligible control overhead . Experimental results demonstrate that Comperity achieves speedups of 1.62 × , 9.0 × , and 2.3 × over state-of-the-art SNN accelerators Prosperity, PTB, and the A100 GPU on SpikeBERT projection layers, respectively. Moreover, Comperity improves geometric mean energy efficiency by 1.35 × , 10.7 × , and 270 × , respectively. Finally, analytical comparisons confirm that Comperity remains within a competitive range in per-synaptic-operation energy against event-driven neuromorphic platforms (e.g., Loihi and TrueNorth), providing a high-throughput and energy-efficient solution for edge neuromorphic computing.
- 1 July 2026
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Architecture and Code Optimization
- Vol. 23 (3) , 1-25
Open AccessComperity: A Stateless Spiking Neural Network Accelerator with Differential Spike Encoding
Spiking Neural Networks (SNNs) leverage discrete, event-driven processing to achieve exceptional energy efficiency through intrinsic bit-sparsity. While current SNN accelerators effectively reduce latency by skipping zero-valued spikes, they often overlook the structural redundancy present in spatially adjacent activation patterns, or rely on expensive global search mechanisms to exploit it. We observe that neighboring spike rows frequently exhibit high similarity due to spatial locality. To exploit this efficiently, we introduce Differential Spike Encoding (DSE) , a logic-driven paradigm that decomposes adjacent rows into a shared Base Vector and sparse Differential Vectors . We present Comperity , a specialized hardware accelerator that implements a stateless, index-driven dataflow . By replacing complex pattern matching with lightweight AND/XOR logic arrays, Comperity significantly reduces arithmetic redundancy with negligible control overhead . Experimental results demonstrate that Comperity achieves speedups of 1.62 × , 9.0 × , and 2.3 × over state-of-the-art SNN accelerators Prosperity, PTB, and the A100 GPU on SpikeBERT projection layers, respectively. Moreover, Comperity improves geometric mean energy efficiency by 1.35 × , 10.7 × , and 270 × , respectively. Finally, analytical comparisons confirm that Comperity remains within a competitive range in per-synaptic-operation energy against event-driven neuromorphic platforms (e.g., Loihi and TrueNorth), providing a high-throughput and energy-efficient solution for edge neuromorphic computing.
- Open Access
- 27 October 2025
- Published by Association for Computing Machinery (ACM) in ACM Computing Surveys
- Vol. 58 (4) , 1-38
A Survey on Deep Learning for Monte Carlo Path Tracing
Recent strides in hardware-accelerated ray tracing have propelled algorithms once deemed suitable only for offline rendering, like Monte Carlo path tracing, into interactive frame rates. While path tracing has been regarded as a practical utility in animating scenes for the film industry, achieving visually noise-free imagery often mandates thousands of samples per pixel and considerable computation time. Regrettably, this poses a difficulty for video games and virtual reality applications, which demand high frame rates and resolutions, thereby constraining the computational overhead of path tracing. Two extant approaches, in-process sampling, and post-processing reconstruction methods, i.e., denoising and upsampling, address this challenge. The giant evolution of deep learning technology has emerged as pivotal in path tracing processing. We explore and advance Monte Carlo path tracing technology based on deep learning. Moreover, we illustrate the merits and demerits of diverse designs and technologies, propose potential future development trends, and aim at providing researchers with a comprehensive understanding of the cutting-edge in deep learning-driven Monte Carlo path tracing.
- 27 October 2025
- Published by Association for Computing Machinery (ACM) in ACM Computing Surveys
- Vol. 58 (4) , 1-38
Open AccessA Survey on Deep Learning for Monte Carlo Path Tracing
Recent strides in hardware-accelerated ray tracing have propelled algorithms once deemed suitable only for offline rendering, like Monte Carlo path tracing, into interactive frame rates. While path tracing has been regarded as a practical utility in animating scenes for the film industry, achieving visually noise-free imagery often mandates thousands of samples per pixel and considerable computation time. Regrettably, this poses a difficulty for video games and virtual reality applications, which demand high frame rates and resolutions, thereby constraining the computational overhead of path tracing. Two extant approaches, in-process sampling, and post-processing reconstruction methods, i.e., denoising and upsampling, address this challenge. The giant evolution of deep learning technology has emerged as pivotal in path tracing processing. We explore and advance Monte Carlo path tracing technology based on deep learning. Moreover, we illustrate the merits and demerits of diverse designs and technologies, propose potential future development trends, and aim at providing researchers with a comprehensive understanding of the cutting-edge in deep learning-driven Monte Carlo path tracing.
- 27 November 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Very Large Scale Integration (VLSI) Systems
- Vol. 32 (2) , 396-400
MPRTA: An Efficient Multilevel Parallel Mobile Accelerator for High-Performance Ray Tracing
Ray tracing has been regarded as the future of graphics rendering technology for a long time. However, interactive ray tracing still faces challenges, especially in mobile devices, such as high computational intensity and multiple branches. In this brief, we aim to maximize overall efficiency by leveraging all forms of potential parallelism, including task, basic block, loop, and pipeline levels. We present multilevel parallel ray tracing accelerator (MPRTA), an innovative mobile accelerator that offers high performance and optimal efficiency for ray tracing. Experimental results indicate that MPRTA is $1.67\times $ more efficient than the currently best-reported mobile accelerator.
- 27 November 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Very Large Scale Integration (VLSI) Systems
- Vol. 32 (2) , 396-400
MPRTA: An Efficient Multilevel Parallel Mobile Accelerator for High-Performance Ray Tracing
Ray tracing has been regarded as the future of graphics rendering technology for a long time. However, interactive ray tracing still faces challenges, especially in mobile devices, such as high computational intensity and multiple branches. In this brief, we aim to maximize overall efficiency by leveraging all forms of potential parallelism, including task, basic block, loop, and pipeline levels. We present multilevel parallel ray tracing accelerator (MPRTA), an innovative mobile accelerator that offers high performance and optimal efficiency for ray tracing. Experimental results indicate that MPRTA is $1.67\times $ more efficient than the currently best-reported mobile accelerator.