Minghua Shen
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Minghua Shen has a total of 23 co-authors. Here are the top co-authors by number of shared publications:
- Nong Xiao
Sun Yat-sen University
8 shared publications - Guojie Luo
Peking University, Beijing, China
4 shared publications - Fuyu Wang
Sun Yat-sen University, Guangzhou, Guangdong, China
3 shared publications - Yutong Lu
Sun Yat-sen University
3 shared publications Nong Xiao
School of Data and Computer Science, Sun Yat-sen University, Guangzhou, China
2 shared publications- Aoxiang Qin
School of Computer Science and Engineering, Sun Yat-sen University, Guangzhou, China
1 shared publication
Latest Publications
- Open Access
- 26 July 2026
- Published by Association for Computing Machinery (ACM)
- p. 1-7
PRICE: An Operator Recomputation and Fission Framework for Computation Graphs via Attainable Arithmetic Intensity
- 26 July 2026
- Published by Association for Computing Machinery (ACM)
- p. 1-7
Open AccessPRICE: An Operator Recomputation and Fission Framework for Computation Graphs via Attainable Arithmetic Intensity
- Open Access
- 30 June 2025
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Architecture and Code Optimization
- Vol. 22 (2) , 1-23
Ceiba: An Efficient and Scalable DNN Scheduler for Spatial Accelerators
Spatial accelerators are domain-specific architectures to elevate performance and energy efficiency for deep neural networks (DNNs). They also bring a large number of schedule parameters to determine computation and data movement patterns of DNNs. Previous works formulate the schedule problem as design space exploration or integer linear programming. However, these advanced techniques face the challenge of efficiency or scalability. In this article, we propose Ceiba, which is a deep reinforcement learning-based DNN scheduler for spatial accelerators. Ceiba observes the running DNN computation as well as the spatial architecture to make schedule decisions. Then, Ceiba receives a reward to learn and produce the best-fit policy. To provide efficient and scalable scheduling, Ceiba constructs a DNN-architecture-specific action space. It is defined by upper and lower bounds to exclude invalid and sub-optimal schedule candidates. Extensive experiments demonstrate that Ceiba generally provides better performance for spatial accelerators under a fixed number of searching steps or a fixed amount of time. Specifically, Ceiba achieves an average 2.2× speedup for the Simba accelerator, compared with the state-of-the-art scheduler. When scaling the batch size and the hardware architecture up by 64×, the performance gains of Ceiba are 1.8× and 1.2× on average, respectively. Moreover, Ceiba exhibits better scalability for the Eyeriss accelerator.
- 30 June 2025
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Architecture and Code Optimization
- Vol. 22 (2) , 1-23
Open AccessCeiba: An Efficient and Scalable DNN Scheduler for Spatial Accelerators
Spatial accelerators are domain-specific architectures to elevate performance and energy efficiency for deep neural networks (DNNs). They also bring a large number of schedule parameters to determine computation and data movement patterns of DNNs. Previous works formulate the schedule problem as design space exploration or integer linear programming. However, these advanced techniques face the challenge of efficiency or scalability. In this article, we propose Ceiba, which is a deep reinforcement learning-based DNN scheduler for spatial accelerators. Ceiba observes the running DNN computation as well as the spatial architecture to make schedule decisions. Then, Ceiba receives a reward to learn and produce the best-fit policy. To provide efficient and scalable scheduling, Ceiba constructs a DNN-architecture-specific action space. It is defined by upper and lower bounds to exclude invalid and sub-optimal schedule candidates. Extensive experiments demonstrate that Ceiba generally provides better performance for spatial accelerators under a fixed number of searching steps or a fixed amount of time. Specifically, Ceiba achieves an average 2.2× speedup for the Simba accelerator, compared with the state-of-the-art scheduler. When scaling the batch size and the hardware architecture up by 64×, the performance gains of Ceiba are 1.8× and 1.2× on average, respectively. Moreover, Ceiba exhibits better scalability for the Eyeriss accelerator.
- Open Access
- 20 June 2025
- Published by Association for Computing Machinery (ACM)
- p. 299-312
Qtenon: Towards Low-Latency Architecture Integration for Accelerating Hybrid Quantum-Classical Computing
- 20 June 2025
- Published by Association for Computing Machinery (ACM)
- p. 299-312
Open AccessQtenon: Towards Low-Latency Architecture Integration for Accelerating Hybrid Quantum-Classical Computing
- 14 January 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems
- Vol. 44 (8) , 3006-3019
PBS: Program Behavior-Aware Scheduling for High-Level Synthesis
Program behavior comprises operation dependency and resource requirement. They impact the performance of scheduling in high-level synthesis (HLS). Most existing scheduling methods focus on one aspect, resulting in poor performance. In this article, we propose PBS, a program behavior-aware scheduling method for HLS. We leverage a hybrid state encoding scheme to facilitate the comprehensive learning of program behaviors. Moreover, we propose bi-directional GNN and multiresolution aggregation schemes for learning complex operation dependency behavior. These schemes are integrated in an RL framework to iteratively improve scheduling solutions toward low latency and resource usage. Experiments show that PBS provides an average 32.7%, 26.3%, and 25.9% latency reductions, compared with the SDC, GNN-based, and RL-based methods, respectively.
- 14 January 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems
- Vol. 44 (8) , 3006-3019
PBS: Program Behavior-Aware Scheduling for High-Level Synthesis
Program behavior comprises operation dependency and resource requirement. They impact the performance of scheduling in high-level synthesis (HLS). Most existing scheduling methods focus on one aspect, resulting in poor performance. In this article, we propose PBS, a program behavior-aware scheduling method for HLS. We leverage a hybrid state encoding scheme to facilitate the comprehensive learning of program behaviors. Moreover, we propose bi-directional GNN and multiresolution aggregation schemes for learning complex operation dependency behavior. These schemes are integrated in an RL framework to iteratively improve scheduling solutions toward low latency and resource usage. Experiments show that PBS provides an average 32.7%, 26.3%, and 25.9% latency reductions, compared with the SDC, GNN-based, and RL-based methods, respectively.
- 9 May 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Computers
- Vol. 73 (8) , 1899-1912
TensorMap: A Deep RL-Based Tensor Mapping Framework for Spatial Accelerators
The mapping of tensor computation is a complex and important process for spatial accelerators. Today's mapping works depend on hand-tuned kernel libraries or search-based heuristics from human experts. The former is time-intensive while the latter easily leads to sub-optimal performance. In this paper, we propose TensorMap, a deep reinforcement learning (RL)-based mapping framework for tensor computations on spatial accelerators. We propose a sequential generation mode for mapping optimization and construct a coarse-grained action space to reduce the complexity of the mapping search space. An efficient policy network is devised to optimize mapping primitives in the RL-based search. We then propose a stop signal that is sampled from Bernoulli distribution to facilitate multi-level loop unrolling for spatial accelerators. Finally, a genetic algorithm is employed to further refine the optimized mappings. In the experiments, we demonstrate TensorMap's ability for different spatial accelerators with various tensor computations. On TPU, TensorMap provides 2.6$\times$ , 2.7$\times$ , and 2.4$\times$ better energy-delay product (EDP) on average compared with FlexTensor, Ansor, and AMOS respectively. On Eyeriss, TensorMap provides 2.1$\times$ , 1.8$\times$ , and 1.7$\times$ better EDP on average compared with FlexTensor, Ansor, and AMOS respectively.- 9 May 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Computers
- Vol. 73 (8) , 1899-1912
TensorMap: A Deep RL-Based Tensor Mapping Framework for Spatial Accelerators
The mapping of tensor computation is a complex and important process for spatial accelerators. Today's mapping works depend on hand-tuned kernel libraries or search-based heuristics from human experts. The former is time-intensive while the latter easily leads to sub-optimal performance. In this paper, we propose TensorMap, a deep reinforcement learning (RL)-based mapping framework for tensor computations on spatial accelerators. We propose a sequential generation mode for mapping optimization and construct a coarse-grained action space to reduce the complexity of the mapping search space. An efficient policy network is devised to optimize mapping primitives in the RL-based search. We then propose a stop signal that is sampled from Bernoulli distribution to facilitate multi-level loop unrolling for spatial accelerators. Finally, a genetic algorithm is employed to further refine the optimized mappings. In the experiments, we demonstrate TensorMap's ability for different spatial accelerators with various tensor computations. On TPU, TensorMap provides 2.6$\times$ , 2.7$\times$ , and 2.4$\times$ better energy-delay product (EDP) on average compared with FlexTensor, Ansor, and AMOS respectively. On Eyeriss, TensorMap provides 2.1$\times$ , 1.8$\times$ , and 1.7$\times$ better EDP on average compared with FlexTensor, Ansor, and AMOS respectively.- 16 March 2022
- Published by Springer Nature in Applied Intelligence
- Vol. 52 (13) , 15447-15461
Improving the exploration efficiency of DQNs via the confidence bound methods
- 16 March 2022
- Published by Springer Nature in Applied Intelligence
- Vol. 52 (13) , 15447-15461
Improving the exploration efficiency of DQNs via the confidence bound methods
- 1 December 2020
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Parallel and Distributed Systems
- Vol. PP (10459219) , 1
Model Parallelism Optimization for Distributed Inference via Decoupled CNN Structure
In this paper, we propose DeCNN, a more effective inference approach that uses decoupled CNN structure to optimize model parallelism for distributed inference on end-user devices. DeCNN is novel consisting of three schemes. Scheme-1 is structure-level optimization. It exploits group convolution and channel shuffle to decouple the original CNN structure for model parallelism. Scheme-2 is partition-level optimization. It is based on channel group to partition the convolutional layers, and then leverages input-based method to partition the fully connected layers, further exposing high degree of parallelism. Scheme-3 is communication-level optimization. It uses inter-sample parallelism to hide communications for better performance and robustness, especially in the weak network connections. We use ImageNet to evaluate the effectiveness of DeCNN for image classification on a distributed multi-ARM platform. Notably, when using the number of devices from 1 to 4, DeCNN can accelerate the inference of large-scale ResNet50 by 3.21x, and reduce 65.3% memory footprint, with 1.29% accuracy improvement.
- 1 December 2020
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Parallel and Distributed Systems
- Vol. PP (10459219) , 1
Model Parallelism Optimization for Distributed Inference via Decoupled CNN Structure
In this paper, we propose DeCNN, a more effective inference approach that uses decoupled CNN structure to optimize model parallelism for distributed inference on end-user devices. DeCNN is novel consisting of three schemes. Scheme-1 is structure-level optimization. It exploits group convolution and channel shuffle to decouple the original CNN structure for model parallelism. Scheme-2 is partition-level optimization. It is based on channel group to partition the convolutional layers, and then leverages input-based method to partition the fully connected layers, further exposing high degree of parallelism. Scheme-3 is communication-level optimization. It uses inter-sample parallelism to hide communications for better performance and robustness, especially in the weak network connections. We use ImageNet to evaluate the effectiveness of DeCNN for image classification on a distributed multi-ARM platform. Notably, when using the number of devices from 1 to 4, DeCNN can accelerate the inference of large-scale ResNet50 by 3.21x, and reduce 65.3% memory footprint, with 1.29% accuracy improvement.
- 4 November 2020
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Parallel and Distributed Systems
- Vol. 32 (4) , 884-899
Coarse-Grained Parallel Routing With Recursive Partitioning for FPGAs
This paper proposes CPRS, a coarse-grained parallel routing scheme in a distributed computing environment. First, we partition entire routing region to guide the assignment of nets for parallel processing. The partitioning is a recursive fashion, and at each recursive partitioning, the region is partitioned into two subregions forming three subsets of nets. The first subset consists of potentially dependent nets and they are distributed in different subregions. The remaining two subsets consist of potentially independent nets and they are distributed in their own subregions. Second, we route the nets of first subset in serial and process the remaining two subsets in parallel. The parallel processing is a coarse-grained fashion, which is implemented by MPI parallel programming model. Finally, we explore the optimization of both partitioning and parallel processing to further improve the overall speedup of parallel routing. In addition, we adopt MPI message to synchronize the intermediate results between different cores in parallel routing for a feasible solution. Experiments use a set of commonly used benchmarks to demonstrate the effectiveness of CPRS. Notably, CPRS achieves about 18x speedup on average using 32 processor cores with minor loss of quality, compared with the VTR 7.0 serial router.
- 4 November 2020
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Parallel and Distributed Systems
- Vol. 32 (4) , 884-899
Coarse-Grained Parallel Routing With Recursive Partitioning for FPGAs
This paper proposes CPRS, a coarse-grained parallel routing scheme in a distributed computing environment. First, we partition entire routing region to guide the assignment of nets for parallel processing. The partitioning is a recursive fashion, and at each recursive partitioning, the region is partitioned into two subregions forming three subsets of nets. The first subset consists of potentially dependent nets and they are distributed in different subregions. The remaining two subsets consist of potentially independent nets and they are distributed in their own subregions. Second, we route the nets of first subset in serial and process the remaining two subsets in parallel. The parallel processing is a coarse-grained fashion, which is implemented by MPI parallel programming model. Finally, we explore the optimization of both partitioning and parallel processing to further improve the overall speedup of parallel routing. In addition, we adopt MPI message to synchronize the intermediate results between different cores in parallel routing for a feasible solution. Experiments use a set of commonly used benchmarks to demonstrate the effectiveness of CPRS. Notably, CPRS achieves about 18x speedup on average using 32 processor cores with minor loss of quality, compared with the VTR 7.0 serial router.
- 15 October 2020
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems
- Vol. 40 (9) , 1850-1863
Combining Static and Dynamic Load Balance in Parallel Routing for FPGAs
Routing is a very complex process in the field programmable gate array (FPGA) CAD flow. The increase of both FPGA size and design complexity leads to a long routing time hindering the productivity. In this article, we propose a more effective parallel router that combines static and dynamic load balance in parallel routing for FPGAs. First, we explore hierarchical region partitioning to assign routing tasks to different cores for static load balance. Then, we coordinate message propagation and task migration at runtime so that load balance between cores can be dynamically maintained in parallel routing. Finally, we combine static and dynamic load balance in the parallel routing for a higher degree of parallelism. Our parallel router performs on the multicore distributed-memory systems and the communication between cores is through message passing interface messages. We demonstrate the effectiveness of our parallel router using large-scale Titan designs. On average, our parallel router can scale up to 32 cores to achieve about 17× speedup with slight loss of quality, compared with the latest VTR 8 router.
- 15 October 2020
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems
- Vol. 40 (9) , 1850-1863
Combining Static and Dynamic Load Balance in Parallel Routing for FPGAs
Routing is a very complex process in the field programmable gate array (FPGA) CAD flow. The increase of both FPGA size and design complexity leads to a long routing time hindering the productivity. In this article, we propose a more effective parallel router that combines static and dynamic load balance in parallel routing for FPGAs. First, we explore hierarchical region partitioning to assign routing tasks to different cores for static load balance. Then, we coordinate message propagation and task migration at runtime so that load balance between cores can be dynamically maintained in parallel routing. Finally, we combine static and dynamic load balance in the parallel routing for a higher degree of parallelism. Our parallel router performs on the multicore distributed-memory systems and the communication between cores is through message passing interface messages. We demonstrate the effectiveness of our parallel router using large-scale Titan designs. On average, our parallel router can scale up to 32 cores to achieve about 17× speedup with slight loss of quality, compared with the latest VTR 8 router.
- 4 September 2019
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems
- Vol. 39 (10) , 2588-2601
Entropy-Directed Scheduling for FPGA High-Level Synthesis
High-level synthesis (HLS) is important for compiling an application design onto FPGA but still faces challenges of balancing scalability and quality of results in the scheduling process. In this paper, we propose an entropy-directed scheduling (EDS) algorithm that efficiently generates high-quality schedules for FPGA HLS. This work is novel in three ways. First, we make the first attempt to adopt entropy in the scheduling of HLS, which is an intuitive and robust measurement with lots of good analytic properties. Second, we creatively leverage the maximum entropy principle to describe the scheduling process, which is proved equivalent to the optimal solution in some particular cases. Third, we make EDS automatically analyze the input graph structure and leverage a three-stage scheduling process to obtain high-quality results. As a result, EDS has the lowest time complexity among existing scheduling algorithms and is flexible to solve both latency-and resource-constrained problems while satisfying other constraints. Experimental results show that for latency-constrained scheduling problem, EDS reduces up to 68% resource usage and is 294x faster than the force-directed scheduling algorithm. For resource-constrained scheduling problem, EDS obtains near-optimal solutions with an average speedup of 16410x compared with ILP. To our best knowledge, this is the first entropy-directed scheduling algorithm for FPGA HLS.
- 4 September 2019
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems
- Vol. 39 (10) , 2588-2601
Entropy-Directed Scheduling for FPGA High-Level Synthesis
High-level synthesis (HLS) is important for compiling an application design onto FPGA but still faces challenges of balancing scalability and quality of results in the scheduling process. In this paper, we propose an entropy-directed scheduling (EDS) algorithm that efficiently generates high-quality schedules for FPGA HLS. This work is novel in three ways. First, we make the first attempt to adopt entropy in the scheduling of HLS, which is an intuitive and robust measurement with lots of good analytic properties. Second, we creatively leverage the maximum entropy principle to describe the scheduling process, which is proved equivalent to the optimal solution in some particular cases. Third, we make EDS automatically analyze the input graph structure and leverage a three-stage scheduling process to obtain high-quality results. As a result, EDS has the lowest time complexity among existing scheduling algorithms and is flexible to solve both latency-and resource-constrained problems while satisfying other constraints. Experimental results show that for latency-constrained scheduling problem, EDS reduces up to 68% resource usage and is 294x faster than the force-directed scheduling algorithm. For resource-constrained scheduling problem, EDS obtains near-optimal solutions with an average speedup of 16410x compared with ILP. To our best knowledge, this is the first entropy-directed scheduling algorithm for FPGA HLS.