ERDSE: efficient reinforcement learning based design space exploration method for CNN accelerator on resource limited platform

ACDSE: A Design Space Exploration Method for CNN Accelerator based on Adaptive Compression Mechanism

“…Heuristic search algorithms can provide beter eiciency than exhaustive search along with the increase of design space scale. Genetic algorithm [12], Bayesian optimization [12,26], Reinforcement Learning (including DQN [17], REINFORCE [1,11,12,38], SAC [12], PPO [7], ERDSE [6]) are popular algorithms for DSE. Among those algorithms, Reinforcement Learning shows beter potential while facing the complicated design space [6,12], because it can dynamically adjust the search policy according to the exploration state.…”

“…In comparison to general-purpose computing engines such as multi-core processors or GPGPUs, customized accelerators, which can achieve higher energy eiciency through elaborately designed data paths and well-organized memory hierarchies, are more suitable for CNN when higher performance or lower power is required. [1,6,11,12,14,15,19,23,27,36,[38][39][40][41][42]. {DI,DO,DK} \ a Parameters about DDR are used to compute the of-chip bandwidth.…”

“…b DI refers to the input data priority scheduling, DO refers to the output data priority scheduling, DK refers to the weight priority scheduling [16]. Although CNN accelerators present superiority in energy eiciency, it is complicated and time-consuming to design CNN accelerators, because various parameters and constraints should be suiciently considered [1,6,11,12,38,40]. Taking the CNN accelerator template from the work of Li et al [16] for instance (Table 1 lists its design parameters and Table 2 deines three constraint conditions according to the commercial FPGA platforms), Fig.…”

“…While facing the complicated design space, DSE methods based on exhaust or random search are hard to guarantee the eiciency. Meanwhile, methods based on heuristic algorithms, such as Genetic Algorithm [12], Bayesian Optimization [22,26], and Reinforcement Learning [1,6,7,12,17,38], may also expose the inferiority of the exploration capacity because low-value samples will misguide the search to local optimal. We choose several algorithms that have been used in previous works to handle accelerator design space exploration with the goal of optimizing forward computing latency.…”

“…Abdelfatah et al introduce a method [1] that divides the search space into hardware and sotware subspaces and implements alternate search to prune the design space, but that method may fall into local optimal solutions, limiting its exploration capacity. he work of Feng et al [6] restricts search in the neighborhood of high-value samples and 2 defines the constraint conditions. The selected methods include random grid search (RGS), Genetic Algorithm (GA), Bayesian Optimization (BO), Deep Q Network (DQN), Sot Actor-Critic (SAC), REINFORCE (RI), Proximal Policy Optimization (PPO), ERDSE and ACDSE.…”