A Survey on System-Level Design of Neural Network Accelerators

In this paper, we present a brief survey on the system-level optimizations used for convolutional neural network (CNN) inference accelerators. For the nested loop of convolutional (CONV) layers, we discuss the effects of loop optimizations such as loop interchange, tiling, unrolling and fusion on CNN accelerators. We also explain memory optimizations that are effective with the loop optimizations. In addition, we discuss streaming architectures and single computation engine architectures that are commonly used in CNN accelerators. Optimizations for CNN models are briefly explained, followed by the recent trends and future directions of the CNN accelerator design.

Research of Register Pressure Aware Loop Unrolling Optimizations for Compiler

Register pressure problem has been a known problem for compiler because of the mismatch between the infinite number of pseudo registers and the finite number of hard registers. Too heavy register pressure may results in register spilling and then leads to performance degradation. There are a lot of optimizations, especially loop optimizations suffer from register spilling in compiler. In order to fight register pressure and therefore improve the effectiveness of compiler, this research takes the register pressure into account to improve loop unrolling optimization during the transformation process. In addition, a register pressure aware transformation is able to reduce the performance overhead of some fine-grained randomization transformations which can be used to defend against ROP attacks. Experiments showed a peak improvement of about 3.6% and an average improvement of about 1% for SPEC CPU 2006 benchmarks and a peak improvement of about 3% and an average improvement of about 1% for the LINPACK benchmark.