Enhancing TVM VTA Simulator Performance Through SIMD Vectorization
Kuo‐Ting Huang, Tzer-Ying Lin, Po-Wen Cheng, Peng‐Sheng Chen · 2024
In this paper, we enhance TVM VTA (Versatile Tensor Accelerator) simulator performance via SIMD vectorization. The VTA simulator is analyzed and identified the performance bottleneck of the whole execution. Subsequently, the hotspot of the VTA simulator was optimized using two SIMD ISAs: ARM Neon and ARM SVE. The approaches of using SIMD intrinsic functions and inline assembly were adopted in the modifications. We evaluate the execution performance of the proposed modifications running on a general-purpose CPU with SIMD instructions. For the tested benchmark programs, the experiment shows that the SIMD technique can significantly improve the execution performance of the VTA simulator, with a maximum speedup of 2.27 times.