Flexible function-level acceleration of embedded vision applications using the Pipelined Vision Processor

摘要

The emerging massive embedded vision market is driving demanding and ever-increasing computationally complex high-performance and low-power MPSoC requirements. To satisfy these requirements innovative solutions are required to deliver high performance pixel processing combined with low energy per pixel execution. These solutions must combine the power efficiency of ASIC style IP while incorporating elements of Instruction-Level Processors flexibility and software ecosystem. This paper introduces Analog Devices BF609's Pipelined Vision Processor (PVP) as a state-of-the-art industrial solution achieving both efficiency and flexibility. The PVP incorporates over 10 function level blocks enabling dozens of programmable functions that can be allocated to implement many algorithms and applications. Additionally, the pipelined style connectivity is programmable enabling many temporal function permutations. Overall, the PVP offers greater than 25 billion operations per second (GOPs) and very low memory bandwidth. These capabilities enable the PVP to execute multiple concurrent ADAS, Industrial, or general vision applications. This paper focuses on the key architecture concepts of the PVP from individual function-block construction to the allocation and chaining of functional blocks to build function based application implementations. The paper also addresses the benefits and challenges of architecting and programming at the function-level granularity and abstractions.