Chris Rowen, Fellow at Cadence, presents the "Designing and Selecting Instruction Sets for Vision" tutorial at the May 2015 Embedded Vision Summit.
Two critical technical trends have reached important inflection points: the massive compute demands of vision processing and the capabilities of specialized vision processors. But what how do you actually select (or even build) a better vision processor?
This talk serves as a step-by-step tutorial on how to dissect a suite of applications, extract the key computational and memory requirements, and determine priorities for processor instruction set and memory organization for optimal vision processing. Chris walks through large scale application profiling, does data type analysis, outlines selection of operations, examines tradeoffs in SIMD, VLIW and specialized execution units, and drills down on resource constraints that impact power, energy, cost, memory system demands and programming flexibility.
He takes a set of real-world applications from ADAS, deep neural network image recognition, and multi-frame imaging as targets and illustrates the process for assessing and selecting vision processors.