Vector-thread architecture and implementation

Abstract

This thesis proposes vector-thread architectures as a performance-efficient solution for all-purpose computing. The VT architectural paradigm unifies the vector and multithreaded compute models. VT provides the programmer with a control processor and a vector of virtual processors. The control processor can use vector-fetch commands to broadcast instructions to all the VPs or each VP can use thread-fetches to direct its own control flow. A seamless intermixing of the vector and threaded control mechanisms allows a VT architecture to flexibly and compactly encode application parallelism and locality. VT architectures can efficiently exploit a wide variety of loop-level parallelism, including non-vectorizable loops with cross-iteration dependencies or internal control flow. The Scale VT architecture is an instantiation of the vector-thread paradigm designed for low-power and high-performance embedded systems. Scale includes a scalar RISC control processor and a four-lane vector-thread unit that can execute 16 operations per cycle and supports up to 128 simultaneously active virtual processor threads. Scale provides unit-stride and strided-segment vector loads and stores, and it implements cache refill/access decoupling. The Scale memory system includes a four-port, non-blocking, 32-way set-associative, 32 KB cache. A prototype Scale VT processor was implemented in 180 nm technology using an ASIC-style design flow. The chip has 7.1 million transistors and a core area of 16.6 mm2, and it runs at 260 MHz while consuming 0.4-1.1 W. This thesis evaluates Scale using a diverse selection of embedded benchmarks, including example kernels for image processing, audio processing, text and data processing, cryptography, network processing, and wireless communication.

(cont.) Larger applications also include a JPEG image encoder and an IEEE 802.11 la wireless transmitter. Scale achieves high performance on a range of different types of codes, generally executing 3-11 compute operations per cycle. Unlike other architectures which improve performance at the expense of increased energy consumption, Scale is generally even more energy efficient than a scalar RISC processor.