Date of Award

Spring 2016

Document Type


Degree Name

Doctor of Philosophy in Electrical Engineering - (Ph.D.)


Electrical and Computer Engineering

First Advisor

Sotirios Ziavras

Second Advisor

Durgamadhab Misra

Third Advisor

Edwin Hou

Fourth Advisor

Roberto Rojas-Cessa

Fifth Advisor

Alexandros V. Gerbessiotis


Taking advantage of DLP (Data-Level Parallelism) is indispensable in most data streaming and multimedia applications. Several architectures have been proposed to improve both the performance and energy consumption for such applications. Superscalar and VLIW (Very Long Instruction Word) processors, along with SIMD (Single-Instruction Multiple-Data) and vector processor (VP) accelerators, are among the available options for designers to accomplish their desired requirements. On the other hand, these choices turn out to be large resource and energy consumers, while also not being always used efficiently due to data dependencies among instructions and limited portion of vectorizable code in single applications that deploy them. This dissertation proposes an innovative architecture for a multithreaded VP which separates the path for performing data shuffle and memory-indexed accesses from the data path for executing other vector instructions that access the memory. This separation speeds up the most common memory access operations by avoiding extra delays and unnecessary stalls. In this multilane-based VP design, each vector lane uses its own private memory to avoid any stalls during memory access instructions. More importantly, the proposed VP has an innovative multithreaded architecture which makes it highly suitable for concurrent sharing in multicore environments. To this end, the VP which is developed in VHDL and prototyped on an FPGA (Field-Programmable Gate Array), serves as a coprocessor for one or more scalar cores in various system architectures presented in the dissertation.

In the first system architecture, the VP is allocated exclusively to a single scalar core. Benchmarking shows that the VP can achieve very high performance. The inclusion of distributed data shuffle engines across vector lanes has a spectacular impact on the execution time, primarily for applications like FFT (Fast-Fourier Transform) that require large amounts of data shuffling.

In the second system architecture, a VP virtualization technique is presented which, when applied, enables the multithreaded VP to simultaneously execute many threads of various vector lengths. The threads compete simultaneously for the VP resources having as a goal an improved aggregate VP utilization. This approach yields high VP utilization even under low utilization for the individual threads. A vector register file (VRF) virtualization technique dynamically allocates physical vector registers to running threads. The technique is implemented for a multi-core processor embedded in an FPGA. Under the dynamic creation of threads, benchmarking demonstrates large VP speedups and drastic energy savings when compared to the first system architecture.

In the last system architecture, further improvements focus on VP virtualization relying exclusively on hardware. Moreover, a pipelined data shuffle network replaces the non-pipelined shuffle engines. The VP can then take advantage of identical instruction flows that may be present in different vector applications by running in a fused instruction mode that increases its utilization. A power dissipation model is introduced as well as two optimization policies towards minimizing the consumed energy, or the product of the energy and runtime for a given application. Benchmarking shows the positive impact of these optimizations.