A Hybrid Hardware/Software Architecture That Combines a 4-wide Very Long Instruction Word Software Processor (VLIW) with Application-specific Super-complex Instruction Set Hardware Functions
Dara Kusic · D-Scholarship@Pitt (University of Pittsburgh) · 2005
Application-driven processor designs are becoming increasingly feasible. Today, advances in field-programmable gate array (FPGA) technology are opening the doors to fast and highly-feasible hardware/software co-designed architectures. Over 100,000 FPGA logic array blocks and nearly 100 ASIC multiply-accumulate cores combine with extensible CPU cores to foster the design of configurable, application-driven hybrid processors.This thesis proposes a hardware/software co-designed architecture targeted to an FPGA. The architecture is a very-long instruction-word (VLIW) processor coupled with super-complex instruction set (SuperCISC) hardware co-processors. Results of the VLIW/SuperCISC show performance speedups over a single-issue processor of 9x to 332x, and entire application speedups from 4x to 127x. Contributions of this research include a 4-way VLIW designed from the ground up, a zero-overhead implementation of a hardware/software interface, evaluation of the scalability of shared data stores, examples of application-specific hardware accelerants, a SystemC simulator, and an evaluation of shared memory configurations.