Font Size: a A A

The Design And Implementation Of A Mechanism For Branch Handling In SIMD

Posted on:2016-01-06Degree:MasterType:Thesis
Country:ChinaCandidate:F ZhouFull Text:PDF
GTID:2348330536467255Subject:Computer technology
Abstract/Summary:PDF Full Text Request
A variety of multi-media applications contain multi-branches,however it is of low efficiency to map them into traditional SIMD structures controlled by one unified instruction stream,which leads to the problem of insufficient exploitation of data level parallelism of applications.Considering the above matter,we proposed an instruction dynamic mapping mechanism for traditional SIMD structures to achieve the purpose of multiple branches executing in SIMD lanes in a parallel way.The main idea of the method is to simultaneously dispatch different instruction streams from all branch paths to desired lanes of SIMD to obtain the effect of lanes in different branch paths function parallelly.In this research subject,I have conducted theoretical research and engineering practice from the following aspects: Firstly,after analyzed the shortage of branch statements in the applications executing serially on the traditional SIMD structure,I proposed the instruction dynamic mapping mechanism for SIMD structure to extend its parallel branch handling capability.And then I designed and implemented the microarchitecture of the scheme based on SIMD structure in FT-X DSP in order to improve the chip's execution efficiency of branch and to explore the largest potential data level parallel in the applications.Secondly,after a comprehensive consideration of the architecture of FT-X DSP,I integrated the microarchitecture of instruction dynamic mapping mechanism into the chip.Finally,I completed the system level functional verification and performance evaluation for the design.Profiling results shows that the proposed method consumes less area overhead and has better scalability and even obtains significant performance increase compared with traditional SIMD structure whose lanes receive the unified instruction stream but execute conditionally through a lane mask vector.Thus,it is of high efficiency to handle multi-branches contained in applications in SIMD structures.
Keywords/Search Tags:SIMD structure, multi-branches, dynamic instruction mapping
PDF Full Text Request
Related items