Read more
Ralf Karrenberg presents Whole-Function Vectorization (WFV), an approach that allows a compiler to automatically create code that exploits data-parallelism using SIMD instructions. Data-parallel applications such as particle simulations, stock option price estimation or video decoding require the same computations to be performed on huge amounts of data. Without WFV, one processor core executes a single instance of a data-parallel function. WFV transforms the function to execute multiple instances at once using SIMD instructions. The author describes an advanced WFV algorithm that includes a variety of analyses and code generation techniques. He shows that this approach improves the performance of the generated code in a variety of use cases.
List of contents
Introduction.- Foundations & Terminology.- Overview.- Related Work.- SIMD Property Analyses.- Whole-Function Vectorization.- Dynamic Code Variants.- Evaluation.- Conclusion.- Outlook.
About the author
Ralf Karrenberg received his PhD in computer science at Saarland University in 2015. His seminal research on compilation techniques for SIMD architectures found wide recognition in both academia and the CPU and GPU industry. Currently, he is working for NVIDIA in Berlin. Prior to that, he contributed to research and development for visual effects in blockbuster movies at Weta Digital, New Zealand.
Summary
Ralf Karrenberg presents Whole-Function Vectorization (WFV), an approach that allows a compiler to automatically create code that exploits data-parallelism using SIMD instructions. Data-parallel applications such as particle simulations, stock option price estimation or video decoding require the same computations to be performed on huge amounts of data. Without WFV, one processor core executes a single instance of a data-parallel function. WFV transforms the function to execute multiple instances at once using SIMD instructions. The author describes an advanced WFV algorithm that includes a variety of analyses and code generation techniques. He shows that this approach improves the performance of the generated code in a variety of use cases.
Additional text
“This dissertation investigates whole function vectorization, which is an automatic procedure to optimize intermediate scalar compiler code for SIMD (single-instruction multiple-date) architectures. … The thesis is well written and easily understandable by anyone with at least some background in compilation. Examples are generously provided to illustrate the major notions and pseudo-code is presented for all major procedures.” (Andreas Maletti, Mathematical Reviews, March, 2016)
Report
"This dissertation investigates whole function vectorization, which is an automatic procedure to optimize intermediate scalar compiler code for SIMD (single-instruction multiple-date) architectures. ... The thesis is well written and easily understandable by anyone with at least some background in compilation. Examples are generously provided to illustrate the major notions and pseudo-code is presented for all major procedures." (Andreas Maletti, Mathematical Reviews, March, 2016)
