AbstractsComputer Science

Optimizing SIMD execution in HW/SW co-designed processors

by Rakesh Kumar

Institution: Universitat Polit├Ęcnica de Catalunya
Year: 2014
Record ID: 1128509
Full text PDF: http://hdl.handle.net/10803/277375


SIMD accelerators are ubiquitous in microprocessors from different computing domains. Their high compute power and hardware simplicity improve overall performance in an energy efficient manner. Moreover, their replicated functional units and simple control mechanism make them amenable to scaling to higher vector lengths. However, code generation for these accelerators has been a challenge from the days of their inception. Compilers generate vector code conservatively to ensure correctness. As a result they lose significant vectorization opportunities and fail to extract maximum benefits out of SIMD accelerators. This thesis proposes to vectorize the program binary at runtime in a speculative manner, in addition to the compile time static vectorization. There are different environments that support runtime profiling and optimization support required for dynamic vectorization, one of most prominent ones being: 1) Dynamic Binary Translators and Optimizers (DBTO) and 2) Hardware/Software (HW/SW) Co-designed Processors. HW/SW co-designed environment provides several advantages over DBTOs like transparent incorporations of new hardware features, binary compatibility, etc. Therefore, we use HW/SW co-designed environment to assess the potential of speculative dynamic vectorization. Furthermore, we analyze vector code generation for wider vector units and find out that even though SIMD accelerators are amenable to scaling from the hardware point of view, vector code generation at higher vector length is even more challenging. The two major factors impeding vectorization for wider SIMD units are: 1) Reduced dynamic instruction stream coverage for vectorization and 2) Large number of permutation instructions. To solve the first problem we propose Variable Length Vectorization that iteratively vectorizes for multiple vector lengths to improve dynamic instruction stream coverage. Secondly, to reduce the number of permutation instructions we propose Selective Writing that selectively writes to different parts of a vector register and avoids permutations. Finally, we tackle the problem of leakage energy in SIMD accelerators. Since SIMD accelerators consume significant amount of real estate on the chip, they become the principle source of leakage if not utilized judiciously. Power gating is one of the most widely used techniques to reduce leakage energy of functional units. However, power gating has its own energy and performance overhead associated with it. We propose to selectively devectorize the vector code when higher SIMD lanes are used intermittently. This selective devectorization keeps the higher SIMD lanes idle and power gated for maximum duration. Therefore, resulting in overall leakage energy reduction.