dslsAndSearchForLinearAlgebraPerformanceOptimization
➤ Gửi thông báo lỗi ⚠️ Báo cáo tài liệu vi phạmNội dung chi tiết: dslsAndSearchForLinearAlgebraPerformanceOptimization
dslsAndSearchForLinearAlgebraPerformanceOptimization
DSLs and Search for Linear Algebra PerfronianceOptimizationbyThomas NelsonB.s. University of Texas, 2007M.S. University of Colorado, 2010A thesis subm dslsAndSearchForLinearAlgebraPerformanceOptimizationmitted to theFaculty of the Graduate School of theUniversity of Colorado in partial fulfillment of the requirements for the degree of Doctor of PhilosophyDepartment of Computuer Science2015This thesis entitled:DSLs and Search for Linear Algebra Perfromance Optimization written by Thomas Nelsonhas be dslsAndSearchForLinearAlgebraPerformanceOptimizationen approved for the Department of Computuer ScienceProf. Elizabeth JessupProf. Jeremy SiekProf. Xiao-Chnan V.Prof. Jed BrownProf. Pavol CernyDateThe fdslsAndSearchForLinearAlgebraPerformanceOptimization
inal copy of this thesis has l>oen examined by the signatories, and we find that both the content, and the form meet acceptable presentation standardsDSLs and Search for Linear Algebra PerfronianceOptimizationbyThomas NelsonB.s. University of Texas, 2007M.S. University of Colorado, 2010A thesis subm dslsAndSearchForLinearAlgebraPerformanceOptimizationebra Pcrfromancc OptimizationThesis directed by Prof. Elizabeth JessupLinear algebradslsAndSearchForLinearAlgebraPerformanceOptimization
rder Bl.AS (B I O), a language for matrix compulation. Scientific programmers often turn to vendor-tuned Basic Linear Algebra Subprograms (Bl.AS) to oDSLs and Search for Linear Algebra PerfronianceOptimizationbyThomas NelsonB.s. University of Texas, 2007M.S. University of Colorado, 2010A thesis subm dslsAndSearchForLinearAlgebraPerformanceOptimizationoptimal performance. The entire sequence needs to be optimized in concert. Instead of vendor-tuned BLAS, a programmer could start with source code in Fortran or c mid use a state-of-the-art optimizing compiler. However, experiments in this thesis show that optimizing compilers often attain only one- dslsAndSearchForLinearAlgebraPerformanceOptimizationquarter of the performance of hand -optimized code. 1 present. a scalable search algorithm for BTO that reliably achieves high performance by choosingdslsAndSearchForLinearAlgebraPerformanceOptimization
the best, combination of loop fusion, array contraction, ami multithreading for data parallelism.The second DSL is t he Optimizing Compiler with TensDSLs and Search for Linear Algebra PerfronianceOptimizationbyThomas NelsonB.s. University of Texas, 2007M.S. University of Colorado, 2010A thesis subm dslsAndSearchForLinearAlgebraPerformanceOptimizationreuse. The best-performing implementation is heavily dependent on the tensor dimensionality and the target architecture. This optimization problem is especially challenging when the computation requires many iterations with tensors of small dimensions. I created a high level search representation an dslsAndSearchForLinearAlgebraPerformanceOptimizationd input language as part of a tookhain to solve these problems. In this thesis. OCTOPI maps such tensor computations to GPUs, starting with a high levdslsAndSearchForLinearAlgebraPerformanceOptimization
el tensor input language and producing efficient CUD A code as output. This approach combines tensor-specific mathematical transformations with a GPU DSLs and Search for Linear Algebra PerfronianceOptimizationbyThomas NelsonB.s. University of Texas, 2007M.S. University of Colorado, 2010A thesis subm dslsAndSearchForLinearAlgebraPerformanceOptimizationode, and a comparison with OpenACC shows the importance of autotuning and otheroptimizations for achieving efficient results.DedicationTo Katie. Thank yon for sharing this adventure with me.VAcknowledgementsFirst I want to thank Katie, for canying me through graduate school with her love and support dslsAndSearchForLinearAlgebraPerformanceOptimization. She kept me on task, helped every way she knew how, and kept me smiling. I also want to thank my parents for encouraging me throughout school, my brdslsAndSearchForLinearAlgebraPerformanceOptimization
other Joe for listening to me ramble on about programming and the meaning of life, ami my whole family for believing that 1 could do this.1 want to thDSLs and Search for Linear Algebra PerfronianceOptimizationbyThomas NelsonB.s. University of Texas, 2007M.S. University of Colorado, 2010A thesis subm dslsAndSearchForLinearAlgebraPerformanceOptimization. Paul Hovland, Prasanna Balaprakesh, and Mary Hall, for teaching me how to do research and giving valuable feedlMwk and insight throughout the PhD process. Thank yon to Xiao-Chuan Cai, Jed Brown, ami Pavol Cerny, forlaiingon my thesis committee and reading this document. dslsAndSearchForLinearAlgebraPerformanceOptimizationDSLs and Search for Linear Algebra PerfronianceOptimizationbyThomas NelsonB.s. University of Texas, 2007M.S. University of Colorado, 2010A thesis submGọi ngay
Chat zalo
Facebook