![[PukiWiki]](image/pukiwiki.png "[PukiWiki]")
Session: Auto-Tuning for Multicore and GPU I
Session Chair: Reiji Suda (The University of Tokyo, Japan)
14:35-15:20 (Invited) Jakub Kurzak (University of Tennessee, Knoxville, USA): BEAST: An Automatic Tuner for Numerical Kernels on Accelerators
Abstract: The objective of the BEAST project (http://icl.cs.utk.edu/beast/) is to provide a tool chain for automatic tuning of numerical kernels on hardware accelerators and coprocessors. BEAST strives to deliver the infrastructure to support the three basic steps of the autotuning process: 1) an engine for setting up the search space and the pruning constraints, 2) a harness for parallel benchmarking and collection of performance metrics, and 3) a toolset for analysis of the collected data. BEAST puts strong emphasis on increasing developer insight into why given kernel / device combinations produce the performance profiles they do.
15:25-15:50 Takahiro Katagiri, Satoshi Ohshima and Masaharu Matsumoto (The University of Tokyo, Japan): Auto-tuning of Computation Kernels from an FDM Code with ppOpen-AT
15:50-16:15 Xiong Xiao, Shoichi Hirasawa, Hiroyuki Takizawa and Hiroaki Kobayashi (Tohoku University, Japan): An Approach to Customization of Compiler Directives for Application-Specific Code Transformations
Session: Auto-Tuning for Multicore and GPU II
Session Co-Chairs: Daisuke Takahashi (University of Tsukuba, Japan)
15:45-16:10 Tadayoshi Horita (Polytechnic University, Japan): A GPGPU-based acceleration of fault-tolerant MLP learnings
16:10-16:35 Satoshi Ohshima, Takahiro Katagiri, Masaharu Matsumoto (The University of Tokyo, Japan): Performance optimization of SpMV using CRS format by considering OpenMP scheduling on CPUs and MIC
16:35-17:00 Qiangqiang Shi (Stony Brook University, USA): Application of GPU to Three Computational Models
17:00-17:25 Bo-Cheng Lai, Ta-Kang Yen, Bo-Yao YU (National Chiao Tung University, Taiwan): A Cache Aware Multithreading Decision Scheme on GPGPUs
Visit MCSoC-14 program page http://www.mcsoc-forum.org/ for more details.
