PPT-Fluidic Kernels: Cooperative Execution of

OpenCL Programs on Multiple Heterogeneous Devices Prasanna Pandit R Govindarajan Supercomputer Education and Research Centre Indian Institute of Science Bangalore CGO2014 Outline Introduction. ij DD ij ij ij ij ij DD DD DD brPage 4br DD ii TATA 4 ij ij TABTAB ij DDD 44 var var AA DD brPage 5br AA DH DH AA 222 TTT GAAA VVV 222 2 4 TTT GAAA DH VVV C S AA CovFS AA Cov brPage 6br DH SC DH DH AA SC AA DH AA DH DH AA brPage 7br DH DH DH Maxims and Implicatures Cooperative Principle - CP • Grice ( 1975 ) proposes that participants in a conversation obey a general ‘Cooperative Principle’ (CP), which is expected t COOPERATIVE GAMESThe arena is a circle about 30 feet in diameter with aclearly de A community journey. Population 710, 2010 census. Non-consolidated grade school and high school. 3 churches. Convenience store. 2 banks. Gosper County seat. Gas station. Restaurant . New library - $350,000 DED grant. Peter Somerville. University of Lincoln. 7. th. May 2014. The case for reform. University governance is increasingly undemocratic and . managerialised. (Deem et al, 2008; Woods, 2011) – a problem of internal power. Shai B. Elbaz and Amir D. Gat. Technion - Israel . Institute of . Technology . Faculty of Mechanical Engineering. 1. 12/01/2013. Background (1/2) – Soft Robotics. Emerging . field . of experimental soft robotics. . Also known as “meta-data”. April 2016. Comments in SPICE Kernels. 2. Comments, also called “meta-data,” are information that describe the context of kernel data, i.e. “data about data”. Comments are provided inside kernels as plain text (prose). April 2016. Porting Kernels. 2. Porting Issues - 1. Data formats vary across platforms, so data files created on platform “X” may not be usable on platform “Y.”. Binary. . formats. : different platforms use different bit patterns to represent numbers (and possibly characters).. Ross McIlroy, Chris Hawblitzel, Galen Hunt. Microsoft Research. Helios: Heterogeneous Multiprocessing with Satellite Kernels. 1. Problem: HW now heterogeneous. Heterogeneity . ignored. by operating systems. . Efficient Intra-SM Slicing through Dynamic Resource Partitioning for GPU Multiprogramming. Qiumin. . Xu*. , . Hyeran. . Jeon. ✝. , . Keunsoo. Kim. ❖. , Won Woo Ro. ❖. . Murali. . Annavaram. Specialization of Kernelization. Daniel Lokshtanov. We. Kernels. ∃. ¬. . Kernels. Why?. What’s Wrong with Kernels. (from a practitioners point of view). Only handles . NP-hard. problems.. Don’t combine well with . Perceptrons. The . perceptron. A. B. instance. . x. i. Compute: . y. i. = . sign(. v. k. . . . x. i. . ). ^. y. i. ^. y. i. If mistake: . v. k+1. = . v. k. + . y. i. . x. i. . x . is a vector. Pipeline . Parallelism from . Multiple . Dependent Kernels for . GPUs. Gwangsun Kim, . Jiyun Jeong, John Kim. Mark Stephenson. GPU Background. CTA. Kernel grid. CTA (Cooperative Thread Array). or Thread block. Ke Wang. Sparse Correspondence Problems. Dense Correspondence Problems. Stereo. Motion. Motion vs. Stereo: Differences. Motion: . Uses velocity: consecutive frames must be close to get good approximate time derivative.