GPU Computing in Python Glib Ivashkevych junior researcher, NSC KIPT 

Parallel revolution The Free Lunch Is Over: A Fundamental Turn Toward Concurrency in Software Herb Sutter, March 2005 

When serial code hits the wall. Power wall. Now, Intel is embarked on a course already adopted by some of its major rivals: obtaining more computing power by stamping multiple processors on a single chip rather than straining to increase the speed of a single processor. Paul S. Otellini, Intel's CEO May 2004 

July 2006 Feb 2007 Nov 2008 Intel launches Core 2 Duo (Conroe) Nvidia releases CUDA SDK Tsubame, first GPU accelerated supercomputer Dec 2008 OpenCL 1.0 specification released Today 50 GPU powered supercomputers in Top500 

It's very clear, that we are close to the tipping point. If we're not at a tipping point, we're racing at it. Jen-Hsun Huang, NVIDIA Co-founder and CEO March 2013 Heterogeneous computing becomes a standard in HPC and programming has changed 

Heterogeneous computing CPU main memory GPU cores GPU memory multiprocessors Host Device 

CPU GPU general purpose sophisticated design and scheduling perfect for task parallelism highly parallel huge memory bandwidth lightweight scheduling perfect for data parallelism 

Anatomy of GPU: multiprocessors GPU MP shared memory GPU is composed of tens of multiprocessors (streaming processors), which are composed of tens of cores = hundreds of cores 

Compute Unified Device Architecture is a hierarchy of computation memory synchronization 

Compute hierarchy software kernel hardware abstractions hardware thread thread block grid of blocks core multiprocessor GPU 

Compute hierarchy thread threadIdx thread block blockIdx, blockDim grid of blocks gridDim 

Python fast development huge # of packages: for data analysis, linear algebra, special functions etc metaprogramming Convenient, but not that fast in number crunching 

PyCUDA Wrapper package around CUDA API Convenient abstractions: GPUArray, random numbers generation, reductions & scans etc Automatic cleanup, initialization and error checking, kernels caching Completeness 

GPUArray NumPy-like interface for GPU arrays Convenient creation and manipulation routines Elementwise operations Cleanup 

SourceModule Abstraction to create, compile and run GPU code GPU code to compile is passed as a string Control over nvcc compiler options Convenient interface to get kernels 

Metaprogramming GPU code can be created at runtime PyCUDA uses mako template engine internally Any template engine is ok to create GPU source code. Remember about codepy Create more flexible and optimized code 

Installation numpy, mako, CUDA driver & toolkit are required Boost.Python is optional Dev packages: if you build from source Also: PyOpenCl, pyfft 

GPU computing resources Documentation Intro to Parallel Programming by David Luebke (Nvidia) and John Owens (UC Davis) Heterogeneous Parallel Programming by Wen-mei W. Hwu (UIUC) Several excellent books