Follow Slashdot stories on Twitter


Forgot your password?
Check out the new SourceForge HTML5 internet speed test! No Flash necessary and runs on all devices. ×

Book Review: OpenCL Programming Guide 40

asgard4 writes "In recent years GPUs have become powerful computing devices whose power is not only used to generate pretty graphics on screen but also to perform heavy computation jobs that were exclusively reserved for high performance super computers in the past. Considering the vast diversity and rapid development cycle of GPUs from different vendors, it is not surprising that the ecosystem of programming environments has flourished fairly quickly as well, with multiple vendors, such as NVIDIA, AMD, and Microsoft, all coming up with their own solutions on how to program GPUs for more general purpose computing (also abbreviated GPGPU) applications. With OpenCL (short for Open Computing Language) the Khronos Group provides an industry standard for programming heavily parallel, heterogeneous systems with a language to write so-called kernels in a C-like language. The OpenCL Programming Guide gives you all the necessary knowledge to get started developing high-performing, parallel applications for such systems with OpenCL 1.1." Keep reading for the rest of asgard4's review.
OpenCL Programming Guide
author Aaftab Munshi, Benedict R. Gaster, Timothy G. Mattson, James Fung, Dan Ginsbur
pages 603
publisher Addison-Wesley Pearson Educatio
rating 9/10
reviewer asgard4
ISBN 0321749642
summary A solid introduction to programming with OpenCL.
The authors of the book certainly know what they are talking about. Most of them have been involved in the standardization effort that went into OpenCL. Munshi, for example, is the editor of the OpenCL specification. So all the information in the book is first-hand knowledge from experts in OpenCL. The reader is expected to be familiar with the C programming language and basic programming concepts. Some experience in parallelizing problems is a benefit but not a requirement.

The book consist of two major parts. The first part is a detailed description of the OpenCL C language and the API used by the host to control the execution of programs written in that language. The second part is comprised of various case studies that show OpenCL in action.
The authors get straight to the point in the introduction, discussing the conceptual foundations of OpenCL in detail. They explain what kernels are (basically functions that are scheduled for execution on a compute device), how the kernel execution model works, how the host manages the command queues that schedule memory transfers or kernel execution on compute devices, and the memory model.

While this first chapter is all prose, the second chapter dives right in with some code and a first HelloWorld example. The following chapters introduce more and more of the OpenCL language and API step-by-step. All API functions are described in somewhat of a reference style with a lot of detail, including possible error codes. However, the text is not a reference. There is always a good explanation with examples or short code listings, the only notable exception being chapter three, which presents the OpenCL C language. A few more examples would have made the text less dry in this chapter.

An important chapter is chapter nine on events and synchronization between multiple compute devices and the host. This chapter is important because — as any experienced parallel programmer knows — getting synchronization right is often tricky but obviously essential for correct execution of a parallel program.

An interesting feature in OpenCL is the built-in interoperability with OpenGL and, surprisingly, Direct3D. Various functions in the OpenCL API allow creating buffers from OpenGL/Direct3D objects, such as textures or vertex buffers, that can be used by an OpenCL kernel. This opens up interesting possibilities for doing a lot more work on the GPU in graphics applications, such as running a fluid simulation on the GPU in OpenCL, which directly writes its results into vertex buffers or textures to be used directly for rendering without the host CPU having to intervene.

Before delving into the case studies the book briefly discusses the embedded profile that is available for OpenCL and the standardized C++ API that the Khronos Group provides in addition to the regular OpenCL API (which is defined exclusively as C functions). The C++ API makes using some of the OpenCL objects a little bit easier and somewhat nicer.

The second part of the book contains various interesting case studies that show off what OpenCL can be used for, such as computing a sobel filter or a histogram for an image, computing FFTs, doing cloth simulation, or multiplying dense and sparse matrices. The choice and variety of case studies is definitely interesting and most will be immediately applicable to the reader when going forward developing applications using OpenCL. All the code for the examples and the case studies in the book are available for download on the book's website.

Overall, the OpenCL Programming Guide succeeds in being a great introduction to OpenCL 1.1. The book covers all of the specification and more, has an easy to read writing style and yet provides all the necessary details to be an all-encompassing guide to OpenCL. The good selection of case studies makes the book even more appealing and demonstrates what can be done with real-life OpenCL code (and also how it needs to be optimized to get the best performance out of current OpenCL platforms, such as GPUs).

Martin Ecker has been involved in real-time graphics programming for more than 15 years and works as a professional game developer for Sony Computer Entertainment America in sunny San Diego, California.

You can purchase OpenCL Programming Guide from Slashdot welcomes readers' book reviews -- to see your own review here, read the book review guidelines, then visit the submission page.


This discussion has been archived. No new comments can be posted.

Book Review: OpenCL Programming Guide

Comments Filter:
  • Great (Score:5, Informative)

    by WilyCoder ( 736280 ) on Friday January 20, 2012 @06:07PM (#38768930)

    I read this book back in August. I've been using OpenGL for almost 10 years now but knew little to nothing about OpenCL.

    This book was really good. There were some typos that I found while reading it (other people had already found and reported them). If you get this book make sure you visit the author's addendum & corrections page.

    I agree with the review, 9/10. If there were NO typos at all, it would be 10/10 for me.

  • This is neither the book nor the review I would have written.

    My book would have started:

    WTF is the Khronos Group? Good question. It sure sounds like one of those faux "we really do talk to each other while going our own separate ways" PR initiatives of the African UNIX warlord alliance of so many bland bodies from ages ago whose names we can no longer recall.

    Circling threat []

    My caption: With the new hipwader hull, Joe had the whole OpenCL stack right at his fingertips.

  • by bored ( 40072 ) on Friday January 20, 2012 @08:38PM (#38770764)

    Is that its not really useful for learning OpenCL. Sure it will teach you the syntax and how to write an OpenCL program. That isn't the problem. The problem is that if your writing something in OpenCL you probably want it to be fast. Learning the language is doable by someone with C experience in just a couple hours with just the SDKs shipped by AMD/Nividia/Intel. Learning how to optimize a routine for a particular GPU/etc is the hard part, and is application specific. It also requires knowledge of how compute device actually work at an extremely low level. I don't believe this book teaches that. Save your money, download the spec and a SDK for your device. Start reading the architecture docs..

    • You mean learning how to actually program, where it's the algorithms that make the difference, is difficult? I thought I could buy this book (I have) and then figure out how to break the NSA's latest encryption standards on my iPhone :(

  • ANOTHER damn thing called a "kernel". Cause that wasn't overloaded enough yet.
  • ...Could I utilise this programming method to say, encode video streams to a common format, in an efficient (ie fully utilising available GPU/CPU cores) manner? Because right now I have a compute cluster comprising a pair of dual core laptops, one of which has an AMD Radeon HD GPU on-die, the other an Intel chipset GPU (but that's not really important), two P4 desktop machines with NVidia GF7 GPUs and a Sempron box with AMD Radeon HD pci-express. Altogether, that's 7 processor cores and 4 (possibly usable b

Kill Ugly Radio - Frank Zappa