16 min read

(For more resources related to this topic, see here.)

The Wikipedia definition says that, Parallel Computing is a form of computation in which many calculations are carried out simultaneously, operating on the principle that large problems can often be divided into smaller ones, which are then solved concurrently (in parallel).

There are many Parallel Computing programming standards or API specifications, such as OpenMP, OpenMPI, Pthreads, and so on. This book is all about OpenCL Parallel Programming. In this article, we will start with a discussion on different types of parallel programming. We will first introduce you to OpenCL with different OpenCL components. We will also take a look at the various hardware and software vendors of OpenCL and their OpenCL installation steps. Finally, at the end of the article we will see an OpenCL program example SAXPY in detail and its implementation.

Advances in computer architecture

All over the 20th century computer architectures have advanced by multiple folds. The trend is continuing in the 21st century and will remain for a long time to come. Some of these trends in architecture follow Moore’s Law. “Moore’s law is the observation that, over the history of computing hardware, the number of transistors on integrated circuits doubles approximately every two years”. Many devices in the computer industry are linked to Moore’s law, whether they are DSPs, memory devices, or digital cameras. All the hardware advances would be of no use if there weren’t any software advances. Algorithms and software applications grow in complexity, as more and more user interaction comes into play. An algorithm can be highly sequential or it may be parallelized, by using any parallel computing framework. Amdahl’s Law is used to predict the speedup for an algorithm, which can be obtained given n threads. This speedup is dependent on the value of the amount of strictly serial or non-parallelizable code (B). The time T(n) an algorithm takes to finish when being executed on n thread(s) of execution corresponds to:

T(n) = T(1) (B + (1-B)/n)

Therefore the theoretical speedup which can be obtained for a given algorithm is given by :

Speedup(n) = 1/(B + (1-B)/n)

Amdahl’s Law has a limitation, that it does not fully exploit the computing power that becomes available as the number of processing core increase.

Gustafson’s Law takes into account the scaling of the platform by adding more processing elements in the platform. This law assumes that the total amount of work that can be done in parallel, varies linearly with the increase in number of processing elements. Let an algorithm be decomposed into (a+b). The variable a is the serial execution time and variable b is the parallel execution time. Then the corresponding speedup for P parallel elements is given by:

(a + P*b)

Speedup = (a + P*b) / (a + b)

Now defining α as a/(a+b), the sequential execution component, as follows, gives the speedup for P processing elements:

Speedup(P) = P – α *(P – 1)

Given a problem which can be solved using OpenCL, the same problem can also be solved on a different hardware with different capabilities. Gustafson’s law suggests that with more number of computing units, the data set should also increase that is, “fixed work per processor”. Whereas Amdahl’s law suggests the speedup which can be obtained for the existing data set if more computing units are added, that is, “Fixed work for all processors”. Let’s take the following example:

Let the serial component and parallel component of execution be of one unit each.

In Amdahl’s Law the strictly serial component of code is B (equals 0.5). For two processors, the speedup T(2) is given by:

T(2) = 1 / (0.5 + (1 – 0.5) / 2) = 1.33

Similarly for four and eight processors, the speedup is given by:

T(4) = 1.6 and T(8) = 1.77

Adding more processors, for example when n tends to infinity, the speedup obtained at max is only 2. On the other hand in Gustafson’s law, Alpha = 1(1+1) = 0.5 (which is also the serial component of code). The speedup for two processors is given by:

Speedup(2) = 2 – 0.5(2 – 1) = 1.5

Similarly for four and eight processors, the speedup is given by:

Speedup(4) = 2.5 and Speedup(8) = 4.5

The following figure shows the work load scaling factor of Gustafson’s law, when compared to Amdahl’s law with a constant workload:

Comparison of Amdahl’s and Gustafson’s Law

OpenCL is all about parallel programming, and Gustafson’s law very well fits into this book as we will be dealing with OpenCL for data parallel applications. Workloads which are data parallel in nature can easily increase the data set and take advantage of the scalable platforms by adding more compute units. For example, more pixels can be computed as more compute units are added.

Different parallel programming techniques

There are several different forms of parallel computing such as bit-level, instruction level, data, and task parallelism. This book will largely focus on data and task parallelism using heterogeneous devices. We just now coined a term, heterogeneous devices. How do we tackle complex tasks “in parallel” using different types of computer architecture? Why do we need OpenCL when there are many (already defined) open standards for Parallel Computing?

To answer this question, let us discuss the pros and cons of different Parallel computing Framework.


OpenMP is an API that supports multi-platform shared memory multiprocessing programming in C, C++, and Fortran. It is prevalent only on a multi-core computer platform with a shared memory subsystem.

A basic OpenMP example implementation of the OpenMP Parallel directive is as follows:

#pragma omp parallel { body; }

When you build the preceding code using the OpenMP shared library, libgomp would expand to something similar to the following code:

void subfunction (void *data) { use data; body; } setup data; GOMP_parallel_start (subfunction, &data, num_threads); subfunction (&data); GOMP_parallel_end (); void GOMP_parallel_start (void (*fn)(void *), void *data, unsigned num_threads)

The OpenMP directives make things easy for the developer to modify the existing code to exploit the multicore architecture. OpenMP, though being a great parallel programming tool, does not support parallel execution on heterogeneous devices, and the use of a multicore architecture with shared memory subsystem does not make it cost effective.


Message Passing Interface (MPI) has an advantage over OpenMP, that it can run on either the shared or distributed memory architecture. Distributed memory computers are less expensive than large shared memory computers. But it has its own drawback with inherent programming and debugging challenges. One major disadvantage of MPI parallel framework is that the performance is limited by the communication network between the nodes.

Supercomputers have a massive number of processors which are interconnected using a high speed network connection or are in computer clusters, where computer processors are in close proximity to each other. In clusters, there is an expensive and dedicated data bus for data transfers across the computers. MPI is extensively used in most of these compute monsters called supercomputers.


The OpenACC Application Program Interface (API) describes a collection of compiler directives to specify loops and regions of code in standard C, C++, and Fortran to be offloaded from a host CPU to an attached accelerator, providing portability across operating systems, host CPUs, and accelerators. OpenACC is similar to OpenMP in terms of program annotation, but unlike OpenMP which can only be accelerated on CPUs, OpenACC programs can be accelerated on a GPU or on other accelerators also. OpenACC aims to overcome the drawbacks of OpenMP by making parallel programming possible across heterogeneous devices. OpenACC standard describes directives and APIs to accelerate the applications. The ease of programming and the ability to scale the existing codes to use the heterogeneous processor, warrantees a great future for OpenACC programming.


Compute Unified Device Architecture (CUDA) is a parallel computing architecture developed by NVIDIA for graphics processing and GPU (General Purpose GPU) programming. There is a fairly good developer community following for the CUDA software framework. Unlike OpenCL, which is supported on GPUs by many vendors and even on many other devices such as IBM’s Cell B.E. processor or TI’s DSP processor and so on, CUDA is supported only for NVIDIA GPUs. Due to this lack of generalization, and focus on a very specific hardware platform from a single vendor, OpenCL is gaining traction.

CUDA or OpenCL?

CUDA is more proprietary and vendor specific but has its own advantages. It is easier to learn and start writing code in CUDA than in OpenCL, due to its simplicity. Optimization of CUDA is more deterministic across a platform, since less number of platforms are supported from a single vendor only. It has simplified few programming constructs and mechanisms. So for a quick start and if you are sure that you can stick to one device (GPU) from a single vendor that is NVIDIA, CUDA can be a good choice.

OpenCL on the other hand is supported for many hardware from several vendors and those hardware vary extensively even in their basic architecture, which created the requirement of understanding a little complicated concepts before starting OpenCL programming. Also, due to the support of a huge range of hardware, although an OpenCL program is portable, it may lose optimization when ported from one platform to another.

The kernel development where most of the effort goes, is practically identical between the two languages. So, one should not worry about which one to choose. Choose the language which is convenient. But remember your OpenCL application will be vendor agnostic. This book aims at attracting more developers to OpenCL.

There are many libraries which use OpenCL programming for acceleration. Some of them are MAGMA, clAMDBLAS, clAMDFFT, BOLT C++ Template library, and JACKET which accelerate MATLAB on GPUs. Besides this, there are C++ and Java bindings available for OpenCL also.

Once you’ve figured out how to write your important “kernels” it’s trivial to port to either OpenCL or CUDA. A kernel is a computation code which is executed by an array of threads. CUDA also has a vast set of CUDA accelerated libraries, that is, CUBLAS, CUFFT, CUSPARSE, Thrust and so on. But it may not take a long time to port these libraries to OpenCL.


Renderscripts is also an API specification which is targeted for 3D rendering and general purpose compute operations in an Android platform. Android apps can accelerate the performance by using these APIs. It is also a cross-platform solution. When an app is run, the scripts are compiled into a machine code of the device. This device can be a CPU, a GPU, or a DSP. The choice of which device to run it on is made at runtime. If a platform does not have a GPU, the code may fall back to the CPU. Only Android supports this API specification as of now. The execution model in Renderscripts is similar to that of OpenCL.

Hybrid parallel computing model

Parallel programming models have their own advantages and disadvantages. With the advent of many different types of computer architectures, there is a need to use multiple programming models to achieve high performance. For example, one may want to use MPI as the message passing framework, and then at each node level one might want to use, OpenCL, CUDA, OpenMP, or OpenACC.

Besides all the above programming models many compilers such as Intel ICC, GCC, and Open64 provide auto parallelization options, which makes the programmers job easy and exploit the underlying hardware architecture without the need of knowing any parallel computing framework. Compilers are known to be good at providing instruction-level parallelism. But tackling data level or task level auto parallelism has its own limitations and complexities.

Introduction to OpenCL

OpenCL standard was first introduced by Apple, and later on became part of the open standards organization “Khronos Group”. It is a non-profit industry consortium, creating open standards for the authoring, and acceleration of parallel computing, graphics, dynamic media, computer vision and sensor processing on a wide variety of platforms and devices.

The goal of OpenCL is to make certain types of parallel programming easier, and to provide vendor agnostic hardware-accelerated parallel execution of code. OpenCL (Open Computing Language) is the first open, royalty-free standard for general-purpose parallel programming of heterogeneous systems. It provides a uniform programming environment for software developers to write efficient, portable code for high-performance compute servers, desktop computer systems, and handheld devices using a diverse mix of multi-core CPUs, GPUs, and DSPs.

OpenCL gives developers a common set of easy-to-use tools to take advantage of any device with an OpenCL driver (processors, graphics cards, and so on) for the processing of parallel code. By creating an efficient, close-to-the-metal programming interface, OpenCL will form the foundation layer of a parallel computing ecosystem of platform-independent tools, middleware, and applications.

We mentioned vendor agnostic, yes that is what OpenCL is about. The different vendors here can be AMD, Intel, NVIDIA, ARM, TI, and so on. The following diagram shows the different vendors and hardware architectures which use the OpenCL specification to leverage the hardware capabilities:

The heterogeneous system

The OpenCL framework defines a language to write “kernels”. These kernels are functions which are capable of running on different compute devices. OpenCL defines an extended C language for writing compute kernels, and a set of APIs for creating and managing these kernels. The compute kernels are compiled with a runtime compiler, which compiles them on-the-fly during host application execution for the targeted device. This enables the host application to take advantage of all the compute devices in the system with a single set of portable compute kernels.

Based on your interest and hardware availability, you might want to do OpenCL programming with a “host and device” combination of “CPU and CPU” or “CPU and GPU”. Both have their own programming strategy. In CPUs you can run very large kernels as the CPU architecture supports out-of-order instruction level parallelism and have large caches. For the GPU you will be better off writing small kernels for better performance.

Hardware and software vendors

There are various hardware vendors who support OpenCL. Every OpenCL vendor provides OpenCL runtime libraries. These runtimes are capable of running only on their specific hardware architectures. Not only across different vendors, but within a vendor there may be different types of architectures which might need a different approach towards OpenCL programming. Now let’s discuss the various hardware vendors who provide an implementation of OpenCL, to exploit their underlying hardware.

Advanced Micro Devices, Inc. (AMD)

With the launch of AMD A Series APU, one of industry’s first Accelerated Processing Unit (APU), AMD is leading the efforts of integrating both the x86_64 CPU and GPU dies in one chip. It has four cores of CPU processing power, and also a four or five graphics SIMD engine, depending on the silicon part which you wish to buy. The following figure shows the block diagram of AMD APU architecture:

AMD architecture diagram—© 2011, Advanced Micro Devices, Inc.

An AMD GPU consist of a number of Compute Engines (CU) and each CU has 16 ALUs. Further, each ALU is a VLIW4 SIMD processor and it could execute a bundle of four or five independent instructions. Each CU could be issued a group of 64 work items which form the work group (wavefront). AMD Radeon ™ HD 6XXX graphics processors uses this design. The following figure shows the HD 6XXX series Compute unit, which has 16 SIMD engines, each of which has four processing elements:

AMD Radeon HD 6xxx Series SIMD Engine—© 2011, Advanced Micro Devices, Inc.

Starting with the AMD Radeon HD 7XXX series of graphics processors from AMD, there were significant architectural changes. AMD introduced the new Graphics Core Next (GCN) architecture. The following figure shows an GCN compute unit which has 4 SIMD engines and each engine is 16 lanes wide:

GCN Compute Unit—© 2011, Advanced Micro Devices, Inc.

A group of these Compute Units forms an AMD HD 7xxx Graphics Processor. In GCN, each CU includes four separate SIMD units for vector processing. Each of these SIMD units simultaneously execute a single operation across 16 work items, but each can be working on a separate wavefront.

Apart from the APUs, AMD also provides discrete graphics cards. The latest family of graphics card, HD 7XXX, and beyond uses the GCN architecture.


One of NVIDIA GPU architectures is codenamed “Kepler”. GeForce® GTX 680 is one Kepler architectural silicon part. Each Kepler GPU consists of different configurations of Graphics Processing Clusters (GPC) and streaming multiprocessors. The GTX 680 consists of four GPCs and eight SMXs as shown in the following figure:

NVIDIA Kepler architecture—GTX 680, © NVIDIA®

Kepler architecture is part of the GTX 6XX and GTX 7XX family of NVIDIA discrete cards. Prior to Kepler, NVIDIA had Fermi architecture which was part of the GTX 5XX family of discrete and mobile graphic processing units.


Intel’s OpenCL implementation is supported in the Sandy Bridge and Ivy Bridge processor families. Sandy Bridge family architecture is also synonymous with the AMD’s APU. These processor architectures also integrated a GPU into the same silicon as the CPU by Intel. Intel changed the design of the L3 cache, and allowed the graphic cores to get access to the L3, which is also called as the last level cache. It is because of this L3 sharing that the graphics performance is good in Intel. Each of the CPUs including the graphics execution unit is connected via Ring Bus. Also each execution unit is a true parallel scalar processor. Sandy Bridge provides the graphics engine HD 2000, with six Execution Units (EU), and HD 3000 (12 EU), and Ivy Bridge provides HD 2500(six EU) and HD 4000 (16 EU). The following figure shows the Sandy bridge architecture with a ring bus, which acts as an interconnect between the cores and the HD graphics:

Intel Sandy Bridge architecture—© Intel®

ARM Mali™ GPUs

ARM also provides GPUs by the name of Mali Graphics processors. The Mali T6XX series of processors come with two, four, or eight graphics cores. These graphic engines deliver graphics compute capability to entry level smartphones, tablets, and Smart TVs. The below diagram shows the Mali T628 graphics processor.

ARM Mali—T628 graphics processor, © ARM

Mali T628 has eight shader cores or graphic cores. These cores also support Renderscripts APIs besides supporting OpenCL.

Besides the four key competitors, companies such as TI (DSP), Altera (FPGA), and Oracle are providing OpenCL implementations for their respective hardware. We suggest you to get hold of the benchmark performance numbers of the different processor architectures we discussed, and try to compare the performance numbers of each of them. This is an important first step towards comparing different architectures, and in the future you might want to select a particular OpenCL platform based on your application workload.

Subscribe to the weekly Packt Hub newsletter

* indicates required


Please enter your comment!
Please enter your name here