General articles on technical subjects.

OpenCL Fireworks

Posted by Vincent Hindriksen on

I like and appreciate differences in the many cultures on our Earth, but also like to recognise different very old traditions everywhere to feel a sort of ancient bond. As an European citizen I’m quite familiar with the replacement of the weekly flowers with a complete tree, each December – and the burning of al those trees in January. Also celebration of New Year falls on different dates, the Chinese new year being the best known (3 February 2011). We – internet-using humans – all know the power of nicely coloured gunpowder: fireworks!

Let’s try to explain the workings of OpenCL in terms of fireworks. The following data is not realistic, but gives a good idea on how it works.

Continue reading “OpenCL Fireworks” →

http://www.flickr.com/photos/imabug/2946930401/

OpenCL Potentials: Medical Imaging

Posted by Vincent Hindriksen on with 2 Comments
Photo by Eugene Mah

When you ever saw a CT or MRI scanner, you might have noticed the full-sized computer next to it (especially the older ones). There is quite some processing power needed to keep up with the data-stream coming from the scanner, to process the data to a 3D-image and to visualize the data on a 2D-screen. Luckily we have OpenCL to make these kinds of medical imaging even faster. Which doctor doesn’t want real-time high-resolution results and which patient doesn’t want to see the results on Apple iPad or Samsung Galaxy Tab?

Architects, bankers and doctors have one thing in common: they get a better feeling for the current subject if they can play with the data. OpenCL makes it possible to process data much faster and thus let the specialist play with it. The interesting part of IT is that it is in every domain now and therefore a new series: OpenCL-potentials.

Continue reading “OpenCL Potentials: Medical Imaging” →

OpenCL on the CPU: AVX and SSE

Posted by Vincent Hindriksen on with 2 Comments

When AMD came out with CPU-support I was the last one who was enthusiastic about it, comparing it as feeding chicken-food to oxen. Now CUDA has CPU-support too, so what was I missing?

This article is a quick overview on OpenCL on CPU-extensions, but expect more to come when the Hybrid X86-Processors actually hit the market. Besides ARM also IBM already has them; also more about their POWER-architecture in an upcoming article to give them the attention they deserve.

CPU extensions

SSE/MMX started in the 90’s extending the IBM-compatible X86-instruction, being able to do an add and a multiplication in one clock-tick. I still remember the discussion in my student-flat that the MP3s I could produce in only 4 minutes on my 166MHz PC just had to be of worse quality than the ones which were encoded in 15 minutes. No, the encoder I “found” on the internet made use of SSE-capabilities. Currently we have reached SSE5 (by AMD) and Intel introduced a new extension called AVX. That’s a lot of abbreviations! MMX stands for “MultiMedia Extension”, SSE for “Streaming SIMD Extensions” with SIMD being “Single Instruction Multiple Data” and AVX for “Advanced Vector Extension”. This sounds actually very interesting, since we saw SIMD and Vectors op the GPU too. Let’s go into SSE (1 to 4) and AVX – both fully supported on the new CPUs by AMD and Intel.

Continue reading “OpenCL on the CPU: AVX and SSE” →

Thalesians talk – OpenCL in financial computations

Posted by Vincent Hindriksen on with 1 Comment

End of October I had a talk for the Thalesians, a group that organises different kind of talks for people working or interested in the financial market. If you live in London, I would certainly recommend you visit one of their talks. But from a personal perspective I had a difficult task: how to make a very diverse public happy? The talks I gave in the past were for a more homogeneous and known public, and now I did not know at all what the level of OpenCL-programming was of the attendants. I chose to give an overview and reserve time for questions.

After starting with some honest remarks about my understanding of the British accent and that I will kill my business for being honest with them, I spoke about 5 subjects. Some of them you might have read here, but not all. You can download the sheets [PDF] via this link: Vincent.Hindriksen.20101027-Thalesians. The below text is to make the sheets more clear, but certainly is not the complete talk. So if you have the feeling I skipped a lot of text, your feeling is right.

Continue reading “Thalesians talk – OpenCL in financial computations” →

Engineering GPGPU into existing software

Posted by Vincent Hindriksen on with 1 Comment

At the Thalesian talk about OpenCL I gave in London it was quite hard to find a way to talk about OpenCL for a very diverse public (without falling back to listing code-samples for 50 minutes); some knew just everything about HPC and other only heard of CUDA and/or OpenCL. One of the subjects I chose to talk about was how to integrate OpenCL (or GPGPU in common) into existing software. The reason is that we all have built nice, cute little programs which were super-fast, but it’s another story when it must be integrated in some enterprise-level software.

Readiness

The most important step is making your software ready. Software engineering can be very hectic; managing this in a nice matter (i.e. PRINCE2) just doesn’t fit in a deadline-mined schedule. We all know it costs less time and money when looking at the total picture, but time is just against.

Let’s exaggerate. New ideas, new updates of algorithms, new tactics and methods arrive on the wrong moment, Murphy-wise. It has to be done yesterday, so testing is only allowed when the code will be in the production-code too. Programmers just have to understand the cost of delay, but luckily is coming to the rescue and says: “It is my responsibility”. And after a year of stress your software is the best in the company and gets labelled as “platform”; meaning that your software is chosen to include all small ideas and scripts your colleagues have come up “which are almost the same as your software does, only little different”. This will turn the platform into something unmanageable. That is a different kind of software-acceptance!

Continue reading “Engineering GPGPU into existing software” →

New grown-ups on the block

Posted by Vincent Hindriksen on with 2 Comments

Members of the band There is one big reason StreamHPC chose for OpenCL and that is (future) hardware-support. I talked about NVIDIA versus AMD a lot, but knowing others would join soon. AMD is correct when they say the future is fusion: hybrid computing with a single chip holding both CPU- and GPU-cores, sharing the same memory and interconnected at high speed. Merging the technologies would also give possible much higher bandwidths to memory for the CPU. Let us see in short which products from experienced companies will appear on the OpenCL-stage.

Continue reading “New grown-ups on the block” →

OpenCL – the battle, part III

Posted by Vincent Hindriksen on with 2 Comments

The first two parts described hardware-companies and operating systems, programming languages and software-companies, written about half a year ago. Now we focus on what has driven NVIDIA and ATI/AMD for decades: games.

Disclaimer: this is an opinion-piece on the current market. We are strong supporters of OpenCL and all companies which support it too. Since our advise on specific hardware in a consult will be based on specific demands on the customer, we could advise differently than would be expected on the below article.

Games

Computer games are cool; merely because you choose from so many different kinds. While Tetris will live forever, the latest games also have something to add: realistic physics simulation. And that’s what’s done by GPUs now. Nintendo has shown us that gameplay and good interaction are far more important than video-quality. The wow-factor for photo-realistic real-time rendering is not as it was years ago.
You might know the basics for falling objects: F = m*g (Force = Mass times Gravity-acceleration), and action = – reaction. If you drop some boxes, you can predict falling speed, interaction, rotation and possible change of centre of gravity from a still image as a human being. A computer has to do a lot more to detect collision, but the idea is very doable on a fast CPU. A very well-known open source library for these purposes is Bullet Physics. The nice thing comes, when there is more than just a few boxes, but thousands of them. Or when you walk through water or under a waterfall, see fire and smoke, break wood but bend metal, etc. The accelerometer of the iPod was a game-changer too in the demand for more realism in graphics. For an example of a “physics puzzle game” not using GPGPU see World of Goo (with free demo) – for the rest we talk more about high-end games. Of current game-ready systems PCs (Apple, Linux and Windows) have OpenCL support, Sony PlayStation 3 is now somewhat vague and the Xbox 360 has none.

The picture is from Crysis 3, which does not use OpenCL, as we know it.

Continue reading “OpenCL – the battle, part III” →

OpenCL in the Clouds

Posted by Vincent Hindriksen on with 4 Comments

Buzz-words are cool; they are loosely defined and are actually formed by the many implementation that use the label. Like Web 2.0 which is cool javascript for the one and interaction for the other. Now we have cloud-computing, which is cluster-computing with “something extra”. More than a year ago clouds were in the data-centre, but now we even have “private clouds”. So how to incorporate GPGPU? A cluster with native nodes to run our OpenCL-code with pre-distributed data is pretty hard to maintain, so what are the other solutions?

Distributed computing

Folding@home now has support for OpenCL to add the power of non-NVIDIA GPUs. While in clusters the server commands the clients what they have to do, here the clients ask the server for jobs. Disadvantage is that the clients are written for a specific job and are not really flexible to take different kind of jobs. There are several solutions for this code-distribution-problem, but still the solution is not suitable for smaller problems and small clusters.

Clusters: MPI

The project SHOC (Scalable HeterOgeneous Computing) is a collection of benchmark programs testing the performance and stability of systems using computing devices with non-traditional architectures for general purpose computing, and the software used to program them. While it is only a benchmark, it can be of great use when designing a cluster. For the rest I only  found CUDA MPI-solutions, which are not ported to OpenCL yet.

Also check out Hoopoe, which is a cloud-computing service to run your OpenCL-kernels in their cloud. It seems to be more limited to .NET and have better support for CUDA, but it is a start. In Europe there is a start-up offering a rent-model for OpenCL-computation-time; please contact us if you want to get in contact with them.

Clusters: OpenMP

MOSIX has added a “Many GPU Package” to their cluster management system, so it now allows applications to transparently use cluster-wide OpenCL devices. When “choosing devices” not only the local GPU pops up, but also all GPUs in the cluster.
It works disk-less, in the way no files are copied to the computation-clients and all stays in-memory. Disk-less computations have an advantage when cloud-computer are not fully trusted. Take note that on most cloud-computers the devices need to be virtualised (see next part).

Below is its layered model, VCL being the “Virtual OpenCL Layer”.

They have chosen to base it on OpenMP; while the kernels don’t need to be altered, some OpenMP-code needs to be added. They are very happy to tell it takes much less code to use openMP instead of MPI.

You see a speed-up between 2.19 and 3.29 on 4 nodes is possible. We see comparable cluster-speed-ups in an old cluster-study. The actual speed-up on clusters depends mostly on the amount of data that needs to be transferred.

The project references to a project called remote CUDA, which only works with NVIDIA-GPUs.

Device Virtualisation

Currently there is no good device virtualisation for OpenCL. The gVirtuS-project currently only supports CUDA, but they claim it is easily rewritten to OpenCL. Code needs to be downloaded with a Mercurius-client (comparable to GIT and in repositories of most Linux-distributions):
> hg clone http://osl.uniparthenope.it/hg/projects/gvirtus/gvirtus gvirtus
Or download it here (dated 7-Oct-2010).

Let me know when you ported it to OpenCL! Actually gVirtuS does not do the whole trick since you need to divide the host-devices between the different guest-OSes, but luckily there is an extension which provides sharing of devices, called fission. More about this later.

We can all agree there still needs to be done a lot in this area of virtualised devices to get OpenCL in the cloud. If you can’t wait, you can theoretically use MOSIX locally.

Afterword

A cloud is the best buzz-word to market a scalable solution to overcome limitations of internet connected personal devices. I personally think the biggest growth will be in personal clouds, so companies will have their own in-house cloud-server (read: clusters); people just want to have a feeling of control, comparable with preference of a daily traffic jam above public transport. But nevertheless shared clouds have potential if it comes to computation-intensive jobs which do not need to be done all year round.

The projects presented here are a start to have OpenCL-power at a larger scale for more demanding cases. Since we can have more power at our fingertips with one desktop-pc stuffed with high-end video-cards than a 4-year-old supercomputer-cluster, there is still time

Please send your comment if I missed a project or method.

Learning both OpenCL and CUDA

Posted by Vincent Hindriksen on with 2 Comments

Be sure to read Taking on OpenCL where I’ve put my latest insights – also for CUDA.

The two¹ “camps” OpenCL and CUDA both claim you should first learn their language first, after which the other would be easy to learn. I’m from the OpenCL-camp, so I say you should learn OpenCL first, but with a strong emphasis on hardware-architecture understanding. If I had chosen for CUDA I would have said the opposite, so in other words it does not matter which you do first. But psychology tells us that you probably like the first language more since there is where you discovered the magic; also most people do not like to learn a second language which is much alike and does not add a real difference. Most programmers just want to get the job done and both camps know that. Be aware of that.

NVIDIA is very good in marketing their products, AMD has – to say it modest – a lower budget for GPGPU-marketing. As a programmer you should be aware of this difference.

The possibilities of OpenCL are larger than those of CUDA, because of task-parallel programming and support for far more different architectures. At the other side CUDA is much more user-friendly and has a lot of convenience built-in.

Continue reading “Learning both OpenCL and CUDA” →

Waiting for Mobile OpenCL

Posted by Vincent Hindriksen on with 3 Comments

People who follow me, know my interest in ARM Cortex CPU & Mali GPU and Imagination Technology’s PowerVR, regarding OpenCL-potential. Here is an overview of what I found out until now and has more open ends than answers. I was very hopeful about access to mobile OpenCL for developers… But now I think it will have to wait until 2011. It seems that the mobile phone makers keep the power to themselves in the form of system-libraries instead of giving direct access. Actually a wise choice, given the fact that a bad kernel could easily crash the phone.

Imagination Technology’s PowerVR SGX product provides the GPU-power for the iPhone 4 and various new Android phones. The company does not provide an OpenCL SDK directly to end-users, but leaves the responsibility to the implementers of their IP. Strangely enough they offer SDKs for OpenGL, and say “no comment” to a pretty direct forum-question. In other words: we don’t know what to expect.

Samsung has the only official implementation for ARM (ARM Cortex-A9 MPCore CPU) for OpenCL. They also just released their home-made Linux-based mobile phone OS “Bada”, but there is not a sign of OpenCL in their API. Samsung being the only one who could open an API to developers on their phones officially, does not make me smile yet.

There are many, many voices that Apple iOS 4 has support for OpenCL, but then only in the system-libraries. Given the fact that Apple is a big fan of OpenCL, we can assume this is true. See the web for a lot of articles about this.

ZiiLabs is open about their support for OpenCL in their ZMS-05 processor, and has an early access program. Early access still means waiting.

Qualcomm’s Snapdragon does not mention OpenCL at all, while it powers the more powerful smartphones. It mentions a recent job-post, so you know the drill: wait.

Android has good support for OpenGL ES, but not officially for OpenCL. You might have heard of the particle experiment for Android, which is actually OpenGL-based. It also mentions the Android-version of the bullet engine for physics-computations, but also no OpenCL. It doesn’t look like Android “Gingerbread” will have support.

So what did you learn after reading this? That developers still won’t have access to OpenCL-API on a mobile platform, and that you have to wait until next year or enter ZiiLabs’ early access program. More about the good and bad of hiding OpenCL on this blog.

Khronos OpenCL presentation at SIGGRAPH 2010

Posted by Vincent Hindriksen on with 1 Comment

Here you find the videos uploaded by Khronos of their presentation about OpenCL. I added the time-line, so you can scroll to the more interesting parts easily. The presentation by Ofer Rosenberg of Intel and Cliff Woolly of NVIDIA were not uploaded (yet). Please note that for non-American people the speech of Affi Munchie is hard to hear; luckily his sheets explain most.

For the first two presentations the sheets can be downloaded from the Khronos-website. The time-line has the sheet-numbers mentioned.

0:00 [sheet 1] Presentation by the president of Khronos and chair of the session: Neill Trevett of NVIDIA.
0:06 [sheet 2] Welcome and a quick overview
1:12 [sheet 3] The prizes for the attendees (not us, online viewers)
1:40 [4] Overview of all members of Khronos. Khronos does not only take care of OpenCL but also the more famous OpenGL and projects like Collada.
2:26 [5] Processor Parallelism. CPUs are getting more parallel and GPUs more programmable. The overlapping area is called Heteregenous Computing and there is where OpenCL pops up.
3:10 [6] OpenCL timeline from version 1.0 to 1.1.
4:44 [7] OpenCL workinggroup with only 30 logos. He mentions missing logos like the one from Apple.
5:18 [8] The Visual Computing Ecosystem, where OpenCL interoperability with other standards are shown. The talk is not complete, so I don;t know if he talks about DirectX.

Continue reading “Khronos OpenCL presentation at SIGGRAPH 2010” →

To think about: an OpenCL Wizard

Posted by Vincent Hindriksen on

A part of reaction on my earlier post was: “VB Programmers span the whole range from hobbyists to scientists and professions. There is no reason that VB programmers should be left out of the loop in GPU programming. (…) CUDA and OpenCL are fundamentally lower level languages that do not have the same level of user-friendliness that VB gives, and so they require a bit more care to program. (…)“. I left out some parts, so I probably got the emphasis very wrong. However, it gave me an idea: what would a “Visual Basic Wizard for OpenCL” look like?

It is actually very interesting to think how to get GPGPU-power to a programmer who’s used to “and then and then and then”. How do you get parallel processing into the mind of an iterative thinker? Simplification is not easy!

By thinking about an OpenCL-wizard, you start to understand the bottlenecks of, and need for OpenCL initialization.

Actually it could better be built into an existing framework like OpenMP, or something alike in Python, Java, etc. Or the IDE could give hints that the standard-function could be replaced by a GPGPU-version. But we just want a wizard which generates “My First OpenCL Program”, which puts a smile on the face of programmers who use their mouse a lot.

Continue reading “To think about: an OpenCL Wizard” →

The rise of the GPGPU-compilers

Posted by Vincent Hindriksen on with 6 Comments

Painting "High Rise" made by Huma Mulji
Painting “High Rise” made by Huma Mulji

If you read The Disasters of Visual Designer Tools you’ll find a common thought about easy programming: many programmers don’t learn the hard-to-grasp backgrounds any more, but twiddle around and click a program together. In the early days of BASIC, you could add Assembly-code to speed up calculations; you only needed tot understand registers, cache and other details of the CPU. The people who did that and learnt about hardware, can actually be considered better programmers than the C++ programmer who completely relies on the compiler’s intelligence. So never be happy if the control is taken out of your hands, because it only speeds up the easy parts. An important side-note is that recompiling easy readable code with a future compiler might give faster code than your optimised well-engineered code; it’s a careful trade-off.

Okay, let’s be honest: OpenCL is not easy fun. It is more a kind of readable Assembly than click-and-play programming. But, oh boy, you learn a lot from it! You learn architectures, capabilities of GPUs, special purpose processors and much more. As blogged before, OpenCL probably is going to be the hidden power for non-CPUs wrapped in something like OpenMP.

Continue reading “The rise of the GPGPU-compilers” →

OpenCL 1.1 changes compared to 1.0

Posted by Vincent Hindriksen on with 1 Comment

This blog-entry is of interest for you, if you don’t want to read the whole new specifications [PDF] for the OpenCL 1.1 changes, but just want an overview of the most important changes differences with 1.0.

The news-release sums up the changes for 1.1 like this:

  1. New datatypes including 3-component vectors and additional formats
  2. Handling command from multiple hosts and processing buffers across multiple devices
  3. Operations on regions of a buffer including read, write and copy of 1D, 2D and 3D rectangular regions
  4. Enhanced use of events to drive and control command execution
  5. Additional OpenCL C built-in functions such as integer clamp, shuffle and asynchronous strided copies
  6. Improved OpenGL interoperability through efficient sharing of images and buffers by linking OpenCL and OpenGL events.

Furthermore we can read the update is completely backwards-compatible with version 1.0. The obvious macro‘s CL_VERSION_1_0 and CL_VERSION_1_1 are added to handle the versioning, but what’s more? This blog-post discusses most changes and with some subjective opinions added to it.

Additional Formats

3-component vectors

We only had 2-, 4-, 8 or 16-component vectors, but not 3 which actually was somewhat strange. In OpenCL 1.1, the functions vload3, vload_half3, vloada_half3 and vstore3, vstore_half3, vstorea_half3 have been added to the family. Watch out, that for the half-functions the offset is calculated somewhat different compared to the even-sized vectors. In version 1.0 you could have chosen for a 4-component vector when using a lot of calculations, or a struct. If you see the new function vec_step below, it seems that it is not more memory-efficient to use this vector instead of a 4-component vector.

RGB with Padding

We have support CL_RGB, CL_RGBa (= RGB with an alpha-channel) and now also RGBx (with padding-channel). The same variants are there for CL_R and CL_RG. Good for graphics-programmers, or for easier reading of 32 bpp BMPs.

Cloud Computing / Multi-user Environments

The support for different hosts gives possibilities for cloud-computing. Side-note: cloud-computing is another word for multi-user environments, with some promotion for big data-centres. All API-functions except clSetKernelArg are thread-safe now; but only when kernels are not shared between hosts; see appendix A.2 for more information. The important part is that you think clearly about how to design your software if you now can assume others can take your resources now too. OpenCL already needed a lot of planning when claiming and releasing resources, so you’re probably already mastering it; now just check more often how much resources are available.

Region-specific Operations

Regions make it possible to split a big buffers to form a queue without having to keeping track of dimensions and offsets during operations, run from host. See clCreateSubBuffer for more information. A real convenience, but watch out when writing to overlapping buffers. The functions clEnqueueCopyBufferRec, clEnqueueReadBufferRect en clEnqueueWriteBufferRect helps synchronising commands to copy, read to or write from a region.

Enhanced Control-room

My favourite description of the host is “the control-room”, since you are not over there on the device but in Houston. The more control and information, the better. The new events are clSetMemObjectDestructorCallback, clCreateUserEvent, clSetUserEventStatus and clSetEventCallback. The first event-listener lets you know when resources a freed, so you can keep track. User-events can be put in the event_wait_list in various functions just like the built-in events; the function will start when all events are CL_COMPLETE. With clSetEventCallback immediate actions-on-events can be programmed; combined with the user-events the programmer got some powerful tools. See the example at clSetUserEventStatus for how to use the user-events.

OpenGL events and Direct3D support

The function clCreateEventFromGLsyncKHR links a CL-event to a GL-event by just giving the name of the OpenGL-event. See gl_sharing for more info.

OpenCL has now support for Direct3D 10, which is great! This might also be a good step to make DirectCompute lighter. See cl_khr_d3d10_sharing for more info. Welcome DirectX-developers! One favour: please be aware that DirectX works on Windows only, not on Apple OSX or iOS, (Embedded) Linux or Symbian. If you use clean calls, it will be more easy to port to other platforms.

Other New Kernel-functions

The following new functions were added to the kernel:

  • get_global_offset: returns the offset of the enqueued kernels.
  • minmag and maxmag: returns the argument with the minimum or maximum distance to zero, falls back to fmin and fmax if distance is equal or an argument is NaN. Example: maxmag(-5, 3) = -5, minmag(-3, 3) = -3.
  • clamp: returns boundary-values if the given number is not between the boundaries.
  • vec_step: returns the number of elements in a scalar or a vector. A scalar returns 1, a vector 2, 4, 8 or 16. If the size is 3, the function returns 4.
  • shuffle and shuffle2: shuffles one or two vectors given another vector with the indices of the new order. Indeed plain old permutations.
  • async_workgroup_strided_copy: buffers between global and local memory on the device. When used correctly, this can overcome some of the hassle when you need to work on global memory objects, but need more speed. Correct usage is described in the reference.

The functions min and max now also work component-wise with a vector as first argument and a scalar as second. Min({2, 4, 6, 8}, 5) will give {2, 4, 5, 5}.

Conclusion

While the many revisions of OpenCL 1.0 were really minor and not a lot attention was paid to them, 1.1 is a big step forward. If you see what has been done to multi-user environments, NVidia and AMD have a lot of work to do with their drivers.

You can read in revision 33 there has been some heated discussion and there was pressure on the decision:

>>Should this extension be KHR or EXT?
 PROPOSED: KHR. If this extension is to be approved by Khronos then it should be KHR, otherwise EXT. Not all platforms can support this extension, but that is also true of OpenGL interop.
RESOLVED: KHR.<<

The part “not all platforms” is very politically written down, since exactly one platform supports this specific extension. I have seen too many of these pressured discussions and I hope Khronos is stronger than i.e. ISO and OpenCL will remain as open as OpenGL.

I’m very happy with the new version, since there is more control with loads of extra events, now multiple hosts are possible, and the forgotten 3-component vector was added. Now let me know in the comments what you think of the new version.

By the way, not all is new. Deprecated are clSetCommandQueueProperty and the __ROUNDING_MODE__ macro.

Let’s enter the Top500 HPC list using GPUs

Posted by Vincent Hindriksen on with 1 Comment

The #500 super-computer has only 24 TFlops (2010-06-06): http://www.top500.org/system/9677

update: scroll down to see the best configuration I have found. In other words: a cluster with at least 30 nodes with 4 high-end GPUs each (costing almost €2000,- per node and giving roughly 5 TFlops single precision, 1 TFLOPS double precision) would enter the Top500. 25 nodes to get to a theoretic 25TFlops and 5 extra for overcoming the overhead. So for about €60 000,- of hardware anyone can be on the list (and add at least €13 000 if you want to use Windows instead of Linux for some reason). Ok, you pay most for the services and actual building when buying such a cluster, but you get the idea it does not cost you a few millions any more. I’m curious: who is building these kind of clusters? Could you tell me the specs (theoretical TFlops, LinPack TFlops and watts/TFlop) of your (theoretical) cluster, which costs the customer less then €100 000,- in total? Or do you know companies who can do this? I’ll make a list of companies who will be building the clusters of tomorrow, the “Top €100.000,- HPC cluster list”. You can mail me via vincent [at] this domain, or put your answer in a comment.

Update: the hardware shopping-list

Nobody told in the remarks it is easy to build a faster machine than the one described above. So I’ll do it. We want the most flops per box, so here’s the wishlist:

  • A motherboard with as many slots as possible for PCI-E, CPU-sockets and memory-banks. This because the lag between the nodes is high.
  • A CPU with at least 4 cores.
  • Focus on the bandwidth, else we will not be able to use all power.
  • Focus on price per GFLOPS.

The following is what I found in local computer stores (which for some reason people there love to talk about extreme machines). AMD currently has the graphics cards with the most double precision power, so I chose for their products. I’m looking around for Intel + Nvidia, but currently they are far behind. Is AMD back on stage after being beaten by Intel’s Core-products for so many years?

The GigaByte GA-890FXA-UD7 (€245,-) has 1 AM3-socket, 6(!) PCI-e slots and supports up to 16GB of memory. We want some power, so we use the AMD Phenom II X6 1090T (€289,-), which I chose for the 6 cores and the low price per FLOPS. And to make it a monster, we add 6 times a AMD HD5970 (€599,-) giving 928 x 6 = 3264 DP-GLOPS. If it can handle 16GB DDR3 (€750,-), so we put it in. It needs about 3 Power-supplies of 700 Watt (€100,-). We add 128GB SSD (€350,-) for working data and a big 2 TB HDD (€100,-). Case needs to house the 3 power supplies (€100,-). Cooling is important and I suggest you compete with a wind-tunnel (€500,-). It will cost you €6228,- for 5,6 Double Precision TFLOPS, and 27 TFLOPS single precision. A cluster would be on the HPC500-list for around €38000,- (pure hardware-price, not taking network-devices too much into account, nor the price for man-hours).

Disclaimer: this is the price of a single node, excluding services, maintenance, software-installation, networking, engineering, etc. Please note that the above price is pure for building a single node for yourself, if you have the knowledge to do so.

Nokia Maemo and OpenCL

Posted by Vincent Hindriksen on with 2 Comments

Update 21-06-2011: Bumped into a project by Nokia: CLEP, “OpenCL Embedded Profile” for the N900.

Maemo is the Debian based Linux-distribution of Nokia for embedded devices. It is on the gadget N900, so you can be root on your own phone and compile your own kernel. In other words: a great developer’s phone.

Which smartphone to buy when you want to toy around with OpenCL “Embedded Profile”? There is more and more evidence that the next iPhone OS will have support for OpenCL, as should be expected Apple being the trademark-owner of OpenCL. This is good, since the mobile market could make the difference for the technique – competing with CUDA and DirectCompute. “The other ARM Cortex-A8 smartphone”, the Nokia N900 does not support it, while the magic of OpenCL attracts to many developers on the Maemo-forums.

The QT-blog that disclosed coming OpenCL-support for QT, spoke about it too:

>>Right now, QtOpenCL works very well with desktop OpenCL implementations, like that from NVIDIA (we’ve tested it under Linux, Mac, and Windows). Embedded devices are currently another matter – OpenCL implementations are still very basic in that space.  The performance improvements on embedded CPU’s are only slightly better than using ARM/NEON instructions for example.  And embedded GPU’s are usually hard-wired for GLSL/ES, lacking many of the features that makes OpenCL really sing.  But like everything in the embedded space, things are likely to change very quickly. By releasing QtOpenCL, hopefully we can stimulate the embedded vendors to accelerate development by giving them something to test with. Be the first embedded device on the block to get the mandelbrot demo running at 10fps, or 20fps, or 60fps!<<

But checking the whole Nokia QT/Maemo-SDK for something like “opencl.h” or words like “opencl” and “khronos” in .h-files did not return anything interesting. The missing reference in the SDK tells me, we cannot expect any OpenCL-implementation on the N900 soon. So do we have to wait for the Nokia N920, Maemo 6 and QT 4.8? Once I know more, by getting deeper into the SDK, you’re the first to know. But first let me show you the documents which tells us OpenCL is coming to the Maemo-platform.

The Maemo Base Port Document, version 1.1

Exhibit number 1. The introduction tells us that the document describes what hardware-designers should do to get Maemo working on their device:

>>When Maemo is ported to a new chipset and HW environment, the majority of the SW worktakes place in the base layer. However, some adjustments may also be needed in the otherlayers. The porting work as a whole is a combined effort by the chipset vendor and Nokia. Thisdocument describes the deliverables expected from the chipset vendor in such an effort. The requirements in this document are expressed in the form of SW component, interface andfunctional requirements. Note that in many cases more detailed discussions are neededbetween Nokia and the chipset vendor to reach a common understanding about the specificsof the system architecture and the required component versions, functionality and interfaces.<<

So the document describes what the hardware must support, to be able to run Maemo. Let’s then find the magic word “OpenCL”:

>>Graphics Adaptation. The Base Port graphics adaptation interfaces consist of X11, OpenGL ES, and OpenVG interfaces. The OpenCL interface is also included in this group since it typically is used to access the GPU for general-purpose parallel computation.<<

And somewhat below:

>>OpenCL 1. The Base Port should provide an implementation of the OpenCL 1.0 interface for general-purpose parallel programming of heterogeneous systems, especially for the use of GPUs for computation (Khronos group standard).<<

That seems to be pretty clear that Maemo-devices must be able to support OpenCL.

http://www.forum.nokia.com/piazza/wiki/images/7/7d/Maemo_Base_Port_v1.1.pdf

Paper “OpenCL on Embedded devices” by Nokia

Exhibit 2 shows tests of a few simple OpenCL-program on an unnamed device with a TI OMAP 3430 (550 MHz ARM Cortex-A8 CPU & 110 MHz POWERVR SGX530 GPU) – which happens to be in the Motorola Droid, Palm Pre, and Nokia N900. So they managed to create a OpenCL-implementation on ARM. If you’re interested in OpenCL for embedded devices, please do read this presentation:

http://www.khronos.org/developers/library/2009-hotchips/Nokia_OpenCL-in-Handheld-Devices.pdf

It is a document from august 2009, which shows they actually were trying POWERVR and OpenCL then. Now with QT and Maemo mentioning it, we can be very sure the N900 or the N920 is eventually going to have OpenCL-support.

Qt Hello World

Posted by Vincent Hindriksen on with 1 Comment

The earlier blog-post was about how to use Qt Creator with OpenCL. The examples are all around Images, but nowhere a simple Hello World. So here it is: AMD’s infamous OpenCL Hello World in Qt. Thank’s to linuxjunk for glueing the parts together.



int main(int argc, char *argv[]) {

    // Define the kernel. Take a good look what it does.
    QByteArray prog(
    "#pragma OPENCL EXTENSION cl_khr_byte_addressable_store : enablen" 
    "__constant char hw[] = "Hello World from Qt!"; n" 
    "__kernel void hello(__global char * out) {n" 
    "  size_t tid = get_global_id(0); n" 
    "  out[tid] = hw[tid]; n" 
    "}n"
    );

     // Get a context on the first available GPU.
     QCLContext context;
     if (!context.create(QCLDevice::GPU))
         qFatal("Could not create OpenCL context");

     // Allocate 100 bytes of memory on the Host.
     size_t mem_size = 100;
     char* outH = new char[mem_size];
     // Allocate buffer on the Device.
     QCLBuffer outCL = context.createBufferHost(outH, sizeof(char) * mem_size,
                                                QCLMemoryObject::WriteOnly);

     // Compile program against device
      QCLProgram program = context.buildProgramFromSourceCode(prog);

     // Create a kernel object, tell it we are using the kernel called "hello".
     QCLKernel kernel = program.createKernel("hello");

     // Set the necessary global memory. In this case it is the buffer-size.
     kernel.setGlobalWorkSize(outCL.size(), 1);

     // Turn on profiling.
     QCLCommandQueue queue = context.commandQueue();
     queue.setProfilingEnabled(true);

     // Queue the kernel up to run.
     // Give it an argument which is the memory we allocated above.
     QCLEvent event = kernel(outCL);

     // Use the event object above to block until processing has completed.
     event.waitForFinished();

     // Timing only works with profiling on. runTime is unsigned.
     printf(" time '%u'n", event.runTime());

     // Read the results out of the shared memory area.
     outCL.read(outH, mem_size);

     // Write to screen.
     printf(" result = '%s'", outH);
}

Have fun!

Using Qt Creator for OpenCL

Posted by Vincent Hindriksen on with 2 Comments

More and more ways are getting available to bring easy OpenCL to you. Most of the convenience libraries are wrappers for other languages, so it seems that C and C++ programmers have the hardest time. Since a while my favourite way to go is Qt: it is multi-platform, has a good IDE, is very extensive, has good multi-core and OpenGL-support and… has an extension for OpenCL: http://labs.trolltech.com/blogs/2010/04/07/using-opencl-with-qt http://blog.qt.digia.com/blog/2010/04/07/using-opencl-with-qt/

Other multi-platform choices are Anjuta, CodeLite, Netbeans and Eclipse. I will discuss them later, but wanted to give Qt an advantage because it also simplifies your OpenCL-development. While it is great for learning OpenCL-concepts, please know that the the commercial version of Qt Creator costs at least €2995,- a year. I must also warn the plugin is still in beta.

streamhpc.com is not affiliated with Qt.

Getting it all

Qt Creator is available in most Linux-repositories: install packages ‘qtcreator’ and ‘qt4-qmake’. For Windows, MAC and the other Linux-distributions there are installers available: http://qt.nokia.com/downloads. People who are not familiar with Qt, really should take a look around on http://qt.nokia.com/.

You can get the source for the plugin QtOpenCL, by using GIT:

git clone http://git.gitorious.org/qt-labs/opencl.git QtOpenCL

See http://qt.gitorious.org/qt-labs/opencl for more information about the status of the project.

You can download it here: https://dl.dropbox.com/u/1118267/QtOpenCL_20110117.zip (version 17 January 2011)

Building the plugin

For Linux and MAC you need to have the ‘build-essentials’. For Windows it might be a lot harder, since you need make, gcc and a lot of other build-tools which are not easily packaged for the Windows-OS. If you’ve made a win32-binary and/or a Windows-specific how-to, let me know.

You might have seen that people have problems building the plugin. The trick is to use the options -qmake and -I (capital i) with the configure-script:

./configure -qmake <location of qmake 4.6 or higher> -I<location of directory CL with OpenCL-headers>

make

Notice the spaces. The program qmake is provided by Qt (package ‘qt4-qmake’), the OpenCL-headers by the SDK of ATI or NVidia (you’ll need the SDK anyway), or by Khronos. By example, on my laptop (NVIDIA, Ubuntu 32bit, with Qt 4.7):

./configure -qmake /usr/bin/qmake-qt4 -I/opt/NVIDIA_GPU_Computing_SDK_3.2/OpenCL/common/inc/

make

This should work. On MAC the directory is not CL, but OpenCL – I haven’t tested it if Qt took that into account.

After building , test it by setting a environment-setting “LD_LIBRARY_PATH” to the lib-directory in the plugin, and run the provided example-app ‘clinfo’. By example, on Linux:

export LD_LIBRARY_PATH=`pwd`/lib:$LD_LIBRARY_PATH

cd util/clinfo/

./clinfo

This should give you information about your OpenCL-setup. If you need further help, please go to the Qt forums.

Configuring Qt Creator

Now it’s time to make a new project with support for OpenCL. This has to be done in two steps.

First make a project and edit the .pro-file by adding the following:

LIBS     += -L<location of opencl-plugin>/lib -L<location of OpenCL-SDK libraries> -lOpenCL -lQtOpenCL

INCLUDEPATH += <location of opencl-plugin>/lib/

<location of OpenCL-SDK include-files>

<location of opencl-plugin>/src/opencl/

By example:

LIBS     += -L/opt/qt-opencl/lib -L/usr/local/cuda/lib -lOpenCL -lQtOpenCL

INCLUDEPATH += /opt/qt-opencl/lib/

/usr/local/cuda/include/

/opt/qt-opencl/src/opencl/

The following screenshot shows how it could look like:

Second we edit (or add) the LD_LIBRARY_PATH in the project-settings (click on ‘Projects’ as seen in screenshot):

/usr/lib/qtcreator:location of opencl-plugin>:<location of OpenCL-SDK libraries>:

By example:

/usr/lib/qtcreator:/opt/qt-opencl/lib:/usr/local/cuda/lib:

As you see, we now also need to have the Qt-creator-libraries and SDK-libraries included.

The following screenshot shows the edit-field for the project-environment:

Testing your setup

Just add something from the clinfo-source to your project:

printf("OpenCL Platforms:n"); 
QList platforms = QCLPlatform::platforms();
foreach (QCLPlatform platform, platforms) { 
   printf("    Platform ID       : %ldn", long(platform.platformId())); 
   printf("    Profile           : %sn", platform.profile().toLatin1().constData()); 
   printf("    Version           : %sn", platform.version().toLatin1().constData()); 
   printf("    Name              : %sn", platform.name().toLatin1().constData()); 
   printf("    Vendor            : %sn", platform.vendor().toLatin1().constData()); 
   printf("    Extension Suffix  : %sn", platform.extensionSuffix().toLatin1().constData());  
   printf("    Extensions        :n");
} QStringList extns = platform.extensions(); 
foreach (QString ext, extns) printf("        %sn", ext.toLatin1().constData()); printf("n");

If it gives errors during programming (underlined includes, etc), focus on INCLUDEPATH in the project-file. If it complaints when building the application, focus on LIBS. If it complaints when running the successfully built application, focus on LD_LIBRARY_PATH.

Ok, it is maybe not that easy to get it running, but I promise it gets easier after this. Check out our Hello World, the provided examples and http://doc.qt.nokia.com/opencl-snapshot/ to start building.

X86 Systems-on-a-Chip and GPGPU

Posted by Vincent Hindriksen on with 3 Comments

The System-on-a-chip (SoC) for X86 will be a revolution for GPGPU. Why? Because currently a big problem is transferring data from CPU-memory to GPU-memory and back, which will be solved with SoCs. Below you can read this architecture-target is very possible.

With AMD+ATI, Intel and its future high-end GPUs, and NVidia with the rumours around its X86-chips, we will certainly get changes in the field. If it is the way to go, what is probable?

  1. Get both CPU and high-end GPU on 1 chip, separated memory
  2. Techniques for sharing memory
  3. Translating OpenCL from and to C on the fly

ARM-processors are combined with GPUs a lot of times, but they don’t have current support for a common shader-languages (read: OpenCL) to make GPGPU in reach. We’ve asked ourselves many times why ARM & friends are involved in OpenCL since the beginning, but still don’t have any public and promoted driver-support. More on ARM, once there is more news on multi-core ARM-CPUs or OpenCL drivers.

1: One chip for everything

The biggest problem with split CPU/GPU-functionality is the bus-speed between the two is limited. The higher this speed, the more useful GPGPU can be. The highest speeds are possible when the signal does not have to leave the chip and there are no concessions made to the architecture of the graphics-card, in other words: glueing CPU and GPU together, but leave the memory-buses the same.

Currently there is Intel’s Nehalem and AMD’s Fusion, but they use DDR3 for both GPU and CPU; this will not really unlock the GPGPU-possibilities of high-end GPUs. It seems these products were designed with lower costs in mind.

But the chances high-end GPUs will be integrated on the CPU is rising. Going to 32nm gives room for more functionality, such as GPUs. Other choices can be smaller chips, more cores and integrating functionality of the north/south-bridge of the motherboard. If GPU-cores can be turned off when not working optimally when being tested in the factory (just like they do with mult-core CPUs), integrating high-end GPU-cores will even become a save choice.

Another way it could go is using optical buses between the GPU and CPU. It’s unknown if it will really see mainstream markets soon enough.

2: Shared memory – new style

Some levels of cache and all memory should be easy accessible by both types of cores. Why? Because eventually you want to switch between CPU- and GPU-instructions continuously. CUDA has a nice feature already, which keeps objects synchronised between CPU and GPU; one step further is leaving out the need of synchronising.

The problem is that video-memory is accessed more parallel to provide higher data-speeds (GDDR5), so we don’t want to limit the GPU by attaching them to slower (=lower bandwidth) DDR3. Doing it the other way would then be the best solution: giving CPUs direct access to GDDR. There is always a probable option that a new type of (replaceable) memory will be used, which has a dual-bus by design.

The hard part is memory-protection; since now more devices get control to memory, the overhead of controlling/arranging the spots can increase enormously and might need a separate core for it – just like the Cell-processor. This need-for-control is a reason I don’t expect access to each other memory before there will be a fast bus between GPU and CPU, since then the access to GDDR via the GPU’s memory-manager will be much faster and maybe even fast enough.

3: Grown up software

If software would be able to easily select devices and use the same code for each device, then we’ve made a giant step forwards. Software has always been one step behind hardware; so when you do not develop such techniques, you just have to wait a while.

Translating OpenCL into normal C and back will be possible in all kinds of ways, once there is more acceptance of (and thus demand for) GPGPU. AMD’s OpenCL-implementation for CPUs is also a way to merge the fields of CPU and GPU. It’s hard to tell how these techniques will merge, but it will certainly happen. Think of situations that some instructions will be sent to the GPU by the OS even when the (OpenCL) programmer did not think of it. Or do you expect to be an ARM-processor integrated in a near-future CPU, when you write an OpenCL-kernel now?

See our article on the bright future of GPGPU to read more about it.

What’s next?

In case this is the way it goes, there will be a lot possible for both OpenCL and CUDA – depending on market demands. Some possibilities will be discussed in an upcoming article about FPGAs, but also let me hear what you think about X86-SoCs. Comment or send an e-mail.

Difference between CUDA and OpenCL 2010

Posted by Vincent Hindriksen on with 11 Comments

THIS ARTICLE IS VERY OUTDATED AND NOW SIMPLY UNTRUE FOR CERTAIN PARTS! NEW ARTICLE COMING UP.

Most GPGPU-enthusiasts have heard of both OpenCL and CUDA. While there are more solutions, these have the most potential. Both techniques are very comparable like a BMW and a Mercedes, but there are some differences. Since the technologies will evolve, we’ll take a look at the differences again next year. We’ve discussed this difference in a with a focus on marketing earlier this year.

Disclaimer: we have a strong focus on OpenCL (but actually for reasons explained in this article).

Terminology

If you have seen kernels of OpenCL and CUDA, you see the biggest difference might be the prefix “cl_” or the prefix “cu_”, but there is also a difference in terminology.

Matt Harvey (developer of Cuda2OpenCL-translator Swan) has summed up the differences in a presentation “Experiences porting from CUDA to OpenCL” (PDF):

CUDA term OpenCL term
GPU Device
Multiprocessor Compute Unit
Scalar core Processing element
Global memory Global memory
Shared (per-block) memory Local memory
Local memory (automatic, or local) Private memory
kernel program
block work-group
thread work item

As far as I know, the kernel-program is also called a kernel in OpenCL. Personally I like Cuda’s terms “thread” and “per-block memory” more. It is very clear CUDA targets the GPU only, while in OpenCL it an be any device.

Edit 2011-01-15: In a talk by Sami Rosendahl the differences are also discussed.

Speed-comparison

We would like to present you a benchmark between OpenCL and CUDA with full comparison, but we don’t have enough hardware in-house to do a full benchmark. Below information is what we’ve found on the net and a little bit based on our own experience.

On NVidia hardware, OpenCL is up to 10% slower (see Matt Harvey’s presentation); this is mainly because OpenCL is implemented on top of CUDA-architecture (this shouldn’t be a reason, but to say NVidia has put more energy in CUDA is just a wild guess also). On ATI 4000-series OpenCL is just slow, but gives very comparable to NVidia if compared to the 5000-series. The specialised streaming processors NVidia’s Tesla and AMD’s FireStream really bite each other, while the Playstation 3 unbelievably still wins on some tasks.

The architecture AMD/ATI-hardware is very different from NVidia’s and that’s why a kernel written with a specific brand or GPU in mind just performs better than a version which is not optimised. So if you do a benchmark, it really depends on which kernels you use for it. To be more precise: any benchmark can be written in favour of a specific architecture. Fine-tuning the software to work a maximum speed in current and future(!) hardware for different kinds of datasets is (still) a specialised task for that reason. This is also one of the current problems of GPGPU, but kernel-optimisers will get better.

If you like pictures, Hugh Merz comes to the rescue, who compared CUDA-FFT against FFTW (“the fastest FFT in the West”). The page is offline now, but you it was clear that the data-transfer from and to the GPU is a huge bottleneck and Hugh Merz was rather sceptical about GPU-computing in 2007. He extended his benchmark with the PS3 and a Tesla-s1070 and now you see bigger differences. Since CPUs go multi-multi-core, you cannot tell how big this gap will be in the future; but you can tell the gap will be bigger and CPUs will more and more be programmed like GPUs (massively parallel).

What we learn from this is 1) that different devices will improve if the demands are more clear, and 2) that it will be all about specialisation, since different manufacturers will hear different demands. The latest GPUs from AMD works much better with OpenCL, the next might beat all others in a many or only specific areas in 2011 – who knows? IBM’s Cell-processor is expected to enter the ring outside the home-brew PS3 render-farms, but with what specialised product? NVidia wants to enter high in the HPC-world, and they might even win it. ARM is developing multiple-core CPUs, but will it support OpenCL for a better FLOP/Watt than competitors?

It’s all about the choices manufacturers make, which way CUDA en OpenCL will develop.

Homogeneous vs Heterogeneous

For us the most important reason to have chosen for OpenCL, even if CUDA is more mature. While CUDA only targets NVidia’s GPUs (homogeneous), OpenCL can target any digital device that has an input and an output (very heterogeneous). AMD/ATI and Intel are both on the path of making architectures that are heterogeneous; just like Systems-on-a-Chip (SoCs) based on an ARM-architecture. Watch for our upcoming article about ARM & SoCs.

While I was searching for more information about this difference, I came across a blog-item by RogueWave, which claims something different. I think they switched Intel’s architectures with NVidia’s or he knew things were going to change. In the near future could bring us an x86-chip from NVidia. This will change a lot in the field, so more about this later. They already have an ARM-chip in their Tegra mobile processor, so NVidia/CUDA still has some big bullets.

Missing language-features

Like Java and .NET are very comparable, developers from both side know very well that their favourite feature is missing at the other camp. Most time such a feature is an external library, just built in. Or is it taste? Or even a stack of soapboxes?

OpenCL has:

  • Task-parallel execution mode (to be used on CPUs) – not needed on NVidia’s GPUs.

CUDA has unique features too:

  • FFT library – so in OpenCL you need to have your own kernels for it.
  • Atomic operations – which make double-write threads easier to implement.
  • Hardware texture interpolation – OpenCL has to fall back to a larger kernel or OpenGL.
  • Templating – in openCL you have to create new kernels for every data-type.

In short CUDA certainly has made a lot of things just easier for the developer, but OpenCL has its potential in support for more than just GPUs. All differences are based on this difference in focus-area.

I’m pretty sure this list is not complete at all, and only explains the type of differences. So please come to the LinkedIn GPGPU Users Group to discuss this.

Last words

THIS ARTICLE IS VERY OUTDATED AND NOW SIMPLY UNTRUE FOR CERTAIN PARTS! NEW ARTICLE COMING UP.

As it is done with more shared standards, there is no win and no gain to promote it. If you promote it, a lot of companies thank you, but the Rreturn-on-Investments is lower than when you have your own standard. So OpenCL is just used-as-it-is-available, while CUDA is highly promoted; for that reason more people invest in partnerships with NVidia to use CUDA instead of non-profit organisation Khronos. And eventually CUDA-drivers can be ported to IBM’s Cell-processors or to ARM, since it is very comparable to OpenCL. It really depends on the profit NVidia will make with such deals, so who can tell what will happen.

We still think OpenCL will win eventually on consumer-markets (desktop and mobile) because of support for more devices, but CUDA will stay a big player in professional and scientific markets because of the legacy software they are currently building up and the more friendly development-support. We hope they will both exist and help each other push forward, just like OpenGL vs DirectX, nVidia vs ATI, Europe vs the USA vs Asia, etc. Time will tell what features will eventually end up in each technology.

Update August 2012: due to higher demand StreamHPC is explicitly offering CUDA to OpenCL porting. 

Does GPGPU have a bright future?

Posted by Vincent Hindriksen on with 4 Comments

This post has a focus towards programmers. The main question “should I invest in learning CUDA/OpenCL?”

Using the video-processor for parallel processing is actually possible since beginning 2006; you just had to know how to use the OpenGL Shader Language. Not long after that (end 2006) CUDA was introduced. A lot has happened after that, which resulted in the introduction of OpenCL in fall 2008. But actually the acceptance of OpenCL is pretty low. Many companies which do use it, want to have it as their own advantage and don’t tell the competition they just saved hundreds of thousands of Euros/Dollars because they could replace their compute-cluster with a single computer which cost them €10 000,- and a rewrite of the calculation-core of their software. Has it become a secret weapon?

This year a lot of effort will be put to integrate OpenCL within the existing programming languages (without all the thousands of tweak-options visible). Think about wizards around pre-built kernels and libraries. Next year everything will be around kernel-development (kernels are the programs which do the actual calculations on the graphics processor). The year after that, the peak is over and nobody knows it is built in their OS or programming-language. It’s just like current programmers use security-protocols, but don’t  know what it actually is.

If I want to slide to the next page on modern mobile phones, I just make a call to a slide-function. A lot is happening when the function is called, such building up the next page in a separate part of memory, calling the GPU-functions to show the slide, possibly unloading the previous page. The same is with OpenCL; I want to calculate a FFT with specified precision and I don’t want to care on which device the calculation is done. The advantage of building blocks (like LEGO) is that we keeps the focus of development on the end-target, while we can tweak it later (if the customer has paid for this extra time). What’s a bright future if nobody knows it?

Has it become a secret weapon?

Yes and no. Companies want to brass about their achievements, but don’t want the competitors to go the same way and don’t want their customers to demand lower prices. AMD and NVidia are pushing OpenCL/CUDA, so it won’t stop growing in the market, but actually this pushing is the biggest growth in the market. NVidia does a good job with marketing their CUDA-platform.

What’s a bright future if nobody knows it?

Everything that has market-wide acceptation has a bright future. It might be replaced by a successor, but acceptance is the key. With acceptance there always will be a demand for (specialised) kernels to be integrated in building blocks.

We also have the new processors with 32+ cores, which actually need to be used; you know the problem with dual-core “support”.

Also the mobile market is growing rapidly. Once that is opened for OpenCL, there will be a huge growth in demand for accelerated software.

My advise: if high performance is very important for your current or future tasks, invest in learning how to write kernels (CUDA or OpenCL, whatever your favourite is). Use wrapper-libraries which make it easy for you, because once you’ve learned how to use the OpenCL-calls they are completely integrated in your favourite programming language.