BrookGPU: General Purpose Programming on GPUs

An anonymous reader writes " BrookGPU is a compiler and runtime system that provides an easy, C-like programming environment (read: No GPU programming experience needed) for today's GPUs. A shader program running on the NVIDIA GeForce FX 5900 Ultra achieves over 20 GFLOPS, roughly equivalent to a 10 GHz Pentium 4. Combine this with the increased memory bandwidth, 25.3 GB/sec peak compared to the Pentium 4's 5.96 GB/sec peak, and you've got a seriously fast compute engine but programming them has been a real pain. BrookGPU adds simple data parallel language additions to C which allow programmers to specify certain parts of their code to run on the GPU. The compiler and runtime takes care of the rest. Here is the Project Page and Sourceforge page."

Comments

High Performance for General Purpose?

[tempfile]

I suspect that this high performance is only attainable for the field the GPU is specialized for, i.e. graphics-related things. Or isn't it?

Re:High Performance for General Purpose?

[fidget42]

Actually, since "graphics-related things" are all matrix operations, this would turn the GPU into a high-end vector (matrix) engine.

Re:High Performance for General Purpose?

[Anonymous Coward]

"graphics-realted" things include things like floating point mathmatics, linear algebra, and vector operations. If you are doing anything computationally intensive, this might be usefull. You don't have to actually use the hardware to do anything graphical if you are just interested in turning numbers.

Re:High Performance for General Purpose?

[Total_Wimp]

I can't help but notice the similarity between shader operations and how neurons interact. These processors might be a good platform for some AI tasks.

I especially like the idea that the GPU and CPU can work together on the task. If the GPU was handling neuron tasks and the CPU was handling other necessary tasks we could get a very big boost to desktop AI

TW

Re:High Performance for General Purpose?

[BrainInAJar]

would the percision be enough though? as far as i know, GPU's do a lot of rounding off

Re:High Performance for General Purpose?

[Anonymous Coward]

NVIDIA's parts are OK, precision wise. You get IEEE floats, more or less. ATI's parts don't quite get you there are the moment, but their next series is planned to.

Re:High Performance for General Purpose?

[Directrix1]

Yes, anything computationally intensive that works over a range of data can usually find a parrallel solution. Such as image/video manipulation/encoding/decoding, encryption, and cracking (and hopefully this will give us a platform for better software RF). I've always wondered why this stuff didn't just become worked into a coprocessor. Because very little new stuff actually happened that was directly related with the video card (as in taking output from the machine and displaying it on a screen). I think the card manufacturers saw this, so they jumped on the 3d acceleration bandwagon toting it as a new video card feature, when it should've just been in the domain of a new math coprocessor.

Re:High Performance for General Purpose?

[axxackall]

Matrix and vector calculations with floating point makes GPU as a very excelent place to host Neural Network (NN) computation.

Of course NN can be used for "graphics-related things", such as image recognition, but not only image, for example voice recognition. And not only recognition, for example forecasting on huge sequences with explicit and implicit (hidden) side-factors.

Stock market trader on GPU, anyone?

Cool, but

[MooCows]

What kind of instructions does the GPU actually accept?
I mean, you probably just can't run any kind of algorithm on there can you?

Re:Cool, but

[scrytch]

> I mean, you probably just can't run any kind of algorithm on there can you?

Probably. I should imagine it has local storage with the corresponding fetch and store instructions, basic math, and ability to jump to arbitrary points in the shader program, which makes it very much turing complete. Everything else is a matter of a compiler backend. Bus latency would be an issue, so it'd be painful for programs that need a lot of I/O, but that's not an issue for a lot of programs.

Basically like having two processors...

[Anonymous Coward]

I wonder how long till we see a (insert worthwhile cause here)-At-Home client that supports this?

Re:Basically like having two processors...

[cybergibbons]

Ha! The C64 disk drive had it's own processor which you could use to run programs as long as you could deal with the painfully slow serial link. Beat that.

Cool ...

[torpor]

... can you say 'software synthesists' wet dream?

Oh, suddenly, that 'game investment' also gives you a few 100 extra voices of polyphony?

Sweet ... $5 to the first person to use Brooke to make a synthesizer. :)

first link is incorrect

[2.246.1010.78]

but the link to the project page is correct.

Like the good old days

[fiskbil]

Reminds me of the good old days when you used the processors in the C64 tapedrive to compute stuff. Wouldn't want to waste those precious cycles.

I'm sure a lot of old farts will tell me how they used some serial controller to compute stuff back in the 60's and that I'm just a little kid. :)

Re:Like the good old days

[Trracer]

I guess you mean in the C1541 floppydrive.

Re:Like the good old days

[tzanger]

Reminds me of the good old days when you used the processors in the C64 tapedrive to compute stuff. Wouldn't want to waste those precious cycles.

Actually it was the old 1540/1541 and later 1571/1581 disk drives. The tape drive did not have a processor in it.

wait a minute

[Janek Kozicki]

A shader program running on the NVIDIA GeForce FX 5900 Ultra achieves over 20 GFLOPS, roughly equivalent to a 10 GHz Pentium 4.

wait, if there is a technology that allows construction of GPU that is 3 times faster than the fastest CPUs, why Intel and AMD do not use this technology to build those 3times faster CPUs?

are you sure that you can compare the speed of GPU and CPU?

Re:wait a minute

[the uNF cola]

You are assuming using the GPU technologies are possible in a CPU. Because something is applicable in one instance doesn't mean it is in all instances. Making some things efficient may take away from the efficiency of others, but in the case of such aa specialized chip, it may not matter.

It may be ok to compare the speed of a GPU and a CPU if they are infact different. If a GPU was a CPU used with cheaper material, yeah, it would be unfair. But as life goes, they both have their merits.. so why not? A GPU is prolly best at some matrix math transforms.. or not. :)

Re:wait a minute

[enigma48]

Definately possible - general purpose CPUs have to do everything where graphics cards can specialize and do what little they can, faster.

Also, good point about comparing GHz to GHz - AMD CPUs do more per cycle than Intel, but are also clocked much lower. You could look at a subset of instructions (ie: FLoating-point OPerations (FLOPS)) but this only gives you a piece of the overall performance picture.

Without having read the article, my guess is they extrapolated (educated, math-based guess) how fast a 10GHz P4 would perform and compared the results that way.

I'd LOVE to see this tech built into a SETI or Folding@Home client (steroids version). (Imagine the kids - "Mom, I need the Radeon 9800XT to find a cure for Grandma's cancer!")

Re:wait a minute

[Entropy_ajb]

Because CPUs are limited to running instructions (for the most part) in serial. GPUs get to run a large number of instructions in parallel. As some above posts mentioned, a lot of the stuff the GPU can do is vector and matrix multiplication, therefore the GPU is really good at multiplying a lot of numbers times a lot of numbers at once. But in everyday life you aren't multiplying a bunch of number times a bunch of numbers at once, you are multiplying one number time another, then multiplying the result times a number, and so on. GPUs are built to a specific task, and at that task they are very fast, but outside that task they won't be able to compete with a real CPU. And on top of all of that I can buy 3 2.4Ghz P4s for the price of a Geforce FX5950.

Re:wait a minute

[mdpye]

And on top of all of that I can buy 3 2.4Ghz P4s for the price of a Geforce FX5950

But you forget the 256MB (at least) RAM on a steaming fast interface that you get with the GeForce... It makes the P4s' cache look pretty paltry in size by comparison.

MP

Re:wait a minute

[Kjella]

wait, if there is a technology that allows construction of GPU that is 3 times faster than the fastest CPUs, why Intel and AMD do not use this technology to build those 3times faster CPUs?

are you sure that you can compare the speed of GPU and CPU?

Well, yes and no. In the same way you can take a render farm and say that "this provides the equivalent of a 100GHz Pentium" Which might be true, for that specific task. You see it already between GPUs, compare Pentium, Xeon, Athlon XP and Athlon 64. Do you get one benchmark "X is 3% faster than Y"? No. Faster at some, slower at others. For a specific benchmark, the difference can be pretty big already among "general" processors.

A specialized processor like a GPU will show much greater variation. It might really shine on some, really suck on others. Which is why it's no good using a GPU as a CPU. Those numbers tell you that it can be much faster than the fastest CPU around. Or better yet, if you can make it run in parallell to the normal CPU, give you a total performance which may theoretically be about 13GHz (10 + 3), where 3 of those can be general-purpose operations. Or it may be a task the GPU runs like a dog, and isn't even worth the overhead.

Kjella

Re:wait a minute

[barik]

Are you sure that you can compare the speed of GPU and CPU?

Professor Pat Hanrahan, of Stanford University, made a stab at answering this question in his presentation 'Why is Graphics Hardware so Fast?'. The first half of the presentation focuses on this question, while the second half of the presentation covers programming languages that utilitize this hardware. Specifically, the Stanford Real-Time Shading Language (RTSL) and Brook are discussed. Overall, it's a good presentation that should get you up to speed with the basics of what's happening in this area of research.

How does this look?

[adrianbaugh]

I'm completely new to meddling with graphics card, so apologies if this is a silly question: when programs utilising the GPU for arbitrary calculations are running does the screen go weird, or is there a way of stopping the output being displayed? A screenfull of junk might not matter to a scientist leaving their computer to crunch numbers for a few months but it wouldn't be good for a general-purpose program.

Re:How does this look?

[Anonymous Coward]

Nope. Nothing appears on your screen until the contents of the area of memory known as the "frame buffer" are rewritten by a program (on either the GPU or CPU). The GPU can execute math code all day and you won't see the results unless it deliberately modifies the frame buffer.

I am not an EE, but...

[unfortunateson]

It would seem to me that the GPU is not going to be as general-purpose as the CPU, but could still attain the high mathematical throughput with vector-oriented processing.

Doing string searches, complex logic analyses, etc. would probably suck, but big data manipulations, such as SETI-style wave transformations, molecular analysis, etc., might be able to take advantage of them.

Good point.

[yoshi_mon]

After taking a quick peek at the language part of the project it seems right now that most of it's functions are all about sets of data and how to move them around.

Makes sence of course as that is what a GPU is all about. (Yes I'm vastly over-simplyifying here.) So I would gather that it might be used for types of data that are streamed alot? Maybe used for video editing, real time video, etc where your trying to deal with a lot of data at once that your trying to move around and not just store or have to perform some more complicated types of functions upon.

However, I'm no 3d programmer and I should would love a more detailed analysis of the potentals for this.

Fast Fourier Transform

[HalfFlat]

I'd love to see an FFT implementation (maybe it's not so hard ... will have to download and play with it.)

A lot of scientific code is constrained by how fast you can do an FFT, perhaps of arbitrary size. And a fast graphics card is a lot cheaper than a high-end processor.

For embarassingly parallel vector problems, this is just the sort of thing for cheap, powerful clusters based around a cheap PC and a fast GPU.

Re:Fast Fourier Transform

[Kazymyr]

Not to mention that you can put several PCI video cards in the same cheap PC. Multiply power by N.

Re:Fast Fourier Transform

[jonsmirl]

http://www.cs.unm.edu/~kmorel/documents/fftgpu/

The FFT on a GPU
This page contains supplemental material for the following paper.

Moreland, K and Angel, E. "The FFT on a GPU." In SIGGRAPH/Eurographics Workshop on Graphics Hardware 2003 Proceedings, pp. 112-119, July 2003.

Re:Fast Fourier Transform

[BiggerIsBetter]

Multiply power by N.

You work for Nvidia, don't you?

The deaf leading the blind...

[Kjella]

...but I assume that in any advanced texturing/shading/bump mapping/other GFX function rendering, you apply all the different effects, and when you're done, specifically call that the frame is to be displayed on screen. (E.g. why your FPS != your monitor refresh rate)

I would assume that this program simply never calls the drawing function, but instead gets the results back from the GPU. The normal screen should be able to run in the meanwhile (I assume you can e.g. build a 3D environment while showing a 2D cutscreen), so I would think you can have a plain GUI, as long as it doesn't need to use anything advanced.

Kjella

Homepage of GPGU research

[zymano]

www.gpgpu.org

Very cool. Vector/Graphics processors could one day overtake General processors. They are way more energy efficient too.

Drawing text with GPU shader units?

[jonsmirl]

Has anyone tried drawing text with GPU shader units? It would work something like this:

1) Each character would have it's own shader program.
2) You would set the shader program, draw a rectange, and the character would appear.
3) The shader programs would be automatically generated by processing TrueType files.

To implement:
1) Break Truetype outline up into a number of convex curve segments.
2) Each of these curve segments would be represented as a set of constants in the shader program
3) For each pixel, test a line from pixel to an edge.
4) If the number of segments crossed is odd the pixel is black else white.
The algorithm can be refined to add antialiasing and hinting.

What you end up with is text that is clear at any resolution. The size of the text is controlled by the rectangle you draw it in. The text can also be clearly rotated and sheared.

An obvious optimization is to get the GPU vendors to add a shader instruction to do the calculation for which side of the bezier curve segment the current point lies.

While not important for games drawing text is critical for desktops. And we all know about the current trends to draw desktops with 3D hardware.

Brook

[belmolis]

This looks like a straightforward and clean extension that experienced C/C++ programmers won't find difficult to learn, but it isn't entirely clear to me whether just using this language, without any knowledge of GPU architecture, will lead to big improvements in performance. Granted, you don't need to know the details, but you've got to have an idea of what it is that you're trying to do and in a general way how the special constructs of the language allow you to do that. As with other such language extensions, you can nominally write in the language but not really use the extensions (how many "C++" programs have you seen that were really C programs with // comments and a few couts?) or use them in unintended ways that prevent the intended optimization. It seems to me that if the project really is aiming at programmers who are not familiar with GPUs, they need at least to provide a brief introduction to the special properties of GPU architecture and some guidelines as to how to use the features of the language to take advantage of them. At present I don't find this either on the web sites or in the distribution.

Research

[dfj225]

I've always wondered why certain research programs (like Folding@home or SETI@home) don't use this type of code. My GPU sees more free time than my CPU plus it would probably get the work done faster. Also, imagine the speed increase of utilizing both the GPU and the CPU to their fullest potential. Now thats some fast folding!

Re:Research

[BiggerIsBetter]

I (and presumably others) have asked some project leaders about this, but it seems to come down to testing and support of various cards. Also, remember that this is relatively unknown technology - Amiga blitting aside ;-) - you have to be pretty sure it's going to give accurate and consistent results before using it seriously. Find-A-Drug was my project of interest, and they have a Linux version too.

GPU opcodes

[Anonymous Coward]

Here is a Beyond3d link that has some opcode info. Look around their site for a NV30 vs R300 architecture document that has lots of great stuff. If you are looking for the best s/n ratio, Beyond3d is one of the best. All meat, little fanboyism.

Nivida CG

[Popsikle]

Nvidia has this already!
"About Cg The Cg Language Specification is a high-level C-like graphics programming language that was developed by NVIDIA in close collaboration with Microsoft Corporation. The Cg environment consists of two components: the Cg Toolkit including the NVIDIA Cg Compiler Beta 1.0 optimized for DirectX(R) and OpenGL(R); and the NVIDIA Cg Browser, a prototyping/visualization environment with a large library of Cg shaders. Developers also have access to user documentation and a range of training classes and online materials being developed for the Cg language."

http://www.nvidia.com/object/IO_20020612_7133.html

Re:The future is the past

[Total_Wimp]

PCI-X can fix this data bus in other ways as well. Motherboards come with one AGP slot, but PCI-X can and will provide many expansion slots.

Picture five high end GPUs on the motherboard eclipsing the single high-end cpu for a fraction of the price. Intel and AMD would be forced to cut the asking price of their products to compete. We could finally see some real four-way competition for "processors".

TW

Re:I've always wondered when this would happen...

[Animats]

But what I'm really looking forward to is a Physics specific processor that sits alongside the graphics processor, and is resposible for collisions detection.

It's been done. The Havok game physics system is available for the Playstation 2, and the physics is running in the vector processors, where most of the PS2's compute power resides.

Collision detection isn't that CPU-intensive. (This may surprise people not familiar with the field. But it's true. If collision detection is using substantial CPU time, you're doing it wrong.) Correct collision resolution is where the time goes.

Physics code works better with double-precision FPUs. You need both dynamic range and long mantissas to do it well. Some of the game consoles, and most of the GPUs, only have single-precision FPUs. It's possible to make physics code work in single precision, but fast-moving objects that cover considerable distance may have problems.

Re:HP for GP?-AGP Bottleneck.

[Nexx]

WARNING: Lots of conjecture involved.

That said, if you can fit your data sets and your program on to the video memory (128MB isn't uncommon on high-end), and you're doing lengthy calculations on these sets while being only interested in the results (again, not uncommon in HPC), then the relative slowness of reading these results back becomes a nonissue.

Does that help? :)

GPU use for scientific programming.

[kiniry]

Researchers at Caltech and other institutions have been looking at this for about three years. See "Sparse Matrix Solvers on the GPU: Conjugate Gradients and Multigrid" by Bolz, Farmer, Grinspun and Schroder (SIGGRAPH 2003), for example. The paper, illustrations, and movies are available from Dr. Grinspun's homepage. The primary problems with the approach at the time this work was done was the limited bandwidth of texture-related operations in OpenGL based upon improper assumptions in pipeline optimization.

Re:Excellent!

[larkost]

2.1 GB/s is very nice, but it only refers to transfers in one direction: to the card. There is a (much) smaller bandwidth back to the motherboard. This is because for their designed purpose, graphics cards do not need to talk back to the system much, they just crunch the numbers and spit out the results to a monitor.

With encryption you are usually looking at processing streams of data. If your encryption method involves a lot of floating point math (almost never) on every bit of information, then it would be nice. But encryption is almost always integer based (GPUs don't' shine in integer like they do in floating point), and involves just as much data going in as coming back.

If you are looking for a great (co) processor for integers, look at the Altivec section of the G4 (and the similar one in the G5.. I forget the IBM name).

Imagine a Beowulf Cluster... no, seriously

[billstewart]

There's a cluster of Sony Playstations at UIUC (BBC) that's using the Emotion Engine to do numbercrunching and running Linux on the main processors to do communications and I/O. It's probably not strictly Beowulf, because it's using the Playstation version of Linux.

This cluster has 70 Playstations (one article said that they'd ordered 100, but only 70 are in the cluster... Obviously the others are being used for "research".)

Re:HP for GP?-Fakeout.

[sql*kitten]

I thought the real reason to get a *professional level* card is to get a guarantee of reliability

Well, ISV certification - a CAD vendor will assert "with this card, our software produces no rendering artifacts".

A real-world example - ray tracing

[ron_ivi]

http://portal.acm.org/citation.cfm?doid=566654.566 640
http://www.theregister.co.uk/content/54/25312.html
http://online.cs.nps.navy.mil/DistanceEducation/on line.siggraph.org/2002/Papers/13_GraphicsHardware/ purcell.ppt

How long until

[Lord Kano]

Someone ports a GPU Linux and some asshole loads 8 PCI cards into his machine and maked a beowulf cluster inside of one case?

Re:AT&T DSP32 Cluster Supercomputer in late 80

[billstewart]

Yes, they were 25 MFLOPS. The chip had a 12.5 MHz cycle rate (I think that was also the clock speed), and each cycle could do a 32-bit multiply, a 32-bit add, and a 24-bit simple integer operation (some integer ops took multiple clocks, I think?)

Your music application sounds like fun. I didn't know anybody was still doing anything quite like that by 1990 - there was a whole range of people around John Cage's time who did lots of prepared piano stuff.

Some of the people who were trying to sell our multi-processor supercomputer flavor came up with a music studio application, doing lots of audio processing and mixing, sort of like your device turned inside out. Don't know if they sold more than one of them before the Lucent spinoff took them away.