Programming in CUDA

Lines of code counter for cuda

2011-08-11T13:32:00.000+01:00

After a quick search I didn't find a fully baked version which had .cu line counting. I looked briefly at:
CLOC
Sonar
sclc
SLOCCount

None had cuda support out of the box but CLOC allowed you to define your own language support.
CLOC requires perl so I installed ActiveState ActivePerl 5.12.4 to run on windows.
I then downloaded CLOC-1.5.3.exe

Next step is to define support for cuda. This was very straightforward.
run
cloc.exe --write-lang-def= <your definition file>

then edit <your definition file> to include a definition for cuda.
The CLOC howto on this is here
The lines I added were:
Cuda
filter remove_matches ^\s*//
filter remove_inline //.*$
filter call_regexp_common C
extension cu
3rd_gen_scale 1.51

To run cloc with your updated definition file:
cloc.exe --read-lang-def= <your definition file> <list of files/folders>

You will then have an output similar to this:

4 text files.
3 unique files.
1 file ignored.

http://cloc.sourceforge.net v 1.53 T=1.0 s (3.0 files/s, 139.0 lines/s)
-------------------------------------------------------------------------------
Language files blank comment code
-------------------------------------------------------------------------------
C++ 1 19 2 51
Cuda 1 13 2 47
C/C++ Header 1 0 0 5
-------------------------------------------------------------------------------
SUM: 3 32 4 103
-------------------------------------------------------------------------------

This has not been extensively tested yet.
As a quick check you can output a version of processed files with comments/white spaces removed
cloc.exe -strip-comments=uncom --read-lang-def=<your definition file > <list of files/folders>
Which will output uncommented versions of your processed files with an additional .uncom extension
You can then compare the correctness of parsing using these files.

Streaming multimedia to an xbox 360 from a laptop on the same subnet with UPC EPC2425 Cisco router in Ireland

2011-04-04T13:21:00.000+01:00

xbox 360 has support for streaming multimedia from a laptop using windows media player 11. Unfortunately, the EPC2425 UPC setup does not support multicasting and therefore windows media player streaming out of the box will not work.

A workaround is to download Tversity free version. You may run into problems starting the Tversity Multimedia server and timeouts.

This is due to conflicts with the flexnet licencing serve and acroread service which can be disabled using services.msc executed from start->run.

Introducing new Tesla Fermi based products webinar review

2009-12-16T19:06:00.000Z

Just after attending Nvidia's "Introducing new Tesla Fermi based products" webinar.
It was light on detail and was more marketing than technical but I did pick up some nuggets of info.
Current planned release for Tesla based Fermi products is Q2 and Q3.
In the Tesla personal supercomputer market current product been C1060,
the Tesla C2050 will be available in Q2: 520 -630 GFlops DP, 3GB mem with ECC,
and C2070 in Q3: 520 -630 GFlops DP, 6GB mem with ECC.
In the Tesla Datacenter product line, current model been the S1070
the Teslo S2060 will be available in Q2: 2.1 -2.5 TFlops DP, 3GB mem with ECC,
and S2070 in Q3: 2.1 -2.5 GFlops DP, 6GB mem with ECC.

There was some information on faster infiniband with DMA between the Mellamox Infiniband and Fermi chips.

Regarding ECC memory performance profiling they are seeing a 0-10% drop in memory bandwidth with ECC and no ECC support on GeForce product range (as expected).
GDDR5 is about 2X faster than GDDR3.

Most other aspects ddiscussed is available in greater detail in the Nvidia Fermi Whitepaper

MIT and Harvard researchers demonstrate a better way for computers to see

2009-12-16T17:21:00.001Z

The video and academic source article are well worth the time. The high throughput screening methodology like all many good ideas is extremely simple and highlights the importance of how clever quantitive research can be a winner when compared to restricted qulaitive research.

in reference to: MIT and Harvard researchers demonstrate a better way for computers to ‘see’ (view on Google Sidewiki)

Coordinating the use of GPU and CPU for improving performance of compute intensive applications

2009-12-16T16:32:00.001Z

Just finished reading this paper. It looks at the allocation of hetrogenous CPU/GPU resources to solve a classification problem.
It uses and extends an event driven parallel framework called Anthill.
The concepts used are well known and well understood parallel models. It uses streams to label data streams through data filter functions. The functions can be implemented on the CPU and/or GPU and the user can ask for filter function weighting to be used i.e. GPU filter function is 3x CPU filter.
The candidate example application of Neuroblastoma image analysis displays some of the power of this type of technique.
The parallel model allows:
* Data parallelism
* Task parallelism
* aschrony. I don't necessarily agree with the aschrony concept. As I would consider it a translation of a recursive type algorithm to something simailar a split and search. Where the arbitrary splitting may lead to redundancy in the search space which is hidden due to the increased parallel throughput.

There is also very minimal discussion on the memory management which I would consider the main complication of hetrogenous systems. Nothin is mentioned on whether the host-->device transactions are synchronous/asynchronous and do they depend on the compute mode of the GPU used.

The overall programmming model reminds me of a DSP model and although leads itself nicely to classifier and data mining applications. I am not sure of its transferability to a wider application domain.
In addition, no discussion is given on the developer cost in developing multiple implementation for CPU/GPU versus performance improvement.

in reference to: Coordinating the Use of GPU and CPU for Improving Performance of Compute Intensive Applications :: GPGPU.org (view on Google Sidewiki)

Third Workshop on General-Purpose Computation on Graphics Processing units

2009-12-16T14:16:00.001Z

I won't be able to attend this myself but may be of interest to others.
GPGPU-3, 14th March in Pittsburgh
Papers due: January 15, 2010

in reference to:

"Papers due: January 15, 2010"
- Workshop on General Purpose Processing on Graphics Processing Units (view on Google Sidewiki)

Frontiers of GPU, Multi- and Many-Core Systems Workshop

2009-12-16T14:09:00.001Z

Workshop on in Melbourne, Australia from the 17th-20 May, 2010. Paper submission 31st December.
Not too sure if I'll be able to attend myself but looks very interesting. Looking at cost effective translation of applications across differing parallel architectures is an admirable but challenging goal.
Let me know if anyone is planning on attending this or has previousily attended any of the related workshops.

in reference to: Salsa Workshop (view on Google Sidewiki)

DirectX 11 used in Colin McRae 2

2009-12-02T15:39:00.001Z

I've became very jealous of those with an ATI Radeon Series GPU after watching the video in the referenced article.
The full game is not released (actually just released today I think) but there is a demo link in the article for those who want to have a real taster. Interestingly the main conclusion in the article was that the DirectX11 functionality in the demo hurt frame rate by 30% and didn't give any appreciable improvements. Still roll on DirectX 11 support in Nvidia's upcoming fermi chip and games titles with DirectX11 built in for the bottom up.This will be when we will be really able to rate the performance of DirectX11 functionality.
Also a big question is what version of DirectX will be in the next generation of consoles, as many PC games these days are console ports and therefore mainly DirectX 9 derived. As the next gen consoles are more than likely going to be locked down sometime next year, and release around 2012-2013 then DirectX 11 looks like the prime candidate.

From a technical perspective, the game Dirt 2 uses Directx 11 in five situations:
* Hardware tesselated dynamic water, animated crowd and dynamic clothing.
* Full floating point HDR lighting, otherwise known as high dynamic range rendering (HDRR or HDR Rendering) carries out light calculations in a higher dynamic range. Therefore bright things can be very dark, light things very bright and details can be seen in both instances.
* High Definition Ambient Occlusion
* Full screen resolution post processing.

All of these features look extremely useful and I can't wait for a Nvidia DirectX 11 chip so I can spend a bit of time playing around with them.

in reference to: bit-tech.net | Feature - DirectX 11 performance first look: Dirt 2 (view on Google Sidewiki)

Interview article with Jen-Hsun Huang

2009-11-30T14:02:00.001Z

Jen-Hsun Huang is the Co-founder and CEO of Nvidia. In the referenced article he gives an an interview to www.tbreak,com

I liked his advice on business plan development as I am currently working on this.
"when I started writing NVIDIA’s Business Plan and after three months of rephrasing and making it sound good, I realized that if I kept working on this, the market opportunity will pass us. I knew what we believed so I just wrote the first principals and list of actions that are basically strategies on what we believe in and the resources we have. To this day, NVIDIA’s business planing is exactly like that- Short and crisp."

A key learning is the way a key architectural decision (limited use of high cost memory) led to initial success but when the price point of memory changed.

His take on the difference of strategy and technology. Two companies can have two similar products but widely varying strategies. Within Nvidia he defines strategies as "things like how do you build a product, how do you pace it, how do you work with developers, what are the things you want to be really good at and lesser-good at etc."

The strategies used by Nvidia to bring geometry processing and SLI to market when large players such as microsoft, OEMs and Intel were not happy about the technologies.

He also gives his take on future direction: ION on the smartphone, Tesla in the cloud, gaming with physx, raytracing and 3D.

in reference to: An interview with NVIDIA's CEO Jen Hsun Huang | t-break: Tech @ Its Fastest (view on Google Sidewiki)

PyCUDA: GPU Run-Time Code Generation for High-Performance Computing

2009-11-26T18:21:00.001Z

Have yet to use pyCUDA but hope to get around to using it at some stage maybe for some rapid prototyping.
The reference article gives some nice technical detail on the use of run-time code generation (RTCG) for the GPU and the use of pyCUDA.

in reference to:

"PyCUDA: GPU Run-Time Code Generation for High-Performance Computing"
- [0911.3456] PyCUDA: GPU Run-Time Code Generation for High-Performance Computing (view on Google Sidewiki)

A fast two-dimensional floodplain inundation model

2009-11-26T16:18:00.001Z

Looks like a very nice application for CUDA and has particular resonance with me currently due to Irelands current record breaking floods. I have yet to find the paper available online though.

in reference to: A fast two-dimensional floodplain inundation model :: GPGPU.org (view on Google Sidewiki)

Nice article on MIT algorithm lectures

2009-11-26T16:13:00.001Z

Goes through the complete lecture set of MIT's introduction to algorithms lecture course.

in reference to: Summary of all the MIT Introduction to Algorithms lectures - good coders code, great reuse (view on Google Sidewiki)

End of Line for IBM's cell

2009-11-26T16:05:00.001Z

IBM's VP of Deep Computing, David Turek, has stated that the Cell processor is at the end of its evolutionary line.
While many of the concepts are planned to live on in other IBM product offerings the cell product will not be developed further.
This has been clarified by an IBM sokesman here http://www.driverheaven.net/news.php?newsid=344
Only one development cycle has been halted for one product in the Cell processor family the PowerXCell-8i cpu. So whether the Cell processor will be used in playstation 4 is still open to debate.

The main sticking point for the cell has been it's programmability. It has the name of a programming model that is considered outright hostel by many developers.

The future of the heterogenous single chip multiprocessors (CMP) does not end with the Cell with a large number of companies trending towards this model.

The Nvidia CUDA model shows a model that works well for programmability and with the future release of extended functionality in the fermi architecture. The future looks interesting. In addition, the rumour mill is still full of talk regarding a Nvidia x86 type chip in the future.

Then we have the planned fusion of the CPU and GPU in next generation offers from Intel and AMD. In addition, we have the planned release of Larrabbe with vectorised x86 based multicore architecture. Oficially, this product is not planned for high end graphics but will be an interesting and potentially very adaptable product.

It will now be interesting to see what the playstation 4 and the future generations of HPC will be based on Intel, Nvidia, IBM and AMD or someone else?

in reference to: End of the line for IBM's Cell (view on Google Sidewiki)

Nvidia Fermi Whitepaper

2009-10-15T20:07:00.000+01:00

Got a chance to go through the Nvidia Fermi whitepaper available here.
This really looks like a very exciting architecture with some extremely nice features.

Heres a list of some that I hope to hope to go into more detail in later posts.
potentially 8x faster double precision
ECC support
64 KB RAM with configurable partitioning for L1 and shared memory
A L2 cache
Faster atomic operations
Faster context switching
Dual wrap scheduler whichs allows two independent warp dispatches
unified address architecture with support for C++
Optimised OpenCL and directCompute execution
Full IEEE 754-2008 32-bit and 64-bit floating point precision
Full 32-bit integer pipeline with 64-bit extensions
Support for 64-bit memory
Faster predication support
Faster atomic operations
Gigathread technology
10x faster context switching
concurrent kernel execution
overlapping bidirectional memory operations
Out of Order thread block execution

Powering GPUs with diesel

2009-10-14T13:52:00.000+01:00

Its very interesting to view the current performance/watt of CPU versus GPU technologies.

A good example is the article here. Where they are trying to achieve radio astronomy observations which are much wider than the current 3 degree limited radio telescopes. It is planned to achieve this through a grid of smaller cheaper antennaes called the MWA [Murchinson Widefield Array]. Unfortunately the telescopes need to be developed in extermely isolated areas to remove as much radio noise as possible. The frequency they are searching at is 300 - 800 Mhz.
A prototype system which had 32 antenna clusters with each cluster formed out of 16 small antennas was developed in the Australian outback (the complete system planned to have 8000 antennaes). A major issue was powering a system in such an isolated area. Looking at various power requirements it became obvious that for there particular processing domain GPUs are currently a much more power effecient alternative than a CPU only solution. For the prototype all that was available for power was a 20Kw diesel engine. In its current prototype implementation it requires 20 teraflops of processing power. Given that you would need around 200 CPUs for the job, as 200 Xeon 5500 CPUs at 3.2 GHz [100 GFLOPS each] would consume a grand total of 24,000Watts [that's not counting the rest of the computer needed for these CPUs to function]. Overall estimate in CPU-based setup was in excess of 55,000 Watts [55kW] and obviously, it was a dead end. On the other hand, in around 1kW ofpower, scientists managed to squeeze 4.5TFLOPS [dual GTX 295 card], meaning "only" 5.5kW is needed for 20TFLOPS. With upcoming Fermi-based cards, Australians expect to build a 20TFLOPS setup using only3.3kW.

But this issue is not closed a fully grown array will generate 1Tbps of data and require 1 EFLOPS [Exa] to process all the generated data. if you would need to build a 1EFLOPS supercomputer today, you would need quite an interesting number: 666,666 Fermi-based Tesla cards with 3.99 PetaBytes of GDDR5 memory. If you require 1EFLOPS from world's most powerful CPUs, the clock would stop with Intel Xeon W5590 at 3.33 GHz, and you would need around 10 million of them. In both cases, EFLOPS is unattainable.

Therefore it will be interesting when MWA opens in 2020 what will be the underlying processing technology.

OpenGL 3.2

2009-08-26T12:07:00.000+01:00

I'm doing a bit of work these days on OpenGL with CUDA.
A decent if a little AMD specific take on the new features of OpenGL is available here.

A good webinar of openGL and its new features is available from Nvidia here.
while the Nvidia OpenGL 3.2 beta drivers are available here.

The AMD article highlights the following new features.
Geometry Shader Support --> geometry primitives can be modified on GPU
Sync objects --> keeps CPU and GPU in sync
Multisample tectures and samplers --> Can apply multisample rendering hardware to extures and render buffer objects, instead of only screen space windows. Shader can read from each sample of a multisampled texture and apply custom blend schemes.

I hope to give an overview of the Nvidia webinar at a later date.

Simple Ray tracing example

2009-08-11T14:11:00.000+01:00

Peter Trier from the computer graphics group at the Alexandria Institute has written a nice simple tutorial for building a CUDA based ray tracer here.

Running pre-built executable
The executable ran without issue when I copied the glut32.dll availble from \C\bin\win32\Release into example directory.

Building available code
The build instructions I have here were applied using Windows XP 32-bit, Visual Studio 2005 and CUDA 2.3.

With regards to building the code.

Move the src folder to \src
When you open the CUDA_raytracer.vcproj you are asked to create a .sln do this. (Another option would be to include this vcproj in an existing .sln such as the release.sln in the \src folder.
Install AntTweakBar. It is used to for the user interface in the ray tracing program. After installing this. Can simply copy the AntTweakBar.h, dll and lib files into the example source folder.
There is an error regarding missing Cg libraries these can be removed as there is no dependencies in example code.
The app will be built at \NVIDIA GPU Computing SDK\C\bin\win<3264>\
To run application copy the AntTweakBar.dll and data folder into build folder.
The application should now run

Very good article on CUDA anf the GT200 architecture

2009-07-29T18:34:00.000+01:00

here is one of the best articles around to get a handle on CUDA with emphasis on the GT200 Architecture.
It has:

A really nice concise overview of the way Directx in many ways drove hardware towards GPGPU capabilities.
Some basic pointers on what does and doesn't map well to the GPU.
A nice concise overview of where memory is located, its scope and potential usage scenarios.
Overview of the CUDA API architecture, toolchain and optional libraries etc.
Describes the extensions of different hardware compute architectures.
A nice description of the GT200 system architecture and contrasts it with other architectures such as x86, Niagra.
Nice analysis of the actual hardware architecture such as the TPC, SM, memory pipelines, clocking domains
A nice analysis of the TPC --> SM --> SP control and computation model.
Useful info such as the cycle times for differenet operations, dual issues and hardware constraints.

Sun's Niagara 3

2009-07-29T16:20:00.000+01:00

Report here on Sun's Niagara 3. It is reported that it will be out by late 2009 and will have 16-cores and 16 threads per core.

Dated but decent article "Nvidia's CUDA: The End of the CPU?"

2009-07-23T19:12:00.000+01:00

The article here. Is a little out dated but well worth a read for newbies to CUDA.

It gives good brief background history of where CUDA evolved from.

Although some of the details are outdated. For example double precision has been available for a while in CUDA and support for SLI has been made available just recently with CUDA 2.3.
It is also quite specific to 8800GTX hardware.

HEXUS.net - Tech explained :: NVIDIA Tegra SoC

2009-07-15T18:14:00.001+01:00

A good overview of Nvidia new SOC for mobile devices.
HEXUS.net - Tech explained :: NVIDIA Tegra SoC : Page - 1/1

Review of CUDA enabled video editing software

2009-07-15T18:11:00.001+01:00

A review of available CUDA enabled video editing software.

NeST ties up with Nvidia

2009-07-15T18:01:00.000+01:00

It is reported here that Network Systems and Technologies (NeST) based at the Technopark has tied up with visual computing initiative major Nvidia
Corporation

Analysis of Risk

2009-07-15T15:57:00.000+01:00

Bruce Schneier discusses the over estimation of terrorist risk here. He adopted the ideas of the a software engineering paper More Risk Analysis Can Lead to Increased Over-Optimism and Over-Confidence to the estimation terrorist attack risk. Both articles make interesting reading especially where they discuss behavioural economics, the cognitive biases that make us think and act.
One is control bias. Where someone underestimates the risk when they are in control and overestimate risk when they are not. A good example of this is peoples estimation of risk when driving themselves compared to been flown by others.
Another is availability heuristic. Where we estimate on what can be easily brought to mind. The example given is we overestimate what we see and hear on the news and underestimate hard to imagine risks.
The peak end rule is basically we give additional weight to the last imagined risk.

Using vi with CUDA

2009-04-16T12:39:00.000+01:00

Here is a nifty configuration file for CUDA syntax highlighting in vi.

cuda and cmake

2009-04-16T12:24:00.000+01:00

cmake is a cross-platform makefile generator. http://www.cmake.org
Abe Stephens has developed a really nifty script to allow the use of cmake with CUDA. FindCuda.cmake.

If you have any issues a good place to ask questions is here.

GPU Gems series of books available online.

2009-04-14T13:45:00.000+01:00

The book series GPU gems is a great source on GPU programming techniques. They are very well written with a wide range of articles. The third book in particular has quiet a number of articles on CUDA.

The three GPU gems books are available free online:
http://developer.nvidia.com/object/gpu_gems_home.html
http://developer.nvidia.com/object/gpu_gems_2_home.html
http://http.developer.nvidia.com/GPUGems3/gpugems3_part01.html