NVIDIA pioneering OpenCL support on top of CUDA

Apple has spun its OpenCL API off to the Khronos Group, which maintains it as an open, royalty free standard that any GPU maker can implement. The first operating system to support OpenCL will be MacOS X 10.6 Snow Leopard, which debuts next year. Khronos also maintains the OpenGL graphics programing API, and the OpenAL API for audio. OpenCL was given a similar name to associate the technologies together.

Why OpenCL?

Clock rates of general purpose CPUs are no longer rapidly increasing; Intel and other CPU makers are now using multiple cores to speed up their processors. In contrast, video card GPUs are gaining tremendous new processing power in addition to gaining multiple core support.

OpenCL is designed to allow developers to spin processor intensive tasks to the often idle GPU to take advantage of all that latent processing power. Additionally, OpenCL works with Snow Leopard's Grand Central scheduling technology to support multiple core architectures to do as much as possible in parallel across both the CPU and any available GPUs.

An OpenCL head start for NVIDIA, via CUDA

NVIDIA has a leading advantage in supporting OpenCL because of the work the company has already completed with its CUDA driver interface, to implement parallel programing and "GPGPU," or general purpose computing performed on graphics processing units.

Manju Hegde, the General Manager of CUDA at NVIDIA, explained to AppleInsider that supporting OpenCL on a GPU requires certain hardware capabilities such as scatter write, as well as certain generality of control flow. Both have already been implemented in NVIDIA's CUDA architecture.

NVIDIA's CUDA ISA and hardware compute engine "were designed to support multiple entry points into the compute power of the GPU including standard computing languages (such as C, Fortran, etc) as well as API style interfaces like OpenCL," Hegde wrote in an email interview.

Rather than being competing technologies, Hegde noted that "OpenCL is a layer on top of the CUDA driver interface. As such, OpenCL is one avenue to GPU computing through CUDA, C for CUDA is another."

OpenCL vs CUDA?

When asked how NVIDIA's CUDA compares with OpenCL, and if NVIDIA is planning to support both in its future products, Hegde explained, "This is probably better put by saying how does C for CUDA compare with OpenCL – this is a language to language comparison."

Hegde added, "The answer is that the two share very similar constructs for defining data parallelism, which is generally the major task, so the code will be very similar and the porting efforts will be minor.

"As OpenCL is another method of accessing the GPU, we wholeheartedly support it. Its sits seamlessly on top of our CUDA architecture and as such, developers using NVIDA hardware have a choice of language and programming environment.

"With regards to product support, we plan to have OpenCL supported on the CUDA architecture which means that any NVIDIA GPU built upon the CUDA architecture will support OpenCL. This means every GPU (including GeForce, Tesla and Quadro lines) from the GeForce 8 series onwards will support OpenCL. This gives OpenCL developers an installed base of more than 100 million GPUs."

Will OpenCL be Compatible?

We asked NVIDIA if it thinks OpenCL (which has also been adopted as the latest GPGPU strategy by AMD) is complete enough to give users a seamless experience when running OpenCL software across different GPU architectures (in the manner of PostScript across various vendors' laser printers), or if it anticipates problems and incompatibilities between vendors' implementations (in the manner of Java across various devices' implementations).

Hegde answered, "OpenCL is a multi-vendor standard and so the expectation is that if a vendor has an OpenCL compliant implementation, code written in OpenCL should run seamlessly across their architectures.

"NVIDIA has followed a very consistent and unfaltering strategy with CUDA. The C for CUDA programming model is being taught in more than 50 schools around the world. We have in excess of 25,000 developers actively working on CUDA today. If you look at www.nvidia.com/cuda, you’ll see hundreds of codes and applications that are using our CUDA architecture today. Moreover, CUDA was designed to natively support all parallel computing interfaces and will seamlessly run OpenCL and future standards as they arise."

The future of OpenCL

We also asked NVIDIA for some immediate examples of applications that can take advantage of OpenCL now, and what future potential it sees in the specification.

Hegde answered, "While the OpenCL spec is announced today, there are conformance tests that need to be developed and then final implementations will be released around Q2 next year. So we are a little way away from having apps that can take advantage of OpenCL today. Of course C for CUDA is available today on the Mac OS, so developers wanting to start developing for the GPU can get started now and as we said before, both C for CUDA and OpenCL share very similar constructs for defining data parallelism, so if they wish, porting that code to OpenCL after its full release, will be easy.

"In terms of potential, it’s huge! The enormous parallel processing power of the GPU has been delivering speed up for 20-200X in many codes, from oil and gas exploration and medical imaging to video transcoding. As more and more developers begin to port their apps to the GPU, you will see a new wave of applications hit the market."

OpenCL cross platform and other GPGPU standards

When asked if any direct OS support is required to implement OpenCL support on other platforms, and if NVIDIA sees momentum clearly building behind OpenCL as the standard for GPGPU computing, Hegde noted that, "yes, the first OS to support OpenCL will be Snow Leopard."

Hegde also said, "In the world of parallel computing, there are a range of standards emerging, C for CUDA from NVIDIA, OpenCL from Khronos, DX11 Compute from Microsoft, and so on. Developers like to have choice, they pick whatever different programming style suits their needs/deployment. And developers will use the interface most comfortable to them, ie: one that supports libraries and OS that they are accustomed to.

"NVIDIA will continue to invest in both its CUDA architecture and its C for CUDA programming environment, while also offering robust support for new standards as they emerge.

"To summarize, we think OpenCL is great and we support any initiative that unleashes the massive power of the GPU. We have worked extremely closely with Apple on the OpenCL spec, OpenCL was developed on NVIDIA GPUs and we were the first to show working OpenCL code so we are confident that our implementation of OpenCL will be second to none. The addition of OpenCL to our industry leading toolkit for GPU Computing means a fantastic array of choices for developers."