OpenCL-Based Linear Algebra Libraries for High-Performance Computing

Despite its promise, OpenCL adoption is slow, owing to a lack of libraries and tools. Vendors have shown few signs of plans to provide OpenCL libraries, and were they to do so they would likely be incompatible with one another, much as NVIDIA's BLAS (CUDA) is presently not interchangeable with Intel's BLAS (MKL). The unified language and environment of OpenCL allows the community to ensure that the spirit of the language?its interchangeability?is reflected in its library and tools ecosystem. EM Photonics is well positioned to lead this effort; we have strong ties to several hardware manufacturers, to application developers, and we maintain a world-class LAPACK library for NVIDIA GPUs. To begin this process, EM Photonics proposes the development of a set of OpenCL BLAS routines and the framework necessary to allow researchers, developers, and hardware manufactures to integrate platform optimized versions of BLAS libraries. This software will be made open source to encourage community involvement and allow it to continue to evolve with the with future hardware technology. Upon completion of this project, EM Photonics will have developed a complete set of OpenCL BLAS routines. In addition, we will have the framework necessary for their efficient execution that also allows new routines to be added by either EM Photonics or third parties. This package will be released under an open source license to encourage community participation and allow for widespread adoption, and EM Photonics will be its ongoing steward. The commercial success of our CULA product has both opened doors for partnership opportunities and provided us commercialization opportunities that can be further leveraged once this project is complete. Based on this combination of technology, experience, partnerships, and commercial momentum, we are convinced this project will successfully meet our SBIR objectives and continue to flourish beyond.