W Randolph Franklin, President and Managing Director
info@quantum-geometry.com
We design and implement parallel algorithms for computational geometry, CAD, and computational cartography executing on multicore and many core (NVIDIA GPU) processors.
Prof Franklin is an Emeritus Professor of Engineering at Rensselaer Polytechnic Institute. His background includes a PhD, Applied Math, Harvard, sabbatical in the EECS Dept, UC Berkeley. He served 3 years as a Program Director in NSF/CISE, which included 2 joint solicitations with DARPA/DSO named Computational Algorithms and Representations for Geometric Objects (CARGO). Here is his detailed website and long resume. His proudest accomplishment at RPI is his 18 PhD students and 70 masters students.
At RPI, he created the first Quantum Computer Programming course in the School of Engineering and taught it three times. Here is its announcement, syllabus, and blog.
He has been designing and implementing efficient graphics, CAD, and GIS algorithms for many years, as detailed on his website and resume. At his level, all those areas are quite similar. His code is freely available for non-profit research and education. Skeptics are encouraged to bang on it and look for problems. His programs' efficiency derives from combining various unusual techniques, as this paper demonstrates:
This paper presents a technique for employing high-performance computing for accelerating the exact evaluation of geometric predicates. Arithmetic filters are implemented using interval arithmetic to reduce the necessity of exact arithmetic while ensuring the results of the predicates are still exact. Furthermore, the computation with interval arithmetic is offloaded to a CUDA-enabled GPU. If the GPU detects that some results cannot be trusted, the corresponding predicates are re-evaluated in parallel on the CPU using arbitrary-precision rational numbers. As a case study, a red-blue 3D triangle intersection algorithm has been implemented. Since the intervals are implemented using floating-point numbers, the parallel computing power of GPUs for processing these numbers led to a speedup of up to 1936 times (when compared against a similar sequential implementation) in the evaluation of these predicates (and up to 414 times if the entire runnning-time of the algorithm is considered). The excellent performance associated to the exactness makes this technique suitable for accelerating geometric operations in fields such as CAD, GIS and 3D modeling.