|
Hardware-accelerated
scientific computing and visualization The GPU has evolved
into a highly parallelized multiprocessor machine over the last several
years. In addition, the GPU supports the high precision computation (up to
32bit float), and is fully programmable. Due to their increasing
computational power and flexibility, current GPUs
are becoming a very powerful computational platform for general-purpose
computation problems as well as traditional computer graphics tasks. In this
work, we focus on developing parallel algorithms to solve expensive
scientific computing problems efficiently. We developed a novel computational
technique, which we call the Fast Iterative Method (FIM), to solve a class of
Hamilton-Jacobi (H-J) equations on massively parallel systems, e.g., GPUs. The proposed method is classified into a class of label-correcting
algorithms, and has suboptimal worst-case performance, but in practice,
performs fewer computations per node than guaranteed-optimal
alternatives. Furthermore, the proposed method uses only local,
synchronous updates and therefore has better cache coherency, is simple to
implement, and scales efficiently on parallel architectures, such as multicore processors or graphics processing units (GPUs). We also implemented a super-fast DT-MRI tractography system as an application of the proposed GPU
H-J solver. Papers and pre-prints
PDE(Partial
Differential Equation)-based, hardware accelerated image processing In this
work, we focus on PDE-based image processing algorithms and their
implementation on parallel hardware, e.g., GPUs
(Graphic Processing Units). PDE-based image processing algorithms can
formulate the problems as a process of diffusion or optimization, and many
interesting algorithms have been introduced during the past decade. The PDE
on a regular grid (image) can be mapped and solved on the GPU very
efficiently. We have developed several fast GPU-based PDE solvers and their
applications in the field of computer graphics and geoscience.
We developed a multigrid solver for anisotropic PDEs on the GPU for edge preserving diffusion on height
field data. For 3D seismic volume processing, we developed a
structure-oriented anisotropic filter which removes random noise and enhance
stratigraphic coherency while preserving geological features, e.g., faults
and channels. We showed the proposed method runs very fast on a commodity graphics
card, up to 15x speed up, and implemented an automatic fault extraction
system for interactive fault detection. Papers
Aug 9, 2007 Won-Ki
Jeong |
