50 likes | 183 Views
CSE 690: GPGPU Lecture 8: Image Processing PDE Solvers . Klaus Mueller Computer Science, Stony Brook University. A Quick Note. The recent generations of GPUs do seem to conform to the IEEE floating point standard Nvidia site explicitely confirms this for Quadro FX
E N D
CSE 690: GPGPULecture 8: Image Processing PDE Solvers Klaus Mueller Computer Science, Stony Brook University
A Quick Note • The recent generations of GPUs do seem to conform to the IEEE floating point standard • Nvidia site explicitely confirms this for Quadro FX • R. Fernando, Eurographics 2004 (general new FX) • Older generations do not (new generations?) • see paper GPU Floating Paranoia by Hillesland/Lastra, GPGPU Workshop 2004 • IEEE sets clear conventions for rounding • round to nearest representable number (basis 2) • when tie, round to nearest even number
Image Processing • Simple mask operations: • smoothing, edge detection, averaging, median • pixels under the mask are processed, combined, and written to output pixel fragment program input texture output texture
Image Processing • Very streamable • no data reuse • fragments in -> fragments out • highly applicable to video processing • GPU/CPU speedup results obtained by Payne et al. (GeForce FX 5900 vs. 2.8GHz P4): • for 3x3 convolution masks and 1k x 1k images: 30 • for 5x5 masks and 1k x 1k images: 50-60 speedup • for 2k x 2k images (double HDTV) get 15 fps, while CPU gets 1.8 fps
Multigrid Solver • Cliff Woolley’s presentation at the Graphics Hardware Workshop 2003 • “A Multigrid Solver for Boundary Value Problems Using Programmable Graphics Hardware” • paper by N. Goodnight, C. Woolley, G. Lewin, D. Luebke, and G. Humphries, U Virginia