GPU activities at FI MUNI and their results Ji Matela, Ji Filipovi - PowerPoint PPT Presentation

GPU activities at FI MUNI and their results Jiří Matela, Jiří Filipovič <matela@ics.muni.cz> , <fila@ics.muni.cz> Laboratory of Advanced Network Technologies MetaCentrum CESNET Grid Computing Seminar 2010 Praha, 2010–10–15 1/11

AutoGrid • Potential maps generation for molecular docking • The most computationally expensive parts accelerated on GPU • CPU part analyzed and modified 2/11

AutoGrid – Speedup Accelerated design shows speedup of up to 400 × 450 disntance-dependent constant 400 350 300 speedup 250 200 150 100 50 0 0 50 100 150 200 250 300 350 grid size 3/11

Discrete Wavelet Transform (DWT) • Digital signal processing technique • Application in diverse areas • digital speech recognition • multi-resolution video processing • data compression 4/11

DWT – Speedup Our GPU implementation shows 68 × speedup 5/11

Real-Time Video and Fast Large-Scale Image Compression • Ongoing project • Real-time compression and transmission of video in HD post-HD resolutions • Fast compression of pathological images of resolutions in order of gigapixels • GPU acceleration of JPEG2000 6/11

JPEG2000 compression process Data compression Context Arithmetic DWT Modeling Encoding 7/11

Context-Modeling in JPEG2000 • Serial algorithm • Redesign to fit specifics of GPUs • 12 × faster compared to JasPer CPU implementation 8/11

Soft tissues simulations • Haptic surgical simulators • Simulations modelled using Finite Element Method (FEM) • FEM discretizes the modeled object as a mesh of elements • Per element computation and system of equations solving • Per element computation is complex problem so that it needs to be decomposed into several GPU functions • Not easy to choose decomposition granularity • Manual development of as small functions as possible • Automatic fusion into lager functions 9/11

Preliminary performance gain of fusion 80% gain compared to non fused approach GPU SMEM/GMEM GPU GMEM 10000 CPU 1 core 8000 thousands elements/s 6000 4000 2000 0 0 10000 20000 30000 40000 50000 60000 # elements 10/11

Thank you for you attention! Q?/A! <matela@ics.muni.cz> , <fila@ics.muni.cz> http://www.sitola.cz/ 11/11