Hi I'm Chenhan Yu (Chenhan D. Yu), a deep learning software engineer at NVIDIA. This site hosts my research results of my PhD study and my complete CV. You can find a full publication list in the CV page.
Please bare with my silly photos!
If you would like to copy or quote, please inform me. The website is built by Google site (Google view is the name now) using html and html5. The webpage is supposed to adjust itself depending on what device you are using to browse.
I'm recently interested in cross-platform high-performance machine learning primitives and scalable hierarchical matrix (H-matrix) approximation algorithms for scientific computing and machine learning tasks.
My latest work GOFMM (Geometry-Oblivious Fast Multipole Method) focuses on compressing generic SPD (Semi Symmetric Positive Definite) matrices using pure algebraic FMM (Fast Multipole Methods), which only requires matrix entries as input. With a log-linear time compression phase, approximate matrix-vector multiplication and matrix inverse can be computed in linear (or log-linear) time. GOFMM employs cross-platform N-body computation primitives and a self-contained task-base runtime system to achieve portable performance across different computing platforms.
The Science of High-Performance Computing (SHPC) Group
@ UT Austin
``Science is knowledge that has been made systematic. The Science of High-Performance Computing Group focuses on the knowledge that underlies scientific software libraries and makes it systematic. On the one hand this has facilitated the software architecture of such libraries. On the other hand it has allowed us to make much of the development of such software mechanical, via goal-oriented programming techniques from formal methods and Design-by-Transformation from software engineering.''' -- Podfather of SHPC.
NSF SPX group
Parallel Algorithms for Data Analysis and Simulation (PADAS) Group
@ UT Austin
``The mission of the Parallel Algorithms for Data Analysis and Simulation (PADAS) group is to integrate applied mathematics and computer science to design and deploy algorithms for grand challenge problems that scale to the largest supercomputing platforms. The group has developed high performance computing technologies for integral equations, fast multipole methods, multigrid solvers, octree data structures, fast solvers for inverse problems, and computational statistics algorithms.'' -- Boss of PADAS.
Site designed by Chenhan D. Yu