Behrooz Zarebavani
I am a fourth-year Computer Science Ph.D. student, focusing on the acceleration of numerical methods used in computer graphics.
About Me
Hello! My name is Behrooz. I don't enjoy waiting, and I love animations! So, I'm focused on speeding up computationally intensive algorithms used in computer graphics. I prefer my research to be directly applicable to real-world problems, so I'm doing my best to develop tools that practitioners can easily use.
I am looking for internship position! Please let me know if we can help each other.
Contact
Research
Physics-based Simulation / Numerical Methods / Optimizer / HPC
Parth
A new tool that integrates with state-of-the-art sparse solvers such as Intel MKL, Apple Accelerate, and CHOLMOD. It significantly enhances the end-to-end performance of these tools, leading to a speedup of up to 3x when applied within complex physics-based simulations involving contact.
The code will be available after the SIGGRAPH 2024 evaluation.
Sparse Matrix Computation / Scheduler / Numerical Acceleration
HDagg
It is an open-source scheduler that accelerates sparse kernel computations with loop-carried dependencies. It optimizes computation sequences based on sparsity patterns, kernel specifics, and hardware type. HDagg's precise adjustments of load-balance, locality and synchronization provide significant efficiency, outperforming current advanced kernels implemented in MKL such as Sprase Triangular Solver and Incomplete Cholesky Factorization by up to 13x speedup.
Github link for the code base and publication: HDagg
GPU programming / Causal Structure Discovery / Bayesian Network / HPC
cuPC
This innovative algorithm offers an efficient implementation of the Peter-Clark (PC) algorithm. This solution provide a fast and efficient method towards uncovering causal relationships in observational data and significantly surpasses the performance of previous methods. cuPC represents the first GPU deployment of this algorithm, which has effectively reduced the runtime from 11 hours to a mere 4 seconds on challenging dataset.
Github link for the code base and publication: cuPC