Research Activities for the year 2016


Robust Control Design with Parallel Quantitative Feedback Theory using GPGPU
    Harsh Purohit, Prof P S V Nataraj, IIT Bombay

Abstract: ​ We will describe an efficient method to compute the value set of any parametric uncertain system which can be modeled into input output transfer function. This value set of frequency response can be obtained by the calculation of the magnitudes and phases for all possible uncertain transfer functions. In robotics and autonomous machines application, these are very useful for controller synthesis. This computation is independent to each other and can be executed in parallel. To show the effectiveness of the proposed method, we have chosen 3­DOF longitudinal aircraft model with large number of parameters.

● BREMICS : Parallelization for Multicore and SIMD Architectures
Mandar Datar, Kulshreshth Dhiman, Prof. Sachin B Patkar, IIT Bombay

Abstract: ​BREMICS is Gauss­Seidal­Newton/Gauss­Jacobi­Newton method based circuit simulator which is highly parallelizable. It can benefit from multi­core (such as Intel Xeon Phi) as well as SIMD parallel hardware (such as GPUs). In BREMICS algorithm, significant time is spent in device model evaluations. Complex device models such as BSIM3 need complex kernel for model evaluation. So, we have built “Look­up­table (LUT) interpolation” based device models. With such LUTs, model evaluation kernel is smaller, and divergence is less too. We have developed GPU version of BREMICS, based on cubic­spline interpolator, which shows 11X speedup compared to serial software version. We also intend to compare
GPU performance and utilization over other approaches like CPU­FPGA and Multicore. We are further optimizing the CUDA kernel and the LUT based model­evaluation kernel, by precomputing the interpolating polynomial for each LUT cell.

Source Code Libraries

A GPU based Library for Bernstein approach to polynomial global optimization ​by Priyadarshan Dhabe, Prof P S V Nataraj, IIT Bombay
Version 1 of this library contains the following important functions:
● GPU parallel implementation of matrix method for explicit computation of all the Bernstein Coefficients (BCs) of a multi­variate polynomial.
● Computing minimum BC of a polynomial on a given domain along with its index on GPU.
● Computing two child Bernstein patches on GPU
● GPU implementation of basic Bernstein Algorithm for polynomial global optimization using maximum width direction as direction of subdivision and mid point as point of subdivision.
● GPU parallel implementation of explicit computation of all the BC's using Smith's implicit Bernstein form (IBF).
● GPU parallel implementation of Bernstein algorithm using IBF.
● GPU parallel implementation of computation of all the BCs using proposed "Modified IBF" (MIBF)
● GPU parallel implementation of Bernstein Algorithm using IBF and MIBF
● GPU parallel implementation of Modified matrix method for explicit computation of all the Bernstein Coefficients (BCs) of a multi­variate polynomial.

Poster Publications

Parallel computation of the value set of frequency response for uncertain systems with GPGPU ​at GPU Technology Conference, 2016,
USA by Harsh Purohit and P S V Nataraj.