PyPacho: A Python library that implements parallel basic operations on GPUs

Python has become an extremely popular programming language that is widely used in many different domains including scientific computing. Numpy is a Python library that includes numerous functions for numerical linear algebra. Numpy uses code that has been parallelized to take advantage of microproc...

Full description

Saved in:
Bibliographic Details
Main Authors: Arcila-Moreno, Juan D., Cifuentes Garcia, Diego Alejandro, Correa Zabala, Francisco Jose, Jaramillo, Esteban Echeverri, Trefftz, Christian, Zapata-Palacio, Andres Felipe
Format: Conference Proceeding
Language:English
Subjects:
Codes
Graphical Processing Units
Jacobian matrices
Linear Algebra
Microcomputers
Microprocessors
Numerical Methods
Parallel processing
Parallel programming
Python
Scientific computing
Software Libraries
Online Access:Request full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Python has become an extremely popular programming language that is widely used in many different domains including scientific computing. Numpy is a Python library that includes numerous functions for numerical linear algebra. Numpy uses code that has been parallelized to take advantage of microprocessors with several cores making Numpy very fast and efficient on current microprocessors. Graphical Processing Units (GPUs) are available on many computers. Originally designed to accelerate the performance of graphical applications, they have become very useful to improve the speed of general purpose applications. In particular, GPUs work very well on large arrays. In this paper we present PyPacho, a library that is, by design, similar to Numpy. This library has been written to accelerate the execution of code on GPUs by using the services provided by PyCuda and OpenCL. Our goal is to create a library that will allow Python code originally written using Numpy methods to execute efficiently on GPUs. Our main strategy is to take advantage of the use of parallelism in basic operations on matrices and vectors. Furthermore, we implemented three different methods to solve systems of linear equations: Jacobi, Gradient Descent and Conjugate Gradient. We compared the execution times of those methods on different platforms. Initial results are encouraging.
ISSN:2644-3163
DOI:10.1109/IEMCON53756.2021.9623197