Python Parallel Programming Cookbook

Using GPU-accelerated libraries with NumbaPro

NumbaPro provides a Python wrap for CUDA libraries for numerical computing. Each code using these libraries will get a significant speedup without writing any GPU-specific code. The libraries are explained as follows:

• cuBLAS: This is a library developed by NVIDIA that provides the main functions of linear algebra to run on a GPU. Like the Basic Linear Algebra Subprograms (BLAS) library that implements the functions of linear algebra on the CPU, the cuBLAS library classifies its functions into three levels:
  • Level 1: Vector operations
  • Level 2: Transactions between a matrix and vector
  • Level 3: Operations between matrices

  The division of these functions in the three levels is based on the number of nested loops ...