Optimizing Rodinia Benchmarks with Nvidia Libraries

This project aims to improve the performance of the Rodinia benchmarks by using Nvidia libraries such as cuSOLVER and cuBLAS.

Table of Contents

• Installation
• Usage
• File Structure

Installation

To use this project, you will need to install the following dependencies:

• Nvidia CUDA Toolkit: Follow the instructions here to install CUDA on your system.
• cuBLAS library: Follow the instructions here to install cuBLAS.
• cuSOLVER library: Follow the instructions here to install cuSOLVER.

You will also need a GPU with CUDA support.

Usage

To use the optimized benchmarks:

1. Run Gaussian with the desired benchmark and input parameters ./<filename> -s <matrix-dimension>

For example: ./gaussian -s 1024

2. Run lud with the desired benchmark and input parameters. ./<filename> -s <matrix-dimension>

For example: ./lud -s 20000.

3. Run combination with the desired benchmark and input parameters. ./<filename> -s <matrix-dimension>

For example: ./cusolver -s 20000.

Also in each folder, there is a corresponding run.sh file that is used to profile the running process for a series of inputs. You can run it using the ./run.sh command.

File Structure

This project mainly consists of the following files and directories:

• cuBLAS: This directory contains the source code for the project.
  • Compilation_Execution_Plot.ipynb: This jupyter notebook file contains compliation, execution and plotting of cu files using cuBLAS library.
  • gaussian:
    • common: This directory required to run the executables.
    • managedm: This directory contains implementation of gaussian using managed memory in gaussian_manegedm.cu.
    • pinnedm: This directory contains implementation of gaussian using pinned memory in gaussian_pinnedm.cu.
    • regular: This directory contains implementation of gaussian using regular memory in gaussian.cu.
    • mdt3_output.txt: This file contains the processed data of the profiling of gaussian_manegedm.cu.
    • pdt3_output.txt: This file contains the processed data of the profiling of gaussian_pinnedm.cu.
    • rdt3_output.txt: This file contains the processed data of the profiling of gaussian.cu.
  • lud/cuda:
    • lud_kernel.cu: This file contains kernel function.
    • lud.cu: This file contains main function.
• cuSOLVER: This directory contains the main function of the program.
  • getrf: cusolver_getrf_example.cu is the cu file using cuSOLVER library to implement both gaussian and lud algorithm.