<aside> ❗ Headings and text with “▶” can be toggled to expand contained info. You can also see the Table of Contents as a collapsible sidebar on the right →

</aside>

Table of Contents

Introduction

This module provides guidance to perform CUDA programming and to use performance counters.

Objectives

• Be able to program and optimize programs.

Readings

Required

Optional

Notes / Things that stand out

Applications Suitable for GPUs

GPUs are good for tasks:

• Involving massive parallel data to take advantage of massive number of threads.
• Applications that have low dominance in host device communication costs are good for running on GPUs.
• Furthermore, applications that have coalesced data accesses in the global memory accesses.

Profiling

• Enables us to identify performance bottlenecks by identifying hotspots in our applications.
• It provides key performance metrics such as:
  • Reported throughput,
  • Number of divergent branches,
  • Number of divergent memory, both coalesced and uncoalesced,
  • Occupancy,
  • Memory bandwidth utilization.
• To achieve this, use profilers provided by GPU vendors to analyze and optimize our kernel functions.
• Profiling is also useful before you convert your applications for GPUs to identify the key hotspots in CPUs.