Skip to content

Latest commit

 

History

History
107 lines (87 loc) · 2.8 KB

outline.markdown

File metadata and controls

107 lines (87 loc) · 2.8 KB
Raw

OpenACC Porting and Best Practices Guide

  1. What is OpenACC?

High level overview of compiler directives

  • Compiler directives basics
  • Benefits and Tradeoffs to compiler directives
  • Abstract accelerator model
    • Multiple Levels of Parallelism
    • Memory hierarchies
  • Should I say something here about the relevance toward OpenMP target directives?
  1. Accelerating an application

Describe the porting process at a high level including APOD

  • Assess the code to identify parallelism
  • Parallelize the code with directives
  • Optimize data locality (Optimize, part 1)
  • Optimize loops (Optimize, part 2)
  • Return to step 1 (Deploy?)
  1. Assess

Where to start

  • Generating a CPU profile using common tools
    • What tools can we point to here?
  • Discuss coarse vs. fine-grained parallelism
  • Utilizing existing OpenMP directives?
  1. Parallelize

Moving computation to the GPU

  • PARALLEL and KERNELS regions
  • LOOP directive
    • private clause
    • reduction clause
  • ROUTINE directive
  1. Optimize Data Locality

Improving data movement

  • Introduction to data regions
  • Data clauses
  • Introduction to unstructured data lifetimes
    • Show usage in C++ classes
  • CACHE directive
  • Asynchronous overlapping
  • Dealing with global data (is this needed?)
  1. Optimize Loops

Loop-level optimizations that make a difference

  • Common loop transformations
  • COLLAPSE directive
  • TILE directive
    • I will leave this out unless someone can provide me a case where it's beneficial
  1. Deploy

What more is there to say at this point?

  1. Interoperability

How to use OpenACC with math libraries and CUDA

  • Reuse examples from my blog post
  1. Interacting with MPI

Expand on previous chapter with specifics around MPI

  • Reuse Jiri's material
  1. Writing Portable OpenACC

Tips on writing code with portability in mind. I'm not sure exactly what will go here, but it seems important to me.

A. Appendix - OpenACC for XX programmers

Tips for understanding OpenACC if you already have understanding of another parallel programming paradigm.

  • OpenACC for CUDA programmers
  • OpenACC for OpenCL programmers
  • OpenACC for OpenMP programmers

NOTES:

  • What tools can/should I show?
    • pgprof/gprof for CPU profiling
      • Identifies hotspots, but not loop info
    • CUDA visual profiler/Nsight
      • Good for data motion insights
      • Good for prioritizing kernels
      • Poor for guiding how to optimize loops
    • Vampir/Tau?
      • CPU/GPU/MPI profiling in 1 tool
    • DDT/Totalview?
      • Commercial products, limited OpenACC support