Practical work with OpenACC

1. Download the move example and the scalar product example
• unzip the files in your account on plexi
• tar xzf T1.tgz
• change into the directory (T1 or skalar)
• cd T1
• make run
  
2. Change the  data clauses  in the files and watch out for changes wrt. run time or results
• What is the difference between copy, copyin, present ?
• How is this related to input, output and global variables in the classical subroutine programming?
3. Take your shared memory codes from module 5 (exercise sheet, description of A-D) and convert them into OpenACC
• mkdir matvec
• copy source files from shm-version into your directory
• copy Makefile and file david.sh from directory OpenAcc/skalar into your directory
• change codes into C (header files, cout --> printf)
• make
  
• Add OpenACC pragma directives
• make run
  
• Check the performance

• General Intro, Quick Reference Card (update: Makefile with profiling)
• What to read first:

• Michael Wolfe: Getting started guide, kernels vs. parallel, web-seminar (1,2)
• Rob Farber: First steps, profiling and improving (incl. OpenACC on shm-host); Phi vs. CUDA
•                     OpenCL vs. CUDA vs. OpenACC (video)
• Mark Harris; Intro, Part1 , Part2
• More: Maximizing Performance
• PGI-Documentation (profiler)

• PGI: 
• compilers: pgcc, pgfortran (OpenACC is not supported for C++)
• see §7 in manual for OpenACC directives. 
• compiler options:  pgcc -help
• Profiling: pgcollect, pgprof
• Debugger: pgdbg
• Info in Accelerator: pgaccelinfo