Best practices for OpenMP

[lucaparisi91]

• Use all the directives ( teams ( more than 1 SIMD processor ) , parallel ( use more than one thread. For the Cray fortran compiler you also need the SIMD directive [ to check on Archer2 ] )
• make sure to have enough parallelism ( O(1000) iterations )
• combine teams distribute parallel for if possible. Increasingly difficult if not in the same function and not inline, and difficult to use if in different compilation unites.
• use environment variables to track data movement ( CRAY_ACC_DEBUG, NVCOMPILER_ACC_NOTIFY , GOMP_DEBUG, LIBOMP_TARGET_INFO, -Minfo=mp,accel , -Rpass=openmp-opt )
• prefers firstprivate for scalar variable( usually sent as a kernel parameter ), private will endup in global or local memory
• merge parallel regions
• use tasks and dependencies
• use omp assumptions ( 5.1 )

[lucaparisi91]

Resources:

• https://www.youtube.com/watch?v=ongYcSq5zH0&list=PLDPdSvR_5-GhR1OWQW3ovK_qGg09tmNuq&index=27