Home

Main Topics

Basic Functionality

• Parallelisation over SMs ( teams )
• Parallelisation over warps ( parallel )
• Data Transfers ( map directives )
• #pragma omp declare

Advanced Topics

• Custom mappers ( #1 )
  • mapping of a class containing pointers
• Custom memory allocators ( #4 )
  • Memory allocations allocation per thread ( firstprivate) vs shared variables ( may live in shared memory/global memory/local memory vs being passed as a kernel argument )
  • Pinned memory allocation on CPU
  • shared memory allocations
• Concurrency
  • submit kernels from multiple threads (#3). Demonstrates using different cudaStreams/hipStreams.
  • use openmp tasking (#7). Demonstrates overlapping memory transfers and execution or multiple small kernels which might be difficult to merge in one bigger kernel.
• Interoperability ( #2 )
  • Dementrate how to use cuda with a variable mapped from openmp and how to use a variable allocated from cuda in openmp.
  • example of using cuFFT ( or any cuda/rocm numerical library ) together with openmp

Performance

• OPenMP generic( CPU ) mode
• Occupancy
• Memory bandwith ( global/shared ), roofline plot
• Coalesced access ( global memory )
• Bank Conflicts ( shared memory )

Exercices

Jacobi tutorial

• Naive implementation [ teams only, teams + parallel ]
• Add mapping directives to control memory transfers
• Use a custom mapper for transferring the whole class
• Split up in subdomains for overlapping transfer and computation ( use streams )
• Use shared memory to improve performance bandwidth ( custom memory allocators ? )