Skip to content

merge aomp-dev into amd-staging#2166

Closed
z1-cciauto wants to merge 4 commits intoamd-stagingfrom
upstream_merge_202604162306
Closed

merge aomp-dev into amd-staging#2166
z1-cciauto wants to merge 4 commits intoamd-stagingfrom
upstream_merge_202604162306

Conversation

@z1-cciauto
Copy link
Copy Markdown
Collaborator

No description provided.

saiislam and others added 4 commits April 16, 2026 08:51
#2140)

ROCm's GPU-aware MPI stack components:
- XPMEM:
  - library for cross-process memory access
  - version: 2.7.4
  - source: https://github.com/openucx/xpmem
- UCX:
  - Communication framework for high-bandwidth, low-latency communication
  - version: 1.20.0
  - source: https://github.com/openucx/ucx
  - dependencies:
    - XPMEM
    - ROCm
- UCC:
  - Collective communication library for high-performance communication
  - version: 1.6.0
  - source: https://github.com/openucx/ucc
  - dependencies:
    - UCX
    - ROCm
- OpenMPI build with ROCm support:
  - version: 5.0.9
  - dependencies:
    - UCX
    - UCC
    - ROCm

Installation:
- Install GPU-aware ROCm OpenMPI with all dependencies:
  - ./build_supp.sh rocmopenmpi
- Install XPMEM
  - ./build_supp.sh xpmem
- Install UCX:
  - ./build_supp.sh ucx
- Install UCC:
  - ./build_supp.sh ucc

Default MCA parameters in $AOMP_SUPP/rocmopenmpi/etc/openmpi-mca-params.conf:
  pml = ucx
  osc = ucx
  coll_ucc_enable = 1
  coll_ucc_priority = 100
Build Options:
  download    Download OSU Micro-Benchmarks tarball
  build       Build OSU Micro-Benchmarks with ROCm support
  install     Download and build (full installation)

Benchmark Options:
  quick       Run quick sanity tests (default)
  full        Run full benchmark suite
  bandwidth   Run bandwidth-focused benchmarks
  latency     Run latency-focused benchmarks
  collective  Run collective operation benchmarks
@ronlieb ronlieb closed this Apr 17, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

4 participants