Repository navigation

Device ghost-exchange plans, fused block exchanges, single-reduction PCG (CUDA-aware MPI) - #104

Merged

spossann merged 16 commits into

fix-cuda-aware-mpi-issues

Oct 24, 2026

Merged

Device ghost-exchange plans, fused block exchanges, single-reduction PCG (CUDA-aware MPI)#104
spossann merged 16 commits into
devel-tinyfrom
fix-cuda-aware-mpi-issues

Commits on Oct 13, 2026
- fixes
  max-models
  committed
  a8d5338
- Merge branch 'devel-tiny' into fix-cuda-aware-mpi-issues
  max-models
  committed
  81bab4c
Commits on Oct 14, 2026
- Build device exchange plans from per-region indices
  max-models
  committed
  0ab90db
- BlockLinearOperator.dot: write the first block of each row directly
  max-models
  committed
  7664ff8
- stencil_dot_3d (CUDA): 8 threads per row, coalesced reads
  max-models
  committed
  c10b33b
- MPI improvements
  max-models
  committed
  6571ebd
- Merge branch 'devel-tiny' into fix-cuda-aware-mpi-issues
  max-models
  committed
  518eab2
Commits on Oct 15, 2026
- Optimization
  max-models
  committed
  6115208
Commits on Oct 16, 2026
- Use the stencil kernels directly (the CachedLaunch module was never committed)
  
  max-models
  and
  claude
  committed
  fcd8c8a
- Device paths: no raw kernels on the fake CuPy; PCG with one host copy per iteration again
  
  max-models
  and
  claude
  committed
  8ae781d
- stencil_dot_3d on the fake CuPy: one thread per entry, no warp shuffles
  
  max-models
  and
  claude
  committed
  b70cb6c
- CachedLaunch: reuse the prepared launch of a kernel for the same buffers
  
  max-models
  and
  claude
  committed
  1d8d47c
Commits on Oct 23, 2026