Tom Stitt
Tom Stitt
- Adds and renames the `display` and `update` methods - Moves `ReadInputStreams` from `glvis.cpp` to `lib/stream_state.*` - `keep_attr` is now a member of `StreamState`
Add basic support for external tpls and add guard for examples/regression tests
This PR adds support for HYPRE's runtime compute/memory policy. HYPRE's runtime policy requires at least HYPRE 2.26.0 and the lazy device library initialization requires https://github.com/hypre-space/hypre/pull/962, which is not yet in...
Switch `hipMemcpy` to `hipMemcpyAsync` for device to device copies to avoid unneeded host-side syncs; matches cuda's performant behavior. See https://github.com/mfem/mfem/pull/2780
Adds `mfem::device_reduce_buf_type` so that we can optional specify the `MemoryType` for the cuda/hip reduction buffers used in `{device}VectorMin` and `{device}VectorDot` | PR | Author | Editor | Reviewers | Assignment...
Adds managed Umpire host and device memory spaces (`ManagedUmpireHostMemorySpace` + `ManagedUmpireDeviceMemorySpace`) and factors out some common stuff to the base class | PR | Author | Editor | Reviewers |...
Zero-copies the reduction buffer if the hip architecture is `gfx94x`, otherwise uses `HOST_PINNED` + `DEVICE`. Alternative to https://github.com/mfem/mfem/pull/4228 | PR | Author | Editor | Reviewers | Assignment | Approval...
| PR | Author | Editor | Reviewers | Assignment | Approval | Merge| | --- | --- | --- | --- | --- | --- | --- | |...
Use HYPRE's SpMV since it is faster than rocSPARSE's. I've compared with ROCm 5.7 and 6.1. | PR | Author | Editor | Reviewers | Assignment | Approval | Merge|...
**Is your feature request related to a problem? Please describe.** The `RAJA::hip_multi_reduce_atomic` and `RAJA::cuda_multi_reduce_atomic` multireducers allocate GPU memory even when they are only used in a CPU kernel. Our application...