Device op cuda#46

Open

devreal wants to merge 76 commits intoICLDisco:mainfrom

devreal:device_op_cuda

Commits on Nov 4, 2023

spack:fix for dlopen missing symbol problem
hppritcha
committed

Commits on Nov 7, 2023

Merge pull request #12050 from hppritcha/fix_for_slack_ompi500_dlopen
hppritcha
authored
Initial draft of CUDA device support for ops

Joseph Schuchart
authored and
devreal
committed
First working version of CUDA op support

Joseph Schuchart
authored and
devreal
committed
Update copyright header

Joseph Schuchart
authored and
devreal
committed
Fix minor bugs to get osu_allreduce working

Joseph Schuchart
authored and
devreal
committed
cuMemAllocAsync is supported since CUDA 11.2.0

Joseph Schuchart
authored and
devreal
committed
coll/base/allreduce: Condition device allocation on op/dtype support

Joseph Schuchart
authored and
devreal
committed
Make sure the device op callbacks are zero-initialized

Joseph Schuchart
authored and
devreal
committed
Be more graceful when creating a context and stream

Joseph Schuchart
authored and
devreal
committed
fix wrong call to memset
devreal
committed
Add detector for cudart
devreal
committed
Add CUDA stream-based allocator and memory pools
devreal
committed
Don't memset the CUDA op component, we need the version
devreal
committed
Set the memory pool release threshold
devreal
committed
Implement device-compatible allocator to cache coll temporaries
devreal
committed
Fix devicebucket allocator for larger sizes
devreal
committed
Fix the RDMA fallback protocol selection.

bosilca
authored and
devreal
committed
Stream-based reduction and ddt copy and 3buff cuda kernels, adopted for allreduce recursive doubling
devreal
committed
Remove extra copies from allreduce redscat and ring
devreal
committed
Allow ops and memcpy on managed memory from the host
devreal
committed
reduce_local: add support for device memory
devreal
committed
Draft of ompi_op_select_device
devreal
committed
Second draft of ompi_op_select_device
devreal
committed
Fix undefined symbols in cuda op component
devreal
committed
Fix off-by-one error in device-bucket allocator
devreal
committed
Heuristic to select op device based on element count
devreal
committed
init op_rocm, not compilable yet

Phuong Nguyen
authored and
devreal
committed
implemented funcs in accelerator_rocm modules

Phuong Nguyen
authored and
devreal
committed
add -I include path to Makefile

Phuong Nguyen
authored and
devreal
committed
added rocm codes into test example

Phuong Nguyen
authored and
devreal
committed
fixed kernel launches in hip

Phuong Nguyen
authored and
devreal
committed
Make headers in reduce_local better parsable
devreal
committed
CUDA: disable internal memory pool (seems broken)
devreal
committed
Op: minor comment correction
devreal
committed
Reduce_local: set hip device during init
devreal
committed
CUDA accelerator: fix compiler warnings
devreal
committed
Device op: pass device to lower-level op to avoid recurring queries
devreal
committed
CUDA/ROCm: Fix vectorized ops and rocm integration
devreal
committed
Reduce_local: use OPAL defines to detect device support
devreal
committed
CUDA op: fix vectorized ops
devreal
committed
Reduce: add vectors to cuda implementation
devreal
committed
Allreduce: cleanup and minor fixes
devreal
committed
Add MCA op_[cuda|rocm]_max_num_[blocks|threads]
devreal
committed
Fix the generation of "unsigned char" ops.

bosilca
authored and
devreal
committed
We need CXX17 for the CUDA ops.

bosilca
authored and
devreal
committed
ROCM: add vectorization of some basic ops
devreal
committed
Device allocators: correctly handle non-zero ID single accelerator
devreal
committed
CUDA op: consistently name unsigned_long functions as ulong
devreal
committed
ROCM op: remove debug output
devreal
committed
Reduce_local test: correctly test for OPAL_CUDA_SUPPORT and OPAL_ROCM_SUPPORT
devreal
committed
More unsigned_long -> ulong fixes in CUDA and ROCm op
devreal
committed
Fix type in ulong conversion
devreal
committed
Reduce_local: access only host-side memory in error message
devreal
committed
Make sure CUDA accelerator is initialized before querying number of devices
devreal
committed
Accelerator: provide peak bandwidth estimate
devreal
committed
accelerator/rocm: regular memory behaves like unified memory
devreal
committed
ROCM: add missing FUNC_FUNC_FN macro
devreal
committed
opal_datatype_accelerator_memcpy: determine device copy type
devreal
committed
accelerator rocm: fix global memcpy stream variable
devreal
committed
Thread base: fix missing include file
devreal
committed
Accelerator: Remove debug output
devreal
committed
Allreduce: don't copy inputs if data can be accessed from the host
devreal
committed
Be more careful when releasing temporary receive buffers
devreal
committed
Remove debug output and dead code
devreal
committed
Bump max devicebucket allocator max size to 1GB
devreal
committed
accelerator/cuda: fix error message
devreal
committed
CUDA: Select compute capability 52 by default
devreal
committed
Sqash const correctness warnings
devreal
committed
Squash warnings about mismatched function pointer types
devreal
committed
Squash printfs
devreal
committed
Replace fprintf with show_help
devreal
committed
Squash compiler warnings
devreal
committed
Clean up cuda and rocm op codes
devreal
committed
Minor tweak to CUDA op configury
devreal
committed

Commits on Nov 8, 2023

Fix rebase errors
devreal
committed