* relax MPI dependency from code
This commit (series) removes the strict dependency on MPI in code base.
rocSHMEM will still be compiled with MPI, but the goal is to make the
code work even if MPI_Init_thread has not been invoked, at least for
certain, well-defined scenarios. Hence, the goal is not remove any
mentioning of MPI from rocSHMEM, but to ensure correct execution of the
ipc conduit even if the library has been initialized using other means.
Details:
- add non-MPI version of remote_heap and WindowInfo classes
- host interfaces work on WindowInfoMPI, they will not work with the
non-MPI code path. Since it is unclear whether we plan to support the
host interfaces at all, this is probably not a major limitation.
* update symmetric_heap structures and backend
* first cut on initialization
and enabling non-MPI initialization of the IPCBackend
* add non-MPI hostInterface methods
at the moment, only barrier_all and sync_all are explicitely supported.
* add non-mpi version of ipc_policy
and a number of smaller fixes required in other files.
A small init/finalize test already passes now with the branch.
* add non-mpi team_split_strided code
* minor fixes for non-MPI use-case
* disable symmetric-heap-window-ionfo test
disable this test for now just to make the compilation pass. Will have
to rework it.
* make no-mpi great again
after rebasing on top of the MPI singleton changes.
* enable running functional tests with uuid init
to run the functional tests using rocshmem_init_attr and the uuid
mechanism requires
a) a PMIx installation on the system
b) setting the environment variable ROCSHMEM_TEST_UUID=1
* fix multi-team creation bug
fix a bug occuring when creating many teams, which was the result of
incorrectly applying two indices in our own implementation of Allreduce.
* make unit tests pass again
* reverse offload was impacted by code change
fix the RO conduit to cope wioth the non-MPI path introduced for the IPC
conduit.
* update to cmake logic to find pmix
* Update src/memory/window_info.hpp
Co-authored-by: Yiltan <ytemucin@amd.com>
* Update CMakeLists.txt
Co-authored-by: Yiltan <ytemucin@amd.com>
* document ROCSHMEM_UNIQUEID_NO_MPI
* rename env. variable to UNIQUEID_WITH_MPI
* update host.cpp to use USE_HDP_FLUSH macro
instead of the deprecated USE_COHERENT_HEAP.
* add note for running example with RO conduit
add a note clarifying that running init_attr_test from the example
directory requires setting an additional environment variable with the
RO conduit.
* Find PMIx in more cases, only apply pmix build options to the test that
needs it, if OMPI_COMM_WORLD_LOCA_RANK is not setenv, abort
---------
Co-authored-by: Yiltan <ytemucin@amd.com>
Co-authored-by: Aurelien Bouteiller <abouteil@amd.com>
* Refactor `Barrier_all` and `Sync_all` to use default context
- Removed context-specific implementations of barrier_all and sync_all
- Added barrier_all and sync_all to the default context implementation
- Updated functional tests to use the default context for barrier_all and sync_all
* Update `Barrier_all` and `Sync_all` API usage in documentation
* Update `CHANGELOG`
---------
Co-authored-by: Yiltan <ytemucin@amd.com>
* Revert "SWDEV-536571 - Include assert header. (#157)"
This reverts commit bcc14b1a34.
* Fix use of assert/abort and required includes
* Disable IPC AMO testers for non-implemented functions
* Use FineGrained allocator for heap by default, consolidate all types of
allocators under saner cmake controls
Co-authored-by: Yiltan <ytemucin@amd.com>
* Uncached may not be only for debug
Need to include the rocshmem config otherwise produce an inconsistent
build with different allocators used in different files
* Undo this pr adding presumably useless hip_host_allocator_noncoherent
* Rename HEAP_IS_COHERENT/USE_COHERENT_HEAP to USE_HDP_FLUSH as the former
was misleading
* Remove unused __roc_inv()
---------
Co-authored-by: Yiltan <ytemucin@amd.com>
* Use a single printf per line (reduce chances of lines being cut in logs)
* team_comm can be an int or a pointer depending on MPI impl.
Received is confusing (since we are on the origin), use submitted
instead
* Print arguments to calls when using DEBUG
---------
Signed-off-by: Aurelien Bouteiller <abouteil@amd.com>
* Add dlmalloc_strat allocator strategy
- Use mspace variant to ease encapsulation
- Make pow2bins and dlmalloc cmake selectable
* Add unit tester for dlmalloc, rework single_heap, pow2bins unit testers
accordingly
- add dlmalloc get_used/get_avail, and have all strats allocators also have a get_used
- Rework memallocator unit tests: bin size is per strat, alignment is verified in singleheap
* bugfix: dlmalloc exposed that the pingpong test would write past end of
allocation with -w 32
* iostream leakage/mixed usage of cerr and fprintf(stderr
---------
Signed-off-by: Aurelien Bouteiller <aurelien.bouteiller@amd.com>
* unify handling of env variables
create a class containing all (most?) environment variables used by rocshmem and an object that is instatiated
before library_init, since some of the environment variables need to be
set before we start the bootstraping process.
This allows us to remove two files from the bootstrap directory.
* replace INFO and TRACE macros with DPRINTF
to be more consistent with the rest of the rocSHMEM code
* Remove dev_mono_linear (followup to removal of slab_heap)
* cleanup: use CHECK_HIP rather than ad-hoc error checking
---------
Signed-off-by: Aurelien Bouteiller <aurelien.bouteiller@amd.com>
* add code for bootstrapping
the bootstrapping code has been extracted from the MSCCLPP library,
which in parts is based on the code from NVIDIA. The code has been
modified to match the specific requirements of the rocSHMEM library.
* add code to use the new uniqueId bootstrapping
* adjust init_attr example
extend the rocshmem_init_attr example to use two disjoint groups
of processe, in order to trigger the new code path.
* add env variable for bootstrap timeout
* Update examples/rocshmem_init_attr_test.cc
Co-authored-by: Aurelien Bouteiller <Aurelien.bouteiller@gmail.com>
* Update src/rocshmem.cpp
Co-authored-by: Aurelien Bouteiller <Aurelien.bouteiller@gmail.com>
---------
Co-authored-by: Aurelien Bouteiller <Aurelien.bouteiller@gmail.com>
* Update backend to use provided MPI communicator during library initialization, default to `MPI_COMM_WORLD`
* Update `rocshmem_my_pe` and `rocshmem_n_pes` host APIs
- Return values from backend if initialized; otherwise, fallback to MPI_Singleton.
* Remove unused forward_list
* Remove unused __read_clock function
* Replace wallClk code with hip function
* Remove unused unit test for ipc
* Remove slab heap
* Remove unused EBO spinlock
* Update the naming convention for collective APIs to ensure consistency across the interface.
* Move all collective API declarations to rocshmem_COLL.hpp
* The following APIs were updated as part of this change:
- `barrier`
- `barrier_all`
- `sync`
- `sync_all`
- `all_to_all`
- `broadcast`
- `fcollect`
- `all_reduce`
* Update header file generation code for collective APIs
* Add thread, wavefront, and workgroup-level `barrier` APIs in IPC and RO conduits; remove collectives on default context
- Implemented `barrier` APIs for thread, wavefront, and workgroup scopes
- Added support into both IPC and RO conduits
- Added functional tests to cover all `barrier` APIs
- Removed collective operations on default context
* Add thread, wavefront, and workgroup-level `sync` APIs in IPC and RO conduits.
- Implemented `sync` APIs for thread, wavefront, and workgroup scopes
- Added support into both IPC and RO conduits
- Added functional tests to cover all `sync` APIs
* update naming convention for context-based `barrier` APIs
* Fix deadlock in `rocshmem_ctx_wg_barrier_all` API in IPC conduit by adding per-context pSync buffers and context IDs
- Added separate pSync buffers for each device context
- Resolved deadlock when invoking barrier API (`rocshmem_ctx_wg_barrier_all`) concurrently from multiple contexts
* Update barrier_all functional tests for multi-context support
* Add thread, wavefront, and workgroup-level barrier_all APIs in IPC and RO conduits
- Implemented barrier_all APIs at thread, wavefront, and workgroup granularity
- Added support in both IPC and RO conduits
- Updated functional tests to cover all `barrier_all` APIs
* Add thread, wavefront, and workgroup-level sync_all APIs in IPC and RO conduits
- Implemented sync_all APIs for thread, wavefront, and workgroup scopes
- Added support into both IPC and RO conduits
- Added functional tests to cover all `sync_all` APIs
add the interfaces required to support rocshmem initialization
through the uniqueID mechanism. At the moment this still maps to
MPI initialization underneath the hood, but adding the functions might
simplify the porting of some applications to rocshmem. In addition, if
we need to transition away from MPI one day, this is also one step into
this direction.
This PR addresses two issues:
- reduce the number of contexts supported by the host-interface by
default to 1, we are not using those at the moment, and hence
we now create fewer MPI_Win at the startup
- introduces a micro-sleep in RO progress engine in case there are no
pending requests. This leads significant performance improvements
observed for inter-node communication with THor2 NICs.
* Allocate default context buffers and initialize queue for management
- Allocated the status flag, g return, and atomic return buffers for
the default context.
- Initialized `AtomicWFQueueProxy` instances to manage these buffers
efficiently for concurrent access.
* Update `BlockHandle` with default context buffers
* Add default context flag and update buffer retrieval functions
- Added a flag to distinguish the default context from other contexts.
- Modified return buffer functionns and `get_status_flag` function to accommodate
the default context
* Add default context primitive tests
- get, put, get_nbi, put_nbi, g, and p APIs.
* feat: Add AtomicWFQueue implementation
- Implemented wavefront-safe atomic FIFO queue ensuring first-come, first-serve order
- Added efficient synchronization using atomics
- Enhanced `dequeue` to wait until an element is available
* test: Add GTest for AtomicWFQueue
- Implemented unit tests for AtomicWFQueue using GoogleTest framework
- Added tests for `enqueue`, `dequeue`, and edge cases
- Ensured synchronization behavior and correctness under concurrent conditions
* Add assert in `enqueue` and update atomics
- Added an assert in the `enqueue` function to ensure it fails if the queue is full
* Update HIP version check for compatibility with versions >= 5.5
* Update memory allocator for context BlockHandle
- Replaced `HIPAllocator` with `HIPDefaultFinegrainedAllocator` for context `BlockHandle`.
* Update run commands for `rocshmem_g` and `rocshmem_p` functional tests
* add team-barrier implementation
add a team-barrier API and implementation in the IPC and RO conduit.
Clean up some of the logic in the RO Conduit to distinguish between
sync, sync_all, barrier, and barrier_all.
* add team_barrier_tests to functional tests
* RO/collectives: add linear algorithms using RPut/Rget
- make broadcast, alltoall and fcollect use a simple linear algorithm
using MPI_RPut/Rget, but without blocking in the execution
- remove the to_all interfaces, since they have been deprecated.
- remove the active-set interfaces, since they have been removed from
rocSHMEM
* avoid notification after barrier
Co-authored-by: Avinash Kethineedi <avinash.kethineedi@amd.com>
* disable allocation of ata_buffer
a temporary buffer of 128MB was allocated when creating a team. In
previous versions of the code, that buffer was used by some collective
operations. This is not the case for now. Therefore, do not allocate the
buffer for now. I am not removing the element itself from teh
structure, since we might need it in future versions again.
---------
Co-authored-by: Avinash Kethineedi <avinash.kethineedi@amd.com>
- Remove hdp and ipc pointers from BlockHandle, align RO stats with RO contexts
- Add run commands for `rocshmem_g` and `rocshmem_p` API tests in driver.sh
- Allocate rocshmem API return buffers based on number of device contexts.
- Associate status flag address with blocking calls and remove threadId dependency
- Associated the status flag address with each blocking call request to notify the GPU thread.
- Removed dependency on threadId for determining the appropriate status flag index.
- Move status flag buffer allocation to backend.
- Initialize allocated memeory to zero
* Update(DeviceProxy): Dynamically Determine Memory Allocation Size & Remove Compile-Time size Calculations
- Modified the Device proxy class to determine memory allocation size at runtime.
- Updated all classes that include the Device proxy to use dynamic memory allocation.
- Removed compile-time memory size calculations.
- Ensured the allocated number of backend queue data structures matches the number of RO device contexts.
Edgar Gabriel
Avinash Kethineedi
Aurelien Bouteiller
Yiltan
Jobbins
Brandon Potter
Yiltan Hassan Temucin