Yiltan
f693c98fb2
* Initial ROCm-docs commit Co-authored-by: Aurélien Bouteiller <bouteill@icl.utk.edu> Co-authored-by: Alex Xu <alex.xu@amd.com> Co-authored-by: yugang-amd <yugang.wang@amd.com>
55 行
2.4 KiB
ReStructuredText
55 行
2.4 KiB
ReStructuredText
.. meta::
|
|
:description: rocSHMEM intra-kernel networking runtime for AMD dGPUs on the ROCm platform.
|
|
:keywords: rocSHMEM, API, ROCm, documentation, HIP, Networking, Communication
|
|
|
|
.. _rocshmem-introduction:
|
|
|
|
---------------------------
|
|
What is rocSHMEM?
|
|
---------------------------
|
|
|
|
The ROCm OpenSHMEM (rocSHMEM) runtime is part of an AMD and AMD Research initiative
|
|
to provide GPU-centric networking through an OpenSHMEM-like interface.
|
|
This intra-kernel networking library simplifies application code complexity and
|
|
enables more fine-grained communication/computation overlap
|
|
than traditional host-driven networking.
|
|
rocSHMEM uses a single symmetric heap (SHEAP) that is allocated on GPU memories.
|
|
|
|
The code is open and hosted at `<https://github.com/ROCm/rocSHMEM>`_.
|
|
|
|
The rocSHMEM Programming Model
|
|
-------------------------------
|
|
|
|
Defining how OpenSHMEM applications interact with GPUs remains an
|
|
ongoing active discussion within the OpenSHMEM community, and the OpenSHMEM
|
|
specification has yet to coalesce on this topic.
|
|
rocSHMEM extends beyond the OpenSHMEM specification to add semantic that
|
|
support GPU kernel communication, while maintaining close resemblance to
|
|
the original OpenSHMEM specification semantics.
|
|
|
|
Applications that use HIP can be easily interface with rocSHMEM.
|
|
As per the HIP programming model,
|
|
rocSHMEM has `__host__` APIs which are to be called from host code,
|
|
and `__device__` APIs which can be called within GPU Kernels.
|
|
Any device APIs which do not have any special suffixes/infixes (e.g. `_wg` or `_wave`)
|
|
must be called by a single thread.
|
|
GPU specific `_wg` and `_wave` APIs are expected to be called from multiple GPU threads
|
|
and block until the calling scope completes.
|
|
These APIs can be called in divergent code paths but this is not recommended.
|
|
|
|
Wavefront APIs
|
|
==============
|
|
The wavefront APIs are any API calls that have the suffix `_wave`.
|
|
The parameters in which these routines are called must be
|
|
the same for every thread in the wavefront.
|
|
If any thread calls these routines with differing parameters, the behavior is undefined.
|
|
These APIs will block until the calling wavefront completes.
|
|
|
|
Workgroup APIs
|
|
==============
|
|
The workgroup APIs are any API calls that have the suffix `_wg` or infix `_wg_`.
|
|
The parameters in which these routines are called must be
|
|
the same for every thread in the workgroup.
|
|
If any thread calls these routines with differing parameters, the behavior is undefined.
|
|
These APIs will block until the calling workgroup completes.
|