32 lines
858 B
ReStructuredText
32 lines
858 B
ReStructuredText
|
.. _libcudacxx-ptx-instructions-mbarrier-red-async:
|
||
|
|
|
||
|
|
red.async
|
||
|
|
=========
|
||
|
|
|
||
|
|
- PTX ISA:
|
||
|
|
`red.async <https://docs.nvidia.com/cuda/parallel-thread-execution/index.html#parallel-synchronization-and-communication-instructions-red-async>`__
|
||
|
|
|
||
|
|
.. _red.async-1:
|
||
|
|
|
||
|
|
red.async
|
||
|
|
---------
|
||
|
|
|
||
|
|
.. include:: generated/red_async.rst
|
||
|
|
|
||
|
|
red.async ``.s64`` emulation
|
||
|
|
----------------------------
|
||
|
|
|
||
|
|
PTX does not currently (CTK 12.3) expose ``red.async.add.s64``. This
|
||
|
|
exposure is emulated in ``cuda::ptx`` using
|
||
|
|
|
||
|
|
.. code:: cuda
|
||
|
|
|
||
|
|
// red.async.relaxed.cluster.shared::cluster.mbarrier::complete_tx::bytes{.op}.u64 [dest], value, [remote_bar]; // .u64 intentional PTX ISA 81, SM_90
|
||
|
|
// .op = { .add }
|
||
|
|
template <typename=void>
|
||
|
|
__device__ static inline void red_async(
|
||
|
|
cuda::ptx::op_add_t,
|
||
|
|
int64_t* dest,
|
||
|
|
const int64_t& value,
|
||
|
|
int64_t* remote_bar);
|