84 lines
2.8 KiB
ReStructuredText
84 lines
2.8 KiB
ReStructuredText
|
.. _libcudacxx-ptx:
|
||
|
|
|
||
|
|
PTX API
|
||
|
|
========
|
||
|
|
|
||
|
|
|
||
|
|
The ``cuda::ptx`` namespace contains functions that map one-to-one to
|
||
|
|
`PTX instructions <https://docs.nvidia.com/cuda/parallel-thread-execution/index.html>`__.
|
||
|
|
These can be used for maximal control of the generated code, or to
|
||
|
|
experiment with new hardware features before a high-level C++ API is
|
||
|
|
available.
|
||
|
|
|
||
|
|
.. toctree::
|
||
|
|
:maxdepth: 1
|
||
|
|
|
||
|
|
ptx/examples
|
||
|
|
ptx/instructions
|
||
|
|
ptx/pragmas
|
||
|
|
|
||
|
|
Versions and compatibility
|
||
|
|
---------------------------
|
||
|
|
|
||
|
|
The ``cuda/ptx`` header is intended to present a stable API within one major
|
||
|
|
version of the CTK on a best effort basis. This means that:
|
||
|
|
|
||
|
|
- All functions are marked static inline.
|
||
|
|
|
||
|
|
- The type of a function parameter can be changed to be more generic if
|
||
|
|
that means that code that called the original version can still be
|
||
|
|
compiled.
|
||
|
|
|
||
|
|
- Good exposure of the PTX should be high priority. If, at a new major
|
||
|
|
version, we face a difficult choice between breaking
|
||
|
|
backward-compatibility and an improvement of the PTX exposure, we
|
||
|
|
will tend to the latter option more easily than in other parts of
|
||
|
|
libcu++.
|
||
|
|
|
||
|
|
The API does not guarantee stability of template parameters. The order and
|
||
|
|
number of template parameters may change. Use arguments to driver overload
|
||
|
|
resolution as in the code below to ensure forward-compatibility:
|
||
|
|
|
||
|
|
.. code:: cuda
|
||
|
|
|
||
|
|
// Use arguments to drive overload resolution:
|
||
|
|
cuda::ptx::mbarrier_arrive_expect_tx(cuda::ptx::sem_release, cuda::ptx::scope_cta, cuda::ptx::space_shared, &bar, 1);
|
||
|
|
|
||
|
|
// Specifying templates directly is not forward-compatible, as order and number
|
||
|
|
// of template parameters may change in a minor release:
|
||
|
|
cuda::ptx::mbarrier_arrive_expect_tx<cuda::ptx::sem_release_t>(
|
||
|
|
cuda::ptx::sem_release, cuda::ptx::scope_cta, cuda::ptx::space_shared, &bar, 1
|
||
|
|
);
|
||
|
|
|
||
|
|
**PTX ISA version and compute capability.** Each binding notes under
|
||
|
|
which PTX ISA version and SM version it may be used. Example:
|
||
|
|
|
||
|
|
.. code:: cuda
|
||
|
|
|
||
|
|
// mbarrier.arrive.shared::cta.b64 state, [addr]; // 1. PTX ISA 70, SM_80
|
||
|
|
__device__ inline uint64_t mbarrier_arrive(
|
||
|
|
cuda::ptx::sem_release_t sem,
|
||
|
|
cuda::ptx::scope_cta_t scope,
|
||
|
|
cuda::ptx::space_shared_t space,
|
||
|
|
uint64_t* addr);
|
||
|
|
|
||
|
|
To check if the current compiler is recent enough, use:
|
||
|
|
|
||
|
|
.. code:: cuda
|
||
|
|
|
||
|
|
#if __cccl_ptx_isa >= 700
|
||
|
|
cuda::ptx::mbarrier_arrive(cuda::ptx::sem_release, cuda::ptx::scope_cta, cuda::ptx::space_shared, &bar, 1);
|
||
|
|
#endif
|
||
|
|
|
||
|
|
Ensure that you only call the function when compiling for a recent
|
||
|
|
enough compute capability (SM version), like this:
|
||
|
|
|
||
|
|
.. code:: cuda
|
||
|
|
|
||
|
|
NV_IF_TARGET(NV_PROVIDES_SM_80,(
|
||
|
|
cuda::ptx::mbarrier_arrive(cuda::ptx::sem_release, cuda::ptx::scope_cta, cuda::ptx::space_shared, &bar, 1);
|
||
|
|
));
|
||
|
|
|
||
|
|
For more information on which compilers correspond to which PTX ISA, see
|
||
|
|
the `PTX ISA release notes <https://docs.nvidia.com/cuda/parallel-thread-execution/index.html#release-notes>`__.
|