Added support for most of <mutex> #113
Conversation
|
TODO: Test ports and heterogeneous tests. |
brycelelbach
added
enhancement
|
@wmaxey @ogiroux is this just waiting on porting tests? I'd like to move forward with exploring NVIDIA/cccl#990, but it depends on having a |
|
@jrhemstad How important is this and should I try to get this rebased? |
It's important, but not urgent. No need to try and get it merged before the next release. There is still a sizable amount of work to add tests before it can be merged anyways. |
miscco
force-pushed
the
mutex
branch
4 times, most recently
from
0f16134 to
2c19f4a
Compare
miscco
force-pushed
the
mutex
branch
2 times, most recently
from
4952f9e to
e8b2e53
Compare
...x/thread.mutex.requirements/thread.mutex.requirements.mutex/thread.mutex.class/lock.pass.cpp
Outdated
Show resolved
Hide resolved
...ream-tests/test/cuda/thread/thread.mutex/thread.once/thread.once.callonce/call_once.pass.cpp
Outdated
Show resolved
Hide resolved
| handler.join_test_thread(); | ||
| handler.syncthreads(); | ||
|
|
||
| assert(init3_called[Sco] == 2); |
There was a problem hiding this comment.
| assert(init3_called[Sco] == 2); | |
| assert(init3_called[Sco] == 1); |
| if (init3_called[Sco] == 1) | ||
| #ifdef __CUDA_ARCH__ | ||
| _LIBCUDACXX_UNREACHABLE(); | ||
| #else | ||
| TEST_THROW(1); | ||
| #endif |
There was a problem hiding this comment.
| if (init3_called[Sco] == 1) | |
| #ifdef __CUDA_ARCH__ | |
| _LIBCUDACXX_UNREACHABLE(); | |
| #else | |
| TEST_THROW(1); | |
| #endif | |
| assert(init3_called[Sco] == 1); |
There was a problem hiding this comment.
Technically that would not be the same. We need to throw to ensure that we do not increment the init3_called variable afterwards but continue the test as expected
|
While everything works fine locally, this PR is incredibly flaky in CI. We need to investigate why that is the case and find a strategy to unflake it |
| if (__m_ == nullptr) | ||
| __throw_system_error(EPERM, "unique_lock::lock: references null mutex"); | ||
| if (__owns_) | ||
| __throw_system_error(EDEADLK, "unique_lock::lock: already locked"); | ||
| #endif // _LIBCUDACXX_NO_EXCEPTIONS |
There was a problem hiding this comment.
Without exceptions, on the host, we should be always printing the error, and aborting.
Without exceptions, on the device, we should be __traping, and when on debug mode, we should be printing the error.
Same for all other error handling in this PR for which we are already producing errors.
| using __libcpp_mutex_base_t = __libcpp_mutex_t; | ||
| #else | ||
| template<int _Sco> | ||
| using __libcpp_mutex_base_t = __atomic_semaphore_base<_Sco,1>; |
There was a problem hiding this comment.
Some comments would help here, but from skimming through the implementation of __atomic_semaphore_base, it seems that its not fair, which will starve threads from different NUMA nodes at system scope?
| @@ -274,6 +345,8 @@ void | |||
| swap(unique_lock<_Mutex>& __x, unique_lock<_Mutex>& __y) _NOEXCEPT | |||
| {__x.swap(__y);} | |||
|
|
|||
| #ifndef _LIBCUDACXX_HAS_THREAD_API_CUDA | |||
There was a problem hiding this comment.
Is condition variable supported?
|
Dropping in favor of NVIDIA/cccl#187 |
Has mutex, timed_mutex, once_flag, call_once, unique_lock, scoped_lock, and varied free functions that go with them. Excludes only condition variable support and the recursive versions of mutex.