f68324cd09
`SRWLOCK`, as used by `std::shared_mutex`, is a inherently unfair mutex
and makes no guarantee whatsoever whether a thread may acquire the lock
in a timely manner. This is problematic for our renderer which relies on
being able to acquire the lock in a timely and predictable manner.
Drawing stalls of up to one minute have been observed in tests.
This issue can be solved with a primitive ticket lock, which is 10x
slower than a `SRWLOCK` but still sufficiently fast for our use case
(10M locks per second per thread). It's likely that any non-trivial lock
duration will diminish the difference to "negligible".
## Validation Steps Performed
* It still blends ✔️
30 lines
1 KiB
C++
30 lines
1 KiB
C++
// Copyright (c) Microsoft Corporation.
|
|
// Licensed under the MIT license.
|
|
|
|
#pragma once
|
|
|
|
namespace til
|
|
{
|
|
template<typename T>
|
|
inline void atomic_wait(const std::atomic<T>& atomic, const T& current, DWORD waitMilliseconds = INFINITE) noexcept
|
|
{
|
|
static_assert(sizeof(atomic) == sizeof(current));
|
|
#pragma warning(suppress : 26492) // Don't use const_cast to cast away const or volatile
|
|
WaitOnAddress(const_cast<std::atomic<T>*>(&atomic), const_cast<T*>(¤t), sizeof(current), waitMilliseconds);
|
|
}
|
|
|
|
template<typename T>
|
|
inline void atomic_notify_one(const std::atomic<T>& atomic) noexcept
|
|
{
|
|
#pragma warning(suppress : 26492) // Don't use const_cast to cast away const or volatile
|
|
WakeByAddressSingle(const_cast<std::atomic<T>*>(&atomic));
|
|
}
|
|
|
|
template<typename T>
|
|
inline void atomic_notify_all(const std::atomic<T>& atomic) noexcept
|
|
{
|
|
#pragma warning(suppress : 26492) // Don't use const_cast to cast away const or volatile
|
|
WakeByAddressAll(const_cast<std::atomic<T>*>(&atomic));
|
|
}
|
|
}
|