PR uses TBB internal atomic's for atomic CAS on non-atomic data
Corresponding PR https://github.com/Project-OSRM/osrm-backend/pull/4199
Other options:
* use sequential update
* use an internal packed vector lock -> makes packed vector non-movable
* use boost.interprocess atomics implementation -> outdated and only 32 bit version
* use glib atomic's -> requires new dependency
* wait for https://isocpp.org/blog/2014/05/n4013 as_atomic
* use c11 _Atomic and atomic_compare_exchange_weak -> not possible to mix c++11 and c11
* use builtin functions gcc __sync_bool_compare_and_swap and msvc _InterlockedCompareExchange64 -> possible, but requires proper testing
boolean CompareAndSwapPointer(volatile * void * ptr,
void * new_value,
void * old_value) {
if defined(_MSC_VER)
if (InterlockedCompareExchange(ptr, new_value, old_value) == old_value) return false;
else return true;
elif (__GNUC__ * 10000 + __GNUC_MINOR__ * 100 + __GNUC_PATCHLEVEL__) > 40100
return __sync_bool_compare_and_swap(ptr, old_value, new_value);
else
error No implementation
endif
}
* use Boost.Atomic -> requires new dependency
WordT local_lower_word = lower_word, new_lower_word;
do
{
new_lower_word = set_lower_value<WordT, T>(local_lower_word,
lower_mask[internal_index.element],
lower_offset[internal_index.element],
value);
} while (!boost::atomics::detail::operations<sizeof(WordT), false>::compare_exchange_weak(
lower_word,
local_lower_word,
new_lower_word,
boost::memory_order_release,
boost::memory_order_relaxed));
- template function for tile functionality with edge finder operator
- refactors unit tests into single function (reduce code duplication)
- adds unit tests for core-ch
The alpha constant is for the local optimality T-Test threshold.
Before we used epsilon for the T-Test threshold, but the epsilon
constant is meant to be used for the stretch test(s) only.
This changeset fixes the local optimality T-Test and uses the
epsilon constant for the two stretch tests:
- We test the stretch for the total route against epsilon and
- We test the detour against the epsilon now, too
We can discuss if the second stretch test should actually use
epsilon, too, or a adapted value of it - but definitly not alpha.
The new numbering uses the partition information
to sort border nodes first to compactify storages
that need access indexed by border node ID.
We also get an optimized cache performance for free
sincr we can also recursively sort the nodes by cell ID.
This implements issue #3779.
