You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Atomic Refs Bound to Memory Orderings & Atomic Accessors
The mdspan paper P0009 listed atomic accessors as a reason for having accessors in the first place.
One of the use cases is for parallel algorithms updating data in a way which has data races.
Consider the histogram computation:
template<classExecT>
voidcompute_histogram(ExecT exec, float bin_size,
std::mdspan<int, std::dextents<size_t,1>> output,
std::mdspan<float, std::dextents<size_t,1>> data) {
static_assert(std::is_execution_policy_v<ExecT>);
std::for_each(exec, data.data_handle(), data.data_handle()+data.extent(0), [=](float val) {
int bin = std::abs(val)/bin_size;
bin = std::clamp(bin, size_t(0), output.extent(0));
output[bin]++;
});
Depending on whether ExecT is sequenced_policy or not, the update needs to happen atomically.
With just atomic_ref one could do:
std::for_each(exec, data.data_handle(), data.data_handle()+data.extent(0), [=](float val) {
int bin = std::abs(val)/bin_size;
bin = std::clamp(bin, size_t(0), output.extent(0));
atomic_ref(output[bin])++;
});
This paper proposes a way for doing that for general mdspans without writing out atomic_ref everywhere inside a complex algorithm and also fixes the fact that you can't do the simple ++ with relaxed memory order - enough for these kind of accumulation cases.
What this paper proposes
"Bounded" versions of atomic_ref with a fixed (bounded) memory order
Conceptually just like atomic_ref with a fixed memory order - all the same members
Corresponding atomic accessors which use the various atomic_ref types as reference type
atomic_accessor => atomic_ref
atomic_accessor_relaxed => atomic_ref_relaxed
atomic_accessor_acq_rel => atomic_ref_acq_rel
atomic_accessor_seq_cst => atomic_ref_seq_cst
Histogram with atomic accessor
// sequenced_policy does not attach atomic accessortemplate<classT, classExtents, classLayoutPolicy>
autoadd_atomic_accessor_if_needed(
std::execution::sequenced_policy, mdspan<T, Extents, LayoutPolicy> m) {
return m;
}
// parallel policies attach atomic accessor:template<classExecutionPolicy, classT, classExtents, classLayoutPolicy>
autoadd_atomic_accessor_if_needed(
ExecutionPolicy, mdspan<T, Extents, LayoutPolicy> m) {
returnmdspan(m.data_handle(), m.mapping(), atomic_accessor<T>());
}
template<classExecT>
voidcompute_histogram(ExecT exec, float bin_size,
std::mdspan<int, std::dextents<size_t,1>> output,
std::mdspan<float, std::dextents<size_t,1>> data) {
static_assert(std::is_execution_policy_v<ExecT>);
auto accumulator = add_atomic_accessor_if_needed(exec, output);
std::for_each(exec, data.data_handle(), data.data_handle()+data.extent(0), [=](float val) {
int bin = std::abs(val)/bin_size;
bin = std::clamp(bin, size_t(0), output.extent(0));
// this is now atomic for parallel policies
accumulator[bin]++;
});
Previous Question: Why not expose the templated type?
There is really no use-case where one would want to write something generic over the memory order!
The algorithm determines the memory order you need - the only generic question we identified is whether you need atomics at all (i.e. what execution policy is used)
SG1 also felt that its easier to see the memory order with the explicit names
We don't feel there is any real risk here and mdspan somewhat hides its accessor anyway
Wording Considerations
Used exposition only atomic-ref-bound<T,MemoryOrder> for wording
atomic_ref_MEM_ORDER are template aliases - see open question whether this is ok
atomic_ref has 4 specializations: general, integral (not bool), floating point and pointer
atomic-ref-bound needs only 2 (general and arithmetic/pointer{not-bool} versions)
integral, floating point and pointer type versions can be defined in terms of atomic_ref with constraints on member functions
the only reason we need general is difference_type member typedef -> doesn't exist in atomic_ref general
Open Questions
Is reasoning for having atomic-ref-bound exposition only good enough, or does LEWG want it as a public thing?
Please Poll: atomic-ref-bound should be exposition only.
Author position: in favor
Converting between atomic-ref-bound and basic-atomic-accessors of different memory orders
We don't see many good reasons to have them:
atomic_ref is meant to be almost used in-place for individual operations e.g. atomic_ref(a[i])++
atomic accessors are something you are supposed to add in a fairly local scope
we don't expect folks to pass mdspan with atomic accessors around
There is a reason not to permit it:
Writing functions which take atomic_ref_MEM_ORDER imply ordering behavior: allowing conversions would change that behavior
Please Poll: atomic-ref-bound and basic-atomic-accessor should have converting constructors for different memory-orders
Author position: against
The text was updated successfully, but these errors were encountered:
Atomic Refs Bound to Memory Orderings & Atomic Accessors
The
mdspan
paper P0009 listed atomic accessors as a reason for having accessors in the first place.One of the use cases is for parallel algorithms updating data in a way which has data races.
Consider the histogram computation:
Depending on whether
ExecT
issequenced_policy
or not, the update needs to happen atomically.With just
atomic_ref
one could do:This paper proposes a way for doing that for general
mdspan
s without writing outatomic_ref
everywhere inside a complex algorithm and also fixes the fact that you can't do the simple++
with relaxed memory order - enough for these kind of accumulation cases.What this paper proposes
atomic_ref
with a fixed (bounded) memory orderatomic_ref_relaxed
,atomic_ref_seq_cst
,atomic_ref_acq_rel
atomic_ref
with a fixed memory order - all the same membersatomic_ref
types as reference typeatomic_accessor
=>atomic_ref
atomic_accessor_relaxed
=>atomic_ref_relaxed
atomic_accessor_acq_rel
=>atomic_ref_acq_rel
atomic_accessor_seq_cst
=>atomic_ref_seq_cst
Histogram with atomic accessor
Previous Question: Why not expose the templated type?
mdspan
somewhat hides its accessor anywayWording Considerations
atomic-ref-bound<T,MemoryOrder>
for wordingatomic_ref_MEM_ORDER
are template aliases - see open question whether this is okatomic_ref
has 4 specializations: general, integral (not bool), floating point and pointeratomic-ref-bound
needs only 2 (general and arithmetic/pointer{not-bool} versions)atomic_ref
with constraints on member functionsdifference_type
member typedef -> doesn't exist inatomic_ref
generalOpen Questions
atomic-ref-bound
exposition only good enough, or does LEWG want it as a public thing?atomic-ref-bound
should be exposition only.atomic-ref-bound
andbasic-atomic-accessors
of different memory ordersatomic_ref
is meant to be almost usedin-place
for individual operations e.g.atomic_ref(a[i])++
mdspan
with atomic accessors aroundatomic_ref_MEM_ORDER
imply ordering behavior: allowing conversions would change that behavioratomic-ref-bound
andbasic-atomic-accessor
should have converting constructors for different memory-ordersThe text was updated successfully, but these errors were encountered: