From 1ef3fe0a1f026689e64963ec9ab0b04b7e6b1bde Mon Sep 17 00:00:00 2001
From: waffl3x <waffl3x@baylibre.com>
Date: Tue, 15 Apr 2025 04:12:55 -0600
Subject: [PATCH] OpenMP: omp.h omp::allocator C++ Allocator interface
The implementation of each allocator is simplified by inheriting from
__detail::__allocator_templ. At the moment, none of the implementations
diverge in any way, simply passing in the allocator handle to be used when
an allocation is made. In the future, const_mem will need special handling
added to it to support constant memory space.
libgomp/ChangeLog:
* omp.h.in: Add omp::allocator::* and ompx::allocator::* allocators.
(__detail::__allocator_templ<T, omp_allocator_handle_t>):
New struct template.
(null_allocator<T>): New struct template.
(default_mem<T>): Likewise.
(large_cap_mem<T>): Likewise.
(const_mem<T>): Likewise.
(high_bw_mem<T>): Likewise.
(low_lat_mem<T>): Likewise.
(cgroup_mem<T>): Likewise.
(pteam_mem<T>): Likewise.
(thread_mem<T>): Likewise.
(ompx::allocator::gnu_pinned_mem<T>): Likewise.
* testsuite/libgomp.c++/allocator-1.C: New test.
* testsuite/libgomp.c++/allocator-2.C: New test.
Signed-off-by: waffl3x <waffl3x@baylibre.com>
---
libgomp/omp.h.in | 132 ++++++++++++++++
libgomp/testsuite/libgomp.c++/allocator-1.C | 158 ++++++++++++++++++++
libgomp/testsuite/libgomp.c++/allocator-2.C | 132 ++++++++++++++++
3 files changed, 422 insertions(+)
create mode 100644 libgomp/testsuite/libgomp.c++/allocator-1.C
create mode 100644 libgomp/testsuite/libgomp.c++/allocator-2.C
@@ -432,4 +432,136 @@ extern const char *omp_get_uid_from_device (int) __GOMP_NOTHROW;
}
#endif
+#if __cplusplus >= 201103L
+
+/* std::__throw_bad_alloc and std::__throw_bad_array_new_length. */
+#include <bits/functexcept.h>
+
+namespace omp
+{
+namespace allocator
+{
+
+namespace __detail
+{
+
+template<typename __T, omp_allocator_handle_t __Handle>
+struct __allocator_templ
+{
+ using value_type = __T;
+ using pointer = __T*;
+ using const_pointer = const __T*;
+ using size_type = __SIZE_TYPE__;
+ using difference_type = __PTRDIFF_TYPE__;
+
+ __T*
+ allocate (size_type __n)
+ {
+ if (__SIZE_MAX__ / sizeof(__T) < __n)
+ std::__throw_bad_array_new_length ();
+ void *__p = omp_aligned_alloc (alignof(__T), __n * sizeof(__T), __Handle);
+ if (!__p)
+ std::__throw_bad_alloc ();
+ return static_cast<__T*>(__p);
+ }
+
+ void
+ deallocate (__T *__p, size_type) __GOMP_NOTHROW
+ {
+ omp_free (static_cast<void*>(__p), __Handle);
+ }
+};
+
+template<typename __T, typename __U, omp_allocator_handle_t __Handle>
+constexpr bool
+operator== (const __allocator_templ<__T, __Handle>&,
+ const __allocator_templ<__U, __Handle>&) __GOMP_NOTHROW
+{
+ return true;
+}
+
+template<typename __T, omp_allocator_handle_t __Handle,
+ typename __U, omp_allocator_handle_t __UHandle>
+constexpr bool
+operator== (const __allocator_templ<__T, __Handle>&,
+ const __allocator_templ<__U, __UHandle>&) __GOMP_NOTHROW
+{
+ return false;
+}
+
+template<typename __T, typename __U, omp_allocator_handle_t __Handle>
+constexpr bool
+operator!= (const __allocator_templ<__T, __Handle>&,
+ const __allocator_templ<__U, __Handle>&) __GOMP_NOTHROW
+{
+ return false;
+}
+
+template<typename __T, omp_allocator_handle_t __Handle,
+ typename __U, omp_allocator_handle_t __UHandle>
+constexpr bool
+operator!= (const __allocator_templ<__T, __Handle>&,
+ const __allocator_templ<__U, __UHandle>&) __GOMP_NOTHROW
+{
+ return true;
+}
+
+} /* namespace __detail */
+
+template<typename __T>
+struct null_allocator
+ : __detail::__allocator_templ<__T, omp_null_allocator> {};
+
+template<typename __T>
+struct default_mem
+ : __detail::__allocator_templ<__T, omp_default_mem_alloc> {};
+
+template<typename __T>
+struct large_cap_mem
+ : __detail::__allocator_templ<__T, omp_large_cap_mem_alloc> {};
+
+template<typename __T>
+struct const_mem
+ : __detail::__allocator_templ<__T, omp_const_mem_alloc> {};
+
+template<typename __T>
+struct high_bw_mem
+ : __detail::__allocator_templ<__T, omp_high_bw_mem_alloc> {};
+
+template<typename __T>
+struct low_lat_mem
+ : __detail::__allocator_templ<__T, omp_low_lat_mem_alloc> {};
+
+template<typename __T>
+struct cgroup_mem
+ : __detail::__allocator_templ<__T, omp_cgroup_mem_alloc> {};
+
+template<typename __T>
+struct pteam_mem
+ : __detail::__allocator_templ<__T, omp_pteam_mem_alloc> {};
+
+template<typename __T>
+struct thread_mem
+ : __detail::__allocator_templ<__T, omp_thread_mem_alloc> {};
+
+} /* namespace allocator */
+
+} /* namespace omp */
+
+namespace ompx
+{
+
+namespace allocator
+{
+
+template<typename __T>
+struct gnu_pinned_mem
+ : omp::allocator::__detail::__allocator_templ<__T, ompx_gnu_pinned_mem_alloc> {};
+
+} /* namespace allocator */
+
+} /* namespace ompx */
+
+#endif /* __cplusplus */
+
#endif /* _OMP_H */
new file mode 100644
@@ -0,0 +1,158 @@
+// { dg-do run }
+
+#include <omp.h>
+#include <memory>
+#include <limits>
+
+template<typename T, template<typename> class Alloc>
+void test (T const initial_value = T())
+{
+ using Allocator = Alloc<T>;
+ Allocator a;
+ using Traits = std::allocator_traits<Allocator>;
+ static_assert (__is_same(typename Traits::allocator_type, Allocator ));
+ static_assert (__is_same(typename Traits::value_type, T ));
+ static_assert (__is_same(typename Traits::pointer, T* ));
+ static_assert (__is_same(typename Traits::const_pointer, T const* ));
+ static_assert (__is_same(typename Traits::void_pointer, void* ));
+ static_assert (__is_same(typename Traits::const_void_pointer, void const* ));
+ static_assert (__is_same(typename Traits::difference_type, __PTRDIFF_TYPE__));
+ static_assert (__is_same(typename Traits::size_type, __SIZE_TYPE__ ));
+ static_assert (Traits::propagate_on_container_copy_assignment::value == false);
+ static_assert (Traits::propagate_on_container_move_assignment::value == false);
+ static_assert (Traits::propagate_on_container_swap::value == false);
+ static_assert (Traits::is_always_equal::value == true);
+
+ static constexpr __SIZE_TYPE__ correct_max_size
+ = std::numeric_limits<__SIZE_TYPE__>::max () / sizeof (T);
+ if (Traits::max_size (a) != correct_max_size)
+ __builtin_abort ();
+
+ static constexpr __SIZE_TYPE__ alloc_count = 1;
+ T *p = Traits::allocate (a, alloc_count);
+ if (p == nullptr)
+ __builtin_abort ();
+ Traits::construct (a, p, initial_value);
+ if (*p != initial_value)
+ __builtin_abort ();
+ Traits::destroy (a, p);
+ Traits::deallocate (a, p, alloc_count);
+ /* Not interesting but might as well test it. */
+ static_cast<void>(Traits::select_on_container_copy_construction (a));
+
+ if (!(a == Allocator()))
+ __builtin_abort ();
+ if (a != Allocator())
+ __builtin_abort ();
+ if (!(a == Alloc<void>()))
+ __builtin_abort ();
+ if (a != Alloc<void>())
+ __builtin_abort ();
+}
+
+#define CHECK_INEQUALITY(other_alloc_templ, type) \
+do { \
+ /* Skip tests for itself, those are equal. Intantiate each */ \
+ /* one with void so we can easily tell if they are the same. */ \
+ if (!__is_same (AllocTempl<void>, other_alloc_templ<void>)) \
+ { \
+ other_alloc_templ<type> other; \
+ if (a == other) \
+ __builtin_abort (); \
+ if (!(a != other)) \
+ __builtin_abort (); \
+ } \
+} while (false)
+
+template<typename T, template<typename> class AllocTempl>
+void test_inequality ()
+{
+ using Allocator = AllocTempl<T>;
+ Allocator a;
+ CHECK_INEQUALITY (omp::allocator::null_allocator, void);
+ CHECK_INEQUALITY (omp::allocator::default_mem, void);
+ CHECK_INEQUALITY (omp::allocator::large_cap_mem, void);
+ CHECK_INEQUALITY (omp::allocator::const_mem, void);
+ CHECK_INEQUALITY (omp::allocator::high_bw_mem, void);
+ CHECK_INEQUALITY (omp::allocator::low_lat_mem, void);
+ CHECK_INEQUALITY (omp::allocator::cgroup_mem, void);
+ CHECK_INEQUALITY (omp::allocator::pteam_mem, void);
+ CHECK_INEQUALITY (omp::allocator::thread_mem, void);
+ CHECK_INEQUALITY (ompx::allocator::gnu_pinned_mem, void);
+ /* And again with the same type passed to the allocator. */
+ CHECK_INEQUALITY (omp::allocator::null_allocator, T);
+ CHECK_INEQUALITY (omp::allocator::default_mem, T);
+ CHECK_INEQUALITY (omp::allocator::large_cap_mem, T);
+ CHECK_INEQUALITY (omp::allocator::const_mem, T);
+ CHECK_INEQUALITY (omp::allocator::high_bw_mem, T);
+ CHECK_INEQUALITY (omp::allocator::low_lat_mem, T);
+ CHECK_INEQUALITY (omp::allocator::cgroup_mem, T);
+ CHECK_INEQUALITY (omp::allocator::pteam_mem, T);
+ CHECK_INEQUALITY (omp::allocator::thread_mem, T);
+ CHECK_INEQUALITY (ompx::allocator::gnu_pinned_mem, T);
+}
+
+#undef CHECK_INEQUALITY
+
+struct S
+{
+ int _v0;
+ bool _v1;
+ float _v2;
+
+ bool operator== (S const& other) const noexcept {
+ return _v0 == other._v0
+ && _v1 == other._v1
+ && _v2 == other._v2;
+ }
+ bool operator!= (S const& other) const noexcept {
+ return !this->operator==(other);
+ }
+};
+
+int main ()
+{
+ test<int, omp::allocator::null_allocator>(42);
+ test<int, omp::allocator::default_mem>(42);
+ test<int, omp::allocator::large_cap_mem>(42);
+ test<int, omp::allocator::const_mem>(42);
+ test<int, omp::allocator::high_bw_mem>(42);
+ test<int, omp::allocator::low_lat_mem>(42);
+ test<int, omp::allocator::cgroup_mem>(42);
+ test<int, omp::allocator::pteam_mem>(42);
+ test<int, omp::allocator::thread_mem>(42);
+ test<int, ompx::allocator::gnu_pinned_mem>(42);
+
+ test<long long, omp::allocator::null_allocator>(42);
+ test<long long, omp::allocator::default_mem>(42);
+ test<long long, omp::allocator::large_cap_mem>(42);
+ test<long long, omp::allocator::const_mem>(42);
+ test<long long, omp::allocator::high_bw_mem>(42);
+ test<long long, omp::allocator::low_lat_mem>(42);
+ test<long long, omp::allocator::cgroup_mem>(42);
+ test<long long, omp::allocator::pteam_mem>(42);
+ test<long long, omp::allocator::thread_mem>(42);
+ test<long long, ompx::allocator::gnu_pinned_mem>(42);
+
+ test<S, omp::allocator::null_allocator>( S{42, true, 128.f});
+ test<S, omp::allocator::default_mem>( S{42, true, 128.f});
+ test<S, omp::allocator::large_cap_mem>( S{42, true, 128.f});
+ test<S, omp::allocator::const_mem>( S{42, true, 128.f});
+ test<S, omp::allocator::high_bw_mem>( S{42, true, 128.f});
+ test<S, omp::allocator::low_lat_mem>( S{42, true, 128.f});
+ test<S, omp::allocator::cgroup_mem>( S{42, true, 128.f});
+ test<S, omp::allocator::pteam_mem>( S{42, true, 128.f});
+ test<S, omp::allocator::thread_mem>( S{42, true, 128.f});
+ test<S, ompx::allocator::gnu_pinned_mem>(S{42, true, 128.f});
+
+ test_inequality<int, omp::allocator::null_allocator>();
+ test_inequality<int, omp::allocator::default_mem>();
+ test_inequality<int, omp::allocator::large_cap_mem>();
+ test_inequality<int, omp::allocator::const_mem>();
+ test_inequality<int, omp::allocator::high_bw_mem>();
+ test_inequality<int, omp::allocator::low_lat_mem>();
+ test_inequality<int, omp::allocator::cgroup_mem>();
+ test_inequality<int, omp::allocator::pteam_mem>();
+ test_inequality<int, omp::allocator::thread_mem>();
+ test_inequality<int, ompx::allocator::gnu_pinned_mem>();
+}
new file mode 100644
@@ -0,0 +1,132 @@
+// { dg-do run }
+// { dg-additional-options "-Wno-psabi" }
+
+#include <omp.h>
+#include <vector>
+
+template<typename T>
+bool ptr_is_aligned(T *ptr, std::size_t alignment)
+{
+ /* ALIGNMENT must be a power of 2. */
+ if ((alignment & (alignment - 1)) != 0)
+ __builtin_abort ();
+ __UINTPTR_TYPE__ ptr_value
+ = reinterpret_cast<__UINTPTR_TYPE__>(static_cast<void*>(ptr));
+ return (ptr_value % alignment) == 0;
+}
+
+template<typename T, template<typename> class Alloc>
+void f (T v0, T v1, T v2, T v3)
+{
+ std::vector<T, Alloc<T>> vec;
+ vec.push_back (v0);
+ vec.push_back (v1);
+ vec.push_back (v2);
+ vec.push_back (v3);
+ if (vec.at (0) != v0)
+ __builtin_abort ();
+ if (vec.at (1) != v1)
+ __builtin_abort ();
+ if (vec.at (2) != v2)
+ __builtin_abort ();
+ if (vec.at (3) != v3)
+ __builtin_abort ();
+ if (!ptr_is_aligned (&vec.at (0), alignof (T)))
+ __builtin_abort ();
+ if (!ptr_is_aligned (&vec.at (1), alignof (T)))
+ __builtin_abort ();
+ if (!ptr_is_aligned (&vec.at (2), alignof (T)))
+ __builtin_abort ();
+ if (!ptr_is_aligned (&vec.at (3), alignof (T)))
+ __builtin_abort ();
+}
+
+struct S0
+{
+ int _v0;
+ bool _v1;
+ float _v2;
+
+ bool operator== (S0 const& other) const noexcept {
+ return _v0 == other._v0
+ && _v1 == other._v1
+ && _v2 == other._v2;
+ }
+ bool operator!= (S0 const& other) const noexcept {
+ return !this->operator==(other);
+ }
+};
+
+struct alignas(128) S1
+{
+ int _v0;
+ bool _v1;
+ float _v2;
+
+ bool operator== (S1 const& other) const noexcept {
+ return _v0 == other._v0
+ && _v1 == other._v1
+ && _v2 == other._v2;
+ }
+ bool operator!= (S1 const& other) const noexcept {
+ return !this->operator==(other);
+ }
+};
+
+/* Note: the test for const_mem should be disabled in the future. */
+
+int main ()
+{
+ f<int, omp::allocator::null_allocator >(0, 1, 2, 3);
+ f<int, omp::allocator::default_mem >(0, 1, 2, 3);
+ f<int, omp::allocator::large_cap_mem >(0, 1, 2, 3);
+ f<int, omp::allocator::const_mem >(0, 1, 2, 3);
+ f<int, omp::allocator::high_bw_mem >(0, 1, 2, 3);
+ f<int, omp::allocator::low_lat_mem >(0, 1, 2, 3);
+ f<int, omp::allocator::cgroup_mem >(0, 1, 2, 3);
+ f<int, omp::allocator::pteam_mem >(0, 1, 2, 3);
+ f<int, omp::allocator::thread_mem >(0, 1, 2, 3);
+ f<int, ompx::allocator::gnu_pinned_mem>(0, 1, 2, 3);
+
+ f<long long, omp::allocator::null_allocator >(0, 1, 2, 3);
+ f<long long, omp::allocator::default_mem >(0, 1, 2, 3);
+ f<long long, omp::allocator::large_cap_mem >(0, 1, 2, 3);
+ f<long long, omp::allocator::const_mem >(0, 1, 2, 3);
+ f<long long, omp::allocator::high_bw_mem >(0, 1, 2, 3);
+ f<long long, omp::allocator::low_lat_mem >(0, 1, 2, 3);
+ f<long long, omp::allocator::cgroup_mem >(0, 1, 2, 3);
+ f<long long, omp::allocator::pteam_mem >(0, 1, 2, 3);
+ f<long long, omp::allocator::thread_mem >(0, 1, 2, 3);
+ f<long long, ompx::allocator::gnu_pinned_mem>(0, 1, 2, 3);
+
+ S0 s0_0{ 42, true, 111128.f};
+ S0 s0_1{ 142, false, 11128.f};
+ S0 s0_2{ 1142, true, 1128.f};
+ S0 s0_3{11142, false, 128.f};
+ f<S0, omp::allocator::null_allocator >(s0_0, s0_1, s0_2, s0_3);
+ f<S0, omp::allocator::default_mem >(s0_0, s0_1, s0_2, s0_3);
+ f<S0, omp::allocator::large_cap_mem >(s0_0, s0_1, s0_2, s0_3);
+ f<S0, omp::allocator::const_mem >(s0_0, s0_1, s0_2, s0_3);
+ f<S0, omp::allocator::high_bw_mem >(s0_0, s0_1, s0_2, s0_3);
+ f<S0, omp::allocator::low_lat_mem >(s0_0, s0_1, s0_2, s0_3);
+ f<S0, omp::allocator::cgroup_mem >(s0_0, s0_1, s0_2, s0_3);
+ f<S0, omp::allocator::pteam_mem >(s0_0, s0_1, s0_2, s0_3);
+ f<S0, omp::allocator::thread_mem >(s0_0, s0_1, s0_2, s0_3);
+ f<S0, ompx::allocator::gnu_pinned_mem>(s0_0, s0_1, s0_2, s0_3);
+
+ S1 s1_0{ 42, true, 111128.f};
+ S1 s1_1{ 142, false, 11128.f};
+ S1 s1_2{ 1142, true, 1128.f};
+ S1 s1_3{11142, false, 128.f};
+
+ f<S1, omp::allocator::null_allocator >(s1_0, s1_1, s1_2, s1_3);
+ f<S1, omp::allocator::default_mem >(s1_0, s1_1, s1_2, s1_3);
+ f<S1, omp::allocator::large_cap_mem >(s1_0, s1_1, s1_2, s1_3);
+ f<S1, omp::allocator::const_mem >(s1_0, s1_1, s1_2, s1_3);
+ f<S1, omp::allocator::high_bw_mem >(s1_0, s1_1, s1_2, s1_3);
+ f<S1, omp::allocator::low_lat_mem >(s1_0, s1_1, s1_2, s1_3);
+ f<S1, omp::allocator::cgroup_mem >(s1_0, s1_1, s1_2, s1_3);
+ f<S1, omp::allocator::pteam_mem >(s1_0, s1_1, s1_2, s1_3);
+ f<S1, omp::allocator::thread_mem >(s1_0, s1_1, s1_2, s1_3);
+ f<S1, ompx::allocator::gnu_pinned_mem>(s1_0, s1_1, s1_2, s1_3);
+}
--
2.47.1