.. MIT License Copyright (c) 2024-2026 Advanced Micro Devices, Inc. Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions: The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software. THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. .. meta:: :description: API reference for the cuda::mr::resource and cuda::mr::async_resource concepts, which constrain memory resource types for synchronous and stream-ordered allocation in libhipcxx. :keywords: libhipcxx, ROCm, HIP, C++, resource concept, async_resource, memory resource, allocate, deallocate, stream-ordered .. _libcudacxx-extended-api-memory-resources-resource: The ``cuda::resource`` concept ------------------------------ This page documents the ``cuda::mr::resource`` and ``cuda::mr::async_resource`` concepts, which constrain memory resource types for synchronous and stream-ordered allocation. The `std::pmr::memory_resource `__ feature provides only a single ``allocate`` interface, which is sufficient for homogeneous memory systems. However, CUDA provides both synchronous and `stream-ordered allocation `__. With `std::pmr::memory_resource `__ there is no way to tell whether a memory resource can utilize stream-ordered allocations. Even if the application knows it can, there is no way to properly tell the memory resource to use stream-ordered allocation. Ideally, this should not be something discovered through an assert at run time, but should be checked by the compiler. The ``cuda::mr::resource`` concept provides basic type checks to ensure that a given memory resource provides the expected ``allocate`` / ``deallocate`` interface and is also equality comparable, which covers the whole API surface of `std::pmr::memory_resource `__. See below for different memory resources and potential pitfalls. To demonstrate, the following example defines several resources, only some of which are valid implementations of the ``cuda::mr::resource`` concept. The ``static_assertion``'s will result in compile-time errors for the invalid resources. .. code:: cpp struct valid_resource { void* allocate(std::size_t, std::size_t) { return nullptr; } void deallocate(void*, std::size_t, std::size_t) noexcept {} bool operator==(const valid_resource&) const { return true; } // NOTE: C++20 thankfully added operator rewrite rules so defining operator!= is not required. // However, if compiled with C++14 / C++17, operator != must also be defined. bool operator!=(const valid_resource&) const { return false; } }; static_assert(cuda::mr::resource, ""); struct invalid_argument {}; struct invalid_allocate_argument { void* allocate(invalid_argument, std::size_t) { return nullptr; } void deallocate(void*, std::size_t, std::size_t) noexcept {} bool operator==(const invalid_allocate_argument&) { return true; } }; static_assert(!cuda::mr::resource, ""); struct invalid_allocate_return { int allocate(std::size_t, std::size_t) { return 42; } void deallocate(void*, std::size_t, std::size_t) noexcept {} bool operator==(const invalid_allocate_return&) { return true; } }; static_assert(!cuda::mr::resource, ""); struct invalid_deallocate_argument { void* allocate(std::size_t, std::size_t) { return nullptr; } void deallocate(void*, invalid_argument, std::size_t) noexcept {} bool operator==(const invalid_deallocate_argument&) { return true; } }; static_assert(!cuda::mr::resource, ""); struct non_comparable { void* allocate(std::size_t, std::size_t) { return nullptr; } void deallocate(void*, std::size_t, std::size_t) noexcept {} }; static_assert(!cuda::mr::resource, ""); struct non_eq_comparable { void* allocate(std::size_t, std::size_t) { return nullptr; } void deallocate(void*, std::size_t, std::size_t) noexcept {} bool operator!=(const non_eq_comparable&) { return false; } }; static_assert(!cuda::mr::resource, ""); In addition to the `std::pmr::memory_resource `_ interface the ``cuda::mr::async_resource`` concept verifies that a memory resource also satisfies the ``allocate_async`` / ``deallocate_async`` interface. Requiring both the PMR interface and the async interface is a deliberate design decision. .. code:: cpp struct valid_resource { void* allocate(std::size_t, std::size_t) { return nullptr; } void deallocate(void*, std::size_t, std::size_t) noexcept {} void* allocate_async(std::size_t, std::size_t, cuda::stream_ref) { return nullptr; } void deallocate_async(void*, std::size_t, std::size_t, cuda::stream_ref) {} bool operator==(const valid_resource&) const { return true; } bool operator!=(const valid_resource&) const { return false; } }; static_assert(cuda::mr::async_resource, ""); A library can easily decide whether to use the async interface: .. code:: cpp template requires cuda::mr::resource void* maybe_allocate_async(MemoryResource& resource, std::size_t size, std::size_t align, cuda::stream_ref stream) { if constexpr(cuda::mr::async_resource) { return resource.allocate_async(size, align, stream); } else { return resource.allocate(size, align); } } .. rubric:: Putting them together Applications and libraries may want to combine type checks for arbitrary properties with the ``{async_}resource`` concept. The ``{async_}resource_with`` concept allows checking resources for arbitrary properties. .. code:: cpp struct required_alignment{ using value_type = std::size_t; }; struct my_memory_resource { void* allocate(std::size_t, std::size_t) { return nullptr; } void deallocate(void*, std::size_t, std::size_t) noexcept {} bool operator==(const my_memory_resource&) const { return true; } bool operator!=(const my_memory_resource&) const { return false; } friend constexpr std::size_t get_property(const my_memory_resource& resource, required_alignment) noexcept { return resource.required_alignment; } std::size_t required_alignment; }; template requires cuda::mr::resource void* maybe_allocate_async_check_alignment(MemoryResource& resource, std::size_t size, cuda::stream_ref stream) { if constexpr(cuda::mr::async_resource_with) { return resource.allocate_async(size, get_property(resource, required_alignment), stream); } else if constexpr (cuda::mr::async_resource) { return resource.allocate_async(size, my_default_alignment, stream); } else if constexpr (cuda::mr::resource_with) { return resource.allocate(size, get_property(resource, required_alignment)); } else { return resource.allocate(size, my_default_alignment); } } // Potentially more concise template requires cuda::mr::resource void* maybe_allocate_async_check_alignment2(MemoryResource& resource, std::size_t size, cuda::stream_ref stream) { constexpr std::size_t align = cuda::mr::resource_with ? get_property(resource, required_alignment) : my_default_alignment; if constexpr(cuda::mr::async_resource) { return resource.allocate_async(size, align, stream); } else { return resource.allocate(size, align); } }