2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2017,2018,2019, by the GROMACS development team, led by
5 * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
6 * and including many others, as listed in the AUTHORS file in the
7 * top-level source directory and at http://www.gromacs.org.
9 * GROMACS is free software; you can redistribute it and/or
10 * modify it under the terms of the GNU Lesser General Public License
11 * as published by the Free Software Foundation; either version 2.1
12 * of the License, or (at your option) any later version.
14 * GROMACS is distributed in the hope that it will be useful,
15 * but WITHOUT ANY WARRANTY; without even the implied warranty of
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
17 * Lesser General Public License for more details.
19 * You should have received a copy of the GNU Lesser General Public
20 * License along with GROMACS; if not, see
21 * http://www.gnu.org/licenses, or write to the Free Software Foundation,
22 * Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
24 * If you want to redistribute modifications to GROMACS, please
25 * consider that scientific software is very special. Version
26 * control is crucial - bugs must be traceable. We will be happy to
27 * consider code for inclusion in the official distribution, but
28 * derived work must not be called official GROMACS. Details are found
29 * in the README & COPYING files - if they are missing, get the
30 * official version at http://www.gromacs.org.
32 * To help us fund GROMACS development, we humbly ask that you cite
33 * the research papers on the package. Check out http://www.gromacs.org.
36 * \brief Declares gmx::HostAllocationPolicy, gmx::HostAllocator,
37 * gmx::HostVector and gmx::PaddedHostVector, which are used to make/be
38 * standard library containers that can allocate memory suitable for transfers.
39 * Currently the only supported transfers using pinned memory are
40 * to CUDA GPUs, but other possibilities exist in future.
42 * \todo This should not be in the public API, but it needs to be
43 * for the moment because state.h is in that API.
45 * \author Mark Abraham <mark.j.abraham@gmail.com>
48 #ifndef GMX_GPU_UTILS_HOSTALLOCATOR_H
49 #define GMX_GPU_UTILS_HOSTALLOCATOR_H
55 #include "gromacs/math/paddedvector.h"
56 #include "gromacs/utility/alignedallocator.h"
57 #include "gromacs/utility/exceptions.h"
62 /*! \brief Helper enum for pinning policy of the allocation of
63 * HostAllocationPolicy.
65 * For an efficient non-blocking transfer (e.g. to a GPU), the memory
66 * pages for a buffer need to be pinned to a physical page. Aligning
67 * such buffers to a physical page should miminize the number of pages
68 * that need to be pinned. However, some buffers that may be used for
69 * such transfers may also be used in either GROMACS builds or run
70 * paths that cannot use such a device, so the policy can be
71 * configured so that the resource consumption is no higher than
72 * required for correct, efficient operation in all cases. */
73 enum class PinningPolicy : int
75 CannotBePinned, // Memory is not known to be suitable for pinning.
76 PinnedIfSupported, // Memory is suitable for efficient pinning, e.g. because it is
77 // allocated to be page aligned, and will be pinned when supported.
80 //! Forward declaration of host allocation policy class.
81 class HostAllocationPolicy;
83 /*! \brief Memory allocator that uses HostAllocationPolicy.
85 * \tparam T Type of objects to allocate
87 * This convenience partial specialization can be used for the
88 * optional allocator template parameter in standard library
89 * containers whose memory may be used for e.g. GPU transfers. The
90 * memory will always be allocated according to the behavior of
91 * HostAllocationPolicy.
94 using HostAllocator = Allocator<T, HostAllocationPolicy>;
96 //! Convenience alias for std::vector that uses HostAllocator.
98 using HostVector = std::vector<T, HostAllocator<T>>;
100 //! Convenience alias for PaddedVector that uses HostAllocator.
102 using PaddedHostVector = PaddedVector<T, HostAllocator<T>>;
105 * \brief Policy class for configuring gmx::Allocator, to manage
106 * allocations of memory that may be needed for e.g. GPU transfers.
108 * This allocator has state, so is most useful in cases where it is
109 * not known at compile time whether the allocated memory will be
110 * transferred to some device. It will increase the size of containers
111 * that use it. If the GROMACS build is configured with CUDA support,
112 * then memory will be allocated with PageAlignedAllocator, and that
113 * page pinned to physical memory if the pinning mode has been
114 * activated. If pinning mode is deactivated, or the GROMACS build
115 * does not support CUDA, then the memory will be allocated with
116 * AlignedAllocator. The pin() and unpin() methods work with the CUDA
117 * build, and silently do nothing otherwise. In future, we may modify
118 * or generalize this to work differently in other cases.
120 * The intended use is to configure gmx::Allocator with this class as
121 * its policy class, and then to use e.g.
122 * std::vector::get_allocator().getPolicy() to control whether the
123 * allocation policy should activate its pinning mode. The policy
124 * object can also be used to explicitly pin() and unpin() the buffer
125 * when it is using PinningPolicy::PinnedIfSupported. The policy object is
126 * returned by value (as required by the C++ standard for
127 * get_allocator(), which copies a std::shared_ptr, so the policy
128 * object should be retrieved sparingly, e.g. only upon resize of the
129 * allocation. (Normal operation of the vector, e.g. during resize,
130 * incurs only the cost of the pointer indirection needed to consult
131 * the current state of the allocation policy.)
133 * \todo As a minor optimization, consider also having a stateless
134 * version of this policy, which might be slightly faster or more
135 * convenient to use in the cases where it is known at compile time
136 * that the allocation will be used to transfer to a GPU.
138 class HostAllocationPolicy
142 HostAllocationPolicy(PinningPolicy policy = PinningPolicy::CannotBePinned);
143 /*! \brief Return the alignment size currently used by the active pinning policy. */
144 std::size_t alignment() const noexcept;
145 /*! \brief Allocate and perhaps pin page-aligned memory suitable for
146 * e.g. GPU transfers.
148 * Before attempting to allocate, unpin() is called. After a
149 * successful allocation, pin() is called. (Whether these do
150 * things depends on the PinningPolicy that is in effect.)
152 * \param bytes Amount of memory (bytes) to allocate. It is valid to ask for
153 * 0 bytes, which will return a non-null pointer that is properly
154 * aligned and padded (but that you should not use).
156 * \return Valid pointer if the allocation+optional pinning worked, otherwise nullptr.
158 * \note Memory allocated with this routine must be released
159 * with gmx::HostAllocationPolicy::free(), and
160 * absolutely not the system free().
164 void* malloc(std::size_t bytes) const noexcept;
165 /*! \brief Free the memory, after unpinning (if appropriate).
167 * \param buffer Memory pointer previously returned from gmx::HostAllocationPolicy::malloc()
169 * \note This routine should only be called with pointers
170 * obtained from gmx:HostAllocationPolicy::malloc(),
171 * and absolutely not any pointers obtained the system
176 void free(void* buffer) const noexcept;
177 /*! \brief Return the active pinning policy.
181 PinningPolicy pinningPolicy() const { return pinningPolicy_; }
182 //! Don't propagate for copy
183 using propagate_on_container_copy_assignment = std::false_type;
184 //! Propagate for move
185 using propagate_on_container_move_assignment = std::true_type;
186 //! Propagate for move
187 using propagate_on_container_swap = std::true_type;
188 //! Use default allocator for copy (same as construct+copy)
189 HostAllocationPolicy select_on_container_copy_construction() const { return {}; }
193 PinningPolicy pinningPolicy_;
196 /*! \brief Return true if two allocators are identical
198 * True if pinning policy is the same.
200 template<class T1, class T2>
201 bool operator==(const Allocator<T1, HostAllocationPolicy>& a, const Allocator<T2, HostAllocationPolicy>& b)
203 return a.pinningPolicy() == b.pinningPolicy();
206 /*! \brief Helper function for changing the pinning policy of a pinnable vector.
208 * If the vector has contents, then a full reallocation and buffer
209 * copy are needed if the policy change requires tighter restrictions,
210 * and desirable even if the policy change requires looser
211 * restrictions. That cost is OK, because GROMACS will do this
212 * operation very rarely (e.g. when auto-tuning and deciding to switch
213 * whether a task will run on a GPU, or not). */
214 template<typename PinnableVector>
215 void changePinningPolicy(PinnableVector* v, PinningPolicy pinningPolicy)
217 // Force reallocation by element-wise move (because policy is
218 // different container is forced to realloc). Does nothing if
219 // policy is the same.
220 *v = PinnableVector(std::move(*v), { pinningPolicy });