2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2017,2018, 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, and
37 * gmx::HostVector, which are used to make/be standard library
38 * 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
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> >;
101 * \brief Policy class for configuring gmx::Allocator, to manage
102 * allocations of memory that may be needed for e.g. GPU transfers.
104 * This allocator has state, so is most useful in cases where it is
105 * not known at compile time whether the allocated memory will be
106 * transferred to some device. It will increase the size of containers
107 * that use it. If the GROMACS build is configured with CUDA support,
108 * then memory will be allocated with PageAlignedAllocator, and that
109 * page pinned to physical memory if the pinning mode has been
110 * activated. If pinning mode is deactivated, or the GROMACS build
111 * does not support CUDA, then the memory will be allocated with
112 * AlignedAllocator. The pin() and unpin() methods work with the CUDA
113 * build, and silently do nothing otherwise. In future, we may modify
114 * or generalize this to work differently in other cases.
116 * The intended use is to configure gmx::Allocator with this class as
117 * its policy class, and then to use e.g.
118 * std::vector::get_allocator().getPolicy() to control whether the
119 * allocation policy should activate its pinning mode. The policy
120 * object can also be used to explicitly pin() and unpin() the buffer
121 * when it is using PinningPolicy::PinnedIfSupported. The policy object is
122 * returned by value (as required by the C++ standard for
123 * get_allocator(), which copies a std::shared_ptr, so the policy
124 * object should be retrieved sparingly, e.g. only upon resize of the
125 * allocation. (Normal operation of the vector, e.g. during resize,
126 * incurs only the cost of the pointer indirection needed to consult
127 * the current state of the allocation policy.)
129 * \todo As a minor optimization, consider also having a stateless
130 * version of this policy, which might be slightly faster or more
131 * convenient to use in the cases where it is known at compile time
132 * that the allocation will be used to transfer to a GPU.
134 class HostAllocationPolicy
138 HostAllocationPolicy(PinningPolicy policy = PinningPolicy::CannotBePinned);
139 /*! \brief Return the alignment size currently used by the active pinning policy. */
140 std::size_t alignment();
141 /*! \brief Allocate and perhaps pin page-aligned memory suitable for
142 * e.g. GPU transfers.
144 * Before attempting to allocate, unpin() is called. After a
145 * successful allocation, pin() is called. (Whether these do
146 * things depends on the PinningPolicy that is in effect.)
148 * \param bytes Amount of memory (bytes) to allocate. It is valid to ask for
149 * 0 bytes, which will return a non-null pointer that is properly
150 * aligned and padded (but that you should not use).
152 * \return Valid pointer if the allocation+optional pinning worked, otherwise nullptr.
154 * \note Memory allocated with this routine must be released
155 * with gmx::HostAllocationPolicy::free(), and
156 * absolutely not the system free().
160 void *malloc(std::size_t bytes) const noexcept;
161 /*! \brief Free the memory, after unpinning (if appropriate).
163 * \param buffer Memory pointer previously returned from gmx::HostAllocationPolicy::malloc()
165 * \note This routine should only be called with pointers
166 * obtained from gmx:HostAllocationPolicy::malloc(),
167 * and absolutely not any pointers obtained the system
172 void free(void *buffer) const noexcept;
173 /*! \brief Pin the allocation to physical memory, if appropriate.
175 * If the allocation policy is not in pinning mode, or the
176 * allocation is empty, ot the allocation is already pinned,
181 void pin(void* p, size_t n) const noexcept;
182 /*! \brief Unpin the allocation, if appropriate.
184 * Regardless of the allocation policy, unpin the memory if
185 * previously pinned, otherwise do nothing.
189 void unpin(void* p) const noexcept;
190 /*! \brief Return the current pinning policy (which is semi-independent
191 * of whether the buffer is actually pinned).
195 PinningPolicy pinningPolicy() const { return pinningPolicy_; }
196 //! Don't propagate for copy
197 using propagate_on_container_copy_assignment = std::false_type;
198 //! Propagate for move
199 using propagate_on_container_move_assignment = std::true_type;
200 //! Propagate for move
201 using propagate_on_container_swap = std::true_type;
202 //! Use default allocator for copy (same as construct+copy)
203 template<typename U = void>
204 HostAllocationPolicy select_on_container_copy_construction() const
210 PinningPolicy pinningPolicy_;
213 /*! \brief Helper function for changing the pinning policy of a HostVector.
215 * If the vector has contents, then a full reallocation and buffer
216 * copy are needed if the policy change requires tighter restrictions,
217 * and desirable even if the policy change requires looser
218 * restrictions. That cost is OK, because GROMACS will do this
219 * operation very rarely (e.g. when auto-tuning and deciding to switch
220 * whether a task will run on a GPU, or not). */
222 void changePinningPolicy(HostVector<T> *v, PinningPolicy pinningPolicy)
224 if (v->get_allocator().pinningPolicy() == pinningPolicy)
228 //Force reallocation by creating copy
229 *v = HostVector<T>(*v, {pinningPolicy});