2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2020,2021, 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.
35 /*! \libinternal \file
36 * \brief Implements a GpuEventSynchronizer class for SYCL.
38 * This implementation relies on SYCL_INTEL_enqueue_barrier proposal,
39 * https://github.com/intel/llvm/blob/sycl/sycl/doc/extensions/EnqueueBarrier/enqueue_barrier.asciidoc
41 * Using event-based synchronization is not recommended for SYCL.
42 * SYCL queues are out-of-order and rely on data dependencies, allowing only to wait
43 * for a specific kernel (by capturing the \c event returned from \c queue.submit) or for all
44 * the tasks in the queue (\c queue.wait).
46 * \author Erik Lindahl <erik.lindahl@gmail.com>
47 * \author Andrey Alekseenko <al42and@gmail.com>
50 #ifndef GMX_GPU_UTILS_GPUEVENTSYNCHRONIZER_SYCL_H
51 #define GMX_GPU_UTILS_GPUEVENTSYNCHRONIZER_SYCL_H
55 #include "gromacs/gpu_utils/device_stream.h"
56 #include "gromacs/gpu_utils/gmxsycl.h"
57 #include "gromacs/utility/exceptions.h"
58 #include "gromacs/utility/gmxassert.h"
61 /*! \libinternal \brief
62 * A class which allows for CPU thread to mark and wait for certain GPU stream execution point.
63 * The event can be put into the stream with markEvent() and then later waited on with waitForEvent().
64 * This can be repeated as necessary, but the current implementation does not allow waiting on
65 * completed event more than once, expecting only exact pairs of markEvent(stream); waitForEvent().
66 * The class generally attempts to track the correctness of its state transitions, but
67 * please note that calling waitForEvent() right after the construction will fail with OpenCL
68 * and SYCL but succeed with CUDA.
70 * Another possible mode of operation can be implemented if needed:
71 * multiple calls to waitForEvent() after a single markEvent(). For this, event.reset() call
72 * from waitForEvent() should instead happen conditionally at the beginning of markEvent(), replacing
73 * the GMX_ASSERT(). This was tested to work both with CUDA, NVidia OpenCL, and Intel SYCL,
74 * but not with AMD/Intel OpenCl.
76 * \warning This class is offered for uniformity with other GPU implementations, but expect it to
77 * be deprecated in the future.
80 class GpuEventSynchronizer
84 GpuEventSynchronizer()
86 doNotSynchronizeBetweenStreams_ = (std::getenv("GMX_GPU_SYCL_NO_SYNCHRONIZE") != nullptr);
89 //! A constructor from an existing event.
90 GpuEventSynchronizer(const cl::sycl::event& event) : events_{ event } {}
92 ~GpuEventSynchronizer() = default;
94 GpuEventSynchronizer(const GpuEventSynchronizer&) = delete;
96 GpuEventSynchronizer& operator=(GpuEventSynchronizer&&) = delete;
97 //! Moving is disabled but can be considered in the future if needed
98 GpuEventSynchronizer(GpuEventSynchronizer&&) = delete;
100 /*! \brief Marks the synchronization point in the \p deviceStream.
101 * Should be called first and then followed by waitForEvent() or enqueueWaitEvent().
103 inline void markEvent(const DeviceStream& deviceStream)
105 GMX_ASSERT(!isMarked(), "Do not call markEvent more than once!");
106 # if GMX_SYCL_HIPSYCL
107 // Relies on HIPSYCL_EXT_QUEUE_WAIT_LIST extension
108 events_ = deviceStream.stream().get_wait_list();
110 // Relies on SYCL_INTEL_enqueue_barrier
111 events_ = { deviceStream.stream().submit_barrier() };
114 /*! \brief Synchronizes the host thread on the marked event.
115 * As in the OpenCL implementation, the event is released.
117 inline void waitForEvent()
119 GMX_ASSERT(isMarked(), "Don't call waitForEvent before marking the event!");
121 GMX_ASSERT(events_.size() == 1, "One event expected in DPCPP, but we have several!");
123 for (auto& event : events_)
125 event.wait_and_throw();
129 /*! \brief Checks the completion of the underlying event and resets the object if it was. */
130 inline bool isReady()
132 bool allReady = std::all_of(events_.begin(), events_.end(), [](cl::sycl::event& event) {
133 auto info = event.get_info<cl::sycl::info::event::command_execution_status>();
134 bool isComplete = (info == cl::sycl::info::event_command_status::complete);
143 /*! \brief Enqueues a wait for the recorded event in stream \p deviceStream.
144 * As in the OpenCL implementation, the event is released.
146 inline void enqueueWaitEvent(const DeviceStream& deviceStream)
148 if (!doNotSynchronizeBetweenStreams_)
150 # if GMX_SYCL_HIPSYCL
151 // Submit an empty kernel that depends on all the events recorded.
152 deviceStream.stream().single_task(events_, [=]() {});
154 // Relies on SYCL_INTEL_enqueue_barrier extensions
155 GMX_ASSERT(events_.size() == 1, "Only one event expected in DPCPP!");
156 deviceStream.stream().submit_barrier(events_);
161 //! Reset the event to unmarked state.
162 inline void reset() { events_.clear(); }
163 //! Check if the event is marked. Needed for some workarounds for #3988
164 inline bool isMarked() const { return !events_.empty(); }
167 std::vector<cl::sycl::event> events_;
168 /*! \brief Dev. setting to no-op enqueueWaitEvent
170 * In SYCL, dependencies between the GPU tasks are managed by the runtime, so manual
171 * synchronization between GPU streams should be redundant, but we keep it on by default.
173 * Setting this to \c true via \c GMX_GPU_SYCL_NO_SYNCHRONIZE environment variable will
174 * immediately return from \ref enqueueWaitEvent, without placing a barrier into the stream.
176 bool doNotSynchronizeBetweenStreams_;
179 #endif // !defined DOXYGEN
181 #endif // GMX_GPU_UTILS_GPUEVENTSYNCHRONIZER_SYCL_H