Simplified uniform GPU selection in CMake
[gromacs.git] / src / gromacs / gpu_utils / gpuregiontimer.h
blobf0860b164d9cbdbe61aade8bf3d5e127c5ed08cc
1 /*
2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2016,2017,2018,2019,2020, 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 /*! \libinternal \file
37 * \brief Defines the GPU region timer implementation/wrapper classes.
38 * The implementations live in gpuregiontimer.cuh for CUDA and gpuregiontimer_ocl.h for OpenCL.
40 * \author Aleksei Iupinov <a.yupinov@gmail.com>
43 #ifndef GMX_GPU_UTILS_GPUREGIONTIMER_H
44 #define GMX_GPU_UTILS_GPUREGIONTIMER_H
46 #include <string>
48 #include "gromacs/utility/gmxassert.h"
50 //! Debug GPU timers in debug builds only
51 #if defined(NDEBUG)
52 static const bool c_debugTimerState = false;
53 #else
54 static const bool c_debugTimerState = true;
55 #endif
57 /*! \libinternal \brief
58 * This is a GPU region timing wrapper class.
59 * It allows for host-side tracking of the accumulated execution timespans in GPU code
60 * (measuring kernel or transfers duration).
61 * It also partially tracks the correctness of the timer state transitions,
62 * as far as current implementation allows (see TODO in getLastRangeTime() for a disabled check).
63 * Internally it uses GpuRegionTimerImpl for measuring regions.
65 template<typename GpuRegionTimerImpl>
66 class GpuRegionTimerWrapper
68 //! The timer state used for debug-only assertions
69 enum class TimerState
71 Idle,
72 Recording,
73 Stopped
74 } debugState_ = TimerState::Idle;
76 //! The number of times the timespan has been measured
77 unsigned int callCount_ = 0;
78 //! The accumulated duration of the timespans measured (milliseconds)
79 double totalMilliseconds_ = 0.0;
80 //! The underlying region timer implementation
81 GpuRegionTimerImpl impl_;
83 public:
84 /*! \brief
85 * To be called before the region start.
87 * \param[in] deviceStream The GPU command stream where the event being measured takes place.
89 void openTimingRegion(const DeviceStream& deviceStream)
91 if (c_debugTimerState)
93 std::string error = "GPU timer should be idle, but is "
94 + std::string((debugState_ == TimerState::Stopped) ? "stopped" : "recording")
95 + ".";
96 GMX_ASSERT(debugState_ == TimerState::Idle, error.c_str());
97 debugState_ = TimerState::Recording;
99 impl_.openTimingRegion(deviceStream);
101 /*! \brief
102 * To be called after the region end.
104 * \param[in] deviceStream The GPU command stream where the event being measured takes place.
106 void closeTimingRegion(const DeviceStream& deviceStream)
108 if (c_debugTimerState)
110 std::string error = "GPU timer should be recording, but is "
111 + std::string((debugState_ == TimerState::Idle) ? "idle" : "stopped")
112 + ".";
113 GMX_ASSERT(debugState_ == TimerState::Recording, error.c_str());
114 debugState_ = TimerState::Stopped;
116 callCount_++;
117 impl_.closeTimingRegion(deviceStream);
119 /*! \brief
120 * Accumulates the last timespan of all the events used into the total duration,
121 * and resets the internal timer state.
122 * To be called after closeTimingRegion() and the command stream of the event having been
123 * synchronized. \returns The last timespan (in milliseconds).
125 double getLastRangeTime()
127 if (c_debugTimerState)
129 /* The assertion below is commented because it is currently triggered on a special case:
130 * the early return before the local non-bonded kernel launch if there is nothing to do.
131 * This can be reproduced in OpenCL build by running
132 * mdrun-mpi-test -ntmpi 2 --gtest_filter=*Empty*
133 * Similarly, the GpuRegionTimerImpl suffers from non-nullptr
134 * cl_event conditionals which ideally should only be asserts.
135 * TODO: improve internal task scheduling, re-enable the assert, turn conditionals into asserts
138 std::string error = "GPU timer should be stopped, but is " + std::string((debugState_ == TimerState::Idle) ? "idle" : "recording") + ".";
139 GMX_ASSERT(debugState_ == TimerState::Stopped, error.c_str());
141 debugState_ = TimerState::Idle;
143 double milliseconds = impl_.getLastRangeTime();
144 totalMilliseconds_ += milliseconds;
145 return milliseconds;
147 /*! \brief Resets the implementation and total time/call count to zeroes. */
148 void reset()
150 if (c_debugTimerState)
152 debugState_ = TimerState::Idle;
154 totalMilliseconds_ = 0.0;
155 callCount_ = 0;
156 impl_.reset();
158 /*! \brief Gets total time recorded (in milliseconds). */
159 double getTotalTime() const { return totalMilliseconds_; }
160 /*! \brief Gets total call count recorded. */
161 unsigned int getCallCount() const { return callCount_; }
162 /*! \brief
163 * Gets a pointer to a new timing event for passing into individual GPU API calls
164 * within the region if they require it (e.g. on OpenCL).
165 * \returns The pointer to the underlying single command timing event.
167 CommandEvent* fetchNextEvent()
169 if (c_debugTimerState)
171 std::string error = "GPU timer should be recording, but is "
172 + std::string((debugState_ == TimerState::Idle) ? "idle" : "stopped")
173 + ".";
174 GMX_ASSERT(debugState_ == TimerState::Recording, error.c_str());
176 return impl_.fetchNextEvent();
180 #endif