2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2010,2011,2012,2013,2014,2015,2016, The GROMACS development team.
5 * Copyright (c) 2017,2018,2019,2020,2021, by the GROMACS development team, led by
6 * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
7 * and including many others, as listed in the AUTHORS file in the
8 * top-level source directory and at http://www.gromacs.org.
10 * GROMACS is free software; you can redistribute it and/or
11 * modify it under the terms of the GNU Lesser General Public License
12 * as published by the Free Software Foundation; either version 2.1
13 * of the License, or (at your option) any later version.
15 * GROMACS is distributed in the hope that it will be useful,
16 * but WITHOUT ANY WARRANTY; without even the implied warranty of
17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
18 * Lesser General Public License for more details.
20 * You should have received a copy of the GNU Lesser General Public
21 * License along with GROMACS; if not, see
22 * http://www.gnu.org/licenses, or write to the Free Software Foundation,
23 * Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
25 * If you want to redistribute modifications to GROMACS, please
26 * consider that scientific software is very special. Version
27 * control is crucial - bugs must be traceable. We will be happy to
28 * consider code for inclusion in the official distribution, but
29 * derived work must not be called official GROMACS. Details are found
30 * in the README & COPYING files - if they are missing, get the
31 * official version at http://www.gromacs.org.
33 * To help us fund GROMACS development, we humbly ask that you cite
34 * the research papers on the package. Check out http://www.gromacs.org.
37 * \brief Define functions for detection and initialization for CUDA devices.
39 * \author Szilard Pall <pall.szilard@gmail.com>
44 #include "gpu_utils.h"
50 #include <cuda_profiler_api.h>
52 #include "gromacs/gpu_utils/cudautils.cuh"
53 #include "gromacs/gpu_utils/device_context.h"
54 #include "gromacs/gpu_utils/device_stream.h"
55 #include "gromacs/hardware/device_information.h"
56 #include "gromacs/hardware/device_management.h"
57 #include "gromacs/utility/basedefinitions.h"
58 #include "gromacs/utility/cstringutil.h"
59 #include "gromacs/utility/exceptions.h"
60 #include "gromacs/utility/fatalerror.h"
61 #include "gromacs/utility/gmxassert.h"
62 #include "gromacs/utility/logger.h"
63 #include "gromacs/utility/programcontext.h"
64 #include "gromacs/utility/smalloc.h"
65 #include "gromacs/utility/snprintf.h"
66 #include "gromacs/utility/stringutil.h"
68 static bool cudaProfilerRun = ((getenv("NVPROF_ID") != nullptr));
70 bool isHostMemoryPinned(const void* h_ptr)
72 cudaPointerAttributes memoryAttributes;
73 cudaError_t stat = cudaPointerGetAttributes(&memoryAttributes, h_ptr);
75 bool isPinned = false;
79 // In CUDA 11.0, the field called memoryType in
80 // cudaPointerAttributes was replaced by a field called
81 // type, along with a documented change of behavior when the
82 // pointer passed to cudaPointerGetAttributes is to
83 // non-registered host memory. That change means that this
84 // code needs conditional compilation and different
85 // execution paths to function with all supported versions.
86 #if CUDART_VERSION < 11 * 1000
89 isPinned = (memoryAttributes.type == cudaMemoryTypeHost);
93 case cudaErrorInvalidValue:
94 // If the buffer was not pinned, then it will not be recognized by CUDA at all
96 // Reset the last error status
100 default: CU_RET_ERR(stat, "Unexpected CUDA error");
105 void startGpuProfiler(void)
107 /* The NVPROF_ID environment variable is set by nvprof and indicates that
108 mdrun is executed in the CUDA profiler.
109 If nvprof was run is with "--profile-from-start off", the profiler will
110 be started here. This way we can avoid tracing the CUDA events from the
111 first part of the run. Starting the profiler again does nothing.
116 stat = cudaProfilerStart();
117 CU_RET_ERR(stat, "cudaProfilerStart failed");
121 void stopGpuProfiler(void)
123 /* Stopping the nvidia here allows us to eliminate the subsequent
124 API calls from the trace, e.g. uninitialization and cleanup. */
128 stat = cudaProfilerStop();
129 CU_RET_ERR(stat, "cudaProfilerStop failed");
133 void resetGpuProfiler(void)
135 /* With CUDA <=7.5 the profiler can't be properly reset; we can only start
136 * the profiling here (can't stop it) which will achieve the desired effect if
137 * the run was started with the profiling disabled.
139 * TODO: add a stop (or replace it with reset) when this will work correctly in CUDA.
148 /*! \brief Check and act on status returned from peer access CUDA call
150 * If status is "cudaSuccess", we continue. If
151 * "cudaErrorPeerAccessAlreadyEnabled", then peer access has already
152 * been enabled so we ignore. If "cudaErrorInvalidDevice" then the
153 * run is trying to access an invalid GPU, so we throw an error. If
154 * "cudaErrorInvalidValue" then there is a problem with the arguments
155 * to the CUDA call, and we throw an error. These cover all expected
156 * statuses, but if any other is returned we issue a warning and
159 * \param[in] stat CUDA call return status
160 * \param[in] gpuA ID for GPU initiating peer access call
161 * \param[in] gpuB ID for remote GPU
162 * \param[in] mdlog Logger object
163 * \param[in] cudaCallName name of CUDA peer access call
165 static void peerAccessCheckStat(const cudaError_t stat,
168 const gmx::MDLogger& mdlog,
169 const char* cudaCallName)
172 if (stat == cudaErrorPeerAccessAlreadyEnabled)
174 // Since peer access has already been enabled, this error can safely be ignored.
175 // Now clear the error internally within CUDA:
179 if ((stat == cudaErrorInvalidDevice) || (stat == cudaErrorInvalidValue))
181 std::string errorString =
182 gmx::formatString("%s from GPU %d to GPU %d failed", cudaCallName, gpuA, gpuB);
183 CU_RET_ERR(stat, errorString.c_str());
185 if (stat != cudaSuccess)
187 GMX_LOG(mdlog.warning)
189 .appendTextFormatted(
190 "GPU peer access not enabled between GPUs %d and %d due to unexpected "
191 "return value from %s. %s",
195 gmx::getDeviceErrorString(stat).c_str());
196 // Clear the error internally within CUDA
201 void setupGpuDevicePeerAccess(const std::vector<int>& gpuIdsToUse, const gmx::MDLogger& mdlog)
205 // take a note of currently-set GPU
207 stat = cudaGetDevice(¤tGpu);
208 CU_RET_ERR(stat, "cudaGetDevice in setupGpuDevicePeerAccess failed");
210 std::string message = gmx::formatString(
211 "Note: Peer access enabled between the following GPU pairs in the node:\n ");
212 bool peerAccessEnabled = false;
214 for (unsigned int i = 0; i < gpuIdsToUse.size(); i++)
216 int gpuA = gpuIdsToUse[i];
217 stat = cudaSetDevice(gpuA);
218 if (stat != cudaSuccess)
220 GMX_LOG(mdlog.warning)
222 .appendTextFormatted(
223 "GPU peer access not enabled due to unexpected return value from "
224 "cudaSetDevice(%d). %s",
226 gmx::getDeviceErrorString(stat).c_str());
229 for (unsigned int j = 0; j < gpuIdsToUse.size(); j++)
233 int gpuB = gpuIdsToUse[j];
234 int canAccessPeer = 0;
235 stat = cudaDeviceCanAccessPeer(&canAccessPeer, gpuA, gpuB);
236 peerAccessCheckStat(stat, gpuA, gpuB, mdlog, "cudaDeviceCanAccessPeer");
240 stat = cudaDeviceEnablePeerAccess(gpuB, 0);
241 peerAccessCheckStat(stat, gpuA, gpuB, mdlog, "cudaDeviceEnablePeerAccess");
243 message = gmx::formatString("%s%d->%d ", message.c_str(), gpuA, gpuB);
244 peerAccessEnabled = true;
250 // re-set GPU to that originally set
251 stat = cudaSetDevice(currentGpu);
252 if (stat != cudaSuccess)
254 CU_RET_ERR(stat, "cudaSetDevice in setupGpuDevicePeerAccess failed");
258 if (peerAccessEnabled)
260 GMX_LOG(mdlog.info).asParagraph().appendTextFormatted("%s", message.c_str());