2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2012,2014,2015,2016,2017 by the GROMACS development team.
5 * Copyright (c) 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.
36 #ifndef GMX_GPU_UTILS_CUDAUTILS_CUH
37 #define GMX_GPU_UTILS_CUDAUTILS_CUH
43 #include <type_traits>
45 #include "gromacs/gpu_utils/device_stream.h"
46 #include "gromacs/gpu_utils/gputraits.cuh"
47 #include "gromacs/math/vec.h"
48 #include "gromacs/math/vectypes.h"
49 #include "gromacs/utility/exceptions.h"
50 #include "gromacs/utility/fatalerror.h"
51 #include "gromacs/utility/gmxassert.h"
52 #include "gromacs/utility/stringutil.h"
59 /*! \brief Add the API information on the specific error to the error message.
61 * \param[in] deviceError The error to assert cudaSuccess on.
63 * \returns A description of the API error. Returns '(CUDA error #0 (cudaSuccess): no error)' in case deviceError is cudaSuccess.
65 inline std::string getDeviceErrorString(const cudaError_t deviceError)
67 return formatString("CUDA error #%d (%s): %s.",
69 cudaGetErrorName(deviceError),
70 cudaGetErrorString(deviceError));
73 /*! \brief Check if API returned an error and throw an exception with information on it.
75 * \param[in] deviceError The error to assert cudaSuccess on.
76 * \param[in] errorMessage Undecorated error message.
78 * \throws InternalError if deviceError is not a success.
80 inline void checkDeviceError(const cudaError_t deviceError, const std::string& errorMessage)
82 if (deviceError != cudaSuccess)
84 GMX_THROW(gmx::InternalError(errorMessage + " " + getDeviceErrorString(deviceError)));
88 /*! \brief Helper function to ensure no pending error silently
89 * disrupts error handling.
91 * Asserts in a debug build if an unhandled error is present. Issues a
92 * warning at run time otherwise.
94 * \param[in] errorMessage Undecorated error message.
96 inline void ensureNoPendingDeviceError(const std::string& errorMessage)
98 // Ensure there is no pending error that would otherwise affect
99 // the behaviour of future error handling.
100 cudaError_t deviceError = cudaGetLastError();
101 if (deviceError == cudaSuccess)
106 // If we would find an error in a release build, we do not know
107 // what is appropriate to do about it, so assert only for debug
109 const std::string fullErrorMessage =
110 errorMessage + " An unhandled error from a previous CUDA operation was detected. "
111 + gmx::getDeviceErrorString(deviceError);
112 GMX_ASSERT(deviceError == cudaSuccess, fullErrorMessage.c_str());
113 // TODO When we evolve a better logging framework, use that
114 // for release-build error reporting.
115 gmx_warning("%s", fullErrorMessage.c_str());
121 enum class GpuApiCallBehavior;
123 /* TODO error checking needs to be rewritten. We have 2 types of error checks needed
124 based on where they occur in the code:
125 - non performance-critical: these errors are unsafe to be ignored and must be
126 _always_ checked for, e.g. initializations
127 - performance critical: handling errors might hurt performance so care need to be taken
128 when/if we should check for them at all, e.g. in cu_upload_X. However, we should be
129 able to turn the check for these errors on!
131 Probably we'll need two sets of the macros below...
134 #define CHECK_CUDA_ERRORS
136 #ifdef CHECK_CUDA_ERRORS
138 /*! Check for CUDA error on the return status of a CUDA RT API call. */
139 # define CU_RET_ERR(deviceError, msg) \
142 if ((deviceError) != cudaSuccess) \
144 gmx_fatal(FARGS, "%s\n", ((msg) + gmx::getDeviceErrorString(deviceError)).c_str()); \
148 #else /* CHECK_CUDA_ERRORS */
150 # define CU_RET_ERR(status, msg) \
155 #endif /* CHECK_CUDA_ERRORS */
157 // TODO: the 2 functions below are pretty much a constructor/destructor of a simple
158 // GPU table object. There is also almost self-contained fetchFromParamLookupTable()
159 // in cuda_kernel_utils.cuh. They could all live in a separate class/struct file.
161 /*! \brief Add a triplets stored in a float3 to an rvec variable.
163 * \param[out] a Rvec to increment
164 * \param[in] b Float triplet to increment with.
166 static inline void rvec_inc(rvec a, const float3 b)
168 rvec tmp = { b.x, b.y, b.z };
172 /*! \brief Returns true if all tasks in \p s have completed.
174 * \param[in] deviceStream CUDA stream to check.
176 * \returns True if all tasks enqueued in the stream \p deviceStream (at the time of this call) have completed.
178 static inline bool haveStreamTasksCompleted(const DeviceStream& deviceStream)
180 cudaError_t stat = cudaStreamQuery(deviceStream.stream());
182 if (stat == cudaErrorNotReady)
184 // work is still in progress in the stream
188 GMX_ASSERT(stat != cudaErrorInvalidResourceHandle,
189 ("Stream identifier not valid. " + gmx::getDeviceErrorString(stat)).c_str());
191 // cudaSuccess and cudaErrorNotReady are the expected return values
192 CU_RET_ERR(stat, "Unexpected cudaStreamQuery failure. ");
194 GMX_ASSERT(stat == cudaSuccess,
195 ("Values other than cudaSuccess should have been explicitly handled. "
196 + gmx::getDeviceErrorString(stat))
202 /* Kernel launch helpers */
205 * A function for setting up a single CUDA kernel argument.
206 * This is the tail of the compile-time recursive function below.
207 * It has to be seen by the compiler first.
209 * \tparam totalArgsCount Number of the kernel arguments
210 * \tparam KernelPtr Kernel function handle type
211 * \param[in] argIndex Index of the current argument
213 template<size_t totalArgsCount, typename KernelPtr>
214 void prepareGpuKernelArgument(KernelPtr /*kernel*/,
215 std::array<void*, totalArgsCount>* /* kernelArgsPtr */,
216 size_t gmx_used_in_debug argIndex)
218 GMX_ASSERT(argIndex == totalArgsCount, "Tail expansion");
222 * Compile-time recursive function for setting up a single CUDA kernel argument.
223 * This function copies a kernel argument pointer \p argPtr into \p kernelArgsPtr,
224 * and calls itself on the next argument, eventually calling the tail function above.
226 * \tparam CurrentArg Type of the current argument
227 * \tparam RemainingArgs Types of remaining arguments after the current one
228 * \tparam totalArgsCount Number of the kernel arguments
229 * \tparam KernelPtr Kernel function handle type
230 * \param[in] kernel Kernel function handle
231 * \param[in,out] kernelArgsPtr Pointer to the argument array to be filled in
232 * \param[in] argIndex Index of the current argument
233 * \param[in] argPtr Pointer to the current argument
234 * \param[in] otherArgsPtrs Pack of pointers to arguments remaining to process after the current one
236 template<typename CurrentArg, typename... RemainingArgs, size_t totalArgsCount, typename KernelPtr>
237 void prepareGpuKernelArgument(KernelPtr kernel,
238 std::array<void*, totalArgsCount>* kernelArgsPtr,
240 const CurrentArg* argPtr,
241 const RemainingArgs*... otherArgsPtrs)
243 (*kernelArgsPtr)[argIndex] = const_cast<void*>(static_cast<const void*>(argPtr));
244 prepareGpuKernelArgument(kernel, kernelArgsPtr, argIndex + 1, otherArgsPtrs...);
248 * A wrapper function for setting up all the CUDA kernel arguments.
249 * Calls the recursive functions above.
251 * \tparam KernelPtr Kernel function handle type
252 * \tparam Args Types of all the kernel arguments
253 * \param[in] kernel Kernel function handle
254 * \param[in] argsPtrs Pointers to all the kernel arguments
255 * \returns A prepared parameter pack to be used with launchGpuKernel() as the last argument.
257 template<typename KernelPtr, typename... Args>
258 std::array<void*, sizeof...(Args)> prepareGpuKernelArguments(KernelPtr kernel,
259 const KernelLaunchConfig& /*config */,
260 const Args*... argsPtrs)
262 std::array<void*, sizeof...(Args)> kernelArgs;
263 prepareGpuKernelArgument(kernel, &kernelArgs, 0, argsPtrs...);
267 /*! \brief Launches the CUDA kernel and handles the errors.
269 * \tparam Args Types of all the kernel arguments
270 * \param[in] kernel Kernel function handle
271 * \param[in] config Kernel configuration for launching
272 * \param[in] deviceStream GPU stream to launch kernel in
273 * \param[in] kernelName Human readable kernel description, for error handling only
274 * \param[in] kernelArgs Array of the pointers to the kernel arguments, prepared by
275 * prepareGpuKernelArguments() \throws gmx::InternalError on kernel launch failure
277 template<typename... Args>
278 void launchGpuKernel(void (*kernel)(Args...),
279 const KernelLaunchConfig& config,
280 const DeviceStream& deviceStream,
281 CommandEvent* /*timingEvent */,
282 const char* kernelName,
283 const std::array<void*, sizeof...(Args)>& kernelArgs)
285 dim3 blockSize(config.blockSize[0], config.blockSize[1], config.blockSize[2]);
286 dim3 gridSize(config.gridSize[0], config.gridSize[1], config.gridSize[2]);
287 cudaLaunchKernel(reinterpret_cast<void*>(kernel),
290 const_cast<void**>(kernelArgs.data()),
291 config.sharedMemorySize,
292 deviceStream.stream());
294 gmx::ensureNoPendingDeviceError("GPU kernel (" + std::string(kernelName)
295 + ") failed to launch.");