2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2014,2015,2016,2017,2018 by the GROMACS development team.
5 * Copyright (c) 2019,2020, 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 /*! \libinternal \file
37 * \brief Declare utility routines for OpenCL
39 * \author Anca Hamuraru <anca@streamcomputing.eu>
42 #ifndef GMX_GPU_UTILS_OCLUTILS_H
43 #define GMX_GPU_UTILS_OCLUTILS_H
47 #include "gromacs/gpu_utils/device_context.h"
48 #include "gromacs/gpu_utils/device_stream.h"
49 #include "gromacs/gpu_utils/gmxopencl.h"
50 #include "gromacs/gpu_utils/gputraits_ocl.h"
51 #include "gromacs/utility/exceptions.h"
52 #include "gromacs/utility/gmxassert.h"
53 #include "gromacs/utility/stringutil.h"
55 enum class GpuApiCallBehavior;
58 * \brief OpenCL GPU runtime data
60 * The device runtime data is meant to hold objects associated with a GROMACS rank's
61 * (thread or process) use of a single device (multiple devices per rank is not
62 * implemented). These objects should be constructed at ther point where a device
63 * dets assigned to a rank and released at when this assignment is no longer valid
64 * (i.e. at cleanup in the current implementation).
67 struct gmx_device_runtime_data_t
73 /*! \brief Allocate host memory in malloc style */
74 void pmalloc(void** h_ptr, size_t nbytes);
76 /*! \brief Free host memory in malloc style */
77 void pfree(void* h_ptr);
79 /*! \brief Convert error code to diagnostic string */
80 std::string ocl_get_error_string(cl_int error);
82 //! A debug checker to track cl_events being released correctly
83 inline void ensureReferenceCount(const cl_event& event, unsigned int expectedRefCount)
87 cl_int clError = clGetEventInfo(event, CL_EVENT_REFERENCE_COUNT, sizeof(refCount), &refCount, nullptr);
88 GMX_ASSERT(CL_SUCCESS == clError, ocl_get_error_string(clError).c_str());
89 GMX_ASSERT(refCount == expectedRefCount, "Unexpected reference count");
91 GMX_UNUSED_VALUE(event);
92 GMX_UNUSED_VALUE(expectedRefCount);
96 /*! \brief Pretend to synchronize an OpenCL stream (dummy implementation).
98 * \returns Not implemented in OpenCL.
100 static inline bool haveStreamTasksCompleted(const DeviceStream& /* deviceStream */)
102 GMX_RELEASE_ASSERT(false, "haveStreamTasksCompleted is not implemented for OpenCL");
106 /* Kernel launch helpers */
109 * A function for setting up a single OpenCL kernel argument.
110 * This is the tail of the compile-time recursive function below.
111 * It has to be seen by the compiler first.
112 * As NB kernels might be using dynamic local memory as the last argument,
113 * this function also manages that, using sharedMemorySize from \p config.
115 * \param[in] kernel Kernel function handle
116 * \param[in] config Kernel configuration for launching
117 * \param[in] argIndex Index of the current argument
119 void inline prepareGpuKernelArgument(cl_kernel kernel, const KernelLaunchConfig& config, size_t argIndex)
121 if (config.sharedMemorySize > 0)
123 cl_int gmx_used_in_debug clError =
124 clSetKernelArg(kernel, argIndex, config.sharedMemorySize, nullptr);
125 GMX_ASSERT(CL_SUCCESS == clError, ocl_get_error_string(clError).c_str());
130 * Compile-time recursive function for setting up a single OpenCL kernel argument.
131 * This function uses one kernel argument pointer \p argPtr to call clSetKernelArg(),
132 * and calls itself on the next argument, eventually calling the tail function above.
134 * \tparam CurrentArg Type of the current argument
135 * \tparam RemainingArgs Types of remaining arguments after the current one
136 * \param[in] kernel Kernel function handle
137 * \param[in] config Kernel configuration for launching
138 * \param[in] argIndex Index of the current argument
139 * \param[in] argPtr Pointer to the current argument
140 * \param[in] otherArgsPtrs Pack of pointers to arguments remaining to process after the current one
142 template<typename CurrentArg, typename... RemainingArgs>
143 void prepareGpuKernelArgument(cl_kernel kernel,
144 const KernelLaunchConfig& config,
146 const CurrentArg* argPtr,
147 const RemainingArgs*... otherArgsPtrs)
149 cl_int gmx_used_in_debug clError = clSetKernelArg(kernel, argIndex, sizeof(CurrentArg), argPtr);
150 GMX_ASSERT(CL_SUCCESS == clError, ocl_get_error_string(clError).c_str());
152 // Assert on types not allowed to be passed to a kernel
153 // (as per section 6.9 of the OpenCL spec).
155 !std::is_same_v<CurrentArg,
156 bool> && !std::is_same_v<CurrentArg, size_t> && !std::is_same_v<CurrentArg, ptrdiff_t> && !std::is_same_v<CurrentArg, intptr_t> && !std::is_same_v<CurrentArg, uintptr_t>,
157 "Invalid type passed to OpenCL kernel functions (see OpenCL spec section 6.9).");
159 prepareGpuKernelArgument(kernel, config, argIndex + 1, otherArgsPtrs...);
163 * A wrapper function for setting up all the OpenCL kernel arguments.
164 * Calls the recursive functions above.
166 * \tparam Args Types of all the kernel arguments
167 * \param[in] kernel Kernel function handle
168 * \param[in] config Kernel configuration for launching
169 * \param[in] argsPtrs Pointers to all the kernel arguments
170 * \returns A handle for the prepared parameter pack to be used with launchGpuKernel() as the last argument
171 * - currently always nullptr for OpenCL, as it manages kernel/arguments association by itself.
173 template<typename... Args>
174 void* prepareGpuKernelArguments(cl_kernel kernel, const KernelLaunchConfig& config, const Args*... argsPtrs)
176 prepareGpuKernelArgument(kernel, config, 0, argsPtrs...);
180 /*! \brief Launches the OpenCL kernel and handles the errors.
182 * \param[in] kernel Kernel function handle
183 * \param[in] config Kernel configuration for launching
184 * \param[in] deviceStream GPU stream to launch kernel in
185 * \param[in] timingEvent Timing event, fetched from GpuRegionTimer
186 * \param[in] kernelName Human readable kernel description, for error handling only
187 * \throws gmx::InternalError on kernel launch failure
189 inline void launchGpuKernel(cl_kernel kernel,
190 const KernelLaunchConfig& config,
191 const DeviceStream& deviceStream,
192 CommandEvent* timingEvent,
193 const char* kernelName,
194 const void* /*kernelArgs*/)
196 const int workDimensions = 3;
197 const size_t* globalWorkOffset = nullptr;
198 const size_t waitListSize = 0;
199 const cl_event* waitList = nullptr;
200 size_t globalWorkSize[3];
201 for (int i = 0; i < workDimensions; i++)
203 globalWorkSize[i] = config.gridSize[i] * config.blockSize[i];
205 cl_int clError = clEnqueueNDRangeKernel(deviceStream.stream(), kernel, workDimensions,
206 globalWorkOffset, globalWorkSize, config.blockSize,
207 waitListSize, waitList, timingEvent);
208 if (CL_SUCCESS != clError)
210 const std::string errorMessage = "GPU kernel (" + std::string(kernelName)
211 + ") failed to launch: " + ocl_get_error_string(clError);
212 GMX_THROW(gmx::InternalError(errorMessage));