2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2014,2015,2016,2017,2018,2019, 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 Declare utility routines for OpenCL
38 * \author Anca Hamuraru <anca@streamcomputing.eu>
41 #ifndef GMX_GPU_UTILS_OCLUTILS_H
42 #define GMX_GPU_UTILS_OCLUTILS_H
46 #include "gromacs/gpu_utils/gmxopencl.h"
47 #include "gromacs/gpu_utils/gputraits_ocl.h"
48 #include "gromacs/utility/exceptions.h"
49 #include "gromacs/utility/gmxassert.h"
51 enum class GpuApiCallBehavior;
53 /*! \brief OpenCL vendor IDs */
56 OCL_VENDOR_NVIDIA = 0,
63 * \brief OpenCL GPU device identificator
65 * An OpenCL device is identified by its ID.
66 * The platform ID is also included for caching reasons.
70 cl_platform_id ocl_platform_id; /**< Platform ID */
71 cl_device_id ocl_device_id; /**< Device ID */
75 * \brief OpenCL device information.
77 * The OpenCL device information is queried and set at detection and contains
78 * both information about the device/hardware returned by the runtime as well
79 * as additional data like support status.
81 struct gmx_device_info_t
83 ocl_gpu_id_t ocl_gpu_id; /**< device ID assigned at detection */
84 char device_name[256]; /**< device name */
85 char device_version[256]; /**< device version */
86 char device_vendor[256]; /**< device vendor */
87 int compute_units; /**< number of compute units */
88 int adress_bits; /**< number of adress bits the device is capable of */
89 int stat; /**< device status takes values of e_gpu_detect_res_t */
90 ocl_vendor_id_t vendor_e; /**< device vendor as defined by ocl_vendor_id_t */
91 size_t maxWorkItemSizes[3]; /**< workgroup size limits (CL_DEVICE_MAX_WORK_ITEM_SIZES) */
92 size_t maxWorkGroupSize; /**< workgroup total size limit (CL_DEVICE_MAX_WORK_GROUP_SIZE) */
96 * \brief OpenCL GPU runtime data
98 * The device runtime data is meant to hold objects associated with a GROMACS rank's
99 * (thread or process) use of a single device (multiple devices per rank is not
100 * implemented). These objects should be constructed at ther point where a device
101 * dets assigned to a rank and released at when this assignment is no longer valid
102 * (i.e. at cleanup in the current implementation).
105 struct gmx_device_runtime_data_t
107 cl_context context; /**< OpenCL context */
108 cl_program program; /**< OpenCL program */
111 /*! \brief Launches synchronous or asynchronous device to host memory copy.
113 * If copy_event is not NULL, on return it will contain an event object
114 * identifying this particular device to host operation. The event can further
115 * be used to queue a wait for this operation or to query profiling information.
117 int ocl_copy_D2H(void* h_dest,
121 GpuApiCallBehavior transferKind,
122 cl_command_queue command_queue,
123 cl_event* copy_event);
126 /*! \brief Launches asynchronous device to host memory copy. */
127 int ocl_copy_D2H_async(void* h_dest,
131 cl_command_queue command_queue,
132 cl_event* copy_event);
134 /*! \brief Launches synchronous or asynchronous host to device memory copy.
136 * If copy_event is not NULL, on return it will contain an event object
137 * identifying this particular host to device operation. The event can further
138 * be used to queue a wait for this operation or to query profiling information.
140 int ocl_copy_H2D(cl_mem d_dest,
144 GpuApiCallBehavior transferKind,
145 cl_command_queue command_queue,
146 cl_event* copy_event);
148 /*! \brief Launches asynchronous host to device memory copy. */
149 int ocl_copy_H2D_async(cl_mem d_dest,
153 cl_command_queue command_queue,
154 cl_event* copy_event);
156 /*! \brief Launches synchronous host to device memory copy. */
157 int ocl_copy_H2D_sync(cl_mem d_dest, const void* h_src, size_t offset, size_t bytes, cl_command_queue command_queue);
159 /*! \brief Allocate host memory in malloc style */
160 void pmalloc(void** h_ptr, size_t nbytes);
162 /*! \brief Free host memory in malloc style */
163 void pfree(void* h_ptr);
165 /*! \brief Convert error code to diagnostic string */
166 std::string ocl_get_error_string(cl_int error);
168 /*! \brief Calls clFinish() in the stream \p s.
170 * \param[in] s stream to synchronize with
172 static inline void gpuStreamSynchronize(cl_command_queue s)
174 cl_int cl_error = clFinish(s);
175 GMX_RELEASE_ASSERT(CL_SUCCESS == cl_error,
176 ("Error caught during clFinish:" + ocl_get_error_string(cl_error)).c_str());
179 //! A debug checker to track cl_events being released correctly
180 inline void ensureReferenceCount(const cl_event& event, unsigned int refCount)
183 cl_int clError = clGetEventInfo(event, CL_EVENT_REFERENCE_COUNT, sizeof(refCount), &refCount, nullptr);
184 GMX_ASSERT(CL_SUCCESS == clError, ocl_get_error_string(clError).c_str());
185 GMX_ASSERT(refCount == refCount, "Unexpected reference count");
187 GMX_UNUSED_VALUE(event);
188 GMX_UNUSED_VALUE(refCount);
192 /*! \brief Pretend to synchronize an OpenCL stream (dummy implementation).
194 * \param[in] s queue to check
196 * \returns True if all tasks enqueued in the stream \p s (at the time of this call) have completed.
198 static inline bool haveStreamTasksCompleted(cl_command_queue gmx_unused s)
200 GMX_RELEASE_ASSERT(false, "haveStreamTasksCompleted is not implemented for OpenCL");
204 /* Kernel launch helpers */
207 * A function for setting up a single OpenCL kernel argument.
208 * This is the tail of the compile-time recursive function below.
209 * It has to be seen by the compiler first.
210 * As NB kernels might be using dynamic local memory as the last argument,
211 * this function also manages that, using sharedMemorySize from \p config.
213 * \param[in] kernel Kernel function handle
214 * \param[in] config Kernel configuration for launching
215 * \param[in] argIndex Index of the current argument
217 void inline prepareGpuKernelArgument(cl_kernel kernel, const KernelLaunchConfig& config, size_t argIndex)
219 if (config.sharedMemorySize > 0)
221 cl_int gmx_used_in_debug clError =
222 clSetKernelArg(kernel, argIndex, config.sharedMemorySize, nullptr);
223 GMX_ASSERT(CL_SUCCESS == clError, ocl_get_error_string(clError).c_str());
228 * Compile-time recursive function for setting up a single OpenCL kernel argument.
229 * This function uses one kernel argument pointer \p argPtr to call clSetKernelArg(),
230 * and calls itself on the next argument, eventually calling the tail function above.
232 * \tparam CurrentArg Type of the current argument
233 * \tparam RemainingArgs Types of remaining arguments after the current one
234 * \param[in] kernel Kernel function handle
235 * \param[in] config Kernel configuration for launching
236 * \param[in] argIndex Index of the current argument
237 * \param[in] argPtr Pointer to the current argument
238 * \param[in] otherArgsPtrs Pack of pointers to arguments remaining to process after the current one
240 template<typename CurrentArg, typename... RemainingArgs>
241 void prepareGpuKernelArgument(cl_kernel kernel,
242 const KernelLaunchConfig& config,
244 const CurrentArg* argPtr,
245 const RemainingArgs*... otherArgsPtrs)
247 cl_int gmx_used_in_debug clError = clSetKernelArg(kernel, argIndex, sizeof(CurrentArg), argPtr);
248 GMX_ASSERT(CL_SUCCESS == clError, ocl_get_error_string(clError).c_str());
250 // Assert on types not allowed to be passed to a kernel
251 // (as per section 6.9 of the OpenCL spec).
252 static_assert(!std::is_same<CurrentArg, bool>::value && !std::is_same<CurrentArg, size_t>::value
253 && !std::is_same<CurrentArg, ptrdiff_t>::value
254 && !std::is_same<CurrentArg, intptr_t>::value
255 && !std::is_same<CurrentArg, uintptr_t>::value,
256 "Invalid type passed to OpenCL kernel functions (see OpenCL spec section 6.9).");
258 prepareGpuKernelArgument(kernel, config, argIndex + 1, otherArgsPtrs...);
262 * A wrapper function for setting up all the OpenCL kernel arguments.
263 * Calls the recursive functions above.
265 * \tparam Args Types of all the kernel arguments
266 * \param[in] kernel Kernel function handle
267 * \param[in] config Kernel configuration for launching
268 * \param[in] argsPtrs Pointers to all the kernel arguments
269 * \returns A handle for the prepared parameter pack to be used with launchGpuKernel() as the last argument
270 * - currently always nullptr for OpenCL, as it manages kernel/arguments association by itself.
272 template<typename... Args>
273 void* prepareGpuKernelArguments(cl_kernel kernel, const KernelLaunchConfig& config, const Args*... argsPtrs)
275 prepareGpuKernelArgument(kernel, config, 0, argsPtrs...);
279 /*! \brief Launches the OpenCL kernel and handles the errors.
281 * \param[in] kernel Kernel function handle
282 * \param[in] config Kernel configuration for launching
283 * \param[in] timingEvent Timing event, fetched from GpuRegionTimer
284 * \param[in] kernelName Human readable kernel description, for error handling only
285 * \throws gmx::InternalError on kernel launch failure
287 inline void launchGpuKernel(cl_kernel kernel,
288 const KernelLaunchConfig& config,
289 CommandEvent* timingEvent,
290 const char* kernelName,
291 const void* /*kernelArgs*/)
293 const int workDimensions = 3;
294 const size_t* globalWorkOffset = nullptr;
295 const size_t waitListSize = 0;
296 const cl_event* waitList = nullptr;
297 size_t globalWorkSize[3];
298 for (int i = 0; i < workDimensions; i++)
300 globalWorkSize[i] = config.gridSize[i] * config.blockSize[i];
302 cl_int clError = clEnqueueNDRangeKernel(config.stream, kernel, workDimensions, globalWorkOffset,
303 globalWorkSize, config.blockSize, waitListSize,
304 waitList, timingEvent);
305 if (CL_SUCCESS != clError)
307 const std::string errorMessage = "GPU kernel (" + std::string(kernelName)
308 + ") failed to launch: " + ocl_get_error_string(clError);
309 GMX_THROW(gmx::InternalError(errorMessage));