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.
37 * \brief Define utility routines for OpenCL
39 * \author Anca Hamuraru <anca@streamcomputing.eu>
52 #include "gromacs/gpu_utils/gpu_utils.h"
53 #include "gromacs/utility/fatalerror.h"
54 #include "gromacs/utility/smalloc.h"
56 int ocl_copy_H2D(cl_mem d_dest,
60 GpuApiCallBehavior transferKind,
61 cl_command_queue command_queue,
64 cl_int gmx_unused cl_error;
66 if (d_dest == nullptr || h_src == nullptr || bytes == 0)
73 case GpuApiCallBehavior::Async:
74 cl_error = clEnqueueWriteBuffer(command_queue, d_dest, CL_FALSE, offset, bytes, h_src,
75 0, nullptr, copy_event);
78 case GpuApiCallBehavior::Sync:
79 cl_error = clEnqueueWriteBuffer(command_queue, d_dest, CL_TRUE, offset, bytes, h_src, 0,
85 GMX_ASSERT(cl_error == CL_SUCCESS,
86 ("clEnqueueWriteBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
91 /*! \brief Launches asynchronous host to device memory copy.
93 * If copy_event is not nullptr, on return it will contain an event object
94 * identifying this particular host to device operation. The event can further
95 * be used to queue a wait for this operation or to query profiling information.
97 int ocl_copy_H2D_async(cl_mem d_dest,
101 cl_command_queue command_queue,
102 cl_event* copy_event)
104 return ocl_copy_H2D(d_dest, h_src, offset, bytes, GpuApiCallBehavior::Async, command_queue, copy_event);
107 /*! \brief Launches synchronous host to device memory copy.
109 int ocl_copy_H2D_sync(cl_mem d_dest, const void* h_src, size_t offset, size_t bytes, cl_command_queue command_queue)
111 return ocl_copy_H2D(d_dest, h_src, offset, bytes, GpuApiCallBehavior::Sync, command_queue, nullptr);
114 int ocl_copy_D2H(void* h_dest,
118 GpuApiCallBehavior transferKind,
119 cl_command_queue command_queue,
120 cl_event* copy_event)
122 cl_int gmx_unused cl_error;
124 if (h_dest == nullptr || d_src == nullptr || bytes == 0)
129 switch (transferKind)
131 case GpuApiCallBehavior::Async:
132 cl_error = clEnqueueReadBuffer(command_queue, d_src, CL_FALSE, offset, bytes, h_dest, 0,
133 nullptr, copy_event);
136 case GpuApiCallBehavior::Sync:
137 cl_error = clEnqueueReadBuffer(command_queue, d_src, CL_TRUE, offset, bytes, h_dest, 0,
138 nullptr, copy_event);
143 GMX_ASSERT(cl_error == CL_SUCCESS,
144 ("clEnqueueWriteBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
150 /*! \brief Launches asynchronous device to host memory copy.
152 * If copy_event is not nullptr, on return it will contain an event object
153 * identifying this particular host to device operation. The event can further
154 * be used to queue a wait for this operation or to query profiling information.
156 int ocl_copy_D2H_async(void* h_dest,
160 cl_command_queue command_queue,
161 cl_event* copy_event)
163 return ocl_copy_D2H(h_dest, d_src, offset, bytes, GpuApiCallBehavior::Async, command_queue, copy_event);
166 /*! \brief \brief Allocates nbytes of host memory. Use ocl_free to free memory allocated with this function.
169 * This function should allocate page-locked memory to help reduce D2H and H2D
170 * transfer times, similar with pmalloc from pmalloc_cuda.cu.
172 * \param[in,out] h_ptr Pointer where to store the address of the newly allocated buffer.
173 * \param[in] nbytes Size in bytes of the buffer to be allocated.
175 void pmalloc(void** h_ptr, size_t nbytes)
177 /* Need a temporary type whose size is 1 byte, so that the
178 * implementation of snew_aligned can cope without issuing
180 char** temporary = reinterpret_cast<char**>(h_ptr);
182 /* 16-byte alignment is required by the neighbour-searching code,
183 * because it uses four-wide SIMD for bounding-box calculation.
184 * However, when we organize using page-locked memory for
185 * device-host transfers, it will probably need to be aligned to a
186 * 4kb page, like CUDA does. */
187 snew_aligned(*temporary, nbytes, 16);
190 /*! \brief Frees memory allocated with pmalloc.
192 * \param[in] h_ptr Buffer allocated with pmalloc that needs to be freed.
194 void pfree(void* h_ptr)
199 sfree_aligned(h_ptr);
203 /*! \brief Convert error code to diagnostic string */
204 std::string ocl_get_error_string(cl_int error)
208 // run-time and JIT compiler errors
209 case 0: return "CL_SUCCESS";
210 case -1: return "CL_DEVICE_NOT_FOUND";
211 case -2: return "CL_DEVICE_NOT_AVAILABLE";
212 case -3: return "CL_COMPILER_NOT_AVAILABLE";
213 case -4: return "CL_MEM_OBJECT_ALLOCATION_FAILURE";
214 case -5: return "CL_OUT_OF_RESOURCES";
215 case -6: return "CL_OUT_OF_HOST_MEMORY";
216 case -7: return "CL_PROFILING_INFO_NOT_AVAILABLE";
217 case -8: return "CL_MEM_COPY_OVERLAP";
218 case -9: return "CL_IMAGE_FORMAT_MISMATCH";
219 case -10: return "CL_IMAGE_FORMAT_NOT_SUPPORTED";
220 case -11: return "CL_BUILD_PROGRAM_FAILURE";
221 case -12: return "CL_MAP_FAILURE";
222 case -13: return "CL_MISALIGNED_SUB_BUFFER_OFFSET";
223 case -14: return "CL_EXEC_STATUS_ERROR_FOR_EVENTS_IN_WAIT_LIST";
224 case -15: return "CL_COMPILE_PROGRAM_FAILURE";
225 case -16: return "CL_LINKER_NOT_AVAILABLE";
226 case -17: return "CL_LINK_PROGRAM_FAILURE";
227 case -18: return "CL_DEVICE_PARTITION_FAILED";
228 case -19: return "CL_KERNEL_ARG_INFO_NOT_AVAILABLE";
230 // compile-time errors
231 case -30: return "CL_INVALID_VALUE";
232 case -31: return "CL_INVALID_DEVICE_TYPE";
233 case -32: return "CL_INVALID_PLATFORM";
234 case -33: return "CL_INVALID_DEVICE";
235 case -34: return "CL_INVALID_CONTEXT";
236 case -35: return "CL_INVALID_QUEUE_PROPERTIES";
237 case -36: return "CL_INVALID_COMMAND_QUEUE";
238 case -37: return "CL_INVALID_HOST_PTR";
239 case -38: return "CL_INVALID_MEM_OBJECT";
240 case -39: return "CL_INVALID_IMAGE_FORMAT_DESCRIPTOR";
241 case -40: return "CL_INVALID_IMAGE_SIZE";
242 case -41: return "CL_INVALID_SAMPLER";
243 case -42: return "CL_INVALID_BINARY";
244 case -43: return "CL_INVALID_BUILD_OPTIONS";
245 case -44: return "CL_INVALID_PROGRAM";
246 case -45: return "CL_INVALID_PROGRAM_EXECUTABLE";
247 case -46: return "CL_INVALID_KERNEL_NAME";
248 case -47: return "CL_INVALID_KERNEL_DEFINITION";
249 case -48: return "CL_INVALID_KERNEL";
250 case -49: return "CL_INVALID_ARG_INDEX";
251 case -50: return "CL_INVALID_ARG_VALUE";
252 case -51: return "CL_INVALID_ARG_SIZE";
253 case -52: return "CL_INVALID_KERNEL_ARGS";
254 case -53: return "CL_INVALID_WORK_DIMENSION";
255 case -54: return "CL_INVALID_WORK_GROUP_SIZE";
256 case -55: return "CL_INVALID_WORK_ITEM_SIZE";
257 case -56: return "CL_INVALID_GLOBAL_OFFSET";
258 case -57: return "CL_INVALID_EVENT_WAIT_LIST";
259 case -58: return "CL_INVALID_EVENT";
260 case -59: return "CL_INVALID_OPERATION";
261 case -60: return "CL_INVALID_GL_OBJECT";
262 case -61: return "CL_INVALID_BUFFER_SIZE";
263 case -62: return "CL_INVALID_MIP_LEVEL";
264 case -63: return "CL_INVALID_GLOBAL_WORK_SIZE";
265 case -64: return "CL_INVALID_PROPERTY";
266 case -65: return "CL_INVALID_IMAGE_DESCRIPTOR";
267 case -66: return "CL_INVALID_COMPILER_OPTIONS";
268 case -67: return "CL_INVALID_LINKER_OPTIONS";
269 case -68: return "CL_INVALID_DEVICE_PARTITION_COUNT";
272 case -1000: return "CL_INVALID_GL_SHAREGROUP_REFERENCE_KHR";
273 case -1001: return "CL_PLATFORM_NOT_FOUND_KHR";
274 case -1002: return "CL_INVALID_D3D10_DEVICE_KHR";
275 case -1003: return "CL_INVALID_D3D10_RESOURCE_KHR";
276 case -1004: return "CL_D3D10_RESOURCE_ALREADY_ACQUIRED_KHR";
277 case -1005: return "CL_D3D10_RESOURCE_NOT_ACQUIRED_KHR";
278 default: return "Unknown OpenCL error: " + std::to_string(static_cast<int32_t>(error));