src/gromacs/gpu_utils/devicebuffer_ocl.h

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2018,2019,2020, 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.
   8  *
   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.
  13  *
  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.
  18  *
  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.
  23  *
  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.
  31  *
  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.
  34  */
  35 #ifndef GMX_GPU_UTILS_DEVICEBUFFER_OCL_H
  36 #define GMX_GPU_UTILS_DEVICEBUFFER_OCL_H
  37
  38 /*! \libinternal \file
  39  *  \brief Implements the DeviceBuffer type and routines for OpenCL.
  40  *  Should only be included directly by the main DeviceBuffer file devicebuffer.h.
  41  *  TODO: the intent is for DeviceBuffer to become a class.
  42  *
  43  *  \author Aleksei Iupinov <a.yupinov@gmail.com>
  44  *
  45  *  \inlibraryapi
  46  */
  47
  48 #include "gromacs/gpu_utils/device_context.h"
  49 #include "gromacs/gpu_utils/devicebuffer_datatype.h"
  50 #include "gromacs/gpu_utils/gpu_utils.h" //only for GpuApiCallBehavior
  51 #include "gromacs/gpu_utils/gputraits_ocl.h"
  52 #include "gromacs/gpu_utils/oclutils.h"
  53 #include "gromacs/utility/gmxassert.h"
  54 #include "gromacs/utility/stringutil.h"
  55
  56 /*! \libinternal \brief
  57  * Allocates a device-side buffer.
  58  * It is currently a caller's responsibility to call it only on not-yet allocated buffers.
  59  *
  60  * \tparam        ValueType            Raw value type of the \p buffer.
  61  * \param[in,out] buffer               Pointer to the device-side buffer.
  62  * \param[in]     numValues            Number of values to accomodate.
  63  * \param[in]     deviceContext        The buffer's device context-to-be.
  64  */
  65 template<typename ValueType>
  66 void allocateDeviceBuffer(DeviceBuffer<ValueType>* buffer, size_t numValues, const DeviceContext& deviceContext)
  67 {
  68     GMX_ASSERT(buffer, "needs a buffer pointer");
  69     void*  hostPtr = nullptr;
  70     cl_int clError;
  71     *buffer = clCreateBuffer(deviceContext.context(), CL_MEM_READ_WRITE,
  72                              numValues * sizeof(ValueType), hostPtr, &clError);
  73     GMX_RELEASE_ASSERT(clError == CL_SUCCESS,
  74                        gmx::formatString("clCreateBuffer failure (OpenCL error %d: %s)", clError,
  75                                          ocl_get_error_string(clError).c_str())
  76                                .c_str());
  77 }
  78
  79 /*! \brief
  80  * Frees a device-side buffer.
  81  * This does not reset separately stored size/capacity integers,
  82  * as this is planned to be a destructor of DeviceBuffer as a proper class,
  83  * and no calls on \p buffer should be made afterwards.
  84  *
  85  * \param[in] buffer  Pointer to the buffer to free.
  86  */
  87 template<typename DeviceBuffer>
  88 void freeDeviceBuffer(DeviceBuffer* buffer)
  89 {
  90     GMX_ASSERT(buffer, "needs a buffer pointer");
  91     if (*buffer)
  92     {
  93         cl_int clError = clReleaseMemObject(*buffer);
  94         GMX_RELEASE_ASSERT(clError == CL_SUCCESS,
  95                            gmx::formatString("clReleaseMemObject failed (OpenCL error %d: %s)",
  96                                              clError, ocl_get_error_string(clError).c_str())
  97                                    .c_str());
  98     }
  99 }
 100
 101 /*! \brief
 102  * Performs the host-to-device data copy, synchronous or asynchronously on request.
 103  *
 104  * TODO: This is meant to gradually replace cu/ocl_copy_h2d.
 105  *
 106  * \tparam        ValueType            Raw value type of the \p buffer.
 107  * \param[in,out] buffer               Pointer to the device-side buffer
 108  * \param[in]     hostBuffer           Pointer to the raw host-side memory, also typed \p ValueType
 109  * \param[in]     startingOffset       Offset (in values) at the device-side buffer to copy into.
 110  * \param[in]     numValues            Number of values to copy.
 111  * \param[in]     deviceStream         GPU stream to perform asynchronous copy in.
 112  * \param[in]     transferKind         Copy type: synchronous or asynchronous.
 113  * \param[out]    timingEvent          A pointer to the H2D copy timing event to be filled in.
 114  *                                     If the pointer is not null, the event can further be used
 115  *                                     to queue a wait for this operation or to query profiling information.
 116  */
 117 template<typename ValueType>
 118 void copyToDeviceBuffer(DeviceBuffer<ValueType>* buffer,
 119                         const ValueType*         hostBuffer,
 120                         size_t                   startingOffset,
 121                         size_t                   numValues,
 122                         const DeviceStream&      deviceStream,
 123                         GpuApiCallBehavior       transferKind,
 124                         CommandEvent*            timingEvent)
 125 {
 126     if (numValues == 0)
 127     {
 128         return; // such calls are actually made with empty domains
 129     }
 130     GMX_ASSERT(buffer, "needs a buffer pointer");
 131     GMX_ASSERT(hostBuffer, "needs a host buffer pointer");
 132     cl_int       clError;
 133     const size_t offset = startingOffset * sizeof(ValueType);
 134     const size_t bytes  = numValues * sizeof(ValueType);
 135     switch (transferKind)
 136     {
 137         case GpuApiCallBehavior::Async:
 138             clError = clEnqueueWriteBuffer(deviceStream.stream(), *buffer, CL_FALSE, offset, bytes,
 139                                            hostBuffer, 0, nullptr, timingEvent);
 140             GMX_RELEASE_ASSERT(
 141                     clError == CL_SUCCESS,
 142                     gmx::formatString("Asynchronous H2D copy failed (OpenCL error %d: %s)", clError,
 143                                       ocl_get_error_string(clError).c_str())
 144                             .c_str());
 145             break;
 146
 147         case GpuApiCallBehavior::Sync:
 148             clError = clEnqueueWriteBuffer(deviceStream.stream(), *buffer, CL_TRUE, offset, bytes,
 149                                            hostBuffer, 0, nullptr, timingEvent);
 150             GMX_RELEASE_ASSERT(
 151                     clError == CL_SUCCESS,
 152                     gmx::formatString("Synchronous H2D copy failed (OpenCL error %d: %s)", clError,
 153                                       ocl_get_error_string(clError).c_str())
 154                             .c_str());
 155             break;
 156
 157         default: throw;
 158     }
 159 }
 160
 161 /*! \brief
 162  * Performs the device-to-host data copy, synchronous or asynchronously on request.
 163  *
 164  * TODO: This is meant to gradually replace cu/ocl_copy_d2h.
 165  *
 166  * \tparam        ValueType            Raw value type of the \p buffer.
 167  * \param[in,out] hostBuffer           Pointer to the raw host-side memory, also typed \p ValueType
 168  * \param[in]     buffer               Pointer to the device-side buffer
 169  * \param[in]     startingOffset       Offset (in values) at the device-side buffer to copy from.
 170  * \param[in]     numValues            Number of values to copy.
 171  * \param[in]     deviceStream         GPU stream to perform asynchronous copy in.
 172  * \param[in]     transferKind         Copy type: synchronous or asynchronous.
 173  * \param[out]    timingEvent          A pointer to the H2D copy timing event to be filled in.
 174  *                                     If the pointer is not null, the event can further be used
 175  *                                     to queue a wait for this operation or to query profiling information.
 176  */
 177 template<typename ValueType>
 178 void copyFromDeviceBuffer(ValueType*               hostBuffer,
 179                           DeviceBuffer<ValueType>* buffer,
 180                           size_t                   startingOffset,
 181                           size_t                   numValues,
 182                           const DeviceStream&      deviceStream,
 183                           GpuApiCallBehavior       transferKind,
 184                           CommandEvent*            timingEvent)
 185 {
 186     GMX_ASSERT(buffer, "needs a buffer pointer");
 187     GMX_ASSERT(hostBuffer, "needs a host buffer pointer");
 188     cl_int       clError;
 189     const size_t offset = startingOffset * sizeof(ValueType);
 190     const size_t bytes  = numValues * sizeof(ValueType);
 191     switch (transferKind)
 192     {
 193         case GpuApiCallBehavior::Async:
 194             clError = clEnqueueReadBuffer(deviceStream.stream(), *buffer, CL_FALSE, offset, bytes,
 195                                           hostBuffer, 0, nullptr, timingEvent);
 196             GMX_RELEASE_ASSERT(
 197                     clError == CL_SUCCESS,
 198                     gmx::formatString("Asynchronous D2H copy failed (OpenCL error %d: %s)", clError,
 199                                       ocl_get_error_string(clError).c_str())
 200                             .c_str());
 201             break;
 202
 203         case GpuApiCallBehavior::Sync:
 204             clError = clEnqueueReadBuffer(deviceStream.stream(), *buffer, CL_TRUE, offset, bytes,
 205                                           hostBuffer, 0, nullptr, timingEvent);
 206             GMX_RELEASE_ASSERT(
 207                     clError == CL_SUCCESS,
 208                     gmx::formatString("Synchronous D2H copy failed (OpenCL error %d: %s)", clError,
 209                                       ocl_get_error_string(clError).c_str())
 210                             .c_str());
 211             break;
 212
 213         default: throw;
 214     }
 215 }
 216
 217 /*! \brief
 218  * Clears the device buffer asynchronously.
 219  *
 220  * \tparam        ValueType       Raw value type of the \p buffer.
 221  * \param[in,out] buffer          Pointer to the device-side buffer
 222  * \param[in]     startingOffset  Offset (in values) at the device-side buffer to start clearing at.
 223  * \param[in]     numValues       Number of values to clear.
 224  * \param[in]     deviceStream    GPU stream.
 225  */
 226 template<typename ValueType>
 227 void clearDeviceBufferAsync(DeviceBuffer<ValueType>* buffer,
 228                             size_t                   startingOffset,
 229                             size_t                   numValues,
 230                             const DeviceStream&      deviceStream)
 231 {
 232     GMX_ASSERT(buffer, "needs a buffer pointer");
 233     const size_t    offset        = startingOffset * sizeof(ValueType);
 234     const size_t    bytes         = numValues * sizeof(ValueType);
 235     const int       pattern       = 0;
 236     const cl_uint   numWaitEvents = 0;
 237     const cl_event* waitEvents    = nullptr;
 238     cl_event        commandEvent;
 239     cl_int clError = clEnqueueFillBuffer(deviceStream.stream(), *buffer, &pattern, sizeof(pattern),
 240                                          offset, bytes, numWaitEvents, waitEvents, &commandEvent);
 241     GMX_RELEASE_ASSERT(clError == CL_SUCCESS,
 242                        gmx::formatString("Couldn't clear the device buffer (OpenCL error %d: %s)",
 243                                          clError, ocl_get_error_string(clError).c_str())
 244                                .c_str());
 245 }
 246
 247 #if defined(__clang__)
 248 #    pragma clang diagnostic push
 249 #    pragma clang diagnostic ignored "-Wunused-template"
 250 #endif
 251
 252 /*! \brief Check the validity of the device buffer.
 253  *
 254  * Checks if the buffer is not nullptr and if its allocation is big enough.
 255  *
 256  * \param[in] buffer        Device buffer to be checked.
 257  * \param[in] requiredSize  Number of elements that the buffer will have to accommodate.
 258  *
 259  * \returns Whether the device buffer can be set.
 260  */
 261 template<typename T>
 262 static bool checkDeviceBuffer(DeviceBuffer<T> buffer, int requiredSize)
 263 {
 264     size_t size;
 265     int    retval = clGetMemObjectInfo(buffer, CL_MEM_SIZE, sizeof(size), &size, nullptr);
 266     GMX_ASSERT(retval == CL_SUCCESS,
 267                gmx::formatString("clGetMemObjectInfo failed with error code #%d", retval).c_str());
 268     GMX_ASSERT(static_cast<int>(size) >= requiredSize,
 269                "Number of atoms in device buffer is smaller then required size.");
 270     return retval == CL_SUCCESS && static_cast<int>(size) >= requiredSize;
 271 }
 272
 273 #if defined(__clang__)
 274 #    pragma clang diagnostic pop
 275 #endif
 276
 277 #endif