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  * Note that synchronous copy will not synchronize the stream in case of zero \p numValues
 105  * because of the early return.
 106  *
 107  * \tparam        ValueType            Raw value type of the \p buffer.
 108  * \param[in,out] buffer               Pointer to the device-side buffer
 109  * \param[in]     hostBuffer           Pointer to the raw host-side memory, also typed \p ValueType
 110  * \param[in]     startingOffset       Offset (in values) at the device-side buffer to copy into.
 111  * \param[in]     numValues            Number of values to copy.
 112  * \param[in]     deviceStream         GPU stream to perform asynchronous copy in.
 113  * \param[in]     transferKind         Copy type: synchronous or asynchronous.
 114  * \param[out]    timingEvent          A pointer to the H2D copy timing event to be filled in.
 115  *                                     If the pointer is not null, the event can further be used
 116  *                                     to queue a wait for this operation or to query profiling information.
 117  */
 118 template<typename ValueType>
 119 void copyToDeviceBuffer(DeviceBuffer<ValueType>* buffer,
 120                         const ValueType*         hostBuffer,
 121                         size_t                   startingOffset,
 122                         size_t                   numValues,
 123                         const DeviceStream&      deviceStream,
 124                         GpuApiCallBehavior       transferKind,
 125                         CommandEvent*            timingEvent)
 126 {
 127     if (numValues == 0)
 128     {
 129         return; // such calls are actually made with empty domains
 130     }
 131     GMX_ASSERT(buffer, "needs a buffer pointer");
 132     GMX_ASSERT(hostBuffer, "needs a host buffer pointer");
 133     cl_int       clError;
 134     const size_t offset = startingOffset * sizeof(ValueType);
 135     const size_t bytes  = numValues * sizeof(ValueType);
 136     switch (transferKind)
 137     {
 138         case GpuApiCallBehavior::Async:
 139             clError = clEnqueueWriteBuffer(deviceStream.stream(), *buffer, CL_FALSE, offset, bytes,
 140                                            hostBuffer, 0, nullptr, timingEvent);
 141             GMX_RELEASE_ASSERT(
 142                     clError == CL_SUCCESS,
 143                     gmx::formatString("Asynchronous H2D copy failed (OpenCL error %d: %s)", clError,
 144                                       ocl_get_error_string(clError).c_str())
 145                             .c_str());
 146             break;
 147
 148         case GpuApiCallBehavior::Sync:
 149             clError = clEnqueueWriteBuffer(deviceStream.stream(), *buffer, CL_TRUE, offset, bytes,
 150                                            hostBuffer, 0, nullptr, timingEvent);
 151             GMX_RELEASE_ASSERT(
 152                     clError == CL_SUCCESS,
 153                     gmx::formatString("Synchronous H2D copy failed (OpenCL error %d: %s)", clError,
 154                                       ocl_get_error_string(clError).c_str())
 155                             .c_str());
 156             break;
 157
 158         default: throw;
 159     }
 160 }
 161
 162 /*! \brief
 163  * Performs the device-to-host data copy, synchronous or asynchronously on request.
 164  *
 165  * Note that synchronous copy will not synchronize the stream in case of zero \p numValues
 166  * because of the early return.
 167  *
 168  * \tparam        ValueType            Raw value type of the \p buffer.
 169  * \param[in,out] hostBuffer           Pointer to the raw host-side memory, also typed \p ValueType
 170  * \param[in]     buffer               Pointer to the device-side buffer
 171  * \param[in]     startingOffset       Offset (in values) at the device-side buffer to copy from.
 172  * \param[in]     numValues            Number of values to copy.
 173  * \param[in]     deviceStream         GPU stream to perform asynchronous copy in.
 174  * \param[in]     transferKind         Copy type: synchronous or asynchronous.
 175  * \param[out]    timingEvent          A pointer to the H2D copy timing event to be filled in.
 176  *                                     If the pointer is not null, the event can further be used
 177  *                                     to queue a wait for this operation or to query profiling information.
 178  */
 179 template<typename ValueType>
 180 void copyFromDeviceBuffer(ValueType*               hostBuffer,
 181                           DeviceBuffer<ValueType>* buffer,
 182                           size_t                   startingOffset,
 183                           size_t                   numValues,
 184                           const DeviceStream&      deviceStream,
 185                           GpuApiCallBehavior       transferKind,
 186                           CommandEvent*            timingEvent)
 187 {
 188     if (numValues == 0)
 189     {
 190         return;
 191     }
 192     GMX_ASSERT(buffer, "needs a buffer pointer");
 193     GMX_ASSERT(hostBuffer, "needs a host buffer pointer");
 194     cl_int       clError;
 195     const size_t offset = startingOffset * sizeof(ValueType);
 196     const size_t bytes  = numValues * sizeof(ValueType);
 197     switch (transferKind)
 198     {
 199         case GpuApiCallBehavior::Async:
 200             clError = clEnqueueReadBuffer(deviceStream.stream(), *buffer, CL_FALSE, offset, bytes,
 201                                           hostBuffer, 0, nullptr, timingEvent);
 202             GMX_RELEASE_ASSERT(
 203                     clError == CL_SUCCESS,
 204                     gmx::formatString("Asynchronous D2H copy failed (OpenCL error %d: %s)", clError,
 205                                       ocl_get_error_string(clError).c_str())
 206                             .c_str());
 207             break;
 208
 209         case GpuApiCallBehavior::Sync:
 210             clError = clEnqueueReadBuffer(deviceStream.stream(), *buffer, CL_TRUE, offset, bytes,
 211                                           hostBuffer, 0, nullptr, timingEvent);
 212             GMX_RELEASE_ASSERT(
 213                     clError == CL_SUCCESS,
 214                     gmx::formatString("Synchronous D2H copy failed (OpenCL error %d: %s)", clError,
 215                                       ocl_get_error_string(clError).c_str())
 216                             .c_str());
 217             break;
 218
 219         default: throw;
 220     }
 221 }
 222
 223 /*! \brief
 224  * Clears the device buffer asynchronously.
 225  *
 226  * \tparam        ValueType       Raw value type of the \p buffer.
 227  * \param[in,out] buffer          Pointer to the device-side buffer
 228  * \param[in]     startingOffset  Offset (in values) at the device-side buffer to start clearing at.
 229  * \param[in]     numValues       Number of values to clear.
 230  * \param[in]     deviceStream    GPU stream.
 231  */
 232 template<typename ValueType>
 233 void clearDeviceBufferAsync(DeviceBuffer<ValueType>* buffer,
 234                             size_t                   startingOffset,
 235                             size_t                   numValues,
 236                             const DeviceStream&      deviceStream)
 237 {
 238     GMX_ASSERT(buffer, "needs a buffer pointer");
 239     const size_t    offset        = startingOffset * sizeof(ValueType);
 240     const size_t    bytes         = numValues * sizeof(ValueType);
 241     const int       pattern       = 0;
 242     const cl_uint   numWaitEvents = 0;
 243     const cl_event* waitEvents    = nullptr;
 244     cl_event        commandEvent;
 245     cl_int clError = clEnqueueFillBuffer(deviceStream.stream(), *buffer, &pattern, sizeof(pattern),
 246                                          offset, bytes, numWaitEvents, waitEvents, &commandEvent);
 247     GMX_RELEASE_ASSERT(clError == CL_SUCCESS,
 248                        gmx::formatString("Couldn't clear the device buffer (OpenCL error %d: %s)",
 249                                          clError, ocl_get_error_string(clError).c_str())
 250                                .c_str());
 251 }
 252
 253 #if defined(__clang__)
 254 #    pragma clang diagnostic push
 255 #    pragma clang diagnostic ignored "-Wunused-template"
 256 #endif
 257
 258 /*! \brief Check the validity of the device buffer.
 259  *
 260  * Checks if the buffer is not nullptr and if its allocation is big enough.
 261  *
 262  * \param[in] buffer        Device buffer to be checked.
 263  * \param[in] requiredSize  Number of elements that the buffer will have to accommodate.
 264  *
 265  * \returns Whether the device buffer can be set.
 266  */
 267 template<typename T>
 268 static bool checkDeviceBuffer(DeviceBuffer<T> buffer, int requiredSize)
 269 {
 270     size_t size;
 271     int    retval = clGetMemObjectInfo(buffer, CL_MEM_SIZE, sizeof(size), &size, nullptr);
 272     GMX_ASSERT(retval == CL_SUCCESS,
 273                gmx::formatString("clGetMemObjectInfo failed with error code #%d", retval).c_str());
 274     GMX_ASSERT(static_cast<int>(size) >= requiredSize,
 275                "Number of atoms in device buffer is smaller then required size.");
 276     return retval == CL_SUCCESS && static_cast<int>(size) >= requiredSize;
 277 }
 278
 279 //! Device texture wrapper.
 280 using DeviceTexture = void*;
 281
 282 /*! \brief Create a texture object for an array of type ValueType.
 283  *
 284  * Creates the device buffer and copies read-only data for an array of type ValueType.
 285  *
 286  * \todo Decide if using image2d is most efficient.
 287  *
 288  * \tparam      ValueType      Raw data type.
 289  *
 290  * \param[out]  deviceBuffer   Device buffer to store data in.
 291  * \param[in]   hostBuffer     Host buffer to get date from.
 292  * \param[in]   numValues      Number of elements in the buffer.
 293  * \param[in]   deviceContext  GPU device context.
 294  */
 295 template<typename ValueType>
 296 void initParamLookupTable(DeviceBuffer<ValueType>* deviceBuffer,
 297                           DeviceTexture* /* deviceTexture */,
 298                           const ValueType*     hostBuffer,
 299                           int                  numValues,
 300                           const DeviceContext& deviceContext)
 301 {
 302     GMX_ASSERT(hostBuffer, "Host buffer pointer can not be null");
 303     const size_t bytes = numValues * sizeof(ValueType);
 304     cl_int       clError;
 305     *deviceBuffer = clCreateBuffer(deviceContext.context(),
 306                                    CL_MEM_READ_ONLY | CL_MEM_HOST_WRITE_ONLY | CL_MEM_COPY_HOST_PTR,
 307                                    bytes, const_cast<ValueType*>(hostBuffer), &clError);
 308
 309     GMX_RELEASE_ASSERT(clError == CL_SUCCESS,
 310                        gmx::formatString("Constant memory allocation failed (OpenCL error %d: %s)",
 311                                          clError, ocl_get_error_string(clError).c_str())
 312                                .c_str());
 313 }
 314
 315 /*! \brief Release the OpenCL device buffer.
 316  *
 317  * \tparam        ValueType     Raw data type.
 318  *
 319  * \param[in,out] deviceBuffer  Device buffer to store data in.
 320  */
 321 template<typename ValueType>
 322 void destroyParamLookupTable(DeviceBuffer<ValueType>* deviceBuffer, DeviceTexture& /* deviceTexture*/)
 323 {
 324     freeDeviceBuffer(deviceBuffer);
 325 }
 326 #if defined(__clang__)
 327 #    pragma clang diagnostic pop
 328 #endif
 329
 330 #endif