src/gromacs/gpu_utils/devicebuffer_ocl.h

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2018, 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/gpu_utils.h" //only for GpuApiCallBehavior
  49 #include "gromacs/gpu_utils/gputraits_ocl.h"
  50 #include "gromacs/utility/gmxassert.h"
  51
  52 /*! \libinternal \brief
  53  * A minimal cl_mem wrapper that remembers its allocation type.
  54  * The only point is making template type deduction possible.
  55  */
  56 template<typename ValueType>
  57 class TypedClMemory
  58 {
  59     private:
  60         //! \brief Underlying data - not nulled right here only because we still have some snew()'s around
  61         cl_mem data_;
  62     public:
  63         //! \brief An assignment operator - the purpose is to make allocation/zeroing work
  64         TypedClMemory &operator=(cl_mem data){data_ = data; return *this; }
  65         //! \brief Returns underlying cl_mem transparently
  66         operator cl_mem() {return data_; }
  67 };
  68
  69 //! \libinternal \brief A device-side buffer of ValueTypes
  70 template<typename ValueType>
  71 using DeviceBuffer = TypedClMemory<ValueType>;
  72
  73 /*! \libinternal \brief
  74  * Allocates a device-side buffer.
  75  * It is currently a caller's responsibility to call it only on not-yet allocated buffers.
  76  *
  77  * \tparam        ValueType            Raw value type of the \p buffer.
  78  * \param[in,out] buffer               Pointer to the device-side buffer.
  79  * \param[in]     numValues            Number of values to accomodate.
  80  * \param[in]     context              The buffer's context-to-be.
  81  */
  82 template <typename ValueType>
  83 void allocateDeviceBuffer(DeviceBuffer<ValueType> *buffer,
  84                           size_t                   numValues,
  85                           Context                  context)
  86 {
  87     GMX_ASSERT(buffer, "needs a buffer pointer");
  88     void  *hostPtr = nullptr;
  89     cl_int clError;
  90     *buffer = clCreateBuffer(context, CL_MEM_READ_WRITE, numValues * sizeof(ValueType), hostPtr, &clError);
  91     GMX_RELEASE_ASSERT(clError == CL_SUCCESS, "clCreateBuffer failure");
  92 }
  93
  94 /*! \brief
  95  * Frees a device-side buffer.
  96  * This does not reset separately stored size/capacity integers,
  97  * as this is planned to be a destructor of DeviceBuffer as a proper class,
  98  * and no calls on \p buffer should be made afterwards.
  99  *
 100  * \param[in] buffer  Pointer to the buffer to free.
 101  */
 102 template <typename DeviceBuffer>
 103 void freeDeviceBuffer(DeviceBuffer *buffer)
 104 {
 105     GMX_ASSERT(buffer, "needs a buffer pointer");
 106     if (*buffer)
 107     {
 108         GMX_RELEASE_ASSERT(clReleaseMemObject(*buffer) == CL_SUCCESS, "clReleaseMemObject failed");
 109     }
 110 }
 111
 112 /*! \brief
 113  * Performs the host-to-device data copy, synchronous or asynchronously on request.
 114  *
 115  * TODO: This is meant to gradually replace cu/ocl_copy_h2d.
 116  *
 117  * \tparam        ValueType            Raw value type of the \p buffer.
 118  * \param[in,out] buffer               Pointer to the device-side buffer
 119  * \param[in]     hostBuffer           Pointer to the raw host-side memory, also typed \p ValueType
 120  * \param[in]     startingOffset       Offset (in values) at the device-side buffer to copy into.
 121  * \param[in]     numValues            Number of values to copy.
 122  * \param[in]     stream               GPU stream to perform asynchronous copy in.
 123  * \param[in]     transferKind         Copy type: synchronous or asynchronous.
 124  * \param[out]    timingEvent          A pointer to the H2D copy timing event to be filled in.
 125  *                                     If the pointer is not null, the event can further be used
 126  *                                     to queue a wait for this operation or to query profiling information.
 127  */
 128 template <typename ValueType>
 129 void copyToDeviceBuffer(DeviceBuffer<ValueType> *buffer,
 130                         const ValueType         *hostBuffer,
 131                         size_t                   startingOffset,
 132                         size_t                   numValues,
 133                         CommandStream            stream,
 134                         GpuApiCallBehavior       transferKind,
 135                         CommandEvent            *timingEvent)
 136 {
 137     if (numValues == 0)
 138     {
 139         return; // such calls are actually made with empty domains
 140     }
 141     GMX_ASSERT(buffer, "needs a buffer pointer");
 142     GMX_ASSERT(hostBuffer, "needs a host buffer pointer");
 143     cl_int       clError;
 144     const size_t offset = startingOffset * sizeof(ValueType);
 145     const size_t bytes  = numValues * sizeof(ValueType);
 146     switch (transferKind)
 147     {
 148         case GpuApiCallBehavior::Async:
 149             clError = clEnqueueWriteBuffer(stream, *buffer, CL_FALSE, offset, bytes, hostBuffer, 0, nullptr, timingEvent);
 150             GMX_RELEASE_ASSERT(clError == CL_SUCCESS, "Asynchronous H2D copy failed");
 151             break;
 152
 153         case GpuApiCallBehavior::Sync:
 154             clError = clEnqueueWriteBuffer(stream, *buffer, CL_TRUE, offset, bytes, hostBuffer, 0, nullptr, timingEvent);
 155             GMX_RELEASE_ASSERT(clError == CL_SUCCESS, "Synchronous H2D copy failed");
 156             break;
 157
 158         default:
 159             throw;
 160     }
 161 }
 162
 163 /*! \brief
 164  * Performs the device-to-host data copy, synchronous or asynchronously on request.
 165  *
 166  * TODO: This is meant to gradually replace cu/ocl_copy_d2h.
 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]     stream               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                           CommandStream                  stream,
 185                           GpuApiCallBehavior             transferKind,
 186                           CommandEvent                  *timingEvent)
 187 {
 188     GMX_ASSERT(buffer, "needs a buffer pointer");
 189     GMX_ASSERT(hostBuffer, "needs a host buffer pointer");
 190     cl_int       clError;
 191     const size_t offset = startingOffset * sizeof(ValueType);
 192     const size_t bytes  = numValues * sizeof(ValueType);
 193     switch (transferKind)
 194     {
 195         case GpuApiCallBehavior::Async:
 196             clError = clEnqueueReadBuffer(stream, *buffer, CL_FALSE, offset, bytes, hostBuffer, 0, nullptr, timingEvent);
 197             GMX_RELEASE_ASSERT(clError == CL_SUCCESS, "Asynchronous D2H copy failed");
 198             break;
 199
 200         case GpuApiCallBehavior::Sync:
 201             clError = clEnqueueReadBuffer(stream, *buffer, CL_TRUE, offset, bytes, hostBuffer, 0, nullptr, timingEvent);
 202             GMX_RELEASE_ASSERT(clError == CL_SUCCESS, "Synchronous D2H copy failed");
 203             break;
 204
 205         default:
 206             throw;
 207     }
 208 }
 209
 210 /*! \brief
 211  * Clears the device buffer asynchronously.
 212  *
 213  * \tparam        ValueType        Raw value type of the \p buffer.
 214  * \param[in,out] buffer           Pointer to the device-side buffer
 215  * \param[in]     startingOffset   Offset (in values) at the device-side buffer to start clearing at.
 216  * \param[in]     numValues        Number of values to clear.
 217  * \param[in]     stream           GPU stream.
 218  */
 219 template <typename ValueType>
 220 void clearDeviceBufferAsync(DeviceBuffer<ValueType> *buffer,
 221                             size_t                   startingOffset,
 222                             size_t                   numValues,
 223                             CommandStream            stream)
 224 {
 225     GMX_ASSERT(buffer, "needs a buffer pointer");
 226     const size_t    offset        = startingOffset * sizeof(ValueType);
 227     const size_t    bytes         = numValues * sizeof(ValueType);
 228     const ValueType pattern       = 0;
 229     const cl_uint   numWaitEvents = 0;
 230     const cl_event *waitEvents    = nullptr;
 231     cl_event        commandEvent;
 232     cl_int          clError = clEnqueueFillBuffer(stream, *buffer, &pattern, sizeof(pattern),
 233                                                   offset, bytes,
 234                                                   numWaitEvents, waitEvents, &commandEvent);
 235     GMX_RELEASE_ASSERT(clError == CL_SUCCESS, "Couldn't clear the device buffer");
 236 }
 237
 238 #endif