src/gromacs/gpu_utils/oclutils.cpp

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2014,2015, 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 /*! \internal \file
  36  *  \brief Define utility routines for OpenCL
  37  *
  38  *  \author Anca Hamuraru <anca@streamcomputing.eu>
  39  */
  40 #include "gmxpre.h"
  41
  42 #include "oclutils.h"
  43
  44 #include <stdlib.h>
  45
  46 #include <cassert>
  47 #include <cstdio>
  48
  49 #include "gromacs/utility/fatalerror.h"
  50 #include "gromacs/utility/smalloc.h"
  51
  52 /*! \brief Launches synchronous or asynchronous host to device memory copy.
  53  *
  54  *  If copy_event is not NULL, on return it will contain an event object
  55  *  identifying this particular host to device operation. The event can further
  56  *  be used to queue a wait for this operation or to query profiling information.
  57  */
  58 static int ocl_copy_H2D_generic(cl_mem d_dest, void* h_src,
  59                                 size_t offset, size_t bytes,
  60                                 bool bAsync /* = false*/,
  61                                 cl_command_queue command_queue,
  62                                 cl_event *copy_event)
  63 {
  64     cl_int gmx_unused cl_error;
  65
  66     if (d_dest == NULL || h_src == NULL || bytes == 0)
  67     {
  68         return -1;
  69     }
  70
  71     if (bAsync)
  72     {
  73         cl_error = clEnqueueWriteBuffer(command_queue, d_dest, CL_FALSE, offset, bytes, h_src, 0, NULL, copy_event);
  74         assert(cl_error == CL_SUCCESS);
  75         // TODO: handle errors
  76     }
  77     else
  78     {
  79         cl_error = clEnqueueWriteBuffer(command_queue, d_dest, CL_TRUE, offset, bytes, h_src, 0, NULL, copy_event);
  80         assert(cl_error == CL_SUCCESS);
  81         // TODO: handle errors
  82     }
  83
  84     return 0;
  85 }
  86
  87 /*! \brief Launches asynchronous host to device memory copy.
  88  *
  89  *  If copy_event is not NULL, on return it will contain an event object
  90  *  identifying this particular host to device operation. The event can further
  91  *  be used to queue a wait for this operation or to query profiling information.
  92  */
  93 int ocl_copy_H2D_async(cl_mem d_dest, void * h_src,
  94                        size_t offset, size_t bytes,
  95                        cl_command_queue command_queue,
  96                        cl_event *copy_event)
  97 {
  98     return ocl_copy_H2D_generic(d_dest, h_src, offset, bytes, true, command_queue, copy_event);
  99 }
 100
 101 /*! \brief Launches synchronous host to device memory copy.
 102  */
 103 int ocl_copy_H2D(cl_mem d_dest, void * h_src,
 104                  size_t offset, size_t bytes,
 105                  cl_command_queue command_queue)
 106 {
 107     return ocl_copy_H2D_generic(d_dest, h_src, offset, bytes, false, command_queue, NULL);
 108 }
 109
 110 /*! \brief Launches synchronous or asynchronous device to host memory copy.
 111  *
 112  *  If copy_event is not NULL, on return it will contain an event object
 113  *  identifying this particular device to host operation. The event can further
 114  *  be used to queue a wait for this operation or to query profiling information.
 115  */
 116 int ocl_copy_D2H_generic(void * h_dest, cl_mem d_src,
 117                          size_t offset, size_t bytes,
 118                          bool bAsync,
 119                          cl_command_queue command_queue,
 120                          cl_event *copy_event)
 121 {
 122     cl_int gmx_unused cl_error;
 123
 124     if (h_dest == NULL || d_src == NULL || bytes == 0)
 125     {
 126         return -1;
 127     }
 128
 129     if (bAsync)
 130     {
 131         cl_error = clEnqueueReadBuffer(command_queue, d_src, CL_FALSE, offset, bytes, h_dest, 0, NULL, copy_event);
 132         assert(cl_error == CL_SUCCESS);
 133         // TODO: handle errors
 134     }
 135     else
 136     {
 137         cl_error = clEnqueueReadBuffer(command_queue, d_src, CL_TRUE, offset, bytes, h_dest, 0, NULL, copy_event);
 138         assert(cl_error == CL_SUCCESS);
 139         // TODO: handle errors
 140     }
 141
 142     return 0;
 143 }
 144
 145 /*! \brief Launches asynchronous device to host memory copy.
 146  *
 147  *  If copy_event is not NULL, on return it will contain an event object
 148  *  identifying this particular host to device operation. The event can further
 149  *  be used to queue a wait for this operation or to query profiling information.
 150  */
 151 int ocl_copy_D2H_async(void * h_dest, cl_mem d_src,
 152                        size_t offset, size_t bytes,
 153                        cl_command_queue command_queue,
 154                        cl_event *copy_event)
 155 {
 156     return ocl_copy_D2H_generic(h_dest, d_src, offset, bytes, true, command_queue, copy_event);
 157 }
 158
 159 /*! \brief \brief Allocates nbytes of host memory. Use ocl_free to free memory allocated with this function.
 160  *
 161  *  \todo
 162  *  This function should allocate page-locked memory to help reduce D2H and H2D
 163  *  transfer times, similar with pmalloc from pmalloc_cuda.cu.
 164  *
 165  * \param[in,out]    h_ptr   Pointer where to store the address of the newly allocated buffer.
 166  * \param[in]        nbytes  Size in bytes of the buffer to be allocated.
 167  */
 168 void ocl_pmalloc(void **h_ptr, size_t nbytes)
 169 {
 170     /* Need a temporary type whose size is 1 byte, so that the
 171      * implementation of snew_aligned can cope without issuing
 172      * warnings. */
 173     char **temporary = reinterpret_cast<char **>(h_ptr);
 174
 175     /* 16-byte alignment is required by the neighbour-searching code,
 176      * because it uses four-wide SIMD for bounding-box calculation.
 177      * However, when we organize using page-locked memory for
 178      * device-host transfers, it will probably need to be aligned to a
 179      * 4kb page, like CUDA does. */
 180     snew_aligned(*temporary, nbytes, 16);
 181 }
 182
 183 /*! \brief Frees memory allocated with ocl_pmalloc.
 184  *
 185  * \param[in]    h_ptr   Buffer allocated with ocl_pmalloc that needs to be freed.
 186  */
 187 void ocl_pfree(void *h_ptr)
 188 {
 189
 190     if (h_ptr)
 191     {
 192         sfree_aligned(h_ptr);
 193     }
 194     return;
 195 }