src/gromacs/gpu_utils/gpueventsynchronizer_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 /*! \libinternal \file
  36  *  \brief Implements a GpuEventSynchronizer class for OpenCL.
  37  *
  38  *  \author Aleksei Iupinov <a.yupinov@gmail.com>
  39  * \inlibraryapi
  40  */
  41 #ifndef GMX_GPU_UTILS_GPUEVENTSYNCHRONIZER_OCL_H
  42 #define GMX_GPU_UTILS_GPUEVENTSYNCHRONIZER_OCL_H
  43
  44 #include "gromacs/gpu_utils/gputraits_ocl.h"
  45 #include "gromacs/gpu_utils/oclutils.h"
  46 #include "gromacs/utility/exceptions.h"
  47 #include "gromacs/utility/gmxassert.h"
  48
  49 /*! \libinternal \brief
  50  * A class which allows for CPU thread to mark and wait for certain GPU stream execution point.
  51  * The event can be put into the stream with markEvent() and then later waited on with waitForEvent().
  52  * This can be repeated as necessary, but the current implementation does not allow waiting on
  53  * completed event more than once, expecting only exact pairs of markEvent(stream); waitForEvent().
  54  * The class generally attempts to track the correctness of its state transitions, but
  55  * please note that calling waitForEvent() right after the construction will fail with OpenCL but succeed with CUDA.
  56  *
  57  * Another possible mode of operation can be implemented if needed:
  58  * multiple calls to waitForEvent() after a single markEvent(). For this, clReleaseEvent() call
  59  * from waitForEvent() should instead happen conditionally at the beginning of markEvent(), replacing
  60  * the GMX_ASSERT(). This was tested to work both with CUDA and NVidia OpenCL, but not with AMD/Intel OpenCl.
  61  */
  62 class GpuEventSynchronizer
  63 {
  64     public:
  65         //! A constructor
  66         GpuEventSynchronizer() : event_(nullptr){}
  67         //! A destructor
  68         ~GpuEventSynchronizer()
  69         {
  70             // This additional code only prevents cl_event leak in an unlikely situation of destructor
  71             // being called after markEvent() but before waitForEvent().
  72             if (event_)
  73             {
  74                 ensureReferenceCount(event_, 1);
  75                 clReleaseEvent(event_);
  76             }
  77         }
  78         //! No copying
  79         GpuEventSynchronizer(const GpuEventSynchronizer &)       = delete;
  80         //! No assignment
  81         GpuEventSynchronizer &operator=(GpuEventSynchronizer &&) = delete;
  82         //! Moving is disabled but can be considered in the future if needed
  83         GpuEventSynchronizer(GpuEventSynchronizer &&)            = delete;
  84
  85         /*! \brief Marks the synchronization point in the \p stream.
  86          * Should be called first and then followed by waitForEvent().
  87          */
  88         inline void markEvent(CommandStream stream)
  89         {
  90             GMX_ASSERT(nullptr == event_, "Do not call markEvent more than once!");
  91             cl_int clError = clEnqueueMarkerWithWaitList(stream, 0, nullptr, &event_);
  92             if (CL_SUCCESS != clError)
  93             {
  94                 GMX_THROW(gmx::InternalError("Failed to enqueue the GPU synchronization event: " + ocl_get_error_string(clError)));
  95             }
  96         }
  97         /*! \brief Synchronizes the host thread on the marked event. */
  98         inline void waitForEvent()
  99         {
 100             cl_int clError = clWaitForEvents(1, &event_);
 101             if (CL_SUCCESS != clError)
 102             {
 103                 GMX_THROW(gmx::InternalError("Failed to synchronize on the GPU event: " + ocl_get_error_string(clError)));
 104             }
 105
 106             // Reference count can't be checked after the event's released, it seems (segfault on NVIDIA).
 107             ensureReferenceCount(event_, 1);
 108             clError = clReleaseEvent(event_);
 109             if (CL_SUCCESS != clError)
 110             {
 111                 GMX_THROW(gmx::InternalError("Failed to release the GPU event: " + ocl_get_error_string(clError)));
 112             }
 113             event_ = nullptr;
 114         }
 115
 116     private:
 117         cl_event event_;
 118 };
 119
 120 #endif