src/gromacs/ewald/pme_gpu_program_impl_sycl.cpp

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2021, 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
  36 /*! \internal \file
  37  * \brief
  38  * Implements PmeGpuProgramImpl, which stores permanent PME GPU context-derived data,
  39  * such as (compiled) kernel handles.
  40  *
  41  * \author Aleksei Iupinov <a.yupinov@gmail.com>
  42  * \author Andrey Alekseenko <al42and@gmail.com>
  43  * \ingroup module_ewald
  44  */
  45 #include "gmxpre.h"
  46
  47 #include "gromacs/hardware/device_information.h"
  48 #include "gromacs/gpu_utils/gmxsycl.h"
  49 #include "gromacs/gpu_utils/syclutils.h"
  50
  51 #include "pme_gpu_program_impl.h"
  52 #include "pme_gather_sycl.h"
  53 #include "pme_spread_sycl.h"
  54
  55 #include "pme_gpu_constants.h"
  56 #include "pme_gpu_internal.h" // for GridOrdering enum
  57 #include "pme_gpu_types_host.h"
  58
  59 // PME interpolation order
  60 constexpr int c_pmeOrder = 4;
  61 // These hardcoded spread/gather parameters refer to not-implemented PME GPU 2D decomposition in X/Y
  62 constexpr bool c_wrapX = true;
  63 constexpr bool c_wrapY = true;
  64
  65 static int subGroupSizeFromVendor(const DeviceInformation& deviceInfo)
  66 {
  67     switch (deviceInfo.deviceVendor)
  68     {
  69         case DeviceVendor::Amd: return 64;   // Handle RDNA2 devices, Issue #3972.
  70         case DeviceVendor::Intel: return 16; // TODO: Choose best value, Issue #4153.
  71         case DeviceVendor::Nvidia: return 32;
  72         default: GMX_RELEASE_ASSERT(false, "Unknown device vendor"); return 0;
  73     }
  74 }
  75
  76 #define INSTANTIATE_SPREAD_2(                                                                      \
  77         order, computeSplines, spreadCharges, numGrids, writeGlobal, threadsPerAtom, subGroupSize) \
  78     extern template class PmeSplineAndSpreadKernel<order, computeSplines, spreadCharges, true, true, numGrids, writeGlobal, threadsPerAtom, subGroupSize>;
  79
  80 #define INSTANTIATE_SPREAD(order, numGrids, threadsPerAtom, subGroupSize)                   \
  81     INSTANTIATE_SPREAD_2(order, true, true, numGrids, true, threadsPerAtom, subGroupSize);  \
  82     INSTANTIATE_SPREAD_2(order, true, false, numGrids, true, threadsPerAtom, subGroupSize); \
  83     INSTANTIATE_SPREAD_2(order, false, true, numGrids, true, threadsPerAtom, subGroupSize); \
  84     INSTANTIATE_SPREAD_2(order, true, true, numGrids, false, threadsPerAtom, subGroupSize);
  85
  86 #define INSTANTIATE_GATHER_2(order, numGrids, readGlobal, threadsPerAtom, subGroupSize) \
  87     extern template class PmeGatherKernel<order, true, true, numGrids, readGlobal, threadsPerAtom, subGroupSize>;
  88
  89 #define INSTANTIATE_GATHER(order, numGrids, threadsPerAtom, subGroupSize)      \
  90     INSTANTIATE_GATHER_2(order, numGrids, true, threadsPerAtom, subGroupSize); \
  91     INSTANTIATE_GATHER_2(order, numGrids, false, threadsPerAtom, subGroupSize);
  92
  93 #define INSTANTIATE_X(x, order, subGroupSize)                              \
  94     INSTANTIATE_##x(order, 1, ThreadsPerAtom::Order, subGroupSize);        \
  95     INSTANTIATE_##x(order, 1, ThreadsPerAtom::OrderSquared, subGroupSize); \
  96     INSTANTIATE_##x(order, 2, ThreadsPerAtom::Order, subGroupSize);        \
  97     INSTANTIATE_##x(order, 2, ThreadsPerAtom::OrderSquared, subGroupSize);
  98
  99 #define INSTANTIATE(order, subGroupSize)        \
 100     INSTANTIATE_X(SPREAD, order, subGroupSize); \
 101     INSTANTIATE_X(GATHER, order, subGroupSize);
 102
 103 #if GMX_SYCL_DPCPP
 104 INSTANTIATE(4, 16);
 105 #elif GMX_SYCL_HIPSYCL
 106 INSTANTIATE(4, 32);
 107 INSTANTIATE(4, 64);
 108 #endif
 109
 110
 111 //! Helper function to set proper kernel functor pointers
 112 template<int subGroupSize>
 113 static void setKernelPointers(struct PmeGpuProgramImpl* pmeGpuProgram)
 114 {
 115     /* Not all combinations of the splineAndSpread, spline and Spread kernels are required
 116      * If only the spline (without the spread) then it does not make sense not to write the data to global memory
 117      * Similarly the spread kernel (without the spline) implies that we should read the spline data from global memory
 118      */
 119     pmeGpuProgram->splineAndSpreadKernelSingle =
 120             new PmeSplineAndSpreadKernel<c_pmeOrder, true, true, c_wrapX, c_wrapY, 1, false, ThreadsPerAtom::OrderSquared, subGroupSize>();
 121     pmeGpuProgram->splineAndSpreadKernelThPerAtom4Single =
 122             new PmeSplineAndSpreadKernel<c_pmeOrder, true, true, c_wrapX, c_wrapY, 1, false, ThreadsPerAtom::Order, subGroupSize>();
 123     pmeGpuProgram->splineAndSpreadKernelWriteSplinesSingle =
 124             new PmeSplineAndSpreadKernel<c_pmeOrder, true, true, c_wrapX, c_wrapY, 1, true, ThreadsPerAtom::OrderSquared, subGroupSize>();
 125     pmeGpuProgram->splineAndSpreadKernelWriteSplinesThPerAtom4Single =
 126             new PmeSplineAndSpreadKernel<c_pmeOrder, true, true, c_wrapX, c_wrapY, 1, true, ThreadsPerAtom::Order, subGroupSize>();
 127     pmeGpuProgram->splineKernelSingle =
 128             new PmeSplineAndSpreadKernel<c_pmeOrder, true, false, c_wrapX, c_wrapY, 1, true, ThreadsPerAtom::OrderSquared, subGroupSize>();
 129     pmeGpuProgram->splineKernelThPerAtom4Single =
 130             new PmeSplineAndSpreadKernel<c_pmeOrder, true, false, c_wrapX, c_wrapY, 1, true, ThreadsPerAtom::Order, subGroupSize>();
 131     pmeGpuProgram->spreadKernelSingle =
 132             new PmeSplineAndSpreadKernel<c_pmeOrder, false, true, c_wrapX, c_wrapY, 1, true, ThreadsPerAtom::OrderSquared, subGroupSize>();
 133     pmeGpuProgram->spreadKernelThPerAtom4Single =
 134             new PmeSplineAndSpreadKernel<c_pmeOrder, false, true, c_wrapX, c_wrapY, 1, true, ThreadsPerAtom::Order, subGroupSize>();
 135     pmeGpuProgram->splineAndSpreadKernelDual =
 136             new PmeSplineAndSpreadKernel<c_pmeOrder, true, true, c_wrapX, c_wrapY, 2, false, ThreadsPerAtom::OrderSquared, subGroupSize>();
 137     pmeGpuProgram->splineAndSpreadKernelThPerAtom4Dual =
 138             new PmeSplineAndSpreadKernel<c_pmeOrder, true, true, c_wrapX, c_wrapY, 2, false, ThreadsPerAtom::Order, subGroupSize>();
 139     pmeGpuProgram->splineAndSpreadKernelWriteSplinesDual =
 140             new PmeSplineAndSpreadKernel<c_pmeOrder, true, true, c_wrapX, c_wrapY, 2, true, ThreadsPerAtom::OrderSquared, subGroupSize>();
 141     pmeGpuProgram->splineAndSpreadKernelWriteSplinesThPerAtom4Dual =
 142             new PmeSplineAndSpreadKernel<c_pmeOrder, true, true, c_wrapX, c_wrapY, 2, true, ThreadsPerAtom::Order, subGroupSize>();
 143     pmeGpuProgram->splineKernelDual =
 144             new PmeSplineAndSpreadKernel<c_pmeOrder, true, false, c_wrapX, c_wrapY, 2, true, ThreadsPerAtom::OrderSquared, subGroupSize>();
 145     pmeGpuProgram->splineKernelThPerAtom4Dual =
 146             new PmeSplineAndSpreadKernel<c_pmeOrder, true, false, c_wrapX, c_wrapY, 2, true, ThreadsPerAtom::Order, subGroupSize>();
 147     pmeGpuProgram->spreadKernelDual =
 148             new PmeSplineAndSpreadKernel<c_pmeOrder, false, true, c_wrapX, c_wrapY, 2, true, ThreadsPerAtom::OrderSquared, subGroupSize>();
 149     pmeGpuProgram->spreadKernelThPerAtom4Dual =
 150             new PmeSplineAndSpreadKernel<c_pmeOrder, false, true, c_wrapX, c_wrapY, 2, true, ThreadsPerAtom::Order, subGroupSize>();
 151     pmeGpuProgram->gatherKernelSingle =
 152             new PmeGatherKernel<c_pmeOrder, c_wrapX, c_wrapY, 1, false, ThreadsPerAtom::OrderSquared, subGroupSize>();
 153     pmeGpuProgram->gatherKernelThPerAtom4Single =
 154             new PmeGatherKernel<c_pmeOrder, c_wrapX, c_wrapY, 1, false, ThreadsPerAtom::Order, subGroupSize>();
 155     pmeGpuProgram->gatherKernelReadSplinesSingle =
 156             new PmeGatherKernel<c_pmeOrder, c_wrapX, c_wrapY, 1, true, ThreadsPerAtom::OrderSquared, subGroupSize>();
 157     pmeGpuProgram->gatherKernelReadSplinesThPerAtom4Single =
 158             new PmeGatherKernel<c_pmeOrder, c_wrapX, c_wrapY, 1, true, ThreadsPerAtom::Order, subGroupSize>();
 159     pmeGpuProgram->gatherKernelDual =
 160             new PmeGatherKernel<c_pmeOrder, c_wrapX, c_wrapY, 2, false, ThreadsPerAtom::OrderSquared, subGroupSize>();
 161     pmeGpuProgram->gatherKernelThPerAtom4Dual =
 162             new PmeGatherKernel<c_pmeOrder, c_wrapX, c_wrapY, 2, false, ThreadsPerAtom::Order, subGroupSize>();
 163     pmeGpuProgram->gatherKernelReadSplinesDual =
 164             new PmeGatherKernel<c_pmeOrder, c_wrapX, c_wrapY, 2, true, ThreadsPerAtom::OrderSquared, subGroupSize>();
 165     pmeGpuProgram->gatherKernelReadSplinesThPerAtom4Dual =
 166             new PmeGatherKernel<c_pmeOrder, c_wrapX, c_wrapY, 2, true, ThreadsPerAtom::Order, subGroupSize>();
 167 }
 168
 169 PmeGpuProgramImpl::PmeGpuProgramImpl(const DeviceContext& deviceContext) :
 170     deviceContext_(deviceContext)
 171 {
 172     // kernel parameters
 173     warpSize_             = subGroupSizeFromVendor(deviceContext.deviceInfo());
 174     spreadWorkGroupSize   = c_spreadMaxWarpsPerBlock * warpSize_;
 175     solveMaxWorkGroupSize = c_solveMaxWarpsPerBlock * warpSize_;
 176     gatherWorkGroupSize   = c_gatherMaxWarpsPerBlock * warpSize_;
 177
 178     switch (warpSize_)
 179     {
 180 #if GMX_SYCL_DPCPP
 181         case 16: setKernelPointers<16>(this); break;
 182 #elif GMX_SYCL_HIPSYCL
 183         case 32: setKernelPointers<32>(this); break;
 184         case 64: setKernelPointers<64>(this); break;
 185 #endif
 186         default: GMX_RELEASE_ASSERT(false, "Invalid sub group size");
 187     }
 188 }
 189
 190 PmeGpuProgramImpl::~PmeGpuProgramImpl()
 191 {
 192     delete splineKernelSingle;
 193     delete splineKernelThPerAtom4Single;
 194     delete spreadKernelSingle;
 195     delete spreadKernelThPerAtom4Single;
 196     delete splineAndSpreadKernelSingle;
 197     delete splineAndSpreadKernelThPerAtom4Single;
 198     delete splineAndSpreadKernelWriteSplinesSingle;
 199     delete splineAndSpreadKernelWriteSplinesThPerAtom4Single;
 200     delete splineKernelDual;
 201     delete splineKernelThPerAtom4Dual;
 202     delete spreadKernelDual;
 203     delete spreadKernelThPerAtom4Dual;
 204     delete splineAndSpreadKernelDual;
 205     delete splineAndSpreadKernelThPerAtom4Dual;
 206     delete splineAndSpreadKernelWriteSplinesDual;
 207     delete splineAndSpreadKernelWriteSplinesThPerAtom4Dual;
 208 }