2 * This file is part of the GROMACS molecular simulation package.
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.
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.
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.
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.
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.
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.
38 * Implements PmeGpuProgramImpl, which stores permanent PME GPU context-derived data,
39 * such as (compiled) kernel handles.
41 * \author Aleksei Iupinov <a.yupinov@gmail.com>
42 * \author Andrey Alekseenko <al42and@gmail.com>
43 * \ingroup module_ewald
47 #include "gromacs/hardware/device_information.h"
48 #include "gromacs/gpu_utils/gmxsycl.h"
49 #include "gromacs/gpu_utils/syclutils.h"
51 #include "pme_gpu_program_impl.h"
52 #include "pme_spread_sycl.h"
54 #include "pme_gpu_constants.h"
55 #include "pme_gpu_internal.h" // for GridOrdering enum
56 #include "pme_gpu_types_host.h"
58 // PME interpolation order
59 constexpr int c_pmeOrder = 4;
60 // These hardcoded spread/gather parameters refer to not-implemented PME GPU 2D decomposition in X/Y
61 constexpr bool c_wrapX = true;
62 constexpr bool c_wrapY = true;
64 static int subGroupSizeFromVendor(const DeviceInformation& deviceInfo)
66 switch (deviceInfo.deviceVendor)
68 case DeviceVendor::Amd: return 64; // Handle RDNA2 devices, Issue #3972.
69 case DeviceVendor::Intel: return 16; // TODO: Choose best value, Issue #4153.
70 case DeviceVendor::Nvidia: return 32;
71 default: GMX_RELEASE_ASSERT(false, "Unknown device vendor"); return 0;
75 #define INSTANTIATE_3(order, computeSplines, spreadCharges, numGrids, writeGlobal, threadsPerAtom, subGroupSize) \
76 extern template class PmeSplineAndSpreadKernel<order, computeSplines, spreadCharges, true, true, numGrids, writeGlobal, threadsPerAtom, subGroupSize>;
78 #define INSTANTIATE_2(order, numGrids, threadsPerAtom, subGroupSize) \
79 INSTANTIATE_3(order, true, true, numGrids, true, threadsPerAtom, subGroupSize); \
80 INSTANTIATE_3(order, true, false, numGrids, true, threadsPerAtom, subGroupSize); \
81 INSTANTIATE_3(order, false, true, numGrids, true, threadsPerAtom, subGroupSize); \
82 INSTANTIATE_3(order, true, true, numGrids, false, threadsPerAtom, subGroupSize);
84 #define INSTANTIATE(order, subGroupSize) \
85 INSTANTIATE_2(order, 1, ThreadsPerAtom::Order, subGroupSize); \
86 INSTANTIATE_2(order, 1, ThreadsPerAtom::OrderSquared, subGroupSize); \
87 INSTANTIATE_2(order, 2, ThreadsPerAtom::Order, subGroupSize); \
88 INSTANTIATE_2(order, 2, ThreadsPerAtom::OrderSquared, subGroupSize);
92 #elif GMX_SYCL_HIPSYCL
98 //! Helper function to set proper kernel functor pointers
99 template<int subGroupSize>
100 static void setKernelPointers(struct PmeGpuProgramImpl* pmeGpuProgram)
102 /* Not all combinations of the splineAndSpread, spline and Spread kernels are required
103 * If only the spline (without the spread) then it does not make sense not to write the data to global memory
104 * Similarly the spread kernel (without the spline) implies that we should read the spline data from global memory
106 pmeGpuProgram->splineAndSpreadKernelSingle =
107 new PmeSplineAndSpreadKernel<c_pmeOrder, true, true, c_wrapX, c_wrapY, 1, false, ThreadsPerAtom::OrderSquared, subGroupSize>();
108 pmeGpuProgram->splineAndSpreadKernelThPerAtom4Single =
109 new PmeSplineAndSpreadKernel<c_pmeOrder, true, true, c_wrapX, c_wrapY, 1, false, ThreadsPerAtom::Order, subGroupSize>();
110 pmeGpuProgram->splineAndSpreadKernelWriteSplinesSingle =
111 new PmeSplineAndSpreadKernel<c_pmeOrder, true, true, c_wrapX, c_wrapY, 1, true, ThreadsPerAtom::OrderSquared, subGroupSize>();
112 pmeGpuProgram->splineAndSpreadKernelWriteSplinesThPerAtom4Single =
113 new PmeSplineAndSpreadKernel<c_pmeOrder, true, true, c_wrapX, c_wrapY, 1, true, ThreadsPerAtom::Order, subGroupSize>();
114 pmeGpuProgram->splineKernelSingle =
115 new PmeSplineAndSpreadKernel<c_pmeOrder, true, false, c_wrapX, c_wrapY, 1, true, ThreadsPerAtom::OrderSquared, subGroupSize>();
116 pmeGpuProgram->splineKernelThPerAtom4Single =
117 new PmeSplineAndSpreadKernel<c_pmeOrder, true, false, c_wrapX, c_wrapY, 1, true, ThreadsPerAtom::Order, subGroupSize>();
118 pmeGpuProgram->spreadKernelSingle =
119 new PmeSplineAndSpreadKernel<c_pmeOrder, false, true, c_wrapX, c_wrapY, 1, true, ThreadsPerAtom::OrderSquared, subGroupSize>();
120 pmeGpuProgram->spreadKernelThPerAtom4Single =
121 new PmeSplineAndSpreadKernel<c_pmeOrder, false, true, c_wrapX, c_wrapY, 1, true, ThreadsPerAtom::Order, subGroupSize>();
122 pmeGpuProgram->splineAndSpreadKernelDual =
123 new PmeSplineAndSpreadKernel<c_pmeOrder, true, true, c_wrapX, c_wrapY, 2, false, ThreadsPerAtom::OrderSquared, subGroupSize>();
124 pmeGpuProgram->splineAndSpreadKernelThPerAtom4Dual =
125 new PmeSplineAndSpreadKernel<c_pmeOrder, true, true, c_wrapX, c_wrapY, 2, false, ThreadsPerAtom::Order, subGroupSize>();
126 pmeGpuProgram->splineAndSpreadKernelWriteSplinesDual =
127 new PmeSplineAndSpreadKernel<c_pmeOrder, true, true, c_wrapX, c_wrapY, 2, true, ThreadsPerAtom::OrderSquared, subGroupSize>();
128 pmeGpuProgram->splineAndSpreadKernelWriteSplinesThPerAtom4Dual =
129 new PmeSplineAndSpreadKernel<c_pmeOrder, true, true, c_wrapX, c_wrapY, 2, true, ThreadsPerAtom::Order, subGroupSize>();
130 pmeGpuProgram->splineKernelDual =
131 new PmeSplineAndSpreadKernel<c_pmeOrder, true, false, c_wrapX, c_wrapY, 2, true, ThreadsPerAtom::OrderSquared, subGroupSize>();
132 pmeGpuProgram->splineKernelThPerAtom4Dual =
133 new PmeSplineAndSpreadKernel<c_pmeOrder, true, false, c_wrapX, c_wrapY, 2, true, ThreadsPerAtom::Order, subGroupSize>();
134 pmeGpuProgram->spreadKernelDual =
135 new PmeSplineAndSpreadKernel<c_pmeOrder, false, true, c_wrapX, c_wrapY, 2, true, ThreadsPerAtom::OrderSquared, subGroupSize>();
136 pmeGpuProgram->spreadKernelThPerAtom4Dual =
137 new PmeSplineAndSpreadKernel<c_pmeOrder, false, true, c_wrapX, c_wrapY, 2, true, ThreadsPerAtom::Order, subGroupSize>();
140 PmeGpuProgramImpl::PmeGpuProgramImpl(const DeviceContext& deviceContext) :
141 deviceContext_(deviceContext)
144 warpSize_ = subGroupSizeFromVendor(deviceContext.deviceInfo());
145 spreadWorkGroupSize = c_spreadMaxWarpsPerBlock * warpSize_;
146 solveMaxWorkGroupSize = c_solveMaxWarpsPerBlock * warpSize_;
147 gatherWorkGroupSize = c_gatherMaxWarpsPerBlock * warpSize_;
152 case 16: setKernelPointers<16>(this); break;
153 #elif GMX_SYCL_HIPSYCL
154 case 32: setKernelPointers<32>(this); break;
155 case 64: setKernelPointers<64>(this); break;
157 default: GMX_RELEASE_ASSERT(false, "Invalid sub group size");
161 PmeGpuProgramImpl::~PmeGpuProgramImpl()
163 delete splineKernelSingle;
164 delete splineKernelThPerAtom4Single;
165 delete spreadKernelSingle;
166 delete spreadKernelThPerAtom4Single;
167 delete splineAndSpreadKernelSingle;
168 delete splineAndSpreadKernelThPerAtom4Single;
169 delete splineAndSpreadKernelWriteSplinesSingle;
170 delete splineAndSpreadKernelWriteSplinesThPerAtom4Single;
171 delete splineKernelDual;
172 delete splineKernelThPerAtom4Dual;
173 delete spreadKernelDual;
174 delete spreadKernelThPerAtom4Dual;
175 delete splineAndSpreadKernelDual;
176 delete splineAndSpreadKernelThPerAtom4Dual;
177 delete splineAndSpreadKernelWriteSplinesDual;
178 delete splineAndSpreadKernelWriteSplinesThPerAtom4Dual;