2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2020,2021, 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.
40 * \ingroup module_nbnxm
44 #include "nbnxm_sycl_kernel_pruneonly.h"
46 #include "gromacs/gpu_utils/devicebuffer.h"
47 #include "gromacs/gpu_utils/gmxsycl.h"
48 #include "gromacs/utility/template_mp.h"
50 #include "nbnxm_sycl_kernel_utils.h"
51 #include "nbnxm_sycl_types.h"
53 using cl::sycl::access::fence_space;
54 using cl::sycl::access::mode;
55 using cl::sycl::access::target;
57 //! \brief Class name for NBNXM prune-only kernel
58 template<bool haveFreshList>
59 class NbnxmKernelPruneOnly;
64 /*! \brief Prune-only kernel for NBNXM.
67 template<bool haveFreshList>
68 auto nbnxmKernelPruneOnly(cl::sycl::handler& cgh,
69 DeviceAccessor<Float4, mode::read> a_xq,
70 DeviceAccessor<Float3, mode::read> a_shiftVec,
71 DeviceAccessor<nbnxn_cj4_t, mode::read_write> a_plistCJ4,
72 DeviceAccessor<nbnxn_sci_t, mode::read> a_plistSci,
73 DeviceAccessor<unsigned int, haveFreshList ? mode::write : mode::read> a_plistIMask,
74 const float rlistOuterSq,
75 const float rlistInnerSq,
80 cgh.require(a_shiftVec);
81 cgh.require(a_plistCJ4);
82 cgh.require(a_plistSci);
83 cgh.require(a_plistIMask);
85 /* shmem buffer for i x+q pre-loading */
86 cl::sycl::accessor<Float4, 2, mode::read_write, target::local> sm_xq(
87 cl::sycl::range<2>(c_nbnxnGpuNumClusterPerSupercluster, c_clSize), cgh);
89 constexpr int warpSize = c_clSize * c_clSize / 2;
91 /* Somewhat weird behavior inherited from OpenCL.
92 * With clSize == 4, we use sub_group size of 16 (not enforced in OpenCL implementation, but chosen
93 * by the IGC compiler), however for data layout we consider it to be 8.
94 * Setting sub_group size to 8 slows down the prune-only kernel 1.5-2 times.
95 * For clSize == But we need to set specific sub_group size >= 32 for clSize == 8 for correctness,
96 * but it causes very poor performance.
98 constexpr int gmx_unused requiredSubGroupSize = (c_clSize == 4) ? 16 : warpSize;
101 * Work group (block) must have range (c_clSize, c_clSize, ...) (for itemIdx calculation, easy
103 return [=](cl::sycl::nd_item<3> itemIdx) [[intel::reqd_sub_group_size(requiredSubGroupSize)]]
105 // thread/block/warp id-s
106 const unsigned tidxi = itemIdx.get_local_id(2);
107 const unsigned tidxj = itemIdx.get_local_id(1);
108 const int tidx = tidxj * c_clSize + tidxi;
109 const unsigned tidxz = itemIdx.get_local_id(0);
110 const unsigned bidx = itemIdx.get_group(0);
112 const sycl_2020::sub_group sg = itemIdx.get_sub_group();
113 const unsigned widx = tidx / warpSize;
115 // my i super-cluster's index = sciOffset + current bidx * numParts + part
116 const nbnxn_sci_t nbSci = a_plistSci[bidx * numParts + part];
117 const int sci = nbSci.sci; /* super-cluster */
118 const int cij4Start = nbSci.cj4_ind_start; /* first ...*/
119 const int cij4End = nbSci.cj4_ind_end; /* and last index of j clusters */
123 for (int i = 0; i < c_nbnxnGpuNumClusterPerSupercluster; i += c_clSize)
125 /* Pre-load i-atom x and q into shared memory */
126 const int ci = sci * c_nbnxnGpuNumClusterPerSupercluster + tidxj + i;
127 const int ai = ci * c_clSize + tidxi;
129 /* We don't need q, but using float4 in shmem avoids bank conflicts.
130 (but it also wastes L2 bandwidth). */
131 const Float4 xq = a_xq[ai];
132 const Float3 shift = a_shiftVec[nbSci.shift];
133 const Float4 xi(xq[0] + shift[0], xq[1] + shift[1], xq[2] + shift[2], xq[3]);
134 sm_xq[tidxj + i][tidxi] = xi;
137 itemIdx.barrier(fence_space::local_space);
139 /* loop over the j clusters = seen by any of the atoms in the current super-cluster.
140 * The loop stride c_syclPruneKernelJ4Concurrency ensures that consecutive warps-pairs are
141 * assigned consecutive j4's entries. */
142 for (int j4 = cij4Start + tidxz; j4 < cij4End; j4 += c_syclPruneKernelJ4Concurrency)
144 unsigned imaskFull, imaskCheck, imaskNew;
146 if constexpr (haveFreshList)
148 /* Read the mask from the list transferred from the CPU */
149 imaskFull = a_plistCJ4[j4].imei[widx].imask;
150 /* We attempt to prune all pairs present in the original list */
151 imaskCheck = imaskFull;
156 /* Read the mask from the "warp-pruned" by rlistOuter mask array */
157 imaskFull = a_plistIMask[j4 * c_nbnxnGpuClusterpairSplit + widx];
158 /* Read the old rolling pruned mask, use as a base for new */
159 imaskNew = a_plistCJ4[j4].imei[widx].imask;
160 /* We only need to check pairs with different mask */
161 imaskCheck = (imaskNew ^ imaskFull);
166 for (int jm = 0; jm < c_nbnxnGpuJgroupSize; jm++)
168 if (imaskCheck & (superClInteractionMask << (jm * c_nbnxnGpuNumClusterPerSupercluster)))
170 unsigned mask_ji = (1U << (jm * c_nbnxnGpuNumClusterPerSupercluster));
171 // SYCL-TODO: Reevaluate prefetching methods
172 const int cj = a_plistCJ4[j4].cj[jm];
173 const int aj = cj * c_clSize + tidxj;
175 /* load j atom data */
176 const Float4 tmp = a_xq[aj];
177 const Float3 xj(tmp[0], tmp[1], tmp[2]);
179 for (int i = 0; i < c_nbnxnGpuNumClusterPerSupercluster; i++)
181 if (imaskCheck & mask_ji)
183 // load i-cluster coordinates from shmem
184 const Float4 xi = sm_xq[i][tidxi];
185 // distance between i and j atoms
186 Float3 rv(xi[0], xi[1], xi[2]);
188 const float r2 = norm2(rv);
190 /* If _none_ of the atoms pairs are in rlistOuter
191 * range, the bit corresponding to the current
192 * cluster-pair in imask gets set to 0. */
193 if (haveFreshList && !(sycl_2020::group_any_of(sg, r2 < rlistOuterSq)))
195 imaskFull &= ~mask_ji;
197 /* If any atom pair is within range, set the bit
198 * corresponding to the current cluster-pair. */
199 if (sycl_2020::group_any_of(sg, r2 < rlistInnerSq))
203 } // (imaskCheck & mask_ji)
204 /* shift the mask bit by 1 */
206 } // (int i = 0; i < c_nbnxnGpuNumClusterPerSupercluster; i++)
207 } // (imaskCheck & (superClInteractionMask << (jm * c_nbnxnGpuNumClusterPerSupercluster)))
208 } // for (int jm = 0; jm < c_nbnxnGpuJgroupSize; jm++)
210 if constexpr (haveFreshList)
212 /* copy the list pruned to rlistOuter to a separate buffer */
213 a_plistIMask[j4 * c_nbnxnGpuClusterpairSplit + widx] = imaskFull;
215 /* update the imask with only the pairs up to rlistInner */
216 a_plistCJ4[j4].imei[widx].imask = imaskNew;
218 } // for (int j4 = cij4_start + tidxz; j4 < cij4_end; j4 += c_syclPruneKernelJ4Concurrency)
222 //! \brief Leap Frog SYCL prune-only kernel launch code.
223 template<bool haveFreshList, class... Args>
224 cl::sycl::event launchNbnxmKernelPruneOnly(const DeviceStream& deviceStream,
225 const int numSciInPart,
228 using kernelNameType = NbnxmKernelPruneOnly<haveFreshList>;
230 /* Kernel launch config:
231 * - The thread block dimensions match the size of i-clusters, j-clusters,
232 * and j-cluster concurrency, in x, y, and z, respectively.
233 * - The 1D block-grid contains as many blocks as super-clusters.
235 const unsigned long numBlocks = numSciInPart;
236 const cl::sycl::range<3> blockSize{ c_syclPruneKernelJ4Concurrency, c_clSize, c_clSize };
237 const cl::sycl::range<3> globalSize{ numBlocks * blockSize[0], blockSize[1], blockSize[2] };
238 const cl::sycl::nd_range<3> range{ globalSize, blockSize };
240 cl::sycl::queue q = deviceStream.stream();
242 cl::sycl::event e = q.submit([&](cl::sycl::handler& cgh) {
243 auto kernel = nbnxmKernelPruneOnly<haveFreshList>(cgh, std::forward<Args>(args)...);
244 cgh.parallel_for<kernelNameType>(range, kernel);
250 //! \brief Select templated kernel and launch it.
251 template<class... Args>
252 cl::sycl::event chooseAndLaunchNbnxmKernelPruneOnly(bool haveFreshList, Args&&... args)
254 return gmx::dispatchTemplatedFunction(
255 [&](auto haveFreshList_) {
256 return launchNbnxmKernelPruneOnly<haveFreshList_>(std::forward<Args>(args)...);
261 void launchNbnxmKernelPruneOnly(NbnxmGpu* nb,
262 const InteractionLocality iloc,
265 const int numSciInPart)
267 NBAtomDataGpu* adat = nb->atdat;
268 NBParamGpu* nbp = nb->nbparam;
269 gpu_plist* plist = nb->plist[iloc];
270 const bool haveFreshList = plist->haveFreshList;
271 const DeviceStream& deviceStream = *nb->deviceStreams[iloc];
273 cl::sycl::event e = chooseAndLaunchNbnxmKernelPruneOnly(haveFreshList,