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.
38 * Data management and kernel launch functions for nbnxm sycl.
40 * \ingroup module_nbnxm
44 #include "gromacs/nbnxm/gpu_common.h"
45 #include "gromacs/utility/exceptions.h"
47 #include "nbnxm_sycl_kernel.h"
48 #include "nbnxm_sycl_kernel_pruneonly.h"
49 #include "nbnxm_sycl_types.h"
55 * Launch asynchronously the download of nonbonded forces from the GPU
56 * (and energies/shift forces if required).
58 void gpu_launch_cpyback(NbnxmGpu* nb,
59 struct nbnxn_atomdata_t* nbatom,
60 const gmx::StepWorkload& stepWork,
61 const AtomLocality atomLocality)
63 GMX_ASSERT(nb, "Need a valid nbnxn_gpu object");
65 const InteractionLocality iloc = gpuAtomToInteractionLocality(atomLocality);
66 const DeviceStream& deviceStream = *nb->deviceStreams[iloc];
67 sycl_atomdata_t* adat = nb->atdat;
69 /* don't launch non-local copy-back if there was no non-local work to do */
70 if ((iloc == InteractionLocality::NonLocal) && !haveGpuShortRangeWork(*nb, iloc))
72 nb->bNonLocalStreamActive = false;
76 int adatBegin, adatLen;
77 getGpuAtomRange(adat, atomLocality, &adatBegin, &adatLen);
79 // With DD the local D2H transfer can only start after the non-local kernel has finished.
80 if (iloc == InteractionLocality::Local && nb->bNonLocalStreamActive)
82 nb->nonlocal_done.waitForEvent();
86 * Skip if buffer ops / reduction is offloaded to the GPU.
88 if (!stepWork.useGpuFBufferOps)
90 GMX_ASSERT(adat->f.elementSize() == sizeof(float3),
91 "The size of the force buffer element should be equal to the size of float3.");
92 copyFromDeviceBuffer(reinterpret_cast<float3*>(nbatom->out[0].f.data()) + adatBegin,
97 GpuApiCallBehavior::Async,
101 /* After the non-local D2H is launched the nonlocal_done event can be
102 recorded which signals that the local D2H can proceed. This event is not
103 placed after the non-local kernel because we want the non-local data
105 if (iloc == InteractionLocality::NonLocal)
107 nb->nonlocal_done.markEvent(deviceStream);
108 nb->bNonLocalStreamActive = true;
111 /* only transfer energies in the local stream */
112 if (iloc == InteractionLocality::Local)
114 /* DtoH fshift when virial is needed */
115 if (stepWork.computeVirial)
117 GMX_ASSERT(sizeof(*nb->nbst.fshift) == adat->fShift.elementSize(),
118 "Sizes of host- and device-side shift vector elements should be the same.");
119 copyFromDeviceBuffer(
120 nb->nbst.fshift, &adat->fShift, 0, SHIFTS, deviceStream, GpuApiCallBehavior::Async, nullptr);
124 if (stepWork.computeEnergy)
126 GMX_ASSERT(sizeof(*nb->nbst.e_lj) == sizeof(float),
127 "Sizes of host- and device-side LJ energy terms should be the same.");
128 copyFromDeviceBuffer(
129 nb->nbst.e_lj, &adat->eLJ, 0, 1, deviceStream, GpuApiCallBehavior::Async, nullptr);
130 GMX_ASSERT(sizeof(*nb->nbst.e_el) == sizeof(float),
131 "Sizes of host- and device-side electrostatic energy terms should be the "
133 copyFromDeviceBuffer(
134 nb->nbst.e_el, &adat->eElec, 0, 1, deviceStream, GpuApiCallBehavior::Async, nullptr);
139 /*! \brief Launch asynchronously the xq buffer host to device copy. */
140 void gpu_copy_xq_to_gpu(NbnxmGpu* nb, const nbnxn_atomdata_t* nbatom, const AtomLocality atomLocality)
142 GMX_ASSERT(nb, "Need a valid nbnxn_gpu object");
143 validateGpuAtomLocality(atomLocality);
145 const InteractionLocality iloc = gpuAtomToInteractionLocality(atomLocality);
147 sycl_atomdata_t* adat = nb->atdat;
148 gpu_plist* plist = nb->plist[iloc];
149 const DeviceStream& deviceStream = *nb->deviceStreams[iloc];
151 /* Don't launch the non-local H2D copy if there is no dependent
152 work to do: neither non-local nor other (e.g. bonded) work
153 to do that has as input the nbnxn coordinates.
154 Doing the same for the local kernel is more complicated, since the
155 local part of the force array also depends on the non-local kernel.
156 So to avoid complicating the code and to reduce the risk of bugs,
157 we always call the local local x+q copy (and the rest of the local
158 work in nbnxn_gpu_launch_kernel().
160 if ((iloc == InteractionLocality::NonLocal) && !haveGpuShortRangeWork(*nb, iloc))
162 plist->haveFreshList = false;
166 int adatBegin, adatLen;
167 getGpuAtomRange(adat, atomLocality, &adatBegin, &adatLen);
170 GMX_ASSERT(adat->xq.elementSize() == sizeof(float4),
171 "The size of the xyzq buffer element should be equal to the size of float4.");
172 copyToDeviceBuffer(&adat->xq,
173 reinterpret_cast<const float4*>(nbatom->x().data()) + adatBegin,
177 GpuApiCallBehavior::Async,
180 /* No need to enforce stream synchronization with events like we do in CUDA/OpenCL.
181 * Runtime should do the scheduling correctly based on data dependencies. */
184 void gpu_launch_kernel_pruneonly(NbnxmGpu* nb, const InteractionLocality iloc, const int numParts)
186 gpu_plist* plist = nb->plist[iloc];
188 if (plist->haveFreshList)
190 GMX_ASSERT(numParts == 1, "With first pruning we expect 1 part");
192 /* Set rollingPruningNumParts to signal that it is not set */
193 plist->rollingPruningNumParts = 0;
194 plist->rollingPruningPart = 0;
198 if (plist->rollingPruningNumParts == 0)
200 plist->rollingPruningNumParts = numParts;
204 GMX_ASSERT(numParts == plist->rollingPruningNumParts,
205 "It is not allowed to change numParts in between list generation steps");
209 /* Use a local variable for part and update in plist, so we can return here
210 * without duplicating the part increment code.
212 const int part = plist->rollingPruningPart;
214 plist->rollingPruningPart++;
215 if (plist->rollingPruningPart >= plist->rollingPruningNumParts)
217 plist->rollingPruningPart = 0;
220 /* Compute the number of list entries to prune in this pass */
221 const int numSciInPart = (plist->nsci - part) / numParts;
223 /* Don't launch the kernel if there is no work to do */
224 if (numSciInPart <= 0)
226 plist->haveFreshList = false;
230 launchNbnxmKernelPruneOnly(nb, iloc, numParts, part, numSciInPart);
232 if (plist->haveFreshList)
234 plist->haveFreshList = false;
235 nb->didPrune[iloc] = true; // Mark that pruning has been done
239 nb->didRollingPrune[iloc] = true; // Mark that rolling pruning has been done
243 void gpu_launch_kernel(NbnxmGpu* nb, const gmx::StepWorkload& stepWork, const Nbnxm::InteractionLocality iloc)
245 const NBParamGpu* nbp = nb->nbparam;
246 gpu_plist* plist = nb->plist[iloc];
248 if (canSkipNonbondedWork(*nb, iloc))
250 plist->haveFreshList = false;
254 if (nbp->useDynamicPruning && plist->haveFreshList)
256 // Prunes for rlistOuter and rlistInner, sets plist->haveFreshList=false
257 gpu_launch_kernel_pruneonly(nb, iloc, 1);
260 if (plist->nsci == 0)
262 /* Don't launch an empty local kernel */
266 launchNbnxmKernel(nb, stepWork, iloc);