2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2016,2017,2018, 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.
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.
37 * \brief This file contains internal CUDA function implementations
38 * for performing the PME calculations on GPU.
40 * \author Aleksei Iupinov <a.yupinov@gmail.com>
50 #include "gromacs/gpu_utils/cudautils.cuh"
51 #include "gromacs/gpu_utils/pmalloc_cuda.h"
52 #include "gromacs/utility/gmxassert.h"
53 #include "gromacs/utility/smalloc.h"
56 #include "pme-3dfft.cuh"
59 int pme_gpu_get_atom_data_alignment(const PmeGpu *pmeGpu)
61 const int order = pmeGpu->common->pme_order;
62 GMX_ASSERT(order > 0, "Invalid PME order");
63 return PME_ATOM_DATA_ALIGNMENT;
66 int pme_gpu_get_atoms_per_warp(const PmeGpu *pmeGpu)
68 const int order = pmeGpu->common->pme_order;
69 GMX_ASSERT(order > 0, "Invalid PME order");
70 return PME_SPREADGATHER_ATOMS_PER_WARP;
73 void pme_gpu_synchronize(const PmeGpu *pmeGpu)
75 cudaError_t stat = cudaStreamSynchronize(pmeGpu->archSpecific->pmeStream);
76 CU_RET_ERR(stat, "Failed to synchronize the PME GPU stream!");
79 void pme_gpu_alloc_energy_virial(const PmeGpu *pmeGpu)
81 const size_t energyAndVirialSize = c_virialAndEnergyCount * sizeof(float);
82 cudaError_t stat = cudaMalloc((void **)&pmeGpu->kernelParams->constants.d_virialAndEnergy, energyAndVirialSize);
83 CU_RET_ERR(stat, "cudaMalloc failed on PME energy and virial");
84 pmalloc((void **)&pmeGpu->staging.h_virialAndEnergy, energyAndVirialSize);
87 void pme_gpu_free_energy_virial(PmeGpu *pmeGpu)
89 cudaError_t stat = cudaFree(pmeGpu->kernelParams->constants.d_virialAndEnergy);
90 CU_RET_ERR(stat, "cudaFree failed on PME energy and virial");
91 pmeGpu->kernelParams->constants.d_virialAndEnergy = nullptr;
92 pfree(pmeGpu->staging.h_virialAndEnergy);
93 pmeGpu->staging.h_virialAndEnergy = nullptr;
96 void pme_gpu_clear_energy_virial(const PmeGpu *pmeGpu)
98 cudaError_t stat = cudaMemsetAsync(pmeGpu->kernelParams->constants.d_virialAndEnergy, 0,
99 c_virialAndEnergyCount * sizeof(float), pmeGpu->archSpecific->pmeStream);
100 CU_RET_ERR(stat, "PME energy/virial cudaMemsetAsync error");
103 void pme_gpu_realloc_and_copy_bspline_values(const PmeGpu *pmeGpu)
105 const int splineValuesOffset[DIM] = {
107 pmeGpu->kernelParams->grid.realGridSize[XX],
108 pmeGpu->kernelParams->grid.realGridSize[XX] + pmeGpu->kernelParams->grid.realGridSize[YY]
110 memcpy((void *)&pmeGpu->kernelParams->grid.splineValuesOffset, &splineValuesOffset, sizeof(splineValuesOffset));
112 const int newSplineValuesSize = pmeGpu->kernelParams->grid.realGridSize[XX] +
113 pmeGpu->kernelParams->grid.realGridSize[YY] +
114 pmeGpu->kernelParams->grid.realGridSize[ZZ];
115 const bool shouldRealloc = (newSplineValuesSize > pmeGpu->archSpecific->splineValuesSize);
116 cu_realloc_buffered((void **)&pmeGpu->kernelParams->grid.d_splineModuli, nullptr, sizeof(float),
117 &pmeGpu->archSpecific->splineValuesSize, &pmeGpu->archSpecific->splineValuesSizeAlloc, newSplineValuesSize, pmeGpu->archSpecific->pmeStream, true);
120 /* Reallocate the host buffer */
121 pfree(pmeGpu->staging.h_splineModuli);
122 pmalloc((void **)&pmeGpu->staging.h_splineModuli, newSplineValuesSize * sizeof(float));
124 for (int i = 0; i < DIM; i++)
126 memcpy(pmeGpu->staging.h_splineModuli + splineValuesOffset[i], pmeGpu->common->bsp_mod[i].data(), pmeGpu->common->bsp_mod[i].size() * sizeof(float));
128 /* TODO: pin original buffer instead! */
129 cu_copy_H2D(pmeGpu->kernelParams->grid.d_splineModuli, pmeGpu->staging.h_splineModuli,
130 newSplineValuesSize * sizeof(float), pmeGpu->settings.transferKind, pmeGpu->archSpecific->pmeStream);
133 void pme_gpu_free_bspline_values(const PmeGpu *pmeGpu)
135 pfree(pmeGpu->staging.h_splineModuli);
136 cu_free_buffered(pmeGpu->kernelParams->grid.d_splineModuli, &pmeGpu->archSpecific->splineValuesSize,
137 &pmeGpu->archSpecific->splineValuesSizeAlloc);
140 void pme_gpu_realloc_forces(PmeGpu *pmeGpu)
142 const size_t newForcesSize = pmeGpu->nAtomsAlloc * DIM;
143 GMX_ASSERT(newForcesSize > 0, "Bad number of atoms in PME GPU");
144 cu_realloc_buffered((void **)&pmeGpu->kernelParams->atoms.d_forces, nullptr, sizeof(float),
145 &pmeGpu->archSpecific->forcesSize, &pmeGpu->archSpecific->forcesSizeAlloc, newForcesSize, pmeGpu->archSpecific->pmeStream, true);
146 pmeGpu->staging.h_forces.reserve(pmeGpu->nAtomsAlloc);
147 pmeGpu->staging.h_forces.resize(pmeGpu->kernelParams->atoms.nAtoms);
150 void pme_gpu_free_forces(const PmeGpu *pmeGpu)
152 cu_free_buffered(pmeGpu->kernelParams->atoms.d_forces, &pmeGpu->archSpecific->forcesSize, &pmeGpu->archSpecific->forcesSizeAlloc);
155 void pme_gpu_copy_input_forces(PmeGpu *pmeGpu)
157 const size_t forcesSize = DIM * pmeGpu->kernelParams->atoms.nAtoms * sizeof(float);
158 GMX_ASSERT(forcesSize > 0, "Bad number of atoms in PME GPU");
159 cu_copy_H2D(pmeGpu->kernelParams->atoms.d_forces, pmeGpu->staging.h_forces.data(), forcesSize, pmeGpu->settings.transferKind, pmeGpu->archSpecific->pmeStream);
162 void pme_gpu_copy_output_forces(PmeGpu *pmeGpu)
164 const size_t forcesSize = DIM * pmeGpu->kernelParams->atoms.nAtoms * sizeof(float);
165 GMX_ASSERT(forcesSize > 0, "Bad number of atoms in PME GPU");
166 cu_copy_D2H(pmeGpu->staging.h_forces.data(), pmeGpu->kernelParams->atoms.d_forces, forcesSize, pmeGpu->settings.transferKind, pmeGpu->archSpecific->pmeStream);
169 void pme_gpu_realloc_coordinates(const PmeGpu *pmeGpu)
171 const size_t newCoordinatesSize = pmeGpu->nAtomsAlloc * DIM;
172 GMX_ASSERT(newCoordinatesSize > 0, "Bad number of atoms in PME GPU");
173 cu_realloc_buffered((void **)&pmeGpu->kernelParams->atoms.d_coordinates, nullptr, sizeof(float),
174 &pmeGpu->archSpecific->coordinatesSize, &pmeGpu->archSpecific->coordinatesSizeAlloc, newCoordinatesSize, pmeGpu->archSpecific->pmeStream, true);
177 const size_t paddingIndex = DIM * pmeGpu->kernelParams->atoms.nAtoms;
178 const size_t paddingCount = DIM * pmeGpu->nAtomsAlloc - paddingIndex;
179 if (paddingCount > 0)
181 cudaError_t stat = cudaMemsetAsync(pmeGpu->kernelParams->atoms.d_coordinates + paddingIndex, 0, paddingCount * sizeof(float), pmeGpu->archSpecific->pmeStream);
182 CU_RET_ERR(stat, "PME failed to clear the padded coordinates");
187 void pme_gpu_copy_input_coordinates(const PmeGpu *pmeGpu, const rvec *h_coordinates)
189 GMX_ASSERT(h_coordinates, "Bad host-side coordinate buffer in PME GPU");
191 GMX_RELEASE_ASSERT(false, "Only single precision is supported");
192 GMX_UNUSED_VALUE(h_coordinates);
194 cu_copy_H2D(pmeGpu->kernelParams->atoms.d_coordinates, const_cast<rvec *>(h_coordinates),
195 pmeGpu->kernelParams->atoms.nAtoms * sizeof(rvec), pmeGpu->settings.transferKind, pmeGpu->archSpecific->pmeStream);
199 void pme_gpu_free_coordinates(const PmeGpu *pmeGpu)
201 cu_free_buffered(pmeGpu->kernelParams->atoms.d_coordinates, &pmeGpu->archSpecific->coordinatesSize, &pmeGpu->archSpecific->coordinatesSizeAlloc);
204 void pme_gpu_realloc_and_copy_input_coefficients(const PmeGpu *pmeGpu, const float *h_coefficients)
206 GMX_ASSERT(h_coefficients, "Bad host-side charge buffer in PME GPU");
207 const size_t newCoefficientsSize = pmeGpu->nAtomsAlloc;
208 GMX_ASSERT(newCoefficientsSize > 0, "Bad number of atoms in PME GPU");
209 cu_realloc_buffered((void **)&pmeGpu->kernelParams->atoms.d_coefficients, nullptr, sizeof(float),
210 &pmeGpu->archSpecific->coefficientsSize, &pmeGpu->archSpecific->coefficientsSizeAlloc,
211 newCoefficientsSize, pmeGpu->archSpecific->pmeStream, true);
212 cu_copy_H2D(pmeGpu->kernelParams->atoms.d_coefficients, const_cast<float *>(h_coefficients),
213 pmeGpu->kernelParams->atoms.nAtoms * sizeof(float), pmeGpu->settings.transferKind, pmeGpu->archSpecific->pmeStream);
216 const size_t paddingIndex = pmeGpu->kernelParams->atoms.nAtoms;
217 const size_t paddingCount = pmeGpu->nAtomsAlloc - paddingIndex;
218 if (paddingCount > 0)
220 cudaError_t stat = cudaMemsetAsync(pmeGpu->kernelParams->atoms.d_coefficients + paddingIndex, 0, paddingCount * sizeof(float), pmeGpu->archSpecific->pmeStream);
221 CU_RET_ERR(stat, "PME failed to clear the padded charges");
226 void pme_gpu_free_coefficients(const PmeGpu *pmeGpu)
228 cu_free_buffered(pmeGpu->kernelParams->atoms.d_coefficients, &pmeGpu->archSpecific->coefficientsSize, &pmeGpu->archSpecific->coefficientsSizeAlloc);
231 void pme_gpu_realloc_spline_data(const PmeGpu *pmeGpu)
233 const int order = pmeGpu->common->pme_order;
234 const int alignment = pme_gpu_get_atoms_per_warp(pmeGpu);
235 const size_t nAtomsPadded = ((pmeGpu->nAtomsAlloc + alignment - 1) / alignment) * alignment;
236 const int newSplineDataSize = DIM * order * nAtomsPadded;
237 GMX_ASSERT(newSplineDataSize > 0, "Bad number of atoms in PME GPU");
238 /* Two arrays of the same size */
239 const bool shouldRealloc = (newSplineDataSize > pmeGpu->archSpecific->splineDataSize);
240 int currentSizeTemp = pmeGpu->archSpecific->splineDataSize;
241 int currentSizeTempAlloc = pmeGpu->archSpecific->splineDataSizeAlloc;
242 cu_realloc_buffered((void **)&pmeGpu->kernelParams->atoms.d_theta, nullptr, sizeof(float),
243 ¤tSizeTemp, ¤tSizeTempAlloc, newSplineDataSize, pmeGpu->archSpecific->pmeStream, true);
244 cu_realloc_buffered((void **)&pmeGpu->kernelParams->atoms.d_dtheta, nullptr, sizeof(float),
245 &pmeGpu->archSpecific->splineDataSize, &pmeGpu->archSpecific->splineDataSizeAlloc, newSplineDataSize, pmeGpu->archSpecific->pmeStream, true);
246 // the host side reallocation
249 pfree(pmeGpu->staging.h_theta);
250 pmalloc((void **)&pmeGpu->staging.h_theta, newSplineDataSize * sizeof(float));
251 pfree(pmeGpu->staging.h_dtheta);
252 pmalloc((void **)&pmeGpu->staging.h_dtheta, newSplineDataSize * sizeof(float));
256 void pme_gpu_free_spline_data(const PmeGpu *pmeGpu)
258 /* Two arrays of the same size */
259 cu_free_buffered(pmeGpu->kernelParams->atoms.d_theta);
260 cu_free_buffered(pmeGpu->kernelParams->atoms.d_dtheta, &pmeGpu->archSpecific->splineDataSize, &pmeGpu->archSpecific->splineDataSizeAlloc);
261 pfree(pmeGpu->staging.h_theta);
262 pfree(pmeGpu->staging.h_dtheta);
265 void pme_gpu_realloc_grid_indices(const PmeGpu *pmeGpu)
267 const size_t newIndicesSize = DIM * pmeGpu->nAtomsAlloc;
268 GMX_ASSERT(newIndicesSize > 0, "Bad number of atoms in PME GPU");
269 cu_realloc_buffered((void **)&pmeGpu->kernelParams->atoms.d_gridlineIndices, nullptr, sizeof(int),
270 &pmeGpu->archSpecific->gridlineIndicesSize, &pmeGpu->archSpecific->gridlineIndicesSizeAlloc, newIndicesSize, pmeGpu->archSpecific->pmeStream, true);
271 pfree(pmeGpu->staging.h_gridlineIndices);
272 pmalloc((void **)&pmeGpu->staging.h_gridlineIndices, newIndicesSize * sizeof(int));
275 void pme_gpu_free_grid_indices(const PmeGpu *pmeGpu)
277 cu_free_buffered(pmeGpu->kernelParams->atoms.d_gridlineIndices, &pmeGpu->archSpecific->gridlineIndicesSize, &pmeGpu->archSpecific->gridlineIndicesSizeAlloc);
278 pfree(pmeGpu->staging.h_gridlineIndices);
281 void pme_gpu_realloc_grids(PmeGpu *pmeGpu)
283 auto *kernelParamsPtr = pmeGpu->kernelParams.get();
284 const int newRealGridSize = kernelParamsPtr->grid.realGridSizePadded[XX] *
285 kernelParamsPtr->grid.realGridSizePadded[YY] *
286 kernelParamsPtr->grid.realGridSizePadded[ZZ];
287 const int newComplexGridSize = kernelParamsPtr->grid.complexGridSizePadded[XX] *
288 kernelParamsPtr->grid.complexGridSizePadded[YY] *
289 kernelParamsPtr->grid.complexGridSizePadded[ZZ] * 2;
290 // Multiplied by 2 because we count complex grid size for complex numbers, but all allocations/pointers are float
291 if (pmeGpu->archSpecific->performOutOfPlaceFFT)
293 /* 2 separate grids */
294 cu_realloc_buffered((void **)&kernelParamsPtr->grid.d_fourierGrid, nullptr, sizeof(float),
295 &pmeGpu->archSpecific->complexGridSize, &pmeGpu->archSpecific->complexGridSizeAlloc,
296 newComplexGridSize, pmeGpu->archSpecific->pmeStream, true);
297 cu_realloc_buffered((void **)&kernelParamsPtr->grid.d_realGrid, nullptr, sizeof(float),
298 &pmeGpu->archSpecific->realGridSize, &pmeGpu->archSpecific->realGridSizeAlloc,
299 newRealGridSize, pmeGpu->archSpecific->pmeStream, true);
303 /* A single buffer so that any grid will fit */
304 const int newGridsSize = std::max(newRealGridSize, newComplexGridSize);
305 cu_realloc_buffered((void **)&kernelParamsPtr->grid.d_realGrid, nullptr, sizeof(float),
306 &pmeGpu->archSpecific->realGridSize, &pmeGpu->archSpecific->realGridSizeAlloc,
307 newGridsSize, pmeGpu->archSpecific->pmeStream, true);
308 kernelParamsPtr->grid.d_fourierGrid = kernelParamsPtr->grid.d_realGrid;
309 pmeGpu->archSpecific->complexGridSize = pmeGpu->archSpecific->realGridSize;
310 // the size might get used later for copying the grid
314 void pme_gpu_free_grids(const PmeGpu *pmeGpu)
316 if (pmeGpu->archSpecific->performOutOfPlaceFFT)
318 cu_free_buffered(pmeGpu->kernelParams->grid.d_fourierGrid);
320 cu_free_buffered(pmeGpu->kernelParams->grid.d_realGrid,
321 &pmeGpu->archSpecific->realGridSize, &pmeGpu->archSpecific->realGridSizeAlloc);
324 void pme_gpu_clear_grids(const PmeGpu *pmeGpu)
326 cudaError_t stat = cudaMemsetAsync(pmeGpu->kernelParams->grid.d_realGrid, 0,
327 pmeGpu->archSpecific->realGridSize * sizeof(float), pmeGpu->archSpecific->pmeStream);
328 /* Should the complex grid be cleared in some weird case? */
329 CU_RET_ERR(stat, "cudaMemsetAsync on the PME grid error");
332 void pme_gpu_realloc_and_copy_fract_shifts(PmeGpu *pmeGpu)
334 pme_gpu_free_fract_shifts(pmeGpu);
336 auto *kernelParamsPtr = pmeGpu->kernelParams.get();
338 const int nx = kernelParamsPtr->grid.realGridSize[XX];
339 const int ny = kernelParamsPtr->grid.realGridSize[YY];
340 const int nz = kernelParamsPtr->grid.realGridSize[ZZ];
341 const int cellCount = c_pmeNeighborUnitcellCount;
342 const int gridDataOffset[DIM] = {0, cellCount * nx, cellCount * (nx + ny)};
344 memcpy(kernelParamsPtr->grid.tablesOffsets, &gridDataOffset, sizeof(gridDataOffset));
346 const int newFractShiftsSize = cellCount * (nx + ny + nz);
348 initParamLookupTable(kernelParamsPtr->grid.d_fractShiftsTable,
349 kernelParamsPtr->fractShiftsTableTexture,
350 pmeGpu->common->fsh.data(),
354 initParamLookupTable(kernelParamsPtr->grid.d_gridlineIndicesTable,
355 kernelParamsPtr->gridlineIndicesTableTexture,
356 pmeGpu->common->nn.data(),
361 void pme_gpu_free_fract_shifts(const PmeGpu *pmeGpu)
363 auto *kernelParamsPtr = pmeGpu->kernelParams.get();
364 destroyParamLookupTable(kernelParamsPtr->grid.d_fractShiftsTable,
365 kernelParamsPtr->fractShiftsTableTexture,
367 destroyParamLookupTable(kernelParamsPtr->grid.d_gridlineIndicesTable,
368 kernelParamsPtr->gridlineIndicesTableTexture,
372 bool pme_gpu_stream_query(const PmeGpu *pmeGpu)
374 return haveStreamTasksCompleted(pmeGpu->archSpecific->pmeStream);
377 void pme_gpu_copy_input_gather_grid(const PmeGpu *pmeGpu, float *h_grid)
379 const size_t gridSize = pmeGpu->archSpecific->realGridSize * sizeof(float);
380 cu_copy_H2D(pmeGpu->kernelParams->grid.d_realGrid, h_grid, gridSize, pmeGpu->settings.transferKind, pmeGpu->archSpecific->pmeStream);
383 void pme_gpu_copy_output_spread_grid(const PmeGpu *pmeGpu, float *h_grid)
385 const size_t gridSize = pmeGpu->archSpecific->realGridSize * sizeof(float);
386 cu_copy_D2H(h_grid, pmeGpu->kernelParams->grid.d_realGrid, gridSize, pmeGpu->settings.transferKind, pmeGpu->archSpecific->pmeStream);
387 cudaError_t stat = cudaEventRecord(pmeGpu->archSpecific->syncSpreadGridD2H, pmeGpu->archSpecific->pmeStream);
388 CU_RET_ERR(stat, "PME spread grid sync event record failure");
391 void pme_gpu_copy_output_spread_atom_data(const PmeGpu *pmeGpu)
393 const int alignment = pme_gpu_get_atoms_per_warp(pmeGpu);
394 const size_t nAtomsPadded = ((pmeGpu->nAtomsAlloc + alignment - 1) / alignment) * alignment;
395 const size_t splinesSize = DIM * nAtomsPadded * pmeGpu->common->pme_order * sizeof(float);
396 auto *kernelParamsPtr = pmeGpu->kernelParams.get();
397 cu_copy_D2H(pmeGpu->staging.h_dtheta, kernelParamsPtr->atoms.d_dtheta, splinesSize, pmeGpu->settings.transferKind, pmeGpu->archSpecific->pmeStream);
398 cu_copy_D2H(pmeGpu->staging.h_theta, kernelParamsPtr->atoms.d_theta, splinesSize, pmeGpu->settings.transferKind, pmeGpu->archSpecific->pmeStream);
399 cu_copy_D2H(pmeGpu->staging.h_gridlineIndices, kernelParamsPtr->atoms.d_gridlineIndices,
400 kernelParamsPtr->atoms.nAtoms * DIM * sizeof(int), pmeGpu->settings.transferKind, pmeGpu->archSpecific->pmeStream);
403 void pme_gpu_copy_input_gather_atom_data(const PmeGpu *pmeGpu)
405 const int alignment = pme_gpu_get_atoms_per_warp(pmeGpu);
406 const size_t nAtomsPadded = ((pmeGpu->nAtomsAlloc + alignment - 1) / alignment) * alignment;
407 const size_t splinesSize = DIM * nAtomsPadded * pmeGpu->common->pme_order * sizeof(float);
408 auto *kernelParamsPtr = pmeGpu->kernelParams.get();
411 const size_t gridlineIndicesSizePerAtom = DIM * sizeof(int);
412 const size_t splineDataSizePerAtom = pmeGpu->common->pme_order * DIM * sizeof(float);
413 // TODO: could clear only the padding and not the whole thing, but this is a test-exclusive code anyway
414 CU_RET_ERR(cudaMemsetAsync(kernelParamsPtr->atoms.d_gridlineIndices, 0, pmeGpu->nAtomsAlloc * gridlineIndicesSizePerAtom, pmeGpu->archSpecific->pmeStream),
415 "PME failed to clear the gridline indices");
416 CU_RET_ERR(cudaMemsetAsync(kernelParamsPtr->atoms.d_dtheta, 0, pmeGpu->nAtomsAlloc * splineDataSizePerAtom, pmeGpu->archSpecific->pmeStream),
417 "PME failed to clear the spline derivatives");
418 CU_RET_ERR(cudaMemsetAsync(kernelParamsPtr->atoms.d_theta, 0, pmeGpu->nAtomsAlloc * splineDataSizePerAtom, pmeGpu->archSpecific->pmeStream),
419 "PME failed to clear the spline values");
421 cu_copy_H2D(kernelParamsPtr->atoms.d_dtheta, pmeGpu->staging.h_dtheta, splinesSize, pmeGpu->settings.transferKind, pmeGpu->archSpecific->pmeStream);
422 cu_copy_H2D(kernelParamsPtr->atoms.d_theta, pmeGpu->staging.h_theta, splinesSize, pmeGpu->settings.transferKind, pmeGpu->archSpecific->pmeStream);
423 cu_copy_H2D(kernelParamsPtr->atoms.d_gridlineIndices, pmeGpu->staging.h_gridlineIndices,
424 kernelParamsPtr->atoms.nAtoms * DIM * sizeof(int), pmeGpu->settings.transferKind, pmeGpu->archSpecific->pmeStream);
427 void pme_gpu_sync_spread_grid(const PmeGpu *pmeGpu)
429 cudaError_t stat = cudaEventSynchronize(pmeGpu->archSpecific->syncSpreadGridD2H);
430 CU_RET_ERR(stat, "Error while waiting for the PME GPU spread grid to be copied to the host");
433 void pme_gpu_init_internal(PmeGpu *pmeGpu)
435 /* Allocate the target-specific structures */
436 pmeGpu->archSpecific.reset(new PmeGpuSpecific());
437 pmeGpu->kernelParams.reset(new PmeGpuKernelParams());
439 pmeGpu->archSpecific->performOutOfPlaceFFT = true;
440 /* This should give better performance, according to the cuFFT documentation.
441 * The performance seems to be the same though.
442 * TODO: PME could also try to pick up nice grid sizes (with factors of 2, 3, 5, 7).
445 /* WARNING: CUDA timings are incorrect with multiple streams.
446 * This is the main reason why they are disabled by default.
448 // TODO: Consider turning on by default when we can detect nr of streams.
449 pmeGpu->archSpecific->useTiming = (getenv("GMX_ENABLE_GPU_TIMING") != nullptr);
451 /* Creating a PME CUDA stream */
453 int highest_priority, lowest_priority;
454 stat = cudaDeviceGetStreamPriorityRange(&lowest_priority, &highest_priority);
455 CU_RET_ERR(stat, "PME cudaDeviceGetStreamPriorityRange failed");
456 stat = cudaStreamCreateWithPriority(&pmeGpu->archSpecific->pmeStream,
457 cudaStreamDefault, //cudaStreamNonBlocking,
459 CU_RET_ERR(stat, "cudaStreamCreateWithPriority on the PME stream failed");
462 void pme_gpu_destroy_specific(const PmeGpu *pmeGpu)
464 /* Destroy the CUDA stream */
465 cudaError_t stat = cudaStreamDestroy(pmeGpu->archSpecific->pmeStream);
466 CU_RET_ERR(stat, "PME cudaStreamDestroy error");
469 void pme_gpu_init_sync_events(const PmeGpu *pmeGpu)
471 const auto eventFlags = cudaEventDisableTiming;
472 CU_RET_ERR(cudaEventCreateWithFlags(&pmeGpu->archSpecific->syncSpreadGridD2H, eventFlags), "cudaEventCreate on syncSpreadGridD2H failed");
475 void pme_gpu_destroy_sync_events(const PmeGpu *pmeGpu)
477 CU_RET_ERR(cudaEventDestroy(pmeGpu->archSpecific->syncSpreadGridD2H), "cudaEventDestroy failed on syncSpreadGridD2H");
480 void pme_gpu_reinit_3dfft(const PmeGpu *pmeGpu)
482 if (pme_gpu_performs_FFT(pmeGpu))
484 pmeGpu->archSpecific->fftSetup.resize(0);
485 for (int i = 0; i < pmeGpu->common->ngrids; i++)
487 pmeGpu->archSpecific->fftSetup.push_back(std::unique_ptr<GpuParallel3dFft>(new GpuParallel3dFft(pmeGpu)));
492 void pme_gpu_destroy_3dfft(const PmeGpu *pmeGpu)
494 pmeGpu->archSpecific->fftSetup.resize(0);