2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2016,2017,2018,2019, 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 * Implements common routines for PME tests.
39 * \author Aleksei Iupinov <a.yupinov@gmail.com>
40 * \ingroup module_ewald
44 #include "pmetestcommon.h"
50 #include "gromacs/domdec/domdec.h"
51 #include "gromacs/ewald/pme_gather.h"
52 #include "gromacs/ewald/pme_gpu_internal.h"
53 #include "gromacs/ewald/pme_grid.h"
54 #include "gromacs/ewald/pme_internal.h"
55 #include "gromacs/ewald/pme_redistribute.h"
56 #include "gromacs/ewald/pme_solve.h"
57 #include "gromacs/ewald/pme_spread.h"
58 #include "gromacs/fft/parallel_3dfft.h"
59 #include "gromacs/gpu_utils/gpu_utils.h"
60 #include "gromacs/math/invertmatrix.h"
61 #include "gromacs/mdtypes/commrec.h"
62 #include "gromacs/pbcutil/pbc.h"
63 #include "gromacs/topology/topology.h"
64 #include "gromacs/utility/exceptions.h"
65 #include "gromacs/utility/gmxassert.h"
66 #include "gromacs/utility/logger.h"
67 #include "gromacs/utility/stringutil.h"
69 #include "testutils/testasserts.h"
76 bool pmeSupportsInputForMode(const gmx_hw_info_t &hwinfo,
77 const t_inputrec *inputRec,
89 implemented = (pme_gpu_supports_build(nullptr) &&
90 pme_gpu_supports_hardware(hwinfo, nullptr) &&
91 pme_gpu_supports_input(*inputRec, mtop, nullptr));
95 GMX_THROW(InternalError("Test not implemented for this mode"));
100 uint64_t getSplineModuliDoublePrecisionUlps(int splineOrder)
102 /* Arbitrary ulp tolerance for sine/cosine implementation. It's
103 * hard to know what to pick without testing lots of
104 * implementations. */
105 const uint64_t sineUlps = 10;
106 return 4 * (splineOrder - 2) + 2 * sineUlps * splineOrder;
109 //! PME initialization - internal
110 static PmeSafePointer pmeInitInternal(const t_inputrec *inputRec,
112 const gmx_device_info_t *gpuInfo,
113 PmeGpuProgramHandle pmeGpuProgram,
114 const Matrix3x3 &box,
115 real ewaldCoeff_q = 1.0F,
116 real ewaldCoeff_lj = 1.0F
119 const MDLogger dummyLogger;
120 const auto runMode = (mode == CodePath::CPU) ? PmeRunMode::CPU : PmeRunMode::Mixed;
121 t_commrec dummyCommrec = {0};
122 NumPmeDomains numPmeDomains = { 1, 1 };
123 gmx_pme_t *pmeDataRaw = gmx_pme_init(&dummyCommrec, numPmeDomains, inputRec, false, false, true,
124 ewaldCoeff_q, ewaldCoeff_lj, 1, runMode, nullptr, gpuInfo, pmeGpuProgram, dummyLogger);
125 PmeSafePointer pme(pmeDataRaw); // taking ownership
127 // TODO get rid of this with proper matrix type
129 for (int i = 0; i < DIM; i++)
131 for (int j = 0; j < DIM; j++)
133 boxTemp[i][j] = box[i * DIM + j];
136 const char *boxError = check_box(-1, boxTemp);
137 GMX_RELEASE_ASSERT(boxError == nullptr, boxError);
142 invertBoxMatrix(boxTemp, pme->recipbox);
146 pme_gpu_set_testing(pme->gpu, true);
147 pme_gpu_update_input_box(pme->gpu, boxTemp);
151 GMX_THROW(InternalError("Test not implemented for this mode"));
157 //! Simple PME initialization based on input, no atom data
158 PmeSafePointer pmeInitEmpty(const t_inputrec *inputRec,
160 const gmx_device_info_t *gpuInfo,
161 PmeGpuProgramHandle pmeGpuProgram,
162 const Matrix3x3 &box,
167 return pmeInitInternal(inputRec, mode, gpuInfo, pmeGpuProgram, box, ewaldCoeff_q, ewaldCoeff_lj);
168 // hiding the fact that PME actually needs to know the number of atoms in advance
171 //! PME initialization with atom data
172 PmeSafePointer pmeInitAtoms(const t_inputrec *inputRec,
174 const gmx_device_info_t *gpuInfo,
175 PmeGpuProgramHandle pmeGpuProgram,
176 const CoordinatesVector &coordinates,
177 const ChargesVector &charges,
181 const index atomCount = coordinates.size();
182 GMX_RELEASE_ASSERT(atomCount == charges.ssize(), "Mismatch in atom data");
183 PmeSafePointer pmeSafe = pmeInitInternal(inputRec, mode, gpuInfo, pmeGpuProgram, box);
184 PmeAtomComm *atc = nullptr;
189 atc = &(pmeSafe->atc[0]);
190 atc->x = coordinates;
191 atc->coefficient = charges;
192 gmx_pme_reinit_atoms(pmeSafe.get(), atomCount, charges.data());
193 /* With decomposition there would be more boilerplate atc code here, e.g. do_redist_pos_coeffs */
197 // TODO: Avoid use of atc in the GPU code path
198 atc = &(pmeSafe->atc[0]);
199 // We need to set atc->n for passing the size in the tests
200 atc->setNumAtoms(atomCount);
201 gmx_pme_reinit_atoms(pmeSafe.get(), atomCount, charges.data());
202 pme_gpu_copy_input_coordinates(pmeSafe->gpu, as_rvec_array(coordinates.data()));
206 GMX_THROW(InternalError("Test not implemented for this mode"));
212 //! Getting local PME real grid pointer for test I/O
213 static real *pmeGetRealGridInternal(const gmx_pme_t *pme)
215 const size_t gridIndex = 0;
216 return pme->fftgrid[gridIndex];
219 //! Getting local PME real grid dimensions
220 static void pmeGetRealGridSizesInternal(const gmx_pme_t *pme,
222 IVec &gridSize, //NOLINT(google-runtime-references)
223 IVec &paddedGridSize) //NOLINT(google-runtime-references)
225 const size_t gridIndex = 0;
226 IVec gridOffsetUnused;
230 gmx_parallel_3dfft_real_limits(pme->pfft_setup[gridIndex], gridSize, gridOffsetUnused, paddedGridSize);
234 pme_gpu_get_real_grid_sizes(pme->gpu, &gridSize, &paddedGridSize);
238 GMX_THROW(InternalError("Test not implemented for this mode"));
242 //! Getting local PME complex grid pointer for test I/O
243 static t_complex *pmeGetComplexGridInternal(const gmx_pme_t *pme)
245 const size_t gridIndex = 0;
246 return pme->cfftgrid[gridIndex];
249 //! Getting local PME complex grid dimensions
250 static void pmeGetComplexGridSizesInternal(const gmx_pme_t *pme,
251 IVec &gridSize, //NOLINT(google-runtime-references)
252 IVec &paddedGridSize) //NOLINT(google-runtime-references)
254 const size_t gridIndex = 0;
255 IVec gridOffsetUnused, complexOrderUnused;
256 gmx_parallel_3dfft_complex_limits(pme->pfft_setup[gridIndex], complexOrderUnused, gridSize, gridOffsetUnused, paddedGridSize); //TODO: what about YZX ordering?
259 //! Getting the PME grid memory buffer and its sizes - template definition
260 template<typename ValueType> static void pmeGetGridAndSizesInternal(const gmx_pme_t * /*unused*/, CodePath /*unused*/, ValueType * & /*unused*/, IVec & /*unused*/, IVec & /*unused*/) //NOLINT(google-runtime-references)
262 GMX_THROW(InternalError("Deleted function call"));
263 // explicitly deleting general template does not compile in clang/icc, see https://llvm.org/bugs/show_bug.cgi?id=17537
266 //! Getting the PME real grid memory buffer and its sizes
267 template<> void pmeGetGridAndSizesInternal<real>(const gmx_pme_t *pme, CodePath mode, real * &grid, IVec &gridSize, IVec &paddedGridSize)
269 grid = pmeGetRealGridInternal(pme);
270 pmeGetRealGridSizesInternal(pme, mode, gridSize, paddedGridSize);
273 //! Getting the PME complex grid memory buffer and its sizes
274 template<> void pmeGetGridAndSizesInternal<t_complex>(const gmx_pme_t *pme, CodePath /*unused*/, t_complex * &grid, IVec &gridSize, IVec &paddedGridSize)
276 grid = pmeGetComplexGridInternal(pme);
277 pmeGetComplexGridSizesInternal(pme, gridSize, paddedGridSize);
280 //! PME spline calculation and charge spreading
281 void pmePerformSplineAndSpread(gmx_pme_t *pme, CodePath mode, // TODO const qualifiers elsewhere
282 bool computeSplines, bool spreadCharges)
284 GMX_RELEASE_ASSERT(pme != nullptr, "PME data is not initialized");
285 PmeAtomComm *atc = &(pme->atc[0]);
286 const size_t gridIndex = 0;
287 const bool computeSplinesForZeroCharges = true;
288 real *fftgrid = spreadCharges ? pme->fftgrid[gridIndex] : nullptr;
289 real *pmegrid = pme->pmegrid[gridIndex].grid.grid;
294 spread_on_grid(pme, atc, &pme->pmegrid[gridIndex], computeSplines, spreadCharges,
295 fftgrid, computeSplinesForZeroCharges, gridIndex);
296 if (spreadCharges && !pme->bUseThreads)
298 wrap_periodic_pmegrid(pme, pmegrid);
299 copy_pmegrid_to_fftgrid(pme, pmegrid, fftgrid, gridIndex);
304 pme_gpu_spread(pme->gpu, gridIndex, fftgrid, computeSplines, spreadCharges);
308 GMX_THROW(InternalError("Test not implemented for this mode"));
312 //! Getting the internal spline data buffer pointer
313 static real *pmeGetSplineDataInternal(const gmx_pme_t *pme, PmeSplineDataType type, int dimIndex)
315 GMX_ASSERT((0 <= dimIndex) && (dimIndex < DIM), "Invalid dimension index");
316 const PmeAtomComm *atc = &(pme->atc[0]);
317 const size_t threadIndex = 0;
318 real *splineBuffer = nullptr;
321 case PmeSplineDataType::Values:
322 splineBuffer = atc->spline[threadIndex].theta.coefficients[dimIndex];
325 case PmeSplineDataType::Derivatives:
326 splineBuffer = atc->spline[threadIndex].dtheta.coefficients[dimIndex];
330 GMX_THROW(InternalError("Unknown spline data type"));
336 void pmePerformSolve(const gmx_pme_t *pme, CodePath mode,
337 PmeSolveAlgorithm method, real cellVolume,
338 GridOrdering gridOrdering, bool computeEnergyAndVirial)
340 t_complex *h_grid = pmeGetComplexGridInternal(pme);
341 const bool useLorentzBerthelot = false;
342 const size_t threadIndex = 0;
346 if (gridOrdering != GridOrdering::YZX)
348 GMX_THROW(InternalError("Test not implemented for this mode"));
352 case PmeSolveAlgorithm::Coulomb:
353 solve_pme_yzx(pme, h_grid, cellVolume,
354 computeEnergyAndVirial, pme->nthread, threadIndex);
357 case PmeSolveAlgorithm::LennardJones:
358 solve_pme_lj_yzx(pme, &h_grid, useLorentzBerthelot,
359 cellVolume, computeEnergyAndVirial, pme->nthread, threadIndex);
363 GMX_THROW(InternalError("Test not implemented for this mode"));
370 case PmeSolveAlgorithm::Coulomb:
371 pme_gpu_solve(pme->gpu, h_grid, gridOrdering, computeEnergyAndVirial);
375 GMX_THROW(InternalError("Test not implemented for this mode"));
380 GMX_THROW(InternalError("Test not implemented for this mode"));
384 //! PME force gathering
385 void pmePerformGather(gmx_pme_t *pme, CodePath mode,
386 PmeForceOutputHandling inputTreatment, ForcesVector &forces)
388 PmeAtomComm *atc = &(pme->atc[0]);
389 const index atomCount = atc->numAtoms();
390 GMX_RELEASE_ASSERT(forces.ssize() == atomCount, "Invalid force buffer size");
391 const bool forceReductionWithInput = (inputTreatment == PmeForceOutputHandling::ReduceWithInput);
392 const real scale = 1.0;
393 const size_t threadIndex = 0;
394 const size_t gridIndex = 0;
395 real *pmegrid = pme->pmegrid[gridIndex].grid.grid;
396 real *fftgrid = pme->fftgrid[gridIndex];
402 if (atc->nthread == 1)
404 // something which is normally done in serial spline computation (make_thread_local_ind())
405 atc->spline[threadIndex].n = atomCount;
407 copy_fftgrid_to_pmegrid(pme, fftgrid, pmegrid, gridIndex, pme->nthread, threadIndex);
408 unwrap_periodic_pmegrid(pme, pmegrid);
409 gather_f_bsplines(pme, pmegrid, !forceReductionWithInput, atc, &atc->spline[threadIndex], scale);
414 // Variable initialization needs a non-switch scope
415 PmeOutput output = pme_gpu_getOutput(*pme, GMX_PME_CALC_F);
416 GMX_ASSERT(forces.size() == output.forces_.size(), "Size of force buffers did not match");
417 if (forceReductionWithInput)
419 std::copy(std::begin(forces), std::end(forces), std::begin(output.forces_));
421 pme_gpu_gather(pme->gpu, inputTreatment, reinterpret_cast<float *>(fftgrid));
422 std::copy(std::begin(output.forces_), std::end(output.forces_), std::begin(forces));
427 GMX_THROW(InternalError("Test not implemented for this mode"));
431 //! PME test finalization before fetching the outputs
432 void pmeFinalizeTest(const gmx_pme_t *pme, CodePath mode)
440 pme_gpu_synchronize(pme->gpu);
444 GMX_THROW(InternalError("Test not implemented for this mode"));
448 //! Setting atom spline values/derivatives to be used in spread/gather
449 void pmeSetSplineData(const gmx_pme_t *pme, CodePath mode,
450 const SplineParamsDimVector &splineValues, PmeSplineDataType type, int dimIndex)
452 const PmeAtomComm *atc = &(pme->atc[0]);
453 const index atomCount = atc->numAtoms();
454 const index pmeOrder = pme->pme_order;
455 const index dimSize = pmeOrder * atomCount;
456 GMX_RELEASE_ASSERT(dimSize == splineValues.ssize(), "Mismatch in spline data");
457 real *splineBuffer = pmeGetSplineDataInternal(pme, type, dimIndex);
462 std::copy(splineValues.begin(), splineValues.end(), splineBuffer);
466 std::copy(splineValues.begin(), splineValues.end(), splineBuffer);
467 pme_gpu_transform_spline_atom_data(pme->gpu, atc, type, dimIndex, PmeLayoutTransform::HostToGpu);
471 GMX_THROW(InternalError("Test not implemented for this mode"));
475 //! Setting gridline indices to be used in spread/gather
476 void pmeSetGridLineIndices(gmx_pme_t *pme, CodePath mode,
477 const GridLineIndicesVector &gridLineIndices)
479 PmeAtomComm *atc = &(pme->atc[0]);
480 const index atomCount = atc->numAtoms();
481 GMX_RELEASE_ASSERT(atomCount == gridLineIndices.ssize(), "Mismatch in gridline indices size");
483 IVec paddedGridSizeUnused, gridSize(0, 0, 0);
484 pmeGetRealGridSizesInternal(pme, mode, gridSize, paddedGridSizeUnused);
486 for (const auto &index : gridLineIndices)
488 for (int i = 0; i < DIM; i++)
490 GMX_RELEASE_ASSERT((0 <= index[i]) && (index[i] < gridSize[i]), "Invalid gridline index");
497 memcpy(pme->gpu->staging.h_gridlineIndices, gridLineIndices.data(), atomCount * sizeof(gridLineIndices[0]));
501 atc->idx.resize(gridLineIndices.size());
502 std::copy(gridLineIndices.begin(), gridLineIndices.end(), atc->idx.begin());
505 GMX_THROW(InternalError("Test not implemented for this mode"));
509 //! Getting plain index into the complex 3d grid
510 inline size_t pmeGetGridPlainIndexInternal(const IVec &index, const IVec &paddedGridSize, GridOrdering gridOrdering)
513 switch (gridOrdering)
515 case GridOrdering::YZX:
516 result = (index[YY] * paddedGridSize[ZZ] + index[ZZ]) * paddedGridSize[XX] + index[XX];
519 case GridOrdering::XYZ:
520 result = (index[XX] * paddedGridSize[YY] + index[YY]) * paddedGridSize[ZZ] + index[ZZ];
524 GMX_THROW(InternalError("Test not implemented for this mode"));
529 //! Setting real or complex grid
530 template<typename ValueType>
531 static void pmeSetGridInternal(const gmx_pme_t *pme, CodePath mode,
532 GridOrdering gridOrdering,
533 const SparseGridValuesInput<ValueType> &gridValues)
535 IVec gridSize(0, 0, 0), paddedGridSize(0, 0, 0);
537 pmeGetGridAndSizesInternal<ValueType>(pme, mode, grid, gridSize, paddedGridSize);
541 case CodePath::GPU: // intentional absence of break, the grid will be copied from the host buffer in testing mode
543 std::memset(grid, 0, paddedGridSize[XX] * paddedGridSize[YY] * paddedGridSize[ZZ] * sizeof(ValueType));
544 for (const auto &gridValue : gridValues)
546 for (int i = 0; i < DIM; i++)
548 GMX_RELEASE_ASSERT((0 <= gridValue.first[i]) && (gridValue.first[i] < gridSize[i]), "Invalid grid value index");
550 const size_t gridValueIndex = pmeGetGridPlainIndexInternal(gridValue.first, paddedGridSize, gridOrdering);
551 grid[gridValueIndex] = gridValue.second;
556 GMX_THROW(InternalError("Test not implemented for this mode"));
560 //! Setting real grid to be used in gather
561 void pmeSetRealGrid(const gmx_pme_t *pme, CodePath mode,
562 const SparseRealGridValuesInput &gridValues)
564 pmeSetGridInternal<real>(pme, mode, GridOrdering::XYZ, gridValues);
567 //! Setting complex grid to be used in solve
568 void pmeSetComplexGrid(const gmx_pme_t *pme, CodePath mode,
569 GridOrdering gridOrdering,
570 const SparseComplexGridValuesInput &gridValues)
572 pmeSetGridInternal<t_complex>(pme, mode, gridOrdering, gridValues);
575 //! Getting the single dimension's spline values or derivatives
576 SplineParamsDimVector pmeGetSplineData(const gmx_pme_t *pme, CodePath mode,
577 PmeSplineDataType type, int dimIndex)
579 GMX_RELEASE_ASSERT(pme != nullptr, "PME data is not initialized");
580 const PmeAtomComm *atc = &(pme->atc[0]);
581 const size_t atomCount = atc->numAtoms();
582 const size_t pmeOrder = pme->pme_order;
583 const size_t dimSize = pmeOrder * atomCount;
585 real *sourceBuffer = pmeGetSplineDataInternal(pme, type, dimIndex);
586 SplineParamsDimVector result;
590 pme_gpu_transform_spline_atom_data(pme->gpu, atc, type, dimIndex, PmeLayoutTransform::GpuToHost);
591 result = arrayRefFromArray(sourceBuffer, dimSize);
595 result = arrayRefFromArray(sourceBuffer, dimSize);
599 GMX_THROW(InternalError("Test not implemented for this mode"));
604 //! Getting the gridline indices
605 GridLineIndicesVector pmeGetGridlineIndices(const gmx_pme_t *pme, CodePath mode)
607 GMX_RELEASE_ASSERT(pme != nullptr, "PME data is not initialized");
608 const PmeAtomComm *atc = &(pme->atc[0]);
609 const size_t atomCount = atc->numAtoms();
611 GridLineIndicesVector gridLineIndices;
615 gridLineIndices = arrayRefFromArray(reinterpret_cast<IVec *>(pme->gpu->staging.h_gridlineIndices), atomCount);
619 gridLineIndices = atc->idx;
623 GMX_THROW(InternalError("Test not implemented for this mode"));
625 return gridLineIndices;
628 //! Getting real or complex grid - only non zero values
629 template<typename ValueType>
630 static SparseGridValuesOutput<ValueType> pmeGetGridInternal(const gmx_pme_t *pme, CodePath mode, GridOrdering gridOrdering)
632 IVec gridSize(0, 0, 0), paddedGridSize(0, 0, 0);
634 pmeGetGridAndSizesInternal<ValueType>(pme, mode, grid, gridSize, paddedGridSize);
635 SparseGridValuesOutput<ValueType> gridValues;
638 case CodePath::GPU: // intentional absence of break
641 for (int ix = 0; ix < gridSize[XX]; ix++)
643 for (int iy = 0; iy < gridSize[YY]; iy++)
645 for (int iz = 0; iz < gridSize[ZZ]; iz++)
647 IVec temp(ix, iy, iz);
648 const size_t gridValueIndex = pmeGetGridPlainIndexInternal(temp, paddedGridSize, gridOrdering);
649 const ValueType value = grid[gridValueIndex];
650 if (value != ValueType {})
652 auto key = formatString("Cell %d %d %d", ix, iy, iz);
653 gridValues[key] = value;
661 GMX_THROW(InternalError("Test not implemented for this mode"));
666 //! Getting the real grid (spreading output of pmePerformSplineAndSpread())
667 SparseRealGridValuesOutput pmeGetRealGrid(const gmx_pme_t *pme, CodePath mode)
669 return pmeGetGridInternal<real>(pme, mode, GridOrdering::XYZ);
672 //! Getting the complex grid output of pmePerformSolve()
673 SparseComplexGridValuesOutput pmeGetComplexGrid(const gmx_pme_t *pme, CodePath mode,
674 GridOrdering gridOrdering)
676 return pmeGetGridInternal<t_complex>(pme, mode, gridOrdering);
679 //! Getting the reciprocal energy and virial
680 PmeOutput pmeGetReciprocalEnergyAndVirial(const gmx_pme_t *pme, CodePath mode,
681 PmeSolveAlgorithm method)
689 case PmeSolveAlgorithm::Coulomb:
690 get_pme_ener_vir_q(pme->solve_work, pme->nthread, &output);
693 case PmeSolveAlgorithm::LennardJones:
694 get_pme_ener_vir_lj(pme->solve_work, pme->nthread, &output);
698 GMX_THROW(InternalError("Test not implemented for this mode"));
704 case PmeSolveAlgorithm::Coulomb:
705 pme_gpu_getEnergyAndVirial(*pme, &output);
709 GMX_THROW(InternalError("Test not implemented for this mode"));
714 GMX_THROW(InternalError("Test not implemented for this mode"));