src/gromacs/ewald/tests/pmetestcommon.cpp

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2016,2017,2018,2019,2020, 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.
   8  *
   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.
  13  *
  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.
  18  *
  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.
  23  *
  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.
  31  *
  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.
  34  */
  35 /*! \internal \file
  36  * \brief
  37  * Implements common routines for PME tests.
  38  *
  39  * \author Aleksei Iupinov <a.yupinov@gmail.com>
  40  * \ingroup module_ewald
  41  */
  42 #include "gmxpre.h"
  43
  44 #include "pmetestcommon.h"
  45
  46 #include <cstring>
  47
  48 #include <algorithm>
  49
  50 #include "gromacs/domdec/domdec.h"
  51 #include "gromacs/ewald/pme_gather.h"
  52 #include "gromacs/ewald/pme_gpu_calculate_splines.h"
  53 #include "gromacs/ewald/pme_gpu_constants.h"
  54 #include "gromacs/ewald/pme_gpu_internal.h"
  55 #include "gromacs/ewald/pme_gpu_staging.h"
  56 #include "gromacs/ewald/pme_grid.h"
  57 #include "gromacs/ewald/pme_internal.h"
  58 #include "gromacs/ewald/pme_redistribute.h"
  59 #include "gromacs/ewald/pme_solve.h"
  60 #include "gromacs/ewald/pme_spread.h"
  61 #include "gromacs/fft/parallel_3dfft.h"
  62 #include "gromacs/gpu_utils/gpu_utils.h"
  63 #include "gromacs/math/invertmatrix.h"
  64 #include "gromacs/mdtypes/commrec.h"
  65 #include "gromacs/pbcutil/pbc.h"
  66 #include "gromacs/topology/topology.h"
  67 #include "gromacs/utility/exceptions.h"
  68 #include "gromacs/utility/gmxassert.h"
  69 #include "gromacs/utility/logger.h"
  70 #include "gromacs/utility/stringutil.h"
  71
  72 #include "testutils/testasserts.h"
  73
  74 #include "testhardwarecontexts.h"
  75
  76 namespace gmx
  77 {
  78 namespace test
  79 {
  80
  81 bool pmeSupportsInputForMode(const gmx_hw_info_t& hwinfo, const t_inputrec* inputRec, CodePath mode)
  82 {
  83     bool       implemented;
  84     gmx_mtop_t mtop;
  85     switch (mode)
  86     {
  87         case CodePath::CPU: implemented = true; break;
  88
  89         case CodePath::GPU:
  90             implemented = (pme_gpu_supports_build(nullptr) && pme_gpu_supports_hardware(hwinfo, nullptr)
  91                            && pme_gpu_supports_input(*inputRec, mtop, nullptr));
  92             break;
  93
  94         default: GMX_THROW(InternalError("Test not implemented for this mode"));
  95     }
  96     return implemented;
  97 }
  98
  99 uint64_t getSplineModuliDoublePrecisionUlps(int splineOrder)
 100 {
 101     /* Arbitrary ulp tolerance for sine/cosine implementation. It's
 102      * hard to know what to pick without testing lots of
 103      * implementations. */
 104     const uint64_t sineUlps = 10;
 105     return 4 * (splineOrder - 2) + 2 * sineUlps * splineOrder;
 106 }
 107
 108 //! PME initialization
 109 PmeSafePointer pmeInitWrapper(const t_inputrec*        inputRec,
 110                               const CodePath           mode,
 111                               const DeviceInformation* deviceInfo,
 112                               const PmeGpuProgram*     pmeGpuProgram,
 113                               const Matrix3x3&         box,
 114                               const real               ewaldCoeff_q,
 115                               const real               ewaldCoeff_lj)
 116 {
 117     const MDLogger dummyLogger;
 118     const auto     runMode       = (mode == CodePath::CPU) ? PmeRunMode::CPU : PmeRunMode::Mixed;
 119     t_commrec      dummyCommrec  = { 0 };
 120     NumPmeDomains  numPmeDomains = { 1, 1 };
 121     gmx_pme_t*     pmeDataRaw =
 122             gmx_pme_init(&dummyCommrec, numPmeDomains, inputRec, false, false, true, ewaldCoeff_q,
 123                          ewaldCoeff_lj, 1, runMode, nullptr, deviceInfo, pmeGpuProgram, dummyLogger);
 124     PmeSafePointer pme(pmeDataRaw); // taking ownership
 125
 126     // TODO get rid of this with proper matrix type
 127     matrix boxTemp;
 128     for (int i = 0; i < DIM; i++)
 129     {
 130         for (int j = 0; j < DIM; j++)
 131         {
 132             boxTemp[i][j] = box[i * DIM + j];
 133         }
 134     }
 135     const char* boxError = check_box(PbcType::Unset, boxTemp);
 136     GMX_RELEASE_ASSERT(boxError == nullptr, boxError);
 137
 138     switch (mode)
 139     {
 140         case CodePath::CPU: invertBoxMatrix(boxTemp, pme->recipbox); break;
 141
 142         case CodePath::GPU:
 143             pme_gpu_set_testing(pme->gpu, true);
 144             pme_gpu_update_input_box(pme->gpu, boxTemp);
 145             break;
 146
 147         default: GMX_THROW(InternalError("Test not implemented for this mode"));
 148     }
 149
 150     return pme;
 151 }
 152
 153 //! Simple PME initialization based on input, no atom data
 154 PmeSafePointer pmeInitEmpty(const t_inputrec*        inputRec,
 155                             const CodePath           mode,
 156                             const DeviceInformation* deviceInfo,
 157                             const PmeGpuProgram*     pmeGpuProgram,
 158                             const Matrix3x3&         box,
 159                             const real               ewaldCoeff_q,
 160                             const real               ewaldCoeff_lj)
 161 {
 162     return pmeInitWrapper(inputRec, mode, deviceInfo, pmeGpuProgram, box, ewaldCoeff_q, ewaldCoeff_lj);
 163     // hiding the fact that PME actually needs to know the number of atoms in advance
 164 }
 165
 166 PmeSafePointer pmeInitEmpty(const t_inputrec* inputRec)
 167 {
 168     const Matrix3x3 defaultBox = { { 1.0F, 0.0F, 0.0F, 0.0F, 1.0F, 0.0F, 0.0F, 0.0F, 1.0F } };
 169     return pmeInitWrapper(inputRec, CodePath::CPU, nullptr, nullptr, defaultBox, 0.0F, 0.0F);
 170 }
 171
 172 //! Make a GPU state-propagator manager
 173 std::unique_ptr<StatePropagatorDataGpu> makeStatePropagatorDataGpu(const gmx_pme_t&     pme,
 174                                                                    const DeviceContext& deviceContext)
 175 {
 176     // TODO: Pin the host buffer and use async memory copies
 177     // TODO: Special constructor for PME-only rank / PME-tests is used here. There should be a mechanism to
 178     //       restrict one from using other constructor here.
 179     return std::make_unique<StatePropagatorDataGpu>(pme_gpu_get_device_stream(&pme), deviceContext,
 180                                                     GpuApiCallBehavior::Sync,
 181                                                     pme_gpu_get_block_size(&pme), nullptr);
 182 }
 183
 184 //! PME initialization with atom data
 185 void pmeInitAtoms(gmx_pme_t*               pme,
 186                   StatePropagatorDataGpu*  stateGpu,
 187                   const CodePath           mode,
 188                   const CoordinatesVector& coordinates,
 189                   const ChargesVector&     charges)
 190 {
 191     const index atomCount = coordinates.size();
 192     GMX_RELEASE_ASSERT(atomCount == charges.ssize(), "Mismatch in atom data");
 193     PmeAtomComm* atc = nullptr;
 194
 195     switch (mode)
 196     {
 197         case CodePath::CPU:
 198             atc              = &(pme->atc[0]);
 199             atc->x           = coordinates;
 200             atc->coefficient = charges;
 201             gmx_pme_reinit_atoms(pme, atomCount, charges.data());
 202             /* With decomposition there would be more boilerplate atc code here, e.g. do_redist_pos_coeffs */
 203             break;
 204
 205         case CodePath::GPU:
 206             // TODO: Avoid use of atc in the GPU code path
 207             atc = &(pme->atc[0]);
 208             // We need to set atc->n for passing the size in the tests
 209             atc->setNumAtoms(atomCount);
 210             gmx_pme_reinit_atoms(pme, atomCount, charges.data());
 211
 212             stateGpu->reinit(atomCount, atomCount);
 213             stateGpu->copyCoordinatesToGpu(arrayRefFromArray(coordinates.data(), coordinates.size()),
 214                                            gmx::AtomLocality::All);
 215             pme_gpu_set_kernelparam_coordinates(pme->gpu, stateGpu->getCoordinates());
 216
 217             break;
 218
 219         default: GMX_THROW(InternalError("Test not implemented for this mode"));
 220     }
 221 }
 222
 223 //! Getting local PME real grid pointer for test I/O
 224 static real* pmeGetRealGridInternal(const gmx_pme_t* pme)
 225 {
 226     const size_t gridIndex = 0;
 227     return pme->fftgrid[gridIndex];
 228 }
 229
 230 //! Getting local PME real grid dimensions
 231 static void pmeGetRealGridSizesInternal(const gmx_pme_t* pme,
 232                                         CodePath         mode,
 233                                         IVec& gridSize,       //NOLINT(google-runtime-references)
 234                                         IVec& paddedGridSize) //NOLINT(google-runtime-references)
 235 {
 236     const size_t gridIndex = 0;
 237     IVec         gridOffsetUnused;
 238     switch (mode)
 239     {
 240         case CodePath::CPU:
 241             gmx_parallel_3dfft_real_limits(pme->pfft_setup[gridIndex], gridSize, gridOffsetUnused,
 242                                            paddedGridSize);
 243             break;
 244
 245         case CodePath::GPU:
 246             pme_gpu_get_real_grid_sizes(pme->gpu, &gridSize, &paddedGridSize);
 247             break;
 248
 249         default: GMX_THROW(InternalError("Test not implemented for this mode"));
 250     }
 251 }
 252
 253 //! Getting local PME complex grid pointer for test I/O
 254 static t_complex* pmeGetComplexGridInternal(const gmx_pme_t* pme)
 255 {
 256     const size_t gridIndex = 0;
 257     return pme->cfftgrid[gridIndex];
 258 }
 259
 260 //! Getting local PME complex grid dimensions
 261 static void pmeGetComplexGridSizesInternal(const gmx_pme_t* pme,
 262                                            IVec& gridSize,       //NOLINT(google-runtime-references)
 263                                            IVec& paddedGridSize) //NOLINT(google-runtime-references)
 264 {
 265     const size_t gridIndex = 0;
 266     IVec         gridOffsetUnused, complexOrderUnused;
 267     gmx_parallel_3dfft_complex_limits(pme->pfft_setup[gridIndex], complexOrderUnused, gridSize,
 268                                       gridOffsetUnused, paddedGridSize); // TODO: what about YZX ordering?
 269 }
 270
 271 //! Getting the PME grid memory buffer and its sizes - template definition
 272 template<typename ValueType>
 273 static void pmeGetGridAndSizesInternal(const gmx_pme_t* /*unused*/,
 274                                        CodePath /*unused*/,
 275                                        ValueType*& /*unused*/, //NOLINT(google-runtime-references)
 276                                        IVec& /*unused*/,       //NOLINT(google-runtime-references)
 277                                        IVec& /*unused*/)       //NOLINT(google-runtime-references)
 278 {
 279     GMX_THROW(InternalError("Deleted function call"));
 280     // explicitly deleting general template does not compile in clang/icc, see https://llvm.org/bugs/show_bug.cgi?id=17537
 281 }
 282
 283 //! Getting the PME real grid memory buffer and its sizes
 284 template<>
 285 void pmeGetGridAndSizesInternal<real>(const gmx_pme_t* pme, CodePath mode, real*& grid, IVec& gridSize, IVec& paddedGridSize)
 286 {
 287     grid = pmeGetRealGridInternal(pme);
 288     pmeGetRealGridSizesInternal(pme, mode, gridSize, paddedGridSize);
 289 }
 290
 291 //! Getting the PME complex grid memory buffer and its sizes
 292 template<>
 293 void pmeGetGridAndSizesInternal<t_complex>(const gmx_pme_t* pme,
 294                                            CodePath /*unused*/,
 295                                            t_complex*& grid,
 296                                            IVec&       gridSize,
 297                                            IVec&       paddedGridSize)
 298 {
 299     grid = pmeGetComplexGridInternal(pme);
 300     pmeGetComplexGridSizesInternal(pme, gridSize, paddedGridSize);
 301 }
 302
 303 //! PME spline calculation and charge spreading
 304 void pmePerformSplineAndSpread(gmx_pme_t* pme,
 305                                CodePath   mode, // TODO const qualifiers elsewhere
 306                                bool       computeSplines,
 307                                bool       spreadCharges)
 308 {
 309     GMX_RELEASE_ASSERT(pme != nullptr, "PME data is not initialized");
 310     PmeAtomComm* atc                          = &(pme->atc[0]);
 311     const size_t gridIndex                    = 0;
 312     const bool   computeSplinesForZeroCharges = true;
 313     real*        fftgrid                      = spreadCharges ? pme->fftgrid[gridIndex] : nullptr;
 314     real*        pmegrid                      = pme->pmegrid[gridIndex].grid.grid;
 315
 316     switch (mode)
 317     {
 318         case CodePath::CPU:
 319             spread_on_grid(pme, atc, &pme->pmegrid[gridIndex], computeSplines, spreadCharges,
 320                            fftgrid, computeSplinesForZeroCharges, gridIndex);
 321             if (spreadCharges && !pme->bUseThreads)
 322             {
 323                 wrap_periodic_pmegrid(pme, pmegrid);
 324                 copy_pmegrid_to_fftgrid(pme, pmegrid, fftgrid, gridIndex);
 325             }
 326             break;
 327
 328         case CodePath::GPU:
 329         {
 330             // no synchronization needed as x is transferred in the PME stream
 331             GpuEventSynchronizer* xReadyOnDevice = nullptr;
 332             pme_gpu_spread(pme->gpu, xReadyOnDevice, gridIndex, fftgrid, computeSplines, spreadCharges);
 333         }
 334         break;
 335
 336         default: GMX_THROW(InternalError("Test not implemented for this mode"));
 337     }
 338 }
 339
 340 //! Getting the internal spline data buffer pointer
 341 static real* pmeGetSplineDataInternal(const gmx_pme_t* pme, PmeSplineDataType type, int dimIndex)
 342 {
 343     GMX_ASSERT((0 <= dimIndex) && (dimIndex < DIM), "Invalid dimension index");
 344     const PmeAtomComm* atc          = &(pme->atc[0]);
 345     const size_t       threadIndex  = 0;
 346     real*              splineBuffer = nullptr;
 347     switch (type)
 348     {
 349         case PmeSplineDataType::Values:
 350             splineBuffer = atc->spline[threadIndex].theta.coefficients[dimIndex];
 351             break;
 352
 353         case PmeSplineDataType::Derivatives:
 354             splineBuffer = atc->spline[threadIndex].dtheta.coefficients[dimIndex];
 355             break;
 356
 357         default: GMX_THROW(InternalError("Unknown spline data type"));
 358     }
 359     return splineBuffer;
 360 }
 361
 362 //! PME solving
 363 void pmePerformSolve(const gmx_pme_t*  pme,
 364                      CodePath          mode,
 365                      PmeSolveAlgorithm method,
 366                      real              cellVolume,
 367                      GridOrdering      gridOrdering,
 368                      bool              computeEnergyAndVirial)
 369 {
 370     t_complex*   h_grid              = pmeGetComplexGridInternal(pme);
 371     const bool   useLorentzBerthelot = false;
 372     const size_t threadIndex         = 0;
 373     switch (mode)
 374     {
 375         case CodePath::CPU:
 376             if (gridOrdering != GridOrdering::YZX)
 377             {
 378                 GMX_THROW(InternalError("Test not implemented for this mode"));
 379             }
 380             switch (method)
 381             {
 382                 case PmeSolveAlgorithm::Coulomb:
 383                     solve_pme_yzx(pme, h_grid, cellVolume, computeEnergyAndVirial, pme->nthread, threadIndex);
 384                     break;
 385
 386                 case PmeSolveAlgorithm::LennardJones:
 387                     solve_pme_lj_yzx(pme, &h_grid, useLorentzBerthelot, cellVolume,
 388                                      computeEnergyAndVirial, pme->nthread, threadIndex);
 389                     break;
 390
 391                 default: GMX_THROW(InternalError("Test not implemented for this mode"));
 392             }
 393             break;
 394
 395         case CodePath::GPU:
 396             switch (method)
 397             {
 398                 case PmeSolveAlgorithm::Coulomb:
 399                     pme_gpu_solve(pme->gpu, h_grid, gridOrdering, computeEnergyAndVirial);
 400                     break;
 401
 402                 default: GMX_THROW(InternalError("Test not implemented for this mode"));
 403             }
 404             break;
 405
 406         default: GMX_THROW(InternalError("Test not implemented for this mode"));
 407     }
 408 }
 409
 410 //! PME force gathering
 411 void pmePerformGather(gmx_pme_t* pme, CodePath mode, ForcesVector& forces)
 412 {
 413     PmeAtomComm* atc       = &(pme->atc[0]);
 414     const index  atomCount = atc->numAtoms();
 415     GMX_RELEASE_ASSERT(forces.ssize() == atomCount, "Invalid force buffer size");
 416     const real   scale       = 1.0;
 417     const size_t threadIndex = 0;
 418     const size_t gridIndex   = 0;
 419     real*        pmegrid     = pme->pmegrid[gridIndex].grid.grid;
 420     real*        fftgrid     = pme->fftgrid[gridIndex];
 421
 422     switch (mode)
 423     {
 424         case CodePath::CPU:
 425             atc->f = forces;
 426             if (atc->nthread == 1)
 427             {
 428                 // something which is normally done in serial spline computation (make_thread_local_ind())
 429                 atc->spline[threadIndex].n = atomCount;
 430             }
 431             copy_fftgrid_to_pmegrid(pme, fftgrid, pmegrid, gridIndex, pme->nthread, threadIndex);
 432             unwrap_periodic_pmegrid(pme, pmegrid);
 433             gather_f_bsplines(pme, pmegrid, true, atc, &atc->spline[threadIndex], scale);
 434             break;
 435
 436         case CodePath::GPU:
 437         {
 438             // Variable initialization needs a non-switch scope
 439             const bool computeEnergyAndVirial = false;
 440             PmeOutput  output                 = pme_gpu_getOutput(*pme, computeEnergyAndVirial);
 441             GMX_ASSERT(forces.size() == output.forces_.size(),
 442                        "Size of force buffers did not match");
 443             pme_gpu_gather(pme->gpu, reinterpret_cast<float*>(fftgrid));
 444             std::copy(std::begin(output.forces_), std::end(output.forces_), std::begin(forces));
 445         }
 446         break;
 447
 448         default: GMX_THROW(InternalError("Test not implemented for this mode"));
 449     }
 450 }
 451
 452 //! PME test finalization before fetching the outputs
 453 void pmeFinalizeTest(const gmx_pme_t* pme, CodePath mode)
 454 {
 455     switch (mode)
 456     {
 457         case CodePath::CPU: break;
 458
 459         case CodePath::GPU: pme_gpu_synchronize(pme->gpu); break;
 460
 461         default: GMX_THROW(InternalError("Test not implemented for this mode"));
 462     }
 463 }
 464
 465 //! A binary enum for spline data layout transformation
 466 enum class PmeLayoutTransform
 467 {
 468     GpuToHost,
 469     HostToGpu
 470 };
 471
 472 /*! \brief Gets a unique index to an element in a spline parameter buffer.
 473  *
 474  * These theta/dtheta buffers are laid out for GPU spread/gather
 475  * kernels. The index is wrt the execution block, in range(0,
 476  * atomsPerBlock * order * DIM).
 477  *
 478  * This is a wrapper, only used in unit tests.
 479  * \param[in] order            PME order
 480  * \param[in] splineIndex      Spline contribution index (from 0 to \p order - 1)
 481  * \param[in] dimIndex         Dimension index (from 0 to 2)
 482  * \param[in] atomIndex        Atom index wrt the block.
 483  * \param[in] atomsPerWarp     Number of atoms processed by a warp.
 484  *
 485  * \returns Index into theta or dtheta array using GPU layout.
 486  */
 487 static int getSplineParamFullIndex(int order, int splineIndex, int dimIndex, int atomIndex, int atomsPerWarp)
 488 {
 489     if (order != c_pmeGpuOrder)
 490     {
 491         throw order;
 492     }
 493     constexpr int fixedOrder = c_pmeGpuOrder;
 494     GMX_UNUSED_VALUE(fixedOrder);
 495
 496     const int atomWarpIndex = atomIndex % atomsPerWarp;
 497     const int warpIndex     = atomIndex / atomsPerWarp;
 498     int       indexBase, result;
 499     switch (atomsPerWarp)
 500     {
 501         case 1:
 502             indexBase = getSplineParamIndexBase<fixedOrder, 1>(warpIndex, atomWarpIndex);
 503             result    = getSplineParamIndex<fixedOrder, 1>(indexBase, dimIndex, splineIndex);
 504             break;
 505
 506         case 2:
 507             indexBase = getSplineParamIndexBase<fixedOrder, 2>(warpIndex, atomWarpIndex);
 508             result    = getSplineParamIndex<fixedOrder, 2>(indexBase, dimIndex, splineIndex);
 509             break;
 510
 511         case 4:
 512             indexBase = getSplineParamIndexBase<fixedOrder, 4>(warpIndex, atomWarpIndex);
 513             result    = getSplineParamIndex<fixedOrder, 4>(indexBase, dimIndex, splineIndex);
 514             break;
 515
 516         case 8:
 517             indexBase = getSplineParamIndexBase<fixedOrder, 8>(warpIndex, atomWarpIndex);
 518             result    = getSplineParamIndex<fixedOrder, 8>(indexBase, dimIndex, splineIndex);
 519             break;
 520
 521         default:
 522             GMX_THROW(NotImplementedError(
 523                     formatString("Test function call not unrolled for atomsPerWarp = %d in "
 524                                  "getSplineParamFullIndex",
 525                                  atomsPerWarp)));
 526     }
 527     return result;
 528 }
 529
 530 /*!\brief Return the number of atoms per warp */
 531 static int pme_gpu_get_atoms_per_warp(const PmeGpu* pmeGpu)
 532 {
 533     const int order = pmeGpu->common->pme_order;
 534     const int threadsPerAtom =
 535             (pmeGpu->settings.threadsPerAtom == ThreadsPerAtom::Order ? order : order * order);
 536     return pmeGpu->programHandle_->warpSize() / threadsPerAtom;
 537 }
 538
 539 /*! \brief Rearranges the atom spline data between the GPU and host layouts.
 540  * Only used for test purposes so far, likely to be horribly slow.
 541  *
 542  * \param[in]  pmeGpu     The PME GPU structure.
 543  * \param[out] atc        The PME CPU atom data structure (with a single-threaded layout).
 544  * \param[in]  type       The spline data type (values or derivatives).
 545  * \param[in]  dimIndex   Dimension index.
 546  * \param[in]  transform  Layout transform type
 547  */
 548 static void pme_gpu_transform_spline_atom_data(const PmeGpu*      pmeGpu,
 549                                                const PmeAtomComm* atc,
 550                                                PmeSplineDataType  type,
 551                                                int                dimIndex,
 552                                                PmeLayoutTransform transform)
 553 {
 554     // The GPU atom spline data is laid out in a different way currently than the CPU one.
 555     // This function converts the data from GPU to CPU layout (in the host memory).
 556     // It is only intended for testing purposes so far.
 557     // Ideally we should use similar layouts on CPU and GPU if we care about mixed modes and their
 558     // performance (e.g. spreading on GPU, gathering on CPU).
 559     GMX_RELEASE_ASSERT(atc->nthread == 1, "Only the serial PME data layout is supported");
 560     const uintmax_t threadIndex  = 0;
 561     const auto      atomCount    = atc->numAtoms();
 562     const auto      atomsPerWarp = pme_gpu_get_atoms_per_warp(pmeGpu);
 563     const auto      pmeOrder     = pmeGpu->common->pme_order;
 564     GMX_ASSERT(pmeOrder == c_pmeGpuOrder, "Only PME order 4 is implemented");
 565
 566     real*  cpuSplineBuffer;
 567     float* h_splineBuffer;
 568     switch (type)
 569     {
 570         case PmeSplineDataType::Values:
 571             cpuSplineBuffer = atc->spline[threadIndex].theta.coefficients[dimIndex];
 572             h_splineBuffer  = pmeGpu->staging.h_theta;
 573             break;
 574
 575         case PmeSplineDataType::Derivatives:
 576             cpuSplineBuffer = atc->spline[threadIndex].dtheta.coefficients[dimIndex];
 577             h_splineBuffer  = pmeGpu->staging.h_dtheta;
 578             break;
 579
 580         default: GMX_THROW(InternalError("Unknown spline data type"));
 581     }
 582
 583     for (auto atomIndex = 0; atomIndex < atomCount; atomIndex++)
 584     {
 585         for (auto orderIndex = 0; orderIndex < pmeOrder; orderIndex++)
 586         {
 587             const auto gpuValueIndex =
 588                     getSplineParamFullIndex(pmeOrder, orderIndex, dimIndex, atomIndex, atomsPerWarp);
 589             const auto cpuValueIndex = atomIndex * pmeOrder + orderIndex;
 590             GMX_ASSERT(cpuValueIndex < atomCount * pmeOrder,
 591                        "Atom spline data index out of bounds (while transforming GPU data layout "
 592                        "for host)");
 593             switch (transform)
 594             {
 595                 case PmeLayoutTransform::GpuToHost:
 596                     cpuSplineBuffer[cpuValueIndex] = h_splineBuffer[gpuValueIndex];
 597                     break;
 598
 599                 case PmeLayoutTransform::HostToGpu:
 600                     h_splineBuffer[gpuValueIndex] = cpuSplineBuffer[cpuValueIndex];
 601                     break;
 602
 603                 default: GMX_THROW(InternalError("Unknown layout transform"));
 604             }
 605         }
 606     }
 607 }
 608
 609 //! Setting atom spline values/derivatives to be used in spread/gather
 610 void pmeSetSplineData(const gmx_pme_t*             pme,
 611                       CodePath                     mode,
 612                       const SplineParamsDimVector& splineValues,
 613                       PmeSplineDataType            type,
 614                       int                          dimIndex)
 615 {
 616     const PmeAtomComm* atc       = &(pme->atc[0]);
 617     const index        atomCount = atc->numAtoms();
 618     const index        pmeOrder  = pme->pme_order;
 619     const index        dimSize   = pmeOrder * atomCount;
 620     GMX_RELEASE_ASSERT(dimSize == splineValues.ssize(), "Mismatch in spline data");
 621     real* splineBuffer = pmeGetSplineDataInternal(pme, type, dimIndex);
 622
 623     switch (mode)
 624     {
 625         case CodePath::CPU:
 626             std::copy(splineValues.begin(), splineValues.end(), splineBuffer);
 627             break;
 628
 629         case CodePath::GPU:
 630             std::copy(splineValues.begin(), splineValues.end(), splineBuffer);
 631             pme_gpu_transform_spline_atom_data(pme->gpu, atc, type, dimIndex, PmeLayoutTransform::HostToGpu);
 632             break;
 633
 634         default: GMX_THROW(InternalError("Test not implemented for this mode"));
 635     }
 636 }
 637
 638 //! Setting gridline indices to be used in spread/gather
 639 void pmeSetGridLineIndices(gmx_pme_t* pme, CodePath mode, const GridLineIndicesVector& gridLineIndices)
 640 {
 641     PmeAtomComm* atc       = &(pme->atc[0]);
 642     const index  atomCount = atc->numAtoms();
 643     GMX_RELEASE_ASSERT(atomCount == gridLineIndices.ssize(), "Mismatch in gridline indices size");
 644
 645     IVec paddedGridSizeUnused, gridSize(0, 0, 0);
 646     pmeGetRealGridSizesInternal(pme, mode, gridSize, paddedGridSizeUnused);
 647
 648     for (const auto& index : gridLineIndices)
 649     {
 650         for (int i = 0; i < DIM; i++)
 651         {
 652             GMX_RELEASE_ASSERT((0 <= index[i]) && (index[i] < gridSize[i]),
 653                                "Invalid gridline index");
 654         }
 655     }
 656
 657     switch (mode)
 658     {
 659         case CodePath::GPU:
 660             memcpy(pme_gpu_staging(pme->gpu).h_gridlineIndices, gridLineIndices.data(),
 661                    atomCount * sizeof(gridLineIndices[0]));
 662             break;
 663
 664         case CodePath::CPU:
 665             atc->idx.resize(gridLineIndices.size());
 666             std::copy(gridLineIndices.begin(), gridLineIndices.end(), atc->idx.begin());
 667             break;
 668         default: GMX_THROW(InternalError("Test not implemented for this mode"));
 669     }
 670 }
 671
 672 //! Getting plain index into the complex 3d grid
 673 inline size_t pmeGetGridPlainIndexInternal(const IVec& index, const IVec& paddedGridSize, GridOrdering gridOrdering)
 674 {
 675     size_t result;
 676     switch (gridOrdering)
 677     {
 678         case GridOrdering::YZX:
 679             result = (index[YY] * paddedGridSize[ZZ] + index[ZZ]) * paddedGridSize[XX] + index[XX];
 680             break;
 681
 682         case GridOrdering::XYZ:
 683             result = (index[XX] * paddedGridSize[YY] + index[YY]) * paddedGridSize[ZZ] + index[ZZ];
 684             break;
 685
 686         default: GMX_THROW(InternalError("Test not implemented for this mode"));
 687     }
 688     return result;
 689 }
 690
 691 //! Setting real or complex grid
 692 template<typename ValueType>
 693 static void pmeSetGridInternal(const gmx_pme_t*                        pme,
 694                                CodePath                                mode,
 695                                GridOrdering                            gridOrdering,
 696                                const SparseGridValuesInput<ValueType>& gridValues)
 697 {
 698     IVec       gridSize(0, 0, 0), paddedGridSize(0, 0, 0);
 699     ValueType* grid;
 700     pmeGetGridAndSizesInternal<ValueType>(pme, mode, grid, gridSize, paddedGridSize);
 701
 702     switch (mode)
 703     {
 704         case CodePath::GPU: // intentional absence of break, the grid will be copied from the host buffer in testing mode
 705         case CodePath::CPU:
 706             std::memset(grid, 0,
 707                         paddedGridSize[XX] * paddedGridSize[YY] * paddedGridSize[ZZ] * sizeof(ValueType));
 708             for (const auto& gridValue : gridValues)
 709             {
 710                 for (int i = 0; i < DIM; i++)
 711                 {
 712                     GMX_RELEASE_ASSERT((0 <= gridValue.first[i]) && (gridValue.first[i] < gridSize[i]),
 713                                        "Invalid grid value index");
 714                 }
 715                 const size_t gridValueIndex =
 716                         pmeGetGridPlainIndexInternal(gridValue.first, paddedGridSize, gridOrdering);
 717                 grid[gridValueIndex] = gridValue.second;
 718             }
 719             break;
 720
 721         default: GMX_THROW(InternalError("Test not implemented for this mode"));
 722     }
 723 }
 724
 725 //! Setting real grid to be used in gather
 726 void pmeSetRealGrid(const gmx_pme_t* pme, CodePath mode, const SparseRealGridValuesInput& gridValues)
 727 {
 728     pmeSetGridInternal<real>(pme, mode, GridOrdering::XYZ, gridValues);
 729 }
 730
 731 //! Setting complex grid to be used in solve
 732 void pmeSetComplexGrid(const gmx_pme_t*                    pme,
 733                        CodePath                            mode,
 734                        GridOrdering                        gridOrdering,
 735                        const SparseComplexGridValuesInput& gridValues)
 736 {
 737     pmeSetGridInternal<t_complex>(pme, mode, gridOrdering, gridValues);
 738 }
 739
 740 //! Getting the single dimension's spline values or derivatives
 741 SplineParamsDimVector pmeGetSplineData(const gmx_pme_t* pme, CodePath mode, PmeSplineDataType type, int dimIndex)
 742 {
 743     GMX_RELEASE_ASSERT(pme != nullptr, "PME data is not initialized");
 744     const PmeAtomComm* atc       = &(pme->atc[0]);
 745     const size_t       atomCount = atc->numAtoms();
 746     const size_t       pmeOrder  = pme->pme_order;
 747     const size_t       dimSize   = pmeOrder * atomCount;
 748
 749     real*                 sourceBuffer = pmeGetSplineDataInternal(pme, type, dimIndex);
 750     SplineParamsDimVector result;
 751     switch (mode)
 752     {
 753         case CodePath::GPU:
 754             pme_gpu_transform_spline_atom_data(pme->gpu, atc, type, dimIndex, PmeLayoutTransform::GpuToHost);
 755             result = arrayRefFromArray(sourceBuffer, dimSize);
 756             break;
 757
 758         case CodePath::CPU: result = arrayRefFromArray(sourceBuffer, dimSize); break;
 759
 760         default: GMX_THROW(InternalError("Test not implemented for this mode"));
 761     }
 762     return result;
 763 }
 764
 765 //! Getting the gridline indices
 766 GridLineIndicesVector pmeGetGridlineIndices(const gmx_pme_t* pme, CodePath mode)
 767 {
 768     GMX_RELEASE_ASSERT(pme != nullptr, "PME data is not initialized");
 769     const PmeAtomComm* atc       = &(pme->atc[0]);
 770     const size_t       atomCount = atc->numAtoms();
 771
 772     GridLineIndicesVector gridLineIndices;
 773     switch (mode)
 774     {
 775         case CodePath::GPU:
 776             gridLineIndices = arrayRefFromArray(
 777                     reinterpret_cast<IVec*>(pme_gpu_staging(pme->gpu).h_gridlineIndices), atomCount);
 778             break;
 779
 780         case CodePath::CPU: gridLineIndices = atc->idx; break;
 781
 782         default: GMX_THROW(InternalError("Test not implemented for this mode"));
 783     }
 784     return gridLineIndices;
 785 }
 786
 787 //! Getting real or complex grid - only non zero values
 788 template<typename ValueType>
 789 static SparseGridValuesOutput<ValueType> pmeGetGridInternal(const gmx_pme_t* pme,
 790                                                             CodePath         mode,
 791                                                             GridOrdering     gridOrdering)
 792 {
 793     IVec       gridSize(0, 0, 0), paddedGridSize(0, 0, 0);
 794     ValueType* grid;
 795     pmeGetGridAndSizesInternal<ValueType>(pme, mode, grid, gridSize, paddedGridSize);
 796     SparseGridValuesOutput<ValueType> gridValues;
 797     switch (mode)
 798     {
 799         case CodePath::GPU: // intentional absence of break
 800         case CodePath::CPU:
 801             gridValues.clear();
 802             for (int ix = 0; ix < gridSize[XX]; ix++)
 803             {
 804                 for (int iy = 0; iy < gridSize[YY]; iy++)
 805                 {
 806                     for (int iz = 0; iz < gridSize[ZZ]; iz++)
 807                     {
 808                         IVec         temp(ix, iy, iz);
 809                         const size_t gridValueIndex =
 810                                 pmeGetGridPlainIndexInternal(temp, paddedGridSize, gridOrdering);
 811                         const ValueType value = grid[gridValueIndex];
 812                         if (value != ValueType{})
 813                         {
 814                             auto key        = formatString("Cell %d %d %d", ix, iy, iz);
 815                             gridValues[key] = value;
 816                         }
 817                     }
 818                 }
 819             }
 820             break;
 821
 822         default: GMX_THROW(InternalError("Test not implemented for this mode"));
 823     }
 824     return gridValues;
 825 }
 826
 827 //! Getting the real grid (spreading output of pmePerformSplineAndSpread())
 828 SparseRealGridValuesOutput pmeGetRealGrid(const gmx_pme_t* pme, CodePath mode)
 829 {
 830     return pmeGetGridInternal<real>(pme, mode, GridOrdering::XYZ);
 831 }
 832
 833 //! Getting the complex grid output of pmePerformSolve()
 834 SparseComplexGridValuesOutput pmeGetComplexGrid(const gmx_pme_t* pme, CodePath mode, GridOrdering gridOrdering)
 835 {
 836     return pmeGetGridInternal<t_complex>(pme, mode, gridOrdering);
 837 }
 838
 839 //! Getting the reciprocal energy and virial
 840 PmeOutput pmeGetReciprocalEnergyAndVirial(const gmx_pme_t* pme, CodePath mode, PmeSolveAlgorithm method)
 841 {
 842     PmeOutput output;
 843     switch (mode)
 844     {
 845         case CodePath::CPU:
 846             switch (method)
 847             {
 848                 case PmeSolveAlgorithm::Coulomb:
 849                     get_pme_ener_vir_q(pme->solve_work, pme->nthread, &output);
 850                     break;
 851
 852                 case PmeSolveAlgorithm::LennardJones:
 853                     get_pme_ener_vir_lj(pme->solve_work, pme->nthread, &output);
 854                     break;
 855
 856                 default: GMX_THROW(InternalError("Test not implemented for this mode"));
 857             }
 858             break;
 859         case CodePath::GPU:
 860             switch (method)
 861             {
 862                 case PmeSolveAlgorithm::Coulomb: pme_gpu_getEnergyAndVirial(*pme, &output); break;
 863
 864                 default: GMX_THROW(InternalError("Test not implemented for this mode"));
 865             }
 866             break;
 867
 868         default: GMX_THROW(InternalError("Test not implemented for this mode"));
 869     }
 870     return output;
 871 }
 872
 873 } // namespace test
 874 } // namespace gmx