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/hardware/device_management.h"
  64 #include "gromacs/math/invertmatrix.h"
  65 #include "gromacs/mdtypes/commrec.h"
  66 #include "gromacs/pbcutil/pbc.h"
  67 #include "gromacs/topology/topology.h"
  68 #include "gromacs/utility/exceptions.h"
  69 #include "gromacs/utility/gmxassert.h"
  70 #include "gromacs/utility/logger.h"
  71 #include "gromacs/utility/stringutil.h"
  72
  73 #include "testutils/test_hardware_environment.h"
  74 #include "testutils/testasserts.h"
  75
  76 class DeviceContext;
  77
  78 namespace gmx
  79 {
  80 namespace test
  81 {
  82
  83 bool pmeSupportsInputForMode(const gmx_hw_info_t& hwinfo, const t_inputrec* inputRec, CodePath mode)
  84 {
  85     bool implemented;
  86     switch (mode)
  87     {
  88         case CodePath::CPU: implemented = true; break;
  89
  90         case CodePath::GPU:
  91             implemented = (pme_gpu_supports_build(nullptr) && pme_gpu_supports_hardware(hwinfo, nullptr)
  92                            && pme_gpu_supports_input(*inputRec, nullptr));
  93             break;
  94
  95         default: GMX_THROW(InternalError("Test not implemented for this mode"));
  96     }
  97     return implemented;
  98 }
  99
 100 uint64_t getSplineModuliDoublePrecisionUlps(int splineOrder)
 101 {
 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;
 107 }
 108
 109 //! PME initialization
 110 PmeSafePointer pmeInitWrapper(const t_inputrec*    inputRec,
 111                               const CodePath       mode,
 112                               const DeviceContext* deviceContext,
 113                               const DeviceStream*  deviceStream,
 114                               const PmeGpuProgram* pmeGpuProgram,
 115                               const Matrix3x3&     box,
 116                               const real           ewaldCoeff_q,
 117                               const real           ewaldCoeff_lj)
 118 {
 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,
 125                                          deviceContext, deviceStream, pmeGpuProgram, dummyLogger);
 126     PmeSafePointer pme(pmeDataRaw); // taking ownership
 127
 128     // TODO get rid of this with proper matrix type
 129     matrix boxTemp;
 130     for (int i = 0; i < DIM; i++)
 131     {
 132         for (int j = 0; j < DIM; j++)
 133         {
 134             boxTemp[i][j] = box[i * DIM + j];
 135         }
 136     }
 137     const char* boxError = check_box(PbcType::Unset, boxTemp);
 138     GMX_RELEASE_ASSERT(boxError == nullptr, boxError);
 139
 140     switch (mode)
 141     {
 142         case CodePath::CPU: invertBoxMatrix(boxTemp, pme->recipbox); break;
 143
 144         case CodePath::GPU:
 145             pme_gpu_set_testing(pme->gpu, true);
 146             pme_gpu_update_input_box(pme->gpu, boxTemp);
 147             break;
 148
 149         default: GMX_THROW(InternalError("Test not implemented for this mode"));
 150     }
 151
 152     return pme;
 153 }
 154
 155 PmeSafePointer pmeInitEmpty(const t_inputrec* inputRec)
 156 {
 157     const Matrix3x3 defaultBox = { { 1.0F, 0.0F, 0.0F, 0.0F, 1.0F, 0.0F, 0.0F, 0.0F, 1.0F } };
 158     return pmeInitWrapper(inputRec, CodePath::CPU, nullptr, nullptr, nullptr, defaultBox, 0.0F, 0.0F);
 159 }
 160
 161 //! Make a GPU state-propagator manager
 162 std::unique_ptr<StatePropagatorDataGpu> makeStatePropagatorDataGpu(const gmx_pme_t&     pme,
 163                                                                    const DeviceContext* deviceContext,
 164                                                                    const DeviceStream* deviceStream)
 165 {
 166     // TODO: Pin the host buffer and use async memory copies
 167     // TODO: Special constructor for PME-only rank / PME-tests is used here. There should be a mechanism to
 168     //       restrict one from using other constructor here.
 169     return std::make_unique<StatePropagatorDataGpu>(deviceStream, *deviceContext, GpuApiCallBehavior::Sync,
 170                                                     pme_gpu_get_block_size(&pme), nullptr);
 171 }
 172
 173 //! PME initialization with atom data
 174 void pmeInitAtoms(gmx_pme_t*               pme,
 175                   StatePropagatorDataGpu*  stateGpu,
 176                   const CodePath           mode,
 177                   const CoordinatesVector& coordinates,
 178                   const ChargesVector&     charges)
 179 {
 180     const index atomCount = coordinates.size();
 181     GMX_RELEASE_ASSERT(atomCount == charges.ssize(), "Mismatch in atom data");
 182     PmeAtomComm* atc = nullptr;
 183
 184     switch (mode)
 185     {
 186         case CodePath::CPU:
 187             atc              = &(pme->atc[0]);
 188             atc->x           = coordinates;
 189             atc->coefficient = charges;
 190             gmx_pme_reinit_atoms(pme, atomCount, charges.data(), nullptr);
 191             /* With decomposition there would be more boilerplate atc code here, e.g. do_redist_pos_coeffs */
 192             break;
 193
 194         case CodePath::GPU:
 195             // TODO: Avoid use of atc in the GPU code path
 196             atc = &(pme->atc[0]);
 197             // We need to set atc->n for passing the size in the tests
 198             atc->setNumAtoms(atomCount);
 199             gmx_pme_reinit_atoms(pme, atomCount, charges.data(), nullptr);
 200
 201             stateGpu->reinit(atomCount, atomCount);
 202             stateGpu->copyCoordinatesToGpu(arrayRefFromArray(coordinates.data(), coordinates.size()),
 203                                            gmx::AtomLocality::All);
 204             pme_gpu_set_kernelparam_coordinates(pme->gpu, stateGpu->getCoordinates());
 205
 206             break;
 207
 208         default: GMX_THROW(InternalError("Test not implemented for this mode"));
 209     }
 210 }
 211
 212 //! Getting local PME real grid pointer for test I/O
 213 static real* pmeGetRealGridInternal(const gmx_pme_t* pme)
 214 {
 215     const size_t gridIndex = 0;
 216     return pme->fftgrid[gridIndex];
 217 }
 218
 219 //! Getting local PME real grid dimensions
 220 static void pmeGetRealGridSizesInternal(const gmx_pme_t* pme,
 221                                         CodePath         mode,
 222                                         IVec& gridSize,       //NOLINT(google-runtime-references)
 223                                         IVec& paddedGridSize) //NOLINT(google-runtime-references)
 224 {
 225     const size_t gridIndex = 0;
 226     IVec         gridOffsetUnused;
 227     switch (mode)
 228     {
 229         case CodePath::CPU:
 230             gmx_parallel_3dfft_real_limits(pme->pfft_setup[gridIndex], gridSize, gridOffsetUnused,
 231                                            paddedGridSize);
 232             break;
 233
 234         case CodePath::GPU:
 235             pme_gpu_get_real_grid_sizes(pme->gpu, &gridSize, &paddedGridSize);
 236             break;
 237
 238         default: GMX_THROW(InternalError("Test not implemented for this mode"));
 239     }
 240 }
 241
 242 //! Getting local PME complex grid pointer for test I/O
 243 static t_complex* pmeGetComplexGridInternal(const gmx_pme_t* pme)
 244 {
 245     const size_t gridIndex = 0;
 246     return pme->cfftgrid[gridIndex];
 247 }
 248
 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)
 253 {
 254     const size_t gridIndex = 0;
 255     IVec         gridOffsetUnused, complexOrderUnused;
 256     gmx_parallel_3dfft_complex_limits(pme->pfft_setup[gridIndex], complexOrderUnused, gridSize,
 257                                       gridOffsetUnused, paddedGridSize); // TODO: what about YZX ordering?
 258 }
 259
 260 //! Getting the PME grid memory buffer and its sizes - template definition
 261 template<typename ValueType>
 262 static void pmeGetGridAndSizesInternal(const gmx_pme_t* /*unused*/,
 263                                        CodePath /*unused*/,
 264                                        ValueType*& /*unused*/, //NOLINT(google-runtime-references)
 265                                        IVec& /*unused*/,       //NOLINT(google-runtime-references)
 266                                        IVec& /*unused*/)       //NOLINT(google-runtime-references)
 267 {
 268     GMX_THROW(InternalError("Deleted function call"));
 269     // explicitly deleting general template does not compile in clang/icc, see https://llvm.org/bugs/show_bug.cgi?id=17537
 270 }
 271
 272 //! Getting the PME real grid memory buffer and its sizes
 273 template<>
 274 void pmeGetGridAndSizesInternal<real>(const gmx_pme_t* pme, CodePath mode, real*& grid, IVec& gridSize, IVec& paddedGridSize)
 275 {
 276     grid = pmeGetRealGridInternal(pme);
 277     pmeGetRealGridSizesInternal(pme, mode, gridSize, paddedGridSize);
 278 }
 279
 280 //! Getting the PME complex grid memory buffer and its sizes
 281 template<>
 282 void pmeGetGridAndSizesInternal<t_complex>(const gmx_pme_t* pme,
 283                                            CodePath /*unused*/,
 284                                            t_complex*& grid,
 285                                            IVec&       gridSize,
 286                                            IVec&       paddedGridSize)
 287 {
 288     grid = pmeGetComplexGridInternal(pme);
 289     pmeGetComplexGridSizesInternal(pme, gridSize, paddedGridSize);
 290 }
 291
 292 //! PME spline calculation and charge spreading
 293 void pmePerformSplineAndSpread(gmx_pme_t* pme,
 294                                CodePath   mode, // TODO const qualifiers elsewhere
 295                                bool       computeSplines,
 296                                bool       spreadCharges)
 297 {
 298     GMX_RELEASE_ASSERT(pme != nullptr, "PME data is not initialized");
 299     PmeAtomComm* atc                          = &(pme->atc[0]);
 300     const size_t gridIndex                    = 0;
 301     const bool   computeSplinesForZeroCharges = true;
 302     real**       fftgrid                      = spreadCharges ? pme->fftgrid : nullptr;
 303     real*        pmegrid                      = pme->pmegrid[gridIndex].grid.grid;
 304
 305     switch (mode)
 306     {
 307         case CodePath::CPU:
 308             spread_on_grid(pme, atc, &pme->pmegrid[gridIndex], computeSplines, spreadCharges,
 309                            fftgrid != nullptr ? fftgrid[gridIndex] : nullptr,
 310                            computeSplinesForZeroCharges, gridIndex);
 311             if (spreadCharges && !pme->bUseThreads)
 312             {
 313                 wrap_periodic_pmegrid(pme, pmegrid);
 314                 copy_pmegrid_to_fftgrid(
 315                         pme, pmegrid, fftgrid != nullptr ? fftgrid[gridIndex] : nullptr, gridIndex);
 316             }
 317             break;
 318
 319 /* The compiler will complain about passing fftgrid (converting double ** to float **) if using
 320  * double precision. GPUs are not used with double precision anyhow. */
 321 #if !GMX_DOUBLE
 322         case CodePath::GPU:
 323         {
 324             const real lambdaQ = 1.0;
 325             // no synchronization needed as x is transferred in the PME stream
 326             GpuEventSynchronizer* xReadyOnDevice = nullptr;
 327             pme_gpu_spread(pme->gpu, xReadyOnDevice, fftgrid, computeSplines, spreadCharges, lambdaQ);
 328         }
 329         break;
 330 #endif
 331
 332         default: GMX_THROW(InternalError("Test not implemented for this mode"));
 333     }
 334 }
 335
 336 //! Getting the internal spline data buffer pointer
 337 static real* pmeGetSplineDataInternal(const gmx_pme_t* pme, PmeSplineDataType type, int dimIndex)
 338 {
 339     GMX_ASSERT((0 <= dimIndex) && (dimIndex < DIM), "Invalid dimension index");
 340     const PmeAtomComm* atc          = &(pme->atc[0]);
 341     const size_t       threadIndex  = 0;
 342     real*              splineBuffer = nullptr;
 343     switch (type)
 344     {
 345         case PmeSplineDataType::Values:
 346             splineBuffer = atc->spline[threadIndex].theta.coefficients[dimIndex];
 347             break;
 348
 349         case PmeSplineDataType::Derivatives:
 350             splineBuffer = atc->spline[threadIndex].dtheta.coefficients[dimIndex];
 351             break;
 352
 353         default: GMX_THROW(InternalError("Unknown spline data type"));
 354     }
 355     return splineBuffer;
 356 }
 357
 358 //! PME solving
 359 void pmePerformSolve(const gmx_pme_t*  pme,
 360                      CodePath          mode,
 361                      PmeSolveAlgorithm method,
 362                      real              cellVolume,
 363                      GridOrdering      gridOrdering,
 364                      bool              computeEnergyAndVirial)
 365 {
 366     t_complex*   h_grid              = pmeGetComplexGridInternal(pme);
 367     const bool   useLorentzBerthelot = false;
 368     const size_t threadIndex         = 0;
 369     const size_t gridIndex           = 0;
 370     switch (mode)
 371     {
 372         case CodePath::CPU:
 373             if (gridOrdering != GridOrdering::YZX)
 374             {
 375                 GMX_THROW(InternalError("Test not implemented for this mode"));
 376             }
 377             switch (method)
 378             {
 379                 case PmeSolveAlgorithm::Coulomb:
 380                     solve_pme_yzx(pme, h_grid, cellVolume, computeEnergyAndVirial, pme->nthread, threadIndex);
 381                     break;
 382
 383                 case PmeSolveAlgorithm::LennardJones:
 384                     solve_pme_lj_yzx(pme, &h_grid, useLorentzBerthelot, cellVolume,
 385                                      computeEnergyAndVirial, pme->nthread, threadIndex);
 386                     break;
 387
 388                 default: GMX_THROW(InternalError("Test not implemented for this mode"));
 389             }
 390             break;
 391
 392         case CodePath::GPU:
 393             switch (method)
 394             {
 395                 case PmeSolveAlgorithm::Coulomb:
 396                     pme_gpu_solve(pme->gpu, gridIndex, h_grid, gridOrdering, computeEnergyAndVirial);
 397                     break;
 398
 399                 default: GMX_THROW(InternalError("Test not implemented for this mode"));
 400             }
 401             break;
 402
 403         default: GMX_THROW(InternalError("Test not implemented for this mode"));
 404     }
 405 }
 406
 407 //! PME force gathering
 408 void pmePerformGather(gmx_pme_t* pme, CodePath mode, ForcesVector& forces)
 409 {
 410     PmeAtomComm* atc       = &(pme->atc[0]);
 411     const index  atomCount = atc->numAtoms();
 412     GMX_RELEASE_ASSERT(forces.ssize() == atomCount, "Invalid force buffer size");
 413     const real   scale       = 1.0;
 414     const size_t threadIndex = 0;
 415     const size_t gridIndex   = 0;
 416     real*        pmegrid     = pme->pmegrid[gridIndex].grid.grid;
 417     real**       fftgrid     = pme->fftgrid;
 418
 419     switch (mode)
 420     {
 421         case CodePath::CPU:
 422             atc->f = forces;
 423             if (atc->nthread == 1)
 424             {
 425                 // something which is normally done in serial spline computation (make_thread_local_ind())
 426                 atc->spline[threadIndex].n = atomCount;
 427             }
 428             copy_fftgrid_to_pmegrid(pme, fftgrid[gridIndex], pmegrid, gridIndex, pme->nthread, threadIndex);
 429             unwrap_periodic_pmegrid(pme, pmegrid);
 430             gather_f_bsplines(pme, pmegrid, true, atc, &atc->spline[threadIndex], scale);
 431             break;
 432
 433 /* The compiler will complain about passing fftgrid (converting double ** to float **) if using
 434  * double precision. GPUs are not used with double precision anyhow. */
 435 #if !GMX_DOUBLE
 436         case CodePath::GPU:
 437         {
 438             // Variable initialization needs a non-switch scope
 439             const bool computeEnergyAndVirial = false;
 440             const real lambdaQ                = 1.0;
 441             PmeOutput  output = pme_gpu_getOutput(*pme, computeEnergyAndVirial, lambdaQ);
 442             GMX_ASSERT(forces.size() == output.forces_.size(),
 443                        "Size of force buffers did not match");
 444             pme_gpu_gather(pme->gpu, fftgrid, lambdaQ);
 445             std::copy(std::begin(output.forces_), std::end(output.forces_), std::begin(forces));
 446         }
 447         break;
 448 #endif
 449
 450         default: GMX_THROW(InternalError("Test not implemented for this mode"));
 451     }
 452 }
 453
 454 //! PME test finalization before fetching the outputs
 455 void pmeFinalizeTest(const gmx_pme_t* pme, CodePath mode)
 456 {
 457     switch (mode)
 458     {
 459         case CodePath::CPU: break;
 460
 461         case CodePath::GPU: pme_gpu_synchronize(pme->gpu); break;
 462
 463         default: GMX_THROW(InternalError("Test not implemented for this mode"));
 464     }
 465 }
 466
 467 //! A binary enum for spline data layout transformation
 468 enum class PmeLayoutTransform
 469 {
 470     GpuToHost,
 471     HostToGpu
 472 };
 473
 474 /*! \brief Gets a unique index to an element in a spline parameter buffer.
 475  *
 476  * These theta/dtheta buffers are laid out for GPU spread/gather
 477  * kernels. The index is wrt the execution block, in range(0,
 478  * atomsPerBlock * order * DIM).
 479  *
 480  * This is a wrapper, only used in unit tests.
 481  * \param[in] order            PME order
 482  * \param[in] splineIndex      Spline contribution index (from 0 to \p order - 1)
 483  * \param[in] dimIndex         Dimension index (from 0 to 2)
 484  * \param[in] atomIndex        Atom index wrt the block.
 485  * \param[in] atomsPerWarp     Number of atoms processed by a warp.
 486  *
 487  * \returns Index into theta or dtheta array using GPU layout.
 488  */
 489 static int getSplineParamFullIndex(int order, int splineIndex, int dimIndex, int atomIndex, int atomsPerWarp)
 490 {
 491     if (order != c_pmeGpuOrder)
 492     {
 493         throw order;
 494     }
 495     constexpr int fixedOrder = c_pmeGpuOrder;
 496     GMX_UNUSED_VALUE(fixedOrder);
 497
 498     const int atomWarpIndex = atomIndex % atomsPerWarp;
 499     const int warpIndex     = atomIndex / atomsPerWarp;
 500     int       indexBase, result;
 501     switch (atomsPerWarp)
 502     {
 503         case 1:
 504             indexBase = getSplineParamIndexBase<fixedOrder, 1>(warpIndex, atomWarpIndex);
 505             result    = getSplineParamIndex<fixedOrder, 1>(indexBase, dimIndex, splineIndex);
 506             break;
 507
 508         case 2:
 509             indexBase = getSplineParamIndexBase<fixedOrder, 2>(warpIndex, atomWarpIndex);
 510             result    = getSplineParamIndex<fixedOrder, 2>(indexBase, dimIndex, splineIndex);
 511             break;
 512
 513         case 4:
 514             indexBase = getSplineParamIndexBase<fixedOrder, 4>(warpIndex, atomWarpIndex);
 515             result    = getSplineParamIndex<fixedOrder, 4>(indexBase, dimIndex, splineIndex);
 516             break;
 517
 518         case 8:
 519             indexBase = getSplineParamIndexBase<fixedOrder, 8>(warpIndex, atomWarpIndex);
 520             result    = getSplineParamIndex<fixedOrder, 8>(indexBase, dimIndex, splineIndex);
 521             break;
 522
 523         default:
 524             GMX_THROW(NotImplementedError(
 525                     formatString("Test function call not unrolled for atomsPerWarp = %d in "
 526                                  "getSplineParamFullIndex",
 527                                  atomsPerWarp)));
 528     }
 529     return result;
 530 }
 531
 532 /*!\brief Return the number of atoms per warp */
 533 static int pme_gpu_get_atoms_per_warp(const PmeGpu* pmeGpu)
 534 {
 535     const int order = pmeGpu->common->pme_order;
 536     const int threadsPerAtom =
 537             (pmeGpu->settings.threadsPerAtom == ThreadsPerAtom::Order ? order : order * order);
 538     return pmeGpu->programHandle_->warpSize() / threadsPerAtom;
 539 }
 540
 541 /*! \brief Rearranges the atom spline data between the GPU and host layouts.
 542  * Only used for test purposes so far, likely to be horribly slow.
 543  *
 544  * \param[in]  pmeGpu     The PME GPU structure.
 545  * \param[out] atc        The PME CPU atom data structure (with a single-threaded layout).
 546  * \param[in]  type       The spline data type (values or derivatives).
 547  * \param[in]  dimIndex   Dimension index.
 548  * \param[in]  transform  Layout transform type
 549  */
 550 static void pme_gpu_transform_spline_atom_data(const PmeGpu*      pmeGpu,
 551                                                const PmeAtomComm* atc,
 552                                                PmeSplineDataType  type,
 553                                                int                dimIndex,
 554                                                PmeLayoutTransform transform)
 555 {
 556     // The GPU atom spline data is laid out in a different way currently than the CPU one.
 557     // This function converts the data from GPU to CPU layout (in the host memory).
 558     // It is only intended for testing purposes so far.
 559     // Ideally we should use similar layouts on CPU and GPU if we care about mixed modes and their
 560     // performance (e.g. spreading on GPU, gathering on CPU).
 561     GMX_RELEASE_ASSERT(atc->nthread == 1, "Only the serial PME data layout is supported");
 562     const uintmax_t threadIndex  = 0;
 563     const auto      atomCount    = atc->numAtoms();
 564     const auto      atomsPerWarp = pme_gpu_get_atoms_per_warp(pmeGpu);
 565     const auto      pmeOrder     = pmeGpu->common->pme_order;
 566     GMX_ASSERT(pmeOrder == c_pmeGpuOrder, "Only PME order 4 is implemented");
 567
 568     real*  cpuSplineBuffer;
 569     float* h_splineBuffer;
 570     switch (type)
 571     {
 572         case PmeSplineDataType::Values:
 573             cpuSplineBuffer = atc->spline[threadIndex].theta.coefficients[dimIndex];
 574             h_splineBuffer  = pmeGpu->staging.h_theta;
 575             break;
 576
 577         case PmeSplineDataType::Derivatives:
 578             cpuSplineBuffer = atc->spline[threadIndex].dtheta.coefficients[dimIndex];
 579             h_splineBuffer  = pmeGpu->staging.h_dtheta;
 580             break;
 581
 582         default: GMX_THROW(InternalError("Unknown spline data type"));
 583     }
 584
 585     for (auto atomIndex = 0; atomIndex < atomCount; atomIndex++)
 586     {
 587         for (auto orderIndex = 0; orderIndex < pmeOrder; orderIndex++)
 588         {
 589             const auto gpuValueIndex =
 590                     getSplineParamFullIndex(pmeOrder, orderIndex, dimIndex, atomIndex, atomsPerWarp);
 591             const auto cpuValueIndex = atomIndex * pmeOrder + orderIndex;
 592             GMX_ASSERT(cpuValueIndex < atomCount * pmeOrder,
 593                        "Atom spline data index out of bounds (while transforming GPU data layout "
 594                        "for host)");
 595             switch (transform)
 596             {
 597                 case PmeLayoutTransform::GpuToHost:
 598                     cpuSplineBuffer[cpuValueIndex] = h_splineBuffer[gpuValueIndex];
 599                     break;
 600
 601                 case PmeLayoutTransform::HostToGpu:
 602                     h_splineBuffer[gpuValueIndex] = cpuSplineBuffer[cpuValueIndex];
 603                     break;
 604
 605                 default: GMX_THROW(InternalError("Unknown layout transform"));
 606             }
 607         }
 608     }
 609 }
 610
 611 //! Setting atom spline values/derivatives to be used in spread/gather
 612 void pmeSetSplineData(const gmx_pme_t*             pme,
 613                       CodePath                     mode,
 614                       const SplineParamsDimVector& splineValues,
 615                       PmeSplineDataType            type,
 616                       int                          dimIndex)
 617 {
 618     const PmeAtomComm* atc       = &(pme->atc[0]);
 619     const index        atomCount = atc->numAtoms();
 620     const index        pmeOrder  = pme->pme_order;
 621     const index        dimSize   = pmeOrder * atomCount;
 622     GMX_RELEASE_ASSERT(dimSize == splineValues.ssize(), "Mismatch in spline data");
 623     real* splineBuffer = pmeGetSplineDataInternal(pme, type, dimIndex);
 624
 625     switch (mode)
 626     {
 627         case CodePath::CPU:
 628             std::copy(splineValues.begin(), splineValues.end(), splineBuffer);
 629             break;
 630
 631         case CodePath::GPU:
 632             std::copy(splineValues.begin(), splineValues.end(), splineBuffer);
 633             pme_gpu_transform_spline_atom_data(pme->gpu, atc, type, dimIndex, PmeLayoutTransform::HostToGpu);
 634             break;
 635
 636         default: GMX_THROW(InternalError("Test not implemented for this mode"));
 637     }
 638 }
 639
 640 //! Setting gridline indices to be used in spread/gather
 641 void pmeSetGridLineIndices(gmx_pme_t* pme, CodePath mode, const GridLineIndicesVector& gridLineIndices)
 642 {
 643     PmeAtomComm* atc       = &(pme->atc[0]);
 644     const index  atomCount = atc->numAtoms();
 645     GMX_RELEASE_ASSERT(atomCount == gridLineIndices.ssize(), "Mismatch in gridline indices size");
 646
 647     IVec paddedGridSizeUnused, gridSize(0, 0, 0);
 648     pmeGetRealGridSizesInternal(pme, mode, gridSize, paddedGridSizeUnused);
 649
 650     for (const auto& index : gridLineIndices)
 651     {
 652         for (int i = 0; i < DIM; i++)
 653         {
 654             GMX_RELEASE_ASSERT((0 <= index[i]) && (index[i] < gridSize[i]),
 655                                "Invalid gridline index");
 656         }
 657     }
 658
 659     switch (mode)
 660     {
 661         case CodePath::GPU:
 662             memcpy(pme_gpu_staging(pme->gpu).h_gridlineIndices, gridLineIndices.data(),
 663                    atomCount * sizeof(gridLineIndices[0]));
 664             break;
 665
 666         case CodePath::CPU:
 667             atc->idx.resize(gridLineIndices.size());
 668             std::copy(gridLineIndices.begin(), gridLineIndices.end(), atc->idx.begin());
 669             break;
 670         default: GMX_THROW(InternalError("Test not implemented for this mode"));
 671     }
 672 }
 673
 674 //! Getting plain index into the complex 3d grid
 675 inline size_t pmeGetGridPlainIndexInternal(const IVec& index, const IVec& paddedGridSize, GridOrdering gridOrdering)
 676 {
 677     size_t result;
 678     switch (gridOrdering)
 679     {
 680         case GridOrdering::YZX:
 681             result = (index[YY] * paddedGridSize[ZZ] + index[ZZ]) * paddedGridSize[XX] + index[XX];
 682             break;
 683
 684         case GridOrdering::XYZ:
 685             result = (index[XX] * paddedGridSize[YY] + index[YY]) * paddedGridSize[ZZ] + index[ZZ];
 686             break;
 687
 688         default: GMX_THROW(InternalError("Test not implemented for this mode"));
 689     }
 690     return result;
 691 }
 692
 693 //! Setting real or complex grid
 694 template<typename ValueType>
 695 static void pmeSetGridInternal(const gmx_pme_t*                        pme,
 696                                CodePath                                mode,
 697                                GridOrdering                            gridOrdering,
 698                                const SparseGridValuesInput<ValueType>& gridValues)
 699 {
 700     IVec       gridSize(0, 0, 0), paddedGridSize(0, 0, 0);
 701     ValueType* grid;
 702     pmeGetGridAndSizesInternal<ValueType>(pme, mode, grid, gridSize, paddedGridSize);
 703
 704     switch (mode)
 705     {
 706         case CodePath::GPU: // intentional absence of break, the grid will be copied from the host buffer in testing mode
 707         case CodePath::CPU:
 708             std::memset(grid, 0,
 709                         paddedGridSize[XX] * paddedGridSize[YY] * paddedGridSize[ZZ] * sizeof(ValueType));
 710             for (const auto& gridValue : gridValues)
 711             {
 712                 for (int i = 0; i < DIM; i++)
 713                 {
 714                     GMX_RELEASE_ASSERT((0 <= gridValue.first[i]) && (gridValue.first[i] < gridSize[i]),
 715                                        "Invalid grid value index");
 716                 }
 717                 const size_t gridValueIndex =
 718                         pmeGetGridPlainIndexInternal(gridValue.first, paddedGridSize, gridOrdering);
 719                 grid[gridValueIndex] = gridValue.second;
 720             }
 721             break;
 722
 723         default: GMX_THROW(InternalError("Test not implemented for this mode"));
 724     }
 725 }
 726
 727 //! Setting real grid to be used in gather
 728 void pmeSetRealGrid(const gmx_pme_t* pme, CodePath mode, const SparseRealGridValuesInput& gridValues)
 729 {
 730     pmeSetGridInternal<real>(pme, mode, GridOrdering::XYZ, gridValues);
 731 }
 732
 733 //! Setting complex grid to be used in solve
 734 void pmeSetComplexGrid(const gmx_pme_t*                    pme,
 735                        CodePath                            mode,
 736                        GridOrdering                        gridOrdering,
 737                        const SparseComplexGridValuesInput& gridValues)
 738 {
 739     pmeSetGridInternal<t_complex>(pme, mode, gridOrdering, gridValues);
 740 }
 741
 742 //! Getting the single dimension's spline values or derivatives
 743 SplineParamsDimVector pmeGetSplineData(const gmx_pme_t* pme, CodePath mode, PmeSplineDataType type, int dimIndex)
 744 {
 745     GMX_RELEASE_ASSERT(pme != nullptr, "PME data is not initialized");
 746     const PmeAtomComm* atc       = &(pme->atc[0]);
 747     const size_t       atomCount = atc->numAtoms();
 748     const size_t       pmeOrder  = pme->pme_order;
 749     const size_t       dimSize   = pmeOrder * atomCount;
 750
 751     real*                 sourceBuffer = pmeGetSplineDataInternal(pme, type, dimIndex);
 752     SplineParamsDimVector result;
 753     switch (mode)
 754     {
 755         case CodePath::GPU:
 756             pme_gpu_transform_spline_atom_data(pme->gpu, atc, type, dimIndex, PmeLayoutTransform::GpuToHost);
 757             result = arrayRefFromArray(sourceBuffer, dimSize);
 758             break;
 759
 760         case CodePath::CPU: result = arrayRefFromArray(sourceBuffer, dimSize); break;
 761
 762         default: GMX_THROW(InternalError("Test not implemented for this mode"));
 763     }
 764     return result;
 765 }
 766
 767 //! Getting the gridline indices
 768 GridLineIndicesVector pmeGetGridlineIndices(const gmx_pme_t* pme, CodePath mode)
 769 {
 770     GMX_RELEASE_ASSERT(pme != nullptr, "PME data is not initialized");
 771     const PmeAtomComm* atc       = &(pme->atc[0]);
 772     const size_t       atomCount = atc->numAtoms();
 773
 774     GridLineIndicesVector gridLineIndices;
 775     switch (mode)
 776     {
 777         case CodePath::GPU:
 778             gridLineIndices = arrayRefFromArray(
 779                     reinterpret_cast<IVec*>(pme_gpu_staging(pme->gpu).h_gridlineIndices), atomCount);
 780             break;
 781
 782         case CodePath::CPU: gridLineIndices = atc->idx; break;
 783
 784         default: GMX_THROW(InternalError("Test not implemented for this mode"));
 785     }
 786     return gridLineIndices;
 787 }
 788
 789 //! Getting real or complex grid - only non zero values
 790 template<typename ValueType>
 791 static SparseGridValuesOutput<ValueType> pmeGetGridInternal(const gmx_pme_t* pme,
 792                                                             CodePath         mode,
 793                                                             GridOrdering     gridOrdering)
 794 {
 795     IVec       gridSize(0, 0, 0), paddedGridSize(0, 0, 0);
 796     ValueType* grid;
 797     pmeGetGridAndSizesInternal<ValueType>(pme, mode, grid, gridSize, paddedGridSize);
 798     SparseGridValuesOutput<ValueType> gridValues;
 799     switch (mode)
 800     {
 801         case CodePath::GPU: // intentional absence of break
 802         case CodePath::CPU:
 803             gridValues.clear();
 804             for (int ix = 0; ix < gridSize[XX]; ix++)
 805             {
 806                 for (int iy = 0; iy < gridSize[YY]; iy++)
 807                 {
 808                     for (int iz = 0; iz < gridSize[ZZ]; iz++)
 809                     {
 810                         IVec         temp(ix, iy, iz);
 811                         const size_t gridValueIndex =
 812                                 pmeGetGridPlainIndexInternal(temp, paddedGridSize, gridOrdering);
 813                         const ValueType value = grid[gridValueIndex];
 814                         if (value != ValueType{})
 815                         {
 816                             auto key        = formatString("Cell %d %d %d", ix, iy, iz);
 817                             gridValues[key] = value;
 818                         }
 819                     }
 820                 }
 821             }
 822             break;
 823
 824         default: GMX_THROW(InternalError("Test not implemented for this mode"));
 825     }
 826     return gridValues;
 827 }
 828
 829 //! Getting the real grid (spreading output of pmePerformSplineAndSpread())
 830 SparseRealGridValuesOutput pmeGetRealGrid(const gmx_pme_t* pme, CodePath mode)
 831 {
 832     return pmeGetGridInternal<real>(pme, mode, GridOrdering::XYZ);
 833 }
 834
 835 //! Getting the complex grid output of pmePerformSolve()
 836 SparseComplexGridValuesOutput pmeGetComplexGrid(const gmx_pme_t* pme, CodePath mode, GridOrdering gridOrdering)
 837 {
 838     return pmeGetGridInternal<t_complex>(pme, mode, gridOrdering);
 839 }
 840
 841 //! Getting the reciprocal energy and virial
 842 PmeOutput pmeGetReciprocalEnergyAndVirial(const gmx_pme_t* pme, CodePath mode, PmeSolveAlgorithm method)
 843 {
 844     PmeOutput  output;
 845     const real lambdaQ = 1.0;
 846     switch (mode)
 847     {
 848         case CodePath::CPU:
 849             switch (method)
 850             {
 851                 case PmeSolveAlgorithm::Coulomb:
 852                     get_pme_ener_vir_q(pme->solve_work, pme->nthread, &output);
 853                     break;
 854
 855                 case PmeSolveAlgorithm::LennardJones:
 856                     get_pme_ener_vir_lj(pme->solve_work, pme->nthread, &output);
 857                     break;
 858
 859                 default: GMX_THROW(InternalError("Test not implemented for this mode"));
 860             }
 861             break;
 862         case CodePath::GPU:
 863             switch (method)
 864             {
 865                 case PmeSolveAlgorithm::Coulomb:
 866                     pme_gpu_getEnergyAndVirial(*pme, lambdaQ, &output);
 867                     break;
 868
 869                 default: GMX_THROW(InternalError("Test not implemented for this mode"));
 870             }
 871             break;
 872
 873         default: GMX_THROW(InternalError("Test not implemented for this mode"));
 874     }
 875     return output;
 876 }
 877
 878 const char* codePathToString(CodePath codePath)
 879 {
 880     switch (codePath)
 881     {
 882         case CodePath::CPU: return "CPU";
 883         case CodePath::GPU: return "GPU";
 884         default: GMX_THROW(NotImplementedError("This CodePath should support codePathToString"));
 885     }
 886 }
 887
 888 PmeTestHardwareContext::PmeTestHardwareContext() : codePath_(CodePath::CPU) {}
 889
 890 PmeTestHardwareContext::PmeTestHardwareContext(TestDevice* testDevice) :
 891     codePath_(CodePath::CPU),
 892     testDevice_(testDevice)
 893 {
 894     setActiveDevice(testDevice_->deviceInfo());
 895     pmeGpuProgram_ = buildPmeGpuProgram(testDevice_->deviceContext());
 896 }
 897
 898 //! Returns a human-readable context description line
 899 std::string PmeTestHardwareContext::description() const
 900 {
 901     switch (codePath_)
 902     {
 903         case CodePath::CPU: return "CPU";
 904         case CodePath::GPU: return "GPU (" + testDevice_->description() + ")";
 905         default: return "Unknown code path.";
 906     }
 907 }
 908
 909 void PmeTestHardwareContext::activate() const
 910 {
 911     if (codePath_ == CodePath::GPU)
 912     {
 913         setActiveDevice(testDevice_->deviceInfo());
 914     }
 915 }
 916
 917 std::vector<std::unique_ptr<PmeTestHardwareContext>> createPmeTestHardwareContextList()
 918 {
 919     std::vector<std::unique_ptr<PmeTestHardwareContext>> pmeTestHardwareContextList;
 920     // Add CPU
 921     pmeTestHardwareContextList.emplace_back(std::make_unique<PmeTestHardwareContext>());
 922     // Add GPU devices
 923     const auto& testDeviceList = getTestHardwareEnvironment()->getTestDeviceList();
 924     for (const auto& testDevice : testDeviceList)
 925     {
 926         pmeTestHardwareContextList.emplace_back(std::make_unique<PmeTestHardwareContext>(testDevice.get()));
 927     }
 928     return pmeTestHardwareContextList;
 929 }
 930
 931 } // namespace test
 932 } // namespace gmx