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/device_stream_manager.h"
  63 #include "gromacs/gpu_utils/gpu_utils.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/testasserts.h"
  74
  75 #include "testhardwarecontexts.h"
  76
  77 namespace gmx
  78 {
  79 namespace test
  80 {
  81
  82 bool pmeSupportsInputForMode(const gmx_hw_info_t& hwinfo, const t_inputrec* inputRec, CodePath mode)
  83 {
  84     bool implemented;
  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, 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 DeviceContext* deviceContext,
 112                               const DeviceStream*  deviceStream,
 113                               const PmeGpuProgram* pmeGpuProgram,
 114                               const Matrix3x3&     box,
 115                               const real           ewaldCoeff_q,
 116                               const real           ewaldCoeff_lj)
 117 {
 118     const MDLogger dummyLogger;
 119     const auto     runMode       = (mode == CodePath::CPU) ? PmeRunMode::CPU : PmeRunMode::Mixed;
 120     t_commrec      dummyCommrec  = { 0 };
 121     NumPmeDomains  numPmeDomains = { 1, 1 };
 122     gmx_pme_t* pmeDataRaw = gmx_pme_init(&dummyCommrec, numPmeDomains, inputRec, false, false, true,
 123                                          ewaldCoeff_q, ewaldCoeff_lj, 1, runMode, nullptr,
 124                                          deviceContext, deviceStream, pmeGpuProgram, dummyLogger);
 125     PmeSafePointer pme(pmeDataRaw); // taking ownership
 126
 127     // TODO get rid of this with proper matrix type
 128     matrix boxTemp;
 129     for (int i = 0; i < DIM; i++)
 130     {
 131         for (int j = 0; j < DIM; j++)
 132         {
 133             boxTemp[i][j] = box[i * DIM + j];
 134         }
 135     }
 136     const char* boxError = check_box(PbcType::Unset, boxTemp);
 137     GMX_RELEASE_ASSERT(boxError == nullptr, boxError);
 138
 139     switch (mode)
 140     {
 141         case CodePath::CPU: invertBoxMatrix(boxTemp, pme->recipbox); break;
 142
 143         case CodePath::GPU:
 144             pme_gpu_set_testing(pme->gpu, true);
 145             pme_gpu_update_input_box(pme->gpu, boxTemp);
 146             break;
 147
 148         default: GMX_THROW(InternalError("Test not implemented for this mode"));
 149     }
 150
 151     return pme;
 152 }
 153
 154 //! Simple PME initialization based on input, no atom data
 155 PmeSafePointer pmeInitEmpty(const t_inputrec*    inputRec,
 156                             const CodePath       mode,
 157                             const DeviceContext* deviceContext,
 158                             const DeviceStream*  deviceStream,
 159                             const PmeGpuProgram* pmeGpuProgram,
 160                             const Matrix3x3&     box,
 161                             const real           ewaldCoeff_q,
 162                             const real           ewaldCoeff_lj)
 163 {
 164     return pmeInitWrapper(inputRec, mode, deviceContext, deviceStream, pmeGpuProgram, box,
 165                           ewaldCoeff_q, ewaldCoeff_lj);
 166     // hiding the fact that PME actually needs to know the number of atoms in advance
 167 }
 168
 169 PmeSafePointer pmeInitEmpty(const t_inputrec* inputRec)
 170 {
 171     const Matrix3x3 defaultBox = { { 1.0F, 0.0F, 0.0F, 0.0F, 1.0F, 0.0F, 0.0F, 0.0F, 1.0F } };
 172     return pmeInitWrapper(inputRec, CodePath::CPU, nullptr, nullptr, nullptr, defaultBox, 0.0F, 0.0F);
 173 }
 174
 175 //! Make a GPU state-propagator manager
 176 std::unique_ptr<StatePropagatorDataGpu> makeStatePropagatorDataGpu(const gmx_pme_t&     pme,
 177                                                                    const DeviceContext* deviceContext,
 178                                                                    const DeviceStream* deviceStream)
 179 {
 180     // TODO: Pin the host buffer and use async memory copies
 181     // TODO: Special constructor for PME-only rank / PME-tests is used here. There should be a mechanism to
 182     //       restrict one from using other constructor here.
 183     return std::make_unique<StatePropagatorDataGpu>(deviceStream, *deviceContext, GpuApiCallBehavior::Sync,
 184                                                     pme_gpu_get_block_size(&pme), nullptr);
 185 }
 186
 187 //! PME initialization with atom data
 188 void pmeInitAtoms(gmx_pme_t*               pme,
 189                   StatePropagatorDataGpu*  stateGpu,
 190                   const CodePath           mode,
 191                   const CoordinatesVector& coordinates,
 192                   const ChargesVector&     charges)
 193 {
 194     const index atomCount = coordinates.size();
 195     GMX_RELEASE_ASSERT(atomCount == charges.ssize(), "Mismatch in atom data");
 196     PmeAtomComm* atc = nullptr;
 197
 198     switch (mode)
 199     {
 200         case CodePath::CPU:
 201             atc              = &(pme->atc[0]);
 202             atc->x           = coordinates;
 203             atc->coefficient = charges;
 204             gmx_pme_reinit_atoms(pme, atomCount, charges.data(), nullptr);
 205             /* With decomposition there would be more boilerplate atc code here, e.g. do_redist_pos_coeffs */
 206             break;
 207
 208         case CodePath::GPU:
 209             // TODO: Avoid use of atc in the GPU code path
 210             atc = &(pme->atc[0]);
 211             // We need to set atc->n for passing the size in the tests
 212             atc->setNumAtoms(atomCount);
 213             gmx_pme_reinit_atoms(pme, atomCount, charges.data(), nullptr);
 214
 215             stateGpu->reinit(atomCount, atomCount);
 216             stateGpu->copyCoordinatesToGpu(arrayRefFromArray(coordinates.data(), coordinates.size()),
 217                                            gmx::AtomLocality::All);
 218             pme_gpu_set_kernelparam_coordinates(pme->gpu, stateGpu->getCoordinates());
 219
 220             break;
 221
 222         default: GMX_THROW(InternalError("Test not implemented for this mode"));
 223     }
 224 }
 225
 226 //! Getting local PME real grid pointer for test I/O
 227 static real* pmeGetRealGridInternal(const gmx_pme_t* pme)
 228 {
 229     const size_t gridIndex = 0;
 230     return pme->fftgrid[gridIndex];
 231 }
 232
 233 //! Getting local PME real grid dimensions
 234 static void pmeGetRealGridSizesInternal(const gmx_pme_t* pme,
 235                                         CodePath         mode,
 236                                         IVec& gridSize,       //NOLINT(google-runtime-references)
 237                                         IVec& paddedGridSize) //NOLINT(google-runtime-references)
 238 {
 239     const size_t gridIndex = 0;
 240     IVec         gridOffsetUnused;
 241     switch (mode)
 242     {
 243         case CodePath::CPU:
 244             gmx_parallel_3dfft_real_limits(pme->pfft_setup[gridIndex], gridSize, gridOffsetUnused,
 245                                            paddedGridSize);
 246             break;
 247
 248         case CodePath::GPU:
 249             pme_gpu_get_real_grid_sizes(pme->gpu, &gridSize, &paddedGridSize);
 250             break;
 251
 252         default: GMX_THROW(InternalError("Test not implemented for this mode"));
 253     }
 254 }
 255
 256 //! Getting local PME complex grid pointer for test I/O
 257 static t_complex* pmeGetComplexGridInternal(const gmx_pme_t* pme)
 258 {
 259     const size_t gridIndex = 0;
 260     return pme->cfftgrid[gridIndex];
 261 }
 262
 263 //! Getting local PME complex grid dimensions
 264 static void pmeGetComplexGridSizesInternal(const gmx_pme_t* pme,
 265                                            IVec& gridSize,       //NOLINT(google-runtime-references)
 266                                            IVec& paddedGridSize) //NOLINT(google-runtime-references)
 267 {
 268     const size_t gridIndex = 0;
 269     IVec         gridOffsetUnused, complexOrderUnused;
 270     gmx_parallel_3dfft_complex_limits(pme->pfft_setup[gridIndex], complexOrderUnused, gridSize,
 271                                       gridOffsetUnused, paddedGridSize); // TODO: what about YZX ordering?
 272 }
 273
 274 //! Getting the PME grid memory buffer and its sizes - template definition
 275 template<typename ValueType>
 276 static void pmeGetGridAndSizesInternal(const gmx_pme_t* /*unused*/,
 277                                        CodePath /*unused*/,
 278                                        ValueType*& /*unused*/, //NOLINT(google-runtime-references)
 279                                        IVec& /*unused*/,       //NOLINT(google-runtime-references)
 280                                        IVec& /*unused*/)       //NOLINT(google-runtime-references)
 281 {
 282     GMX_THROW(InternalError("Deleted function call"));
 283     // explicitly deleting general template does not compile in clang/icc, see https://llvm.org/bugs/show_bug.cgi?id=17537
 284 }
 285
 286 //! Getting the PME real grid memory buffer and its sizes
 287 template<>
 288 void pmeGetGridAndSizesInternal<real>(const gmx_pme_t* pme, CodePath mode, real*& grid, IVec& gridSize, IVec& paddedGridSize)
 289 {
 290     grid = pmeGetRealGridInternal(pme);
 291     pmeGetRealGridSizesInternal(pme, mode, gridSize, paddedGridSize);
 292 }
 293
 294 //! Getting the PME complex grid memory buffer and its sizes
 295 template<>
 296 void pmeGetGridAndSizesInternal<t_complex>(const gmx_pme_t* pme,
 297                                            CodePath /*unused*/,
 298                                            t_complex*& grid,
 299                                            IVec&       gridSize,
 300                                            IVec&       paddedGridSize)
 301 {
 302     grid = pmeGetComplexGridInternal(pme);
 303     pmeGetComplexGridSizesInternal(pme, gridSize, paddedGridSize);
 304 }
 305
 306 //! PME spline calculation and charge spreading
 307 void pmePerformSplineAndSpread(gmx_pme_t* pme,
 308                                CodePath   mode, // TODO const qualifiers elsewhere
 309                                bool       computeSplines,
 310                                bool       spreadCharges)
 311 {
 312     GMX_RELEASE_ASSERT(pme != nullptr, "PME data is not initialized");
 313     PmeAtomComm* atc                          = &(pme->atc[0]);
 314     const size_t gridIndex                    = 0;
 315     const bool   computeSplinesForZeroCharges = true;
 316     real**       fftgrid                      = spreadCharges ? pme->fftgrid : nullptr;
 317     real*        pmegrid                      = pme->pmegrid[gridIndex].grid.grid;
 318
 319     switch (mode)
 320     {
 321         case CodePath::CPU:
 322             spread_on_grid(pme, atc, &pme->pmegrid[gridIndex], computeSplines, spreadCharges,
 323                            fftgrid != nullptr ? fftgrid[gridIndex] : nullptr,
 324                            computeSplinesForZeroCharges, gridIndex);
 325             if (spreadCharges && !pme->bUseThreads)
 326             {
 327                 wrap_periodic_pmegrid(pme, pmegrid);
 328                 copy_pmegrid_to_fftgrid(
 329                         pme, pmegrid, fftgrid != nullptr ? fftgrid[gridIndex] : nullptr, gridIndex);
 330             }
 331             break;
 332
 333 /* The compiler will complain about passing fftgrid (converting double ** to float **) if using
 334  * double precision. GPUs are not used with double precision anyhow. */
 335 #if !GMX_DOUBLE
 336         case CodePath::GPU:
 337         {
 338             const real lambdaQ = 1.0;
 339             // no synchronization needed as x is transferred in the PME stream
 340             GpuEventSynchronizer* xReadyOnDevice = nullptr;
 341             pme_gpu_spread(pme->gpu, xReadyOnDevice, fftgrid, computeSplines, spreadCharges, lambdaQ);
 342         }
 343         break;
 344 #endif
 345
 346         default: GMX_THROW(InternalError("Test not implemented for this mode"));
 347     }
 348 }
 349
 350 //! Getting the internal spline data buffer pointer
 351 static real* pmeGetSplineDataInternal(const gmx_pme_t* pme, PmeSplineDataType type, int dimIndex)
 352 {
 353     GMX_ASSERT((0 <= dimIndex) && (dimIndex < DIM), "Invalid dimension index");
 354     const PmeAtomComm* atc          = &(pme->atc[0]);
 355     const size_t       threadIndex  = 0;
 356     real*              splineBuffer = nullptr;
 357     switch (type)
 358     {
 359         case PmeSplineDataType::Values:
 360             splineBuffer = atc->spline[threadIndex].theta.coefficients[dimIndex];
 361             break;
 362
 363         case PmeSplineDataType::Derivatives:
 364             splineBuffer = atc->spline[threadIndex].dtheta.coefficients[dimIndex];
 365             break;
 366
 367         default: GMX_THROW(InternalError("Unknown spline data type"));
 368     }
 369     return splineBuffer;
 370 }
 371
 372 //! PME solving
 373 void pmePerformSolve(const gmx_pme_t*  pme,
 374                      CodePath          mode,
 375                      PmeSolveAlgorithm method,
 376                      real              cellVolume,
 377                      GridOrdering      gridOrdering,
 378                      bool              computeEnergyAndVirial)
 379 {
 380     t_complex*   h_grid              = pmeGetComplexGridInternal(pme);
 381     const bool   useLorentzBerthelot = false;
 382     const size_t threadIndex         = 0;
 383     const size_t gridIndex           = 0;
 384     switch (mode)
 385     {
 386         case CodePath::CPU:
 387             if (gridOrdering != GridOrdering::YZX)
 388             {
 389                 GMX_THROW(InternalError("Test not implemented for this mode"));
 390             }
 391             switch (method)
 392             {
 393                 case PmeSolveAlgorithm::Coulomb:
 394                     solve_pme_yzx(pme, h_grid, cellVolume, computeEnergyAndVirial, pme->nthread, threadIndex);
 395                     break;
 396
 397                 case PmeSolveAlgorithm::LennardJones:
 398                     solve_pme_lj_yzx(pme, &h_grid, useLorentzBerthelot, cellVolume,
 399                                      computeEnergyAndVirial, pme->nthread, threadIndex);
 400                     break;
 401
 402                 default: GMX_THROW(InternalError("Test not implemented for this mode"));
 403             }
 404             break;
 405
 406         case CodePath::GPU:
 407             switch (method)
 408             {
 409                 case PmeSolveAlgorithm::Coulomb:
 410                     pme_gpu_solve(pme->gpu, gridIndex, h_grid, gridOrdering, computeEnergyAndVirial);
 411                     break;
 412
 413                 default: GMX_THROW(InternalError("Test not implemented for this mode"));
 414             }
 415             break;
 416
 417         default: GMX_THROW(InternalError("Test not implemented for this mode"));
 418     }
 419 }
 420
 421 //! PME force gathering
 422 void pmePerformGather(gmx_pme_t* pme, CodePath mode, ForcesVector& forces)
 423 {
 424     PmeAtomComm* atc       = &(pme->atc[0]);
 425     const index  atomCount = atc->numAtoms();
 426     GMX_RELEASE_ASSERT(forces.ssize() == atomCount, "Invalid force buffer size");
 427     const real   scale       = 1.0;
 428     const size_t threadIndex = 0;
 429     const size_t gridIndex   = 0;
 430     real*        pmegrid     = pme->pmegrid[gridIndex].grid.grid;
 431     real**       fftgrid     = pme->fftgrid;
 432
 433     switch (mode)
 434     {
 435         case CodePath::CPU:
 436             atc->f = forces;
 437             if (atc->nthread == 1)
 438             {
 439                 // something which is normally done in serial spline computation (make_thread_local_ind())
 440                 atc->spline[threadIndex].n = atomCount;
 441             }
 442             copy_fftgrid_to_pmegrid(pme, fftgrid[gridIndex], pmegrid, gridIndex, pme->nthread, threadIndex);
 443             unwrap_periodic_pmegrid(pme, pmegrid);
 444             gather_f_bsplines(pme, pmegrid, true, atc, &atc->spline[threadIndex], scale);
 445             break;
 446
 447 /* The compiler will complain about passing fftgrid (converting double ** to float **) if using
 448  * double precision. GPUs are not used with double precision anyhow. */
 449 #if !GMX_DOUBLE
 450         case CodePath::GPU:
 451         {
 452             // Variable initialization needs a non-switch scope
 453             const bool computeEnergyAndVirial = false;
 454             const real lambdaQ                = 1.0;
 455             PmeOutput  output = pme_gpu_getOutput(*pme, computeEnergyAndVirial, lambdaQ);
 456             GMX_ASSERT(forces.size() == output.forces_.size(),
 457                        "Size of force buffers did not match");
 458             pme_gpu_gather(pme->gpu, fftgrid, lambdaQ);
 459             std::copy(std::begin(output.forces_), std::end(output.forces_), std::begin(forces));
 460         }
 461         break;
 462 #endif
 463
 464         default: GMX_THROW(InternalError("Test not implemented for this mode"));
 465     }
 466 }
 467
 468 //! PME test finalization before fetching the outputs
 469 void pmeFinalizeTest(const gmx_pme_t* pme, CodePath mode)
 470 {
 471     switch (mode)
 472     {
 473         case CodePath::CPU: break;
 474
 475         case CodePath::GPU: pme_gpu_synchronize(pme->gpu); break;
 476
 477         default: GMX_THROW(InternalError("Test not implemented for this mode"));
 478     }
 479 }
 480
 481 //! A binary enum for spline data layout transformation
 482 enum class PmeLayoutTransform
 483 {
 484     GpuToHost,
 485     HostToGpu
 486 };
 487
 488 /*! \brief Gets a unique index to an element in a spline parameter buffer.
 489  *
 490  * These theta/dtheta buffers are laid out for GPU spread/gather
 491  * kernels. The index is wrt the execution block, in range(0,
 492  * atomsPerBlock * order * DIM).
 493  *
 494  * This is a wrapper, only used in unit tests.
 495  * \param[in] order            PME order
 496  * \param[in] splineIndex      Spline contribution index (from 0 to \p order - 1)
 497  * \param[in] dimIndex         Dimension index (from 0 to 2)
 498  * \param[in] atomIndex        Atom index wrt the block.
 499  * \param[in] atomsPerWarp     Number of atoms processed by a warp.
 500  *
 501  * \returns Index into theta or dtheta array using GPU layout.
 502  */
 503 static int getSplineParamFullIndex(int order, int splineIndex, int dimIndex, int atomIndex, int atomsPerWarp)
 504 {
 505     if (order != c_pmeGpuOrder)
 506     {
 507         throw order;
 508     }
 509     constexpr int fixedOrder = c_pmeGpuOrder;
 510     GMX_UNUSED_VALUE(fixedOrder);
 511
 512     const int atomWarpIndex = atomIndex % atomsPerWarp;
 513     const int warpIndex     = atomIndex / atomsPerWarp;
 514     int       indexBase, result;
 515     switch (atomsPerWarp)
 516     {
 517         case 1:
 518             indexBase = getSplineParamIndexBase<fixedOrder, 1>(warpIndex, atomWarpIndex);
 519             result    = getSplineParamIndex<fixedOrder, 1>(indexBase, dimIndex, splineIndex);
 520             break;
 521
 522         case 2:
 523             indexBase = getSplineParamIndexBase<fixedOrder, 2>(warpIndex, atomWarpIndex);
 524             result    = getSplineParamIndex<fixedOrder, 2>(indexBase, dimIndex, splineIndex);
 525             break;
 526
 527         case 4:
 528             indexBase = getSplineParamIndexBase<fixedOrder, 4>(warpIndex, atomWarpIndex);
 529             result    = getSplineParamIndex<fixedOrder, 4>(indexBase, dimIndex, splineIndex);
 530             break;
 531
 532         case 8:
 533             indexBase = getSplineParamIndexBase<fixedOrder, 8>(warpIndex, atomWarpIndex);
 534             result    = getSplineParamIndex<fixedOrder, 8>(indexBase, dimIndex, splineIndex);
 535             break;
 536
 537         default:
 538             GMX_THROW(NotImplementedError(
 539                     formatString("Test function call not unrolled for atomsPerWarp = %d in "
 540                                  "getSplineParamFullIndex",
 541                                  atomsPerWarp)));
 542     }
 543     return result;
 544 }
 545
 546 /*!\brief Return the number of atoms per warp */
 547 static int pme_gpu_get_atoms_per_warp(const PmeGpu* pmeGpu)
 548 {
 549     const int order = pmeGpu->common->pme_order;
 550     const int threadsPerAtom =
 551             (pmeGpu->settings.threadsPerAtom == ThreadsPerAtom::Order ? order : order * order);
 552     return pmeGpu->programHandle_->warpSize() / threadsPerAtom;
 553 }
 554
 555 /*! \brief Rearranges the atom spline data between the GPU and host layouts.
 556  * Only used for test purposes so far, likely to be horribly slow.
 557  *
 558  * \param[in]  pmeGpu     The PME GPU structure.
 559  * \param[out] atc        The PME CPU atom data structure (with a single-threaded layout).
 560  * \param[in]  type       The spline data type (values or derivatives).
 561  * \param[in]  dimIndex   Dimension index.
 562  * \param[in]  transform  Layout transform type
 563  */
 564 static void pme_gpu_transform_spline_atom_data(const PmeGpu*      pmeGpu,
 565                                                const PmeAtomComm* atc,
 566                                                PmeSplineDataType  type,
 567                                                int                dimIndex,
 568                                                PmeLayoutTransform transform)
 569 {
 570     // The GPU atom spline data is laid out in a different way currently than the CPU one.
 571     // This function converts the data from GPU to CPU layout (in the host memory).
 572     // It is only intended for testing purposes so far.
 573     // Ideally we should use similar layouts on CPU and GPU if we care about mixed modes and their
 574     // performance (e.g. spreading on GPU, gathering on CPU).
 575     GMX_RELEASE_ASSERT(atc->nthread == 1, "Only the serial PME data layout is supported");
 576     const uintmax_t threadIndex  = 0;
 577     const auto      atomCount    = atc->numAtoms();
 578     const auto      atomsPerWarp = pme_gpu_get_atoms_per_warp(pmeGpu);
 579     const auto      pmeOrder     = pmeGpu->common->pme_order;
 580     GMX_ASSERT(pmeOrder == c_pmeGpuOrder, "Only PME order 4 is implemented");
 581
 582     real*  cpuSplineBuffer;
 583     float* h_splineBuffer;
 584     switch (type)
 585     {
 586         case PmeSplineDataType::Values:
 587             cpuSplineBuffer = atc->spline[threadIndex].theta.coefficients[dimIndex];
 588             h_splineBuffer  = pmeGpu->staging.h_theta;
 589             break;
 590
 591         case PmeSplineDataType::Derivatives:
 592             cpuSplineBuffer = atc->spline[threadIndex].dtheta.coefficients[dimIndex];
 593             h_splineBuffer  = pmeGpu->staging.h_dtheta;
 594             break;
 595
 596         default: GMX_THROW(InternalError("Unknown spline data type"));
 597     }
 598
 599     for (auto atomIndex = 0; atomIndex < atomCount; atomIndex++)
 600     {
 601         for (auto orderIndex = 0; orderIndex < pmeOrder; orderIndex++)
 602         {
 603             const auto gpuValueIndex =
 604                     getSplineParamFullIndex(pmeOrder, orderIndex, dimIndex, atomIndex, atomsPerWarp);
 605             const auto cpuValueIndex = atomIndex * pmeOrder + orderIndex;
 606             GMX_ASSERT(cpuValueIndex < atomCount * pmeOrder,
 607                        "Atom spline data index out of bounds (while transforming GPU data layout "
 608                        "for host)");
 609             switch (transform)
 610             {
 611                 case PmeLayoutTransform::GpuToHost:
 612                     cpuSplineBuffer[cpuValueIndex] = h_splineBuffer[gpuValueIndex];
 613                     break;
 614
 615                 case PmeLayoutTransform::HostToGpu:
 616                     h_splineBuffer[gpuValueIndex] = cpuSplineBuffer[cpuValueIndex];
 617                     break;
 618
 619                 default: GMX_THROW(InternalError("Unknown layout transform"));
 620             }
 621         }
 622     }
 623 }
 624
 625 //! Setting atom spline values/derivatives to be used in spread/gather
 626 void pmeSetSplineData(const gmx_pme_t*             pme,
 627                       CodePath                     mode,
 628                       const SplineParamsDimVector& splineValues,
 629                       PmeSplineDataType            type,
 630                       int                          dimIndex)
 631 {
 632     const PmeAtomComm* atc       = &(pme->atc[0]);
 633     const index        atomCount = atc->numAtoms();
 634     const index        pmeOrder  = pme->pme_order;
 635     const index        dimSize   = pmeOrder * atomCount;
 636     GMX_RELEASE_ASSERT(dimSize == splineValues.ssize(), "Mismatch in spline data");
 637     real* splineBuffer = pmeGetSplineDataInternal(pme, type, dimIndex);
 638
 639     switch (mode)
 640     {
 641         case CodePath::CPU:
 642             std::copy(splineValues.begin(), splineValues.end(), splineBuffer);
 643             break;
 644
 645         case CodePath::GPU:
 646             std::copy(splineValues.begin(), splineValues.end(), splineBuffer);
 647             pme_gpu_transform_spline_atom_data(pme->gpu, atc, type, dimIndex, PmeLayoutTransform::HostToGpu);
 648             break;
 649
 650         default: GMX_THROW(InternalError("Test not implemented for this mode"));
 651     }
 652 }
 653
 654 //! Setting gridline indices to be used in spread/gather
 655 void pmeSetGridLineIndices(gmx_pme_t* pme, CodePath mode, const GridLineIndicesVector& gridLineIndices)
 656 {
 657     PmeAtomComm* atc       = &(pme->atc[0]);
 658     const index  atomCount = atc->numAtoms();
 659     GMX_RELEASE_ASSERT(atomCount == gridLineIndices.ssize(), "Mismatch in gridline indices size");
 660
 661     IVec paddedGridSizeUnused, gridSize(0, 0, 0);
 662     pmeGetRealGridSizesInternal(pme, mode, gridSize, paddedGridSizeUnused);
 663
 664     for (const auto& index : gridLineIndices)
 665     {
 666         for (int i = 0; i < DIM; i++)
 667         {
 668             GMX_RELEASE_ASSERT((0 <= index[i]) && (index[i] < gridSize[i]),
 669                                "Invalid gridline index");
 670         }
 671     }
 672
 673     switch (mode)
 674     {
 675         case CodePath::GPU:
 676             memcpy(pme_gpu_staging(pme->gpu).h_gridlineIndices, gridLineIndices.data(),
 677                    atomCount * sizeof(gridLineIndices[0]));
 678             break;
 679
 680         case CodePath::CPU:
 681             atc->idx.resize(gridLineIndices.size());
 682             std::copy(gridLineIndices.begin(), gridLineIndices.end(), atc->idx.begin());
 683             break;
 684         default: GMX_THROW(InternalError("Test not implemented for this mode"));
 685     }
 686 }
 687
 688 //! Getting plain index into the complex 3d grid
 689 inline size_t pmeGetGridPlainIndexInternal(const IVec& index, const IVec& paddedGridSize, GridOrdering gridOrdering)
 690 {
 691     size_t result;
 692     switch (gridOrdering)
 693     {
 694         case GridOrdering::YZX:
 695             result = (index[YY] * paddedGridSize[ZZ] + index[ZZ]) * paddedGridSize[XX] + index[XX];
 696             break;
 697
 698         case GridOrdering::XYZ:
 699             result = (index[XX] * paddedGridSize[YY] + index[YY]) * paddedGridSize[ZZ] + index[ZZ];
 700             break;
 701
 702         default: GMX_THROW(InternalError("Test not implemented for this mode"));
 703     }
 704     return result;
 705 }
 706
 707 //! Setting real or complex grid
 708 template<typename ValueType>
 709 static void pmeSetGridInternal(const gmx_pme_t*                        pme,
 710                                CodePath                                mode,
 711                                GridOrdering                            gridOrdering,
 712                                const SparseGridValuesInput<ValueType>& gridValues)
 713 {
 714     IVec       gridSize(0, 0, 0), paddedGridSize(0, 0, 0);
 715     ValueType* grid;
 716     pmeGetGridAndSizesInternal<ValueType>(pme, mode, grid, gridSize, paddedGridSize);
 717
 718     switch (mode)
 719     {
 720         case CodePath::GPU: // intentional absence of break, the grid will be copied from the host buffer in testing mode
 721         case CodePath::CPU:
 722             std::memset(grid, 0,
 723                         paddedGridSize[XX] * paddedGridSize[YY] * paddedGridSize[ZZ] * sizeof(ValueType));
 724             for (const auto& gridValue : gridValues)
 725             {
 726                 for (int i = 0; i < DIM; i++)
 727                 {
 728                     GMX_RELEASE_ASSERT((0 <= gridValue.first[i]) && (gridValue.first[i] < gridSize[i]),
 729                                        "Invalid grid value index");
 730                 }
 731                 const size_t gridValueIndex =
 732                         pmeGetGridPlainIndexInternal(gridValue.first, paddedGridSize, gridOrdering);
 733                 grid[gridValueIndex] = gridValue.second;
 734             }
 735             break;
 736
 737         default: GMX_THROW(InternalError("Test not implemented for this mode"));
 738     }
 739 }
 740
 741 //! Setting real grid to be used in gather
 742 void pmeSetRealGrid(const gmx_pme_t* pme, CodePath mode, const SparseRealGridValuesInput& gridValues)
 743 {
 744     pmeSetGridInternal<real>(pme, mode, GridOrdering::XYZ, gridValues);
 745 }
 746
 747 //! Setting complex grid to be used in solve
 748 void pmeSetComplexGrid(const gmx_pme_t*                    pme,
 749                        CodePath                            mode,
 750                        GridOrdering                        gridOrdering,
 751                        const SparseComplexGridValuesInput& gridValues)
 752 {
 753     pmeSetGridInternal<t_complex>(pme, mode, gridOrdering, gridValues);
 754 }
 755
 756 //! Getting the single dimension's spline values or derivatives
 757 SplineParamsDimVector pmeGetSplineData(const gmx_pme_t* pme, CodePath mode, PmeSplineDataType type, int dimIndex)
 758 {
 759     GMX_RELEASE_ASSERT(pme != nullptr, "PME data is not initialized");
 760     const PmeAtomComm* atc       = &(pme->atc[0]);
 761     const size_t       atomCount = atc->numAtoms();
 762     const size_t       pmeOrder  = pme->pme_order;
 763     const size_t       dimSize   = pmeOrder * atomCount;
 764
 765     real*                 sourceBuffer = pmeGetSplineDataInternal(pme, type, dimIndex);
 766     SplineParamsDimVector result;
 767     switch (mode)
 768     {
 769         case CodePath::GPU:
 770             pme_gpu_transform_spline_atom_data(pme->gpu, atc, type, dimIndex, PmeLayoutTransform::GpuToHost);
 771             result = arrayRefFromArray(sourceBuffer, dimSize);
 772             break;
 773
 774         case CodePath::CPU: result = arrayRefFromArray(sourceBuffer, dimSize); break;
 775
 776         default: GMX_THROW(InternalError("Test not implemented for this mode"));
 777     }
 778     return result;
 779 }
 780
 781 //! Getting the gridline indices
 782 GridLineIndicesVector pmeGetGridlineIndices(const gmx_pme_t* pme, CodePath mode)
 783 {
 784     GMX_RELEASE_ASSERT(pme != nullptr, "PME data is not initialized");
 785     const PmeAtomComm* atc       = &(pme->atc[0]);
 786     const size_t       atomCount = atc->numAtoms();
 787
 788     GridLineIndicesVector gridLineIndices;
 789     switch (mode)
 790     {
 791         case CodePath::GPU:
 792             gridLineIndices = arrayRefFromArray(
 793                     reinterpret_cast<IVec*>(pme_gpu_staging(pme->gpu).h_gridlineIndices), atomCount);
 794             break;
 795
 796         case CodePath::CPU: gridLineIndices = atc->idx; break;
 797
 798         default: GMX_THROW(InternalError("Test not implemented for this mode"));
 799     }
 800     return gridLineIndices;
 801 }
 802
 803 //! Getting real or complex grid - only non zero values
 804 template<typename ValueType>
 805 static SparseGridValuesOutput<ValueType> pmeGetGridInternal(const gmx_pme_t* pme,
 806                                                             CodePath         mode,
 807                                                             GridOrdering     gridOrdering)
 808 {
 809     IVec       gridSize(0, 0, 0), paddedGridSize(0, 0, 0);
 810     ValueType* grid;
 811     pmeGetGridAndSizesInternal<ValueType>(pme, mode, grid, gridSize, paddedGridSize);
 812     SparseGridValuesOutput<ValueType> gridValues;
 813     switch (mode)
 814     {
 815         case CodePath::GPU: // intentional absence of break
 816         case CodePath::CPU:
 817             gridValues.clear();
 818             for (int ix = 0; ix < gridSize[XX]; ix++)
 819             {
 820                 for (int iy = 0; iy < gridSize[YY]; iy++)
 821                 {
 822                     for (int iz = 0; iz < gridSize[ZZ]; iz++)
 823                     {
 824                         IVec         temp(ix, iy, iz);
 825                         const size_t gridValueIndex =
 826                                 pmeGetGridPlainIndexInternal(temp, paddedGridSize, gridOrdering);
 827                         const ValueType value = grid[gridValueIndex];
 828                         if (value != ValueType{})
 829                         {
 830                             auto key        = formatString("Cell %d %d %d", ix, iy, iz);
 831                             gridValues[key] = value;
 832                         }
 833                     }
 834                 }
 835             }
 836             break;
 837
 838         default: GMX_THROW(InternalError("Test not implemented for this mode"));
 839     }
 840     return gridValues;
 841 }
 842
 843 //! Getting the real grid (spreading output of pmePerformSplineAndSpread())
 844 SparseRealGridValuesOutput pmeGetRealGrid(const gmx_pme_t* pme, CodePath mode)
 845 {
 846     return pmeGetGridInternal<real>(pme, mode, GridOrdering::XYZ);
 847 }
 848
 849 //! Getting the complex grid output of pmePerformSolve()
 850 SparseComplexGridValuesOutput pmeGetComplexGrid(const gmx_pme_t* pme, CodePath mode, GridOrdering gridOrdering)
 851 {
 852     return pmeGetGridInternal<t_complex>(pme, mode, gridOrdering);
 853 }
 854
 855 //! Getting the reciprocal energy and virial
 856 PmeOutput pmeGetReciprocalEnergyAndVirial(const gmx_pme_t* pme, CodePath mode, PmeSolveAlgorithm method)
 857 {
 858     PmeOutput  output;
 859     const real lambdaQ = 1.0;
 860     switch (mode)
 861     {
 862         case CodePath::CPU:
 863             switch (method)
 864             {
 865                 case PmeSolveAlgorithm::Coulomb:
 866                     get_pme_ener_vir_q(pme->solve_work, pme->nthread, &output);
 867                     break;
 868
 869                 case PmeSolveAlgorithm::LennardJones:
 870                     get_pme_ener_vir_lj(pme->solve_work, pme->nthread, &output);
 871                     break;
 872
 873                 default: GMX_THROW(InternalError("Test not implemented for this mode"));
 874             }
 875             break;
 876         case CodePath::GPU:
 877             switch (method)
 878             {
 879                 case PmeSolveAlgorithm::Coulomb:
 880                     pme_gpu_getEnergyAndVirial(*pme, lambdaQ, &output);
 881                     break;
 882
 883                 default: GMX_THROW(InternalError("Test not implemented for this mode"));
 884             }
 885             break;
 886
 887         default: GMX_THROW(InternalError("Test not implemented for this mode"));
 888     }
 889     return output;
 890 }
 891
 892 } // namespace test
 893 } // namespace gmx