CudaKernels.cpp 69.2 KB
Newer Older
1
2
3
4
5
6
7
8
/* -------------------------------------------------------------------------- *
 *                                   OpenMM                                   *
 * -------------------------------------------------------------------------- *
 * This is part of the OpenMM molecular simulation toolkit originating from   *
 * Simbios, the NIH National Center for Physics-Based Simulation of           *
 * Biological Structures at Stanford, funded under the NIH Roadmap for        *
 * Medical Research, grant U54 GM072970. See https://simtk.org.               *
 *                                                                            *
9
 * Portions copyright (c) 2008-2025 Stanford University and the Authors.      *
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
 * Authors: Peter Eastman                                                     *
 * Contributors:                                                              *
 *                                                                            *
 * This program is free software: you can redistribute it and/or modify       *
 * it under the terms of the GNU Lesser General Public License as published   *
 * by the Free Software Foundation, either version 3 of the License, or       *
 * (at your option) any later version.                                        *
 *                                                                            *
 * This program is distributed in the hope that it will be useful,            *
 * but WITHOUT ANY WARRANTY; without even the implied warranty of             *
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the              *
 * GNU Lesser General Public License for more details.                        *
 *                                                                            *
 * You should have received a copy of the GNU Lesser General Public License   *
 * along with this program.  If not, see <http://www.gnu.org/licenses/>.      *
 * -------------------------------------------------------------------------- */

#include "CudaKernels.h"
#include "CudaForceInfo.h"
#include "openmm/Context.h"
#include "openmm/internal/ContextImpl.h"
#include "openmm/internal/NonbondedForceImpl.h"
32
#include "openmm/common/ContextSelector.h"
33
#include "CommonKernelSources.h"
34
35
36
#include "CudaBondedUtilities.h"
#include "CudaExpressionUtilities.h"
#include "CudaIntegrationUtilities.h"
37
#include "CudaNonbondedUtilities.h"
38
#include "CudaKernelSources.h"
39
40
#include "SimTKOpenMMRealType.h"
#include "SimTKOpenMMUtilities.h"
41
#include <algorithm>
42
#include <cmath>
43
#include <iterator>
44
#include <set>
Andy Simmonett's avatar
Andy Simmonett committed
45
#include <assert.h>
46
47
48
49

using namespace OpenMM;
using namespace std;

50
51
52
53
54
55
56
#define CHECK_RESULT(result, prefix) \
    if (result != CUDA_SUCCESS) { \
        std::stringstream m; \
        m<<prefix<<": "<<CudaContext::getErrorString(result)<<" ("<<result<<")"<<" at "<<__FILE__<<":"<<__LINE__; \
        throw OpenMMException(m.str());\
    }

57
58
59
60
void CudaCalcForcesAndEnergyKernel::initialize(const System& system) {
}

void CudaCalcForcesAndEnergyKernel::beginComputation(ContextImpl& context, bool includeForces, bool includeEnergy, int groups) {
61
    cu.setForcesValid(true);
62
    ContextSelector selector(cu);
63
    cu.clearAutoclearBuffers();
peastman's avatar
peastman committed
64
65
    for (auto computation : cu.getPreComputations())
        computation->computeForceAndEnergy(includeForces, includeEnergy, groups);
66
    CudaNonbondedUtilities& nb = cu.getNonbondedUtilities();
67
    cu.setComputeForceCount(cu.getComputeForceCount()+1);
68
    nb.prepareInteractions(groups);
69
    map<string, double>& derivs = cu.getEnergyParamDerivWorkspace();
peastman's avatar
peastman committed
70
71
    for (auto& param : context.getParameters())
        derivs[param.first] = 0;
72
73
}

74
double CudaCalcForcesAndEnergyKernel::finishComputation(ContextImpl& context, bool includeForces, bool includeEnergy, int groups, bool& valid) {
75
    ContextSelector selector(cu);
76
    cu.getBondedUtilities().computeInteractions(groups);
77
    cu.getNonbondedUtilities().computeInteractions(groups, includeForces, includeEnergy);
78
    double sum = 0.0;
peastman's avatar
peastman committed
79
80
    for (auto computation : cu.getPostComputations())
        sum += computation->computeForceAndEnergy(includeForces, includeEnergy, groups);
81
    cu.getIntegrationUtilities().distributeForcesFromVirtualSites();
Peter Eastman's avatar
Peter Eastman committed
82
83
    if (includeEnergy)
        sum += cu.reduceEnergy();
84
85
    if (!cu.getForcesValid())
        valid = false;
86
87
88
    return sum;
}

89
class CudaCalcNonbondedForceKernel::ForceInfo : public CudaForceInfo {
90
public:
91
    ForceInfo(const NonbondedForce& force) : force(force) {
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
        particleOffset.resize(force.getNumParticles(), -1);
        exceptionOffset.resize(force.getNumExceptions(), -1);
        for (int i = 0; i < force.getNumParticleParameterOffsets(); i++) {
            string parameter;
            int particleIndex;
            double chargeScale, sigmaScale, epsilonScale;
            force.getParticleParameterOffset(i, parameter, particleIndex, chargeScale, sigmaScale, epsilonScale);
            particleOffset[particleIndex] = i;
        }
        for (int i = 0; i < force.getNumExceptionParameterOffsets(); i++) {
            string parameter;
            int exceptionIndex;
            double chargeProdScale, sigmaScale, epsilonScale;
            force.getExceptionParameterOffset(i, parameter, exceptionIndex, chargeProdScale, sigmaScale, epsilonScale);
            exceptionOffset[exceptionIndex] = i;
        }
108
109
    }
    bool areParticlesIdentical(int particle1, int particle2) {
110
111
112
113
114
115
116
117
118
119
120
        if (particleOffset[particle1] != -1 || particleOffset[particle2] != -1) {
            if (particleOffset[particle1] == -1 || particleOffset[particle2] == -1)
                return false;
            string parameter1, parameter2;
            int particleIndex1, particleIndex2;
            double chargeScale1, sigmaScale1, epsilonScale1, chargeScale2, sigmaScale2, epsilonScale2;
            force.getParticleParameterOffset(particleOffset[particle1], parameter1, particleIndex1, chargeScale1, sigmaScale1, epsilonScale1);
            force.getParticleParameterOffset(particleOffset[particle2], parameter2, particleIndex2, chargeScale2, sigmaScale2, epsilonScale2);
            if (parameter1 != parameter2 || chargeScale1 != chargeScale2 || sigmaScale1 != sigmaScale2 || epsilonScale1 != epsilonScale2)
                return false;
        }
121
122
123
124
        double charge1, charge2, sigma1, sigma2, epsilon1, epsilon2;
        force.getParticleParameters(particle1, charge1, sigma1, epsilon1);
        force.getParticleParameters(particle2, charge2, sigma2, epsilon2);
        return (charge1 == charge2 && sigma1 == sigma2 && epsilon1 == epsilon2);
125
126
    }
    int getNumParticleGroups() {
127
        return force.getNumExceptions();
128
129
130
    }
    void getParticlesInGroup(int index, vector<int>& particles) {
        int particle1, particle2;
131
132
        double chargeProd, sigma, epsilon;
        force.getExceptionParameters(index, particle1, particle2, chargeProd, sigma, epsilon);
133
134
135
136
137
        particles.resize(2);
        particles[0] = particle1;
        particles[1] = particle2;
    }
    bool areGroupsIdentical(int group1, int group2) {
138
139
140
141
142
143
144
145
146
147
148
        if (exceptionOffset[group1] != -1 || exceptionOffset[group2] != -1) {
            if (exceptionOffset[group1] == -1 || exceptionOffset[group2] == -1)
                return false;
            string parameter1, parameter2;
            int exceptionIndex1, exceptionIndex2;
            double chargeProdScale1, sigmaScale1, epsilonScale1, chargeProdScale2, sigmaScale2, epsilonScale2;
            force.getExceptionParameterOffset(exceptionOffset[group1], parameter1, exceptionIndex1, chargeProdScale1, sigmaScale1, epsilonScale1);
            force.getExceptionParameterOffset(exceptionOffset[group2], parameter2, exceptionIndex2, chargeProdScale2, sigmaScale2, epsilonScale2);
            if (parameter1 != parameter2 || chargeProdScale1 != chargeProdScale2 || sigmaScale1 != sigmaScale2 || epsilonScale1 != epsilonScale2)
                return false;
        }
149
        int particle1, particle2;
150
151
152
153
        double chargeProd1, chargeProd2, sigma1, sigma2, epsilon1, epsilon2;
        force.getExceptionParameters(group1, particle1, particle2, chargeProd1, sigma1, epsilon1);
        force.getExceptionParameters(group2, particle1, particle2, chargeProd2, sigma2, epsilon2);
        return (chargeProd1 == chargeProd2 && sigma1 == sigma2 && epsilon1 == epsilon2);
154
155
    }
private:
156
    const NonbondedForce& force;
157
    vector<int> particleOffset, exceptionOffset;
158
159
};

160
class CudaCalcNonbondedForceKernel::PmeIO : public CalcPmeReciprocalForceKernel::IO {
161
public:
162
163
    PmeIO(CudaContext& cu, CUfunction addForcesKernel) : cu(cu), addForcesKernel(addForcesKernel) {
        forceTemp.initialize<float4>(cu, cu.getNumAtoms(), "PmeForce");
164
    }
165
    float* getPosq() {
166
        ContextSelector selector(cu);
167
168
        cu.getPosq().download(posq);
        return (float*) &posq[0];
169
    }
170
171
172
173
    void setForce(float* force) {
        forceTemp.upload(force);
        void* args[] = {&forceTemp.getDevicePointer(), &cu.getForce().getDevicePointer()};
        cu.executeKernel(addForcesKernel, args, cu.getNumAtoms());
174
175
    }
private:
176
177
178
179
    CudaContext& cu;
    vector<float4> posq;
    CudaArray forceTemp;
    CUfunction addForcesKernel;
180
181
};

182
183
184
class CudaCalcNonbondedForceKernel::PmePreComputation : public CudaContext::ForcePreComputation {
public:
    PmePreComputation(CudaContext& cu, Kernel& pme, CalcPmeReciprocalForceKernel::IO& io) : cu(cu), pme(pme), io(io) {
185
    }
186
187
188
    void computeForceAndEnergy(bool includeForces, bool includeEnergy, int groups) {
        Vec3 boxVectors[3] = {Vec3(cu.getPeriodicBoxSize().x, 0, 0), Vec3(0, cu.getPeriodicBoxSize().y, 0), Vec3(0, 0, cu.getPeriodicBoxSize().z)};
        pme.getAs<CalcPmeReciprocalForceKernel>().beginComputation(io, boxVectors, includeEnergy);
189
    }
190
191
192
193
194
private:
    CudaContext& cu;
    Kernel pme;
    CalcPmeReciprocalForceKernel::IO& io;
};
195

196
class CudaCalcNonbondedForceKernel::PmePostComputation : public CudaContext::ForcePostComputation {
197
public:
198
    PmePostComputation(Kernel& pme, CalcPmeReciprocalForceKernel::IO& io) : pme(pme), io(io) {
199
    }
200
201
    double computeForceAndEnergy(bool includeForces, bool includeEnergy, int groups) {
        return pme.getAs<CalcPmeReciprocalForceKernel>().finishComputation(io);
202
203
    }
private:
204
205
    Kernel pme;
    CalcPmeReciprocalForceKernel::IO& io;
206
207
};

208
209
210
class CudaCalcNonbondedForceKernel::SyncStreamPreComputation : public CudaContext::ForcePreComputation {
public:
    SyncStreamPreComputation(CudaContext& cu, CUstream stream, CUevent event, int forceGroup) : cu(cu), stream(stream), event(event), forceGroup(forceGroup) {
211
    }
212
213
214
215
216
    void computeForceAndEnergy(bool includeForces, bool includeEnergy, int groups) {
        if ((groups&(1<<forceGroup)) != 0) {
            cuEventRecord(event, cu.getCurrentStream());
            cuStreamWaitEvent(stream, event, 0);
        }
217
    }
218
219
220
221
222
223
private:
    CudaContext& cu;
    CUstream stream;
    CUevent event;
    int forceGroup;
};
224

225
class CudaCalcNonbondedForceKernel::SyncStreamPostComputation : public CudaContext::ForcePostComputation {
226
public:
227
228
    SyncStreamPostComputation(CudaContext& cu, CUevent event, CUfunction addEnergyKernel, CudaArray& pmeEnergyBuffer, int forceGroup) : cu(cu), event(event),
            addEnergyKernel(addEnergyKernel), pmeEnergyBuffer(pmeEnergyBuffer), forceGroup(forceGroup) {
229
    }
230
231
232
233
234
235
236
237
238
239
    double computeForceAndEnergy(bool includeForces, bool includeEnergy, int groups) {
        if ((groups&(1<<forceGroup)) != 0) {
            cuStreamWaitEvent(cu.getCurrentStream(), event, 0);
            if (includeEnergy) {
                int bufferSize = pmeEnergyBuffer.getSize();
                void* args[] = {&pmeEnergyBuffer.getDevicePointer(), &cu.getEnergyBuffer().getDevicePointer(), &bufferSize};
                cu.executeKernel(addEnergyKernel, args, bufferSize);
            }
        }
        return 0.0;
240
241
    }
private:
242
243
244
245
246
    CudaContext& cu;
    CUevent event;
    CUfunction addEnergyKernel;
    CudaArray& pmeEnergyBuffer;
    int forceGroup;
247
248
};

249
CudaCalcNonbondedForceKernel::~CudaCalcNonbondedForceKernel() {
250
    ContextSelector selector(cu);
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
    if (sort != NULL)
        delete sort;
    if (fft != NULL)
        delete fft;
    if (dispersionFft != NULL)
        delete dispersionFft;
    if (pmeio != NULL)
        delete pmeio;
    if (hasInitializedFFT) {
        if (useCudaFFT) {
            cufftDestroy(fftForward);
            cufftDestroy(fftBackward);
            if (doLJPME) {
                cufftDestroy(dispersionFftForward);
                cufftDestroy(dispersionFftBackward);                
            }
        }
        if (usePmeStream) {
            cuStreamDestroy(pmeStream);
            cuEventDestroy(pmeSyncEvent);
            cuEventDestroy(paramsSyncEvent);
        }
    }
274
275
}

276
void CudaCalcNonbondedForceKernel::initialize(const System& system, const NonbondedForce& force) {
277
    ContextSelector selector(cu);
278
279
280
281
    int forceIndex;
    for (forceIndex = 0; forceIndex < system.getNumForces() && &system.getForce(forceIndex) != &force; ++forceIndex)
        ;
    string prefix = "nonbonded"+cu.intToString(forceIndex)+"_";
282

283
    // Identify which exceptions are 1-4 interactions.
284

285
286
287
288
289
290
291
    set<int> exceptionsWithOffsets;
    for (int i = 0; i < force.getNumExceptionParameterOffsets(); i++) {
        string param;
        int exception;
        double charge, sigma, epsilon;
        force.getExceptionParameterOffset(i, param, exception, charge, sigma, epsilon);
        exceptionsWithOffsets.insert(exception);
292
    }
293
294
295
296
297
298
299
300
301
302
    vector<pair<int, int> > exclusions;
    vector<int> exceptions;
    for (int i = 0; i < force.getNumExceptions(); i++) {
        int particle1, particle2;
        double chargeProd, sigma, epsilon;
        force.getExceptionParameters(i, particle1, particle2, chargeProd, sigma, epsilon);
        exclusions.push_back(pair<int, int>(particle1, particle2));
        if (chargeProd != 0.0 || epsilon != 0.0 || exceptionsWithOffsets.find(i) != exceptionsWithOffsets.end()) {
            exceptionIndex[i] = exceptions.size();
            exceptions.push_back(i);
303
304
305
        }
    }

306
    // Initialize nonbonded interactions.
307

308
309
310
311
312
313
314
315
316
317
318
319
320
321
    int numParticles = force.getNumParticles();
    vector<float4> baseParticleParamVec(cu.getPaddedNumAtoms(), make_float4(0, 0, 0, 0));
    vector<vector<int> > exclusionList(numParticles);
    hasCoulomb = false;
    hasLJ = false;
    for (int i = 0; i < numParticles; i++) {
        double charge, sigma, epsilon;
        force.getParticleParameters(i, charge, sigma, epsilon);
        baseParticleParamVec[i] = make_float4(charge, sigma, epsilon, 0);
        exclusionList[i].push_back(i);
        if (charge != 0.0)
            hasCoulomb = true;
        if (epsilon != 0.0)
            hasLJ = true;
322
    }
323
324
325
326
327
328
329
330
331
    for (int i = 0; i < force.getNumParticleParameterOffsets(); i++) {
        string param;
        int particle;
        double charge, sigma, epsilon;
        force.getParticleParameterOffset(i, param, particle, charge, sigma, epsilon);
        if (charge != 0.0)
            hasCoulomb = true;
        if (epsilon != 0.0)
            hasLJ = true;
332
    }
333
334
335
    for (auto exclusion : exclusions) {
        exclusionList[exclusion.first].push_back(exclusion.second);
        exclusionList[exclusion.second].push_back(exclusion.first);
336
    }
337
338
339
340
341
    nonbondedMethod = CalcNonbondedForceKernel::NonbondedMethod(force.getNonbondedMethod());
    bool useCutoff = (nonbondedMethod != NoCutoff);
    bool usePeriodic = (nonbondedMethod != NoCutoff && nonbondedMethod != CutoffNonPeriodic);
    doLJPME = (nonbondedMethod == LJPME && hasLJ);
    usePosqCharges = hasCoulomb ? cu.requestPosqCharges() : false;
342

343
344
345
346
347
348
    map<string, string> defines;
    defines["HAS_COULOMB"] = (hasCoulomb ? "1" : "0");
    defines["HAS_LENNARD_JONES"] = (hasLJ ? "1" : "0");
    defines["USE_LJ_SWITCH"] = (useCutoff && force.getUseSwitchingFunction() ? "1" : "0");
    if (useCutoff) {
        // Compute the reaction field constants.
349

350
351
352
353
354
355
356
357
358
359
360
361
        double reactionFieldK = pow(force.getCutoffDistance(), -3.0)*(force.getReactionFieldDielectric()-1.0)/(2.0*force.getReactionFieldDielectric()+1.0);
        double reactionFieldC = (1.0 / force.getCutoffDistance())*(3.0*force.getReactionFieldDielectric())/(2.0*force.getReactionFieldDielectric()+1.0);
        defines["REACTION_FIELD_K"] = cu.doubleToString(reactionFieldK);
        defines["REACTION_FIELD_C"] = cu.doubleToString(reactionFieldC);
        
        // Compute the switching coefficients.
        
        if (force.getUseSwitchingFunction()) {
            defines["LJ_SWITCH_CUTOFF"] = cu.doubleToString(force.getSwitchingDistance());
            defines["LJ_SWITCH_C3"] = cu.doubleToString(10/pow(force.getSwitchingDistance()-force.getCutoffDistance(), 3.0));
            defines["LJ_SWITCH_C4"] = cu.doubleToString(15/pow(force.getSwitchingDistance()-force.getCutoffDistance(), 4.0));
            defines["LJ_SWITCH_C5"] = cu.doubleToString(6/pow(force.getSwitchingDistance()-force.getCutoffDistance(), 5.0));
362
363
        }
    }
364
365
366
367
368
369
    if (force.getUseDispersionCorrection() && cu.getContextIndex() == 0 && !doLJPME)
        dispersionCoefficient = NonbondedForceImpl::calcDispersionCorrection(system, force);
    else
        dispersionCoefficient = 0.0;
    alpha = 0;
    ewaldSelfEnergy = 0.0;
370
    totalCharge = 0.0;
371
    map<string, string> paramsDefines;
372
    paramsDefines["ONE_4PI_EPS0"] = cu.doubleToString(ONE_4PI_EPS0);
373
374
    paramsDefines["EPSILON0"] = cu.doubleToString(EPSILON0);
    paramsDefines["WORK_GROUP_SIZE"] = cu.intToString(CudaContext::ThreadBlockSize);
375
376
377
    hasOffsets = (force.getNumParticleParameterOffsets() > 0 || force.getNumExceptionParameterOffsets() > 0);
    if (hasOffsets)
        paramsDefines["HAS_OFFSETS"] = "1";
378
379
380
381
    if (force.getNumParticleParameterOffsets() > 0)
        paramsDefines["HAS_PARTICLE_OFFSETS"] = "1";
    if (force.getNumExceptionParameterOffsets() > 0)
        paramsDefines["HAS_EXCEPTION_OFFSETS"] = "1";
382
383
    if (usePosqCharges)
        paramsDefines["USE_POSQ_CHARGES"] = "1";
384
385
    if (doLJPME)
        paramsDefines["INCLUDE_LJPME_EXCEPTIONS"] = "1";
386
387
    if (nonbondedMethod == Ewald) {
        // Compute the Ewald parameters.
388

389
390
391
392
393
394
395
396
        int kmaxx, kmaxy, kmaxz;
        NonbondedForceImpl::calcEwaldParameters(system, force, alpha, kmaxx, kmaxy, kmaxz);
        defines["EWALD_ALPHA"] = cu.doubleToString(alpha);
        defines["TWO_OVER_SQRT_PI"] = cu.doubleToString(2.0/sqrt(M_PI));
        defines["USE_EWALD"] = "1";
        if (cu.getContextIndex() == 0) {
            paramsDefines["INCLUDE_EWALD"] = "1";
            paramsDefines["EWALD_SELF_ENERGY_SCALE"] = cu.doubleToString(ONE_4PI_EPS0*alpha/sqrt(M_PI));
397
            for (int i = 0; i < numParticles; i++) {
398
                ewaldSelfEnergy -= baseParticleParamVec[i].x*baseParticleParamVec[i].x*ONE_4PI_EPS0*alpha/sqrt(M_PI);
399
400
                totalCharge += baseParticleParamVec[i].x;
            }
401

402
            // Create the reciprocal space kernels.
403

404
405
406
407
408
409
410
411
412
            map<string, string> replacements;
            replacements["NUM_ATOMS"] = cu.intToString(numParticles);
            replacements["PADDED_NUM_ATOMS"] = cu.intToString(cu.getPaddedNumAtoms());
            replacements["KMAX_X"] = cu.intToString(kmaxx);
            replacements["KMAX_Y"] = cu.intToString(kmaxy);
            replacements["KMAX_Z"] = cu.intToString(kmaxz);
            replacements["EXP_COEFFICIENT"] = cu.doubleToString(-1.0/(4.0*alpha*alpha));
            replacements["ONE_4PI_EPS0"] = cu.doubleToString(ONE_4PI_EPS0);
            replacements["M_PI"] = cu.doubleToString(M_PI);
413
            CUmodule module = cu.createModule(CudaKernelSources::vectorOps+CommonKernelSources::ewald, replacements);
414
415
416
417
            ewaldSumsKernel = cu.getKernel(module, "calculateEwaldCosSinSums");
            ewaldForcesKernel = cu.getKernel(module, "calculateEwaldForces");
            int elementSize = (cu.getUseDoublePrecision() ? sizeof(double2) : sizeof(float2));
            cosSinSums.initialize(cu, (2*kmaxx-1)*(2*kmaxy-1)*(2*kmaxz-1), elementSize, "cosSinSums");
418
419
        }
    }
420
421
    else if (((nonbondedMethod == PME || nonbondedMethod == LJPME) && hasCoulomb) || doLJPME) {
        // Compute the PME parameters.
422

423
424
425
426
427
428
429
430
431
432
        NonbondedForceImpl::calcPMEParameters(system, force, alpha, gridSizeX, gridSizeY, gridSizeZ, false);
        gridSizeX = CudaFFT3D::findLegalDimension(gridSizeX);
        gridSizeY = CudaFFT3D::findLegalDimension(gridSizeY);
        gridSizeZ = CudaFFT3D::findLegalDimension(gridSizeZ);
        if (doLJPME) {
            NonbondedForceImpl::calcPMEParameters(system, force, dispersionAlpha, dispersionGridSizeX,
                                                  dispersionGridSizeY, dispersionGridSizeZ, true);
            dispersionGridSizeX = CudaFFT3D::findLegalDimension(dispersionGridSizeX);
            dispersionGridSizeY = CudaFFT3D::findLegalDimension(dispersionGridSizeY);
            dispersionGridSizeZ = CudaFFT3D::findLegalDimension(dispersionGridSizeZ);
433
434
        }

435
436
437
        defines["EWALD_ALPHA"] = cu.doubleToString(alpha);
        defines["TWO_OVER_SQRT_PI"] = cu.doubleToString(2.0/sqrt(M_PI));
        defines["USE_EWALD"] = "1";
438
        defines["DO_LJPME"] = doLJPME ? "1" : "0";
439
        if (doLJPME) {
440
            defines["EWALD_DISPERSION_ALPHA"] = cu.doubleToString(dispersionAlpha);
441
442
443
444
445
446
447
448
            double invRCut6 = pow(force.getCutoffDistance(), -6);
            double dalphaR = dispersionAlpha * force.getCutoffDistance();
            double dar2 = dalphaR*dalphaR;
            double dar4 = dar2*dar2;
            double multShift6 = -invRCut6*(1.0 - exp(-dar2) * (1.0 + dar2 + 0.5*dar4));
            defines["INVCUT6"] = cu.doubleToString(invRCut6);
            defines["MULTSHIFT6"] = cu.doubleToString(multShift6);
        }
449
        if (cu.getContextIndex() == 0) {
450
451
            paramsDefines["INCLUDE_EWALD"] = "1";
            paramsDefines["EWALD_SELF_ENERGY_SCALE"] = cu.doubleToString(ONE_4PI_EPS0*alpha/sqrt(M_PI));
452
            for (int i = 0; i < numParticles; i++) {
453
                ewaldSelfEnergy -= baseParticleParamVec[i].x*baseParticleParamVec[i].x*ONE_4PI_EPS0*alpha/sqrt(M_PI);
454
455
                totalCharge += baseParticleParamVec[i].x;
            }
456
457
458
459
460
461
            if (doLJPME) {
                paramsDefines["INCLUDE_LJPME"] = "1";
                paramsDefines["LJPME_SELF_ENERGY_SCALE"] = cu.doubleToString(pow(dispersionAlpha, 6)/3.0);
                for (int i = 0; i < numParticles; i++)
                    ewaldSelfEnergy += baseParticleParamVec[i].z*pow(baseParticleParamVec[i].y*dispersionAlpha, 6)/3.0;
            }
Peter Eastman's avatar
Peter Eastman committed
462
463
            char deviceName[100];
            cuDeviceGetName(deviceName, 100, cu.getDevice());
464
            usePmeStream = (!cu.getPlatformData().disablePmeStream && !cu.getPlatformData().useCpuPme && string(deviceName) != "GeForce GTX 980"); // Using a separate stream is slower on GTX 980
465
            map<string, string> pmeDefines;
466
467
468
469
470
471
472
473
474
            pmeDefines["PME_ORDER"] = cu.intToString(PmeOrder);
            pmeDefines["NUM_ATOMS"] = cu.intToString(numParticles);
            pmeDefines["PADDED_NUM_ATOMS"] = cu.intToString(cu.getPaddedNumAtoms());
            pmeDefines["RECIP_EXP_FACTOR"] = cu.doubleToString(M_PI*M_PI/(alpha*alpha));
            pmeDefines["GRID_SIZE_X"] = cu.intToString(gridSizeX);
            pmeDefines["GRID_SIZE_Y"] = cu.intToString(gridSizeY);
            pmeDefines["GRID_SIZE_Z"] = cu.intToString(gridSizeZ);
            pmeDefines["EPSILON_FACTOR"] = cu.doubleToString(sqrt(ONE_4PI_EPS0));
            pmeDefines["M_PI"] = cu.doubleToString(M_PI);
475
476
            useFixedPointChargeSpreading = cu.getUseDoublePrecision() || cu.getPlatformData().deterministicForces;
            if (useFixedPointChargeSpreading)
477
                pmeDefines["USE_FIXED_POINT_CHARGE_SPREADING"] = "1";
Peter Eastman's avatar
Peter Eastman committed
478
479
            if (usePmeStream)
                pmeDefines["USE_PME_STREAM"] = "1";
480
481
            map<string, string> replacements;
            replacements["CHARGE"] = (usePosqCharges ? "pos.w" : "charges[atom]");
482
            CUmodule module = cu.createModule(CudaKernelSources::vectorOps+cu.replaceStrings(CommonKernelSources::pme, replacements), pmeDefines);
483
            if (cu.getPlatformData().useCpuPme && !doLJPME && usePosqCharges) {
484
485
                // Create the CPU PME kernel.

486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
                try {
                    cpuPme = getPlatform().createKernel(CalcPmeReciprocalForceKernel::Name(), *cu.getPlatformData().context);
                    cpuPme.getAs<CalcPmeReciprocalForceKernel>().initialize(gridSizeX, gridSizeY, gridSizeZ, numParticles, alpha, cu.getPlatformData().deterministicForces);
                    CUfunction addForcesKernel = cu.getKernel(module, "addForces");
                    pmeio = new PmeIO(cu, addForcesKernel);
                    cu.addPreComputation(new PmePreComputation(cu, cpuPme, *pmeio));
                    cu.addPostComputation(new PmePostComputation(cpuPme, *pmeio));
                }
                catch (OpenMMException& ex) {
                    // The CPU PME plugin isn't available.
                }
            }
            if (pmeio == NULL) {
                pmeGridIndexKernel = cu.getKernel(module, "findAtomGridIndex");
                pmeSpreadChargeKernel = cu.getKernel(module, "gridSpreadCharge");
                pmeConvolutionKernel = cu.getKernel(module, "reciprocalConvolution");
                pmeInterpolateForceKernel = cu.getKernel(module, "gridInterpolateForce");
                pmeEvalEnergyKernel = cu.getKernel(module, "gridEvaluateEnergy");
504
505
                if (useFixedPointChargeSpreading)
                    pmeFinishSpreadChargeKernel = cu.getKernel(module, "finishSpreadCharge");
506
                cuFuncSetCacheConfig(pmeSpreadChargeKernel, CU_FUNC_CACHE_PREFER_SHARED);
507
508
509
510
511
512
513
514
515
                cuFuncSetCacheConfig(pmeInterpolateForceKernel, CU_FUNC_CACHE_PREFER_L1);
                if (doLJPME) {
                    pmeDefines["EWALD_ALPHA"] = cu.doubleToString(dispersionAlpha);
                    pmeDefines["GRID_SIZE_X"] = cu.intToString(dispersionGridSizeX);
                    pmeDefines["GRID_SIZE_Y"] = cu.intToString(dispersionGridSizeY);
                    pmeDefines["GRID_SIZE_Z"] = cu.intToString(dispersionGridSizeZ);
                    pmeDefines["RECIP_EXP_FACTOR"] = cu.doubleToString(M_PI*M_PI/(dispersionAlpha*dispersionAlpha));
                    pmeDefines["USE_LJPME"] = "1";
                    pmeDefines["CHARGE_FROM_SIGEPS"] = "1";
516
                    module = cu.createModule(CudaKernelSources::vectorOps+CommonKernelSources::pme, pmeDefines);
517
518
                    if (useFixedPointChargeSpreading)
                        pmeDispersionFinishSpreadChargeKernel = cu.getKernel(module, "finishSpreadCharge");
519
520
521
522
523
524
525
                    pmeDispersionGridIndexKernel = cu.getKernel(module, "findAtomGridIndex");
                    pmeDispersionSpreadChargeKernel = cu.getKernel(module, "gridSpreadCharge");
                    pmeDispersionConvolutionKernel = cu.getKernel(module, "reciprocalConvolution");
                    pmeEvalDispersionEnergyKernel = cu.getKernel(module, "gridEvaluateEnergy");
                    pmeInterpolateDispersionForceKernel = cu.getKernel(module, "gridInterpolateForce");
                    cuFuncSetCacheConfig(pmeDispersionSpreadChargeKernel, CU_FUNC_CACHE_PREFER_L1);
                }
peastman's avatar
peastman committed
526

527
                // Create required data structures.
peastman's avatar
peastman committed
528

529
                int elementSize = (cu.getUseDoublePrecision() ? sizeof(double) : sizeof(float));
530
                int gridElements = gridSizeX*gridSizeY*gridSizeZ;
531
                if (doLJPME) {
532
                    gridElements = max(gridElements, dispersionGridSizeX*dispersionGridSizeY*dispersionGridSizeZ);
533
534
535
                }
                pmeGrid1.initialize(cu, gridElements, 2*elementSize, "pmeGrid1");
                pmeGrid2.initialize(cu, gridElements, 2*elementSize, "pmeGrid2");
536
537
538
539
                if (useFixedPointChargeSpreading)
                    cu.addAutoclearBuffer(pmeGrid2);
                else
                    cu.addAutoclearBuffer(pmeGrid1);
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
                pmeBsplineModuliX.initialize(cu, gridSizeX, elementSize, "pmeBsplineModuliX");
                pmeBsplineModuliY.initialize(cu, gridSizeY, elementSize, "pmeBsplineModuliY");
                pmeBsplineModuliZ.initialize(cu, gridSizeZ, elementSize, "pmeBsplineModuliZ");
                if (doLJPME) {
                    pmeDispersionBsplineModuliX.initialize(cu, dispersionGridSizeX, elementSize, "pmeDispersionBsplineModuliX");
                    pmeDispersionBsplineModuliY.initialize(cu, dispersionGridSizeY, elementSize, "pmeDispersionBsplineModuliY");
                    pmeDispersionBsplineModuliZ.initialize(cu, dispersionGridSizeZ, elementSize, "pmeDispersionBsplineModuliZ");
                }
                pmeAtomGridIndex.initialize<int2>(cu, numParticles, "pmeAtomGridIndex");
                int energyElementSize = (cu.getUseDoublePrecision() || cu.getUseMixedPrecision() ? sizeof(double) : sizeof(float));
                pmeEnergyBuffer.initialize(cu, cu.getNumThreadBlocks()*CudaContext::ThreadBlockSize, energyElementSize, "pmeEnergyBuffer");
                cu.clearBuffer(pmeEnergyBuffer);
                sort = new CudaSort(cu, new SortTrait(), cu.getNumAtoms());
                int cufftVersion;
                cufftGetVersion(&cufftVersion);
                useCudaFFT = (cufftVersion >= 7050); // There was a critical bug in version 7.0
                if (useCudaFFT) {
                    cufftResult result = cufftPlan3d(&fftForward, gridSizeX, gridSizeY, gridSizeZ, cu.getUseDoublePrecision() ? CUFFT_D2Z : CUFFT_R2C);
                    if (result != CUFFT_SUCCESS)
                        throw OpenMMException("Error initializing FFT: "+cu.intToString(result));
                    result = cufftPlan3d(&fftBackward, gridSizeX, gridSizeY, gridSizeZ, cu.getUseDoublePrecision() ? CUFFT_Z2D : CUFFT_C2R);
                    if (result != CUFFT_SUCCESS)
                        throw OpenMMException("Error initializing FFT: "+cu.intToString(result));
                    if (doLJPME) {
                        result = cufftPlan3d(&dispersionFftForward, dispersionGridSizeX, dispersionGridSizeY, 
                                                dispersionGridSizeZ, cu.getUseDoublePrecision() ? CUFFT_D2Z : CUFFT_R2C);
                        if (result != CUFFT_SUCCESS)
                            throw OpenMMException("Error initializing disperison FFT: "+cu.intToString(result));
                        result = cufftPlan3d(&dispersionFftBackward, dispersionGridSizeX, dispersionGridSizeY,
                                             dispersionGridSizeZ, cu.getUseDoublePrecision() ? CUFFT_Z2D : CUFFT_C2R);
                        if (result != CUFFT_SUCCESS)
                            throw OpenMMException("Error initializing disperison FFT: "+cu.intToString(result));
                    }
                }
                else {
                    fft = new CudaFFT3D(cu, gridSizeX, gridSizeY, gridSizeZ, true);
                    if (doLJPME)
                        dispersionFft = new CudaFFT3D(cu, dispersionGridSizeX, dispersionGridSizeY, dispersionGridSizeZ, true);
                }
579

580
                // Prepare for doing PME on its own stream.
581

582
583
584
585
586
587
588
589
590
591
                if (usePmeStream) {
                    cuStreamCreate(&pmeStream, CU_STREAM_NON_BLOCKING);
                    if (useCudaFFT) {
                        cufftSetStream(fftForward, pmeStream);
                        cufftSetStream(fftBackward, pmeStream);
                        if (doLJPME) {
                            cufftSetStream(dispersionFftForward, pmeStream);
                            cufftSetStream(dispersionFftBackward, pmeStream);
                        }
                    }
592
593
                    CHECK_RESULT(cuEventCreate(&pmeSyncEvent, cu.getEventFlags()), "Error creating event for NonbondedForce");
                    CHECK_RESULT(cuEventCreate(&paramsSyncEvent, cu.getEventFlags()), "Error creating event for NonbondedForce");
594
595
596
597
598
599
600
                    int recipForceGroup = force.getReciprocalSpaceForceGroup();
                    if (recipForceGroup < 0)
                        recipForceGroup = force.getForceGroup();
                    cu.addPreComputation(new SyncStreamPreComputation(cu, pmeStream, pmeSyncEvent, recipForceGroup));
                    cu.addPostComputation(new SyncStreamPostComputation(cu, pmeSyncEvent, cu.getKernel(module, "addEnergy"), pmeEnergyBuffer, recipForceGroup));
                }
                hasInitializedFFT = true;
601

602
                // Initialize the b-spline moduli.
603

604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
                for (int grid = 0; grid < 2; grid++) {
                    int xsize, ysize, zsize;
                    CudaArray *xmoduli, *ymoduli, *zmoduli;
                    if (grid == 0) {
                        xsize = gridSizeX;
                        ysize = gridSizeY;
                        zsize = gridSizeZ;
                        xmoduli = &pmeBsplineModuliX;
                        ymoduli = &pmeBsplineModuliY;
                        zmoduli = &pmeBsplineModuliZ;
                    }
                    else {
                        if (!doLJPME)
                            continue;
                        xsize = dispersionGridSizeX;
                        ysize = dispersionGridSizeY;
                        zsize = dispersionGridSizeZ;
                        xmoduli = &pmeDispersionBsplineModuliX;
                        ymoduli = &pmeDispersionBsplineModuliY;
                        zmoduli = &pmeDispersionBsplineModuliZ;
                    }
                    int maxSize = max(max(xsize, ysize), zsize);
                    vector<double> data(PmeOrder);
                    vector<double> ddata(PmeOrder);
                    vector<double> bsplines_data(maxSize);
                    data[PmeOrder-1] = 0.0;
                    data[1] = 0.0;
                    data[0] = 1.0;
                    for (int i = 3; i < PmeOrder; i++) {
                        double div = 1.0/(i-1.0);
                        data[i-1] = 0.0;
                        for (int j = 1; j < (i-1); j++)
                            data[i-j-1] = div*(j*data[i-j-2]+(i-j)*data[i-j-1]);
                        data[0] = div*data[0];
                    }
639

640
                    // Differentiate.
641

642
643
644
645
646
647
648
649
650
651
652
653
                    ddata[0] = -data[0];
                    for (int i = 1; i < PmeOrder; i++)
                        ddata[i] = data[i-1]-data[i];
                    double div = 1.0/(PmeOrder-1);
                    data[PmeOrder-1] = 0.0;
                    for (int i = 1; i < (PmeOrder-1); i++)
                        data[PmeOrder-i-1] = div*(i*data[PmeOrder-i-2]+(PmeOrder-i)*data[PmeOrder-i-1]);
                    data[0] = div*data[0];
                    for (int i = 0; i < maxSize; i++)
                        bsplines_data[i] = 0.0;
                    for (int i = 1; i <= PmeOrder; i++)
                        bsplines_data[i] = data[i-1];
654

655
                    // Evaluate the actual bspline moduli for X/Y/Z.
656

657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
                    for (int dim = 0; dim < 3; dim++) {
                        int ndata = (dim == 0 ? xsize : dim == 1 ? ysize : zsize);
                        vector<double> moduli(ndata);
                        for (int i = 0; i < ndata; i++) {
                            double sc = 0.0;
                            double ss = 0.0;
                            for (int j = 0; j < ndata; j++) {
                                double arg = (2.0*M_PI*i*j)/ndata;
                                sc += bsplines_data[j]*cos(arg);
                                ss += bsplines_data[j]*sin(arg);
                            }
                            moduli[i] = sc*sc+ss*ss;
                        }
                        for (int i = 0; i < ndata; i++)
                            if (moduli[i] < 1.0e-7)
672
                                moduli[i] = (moduli[(i-1+ndata)%ndata]+moduli[(i+1)%ndata])*0.5;
673
674
675
676
677
678
679
680
681
682
683
                        if (dim == 0)
                            xmoduli->upload(moduli, true);
                        else if (dim == 1)
                            ymoduli->upload(moduli, true);
                        else
                            zmoduli->upload(moduli, true);
                    }
                }
            }
        }
    }
684

685
    // Add code to subtract off the reciprocal part of excluded interactions.
686

687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
    if ((nonbondedMethod == Ewald || nonbondedMethod == PME || nonbondedMethod == LJPME) && pmeio == NULL) {
        int numContexts = cu.getPlatformData().contexts.size();
        int startIndex = cu.getContextIndex()*force.getNumExceptions()/numContexts;
        int endIndex = (cu.getContextIndex()+1)*force.getNumExceptions()/numContexts;
        int numExclusions = endIndex-startIndex;
        if (numExclusions > 0) {
            paramsDefines["HAS_EXCLUSIONS"] = "1";
            vector<vector<int> > atoms(numExclusions, vector<int>(2));
            exclusionAtoms.initialize<int2>(cu, numExclusions, "exclusionAtoms");
            exclusionParams.initialize<float4>(cu, numExclusions, "exclusionParams");
            vector<int2> exclusionAtomsVec(numExclusions);
            for (int i = 0; i < numExclusions; i++) {
                int j = i+startIndex;
                exclusionAtomsVec[i] = make_int2(exclusions[j].first, exclusions[j].second);
                atoms[i][0] = exclusions[j].first;
                atoms[i][1] = exclusions[j].second;
            }
            exclusionAtoms.upload(exclusionAtomsVec);
            map<string, string> replacements;
706
            replacements["PARAMS"] = cu.getBondedUtilities().addArgument(exclusionParams, "float4");
707
708
709
            replacements["EWALD_ALPHA"] = cu.doubleToString(alpha);
            replacements["TWO_OVER_SQRT_PI"] = cu.doubleToString(2.0/sqrt(M_PI));
            replacements["DO_LJPME"] = doLJPME ? "1" : "0";
710
            replacements["USE_PERIODIC"] = force.getExceptionsUsePeriodicBoundaryConditions() ? "1" : "0";
711
712
            if (doLJPME)
                replacements["EWALD_DISPERSION_ALPHA"] = cu.doubleToString(dispersionAlpha);
713
714
            if (force.getIncludeDirectSpace())
                cu.getBondedUtilities().addInteraction(atoms, cu.replaceStrings(CommonKernelSources::pmeExclusions, replacements), force.getForceGroup());
715
        }
716
    }
717

718
    // Add the interaction to the default nonbonded kernel.
719

720
    string source = cu.replaceStrings(CommonKernelSources::coulombLennardJones, defines);
721
722
723
724
    charges.initialize(cu, cu.getPaddedNumAtoms(), cu.getUseDoublePrecision() ? sizeof(double) : sizeof(float), "charges");
    baseParticleParams.initialize<float4>(cu, cu.getPaddedNumAtoms(), "baseParticleParams");
    baseParticleParams.upload(baseParticleParamVec);
    map<string, string> replacements;
725
    replacements["ONE_4PI_EPS0"] = cu.doubleToString(ONE_4PI_EPS0);
726
727
728
    if (usePosqCharges) {
        replacements["CHARGE1"] = "posq1.w";
        replacements["CHARGE2"] = "posq2.w";
729
    }
730
731
732
    else {
        replacements["CHARGE1"] = prefix+"charge1";
        replacements["CHARGE2"] = prefix+"charge2";
733
    }
734
    if (hasCoulomb && !usePosqCharges)
735
736
737
738
739
740
        cu.getNonbondedUtilities().addParameter(CudaNonbondedUtilities::ParameterInfo(prefix+"charge", "real", 1, charges.getElementSize(), charges.getDevicePointer()));
    sigmaEpsilon.initialize<float2>(cu, cu.getPaddedNumAtoms(), "sigmaEpsilon");
    if (hasLJ) {
        replacements["SIGMA_EPSILON1"] = prefix+"sigmaEpsilon1";
        replacements["SIGMA_EPSILON2"] = prefix+"sigmaEpsilon2";
        cu.getNonbondedUtilities().addParameter(CudaNonbondedUtilities::ParameterInfo(prefix+"sigmaEpsilon", "float", 2, sizeof(float2), sigmaEpsilon.getDevicePointer()));
741
    }
742
    source = cu.replaceStrings(source, replacements);
743
    if (force.getIncludeDirectSpace())
744
        cu.getNonbondedUtilities().addInteraction(useCutoff, usePeriodic, true, force.getCutoffDistance(), exclusionList, source, force.getForceGroup(), numParticles > 3000, true);
745

746
    // Initialize the exceptions.
747

748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
    int numContexts = cu.getPlatformData().contexts.size();
    int startIndex = cu.getContextIndex()*exceptions.size()/numContexts;
    int endIndex = (cu.getContextIndex()+1)*exceptions.size()/numContexts;
    int numExceptions = endIndex-startIndex;
    if (numExceptions > 0) {
        paramsDefines["HAS_EXCEPTIONS"] = "1";
        exceptionAtoms.resize(numExceptions);
        vector<vector<int> > atoms(numExceptions, vector<int>(2));
        exceptionParams.initialize<float4>(cu, numExceptions, "exceptionParams");
        baseExceptionParams.initialize<float4>(cu, numExceptions, "baseExceptionParams");
        vector<float4> baseExceptionParamsVec(numExceptions);
        for (int i = 0; i < numExceptions; i++) {
            double chargeProd, sigma, epsilon;
            force.getExceptionParameters(exceptions[startIndex+i], atoms[i][0], atoms[i][1], chargeProd, sigma, epsilon);
            baseExceptionParamsVec[i] = make_float4(chargeProd, sigma, epsilon, 0);
            exceptionAtoms[i] = make_pair(atoms[i][0], atoms[i][1]);
        }
        baseExceptionParams.upload(baseExceptionParamsVec);
        map<string, string> replacements;
767
        replacements["APPLY_PERIODIC"] = (usePeriodic && force.getExceptionsUsePeriodicBoundaryConditions() ? "1" : "0");
768
        replacements["PARAMS"] = cu.getBondedUtilities().addArgument(exceptionParams, "float4");
769
770
        if (force.getIncludeDirectSpace())
            cu.getBondedUtilities().addInteraction(atoms, cu.replaceStrings(CommonKernelSources::nonbondedExceptions, replacements), force.getForceGroup());
771
772
773
    }
    
    // Initialize parameter offsets.
774

775
    vector<vector<float4> > particleOffsetVec(force.getNumParticles());
776
    vector<vector<float4> > exceptionOffsetVec(numExceptions);
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
    for (int i = 0; i < force.getNumParticleParameterOffsets(); i++) {
        string param;
        int particle;
        double charge, sigma, epsilon;
        force.getParticleParameterOffset(i, param, particle, charge, sigma, epsilon);
        auto paramPos = find(paramNames.begin(), paramNames.end(), param);
        int paramIndex;
        if (paramPos == paramNames.end()) {
            paramIndex = paramNames.size();
            paramNames.push_back(param);
        }
        else
            paramIndex = paramPos-paramNames.begin();
        particleOffsetVec[particle].push_back(make_float4(charge, sigma, epsilon, paramIndex));
    }
    for (int i = 0; i < force.getNumExceptionParameterOffsets(); i++) {
        string param;
        int exception;
        double charge, sigma, epsilon;
        force.getExceptionParameterOffset(i, param, exception, charge, sigma, epsilon);
797
798
799
        int index = exceptionIndex[exception];
        if (index < startIndex || index >= endIndex)
            continue;
800
801
802
803
804
805
806
807
        auto paramPos = find(paramNames.begin(), paramNames.end(), param);
        int paramIndex;
        if (paramPos == paramNames.end()) {
            paramIndex = paramNames.size();
            paramNames.push_back(param);
        }
        else
            paramIndex = paramPos-paramNames.begin();
808
        exceptionOffsetVec[index-startIndex].push_back(make_float4(charge, sigma, epsilon, paramIndex));
809
    }
810
811
812
813
814
815
816
817
818
    paramValues.resize(paramNames.size(), 0.0);
    particleParamOffsets.initialize<float4>(cu, max(force.getNumParticleParameterOffsets(), 1), "particleParamOffsets");
    particleOffsetIndices.initialize<int>(cu, cu.getPaddedNumAtoms()+1, "particleOffsetIndices");
    vector<int> particleOffsetIndicesVec, exceptionOffsetIndicesVec;
    vector<float4> p, e;
    for (int i = 0; i < particleOffsetVec.size(); i++) {
        particleOffsetIndicesVec.push_back(p.size());
        for (int j = 0; j < particleOffsetVec[i].size(); j++)
            p.push_back(particleOffsetVec[i][j]);
819
    }
820
821
822
823
824
825
    while (particleOffsetIndicesVec.size() < particleOffsetIndices.getSize())
        particleOffsetIndicesVec.push_back(p.size());
    for (int i = 0; i < exceptionOffsetVec.size(); i++) {
        exceptionOffsetIndicesVec.push_back(e.size());
        for (int j = 0; j < exceptionOffsetVec[i].size(); j++)
            e.push_back(exceptionOffsetVec[i][j]);
826
    }
827
828
829
830
    exceptionOffsetIndicesVec.push_back(e.size());
    if (force.getNumParticleParameterOffsets() > 0) {
        particleParamOffsets.upload(p);
        particleOffsetIndices.upload(particleOffsetIndicesVec);
831
    }
832
833
834
    exceptionParamOffsets.initialize<float4>(cu, max((int) e.size(), 1), "exceptionParamOffsets");
    exceptionOffsetIndices.initialize<int>(cu, exceptionOffsetIndicesVec.size(), "exceptionOffsetIndices");
    if (e.size() > 0) {
835
836
        exceptionParamOffsets.upload(e);
        exceptionOffsetIndices.upload(exceptionOffsetIndicesVec);
837
    }
838
    globalParams.initialize(cu, max((int) paramValues.size(), 1), cu.getUseDoublePrecision() ? sizeof(double) : sizeof(float), "globalParams");
peastman's avatar
peastman committed
839
840
    if (paramValues.size() > 0)
        globalParams.upload(paramValues, true);
841
842
    chargeBuffer.initialize(cu, cu.getNumThreadBlocks(), cu.getUseDoublePrecision() ? sizeof(double) : sizeof(float), "chargeBuffer");
    cu.clearBuffer(chargeBuffer);
843
    recomputeParams = true;
844
    
845
    // Initialize the kernel for updating parameters.
846
    
847
    CUmodule module = cu.createModule(CommonKernelSources::nonbondedParameters, paramsDefines);
848
849
    computeParamsKernel = cu.getKernel(module, "computeParameters");
    computeExclusionParamsKernel = cu.getKernel(module, "computeExclusionParameters");
850
    computePlasmaCorrectionKernel = cu.getKernel(module, "computePlasmaCorrection");
851
852
    info = new ForceInfo(force);
    cu.addForce(info);
853
854
}

855
856
double CudaCalcNonbondedForceKernel::execute(ContextImpl& context, bool includeForces, bool includeEnergy, bool includeDirect, bool includeReciprocal) {
    // Update particle and exception parameters.
857

858
    ContextSelector selector(cu);
859
860
861
862
863
864
    bool paramChanged = false;
    for (int i = 0; i < paramNames.size(); i++) {
        double value = context.getParameter(paramNames[i]);
        if (value != paramValues[i]) {
            paramValues[i] = value;;
            paramChanged = true;
865
        }
866
    }
867
868
869
870
    if (paramChanged) {
        recomputeParams = true;
        globalParams.upload(paramValues, true);
    }
871
872
873
874
875
876
    double energy = 0.0;
    if (includeReciprocal && (pmeGrid1.isInitialized() || cosSinSums.isInitialized())) {
        double4 boxSize = cu.getPeriodicBoxSize();
        double volume = boxSize.x*boxSize.y*boxSize.z;
        energy = ewaldSelfEnergy - totalCharge*totalCharge/(8*EPSILON0*volume*alpha*alpha);
    }
877
    if (recomputeParams || hasOffsets) {
878
        int computeSelfEnergy = (includeEnergy && includeReciprocal);
879
        int numAtoms = cu.getNumAtoms();
880
881
        vector<void*> paramsArgs = {&cu.getEnergyBuffer().getDevicePointer(), &computeSelfEnergy, &globalParams.getDevicePointer(), &numAtoms,
                &baseParticleParams.getDevicePointer(), &cu.getPosq().getDevicePointer(), &charges.getDevicePointer(), &sigmaEpsilon.getDevicePointer(),
882
                &particleParamOffsets.getDevicePointer(), &particleOffsetIndices.getDevicePointer(), &chargeBuffer.getDevicePointer()};
883
884
885
886
887
888
889
890
        int numExceptions;
        if (exceptionParams.isInitialized()) {
            numExceptions = exceptionParams.getSize();
            paramsArgs.push_back(&numExceptions);
            paramsArgs.push_back(&baseExceptionParams.getDevicePointer());
            paramsArgs.push_back(&exceptionParams.getDevicePointer());
            paramsArgs.push_back(&exceptionParamOffsets.getDevicePointer());
            paramsArgs.push_back(&exceptionOffsetIndices.getDevicePointer());
891
        }
892
893
894
895
896
897
        cu.executeKernel(computeParamsKernel, &paramsArgs[0], cu.getPaddedNumAtoms());
        if (exclusionParams.isInitialized()) {
            int numExclusions = exclusionParams.getSize();
            vector<void*> exclusionParamsArgs = {&cu.getPosq().getDevicePointer(), &charges.getDevicePointer(), &sigmaEpsilon.getDevicePointer(),
                    &numExclusions, &exclusionAtoms.getDevicePointer(), &exclusionParams.getDevicePointer()};
            cu.executeKernel(computeExclusionParamsKernel, &exclusionParamsArgs[0], numExclusions);
898
        }
899
900
901
902
        if (usePmeStream) {
            cuEventRecord(paramsSyncEvent, cu.getCurrentStream());
            cuStreamWaitEvent(pmeStream, paramsSyncEvent, 0);
        }
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
        if (hasOffsets) {
            // The Ewald self energy was computed in the kernel.

            energy = 0.0;
            if (pmeGrid1.isInitialized() || cosSinSums.isInitialized()) {
                // Invoke a kernel to compute the correction for the neutralizing plasma.

                double4 boxSize = cu.getPeriodicBoxSize();
                if (cu.getUseDoublePrecision()) {
                    double volume = boxSize.x*boxSize.y*boxSize.z;
                    vector<void*> correctionArgs = {&chargeBuffer.getDevicePointer(), &cu.getEnergyBuffer().getDevicePointer(), &alpha, &volume};
                    cu.executeKernel(computePlasmaCorrectionKernel, &correctionArgs[0], CudaContext::ThreadBlockSize, CudaContext::ThreadBlockSize);
                }
                else {
                    float alphaFloat = (float) alpha;
                    float volume = boxSize.x*boxSize.y*boxSize.z;
                    vector<void*> correctionArgs = {&chargeBuffer.getDevicePointer(), &cu.getEnergyBuffer().getDevicePointer(), &alphaFloat, &volume};
                    cu.executeKernel(computePlasmaCorrectionKernel, &correctionArgs[0], CudaContext::ThreadBlockSize, CudaContext::ThreadBlockSize);
                }
            }
        }
924
        recomputeParams = false;
925
    }
926
927
928
929
930
931
932
933
934
935
936
937
    
    // Do reciprocal space calculations.
    
    if (cosSinSums.isInitialized() && includeReciprocal) {
        void* sumsArgs[] = {&cu.getEnergyBuffer().getDevicePointer(), &cu.getPosq().getDevicePointer(), &cosSinSums.getDevicePointer(), cu.getPeriodicBoxSizePointer()};
        cu.executeKernel(ewaldSumsKernel, sumsArgs, cosSinSums.getSize());
        void* forcesArgs[] = {&cu.getForce().getDevicePointer(), &cu.getPosq().getDevicePointer(), &cosSinSums.getDevicePointer(), cu.getPeriodicBoxSizePointer()};
        cu.executeKernel(ewaldForcesKernel, forcesArgs, cu.getNumAtoms());
    }
    if (pmeGrid1.isInitialized() && includeReciprocal) {
        if (usePmeStream)
            cu.setCurrentStream(pmeStream);
938

939
        // Invert the periodic box vectors.
940

941
942
943
944
        Vec3 boxVectors[3];
        cu.getPeriodicBoxVectors(boxVectors[0], boxVectors[1], boxVectors[2]);
        double determinant = boxVectors[0][0]*boxVectors[1][1]*boxVectors[2][2];
        double scale = 1.0/determinant;
945
946
947
948
949
        double4 recipBoxVectors[3];
        recipBoxVectors[0] = make_double4(boxVectors[1][1]*boxVectors[2][2]*scale, 0, 0, 0);
        recipBoxVectors[1] = make_double4(-boxVectors[1][0]*boxVectors[2][2]*scale, boxVectors[0][0]*boxVectors[2][2]*scale, 0, 0);
        recipBoxVectors[2] = make_double4((boxVectors[1][0]*boxVectors[2][1]-boxVectors[1][1]*boxVectors[2][0])*scale, -boxVectors[0][0]*boxVectors[2][1]*scale, boxVectors[0][0]*boxVectors[1][1]*scale, 0);
        float4 recipBoxVectorsFloat[3];
950
951
952
953
954
955
956
        void* recipBoxVectorPointer[3];
        if (cu.getUseDoublePrecision()) {
            recipBoxVectorPointer[0] = &recipBoxVectors[0];
            recipBoxVectorPointer[1] = &recipBoxVectors[1];
            recipBoxVectorPointer[2] = &recipBoxVectors[2];
        }
        else {
957
958
959
            recipBoxVectorsFloat[0] = make_float4((float) recipBoxVectors[0].x, 0, 0, 0);
            recipBoxVectorsFloat[1] = make_float4((float) recipBoxVectors[1].x, (float) recipBoxVectors[1].y, 0, 0);
            recipBoxVectorsFloat[2] = make_float4((float) recipBoxVectors[2].x, (float) recipBoxVectors[2].y, (float) recipBoxVectors[2].z, 0);
960
961
962
963
            recipBoxVectorPointer[0] = &recipBoxVectorsFloat[0];
            recipBoxVectorPointer[1] = &recipBoxVectorsFloat[1];
            recipBoxVectorPointer[2] = &recipBoxVectorsFloat[2];
        }
964

965
        // Execute the reciprocal space kernels.
966

967
968
969
970
        // With fixed point charge spreading, finishSpreadCharge kernel is not needed as
        // gridSpreadCharge can write directly to pmeGrid1.
        CudaArray& pmeSpreadDstGrid = useFixedPointChargeSpreading ? pmeGrid2 : pmeGrid1;

971
972
973
974
975
        if (hasCoulomb) {
            void* gridIndexArgs[] = {&cu.getPosq().getDevicePointer(), &pmeAtomGridIndex.getDevicePointer(), cu.getPeriodicBoxSizePointer(),
                    cu.getInvPeriodicBoxSizePointer(), cu.getPeriodicBoxVecXPointer(), cu.getPeriodicBoxVecYPointer(), cu.getPeriodicBoxVecZPointer(),
                    recipBoxVectorPointer[0], recipBoxVectorPointer[1], recipBoxVectorPointer[2]};
            cu.executeKernel(pmeGridIndexKernel, gridIndexArgs, cu.getNumAtoms());
976

977
            sort->sort(pmeAtomGridIndex);
978

979
            void* spreadArgs[] = {&cu.getPosq().getDevicePointer(), &pmeSpreadDstGrid.getDevicePointer(), cu.getPeriodicBoxSizePointer(),
980
981
982
                    cu.getInvPeriodicBoxSizePointer(), cu.getPeriodicBoxVecXPointer(), cu.getPeriodicBoxVecYPointer(), cu.getPeriodicBoxVecZPointer(),
                    recipBoxVectorPointer[0], recipBoxVectorPointer[1], recipBoxVectorPointer[2], &pmeAtomGridIndex.getDevicePointer(),
                    &charges.getDevicePointer()};
983
            cu.executeKernel(pmeSpreadChargeKernel, spreadArgs, PmeOrder*cu.getNumAtoms(), 128);
984

985
986
987
988
            if (useFixedPointChargeSpreading) {
                void* finishSpreadArgs[] = {&pmeGrid2.getDevicePointer(), &pmeGrid1.getDevicePointer()};
                cu.executeKernel(pmeFinishSpreadChargeKernel, finishSpreadArgs, gridSizeX*gridSizeY*gridSizeZ, 256);
            }
989

990
            if (useCudaFFT) {
991
992
993
994
995
996
997
998
999
                if (cu.getUseDoublePrecision()) {
                    cufftResult result = cufftExecD2Z(fftForward, (double*) pmeGrid1.getDevicePointer(), (double2*) pmeGrid2.getDevicePointer());
                    if (result != CUFFT_SUCCESS)
                        throw OpenMMException("Error executing FFT: "+cu.intToString(result));
                } else {
                    cufftResult result = cufftExecR2C(fftForward, (float*) pmeGrid1.getDevicePointer(), (float2*) pmeGrid2.getDevicePointer());
                    if (result != CUFFT_SUCCESS)
                        throw OpenMMException("Error executing FFT: "+cu.intToString(result));
                }
1000
1001
1002
1003
            }
            else {
                fft->execFFT(pmeGrid1, pmeGrid2, true);
            }
1004

1005
1006
1007
            if (includeEnergy) {
                void* computeEnergyArgs[] = {&pmeGrid2.getDevicePointer(), usePmeStream ? &pmeEnergyBuffer.getDevicePointer() : &cu.getEnergyBuffer().getDevicePointer(),
                        &pmeBsplineModuliX.getDevicePointer(), &pmeBsplineModuliY.getDevicePointer(), &pmeBsplineModuliZ.getDevicePointer(),
1008
                        recipBoxVectorPointer[0], recipBoxVectorPointer[1], recipBoxVectorPointer[2]};
1009
1010
                cu.executeKernel(pmeEvalEnergyKernel, computeEnergyArgs, gridSizeX*gridSizeY*gridSizeZ);
            }
1011

1012
1013
1014
            void* convolutionArgs[] = {&pmeGrid2.getDevicePointer(), &pmeBsplineModuliX.getDevicePointer(),
                    &pmeBsplineModuliY.getDevicePointer(), &pmeBsplineModuliZ.getDevicePointer(),
                    recipBoxVectorPointer[0], recipBoxVectorPointer[1], recipBoxVectorPointer[2]};
1015
            cu.executeKernel(pmeConvolutionKernel, convolutionArgs, gridSizeX*gridSizeY*gridSizeZ, 256);
1016

1017
            if (useCudaFFT) {
1018
1019
1020
1021
1022
1023
1024
1025
1026
                if (cu.getUseDoublePrecision()) {
                    cufftResult result = cufftExecZ2D(fftBackward, (double2*) pmeGrid2.getDevicePointer(), (double*) pmeGrid1.getDevicePointer());
                    if (result != CUFFT_SUCCESS)
                        throw OpenMMException("Error executing FFT: "+cu.intToString(result));
                } else {
                    cufftResult result = cufftExecC2R(fftBackward, (float2*) pmeGrid2.getDevicePointer(), (float*)  pmeGrid1.getDevicePointer());
                    if (result != CUFFT_SUCCESS)
                        throw OpenMMException("Error executing FFT: "+cu.intToString(result));
                }
1027
1028
1029
1030
            }
            else {
                fft->execFFT(pmeGrid2, pmeGrid1, false);
            }
1031

1032
1033
1034
1035
1036
1037
            void* interpolateArgs[] = {&cu.getPosq().getDevicePointer(), &cu.getForce().getDevicePointer(), &pmeGrid1.getDevicePointer(), cu.getPeriodicBoxSizePointer(),
                    cu.getInvPeriodicBoxSizePointer(), cu.getPeriodicBoxVecXPointer(), cu.getPeriodicBoxVecYPointer(), cu.getPeriodicBoxVecZPointer(),
                    recipBoxVectorPointer[0], recipBoxVectorPointer[1], recipBoxVectorPointer[2], &pmeAtomGridIndex.getDevicePointer(),
                    &charges.getDevicePointer()};
            cu.executeKernel(pmeInterpolateForceKernel, interpolateArgs, cu.getNumAtoms(), 128);
        }
1038

1039
1040
1041
1042
1043
1044
        if (doLJPME && hasLJ) {
            if (!hasCoulomb) {
                void* gridIndexArgs[] = {&cu.getPosq().getDevicePointer(), &pmeAtomGridIndex.getDevicePointer(), cu.getPeriodicBoxSizePointer(),
                        cu.getInvPeriodicBoxSizePointer(), cu.getPeriodicBoxVecXPointer(), cu.getPeriodicBoxVecYPointer(), cu.getPeriodicBoxVecZPointer(),
                        recipBoxVectorPointer[0], recipBoxVectorPointer[1], recipBoxVectorPointer[2]};
                cu.executeKernel(pmeDispersionGridIndexKernel, gridIndexArgs, cu.getNumAtoms());
1045

1046
1047
1048
                sort->sort(pmeAtomGridIndex);
                cu.clearBuffer(pmeEnergyBuffer);
            }
1049

1050
1051
            cu.clearBuffer(pmeSpreadDstGrid);
            void* spreadArgs[] = {&cu.getPosq().getDevicePointer(), &pmeSpreadDstGrid.getDevicePointer(), cu.getPeriodicBoxSizePointer(),
1052
1053
1054
                    cu.getInvPeriodicBoxSizePointer(), cu.getPeriodicBoxVecXPointer(), cu.getPeriodicBoxVecYPointer(), cu.getPeriodicBoxVecZPointer(),
                    recipBoxVectorPointer[0], recipBoxVectorPointer[1], recipBoxVectorPointer[2], &pmeAtomGridIndex.getDevicePointer(),
                    &sigmaEpsilon.getDevicePointer()};
1055
            cu.executeKernel(pmeDispersionSpreadChargeKernel, spreadArgs, PmeOrder*cu.getNumAtoms(), 128);
1056

1057
1058
1059
1060
            if (useFixedPointChargeSpreading) {
                void* finishSpreadArgs[] = {&pmeGrid2.getDevicePointer(), &pmeGrid1.getDevicePointer()};
                cu.executeKernel(pmeDispersionFinishSpreadChargeKernel, finishSpreadArgs, dispersionGridSizeX*dispersionGridSizeY*dispersionGridSizeZ, 256);
            }
1061

1062
            if (useCudaFFT) {
1063
1064
1065
1066
1067
1068
1069
1070
1071
                if (cu.getUseDoublePrecision()) {
                    cufftResult result = cufftExecD2Z(dispersionFftForward, (double*) pmeGrid1.getDevicePointer(), (double2*) pmeGrid2.getDevicePointer());
                    if (result != CUFFT_SUCCESS)
                        throw OpenMMException("Error executing FFT: "+cu.intToString(result));
                } else {
                    cufftResult result = cufftExecR2C(dispersionFftForward, (float*) pmeGrid1.getDevicePointer(), (float2*) pmeGrid2.getDevicePointer());
                    if (result != CUFFT_SUCCESS)
                        throw OpenMMException("Error executing FFT: "+cu.intToString(result));
                }
1072
1073
1074
1075
            }
            else {
                dispersionFft->execFFT(pmeGrid1, pmeGrid2, true);
            }
1076

1077
1078
1079
            if (includeEnergy) {
                void* computeEnergyArgs[] = {&pmeGrid2.getDevicePointer(), usePmeStream ? &pmeEnergyBuffer.getDevicePointer() : &cu.getEnergyBuffer().getDevicePointer(),
                        &pmeDispersionBsplineModuliX.getDevicePointer(), &pmeDispersionBsplineModuliY.getDevicePointer(), &pmeDispersionBsplineModuliZ.getDevicePointer(),
1080
                        recipBoxVectorPointer[0], recipBoxVectorPointer[1], recipBoxVectorPointer[2]};
1081
1082
                cu.executeKernel(pmeEvalDispersionEnergyKernel, computeEnergyArgs, dispersionGridSizeX*dispersionGridSizeY*dispersionGridSizeZ);
            }
1083

1084
1085
1086
            void* convolutionArgs[] = {&pmeGrid2.getDevicePointer(), &pmeDispersionBsplineModuliX.getDevicePointer(),
                    &pmeDispersionBsplineModuliY.getDevicePointer(), &pmeDispersionBsplineModuliZ.getDevicePointer(),
                    recipBoxVectorPointer[0], recipBoxVectorPointer[1], recipBoxVectorPointer[2]};
1087
            cu.executeKernel(pmeDispersionConvolutionKernel, convolutionArgs, dispersionGridSizeX*dispersionGridSizeY*dispersionGridSizeZ, 256);
1088

1089
            if (useCudaFFT) {
1090
1091
1092
1093
1094
1095
1096
1097
1098
                if (cu.getUseDoublePrecision()) {
                    cufftResult result = cufftExecZ2D(dispersionFftBackward, (double2*) pmeGrid2.getDevicePointer(), (double*) pmeGrid1.getDevicePointer());
                    if (result != CUFFT_SUCCESS)
                        throw OpenMMException("Error executing FFT: "+cu.intToString(result));
                } else {
                    cufftResult result = cufftExecC2R(dispersionFftBackward, (float2*) pmeGrid2.getDevicePointer(), (float*)  pmeGrid1.getDevicePointer());
                    if (result != CUFFT_SUCCESS)
                        throw OpenMMException("Error executing FFT: "+cu.intToString(result));
                }
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
            }
            else {
                dispersionFft->execFFT(pmeGrid2, pmeGrid1, false);
            }

            void* interpolateArgs[] = {&cu.getPosq().getDevicePointer(), &cu.getForce().getDevicePointer(), &pmeGrid1.getDevicePointer(), cu.getPeriodicBoxSizePointer(),
                    cu.getInvPeriodicBoxSizePointer(), cu.getPeriodicBoxVecXPointer(), cu.getPeriodicBoxVecYPointer(), cu.getPeriodicBoxVecZPointer(),
                    recipBoxVectorPointer[0], recipBoxVectorPointer[1], recipBoxVectorPointer[2], &pmeAtomGridIndex.getDevicePointer(),
                    &sigmaEpsilon.getDevicePointer()};
            cu.executeKernel(pmeInterpolateDispersionForceKernel, interpolateArgs, cu.getNumAtoms(), 128);
        }
        if (usePmeStream) {
            cuEventRecord(pmeSyncEvent, pmeStream);
            cu.restoreDefaultStream();
        }
    }
1115

1116
1117
1118
1119
1120
    if (dispersionCoefficient != 0.0 && includeDirect) {
        double4 boxSize = cu.getPeriodicBoxSize();
        energy += dispersionCoefficient/(boxSize.x*boxSize.y*boxSize.z);
    }
    return energy;
1121
1122
}

1123
void CudaCalcNonbondedForceKernel::copyParametersToContext(ContextImpl& context, const NonbondedForce& force, int firstParticle, int lastParticle, int firstException, int lastException) {
1124
1125
    // Make sure the new parameters are acceptable.
    
1126
    ContextSelector selector(cu);
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
    if (force.getNumParticles() != cu.getNumAtoms())
        throw OpenMMException("updateParametersInContext: The number of particles has changed");
    if (!hasCoulomb || !hasLJ) {
        for (int i = 0; i < force.getNumParticles(); i++) {
            double charge, sigma, epsilon;
            force.getParticleParameters(i, charge, sigma, epsilon);
            if (!hasCoulomb && charge != 0.0)
                throw OpenMMException("updateParametersInContext: The nonbonded force kernel does not include Coulomb interactions, because all charges were originally 0");
            if (!hasLJ && epsilon != 0.0)
                throw OpenMMException("updateParametersInContext: The nonbonded force kernel does not include Lennard-Jones interactions, because all epsilons were originally 0");
        }
1138
    }
1139
1140
1141
1142
1143
1144
1145
1146
    set<int> exceptionsWithOffsets;
    for (int i = 0; i < force.getNumExceptionParameterOffsets(); i++) {
        string param;
        int exception;
        double charge, sigma, epsilon;
        force.getExceptionParameterOffset(i, param, exception, charge, sigma, epsilon);
        exceptionsWithOffsets.insert(exception);
    }
1147
1148
1149
1150
1151
    vector<int> exceptions;
    for (int i = 0; i < force.getNumExceptions(); i++) {
        int particle1, particle2;
        double chargeProd, sigma, epsilon;
        force.getExceptionParameters(i, particle1, particle2, chargeProd, sigma, epsilon);
1152
1153
1154
1155
1156
        if (exceptionIndex.find(i) == exceptionIndex.end()) {
            if (chargeProd != 0.0 || epsilon != 0.0 || exceptionsWithOffsets.find(i) != exceptionsWithOffsets.end())
                throw OpenMMException("updateParametersInContext: The set of non-excluded exceptions has changed");
        }
        else
1157
            exceptions.push_back(i);
1158
    }
1159
1160
1161
1162
    int numContexts = cu.getPlatformData().contexts.size();
    int startIndex = cu.getContextIndex()*exceptions.size()/numContexts;
    int endIndex = (cu.getContextIndex()+1)*exceptions.size()/numContexts;
    int numExceptions = endIndex-startIndex;
1163
1164
    if (numExceptions != exceptionAtoms.size())
        throw OpenMMException("updateParametersInContext: The set of non-excluded exceptions has changed");
1165
1166
1167
    
    // Record the per-particle parameters.
    
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
    if (firstParticle <= lastParticle) {
        vector<float4> baseParticleParamVec(cu.getPaddedNumAtoms(), make_float4(0, 0, 0, 0));
        for (int i = 0; i < force.getNumParticles(); i++) {
            double charge, sigma, epsilon;
            force.getParticleParameters(i, charge, sigma, epsilon);
            baseParticleParamVec[i] = make_float4(charge, sigma, epsilon, 0);
        }
        baseParticleParams.uploadSubArray(&baseParticleParamVec[firstParticle], firstParticle, lastParticle-firstParticle+1);

        // Compute the self energy.

        ewaldSelfEnergy = 0.0;
1180
        totalCharge = 0.0;
1181
1182
1183
1184
        if (nonbondedMethod == Ewald || nonbondedMethod == PME || nonbondedMethod == LJPME) {
            if (cu.getContextIndex() == 0) {
                for (int i = 0; i < force.getNumParticles(); i++) {
                    ewaldSelfEnergy -= baseParticleParamVec[i].x*baseParticleParamVec[i].x*ONE_4PI_EPS0*alpha/sqrt(M_PI);
1185
                    totalCharge += baseParticleParamVec[i].x;
1186
1187
1188
1189
1190
                    if (doLJPME)
                        ewaldSelfEnergy += baseParticleParamVec[i].z*pow(baseParticleParamVec[i].y*dispersionAlpha, 6)/3.0;
                }
            }
        }
1191
    }
1192
1193
1194
    
    // Record the exceptions.
    
1195
    if (firstException <= lastException) {
1196
1197
        vector<float4> baseExceptionParamsVec(numExceptions);
        for (int i = 0; i < numExceptions; i++) {
1198
            int particle1, particle2;
1199
            double chargeProd, sigma, epsilon;
1200
1201
1202
            force.getExceptionParameters(exceptions[startIndex+i], particle1, particle2, chargeProd, sigma, epsilon);
            if (make_pair(particle1, particle2) != exceptionAtoms[i])
                throw OpenMMException("updateParametersInContext: The set of non-excluded exceptions has changed");
1203
1204
1205
            baseExceptionParamsVec[i] = make_float4(chargeProd, sigma, epsilon, 0);
        }
        baseExceptionParams.upload(baseExceptionParamsVec);
1206
    }
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252

    // Record parameter offsets.

    vector<vector<mm_float4> > particleOffsetVec(force.getNumParticles());
    vector<vector<mm_float4> > exceptionOffsetVec(numExceptions);
    for (int i = 0; i < force.getNumParticleParameterOffsets(); i++) {
        string param;
        int particle;
        double charge, sigma, epsilon;
        force.getParticleParameterOffset(i, param, particle, charge, sigma, epsilon);
        auto paramPos = find(paramNames.begin(), paramNames.end(), param);
        if (paramPos == paramNames.end())
            throw OpenMMException("updateParametersInContext: The parameter of a particle parameter offset has changed");
        int paramIndex = paramPos-paramNames.begin();
        particleOffsetVec[particle].push_back(mm_float4(charge, sigma, epsilon, paramIndex));
    }
    for (int i = 0; i < force.getNumExceptionParameterOffsets(); i++) {
        string param;
        int exception;
        double charge, sigma, epsilon;
        force.getExceptionParameterOffset(i, param, exception, charge, sigma, epsilon);
        int index = exceptionIndex[exception];
        if (index < startIndex || index >= endIndex)
            continue;
        auto paramPos = find(paramNames.begin(), paramNames.end(), param);
        if (paramPos == paramNames.end())
            throw OpenMMException("updateParametersInContext: The parameter of an exception parameter offset has changed");
        int paramIndex = paramPos-paramNames.begin();
        exceptionOffsetVec[index-startIndex].push_back(mm_float4(charge, sigma, epsilon, paramIndex));
    }
    if (max(force.getNumParticleParameterOffsets(), 1) != particleParamOffsets.getSize())
        throw OpenMMException("updateParametersInContext: The number of particle parameter offsets has changed");
    vector<mm_float4> p, e;
    for (int i = 0; i < particleOffsetVec.size(); i++)
        for (int j = 0; j < particleOffsetVec[i].size(); j++)
            p.push_back(particleOffsetVec[i][j]);
    for (int i = 0; i < exceptionOffsetVec.size(); i++)
        for (int j = 0; j < exceptionOffsetVec[i].size(); j++)
            e.push_back(exceptionOffsetVec[i][j]);
    if (force.getNumParticleParameterOffsets() > 0)
        particleParamOffsets.upload(p);
    if (max((int) e.size(), 1) != exceptionParamOffsets.getSize())
        throw OpenMMException("updateParametersInContext: The number of exception parameter offsets has changed");
    if (e.size() > 0)
        exceptionParamOffsets.upload(e);

1253
    // Compute other values.
1254

1255
1256
    if (force.getUseDispersionCorrection() && cu.getContextIndex() == 0 && (nonbondedMethod == CutoffPeriodic || nonbondedMethod == Ewald || nonbondedMethod == PME))
        dispersionCoefficient = NonbondedForceImpl::calcDispersionCorrection(context.getSystem(), force);
1257
1258
    cu.invalidateMolecules(info, firstParticle <= lastParticle || force.getNumParticleParameterOffsets() > 0,
                           firstException <= lastException || force.getNumExceptionParameterOffsets() > 0);
1259
    recomputeParams = true;
1260
1261
}

1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
void CudaCalcNonbondedForceKernel::getPMEParameters(double& alpha, int& nx, int& ny, int& nz) const {
    if (nonbondedMethod != PME)
        throw OpenMMException("getPMEParametersInContext: This Context is not using PME");
    if (cu.getPlatformData().useCpuPme)
        cpuPme.getAs<CalcPmeReciprocalForceKernel>().getPMEParameters(alpha, nx, ny, nz);
    else {
        alpha = this->alpha;
        nx = gridSizeX;
        ny = gridSizeY;
        nz = gridSizeZ;
    }
1273
1274
}

1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
void CudaCalcNonbondedForceKernel::getLJPMEParameters(double& alpha, int& nx, int& ny, int& nz) const {
    if (!doLJPME)
        throw OpenMMException("getPMEParametersInContext: This Context is not using PME");
    if (cu.getPlatformData().useCpuPme)
        //cpuPme.getAs<CalcPmeReciprocalForceKernel>().getLJPMEParameters(alpha, nx, ny, nz);
        throw OpenMMException("getPMEParametersInContext: CPUPME has not been implemented for LJPME yet.");
    else {
        alpha = this->dispersionAlpha;
        nx = dispersionGridSizeX;
        ny = dispersionGridSizeY;
        nz = dispersionGridSizeZ;
    }
1287
}