AmoebaCudaKernels.cpp 142 KB
Newer Older
1
2
3
4
5
6
7
8
/* -------------------------------------------------------------------------- *
 *                               OpenMMAmoeba                                 *
 * -------------------------------------------------------------------------- *
 * 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) 2008-2015 Stanford University and the Authors.      *
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
 * Authors: Peter Eastman, Mark Friedrichs                                    *
 * 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/>.      *
 * -------------------------------------------------------------------------- */

27
28
29
#ifdef WIN32
  #define _USE_MATH_DEFINES // Needed to get M_PI
#endif
30
31
32
#include "AmoebaCudaKernels.h"
#include "CudaAmoebaKernelSources.h"
#include "openmm/internal/ContextImpl.h"
33
#include "openmm/internal/AmoebaGeneralizedKirkwoodForceImpl.h"
34
35
36
#include "openmm/internal/AmoebaMultipoleForceImpl.h"
#include "openmm/internal/AmoebaWcaDispersionForceImpl.h"
#include "openmm/internal/AmoebaTorsionTorsionForceImpl.h"
37
#include "openmm/internal/AmoebaVdwForceImpl.h"
38
39
#include "openmm/internal/NonbondedForceImpl.h"
#include "CudaBondedUtilities.h"
40
#include "CudaFFT3D.h"
41
42
#include "CudaForceInfo.h"
#include "CudaKernelSources.h"
43
#include "CudaNonbondedUtilities.h"
peastman's avatar
peastman committed
44
#include "jama_svd.h"
45

46
#include <algorithm>
47
48
49
50
51
52
53
54
#include <cmath>
#ifdef _MSC_VER
#include <windows.h>
#endif

using namespace OpenMM;
using namespace std;

55
56
57
58
59
60
61
#define CHECK_RESULT(result) \
    if (result != CUDA_SUCCESS) { \
        std::stringstream m; \
        m<<errorMessage<<": "<<cu.getErrorString(result)<<" ("<<result<<")"<<" at "<<__FILE__<<":"<<__LINE__; \
        throw OpenMMException(m.str());\
    }

62
/* -------------------------------------------------------------------------- *
63
 *                            AmoebaBondForce                                 *
64
65
 * -------------------------------------------------------------------------- */

66
class CudaCalcAmoebaBondForceKernel::ForceInfo : public CudaForceInfo {
67
public:
68
    ForceInfo(const AmoebaBondForce& force) : force(force) {
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
    }
    int getNumParticleGroups() {
        return force.getNumBonds();
    }
    void getParticlesInGroup(int index, std::vector<int>& particles) {
        int particle1, particle2;
        double length, k;
        force.getBondParameters(index, particle1, particle2, length, k);
        particles.resize(2);
        particles[0] = particle1;
        particles[1] = particle2;
    }
    bool areGroupsIdentical(int group1, int group2) {
        int particle1, particle2;
        double length1, length2, k1, k2;
        force.getBondParameters(group1, particle1, particle2, length1, k1);
        force.getBondParameters(group2, particle1, particle2, length2, k2);
        return (length1 == length2 && k1 == k2);
    }
private:
89
    const AmoebaBondForce& force;
90
91
};

92
CudaCalcAmoebaBondForceKernel::CudaCalcAmoebaBondForceKernel(std::string name, const Platform& platform, CudaContext& cu, const System& system) : 
93
                CalcAmoebaBondForceKernel(name, platform), cu(cu), system(system), params(NULL) {
94
95
}

96
CudaCalcAmoebaBondForceKernel::~CudaCalcAmoebaBondForceKernel() {
97
98
99
100
101
    cu.setAsCurrent();
    if (params != NULL)
        delete params;
}

102
void CudaCalcAmoebaBondForceKernel::initialize(const System& system, const AmoebaBondForce& force) {
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
    cu.setAsCurrent();
    int numContexts = cu.getPlatformData().contexts.size();
    int startIndex = cu.getContextIndex()*force.getNumBonds()/numContexts;
    int endIndex = (cu.getContextIndex()+1)*force.getNumBonds()/numContexts;
    numBonds = endIndex-startIndex;
    if (numBonds == 0)
        return;
    vector<vector<int> > atoms(numBonds, vector<int>(2));
    params = CudaArray::create<float2>(cu, numBonds, "bondParams");
    vector<float2> paramVector(numBonds);
    for (int i = 0; i < numBonds; i++) {
        double length, k;
        force.getBondParameters(startIndex+i, atoms[i][0], atoms[i][1], length, k);
        paramVector[i] = make_float2((float) length, (float) k);
    }
    params->upload(paramVector);
    map<string, string> replacements;
    replacements["COMPUTE_FORCE"] = CudaAmoebaKernelSources::amoebaBondForce;
    replacements["PARAMS"] = cu.getBondedUtilities().addArgument(params->getDevicePointer(), "float2");
122
123
    replacements["CUBIC_K"] = cu.doubleToString(force.getAmoebaGlobalBondCubic());
    replacements["QUARTIC_K"] = cu.doubleToString(force.getAmoebaGlobalBondQuartic());
124
125
126
127
    cu.getBondedUtilities().addInteraction(atoms, cu.replaceStrings(CudaKernelSources::bondForce, replacements), force.getForceGroup());
    cu.addForce(new ForceInfo(force));
}

128
double CudaCalcAmoebaBondForceKernel::execute(ContextImpl& context, bool includeForces, bool includeEnergy) {
129
130
131
    return 0.0;
}

132
133
134
135
136
137
138
void CudaCalcAmoebaBondForceKernel::copyParametersToContext(ContextImpl& context, const AmoebaBondForce& force) {
    cu.setAsCurrent();
    int numContexts = cu.getPlatformData().contexts.size();
    int startIndex = cu.getContextIndex()*force.getNumBonds()/numContexts;
    int endIndex = (cu.getContextIndex()+1)*force.getNumBonds()/numContexts;
    if (numBonds != endIndex-startIndex)
        throw OpenMMException("updateParametersInContext: The number of bonds has changed");
139
140
    if (numBonds == 0)
        return;
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
    
    // Record the per-bond parameters.
    
    vector<float2> paramVector(numBonds);
    for (int i = 0; i < numBonds; i++) {
        int atom1, atom2;
        double length, k;
        force.getBondParameters(startIndex+i, atom1, atom2, length, k);
        paramVector[i] = make_float2((float) length, (float) k);
    }
    params->upload(paramVector);
    
    // Mark that the current reordering may be invalid.
    
    cu.invalidateMolecules();
}

158
/* -------------------------------------------------------------------------- *
159
 *                            AmoebaAngleForce                                *
160
161
 * -------------------------------------------------------------------------- */

162
class CudaCalcAmoebaAngleForceKernel::ForceInfo : public CudaForceInfo {
163
public:
164
    ForceInfo(const AmoebaAngleForce& force) : force(force) {
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
    }
    int getNumParticleGroups() {
        return force.getNumAngles();
    }
    void getParticlesInGroup(int index, std::vector<int>& particles) {
        int particle1, particle2, particle3;
        double angle, k;
        force.getAngleParameters(index, particle1, particle2, particle3, angle, k);
        particles.resize(3);
        particles[0] = particle1;
        particles[1] = particle2;
        particles[2] = particle3;
    }
    bool areGroupsIdentical(int group1, int group2) {
        int particle1, particle2, particle3;
        double angle1, angle2, k1, k2;
        force.getAngleParameters(group1, particle1, particle2, particle3, angle1, k1);
        force.getAngleParameters(group2, particle1, particle2, particle3, angle2, k2);
        return (angle1 == angle2 && k1 == k2);
    }
private:
186
    const AmoebaAngleForce& force;
187
188
};

189
CudaCalcAmoebaAngleForceKernel::CudaCalcAmoebaAngleForceKernel(std::string name, const Platform& platform, CudaContext& cu, const System& system) :
190
            CalcAmoebaAngleForceKernel(name, platform), cu(cu), system(system), params(NULL) {
191
192
}

193
CudaCalcAmoebaAngleForceKernel::~CudaCalcAmoebaAngleForceKernel() {
194
195
196
197
198
    cu.setAsCurrent();
    if (params != NULL)
        delete params;
}

199
void CudaCalcAmoebaAngleForceKernel::initialize(const System& system, const AmoebaAngleForce& force) {
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
    cu.setAsCurrent();
    int numContexts = cu.getPlatformData().contexts.size();
    int startIndex = cu.getContextIndex()*force.getNumAngles()/numContexts;
    int endIndex = (cu.getContextIndex()+1)*force.getNumAngles()/numContexts;
    numAngles = endIndex-startIndex;
    if (numAngles == 0)
        return;
    vector<vector<int> > atoms(numAngles, vector<int>(3));
    params = CudaArray::create<float2>(cu, numAngles, "angleParams");
    vector<float2> paramVector(numAngles);
    for (int i = 0; i < numAngles; i++) {
        double angle, k;
        force.getAngleParameters(startIndex+i, atoms[i][0], atoms[i][1], atoms[i][2], angle, k);
        paramVector[i] = make_float2((float) angle, (float) k);
    }
    params->upload(paramVector);
    map<string, string> replacements;
    replacements["COMPUTE_FORCE"] = CudaAmoebaKernelSources::amoebaAngleForce;
    replacements["PARAMS"] = cu.getBondedUtilities().addArgument(params->getDevicePointer(), "float2");
219
220
221
222
    replacements["CUBIC_K"] = cu.doubleToString(force.getAmoebaGlobalAngleCubic());
    replacements["QUARTIC_K"] = cu.doubleToString(force.getAmoebaGlobalAngleQuartic());
    replacements["PENTIC_K"] = cu.doubleToString(force.getAmoebaGlobalAnglePentic());
    replacements["SEXTIC_K"] = cu.doubleToString(force.getAmoebaGlobalAngleSextic());
223
224
225
226
227
    replacements["RAD_TO_DEG"] = cu.doubleToString(180/M_PI);
    cu.getBondedUtilities().addInteraction(atoms, cu.replaceStrings(CudaKernelSources::angleForce, replacements), force.getForceGroup());
    cu.addForce(new ForceInfo(force));
}

228
double CudaCalcAmoebaAngleForceKernel::execute(ContextImpl& context, bool includeForces, bool includeEnergy) {
229
230
231
    return 0.0;
}

232
233
234
235
236
237
238
void CudaCalcAmoebaAngleForceKernel::copyParametersToContext(ContextImpl& context, const AmoebaAngleForce& force) {
    cu.setAsCurrent();
    int numContexts = cu.getPlatformData().contexts.size();
    int startIndex = cu.getContextIndex()*force.getNumAngles()/numContexts;
    int endIndex = (cu.getContextIndex()+1)*force.getNumAngles()/numContexts;
    if (numAngles != endIndex-startIndex)
        throw OpenMMException("updateParametersInContext: The number of angles has changed");
239
240
    if (numAngles == 0)
        return;
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
    
    // Record the per-angle parameters.
    
    vector<float2> paramVector(numAngles);
    for (int i = 0; i < numAngles; i++) {
        int atom1, atom2, atom3;
        double angle, k;
        force.getAngleParameters(startIndex+i, atom1, atom2, atom3, angle, k);
        paramVector[i] = make_float2((float) angle, (float) k);
    }
    params->upload(paramVector);
    
    // Mark that the current reordering may be invalid.
    
    cu.invalidateMolecules();
}

258
/* -------------------------------------------------------------------------- *
259
 *                            AmoebaInPlaneAngleForce                         *
260
261
 * -------------------------------------------------------------------------- */

262
class CudaCalcAmoebaInPlaneAngleForceKernel::ForceInfo : public CudaForceInfo {
263
public:
264
    ForceInfo(const AmoebaInPlaneAngleForce& force) : force(force) {
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
    }
    int getNumParticleGroups() {
        return force.getNumAngles();
    }
    void getParticlesInGroup(int index, std::vector<int>& particles) {
        int particle1, particle2, particle3, particle4;
        double angle, k;
        force.getAngleParameters(index, particle1, particle2, particle3, particle4, angle, k);
        particles.resize(4);
        particles[0] = particle1;
        particles[1] = particle2;
        particles[2] = particle3;
        particles[3] = particle4;
    }
    bool areGroupsIdentical(int group1, int group2) {
        int particle1, particle2, particle3, particle4;
        double angle1, angle2, k1, k2;
        force.getAngleParameters(group1, particle1, particle2, particle3, particle4, angle1, k1);
        force.getAngleParameters(group2, particle1, particle2, particle3, particle4, angle2, k2);
        return (angle1 == angle2 && k1 == k2);
    }
private:
287
    const AmoebaInPlaneAngleForce& force;
288
289
};

290
CudaCalcAmoebaInPlaneAngleForceKernel::CudaCalcAmoebaInPlaneAngleForceKernel(std::string name, const Platform& platform, CudaContext& cu, const System& system) : 
291
          CalcAmoebaInPlaneAngleForceKernel(name, platform), cu(cu), system(system), params(NULL) {
292
293
}

294
CudaCalcAmoebaInPlaneAngleForceKernel::~CudaCalcAmoebaInPlaneAngleForceKernel() {
295
296
297
298
299
    cu.setAsCurrent();
    if (params != NULL)
        delete params;
}

300
void CudaCalcAmoebaInPlaneAngleForceKernel::initialize(const System& system, const AmoebaInPlaneAngleForce& force) {
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
    cu.setAsCurrent();
    int numContexts = cu.getPlatformData().contexts.size();
    int startIndex = cu.getContextIndex()*force.getNumAngles()/numContexts;
    int endIndex = (cu.getContextIndex()+1)*force.getNumAngles()/numContexts;
    numAngles = endIndex-startIndex;
    if (numAngles == 0)
        return;
    vector<vector<int> > atoms(numAngles, vector<int>(4));
    params = CudaArray::create<float2>(cu, numAngles, "angleParams");
    vector<float2> paramVector(numAngles);
    for (int i = 0; i < numAngles; i++) {
        double angle, k;
        force.getAngleParameters(startIndex+i, atoms[i][0], atoms[i][1], atoms[i][2], atoms[i][3], angle, k);
        paramVector[i] = make_float2((float) angle, (float) k);
    }
    params->upload(paramVector);
    map<string, string> replacements;
    replacements["PARAMS"] = cu.getBondedUtilities().addArgument(params->getDevicePointer(), "float2");
319
320
321
322
    replacements["CUBIC_K"] = cu.doubleToString(force.getAmoebaGlobalInPlaneAngleCubic());
    replacements["QUARTIC_K"] = cu.doubleToString(force.getAmoebaGlobalInPlaneAngleQuartic());
    replacements["PENTIC_K"] = cu.doubleToString(force.getAmoebaGlobalInPlaneAnglePentic());
    replacements["SEXTIC_K"] = cu.doubleToString(force.getAmoebaGlobalInPlaneAngleSextic());
323
324
325
326
327
    replacements["RAD_TO_DEG"] = cu.doubleToString(180/M_PI);
    cu.getBondedUtilities().addInteraction(atoms, cu.replaceStrings(CudaAmoebaKernelSources::amoebaInPlaneForce, replacements), force.getForceGroup());
    cu.addForce(new ForceInfo(force));
}

328
double CudaCalcAmoebaInPlaneAngleForceKernel::execute(ContextImpl& context, bool includeForces, bool includeEnergy) {
329
330
331
    return 0.0;
}

332
333
334
335
336
337
338
void CudaCalcAmoebaInPlaneAngleForceKernel::copyParametersToContext(ContextImpl& context, const AmoebaInPlaneAngleForce& force) {
    cu.setAsCurrent();
    int numContexts = cu.getPlatformData().contexts.size();
    int startIndex = cu.getContextIndex()*force.getNumAngles()/numContexts;
    int endIndex = (cu.getContextIndex()+1)*force.getNumAngles()/numContexts;
    if (numAngles != endIndex-startIndex)
        throw OpenMMException("updateParametersInContext: The number of in-plane angles has changed");
339
340
    if (numAngles == 0)
        return;
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
    
    // Record the per-angle parameters.
    
    vector<float2> paramVector(numAngles);
    for (int i = 0; i < numAngles; i++) {
        int atom1, atom2, atom3, atom4;
        double angle, k;
        force.getAngleParameters(startIndex+i, atom1, atom2, atom3, atom4, angle, k);
        paramVector[i] = make_float2((float) angle, (float) k);
    }
    params->upload(paramVector);
    
    // Mark that the current reordering may be invalid.
    
    cu.invalidateMolecules();
}

358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
/* -------------------------------------------------------------------------- *
  *                              AmoebaPiTorsion                              *
 * -------------------------------------------------------------------------- */

class CudaCalcAmoebaPiTorsionForceKernel::ForceInfo : public CudaForceInfo {
public:
    ForceInfo(const AmoebaPiTorsionForce& force) : force(force) {
    }
    int getNumParticleGroups() {
        return force.getNumPiTorsions();
    }
    void getParticlesInGroup(int index, std::vector<int>& particles) {
        int particle1, particle2, particle3, particle4, particle5, particle6;
        double k;
        force.getPiTorsionParameters(index, particle1, particle2, particle3, particle4, particle5, particle6, k);
        particles.resize(6);
        particles[0] = particle1;
        particles[1] = particle2;
        particles[2] = particle3;
        particles[3] = particle4;
        particles[4] = particle5;
        particles[5] = particle6;
    }
    bool areGroupsIdentical(int group1, int group2) {
        int particle1, particle2, particle3, particle4, particle5, particle6;
        double k1, k2;
        force.getPiTorsionParameters(group1, particle1, particle2, particle3, particle4, particle5, particle6, k1);
        force.getPiTorsionParameters(group2, particle1, particle2, particle3, particle4, particle5, particle6, k2);
        return (k1 == k2);
    }
private:
    const AmoebaPiTorsionForce& force;
};

392
CudaCalcAmoebaPiTorsionForceKernel::CudaCalcAmoebaPiTorsionForceKernel(std::string name, const Platform& platform, CudaContext& cu, const System& system) :
393
         CalcAmoebaPiTorsionForceKernel(name, platform), cu(cu), system(system), params(NULL) {
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
}

CudaCalcAmoebaPiTorsionForceKernel::~CudaCalcAmoebaPiTorsionForceKernel() {
    cu.setAsCurrent();
    if (params != NULL)
        delete params;
}

void CudaCalcAmoebaPiTorsionForceKernel::initialize(const System& system, const AmoebaPiTorsionForce& force) {
    cu.setAsCurrent();
    int numContexts = cu.getPlatformData().contexts.size();
    int startIndex = cu.getContextIndex()*force.getNumPiTorsions()/numContexts;
    int endIndex = (cu.getContextIndex()+1)*force.getNumPiTorsions()/numContexts;
    numPiTorsions = endIndex-startIndex;
    if (numPiTorsions == 0)
        return;
    vector<vector<int> > atoms(numPiTorsions, vector<int>(6));
    params = CudaArray::create<float>(cu, numPiTorsions, "piTorsionParams");
    vector<float> paramVector(numPiTorsions);
    for (int i = 0; i < numPiTorsions; i++) {
        double k;
        force.getPiTorsionParameters(startIndex+i, atoms[i][0], atoms[i][1], atoms[i][2], atoms[i][3], atoms[i][4], atoms[i][5], k);
        paramVector[i] = (float) k;
    }
    params->upload(paramVector);
    map<string, string> replacements;
    replacements["PARAMS"] = cu.getBondedUtilities().addArgument(params->getDevicePointer(), "float");
    cu.getBondedUtilities().addInteraction(atoms, cu.replaceStrings(CudaAmoebaKernelSources::amoebaPiTorsionForce, replacements), force.getForceGroup());
    cu.addForce(new ForceInfo(force));
}

double CudaCalcAmoebaPiTorsionForceKernel::execute(ContextImpl& context, bool includeForces, bool includeEnergy) {
    return 0.0;
}

429
430
431
432
433
434
435
void CudaCalcAmoebaPiTorsionForceKernel::copyParametersToContext(ContextImpl& context, const AmoebaPiTorsionForce& force) {
    cu.setAsCurrent();
    int numContexts = cu.getPlatformData().contexts.size();
    int startIndex = cu.getContextIndex()*force.getNumPiTorsions()/numContexts;
    int endIndex = (cu.getContextIndex()+1)*force.getNumPiTorsions()/numContexts;
    if (numPiTorsions != endIndex-startIndex)
        throw OpenMMException("updateParametersInContext: The number of torsions has changed");
436
437
    if (numPiTorsions == 0)
        return;
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
    
    // Record the per-torsion parameters.
    
    vector<float> paramVector(numPiTorsions);
    for (int i = 0; i < numPiTorsions; i++) {
        int atom1, atom2, atom3, atom4, atom5, atom6;
        double k;
        force.getPiTorsionParameters(startIndex+i, atom1, atom2, atom3, atom4, atom5, atom6, k);
        paramVector[i] = (float) k;
    }
    params->upload(paramVector);
    
    // Mark that the current reordering may be invalid.
    
    cu.invalidateMolecules();
}

455
456
457
458
459
460
461
462
463
464
465
466
467
/* -------------------------------------------------------------------------- *
 *                           AmoebaStretchBend                                *
 * -------------------------------------------------------------------------- */

class CudaCalcAmoebaStretchBendForceKernel::ForceInfo : public CudaForceInfo {
public:
    ForceInfo(const AmoebaStretchBendForce& force) : force(force) {
    }
    int getNumParticleGroups() {
        return force.getNumStretchBends();
    }
    void getParticlesInGroup(int index, std::vector<int>& particles) {
        int particle1, particle2, particle3;
468
469
        double lengthAB, lengthCB, angle, k1, k2;
        force.getStretchBendParameters(index, particle1, particle2, particle3, lengthAB, lengthCB, angle, k1, k2);
470
471
472
473
474
475
476
        particles.resize(3);
        particles[0] = particle1;
        particles[1] = particle2;
        particles[2] = particle3;
    }
    bool areGroupsIdentical(int group1, int group2) {
        int particle1, particle2, particle3;
477
478
479
480
        double lengthAB1, lengthAB2, lengthCB1, lengthCB2, angle1, angle2, k11, k12, k21, k22;
        force.getStretchBendParameters(group1, particle1, particle2, particle3, lengthAB1, lengthCB1, angle1, k11, k12);
        force.getStretchBendParameters(group2, particle1, particle2, particle3, lengthAB2, lengthCB2, angle2, k21, k22);
        return (lengthAB1 == lengthAB2 && lengthCB1 == lengthCB2 && angle1 == angle2 && k11 == k21 && k12 == k22);
481
482
483
484
485
    }
private:
    const AmoebaStretchBendForce& force;
};

486
CudaCalcAmoebaStretchBendForceKernel::CudaCalcAmoebaStretchBendForceKernel(std::string name, const Platform& platform, CudaContext& cu, const System& system) :
Jason Swails's avatar
Jason Swails committed
487
                   CalcAmoebaStretchBendForceKernel(name, platform), cu(cu), system(system), params1(NULL), params2(NULL) {
488
489
490
491
}

CudaCalcAmoebaStretchBendForceKernel::~CudaCalcAmoebaStretchBendForceKernel() {
    cu.setAsCurrent();
492
493
494
495
    if (params1 != NULL)
        delete params1;
    if (params2 != NULL)
        delete params2;
496
497
498
499
500
501
502
503
504
505
506
}

void CudaCalcAmoebaStretchBendForceKernel::initialize(const System& system, const AmoebaStretchBendForce& force) {
    cu.setAsCurrent();
    int numContexts = cu.getPlatformData().contexts.size();
    int startIndex = cu.getContextIndex()*force.getNumStretchBends()/numContexts;
    int endIndex = (cu.getContextIndex()+1)*force.getNumStretchBends()/numContexts;
    numStretchBends = endIndex-startIndex;
    if (numStretchBends == 0)
        return;
    vector<vector<int> > atoms(numStretchBends, vector<int>(3));
507
508
509
510
    params1 = CudaArray::create<float3>(cu, numStretchBends, "stretchBendParams");
    params2 = CudaArray::create<float2>(cu, numStretchBends, "stretchBendForceConstants");
    vector<float3> paramVector(numStretchBends);
    vector<float2> paramVectorK(numStretchBends);
511
    for (int i = 0; i < numStretchBends; i++) {
512
513
514
515
        double lengthAB, lengthCB, angle, k1, k2;
        force.getStretchBendParameters(startIndex+i, atoms[i][0], atoms[i][1], atoms[i][2], lengthAB, lengthCB, angle, k1, k2);
        paramVector[i] = make_float3((float) lengthAB, (float) lengthCB, (float) angle);
        paramVectorK[i] = make_float2((float) k1, (float) k2);
516
    }
517
518
    params1->upload(paramVector);
    params2->upload(paramVectorK);
519
    map<string, string> replacements;
520
521
    replacements["PARAMS"] = cu.getBondedUtilities().addArgument(params1->getDevicePointer(), "float3");
    replacements["FORCE_CONSTANTS"] = cu.getBondedUtilities().addArgument(params2->getDevicePointer(), "float2");
522
523
524
525
526
527
    replacements["RAD_TO_DEG"] = cu.doubleToString(180/M_PI);
    cu.getBondedUtilities().addInteraction(atoms, cu.replaceStrings(CudaAmoebaKernelSources::amoebaStretchBendForce, replacements), force.getForceGroup());
    cu.addForce(new ForceInfo(force));
}

double CudaCalcAmoebaStretchBendForceKernel::execute(ContextImpl& context, bool includeForces, bool includeEnergy) {
528
    return 0.0;
529
530
}

531
532
533
534
535
536
537
void CudaCalcAmoebaStretchBendForceKernel::copyParametersToContext(ContextImpl& context, const AmoebaStretchBendForce& force) {
    cu.setAsCurrent();
    int numContexts = cu.getPlatformData().contexts.size();
    int startIndex = cu.getContextIndex()*force.getNumStretchBends()/numContexts;
    int endIndex = (cu.getContextIndex()+1)*force.getNumStretchBends()/numContexts;
    if (numStretchBends != endIndex-startIndex)
        throw OpenMMException("updateParametersInContext: The number of bend-stretch terms has changed");
538
539
    if (numStretchBends == 0)
        return;
540
541
542
    
    // Record the per-stretch-bend parameters.
    
Jason Swails's avatar
Jason Swails committed
543
544
    vector<float3> paramVector(numStretchBends);
    vector<float2> paramVector1(numStretchBends);
545
546
    for (int i = 0; i < numStretchBends; i++) {
        int atom1, atom2, atom3;
Jason Swails's avatar
Jason Swails committed
547
548
549
        double lengthAB, lengthCB, angle, k1, k2;
        force.getStretchBendParameters(startIndex+i, atom1, atom2, atom3, lengthAB, lengthCB, angle, k1, k2);
        paramVector[i] = make_float3((float) lengthAB, (float) lengthCB, (float) angle);
Jason Swails's avatar
Jason Swails committed
550
        paramVector1[i] = make_float2((float) k1, (float) k2);
551
    }
Jason Swails's avatar
Jason Swails committed
552
553
    params1->upload(paramVector);
    params2->upload(paramVector1);
554
555
556
557
558
559
    
    // Mark that the current reordering may be invalid.
    
    cu.invalidateMolecules();
}

560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
/* -------------------------------------------------------------------------- *
 *                           AmoebaOutOfPlaneBend                             *
 * -------------------------------------------------------------------------- */

class CudaCalcAmoebaOutOfPlaneBendForceKernel::ForceInfo : public CudaForceInfo {
public:
    ForceInfo(const AmoebaOutOfPlaneBendForce& force) : force(force) {
    }
    int getNumParticleGroups() {
        return force.getNumOutOfPlaneBends();
    }
    void getParticlesInGroup(int index, std::vector<int>& particles) {
        int particle1, particle2, particle3, particle4;
        double k;
        force.getOutOfPlaneBendParameters(index, particle1, particle2, particle3, particle4, k);
        particles.resize(4);
        particles[0] = particle1;
        particles[1] = particle2;
        particles[2] = particle3;
        particles[3] = particle4;
    }
    bool areGroupsIdentical(int group1, int group2) {
        int particle1, particle2, particle3, particle4;
        double k1, k2;
        force.getOutOfPlaneBendParameters(group1, particle1, particle2, particle3, particle4, k1);
        force.getOutOfPlaneBendParameters(group2, particle1, particle2, particle3, particle4, k2);
        return (k1 == k2);
    }
private:
    const AmoebaOutOfPlaneBendForce& force;
};

592
CudaCalcAmoebaOutOfPlaneBendForceKernel::CudaCalcAmoebaOutOfPlaneBendForceKernel(std::string name, const Platform& platform, CudaContext& cu, const System& system) :
593
594
595
596
597
598
599
600
601
602
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
          CalcAmoebaOutOfPlaneBendForceKernel(name, platform), cu(cu), system(system), params(NULL) {
}

CudaCalcAmoebaOutOfPlaneBendForceKernel::~CudaCalcAmoebaOutOfPlaneBendForceKernel() {
    cu.setAsCurrent();
    if (params != NULL)
        delete params;
}

void CudaCalcAmoebaOutOfPlaneBendForceKernel::initialize(const System& system, const AmoebaOutOfPlaneBendForce& force) {
    cu.setAsCurrent();
    int numContexts = cu.getPlatformData().contexts.size();
    int startIndex = cu.getContextIndex()*force.getNumOutOfPlaneBends()/numContexts;
    int endIndex = (cu.getContextIndex()+1)*force.getNumOutOfPlaneBends()/numContexts;
    numOutOfPlaneBends = endIndex-startIndex;
    if (numOutOfPlaneBends == 0)
        return;
    vector<vector<int> > atoms(numOutOfPlaneBends, vector<int>(4));
    params = CudaArray::create<float>(cu, numOutOfPlaneBends, "outOfPlaneParams");
    vector<float> paramVector(numOutOfPlaneBends);
    for (int i = 0; i < numOutOfPlaneBends; i++) {
        double k;
        force.getOutOfPlaneBendParameters(startIndex+i, atoms[i][0], atoms[i][1], atoms[i][2], atoms[i][3], k);
        paramVector[i] = (float) k;
    }
    params->upload(paramVector);
    map<string, string> replacements;
    replacements["PARAMS"] = cu.getBondedUtilities().addArgument(params->getDevicePointer(), "float");
    replacements["CUBIC_K"] = cu.doubleToString(force.getAmoebaGlobalOutOfPlaneBendCubic());
    replacements["QUARTIC_K"] = cu.doubleToString(force.getAmoebaGlobalOutOfPlaneBendQuartic());
    replacements["PENTIC_K"] = cu.doubleToString(force.getAmoebaGlobalOutOfPlaneBendPentic());
    replacements["SEXTIC_K"] = cu.doubleToString(force.getAmoebaGlobalOutOfPlaneBendSextic());
    replacements["RAD_TO_DEG"] = cu.doubleToString(180/M_PI);
    cu.getBondedUtilities().addInteraction(atoms, cu.replaceStrings(CudaAmoebaKernelSources::amoebaOutOfPlaneBendForce, replacements), force.getForceGroup());
    cu.addForce(new ForceInfo(force));
}

double CudaCalcAmoebaOutOfPlaneBendForceKernel::execute(ContextImpl& context, bool includeForces, bool includeEnergy) {
    return 0.0;
}

634
635
636
637
638
639
640
void CudaCalcAmoebaOutOfPlaneBendForceKernel::copyParametersToContext(ContextImpl& context, const AmoebaOutOfPlaneBendForce& force) {
    cu.setAsCurrent();
    int numContexts = cu.getPlatformData().contexts.size();
    int startIndex = cu.getContextIndex()*force.getNumOutOfPlaneBends()/numContexts;
    int endIndex = (cu.getContextIndex()+1)*force.getNumOutOfPlaneBends()/numContexts;
    if (numOutOfPlaneBends != endIndex-startIndex)
        throw OpenMMException("updateParametersInContext: The number of out-of-plane bends has changed");
641
642
    if (numOutOfPlaneBends == 0)
        return;
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
    
    // Record the per-bend parameters.
    
    vector<float> paramVector(numOutOfPlaneBends);
    for (int i = 0; i < numOutOfPlaneBends; i++) {
        int atom1, atom2, atom3, atom4;
        double k;
        force.getOutOfPlaneBendParameters(startIndex+i, atom1, atom2, atom3, atom4, k);
        paramVector[i] = (float) k;
    }
    params->upload(paramVector);
    
    // Mark that the current reordering may be invalid.
    
    cu.invalidateMolecules();
}

660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
/* -------------------------------------------------------------------------- *
 *                           AmoebaTorsionTorsion                             *
 * -------------------------------------------------------------------------- */

class CudaCalcAmoebaTorsionTorsionForceKernel::ForceInfo : public CudaForceInfo {
public:
    ForceInfo(const AmoebaTorsionTorsionForce& force) : force(force) {
    }
    int getNumParticleGroups() {
        return force.getNumTorsionTorsions();
    }
    void getParticlesInGroup(int index, std::vector<int>& particles) {
        int particle1, particle2, particle3, particle4, particle5, chiralCheckAtomIndex, gridIndex;
        force.getTorsionTorsionParameters(index, particle1, particle2, particle3, particle4, particle5, chiralCheckAtomIndex, gridIndex);
        particles.resize(5);
        particles[0] = particle1;
        particles[1] = particle2;
        particles[2] = particle3;
        particles[3] = particle4;
        particles[4] = particle5;
    }
    bool areGroupsIdentical(int group1, int group2) {
        int particle1, particle2, particle3, particle4, particle5;
        int chiral1, chiral2, grid1, grid2;
        force.getTorsionTorsionParameters(group1, particle1, particle2, particle3, particle4, particle5, chiral1, grid1);
        force.getTorsionTorsionParameters(group2, particle1, particle2, particle3, particle4, particle5, chiral2, grid2);
        return (grid1 == grid2);
    }
private:
    const AmoebaTorsionTorsionForce& force;
};

692
CudaCalcAmoebaTorsionTorsionForceKernel::CudaCalcAmoebaTorsionTorsionForceKernel(std::string name, const Platform& platform, CudaContext& cu, const System& system) :
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
                CalcAmoebaTorsionTorsionForceKernel(name, platform), cu(cu), system(system), gridValues(NULL), gridParams(NULL), torsionParams(NULL) {
}

CudaCalcAmoebaTorsionTorsionForceKernel::~CudaCalcAmoebaTorsionTorsionForceKernel() {
    cu.setAsCurrent();
    if (gridValues != NULL)
        delete gridValues;
    if (gridParams != NULL)
        delete gridParams;
    if (torsionParams != NULL)
        delete torsionParams;
}

void CudaCalcAmoebaTorsionTorsionForceKernel::initialize(const System& system, const AmoebaTorsionTorsionForce& force) {
    cu.setAsCurrent();
    int numContexts = cu.getPlatformData().contexts.size();
    int startIndex = cu.getContextIndex()*force.getNumTorsionTorsions()/numContexts;
    int endIndex = (cu.getContextIndex()+1)*force.getNumTorsionTorsions()/numContexts;
    numTorsionTorsions = endIndex-startIndex;
    if (numTorsionTorsions == 0)
        return;
    
    // Record torsion parameters.
    
    vector<vector<int> > atoms(numTorsionTorsions, vector<int>(5));
    vector<int2> torsionParamsVec(numTorsionTorsions);
    torsionParams = CudaArray::create<int2>(cu, numTorsionTorsions, "torsionTorsionParams");
    for (int i = 0; i < numTorsionTorsions; i++)
        force.getTorsionTorsionParameters(startIndex+i, atoms[i][0], atoms[i][1], atoms[i][2], atoms[i][3], atoms[i][4], torsionParamsVec[i].x, torsionParamsVec[i].y);
    torsionParams->upload(torsionParamsVec);
    
    // Record the grids.
    
    vector<float4> gridValuesVec;
    vector<float4> gridParamsVec;
    for (int i = 0; i < force.getNumTorsionTorsionGrids(); i++) {
        const TorsionTorsionGrid& initialGrid = force.getTorsionTorsionGrid(i);

        // check if grid needs to be reordered: x-angle should be 'slow' index

        bool reordered = false;
        TorsionTorsionGrid reorderedGrid;
        if (initialGrid[0][0][0] != initialGrid[0][1][0]) {
            AmoebaTorsionTorsionForceImpl::reorderGrid(initialGrid, reorderedGrid);
            reordered = true;
        }
        const TorsionTorsionGrid& grid = (reordered ? reorderedGrid : initialGrid);
        float range = grid[0][grid[0].size()-1][1] - grid[0][0][1];
        gridParamsVec.push_back(make_float4(gridValuesVec.size(), grid[0][0][0], range/(grid.size()-1), grid.size()));
        for (int j = 0; j < grid.size(); j++)
            for (int k = 0; k < grid[j].size(); k++)
                gridValuesVec.push_back(make_float4((float) grid[j][k][2], (float) grid[j][k][3], (float) grid[j][k][4], (float) grid[j][k][5]));
    }
    gridValues = CudaArray::create<float4>(cu, gridValuesVec.size(), "torsionTorsionGridValues");
    gridParams = CudaArray::create<float4>(cu, gridParamsVec.size(), "torsionTorsionGridParams");
    gridValues->upload(gridValuesVec);
    gridParams->upload(gridParamsVec);
    map<string, string> replacements;
    replacements["GRID_VALUES"] = cu.getBondedUtilities().addArgument(gridValues->getDevicePointer(), "float4");
    replacements["GRID_PARAMS"] = cu.getBondedUtilities().addArgument(gridParams->getDevicePointer(), "float4");
    replacements["TORSION_PARAMS"] = cu.getBondedUtilities().addArgument(torsionParams->getDevicePointer(), "int2");
    replacements["RAD_TO_DEG"] = cu.doubleToString(180/M_PI);
    cu.getBondedUtilities().addInteraction(atoms, cu.replaceStrings(CudaAmoebaKernelSources::amoebaTorsionTorsionForce, replacements), force.getForceGroup());
    cu.getBondedUtilities().addPrefixCode(CudaAmoebaKernelSources::bicubic);
    cu.addForce(new ForceInfo(force));
}

double CudaCalcAmoebaTorsionTorsionForceKernel::execute(ContextImpl& context, bool includeForces, bool includeEnergy) {
    return 0.0;
}

764
765
766
767
768
769
770
771
772
773
774
775
776
777
/* -------------------------------------------------------------------------- *
 *                             AmoebaMultipole                                *
 * -------------------------------------------------------------------------- */

class CudaCalcAmoebaMultipoleForceKernel::ForceInfo : public CudaForceInfo {
public:
    ForceInfo(const AmoebaMultipoleForce& force) : force(force) {
    }
    bool areParticlesIdentical(int particle1, int particle2) {
        double charge1, charge2, thole1, thole2, damping1, damping2, polarity1, polarity2;
        int axis1, axis2, multipole11, multipole12, multipole21, multipole22, multipole31, multipole32;
        vector<double> dipole1, dipole2, quadrupole1, quadrupole2;
        force.getMultipoleParameters(particle1, charge1, dipole1, quadrupole1, axis1, multipole11, multipole21, multipole31, thole1, damping1, polarity1);
        force.getMultipoleParameters(particle2, charge2, dipole2, quadrupole2, axis2, multipole12, multipole22, multipole32, thole2, damping2, polarity2);
778
        if (charge1 != charge2 || thole1 != thole2 || damping1 != damping2 || polarity1 != polarity2 || axis1 != axis2) {
779
780
            return false;
        }
781
782
        for (int i = 0; i < (int) dipole1.size(); ++i) {
            if (dipole1[i] != dipole2[i]) {
783
784
785
                return false;
            }
        }
786
787
        for (int i = 0; i < (int) quadrupole1.size(); ++i) {
            if (quadrupole1[i] != quadrupole2[i]) {
788
789
790
791
792
                return false;
            }
        }
        return true;
    }
793
794
795
796
797
798
799
800
801
802
803
    int getNumParticleGroups() {
        return 7*force.getNumMultipoles();
    }
    void getParticlesInGroup(int index, vector<int>& particles) {
        int particle = index/7;
        int type = index-7*particle;
        force.getCovalentMap(particle, AmoebaMultipoleForce::CovalentType(type), particles);
    }
    bool areGroupsIdentical(int group1, int group2) {
        return ((group1%7) == (group2%7));
    }
804
805
806
807
private:
    const AmoebaMultipoleForce& force;
};

808
CudaCalcAmoebaMultipoleForceKernel::CudaCalcAmoebaMultipoleForceKernel(std::string name, const Platform& platform, CudaContext& cu, const System& system) : 
809
        CalcAmoebaMultipoleForceKernel(name, platform), cu(cu), system(system), hasInitializedScaleFactors(false), hasInitializedFFT(false), multipolesAreValid(false),
810
811
        multipoleParticles(NULL), molecularDipoles(NULL), molecularQuadrupoles(NULL), labFrameDipoles(NULL), labFrameQuadrupoles(NULL), sphericalDipoles(NULL), sphericalQuadrupoles(NULL),
        fracDipoles(NULL), fracQuadrupoles(NULL), field(NULL), fieldPolar(NULL), inducedField(NULL), inducedFieldPolar(NULL), torque(NULL), dampingAndThole(NULL), inducedDipole(NULL),
peastman's avatar
peastman committed
812
        diisCoefficients(NULL), inducedDipolePolar(NULL), inducedDipoleErrors(NULL), prevDipoles(NULL), prevDipolesPolar(NULL), prevDipolesGk(NULL),
813
814
815
816
        prevDipolesGkPolar(NULL), prevErrors(NULL), diisMatrix(NULL), polarizability(NULL), extrapolatedDipole(NULL), extrapolatedDipolePolar(NULL),
        extrapolatedDipoleGk(NULL), extrapolatedDipoleGkPolar(NULL), inducedDipoleFieldGradient(NULL), inducedDipoleFieldGradientPolar(NULL),
        inducedDipoleFieldGradientGk(NULL), inducedDipoleFieldGradientGkPolar(NULL), extrapolatedDipoleFieldGradient(NULL), extrapolatedDipoleFieldGradientPolar(NULL),
        extrapolatedDipoleFieldGradientGk(NULL), extrapolatedDipoleFieldGradientGkPolar(NULL), covalentFlags(NULL), polarizationGroupFlags(NULL),
817
        pmeGrid(NULL), pmeBsplineModuliX(NULL), pmeBsplineModuliY(NULL), pmeBsplineModuliZ(NULL), pmeIgrid(NULL), pmePhi(NULL),
818
        pmePhid(NULL), pmePhip(NULL), pmePhidp(NULL), pmeCphi(NULL), pmeAtomGridIndex(NULL), lastPositions(NULL), sort(NULL), gkKernel(NULL) {
819
820
821
822
823
824
825
826
827
828
829
830
831
832
}

CudaCalcAmoebaMultipoleForceKernel::~CudaCalcAmoebaMultipoleForceKernel() {
    cu.setAsCurrent();
    if (multipoleParticles != NULL)
        delete multipoleParticles;
    if (molecularDipoles != NULL)
        delete molecularDipoles;
    if (molecularQuadrupoles != NULL)
        delete molecularQuadrupoles;
    if (labFrameDipoles != NULL)
        delete labFrameDipoles;
    if (labFrameQuadrupoles != NULL)
        delete labFrameQuadrupoles;
833
834
835
836
    if (sphericalDipoles != NULL)
        delete sphericalDipoles;
    if (sphericalQuadrupoles != NULL)
        delete sphericalQuadrupoles;
837
838
839
840
    if (fracDipoles != NULL)
        delete fracDipoles;
    if (fracQuadrupoles != NULL)
        delete fracQuadrupoles;
841
842
843
844
    if (field != NULL)
        delete field;
    if (fieldPolar != NULL)
        delete fieldPolar;
845
846
847
848
    if (inducedField != NULL)
        delete inducedField;
    if (inducedFieldPolar != NULL)
        delete inducedFieldPolar;
849
850
    if (torque != NULL)
        delete torque;
851
852
853
854
855
856
    if (dampingAndThole != NULL)
        delete dampingAndThole;
    if (inducedDipole != NULL)
        delete inducedDipole;
    if (inducedDipolePolar != NULL)
        delete inducedDipolePolar;
857
858
    if (inducedDipoleErrors != NULL)
        delete inducedDipoleErrors;
peastman's avatar
peastman committed
859
860
861
862
863
864
865
866
867
868
869
870
871
872
    if (prevDipoles != NULL)
        delete prevDipoles;
    if (prevDipolesPolar != NULL)
        delete prevDipolesPolar;
    if (prevDipolesGk != NULL)
        delete prevDipolesGk;
    if (prevDipolesGkPolar != NULL)
        delete prevDipolesGkPolar;
    if (prevErrors != NULL)
        delete prevErrors;
    if (diisMatrix != NULL)
        delete diisMatrix;
    if (diisCoefficients != NULL)
        delete diisCoefficients;
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
    if (extrapolatedDipole != NULL)
        delete extrapolatedDipole;
    if (extrapolatedDipolePolar != NULL)
        delete extrapolatedDipolePolar;
    if (extrapolatedDipoleGk != NULL)
        delete extrapolatedDipoleGk;
    if (extrapolatedDipoleGkPolar != NULL)
        delete extrapolatedDipoleGkPolar;
    if (inducedDipoleFieldGradient != NULL)
        delete inducedDipoleFieldGradient;
    if (inducedDipoleFieldGradientPolar != NULL)
        delete inducedDipoleFieldGradientPolar;
    if (inducedDipoleFieldGradientGk != NULL)
        delete inducedDipoleFieldGradientGk;
    if (inducedDipoleFieldGradientGkPolar != NULL)
        delete inducedDipoleFieldGradientGkPolar;
    if (extrapolatedDipoleFieldGradient != NULL)
        delete extrapolatedDipoleFieldGradient;
    if (extrapolatedDipoleFieldGradientPolar != NULL)
        delete extrapolatedDipoleFieldGradientPolar;
    if (extrapolatedDipoleFieldGradientGk != NULL)
        delete extrapolatedDipoleFieldGradientGk;
    if (extrapolatedDipoleFieldGradientGkPolar != NULL)
        delete extrapolatedDipoleFieldGradientGkPolar;
897
898
899
900
901
902
    if (polarizability != NULL)
        delete polarizability;
    if (covalentFlags != NULL)
        delete covalentFlags;
    if (polarizationGroupFlags != NULL)
        delete polarizationGroupFlags;
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
    if (pmeGrid != NULL)
        delete pmeGrid;
    if (pmeBsplineModuliX != NULL)
        delete pmeBsplineModuliX;
    if (pmeBsplineModuliY != NULL)
        delete pmeBsplineModuliY;
    if (pmeBsplineModuliZ != NULL)
        delete pmeBsplineModuliZ;
    if (pmeIgrid != NULL)
        delete pmeIgrid;
    if (pmePhi != NULL)
        delete pmePhi;
    if (pmePhid != NULL)
        delete pmePhid;
    if (pmePhip != NULL)
        delete pmePhip;
    if (pmePhidp != NULL)
        delete pmePhidp;
921
922
    if (pmeCphi != NULL)
        delete pmeCphi;
923
924
    if (pmeAtomGridIndex != NULL)
        delete pmeAtomGridIndex;
925
926
    if (lastPositions != NULL)
        delete lastPositions;
927
928
929
930
931
932
    if (sort != NULL)
        delete sort;
    if (hasInitializedFFT)
        cufftDestroy(fft);
}

933
934
935
936
937
938
939
void CudaCalcAmoebaMultipoleForceKernel::initialize(const System& system, const AmoebaMultipoleForce& force) {
    cu.setAsCurrent();

    // Initialize multipole parameters.

    numMultipoles = force.getNumMultipoles();
    CudaArray& posq = cu.getPosq();
940
941
942
    vector<double4> temp(posq.getSize());
    float4* posqf = (float4*) &temp[0];
    double4* posqd = (double4*) &temp[0];
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
    vector<float2> dampingAndTholeVec;
    vector<float> polarizabilityVec;
    vector<float> molecularDipolesVec;
    vector<float> molecularQuadrupolesVec;
    vector<int4> multipoleParticlesVec;
    for (int i = 0; i < numMultipoles; i++) {
        double charge, thole, damping, polarity;
        int axisType, atomX, atomY, atomZ;
        vector<double> dipole, quadrupole;
        force.getMultipoleParameters(i, charge, dipole, quadrupole, axisType, atomZ, atomX, atomY, thole, damping, polarity);
        if (cu.getUseDoublePrecision())
            posqd[i] = make_double4(0, 0, 0, charge);
        else
            posqf[i] = make_float4(0, 0, 0, (float) charge);
        dampingAndTholeVec.push_back(make_float2((float) damping, (float) thole));
        polarizabilityVec.push_back((float) polarity);
        multipoleParticlesVec.push_back(make_int4(atomX, atomY, atomZ, axisType));
        for (int j = 0; j < 3; j++)
            molecularDipolesVec.push_back((float) dipole[j]);
962
963
964
965
966
        molecularQuadrupolesVec.push_back((float) quadrupole[0]);
        molecularQuadrupolesVec.push_back((float) quadrupole[1]);
        molecularQuadrupolesVec.push_back((float) quadrupole[2]);
        molecularQuadrupolesVec.push_back((float) quadrupole[4]);
        molecularQuadrupolesVec.push_back((float) quadrupole[5]);
967
    }
968
969
970
971
    hasQuadrupoles = false;
    for (int i = 0; i < (int) molecularQuadrupolesVec.size(); i++)
        if (molecularQuadrupolesVec[i] != 0.0)
            hasQuadrupoles = true;
972
973
974
975
976
977
978
    int paddedNumAtoms = cu.getPaddedNumAtoms();
    for (int i = numMultipoles; i < paddedNumAtoms; i++) {
        dampingAndTholeVec.push_back(make_float2(0, 0));
        polarizabilityVec.push_back(0);
        multipoleParticlesVec.push_back(make_int4(0, 0, 0, 0));
        for (int j = 0; j < 3; j++)
            molecularDipolesVec.push_back(0);
979
        for (int j = 0; j < 5; j++)
980
981
982
983
984
985
            molecularQuadrupolesVec.push_back(0);
    }
    dampingAndThole = CudaArray::create<float2>(cu, paddedNumAtoms, "dampingAndThole");
    polarizability = CudaArray::create<float>(cu, paddedNumAtoms, "polarizability");
    multipoleParticles = CudaArray::create<int4>(cu, paddedNumAtoms, "multipoleParticles");
    molecularDipoles = CudaArray::create<float>(cu, 3*paddedNumAtoms, "molecularDipoles");
986
    molecularQuadrupoles = CudaArray::create<float>(cu, 5*paddedNumAtoms, "molecularQuadrupoles");
987
    lastPositions = new CudaArray(cu, cu.getPosq().getSize(), cu.getPosq().getElementSize(), "lastPositions");
988
989
990
991
992
    dampingAndThole->upload(dampingAndTholeVec);
    polarizability->upload(polarizabilityVec);
    multipoleParticles->upload(multipoleParticlesVec);
    molecularDipoles->upload(molecularDipolesVec);
    molecularQuadrupoles->upload(molecularQuadrupolesVec);
993
    posq.upload(&temp[0]);
994
995
996
    
    // Create workspace arrays.
    
997
    polarizationType = force.getPolarizationType();
998
999
    int elementSize = (cu.getUseDoublePrecision() ? sizeof(double) : sizeof(float));
    labFrameDipoles = new CudaArray(cu, 3*paddedNumAtoms, elementSize, "labFrameDipoles");
1000
    labFrameQuadrupoles = new CudaArray(cu, 5*paddedNumAtoms, elementSize, "labFrameQuadrupoles");
1001
1002
    sphericalDipoles = new CudaArray(cu, 3*paddedNumAtoms, elementSize, "sphericalDipoles");
    sphericalQuadrupoles = new CudaArray(cu, 5*paddedNumAtoms, elementSize, "sphericalQuadrupoles");
1003
1004
    fracDipoles = new CudaArray(cu, 3*paddedNumAtoms, elementSize, "fracDipoles");
    fracQuadrupoles = new CudaArray(cu, 6*paddedNumAtoms, elementSize, "fracQuadrupoles");
1005
1006
    field = new CudaArray(cu, 3*paddedNumAtoms, sizeof(long long), "field");
    fieldPolar = new CudaArray(cu, 3*paddedNumAtoms, sizeof(long long), "fieldPolar");
1007
    torque = new CudaArray(cu, 3*paddedNumAtoms, sizeof(long long), "torque");
1008
1009
    inducedDipole = new CudaArray(cu, 3*paddedNumAtoms, elementSize, "inducedDipole");
    inducedDipolePolar = new CudaArray(cu, 3*paddedNumAtoms, elementSize, "inducedDipolePolar");
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
    if (polarizationType == AmoebaMultipoleForce::Mutual) {
        inducedDipoleErrors = new CudaArray(cu, cu.getNumThreadBlocks(), sizeof(float2), "inducedDipoleErrors");
        prevDipoles = new CudaArray(cu, 3*numMultipoles*MaxPrevDIISDipoles, elementSize, "prevDipoles");
        prevDipolesPolar = new CudaArray(cu, 3*numMultipoles*MaxPrevDIISDipoles, elementSize, "prevDipolesPolar");
        prevErrors = new CudaArray(cu, 3*numMultipoles*MaxPrevDIISDipoles, elementSize, "prevErrors");
        diisMatrix = new CudaArray(cu, MaxPrevDIISDipoles*MaxPrevDIISDipoles, elementSize, "diisMatrix");
        diisCoefficients = new CudaArray(cu, MaxPrevDIISDipoles+1, sizeof(float), "diisMatrix");
    }
    else if (polarizationType == AmoebaMultipoleForce::Extrapolated) {
        int numOrders = force.getExtrapolationCoefficients().size();
        extrapolatedDipole = new CudaArray(cu, 3*numMultipoles*numOrders, elementSize, "extrapolatedDipole");
        extrapolatedDipolePolar = new CudaArray(cu, 3*numMultipoles*numOrders, elementSize, "extrapolatedDipolePolar");
1022
1023
1024
1025
        inducedDipoleFieldGradient = new CudaArray(cu, 6*paddedNumAtoms, sizeof(long long), "inducedDipoleFieldGradient");
        inducedDipoleFieldGradientPolar = new CudaArray(cu, 6*paddedNumAtoms, sizeof(long long), "inducedDipoleFieldGradientPolar");
        extrapolatedDipoleFieldGradient = new CudaArray(cu, 6*numMultipoles*(numOrders-1), elementSize, "extrapolatedDipoleFieldGradient");
        extrapolatedDipoleFieldGradientPolar = new CudaArray(cu, 6*numMultipoles*(numOrders-1), elementSize, "extrapolatedDipoleFieldGradientPolar");
1026
    }
1027
1028
    cu.addAutoclearBuffer(*field);
    cu.addAutoclearBuffer(*fieldPolar);
1029
    cu.addAutoclearBuffer(*torque);
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
    
    // Record which atoms should be flagged as exclusions based on covalent groups, and determine
    // the values for the covalent group flags.
    
    vector<vector<int> > exclusions(numMultipoles);
    for (int i = 0; i < numMultipoles; i++) {
        vector<int> atoms;
        set<int> allAtoms;
        allAtoms.insert(i);
        force.getCovalentMap(i, AmoebaMultipoleForce::Covalent12, atoms);
        allAtoms.insert(atoms.begin(), atoms.end());
        force.getCovalentMap(i, AmoebaMultipoleForce::Covalent13, atoms);
        allAtoms.insert(atoms.begin(), atoms.end());
        for (set<int>::const_iterator iter = allAtoms.begin(); iter != allAtoms.end(); ++iter)
            covalentFlagValues.push_back(make_int3(i, *iter, 0));
        force.getCovalentMap(i, AmoebaMultipoleForce::Covalent14, atoms);
        allAtoms.insert(atoms.begin(), atoms.end());
        for (int j = 0; j < (int) atoms.size(); j++)
            covalentFlagValues.push_back(make_int3(i, atoms[j], 1));
        force.getCovalentMap(i, AmoebaMultipoleForce::Covalent15, atoms);
        for (int j = 0; j < (int) atoms.size(); j++)
            covalentFlagValues.push_back(make_int3(i, atoms[j], 2));
        allAtoms.insert(atoms.begin(), atoms.end());
        force.getCovalentMap(i, AmoebaMultipoleForce::PolarizationCovalent11, atoms);
        allAtoms.insert(atoms.begin(), atoms.end());
        exclusions[i].insert(exclusions[i].end(), allAtoms.begin(), allAtoms.end());
1056
1057
1058
1059
1060
1061

        // Workaround for bug in TINKER: if an atom is listed in both the PolarizationCovalent11
        // and PolarizationCovalent12 maps, the latter takes precedence.

        vector<int> atoms12;
        force.getCovalentMap(i, AmoebaMultipoleForce::PolarizationCovalent12, atoms12);
1062
        for (int j = 0; j < (int) atoms.size(); j++)
1063
1064
            if (find(atoms12.begin(), atoms12.end(), atoms[j]) == atoms12.end())
                polarizationFlagValues.push_back(make_int2(i, atoms[j]));
1065
    }
1066
1067
1068
1069
1070
1071
1072
1073
1074
    set<pair<int, int> > tilesWithExclusions;
    for (int atom1 = 0; atom1 < (int) exclusions.size(); ++atom1) {
        int x = atom1/CudaContext::TileSize;
        for (int j = 0; j < (int) exclusions[atom1].size(); ++j) {
            int atom2 = exclusions[atom1][j];
            int y = atom2/CudaContext::TileSize;
            tilesWithExclusions.insert(make_pair(max(x, y), min(x, y)));
        }
    }
1075
    
1076
1077
    // Record other options.
    
1078
    if (polarizationType == AmoebaMultipoleForce::Mutual) {
1079
1080
1081
1082
        maxInducedIterations = force.getMutualInducedMaxIterations();
        inducedEpsilon = force.getMutualInducedTargetEpsilon();
    }
    else
1083
        maxInducedIterations = 0;
1084
1085
1086
1087
    if (polarizationType != AmoebaMultipoleForce::Direct) {
        inducedField = new CudaArray(cu, 3*paddedNumAtoms, sizeof(long long), "inducedField");
        inducedFieldPolar = new CudaArray(cu, 3*paddedNumAtoms, sizeof(long long), "inducedFieldPolar");
    }
1088
    usePME = (force.getNonbondedMethod() == AmoebaMultipoleForce::PME);
1089
    
1090
1091
1092
1093
1094
1095
1096
    // See whether there's an AmoebaGeneralizedKirkwoodForce in the System.

    const AmoebaGeneralizedKirkwoodForce* gk = NULL;
    for (int i = 0; i < system.getNumForces() && gk == NULL; i++)
        gk = dynamic_cast<const AmoebaGeneralizedKirkwoodForce*>(&system.getForce(i));
    double innerDielectric = (gk == NULL ? 1.0 : gk->getSoluteDielectric());
    
1097
1098
    // Create the kernels.

1099
1100
1101
1102
    bool useShuffle = (cu.getComputeCapability() >= 3.0 && !cu.getUseDoublePrecision());
    double fixedThreadMemory = 19*elementSize+2*sizeof(float)+3*sizeof(int)/(double) cu.TileSize;
    double inducedThreadMemory = 15*elementSize+2*sizeof(float);
    double electrostaticsThreadMemory = 0;
Peter Eastman's avatar
Peter Eastman committed
1103
    if (!useShuffle)
1104
        fixedThreadMemory += 3*elementSize;
1105
1106
1107
1108
    map<string, string> defines;
    defines["NUM_ATOMS"] = cu.intToString(numMultipoles);
    defines["PADDED_NUM_ATOMS"] = cu.intToString(cu.getPaddedNumAtoms());
    defines["NUM_BLOCKS"] = cu.intToString(cu.getNumAtomBlocks());
1109
    defines["ENERGY_SCALE_FACTOR"] = cu.doubleToString(138.9354558456/innerDielectric);
1110
    if (polarizationType == AmoebaMultipoleForce::Direct)
1111
        defines["DIRECT_POLARIZATION"] = "";
1112
1113
1114
1115
    else if (polarizationType == AmoebaMultipoleForce::Mutual)
        defines["MUTUAL_POLARIZATION"] = "";
    else if (polarizationType == AmoebaMultipoleForce::Extrapolated)
        defines["EXTRAPOLATED_POLARIZATION"] = "";
Peter Eastman's avatar
Peter Eastman committed
1116
1117
    if (useShuffle)
        defines["USE_SHUFFLE"] = "";
1118
1119
    if (hasQuadrupoles)
        defines["INCLUDE_QUADRUPOLES"] = "";
1120
1121
1122
1123
1124
1125
1126
1127
    defines["TILE_SIZE"] = cu.intToString(CudaContext::TileSize);
    int numExclusionTiles = tilesWithExclusions.size();
    defines["NUM_TILES_WITH_EXCLUSIONS"] = cu.intToString(numExclusionTiles);
    int numContexts = cu.getPlatformData().contexts.size();
    int startExclusionIndex = cu.getContextIndex()*numExclusionTiles/numContexts;
    int endExclusionIndex = (cu.getContextIndex()+1)*numExclusionTiles/numContexts;
    defines["FIRST_EXCLUSION_TILE"] = cu.intToString(startExclusionIndex);
    defines["LAST_EXCLUSION_TILE"] = cu.intToString(endExclusionIndex);
1128
1129
1130
1131
1132
1133
1134
1135
    maxExtrapolationOrder = force.getExtrapolationCoefficients().size();
    defines["MAX_EXTRAPOLATION_ORDER"] = cu.intToString(maxExtrapolationOrder);
    stringstream coefficients;
    for (int i = 0; i < maxExtrapolationOrder; i++) {
        if (i > 0)
            coefficients << ",";
        double sum = 0;
        for (int j = i; j < maxExtrapolationOrder; j++)
1136
            sum += force.getExtrapolationCoefficients()[j];
1137
1138
1139
        coefficients << cu.doubleToString(sum);
    }
    defines["EXTRAPOLATION_COEFFICIENTS_SUM"] = coefficients.str();
1140
    alpha = force.getAEwald();
1141
1142
1143
1144
1145
1146
1147
1148
    if (usePME) {
        vector<int> pmeGridDimension;
        force.getPmeGridDimensions(pmeGridDimension);
        if (pmeGridDimension[0] == 0 || alpha == 0.0) {
            NonbondedForce nb;
            nb.setEwaldErrorTolerance(force.getEwaldErrorTolerance());
            nb.setCutoffDistance(force.getCutoffDistance());
            NonbondedForceImpl::calcPMEParameters(system, nb, alpha, gridSizeX, gridSizeY, gridSizeZ);
1149
1150
1151
            gridSizeX = CudaFFT3D::findLegalDimension(gridSizeX);
            gridSizeY = CudaFFT3D::findLegalDimension(gridSizeY);
            gridSizeZ = CudaFFT3D::findLegalDimension(gridSizeZ);
1152
        } else {
1153
1154
1155
            gridSizeX = CudaFFT3D::findLegalDimension(pmeGridDimension[0]);
            gridSizeY = CudaFFT3D::findLegalDimension(pmeGridDimension[1]);
            gridSizeZ = CudaFFT3D::findLegalDimension(pmeGridDimension[2]);
1156
1157
1158
1159
1160
1161
1162
1163
        }
        defines["EWALD_ALPHA"] = cu.doubleToString(alpha);
        defines["SQRT_PI"] = cu.doubleToString(sqrt(M_PI));
        defines["USE_EWALD"] = "";
        defines["USE_CUTOFF"] = "";
        defines["USE_PERIODIC"] = "";
        defines["CUTOFF_SQUARED"] = cu.doubleToString(force.getCutoffDistance()*force.getCutoffDistance());
    }
1164
1165
1166
1167
1168
1169
1170
    if (gk != NULL) {
        defines["USE_GK"] = "";
        defines["GK_C"] = cu.doubleToString(2.455);
        double solventDielectric = gk->getSolventDielectric();
        defines["GK_FC"] = cu.doubleToString(1*(1-solventDielectric)/(0+1*solventDielectric));
        defines["GK_FD"] = cu.doubleToString(2*(1-solventDielectric)/(1+2*solventDielectric));
        defines["GK_FQ"] = cu.doubleToString(3*(1-solventDielectric)/(2+3*solventDielectric));
1171
1172
        fixedThreadMemory += 4*elementSize;
        inducedThreadMemory += 13*elementSize;
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
        if (polarizationType == AmoebaMultipoleForce::Mutual) {
            prevDipolesGk = new CudaArray(cu, 3*numMultipoles*MaxPrevDIISDipoles, elementSize, "prevDipolesGk");
            prevDipolesGkPolar = new CudaArray(cu, 3*numMultipoles*MaxPrevDIISDipoles, elementSize, "prevDipolesGkPolar");
        }
        else if (polarizationType == AmoebaMultipoleForce::Extrapolated) {
            int numOrders = force.getExtrapolationCoefficients().size();
            extrapolatedDipoleGk = new CudaArray(cu, 3*numMultipoles*numOrders, elementSize, "extrapolatedDipoleGk");
            extrapolatedDipoleGkPolar = new CudaArray(cu, 3*numMultipoles*numOrders, elementSize, "extrapolatedDipoleGkPolar");
            inducedDipoleFieldGradientGk = new CudaArray(cu, 6*numMultipoles, elementSize, "inducedDipoleFieldGradientGk");
            inducedDipoleFieldGradientGkPolar = new CudaArray(cu, 6*numMultipoles, elementSize, "inducedDipoleFieldGradientGkPolar");
1183
1184
            extrapolatedDipoleFieldGradientGk = new CudaArray(cu, 6*numMultipoles*(numOrders-1), elementSize, "extrapolatedDipoleFieldGradientGk");
            extrapolatedDipoleFieldGradientGkPolar = new CudaArray(cu, 6*numMultipoles*(numOrders-1), elementSize, "extrapolatedDipoleFieldGradientGkPolar");
1185
        }
1186
    }
1187
1188
1189
    int maxThreads = cu.getNonbondedUtilities().getForceThreadBlockSize();
    fixedFieldThreads = min(maxThreads, cu.computeThreadBlockSize(fixedThreadMemory));
    inducedFieldThreads = min(maxThreads, cu.computeThreadBlockSize(inducedThreadMemory));
1190
1191
    CUmodule module = cu.createModule(CudaKernelSources::vectorOps+CudaAmoebaKernelSources::multipoles, defines);
    computeMomentsKernel = cu.getKernel(module, "computeLabFrameMoments");
1192
    recordInducedDipolesKernel = cu.getKernel(module, "recordInducedDipoles");
1193
    mapTorqueKernel = cu.getKernel(module, "mapTorqueToForce");
1194
    computePotentialKernel = cu.getKernel(module, "computePotentialAtPoints");
1195
    defines["THREAD_BLOCK_SIZE"] = cu.intToString(fixedFieldThreads);
1196
1197
    module = cu.createModule(CudaKernelSources::vectorOps+CudaAmoebaKernelSources::multipoleFixedField, defines);
    computeFixedFieldKernel = cu.getKernel(module, "computeFixedField");
1198
    if (polarizationType != AmoebaMultipoleForce::Direct) {
1199
        defines["THREAD_BLOCK_SIZE"] = cu.intToString(inducedFieldThreads);
peastman's avatar
peastman committed
1200
        defines["MAX_PREV_DIIS_DIPOLES"] = cu.intToString(MaxPrevDIISDipoles);
1201
1202
        module = cu.createModule(CudaKernelSources::vectorOps+CudaAmoebaKernelSources::multipoleInducedField, defines);
        computeInducedFieldKernel = cu.getKernel(module, "computeInducedField");
peastman's avatar
peastman committed
1203
1204
1205
        updateInducedFieldKernel = cu.getKernel(module, "updateInducedFieldByDIIS");
        recordDIISDipolesKernel = cu.getKernel(module, "recordInducedDipolesForDIIS");
        buildMatrixKernel = cu.getKernel(module, "computeDIISMatrix");
1206
1207
1208
        initExtrapolatedKernel = cu.getKernel(module, "initExtrapolatedDipoles");
        iterateExtrapolatedKernel = cu.getKernel(module, "iterateExtrapolatedDipoles");
        computeExtrapolatedKernel = cu.getKernel(module, "computeExtrapolatedDipoles");
1209
        addExtrapolatedGradientKernel = cu.getKernel(module, "addExtrapolatedFieldGradientToForce");
1210
    }
1211
    stringstream electrostaticsSource;
1212
1213
1214
    electrostaticsSource << CudaKernelSources::vectorOps;
    electrostaticsSource << CudaAmoebaKernelSources::sphericalMultipoles;
    if (usePME)
1215
        electrostaticsSource << CudaAmoebaKernelSources::pmeMultipoleElectrostatics;
1216
    else
1217
        electrostaticsSource << CudaAmoebaKernelSources::multipoleElectrostatics;
1218
    electrostaticsThreadMemory = 24*elementSize+3*sizeof(float)+3*sizeof(int)/(double) cu.TileSize;
1219
1220
    electrostaticsThreads = min(maxThreads, cu.computeThreadBlockSize(electrostaticsThreadMemory));
    defines["THREAD_BLOCK_SIZE"] = cu.intToString(electrostaticsThreads);
1221
1222
    module = cu.createModule(electrostaticsSource.str(), defines);
    electrostaticsKernel = cu.getKernel(module, "computeElectrostatics");
1223
1224
1225

    // Set up PME.
    
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
    if (usePME) {
        // Create the PME kernels.

        map<string, string> pmeDefines;
        pmeDefines["EWALD_ALPHA"] = cu.doubleToString(alpha);
        pmeDefines["PME_ORDER"] = cu.intToString(PmeOrder);
        pmeDefines["NUM_ATOMS"] = cu.intToString(numMultipoles);
        pmeDefines["PADDED_NUM_ATOMS"] = cu.intToString(cu.getPaddedNumAtoms());
        pmeDefines["EPSILON_FACTOR"] = cu.doubleToString(138.9354558456);
        pmeDefines["GRID_SIZE_X"] = cu.intToString(gridSizeX);
        pmeDefines["GRID_SIZE_Y"] = cu.intToString(gridSizeY);
        pmeDefines["GRID_SIZE_Z"] = cu.intToString(gridSizeZ);
        pmeDefines["M_PI"] = cu.doubleToString(M_PI);
1239
        pmeDefines["SQRT_PI"] = cu.doubleToString(sqrt(M_PI));
1240
        if (polarizationType == AmoebaMultipoleForce::Direct)
1241
            pmeDefines["DIRECT_POLARIZATION"] = "";
1242
1243
        else if (polarizationType == AmoebaMultipoleForce::Extrapolated)
            pmeDefines["EXTRAPOLATED_POLARIZATION"] = "";
1244
        CUmodule module = cu.createModule(CudaKernelSources::vectorOps+CudaAmoebaKernelSources::multipolePme, pmeDefines);
1245
        pmeGridIndexKernel = cu.getKernel(module, "findAtomGridIndex");
1246
        pmeTransformMultipolesKernel = cu.getKernel(module, "transformMultipolesToFractionalCoordinates");
1247
        pmeTransformPotentialKernel = cu.getKernel(module, "transformPotentialToCartesianCoordinates");
1248
        pmeSpreadFixedMultipolesKernel = cu.getKernel(module, "gridSpreadFixedMultipoles");
1249
        pmeSpreadInducedDipolesKernel = cu.getKernel(module, "gridSpreadInducedDipoles");
1250
        pmeFinishSpreadChargeKernel = cu.getKernel(module, "finishSpreadCharge");
1251
1252
        pmeConvolutionKernel = cu.getKernel(module, "reciprocalConvolution");
        pmeFixedPotentialKernel = cu.getKernel(module, "computeFixedPotentialFromGrid");
1253
        pmeInducedPotentialKernel = cu.getKernel(module, "computeInducedPotentialFromGrid");
1254
        pmeFixedForceKernel = cu.getKernel(module, "computeFixedMultipoleForceAndEnergy");
1255
1256
        pmeInducedForceKernel = cu.getKernel(module, "computeInducedDipoleForceAndEnergy");
        pmeRecordInducedFieldDipolesKernel = cu.getKernel(module, "recordInducedFieldDipoles");
1257
1258
1259
1260
        cuFuncSetCacheConfig(pmeSpreadFixedMultipolesKernel, CU_FUNC_CACHE_PREFER_L1);
        cuFuncSetCacheConfig(pmeSpreadInducedDipolesKernel, CU_FUNC_CACHE_PREFER_L1);
        cuFuncSetCacheConfig(pmeFixedPotentialKernel, CU_FUNC_CACHE_PREFER_L1);
        cuFuncSetCacheConfig(pmeInducedPotentialKernel, CU_FUNC_CACHE_PREFER_L1);
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274

        // Create required data structures.

        int elementSize = (cu.getUseDoublePrecision() ? sizeof(double) : sizeof(float));
        pmeGrid = new CudaArray(cu, gridSizeX*gridSizeY*gridSizeZ, 2*elementSize, "pmeGrid");
        cu.addAutoclearBuffer(*pmeGrid);
        pmeBsplineModuliX = new CudaArray(cu, gridSizeX, elementSize, "pmeBsplineModuliX");
        pmeBsplineModuliY = new CudaArray(cu, gridSizeY, elementSize, "pmeBsplineModuliY");
        pmeBsplineModuliZ = new CudaArray(cu, gridSizeZ, elementSize, "pmeBsplineModuliZ");
        pmeIgrid = CudaArray::create<int4>(cu, numMultipoles, "pmeIgrid");
        pmePhi = new CudaArray(cu, 20*numMultipoles, elementSize, "pmePhi");
        pmePhid = new CudaArray(cu, 10*numMultipoles, elementSize, "pmePhid");
        pmePhip = new CudaArray(cu, 10*numMultipoles, elementSize, "pmePhip");
        pmePhidp = new CudaArray(cu, 20*numMultipoles, elementSize, "pmePhidp");
1275
        pmeCphi = new CudaArray(cu, 10*numMultipoles, elementSize, "pmeCphi");
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
1378
        pmeAtomRange = CudaArray::create<int>(cu, gridSizeX*gridSizeY*gridSizeZ+1, "pmeAtomRange");
        pmeAtomGridIndex = CudaArray::create<int2>(cu, numMultipoles, "pmeAtomGridIndex");
        sort = new CudaSort(cu, new SortTrait(), cu.getNumAtoms());
        cufftResult result = cufftPlan3d(&fft, gridSizeX, gridSizeY, gridSizeZ, cu.getUseDoublePrecision() ? CUFFT_Z2Z : CUFFT_C2C);
        if (result != CUFFT_SUCCESS)
            throw OpenMMException("Error initializing FFT: "+cu.intToString(result));
        hasInitializedFFT = true;

        // Initialize the b-spline moduli.

        double data[PmeOrder];
        double x = 0.0;
        data[0] = 1.0 - x;
        data[1] = x;
        for (int i = 2; i < PmeOrder; i++) {
            double denom = 1.0/i;
            data[i] = x*data[i-1]*denom;
            for (int j = 1; j < i; j++)
                data[i-j] = ((x+j)*data[i-j-1] + ((i-j+1)-x)*data[i-j])*denom;
            data[0] = (1.0-x)*data[0]*denom;
        }
        int maxSize = max(max(gridSizeX, gridSizeY), gridSizeZ);
        vector<double> bsplines_data(maxSize+1, 0.0);
        for (int i = 2; i <= PmeOrder+1; i++)
            bsplines_data[i] = data[i-2];
        for (int dim = 0; dim < 3; dim++) {
            int ndata = (dim == 0 ? gridSizeX : dim == 1 ? gridSizeY : gridSizeZ);
            vector<double> moduli(ndata);

            // get the modulus of the discrete Fourier transform

            double factor = 2.0*M_PI/ndata;
            for (int i = 0; i < ndata; i++) {
                double sc = 0.0;
                double ss = 0.0;
                for (int j = 1; j <= ndata; j++) {
                    double arg = factor*i*(j-1);
                    sc += bsplines_data[j]*cos(arg);
                    ss += bsplines_data[j]*sin(arg);
                }
                moduli[i] = sc*sc+ss*ss;
            }

            // Fix for exponential Euler spline interpolation failure.

            double eps = 1.0e-7;
            if (moduli[0] < eps)
                moduli[0] = 0.9*moduli[1];
            for (int i = 1; i < ndata-1; i++)
                if (moduli[i] < eps)
                    moduli[i] = 0.9*(moduli[i-1]+moduli[i+1]);
            if (moduli[ndata-1] < eps)
                moduli[ndata-1] = 0.9*moduli[ndata-2];

            // Compute and apply the optimal zeta coefficient.

            int jcut = 50;
            for (int i = 1; i <= ndata; i++) {
                int k = i - 1;
                if (i > ndata/2)
                    k = k - ndata;
                double zeta;
                if (k == 0)
                    zeta = 1.0;
                else {
                    double sum1 = 1.0;
                    double sum2 = 1.0;
                    factor = M_PI*k/ndata;
                    for (int j = 1; j <= jcut; j++) {
                        double arg = factor/(factor+M_PI*j);
                        sum1 += pow(arg, PmeOrder);
                        sum2 += pow(arg, 2*PmeOrder);
                    }
                    for (int j = 1; j <= jcut; j++) {
                        double arg = factor/(factor-M_PI*j);
                        sum1 += pow(arg, PmeOrder);
                        sum2 += pow(arg, 2*PmeOrder);
                    }
                    zeta = sum2/sum1;
                }
                moduli[i-1] = moduli[i-1]*zeta*zeta;
            }
            if (cu.getUseDoublePrecision()) {
                if (dim == 0)
                    pmeBsplineModuliX->upload(moduli);
                else if (dim == 1)
                    pmeBsplineModuliY->upload(moduli);
                else
                    pmeBsplineModuliZ->upload(moduli);
            }
            else {
                vector<float> modulif(ndata);
                for (int i = 0; i < ndata; i++)
                    modulif[i] = (float) moduli[i];
                if (dim == 0)
                    pmeBsplineModuliX->upload(modulif);
                else if (dim == 1)
                    pmeBsplineModuliY->upload(modulif);
                else
                    pmeBsplineModuliZ->upload(modulif);
            }
        }
    }
1379
1380
1381
1382
1383

    // Add an interaction to the default nonbonded kernel.  This doesn't actually do any calculations.  It's
    // just so that CudaNonbondedUtilities will build the exclusion flags and maintain the neighbor list.
    
    cu.getNonbondedUtilities().addInteraction(usePME, usePME, true, force.getCutoffDistance(), exclusions, "", force.getForceGroup());
1384
    cu.getNonbondedUtilities().setUsePadding(false);
1385
1386
1387
1388
1389
1390
1391
1392
    cu.addForce(new ForceInfo(force));
}

void CudaCalcAmoebaMultipoleForceKernel::initializeScaleFactors() {
    hasInitializedScaleFactors = true;
    CudaNonbondedUtilities& nb = cu.getNonbondedUtilities();
    
    // Figure out the covalent flag values to use for each atom pair.
1393
1394
1395
1396
1397
1398
1399
1400

    vector<ushort2> exclusionTiles;
    nb.getExclusionTiles().download(exclusionTiles);
    map<pair<int, int>, int> exclusionTileMap;
    for (int i = 0; i < (int) exclusionTiles.size(); i++) {
        ushort2 tile = exclusionTiles[i];
        exclusionTileMap[make_pair(tile.x, tile.y)] = i;
    }
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
    covalentFlags = CudaArray::create<uint2>(cu, nb.getExclusions().getSize(), "covalentFlags");
    vector<uint2> covalentFlagsVec(nb.getExclusions().getSize(), make_uint2(0, 0));
    for (int i = 0; i < (int) covalentFlagValues.size(); i++) {
        int atom1 = covalentFlagValues[i].x;
        int atom2 = covalentFlagValues[i].y;
        int value = covalentFlagValues[i].z;
        int x = atom1/CudaContext::TileSize;
        int offset1 = atom1-x*CudaContext::TileSize;
        int y = atom2/CudaContext::TileSize;
        int offset2 = atom2-y*CudaContext::TileSize;
        int f1 = (value == 0 || value == 1 ? 1 : 0);
        int f2 = (value == 0 || value == 2 ? 1 : 0);
1413
        if (x == y) {
1414
            int index = exclusionTileMap[make_pair(x, y)]*CudaContext::TileSize;
1415
1416
1417
1418
1419
1420
            covalentFlagsVec[index+offset1].x |= f1<<offset2;
            covalentFlagsVec[index+offset1].y |= f2<<offset2;
            covalentFlagsVec[index+offset2].x |= f1<<offset1;
            covalentFlagsVec[index+offset2].y |= f2<<offset1;
        }
        else if (x > y) {
1421
            int index = exclusionTileMap[make_pair(x, y)]*CudaContext::TileSize;
1422
1423
1424
1425
            covalentFlagsVec[index+offset1].x |= f1<<offset2;
            covalentFlagsVec[index+offset1].y |= f2<<offset2;
        }
        else {
1426
            int index = exclusionTileMap[make_pair(y, x)]*CudaContext::TileSize;
1427
1428
1429
1430
1431
1432
1433
1434
1435
1436
1437
1438
1439
1440
1441
1442
1443
            covalentFlagsVec[index+offset2].x |= f1<<offset1;
            covalentFlagsVec[index+offset2].y |= f2<<offset1;
        }
    }
    covalentFlags->upload(covalentFlagsVec);
    
    // Do the same for the polarization flags.
    
    polarizationGroupFlags = CudaArray::create<unsigned int>(cu, nb.getExclusions().getSize(), "polarizationGroupFlags");
    vector<unsigned int> polarizationGroupFlagsVec(nb.getExclusions().getSize(), 0);
    for (int i = 0; i < (int) polarizationFlagValues.size(); i++) {
        int atom1 = polarizationFlagValues[i].x;
        int atom2 = polarizationFlagValues[i].y;
        int x = atom1/CudaContext::TileSize;
        int offset1 = atom1-x*CudaContext::TileSize;
        int y = atom2/CudaContext::TileSize;
        int offset2 = atom2-y*CudaContext::TileSize;
1444
        if (x == y) {
1445
            int index = exclusionTileMap[make_pair(x, y)]*CudaContext::TileSize;
1446
1447
1448
1449
            polarizationGroupFlagsVec[index+offset1] |= 1<<offset2;
            polarizationGroupFlagsVec[index+offset2] |= 1<<offset1;
        }
        else if (x > y) {
1450
            int index = exclusionTileMap[make_pair(x, y)]*CudaContext::TileSize;
1451
1452
1453
            polarizationGroupFlagsVec[index+offset1] |= 1<<offset2;
        }
        else {
1454
            int index = exclusionTileMap[make_pair(y, x)]*CudaContext::TileSize;
1455
1456
1457
1458
1459
1460
1461
            polarizationGroupFlagsVec[index+offset2] |= 1<<offset1;
        }
    }
    polarizationGroupFlags->upload(polarizationGroupFlagsVec);
}

double CudaCalcAmoebaMultipoleForceKernel::execute(ContextImpl& context, bool includeForces, bool includeEnergy) {
1462
    if (!hasInitializedScaleFactors) {
1463
        initializeScaleFactors();
1464
1465
1466
1467
1468
1469
        for (int i = 0; i < (int) context.getForceImpls().size() && gkKernel == NULL; i++) {
            AmoebaGeneralizedKirkwoodForceImpl* gkImpl = dynamic_cast<AmoebaGeneralizedKirkwoodForceImpl*>(context.getForceImpls()[i]);
            if (gkImpl != NULL)
                gkKernel = dynamic_cast<CudaCalcAmoebaGeneralizedKirkwoodForceKernel*>(&gkImpl->getKernel().getImpl());
        }
    }
1470
1471
1472
1473
1474
1475
    CudaNonbondedUtilities& nb = cu.getNonbondedUtilities();
    
    // Compute the lab frame moments.

    void* computeMomentsArgs[] = {&cu.getPosq().getDevicePointer(), &multipoleParticles->getDevicePointer(),
        &molecularDipoles->getDevicePointer(), &molecularQuadrupoles->getDevicePointer(),
1476
1477
        &labFrameDipoles->getDevicePointer(), &labFrameQuadrupoles->getDevicePointer(),
        &sphericalDipoles->getDevicePointer(), &sphericalQuadrupoles->getDevicePointer()};
1478
    cu.executeKernel(computeMomentsKernel, computeMomentsArgs, cu.getNumAtoms());
1479
1480
1481
    int startTileIndex = nb.getStartTileIndex();
    int numTileIndices = nb.getNumTiles();
    int numForceThreadBlocks = nb.getNumForceThreadBlocks();
1482
    int elementSize = (cu.getUseDoublePrecision() ? sizeof(double) : sizeof(float));
1483
    if (pmeGrid == NULL) {
1484
1485
        // Compute induced dipoles.
        
1486
1487
        if (gkKernel == NULL) {
            void* computeFixedFieldArgs[] = {&field->getDevicePointer(), &fieldPolar->getDevicePointer(), &cu.getPosq().getDevicePointer(),
1488
                &covalentFlags->getDevicePointer(), &polarizationGroupFlags->getDevicePointer(), &nb.getExclusionTiles().getDevicePointer(), &startTileIndex, &numTileIndices,
1489
                &labFrameDipoles->getDevicePointer(), &labFrameQuadrupoles->getDevicePointer(), &dampingAndThole->getDevicePointer()};
1490
            cu.executeKernel(computeFixedFieldKernel, computeFixedFieldArgs, numForceThreadBlocks*fixedFieldThreads, fixedFieldThreads);
1491
1492
1493
1494
1495
1496
1497
            void* recordInducedDipolesArgs[] = {&field->getDevicePointer(), &fieldPolar->getDevicePointer(),
                &inducedDipole->getDevicePointer(), &inducedDipolePolar->getDevicePointer(), &polarizability->getDevicePointer()};
            cu.executeKernel(recordInducedDipolesKernel, recordInducedDipolesArgs, cu.getNumAtoms());
        }
        else {
            gkKernel->computeBornRadii();
            void* computeFixedFieldArgs[] = {&field->getDevicePointer(), &fieldPolar->getDevicePointer(), &cu.getPosq().getDevicePointer(),
1498
                &covalentFlags->getDevicePointer(), &polarizationGroupFlags->getDevicePointer(), &nb.getExclusionTiles().getDevicePointer(), &startTileIndex, &numTileIndices,
1499
1500
                &gkKernel->getBornRadii()->getDevicePointer(), &gkKernel->getField()->getDevicePointer(),
                &labFrameDipoles->getDevicePointer(), &labFrameQuadrupoles->getDevicePointer(), &dampingAndThole->getDevicePointer()};
1501
            cu.executeKernel(computeFixedFieldKernel, computeFixedFieldArgs, numForceThreadBlocks*fixedFieldThreads, fixedFieldThreads);
1502
1503
1504
1505
1506
1507
            void* recordInducedDipolesArgs[] = {&field->getDevicePointer(), &fieldPolar->getDevicePointer(),
                &gkKernel->getField()->getDevicePointer(), &gkKernel->getInducedDipoles()->getDevicePointer(),
                &gkKernel->getInducedDipolesPolar()->getDevicePointer(), &inducedDipole->getDevicePointer(),
                &inducedDipolePolar->getDevicePointer(), &polarizability->getDevicePointer()};
            cu.executeKernel(recordInducedDipolesKernel, recordInducedDipolesArgs, cu.getNumAtoms());
        }
1508
1509
1510
        
        // Iterate until the dipoles converge.
        
1511
1512
        if (polarizationType == AmoebaMultipoleForce::Extrapolated)
            computeExtrapolatedDipoles(NULL);
1513
        for (int i = 0; i < maxInducedIterations; i++) {
1514
            computeInducedField(NULL);
peastman's avatar
peastman committed
1515
1516
            bool converged = iterateDipolesByDIIS(i);
            if (converged)
1517
                break;
1518
        }
1519
1520
1521
        
        // Compute electrostatic force.
        
1522
        void* electrostaticsArgs[] = {&cu.getForce().getDevicePointer(), &torque->getDevicePointer(), &cu.getEnergyBuffer().getDevicePointer(),
1523
1524
            &cu.getPosq().getDevicePointer(), &covalentFlags->getDevicePointer(), &polarizationGroupFlags->getDevicePointer(),
            &nb.getExclusionTiles().getDevicePointer(), &startTileIndex, &numTileIndices,
1525
            &sphericalDipoles->getDevicePointer(), &sphericalQuadrupoles->getDevicePointer(),
1526
            &inducedDipole->getDevicePointer(), &inducedDipolePolar->getDevicePointer(), &dampingAndThole->getDevicePointer()};
1527
        cu.executeKernel(electrostaticsKernel, electrostaticsArgs, numForceThreadBlocks*electrostaticsThreads, electrostaticsThreads);
1528
1529
        if (gkKernel != NULL)
            gkKernel->finishComputation(*torque, *labFrameDipoles, *labFrameQuadrupoles, *inducedDipole, *inducedDipolePolar, *dampingAndThole, *covalentFlags, *polarizationGroupFlags);
1530
    }
1531
    else {
1532
1533
1534
1535
1536
1537
1538
1539
1540
1541
1542
1543
1544
1545
1546
1547
1548
1549
1550
1551
1552
1553
1554
1555
1556
1557
        // Compute reciprocal box vectors.
        
        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;
        double3 recipBoxVectors[3];
        recipBoxVectors[0] = make_double3(boxVectors[1][1]*boxVectors[2][2]*scale, 0, 0);
        recipBoxVectors[1] = make_double3(-boxVectors[1][0]*boxVectors[2][2]*scale, boxVectors[0][0]*boxVectors[2][2]*scale, 0);
        recipBoxVectors[2] = make_double3((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);
        float3 recipBoxVectorsFloat[3];
        void* recipBoxVectorPointer[3];
        if (cu.getUseDoublePrecision()) {
            recipBoxVectorPointer[0] = &recipBoxVectors[0];
            recipBoxVectorPointer[1] = &recipBoxVectors[1];
            recipBoxVectorPointer[2] = &recipBoxVectors[2];
        }
        else {
            recipBoxVectorsFloat[0] = make_float3((float) recipBoxVectors[0].x, 0, 0);
            recipBoxVectorsFloat[1] = make_float3((float) recipBoxVectors[1].x, (float) recipBoxVectors[1].y, 0);
            recipBoxVectorsFloat[2] = make_float3((float) recipBoxVectors[2].x, (float) recipBoxVectors[2].y, (float) recipBoxVectors[2].z);
            recipBoxVectorPointer[0] = &recipBoxVectorsFloat[0];
            recipBoxVectorPointer[1] = &recipBoxVectorsFloat[1];
            recipBoxVectorPointer[2] = &recipBoxVectorsFloat[2];
        }

1558
        // Reciprocal space calculation.
1559
1560
        
        unsigned int maxTiles = nb.getInteractingTiles().getSize();
1561
        void* gridIndexArgs[] = {&cu.getPosq().getDevicePointer(), &pmeAtomGridIndex->getDevicePointer(),
1562
1563
            cu.getPeriodicBoxVecXPointer(), cu.getPeriodicBoxVecYPointer(), cu.getPeriodicBoxVecZPointer(),
            recipBoxVectorPointer[0], recipBoxVectorPointer[1], recipBoxVectorPointer[2]};
1564
        cu.executeKernel(pmeGridIndexKernel, gridIndexArgs, cu.getNumAtoms(), cu.ThreadBlockSize, cu.ThreadBlockSize*PmeOrder*PmeOrder*elementSize);
1565
        sort->sort(*pmeAtomGridIndex);
1566
1567
1568
1569
        void* pmeTransformMultipolesArgs[] = {&labFrameDipoles->getDevicePointer(), &labFrameQuadrupoles->getDevicePointer(),
            &fracDipoles->getDevicePointer(), &fracQuadrupoles->getDevicePointer(), recipBoxVectorPointer[0], recipBoxVectorPointer[1], recipBoxVectorPointer[2]};
        cu.executeKernel(pmeTransformMultipolesKernel, pmeTransformMultipolesArgs, cu.getNumAtoms());
        void* pmeSpreadFixedMultipolesArgs[] = {&cu.getPosq().getDevicePointer(), &fracDipoles->getDevicePointer(), &fracQuadrupoles->getDevicePointer(),
1570
1571
            &pmeGrid->getDevicePointer(), &pmeAtomGridIndex->getDevicePointer(),  cu.getPeriodicBoxVecXPointer(), cu.getPeriodicBoxVecYPointer(), cu.getPeriodicBoxVecZPointer(),
            recipBoxVectorPointer[0], recipBoxVectorPointer[1], recipBoxVectorPointer[2]};
1572
        cu.executeKernel(pmeSpreadFixedMultipolesKernel, pmeSpreadFixedMultipolesArgs, cu.getNumAtoms());
1573
1574
1575
        void* finishSpreadArgs[] = {&pmeGrid->getDevicePointer()};
        if (cu.getUseDoublePrecision())
            cu.executeKernel(pmeFinishSpreadChargeKernel, finishSpreadArgs, pmeGrid->getSize());
1576
1577
1578
1579
1580
        if (cu.getUseDoublePrecision())
            cufftExecZ2Z(fft, (double2*) pmeGrid->getDevicePointer(), (double2*) pmeGrid->getDevicePointer(), CUFFT_FORWARD);
        else
            cufftExecC2C(fft, (float2*) pmeGrid->getDevicePointer(), (float2*) pmeGrid->getDevicePointer(), CUFFT_FORWARD);
        void* pmeConvolutionArgs[] = {&pmeGrid->getDevicePointer(), &pmeBsplineModuliX->getDevicePointer(), &pmeBsplineModuliY->getDevicePointer(),
1581
            &pmeBsplineModuliZ->getDevicePointer(), cu.getPeriodicBoxSizePointer(), recipBoxVectorPointer[0], recipBoxVectorPointer[1], recipBoxVectorPointer[2]};
1582
1583
1584
1585
1586
1587
        cu.executeKernel(pmeConvolutionKernel, pmeConvolutionArgs, cu.getNumAtoms());
        if (cu.getUseDoublePrecision())
            cufftExecZ2Z(fft, (double2*) pmeGrid->getDevicePointer(), (double2*) pmeGrid->getDevicePointer(), CUFFT_INVERSE);
        else
            cufftExecC2C(fft, (float2*) pmeGrid->getDevicePointer(), (float2*) pmeGrid->getDevicePointer(), CUFFT_INVERSE);
        void* pmeFixedPotentialArgs[] = {&pmeGrid->getDevicePointer(), &pmePhi->getDevicePointer(), &field->getDevicePointer(),
1588
            &fieldPolar ->getDevicePointer(), &cu.getPosq().getDevicePointer(), &labFrameDipoles->getDevicePointer(),
1589
1590
            cu.getPeriodicBoxVecXPointer(), cu.getPeriodicBoxVecYPointer(), cu.getPeriodicBoxVecZPointer(),
            recipBoxVectorPointer[0], recipBoxVectorPointer[1], recipBoxVectorPointer[2], &pmeAtomGridIndex->getDevicePointer()};
1591
        cu.executeKernel(pmeFixedPotentialKernel, pmeFixedPotentialArgs, cu.getNumAtoms());
1592
1593
        void* pmeTransformFixedPotentialArgs[] = {&pmePhi->getDevicePointer(), &pmeCphi->getDevicePointer(), recipBoxVectorPointer[0], recipBoxVectorPointer[1], recipBoxVectorPointer[2]};
        cu.executeKernel(pmeTransformPotentialKernel, pmeTransformFixedPotentialArgs, cu.getNumAtoms());
1594
1595
        void* pmeFixedForceArgs[] = {&cu.getPosq().getDevicePointer(), &cu.getForce().getDevicePointer(), &torque->getDevicePointer(),
            &cu.getEnergyBuffer().getDevicePointer(), &labFrameDipoles->getDevicePointer(), &labFrameQuadrupoles->getDevicePointer(),
1596
            &fracDipoles->getDevicePointer(), &fracQuadrupoles->getDevicePointer(), &pmePhi->getDevicePointer(), &pmeCphi->getDevicePointer(),
1597
            recipBoxVectorPointer[0], recipBoxVectorPointer[1], recipBoxVectorPointer[2]};
1598
        cu.executeKernel(pmeFixedForceKernel, pmeFixedForceArgs, cu.getNumAtoms());
1599
1600
1601
        
        // Direct space calculation.
        
1602
        void* computeFixedFieldArgs[] = {&field->getDevicePointer(), &fieldPolar->getDevicePointer(), &cu.getPosq().getDevicePointer(),
1603
            &covalentFlags->getDevicePointer(), &polarizationGroupFlags->getDevicePointer(), &nb.getExclusionTiles().getDevicePointer(), &startTileIndex, &numTileIndices,
1604
            &nb.getInteractingTiles().getDevicePointer(), &nb.getInteractionCount().getDevicePointer(), cu.getPeriodicBoxSizePointer(),
1605
1606
            cu.getInvPeriodicBoxSizePointer(), cu.getPeriodicBoxVecXPointer(), cu.getPeriodicBoxVecYPointer(), cu.getPeriodicBoxVecZPointer(),
            &maxTiles, &nb.getBlockCenters().getDevicePointer(), &nb.getInteractingAtoms().getDevicePointer(),
1607
            &labFrameDipoles->getDevicePointer(), &labFrameQuadrupoles->getDevicePointer(), &dampingAndThole->getDevicePointer()};
1608
        cu.executeKernel(computeFixedFieldKernel, computeFixedFieldArgs, numForceThreadBlocks*fixedFieldThreads, fixedFieldThreads);
1609
1610
1611
        void* recordInducedDipolesArgs[] = {&field->getDevicePointer(), &fieldPolar->getDevicePointer(),
            &inducedDipole->getDevicePointer(), &inducedDipolePolar->getDevicePointer(), &polarizability->getDevicePointer()};
        cu.executeKernel(recordInducedDipolesKernel, recordInducedDipolesArgs, cu.getNumAtoms());
1612
1613
1614

        // Reciprocal space calculation for the induced dipoles.

1615
        cu.clearBuffer(*pmeGrid);
1616
        void* pmeSpreadInducedDipolesArgs[] = {&cu.getPosq().getDevicePointer(), &inducedDipole->getDevicePointer(), &inducedDipolePolar->getDevicePointer(),
1617
1618
            &pmeGrid->getDevicePointer(), &pmeAtomGridIndex->getDevicePointer(), cu.getPeriodicBoxVecXPointer(), cu.getPeriodicBoxVecYPointer(), cu.getPeriodicBoxVecZPointer(),
            recipBoxVectorPointer[0], recipBoxVectorPointer[1], recipBoxVectorPointer[2]};
1619
        cu.executeKernel(pmeSpreadInducedDipolesKernel, pmeSpreadInducedDipolesArgs, cu.getNumAtoms());
1620
1621
        if (cu.getUseDoublePrecision())
            cu.executeKernel(pmeFinishSpreadChargeKernel, finishSpreadArgs, pmeGrid->getSize());
1622
1623
1624
1625
1626
1627
1628
1629
1630
1631
        if (cu.getUseDoublePrecision())
            cufftExecZ2Z(fft, (double2*) pmeGrid->getDevicePointer(), (double2*) pmeGrid->getDevicePointer(), CUFFT_FORWARD);
        else
            cufftExecC2C(fft, (float2*) pmeGrid->getDevicePointer(), (float2*) pmeGrid->getDevicePointer(), CUFFT_FORWARD);
        cu.executeKernel(pmeConvolutionKernel, pmeConvolutionArgs, cu.getNumAtoms());
        if (cu.getUseDoublePrecision())
            cufftExecZ2Z(fft, (double2*) pmeGrid->getDevicePointer(), (double2*) pmeGrid->getDevicePointer(), CUFFT_INVERSE);
        else
            cufftExecC2C(fft, (float2*) pmeGrid->getDevicePointer(), (float2*) pmeGrid->getDevicePointer(), CUFFT_INVERSE);
        void* pmeInducedPotentialArgs[] = {&pmeGrid->getDevicePointer(), &pmePhid->getDevicePointer(), &pmePhip->getDevicePointer(),
1632
1633
            &pmePhidp->getDevicePointer(), &cu.getPosq().getDevicePointer(), cu.getPeriodicBoxVecXPointer(), cu.getPeriodicBoxVecYPointer(),
            cu.getPeriodicBoxVecZPointer(), recipBoxVectorPointer[0], recipBoxVectorPointer[1], recipBoxVectorPointer[2],
1634
            &pmeAtomGridIndex->getDevicePointer()};
1635
        cu.executeKernel(pmeInducedPotentialKernel, pmeInducedPotentialArgs, cu.getNumAtoms());
1636
        
1637
        // Iterate until the dipoles converge.
1638
        
1639
1640
        if (polarizationType == AmoebaMultipoleForce::Extrapolated)
            computeExtrapolatedDipoles(recipBoxVectorPointer);
1641
        for (int i = 0; i < maxInducedIterations; i++) {
1642
            computeInducedField(recipBoxVectorPointer);
peastman's avatar
peastman committed
1643
1644
            bool converged = iterateDipolesByDIIS(i);
            if (converged)
1645
1646
                break;
        }
1647
1648
1649
1650
        
        // Compute electrostatic force.
        
        void* electrostaticsArgs[] = {&cu.getForce().getDevicePointer(), &torque->getDevicePointer(), &cu.getEnergyBuffer().getDevicePointer(),
1651
1652
            &cu.getPosq().getDevicePointer(), &covalentFlags->getDevicePointer(), &polarizationGroupFlags->getDevicePointer(),
            &nb.getExclusionTiles().getDevicePointer(), &startTileIndex, &numTileIndices,
1653
            &nb.getInteractingTiles().getDevicePointer(), &nb.getInteractionCount().getDevicePointer(),
1654
1655
            cu.getPeriodicBoxSizePointer(), cu.getInvPeriodicBoxSizePointer(), cu.getPeriodicBoxVecXPointer(), cu.getPeriodicBoxVecYPointer(), cu.getPeriodicBoxVecZPointer(),
            &maxTiles, &nb.getBlockCenters().getDevicePointer(), &nb.getInteractingAtoms().getDevicePointer(),
1656
            &sphericalDipoles->getDevicePointer(), &sphericalQuadrupoles->getDevicePointer(),
1657
            &inducedDipole->getDevicePointer(), &inducedDipolePolar->getDevicePointer(), &dampingAndThole->getDevicePointer()};
1658
        cu.executeKernel(electrostaticsKernel, electrostaticsArgs, numForceThreadBlocks*electrostaticsThreads, electrostaticsThreads);
1659
1660
        void* pmeTransformInducedPotentialArgs[] = {&pmePhidp->getDevicePointer(), &pmeCphi->getDevicePointer(), recipBoxVectorPointer[0], recipBoxVectorPointer[1], recipBoxVectorPointer[2]};
        cu.executeKernel(pmeTransformPotentialKernel, pmeTransformInducedPotentialArgs, cu.getNumAtoms());
1661
1662
        void* pmeInducedForceArgs[] = {&cu.getPosq().getDevicePointer(), &cu.getForce().getDevicePointer(), &torque->getDevicePointer(),
            &cu.getEnergyBuffer().getDevicePointer(), &labFrameDipoles->getDevicePointer(), &labFrameQuadrupoles->getDevicePointer(),
1663
            &fracDipoles->getDevicePointer(), &fracQuadrupoles->getDevicePointer(),
1664
            &inducedDipole->getDevicePointer(), &inducedDipolePolar->getDevicePointer(), &pmePhi->getDevicePointer(), &pmePhid->getDevicePointer(),
1665
            &pmePhip->getDevicePointer(), &pmePhidp->getDevicePointer(), &pmeCphi->getDevicePointer(), recipBoxVectorPointer[0], recipBoxVectorPointer[1], recipBoxVectorPointer[2]};
1666
        cu.executeKernel(pmeInducedForceKernel, pmeInducedForceArgs, cu.getNumAtoms());
1667
    }
1668
1669
1670
1671
1672
1673
1674
1675
    
    // If using extrapolated polarization, add in force contributions from µ(m) T µ(n).
    
    if (polarizationType == AmoebaMultipoleForce::Extrapolated) {
        void* extrapolatedArgs[] = {&cu.getForce().getDevicePointer(), &extrapolatedDipole->getDevicePointer(),
            &extrapolatedDipolePolar->getDevicePointer(), &extrapolatedDipoleFieldGradient->getDevicePointer(), &extrapolatedDipoleFieldGradientPolar->getDevicePointer()};
        cu.executeKernel(addExtrapolatedGradientKernel, extrapolatedArgs, numMultipoles);
    }
Peter Eastman's avatar
Peter Eastman committed
1676
1677
1678
1679
1680
1681

    // Map torques to force.

    void* mapTorqueArgs[] = {&cu.getForce().getDevicePointer(), &torque->getDevicePointer(),
        &cu.getPosq().getDevicePointer(), &multipoleParticles->getDevicePointer()};
    cu.executeKernel(mapTorqueKernel, mapTorqueArgs, cu.getNumAtoms());
1682
1683
1684
1685
1686
    
    // Record the current atom positions so we can tell later if they have changed.
    
    cu.getPosq().copyTo(*lastPositions);
    multipolesAreValid = true;
1687
1688
1689
    return 0.0;
}

1690
1691
1692
1693
1694
void CudaCalcAmoebaMultipoleForceKernel::computeInducedField(void** recipBoxVectorPointer) {
    CudaNonbondedUtilities& nb = cu.getNonbondedUtilities();
    int startTileIndex = nb.getStartTileIndex();
    int numTileIndices = nb.getNumTiles();
    int numForceThreadBlocks = nb.getNumForceThreadBlocks();
1695
1696
1697
1698
1699
1700
1701
1702
1703
1704
1705
1706
1707
1708
1709
1710
1711
1712
1713
1714
1715
1716
1717
1718
1719
1720
1721
1722
1723
1724
1725
1726
1727
1728
1729
    unsigned int maxTiles = 0;
    vector<void*> computeInducedFieldArgs;
    computeInducedFieldArgs.push_back(&inducedField->getDevicePointer());
    computeInducedFieldArgs.push_back(&inducedFieldPolar->getDevicePointer());
    computeInducedFieldArgs.push_back(&cu.getPosq().getDevicePointer());
    computeInducedFieldArgs.push_back(&nb.getExclusionTiles().getDevicePointer());
    computeInducedFieldArgs.push_back(&inducedDipole->getDevicePointer());
    computeInducedFieldArgs.push_back(&inducedDipolePolar->getDevicePointer());
    computeInducedFieldArgs.push_back(&startTileIndex);
    computeInducedFieldArgs.push_back(&numTileIndices);
    if (polarizationType == AmoebaMultipoleForce::Extrapolated) {
        computeInducedFieldArgs.push_back(&inducedDipoleFieldGradient->getDevicePointer());
        computeInducedFieldArgs.push_back(&inducedDipoleFieldGradientPolar->getDevicePointer());
    }
    if (pmeGrid != NULL) {
        computeInducedFieldArgs.push_back(&nb.getInteractingTiles().getDevicePointer());
        computeInducedFieldArgs.push_back(&nb.getInteractionCount().getDevicePointer());
        computeInducedFieldArgs.push_back(cu.getPeriodicBoxSizePointer());
        computeInducedFieldArgs.push_back(cu.getInvPeriodicBoxSizePointer());
        computeInducedFieldArgs.push_back(cu.getPeriodicBoxVecXPointer());
        computeInducedFieldArgs.push_back(cu.getPeriodicBoxVecYPointer());
        computeInducedFieldArgs.push_back(cu.getPeriodicBoxVecZPointer());
        computeInducedFieldArgs.push_back(&maxTiles);
        computeInducedFieldArgs.push_back(&nb.getBlockCenters().getDevicePointer());
        computeInducedFieldArgs.push_back(&nb.getInteractingAtoms().getDevicePointer());
    }
    if (gkKernel != NULL) {
        computeInducedFieldArgs.push_back(&gkKernel->getInducedField()->getDevicePointer());
        computeInducedFieldArgs.push_back(&gkKernel->getInducedFieldPolar()->getDevicePointer());
        computeInducedFieldArgs.push_back(&gkKernel->getInducedDipoles()->getDevicePointer());
        computeInducedFieldArgs.push_back(&gkKernel->getInducedDipolesPolar()->getDevicePointer());
        computeInducedFieldArgs.push_back(&gkKernel->getBornRadii()->getDevicePointer());
        if (polarizationType == AmoebaMultipoleForce::Extrapolated) {
            computeInducedFieldArgs.push_back(&inducedDipoleFieldGradientGk->getDevicePointer());
            computeInducedFieldArgs.push_back(&inducedDipoleFieldGradientGkPolar->getDevicePointer());
1730
        }
1731
1732
1733
1734
1735
1736
1737
1738
1739
1740
1741
1742
1743
1744
    }
    computeInducedFieldArgs.push_back(&dampingAndThole->getDevicePointer());
    cu.clearBuffer(*inducedField);
    cu.clearBuffer(*inducedFieldPolar);
    if (polarizationType == AmoebaMultipoleForce::Extrapolated) {
        cu.clearBuffer(*inducedDipoleFieldGradient);
        cu.clearBuffer(*inducedDipoleFieldGradientPolar);
    }
    if (gkKernel != NULL) {
        cu.clearBuffer(*gkKernel->getInducedField());
        cu.clearBuffer(*gkKernel->getInducedFieldPolar());
        if (polarizationType == AmoebaMultipoleForce::Extrapolated) {
            cu.clearBuffer(*inducedDipoleFieldGradientGk);
            cu.clearBuffer(*inducedDipoleFieldGradientGkPolar);
1745
1746
        }
    }
1747
1748
    if (pmeGrid == NULL)
        cu.executeKernel(computeInducedFieldKernel, &computeInducedFieldArgs[0], numForceThreadBlocks*inducedFieldThreads, inducedFieldThreads);
1749
    else {
1750
1751
        maxTiles = nb.getInteractingTiles().getSize();
        cu.executeKernel(computeInducedFieldKernel, &computeInducedFieldArgs[0], numForceThreadBlocks*inducedFieldThreads, inducedFieldThreads);
1752
1753
1754
1755
1756
1757
1758
1759
1760
1761
1762
1763
1764
1765
1766
1767
1768
1769
1770
1771
1772
1773
1774
1775
1776
        cu.clearBuffer(*pmeGrid);
        void* pmeSpreadInducedDipolesArgs[] = {&cu.getPosq().getDevicePointer(), &inducedDipole->getDevicePointer(), &inducedDipolePolar->getDevicePointer(),
            &pmeGrid->getDevicePointer(), &pmeAtomGridIndex->getDevicePointer(), cu.getPeriodicBoxVecXPointer(), cu.getPeriodicBoxVecYPointer(), cu.getPeriodicBoxVecZPointer(),
            recipBoxVectorPointer[0], recipBoxVectorPointer[1], recipBoxVectorPointer[2]};
        cu.executeKernel(pmeSpreadInducedDipolesKernel, pmeSpreadInducedDipolesArgs, cu.getNumAtoms());
        if (cu.getUseDoublePrecision()) {
            void* finishSpreadArgs[] = {&pmeGrid->getDevicePointer()};
            cu.executeKernel(pmeFinishSpreadChargeKernel, finishSpreadArgs, pmeGrid->getSize());
        }
        if (cu.getUseDoublePrecision())
            cufftExecZ2Z(fft, (double2*) pmeGrid->getDevicePointer(), (double2*) pmeGrid->getDevicePointer(), CUFFT_FORWARD);
        else
            cufftExecC2C(fft, (float2*) pmeGrid->getDevicePointer(), (float2*) pmeGrid->getDevicePointer(), CUFFT_FORWARD);
        void* pmeConvolutionArgs[] = {&pmeGrid->getDevicePointer(), &pmeBsplineModuliX->getDevicePointer(), &pmeBsplineModuliY->getDevicePointer(),
            &pmeBsplineModuliZ->getDevicePointer(), cu.getPeriodicBoxSizePointer(), recipBoxVectorPointer[0], recipBoxVectorPointer[1], recipBoxVectorPointer[2]};
        cu.executeKernel(pmeConvolutionKernel, pmeConvolutionArgs, cu.getNumAtoms());
        if (cu.getUseDoublePrecision())
            cufftExecZ2Z(fft, (double2*) pmeGrid->getDevicePointer(), (double2*) pmeGrid->getDevicePointer(), CUFFT_INVERSE);
        else
            cufftExecC2C(fft, (float2*) pmeGrid->getDevicePointer(), (float2*) pmeGrid->getDevicePointer(), CUFFT_INVERSE);
        void* pmeInducedPotentialArgs[] = {&pmeGrid->getDevicePointer(), &pmePhid->getDevicePointer(), &pmePhip->getDevicePointer(),
            &pmePhidp->getDevicePointer(), &cu.getPosq().getDevicePointer(), cu.getPeriodicBoxVecXPointer(), cu.getPeriodicBoxVecYPointer(),
            cu.getPeriodicBoxVecZPointer(), recipBoxVectorPointer[0], recipBoxVectorPointer[1], recipBoxVectorPointer[2],
            &pmeAtomGridIndex->getDevicePointer()};
        cu.executeKernel(pmeInducedPotentialKernel, pmeInducedPotentialArgs, cu.getNumAtoms());
1777
1778
        if (polarizationType == AmoebaMultipoleForce::Extrapolated) {
            void* pmeRecordInducedFieldDipolesArgs[] = {&pmePhid->getDevicePointer(), &pmePhip->getDevicePointer(),
1779
1780
1781
                &inducedField->getDevicePointer(), &inducedFieldPolar->getDevicePointer(), &inducedDipole->getDevicePointer(),
                &inducedDipolePolar->getDevicePointer(), &inducedDipoleFieldGradient->getDevicePointer(), &inducedDipoleFieldGradientPolar->getDevicePointer(),
                recipBoxVectorPointer[0], recipBoxVectorPointer[1], recipBoxVectorPointer[2]};
1782
1783
1784
1785
            cu.executeKernel(pmeRecordInducedFieldDipolesKernel, pmeRecordInducedFieldDipolesArgs, cu.getNumAtoms());
        }
        else {
            void* pmeRecordInducedFieldDipolesArgs[] = {&pmePhid->getDevicePointer(), &pmePhip->getDevicePointer(),
1786
1787
                &inducedField->getDevicePointer(), &inducedFieldPolar->getDevicePointer(), &inducedDipole->getDevicePointer(), &inducedDipolePolar->getDevicePointer(),
                recipBoxVectorPointer[0], recipBoxVectorPointer[1], recipBoxVectorPointer[2]};
1788
1789
            cu.executeKernel(pmeRecordInducedFieldDipolesKernel, pmeRecordInducedFieldDipolesArgs, cu.getNumAtoms());
        }
1790
1791
1792
    }
}

peastman's avatar
peastman committed
1793
bool CudaCalcAmoebaMultipoleForceKernel::iterateDipolesByDIIS(int iteration) {
peastman's avatar
peastman committed
1794
1795
1796
1797
    void* npt = NULL;
    bool trueValue = true, falseValue = false;
    int elementSize = (cu.getUseDoublePrecision() ? sizeof(double) : sizeof(float));
    
peastman's avatar
peastman committed
1798
    // Record the dipoles and errors into the lists of previous dipoles.
peastman's avatar
peastman committed
1799
1800
1801
1802
1803
    
    if (gkKernel != NULL) {
        void* recordDIISDipolesGkArgs[] = {&field->getDevicePointer(), &fieldPolar->getDevicePointer(), &gkKernel->getField()->getDevicePointer(), &gkKernel->getInducedField()->getDevicePointer(),
            &gkKernel->getInducedFieldPolar()->getDevicePointer(), &gkKernel->getInducedDipoles()->getDevicePointer(), &gkKernel->getInducedDipolesPolar()->getDevicePointer(), 
            &polarizability->getDevicePointer(), &inducedDipoleErrors->getDevicePointer(), &prevDipolesGk->getDevicePointer(),
1804
            &prevDipolesGkPolar->getDevicePointer(), &prevErrors->getDevicePointer(), &iteration, &falseValue, &diisMatrix->getDevicePointer()};
peastman's avatar
peastman committed
1805
1806
1807
1808
1809
        cu.executeKernel(recordDIISDipolesKernel, recordDIISDipolesGkArgs, cu.getNumThreadBlocks()*cu.ThreadBlockSize, cu.ThreadBlockSize, cu.ThreadBlockSize*elementSize*2);
    }
    void* recordDIISDipolesArgs[] = {&field->getDevicePointer(), &fieldPolar->getDevicePointer(), &npt, &inducedField->getDevicePointer(),
        &inducedFieldPolar->getDevicePointer(), &inducedDipole->getDevicePointer(), &inducedDipolePolar->getDevicePointer(),
        &polarizability->getDevicePointer(), &inducedDipoleErrors->getDevicePointer(), &prevDipoles->getDevicePointer(),
1810
        &prevDipolesPolar->getDevicePointer(), &prevErrors->getDevicePointer(), &iteration, &trueValue, &diisMatrix->getDevicePointer()};
peastman's avatar
peastman committed
1811
1812
1813
1814
    cu.executeKernel(recordDIISDipolesKernel, recordDIISDipolesArgs, cu.getNumThreadBlocks()*cu.ThreadBlockSize, cu.ThreadBlockSize, cu.ThreadBlockSize*elementSize*2);
    float2* errors = (float2*) cu.getPinnedBuffer();
    inducedDipoleErrors->download(errors, false);
    
peastman's avatar
peastman committed
1815
    // Build the DIIS matrix.
peastman's avatar
peastman committed
1816
1817
    
    int numPrev = (iteration+1 < MaxPrevDIISDipoles ? iteration+1 : MaxPrevDIISDipoles);
1818
1819
1820
    void* buildMatrixArgs[] = {&prevErrors->getDevicePointer(), &iteration, &diisMatrix->getDevicePointer()};
    int threadBlocks = min(numPrev, cu.getNumThreadBlocks());
    cu.executeKernel(buildMatrixKernel, buildMatrixArgs, threadBlocks*128, 128, 128*elementSize);
1821
1822
1823
1824
1825
1826
    vector<float> matrixf;
    vector<double> matrix;
    if (cu.getUseDoublePrecision())
        diisMatrix->download(matrix);
    else
        diisMatrix->download(matrixf);
peastman's avatar
peastman committed
1827
1828
1829
1830
1831
1832
1833
1834
1835
1836
1837
1838
1839
1840
    
    // Determine whether the iteration has converged.
    
    double total1 = 0.0, total2 = 0.0;
    for (int j = 0; j < inducedDipoleErrors->getSize(); j++) {
        total1 += errors[j].x;
        total2 += errors[j].y;
    }
    if (48.033324*sqrt(max(total1, total2)/cu.getNumAtoms()) < inducedEpsilon)
        return true;

    // Compute the coefficients for selecting the new dipoles.

    float* coefficients = (float*) cu.getPinnedBuffer();
peastman's avatar
peastman committed
1841
1842
1843
1844
1845
    if (iteration == 0)
        coefficients[0] = 1;
    else {
        int rank = numPrev+1;
        Array2D<double> b(rank, rank);
1846
1847
1848
        b[0][0] = 0;
        for (int i = 1; i < rank; i++)
            b[i][0] = b[0][i] = -1;
1849
1850
1851
1852
1853
1854
1855
1856
1857
1858
        if (cu.getUseDoublePrecision()) {
            for (int i = 0; i < numPrev; i++)
                for (int j = 0; j < numPrev; j++)
                    b[i+1][j+1] = matrix[i*MaxPrevDIISDipoles+j];
        }
        else {
            for (int i = 0; i < numPrev; i++)
                for (int j = 0; j < numPrev; j++)
                    b[i+1][j+1] = matrixf[i*MaxPrevDIISDipoles+j];
        }
peastman's avatar
peastman committed
1859
1860
1861
1862
1863
1864
1865
1866
1867
1868
1869
1870
1871
1872
1873
1874
1875

        // Solve using SVD.  Since the right hand side is (-1, 0, 0, 0, ...), this is simpler than the general case.

        JAMA::SVD<double> svd(b);
        Array2D<double> u, v;
        svd.getU(u);
        svd.getV(v);
        Array1D<double> s;
        svd.getSingularValues(s);
        int effectiveRank = svd.rank();
        for (int i = 1; i < rank; i++) {
            double d = 0;
            for (int j = 0; j < effectiveRank; j++)
                d -= u[0][j]*v[i][j]/s[j];
            coefficients[i-1] = d;
        }
    }
peastman's avatar
peastman committed
1876
    diisCoefficients->upload(coefficients, false);
peastman's avatar
peastman committed
1877
1878
1879
1880
1881
1882
1883
1884
1885
1886
1887
    
    // Compute the dipoles.
    
    void* updateInducedFieldArgs[] = {&inducedDipole->getDevicePointer(), &inducedDipolePolar->getDevicePointer(),
        &prevDipoles->getDevicePointer(), &prevDipolesPolar->getDevicePointer(), &diisCoefficients->getDevicePointer(), &numPrev};
    cu.executeKernel(updateInducedFieldKernel, updateInducedFieldArgs, cu.getNumThreadBlocks()*cu.ThreadBlockSize);
    if (gkKernel != NULL) {
        void* updateInducedFieldGkArgs[] = {&gkKernel->getInducedDipoles()->getDevicePointer(), &gkKernel->getInducedDipolesPolar()->getDevicePointer(),
            &prevDipolesGk->getDevicePointer(), &prevDipolesGkPolar->getDevicePointer(), &diisCoefficients->getDevicePointer(), &numPrev};
        cu.executeKernel(updateInducedFieldKernel, updateInducedFieldGkArgs, cu.getNumThreadBlocks()*cu.ThreadBlockSize);
    }
peastman's avatar
peastman committed
1888
    return false;
peastman's avatar
peastman committed
1889
1890
}

1891
1892
1893
1894
1895
1896
1897
1898
1899
1900
1901
1902
1903
1904
1905
1906
1907
1908
1909
1910
1911
1912
1913
1914
1915
1916
void CudaCalcAmoebaMultipoleForceKernel::computeExtrapolatedDipoles(void** recipBoxVectorPointer) {
    // Start by storing the direct dipoles as PT0

    void* initArgs[] = {&inducedDipole->getDevicePointer(), &inducedDipolePolar->getDevicePointer(), &extrapolatedDipole->getDevicePointer(),
        &extrapolatedDipolePolar->getDevicePointer(), &inducedDipoleFieldGradient->getDevicePointer(), &inducedDipoleFieldGradientPolar->getDevicePointer()};
    cu.executeKernel(initExtrapolatedKernel, initArgs, extrapolatedDipole->getSize());

    // Recursively apply alpha.Tau to the µ_(n) components to generate µ_(n+1), and store the result

    CudaNonbondedUtilities& nb = cu.getNonbondedUtilities();
    for (int order = 1; order < maxExtrapolationOrder; ++order) {
        cu.clearBuffer(*inducedField);
        cu.clearBuffer(*inducedFieldPolar);
        computeInducedField(recipBoxVectorPointer);
        void* iterateArgs[] = {&order, &inducedDipole->getDevicePointer(), &inducedDipolePolar->getDevicePointer(), &extrapolatedDipole->getDevicePointer(),
            &extrapolatedDipolePolar->getDevicePointer(), &inducedDipoleFieldGradient->getDevicePointer(), &inducedDipoleFieldGradientPolar->getDevicePointer(),
            &inducedField->getDevicePointer(), &inducedFieldPolar->getDevicePointer(), &extrapolatedDipoleFieldGradient->getDevicePointer(), &extrapolatedDipoleFieldGradientPolar->getDevicePointer(),
            &polarizability->getDevicePointer()};
        cu.executeKernel(iterateExtrapolatedKernel, iterateArgs, extrapolatedDipole->getSize());
    }
    
    // Take a linear combination of the µ_(n) components to form the total dipole

    void* computeArgs[] = {&inducedDipole->getDevicePointer(), &inducedDipolePolar->getDevicePointer(), &extrapolatedDipole->getDevicePointer(),
            &extrapolatedDipolePolar->getDevicePointer()};
    cu.executeKernel(computeExtrapolatedKernel, computeArgs, extrapolatedDipole->getSize());
1917
    computeInducedField(recipBoxVectorPointer);
1918
1919
}

1920
1921
1922
1923
1924
1925
1926
1927
1928
1929
1930
1931
1932
1933
1934
1935
1936
1937
1938
1939
1940
1941
1942
1943
1944
1945
1946
1947
void CudaCalcAmoebaMultipoleForceKernel::ensureMultipolesValid(ContextImpl& context) {
    if (multipolesAreValid) {
        int numParticles = cu.getNumAtoms();
        if (cu.getUseDoublePrecision()) {
            vector<double4> pos1, pos2;
            cu.getPosq().download(pos1);
            lastPositions->download(pos2);
            for (int i = 0; i < numParticles; i++)
                if (pos1[i].x != pos2[i].x || pos1[i].y != pos2[i].y || pos1[i].z != pos2[i].z) {
                    multipolesAreValid = false;
                    break;
                }
        }
        else {
            vector<float4> pos1, pos2;
            cu.getPosq().download(pos1);
            lastPositions->download(pos2);
            for (int i = 0; i < numParticles; i++)
                if (pos1[i].x != pos2[i].x || pos1[i].y != pos2[i].y || pos1[i].z != pos2[i].z) {
                    multipolesAreValid = false;
                    break;
                }
        }
    }
    if (!multipolesAreValid)
        context.calcForcesAndEnergy(false, false, -1);
}

1948
1949
1950
1951
void CudaCalcAmoebaMultipoleForceKernel::getInducedDipoles(ContextImpl& context, vector<Vec3>& dipoles) {
    ensureMultipolesValid(context);
    int numParticles = cu.getNumAtoms();
    dipoles.resize(numParticles);
peastman's avatar
peastman committed
1952
    const vector<int>& order = cu.getAtomIndex();
1953
1954
1955
1956
    if (cu.getUseDoublePrecision()) {
        vector<double> d;
        inducedDipole->download(d);
        for (int i = 0; i < numParticles; i++)
peastman's avatar
peastman committed
1957
            dipoles[order[i]] = Vec3(d[3*i], d[3*i+1], d[3*i+2]);
1958
1959
1960
1961
1962
    }
    else {
        vector<float> d;
        inducedDipole->download(d);
        for (int i = 0; i < numParticles; i++)
peastman's avatar
peastman committed
1963
            dipoles[order[i]] = Vec3(d[3*i], d[3*i+1], d[3*i+2]);
1964
1965
1966
    }
}

1967
void CudaCalcAmoebaMultipoleForceKernel::getElectrostaticPotential(ContextImpl& context, const vector<Vec3>& inputGrid, vector<double>& outputElectrostaticPotential) {
1968
    ensureMultipolesValid(context);
1969
1970
1971
1972
1973
1974
1975
1976
1977
1978
1979
1980
1981
1982
1983
1984
1985
1986
1987
1988
1989
1990
1991
1992
    int numPoints = inputGrid.size();
    int elementSize = (cu.getUseDoublePrecision() ? sizeof(double) : sizeof(float));
    CudaArray points(cu, numPoints, 4*elementSize, "points");
    CudaArray potential(cu, numPoints, elementSize, "potential");
    
    // Copy the grid points to the GPU.
    
    if (cu.getUseDoublePrecision()) {
        vector<double4> p(numPoints);
        for (int i = 0; i < numPoints; i++)
            p[i] = make_double4(inputGrid[i][0], inputGrid[i][1], inputGrid[i][2], 0);
        points.upload(p);
    }
    else {
        vector<float4> p(numPoints);
        for (int i = 0; i < numPoints; i++)
            p[i] = make_float4((float) inputGrid[i][0], (float) inputGrid[i][1], (float) inputGrid[i][2], 0);
        points.upload(p);
    }
    
    // Compute the potential.
    
    void* computePotentialArgs[] = {&cu.getPosq().getDevicePointer(), &labFrameDipoles->getDevicePointer(),
        &labFrameQuadrupoles->getDevicePointer(), &inducedDipole->getDevicePointer(), &points.getDevicePointer(),
1993
1994
        &potential.getDevicePointer(), &numPoints, cu.getPeriodicBoxSizePointer(), cu.getInvPeriodicBoxSizePointer(),
        cu.getPeriodicBoxVecXPointer(), cu.getPeriodicBoxVecYPointer(), cu.getPeriodicBoxVecZPointer()};
1995
1996
1997
1998
1999
2000
2001
2002
2003
2004
2005
2006
2007
    int blockSize = 128;
    cu.executeKernel(computePotentialKernel, computePotentialArgs, numPoints, blockSize, blockSize*15*elementSize);
    outputElectrostaticPotential.resize(numPoints);
    if (cu.getUseDoublePrecision())
        potential.download(outputElectrostaticPotential);
    else {
        vector<float> p(numPoints);
        potential.download(p);
        for (int i = 0; i < numPoints; i++)
            outputElectrostaticPotential[i] = p[i];
    }
}

2008
template <class T, class T4, class M4>
Lee-Ping Wang's avatar
Lee-Ping Wang committed
2009
void CudaCalcAmoebaMultipoleForceKernel::computeSystemMultipoleMoments(ContextImpl& context, vector<double>& outputMultipoleMoments) {
2010
2011
    // Compute the local coordinates relative to the center of mass.
    int numAtoms = cu.getNumAtoms();
2012
2013
    vector<T4> posq;
    vector<M4> velm;
2014
2015
2016
2017
2018
2019
2020
2021
2022
2023
2024
2025
2026
2027
2028
2029
2030
2031
2032
2033
2034
2035
2036
2037
2038
2039
2040
2041
2042
2043
2044
2045
2046
2047
2048
2049
2050
2051
2052
2053
    cu.getPosq().download(posq);
    cu.getVelm().download(velm);
    double totalMass = 0.0;
    Vec3 centerOfMass(0, 0, 0);
    for (int i = 0; i < numAtoms; i++) {
        double mass = (velm[i].w > 0 ? 1.0/velm[i].w : 0.0);
        totalMass += mass;
        centerOfMass[0] += mass*posq[i].x;
        centerOfMass[1] += mass*posq[i].y;
        centerOfMass[2] += mass*posq[i].z;
    }
    if (totalMass > 0.0) {
        centerOfMass[0] /= totalMass;
        centerOfMass[1] /= totalMass;
        centerOfMass[2] /= totalMass;
    }
    vector<double4> posqLocal(numAtoms);
    for (int i = 0; i < numAtoms; i++) {
        posqLocal[i].x = posq[i].x - centerOfMass[0];
        posqLocal[i].y = posq[i].y - centerOfMass[1];
        posqLocal[i].z = posq[i].z - centerOfMass[2];
        posqLocal[i].w = posq[i].w;
    }

    // Compute the multipole moments.
    
    double totalCharge = 0.0;
    double xdpl = 0.0;
    double ydpl = 0.0;
    double zdpl = 0.0;
    double xxqdp = 0.0;
    double xyqdp = 0.0;
    double xzqdp = 0.0;
    double yxqdp = 0.0;
    double yyqdp = 0.0;
    double yzqdp = 0.0;
    double zxqdp = 0.0;
    double zyqdp = 0.0;
    double zzqdp = 0.0;
    vector<T> labDipoleVec, inducedDipoleVec, quadrupoleVec;
Lee-Ping Wang's avatar
Lee-Ping Wang committed
2054
2055
2056
    labFrameDipoles->download(labDipoleVec);
    inducedDipole->download(inducedDipoleVec);
    labFrameQuadrupoles->download(quadrupoleVec);
2057
2058
    for (int i = 0; i < numAtoms; i++) {
        totalCharge += posqLocal[i].w;
Lee-Ping Wang's avatar
Lee-Ping Wang committed
2059
2060
2061
2062
2063
2064
2065
2066
2067
2068
2069
2070
2071
2072
2073
        double netDipoleX = (labDipoleVec[3*i]    + inducedDipoleVec[3*i]);
        double netDipoleY = (labDipoleVec[3*i+1]  + inducedDipoleVec[3*i+1]);
        double netDipoleZ = (labDipoleVec[3*i+2]  + inducedDipoleVec[3*i+2]);
        xdpl += posqLocal[i].x*posqLocal[i].w + netDipoleX;
        ydpl += posqLocal[i].y*posqLocal[i].w + netDipoleY;
        zdpl += posqLocal[i].z*posqLocal[i].w + netDipoleZ;
        xxqdp += posqLocal[i].x*posqLocal[i].x*posqLocal[i].w + 2*posqLocal[i].x*netDipoleX;
        xyqdp += posqLocal[i].x*posqLocal[i].y*posqLocal[i].w + posqLocal[i].x*netDipoleY + posqLocal[i].y*netDipoleX;
        xzqdp += posqLocal[i].x*posqLocal[i].z*posqLocal[i].w + posqLocal[i].x*netDipoleZ + posqLocal[i].z*netDipoleX;
        yxqdp += posqLocal[i].y*posqLocal[i].x*posqLocal[i].w + posqLocal[i].y*netDipoleX + posqLocal[i].x*netDipoleY;
        yyqdp += posqLocal[i].y*posqLocal[i].y*posqLocal[i].w + 2*posqLocal[i].y*netDipoleY;
        yzqdp += posqLocal[i].y*posqLocal[i].z*posqLocal[i].w + posqLocal[i].y*netDipoleZ + posqLocal[i].z*netDipoleY;
        zxqdp += posqLocal[i].z*posqLocal[i].x*posqLocal[i].w + posqLocal[i].z*netDipoleX + posqLocal[i].x*netDipoleZ;
        zyqdp += posqLocal[i].z*posqLocal[i].y*posqLocal[i].w + posqLocal[i].z*netDipoleY + posqLocal[i].y*netDipoleZ;
        zzqdp += posqLocal[i].z*posqLocal[i].z*posqLocal[i].w + 2*posqLocal[i].z*netDipoleZ;
2074
2075
2076
2077
    }

    // Convert the quadrupole from traced to traceless form.
 
Lee-Ping Wang's avatar
Lee-Ping Wang committed
2078
2079
2080
2081
2082
2083
2084
2085
2086
2087
    double qave = (xxqdp + yyqdp + zzqdp)/3;
    xxqdp = 1.5*(xxqdp-qave);
    xyqdp = 1.5*xyqdp;
    xzqdp = 1.5*xzqdp;
    yxqdp = 1.5*yxqdp;
    yyqdp = 1.5*(yyqdp-qave);
    yzqdp = 1.5*yzqdp;
    zxqdp = 1.5*zxqdp;
    zyqdp = 1.5*zyqdp;
    zzqdp = 1.5*(zzqdp-qave);
2088
2089
2090

    // Add the traceless atomic quadrupoles to the total quadrupole moment.

Lee-Ping Wang's avatar
Lee-Ping Wang committed
2091
2092
2093
2094
2095
2096
2097
2098
2099
2100
2101
2102
2103
    for (int i = 0; i < numAtoms; i++) {
        xxqdp = xxqdp + 3*quadrupoleVec[5*i];
        xyqdp = xyqdp + 3*quadrupoleVec[5*i+1];
        xzqdp = xzqdp + 3*quadrupoleVec[5*i+2];
        yxqdp = yxqdp + 3*quadrupoleVec[5*i+1];
        yyqdp = yyqdp + 3*quadrupoleVec[5*i+3];
        yzqdp = yzqdp + 3*quadrupoleVec[5*i+4];
        zxqdp = zxqdp + 3*quadrupoleVec[5*i+2];
        zyqdp = zyqdp + 3*quadrupoleVec[5*i+4];
        zzqdp = zzqdp + -3*(quadrupoleVec[5*i]+quadrupoleVec[5*i+3]);
    }
 
    double debye = 4.80321;
2104
2105
    outputMultipoleMoments.resize(13);
    outputMultipoleMoments[0] = totalCharge;
2106
2107
2108
2109
2110
2111
2112
2113
2114
2115
2116
2117
    outputMultipoleMoments[1] = 10.0*xdpl*debye;
    outputMultipoleMoments[2] = 10.0*ydpl*debye;
    outputMultipoleMoments[3] = 10.0*zdpl*debye;
    outputMultipoleMoments[4] = 100.0*xxqdp*debye;
    outputMultipoleMoments[5] = 100.0*xyqdp*debye;
    outputMultipoleMoments[6] = 100.0*xzqdp*debye;
    outputMultipoleMoments[7] = 100.0*yxqdp*debye;
    outputMultipoleMoments[8] = 100.0*yyqdp*debye;
    outputMultipoleMoments[9] = 100.0*yzqdp*debye;
    outputMultipoleMoments[10] = 100.0*zxqdp*debye;
    outputMultipoleMoments[11] = 100.0*zyqdp*debye;
    outputMultipoleMoments[12] = 100.0*zzqdp*debye;
2118
2119
}

Lee-Ping Wang's avatar
Lee-Ping Wang committed
2120
void CudaCalcAmoebaMultipoleForceKernel::getSystemMultipoleMoments(ContextImpl& context, vector<double>& outputMultipoleMoments) {
2121
    ensureMultipolesValid(context);
2122
    if (cu.getUseDoublePrecision())
Lee-Ping Wang's avatar
Lee-Ping Wang committed
2123
        computeSystemMultipoleMoments<double, double4, double4>(context, outputMultipoleMoments);
2124
    else if (cu.getUseMixedPrecision())
Lee-Ping Wang's avatar
Lee-Ping Wang committed
2125
        computeSystemMultipoleMoments<float, float4, double4>(context, outputMultipoleMoments);
2126
    else
Lee-Ping Wang's avatar
Lee-Ping Wang committed
2127
        computeSystemMultipoleMoments<float, float4, float4>(context, outputMultipoleMoments);
2128
2129
}

2130
2131
2132
2133
2134
2135
2136
2137
2138
2139
2140
2141
2142
2143
2144
2145
2146
2147
2148
2149
2150
2151
2152
2153
2154
2155
2156
2157
2158
2159
2160
2161
2162
2163
2164
2165
2166
void CudaCalcAmoebaMultipoleForceKernel::copyParametersToContext(ContextImpl& context, const AmoebaMultipoleForce& force) {
    // Make sure the new parameters are acceptable.
    
    cu.setAsCurrent();
    if (force.getNumMultipoles() != cu.getNumAtoms())
        throw OpenMMException("updateParametersInContext: The number of multipoles has changed");
    
    // Record the per-multipole parameters.
    
    cu.getPosq().download(cu.getPinnedBuffer());
    float4* posqf = (float4*) cu.getPinnedBuffer();
    double4* posqd = (double4*) cu.getPinnedBuffer();
    vector<float2> dampingAndTholeVec;
    vector<float> polarizabilityVec;
    vector<float> molecularDipolesVec;
    vector<float> molecularQuadrupolesVec;
    vector<int4> multipoleParticlesVec;
    for (int i = 0; i < force.getNumMultipoles(); i++) {
        double charge, thole, damping, polarity;
        int axisType, atomX, atomY, atomZ;
        vector<double> dipole, quadrupole;
        force.getMultipoleParameters(i, charge, dipole, quadrupole, axisType, atomZ, atomX, atomY, thole, damping, polarity);
        if (cu.getUseDoublePrecision())
            posqd[i].w = charge;
        else
            posqf[i].w = (float) charge;
        dampingAndTholeVec.push_back(make_float2((float) damping, (float) thole));
        polarizabilityVec.push_back((float) polarity);
        multipoleParticlesVec.push_back(make_int4(atomX, atomY, atomZ, axisType));
        for (int j = 0; j < 3; j++)
            molecularDipolesVec.push_back((float) dipole[j]);
        molecularQuadrupolesVec.push_back((float) quadrupole[0]);
        molecularQuadrupolesVec.push_back((float) quadrupole[1]);
        molecularQuadrupolesVec.push_back((float) quadrupole[2]);
        molecularQuadrupolesVec.push_back((float) quadrupole[4]);
        molecularQuadrupolesVec.push_back((float) quadrupole[5]);
    }
2167
2168
2169
2170
2171
    if (!hasQuadrupoles) {
        for (int i = 0; i < (int) molecularQuadrupolesVec.size(); i++)
            if (molecularQuadrupolesVec[i] != 0.0)
                throw OpenMMException("updateParametersInContext: Cannot set a non-zero quadrupole moment, because quadrupoles were excluded from the kernel");
    }
2172
2173
2174
2175
2176
2177
2178
2179
2180
2181
2182
2183
2184
2185
2186
2187
    for (int i = force.getNumMultipoles(); i < cu.getPaddedNumAtoms(); i++) {
        dampingAndTholeVec.push_back(make_float2(0, 0));
        polarizabilityVec.push_back(0);
        multipoleParticlesVec.push_back(make_int4(0, 0, 0, 0));
        for (int j = 0; j < 3; j++)
            molecularDipolesVec.push_back(0);
        for (int j = 0; j < 5; j++)
            molecularQuadrupolesVec.push_back(0);
    }
    dampingAndThole->upload(dampingAndTholeVec);
    polarizability->upload(polarizabilityVec);
    multipoleParticles->upload(multipoleParticlesVec);
    molecularDipoles->upload(molecularDipolesVec);
    molecularQuadrupoles->upload(molecularQuadrupolesVec);
    cu.getPosq().upload(cu.getPinnedBuffer());
    cu.invalidateMolecules();
2188
    multipolesAreValid = false;
2189
2190
}

2191
2192
2193
2194
2195
2196
2197
2198
2199
void CudaCalcAmoebaMultipoleForceKernel::getPMEParameters(double& alpha, int& nx, int& ny, int& nz) const {
    if (!usePME)
        throw OpenMMException("getPMEParametersInContext: This Context is not using PME");
    alpha = this->alpha;
    nx = gridSizeX;
    ny = gridSizeY;
    nz = gridSizeZ;
}

2200
2201
2202
2203
2204
2205
2206
2207
2208
2209
2210
2211
2212
2213
2214
2215
2216
2217
/* -------------------------------------------------------------------------- *
 *                       AmoebaGeneralizedKirkwood                            *
 * -------------------------------------------------------------------------- */

class CudaCalcAmoebaGeneralizedKirkwoodForceKernel::ForceInfo : public CudaForceInfo {
public:
    ForceInfo(const AmoebaGeneralizedKirkwoodForce& force) : force(force) {
    }
    bool areParticlesIdentical(int particle1, int particle2) {
        double charge1, charge2, radius1, radius2, scale1, scale2;
        force.getParticleParameters(particle1, charge1, radius1, scale1);
        force.getParticleParameters(particle2, charge2, radius2, scale2);
        return (charge1 == charge2 && radius1 == radius2 && scale1 == scale2);
    }
private:
    const AmoebaGeneralizedKirkwoodForce& force;
};

2218
CudaCalcAmoebaGeneralizedKirkwoodForceKernel::CudaCalcAmoebaGeneralizedKirkwoodForceKernel(std::string name, const Platform& platform, CudaContext& cu, const System& system) : 
2219
           CalcAmoebaGeneralizedKirkwoodForceKernel(name, platform), cu(cu), system(system), hasInitializedKernels(false), params(NULL), bornRadii(NULL), field(NULL),
2220
           inducedField(NULL), inducedFieldPolar(NULL), inducedDipoleS(NULL), inducedDipolePolarS(NULL), bornSum(NULL), bornForce(NULL) {
2221
2222
2223
2224
2225
2226
2227
2228
2229
2230
}

CudaCalcAmoebaGeneralizedKirkwoodForceKernel::~CudaCalcAmoebaGeneralizedKirkwoodForceKernel() {
    cu.setAsCurrent();
    if (params != NULL)
        delete params;
    if (bornRadii != NULL)
        delete bornRadii;
    if (field != NULL)
        delete field;
2231
2232
2233
2234
    if (inducedField != NULL)
        delete inducedField;
    if (inducedFieldPolar != NULL)
        delete inducedFieldPolar;
2235
2236
2237
2238
2239
2240
2241
2242
2243
2244
2245
2246
2247
2248
2249
2250
2251
2252
2253
2254
2255
2256
2257
2258
2259
2260
2261
2262
2263
    if (inducedDipoleS != NULL)
        delete inducedDipoleS;
    if (inducedDipolePolarS != NULL)
        delete inducedDipolePolarS;
    if (bornSum != NULL)
        delete bornSum;
    if (bornForce != NULL)
        delete bornForce;
}

void CudaCalcAmoebaGeneralizedKirkwoodForceKernel::initialize(const System& system, const AmoebaGeneralizedKirkwoodForce& force) {
    cu.setAsCurrent();
    if (cu.getPlatformData().contexts.size() > 1)
        throw OpenMMException("AmoebaGeneralizedKirkwoodForce does not support using multiple CUDA devices");
    const AmoebaMultipoleForce* multipoles = NULL;
    for (int i = 0; i < system.getNumForces() && multipoles == NULL; i++)
        multipoles = dynamic_cast<const AmoebaMultipoleForce*>(&system.getForce(i));
    if (multipoles == NULL)
        throw OpenMMException("AmoebaGeneralizedKirkwoodForce requires the System to also contain an AmoebaMultipoleForce");
    CudaNonbondedUtilities& nb = cu.getNonbondedUtilities();
    int paddedNumAtoms = cu.getPaddedNumAtoms();
    int elementSize = (cu.getUseDoublePrecision() ? sizeof(double) : sizeof(float));
    params = CudaArray::create<float2>(cu, paddedNumAtoms, "amoebaGkParams");
    bornRadii = new CudaArray(cu, paddedNumAtoms, elementSize, "bornRadii");
    field = new CudaArray(cu, 3*paddedNumAtoms, sizeof(long long), "gkField");
    bornSum = CudaArray::create<long long>(cu, paddedNumAtoms, "bornSum");
    bornForce = CudaArray::create<long long>(cu, paddedNumAtoms, "bornForce");
    inducedDipoleS = new CudaArray(cu, 3*paddedNumAtoms, elementSize, "inducedDipoleS");
    inducedDipolePolarS = new CudaArray(cu, 3*paddedNumAtoms, elementSize, "inducedDipolePolarS");
2264
2265
2266
2267
    if (multipoles->getPolarizationType() == AmoebaMultipoleForce::Mutual) {
        inducedField = new CudaArray(cu, 3*paddedNumAtoms, sizeof(long long), "gkInducedField");
        inducedFieldPolar = new CudaArray(cu, 3*paddedNumAtoms, sizeof(long long), "gkInducedFieldPolar");
    }
2268
2269
2270
2271
2272
2273
2274
2275
2276
2277
2278
2279
2280
2281
2282
2283
2284
2285
2286
2287
    cu.addAutoclearBuffer(*field);
    cu.addAutoclearBuffer(*bornSum);
    cu.addAutoclearBuffer(*bornForce);
    vector<float2> paramsVector(paddedNumAtoms);
    for (int i = 0; i < force.getNumParticles(); i++) {
        double charge, radius, scalingFactor;
        force.getParticleParameters(i, charge, radius, scalingFactor);
        paramsVector[i] = make_float2((float) radius, (float) (scalingFactor*radius));
        
        // Make sure the charge matches the one specified by the AmoebaMultipoleForce.
        
        double charge2, thole, damping, polarity;
        int axisType, atomX, atomY, atomZ;
        vector<double> dipole, quadrupole;
        multipoles->getMultipoleParameters(i, charge2, dipole, quadrupole, axisType, atomZ, atomX, atomY, thole, damping, polarity);
        if (charge != charge2)
            throw OpenMMException("AmoebaGeneralizedKirkwoodForce and AmoebaMultipoleForce must specify the same charge for every atom");
    }
    params->upload(paramsVector);
    
2288
2289
2290
2291
2292
2293
2294
2295
2296
2297
2298
2299
    // Select the number of threads for each kernel.
    
    double computeBornSumThreadMemory = 4*elementSize+3*sizeof(float);
    double gkForceThreadMemory = 24*elementSize;
    double chainRuleThreadMemory = 10*elementSize;
    double ediffThreadMemory = 28*elementSize+2*sizeof(float)+3*sizeof(int)/(double) cu.TileSize;
    int maxThreads = cu.getNonbondedUtilities().getForceThreadBlockSize();
    computeBornSumThreads = min(maxThreads, cu.computeThreadBlockSize(computeBornSumThreadMemory));
    gkForceThreads = min(maxThreads, cu.computeThreadBlockSize(gkForceThreadMemory));
    chainRuleThreads = min(maxThreads, cu.computeThreadBlockSize(chainRuleThreadMemory));
    ediffThreads = min(maxThreads, cu.computeThreadBlockSize(ediffThreadMemory));
    
2300
    // Set preprocessor macros we will use when we create the kernels.
2301
2302
2303
    
    defines["NUM_ATOMS"] = cu.intToString(cu.getNumAtoms());
    defines["PADDED_NUM_ATOMS"] = cu.intToString(paddedNumAtoms);
2304
2305
2306
2307
    defines["BORN_SUM_THREAD_BLOCK_SIZE"] = cu.intToString(computeBornSumThreads);
    defines["GK_FORCE_THREAD_BLOCK_SIZE"] = cu.intToString(gkForceThreads);
    defines["CHAIN_RULE_THREAD_BLOCK_SIZE"] = cu.intToString(chainRuleThreads);
    defines["EDIFF_THREAD_BLOCK_SIZE"] = cu.intToString(ediffThreads);
2308
2309
2310
2311
2312
2313
    defines["NUM_BLOCKS"] = cu.intToString(cu.getNumAtomBlocks());
    defines["GK_C"] = cu.doubleToString(2.455);
    double solventDielectric = force.getSolventDielectric();
    defines["GK_FC"] = cu.doubleToString(1*(1-solventDielectric)/(0+1*solventDielectric));
    defines["GK_FD"] = cu.doubleToString(2*(1-solventDielectric)/(1+2*solventDielectric));
    defines["GK_FQ"] = cu.doubleToString(3*(1-solventDielectric)/(2+3*solventDielectric));
2314
    defines["EPSILON_FACTOR"] = cu.doubleToString(138.9354558456);
Peter Eastman's avatar
Peter Eastman committed
2315
    defines["M_PI"] = cu.doubleToString(M_PI);
2316
    defines["ENERGY_SCALE_FACTOR"] = cu.doubleToString(138.9354558456/force.getSoluteDielectric());
2317
2318
    if (multipoles->getPolarizationType() == AmoebaMultipoleForce::Direct)
        defines["DIRECT_POLARIZATION"] = "";
2319
2320
2321
2322
2323
2324
    includeSurfaceArea = force.getIncludeCavityTerm();
    if (includeSurfaceArea) {
        defines["SURFACE_AREA_FACTOR"] = cu.doubleToString(force.getSurfaceAreaFactor());
        defines["PROBE_RADIUS"] = cu.doubleToString(force.getProbeRadius());
        defines["DIELECTRIC_OFFSET"] = cu.doubleToString(0.009);
    }
2325
2326
2327
2328
2329
2330
2331
2332
2333
2334
    cu.addForce(new ForceInfo(force));
}

double CudaCalcAmoebaGeneralizedKirkwoodForceKernel::execute(ContextImpl& context, bool includeForces, bool includeEnergy) {
    // Since GK is so tightly entwined with the electrostatics, this method does nothing, and the force calculation
    // is driven by AmoebaMultipoleForce.
    return 0.0;
}

void CudaCalcAmoebaGeneralizedKirkwoodForceKernel::computeBornRadii() {
2335
2336
2337
2338
2339
2340
2341
2342
2343
2344
2345
2346
2347
2348
2349
2350
2351
2352
2353
2354
2355
2356
2357
2358
2359
2360
2361
2362
2363
2364
2365
2366
2367
2368
2369
2370
2371
2372
2373
2374
2375
2376
2377
2378
2379
2380
2381
2382
2383
    if (!hasInitializedKernels) {
        hasInitializedKernels = true;
        
        // Create the kernels.
        
        int numExclusionTiles = cu.getNonbondedUtilities().getExclusionTiles().getSize();
        defines["NUM_TILES_WITH_EXCLUSIONS"] = cu.intToString(numExclusionTiles);
        int numContexts = cu.getPlatformData().contexts.size();
        int startExclusionIndex = cu.getContextIndex()*numExclusionTiles/numContexts;
        int endExclusionIndex = (cu.getContextIndex()+1)*numExclusionTiles/numContexts;
        defines["FIRST_EXCLUSION_TILE"] = cu.intToString(startExclusionIndex);
        defines["LAST_EXCLUSION_TILE"] = cu.intToString(endExclusionIndex);
        stringstream forceSource;
        forceSource << CudaKernelSources::vectorOps;
        forceSource << CudaAmoebaKernelSources::amoebaGk;
        forceSource << "#define F1\n";
        forceSource << CudaAmoebaKernelSources::gkPairForce1;
        forceSource << CudaAmoebaKernelSources::gkPairForce2;
        forceSource << CudaAmoebaKernelSources::gkEDiffPairForce;
        forceSource << "#undef F1\n";
        forceSource << "#define F2\n";
        forceSource << CudaAmoebaKernelSources::gkPairForce1;
        forceSource << CudaAmoebaKernelSources::gkPairForce2;
        forceSource << "#undef F2\n";
        forceSource << "#define T1\n";
        forceSource << CudaAmoebaKernelSources::gkPairForce1;
        forceSource << CudaAmoebaKernelSources::gkPairForce2;
        forceSource << CudaAmoebaKernelSources::gkEDiffPairForce;
        forceSource << "#undef T1\n";
        forceSource << "#define T2\n";
        forceSource << CudaAmoebaKernelSources::gkPairForce1;
        forceSource << CudaAmoebaKernelSources::gkPairForce2;
        forceSource << "#undef T2\n";
        forceSource << "#define T3\n";
        forceSource << CudaAmoebaKernelSources::gkEDiffPairForce;
        forceSource << "#undef T3\n";
        forceSource << "#define B1\n";
        forceSource << "#define B2\n";
        forceSource << CudaAmoebaKernelSources::gkPairForce1;
        forceSource << CudaAmoebaKernelSources::gkPairForce2;
        CUmodule module = cu.createModule(forceSource.str(), defines);
        computeBornSumKernel = cu.getKernel(module, "computeBornSum");
        reduceBornSumKernel = cu.getKernel(module, "reduceBornSum");
        gkForceKernel = cu.getKernel(module, "computeGKForces");
        chainRuleKernel = cu.getKernel(module, "computeChainRuleForce");
        ediffKernel = cu.getKernel(module, "computeEDiffForce");
        if (includeSurfaceArea)
            surfaceAreaKernel = cu.getKernel(module, "computeSurfaceAreaForce");
    }
2384
2385
2386
2387
2388
    CudaNonbondedUtilities& nb = cu.getNonbondedUtilities();
    int numTiles = nb.getNumTiles();
    int numForceThreadBlocks = nb.getNumForceThreadBlocks();
    void* computeBornSumArgs[] = {&bornSum->getDevicePointer(), &cu.getPosq().getDevicePointer(),
        &params->getDevicePointer(), &numTiles};
2389
    cu.executeKernel(computeBornSumKernel, computeBornSumArgs, numForceThreadBlocks*computeBornSumThreads, computeBornSumThreads);
2390
2391
2392
2393
2394
2395
2396
2397
2398
2399
    void* reduceBornSumArgs[] = {&bornSum->getDevicePointer(), &params->getDevicePointer(), &bornRadii->getDevicePointer()};
    cu.executeKernel(reduceBornSumKernel, reduceBornSumArgs, cu.getNumAtoms());
}

void CudaCalcAmoebaGeneralizedKirkwoodForceKernel::finishComputation(CudaArray& torque, CudaArray& labFrameDipoles, CudaArray& labFrameQuadrupoles,
            CudaArray& inducedDipole, CudaArray& inducedDipolePolar, CudaArray& dampingAndThole, CudaArray& covalentFlags, CudaArray& polarizationGroupFlags) {
    CudaNonbondedUtilities& nb = cu.getNonbondedUtilities();
    int startTileIndex = nb.getStartTileIndex();
    int numTileIndices = nb.getNumTiles();
    int numForceThreadBlocks = nb.getNumForceThreadBlocks();
2400
2401
2402
    
    // Compute the GK force.
    
2403
2404
    void* gkForceArgs[] = {&cu.getForce().getDevicePointer(), &torque.getDevicePointer(), &cu.getEnergyBuffer().getDevicePointer(),
        &cu.getPosq().getDevicePointer(), &startTileIndex, &numTileIndices, &labFrameDipoles.getDevicePointer(),
2405
        &labFrameQuadrupoles.getDevicePointer(), &inducedDipoleS->getDevicePointer(), &inducedDipolePolarS->getDevicePointer(),
2406
        &bornRadii->getDevicePointer(), &bornForce->getDevicePointer()};
2407
    cu.executeKernel(gkForceKernel, gkForceArgs, numForceThreadBlocks*gkForceThreads, gkForceThreads);
2408

2409
2410
2411
2412
2413
2414
2415
2416
    // Compute the surface area force.
    
    if (includeSurfaceArea) {
        void* surfaceAreaArgs[] = {&bornForce->getDevicePointer(), &cu.getEnergyBuffer().getDevicePointer(), &params->getDevicePointer(), &bornRadii->getDevicePointer()};
        cu.executeKernel(surfaceAreaKernel, surfaceAreaArgs, cu.getNumAtoms());
    }
    
    // Apply the remaining terms.
2417
2418
2419
    
    void* chainRuleArgs[] = {&cu.getForce().getDevicePointer(), &cu.getPosq().getDevicePointer(), &startTileIndex, &numTileIndices,
        &params->getDevicePointer(), &bornRadii->getDevicePointer(), &bornForce->getDevicePointer()};
2420
    cu.executeKernel(chainRuleKernel, chainRuleArgs, numForceThreadBlocks*chainRuleThreads, chainRuleThreads);    
2421
    void* ediffArgs[] = {&cu.getForce().getDevicePointer(), &torque.getDevicePointer(), &cu.getEnergyBuffer().getDevicePointer(),
2422
2423
        &cu.getPosq().getDevicePointer(), &covalentFlags.getDevicePointer(), &polarizationGroupFlags.getDevicePointer(),
        &nb.getExclusionTiles().getDevicePointer(), &startTileIndex, &numTileIndices,
2424
2425
2426
        &labFrameDipoles.getDevicePointer(), &labFrameQuadrupoles.getDevicePointer(), &inducedDipole.getDevicePointer(),
        &inducedDipolePolar.getDevicePointer(), &inducedDipoleS->getDevicePointer(), &inducedDipolePolarS->getDevicePointer(),
        &dampingAndThole.getDevicePointer()};
2427
    cu.executeKernel(ediffKernel, ediffArgs, numForceThreadBlocks*ediffThreads, ediffThreads);
2428
}
2429

2430
2431
2432
2433
2434
2435
2436
2437
2438
2439
2440
2441
2442
2443
2444
2445
2446
2447
2448
void CudaCalcAmoebaGeneralizedKirkwoodForceKernel::copyParametersToContext(ContextImpl& context, const AmoebaGeneralizedKirkwoodForce& force) {
    // Make sure the new parameters are acceptable.
    
    cu.setAsCurrent();
    if (force.getNumParticles() != cu.getNumAtoms())
        throw OpenMMException("updateParametersInContext: The number of particles has changed");
    
    // Record the per-particle parameters.
    
    vector<float2> paramsVector(cu.getPaddedNumAtoms());
    for (int i = 0; i < force.getNumParticles(); i++) {
        double charge, radius, scalingFactor;
        force.getParticleParameters(i, charge, radius, scalingFactor);
        paramsVector[i] = make_float2((float) radius, (float) (scalingFactor*radius));
    }
    params->upload(paramsVector);
    cu.invalidateMolecules();
}

2449
2450
2451
2452
2453
2454
2455
2456
2457
/* -------------------------------------------------------------------------- *
 *                           AmoebaVdw                                        *
 * -------------------------------------------------------------------------- */

class CudaCalcAmoebaVdwForceKernel::ForceInfo : public CudaForceInfo {
public:
    ForceInfo(const AmoebaVdwForce& force) : force(force) {
    }
    bool areParticlesIdentical(int particle1, int particle2) {
2458
        int iv1, iv2;
2459
        double sigma1, sigma2, epsilon1, epsilon2, reduction1, reduction2;
2460
2461
2462
        force.getParticleParameters(particle1, iv1, sigma1, epsilon1, reduction1);
        force.getParticleParameters(particle2, iv2, sigma2, epsilon2, reduction2);
        return (sigma1 == sigma2 && epsilon1 == epsilon2 && reduction1 == reduction2);
2463
2464
2465
2466
2467
    }
private:
    const AmoebaVdwForce& force;
};

2468
CudaCalcAmoebaVdwForceKernel::CudaCalcAmoebaVdwForceKernel(std::string name, const Platform& platform, CudaContext& cu, const System& system) :
2469
2470
2471
2472
2473
2474
2475
2476
2477
2478
2479
2480
2481
2482
2483
2484
2485
2486
2487
2488
2489
2490
2491
2492
2493
2494
2495
2496
2497
2498
2499
2500
2501
2502
2503
        CalcAmoebaVdwForceKernel(name, platform), cu(cu), system(system), hasInitializedNonbonded(false), sigmaEpsilon(NULL),
        bondReductionAtoms(NULL), bondReductionFactors(NULL), tempPosq(NULL), tempForces(NULL), nonbonded(NULL) {
}

CudaCalcAmoebaVdwForceKernel::~CudaCalcAmoebaVdwForceKernel() {
    cu.setAsCurrent();
    if (sigmaEpsilon != NULL)
        delete sigmaEpsilon;
    if (bondReductionAtoms != NULL)
        delete bondReductionAtoms;
    if (bondReductionFactors != NULL)
        delete bondReductionFactors;
    if (tempPosq != NULL)
        delete tempPosq;
    if (tempForces != NULL)
        delete tempForces;
    if (nonbonded != NULL)
        delete nonbonded;
}

void CudaCalcAmoebaVdwForceKernel::initialize(const System& system, const AmoebaVdwForce& force) {
    cu.setAsCurrent();
    sigmaEpsilon = CudaArray::create<float2>(cu, cu.getPaddedNumAtoms(), "sigmaEpsilon");
    bondReductionAtoms = CudaArray::create<int>(cu, cu.getPaddedNumAtoms(), "bondReductionAtoms");
    bondReductionFactors = CudaArray::create<float>(cu, cu.getPaddedNumAtoms(), "sigmaEpsilon");
    tempPosq = new CudaArray(cu, cu.getPaddedNumAtoms(), cu.getUseDoublePrecision() ? sizeof(double4) : sizeof(float4), "tempPosq");
    tempForces = CudaArray::create<long long>(cu, 3*cu.getPaddedNumAtoms(), "tempForces");
    
    // Record atom parameters.
    
    vector<float2> sigmaEpsilonVec(cu.getPaddedNumAtoms(), make_float2(0, 1));
    vector<int> bondReductionAtomsVec(cu.getPaddedNumAtoms(), 0);
    vector<float> bondReductionFactorsVec(cu.getPaddedNumAtoms(), 0);
    vector<vector<int> > exclusions(cu.getNumAtoms());
    for (int i = 0; i < force.getNumParticles(); i++) {
2504
        int ivIndex;
2505
        double sigma, epsilon, reductionFactor;
2506
        force.getParticleParameters(i, ivIndex, sigma, epsilon, reductionFactor);
2507
2508
2509
2510
2511
2512
2513
2514
2515
        sigmaEpsilonVec[i] = make_float2((float) sigma, (float) epsilon);
        bondReductionAtomsVec[i] = ivIndex;
        bondReductionFactorsVec[i] = (float) reductionFactor;
        force.getParticleExclusions(i, exclusions[i]);
        exclusions[i].push_back(i);
    }
    sigmaEpsilon->upload(sigmaEpsilonVec);
    bondReductionAtoms->upload(bondReductionAtomsVec);
    bondReductionFactors->upload(bondReductionFactorsVec);
2516
2517
2518
2519
    if (force.getUseDispersionCorrection())
        dispersionCoefficient = AmoebaVdwForceImpl::calcDispersionCorrection(system, force);
    else
        dispersionCoefficient = 0.0;               
2520
2521
2522
2523
2524
2525
2526
2527
2528
2529
2530
2531
2532
2533
2534
2535
2536
2537
2538
2539
2540
2541
 
    // This force is applied based on modified atom positions, where hydrogens have been moved slightly
    // closer to their parent atoms.  We therefore create a separate CudaNonbondedUtilities just for
    // this force, so it will have its own neighbor list and interaction kernel.
    
    nonbonded = new CudaNonbondedUtilities(cu);
    nonbonded->addParameter(CudaNonbondedUtilities::ParameterInfo("sigmaEpsilon", "float", 2, sizeof(float2), sigmaEpsilon->getDevicePointer()));
    
    // Create the interaction kernel.
    
    map<string, string> replacements;
    string sigmaCombiningRule = force.getSigmaCombiningRule();
    if (sigmaCombiningRule == "ARITHMETIC")
        replacements["SIGMA_COMBINING_RULE"] = "1";
    else if (sigmaCombiningRule == "GEOMETRIC")
        replacements["SIGMA_COMBINING_RULE"] = "2";
    else if (sigmaCombiningRule == "CUBIC-MEAN")
        replacements["SIGMA_COMBINING_RULE"] = "3";
    else
        throw OpenMMException("Illegal combining rule for sigma: "+sigmaCombiningRule);
    string epsilonCombiningRule = force.getEpsilonCombiningRule();
    if (epsilonCombiningRule == "ARITHMETIC")
2542
        replacements["EPSILON_COMBINING_RULE"] = "1";
2543
    else if (epsilonCombiningRule == "GEOMETRIC")
2544
        replacements["EPSILON_COMBINING_RULE"] = "2";
2545
    else if (epsilonCombiningRule == "HARMONIC")
2546
        replacements["EPSILON_COMBINING_RULE"] = "3";
2547
    else if (epsilonCombiningRule == "HHG")
2548
        replacements["EPSILON_COMBINING_RULE"] = "4";
2549
2550
2551
2552
2553
2554
    else
        throw OpenMMException("Illegal combining rule for sigma: "+sigmaCombiningRule);
    double cutoff = force.getCutoff();
    double taperCutoff = cutoff*0.9;
    replacements["CUTOFF_DISTANCE"] = cu.doubleToString(force.getCutoff());
    replacements["TAPER_CUTOFF"] = cu.doubleToString(taperCutoff);
2555
2556
2557
    replacements["TAPER_C3"] = cu.doubleToString(10/pow(taperCutoff-cutoff, 3.0));
    replacements["TAPER_C4"] = cu.doubleToString(15/pow(taperCutoff-cutoff, 4.0));
    replacements["TAPER_C5"] = cu.doubleToString(6/pow(taperCutoff-cutoff, 5.0));
2558
2559
    bool useCutoff = (force.getNonbondedMethod() != AmoebaVdwForce::NoCutoff);
    nonbonded->addInteraction(useCutoff, useCutoff, true, force.getCutoff(), exclusions,
2560
        cu.replaceStrings(CudaAmoebaKernelSources::amoebaVdwForce2, replacements), 0);
2561
2562
2563
2564
2565
2566
2567
2568
2569
2570
2571
2572
2573
2574
2575
2576
    
    // Create the other kernels.
    
    map<string, string> defines;
    defines["PADDED_NUM_ATOMS"] = cu.intToString(cu.getPaddedNumAtoms());
    CUmodule module = cu.createModule(CudaAmoebaKernelSources::amoebaVdwForce1, defines);
    prepareKernel = cu.getKernel(module, "prepareToComputeForce");
    spreadKernel = cu.getKernel(module, "spreadForces");
    cu.addForce(new ForceInfo(force));
}

double CudaCalcAmoebaVdwForceKernel::execute(ContextImpl& context, bool includeForces, bool includeEnergy) {
    if (!hasInitializedNonbonded) {
        hasInitializedNonbonded = true;
        nonbonded->initialize(system);
    }
2577
2578
    cu.getPosq().copyTo(*tempPosq);
    cu.getForce().copyTo(*tempForces);
2579
2580
2581
    void* prepareArgs[] = {&cu.getForce().getDevicePointer(), &cu.getPosq().getDevicePointer(), &tempPosq->getDevicePointer(),
        &bondReductionAtoms->getDevicePointer(), &bondReductionFactors->getDevicePointer()};
    cu.executeKernel(prepareKernel, prepareArgs, cu.getPaddedNumAtoms());
2582
    nonbonded->prepareInteractions(1);
2583
    nonbonded->computeInteractions(1, includeForces, includeEnergy);
2584
2585
    void* spreadArgs[] = {&cu.getForce().getDevicePointer(), &tempForces->getDevicePointer(), &bondReductionAtoms->getDevicePointer(), &bondReductionFactors->getDevicePointer()};
    cu.executeKernel(spreadKernel, spreadArgs, cu.getPaddedNumAtoms());
2586
2587
    tempPosq->copyTo(cu.getPosq());
    tempForces->copyTo(cu.getForce());
2588
2589
    double4 box = cu.getPeriodicBoxSize();
    return dispersionCoefficient/(box.x*box.y*box.z);
2590
2591
}

2592
2593
2594
2595
2596
2597
2598
2599
2600
2601
2602
2603
2604
2605
2606
2607
2608
2609
2610
2611
2612
2613
2614
2615
2616
2617
2618
2619
2620
2621
void CudaCalcAmoebaVdwForceKernel::copyParametersToContext(ContextImpl& context, const AmoebaVdwForce& force) {
    // Make sure the new parameters are acceptable.
    
    cu.setAsCurrent();
    if (force.getNumParticles() != cu.getNumAtoms())
        throw OpenMMException("updateParametersInContext: The number of particles has changed");
    
    // Record the per-particle parameters.
    
    vector<float2> sigmaEpsilonVec(cu.getPaddedNumAtoms(), make_float2(0, 1));
    vector<int> bondReductionAtomsVec(cu.getPaddedNumAtoms(), 0);
    vector<float> bondReductionFactorsVec(cu.getPaddedNumAtoms(), 0);
    for (int i = 0; i < force.getNumParticles(); i++) {
        int ivIndex;
        double sigma, epsilon, reductionFactor;
        force.getParticleParameters(i, ivIndex, sigma, epsilon, reductionFactor);
        sigmaEpsilonVec[i] = make_float2((float) sigma, (float) epsilon);
        bondReductionAtomsVec[i] = ivIndex;
        bondReductionFactorsVec[i] = (float) reductionFactor;
    }
    sigmaEpsilon->upload(sigmaEpsilonVec);
    bondReductionAtoms->upload(bondReductionAtomsVec);
    bondReductionFactors->upload(bondReductionFactorsVec);
    if (force.getUseDispersionCorrection())
        dispersionCoefficient = AmoebaVdwForceImpl::calcDispersionCorrection(system, force);
    else
        dispersionCoefficient = 0.0;               
    cu.invalidateMolecules();
}

2622
2623
2624
2625
2626
2627
2628
2629
2630
2631
2632
2633
2634
2635
2636
2637
2638
2639
/* -------------------------------------------------------------------------- *
 *                           AmoebaWcaDispersion                              *
 * -------------------------------------------------------------------------- */

class CudaCalcAmoebaWcaDispersionForceKernel::ForceInfo : public CudaForceInfo {
public:
    ForceInfo(const AmoebaWcaDispersionForce& force) : force(force) {
    }
    bool areParticlesIdentical(int particle1, int particle2) {
        double radius1, radius2, epsilon1, epsilon2;
        force.getParticleParameters(particle1, radius1, epsilon1);
        force.getParticleParameters(particle2, radius2, epsilon2);
        return (radius1 == radius2 && epsilon1 == epsilon2);
    }
private:
    const AmoebaWcaDispersionForce& force;
};

2640
CudaCalcAmoebaWcaDispersionForceKernel::CudaCalcAmoebaWcaDispersionForceKernel(std::string name, const Platform& platform, CudaContext& cu, const System& system) : 
2641
2642
2643
2644
           CalcAmoebaWcaDispersionForceKernel(name, platform), cu(cu), system(system), radiusEpsilon(NULL) {
}

CudaCalcAmoebaWcaDispersionForceKernel::~CudaCalcAmoebaWcaDispersionForceKernel() {
2645
    cu.setAsCurrent();
2646
2647
2648
2649
2650
2651
2652
2653
2654
2655
2656
2657
2658
2659
2660
2661
2662
2663
2664
2665
2666
2667
2668
2669
2670
2671
2672
2673
2674
2675
2676
    if (radiusEpsilon != NULL)
        delete radiusEpsilon;
}

void CudaCalcAmoebaWcaDispersionForceKernel::initialize(const System& system, const AmoebaWcaDispersionForce& force) {
    int numParticles = system.getNumParticles();
    int paddedNumAtoms = cu.getPaddedNumAtoms();
    
    // Record parameters.
    
    vector<float2> radiusEpsilonVec(paddedNumAtoms, make_float2(0, 0));
    for (int i = 0; i < numParticles; i++) {
        double radius, epsilon;
        force.getParticleParameters(i, radius, epsilon);
        radiusEpsilonVec[i] = make_float2((float) radius, (float) epsilon);
    }
    radiusEpsilon = CudaArray::create<float2>(cu, paddedNumAtoms, "radiusEpsilon");
    radiusEpsilon->upload(radiusEpsilonVec);
    
    // Create the kernel.
    
    map<string, string> defines;
    defines["NUM_ATOMS"] = cu.intToString(numParticles);
    defines["PADDED_NUM_ATOMS"] = cu.intToString(cu.getPaddedNumAtoms());
    defines["THREAD_BLOCK_SIZE"] = cu.intToString(cu.getNonbondedUtilities().getForceThreadBlockSize());
    defines["NUM_BLOCKS"] = cu.intToString(cu.getNumAtomBlocks());
    defines["EPSO"] = cu.doubleToString(force.getEpso());
    defines["EPSH"] = cu.doubleToString(force.getEpsh());
    defines["RMINO"] = cu.doubleToString(force.getRmino());
    defines["RMINH"] = cu.doubleToString(force.getRminh());
    defines["AWATER"] = cu.doubleToString(force.getAwater());
2677
    defines["SHCTD"] = cu.doubleToString(force.getShctd());
Peter Eastman's avatar
Peter Eastman committed
2678
    defines["M_PI"] = cu.doubleToString(M_PI);
2679
2680
    CUmodule module = cu.createModule(CudaKernelSources::vectorOps+CudaAmoebaKernelSources::amoebaWcaForce, defines);
    forceKernel = cu.getKernel(module, "computeWCAForce");
2681
    totalMaximumDispersionEnergy = AmoebaWcaDispersionForceImpl::getTotalMaximumDispersionEnergy(force);
2682
2683
2684
2685
2686

    // Add an interaction to the default nonbonded kernel.  This doesn't actually do any calculations.  It's
    // just so that CudaNonbondedUtilities will keep track of the tiles.
    
    vector<vector<int> > exclusions;
2687
    cu.getNonbondedUtilities().addInteraction(false, false, false, 1.0, exclusions, "", force.getForceGroup());
2688
2689
2690
2691
2692
2693
2694
2695
2696
2697
2698
2699
2700
2701
    cu.addForce(new ForceInfo(force));
}

double CudaCalcAmoebaWcaDispersionForceKernel::execute(ContextImpl& context, bool includeForces, bool includeEnergy) {
    CudaNonbondedUtilities& nb = cu.getNonbondedUtilities();
    int startTileIndex = nb.getStartTileIndex();
    int numTileIndices = nb.getNumTiles();
    int numForceThreadBlocks = nb.getNumForceThreadBlocks();
    int forceThreadBlockSize = nb.getForceThreadBlockSize();
    void* forceArgs[] = {&cu.getForce().getDevicePointer(), &cu.getEnergyBuffer().getDevicePointer(),
        &cu.getPosq().getDevicePointer(), &startTileIndex, &numTileIndices, &radiusEpsilon->getDevicePointer()};
    cu.executeKernel(forceKernel, forceArgs, numForceThreadBlocks*forceThreadBlockSize, forceThreadBlockSize);
    return totalMaximumDispersionEnergy;
}
2702
2703
2704
2705
2706
2707
2708
2709
2710
2711
2712
2713
2714
2715
2716
2717
2718

void CudaCalcAmoebaWcaDispersionForceKernel::copyParametersToContext(ContextImpl& context, const AmoebaWcaDispersionForce& force) {
    // Make sure the new parameters are acceptable.
    
    cu.setAsCurrent();
    if (force.getNumParticles() != cu.getNumAtoms())
        throw OpenMMException("updateParametersInContext: The number of particles has changed");
    
    // Record the per-particle parameters.
    
    vector<float2> radiusEpsilonVec(cu.getPaddedNumAtoms(), make_float2(0, 0));
    for (int i = 0; i < cu.getNumAtoms(); i++) {
        double radius, epsilon;
        force.getParticleParameters(i, radius, epsilon);
        radiusEpsilonVec[i] = make_float2((float) radius, (float) epsilon);
    }
    radiusEpsilon->upload(radiusEpsilonVec);
2719
    totalMaximumDispersionEnergy = AmoebaWcaDispersionForceImpl::getTotalMaximumDispersionEnergy(force);
2720
2721
    cu.invalidateMolecules();
}