OpenCLKernels.cpp 198 KB
Newer Older
1
2
3
4
5
6
7
8
/* -------------------------------------------------------------------------- *
 *                                   OpenMM                                   *
 * -------------------------------------------------------------------------- *
 * This is part of the OpenMM molecular simulation toolkit originating from   *
 * Simbios, the NIH National Center for Physics-Based Simulation of           *
 * Biological Structures at Stanford, funded under the NIH Roadmap for        *
 * Medical Research, grant U54 GM072970. See https://simtk.org.               *
 *                                                                            *
9
 * Portions copyright (c) 2008-2010 Stanford University and the Authors.      *
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
 * Authors: Peter Eastman                                                     *
 * Contributors:                                                              *
 *                                                                            *
 * This program is free software: you can redistribute it and/or modify       *
 * it under the terms of the GNU Lesser General Public License as published   *
 * by the Free Software Foundation, either version 3 of the License, or       *
 * (at your option) any later version.                                        *
 *                                                                            *
 * This program is distributed in the hope that it will be useful,            *
 * but WITHOUT ANY WARRANTY; without even the implied warranty of             *
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the              *
 * GNU Lesser General Public License for more details.                        *
 *                                                                            *
 * You should have received a copy of the GNU Lesser General Public License   *
 * along with this program.  If not, see <http://www.gnu.org/licenses/>.      *
 * -------------------------------------------------------------------------- */

#include "OpenCLKernels.h"
28
#include "OpenCLForceInfo.h"
29
30
#include "openmm/LangevinIntegrator.h"
#include "openmm/Context.h"
31
#include "openmm/internal/AndersenThermostatImpl.h"
32
#include "openmm/internal/CMAPTorsionForceImpl.h"
33
#include "openmm/internal/ContextImpl.h"
34
#include "openmm/internal/CustomHbondForceImpl.h"
35
#include "openmm/internal/NonbondedForceImpl.h"
36
#include "OpenCLExpressionUtilities.h"
37
#include "OpenCLIntegrationUtilities.h"
38
#include "OpenCLNonbondedUtilities.h"
39
#include "OpenCLKernelSources.h"
40
#include "lepton/Operation.h"
41
42
#include "lepton/Parser.h"
#include "lepton/ParsedExpression.h"
43
#include "../src/SimTKUtilities/SimTKOpenMMRealType.h"
44
#include <cmath>
45
#include <set>
46
47
48
49

using namespace OpenMM;
using namespace std;

50
51
52
53
54
55
56
57
58
59
60
61
62
static string doubleToString(double value) {
    stringstream s;
    s.precision(8);
    s << scientific << value << "f";
    return s.str();
}

static string intToString(int value) {
    stringstream s;
    s << value;
    return s.str();
}

63
64
65
66
67
68
69
static bool isZeroExpression(const Lepton::ParsedExpression& expression) {
    const Lepton::Operation& op = expression.getRootNode().getOperation();
    if (op.getId() != Lepton::Operation::CONSTANT)
        return false;
    return (dynamic_cast<const Lepton::Operation::Constant&>(op).getValue() == 0.0);
}

70
void OpenCLCalcForcesAndEnergyKernel::initialize(const System& system) {
71
72
}

73
void OpenCLCalcForcesAndEnergyKernel::beginComputation(ContextImpl& context, bool includeForces, bool includeEnergy) {
74
    if (cl.getNonbondedUtilities().getUseCutoff() && cl.getComputeForceCount()%100 == 0) {
75
        cl.reorderAtoms();
76
77
        cl.getNonbondedUtilities().updateNeighborListSize();
    }
78
    cl.setComputeForceCount(cl.getComputeForceCount()+1);
79
    cl.clearAutoclearBuffers();
80
    cl.getNonbondedUtilities().prepareInteractions();
81
82
}

83
double OpenCLCalcForcesAndEnergyKernel::finishComputation(ContextImpl& context, bool includeForces, bool includeEnergy) {
84
    cl.getNonbondedUtilities().computeInteractions();
85
    if (includeForces)
86
        cl.reduceForces();
87
    double sum = 0.0f;
88
89
90
91
92
93
    if (includeEnergy) {
        OpenCLArray<cl_float>& energy = cl.getEnergyBuffer();
        energy.download();
        for (int i = 0; i < energy.getSize(); i++)
            sum += energy[i];
    }
94
    return sum;
95
96
}

97
void OpenCLUpdateStateDataKernel::initialize(const System& system) {
98
99
}

100
double OpenCLUpdateStateDataKernel::getTime(const ContextImpl& context) const {
101
    return cl.getTime();
102
103
}

104
void OpenCLUpdateStateDataKernel::setTime(ContextImpl& context, double time) {
105
106
107
    vector<OpenCLContext*>& contexts = cl.getPlatformData().contexts;
    for (int i = 0; i < (int) contexts.size(); i++)
        contexts[i]->setTime(time);
108
109
}

110
void OpenCLUpdateStateDataKernel::getPositions(ContextImpl& context, std::vector<Vec3>& positions) {
111
    OpenCLArray<mm_float4>& posq = cl.getPosq();
112
    posq.download();
113
    OpenCLArray<cl_int>& order = cl.getAtomIndex();
114
115
    int numParticles = context.getSystem().getNumParticles();
    positions.resize(numParticles);
116
    mm_float4 periodicBoxSize = cl.getPeriodicBoxSize();
117
    for (int i = 0; i < numParticles; ++i) {
118
        mm_float4 pos = posq[i];
119
120
        mm_int4 offset = cl.getPosCellOffsets()[i];
        positions[order[i]] = Vec3(pos.x-offset.x*periodicBoxSize.x, pos.y-offset.y*periodicBoxSize.y, pos.z-offset.z*periodicBoxSize.z);
121
122
123
124
    }
}

void OpenCLUpdateStateDataKernel::setPositions(ContextImpl& context, const std::vector<Vec3>& positions) {
125
126
    OpenCLArray<mm_float4>& posq = cl.getPosq();
    OpenCLArray<cl_int>& order = cl.getAtomIndex();
127
128
    int numParticles = context.getSystem().getNumParticles();
    for (int i = 0; i < numParticles; ++i) {
129
        mm_float4& pos = posq[i];
130
        const Vec3& p = positions[order[i]];
131
132
133
        pos.x = (cl_float) p[0];
        pos.y = (cl_float) p[1];
        pos.z = (cl_float) p[2];
134
    }
135
136
    for (int i = numParticles; i < cl.getPaddedNumAtoms(); i++)
        posq[i] = mm_float4(0.0f, 0.0f, 0.0f, 0.0f);
137
    posq.upload();
138
    for (int i = 0; i < (int) cl.getPosCellOffsets().size(); i++)
139
        cl.getPosCellOffsets()[i] = mm_int4(0, 0, 0, 0);
140
141
142
}

void OpenCLUpdateStateDataKernel::getVelocities(ContextImpl& context, std::vector<Vec3>& velocities) {
143
    OpenCLArray<mm_float4>& velm = cl.getVelm();
144
    velm.download();
145
    OpenCLArray<cl_int>& order = cl.getAtomIndex();
146
147
148
    int numParticles = context.getSystem().getNumParticles();
    velocities.resize(numParticles);
    for (int i = 0; i < numParticles; ++i) {
149
150
        mm_float4 vel = velm[i];
        velocities[order[i]] = Vec3(vel.x, vel.y, vel.z);
151
152
153
154
    }
}

void OpenCLUpdateStateDataKernel::setVelocities(ContextImpl& context, const std::vector<Vec3>& velocities) {
155
156
    OpenCLArray<mm_float4>& velm = cl.getVelm();
    OpenCLArray<cl_int>& order = cl.getAtomIndex();
157
158
    int numParticles = context.getSystem().getNumParticles();
    for (int i = 0; i < numParticles; ++i) {
159
        mm_float4& vel = velm[i];
160
        const Vec3& p = velocities[order[i]];
161
162
163
        vel.x = (cl_float) p[0];
        vel.y = (cl_float) p[1];
        vel.z = (cl_float) p[2];
164
    }
165
166
    for (int i = numParticles; i < cl.getPaddedNumAtoms(); i++)
        velm[i] = mm_float4(0.0f, 0.0f, 0.0f, 0.0f);
167
168
169
170
    velm.upload();
}

void OpenCLUpdateStateDataKernel::getForces(ContextImpl& context, std::vector<Vec3>& forces) {
171
    OpenCLArray<mm_float4>& force = cl.getForce();
172
    force.download();
173
    OpenCLArray<cl_int>& order = cl.getAtomIndex();
174
175
176
    int numParticles = context.getSystem().getNumParticles();
    forces.resize(numParticles);
    for (int i = 0; i < numParticles; ++i) {
177
178
        mm_float4 f = force[i];
        forces[order[i]] = Vec3(f.x, f.y, f.z);
179
180
181
    }
}

182
183
184
185
186
187
188
189
void OpenCLUpdateStateDataKernel::getPeriodicBoxVectors(ContextImpl& context, Vec3& a, Vec3& b, Vec3& c) const {
    mm_float4 box = cl.getPeriodicBoxSize();
    a = Vec3(box.x, 0, 0);
    b = Vec3(0, box.y, 0);
    c = Vec3(0, 0, box.z);
}

void OpenCLUpdateStateDataKernel::setPeriodicBoxVectors(ContextImpl& context, const Vec3& a, const Vec3& b, const Vec3& c) const {
190
191
192
    vector<OpenCLContext*>& contexts = cl.getPlatformData().contexts;
    for (int i = 0; i < (int) contexts.size(); i++)
        contexts[i]->setPeriodicBoxSize(a[0], b[1], c[2]);
193
194
}

195
196
197
198
199
200
201
void OpenCLApplyConstraintsKernel::initialize(const System& system) {
}

void OpenCLApplyConstraintsKernel::apply(ContextImpl& context, double tol) {
    cl.getIntegrationUtilities().applyConstraints(tol);
}

202
203
class OpenCLBondForceInfo : public OpenCLForceInfo {
public:
204
    OpenCLBondForceInfo(int requiredBuffers, const HarmonicBondForce& force) : OpenCLForceInfo(requiredBuffers), force(force) {
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
    }
    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:
    const HarmonicBondForce& force;
};

228
229
230
231
232
233
234
OpenCLCalcHarmonicBondForceKernel::~OpenCLCalcHarmonicBondForceKernel() {
    if (params != NULL)
        delete params;
    if (indices != NULL)
        delete indices;
}

235
void OpenCLCalcHarmonicBondForceKernel::initialize(const System& system, const HarmonicBondForce& force) {
236
237
238
239
    int numContexts = cl.getPlatformData().contexts.size();
    int startIndex = cl.getContextIndex()*force.getNumBonds()/numContexts;
    int endIndex = (cl.getContextIndex()+1)*force.getNumBonds()/numContexts;
    numBonds = endIndex-startIndex;
240
241
    if (numBonds == 0)
        return;
242
243
    params = new OpenCLArray<mm_float2>(cl, numBonds, "bondParams");
    indices = new OpenCLArray<mm_int4>(cl, numBonds, "bondIndices");
244
245
246
247
248
249
    vector<int> forceBufferCounter(system.getNumParticles(), 0);
    vector<mm_float2> paramVector(numBonds);
    vector<mm_int4> indicesVector(numBonds);
    for (int i = 0; i < numBonds; i++) {
        int particle1, particle2;
        double length, k;
250
        force.getBondParameters(startIndex+i, particle1, particle2, length, k);
251
252
        paramVector[i] = mm_float2((cl_float) length, (cl_float) k);
        indicesVector[i] = mm_int4(particle1, particle2, forceBufferCounter[particle1]++, forceBufferCounter[particle2]++);
253
254
255
256
    }
    params->upload(paramVector);
    indices->upload(indicesVector);
    int maxBuffers = 1;
257
    for (int i = 0; i < (int) forceBufferCounter.size(); i++)
258
        maxBuffers = max(maxBuffers, forceBufferCounter[i]);
259
    cl.addForce(new OpenCLBondForceInfo(maxBuffers, force));
260
    cl::Program program = cl.createProgram(OpenCLKernelSources::harmonicBondForce);
261
262
263
    kernel = cl::Kernel(program, "calcHarmonicBondForce");
}

264
double OpenCLCalcHarmonicBondForceKernel::execute(ContextImpl& context, bool includeForces, bool includeEnergy) {
265
    if (numBonds == 0)
266
        return 0.0;
267
268
269
270
271
272
273
274
275
276
    if (!hasInitializedKernel) {
        hasInitializedKernel = true;
        kernel.setArg<cl_int>(0, cl.getPaddedNumAtoms());
        kernel.setArg<cl_int>(1, numBonds);
        kernel.setArg<cl::Buffer>(2, cl.getForceBuffers().getDeviceBuffer());
        kernel.setArg<cl::Buffer>(3, cl.getEnergyBuffer().getDeviceBuffer());
        kernel.setArg<cl::Buffer>(4, cl.getPosq().getDeviceBuffer());
        kernel.setArg<cl::Buffer>(5, params->getDeviceBuffer());
        kernel.setArg<cl::Buffer>(6, indices->getDeviceBuffer());
    }
277
    cl.executeKernel(kernel, numBonds);
278
279
    return 0.0;
}
280

281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
class OpenCLCustomBondForceInfo : public OpenCLForceInfo {
public:
    OpenCLCustomBondForceInfo(int requiredBuffers, const CustomBondForce& force) : OpenCLForceInfo(requiredBuffers), force(force) {
    }
    int getNumParticleGroups() {
        return force.getNumBonds();
    }
    void getParticlesInGroup(int index, std::vector<int>& particles) {
        int particle1, particle2;
        vector<double> parameters;
        force.getBondParameters(index, particle1, particle2, parameters);
        particles.resize(2);
        particles[0] = particle1;
        particles[1] = particle2;
    }
    bool areGroupsIdentical(int group1, int group2) {
        int particle1, particle2;
        vector<double> parameters1, parameters2;
        force.getBondParameters(group1, particle1, particle2, parameters1);
        force.getBondParameters(group2, particle1, particle2, parameters2);
        for (int i = 0; i < (int) parameters1.size(); i++)
            if (parameters1[i] != parameters2[i])
                return false;
        return true;
    }
private:
    const CustomBondForce& force;
};

OpenCLCalcCustomBondForceKernel::~OpenCLCalcCustomBondForceKernel() {
    if (params != NULL)
        delete params;
    if (indices != NULL)
        delete indices;
    if (globals != NULL)
        delete globals;
}

void OpenCLCalcCustomBondForceKernel::initialize(const System& system, const CustomBondForce& force) {
320
321
322
323
    int numContexts = cl.getPlatformData().contexts.size();
    int startIndex = cl.getContextIndex()*force.getNumBonds()/numContexts;
    int endIndex = (cl.getContextIndex()+1)*force.getNumBonds()/numContexts;
    numBonds = endIndex-startIndex;
324
325
    if (numBonds == 0)
        return;
326
    params = new OpenCLParameterSet(cl, force.getNumPerBondParameters(), numBonds, "customBondParams");
327
328
329
330
331
332
333
    indices = new OpenCLArray<mm_int4>(cl, numBonds, "customBondIndices");
    string extraArguments;
    if (force.getNumGlobalParameters() > 0) {
        globals = new OpenCLArray<cl_float>(cl, force.getNumGlobalParameters(), "customBondGlobals", false, CL_MEM_READ_ONLY);
        extraArguments += ", __constant float* globals";
    }
    vector<int> forceBufferCounter(system.getNumParticles(), 0);
334
    vector<vector<cl_float> > paramVector(numBonds);
335
336
337
338
    vector<mm_int4> indicesVector(numBonds);
    for (int i = 0; i < numBonds; i++) {
        int particle1, particle2;
        vector<double> parameters;
339
        force.getBondParameters(startIndex+i, particle1, particle2, parameters);
340
        paramVector[i].resize(parameters.size());
341
        for (int j = 0; j < (int) parameters.size(); j++)
342
            paramVector[i][j] = (cl_float) parameters[j];
343
        indicesVector[i] = mm_int4(particle1, particle2, forceBufferCounter[particle1]++, forceBufferCounter[particle2]++);
344
    }
345
    params->setParameterValues(paramVector);
346
347
    indices->upload(indicesVector);
    int maxBuffers = 1;
348
    for (int i = 0; i < (int) forceBufferCounter.size(); i++)
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
        maxBuffers = max(maxBuffers, forceBufferCounter[i]);
    cl.addForce(new OpenCLCustomBondForceInfo(maxBuffers, force));

    // Record information for the expressions.

    globalParamNames.resize(force.getNumGlobalParameters());
    globalParamValues.resize(force.getNumGlobalParameters());
    for (int i = 0; i < force.getNumGlobalParameters(); i++) {
        globalParamNames[i] = force.getGlobalParameterName(i);
        globalParamValues[i] = (cl_float) force.getGlobalParameterDefaultValue(i);
    }
    if (globals != NULL)
        globals->upload(globalParamValues);
    Lepton::ParsedExpression energyExpression = Lepton::Parser::parse(force.getEnergyFunction()).optimize();
    Lepton::ParsedExpression forceExpression = energyExpression.differentiate("r").optimize();
    map<string, Lepton::ParsedExpression> expressions;
    expressions["energy += "] = energyExpression;
    expressions["float dEdR = "] = forceExpression;

    // Create the kernels.

    map<string, string> variables;
    variables["r"] = "r";
    for (int i = 0; i < force.getNumPerBondParameters(); i++) {
        const string& name = force.getPerBondParameterName(i);
374
        variables[name] = "bondParams"+params->getParameterSuffix(i);
375
376
377
378
379
380
381
    }
    for (int i = 0; i < force.getNumGlobalParameters(); i++) {
        const string& name = force.getGlobalParameterName(i);
        string value = "globals["+intToString(i)+"]";
        variables[name] = value;
    }
    stringstream compute;
382
383
384
385
386
    for (int i = 0; i < (int) params->getBuffers().size(); i++) {
        const OpenCLNonbondedUtilities::ParameterInfo& buffer = params->getBuffers()[i];
        extraArguments += ", __global "+buffer.getType()+"* "+buffer.getName();
        compute<<buffer.getType()<<" bondParams"<<(i+1)<<" = "<<buffer.getName()<<"[index];\n";
    }
387
388
    vector<pair<string, string> > functions;
    compute << OpenCLExpressionUtilities::createExpressions(expressions, variables, functions, "temp", "");
389
    map<string, string> replacements;
390
391
    replacements["COMPUTE_FORCE"] = compute.str();
    replacements["EXTRA_ARGUMENTS"] = extraArguments;
392
    cl::Program program = cl.createProgram(cl.replaceStrings(OpenCLKernelSources::customBondForce, replacements));
393
394
395
    kernel = cl::Kernel(program, "computeCustomBondForces");
}

396
double OpenCLCalcCustomBondForceKernel::execute(ContextImpl& context, bool includeForces, bool includeEnergy) {
397
    if (numBonds == 0)
398
        return 0.0;
399
400
    if (globals != NULL) {
        bool changed = false;
401
        for (int i = 0; i < (int) globalParamNames.size(); i++) {
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
            cl_float value = (cl_float) context.getParameter(globalParamNames[i]);
            if (value != globalParamValues[i])
                changed = true;
            globalParamValues[i] = value;
        }
        if (changed)
            globals->upload(globalParamValues);
    }
    if (!hasInitializedKernel) {
        hasInitializedKernel = true;
        kernel.setArg<cl_int>(0, cl.getPaddedNumAtoms());
        kernel.setArg<cl_int>(1, numBonds);
        kernel.setArg<cl::Buffer>(2, cl.getForceBuffers().getDeviceBuffer());
        kernel.setArg<cl::Buffer>(3, cl.getEnergyBuffer().getDeviceBuffer());
        kernel.setArg<cl::Buffer>(4, cl.getPosq().getDeviceBuffer());
417
418
        kernel.setArg<cl::Buffer>(5, indices->getDeviceBuffer());
        int nextIndex = 6;
419
        if (globals != NULL)
420
421
422
            kernel.setArg<cl::Buffer>(nextIndex++, globals->getDeviceBuffer());
        for (int i = 0; i < (int) params->getBuffers().size(); i++) {
            const OpenCLNonbondedUtilities::ParameterInfo& buffer = params->getBuffers()[i];
423
            kernel.setArg<cl::Memory>(nextIndex++, buffer.getMemory());
424
        }
425
426
427
428
429
    }
    cl.executeKernel(kernel, numBonds);
    return 0.0;
}

430
431
class OpenCLAngleForceInfo : public OpenCLForceInfo {
public:
432
    OpenCLAngleForceInfo(int requiredBuffers, const HarmonicAngleForce& force) : OpenCLForceInfo(requiredBuffers), force(force) {
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
    }
    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:
    const HarmonicAngleForce& force;
};

OpenCLCalcHarmonicAngleForceKernel::~OpenCLCalcHarmonicAngleForceKernel() {
    if (params != NULL)
        delete params;
    if (indices != NULL)
        delete indices;
}

void OpenCLCalcHarmonicAngleForceKernel::initialize(const System& system, const HarmonicAngleForce& force) {
465
466
467
468
    int numContexts = cl.getPlatformData().contexts.size();
    int startIndex = cl.getContextIndex()*force.getNumAngles()/numContexts;
    int endIndex = (cl.getContextIndex()+1)*force.getNumAngles()/numContexts;
    numAngles = endIndex-startIndex;
469
470
    if (numAngles == 0)
        return;
471
472
    params = new OpenCLArray<mm_float2>(cl, numAngles, "angleParams");
    indices = new OpenCLArray<mm_int8>(cl, numAngles, "angleIndices");
473
474
475
476
477
478
    vector<int> forceBufferCounter(system.getNumParticles(), 0);
    vector<mm_float2> paramVector(numAngles);
    vector<mm_int8> indicesVector(numAngles);
    for (int i = 0; i < numAngles; i++) {
        int particle1, particle2, particle3;
        double angle, k;
479
        force.getAngleParameters(startIndex+i, particle1, particle2, particle3, angle, k);
480
481
482
        paramVector[i] = mm_float2((cl_float) angle, (cl_float) k);
        indicesVector[i] = mm_int8(particle1, particle2, particle3,
                forceBufferCounter[particle1]++, forceBufferCounter[particle2]++, forceBufferCounter[particle3]++, 0, 0);
483
484
485
486
487

    }
    params->upload(paramVector);
    indices->upload(indicesVector);
    int maxBuffers = 1;
488
    for (int i = 0; i < (int) forceBufferCounter.size(); i++)
489
        maxBuffers = max(maxBuffers, forceBufferCounter[i]);
490
    cl.addForce(new OpenCLAngleForceInfo(maxBuffers, force));
491
    cl::Program program = cl.createProgram(OpenCLKernelSources::harmonicAngleForce);
492
493
494
    kernel = cl::Kernel(program, "calcHarmonicAngleForce");
}

495
double OpenCLCalcHarmonicAngleForceKernel::execute(ContextImpl& context, bool includeForces, bool includeEnergy) {
496
    if (numAngles == 0)
497
        return 0.0;
498
499
500
501
502
503
504
505
506
507
    if (!hasInitializedKernel) {
        hasInitializedKernel = true;
        kernel.setArg<cl_int>(0, cl.getPaddedNumAtoms());
        kernel.setArg<cl_int>(1, numAngles);
        kernel.setArg<cl::Buffer>(2, cl.getForceBuffers().getDeviceBuffer());
        kernel.setArg<cl::Buffer>(3, cl.getEnergyBuffer().getDeviceBuffer());
        kernel.setArg<cl::Buffer>(4, cl.getPosq().getDeviceBuffer());
        kernel.setArg<cl::Buffer>(5, params->getDeviceBuffer());
        kernel.setArg<cl::Buffer>(6, indices->getDeviceBuffer());
    }
508
    cl.executeKernel(kernel, numAngles);
509
510
511
    return 0.0;
}

512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
class OpenCLCustomAngleForceInfo : public OpenCLForceInfo {
public:
    OpenCLCustomAngleForceInfo(int requiredBuffers, const CustomAngleForce& force) : OpenCLForceInfo(requiredBuffers), force(force) {
    }
    int getNumParticleGroups() {
        return force.getNumAngles();
    }
    void getParticlesInGroup(int index, std::vector<int>& particles) {
        int particle1, particle2, particle3;
        vector<double> parameters;
        force.getAngleParameters(index, particle1, particle2, particle3, parameters);
        particles.resize(3);
        particles[0] = particle1;
        particles[1] = particle2;
        particles[2] = particle3;
    }
    bool areGroupsIdentical(int group1, int group2) {
        int particle1, particle2, particle3;
        vector<double> parameters1, parameters2;
        force.getAngleParameters(group1, particle1, particle2, particle3, parameters1);
        force.getAngleParameters(group2, particle1, particle2, particle3, parameters2);
        for (int i = 0; i < (int) parameters1.size(); i++)
            if (parameters1[i] != parameters2[i])
                return false;
        return true;
    }
private:
    const CustomAngleForce& force;
};

OpenCLCalcCustomAngleForceKernel::~OpenCLCalcCustomAngleForceKernel() {
    if (params != NULL)
        delete params;
    if (indices != NULL)
        delete indices;
    if (globals != NULL)
        delete globals;
}

void OpenCLCalcCustomAngleForceKernel::initialize(const System& system, const CustomAngleForce& force) {
552
553
554
555
    int numContexts = cl.getPlatformData().contexts.size();
    int startIndex = cl.getContextIndex()*force.getNumAngles()/numContexts;
    int endIndex = (cl.getContextIndex()+1)*force.getNumAngles()/numContexts;
    numAngles = endIndex-startIndex;
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
    if (numAngles == 0)
        return;
    params = new OpenCLParameterSet(cl, force.getNumPerAngleParameters(), numAngles, "customAngleParams");
    indices = new OpenCLArray<mm_int8>(cl, numAngles, "customAngleIndices");
    string extraArguments;
    if (force.getNumGlobalParameters() > 0) {
        globals = new OpenCLArray<cl_float>(cl, force.getNumGlobalParameters(), "customAngleGlobals", false, CL_MEM_READ_ONLY);
        extraArguments += ", __constant float* globals";
    }
    vector<int> forceBufferCounter(system.getNumParticles(), 0);
    vector<vector<cl_float> > paramVector(numAngles);
    vector<mm_int8> indicesVector(numAngles);
    for (int i = 0; i < numAngles; i++) {
        int particle1, particle2, particle3;
        vector<double> parameters;
571
        force.getAngleParameters(startIndex+i, particle1, particle2, particle3, parameters);
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
        paramVector[i].resize(parameters.size());
        for (int j = 0; j < (int) parameters.size(); j++)
            paramVector[i][j] = (cl_float) parameters[j];
        indicesVector[i] = mm_int8(particle1, particle2, particle3, forceBufferCounter[particle1]++,
                forceBufferCounter[particle2]++, forceBufferCounter[particle3]++, 0, 0);
    }
    params->setParameterValues(paramVector);
    indices->upload(indicesVector);
    int maxBuffers = 1;
    for (int i = 0; i < (int) forceBufferCounter.size(); i++)
        maxBuffers = max(maxBuffers, forceBufferCounter[i]);
    cl.addForce(new OpenCLCustomAngleForceInfo(maxBuffers, force));

    // Record information for the expressions.

    globalParamNames.resize(force.getNumGlobalParameters());
    globalParamValues.resize(force.getNumGlobalParameters());
    for (int i = 0; i < force.getNumGlobalParameters(); i++) {
        globalParamNames[i] = force.getGlobalParameterName(i);
        globalParamValues[i] = (cl_float) force.getGlobalParameterDefaultValue(i);
    }
    if (globals != NULL)
        globals->upload(globalParamValues);
    Lepton::ParsedExpression energyExpression = Lepton::Parser::parse(force.getEnergyFunction()).optimize();
    Lepton::ParsedExpression forceExpression = energyExpression.differentiate("theta").optimize();
    map<string, Lepton::ParsedExpression> expressions;
    expressions["energy += "] = energyExpression;
599
    expressions["float dEdAngle = "] = forceExpression;
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

    // Create the kernels.

    map<string, string> variables;
    variables["theta"] = "theta";
    for (int i = 0; i < force.getNumPerAngleParameters(); i++) {
        const string& name = force.getPerAngleParameterName(i);
        variables[name] = "angleParams"+params->getParameterSuffix(i);
    }
    for (int i = 0; i < force.getNumGlobalParameters(); i++) {
        const string& name = force.getGlobalParameterName(i);
        string value = "globals["+intToString(i)+"]";
        variables[name] = value;
    }
    stringstream compute;
    for (int i = 0; i < (int) params->getBuffers().size(); i++) {
        const OpenCLNonbondedUtilities::ParameterInfo& buffer = params->getBuffers()[i];
        extraArguments += ", __global "+buffer.getType()+"* "+buffer.getName();
        compute<<buffer.getType()<<" angleParams"<<(i+1)<<" = "<<buffer.getName()<<"[index];\n";
    }
    vector<pair<string, string> > functions;
    compute << OpenCLExpressionUtilities::createExpressions(expressions, variables, functions, "temp", "");
    map<string, string> replacements;
    replacements["COMPUTE_FORCE"] = compute.str();
    replacements["EXTRA_ARGUMENTS"] = extraArguments;
    cl::Program program = cl.createProgram(cl.replaceStrings(OpenCLKernelSources::customAngleForce, replacements));
    kernel = cl::Kernel(program, "computeCustomAngleForces");
}

629
double OpenCLCalcCustomAngleForceKernel::execute(ContextImpl& context, bool includeForces, bool includeEnergy) {
630
    if (numAngles == 0)
631
        return 0.0;
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
    if (globals != NULL) {
        bool changed = false;
        for (int i = 0; i < (int) globalParamNames.size(); i++) {
            cl_float value = (cl_float) context.getParameter(globalParamNames[i]);
            if (value != globalParamValues[i])
                changed = true;
            globalParamValues[i] = value;
        }
        if (changed)
            globals->upload(globalParamValues);
    }
    if (!hasInitializedKernel) {
        hasInitializedKernel = true;
        kernel.setArg<cl_int>(0, cl.getPaddedNumAtoms());
        kernel.setArg<cl_int>(1, numAngles);
        kernel.setArg<cl::Buffer>(2, cl.getForceBuffers().getDeviceBuffer());
        kernel.setArg<cl::Buffer>(3, cl.getEnergyBuffer().getDeviceBuffer());
        kernel.setArg<cl::Buffer>(4, cl.getPosq().getDeviceBuffer());
        kernel.setArg<cl::Buffer>(5, indices->getDeviceBuffer());
        int nextIndex = 6;
        if (globals != NULL)
            kernel.setArg<cl::Buffer>(nextIndex++, globals->getDeviceBuffer());
        for (int i = 0; i < (int) params->getBuffers().size(); i++) {
            const OpenCLNonbondedUtilities::ParameterInfo& buffer = params->getBuffers()[i];
656
            kernel.setArg<cl::Memory>(nextIndex++, buffer.getMemory());
657
658
659
660
661
662
        }
    }
    cl.executeKernel(kernel, numAngles);
    return 0.0;
}

663
664
class OpenCLPeriodicTorsionForceInfo : public OpenCLForceInfo {
public:
665
    OpenCLPeriodicTorsionForceInfo(int requiredBuffers, const PeriodicTorsionForce& force) : OpenCLForceInfo(requiredBuffers), force(force) {
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
    }
    int getNumParticleGroups() {
        return force.getNumTorsions();
    }
    void getParticlesInGroup(int index, std::vector<int>& particles) {
        int particle1, particle2, particle3, particle4, periodicity;
        double phase, k;
        force.getTorsionParameters(index, particle1, particle2, particle3, particle4, periodicity, phase, 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, periodicity1, periodicity2;
        double phase1, phase2, k1, k2;
        force.getTorsionParameters(group1, particle1, particle2, particle3, particle4, periodicity1, phase1, k1);
684
        force.getTorsionParameters(group2, particle1, particle2, particle3, particle4, periodicity2, phase2, k2);
685
686
687
688
689
690
691
692
693
694
695
696
697
698
        return (periodicity1 == periodicity2 && phase1 == phase2 && k1 == k2);
    }
private:
    const PeriodicTorsionForce& force;
};

OpenCLCalcPeriodicTorsionForceKernel::~OpenCLCalcPeriodicTorsionForceKernel() {
    if (params != NULL)
        delete params;
    if (indices != NULL)
        delete indices;
}

void OpenCLCalcPeriodicTorsionForceKernel::initialize(const System& system, const PeriodicTorsionForce& force) {
699
700
701
702
    int numContexts = cl.getPlatformData().contexts.size();
    int startIndex = cl.getContextIndex()*force.getNumTorsions()/numContexts;
    int endIndex = (cl.getContextIndex()+1)*force.getNumTorsions()/numContexts;
    numTorsions = endIndex-startIndex;
703
704
    if (numTorsions == 0)
        return;
705
706
    params = new OpenCLArray<mm_float4>(cl, numTorsions, "periodicTorsionParams");
    indices = new OpenCLArray<mm_int8>(cl, numTorsions, "periodicTorsionIndices");
707
708
709
710
711
712
    vector<int> forceBufferCounter(system.getNumParticles(), 0);
    vector<mm_float4> paramVector(numTorsions);
    vector<mm_int8> indicesVector(numTorsions);
    for (int i = 0; i < numTorsions; i++) {
        int particle1, particle2, particle3, particle4, periodicity;
        double phase, k;
713
        force.getTorsionParameters(startIndex+i, particle1, particle2, particle3, particle4, periodicity, phase, k);
714
715
716
        paramVector[i] = mm_float4((cl_float) k, (cl_float) phase, (cl_float) periodicity, 0.0f);
        indicesVector[i] = mm_int8(particle1, particle2, particle3, particle4,
                forceBufferCounter[particle1]++, forceBufferCounter[particle2]++, forceBufferCounter[particle3]++, forceBufferCounter[particle4]++);
717
718
719
720
721

    }
    params->upload(paramVector);
    indices->upload(indicesVector);
    int maxBuffers = 1;
722
    for (int i = 0; i < (int) forceBufferCounter.size(); i++)
723
        maxBuffers = max(maxBuffers, forceBufferCounter[i]);
724
    cl.addForce(new OpenCLPeriodicTorsionForceInfo(maxBuffers, force));
725
    cl::Program program = cl.createProgram(OpenCLKernelSources::periodicTorsionForce);
726
727
728
    kernel = cl::Kernel(program, "calcPeriodicTorsionForce");
}

729
double OpenCLCalcPeriodicTorsionForceKernel::execute(ContextImpl& context, bool includeForces, bool includeEnergy) {
730
    if (numTorsions == 0)
731
        return 0.0;
732
733
734
735
736
737
738
739
740
741
    if (!hasInitializedKernel) {
        hasInitializedKernel = true;
        kernel.setArg<cl_int>(0, cl.getPaddedNumAtoms());
        kernel.setArg<cl_int>(1, numTorsions);
        kernel.setArg<cl::Buffer>(2, cl.getForceBuffers().getDeviceBuffer());
        kernel.setArg<cl::Buffer>(3, cl.getEnergyBuffer().getDeviceBuffer());
        kernel.setArg<cl::Buffer>(4, cl.getPosq().getDeviceBuffer());
        kernel.setArg<cl::Buffer>(5, params->getDeviceBuffer());
        kernel.setArg<cl::Buffer>(6, indices->getDeviceBuffer());
    }
742
    cl.executeKernel(kernel, numTorsions);
743
744
745
    return 0.0;
}

746
747
class OpenCLRBTorsionForceInfo : public OpenCLForceInfo {
public:
748
    OpenCLRBTorsionForceInfo(int requiredBuffers, const RBTorsionForce& force) : OpenCLForceInfo(requiredBuffers), force(force) {
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
    }
    int getNumParticleGroups() {
        return force.getNumTorsions();
    }
    void getParticlesInGroup(int index, std::vector<int>& particles) {
        int particle1, particle2, particle3, particle4;
        double c0, c1, c2, c3, c4, c5;
        force.getTorsionParameters(index, particle1, particle2, particle3, particle4, c0, c1, c2, c3, c4, c5);
        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 c0a, c0b, c1a, c1b, c2a, c2b, c3a, c3b, c4a, c4b, c5a, c5b;
        force.getTorsionParameters(group1, particle1, particle2, particle3, particle4, c0a, c1a, c2a, c3a, c4a, c5a);
767
        force.getTorsionParameters(group2, particle1, particle2, particle3, particle4, c0b, c1b, c2b, c3b, c4b, c5b);
768
769
770
771
772
773
774
775
776
777
778
779
780
781
        return (c0a == c0b && c1a == c1b && c2a == c2b && c3a == c3b && c4a == c4b && c5a == c5b);
    }
private:
    const RBTorsionForce& force;
};

OpenCLCalcRBTorsionForceKernel::~OpenCLCalcRBTorsionForceKernel() {
    if (params != NULL)
        delete params;
    if (indices != NULL)
        delete indices;
}

void OpenCLCalcRBTorsionForceKernel::initialize(const System& system, const RBTorsionForce& force) {
782
783
784
785
    int numContexts = cl.getPlatformData().contexts.size();
    int startIndex = cl.getContextIndex()*force.getNumTorsions()/numContexts;
    int endIndex = (cl.getContextIndex()+1)*force.getNumTorsions()/numContexts;
    numTorsions = endIndex-startIndex;
786
787
    if (numTorsions == 0)
        return;
788
789
    params = new OpenCLArray<mm_float8>(cl, numTorsions, "rbTorsionParams");
    indices = new OpenCLArray<mm_int8>(cl, numTorsions, "rbTorsionIndices");
790
791
792
793
794
795
    vector<int> forceBufferCounter(system.getNumParticles(), 0);
    vector<mm_float8> paramVector(numTorsions);
    vector<mm_int8> indicesVector(numTorsions);
    for (int i = 0; i < numTorsions; i++) {
        int particle1, particle2, particle3, particle4;
        double c0, c1, c2, c3, c4, c5;
796
        force.getTorsionParameters(startIndex+i, particle1, particle2, particle3, particle4, c0, c1, c2, c3, c4, c5);
797
798
799
        paramVector[i] = mm_float8((cl_float) c0, (cl_float) c1, (cl_float) c2, (cl_float) c3, (cl_float) c4, (cl_float) c5, 0.0f, 0.0f);
        indicesVector[i] = mm_int8(particle1, particle2, particle3, particle4,
                forceBufferCounter[particle1]++, forceBufferCounter[particle2]++, forceBufferCounter[particle3]++, forceBufferCounter[particle4]++);
800
801
802
803
804

    }
    params->upload(paramVector);
    indices->upload(indicesVector);
    int maxBuffers = 1;
805
    for (int i = 0; i < (int) forceBufferCounter.size(); i++)
806
        maxBuffers = max(maxBuffers, forceBufferCounter[i]);
807
    cl.addForce(new OpenCLRBTorsionForceInfo(maxBuffers, force));
808
    cl::Program program = cl.createProgram(OpenCLKernelSources::rbTorsionForce);
809
810
811
    kernel = cl::Kernel(program, "calcRBTorsionForce");
}

812
double OpenCLCalcRBTorsionForceKernel::execute(ContextImpl& context, bool includeForces, bool includeEnergy) {
813
    if (numTorsions == 0)
814
        return 0.0;
815
816
817
818
819
820
821
822
823
824
    if (!hasInitializedKernel) {
        hasInitializedKernel = true;
        kernel.setArg<cl_int>(0, cl.getPaddedNumAtoms());
        kernel.setArg<cl_int>(1, numTorsions);
        kernel.setArg<cl::Buffer>(2, cl.getForceBuffers().getDeviceBuffer());
        kernel.setArg<cl::Buffer>(3, cl.getEnergyBuffer().getDeviceBuffer());
        kernel.setArg<cl::Buffer>(4, cl.getPosq().getDeviceBuffer());
        kernel.setArg<cl::Buffer>(5, params->getDeviceBuffer());
        kernel.setArg<cl::Buffer>(6, indices->getDeviceBuffer());
    }
825
    cl.executeKernel(kernel, numTorsions);
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
    return 0.0;
}

class OpenCLCMAPTorsionForceInfo : public OpenCLForceInfo {
public:
    OpenCLCMAPTorsionForceInfo(int requiredBuffers, const CMAPTorsionForce& force) : OpenCLForceInfo(requiredBuffers), force(force) {
    }
    int getNumParticleGroups() {
        return force.getNumTorsions();
    }
    void getParticlesInGroup(int index, std::vector<int>& particles) {
        int map, a1, a2, a3, a4, b1, b2, b3, b4;
        force.getTorsionParameters(index, map, a1, a2, a3, a4, b1, b2, b3, b4);
        particles.resize(8);
        particles[0] = a1;
        particles[1] = a2;
        particles[2] = a3;
        particles[3] = a4;
        particles[4] = b1;
        particles[5] = b2;
        particles[6] = b3;
        particles[7] = b4;
    }
    bool areGroupsIdentical(int group1, int group2) {
        int map1, map2, a1, a2, a3, a4, b1, b2, b3, b4;
        force.getTorsionParameters(group1, map1, a1, a2, a3, a4, b1, b2, b3, b4);
        force.getTorsionParameters(group2, map2, a1, a2, a3, a4, b1, b2, b3, b4);
        return (map1 == map2);
    }
private:
    const CMAPTorsionForce& force;
};

OpenCLCalcCMAPTorsionForceKernel::~OpenCLCalcCMAPTorsionForceKernel() {
    if (coefficients != NULL)
        delete coefficients;
    if (mapPositions != NULL)
        delete mapPositions;
    if (torsionMaps != NULL)
        delete torsionMaps;
    if (torsionIndices != NULL)
        delete torsionIndices;
}

void OpenCLCalcCMAPTorsionForceKernel::initialize(const System& system, const CMAPTorsionForce& force) {
871
872
873
874
    int numContexts = cl.getPlatformData().contexts.size();
    int startIndex = cl.getContextIndex()*force.getNumTorsions()/numContexts;
    int endIndex = (cl.getContextIndex()+1)*force.getNumTorsions()/numContexts;
    numTorsions = endIndex-startIndex;
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
    if (numTorsions == 0)
        return;
    int numMaps = force.getNumMaps();
    vector<mm_float4> coeffVec;
    vector<mm_int2> mapPositionsVec(numMaps);
    vector<double> energy;
    vector<vector<double> > c;
    int currentPosition = 0;
    for (int i = 0; i < numMaps; i++) {
        int size;
        force.getMapParameters(i, size, energy);
        CMAPTorsionForceImpl::calcMapDerivatives(size, energy, c);
        mapPositionsVec[i] = mm_int2(currentPosition, size);
        currentPosition += 4*size*size;
        for (int j = 0; j < size*size; j++) {
890
891
892
893
            coeffVec.push_back(mm_float4((float) c[j][0], (float) c[j][1], (float) c[j][2], (float) c[j][3]));
            coeffVec.push_back(mm_float4((float) c[j][4], (float) c[j][5], (float) c[j][6], (float) c[j][7]));
            coeffVec.push_back(mm_float4((float) c[j][8], (float) c[j][9], (float) c[j][10], (float) c[j][11]));
            coeffVec.push_back(mm_float4((float) c[j][12], (float) c[j][13], (float) c[j][14], (float) c[j][15]));
894
895
896
897
898
899
900
        }
    }
    vector<int> forceBufferCounter(system.getNumParticles(), 0);
    vector<cl_int> torsionMapsVec(numTorsions);
    vector<mm_int16> torsionIndicesVec(numTorsions);
    for (int i = 0; i < numTorsions; i++) {
        mm_int16& ind = torsionIndicesVec[i];
901
        force.getTorsionParameters(startIndex+i, torsionMapsVec[i], ind.s0, ind.s1, ind.s2, ind.s3, ind.s4, ind.s5, ind.s6, ind.s7);
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
        ind.s8 = forceBufferCounter[ind.s0]++;
        ind.s9 = forceBufferCounter[ind.s1]++;
        ind.s10 = forceBufferCounter[ind.s2]++;
        ind.s11 = forceBufferCounter[ind.s3]++;
        ind.s12 = forceBufferCounter[ind.s4]++;
        ind.s13 = forceBufferCounter[ind.s5]++;
        ind.s14 = forceBufferCounter[ind.s6]++;
        ind.s15 = forceBufferCounter[ind.s7]++;
    }
    coefficients = new OpenCLArray<mm_float4>(cl, coeffVec.size(), "cmapTorsionCoefficients");
    mapPositions = new OpenCLArray<mm_int2>(cl, numMaps, "cmapTorsionMapPositions");
    torsionMaps = new OpenCLArray<cl_int>(cl, numTorsions, "cmapTorsionMaps");
    torsionIndices = new OpenCLArray<mm_int16>(cl, numTorsions, "cmapTorsionIndices");
    coefficients->upload(coeffVec);
    mapPositions->upload(mapPositionsVec);
    torsionMaps->upload(torsionMapsVec);
    torsionIndices->upload(torsionIndicesVec);
    int maxBuffers = 1;
    for (int i = 0; i < (int) forceBufferCounter.size(); i++)
        maxBuffers = max(maxBuffers, forceBufferCounter[i]);
    cl.addForce(new OpenCLCMAPTorsionForceInfo(maxBuffers, force));
    cl::Program program = cl.createProgram(OpenCLKernelSources::cmapTorsionForce);
    kernel = cl::Kernel(program, "computeCMAPTorsionForces");
}

927
double OpenCLCalcCMAPTorsionForceKernel::execute(ContextImpl& context, bool includeForces, bool includeEnergy) {
928
    if (numTorsions == 0)
929
        return 0.0;
930
931
932
933
934
935
936
937
938
939
940
941
942
    if (!hasInitializedKernel) {
        hasInitializedKernel = true;
        kernel.setArg<cl_int>(0, cl.getPaddedNumAtoms());
        kernel.setArg<cl_int>(1, numTorsions);
        kernel.setArg<cl::Buffer>(2, cl.getForceBuffers().getDeviceBuffer());
        kernel.setArg<cl::Buffer>(3, cl.getEnergyBuffer().getDeviceBuffer());
        kernel.setArg<cl::Buffer>(4, cl.getPosq().getDeviceBuffer());
        kernel.setArg<cl::Buffer>(5, coefficients->getDeviceBuffer());
        kernel.setArg<cl::Buffer>(6, mapPositions->getDeviceBuffer());
        kernel.setArg<cl::Buffer>(7, torsionIndices->getDeviceBuffer());
        kernel.setArg<cl::Buffer>(8, torsionMaps->getDeviceBuffer());
    }
    cl.executeKernel(kernel, numTorsions);
943
944
945
    return 0.0;
}

946
947
948
949
950
951
952
953
954
955
956
class OpenCLCustomTorsionForceInfo : public OpenCLForceInfo {
public:
    OpenCLCustomTorsionForceInfo(int requiredBuffers, const CustomTorsionForce& force) : OpenCLForceInfo(requiredBuffers), force(force) {
    }
    int getNumParticleGroups() {
        return force.getNumTorsions();
    }
    void getParticlesInGroup(int index, std::vector<int>& particles) {
        int particle1, particle2, particle3, particle4;
        vector<double> parameters;
        force.getTorsionParameters(index, particle1, particle2, particle3, particle4, parameters);
Peter Eastman's avatar
Bug fix  
Peter Eastman committed
957
        particles.resize(4);
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
        particles[0] = particle1;
        particles[1] = particle2;
        particles[2] = particle3;
        particles[3] = particle4;
    }
    bool areGroupsIdentical(int group1, int group2) {
        int particle1, particle2, particle3, particle4;
        vector<double> parameters1, parameters2;
        force.getTorsionParameters(group1, particle1, particle2, particle3, particle4, parameters1);
        force.getTorsionParameters(group2, particle1, particle2, particle3, particle4, parameters2);
        for (int i = 0; i < (int) parameters1.size(); i++)
            if (parameters1[i] != parameters2[i])
                return false;
        return true;
    }
private:
    const CustomTorsionForce& force;
};

OpenCLCalcCustomTorsionForceKernel::~OpenCLCalcCustomTorsionForceKernel() {
    if (params != NULL)
        delete params;
    if (indices != NULL)
        delete indices;
    if (globals != NULL)
        delete globals;
}

void OpenCLCalcCustomTorsionForceKernel::initialize(const System& system, const CustomTorsionForce& force) {
987
988
989
990
    int numContexts = cl.getPlatformData().contexts.size();
    int startIndex = cl.getContextIndex()*force.getNumTorsions()/numContexts;
    int endIndex = (cl.getContextIndex()+1)*force.getNumTorsions()/numContexts;
    numTorsions = endIndex-startIndex;
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
    if (numTorsions == 0)
        return;
    params = new OpenCLParameterSet(cl, force.getNumPerTorsionParameters(), numTorsions, "customTorsionParams");
    indices = new OpenCLArray<mm_int8>(cl, numTorsions, "customTorsionIndices");
    string extraArguments;
    if (force.getNumGlobalParameters() > 0) {
        globals = new OpenCLArray<cl_float>(cl, force.getNumGlobalParameters(), "customTorsionGlobals", false, CL_MEM_READ_ONLY);
        extraArguments += ", __constant float* globals";
    }
    vector<int> forceBufferCounter(system.getNumParticles(), 0);
    vector<vector<cl_float> > paramVector(numTorsions);
    vector<mm_int8> indicesVector(numTorsions);
    for (int i = 0; i < numTorsions; i++) {
        int particle1, particle2, particle3, particle4;
        vector<double> parameters;
1006
        force.getTorsionParameters(startIndex+i, particle1, particle2, particle3, particle4, parameters);
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
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
1056
1057
1058
1059
1060
1061
1062
1063
1064
        paramVector[i].resize(parameters.size());
        for (int j = 0; j < (int) parameters.size(); j++)
            paramVector[i][j] = (cl_float) parameters[j];
        indicesVector[i] = mm_int8(particle1, particle2, particle3, particle4, forceBufferCounter[particle1]++,
                forceBufferCounter[particle2]++, forceBufferCounter[particle3]++, forceBufferCounter[particle4]++);
    }
    params->setParameterValues(paramVector);
    indices->upload(indicesVector);
    int maxBuffers = 1;
    for (int i = 0; i < (int) forceBufferCounter.size(); i++)
        maxBuffers = max(maxBuffers, forceBufferCounter[i]);
    cl.addForce(new OpenCLCustomTorsionForceInfo(maxBuffers, force));

    // Record information for the expressions.

    globalParamNames.resize(force.getNumGlobalParameters());
    globalParamValues.resize(force.getNumGlobalParameters());
    for (int i = 0; i < force.getNumGlobalParameters(); i++) {
        globalParamNames[i] = force.getGlobalParameterName(i);
        globalParamValues[i] = (cl_float) force.getGlobalParameterDefaultValue(i);
    }
    if (globals != NULL)
        globals->upload(globalParamValues);
    Lepton::ParsedExpression energyExpression = Lepton::Parser::parse(force.getEnergyFunction()).optimize();
    Lepton::ParsedExpression forceExpression = energyExpression.differentiate("theta").optimize();
    map<string, Lepton::ParsedExpression> expressions;
    expressions["energy += "] = energyExpression;
    expressions["float dEdAngle = "] = forceExpression;

    // Create the kernels.

    map<string, string> variables;
    variables["theta"] = "theta";
    for (int i = 0; i < force.getNumPerTorsionParameters(); i++) {
        const string& name = force.getPerTorsionParameterName(i);
        variables[name] = "torsionParams"+params->getParameterSuffix(i);
    }
    for (int i = 0; i < force.getNumGlobalParameters(); i++) {
        const string& name = force.getGlobalParameterName(i);
        string value = "globals["+intToString(i)+"]";
        variables[name] = value;
    }
    stringstream compute;
    for (int i = 0; i < (int) params->getBuffers().size(); i++) {
        const OpenCLNonbondedUtilities::ParameterInfo& buffer = params->getBuffers()[i];
        extraArguments += ", __global "+buffer.getType()+"* "+buffer.getName();
        compute<<buffer.getType()<<" torsionParams"<<(i+1)<<" = "<<buffer.getName()<<"[index];\n";
    }
    vector<pair<string, string> > functions;
    compute << OpenCLExpressionUtilities::createExpressions(expressions, variables, functions, "temp", "");
    map<string, string> replacements;
    replacements["COMPUTE_FORCE"] = compute.str();
    replacements["EXTRA_ARGUMENTS"] = extraArguments;
    replacements["M_PI"] = doubleToString(M_PI);
    cl::Program program = cl.createProgram(cl.replaceStrings(OpenCLKernelSources::customTorsionForce, replacements));
    kernel = cl::Kernel(program, "computeCustomTorsionForces");
}

1065
double OpenCLCalcCustomTorsionForceKernel::execute(ContextImpl& context, bool includeForces, bool includeEnergy) {
1066
    if (numTorsions == 0)
1067
        return 0.0;
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
    if (globals != NULL) {
        bool changed = false;
        for (int i = 0; i < (int) globalParamNames.size(); i++) {
            cl_float value = (cl_float) context.getParameter(globalParamNames[i]);
            if (value != globalParamValues[i])
                changed = true;
            globalParamValues[i] = value;
        }
        if (changed)
            globals->upload(globalParamValues);
    }
    if (!hasInitializedKernel) {
        hasInitializedKernel = true;
        kernel.setArg<cl_int>(0, cl.getPaddedNumAtoms());
        kernel.setArg<cl_int>(1, numTorsions);
        kernel.setArg<cl::Buffer>(2, cl.getForceBuffers().getDeviceBuffer());
        kernel.setArg<cl::Buffer>(3, cl.getEnergyBuffer().getDeviceBuffer());
        kernel.setArg<cl::Buffer>(4, cl.getPosq().getDeviceBuffer());
        kernel.setArg<cl::Buffer>(5, indices->getDeviceBuffer());
        int nextIndex = 6;
        if (globals != NULL)
            kernel.setArg<cl::Buffer>(nextIndex++, globals->getDeviceBuffer());
        for (int i = 0; i < (int) params->getBuffers().size(); i++) {
            const OpenCLNonbondedUtilities::ParameterInfo& buffer = params->getBuffers()[i];
1092
            kernel.setArg<cl::Memory>(nextIndex++, buffer.getMemory());
1093
1094
1095
1096
1097
1098
        }
    }
    cl.executeKernel(kernel, numTorsions);
    return 0.0;
}

1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
class OpenCLNonbondedForceInfo : public OpenCLForceInfo {
public:
    OpenCLNonbondedForceInfo(int requiredBuffers, const NonbondedForce& force) : OpenCLForceInfo(requiredBuffers), force(force) {
    }
    bool areParticlesIdentical(int particle1, int particle2) {
        double charge1, charge2, sigma1, sigma2, epsilon1, epsilon2;
        force.getParticleParameters(particle1, charge1, sigma1, epsilon1);
        force.getParticleParameters(particle2, charge2, sigma2, epsilon2);
        return (charge1 == charge2 && sigma1 == sigma2 && epsilon1 == epsilon2);
    }
    int getNumParticleGroups() {
        return force.getNumExceptions();
    }
    void getParticlesInGroup(int index, std::vector<int>& particles) {
        int particle1, particle2;
        double chargeProd, sigma, epsilon;
        force.getExceptionParameters(index, particle1, particle2, chargeProd, sigma, epsilon);
        particles.resize(2);
        particles[0] = particle1;
        particles[1] = particle2;
    }
    bool areGroupsIdentical(int group1, int group2) {
        int particle1, particle2;
        double chargeProd1, chargeProd2, sigma1, sigma2, epsilon1, epsilon2;
        force.getExceptionParameters(group1, particle1, particle2, chargeProd1, sigma1, epsilon1);
        force.getExceptionParameters(group2, particle1, particle2, chargeProd2, sigma2, epsilon2);
        return (chargeProd1 == chargeProd2 && sigma1 == sigma2 && epsilon1 == epsilon2);
    }
private:
    const NonbondedForce& force;
};

1131
1132
1133
OpenCLCalcNonbondedForceKernel::~OpenCLCalcNonbondedForceKernel() {
    if (sigmaEpsilon != NULL)
        delete sigmaEpsilon;
1134
1135
1136
1137
    if (exceptionParams != NULL)
        delete exceptionParams;
    if (exceptionIndices != NULL)
        delete exceptionIndices;
1138
1139
1140
1141
    if (cosSinSums != NULL)
        delete cosSinSums;
    if (pmeGrid != NULL)
        delete pmeGrid;
Peter Eastman's avatar
Peter Eastman committed
1142
1143
    if (pmeGrid2 != NULL)
        delete pmeGrid2;
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
    if (pmeBsplineModuliX != NULL)
        delete pmeBsplineModuliX;
    if (pmeBsplineModuliY != NULL)
        delete pmeBsplineModuliY;
    if (pmeBsplineModuliZ != NULL)
        delete pmeBsplineModuliZ;
    if (pmeBsplineTheta != NULL)
        delete pmeBsplineTheta;
    if (pmeAtomRange != NULL)
        delete pmeAtomRange;
    if (pmeAtomGridIndex != NULL)
        delete pmeAtomGridIndex;
    if (sort != NULL)
        delete sort;
    if (fft != NULL)
        delete fft;
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
}

void OpenCLCalcNonbondedForceKernel::initialize(const System& system, const NonbondedForce& force) {

    // Identify which exceptions are 1-4 interactions.

    vector<pair<int, int> > exclusions;
    vector<int> exceptions;
    for (int i = 0; i < force.getNumExceptions(); i++) {
        int particle1, particle2;
        double chargeProd, sigma, epsilon;
        force.getExceptionParameters(i, particle1, particle2, chargeProd, sigma, epsilon);
        exclusions.push_back(pair<int, int>(particle1, particle2));
        if (chargeProd != 0.0 || epsilon != 0.0)
            exceptions.push_back(i);
    }

    // Initialize nonbonded interactions.

    int numParticles = force.getNumParticles();
    sigmaEpsilon = new OpenCLArray<mm_float2>(cl, numParticles, "sigmaEpsilon");
    OpenCLArray<mm_float4>& posq = cl.getPosq();
    vector<mm_float2> sigmaEpsilonVector(numParticles);
    vector<vector<int> > exclusionList(numParticles);
1184
    double sumSquaredCharges = 0.0;
1185
1186
    bool hasCoulomb = false;
    bool hasLJ = false;
1187
1188
1189
1190
    for (int i = 0; i < numParticles; i++) {
        double charge, sigma, epsilon;
        force.getParticleParameters(i, charge, sigma, epsilon);
        posq[i].w = (float) charge;
1191
        sigmaEpsilonVector[i] = mm_float2((float) (0.5*sigma), (float) (2.0*sqrt(epsilon)));
1192
        exclusionList[i].push_back(i);
1193
        sumSquaredCharges += charge*charge;
1194
1195
1196
1197
        if (charge != 0.0)
            hasCoulomb = true;
        if (epsilon != 0.0)
            hasLJ = true;
1198
1199
1200
1201
1202
1203
1204
1205
1206
    }
    for (int i = 0; i < (int) exclusions.size(); i++) {
        exclusionList[exclusions[i].first].push_back(exclusions[i].second);
        exclusionList[exclusions[i].second].push_back(exclusions[i].first);
    }
    posq.upload();
    sigmaEpsilon->upload(sigmaEpsilonVector);
    bool useCutoff = (force.getNonbondedMethod() != NonbondedForce::NoCutoff);
    bool usePeriodic = (force.getNonbondedMethod() != NonbondedForce::NoCutoff && force.getNonbondedMethod() != NonbondedForce::CutoffNonPeriodic);
1207
    map<string, string> defines;
1208
1209
    defines["HAS_COULOMB"] = (hasCoulomb ? "1" : "0");
    defines["HAS_LENNARD_JONES"] = (hasLJ ? "1" : "0");
1210
    if (useCutoff) {
1211
1212
        // Compute the reaction field constants.

1213
1214
        double reactionFieldK = pow(force.getCutoffDistance(), -3.0)*(force.getReactionFieldDielectric()-1.0)/(2.0*force.getReactionFieldDielectric()+1.0);
        double reactionFieldC = (1.0 / force.getCutoffDistance())*(3.0*force.getReactionFieldDielectric())/(2.0*force.getReactionFieldDielectric()+1.0);
1215
1216
1217
        defines["REACTION_FIELD_K"] = doubleToString(reactionFieldK);
        defines["REACTION_FIELD_C"] = doubleToString(reactionFieldC);
    }
1218
    if (force.getUseDispersionCorrection() && cl.getContextIndex() == 0)
1219
1220
1221
        dispersionCoefficient = NonbondedForceImpl::calcDispersionCorrection(system, force);
    else
        dispersionCoefficient = 0.0;
1222
    double alpha = 0;
1223
1224
1225
1226
1227
1228
1229
1230
    if (force.getNonbondedMethod() == NonbondedForce::Ewald) {
        // Compute the Ewald parameters.

        int kmaxx, kmaxy, kmaxz;
        NonbondedForceImpl::calcEwaldParameters(system, force, alpha, kmaxx, kmaxy, kmaxz);
        defines["EWALD_ALPHA"] = doubleToString(alpha);
        defines["TWO_OVER_SQRT_PI"] = doubleToString(2.0/sqrt(M_PI));
        defines["USE_EWALD"] = "1";
1231
        ewaldSelfEnergy = (cl.getContextIndex() == 0 ? -ONE_4PI_EPS0*alpha*sumSquaredCharges/std::sqrt(M_PI) : 0.0);
1232
1233
1234
1235
1236
1237
1238
1239
1240

        // Create the reciprocal space kernels.

        map<string, string> replacements;
        replacements["NUM_ATOMS"] = intToString(numParticles);
        replacements["KMAX_X"] = intToString(kmaxx);
        replacements["KMAX_Y"] = intToString(kmaxy);
        replacements["KMAX_Z"] = intToString(kmaxz);
        replacements["EXP_COEFFICIENT"] = doubleToString(-1.0/(4.0*alpha*alpha));
1241
        cl::Program program = cl.createProgram(OpenCLKernelSources::ewald, replacements);
1242
1243
1244
1245
        ewaldSumsKernel = cl::Kernel(program, "calculateEwaldCosSinSums");
        ewaldForcesKernel = cl::Kernel(program, "calculateEwaldForces");
        cosSinSums = new OpenCLArray<mm_float2>(cl, (2*kmaxx-1)*(2*kmaxy-1)*(2*kmaxz-1), "cosSinSums");
    }
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
    else if (force.getNonbondedMethod() == NonbondedForce::PME) {
        // Compute the PME parameters.

        int gridSizeX, gridSizeY, gridSizeZ;
        NonbondedForceImpl::calcPMEParameters(system, force, alpha, gridSizeX, gridSizeY, gridSizeZ);
        gridSizeX = OpenCLFFT3D::findLegalDimension(gridSizeX);
        gridSizeY = OpenCLFFT3D::findLegalDimension(gridSizeY);
        gridSizeZ = OpenCLFFT3D::findLegalDimension(gridSizeZ);
        defines["EWALD_ALPHA"] = doubleToString(alpha);
        defines["TWO_OVER_SQRT_PI"] = doubleToString(2.0/sqrt(M_PI));
        defines["USE_EWALD"] = "1";
1257
        ewaldSelfEnergy = (cl.getContextIndex() == 0 ? -ONE_4PI_EPS0*alpha*sumSquaredCharges/std::sqrt(M_PI) : 0.0);
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
        pmeDefines["PME_ORDER"] = intToString(PmeOrder);
        pmeDefines["NUM_ATOMS"] = intToString(numParticles);
        pmeDefines["RECIP_EXP_FACTOR"] = doubleToString(M_PI*M_PI/(alpha*alpha));
        pmeDefines["GRID_SIZE_X"] = intToString(gridSizeX);
        pmeDefines["GRID_SIZE_Y"] = intToString(gridSizeY);
        pmeDefines["GRID_SIZE_Z"] = intToString(gridSizeZ);
        pmeDefines["EPSILON_FACTOR"] = doubleToString(std::sqrt(ONE_4PI_EPS0));

        // Create required data structures.

        pmeGrid = new OpenCLArray<mm_float2>(cl, gridSizeX*gridSizeY*gridSizeZ, "pmeGrid");
Peter Eastman's avatar
Peter Eastman committed
1269
        pmeGrid2 = new OpenCLArray<mm_float2>(cl, gridSizeX*gridSizeY*gridSizeZ, "pmeGrid2");
1270
1271
1272
1273
1274
        pmeBsplineModuliX = new OpenCLArray<cl_float>(cl, gridSizeX, "pmeBsplineModuliX");
        pmeBsplineModuliY = new OpenCLArray<cl_float>(cl, gridSizeY, "pmeBsplineModuliY");
        pmeBsplineModuliZ = new OpenCLArray<cl_float>(cl, gridSizeZ, "pmeBsplineModuliZ");
        pmeBsplineTheta = new OpenCLArray<mm_float4>(cl, PmeOrder*numParticles, "pmeBsplineTheta");
        pmeAtomRange = new OpenCLArray<cl_int>(cl, gridSizeX*gridSizeY*gridSizeZ+1, "pmeAtomRange");
1275
1276
        pmeAtomGridIndex = new OpenCLArray<mm_int2>(cl, numParticles, "pmeAtomGridIndex");
        sort = new OpenCLSort<mm_int2>(cl, cl.getNumAtoms(), "int2", "value.y");
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
        fft = new OpenCLFFT3D(cl, gridSizeX, gridSizeY, gridSizeZ);

        // Initialize the b-spline moduli.

        int maxSize = max(max(gridSizeX, gridSizeY), gridSizeZ);
        vector<double> data(PmeOrder);
        vector<double> ddata(PmeOrder);
        vector<double> bsplines_data(maxSize);
        data[PmeOrder-1] = 0.0;
        data[1] = 0.0;
        data[0] = 1.0;
        for (int i = 3; i < PmeOrder; i++) {
            double div = 1.0/(i-1.0);
            data[i-1] = 0.0;
            for (int j = 1; j < (i-1); j++)
                data[i-j-1] = div*(j*data[i-j-2]+(i-j)*data[i-j-1]);
            data[0] = div*data[0];
        }

        // Differentiate.

        ddata[0] = -data[0];
        for (int i = 1; i < PmeOrder; i++)
            ddata[i] = data[i-1]-data[i];
        double div = 1.0/(PmeOrder-1);
        data[PmeOrder-1] = 0.0;
        for (int i = 1; i < (PmeOrder-1); i++)
            data[PmeOrder-i-1] = div*(i*data[PmeOrder-i-2]+(PmeOrder-i)*data[PmeOrder-i-1]);
        data[0] = div*data[0];
        for (int i = 0; i < maxSize; i++)
            bsplines_data[i] = 0.0;
        for (int i = 1; i <= PmeOrder; i++)
            bsplines_data[i] = data[i-1];

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

        for(int dim = 0; dim < 3; dim++) {
            int ndata = (dim == 0 ? gridSizeX : dim == 1 ? gridSizeY : gridSizeZ);
            vector<cl_float> moduli(ndata);
            for (int i = 0; i < ndata; i++) {
                double sc = 0.0;
                double ss = 0.0;
                for (int j = 0; j < ndata; j++) {
                    double arg = (2.0*M_PI*i*j)/ndata;
                    sc += bsplines_data[j]*cos(arg);
                    ss += bsplines_data[j]*sin(arg);
                }
                moduli[i] = (float) (sc*sc+ss*ss);
            }
            for (int i = 0; i < ndata; i++)
            {
                if (moduli[i] < 1.0e-7)
                    moduli[i] = (moduli[i-1]+moduli[i+1])*0.5f;
            }
            if (dim == 0)
                pmeBsplineModuliX->upload(moduli);
            else if (dim == 1)
                pmeBsplineModuliY->upload(moduli);
            else
                pmeBsplineModuliZ->upload(moduli);
        }
    }
1339
1340
1341
1342
1343
    else
        ewaldSelfEnergy = 0.0;

    // Add the interaction to the default nonbonded kernel.
    
1344
    string source = cl.replaceStrings(OpenCLKernelSources::coulombLennardJones, defines);
1345
    cl.getNonbondedUtilities().addInteraction(useCutoff, usePeriodic, true, force.getCutoffDistance(), exclusionList, source);
Peter Eastman's avatar
Peter Eastman committed
1346
    if (hasLJ)
1347
        cl.getNonbondedUtilities().addParameter(OpenCLNonbondedUtilities::ParameterInfo("sigmaEpsilon", "float", 2, sizeof(cl_float2), sigmaEpsilon->getDeviceBuffer()));
1348

1349
    // Initialize the exceptions.
1350

1351
1352
1353
1354
    int numContexts = cl.getPlatformData().contexts.size();
    int startIndex = cl.getContextIndex()*exceptions.size()/numContexts;
    int endIndex = (cl.getContextIndex()+1)*exceptions.size()/numContexts;
    int numExceptions = endIndex-startIndex;
1355
    int maxBuffers = cl.getNonbondedUtilities().getNumForceBuffers();
1356
1357
1358
1359
1360
1361
    if (numExceptions > 0) {
        exceptionParams = new OpenCLArray<mm_float4>(cl, numExceptions, "exceptionParams");
        exceptionIndices = new OpenCLArray<mm_int4>(cl, numExceptions, "exceptionIndices");
        vector<mm_float4> exceptionParamsVector(numExceptions);
        vector<mm_int4> exceptionIndicesVector(numExceptions);
        vector<int> forceBufferCounter(system.getNumParticles(), 0);
1362
        for (int i = 0; i < numExceptions; i++) {
1363
1364
            int particle1, particle2;
            double chargeProd, sigma, epsilon;
1365
            force.getExceptionParameters(exceptions[startIndex+i], particle1, particle2, chargeProd, sigma, epsilon);
1366
1367
            exceptionParamsVector[i] = mm_float4((float) (ONE_4PI_EPS0*chargeProd), (float) sigma, (float) (4.0*epsilon), 0.0f);
            exceptionIndicesVector[i] = mm_int4(particle1, particle2, forceBufferCounter[particle1]++, forceBufferCounter[particle2]++);
1368
        }
1369
1370
        exceptionParams->upload(exceptionParamsVector);
        exceptionIndices->upload(exceptionIndicesVector);
1371
        for (int i = 0; i < (int) forceBufferCounter.size(); i++)
1372
            maxBuffers = max(maxBuffers, forceBufferCounter[i]);
1373
    }
1374
    cl.addForce(new OpenCLNonbondedForceInfo(maxBuffers, force));
1375
    defines.clear();
Peter Eastman's avatar
Bug fix  
Peter Eastman committed
1376
    defines["NUM_ATOMS"] = intToString(cl.getPaddedNumAtoms());
1377
    defines["NUM_EXCEPTIONS"] = intToString(numExceptions);
1378
    cl::Program program = cl.createProgram(OpenCLKernelSources::nonbondedExceptions, defines);
1379
    exceptionsKernel = cl::Kernel(program, "computeNonbondedExceptions");
1380
1381
}

1382
double OpenCLCalcNonbondedForceKernel::execute(ContextImpl& context, bool includeForces, bool includeEnergy) {
1383
    bool deviceIsCpu = (cl.getDevice().getInfo<CL_DEVICE_TYPE>() == CL_DEVICE_TYPE_CPU);
1384
1385
1386
    if (!hasInitializedKernel) {
        hasInitializedKernel = true;
        if (exceptionIndices != NULL) {
1387
1388
1389
1390
1391
            exceptionsKernel.setArg<cl::Buffer>(0, cl.getForceBuffers().getDeviceBuffer());
            exceptionsKernel.setArg<cl::Buffer>(1, cl.getEnergyBuffer().getDeviceBuffer());
            exceptionsKernel.setArg<cl::Buffer>(2, cl.getPosq().getDeviceBuffer());
            exceptionsKernel.setArg<cl::Buffer>(3, exceptionParams->getDeviceBuffer());
            exceptionsKernel.setArg<cl::Buffer>(4, exceptionIndices->getDeviceBuffer());
1392
1393
1394
1395
1396
1397
1398
1399
1400
        }
        if (cosSinSums != NULL) {
            ewaldSumsKernel.setArg<cl::Buffer>(0, cl.getEnergyBuffer().getDeviceBuffer());
            ewaldSumsKernel.setArg<cl::Buffer>(1, cl.getPosq().getDeviceBuffer());
            ewaldSumsKernel.setArg<cl::Buffer>(2, cosSinSums->getDeviceBuffer());
            ewaldForcesKernel.setArg<cl::Buffer>(0, cl.getForceBuffers().getDeviceBuffer());
            ewaldForcesKernel.setArg<cl::Buffer>(1, cl.getPosq().getDeviceBuffer());
            ewaldForcesKernel.setArg<cl::Buffer>(2, cosSinSums->getDeviceBuffer());
        }
1401
        if (pmeGrid != NULL) {
1402
1403
            string file = (deviceIsCpu ? OpenCLKernelSources::pme_cpu : OpenCLKernelSources::pme);
            cl::Program program = cl.createProgram(file, pmeDefines);
1404
            pmeUpdateBsplinesKernel = cl::Kernel(program, "updateBsplines");
1405
            pmeAtomRangeKernel = cl::Kernel(program, "findAtomRangeForGrid");
1406
1407
1408
1409
1410
            pmeSpreadChargeKernel = cl::Kernel(program, "gridSpreadCharge");
            pmeConvolutionKernel = cl::Kernel(program, "reciprocalConvolution");
            pmeInterpolateForceKernel = cl::Kernel(program, "gridInterpolateForce");
            pmeUpdateBsplinesKernel.setArg<cl::Buffer>(0, cl.getPosq().getDeviceBuffer());
            pmeUpdateBsplinesKernel.setArg<cl::Buffer>(1, pmeBsplineTheta->getDeviceBuffer());
1411
1412
            pmeUpdateBsplinesKernel.setArg(2, OpenCLContext::ThreadBlockSize*PmeOrder*sizeof(mm_float4), NULL);
            pmeUpdateBsplinesKernel.setArg<cl::Buffer>(3, pmeAtomGridIndex->getDeviceBuffer());
1413
1414
1415
            pmeAtomRangeKernel.setArg<cl::Buffer>(0, pmeAtomGridIndex->getDeviceBuffer());
            pmeAtomRangeKernel.setArg<cl::Buffer>(1, pmeAtomRange->getDeviceBuffer());
            pmeAtomRangeKernel.setArg<cl::Buffer>(2, cl.getPosq().getDeviceBuffer());
1416
1417
1418
1419
1420
            pmeSpreadChargeKernel.setArg<cl::Buffer>(0, cl.getPosq().getDeviceBuffer());
            pmeSpreadChargeKernel.setArg<cl::Buffer>(1, pmeAtomGridIndex->getDeviceBuffer());
            pmeSpreadChargeKernel.setArg<cl::Buffer>(2, pmeAtomRange->getDeviceBuffer());
            pmeSpreadChargeKernel.setArg<cl::Buffer>(3, pmeGrid->getDeviceBuffer());
            pmeSpreadChargeKernel.setArg<cl::Buffer>(4, pmeBsplineTheta->getDeviceBuffer());
Peter Eastman's avatar
Peter Eastman committed
1421
            pmeConvolutionKernel.setArg<cl::Buffer>(0, pmeGrid2->getDeviceBuffer());
1422
1423
1424
1425
1426
1427
            pmeConvolutionKernel.setArg<cl::Buffer>(1, cl.getEnergyBuffer().getDeviceBuffer());
            pmeConvolutionKernel.setArg<cl::Buffer>(2, pmeBsplineModuliX->getDeviceBuffer());
            pmeConvolutionKernel.setArg<cl::Buffer>(3, pmeBsplineModuliY->getDeviceBuffer());
            pmeConvolutionKernel.setArg<cl::Buffer>(4, pmeBsplineModuliZ->getDeviceBuffer());
            pmeInterpolateForceKernel.setArg<cl::Buffer>(0, cl.getPosq().getDeviceBuffer());
            pmeInterpolateForceKernel.setArg<cl::Buffer>(1, cl.getForceBuffers().getDeviceBuffer());
1428
1429
            pmeInterpolateForceKernel.setArg<cl::Buffer>(2, pmeGrid->getDeviceBuffer());
            pmeInterpolateForceKernel.setArg(5, 2*128*PmeOrder*sizeof(mm_float4), NULL);
1430
1431
1432
1433
            if (cl.getSupports64BitGlobalAtomics()) {
                pmeFinishSpreadChargeKernel = cl::Kernel(program, "finishSpreadCharge");
                pmeFinishSpreadChargeKernel.setArg<cl::Buffer>(0, pmeGrid->getDeviceBuffer());
            }
1434
       }
1435
    }
1436
1437
    if (exceptionIndices != NULL)
        cl.executeKernel(exceptionsKernel, exceptionIndices->getSize());
1438
    if (cosSinSums != NULL && cl.getContextIndex() == 0) {
1439
1440
        mm_float4 boxSize = cl.getPeriodicBoxSize();
        mm_float4 recipBoxSize = mm_float4((float) (2*M_PI/boxSize.x), (float) (2*M_PI/boxSize.y), (float) (2*M_PI/boxSize.z), 0);
1441
        float recipCoefficient = (float) (ONE_4PI_EPS0*4*M_PI/(boxSize.x*boxSize.y*boxSize.z));
1442
1443
        ewaldSumsKernel.setArg<mm_float4>(3, recipBoxSize);
        ewaldSumsKernel.setArg<cl_float>(4, recipCoefficient);
1444
        cl.executeKernel(ewaldSumsKernel, cosSinSums->getSize());
1445
1446
        ewaldForcesKernel.setArg<mm_float4>(3, recipBoxSize);
        ewaldForcesKernel.setArg<cl_float>(4, recipCoefficient);
1447
1448
        cl.executeKernel(ewaldForcesKernel, cl.getNumAtoms());
    }
1449
    if (pmeGrid != NULL && cl.getContextIndex() == 0) {
1450
1451
        mm_float4 boxSize = cl.getPeriodicBoxSize();
        mm_float4 invBoxSize = cl.getInvPeriodicBoxSize();
1452
1453
        pmeUpdateBsplinesKernel.setArg<mm_float4>(4, boxSize);
        pmeUpdateBsplinesKernel.setArg<mm_float4>(5, invBoxSize);
1454
        cl.executeKernel(pmeUpdateBsplinesKernel, cl.getNumAtoms());
1455
1456
1457
1458
1459
1460
1461
1462
1463
1464
        if (deviceIsCpu) {
            pmeSpreadChargeKernel.setArg<mm_float4>(5, boxSize);
            pmeSpreadChargeKernel.setArg<mm_float4>(6, invBoxSize);
            cl.executeKernel(pmeSpreadChargeKernel, 2*cl.getDevice().getInfo<CL_DEVICE_MAX_COMPUTE_UNITS>(), 1);
        }
        else {
            sort->sort(*pmeAtomGridIndex);
            pmeAtomRangeKernel.setArg<mm_float4>(3, boxSize);
            pmeAtomRangeKernel.setArg<mm_float4>(4, invBoxSize);
            cl.executeKernel(pmeAtomRangeKernel, cl.getNumAtoms());
1465
1466
1467
1468
1469
1470
1471
1472
1473
            if (cl.getSupports64BitGlobalAtomics()) {
                cl.clearBuffer(pmeGrid->getDeviceBuffer(), pmeGrid->getSize()*2);
                pmeSpreadChargeKernel.setArg<mm_float4>(5, boxSize);
                pmeSpreadChargeKernel.setArg<mm_float4>(6, invBoxSize);
                cl.executeKernel(pmeSpreadChargeKernel, cl.getNumAtoms(), PmeOrder*PmeOrder*PmeOrder);
                cl.executeKernel(pmeFinishSpreadChargeKernel, pmeGrid->getSize());
            }
            else
                cl.executeKernel(pmeSpreadChargeKernel, cl.getNumAtoms());
1474
        }
Peter Eastman's avatar
Peter Eastman committed
1475
        fft->execFFT(*pmeGrid, *pmeGrid2, true);
1476
1477
        pmeConvolutionKernel.setArg<mm_float4>(5, invBoxSize);
        pmeConvolutionKernel.setArg<cl_float>(6, (float) (1.0/(M_PI*boxSize.x*boxSize.y*boxSize.z)));
1478
        cl.executeKernel(pmeConvolutionKernel, cl.getNumAtoms());
Peter Eastman's avatar
Peter Eastman committed
1479
        fft->execFFT(*pmeGrid2, *pmeGrid, false);
1480
1481
1482
        pmeInterpolateForceKernel.setArg<mm_float4>(3, boxSize);
        pmeInterpolateForceKernel.setArg<mm_float4>(4, invBoxSize);
        cl.executeKernel(pmeInterpolateForceKernel, cl.getNumAtoms(), 128);
1483
    }
1484
1485
1486
1487
1488
1489
    double energy = ewaldSelfEnergy;
    if (dispersionCoefficient != 0.0) {
        mm_float4 boxSize = cl.getPeriodicBoxSize();
        energy += dispersionCoefficient/(boxSize.x*boxSize.y*boxSize.z);
    }
    return energy;
1490
1491
}

1492
1493
1494
1495
1496
1497
1498
1499
1500
class OpenCLCustomNonbondedForceInfo : public OpenCLForceInfo {
public:
    OpenCLCustomNonbondedForceInfo(int requiredBuffers, const CustomNonbondedForce& force) : OpenCLForceInfo(requiredBuffers), force(force) {
    }
    bool areParticlesIdentical(int particle1, int particle2) {
        vector<double> params1;
        vector<double> params2;
        force.getParticleParameters(particle1, params1);
        force.getParticleParameters(particle2, params2);
1501
        for (int i = 0; i < (int) params1.size(); i++)
1502
1503
1504
1505
1506
            if (params1[i] != params2[i])
                return false;
        return true;
    }
    int getNumParticleGroups() {
1507
        return force.getNumExclusions();
1508
1509
1510
    }
    void getParticlesInGroup(int index, std::vector<int>& particles) {
        int particle1, particle2;
1511
        force.getExclusionParticles(index, particle1, particle2);
1512
1513
1514
1515
1516
1517
1518
1519
1520
1521
1522
1523
1524
1525
1526
1527
        particles.resize(2);
        particles[0] = particle1;
        particles[1] = particle2;
    }
    bool areGroupsIdentical(int group1, int group2) {
        return true;
    }
private:
    const CustomNonbondedForce& force;
};

OpenCLCalcCustomNonbondedForceKernel::~OpenCLCalcCustomNonbondedForceKernel() {
    if (params != NULL)
        delete params;
    if (globals != NULL)
        delete globals;
1528
1529
1530
1531
    if (tabulatedFunctionParams != NULL)
        delete tabulatedFunctionParams;
    for (int i = 0; i < (int) tabulatedFunctions.size(); i++)
        delete tabulatedFunctions[i];
1532
1533
1534
1535
1536
1537
1538
1539
1540
1541
1542
}

void OpenCLCalcCustomNonbondedForceKernel::initialize(const System& system, const CustomNonbondedForce& force) {
    int forceIndex;
    for (forceIndex = 0; forceIndex < system.getNumForces() && &system.getForce(forceIndex) != &force; ++forceIndex)
        ;
    string prefix = "custom"+intToString(forceIndex)+"_";

    // Record parameters and exclusions.

    int numParticles = force.getNumParticles();
1543
    params = new OpenCLParameterSet(cl, force.getNumPerParticleParameters(), numParticles, "customNonbondedParameters");
1544
    if (force.getNumGlobalParameters() > 0)
1545
        globals = new OpenCLArray<cl_float>(cl, force.getNumGlobalParameters(), "customNonbondedGlobals", false, CL_MEM_READ_ONLY);
1546
    vector<vector<cl_float> > paramVector(numParticles);
1547
1548
1549
1550
    vector<vector<int> > exclusionList(numParticles);
    for (int i = 0; i < numParticles; i++) {
        vector<double> parameters;
        force.getParticleParameters(i, parameters);
1551
        paramVector[i].resize(parameters.size());
1552
        for (int j = 0; j < (int) parameters.size(); j++)
1553
            paramVector[i][j] = (cl_float) parameters[j];
1554
1555
        exclusionList[i].push_back(i);
    }
1556
1557
1558
1559
1560
    for (int i = 0; i < force.getNumExclusions(); i++) {
        int particle1, particle2;
        force.getExclusionParticles(i, particle1, particle2);
        exclusionList[particle1].push_back(particle2);
        exclusionList[particle2].push_back(particle1);
1561
    }
1562
    params->setParameterValues(paramVector);
1563
1564
1565

    // Record the tabulated functions.

1566
    OpenCLExpressionUtilities::FunctionPlaceholder fp;
1567
1568
1569
    map<string, Lepton::CustomFunction*> functions;
    vector<pair<string, string> > functionDefinitions;
    vector<mm_float4> tabulatedFunctionParamsVec(force.getNumFunctions());
1570
1571
1572
1573
    for (int i = 0; i < force.getNumFunctions(); i++) {
        string name;
        vector<double> values;
        double min, max;
1574
        force.getFunctionParameters(i, name, values, min, max);
1575
1576
        string arrayName = prefix+"table"+intToString(i);
        functionDefinitions.push_back(make_pair(name, arrayName));
1577
        functions[name] = &fp;
1578
        tabulatedFunctionParamsVec[i] = mm_float4((float) min, (float) max, (float) ((values.size()-1)/(max-min)), (float) values.size()-2);
1579
        vector<mm_float4> f = OpenCLExpressionUtilities::computeFunctionCoefficients(values, min, max);
1580
1581
        tabulatedFunctions.push_back(new OpenCLArray<mm_float4>(cl, values.size()-1, "TabulatedFunction"));
        tabulatedFunctions[tabulatedFunctions.size()-1]->upload(f);
1582
        cl.getNonbondedUtilities().addArgument(OpenCLNonbondedUtilities::ParameterInfo(arrayName, "float", 4, sizeof(cl_float4), tabulatedFunctions[tabulatedFunctions.size()-1]->getDeviceBuffer()));
1583
1584
1585
1586
    }
    if (force.getNumFunctions() > 0) {
        tabulatedFunctionParams = new OpenCLArray<mm_float4>(cl, tabulatedFunctionParamsVec.size(), "tabulatedFunctionParameters", false, CL_MEM_READ_ONLY);
        tabulatedFunctionParams->upload(tabulatedFunctionParamsVec);
1587
        cl.getNonbondedUtilities().addArgument(OpenCLNonbondedUtilities::ParameterInfo(prefix+"functionParams", "float", 4, sizeof(cl_float4), tabulatedFunctionParams->getDeviceBuffer()));
1588
1589
1590
1591
1592
1593
1594
1595
1596
1597
1598
1599
1600
1601
    }

    // Record information for the expressions.

    globalParamNames.resize(force.getNumGlobalParameters());
    globalParamValues.resize(force.getNumGlobalParameters());
    for (int i = 0; i < force.getNumGlobalParameters(); i++) {
        globalParamNames[i] = force.getGlobalParameterName(i);
        globalParamValues[i] = (cl_float) force.getGlobalParameterDefaultValue(i);
    }
    if (globals != NULL)
        globals->upload(globalParamValues);
    bool useCutoff = (force.getNonbondedMethod() != CustomNonbondedForce::NoCutoff);
    bool usePeriodic = (force.getNonbondedMethod() != CustomNonbondedForce::NoCutoff && force.getNonbondedMethod() != CustomNonbondedForce::CutoffNonPeriodic);
1602
    Lepton::ParsedExpression energyExpression = Lepton::Parser::parse(force.getEnergyFunction(), functions).optimize();
1603
    Lepton::ParsedExpression forceExpression = energyExpression.differentiate("r").optimize();
1604
1605
1606
    map<string, Lepton::ParsedExpression> forceExpressions;
    forceExpressions["tempEnergy += "] = energyExpression;
    forceExpressions["tempForce -= "] = forceExpression;
1607
1608
1609

    // Create the kernels.

1610
1611
1612
1613
    map<string, string> variables;
    variables["r"] = "r";
    for (int i = 0; i < force.getNumPerParticleParameters(); i++) {
        const string& name = force.getPerParticleParameterName(i);
1614
1615
        variables[name+"1"] = prefix+"params"+params->getParameterSuffix(i, "1");
        variables[name+"2"] = prefix+"params"+params->getParameterSuffix(i, "2");
1616
1617
1618
1619
    }
    for (int i = 0; i < force.getNumGlobalParameters(); i++) {
        const string& name = force.getGlobalParameterName(i);
        string value = "globals["+intToString(i)+"]";
1620
        variables[name] = prefix+value;
1621
    }
1622
    stringstream compute;
1623
    compute << OpenCLExpressionUtilities::createExpressions(forceExpressions, variables, functionDefinitions, prefix+"temp", prefix+"functionParams");
1624
1625
    map<string, string> replacements;
    replacements["COMPUTE_FORCE"] = compute.str();
1626
    string source = cl.replaceStrings(OpenCLKernelSources::customNonbonded, replacements);
1627
    cl.getNonbondedUtilities().addInteraction(useCutoff, usePeriodic, true, force.getCutoffDistance(), exclusionList, source);
1628
1629
    for (int i = 0; i < (int) params->getBuffers().size(); i++) {
        const OpenCLNonbondedUtilities::ParameterInfo& buffer = params->getBuffers()[i];
1630
        cl.getNonbondedUtilities().addParameter(OpenCLNonbondedUtilities::ParameterInfo(prefix+"params"+intToString(i+1), buffer.getComponentType(), buffer.getNumComponents(), buffer.getSize(), buffer.getMemory()));
1631
    }
1632
1633
    if (globals != NULL) {
        globals->upload(globalParamValues);
1634
        cl.getNonbondedUtilities().addArgument(OpenCLNonbondedUtilities::ParameterInfo(prefix+"globals", "float", 1, sizeof(cl_float), globals->getDeviceBuffer()));
1635
    }
1636
    cl.addForce(new OpenCLCustomNonbondedForceInfo(cl.getNonbondedUtilities().getNumForceBuffers(), force));
1637
1638
}

1639
double OpenCLCalcCustomNonbondedForceKernel::execute(ContextImpl& context, bool includeForces, bool includeEnergy) {
1640
1641
    if (globals != NULL) {
        bool changed = false;
1642
        for (int i = 0; i < (int) globalParamNames.size(); i++) {
1643
1644
1645
1646
1647
1648
1649
1650
            cl_float value = (cl_float) context.getParameter(globalParamNames[i]);
            if (value != globalParamValues[i])
                changed = true;
            globalParamValues[i] = value;
        }
        if (changed)
            globals->upload(globalParamValues);
    }
1651
1652
    return 0.0;
}
Peter Eastman's avatar
Peter Eastman committed
1653
1654
1655
1656
1657
1658
1659
1660
1661
1662
1663
1664
1665
1666
1667

class OpenCLGBSAOBCForceInfo : public OpenCLForceInfo {
public:
    OpenCLGBSAOBCForceInfo(int requiredBuffers, const GBSAOBCForce& force) : OpenCLForceInfo(requiredBuffers), 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 GBSAOBCForce& force;
};

1668
1669
1670
1671
1672
OpenCLCalcGBSAOBCForceKernel::~OpenCLCalcGBSAOBCForceKernel() {
    if (params != NULL)
        delete params;
    if (bornSum != NULL)
        delete bornSum;
1673
1674
    if (longBornSum != NULL)
        delete longBornSum;
1675
1676
1677
1678
    if (bornRadii != NULL)
        delete bornRadii;
    if (bornForce != NULL)
        delete bornForce;
1679
1680
    if (longBornForce != NULL)
        delete longBornForce;
1681
1682
1683
1684
1685
    if (obcChain != NULL)
        delete obcChain;
}

void OpenCLCalcGBSAOBCForceKernel::initialize(const System& system, const GBSAOBCForce& force) {
1686
1687
    if (cl.getPlatformData().contexts.size() > 1)
        throw OpenMMException("GBSAOBCForce does not support using multiple OpenCL devices");
1688
    OpenCLNonbondedUtilities& nb = cl.getNonbondedUtilities();
1689
1690
1691
    params = new OpenCLArray<mm_float2>(cl, cl.getPaddedNumAtoms(), "gbsaObcParams");
    bornRadii = new OpenCLArray<cl_float>(cl, cl.getPaddedNumAtoms(), "bornRadii");
    obcChain = new OpenCLArray<cl_float>(cl, cl.getPaddedNumAtoms(), "obcChain");
1692
1693
1694
1695
1696
1697
1698
1699
1700
1701
1702
1703
1704
    if (cl.getSupports64BitGlobalAtomics()) {
        longBornSum = new OpenCLArray<cl_long>(cl, cl.getPaddedNumAtoms(), "longBornSum");
        longBornForce = new OpenCLArray<cl_long>(cl, cl.getPaddedNumAtoms(), "longBornForce");
        bornForce = new OpenCLArray<cl_float>(cl, cl.getPaddedNumAtoms(), "bornForce");
        cl.addAutoclearBuffer(longBornSum->getDeviceBuffer(), 2*longBornSum->getSize());
        cl.addAutoclearBuffer(longBornForce->getDeviceBuffer(), 2*longBornForce->getSize());
    }
    else {
        bornSum = new OpenCLArray<cl_float>(cl, cl.getPaddedNumAtoms()*nb.getNumForceBuffers(), "bornSum");
        bornForce = new OpenCLArray<cl_float>(cl, cl.getPaddedNumAtoms()*nb.getNumForceBuffers(), "bornForce");
        cl.addAutoclearBuffer(bornSum->getDeviceBuffer(), bornSum->getSize());
        cl.addAutoclearBuffer(bornForce->getDeviceBuffer(), bornForce->getSize());
    }
1705
1706
1707
1708
1709
1710
1711
1712
    OpenCLArray<mm_float4>& posq = cl.getPosq();
    int numParticles = force.getNumParticles();
    vector<mm_float2> paramsVector(numParticles);
    const double dielectricOffset = 0.009;
    for (int i = 0; i < numParticles; i++) {
        double charge, radius, scalingFactor;
        force.getParticleParameters(i, charge, radius, scalingFactor);
        radius -= dielectricOffset;
1713
        paramsVector[i] = mm_float2((float) radius, (float) (scalingFactor*radius));
1714
1715
1716
1717
        posq[i].w = (float) charge;
    }
    posq.upload();
    params->upload(paramsVector);
1718
    prefactor = -ONE_4PI_EPS0*((1.0/force.getSoluteDielectric())-(1.0/force.getSolventDielectric()));
1719
1720
    bool useCutoff = (force.getNonbondedMethod() != GBSAOBCForce::NoCutoff);
    bool usePeriodic = (force.getNonbondedMethod() != GBSAOBCForce::NoCutoff && force.getNonbondedMethod() != GBSAOBCForce::CutoffNonPeriodic);
1721
    string source = OpenCLKernelSources::gbsaObc2;
1722
    nb.addInteraction(useCutoff, usePeriodic, false, force.getCutoffDistance(), vector<vector<int> >(), source);
1723
1724
    nb.addParameter(OpenCLNonbondedUtilities::ParameterInfo("obcParams", "float", 2, sizeof(cl_float2), params->getDeviceBuffer()));;
    nb.addParameter(OpenCLNonbondedUtilities::ParameterInfo("bornForce", "float", 1, sizeof(cl_float), bornForce->getDeviceBuffer()));;
Peter Eastman's avatar
Peter Eastman committed
1725
    cl.addForce(new OpenCLGBSAOBCForceInfo(nb.getNumForceBuffers(), force));
1726
1727
}

1728
double OpenCLCalcGBSAOBCForceKernel::execute(ContextImpl& context, bool includeForces, bool includeEnergy) {
1729
    OpenCLNonbondedUtilities& nb = cl.getNonbondedUtilities();
1730
    bool deviceIsCpu = (cl.getDevice().getInfo<CL_DEVICE_TYPE>() == CL_DEVICE_TYPE_CPU);
1731
1732
1733
1734
    if (!hasCreatedKernels) {
        // These Kernels cannot be created in initialize(), because the OpenCLNonbondedUtilities has not been initialized yet then.

        hasCreatedKernels = true;
1735
        maxTiles = (nb.getUseCutoff() ? nb.getInteractingTiles().getSize() : 0);
1736
1737
1738
1739
1740
1741
1742
        map<string, string> defines;
        if (nb.getForceBufferPerAtomBlock())
            defines["USE_OUTPUT_BUFFER_PER_BLOCK"] = "1";
        if (nb.getUseCutoff())
            defines["USE_CUTOFF"] = "1";
        if (nb.getUsePeriodic())
            defines["USE_PERIODIC"] = "1";
1743
1744
1745
1746
        defines["CUTOFF_SQUARED"] = doubleToString(nb.getCutoffDistance()*nb.getCutoffDistance());
        defines["PREFACTOR"] = doubleToString(prefactor);
        defines["NUM_ATOMS"] = intToString(cl.getNumAtoms());
        defines["PADDED_NUM_ATOMS"] = intToString(cl.getPaddedNumAtoms());
1747
        defines["NUM_BLOCKS"] = OpenCLExpressionUtilities::intToString(cl.getNumAtomBlocks());
1748
1749
        if (cl.getSIMDWidth() == 32)
            defines["WARPS_PER_GROUP"] = OpenCLExpressionUtilities::intToString(cl.getNonbondedUtilities().getForceThreadBlockSize()/OpenCLContext::TileSize);
1750
1751
1752
1753
1754
1755
1756
        string file;
        if (deviceIsCpu)
            file = OpenCLKernelSources::gbsaObc_cpu;
        else if (cl.getSIMDWidth() == 32)
            file = OpenCLKernelSources::gbsaObc_nvidia;
        else
            file = OpenCLKernelSources::gbsaObc_default;
1757
        cl::Program program = cl.createProgram(file, defines);
1758
        bool useLong = (cl.getSupports64BitGlobalAtomics() && !deviceIsCpu);
1759
        int index = 0;
1760
        computeBornSumKernel = cl::Kernel(program, "computeBornSum");
1761
        computeBornSumKernel.setArg<cl::Buffer>(index++, (useLong ? longBornSum->getDeviceBuffer() : bornSum->getDeviceBuffer()));
1762
1763
        computeBornSumKernel.setArg<cl::Buffer>(index++, cl.getPosq().getDeviceBuffer());
        computeBornSumKernel.setArg<cl::Buffer>(index++, params->getDeviceBuffer());
1764
1765
        computeBornSumKernel.setArg(index++, (deviceIsCpu ? OpenCLContext::TileSize : nb.getForceThreadBlockSize())*13*sizeof(cl_float), NULL);
        computeBornSumKernel.setArg(index++, (deviceIsCpu ? 1 : nb.getForceThreadBlockSize())*sizeof(cl_float), NULL);
1766
        if (nb.getUseCutoff()) {
1767
1768
            computeBornSumKernel.setArg<cl::Buffer>(index++, nb.getInteractingTiles().getDeviceBuffer());
            computeBornSumKernel.setArg<cl::Buffer>(index++, nb.getInteractionCount().getDeviceBuffer());
1769
1770
            index += 2; // The periodic box size arguments are set when the kernel is executed.
            computeBornSumKernel.setArg<cl_uint>(index++, maxTiles);
1771
            if (cl.getSIMDWidth() == 32 || deviceIsCpu)
1772
                computeBornSumKernel.setArg<cl::Buffer>(index++, nb.getInteractionFlags().getDeviceBuffer());
1773
        }
1774
1775
        else
            computeBornSumKernel.setArg<cl_uint>(index++, cl.getNumAtomBlocks()*(cl.getNumAtomBlocks()+1)/2);
1776
        force1Kernel = cl::Kernel(program, "computeGBSAForce1");
1777
        index = 0;
1778
1779
        force1Kernel.setArg<cl::Buffer>(index++, (useLong ? cl.getLongForceBuffer().getDeviceBuffer() : cl.getForceBuffers().getDeviceBuffer()));
        force1Kernel.setArg<cl::Buffer>(index++, (useLong ? longBornForce->getDeviceBuffer() : bornForce->getDeviceBuffer()));
1780
1781
1782
        force1Kernel.setArg<cl::Buffer>(index++, cl.getEnergyBuffer().getDeviceBuffer());
        force1Kernel.setArg<cl::Buffer>(index++, cl.getPosq().getDeviceBuffer());
        force1Kernel.setArg<cl::Buffer>(index++, bornRadii->getDeviceBuffer());
1783
1784
        force1Kernel.setArg(index++, (deviceIsCpu ? OpenCLContext::TileSize : nb.getForceThreadBlockSize())*13*sizeof(cl_float), NULL);
        force1Kernel.setArg(index++, (deviceIsCpu ? 1 : nb.getForceThreadBlockSize())*sizeof(mm_float4), NULL);
1785
        if (nb.getUseCutoff()) {
1786
1787
            force1Kernel.setArg<cl::Buffer>(index++, nb.getInteractingTiles().getDeviceBuffer());
            force1Kernel.setArg<cl::Buffer>(index++, nb.getInteractionCount().getDeviceBuffer());
1788
1789
            index += 2; // The periodic box size arguments are set when the kernel is executed.
            force1Kernel.setArg<cl_uint>(index++, maxTiles);
1790
            if (cl.getSIMDWidth() == 32 || deviceIsCpu)
1791
                force1Kernel.setArg<cl::Buffer>(index++, nb.getInteractionFlags().getDeviceBuffer());
1792
        }
1793
1794
        else
            force1Kernel.setArg<cl_uint>(index++, cl.getNumAtomBlocks()*(cl.getNumAtomBlocks()+1)/2);
1795
        program = cl.createProgram(OpenCLKernelSources::gbsaObcReductions, defines);
1796
1797
        reduceBornSumKernel = cl::Kernel(program, "reduceBornSum");
        reduceBornSumKernel.setArg<cl_int>(0, cl.getPaddedNumAtoms());
Peter Eastman's avatar
Peter Eastman committed
1798
        reduceBornSumKernel.setArg<cl_int>(1, nb.getNumForceBuffers());
1799
1800
1801
        reduceBornSumKernel.setArg<cl_float>(2, 1.0f);
        reduceBornSumKernel.setArg<cl_float>(3, 0.8f);
        reduceBornSumKernel.setArg<cl_float>(4, 4.85f);
1802
        reduceBornSumKernel.setArg<cl::Buffer>(5, (useLong ? longBornSum->getDeviceBuffer() : bornSum->getDeviceBuffer()));
1803
1804
1805
        reduceBornSumKernel.setArg<cl::Buffer>(6, params->getDeviceBuffer());
        reduceBornSumKernel.setArg<cl::Buffer>(7, bornRadii->getDeviceBuffer());
        reduceBornSumKernel.setArg<cl::Buffer>(8, obcChain->getDeviceBuffer());
1806
        reduceBornForceKernel = cl::Kernel(program, "reduceBornForce");
1807
1808
1809
1810
1811
1812
1813
1814
1815
1816
        index = 0;
        reduceBornForceKernel.setArg<cl_int>(index++, cl.getPaddedNumAtoms());
        reduceBornForceKernel.setArg<cl_int>(index++, nb.getNumForceBuffers());
        reduceBornForceKernel.setArg<cl::Buffer>(index++, bornForce->getDeviceBuffer());
        if (useLong)
            reduceBornForceKernel.setArg<cl::Buffer>(index++, longBornForce->getDeviceBuffer());
        reduceBornForceKernel.setArg<cl::Buffer>(index++, cl.getEnergyBuffer().getDeviceBuffer());
        reduceBornForceKernel.setArg<cl::Buffer>(index++, params->getDeviceBuffer());
        reduceBornForceKernel.setArg<cl::Buffer>(index++, bornRadii->getDeviceBuffer());
        reduceBornForceKernel.setArg<cl::Buffer>(index++, obcChain->getDeviceBuffer());
1817
    }
1818
    if (nb.getUseCutoff()) {
1819
1820
1821
1822
        computeBornSumKernel.setArg<mm_float4>(7, cl.getPeriodicBoxSize());
        computeBornSumKernel.setArg<mm_float4>(8, cl.getInvPeriodicBoxSize());
        force1Kernel.setArg<mm_float4>(9, cl.getPeriodicBoxSize());
        force1Kernel.setArg<mm_float4>(10, cl.getInvPeriodicBoxSize());
1823
1824
        if (maxTiles < nb.getInteractingTiles().getSize()) {
            maxTiles = nb.getInteractingTiles().getSize();
1825
            computeBornSumKernel.setArg<cl_uint>(10, maxTiles);
1826
            force1Kernel.setArg<cl_uint>(11, maxTiles);
1827
        }
1828
    }
1829
    cl.executeKernel(computeBornSumKernel, nb.getNumForceThreadBlocks()*nb.getForceThreadBlockSize(), nb.getForceThreadBlockSize());
1830
    cl.executeKernel(reduceBornSumKernel, cl.getPaddedNumAtoms());
1831
    cl.executeKernel(force1Kernel, nb.getNumForceThreadBlocks()*nb.getForceThreadBlockSize(), nb.getForceThreadBlockSize());
1832
    cl.executeKernel(reduceBornForceKernel, cl.getPaddedNumAtoms());
1833
    return 0.0;
1834
}
1835
1836
1837
1838
1839
1840
1841
1842
1843
1844

class OpenCLCustomGBForceInfo : public OpenCLForceInfo {
public:
    OpenCLCustomGBForceInfo(int requiredBuffers, const CustomGBForce& force) : OpenCLForceInfo(requiredBuffers), force(force) {
    }
    bool areParticlesIdentical(int particle1, int particle2) {
        vector<double> params1;
        vector<double> params2;
        force.getParticleParameters(particle1, params1);
        force.getParticleParameters(particle2, params2);
1845
        for (int i = 0; i < (int) params1.size(); i++)
1846
1847
1848
1849
1850
1851
1852
1853
1854
1855
1856
1857
1858
1859
1860
1861
1862
1863
1864
1865
1866
1867
1868
1869
1870
1871
            if (params1[i] != params2[i])
                return false;
        return true;
    }
    int getNumParticleGroups() {
        return force.getNumExclusions();
    }
    void getParticlesInGroup(int index, std::vector<int>& particles) {
        int particle1, particle2;
        force.getExclusionParticles(index, particle1, particle2);
        particles.resize(2);
        particles[0] = particle1;
        particles[1] = particle2;
    }
    bool areGroupsIdentical(int group1, int group2) {
        return true;
    }
private:
    const CustomGBForce& force;
};

OpenCLCalcCustomGBForceKernel::~OpenCLCalcCustomGBForceKernel() {
    if (params != NULL)
        delete params;
    if (computedValues != NULL)
        delete computedValues;
1872
1873
    if (energyDerivs != NULL)
        delete energyDerivs;
1874
1875
    if (longEnergyDerivs != NULL)
        delete longEnergyDerivs;
1876
1877
    if (globals != NULL)
        delete globals;
1878
1879
    if (valueBuffers != NULL)
        delete valueBuffers;
1880
1881
    if (longValueBuffers != NULL)
        delete longValueBuffers;
1882
1883
1884
1885
1886
1887
1888
    if (tabulatedFunctionParams != NULL)
        delete tabulatedFunctionParams;
    for (int i = 0; i < (int) tabulatedFunctions.size(); i++)
        delete tabulatedFunctions[i];
}

void OpenCLCalcCustomGBForceKernel::initialize(const System& system, const CustomGBForce& force) {
1889
1890
    if (cl.getPlatformData().contexts.size() > 1)
        throw OpenMMException("CustomGBForce does not support using multiple OpenCL devices");
1891
    bool useExclusionsForValue = false;
1892
    numComputedValues = force.getNumComputedValues();
1893
1894
    vector<string> computedValueNames(force.getNumComputedValues());
    vector<string> computedValueExpressions(force.getNumComputedValues());
1895
1896
    if (force.getNumComputedValues() > 0) {
        CustomGBForce::ComputationType type;
1897
        force.getComputedValueParameters(0, computedValueNames[0], computedValueExpressions[0], type);
1898
1899
1900
1901
        if (type == CustomGBForce::SingleParticle)
            throw OpenMMException("OpenCLPlatform requires that the first computed value for a CustomGBForce be of type ParticlePair or ParticlePairNoExclusions.");
        useExclusionsForValue = (type == CustomGBForce::ParticlePair);
        for (int i = 1; i < force.getNumComputedValues(); i++) {
1902
            force.getComputedValueParameters(i, computedValueNames[i], computedValueExpressions[i], type);
1903
1904
1905
1906
1907
1908
1909
1910
1911
1912
1913
1914
1915
1916
1917
1918
1919
1920
1921
1922
1923
1924
            if (type != CustomGBForce::SingleParticle)
                throw OpenMMException("OpenCLPlatform requires that a CustomGBForce only have one computed value of type ParticlePair or ParticlePairNoExclusions.");
        }
    }
    int forceIndex;
    for (forceIndex = 0; forceIndex < system.getNumForces() && &system.getForce(forceIndex) != &force; ++forceIndex)
        ;
    string prefix = "custom"+intToString(forceIndex)+"_";

    // Record parameters and exclusions.

    int numParticles = force.getNumParticles();
    params = new OpenCLParameterSet(cl, force.getNumPerParticleParameters(), numParticles, "customGBParameters");
    computedValues = new OpenCLParameterSet(cl, force.getNumComputedValues(), numParticles, "customGBComputedValues");
    if (force.getNumGlobalParameters() > 0)
        globals = new OpenCLArray<cl_float>(cl, force.getNumGlobalParameters(), "customGBGlobals", false, CL_MEM_READ_ONLY);
    vector<vector<cl_float> > paramVector(numParticles);
    vector<vector<int> > exclusionList(numParticles);
    for (int i = 0; i < numParticles; i++) {
        vector<double> parameters;
        force.getParticleParameters(i, parameters);
        paramVector[i].resize(parameters.size());
1925
        for (int j = 0; j < (int) parameters.size(); j++)
1926
1927
1928
1929
1930
1931
1932
1933
1934
1935
1936
1937
1938
1939
1940
1941
1942
            paramVector[i][j] = (cl_float) parameters[j];
        exclusionList[i].push_back(i);
    }
    for (int i = 0; i < force.getNumExclusions(); i++) {
        int particle1, particle2;
        force.getExclusionParticles(i, particle1, particle2);
        exclusionList[particle1].push_back(particle2);
        exclusionList[particle2].push_back(particle1);
    }
    params->setParameterValues(paramVector);

    // Record the tabulated functions.

    OpenCLExpressionUtilities::FunctionPlaceholder fp;
    map<string, Lepton::CustomFunction*> functions;
    vector<pair<string, string> > functionDefinitions;
    vector<mm_float4> tabulatedFunctionParamsVec(force.getNumFunctions());
1943
    stringstream tableArgs;
1944
1945
1946
1947
    for (int i = 0; i < force.getNumFunctions(); i++) {
        string name;
        vector<double> values;
        double min, max;
1948
        force.getFunctionParameters(i, name, values, min, max);
1949
1950
1951
        string arrayName = prefix+"table"+intToString(i);
        functionDefinitions.push_back(make_pair(name, arrayName));
        functions[name] = &fp;
1952
        tabulatedFunctionParamsVec[i] = mm_float4((float) min, (float) max, (float) ((values.size()-1)/(max-min)), (float) values.size()-2);
1953
        vector<mm_float4> f = OpenCLExpressionUtilities::computeFunctionCoefficients(values, min, max);
1954
1955
        tabulatedFunctions.push_back(new OpenCLArray<mm_float4>(cl, values.size()-1, "TabulatedFunction"));
        tabulatedFunctions[tabulatedFunctions.size()-1]->upload(f);
1956
        cl.getNonbondedUtilities().addArgument(OpenCLNonbondedUtilities::ParameterInfo(arrayName, "float", 4, sizeof(cl_float4), tabulatedFunctions[tabulatedFunctions.size()-1]->getDeviceBuffer()));
1957
        tableArgs << ", __global float4* " << arrayName;
1958
1959
1960
1961
    }
    if (force.getNumFunctions() > 0) {
        tabulatedFunctionParams = new OpenCLArray<mm_float4>(cl, tabulatedFunctionParamsVec.size(), "tabulatedFunctionParameters", false, CL_MEM_READ_ONLY);
        tabulatedFunctionParams->upload(tabulatedFunctionParamsVec);
1962
        cl.getNonbondedUtilities().addArgument(OpenCLNonbondedUtilities::ParameterInfo(prefix+"functionParams", "float", 4, sizeof(cl_float4), tabulatedFunctionParams->getDeviceBuffer()));
1963
        tableArgs << ", __constant float4* " << prefix << "functionParams";
1964
1965
    }

1966
    // Record the global parameters.
1967
1968
1969
1970
1971
1972
1973
1974
1975

    globalParamNames.resize(force.getNumGlobalParameters());
    globalParamValues.resize(force.getNumGlobalParameters());
    for (int i = 0; i < force.getNumGlobalParameters(); i++) {
        globalParamNames[i] = force.getGlobalParameterName(i);
        globalParamValues[i] = (cl_float) force.getGlobalParameterDefaultValue(i);
    }
    if (globals != NULL)
        globals->upload(globalParamValues);
1976
1977
1978

    // Record derivatives of expressions needed for the chain rule terms.

1979
    vector<vector<Lepton::ParsedExpression> > valueGradientExpressions(force.getNumComputedValues());
1980
    vector<vector<Lepton::ParsedExpression> > valueDerivExpressions(force.getNumComputedValues());
Peter Eastman's avatar
Peter Eastman committed
1981
    needParameterGradient = false;
1982
1983
1984
1985
1986
1987
1988
    for (int i = 1; i < force.getNumComputedValues(); i++) {
        Lepton::ParsedExpression ex = Lepton::Parser::parse(computedValueExpressions[i], functions).optimize();
        valueGradientExpressions[i].push_back(ex.differentiate("x").optimize());
        valueGradientExpressions[i].push_back(ex.differentiate("y").optimize());
        valueGradientExpressions[i].push_back(ex.differentiate("z").optimize());
        if (!isZeroExpression(valueGradientExpressions[i][0]) || !isZeroExpression(valueGradientExpressions[i][1]) || !isZeroExpression(valueGradientExpressions[i][2]))
            needParameterGradient = true;
1989
1990
         for (int j = 0; j < i; j++)
            valueDerivExpressions[i].push_back(ex.differentiate(computedValueNames[j]).optimize());
1991
    }
1992
    vector<vector<Lepton::ParsedExpression> > energyDerivExpressions(force.getNumEnergyTerms());
1993
1994
1995
1996
1997
1998
1999
2000
2001
2002
2003
2004
2005
2006
    for (int i = 0; i < force.getNumEnergyTerms(); i++) {
        string expression;
        CustomGBForce::ComputationType type;
        force.getEnergyTermParameters(i, expression, type);
        Lepton::ParsedExpression ex = Lepton::Parser::parse(expression, functions).optimize();
        for (int j = 0; j < force.getNumComputedValues(); j++) {
            if (type == CustomGBForce::SingleParticle)
                energyDerivExpressions[i].push_back(ex.differentiate(computedValueNames[j]).optimize());
            else {
                energyDerivExpressions[i].push_back(ex.differentiate(computedValueNames[j]+"1").optimize());
                energyDerivExpressions[i].push_back(ex.differentiate(computedValueNames[j]+"2").optimize());
            }
        }
    }
2007
2008
2009
2010
2011
2012
2013
2014
    bool deviceIsCpu = (cl.getDevice().getInfo<CL_DEVICE_TYPE>() == CL_DEVICE_TYPE_CPU);
    bool useLong = (cl.getSupports64BitGlobalAtomics() && !deviceIsCpu);
    if (useLong) {
        longEnergyDerivs = new OpenCLArray<cl_long>(cl, force.getNumComputedValues()*cl.getPaddedNumAtoms(), "customGBLongEnergyDerivatives");
        energyDerivs = new OpenCLParameterSet(cl, force.getNumComputedValues(), cl.getPaddedNumAtoms(), "customGBEnergyDerivatives");
    }
    else
        energyDerivs = new OpenCLParameterSet(cl, force.getNumComputedValues(), cl.getPaddedNumAtoms()*cl.getNonbondedUtilities().getNumForceBuffers(), "customGBEnergyDerivatives");
2015

2016
2017
    // Create the kernels.

2018
2019
    bool useCutoff = (force.getNonbondedMethod() != CustomGBForce::NoCutoff);
    bool usePeriodic = (force.getNonbondedMethod() != CustomGBForce::NoCutoff && force.getNonbondedMethod() != CustomGBForce::CutoffNonPeriodic);
2020
2021
2022
    {
        // Create the N2 value kernel.

2023
2024
2025
        map<string, string> variables;
        map<string, string> rename;
        variables["r"] = "r";
2026
2027
        for (int i = 0; i < force.getNumPerParticleParameters(); i++) {
            const string& name = force.getPerParticleParameterName(i);
2028
2029
2030
2031
            variables[name+"1"] = "params"+params->getParameterSuffix(i, "1");
            variables[name+"2"] = "params"+params->getParameterSuffix(i, "2");
            rename[name+"1"] = name+"2";
            rename[name+"2"] = name+"1";
2032
2033
2034
2035
        }
        for (int i = 0; i < force.getNumGlobalParameters(); i++) {
            const string& name = force.getGlobalParameterName(i);
            string value = "globals["+intToString(i)+"]";
2036
            variables[name] = value;
2037
        }
2038
2039
        map<string, Lepton::ParsedExpression> n2ValueExpressions;
        stringstream n2ValueSource;
2040
2041
2042
2043
        Lepton::ParsedExpression ex = Lepton::Parser::parse(computedValueExpressions[0], functions).optimize();
        n2ValueExpressions["tempValue1 = "] = ex;
        n2ValueExpressions["tempValue2 = "] = ex.renameVariables(rename);
        n2ValueSource << OpenCLExpressionUtilities::createExpressions(n2ValueExpressions, variables, functionDefinitions, "temp", prefix+"functionParams");
2044
2045
2046
2047
2048
2049
2050
        map<string, string> replacements;
        replacements["COMPUTE_VALUE"] = n2ValueSource.str();
        stringstream extraArgs, loadLocal1, loadLocal2, load1, load2;
        if (force.getNumGlobalParameters() > 0)
            extraArgs << ", __constant float* globals";
        for (int i = 0; i < (int) params->getBuffers().size(); i++) {
            const OpenCLNonbondedUtilities::ParameterInfo& buffer = params->getBuffers()[i];
2051
            string paramName = "params"+intToString(i+1);
2052
            extraArgs << ", __global " << buffer.getType() << "* global_" << paramName << ", __local " << buffer.getType() << "* local_" << paramName;
2053
2054
            loadLocal1 << "local_" << paramName << "[localAtomIndex] = " << paramName << "1;\n";
            loadLocal2 << "local_" << paramName << "[localAtomIndex] = global_" << paramName << "[j];\n";
2055
2056
2057
            load1 << buffer.getType() << " " << paramName << "1 = global_" << paramName << "[atom1];\n";
            load2 << buffer.getType() << " " << paramName << "2 = local_" << paramName << "[atom2];\n";
        }
2058
        replacements["PARAMETER_ARGUMENTS"] = extraArgs.str()+tableArgs.str();
2059
2060
2061
2062
2063
2064
2065
2066
2067
2068
2069
2070
2071
        replacements["LOAD_LOCAL_PARAMETERS_FROM_1"] = loadLocal1.str();
        replacements["LOAD_LOCAL_PARAMETERS_FROM_GLOBAL"] = loadLocal2.str();
        replacements["LOAD_ATOM1_PARAMETERS"] = load1.str();
        replacements["LOAD_ATOM2_PARAMETERS"] = load2.str();
        map<string, string> defines;
        if (cl.getNonbondedUtilities().getForceBufferPerAtomBlock())
            defines["USE_OUTPUT_BUFFER_PER_BLOCK"] = "1";
        if (useCutoff)
            defines["USE_CUTOFF"] = "1";
        if (usePeriodic)
            defines["USE_PERIODIC"] = "1";
        if (useExclusionsForValue)
            defines["USE_EXCLUSIONS"] = "1";
2072
2073
        if (cl.getSIMDWidth() == 32)
            defines["WARPS_PER_GROUP"] = OpenCLExpressionUtilities::intToString(cl.getNonbondedUtilities().getForceThreadBlockSize()/OpenCLContext::TileSize);
2074
2075
2076
        defines["CUTOFF_SQUARED"] = doubleToString(force.getCutoffDistance()*force.getCutoffDistance());
        defines["NUM_ATOMS"] = intToString(cl.getNumAtoms());
        defines["PADDED_NUM_ATOMS"] = intToString(cl.getPaddedNumAtoms());
2077
        defines["NUM_BLOCKS"] = OpenCLExpressionUtilities::intToString(cl.getNumAtomBlocks());
2078
2079
2080
2081
2082
2083
        string file;
        if (deviceIsCpu)
            file = OpenCLKernelSources::customGBValueN2_cpu;
        else if (cl.getSIMDWidth() == 32)
            file = OpenCLKernelSources::customGBValueN2_nvidia;
        else
Peter Eastman's avatar
Peter Eastman committed
2084
            file = OpenCLKernelSources::customGBValueN2_default;
2085
        cl::Program program = cl.createProgram(cl.replaceStrings(file, replacements), defines);
2086
        pairValueKernel = cl::Kernel(program, "computeN2Value");
2087
2088
        if (useExclusionsForValue)
            cl.getNonbondedUtilities().requestExclusions(exclusionList);
2089
2090
2091
2092
2093
2094
2095
2096
2097
    }
    {
        // Create the kernel to reduce the N2 value and calculate other values.

        stringstream reductionSource, extraArgs;
        if (force.getNumGlobalParameters() > 0)
            extraArgs << ", __constant float* globals";
        for (int i = 0; i < (int) params->getBuffers().size(); i++) {
            const OpenCLNonbondedUtilities::ParameterInfo& buffer = params->getBuffers()[i];
2098
2099
            string paramName = "params"+intToString(i+1);
            extraArgs << ", __global " << buffer.getType() << "* " << paramName;
2100
2101
2102
        }
        for (int i = 0; i < (int) computedValues->getBuffers().size(); i++) {
            const OpenCLNonbondedUtilities::ParameterInfo& buffer = computedValues->getBuffers()[i];
2103
            string valueName = "values"+intToString(i+1);
2104
2105
2106
2107
            extraArgs << ", __global " << buffer.getType() << "* global_" << valueName;
            reductionSource << buffer.getType() << " local_" << valueName << ";\n";
        }
        reductionSource << "local_values" << computedValues->getParameterSuffix(0) << " = sum;\n";
2108
        map<string, string> variables;
2109
2110
2111
        variables["x"] = "pos.x";
        variables["y"] = "pos.y";
        variables["z"] = "pos.z";
2112
2113
2114
2115
2116
2117
2118
2119
2120
2121
        for (int i = 0; i < force.getNumPerParticleParameters(); i++)
            variables[force.getPerParticleParameterName(i)] = "params"+params->getParameterSuffix(i, "[index]");
        for (int i = 0; i < force.getNumGlobalParameters(); i++)
            variables[force.getGlobalParameterName(i)] = "globals["+intToString(i)+"]";
        for (int i = 1; i < force.getNumComputedValues(); i++) {
            variables[computedValueNames[i-1]] = "local_values"+computedValues->getParameterSuffix(i-1);
            map<string, Lepton::ParsedExpression> valueExpressions;
            valueExpressions["local_values"+computedValues->getParameterSuffix(i)+" = "] = Lepton::Parser::parse(computedValueExpressions[i], functions).optimize();
            reductionSource << OpenCLExpressionUtilities::createExpressions(valueExpressions, variables, functionDefinitions, "value"+intToString(i)+"_temp", "functionParams");
        }
2122
        for (int i = 0; i < (int) computedValues->getBuffers().size(); i++) {
2123
            string valueName = "values"+intToString(i+1);
2124
2125
2126
            reductionSource << "global_" << valueName << "[index] = local_" << valueName << ";\n";
        }
        map<string, string> replacements;
2127
        replacements["PARAMETER_ARGUMENTS"] = extraArgs.str()+tableArgs.str();
2128
2129
2130
        replacements["COMPUTE_VALUES"] = reductionSource.str();
        map<string, string> defines;
        defines["NUM_ATOMS"] = intToString(cl.getNumAtoms());
2131
        cl::Program program = cl.createProgram(cl.replaceStrings(OpenCLKernelSources::customGBValuePerParticle, replacements), defines);
2132
2133
2134
2135
2136
2137
2138
2139
2140
2141
2142
2143
2144
2145
2146
2147
2148
2149
        perParticleValueKernel = cl::Kernel(program, "computePerParticleValues");
    }
    {
        // Create the N2 energy kernel.

        map<string, string> variables;
        variables["r"] = "r";
        for (int i = 0; i < force.getNumPerParticleParameters(); i++) {
            const string& name = force.getPerParticleParameterName(i);
            variables[name+"1"] = "params"+params->getParameterSuffix(i, "1");
            variables[name+"2"] = "params"+params->getParameterSuffix(i, "2");
        }
        for (int i = 0; i < force.getNumComputedValues(); i++) {
            variables[computedValueNames[i]+"1"] = "values"+computedValues->getParameterSuffix(i, "1");
            variables[computedValueNames[i]+"2"] = "values"+computedValues->getParameterSuffix(i, "2");
        }
        for (int i = 0; i < force.getNumGlobalParameters(); i++)
            variables[force.getGlobalParameterName(i)] = "globals["+intToString(i)+"]";
2150
        stringstream n2EnergySource;
2151
        bool anyExclusions = (force.getNumExclusions() > 0);
2152
2153
2154
2155
2156
2157
        for (int i = 0; i < force.getNumEnergyTerms(); i++) {
            string expression;
            CustomGBForce::ComputationType type;
            force.getEnergyTermParameters(i, expression, type);
            if (type == CustomGBForce::SingleParticle)
                continue;
2158
            bool exclude = (anyExclusions && type == CustomGBForce::ParticlePair);
2159
            map<string, Lepton::ParsedExpression> n2EnergyExpressions;
2160
2161
            n2EnergyExpressions["tempEnergy += "] = Lepton::Parser::parse(expression, functions).optimize();
            n2EnergyExpressions["dEdR += "] = Lepton::Parser::parse(expression, functions).differentiate("r").optimize();
2162
2163
2164
2165
2166
2167
2168
2169
2170
2171
2172
2173
            if (useLong) {
                for (int j = 0; j < force.getNumComputedValues(); j++) {
                    string index = intToString(j+1);
                    n2EnergyExpressions["/*"+intToString(i+1)+"*/ deriv"+index+"_1 += "] = energyDerivExpressions[i][2*j];
                    n2EnergyExpressions["/*"+intToString(i+1)+"*/ deriv"+index+"_2 += "] = energyDerivExpressions[i][2*j+1];
                }
            }
            else {
                for (int j = 0; j < force.getNumComputedValues(); j++) {
                    n2EnergyExpressions["/*"+intToString(i+1)+"*/ deriv"+energyDerivs->getParameterSuffix(j, "_1")+" += "] = energyDerivExpressions[i][2*j];
                    n2EnergyExpressions["/*"+intToString(i+1)+"*/ deriv"+energyDerivs->getParameterSuffix(j, "_2")+" += "] = energyDerivExpressions[i][2*j+1];
                }
2174
2175
2176
2177
2178
2179
            }
            if (exclude)
                n2EnergySource << "if (!isExcluded) {\n";
            n2EnergySource << OpenCLExpressionUtilities::createExpressions(n2EnergyExpressions, variables, functionDefinitions, "temp", prefix+"functionParams");
            if (exclude)
                n2EnergySource << "}\n";
2180
2181
        }
        map<string, string> replacements;
2182
        replacements["COMPUTE_INTERACTION"] = n2EnergySource.str();
2183
        stringstream extraArgs, loadLocal1, loadLocal2, clearLocal, load1, load2, declare1, recordDeriv, storeDerivs1, storeDerivs2, declareTemps, setTemps;
2184
2185
2186
2187
2188
2189
        if (force.getNumGlobalParameters() > 0)
            extraArgs << ", __constant float* globals";
        for (int i = 0; i < (int) params->getBuffers().size(); i++) {
            const OpenCLNonbondedUtilities::ParameterInfo& buffer = params->getBuffers()[i];
            string paramName = "params"+intToString(i+1);
            extraArgs << ", __global " << buffer.getType() << "* global_" << paramName << ", __local " << buffer.getType() << "* local_" << paramName;
2190
2191
            loadLocal1 << "local_" << paramName << "[localAtomIndex] = " << paramName << "1;\n";
            loadLocal2 << "local_" << paramName << "[localAtomIndex] = global_" << paramName << "[j];\n";
2192
2193
2194
2195
2196
2197
2198
            load1 << buffer.getType() << " " << paramName << "1 = global_" << paramName << "[atom1];\n";
            load2 << buffer.getType() << " " << paramName << "2 = local_" << paramName << "[atom2];\n";
        }
        for (int i = 0; i < (int) computedValues->getBuffers().size(); i++) {
            const OpenCLNonbondedUtilities::ParameterInfo& buffer = computedValues->getBuffers()[i];
            string valueName = "values"+intToString(i+1);
            extraArgs << ", __global " << buffer.getType() << "* global_" << valueName << ", __local " << buffer.getType() << "* local_" << valueName;
2199
2200
            loadLocal1 << "local_" << valueName << "[localAtomIndex] = " << valueName << "1;\n";
            loadLocal2 << "local_" << valueName << "[localAtomIndex] = global_" << valueName << "[j];\n";
2201
2202
2203
            load1 << buffer.getType() << " " << valueName << "1 = global_" << valueName << "[atom1];\n";
            load2 << buffer.getType() << " " << valueName << "2 = local_" << valueName << "[atom2];\n";
        }
2204
2205
2206
2207
2208
2209
2210
2211
2212
2213
2214
2215
2216
2217
2218
2219
2220
2221
2222
2223
2224
2225
2226
2227
2228
2229
2230
2231
2232
        if (useLong) {
            extraArgs << ", __global long* derivBuffers";
            for (int i = 0; i < force.getNumComputedValues(); i++) {
                string index = intToString(i+1);
                extraArgs << ", __local float* local_deriv" << index;
                clearLocal << "local_deriv" << index << "[localAtomIndex] = 0.0f;\n";
                declare1 << "float deriv" << index << "_1 = 0.0f;\n";
                load2 << "float deriv" << index << "_2 = 0.0f;\n";
                recordDeriv << "local_deriv" << index << "[atom2] += deriv" << index << "_2;\n";
                storeDerivs1 << "STORE_DERIVATIVE_1(" << index << ")\n";
                storeDerivs2 << "STORE_DERIVATIVE_2(" << index << ")\n";
                declareTemps << "__local float tempDerivBuffer" << index << "[64];\n";
                setTemps << "tempDerivBuffer" << index << "[get_local_id(0)] = deriv" << index << "_1;\n";
            }
        }
        else {
            for (int i = 0; i < (int) energyDerivs->getBuffers().size(); i++) {
                const OpenCLNonbondedUtilities::ParameterInfo& buffer = energyDerivs->getBuffers()[i];
                string index = intToString(i+1);
                extraArgs << ", __global " << buffer.getType() << "* derivBuffers" << index << ", __local " << buffer.getType() << "* local_deriv" << index;
                clearLocal << "local_deriv" << index << "[localAtomIndex] = 0.0f;\n";
                declare1 << buffer.getType() << " deriv" << index << "_1 = 0.0f;\n";
                load2 << buffer.getType() << " deriv" << index << "_2 = 0.0f;\n";
                recordDeriv << "local_deriv" << index << "[atom2] += deriv" << index << "_2;\n";
                storeDerivs1 << "STORE_DERIVATIVE_1(" << index << ")\n";
                storeDerivs2 << "STORE_DERIVATIVE_2(" << index << ")\n";
                declareTemps << "__local " << buffer.getType() << " tempDerivBuffer" << index << "[64];\n";
                setTemps << "tempDerivBuffer" << index << "[get_local_id(0)] = deriv" << index << "_1;\n";
            }
2233
        }
2234
2235
2236
        replacements["PARAMETER_ARGUMENTS"] = extraArgs.str()+tableArgs.str();
        replacements["LOAD_LOCAL_PARAMETERS_FROM_1"] = loadLocal1.str();
        replacements["LOAD_LOCAL_PARAMETERS_FROM_GLOBAL"] = loadLocal2.str();
2237
        replacements["CLEAR_LOCAL_DERIVATIVES"] = clearLocal.str();
2238
2239
        replacements["LOAD_ATOM1_PARAMETERS"] = load1.str();
        replacements["LOAD_ATOM2_PARAMETERS"] = load2.str();
2240
        replacements["DECLARE_ATOM1_DERIVATIVES"] = declare1.str();
2241
2242
2243
        replacements["RECORD_DERIVATIVE_2"] = recordDeriv.str();
        replacements["STORE_DERIVATIVES_1"] = storeDerivs1.str();
        replacements["STORE_DERIVATIVES_2"] = storeDerivs2.str();
2244
2245
        replacements["DECLARE_TEMP_BUFFERS"] = declareTemps.str();
        replacements["SET_TEMP_BUFFERS"] = setTemps.str();
2246
2247
2248
2249
2250
2251
2252
2253
2254
        map<string, string> defines;
        if (cl.getNonbondedUtilities().getForceBufferPerAtomBlock())
            defines["USE_OUTPUT_BUFFER_PER_BLOCK"] = "1";
        if (useCutoff)
            defines["USE_CUTOFF"] = "1";
        if (usePeriodic)
            defines["USE_PERIODIC"] = "1";
        if (anyExclusions)
            defines["USE_EXCLUSIONS"] = "1";
2255
2256
        if (cl.getSIMDWidth() == 32)
            defines["WARPS_PER_GROUP"] = OpenCLExpressionUtilities::intToString(cl.getNonbondedUtilities().getForceThreadBlockSize()/OpenCLContext::TileSize);
2257
2258
2259
        defines["CUTOFF_SQUARED"] = doubleToString(force.getCutoffDistance()*force.getCutoffDistance());
        defines["NUM_ATOMS"] = intToString(cl.getNumAtoms());
        defines["PADDED_NUM_ATOMS"] = intToString(cl.getPaddedNumAtoms());
2260
        defines["NUM_BLOCKS"] = OpenCLExpressionUtilities::intToString(cl.getNumAtomBlocks());
2261
2262
2263
2264
2265
2266
2267
        string file;
        if (deviceIsCpu)
            file = OpenCLKernelSources::customGBEnergyN2_cpu;
        else if (cl.getSIMDWidth() == 32)
            file = OpenCLKernelSources::customGBEnergyN2_nvidia;
        else
            file = OpenCLKernelSources::customGBEnergyN2_default;
2268
        cl::Program program = cl.createProgram(cl.replaceStrings(file, replacements), defines);
2269
2270
2271
2272
2273
        pairEnergyKernel = cl::Kernel(program, "computeN2Energy");
    }
    {
        // Create the kernel to reduce the derivatives and calculate per-particle energy terms.

2274
        stringstream compute, extraArgs, reduce;
2275
2276
2277
2278
2279
2280
2281
2282
2283
2284
2285
2286
        if (force.getNumGlobalParameters() > 0)
            extraArgs << ", __constant float* globals";
        for (int i = 0; i < (int) params->getBuffers().size(); i++) {
            const OpenCLNonbondedUtilities::ParameterInfo& buffer = params->getBuffers()[i];
            string paramName = "params"+intToString(i+1);
            extraArgs << ", __global " << buffer.getType() << "* " << paramName;
        }
        for (int i = 0; i < (int) computedValues->getBuffers().size(); i++) {
            const OpenCLNonbondedUtilities::ParameterInfo& buffer = computedValues->getBuffers()[i];
            string valueName = "values"+intToString(i+1);
            extraArgs << ", __global " << buffer.getType() << "* " << valueName;
        }
2287
2288
2289
2290
2291
2292
        for (int i = 0; i < (int) energyDerivs->getBuffers().size(); i++) {
            const OpenCLNonbondedUtilities::ParameterInfo& buffer = energyDerivs->getBuffers()[i];
            string index = intToString(i+1);
            extraArgs << ", __global " << buffer.getType() << "* derivBuffers" << index;
            compute << buffer.getType() << " deriv" << index << " = derivBuffers" << index << "[index];\n";
        }
2293
2294
2295
2296
2297
2298
2299
2300
2301
2302
        if (useLong) {
            extraArgs << ", __global long* derivBuffersIn";
            for (int i = 0; i < energyDerivs->getNumParameters(); ++i)
                reduce << "derivBuffers" << energyDerivs->getParameterSuffix(i, "[index]") <<
                        " = (1.0f/0xFFFFFFFF)*derivBuffersIn[index+PADDED_NUM_ATOMS*" << intToString(i) << "];\n";
        }
        else {
            for (int i = 0; i < (int) energyDerivs->getBuffers().size(); i++)
                reduce << "REDUCE_VALUE(derivBuffers" << intToString(i+1) << ", " << energyDerivs->getBuffers()[i].getType() << ")\n";
        }
2303
        map<string, string> variables;
2304
2305
2306
        variables["x"] = "pos.x";
        variables["y"] = "pos.y";
        variables["z"] = "pos.z";
2307
2308
2309
2310
2311
2312
        for (int i = 0; i < force.getNumPerParticleParameters(); i++)
            variables[force.getPerParticleParameterName(i)] = "params"+params->getParameterSuffix(i, "[index]");
        for (int i = 0; i < force.getNumGlobalParameters(); i++)
            variables[force.getGlobalParameterName(i)] = "globals["+intToString(i)+"]";
        for (int i = 0; i < force.getNumComputedValues(); i++)
            variables[computedValueNames[i]] = "values"+computedValues->getParameterSuffix(i, "[index]");
2313
        map<string, Lepton::ParsedExpression> energyExpressions;
2314
2315
2316
2317
2318
2319
        for (int i = 0; i < force.getNumEnergyTerms(); i++) {
            string expression;
            CustomGBForce::ComputationType type;
            force.getEnergyTermParameters(i, expression, type);
            if (type != CustomGBForce::SingleParticle)
                continue;
2320
2321
            Lepton::ParsedExpression parsed = Lepton::Parser::parse(expression, functions).optimize();
            energyExpressions["/*"+intToString(i+1)+"*/ energy += "] = parsed;
2322
2323
            for (int j = 0; j < force.getNumComputedValues(); j++)
                energyExpressions["/*"+intToString(i+1)+"*/ deriv"+energyDerivs->getParameterSuffix(j)+" += "] = energyDerivExpressions[i][j];
2324
2325
2326
2327
2328
2329
2330
2331
2332
            Lepton::ParsedExpression gradx = parsed.differentiate("x").optimize();
            Lepton::ParsedExpression grady = parsed.differentiate("y").optimize();
            Lepton::ParsedExpression gradz = parsed.differentiate("z").optimize();
            if (!isZeroExpression(gradx))
                energyExpressions["/*"+intToString(i+1)+"*/ force.x -= "] = gradx;
            if (!isZeroExpression(grady))
                energyExpressions["/*"+intToString(i+1)+"*/ force.y -= "] = grady;
            if (!isZeroExpression(gradz))
                energyExpressions["/*"+intToString(i+1)+"*/ force.z -= "] = gradz;
2333
2334
2335
2336
2337
        }
        compute << OpenCLExpressionUtilities::createExpressions(energyExpressions, variables, functionDefinitions, "temp", prefix+"functionParams");
        for (int i = 0; i < (int) energyDerivs->getBuffers().size(); i++) {
            string index = intToString(i+1);
            compute << "derivBuffers" << index << "[index] = deriv" << index << ";\n";
2338
        }
2339
        compute << "forceBuffers[index] = forceBuffers[index]+force;\n";
2340
2341
        map<string, string> replacements;
        replacements["PARAMETER_ARGUMENTS"] = extraArgs.str()+tableArgs.str();
2342
2343
        replacements["REDUCE_DERIVATIVES"] = reduce.str();
        replacements["COMPUTE_ENERGY"] = compute.str();
2344
2345
        map<string, string> defines;
        defines["NUM_ATOMS"] = intToString(cl.getNumAtoms());
2346
        defines["PADDED_NUM_ATOMS"] = intToString(cl.getPaddedNumAtoms());
2347
        cl::Program program = cl.createProgram(cl.replaceStrings(OpenCLKernelSources::customGBEnergyPerParticle, replacements), defines);
2348
        perParticleEnergyKernel = cl::Kernel(program, "computePerParticleEnergy");
2349
    }
Peter Eastman's avatar
Peter Eastman committed
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
    if (needParameterGradient) {
        // Create the kernel to compute chain rule terms for computed values that depend explicitly on particle coordinates.

        stringstream compute, extraArgs;
        if (force.getNumGlobalParameters() > 0)
            extraArgs << ", __constant float* globals";
        for (int i = 0; i < (int) params->getBuffers().size(); i++) {
            const OpenCLNonbondedUtilities::ParameterInfo& buffer = params->getBuffers()[i];
            string paramName = "params"+intToString(i+1);
            extraArgs << ", __global " << buffer.getType() << "* " << paramName;
        }
        for (int i = 0; i < (int) computedValues->getBuffers().size(); i++) {
            const OpenCLNonbondedUtilities::ParameterInfo& buffer = computedValues->getBuffers()[i];
            string valueName = "values"+intToString(i+1);
            extraArgs << ", __global " << buffer.getType() << "* " << valueName;
        }
        for (int i = 0; i < (int) energyDerivs->getBuffers().size(); i++) {
            const OpenCLNonbondedUtilities::ParameterInfo& buffer = energyDerivs->getBuffers()[i];
            string index = intToString(i+1);
            extraArgs << ", __global " << buffer.getType() << "* derivBuffers" << index;
            compute << buffer.getType() << " deriv" << index << " = derivBuffers" << index << "[index];\n";
        }
        map<string, string> variables;
        variables["x"] = "pos.x";
        variables["y"] = "pos.y";
        variables["z"] = "pos.z";
        for (int i = 0; i < force.getNumPerParticleParameters(); i++)
            variables[force.getPerParticleParameterName(i)] = "params"+params->getParameterSuffix(i, "[index]");
        for (int i = 0; i < force.getNumGlobalParameters(); i++)
            variables[force.getGlobalParameterName(i)] = "globals["+intToString(i)+"]";
        for (int i = 0; i < force.getNumComputedValues(); i++)
            variables[computedValueNames[i]] = "values"+computedValues->getParameterSuffix(i, "[index]");
        for (int i = 1; i < force.getNumComputedValues(); i++) {
2383
2384
2385
2386
2387
2388
2389
2390
2391
2392
2393
2394
            string is = intToString(i);
            compute << "float4 dV"<<is<<"dR = (float4) 0;\n";
            for (int j = 1; j < i; j++) {
                if (!isZeroExpression(valueDerivExpressions[i][j])) {
                    map<string, Lepton::ParsedExpression> derivExpressions;
                    string js = intToString(j);
                    derivExpressions["float dV"+is+"dV"+js+" = "] = valueDerivExpressions[i][j];
                    compute << OpenCLExpressionUtilities::createExpressions(derivExpressions, variables, functionDefinitions, "temp_"+is+"_"+js, prefix+"functionParams");
                    compute << "dV"<<is<<"dR += dV"<<is<<"dV"<<js<<"*dV"<<js<<"dR;\n";
                }
            }
            map<string, Lepton::ParsedExpression> gradientExpressions;
Peter Eastman's avatar
Peter Eastman committed
2395
            if (!isZeroExpression(valueGradientExpressions[i][0]))
2396
                gradientExpressions["dV"+is+"dR.x += "] = valueGradientExpressions[i][0];
Peter Eastman's avatar
Peter Eastman committed
2397
            if (!isZeroExpression(valueGradientExpressions[i][1]))
2398
                gradientExpressions["dV"+is+"dR.y += "] = valueGradientExpressions[i][1];
Peter Eastman's avatar
Peter Eastman committed
2399
            if (!isZeroExpression(valueGradientExpressions[i][2]))
2400
2401
2402
2403
2404
2405
                gradientExpressions["dV"+is+"dR.z += "] = valueGradientExpressions[i][2];
            compute << OpenCLExpressionUtilities::createExpressions(gradientExpressions, variables, functionDefinitions, "temp", prefix+"functionParams");
        }
        for (int i = 1; i < force.getNumComputedValues(); i++) {
            string is = intToString(i);
            compute << "force -= deriv"<<energyDerivs->getParameterSuffix(i)<<"*dV"<<is<<"dR;\n";
Peter Eastman's avatar
Peter Eastman committed
2406
2407
2408
2409
2410
2411
2412
2413
2414
        }
        map<string, string> replacements;
        replacements["PARAMETER_ARGUMENTS"] = extraArgs.str()+tableArgs.str();
        replacements["COMPUTE_FORCES"] = compute.str();
        map<string, string> defines;
        defines["NUM_ATOMS"] = intToString(cl.getNumAtoms());
        cl::Program program = cl.createProgram(cl.replaceStrings(OpenCLKernelSources::customGBGradientChainRule, replacements), defines);
        gradientChainRuleKernel = cl::Kernel(program, "computeGradientChainRuleTerms");
    }
2415
    {
Peter Eastman's avatar
Peter Eastman committed
2416
        // Create the code to calculate chain rules terms as part of the default nonbonded kernel.
2417
2418
2419
2420
2421
2422
2423

        map<string, string> globalVariables;
        for (int i = 0; i < force.getNumGlobalParameters(); i++) {
            const string& name = force.getGlobalParameterName(i);
            string value = "globals["+intToString(i)+"]";
            globalVariables[name] = prefix+value;
        }
2424
2425
2426
        map<string, string> variables = globalVariables;
        map<string, string> rename;
        variables["r"] = "r";
2427
2428
        for (int i = 0; i < force.getNumPerParticleParameters(); i++) {
            const string& name = force.getPerParticleParameterName(i);
2429
2430
2431
2432
            variables[name+"1"] = prefix+"params"+params->getParameterSuffix(i, "1");
            variables[name+"2"] = prefix+"params"+params->getParameterSuffix(i, "2");
            rename[name+"1"] =  name+"2";
            rename[name+"2"] =  name+"1";
2433
2434
2435
2436
        }
        map<string, Lepton::ParsedExpression> derivExpressions;
        stringstream chainSource;
        Lepton::ParsedExpression dVdR = Lepton::Parser::parse(computedValueExpressions[0], functions).differentiate("r").optimize();
2437
2438
        derivExpressions["float dV0dR1 = "] = dVdR;
        derivExpressions["float dV0dR2 = "] = dVdR.renameVariables(rename);
2439
        chainSource << OpenCLExpressionUtilities::createExpressions(derivExpressions, variables, functionDefinitions, prefix+"temp0_", prefix+"functionParams");
Peter Eastman's avatar
Peter Eastman committed
2440
2441
2442
2443
2444
2445
        if (useExclusionsForValue)
            chainSource << "if (!isExcluded) {\n";
        chainSource << "tempForce -= dV0dR1*" << prefix << "dEdV" << energyDerivs->getParameterSuffix(0, "1") << ";\n";
        chainSource << "tempForce -= dV0dR2*" << prefix << "dEdV" << energyDerivs->getParameterSuffix(0, "2") << ";\n";
        if (useExclusionsForValue)
            chainSource << "}\n";
2446
2447
2448
        variables = globalVariables;
        map<string, string> rename1;
        map<string, string> rename2;
Peter Eastman's avatar
Peter Eastman committed
2449
2450
2451
2452
2453
2454
2455
2456
2457
2458
2459
2460
        variables["x1"] = "posq1.x";
        variables["y1"] = "posq1.y";
        variables["z1"] = "posq1.z";
        variables["x2"] = "posq2.x";
        variables["y2"] = "posq2.y";
        variables["z2"] = "posq2.z";
        rename1["x"] = "x1";
        rename1["y"] = "y1";
        rename1["z"] = "z1";
        rename2["x"] = "x2";
        rename2["y"] = "y2";
        rename2["z"] = "z2";
2461
2462
        for (int i = 0; i < force.getNumPerParticleParameters(); i++) {
            const string& name = force.getPerParticleParameterName(i);
2463
2464
2465
2466
            variables[name+"1"] = prefix+"params"+params->getParameterSuffix(i, "1");
            variables[name+"2"] = prefix+"params"+params->getParameterSuffix(i, "2");
            rename1[name] = name+"1";
            rename2[name] = name+"2";
2467
2468
2469
        }
        for (int i = 0; i < force.getNumComputedValues(); i++) {
            const string& name = computedValueNames[i];
2470
2471
2472
2473
            variables[name+"1"] = prefix+"values"+computedValues->getParameterSuffix(i, "1");
            variables[name+"2"] = prefix+"values"+computedValues->getParameterSuffix(i, "2");
            rename1[name] = name+"1";
            rename2[name] = name+"2";
2474
2475
            if (i == 0)
                continue;
2476
            string is = intToString(i);
Peter Eastman's avatar
Peter Eastman committed
2477
2478
2479
2480
2481
            chainSource << "float dV"+is+"dR1 = 0;\n";
            chainSource << "float dV"+is+"dR2 = 0;\n";
            for (int j = 0; j < i; j++) {
                string js = intToString(j);
                Lepton::ParsedExpression dVdV = Lepton::Parser::parse(computedValueExpressions[i], functions).differentiate(computedValueNames[j]).optimize();
2482
                derivExpressions.clear();
Peter Eastman's avatar
Peter Eastman committed
2483
2484
2485
                derivExpressions["dV"+is+"dR1 += dV"+js+"dR1*"] = dVdV.renameVariables(rename1);
                derivExpressions["dV"+is+"dR2 += dV"+js+"dR2*"] = dVdV.renameVariables(rename2);
                chainSource << OpenCLExpressionUtilities::createExpressions(derivExpressions, variables, functionDefinitions, prefix+"temp"+is+"_"+js+"_", prefix+"functionParams");
2486
            }
Peter Eastman's avatar
Peter Eastman committed
2487
2488
            chainSource << "tempForce -= dV"<< is << "dR1*" << prefix << "dEdV" << energyDerivs->getParameterSuffix(i, "1") << ";\n";
            chainSource << "tempForce -= dV"<< is << "dR2*" << prefix << "dEdV" << energyDerivs->getParameterSuffix(i, "2") << ";\n";
2489
2490
2491
        }
        map<string, string> replacements;
        replacements["COMPUTE_FORCE"] = chainSource.str();
2492
        string source = cl.replaceStrings(OpenCLKernelSources::customGBChainRule, replacements);
2493
2494
        vector<OpenCLNonbondedUtilities::ParameterInfo> parameters;
        vector<OpenCLNonbondedUtilities::ParameterInfo> arguments;
2495
2496
2497
        for (int i = 0; i < (int) params->getBuffers().size(); i++) {
            const OpenCLNonbondedUtilities::ParameterInfo& buffer = params->getBuffers()[i];
            string paramName = prefix+"params"+intToString(i+1);
2498
            parameters.push_back(OpenCLNonbondedUtilities::ParameterInfo(paramName, buffer.getComponentType(), buffer.getNumComponents(), buffer.getSize(), buffer.getMemory()));
2499
2500
2501
2502
        }
        for (int i = 0; i < (int) computedValues->getBuffers().size(); i++) {
            const OpenCLNonbondedUtilities::ParameterInfo& buffer = computedValues->getBuffers()[i];
            string paramName = prefix+"values"+intToString(i+1);
2503
            parameters.push_back(OpenCLNonbondedUtilities::ParameterInfo(paramName, buffer.getComponentType(), buffer.getNumComponents(), buffer.getSize(), buffer.getMemory()));
2504
2505
2506
2507
        }
        for (int i = 0; i < (int) energyDerivs->getBuffers().size(); i++) {
            const OpenCLNonbondedUtilities::ParameterInfo& buffer = energyDerivs->getBuffers()[i];
            string paramName = prefix+"dEdV"+intToString(i+1);
2508
            parameters.push_back(OpenCLNonbondedUtilities::ParameterInfo(paramName, buffer.getComponentType(), buffer.getNumComponents(), buffer.getSize(), buffer.getMemory()));
2509
2510
2511
        }
        if (globals != NULL) {
            globals->upload(globalParamValues);
2512
2513
            arguments.push_back(OpenCLNonbondedUtilities::ParameterInfo(prefix+"globals", "float", 1, sizeof(cl_float), globals->getDeviceBuffer()));
        }
Peter Eastman's avatar
Peter Eastman committed
2514
2515
2516
2517
2518
        cl.getNonbondedUtilities().addInteraction(useCutoff, usePeriodic, force.getNumExclusions() > 0, force.getCutoffDistance(), exclusionList, source);
        for (int i = 0; i < (int) parameters.size(); i++)
            cl.getNonbondedUtilities().addParameter(parameters[i]);
        for (int i = 0; i < (int) arguments.size(); i++)
            cl.getNonbondedUtilities().addArgument(arguments[i]);
2519
2520
    }
    cl.addForce(new OpenCLCustomGBForceInfo(cl.getNonbondedUtilities().getNumForceBuffers(), force));
2521
2522
2523
2524
    for (int i = 0; i < (int) energyDerivs->getBuffers().size(); i++) {
        const OpenCLNonbondedUtilities::ParameterInfo& buffer = energyDerivs->getBuffers()[i];
        cl.addAutoclearBuffer(buffer.getMemory(), buffer.getSize()*energyDerivs->getNumObjects()/sizeof(cl_float));
    }
2525
2526
    if (useLong)
        cl.addAutoclearBuffer(longEnergyDerivs->getDeviceBuffer(), 2*longEnergyDerivs->getSize());
2527
2528
}

2529
double OpenCLCalcCustomGBForceKernel::execute(ContextImpl& context, bool includeForces, bool includeEnergy) {
2530
    bool deviceIsCpu = (cl.getDevice().getInfo<CL_DEVICE_TYPE>() == CL_DEVICE_TYPE_CPU);
2531
2532
2533
    OpenCLNonbondedUtilities& nb = cl.getNonbondedUtilities();
    if (!hasInitializedKernels) {
        hasInitializedKernels = true;
2534
        maxTiles = (nb.getUseCutoff() ? nb.getInteractingTiles().getSize() : 0);
2535
2536
2537
2538
2539
2540
2541
2542
2543
2544
2545
        bool useLong = (cl.getSupports64BitGlobalAtomics() && !deviceIsCpu);
        if (useLong) {
            longValueBuffers = new OpenCLArray<cl_long>(cl, cl.getPaddedNumAtoms(), "customGBLongValueBuffers");
            cl.addAutoclearBuffer(longValueBuffers->getDeviceBuffer(), 2*longValueBuffers->getSize());
            cl.clearBuffer(longValueBuffers->getDeviceBuffer(), 2*longValueBuffers->getSize());
        }
        else {
            valueBuffers = new OpenCLArray<cl_float>(cl, cl.getPaddedNumAtoms()*nb.getNumForceBuffers(), "customGBValueBuffers");
            cl.addAutoclearBuffer(valueBuffers->getDeviceBuffer(), valueBuffers->getSize());
            cl.clearBuffer(*valueBuffers);
        }
2546
2547
        int index = 0;
        pairValueKernel.setArg<cl::Buffer>(index++, cl.getPosq().getDeviceBuffer());
2548
        pairValueKernel.setArg(index++, nb.getForceThreadBlockSize()*sizeof(cl_float4), NULL);
2549
2550
        pairValueKernel.setArg<cl::Buffer>(index++, cl.getNonbondedUtilities().getExclusions().getDeviceBuffer());
        pairValueKernel.setArg<cl::Buffer>(index++, cl.getNonbondedUtilities().getExclusionIndices().getDeviceBuffer());
2551
        pairValueKernel.setArg<cl::Buffer>(index++, cl.getNonbondedUtilities().getExclusionRowIndices().getDeviceBuffer());
2552
        pairValueKernel.setArg<cl::Buffer>(index++, useLong ? longValueBuffers->getDeviceBuffer() : valueBuffers->getDeviceBuffer());
2553
2554
        pairValueKernel.setArg(index++, (deviceIsCpu ? OpenCLContext::TileSize : nb.getForceThreadBlockSize())*sizeof(cl_float), NULL);
        pairValueKernel.setArg(index++, (deviceIsCpu ? OpenCLContext::TileSize : nb.getForceThreadBlockSize())*sizeof(cl_float), NULL);
2555
2556
2557
        if (nb.getUseCutoff()) {
            pairValueKernel.setArg<cl::Buffer>(index++, nb.getInteractingTiles().getDeviceBuffer());
            pairValueKernel.setArg<cl::Buffer>(index++, nb.getInteractionCount().getDeviceBuffer());
2558
            index += 2; // Periodic box size arguments are set when the kernel is executed.
2559
            pairValueKernel.setArg<cl_uint>(index++, maxTiles);
2560
            if (cl.getSIMDWidth() == 32 || deviceIsCpu)
2561
                pairValueKernel.setArg<cl::Buffer>(index++, nb.getInteractionFlags().getDeviceBuffer());
2562
        }
2563
2564
        else
            pairValueKernel.setArg<cl_uint>(index++, cl.getNumAtomBlocks()*(cl.getNumAtomBlocks()+1)/2);
2565
2566
2567
2568
        if (globals != NULL)
            pairValueKernel.setArg<cl::Buffer>(index++, globals->getDeviceBuffer());
        for (int i = 0; i < (int) params->getBuffers().size(); i++) {
            const OpenCLNonbondedUtilities::ParameterInfo& buffer = params->getBuffers()[i];
2569
            pairValueKernel.setArg<cl::Memory>(index++, buffer.getMemory());
2570
            pairValueKernel.setArg(index++, nb.getForceThreadBlockSize()*buffer.getSize(), NULL);
2571
        }
2572
2573
2574
2575
2576
        if (tabulatedFunctionParams != NULL) {
            for (int i = 0; i < (int) tabulatedFunctions.size(); i++)
                pairValueKernel.setArg<cl::Buffer>(index++, tabulatedFunctions[i]->getDeviceBuffer());
            pairValueKernel.setArg<cl::Buffer>(index++, tabulatedFunctionParams->getDeviceBuffer());
        }
2577
        index = 0;
2578
2579
        perParticleValueKernel.setArg<cl_int>(index++, cl.getPaddedNumAtoms());
        perParticleValueKernel.setArg<cl_int>(index++, nb.getNumForceBuffers());
2580
        perParticleValueKernel.setArg<cl::Buffer>(index++, cl.getPosq().getDeviceBuffer());
2581
        perParticleValueKernel.setArg<cl::Buffer>(index++, useLong ? longValueBuffers->getDeviceBuffer() : valueBuffers->getDeviceBuffer());
2582
        if (globals != NULL)
2583
            perParticleValueKernel.setArg<cl::Buffer>(index++, globals->getDeviceBuffer());
2584
        for (int i = 0; i < (int) params->getBuffers().size(); i++)
2585
            perParticleValueKernel.setArg<cl::Memory>(index++, params->getBuffers()[i].getMemory());
2586
        for (int i = 0; i < (int) computedValues->getBuffers().size(); i++)
2587
            perParticleValueKernel.setArg<cl::Memory>(index++, computedValues->getBuffers()[i].getMemory());
2588
2589
2590
2591
2592
2593
        if (tabulatedFunctionParams != NULL) {
            for (int i = 0; i < (int) tabulatedFunctions.size(); i++)
                perParticleValueKernel.setArg<cl::Buffer>(index++, tabulatedFunctions[i]->getDeviceBuffer());
            perParticleValueKernel.setArg<cl::Buffer>(index++, tabulatedFunctionParams->getDeviceBuffer());
        }
        index = 0;
2594
        pairEnergyKernel.setArg<cl::Buffer>(index++, useLong ? cl.getLongForceBuffer().getDeviceBuffer() : cl.getForceBuffers().getDeviceBuffer());
2595
        pairEnergyKernel.setArg<cl::Buffer>(index++, cl.getEnergyBuffer().getDeviceBuffer());
2596
        pairEnergyKernel.setArg(index++, (deviceIsCpu ? OpenCLContext::TileSize : nb.getForceThreadBlockSize())*sizeof(cl_float4), NULL);
2597
        pairEnergyKernel.setArg<cl::Buffer>(index++, cl.getPosq().getDeviceBuffer());
2598
        pairEnergyKernel.setArg(index++, (deviceIsCpu ? OpenCLContext::TileSize : nb.getForceThreadBlockSize())*sizeof(cl_float4), NULL);
2599
2600
        pairEnergyKernel.setArg<cl::Buffer>(index++, cl.getNonbondedUtilities().getExclusions().getDeviceBuffer());
        pairEnergyKernel.setArg<cl::Buffer>(index++, cl.getNonbondedUtilities().getExclusionIndices().getDeviceBuffer());
2601
        pairEnergyKernel.setArg<cl::Buffer>(index++, cl.getNonbondedUtilities().getExclusionRowIndices().getDeviceBuffer());
2602
        pairEnergyKernel.setArg(index++, (deviceIsCpu ? OpenCLContext::TileSize : nb.getForceThreadBlockSize())*sizeof(cl_float4), NULL);
2603
2604
2605
        if (nb.getUseCutoff()) {
            pairEnergyKernel.setArg<cl::Buffer>(index++, nb.getInteractingTiles().getDeviceBuffer());
            pairEnergyKernel.setArg<cl::Buffer>(index++, nb.getInteractionCount().getDeviceBuffer());
2606
            index += 2; // Periodic box size arguments are set when the kernel is executed.
2607
            pairEnergyKernel.setArg<cl_uint>(index++, maxTiles);
2608
            if (cl.getSIMDWidth() == 32 || deviceIsCpu)
2609
                pairEnergyKernel.setArg<cl::Buffer>(index++, nb.getInteractionFlags().getDeviceBuffer());
2610
        }
2611
2612
        else
            pairEnergyKernel.setArg<cl_uint>(index++, cl.getNumAtomBlocks()*(cl.getNumAtomBlocks()+1)/2);
2613
2614
2615
2616
        if (globals != NULL)
            pairEnergyKernel.setArg<cl::Buffer>(index++, globals->getDeviceBuffer());
        for (int i = 0; i < (int) params->getBuffers().size(); i++) {
            const OpenCLNonbondedUtilities::ParameterInfo& buffer = params->getBuffers()[i];
2617
            pairEnergyKernel.setArg<cl::Memory>(index++, buffer.getMemory());
2618
            pairEnergyKernel.setArg(index++, (deviceIsCpu ? OpenCLContext::TileSize : nb.getForceThreadBlockSize())*buffer.getSize(), NULL);
2619
2620
2621
        }
        for (int i = 0; i < (int) computedValues->getBuffers().size(); i++) {
            const OpenCLNonbondedUtilities::ParameterInfo& buffer = computedValues->getBuffers()[i];
2622
            pairEnergyKernel.setArg<cl::Memory>(index++, buffer.getMemory());
2623
            pairEnergyKernel.setArg(index++, (deviceIsCpu ? OpenCLContext::TileSize : nb.getForceThreadBlockSize())*buffer.getSize(), NULL);
2624
        }
2625
2626
2627
2628
2629
2630
2631
2632
2633
2634
2635
        if (useLong) {
            pairEnergyKernel.setArg<cl::Memory>(index++, longEnergyDerivs->getDeviceBuffer());
            for (int i = 0; i < numComputedValues; ++i)
                pairEnergyKernel.setArg(index++, nb.getForceThreadBlockSize()*sizeof(cl_float), NULL);
        }
        else {
            for (int i = 0; i < (int) energyDerivs->getBuffers().size(); i++) {
                const OpenCLNonbondedUtilities::ParameterInfo& buffer = energyDerivs->getBuffers()[i];
                pairEnergyKernel.setArg<cl::Memory>(index++, buffer.getMemory());
                pairEnergyKernel.setArg(index++, (deviceIsCpu ? OpenCLContext::TileSize : nb.getForceThreadBlockSize())*buffer.getSize(), NULL);
            }
2636
        }
2637
2638
2639
2640
2641
2642
2643
2644
        if (tabulatedFunctionParams != NULL) {
            for (int i = 0; i < (int) tabulatedFunctions.size(); i++)
                pairEnergyKernel.setArg<cl::Buffer>(index++, tabulatedFunctions[i]->getDeviceBuffer());
            pairEnergyKernel.setArg<cl::Buffer>(index++, tabulatedFunctionParams->getDeviceBuffer());
        }
        index = 0;
        perParticleEnergyKernel.setArg<cl_int>(index++, cl.getPaddedNumAtoms());
        perParticleEnergyKernel.setArg<cl_int>(index++, nb.getNumForceBuffers());
2645
        perParticleEnergyKernel.setArg<cl::Buffer>(index++, cl.getForceBuffers().getDeviceBuffer());
2646
        perParticleEnergyKernel.setArg<cl::Buffer>(index++, cl.getEnergyBuffer().getDeviceBuffer());
2647
        perParticleEnergyKernel.setArg<cl::Buffer>(index++, cl.getPosq().getDeviceBuffer());
2648
2649
2650
        if (globals != NULL)
            perParticleEnergyKernel.setArg<cl::Buffer>(index++, globals->getDeviceBuffer());
        for (int i = 0; i < (int) params->getBuffers().size(); i++)
2651
            perParticleEnergyKernel.setArg<cl::Memory>(index++, params->getBuffers()[i].getMemory());
2652
        for (int i = 0; i < (int) computedValues->getBuffers().size(); i++)
2653
            perParticleEnergyKernel.setArg<cl::Memory>(index++, computedValues->getBuffers()[i].getMemory());
2654
        for (int i = 0; i < (int) energyDerivs->getBuffers().size(); i++)
2655
            perParticleEnergyKernel.setArg<cl::Memory>(index++, energyDerivs->getBuffers()[i].getMemory());
2656
2657
        if (useLong)
            perParticleEnergyKernel.setArg<cl::Memory>(index++, longEnergyDerivs->getDeviceBuffer());
2658
2659
2660
2661
2662
        if (tabulatedFunctionParams != NULL) {
            for (int i = 0; i < (int) tabulatedFunctions.size(); i++)
                perParticleEnergyKernel.setArg<cl::Buffer>(index++, tabulatedFunctions[i]->getDeviceBuffer());
            perParticleEnergyKernel.setArg<cl::Buffer>(index++, tabulatedFunctionParams->getDeviceBuffer());
        }
Peter Eastman's avatar
Peter Eastman committed
2663
2664
2665
2666
2667
2668
2669
2670
2671
2672
2673
2674
2675
        if (needParameterGradient) {
            index = 0;
            gradientChainRuleKernel.setArg<cl::Buffer>(index++, cl.getForceBuffers().getDeviceBuffer());
            gradientChainRuleKernel.setArg<cl::Buffer>(index++, cl.getPosq().getDeviceBuffer());
            if (globals != NULL)
                gradientChainRuleKernel.setArg<cl::Buffer>(index++, globals->getDeviceBuffer());
            for (int i = 0; i < (int) params->getBuffers().size(); i++)
                gradientChainRuleKernel.setArg<cl::Memory>(index++, params->getBuffers()[i].getMemory());
            for (int i = 0; i < (int) computedValues->getBuffers().size(); i++)
                gradientChainRuleKernel.setArg<cl::Memory>(index++, computedValues->getBuffers()[i].getMemory());
            for (int i = 0; i < (int) energyDerivs->getBuffers().size(); i++)
                gradientChainRuleKernel.setArg<cl::Memory>(index++, energyDerivs->getBuffers()[i].getMemory());
        }
2676
2677
2678
    }
    if (globals != NULL) {
        bool changed = false;
2679
        for (int i = 0; i < (int) globalParamNames.size(); i++) {
2680
2681
2682
2683
2684
2685
2686
2687
            cl_float value = (cl_float) context.getParameter(globalParamNames[i]);
            if (value != globalParamValues[i])
                changed = true;
            globalParamValues[i] = value;
        }
        if (changed)
            globals->upload(globalParamValues);
    }
2688
    if (nb.getUseCutoff()) {
2689
2690
2691
2692
        pairValueKernel.setArg<mm_float4>(10, cl.getPeriodicBoxSize());
        pairValueKernel.setArg<mm_float4>(11, cl.getInvPeriodicBoxSize());
        pairEnergyKernel.setArg<mm_float4>(11, cl.getPeriodicBoxSize());
        pairEnergyKernel.setArg<mm_float4>(12, cl.getInvPeriodicBoxSize());
2693
2694
        if (maxTiles < nb.getInteractingTiles().getSize()) {
            maxTiles = nb.getInteractingTiles().getSize();
2695
2696
            pairValueKernel.setArg<cl_uint>(12, maxTiles);
            pairEnergyKernel.setArg<cl_uint>(13, maxTiles);
2697
        }
2698
    }
2699
    cl.executeKernel(pairValueKernel, nb.getNumForceThreadBlocks()*nb.getForceThreadBlockSize(), nb.getForceThreadBlockSize());
2700
    cl.executeKernel(perParticleValueKernel, cl.getPaddedNumAtoms());
2701
    cl.executeKernel(pairEnergyKernel, nb.getNumForceThreadBlocks()*nb.getForceThreadBlockSize(), nb.getForceThreadBlockSize());
2702
    cl.executeKernel(perParticleEnergyKernel, cl.getPaddedNumAtoms());
Peter Eastman's avatar
Peter Eastman committed
2703
2704
    if (needParameterGradient)
        cl.executeKernel(gradientChainRuleKernel, cl.getPaddedNumAtoms());
2705
2706
2707
    return 0.0;
}

2708
2709
2710
2711
2712
2713
2714
2715
2716
2717
2718
2719
2720
2721
2722
2723
2724
2725
2726
2727
2728
2729
class OpenCLCustomExternalForceInfo : public OpenCLForceInfo {
public:
    OpenCLCustomExternalForceInfo(const CustomExternalForce& force, int numParticles) : OpenCLForceInfo(1), force(force), indices(numParticles, -1) {
        vector<double> params;
        for (int i = 0; i < force.getNumParticles(); i++) {
            int particle;
            force.getParticleParameters(i, particle, params);
            indices[particle] = i;
        }
    }
    bool areParticlesIdentical(int particle1, int particle2) {
        particle1 = indices[particle1];
        particle2 = indices[particle2];
        if (particle1 == -1 && particle2 == -1)
            return true;
        if (particle1 == -1 || particle2 == -1)
            return false;
        int temp;
        vector<double> params1;
        vector<double> params2;
        force.getParticleParameters(particle1, temp, params1);
        force.getParticleParameters(particle2, temp, params2);
2730
        for (int i = 0; i < (int) params1.size(); i++)
2731
2732
2733
2734
2735
2736
2737
2738
2739
2740
2741
2742
2743
2744
2745
2746
2747
2748
2749
            if (params1[i] != params2[i])
                return false;
        return true;
    }
private:
    const CustomExternalForce& force;
    vector<int> indices;
};

OpenCLCalcCustomExternalForceKernel::~OpenCLCalcCustomExternalForceKernel() {
    if (params != NULL)
        delete params;
    if (indices != NULL)
        delete indices;
    if (globals != NULL)
        delete globals;
}

void OpenCLCalcCustomExternalForceKernel::initialize(const System& system, const CustomExternalForce& force) {
2750
2751
2752
2753
2754
2755
    int numContexts = cl.getPlatformData().contexts.size();
    int startIndex = cl.getContextIndex()*force.getNumParticles()/numContexts;
    int endIndex = (cl.getContextIndex()+1)*force.getNumParticles()/numContexts;
    numParticles = endIndex-startIndex;
    if (numParticles == 0)
        return;
2756
    params = new OpenCLParameterSet(cl, force.getNumPerParticleParameters(), numParticles, "customExternalParams");
2757
2758
2759
2760
2761
2762
    indices = new OpenCLArray<cl_int>(cl, numParticles, "customExternalIndices");
    string extraArguments;
    if (force.getNumGlobalParameters() > 0) {
        globals = new OpenCLArray<cl_float>(cl, force.getNumGlobalParameters(), "customExternalGlobals", false, CL_MEM_READ_ONLY);
        extraArguments += ", __constant float* globals";
    }
2763
    vector<vector<cl_float> > paramVector(numParticles);
2764
2765
2766
    vector<cl_int> indicesVector(numParticles);
    for (int i = 0; i < numParticles; i++) {
        vector<double> parameters;
2767
        force.getParticleParameters(startIndex+i, indicesVector[i], parameters);
2768
        paramVector[i].resize(parameters.size());
2769
        for (int j = 0; j < (int) parameters.size(); j++)
2770
            paramVector[i][j] = (cl_float) parameters[j];
2771
    }
2772
    params->setParameterValues(paramVector);
2773
2774
2775
2776
2777
2778
2779
2780
2781
2782
2783
2784
2785
2786
2787
2788
2789
2790
2791
2792
2793
2794
2795
2796
2797
2798
2799
2800
2801
2802
2803
    indices->upload(indicesVector);
    cl.addForce(new OpenCLCustomExternalForceInfo(force, system.getNumParticles()));

    // Record information for the expressions.

    globalParamNames.resize(force.getNumGlobalParameters());
    globalParamValues.resize(force.getNumGlobalParameters());
    for (int i = 0; i < force.getNumGlobalParameters(); i++) {
        globalParamNames[i] = force.getGlobalParameterName(i);
        globalParamValues[i] = (cl_float) force.getGlobalParameterDefaultValue(i);
    }
    if (globals != NULL)
        globals->upload(globalParamValues);
    Lepton::ParsedExpression energyExpression = Lepton::Parser::parse(force.getEnergyFunction()).optimize();
    Lepton::ParsedExpression forceExpressionX = energyExpression.differentiate("x").optimize();
    Lepton::ParsedExpression forceExpressionY = energyExpression.differentiate("y").optimize();
    Lepton::ParsedExpression forceExpressionZ = energyExpression.differentiate("z").optimize();
    map<string, Lepton::ParsedExpression> expressions;
    expressions["energy += "] = energyExpression;
    expressions["float dEdX = "] = forceExpressionX;
    expressions["float dEdY = "] = forceExpressionY;
    expressions["float dEdZ = "] = forceExpressionZ;

    // Create the kernels.

    map<string, string> variables;
    variables["x"] = "pos.x";
    variables["y"] = "pos.y";
    variables["z"] = "pos.z";
    for (int i = 0; i < force.getNumPerParticleParameters(); i++) {
        const string& name = force.getPerParticleParameterName(i);
2804
        variables[name] = "particleParams"+params->getParameterSuffix(i);
2805
2806
2807
2808
2809
2810
2811
    }
    for (int i = 0; i < force.getNumGlobalParameters(); i++) {
        const string& name = force.getGlobalParameterName(i);
        string value = "globals["+intToString(i)+"]";
        variables[name] = value;
    }
    stringstream compute;
2812
2813
2814
2815
2816
    for (int i = 0; i < (int) params->getBuffers().size(); i++) {
        const OpenCLNonbondedUtilities::ParameterInfo& buffer = params->getBuffers()[i];
        extraArguments += ", __global "+buffer.getType()+"* "+buffer.getName();
        compute<<buffer.getType()<<" particleParams"<<(i+1)<<" = "<<buffer.getName()<<"[index];\n";
    }
2817
2818
    vector<pair<string, string> > functions;
    compute << OpenCLExpressionUtilities::createExpressions(expressions, variables, functions, "temp", "");
2819
    map<string, string> replacements;
2820
2821
    replacements["COMPUTE_FORCE"] = compute.str();
    replacements["EXTRA_ARGUMENTS"] = extraArguments;
2822
    cl::Program program = cl.createProgram(cl.replaceStrings(OpenCLKernelSources::customExternalForce, replacements));
2823
2824
2825
    kernel = cl::Kernel(program, "computeCustomExternalForces");
}

2826
double OpenCLCalcCustomExternalForceKernel::execute(ContextImpl& context, bool includeForces, bool includeEnergy) {
2827
2828
    if (numParticles == 0)
        return 0.0;
2829
2830
    if (globals != NULL) {
        bool changed = false;
2831
        for (int i = 0; i < (int) globalParamNames.size(); i++) {
2832
2833
2834
2835
2836
2837
2838
2839
2840
2841
2842
2843
2844
2845
            cl_float value = (cl_float) context.getParameter(globalParamNames[i]);
            if (value != globalParamValues[i])
                changed = true;
            globalParamValues[i] = value;
        }
        if (changed)
            globals->upload(globalParamValues);
    }
    if (!hasInitializedKernel) {
        hasInitializedKernel = true;
        kernel.setArg<cl_int>(0, numParticles);
        kernel.setArg<cl::Buffer>(1, cl.getForceBuffers().getDeviceBuffer());
        kernel.setArg<cl::Buffer>(2, cl.getEnergyBuffer().getDeviceBuffer());
        kernel.setArg<cl::Buffer>(3, cl.getPosq().getDeviceBuffer());
2846
2847
        kernel.setArg<cl::Buffer>(4, indices->getDeviceBuffer());
        int nextIndex = 5;
2848
        if (globals != NULL)
2849
2850
2851
            kernel.setArg<cl::Buffer>(nextIndex++, globals->getDeviceBuffer());
        for (int i = 0; i < (int) params->getBuffers().size(); i++) {
            const OpenCLNonbondedUtilities::ParameterInfo& buffer = params->getBuffers()[i];
2852
            kernel.setArg<cl::Memory>(nextIndex++, buffer.getMemory());
2853
        }
2854
2855
2856
    }
    cl.executeKernel(kernel, numParticles);
    return 0.0;
2857
}
2858

2859
2860
2861
2862
2863
2864
2865
2866
2867
2868
2869
2870
2871
2872
2873
class OpenCLCustomHbondForceInfo : public OpenCLForceInfo {
public:
    OpenCLCustomHbondForceInfo(int requiredBuffers, const CustomHbondForce& force) : OpenCLForceInfo(requiredBuffers), force(force) {
    }
    bool areParticlesIdentical(int particle1, int particle2) {
        return true;
    }
    int getNumParticleGroups() {
        return force.getNumDonors()+force.getNumAcceptors()+force.getNumExclusions();
    }
    void getParticlesInGroup(int index, std::vector<int>& particles) {
        int p1, p2, p3;
        vector<double> parameters;
        if (index < force.getNumDonors()) {
            force.getDonorParameters(index, p1, p2, p3, parameters);
2874
2875
2876
2877
2878
2879
            particles.clear();
            particles.push_back(p1);
            if (p2 > -1)
                particles.push_back(p2);
            if (p3 > -1)
                particles.push_back(p3);
2880
2881
2882
2883
2884
            return;
        }
        index -= force.getNumDonors();
        if (index < force.getNumAcceptors()) {
            force.getAcceptorParameters(index, p1, p2, p3, parameters);
2885
2886
2887
2888
2889
2890
            particles.clear();
            particles.push_back(p1);
            if (p2 > -1)
                particles.push_back(p2);
            if (p3 > -1)
                particles.push_back(p3);
2891
2892
2893
2894
2895
            return;
        }
        index -= force.getNumAcceptors();
        int donor, acceptor;
        force.getExclusionParticles(index, donor, acceptor);
2896
        particles.clear();
2897
        force.getDonorParameters(donor, p1, p2, p3, parameters);
2898
2899
2900
2901
2902
        particles.push_back(p1);
        if (p2 > -1)
            particles.push_back(p2);
        if (p3 > -1)
            particles.push_back(p3);
2903
        force.getAcceptorParameters(acceptor, p1, p2, p3, parameters);
2904
2905
2906
2907
2908
        particles.push_back(p1);
        if (p2 > -1)
            particles.push_back(p2);
        if (p3 > -1)
            particles.push_back(p3);
2909
2910
2911
2912
2913
2914
2915
2916
2917
2918
2919
2920
2921
2922
2923
2924
2925
2926
2927
2928
2929
2930
2931
2932
2933
2934
2935
2936
2937
2938
2939
2940
2941
2942
2943
2944
2945
2946
2947
2948
2949
    }
    bool areGroupsIdentical(int group1, int group2) {
        int p1, p2, p3;
        vector<double> params1, params2;
        if (group1 < force.getNumDonors() && group2 < force.getNumDonors()) {
            force.getDonorParameters(group1, p1, p2, p3, params1);
            force.getDonorParameters(group2, p1, p2, p3, params2);
            return (params1 == params2 && params1 == params2);
        }
        if (group1 < force.getNumDonors() || group2 < force.getNumDonors())
            return false;
        group1 -= force.getNumDonors();
        group2 -= force.getNumDonors();
        if (group1 < force.getNumAcceptors() && group2 < force.getNumAcceptors()) {
            force.getAcceptorParameters(group1, p1, p2, p3, params1);
            force.getAcceptorParameters(group2, p1, p2, p3, params2);
            return (params1 == params2 && params1 == params2);
        }
        if (group1 < force.getNumAcceptors() || group2 < force.getNumAcceptors())
            return false;
        return true;
    }
private:
    const CustomHbondForce& force;
};

OpenCLCalcCustomHbondForceKernel::~OpenCLCalcCustomHbondForceKernel() {
    if (donorParams != NULL)
        delete donorParams;
    if (acceptorParams != NULL)
        delete acceptorParams;
    if (donors != NULL)
        delete donors;
    if (acceptors != NULL)
        delete acceptors;
    if (donorBufferIndices != NULL)
        delete donorBufferIndices;
    if (acceptorBufferIndices != NULL)
        delete acceptorBufferIndices;
    if (globals != NULL)
        delete globals;
2950
2951
2952
2953
    if (donorExclusions != NULL)
        delete donorExclusions;
    if (acceptorExclusions != NULL)
        delete acceptorExclusions;
2954
2955
2956
2957
2958
2959
    if (tabulatedFunctionParams != NULL)
        delete tabulatedFunctionParams;
    for (int i = 0; i < (int) tabulatedFunctions.size(); i++)
        delete tabulatedFunctions[i];
}

2960
2961
2962
2963
2964
2965
2966
2967
2968
2969
2970
2971
static void addDonorAndAcceptorCode(stringstream& computeDonor, stringstream& computeAcceptor, const string& value) {
    computeDonor << value;
    computeAcceptor << value;
}

static void applyDonorAndAcceptorForces(stringstream& applyToDonor, stringstream& applyToAcceptor, int atom, const string& value) {
    string forceNames[] = {"f1", "f2", "f3"};
    if (atom < 3)
        applyToAcceptor << forceNames[atom]<<".xyz += "<<value<<";\n";
    else
        applyToDonor << forceNames[atom-3]<<".xyz += "<<value<<";\n";
}
2972

2973
void OpenCLCalcCustomHbondForceKernel::initialize(const System& system, const CustomHbondForce& force) {
2974
2975
    // Record the lists of donors and acceptors, and the parameters for each one.

2976
2977
2978
2979
    int numContexts = cl.getPlatformData().contexts.size();
    int startIndex = cl.getContextIndex()*force.getNumDonors()/numContexts;
    int endIndex = (cl.getContextIndex()+1)*force.getNumDonors()/numContexts;
    numDonors = endIndex-startIndex;
2980
    numAcceptors = force.getNumAcceptors();
2981
2982
    if (numDonors == 0 || numAcceptors == 0)
        return;
2983
2984
2985
2986
2987
2988
2989
2990
2991
2992
2993
    int numParticles = system.getNumParticles();
    donors = new OpenCLArray<mm_int4>(cl, numDonors, "customHbondDonors");
    acceptors = new OpenCLArray<mm_int4>(cl, numAcceptors, "customHbondAcceptors");
    donorParams = new OpenCLParameterSet(cl, force.getNumPerDonorParameters(), numDonors, "customHbondDonorParameters");
    acceptorParams = new OpenCLParameterSet(cl, force.getNumPerAcceptorParameters(), numAcceptors, "customHbondAcceptorParameters");
    if (force.getNumGlobalParameters() > 0)
        globals = new OpenCLArray<cl_float>(cl, force.getNumGlobalParameters(), "customHbondGlobals", false, CL_MEM_READ_ONLY);
    vector<vector<cl_float> > donorParamVector(numDonors);
    vector<mm_int4> donorVector(numDonors);
    for (int i = 0; i < numDonors; i++) {
        vector<double> parameters;
2994
        force.getDonorParameters(startIndex+i, donorVector[i].x, donorVector[i].y, donorVector[i].z, parameters);
2995
2996
2997
2998
2999
3000
3001
3002
3003
3004
3005
3006
3007
3008
3009
3010
3011
3012
        donorParamVector[i].resize(parameters.size());
        for (int j = 0; j < (int) parameters.size(); j++)
            donorParamVector[i][j] = (cl_float) parameters[j];
    }
    donors->upload(donorVector);
    donorParams->setParameterValues(donorParamVector);
    vector<vector<cl_float> > acceptorParamVector(numAcceptors);
    vector<mm_int4> acceptorVector(numAcceptors);
    for (int i = 0; i < numAcceptors; i++) {
        vector<double> parameters;
        force.getAcceptorParameters(i, acceptorVector[i].x, acceptorVector[i].y, acceptorVector[i].z, parameters);
        acceptorParamVector[i].resize(parameters.size());
        for (int j = 0; j < (int) parameters.size(); j++)
            acceptorParamVector[i][j] = (cl_float) parameters[j];
    }
    acceptors->upload(acceptorVector);
    acceptorParams->setParameterValues(acceptorParamVector);

3013
    // Select an output buffer index for each donor and acceptor.
3014
3015
3016
3017
3018

    donorBufferIndices = new OpenCLArray<mm_int4>(cl, numDonors, "customHbondDonorBuffers");
    acceptorBufferIndices = new OpenCLArray<mm_int4>(cl, numAcceptors, "customHbondAcceptorBuffers");
    vector<mm_int4> donorBufferVector(numDonors);
    vector<mm_int4> acceptorBufferVector(numAcceptors);
3019
    vector<int> donorBufferCounter(numParticles, 0);
3020
    for (int i = 0; i < numDonors; i++)
3021
3022
3023
        donorBufferVector[i] = mm_int4(donorVector[i].x > -1 ? donorBufferCounter[donorVector[i].x]++ : 0,
                                       donorVector[i].y > -1 ? donorBufferCounter[donorVector[i].y]++ : 0,
                                       donorVector[i].z > -1 ? donorBufferCounter[donorVector[i].z]++ : 0, 0);
3024
    vector<int> acceptorBufferCounter(numParticles, 0);
3025
    for (int i = 0; i < numAcceptors; i++)
3026
3027
3028
        acceptorBufferVector[i] = mm_int4(acceptorVector[i].x > -1 ? acceptorBufferCounter[acceptorVector[i].x]++ : 0,
                                       acceptorVector[i].y > -1 ? acceptorBufferCounter[acceptorVector[i].y]++ : 0,
                                       acceptorVector[i].z > -1 ? acceptorBufferCounter[acceptorVector[i].z]++ : 0, 0);
3029
3030
    donorBufferIndices->upload(donorBufferVector);
    acceptorBufferIndices->upload(acceptorBufferVector);
3031
3032
3033
3034
3035
3036
    int maxBuffers = 1;
    for (int i = 0; i < (int) donorBufferCounter.size(); i++)
        maxBuffers = max(maxBuffers, donorBufferCounter[i]);
    for (int i = 0; i < (int) acceptorBufferCounter.size(); i++)
        maxBuffers = max(maxBuffers, acceptorBufferCounter[i]);
    cl.addForce(new OpenCLCustomHbondForceInfo(maxBuffers, force));
3037
3038
3039

    // Record exclusions.

3040
3041
    vector<mm_int4> donorExclusionVector(numDonors, mm_int4(-1, -1, -1, -1));
    vector<mm_int4> acceptorExclusionVector(numAcceptors, mm_int4(-1, -1, -1, -1));
3042
3043
3044
    for (int i = 0; i < force.getNumExclusions(); i++) {
        int donor, acceptor;
        force.getExclusionParticles(i, donor, acceptor);
3045
3046
3047
        if (donor < startIndex || donor >= endIndex)
            continue;
        donor -= startIndex;
3048
3049
3050
3051
3052
3053
3054
3055
3056
3057
3058
3059
3060
3061
3062
3063
3064
3065
3066
3067
        if (donorExclusionVector[donor].x == -1)
            donorExclusionVector[donor].x = acceptor;
        else if (donorExclusionVector[donor].y == -1)
            donorExclusionVector[donor].y = acceptor;
        else if (donorExclusionVector[donor].z == -1)
            donorExclusionVector[donor].z = acceptor;
        else if (donorExclusionVector[donor].w == -1)
            donorExclusionVector[donor].w = acceptor;
        else
            throw OpenMMException("CustomHbondForce: OpenCLPlatform does not support more than four exclusions per donor");
        if (acceptorExclusionVector[acceptor].x == -1)
            acceptorExclusionVector[acceptor].x = donor;
        else if (acceptorExclusionVector[acceptor].y == -1)
            acceptorExclusionVector[acceptor].y = donor;
        else if (acceptorExclusionVector[acceptor].z == -1)
            acceptorExclusionVector[acceptor].z = donor;
        else if (acceptorExclusionVector[acceptor].w == -1)
            acceptorExclusionVector[acceptor].w = donor;
        else
            throw OpenMMException("CustomHbondForce: OpenCLPlatform does not support more than four exclusions per acceptor");
3068
    }
3069
3070
3071
3072
    donorExclusions = new OpenCLArray<mm_int4>(cl, numDonors, "customHbondDonorExclusions");
    acceptorExclusions = new OpenCLArray<mm_int4>(cl, numDonors, "customHbondAcceptorExclusions");
    donorExclusions->upload(donorExclusionVector);
    acceptorExclusions->upload(acceptorExclusionVector);
3073
3074
3075
3076
3077
3078
3079
3080
3081
3082
3083
3084

    // Record the tabulated functions.

    OpenCLExpressionUtilities::FunctionPlaceholder fp;
    map<string, Lepton::CustomFunction*> functions;
    vector<pair<string, string> > functionDefinitions;
    vector<mm_float4> tabulatedFunctionParamsVec(force.getNumFunctions());
    stringstream tableArgs;
    for (int i = 0; i < force.getNumFunctions(); i++) {
        string name;
        vector<double> values;
        double min, max;
3085
        force.getFunctionParameters(i, name, values, min, max);
3086
        string arrayName = "table"+intToString(i);
3087
3088
        functionDefinitions.push_back(make_pair(name, arrayName));
        functions[name] = &fp;
3089
        tabulatedFunctionParamsVec[i] = mm_float4((float) min, (float) max, (float) ((values.size()-1)/(max-min)), (float) values.size()-2);
3090
        vector<mm_float4> f = OpenCLExpressionUtilities::computeFunctionCoefficients(values, min, max);
3091
3092
3093
3094
3095
3096
3097
        tabulatedFunctions.push_back(new OpenCLArray<mm_float4>(cl, values.size()-1, "TabulatedFunction"));
        tabulatedFunctions[tabulatedFunctions.size()-1]->upload(f);
        tableArgs << ", __global float4* " << arrayName;
    }
    if (force.getNumFunctions() > 0) {
        tabulatedFunctionParams = new OpenCLArray<mm_float4>(cl, tabulatedFunctionParamsVec.size(), "tabulatedFunctionParameters", false, CL_MEM_READ_ONLY);
        tabulatedFunctionParams->upload(tabulatedFunctionParamsVec);
3098
        tableArgs << ", __global float4* functionParams";
3099
3100
    }

3101
    // Record information about parameters.
3102
3103
3104
3105
3106
3107
3108
3109
3110
3111
3112
3113
3114
3115
3116
3117
3118
3119
3120
3121
3122
3123

    globalParamNames.resize(force.getNumGlobalParameters());
    globalParamValues.resize(force.getNumGlobalParameters());
    for (int i = 0; i < force.getNumGlobalParameters(); i++) {
        globalParamNames[i] = force.getGlobalParameterName(i);
        globalParamValues[i] = (cl_float) force.getGlobalParameterDefaultValue(i);
    }
    if (globals != NULL)
        globals->upload(globalParamValues);
    map<string, string> variables;
    for (int i = 0; i < force.getNumPerDonorParameters(); i++) {
        const string& name = force.getPerDonorParameterName(i);
        variables[name] = "donorParams"+donorParams->getParameterSuffix(i);
    }
    for (int i = 0; i < force.getNumPerAcceptorParameters(); i++) {
        const string& name = force.getPerAcceptorParameterName(i);
        variables[name] = "acceptorParams"+acceptorParams->getParameterSuffix(i);
    }
    for (int i = 0; i < force.getNumGlobalParameters(); i++) {
        const string& name = force.getGlobalParameterName(i);
        variables[name] = "globals["+intToString(i)+"]";
    }
3124
3125
3126
3127
3128
3129
3130
3131
3132
3133
3134
3135
3136
3137
3138
3139
3140
3141
3142
3143
3144
3145
3146
3147
3148
3149
3150
3151
3152
3153
3154
3155
3156
3157
3158
3159
3160
3161
3162
3163
3164
3165
3166
3167
3168
3169
3170
3171
3172
3173
3174
3175
3176
3177
3178
3179
3180
3181
3182
3183
3184
3185
3186
3187
3188
3189
3190
3191
3192
3193
3194
3195
3196
3197
3198

    // Now to generate the kernel.  First, it needs to calculate all distances, angles,
    // and dihedrals the expression depends on.

    map<string, vector<int> > distances;
    map<string, vector<int> > angles;
    map<string, vector<int> > dihedrals;
    Lepton::ParsedExpression energyExpression = CustomHbondForceImpl::prepareExpression(force, functions, distances, angles, dihedrals);
    map<string, Lepton::ParsedExpression> forceExpressions;
    set<string> computedDeltas;
    computedDeltas.insert("D1A1");
    string atomNames[] = {"A1", "A2", "A3", "D1", "D2", "D3"};
    string atomNamesLower[] = {"a1", "a2", "a3", "d1", "d2", "d3"};
    stringstream computeDonor, computeAcceptor, extraArgs;
    int index = 0;
    for (map<string, vector<int> >::const_iterator iter = distances.begin(); iter != distances.end(); ++iter, ++index) {
        const vector<int>& atoms = iter->second;
        string deltaName = atomNames[atoms[0]]+atomNames[atoms[1]];
        if (computedDeltas.count(deltaName) == 0) {
            addDonorAndAcceptorCode(computeDonor, computeAcceptor, "float4 delta"+deltaName+" = delta("+atomNamesLower[atoms[0]]+", "+atomNamesLower[atoms[1]]+");\n");
            computedDeltas.insert(deltaName);
        }
        addDonorAndAcceptorCode(computeDonor, computeAcceptor, "float r_"+deltaName+" = sqrt(delta"+deltaName+".w);\n");
        variables[iter->first] = "r_"+deltaName;
        forceExpressions["float dEdDistance"+intToString(index)+" = "] = energyExpression.differentiate(iter->first).optimize();
    }
    index = 0;
    for (map<string, vector<int> >::const_iterator iter = angles.begin(); iter != angles.end(); ++iter, ++index) {
        const vector<int>& atoms = iter->second;
        string deltaName1 = atomNames[atoms[1]]+atomNames[atoms[0]];
        string deltaName2 = atomNames[atoms[1]]+atomNames[atoms[2]];
        string angleName = "angle_"+atomNames[atoms[0]]+atomNames[atoms[1]]+atomNames[atoms[2]];
        if (computedDeltas.count(deltaName1) == 0) {
            addDonorAndAcceptorCode(computeDonor, computeAcceptor, "float4 delta"+deltaName1+" = delta("+atomNamesLower[atoms[1]]+", "+atomNamesLower[atoms[0]]+");\n");
            computedDeltas.insert(deltaName1);
        }
        if (computedDeltas.count(deltaName2) == 0) {
            addDonorAndAcceptorCode(computeDonor, computeAcceptor, "float4 delta"+deltaName2+" = delta("+atomNamesLower[atoms[1]]+", "+atomNamesLower[atoms[2]]+");\n");
            computedDeltas.insert(deltaName2);
        }
        addDonorAndAcceptorCode(computeDonor, computeAcceptor, "float "+angleName+" = computeAngle(delta"+deltaName1+", delta"+deltaName2+");\n");
        variables[iter->first] = angleName;
        forceExpressions["float dEdAngle"+intToString(index)+" = "] = energyExpression.differentiate(iter->first).optimize();
    }
    index = 0;
    for (map<string, vector<int> >::const_iterator iter = dihedrals.begin(); iter != dihedrals.end(); ++iter, ++index) {
        const vector<int>& atoms = iter->second;
        string deltaName1 = atomNames[atoms[0]]+atomNames[atoms[1]];
        string deltaName2 = atomNames[atoms[2]]+atomNames[atoms[1]];
        string deltaName3 = atomNames[atoms[2]]+atomNames[atoms[3]];
        string crossName1 = "cross_"+deltaName1+"_"+deltaName2;
        string crossName2 = "cross_"+deltaName2+"_"+deltaName3;
        string dihedralName = "dihedral_"+atomNames[atoms[0]]+atomNames[atoms[1]]+atomNames[atoms[2]]+atomNames[atoms[3]];
        if (computedDeltas.count(deltaName1) == 0) {
            addDonorAndAcceptorCode(computeDonor, computeAcceptor, "float4 delta"+deltaName1+" = delta("+atomNamesLower[atoms[0]]+", "+atomNamesLower[atoms[1]]+");\n");
            computedDeltas.insert(deltaName1);
        }
        if (computedDeltas.count(deltaName2) == 0) {
            addDonorAndAcceptorCode(computeDonor, computeAcceptor, "float4 delta"+deltaName2+" = delta("+atomNamesLower[atoms[2]]+", "+atomNamesLower[atoms[1]]+");\n");
            computedDeltas.insert(deltaName2);
        }
        if (computedDeltas.count(deltaName3) == 0) {
            addDonorAndAcceptorCode(computeDonor, computeAcceptor, "float4 delta"+deltaName3+" = delta("+atomNamesLower[atoms[2]]+", "+atomNamesLower[atoms[3]]+");\n");
            computedDeltas.insert(deltaName3);
        }
        addDonorAndAcceptorCode(computeDonor, computeAcceptor, "float4 "+crossName1+" = computeCross(delta"+deltaName1+", delta"+deltaName2+");\n");
        addDonorAndAcceptorCode(computeDonor, computeAcceptor, "float4 "+crossName2+" = computeCross(delta"+deltaName2+", delta"+deltaName3+");\n");
        addDonorAndAcceptorCode(computeDonor, computeAcceptor, "float "+dihedralName+" = computeAngle("+crossName1+", "+crossName2+");\n");
        addDonorAndAcceptorCode(computeDonor, computeAcceptor, dihedralName+" *= (delta"+deltaName1+".x*"+crossName2+".x + delta"+deltaName1+".y*"+crossName2+".y + delta"+deltaName1+".z*"+crossName2+".z < 0 ? -1 : 1);\n");
        variables[iter->first] = dihedralName;
        forceExpressions["float dEdDihedral"+intToString(index)+" = "] = energyExpression.differentiate(iter->first).optimize();
    }

    // Next it needs to load parameters from global memory.

3199
    if (force.getNumGlobalParameters() > 0)
3200
        extraArgs << ", __global float* globals";
3201
3202
3203
    for (int i = 0; i < (int) donorParams->getBuffers().size(); i++) {
        const OpenCLNonbondedUtilities::ParameterInfo& buffer = donorParams->getBuffers()[i];
        extraArgs << ", __global "+buffer.getType()+"* donor"+buffer.getName();
3204
        addDonorAndAcceptorCode(computeDonor, computeAcceptor, buffer.getType()+" donorParams"+intToString(i+1)+" = donor"+buffer.getName()+"[index];\n");
3205
3206
3207
3208
    }
    for (int i = 0; i < (int) acceptorParams->getBuffers().size(); i++) {
        const OpenCLNonbondedUtilities::ParameterInfo& buffer = acceptorParams->getBuffers()[i];
        extraArgs << ", __global "+buffer.getType()+"* acceptor"+buffer.getName();
3209
        addDonorAndAcceptorCode(computeDonor, computeAcceptor, buffer.getType()+" acceptorParams"+intToString(i+1)+" = acceptor"+buffer.getName()+"[index];\n");
3210
    }
3211
3212
3213
3214
3215
3216
3217
3218
3219
3220
3221
3222
3223
3224
3225
3226
3227
3228
3229
3230
3231
3232
3233
3234
3235
3236
3237
3238
3239
3240
3241
3242
3243
3244
3245
3246
3247
3248
3249
3250
3251
3252
3253
3254
3255
3256
3257
3258
3259
3260
3261
3262
3263
3264
3265
3266
3267
3268
3269
3270

    // Now evaluate the expressions.

    computeAcceptor << OpenCLExpressionUtilities::createExpressions(forceExpressions, variables, functionDefinitions, "temp", "functionParams");
    forceExpressions["energy += "] = energyExpression;
    computeDonor << OpenCLExpressionUtilities::createExpressions(forceExpressions, variables, functionDefinitions, "temp", "functionParams");

    // Finally, apply forces to atoms.

    index = 0;
    for (map<string, vector<int> >::const_iterator iter = distances.begin(); iter != distances.end(); ++iter, ++index) {
        const vector<int>& atoms = iter->second;
        string deltaName = atomNames[atoms[0]]+atomNames[atoms[1]];
        string value = "(dEdDistance"+intToString(index)+"/r_"+deltaName+")*delta"+deltaName+".xyz";
        applyDonorAndAcceptorForces(computeDonor, computeAcceptor, atoms[0], "-"+value);
        applyDonorAndAcceptorForces(computeDonor, computeAcceptor, atoms[1], value);
    }
    index = 0;
    for (map<string, vector<int> >::const_iterator iter = angles.begin(); iter != angles.end(); ++iter, ++index) {
        const vector<int>& atoms = iter->second;
        string deltaName1 = atomNames[atoms[1]]+atomNames[atoms[0]];
        string deltaName2 = atomNames[atoms[1]]+atomNames[atoms[2]];
        addDonorAndAcceptorCode(computeDonor, computeAcceptor, "{\n");
        addDonorAndAcceptorCode(computeDonor, computeAcceptor, "float4 crossProd = cross(delta"+deltaName2+", delta"+deltaName1+");\n");
        addDonorAndAcceptorCode(computeDonor, computeAcceptor, "float lengthCross = max(length(crossProd), 1e-6f);\n");
        addDonorAndAcceptorCode(computeDonor, computeAcceptor, "float4 deltaCross0 = -cross(delta"+deltaName1+", crossProd)*dEdAngle"+intToString(index)+"/(delta"+deltaName1+".w*lengthCross);\n");
        addDonorAndAcceptorCode(computeDonor, computeAcceptor, "float4 deltaCross2 = cross(delta"+deltaName2+", crossProd)*dEdAngle"+intToString(index)+"/(delta"+deltaName2+".w*lengthCross);\n");
        addDonorAndAcceptorCode(computeDonor, computeAcceptor, "float4 deltaCross1 = -(deltaCross0+deltaCross2);\n");
        applyDonorAndAcceptorForces(computeDonor, computeAcceptor, atoms[0], "deltaCross0.xyz");
        applyDonorAndAcceptorForces(computeDonor, computeAcceptor, atoms[1], "deltaCross1.xyz");
        applyDonorAndAcceptorForces(computeDonor, computeAcceptor, atoms[2], "deltaCross2.xyz");
        addDonorAndAcceptorCode(computeDonor, computeAcceptor, "}\n");
    }
    index = 0;
    for (map<string, vector<int> >::const_iterator iter = dihedrals.begin(); iter != dihedrals.end(); ++iter, ++index) {
        const vector<int>& atoms = iter->second;
        string deltaName1 = atomNames[atoms[0]]+atomNames[atoms[1]];
        string deltaName2 = atomNames[atoms[2]]+atomNames[atoms[1]];
        string deltaName3 = atomNames[atoms[2]]+atomNames[atoms[3]];
        string crossName1 = "cross_"+deltaName1+"_"+deltaName2;
        string crossName2 = "cross_"+deltaName2+"_"+deltaName3;
        addDonorAndAcceptorCode(computeDonor, computeAcceptor, "{\n");
        addDonorAndAcceptorCode(computeDonor, computeAcceptor, "float r = sqrt(delta"+deltaName2+".w);\n");
        addDonorAndAcceptorCode(computeDonor, computeAcceptor, "float4 ff;\n");
        addDonorAndAcceptorCode(computeDonor, computeAcceptor, "ff.x = (-dEdDihedral"+intToString(index)+"*r)/"+crossName1+".w;\n");
        addDonorAndAcceptorCode(computeDonor, computeAcceptor, "ff.y = (delta"+deltaName1+".x*delta"+deltaName2+".x + delta"+deltaName1+".y*delta"+deltaName2+".y + delta"+deltaName1+".z*delta"+deltaName2+".z)/delta"+deltaName2+".w;\n");
        addDonorAndAcceptorCode(computeDonor, computeAcceptor, "ff.z = (delta"+deltaName3+".x*delta"+deltaName2+".x + delta"+deltaName3+".y*delta"+deltaName2+".y + delta"+deltaName3+".z*delta"+deltaName2+".z)/delta"+deltaName2+".w;\n");
        addDonorAndAcceptorCode(computeDonor, computeAcceptor, "ff.w = (dEdDihedral"+intToString(index)+"*r)/"+crossName2+".w;\n");
        addDonorAndAcceptorCode(computeDonor, computeAcceptor, "float4 internalF0 = ff.x*"+crossName1+";\n");
        addDonorAndAcceptorCode(computeDonor, computeAcceptor, "float4 internalF3 = ff.w*"+crossName2+";\n");
        addDonorAndAcceptorCode(computeDonor, computeAcceptor, "float4 s = ff.y*internalF0 - ff.z*internalF3;\n");
        applyDonorAndAcceptorForces(computeDonor, computeAcceptor, atoms[0], "internalF0.xyz");
        applyDonorAndAcceptorForces(computeDonor, computeAcceptor, atoms[1], "s.xyz-internalF0.xyz");
        applyDonorAndAcceptorForces(computeDonor, computeAcceptor, atoms[2], "-s.xyz-internalF3.xyz");
        applyDonorAndAcceptorForces(computeDonor, computeAcceptor, atoms[3], "internalF3.xyz");
        addDonorAndAcceptorCode(computeDonor, computeAcceptor, "}\n");
    }

    // Generate the kernels.

3271
    map<string, string> replacements;
3272
3273
    replacements["COMPUTE_DONOR_FORCE"] = computeDonor.str();
    replacements["COMPUTE_ACCEPTOR_FORCE"] = computeAcceptor.str();
3274
3275
3276
    replacements["PARAMETER_ARGUMENTS"] = extraArgs.str()+tableArgs.str();
    map<string, string> defines;
    defines["PADDED_NUM_ATOMS"] = intToString(cl.getPaddedNumAtoms());
3277
3278
    defines["NUM_DONORS"] = intToString(numDonors);
    defines["NUM_ACCEPTORS"] = intToString(numAcceptors);
3279
    defines["M_PI"] = doubleToString(M_PI);
3280
3281
3282
3283
3284
3285
    if (force.getNonbondedMethod() != CustomHbondForce::NoCutoff) {
        defines["USE_CUTOFF"] = "1";
        defines["CUTOFF_SQUARED"] = doubleToString(force.getCutoffDistance()*force.getCutoffDistance());
    }
    if (force.getNonbondedMethod() != CustomHbondForce::NoCutoff && force.getNonbondedMethod() != CustomHbondForce::CutoffNonPeriodic)
        defines["USE_PERIODIC"] = "1";
3286
3287
    if (force.getNumExclusions() > 0)
        defines["USE_EXCLUSIONS"] = "1";
3288
    cl::Program program = cl.createProgram(cl.replaceStrings(OpenCLKernelSources::customHbondForce, replacements), defines);
3289
3290
    donorKernel = cl::Kernel(program, "computeDonorForces");
    acceptorKernel = cl::Kernel(program, "computeAcceptorForces");
3291
3292
}

3293
double OpenCLCalcCustomHbondForceKernel::execute(ContextImpl& context, bool includeForces, bool includeEnergy) {
3294
3295
    if (numDonors == 0 || numAcceptors == 0)
        return 0.0;
3296
3297
3298
3299
3300
3301
3302
3303
3304
3305
3306
3307
3308
3309
    if (globals != NULL) {
        bool changed = false;
        for (int i = 0; i < (int) globalParamNames.size(); i++) {
            cl_float value = (cl_float) context.getParameter(globalParamNames[i]);
            if (value != globalParamValues[i])
                changed = true;
            globalParamValues[i] = value;
        }
        if (changed)
            globals->upload(globalParamValues);
    }
    if (!hasInitializedKernel) {
        hasInitializedKernel = true;
        int index = 0;
3310
3311
3312
        donorKernel.setArg<cl::Buffer>(index++, cl.getForceBuffers().getDeviceBuffer());
        donorKernel.setArg<cl::Buffer>(index++, cl.getEnergyBuffer().getDeviceBuffer());
        donorKernel.setArg<cl::Buffer>(index++, cl.getPosq().getDeviceBuffer());
3313
        donorKernel.setArg<cl::Buffer>(index++, donorExclusions->getDeviceBuffer());
3314
3315
3316
3317
        donorKernel.setArg<cl::Buffer>(index++, donors->getDeviceBuffer());
        donorKernel.setArg<cl::Buffer>(index++, acceptors->getDeviceBuffer());
        donorKernel.setArg<cl::Buffer>(index++, donorBufferIndices->getDeviceBuffer());
        donorKernel.setArg(index++, 3*OpenCLContext::ThreadBlockSize*sizeof(mm_float4), NULL);
3318
        index += 2; // Periodic box size arguments are set when the kernel is executed.
3319
        if (globals != NULL)
3320
            donorKernel.setArg<cl::Buffer>(index++, globals->getDeviceBuffer());
3321
3322
        for (int i = 0; i < (int) donorParams->getBuffers().size(); i++) {
            const OpenCLNonbondedUtilities::ParameterInfo& buffer = donorParams->getBuffers()[i];
3323
            donorKernel.setArg<cl::Memory>(index++, buffer.getMemory());
3324
3325
3326
        }
        for (int i = 0; i < (int) acceptorParams->getBuffers().size(); i++) {
            const OpenCLNonbondedUtilities::ParameterInfo& buffer = acceptorParams->getBuffers()[i];
3327
            donorKernel.setArg<cl::Memory>(index++, buffer.getMemory());
3328
3329
3330
3331
3332
3333
3334
3335
3336
3337
        }
        if (tabulatedFunctionParams != NULL) {
            for (int i = 0; i < (int) tabulatedFunctions.size(); i++)
                donorKernel.setArg<cl::Buffer>(index++, tabulatedFunctions[i]->getDeviceBuffer());
            donorKernel.setArg<cl::Buffer>(index++, tabulatedFunctionParams->getDeviceBuffer());
        }
        index = 0;
        acceptorKernel.setArg<cl::Buffer>(index++, cl.getForceBuffers().getDeviceBuffer());
        acceptorKernel.setArg<cl::Buffer>(index++, cl.getEnergyBuffer().getDeviceBuffer());
        acceptorKernel.setArg<cl::Buffer>(index++, cl.getPosq().getDeviceBuffer());
3338
        acceptorKernel.setArg<cl::Buffer>(index++, acceptorExclusions->getDeviceBuffer());
3339
3340
3341
3342
        acceptorKernel.setArg<cl::Buffer>(index++, donors->getDeviceBuffer());
        acceptorKernel.setArg<cl::Buffer>(index++, acceptors->getDeviceBuffer());
        acceptorKernel.setArg<cl::Buffer>(index++, acceptorBufferIndices->getDeviceBuffer());
        acceptorKernel.setArg(index++, 3*OpenCLContext::ThreadBlockSize*sizeof(mm_float4), NULL);
3343
        index += 2; // Periodic box size arguments are set when the kernel is executed.
3344
3345
3346
3347
        if (globals != NULL)
            acceptorKernel.setArg<cl::Buffer>(index++, globals->getDeviceBuffer());
        for (int i = 0; i < (int) donorParams->getBuffers().size(); i++) {
            const OpenCLNonbondedUtilities::ParameterInfo& buffer = donorParams->getBuffers()[i];
3348
            acceptorKernel.setArg<cl::Memory>(index++, buffer.getMemory());
3349
3350
3351
        }
        for (int i = 0; i < (int) acceptorParams->getBuffers().size(); i++) {
            const OpenCLNonbondedUtilities::ParameterInfo& buffer = acceptorParams->getBuffers()[i];
3352
            acceptorKernel.setArg<cl::Memory>(index++, buffer.getMemory());
3353
3354
3355
3356
3357
        }
        if (tabulatedFunctionParams != NULL) {
            for (int i = 0; i < (int) tabulatedFunctions.size(); i++)
                acceptorKernel.setArg<cl::Buffer>(index++, tabulatedFunctions[i]->getDeviceBuffer());
            acceptorKernel.setArg<cl::Buffer>(index++, tabulatedFunctionParams->getDeviceBuffer());
3358
3359
        }
    }
3360
3361
    donorKernel.setArg<mm_float4>(8, cl.getPeriodicBoxSize());
    donorKernel.setArg<mm_float4>(9, cl.getInvPeriodicBoxSize());
3362
    cl.executeKernel(donorKernel, std::max(numDonors, numAcceptors));
3363
3364
    acceptorKernel.setArg<mm_float4>(8, cl.getPeriodicBoxSize());
    acceptorKernel.setArg<mm_float4>(9, cl.getInvPeriodicBoxSize());
3365
    cl.executeKernel(acceptorKernel, std::max(numDonors, numAcceptors));
3366
3367
3368
    return 0.0;
}

3369
3370
3371
3372
OpenCLIntegrateVerletStepKernel::~OpenCLIntegrateVerletStepKernel() {
}

void OpenCLIntegrateVerletStepKernel::initialize(const System& system, const VerletIntegrator& integrator) {
3373
    cl.getPlatformData().initializeContexts(system);
3374
    cl::Program program = cl.createProgram(OpenCLKernelSources::verlet, "");
3375
3376
    kernel1 = cl::Kernel(program, "integrateVerletPart1");
    kernel2 = cl::Kernel(program, "integrateVerletPart2");
3377
    prevStepSize = -1.0;
3378
3379
3380
}

void OpenCLIntegrateVerletStepKernel::execute(ContextImpl& context, const VerletIntegrator& integrator) {
3381
    OpenCLIntegrationUtilities& integration = cl.getIntegrationUtilities();
3382
3383
    int numAtoms = cl.getNumAtoms();
    double dt = integrator.getStepSize();
3384
3385
3386
    if (!hasInitializedKernels) {
        hasInitializedKernels = true;
        kernel1.setArg<cl_int>(0, numAtoms);
3387
        kernel1.setArg<cl::Buffer>(1, cl.getIntegrationUtilities().getStepSize().getDeviceBuffer());
3388
3389
3390
3391
3392
        kernel1.setArg<cl::Buffer>(2, cl.getPosq().getDeviceBuffer());
        kernel1.setArg<cl::Buffer>(3, cl.getVelm().getDeviceBuffer());
        kernel1.setArg<cl::Buffer>(4, cl.getForce().getDeviceBuffer());
        kernel1.setArg<cl::Buffer>(5, integration.getPosDelta().getDeviceBuffer());
        kernel2.setArg<cl_int>(0, numAtoms);
3393
        kernel2.setArg<cl::Buffer>(1, cl.getIntegrationUtilities().getStepSize().getDeviceBuffer());
3394
3395
3396
3397
        kernel2.setArg<cl::Buffer>(2, cl.getPosq().getDeviceBuffer());
        kernel2.setArg<cl::Buffer>(3, cl.getVelm().getDeviceBuffer());
        kernel2.setArg<cl::Buffer>(4, integration.getPosDelta().getDeviceBuffer());
    }
3398
3399
    if (dt != prevStepSize) {
        vector<mm_float2> stepSizeVec(1);
3400
        stepSizeVec[0] = mm_float2((cl_float) dt, (cl_float) dt);
3401
        cl.getIntegrationUtilities().getStepSize().upload(stepSizeVec);
3402
3403
        prevStepSize = dt;
    }
3404
3405
3406
3407
3408
3409
3410

    // Call the first integration kernel.

    cl.executeKernel(kernel1, numAtoms);

    // Apply constraints.

3411
    integration.applyConstraints(integrator.getConstraintTolerance());
3412
3413
3414
3415
3416
3417
3418
3419
3420

    // Call the second integration kernel.

    cl.executeKernel(kernel2, numAtoms);

    // Update the time and step count.

    cl.setTime(cl.getTime()+dt);
    cl.setStepCount(cl.getStepCount()+1);
3421
3422
}

3423
3424
3425
3426
3427
3428
OpenCLIntegrateLangevinStepKernel::~OpenCLIntegrateLangevinStepKernel() {
    if (params != NULL)
        delete params;
}

void OpenCLIntegrateLangevinStepKernel::initialize(const System& system, const LangevinIntegrator& integrator) {
3429
    cl.getPlatformData().initializeContexts(system);
3430
3431
3432
3433
    cl.getIntegrationUtilities().initRandomNumberGenerator(integrator.getRandomNumberSeed());
    map<string, string> defines;
    defines["NUM_ATOMS"] = intToString(cl.getNumAtoms());
    defines["PADDED_NUM_ATOMS"] = intToString(cl.getPaddedNumAtoms());
3434
    cl::Program program = cl.createProgram(OpenCLKernelSources::langevin, defines, "");
3435
3436
    kernel1 = cl::Kernel(program, "integrateLangevinPart1");
    kernel2 = cl::Kernel(program, "integrateLangevinPart2");
3437
    params = new OpenCLArray<cl_float>(cl, 3, "langevinParams");
3438
3439
3440
3441
    prevStepSize = -1.0;
}

void OpenCLIntegrateLangevinStepKernel::execute(ContextImpl& context, const LangevinIntegrator& integrator) {
3442
    OpenCLIntegrationUtilities& integration = cl.getIntegrationUtilities();
3443
    int numAtoms = cl.getNumAtoms();
3444
3445
    if (!hasInitializedKernels) {
        hasInitializedKernels = true;
3446
3447
3448
3449
        kernel1.setArg<cl::Buffer>(0, cl.getVelm().getDeviceBuffer());
        kernel1.setArg<cl::Buffer>(1, cl.getForce().getDeviceBuffer());
        kernel1.setArg<cl::Buffer>(2, integration.getPosDelta().getDeviceBuffer());
        kernel1.setArg<cl::Buffer>(3, params->getDeviceBuffer());
3450
3451
3452
        kernel1.setArg<cl::Buffer>(4, integration.getStepSize().getDeviceBuffer());
        kernel1.setArg<cl::Buffer>(5, integration.getRandom().getDeviceBuffer());
        kernel2.setArg<cl::Buffer>(0, cl.getPosq().getDeviceBuffer());
3453
        kernel2.setArg<cl::Buffer>(1, integration.getPosDelta().getDeviceBuffer());
3454
3455
        kernel2.setArg<cl::Buffer>(2, cl.getVelm().getDeviceBuffer());
        kernel2.setArg<cl::Buffer>(3, integration.getStepSize().getDeviceBuffer());
3456
    }
3457
3458
3459
3460
3461
3462
3463
3464
    double temperature = integrator.getTemperature();
    double friction = integrator.getFriction();
    double stepSize = integrator.getStepSize();
    if (temperature != prevTemp || friction != prevFriction || stepSize != prevStepSize) {
        // Calculate the integration parameters.

        double tau = (friction == 0.0 ? 0.0 : 1.0/friction);
        double kT = BOLTZ*temperature;
3465
3466
3467
        double vscale = std::exp(-stepSize/tau);
        double fscale = (1-vscale)*tau;
        double noisescale = std::sqrt(2*kT/tau)*std::sqrt(0.5*(1-vscale*vscale)*tau);
3468
        vector<cl_float> p(params->getSize());
3469
3470
3471
        p[0] = (cl_float) vscale;
        p[1] = (cl_float) fscale;
        p[2] = (cl_float) noisescale;
3472
        params->upload(p);
3473
        integration.getStepSize()[0].y = (cl_float) stepSize;
3474
        integration.getStepSize().upload();
3475
3476
3477
3478
3479
3480
3481
        prevTemp = temperature;
        prevFriction = friction;
        prevStepSize = stepSize;
    }

    // Call the first integration kernel.

3482
    kernel1.setArg<cl_uint>(6, integration.prepareRandomNumbers(cl.getPaddedNumAtoms()));
3483
3484
3485
3486
    cl.executeKernel(kernel1, numAtoms);

    // Apply constraints.

3487
    integration.applyConstraints(integrator.getConstraintTolerance());
3488
3489
3490
3491
3492
3493
3494
3495
3496
3497

    // Call the second integration kernel.

    cl.executeKernel(kernel2, numAtoms);

    // Update the time and step count.

    cl.setTime(cl.getTime()+stepSize);
    cl.setStepCount(cl.getStepCount()+1);
}
3498
3499
3500
3501
3502

OpenCLIntegrateBrownianStepKernel::~OpenCLIntegrateBrownianStepKernel() {
}

void OpenCLIntegrateBrownianStepKernel::initialize(const System& system, const BrownianIntegrator& integrator) {
3503
    cl.getPlatformData().initializeContexts(system);
3504
3505
3506
    cl.getIntegrationUtilities().initRandomNumberGenerator(integrator.getRandomNumberSeed());
    map<string, string> defines;
    defines["NUM_ATOMS"] = intToString(cl.getNumAtoms());
3507
    cl::Program program = cl.createProgram(OpenCLKernelSources::brownian, defines, "");
3508
3509
3510
3511
3512
3513
3514
3515
3516
3517
3518
3519
    kernel1 = cl::Kernel(program, "integrateBrownianPart1");
    kernel2 = cl::Kernel(program, "integrateBrownianPart2");
    prevStepSize = -1.0;
}

void OpenCLIntegrateBrownianStepKernel::execute(ContextImpl& context, const BrownianIntegrator& integrator) {
    OpenCLIntegrationUtilities& integration = cl.getIntegrationUtilities();
    int numAtoms = cl.getNumAtoms();
    if (!hasInitializedKernels) {
        hasInitializedKernels = true;
        kernel1.setArg<cl::Buffer>(2, cl.getForce().getDeviceBuffer());
        kernel1.setArg<cl::Buffer>(3, integration.getPosDelta().getDeviceBuffer());
3520
3521
        kernel1.setArg<cl::Buffer>(4, cl.getVelm().getDeviceBuffer());
        kernel1.setArg<cl::Buffer>(5, integration.getRandom().getDeviceBuffer());
3522
3523
3524
3525
3526
3527
3528
3529
3530
3531
3532
3533
3534
3535
3536
3537
3538
3539
3540
        kernel2.setArg<cl::Buffer>(1, cl.getPosq().getDeviceBuffer());
        kernel2.setArg<cl::Buffer>(2, cl.getVelm().getDeviceBuffer());
        kernel2.setArg<cl::Buffer>(3, integration.getPosDelta().getDeviceBuffer());
    }
    double temperature = integrator.getTemperature();
    double friction = integrator.getFriction();
    double stepSize = integrator.getStepSize();
    if (temperature != prevTemp || friction != prevFriction || stepSize != prevStepSize) {
        double tau = (friction == 0.0 ? 0.0 : 1.0/friction);
        kernel1.setArg<cl_float>(0, (cl_float) (tau*stepSize));
        kernel1.setArg<cl_float>(1, (cl_float) (sqrt(2.0f*BOLTZ*temperature*stepSize*tau)));
        kernel2.setArg<cl_float>(0, (cl_float) (1.0/stepSize));
        prevTemp = temperature;
        prevFriction = friction;
        prevStepSize = stepSize;
    }

    // Call the first integration kernel.

3541
    kernel1.setArg<cl_uint>(6, integration.prepareRandomNumbers(cl.getPaddedNumAtoms()));
3542
3543
3544
3545
3546
3547
3548
3549
3550
3551
3552
3553
3554
3555
3556
    cl.executeKernel(kernel1, numAtoms);

    // Apply constraints.

    integration.applyConstraints(integrator.getConstraintTolerance());

    // Call the second integration kernel.

    cl.executeKernel(kernel2, numAtoms);

    // Update the time and step count.

    cl.setTime(cl.getTime()+stepSize);
    cl.setStepCount(cl.getStepCount()+1);
}
3557
3558
3559
3560
3561

OpenCLIntegrateVariableVerletStepKernel::~OpenCLIntegrateVariableVerletStepKernel() {
}

void OpenCLIntegrateVariableVerletStepKernel::initialize(const System& system, const VariableVerletIntegrator& integrator) {
3562
    cl.getPlatformData().initializeContexts(system);
3563
    cl::Program program = cl.createProgram(OpenCLKernelSources::verlet, "");
3564
3565
3566
3567
3568
3569
3570
    kernel1 = cl::Kernel(program, "integrateVerletPart1");
    kernel2 = cl::Kernel(program, "integrateVerletPart2");
    selectSizeKernel = cl::Kernel(program, "selectVerletStepSize");
    blockSize = std::min(std::min(256, system.getNumParticles()), (int) cl.getDevice().getInfo<CL_DEVICE_MAX_WORK_GROUP_SIZE>());
}

void OpenCLIntegrateVariableVerletStepKernel::execute(ContextImpl& context, const VariableVerletIntegrator& integrator, double maxTime) {
3571
    OpenCLIntegrationUtilities& integration = cl.getIntegrationUtilities();
3572
3573
3574
3575
    int numAtoms = cl.getNumAtoms();
    if (!hasInitializedKernels) {
        hasInitializedKernels = true;
        kernel1.setArg<cl_int>(0, numAtoms);
3576
        kernel1.setArg<cl::Buffer>(1, cl.getIntegrationUtilities().getStepSize().getDeviceBuffer());
3577
3578
3579
3580
3581
        kernel1.setArg<cl::Buffer>(2, cl.getPosq().getDeviceBuffer());
        kernel1.setArg<cl::Buffer>(3, cl.getVelm().getDeviceBuffer());
        kernel1.setArg<cl::Buffer>(4, cl.getForce().getDeviceBuffer());
        kernel1.setArg<cl::Buffer>(5, integration.getPosDelta().getDeviceBuffer());
        kernel2.setArg<cl_int>(0, numAtoms);
3582
        kernel2.setArg<cl::Buffer>(1, cl.getIntegrationUtilities().getStepSize().getDeviceBuffer());
3583
3584
3585
3586
        kernel2.setArg<cl::Buffer>(2, cl.getPosq().getDeviceBuffer());
        kernel2.setArg<cl::Buffer>(3, cl.getVelm().getDeviceBuffer());
        kernel2.setArg<cl::Buffer>(4, integration.getPosDelta().getDeviceBuffer());
        selectSizeKernel.setArg<cl_int>(0, numAtoms);
3587
        selectSizeKernel.setArg<cl::Buffer>(3, cl.getIntegrationUtilities().getStepSize().getDeviceBuffer());
3588
3589
3590
3591
3592
3593
3594
3595
3596
        selectSizeKernel.setArg<cl::Buffer>(4, cl.getVelm().getDeviceBuffer());
        selectSizeKernel.setArg<cl::Buffer>(5, cl.getForce().getDeviceBuffer());
        selectSizeKernel.setArg(6, blockSize*sizeof(cl_float), NULL);
    }

    // Select the step size to use.

    float maxStepSize = (float)(maxTime-cl.getTime());
    selectSizeKernel.setArg<cl_float>(1, maxStepSize);
3597
    selectSizeKernel.setArg<cl_float>(2, (cl_float) integrator.getErrorTolerance());
3598
3599
3600
3601
3602
3603
3604
3605
3606
3607
3608
3609
3610
3611
3612
3613
    cl.executeKernel(selectSizeKernel, blockSize, blockSize);

    // Call the first integration kernel.

    cl.executeKernel(kernel1, numAtoms);

    // Apply constraints.

    integration.applyConstraints(integrator.getConstraintTolerance());

    // Call the second integration kernel.

    cl.executeKernel(kernel2, numAtoms);

    // Update the time and step count.

3614
3615
    cl.getIntegrationUtilities().getStepSize().download();
    double dt = cl.getIntegrationUtilities().getStepSize()[0].y;
3616
3617
3618
3619
3620
3621
3622
3623
3624
3625
3626
3627
3628
    double time = cl.getTime()+dt;
    if (dt == maxStepSize)
        time = maxTime; // Avoid round-off error
    cl.setTime(time);
    cl.setStepCount(cl.getStepCount()+1);
}

OpenCLIntegrateVariableLangevinStepKernel::~OpenCLIntegrateVariableLangevinStepKernel() {
    if (params != NULL)
        delete params;
}

void OpenCLIntegrateVariableLangevinStepKernel::initialize(const System& system, const VariableLangevinIntegrator& integrator) {
3629
    cl.getPlatformData().initializeContexts(system);
3630
3631
3632
3633
    cl.getIntegrationUtilities().initRandomNumberGenerator(integrator.getRandomNumberSeed());
    map<string, string> defines;
    defines["NUM_ATOMS"] = intToString(cl.getNumAtoms());
    defines["PADDED_NUM_ATOMS"] = intToString(cl.getPaddedNumAtoms());
3634
    cl::Program program = cl.createProgram(OpenCLKernelSources::langevin, defines, "");
3635
3636
3637
    kernel1 = cl::Kernel(program, "integrateLangevinPart1");
    kernel2 = cl::Kernel(program, "integrateLangevinPart2");
    selectSizeKernel = cl::Kernel(program, "selectLangevinStepSize");
3638
    params = new OpenCLArray<cl_float>(cl, 3, "langevinParams");
3639
3640
3641
3642
3643
3644
    blockSize = std::min(256, system.getNumParticles());
    blockSize = std::max(blockSize, params->getSize());
    blockSize = std::min(blockSize, (int) cl.getDevice().getInfo<CL_DEVICE_MAX_WORK_GROUP_SIZE>());
}

void OpenCLIntegrateVariableLangevinStepKernel::execute(ContextImpl& context, const VariableLangevinIntegrator& integrator, double maxTime) {
3645
    OpenCLIntegrationUtilities& integration = cl.getIntegrationUtilities();
3646
3647
3648
    int numAtoms = cl.getNumAtoms();
    if (!hasInitializedKernels) {
        hasInitializedKernels = true;
3649
3650
3651
3652
        kernel1.setArg<cl::Buffer>(0, cl.getVelm().getDeviceBuffer());
        kernel1.setArg<cl::Buffer>(1, cl.getForce().getDeviceBuffer());
        kernel1.setArg<cl::Buffer>(2, integration.getPosDelta().getDeviceBuffer());
        kernel1.setArg<cl::Buffer>(3, params->getDeviceBuffer());
3653
3654
3655
        kernel1.setArg<cl::Buffer>(4, integration.getStepSize().getDeviceBuffer());
        kernel1.setArg<cl::Buffer>(5, integration.getRandom().getDeviceBuffer());
        kernel2.setArg<cl::Buffer>(0, cl.getPosq().getDeviceBuffer());
3656
        kernel2.setArg<cl::Buffer>(1, integration.getPosDelta().getDeviceBuffer());
3657
3658
        kernel2.setArg<cl::Buffer>(2, cl.getVelm().getDeviceBuffer());
        kernel2.setArg<cl::Buffer>(3, integration.getStepSize().getDeviceBuffer());
3659
        selectSizeKernel.setArg<cl::Buffer>(4, integration.getStepSize().getDeviceBuffer());
3660
3661
3662
3663
3664
        selectSizeKernel.setArg<cl::Buffer>(5, cl.getVelm().getDeviceBuffer());
        selectSizeKernel.setArg<cl::Buffer>(6, cl.getForce().getDeviceBuffer());
        selectSizeKernel.setArg<cl::Buffer>(7, params->getDeviceBuffer());
        selectSizeKernel.setArg(8, params->getSize()*sizeof(cl_float), NULL);
        selectSizeKernel.setArg(9, blockSize*sizeof(cl_float), NULL);
3665
3666
3667
3668
3669
    }

    // Select the step size to use.

    float maxStepSize = (float)(maxTime-cl.getTime());
3670
3671
3672
3673
    selectSizeKernel.setArg<cl_float>(0, maxStepSize);
    selectSizeKernel.setArg<cl_float>(1, (cl_float) integrator.getErrorTolerance());
    selectSizeKernel.setArg<cl_float>(2, (cl_float) (integrator.getFriction() == 0.0 ? 0.0 : 1.0/integrator.getFriction()));
    selectSizeKernel.setArg<cl_float>(3, (cl_float) (BOLTZ*integrator.getTemperature()));
3674
3675
3676
3677
    cl.executeKernel(selectSizeKernel, blockSize, blockSize);

    // Call the first integration kernel.

3678
    kernel1.setArg<cl_uint>(6, integration.prepareRandomNumbers(cl.getPaddedNumAtoms()));
3679
3680
3681
3682
3683
3684
3685
3686
3687
3688
3689
3690
    cl.executeKernel(kernel1, numAtoms);

    // Apply constraints.

    integration.applyConstraints(integrator.getConstraintTolerance());

    // Call the second integration kernel.

    cl.executeKernel(kernel2, numAtoms);

    // Update the time and step count.

3691
3692
    cl.getIntegrationUtilities().getStepSize().download();
    double dt = cl.getIntegrationUtilities().getStepSize()[0].y;
3693
3694
3695
3696
3697
3698
3699
    double time = cl.getTime()+dt;
    if (dt == maxStepSize)
        time = maxTime; // Avoid round-off error
    cl.setTime(time);
    cl.setStepCount(cl.getStepCount()+1);
}

3700
OpenCLApplyAndersenThermostatKernel::~OpenCLApplyAndersenThermostatKernel() {
3701
3702
    if (atomGroups != NULL)
        delete atomGroups;
3703
3704
3705
3706
3707
3708
}

void OpenCLApplyAndersenThermostatKernel::initialize(const System& system, const AndersenThermostat& thermostat) {
    randomSeed = thermostat.getRandomNumberSeed();
    map<string, string> defines;
    defines["NUM_ATOMS"] = intToString(cl.getNumAtoms());
3709
    cl::Program program = cl.createProgram(OpenCLKernelSources::andersenThermostat, defines);
3710
    kernel = cl::Kernel(program, "applyAndersenThermostat");
3711
3712
3713
3714
3715
3716
3717
3718
3719
3720
3721

    // Create the arrays with the group definitions.

    vector<vector<int> > groups = AndersenThermostatImpl::calcParticleGroups(system);
    atomGroups = new OpenCLArray<int>(cl, cl.getNumAtoms(), "atomGroups");
    vector<int> atoms(atomGroups->getSize());
    for (int i = 0; i < (int) groups.size(); i++) {
        for (int j = 0; j < (int) groups[i].size(); j++)
            atoms[groups[i][j]] = i;
    }
    atomGroups->upload(atoms);
3722
3723
3724
3725
3726
3727
3728
3729
3730
}

void OpenCLApplyAndersenThermostatKernel::execute(ContextImpl& context) {
    if (!hasInitializedKernels) {
        hasInitializedKernels = true;
        cl.getIntegrationUtilities().initRandomNumberGenerator(randomSeed);
        kernel.setArg<cl::Buffer>(2, cl.getVelm().getDeviceBuffer());
        kernel.setArg<cl::Buffer>(3, cl.getIntegrationUtilities().getStepSize().getDeviceBuffer());
        kernel.setArg<cl::Buffer>(4, cl.getIntegrationUtilities().getRandom().getDeviceBuffer());
3731
        kernel.setArg<cl::Buffer>(6, atomGroups->getDeviceBuffer());
3732
3733
3734
3735
3736
3737
3738
    }
    kernel.setArg<cl_float>(0, (cl_float) context.getParameter(AndersenThermostat::CollisionFrequency()));
    kernel.setArg<cl_float>(1, (cl_float) (BOLTZ*context.getParameter(AndersenThermostat::Temperature())));
    kernel.setArg<cl_uint>(5, cl.getIntegrationUtilities().prepareRandomNumbers(cl.getPaddedNumAtoms()));
    cl.executeKernel(kernel, cl.getNumAtoms());
}

3739
3740
3741
3742
3743
3744
3745
3746
3747
3748
3749
3750
3751
3752
3753
3754
3755
3756
3757
3758
3759
3760
3761
3762
3763
3764
3765
3766
3767
3768
3769
3770
3771
3772
3773
3774
3775
3776
3777
3778
3779
3780
3781
3782
3783
3784
3785
3786
3787
OpenCLApplyMonteCarloBarostatKernel::~OpenCLApplyMonteCarloBarostatKernel() {
    if (savedPositions != NULL)
        delete savedPositions;
    if (moleculeAtoms != NULL)
        delete moleculeAtoms;
    if (moleculeStartIndex != NULL)
        delete moleculeStartIndex;
}

void OpenCLApplyMonteCarloBarostatKernel::initialize(const System& system, const MonteCarloBarostat& thermostat) {
    savedPositions = new OpenCLArray<mm_float4>(cl, cl.getPaddedNumAtoms(), "savedPositions");
    cl::Program program = cl.createProgram(OpenCLKernelSources::monteCarloBarostat);
    kernel = cl::Kernel(program, "scalePositions");
}

void OpenCLApplyMonteCarloBarostatKernel::scaleCoordinates(ContextImpl& context, double scale) {
    if (!hasInitializedKernels) {
        hasInitializedKernels = true;

        // Create the arrays with the molecule definitions.

        vector<vector<int> > molecules = context.getMolecules();
        numMolecules = molecules.size();
        moleculeAtoms = new OpenCLArray<int>(cl, cl.getNumAtoms(), "moleculeAtoms");
        moleculeStartIndex = new OpenCLArray<int>(cl, numMolecules+1, "moleculeStartIndex");
        vector<int> atoms(moleculeAtoms->getSize());
        vector<int> startIndex(moleculeStartIndex->getSize());
        int index = 0;
        for (int i = 0; i < numMolecules; i++) {
            startIndex[i] = index;
            for (int j = 0; j < (int) molecules[i].size(); j++)
                atoms[index++] = molecules[i][j];
        }
        startIndex[numMolecules] = index;
        moleculeAtoms->upload(atoms);
        moleculeStartIndex->upload(startIndex);

        // Initialize the kernel arguments.
        
        kernel.setArg<cl_int>(1, numMolecules);
        kernel.setArg<cl::Buffer>(4, cl.getPosq().getDeviceBuffer());
        kernel.setArg<cl::Buffer>(5, moleculeAtoms->getDeviceBuffer());
        kernel.setArg<cl::Buffer>(6, moleculeStartIndex->getDeviceBuffer());
    }
    cl.getQueue().enqueueCopyBuffer(cl.getPosq().getDeviceBuffer(), savedPositions->getDeviceBuffer(), 0, 0, cl.getPosq().getSize()*sizeof(mm_float4));
    kernel.setArg<cl_float>(0, (cl_float) scale);
    kernel.setArg<mm_float4>(2, cl.getPeriodicBoxSize());
    kernel.setArg<mm_float4>(3, cl.getInvPeriodicBoxSize());
    cl.executeKernel(kernel, cl.getNumAtoms());
3788
3789
    for (int i = 0; i < (int) cl.getPosCellOffsets().size(); i++)
        cl.getPosCellOffsets()[i] = mm_int4(0, 0, 0, 0);
3790
3791
3792
3793
3794
3795
}

void OpenCLApplyMonteCarloBarostatKernel::restoreCoordinates(ContextImpl& context) {
    cl.getQueue().enqueueCopyBuffer(savedPositions->getDeviceBuffer(), cl.getPosq().getDeviceBuffer(), 0, 0, cl.getPosq().getSize()*sizeof(mm_float4));
}

3796
3797
3798
3799
3800
3801
3802
3803
3804
3805
3806
void OpenCLCalcKineticEnergyKernel::initialize(const System& system) {
    int numParticles = system.getNumParticles();
    masses.resize(numParticles);
    for (int i = 0; i < numParticles; ++i)
        masses[i] = system.getParticleMass(i);
}

double OpenCLCalcKineticEnergyKernel::execute(ContextImpl& context) {
    // We don't currently have a GPU kernel to do this, so we retrieve the velocities and calculate the energy
    // on the CPU.

3807
    OpenCLArray<mm_float4>& velm = cl.getVelm();
3808
    velm.download();
3809
    double energy = 0.0;
3810
    for (size_t i = 0; i < masses.size(); ++i) {
3811
3812
        mm_float4 v = velm[i];
        energy += masses[i]*(v.x*v.x+v.y*v.y+v.z*v.z);
3813
    }
3814
3815
    return 0.5*energy;
}
3816
3817
3818
3819
3820
3821
3822
3823
3824
3825
3826
3827
3828
3829
3830

OpenCLRemoveCMMotionKernel::~OpenCLRemoveCMMotionKernel() {
    if (cmMomentum != NULL)
        delete cmMomentum;
}

void OpenCLRemoveCMMotionKernel::initialize(const System& system, const CMMotionRemover& force) {
    frequency = force.getFrequency();
    int numAtoms = cl.getNumAtoms();
    cmMomentum = new OpenCLArray<mm_float4>(cl, (numAtoms+OpenCLContext::ThreadBlockSize-1)/OpenCLContext::ThreadBlockSize, "cmMomentum");
    double totalMass = 0.0;
    for (int i = 0; i < numAtoms; i++)
        totalMass += system.getParticleMass(i);
    map<string, string> defines;
    defines["INVERSE_TOTAL_MASS"] = doubleToString(1.0/totalMass);
3831
    cl::Program program = cl.createProgram(OpenCLKernelSources::removeCM, defines);
3832
3833
3834
3835
3836
3837
3838
3839
3840
3841
3842
3843
3844
3845
3846
3847
    kernel1 = cl::Kernel(program, "calcCenterOfMassMomentum");
    kernel1.setArg<cl_int>(0, numAtoms);
    kernel1.setArg<cl::Buffer>(1, cl.getVelm().getDeviceBuffer());
    kernel1.setArg<cl::Buffer>(2, cmMomentum->getDeviceBuffer());
    kernel1.setArg(3, OpenCLContext::ThreadBlockSize*sizeof(mm_float4), NULL);
    kernel2 = cl::Kernel(program, "removeCenterOfMassMomentum");
    kernel2.setArg<cl_int>(0, numAtoms);
    kernel2.setArg<cl::Buffer>(1, cl.getVelm().getDeviceBuffer());
    kernel2.setArg<cl::Buffer>(2, cmMomentum->getDeviceBuffer());
    kernel2.setArg(3, OpenCLContext::ThreadBlockSize*sizeof(mm_float4), NULL);
}

void OpenCLRemoveCMMotionKernel::execute(ContextImpl& context) {
    cl.executeKernel(kernel1, cl.getNumAtoms());
    cl.executeKernel(kernel2, cl.getNumAtoms());
}