"platforms/cpu/src/CpuNonbondedForce.cpp" did not exist on "21847f4a12a3a2e38d03933f737e68a67559ad2c"
OpenCLKernels.cpp 192 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 "openmm/internal/MSVC_erfc.h"
45
#include <cmath>
46
#include <set>
47
48
49
50

using namespace OpenMM;
using namespace std;

51
52
53
54
55
56
57
58
59
60
61
62
63
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();
}

64
65
66
67
68
69
70
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);
}

71
void OpenCLCalcForcesAndEnergyKernel::initialize(const System& system) {
72
73
}

74
void OpenCLCalcForcesAndEnergyKernel::beginComputation(ContextImpl& context, bool includeForces, bool includeEnergy) {
75
    if (cl.getNonbondedUtilities().getUseCutoff() && cl.getComputeForceCount()%100 == 0) {
76
        cl.reorderAtoms();
77
78
        cl.getNonbondedUtilities().updateNeighborListSize();
    }
79
    cl.setComputeForceCount(cl.getComputeForceCount()+1);
80
    cl.clearAutoclearBuffers();
81
    cl.getNonbondedUtilities().prepareInteractions();
82
83
}

84
double OpenCLCalcForcesAndEnergyKernel::finishComputation(ContextImpl& context, bool includeForces, bool includeEnergy) {
85
    cl.getNonbondedUtilities().computeInteractions();
86
87
    if (includeForces)
        cl.reduceBuffer(cl.getForceBuffers(), cl.getNumForceBuffers());
88
    double sum = 0.0f;
89
90
91
92
93
94
    if (includeEnergy) {
        OpenCLArray<cl_float>& energy = cl.getEnergyBuffer();
        energy.download();
        for (int i = 0; i < energy.getSize(); i++)
            sum += energy[i];
    }
95
    return sum;
96
97
}

98
void OpenCLUpdateStateDataKernel::initialize(const System& system) {
99
100
}

101
double OpenCLUpdateStateDataKernel::getTime(const ContextImpl& context) const {
102
    return cl.getTime();
103
104
}

105
void OpenCLUpdateStateDataKernel::setTime(ContextImpl& context, double time) {
106
107
108
    vector<OpenCLContext*>& contexts = cl.getPlatformData().contexts;
    for (int i = 0; i < (int) contexts.size(); i++)
        contexts[i]->setTime(time);
109
110
}

111
void OpenCLUpdateStateDataKernel::getPositions(ContextImpl& context, std::vector<Vec3>& positions) {
112
    OpenCLArray<mm_float4>& posq = cl.getPosq();
113
    posq.download();
114
    OpenCLArray<cl_int>& order = cl.getAtomIndex();
115
116
    int numParticles = context.getSystem().getNumParticles();
    positions.resize(numParticles);
117
    mm_float4 periodicBoxSize = cl.getPeriodicBoxSize();
118
    for (int i = 0; i < numParticles; ++i) {
119
        mm_float4 pos = posq[i];
120
121
        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);
122
123
124
125
    }
}

void OpenCLUpdateStateDataKernel::setPositions(ContextImpl& context, const std::vector<Vec3>& positions) {
126
127
    OpenCLArray<mm_float4>& posq = cl.getPosq();
    OpenCLArray<cl_int>& order = cl.getAtomIndex();
128
129
    int numParticles = context.getSystem().getNumParticles();
    for (int i = 0; i < numParticles; ++i) {
130
        mm_float4& pos = posq[i];
131
        const Vec3& p = positions[order[i]];
132
133
134
        pos.x = (cl_float) p[0];
        pos.y = (cl_float) p[1];
        pos.z = (cl_float) p[2];
135
    }
136
137
    for (int i = numParticles; i < cl.getPaddedNumAtoms(); i++)
        posq[i] = mm_float4(0.0f, 0.0f, 0.0f, 0.0f);
138
    posq.upload();
139
    for (int i = 0; i < (int) cl.getPosCellOffsets().size(); i++)
140
        cl.getPosCellOffsets()[i] = mm_int4(0, 0, 0, 0);
141
142
143
}

void OpenCLUpdateStateDataKernel::getVelocities(ContextImpl& context, std::vector<Vec3>& velocities) {
144
    OpenCLArray<mm_float4>& velm = cl.getVelm();
145
    velm.download();
146
    OpenCLArray<cl_int>& order = cl.getAtomIndex();
147
148
149
    int numParticles = context.getSystem().getNumParticles();
    velocities.resize(numParticles);
    for (int i = 0; i < numParticles; ++i) {
150
151
        mm_float4 vel = velm[i];
        velocities[order[i]] = Vec3(vel.x, vel.y, vel.z);
152
153
154
155
    }
}

void OpenCLUpdateStateDataKernel::setVelocities(ContextImpl& context, const std::vector<Vec3>& velocities) {
156
157
    OpenCLArray<mm_float4>& velm = cl.getVelm();
    OpenCLArray<cl_int>& order = cl.getAtomIndex();
158
159
    int numParticles = context.getSystem().getNumParticles();
    for (int i = 0; i < numParticles; ++i) {
160
        mm_float4& vel = velm[i];
161
        const Vec3& p = velocities[order[i]];
162
163
164
        vel.x = (cl_float) p[0];
        vel.y = (cl_float) p[1];
        vel.z = (cl_float) p[2];
165
    }
166
167
    for (int i = numParticles; i < cl.getPaddedNumAtoms(); i++)
        velm[i] = mm_float4(0.0f, 0.0f, 0.0f, 0.0f);
168
169
170
171
    velm.upload();
}

void OpenCLUpdateStateDataKernel::getForces(ContextImpl& context, std::vector<Vec3>& forces) {
172
    OpenCLArray<mm_float4>& force = cl.getForce();
173
    force.download();
174
    OpenCLArray<cl_int>& order = cl.getAtomIndex();
175
176
177
    int numParticles = context.getSystem().getNumParticles();
    forces.resize(numParticles);
    for (int i = 0; i < numParticles; ++i) {
178
179
        mm_float4 f = force[i];
        forces[order[i]] = Vec3(f.x, f.y, f.z);
180
181
182
    }
}

183
184
185
186
187
188
189
190
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 {
191
192
193
    vector<OpenCLContext*>& contexts = cl.getPlatformData().contexts;
    for (int i = 0; i < (int) contexts.size(); i++)
        contexts[i]->setPeriodicBoxSize(a[0], b[1], c[2]);
194
195
}

196
197
198
199
200
201
202
void OpenCLApplyConstraintsKernel::initialize(const System& system) {
}

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

203
204
class OpenCLBondForceInfo : public OpenCLForceInfo {
public:
205
    OpenCLBondForceInfo(int requiredBuffers, const HarmonicBondForce& force) : OpenCLForceInfo(requiredBuffers), force(force) {
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
    }
    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;
};

229
230
231
232
233
234
235
OpenCLCalcHarmonicBondForceKernel::~OpenCLCalcHarmonicBondForceKernel() {
    if (params != NULL)
        delete params;
    if (indices != NULL)
        delete indices;
}

236
void OpenCLCalcHarmonicBondForceKernel::initialize(const System& system, const HarmonicBondForce& force) {
237
238
239
240
    int numContexts = cl.getPlatformData().contexts.size();
    int startIndex = cl.getContextIndex()*force.getNumBonds()/numContexts;
    int endIndex = (cl.getContextIndex()+1)*force.getNumBonds()/numContexts;
    numBonds = endIndex-startIndex;
241
242
    if (numBonds == 0)
        return;
243
244
    params = new OpenCLArray<mm_float2>(cl, numBonds, "bondParams");
    indices = new OpenCLArray<mm_int4>(cl, numBonds, "bondIndices");
245
246
247
248
249
250
    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;
251
        force.getBondParameters(startIndex+i, particle1, particle2, length, k);
252
253
        paramVector[i] = mm_float2((cl_float) length, (cl_float) k);
        indicesVector[i] = mm_int4(particle1, particle2, forceBufferCounter[particle1]++, forceBufferCounter[particle2]++);
254
255
256
257
    }
    params->upload(paramVector);
    indices->upload(indicesVector);
    int maxBuffers = 1;
258
    for (int i = 0; i < (int) forceBufferCounter.size(); i++)
259
        maxBuffers = max(maxBuffers, forceBufferCounter[i]);
260
    cl.addForce(new OpenCLBondForceInfo(maxBuffers, force));
261
    cl::Program program = cl.createProgram(OpenCLKernelSources::harmonicBondForce);
262
263
264
    kernel = cl::Kernel(program, "calcHarmonicBondForce");
}

265
double OpenCLCalcHarmonicBondForceKernel::execute(ContextImpl& context, bool includeForces, bool includeEnergy) {
266
    if (numBonds == 0)
267
        return 0.0;
268
269
270
271
272
273
274
275
276
277
    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());
    }
278
    cl.executeKernel(kernel, numBonds);
279
280
    return 0.0;
}
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
320
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) {
321
322
323
324
    int numContexts = cl.getPlatformData().contexts.size();
    int startIndex = cl.getContextIndex()*force.getNumBonds()/numContexts;
    int endIndex = (cl.getContextIndex()+1)*force.getNumBonds()/numContexts;
    numBonds = endIndex-startIndex;
325
326
    if (numBonds == 0)
        return;
327
    params = new OpenCLParameterSet(cl, force.getNumPerBondParameters(), numBonds, "customBondParams");
328
329
330
331
332
333
334
    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);
335
    vector<vector<cl_float> > paramVector(numBonds);
336
337
338
339
    vector<mm_int4> indicesVector(numBonds);
    for (int i = 0; i < numBonds; i++) {
        int particle1, particle2;
        vector<double> parameters;
340
        force.getBondParameters(startIndex+i, particle1, particle2, parameters);
341
        paramVector[i].resize(parameters.size());
342
        for (int j = 0; j < (int) parameters.size(); j++)
343
            paramVector[i][j] = (cl_float) parameters[j];
344
        indicesVector[i] = mm_int4(particle1, particle2, forceBufferCounter[particle1]++, forceBufferCounter[particle2]++);
345
    }
346
    params->setParameterValues(paramVector);
347
348
    indices->upload(indicesVector);
    int maxBuffers = 1;
349
    for (int i = 0; i < (int) forceBufferCounter.size(); i++)
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
        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);
375
        variables[name] = "bondParams"+params->getParameterSuffix(i);
376
377
378
379
380
381
382
    }
    for (int i = 0; i < force.getNumGlobalParameters(); i++) {
        const string& name = force.getGlobalParameterName(i);
        string value = "globals["+intToString(i)+"]";
        variables[name] = value;
    }
    stringstream compute;
383
384
385
386
387
    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";
    }
388
389
    vector<pair<string, string> > functions;
    compute << OpenCLExpressionUtilities::createExpressions(expressions, variables, functions, "temp", "");
390
    map<string, string> replacements;
391
392
    replacements["COMPUTE_FORCE"] = compute.str();
    replacements["EXTRA_ARGUMENTS"] = extraArguments;
393
    cl::Program program = cl.createProgram(cl.replaceStrings(OpenCLKernelSources::customBondForce, replacements));
394
395
396
    kernel = cl::Kernel(program, "computeCustomBondForces");
}

397
double OpenCLCalcCustomBondForceKernel::execute(ContextImpl& context, bool includeForces, bool includeEnergy) {
398
    if (numBonds == 0)
399
        return 0.0;
400
401
    if (globals != NULL) {
        bool changed = false;
402
        for (int i = 0; i < (int) globalParamNames.size(); i++) {
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
            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());
418
419
        kernel.setArg<cl::Buffer>(5, indices->getDeviceBuffer());
        int nextIndex = 6;
420
        if (globals != NULL)
421
422
423
            kernel.setArg<cl::Buffer>(nextIndex++, globals->getDeviceBuffer());
        for (int i = 0; i < (int) params->getBuffers().size(); i++) {
            const OpenCLNonbondedUtilities::ParameterInfo& buffer = params->getBuffers()[i];
424
            kernel.setArg<cl::Memory>(nextIndex++, buffer.getMemory());
425
        }
426
427
428
429
430
    }
    cl.executeKernel(kernel, numBonds);
    return 0.0;
}

431
432
class OpenCLAngleForceInfo : public OpenCLForceInfo {
public:
433
    OpenCLAngleForceInfo(int requiredBuffers, const HarmonicAngleForce& force) : OpenCLForceInfo(requiredBuffers), force(force) {
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
465
    }
    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) {
466
467
468
469
    int numContexts = cl.getPlatformData().contexts.size();
    int startIndex = cl.getContextIndex()*force.getNumAngles()/numContexts;
    int endIndex = (cl.getContextIndex()+1)*force.getNumAngles()/numContexts;
    numAngles = endIndex-startIndex;
470
471
    if (numAngles == 0)
        return;
472
473
    params = new OpenCLArray<mm_float2>(cl, numAngles, "angleParams");
    indices = new OpenCLArray<mm_int8>(cl, numAngles, "angleIndices");
474
475
476
477
478
479
    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;
480
        force.getAngleParameters(startIndex+i, particle1, particle2, particle3, angle, k);
481
482
483
        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);
484
485
486
487
488

    }
    params->upload(paramVector);
    indices->upload(indicesVector);
    int maxBuffers = 1;
489
    for (int i = 0; i < (int) forceBufferCounter.size(); i++)
490
        maxBuffers = max(maxBuffers, forceBufferCounter[i]);
491
    cl.addForce(new OpenCLAngleForceInfo(maxBuffers, force));
492
    cl::Program program = cl.createProgram(OpenCLKernelSources::harmonicAngleForce);
493
494
495
    kernel = cl::Kernel(program, "calcHarmonicAngleForce");
}

496
double OpenCLCalcHarmonicAngleForceKernel::execute(ContextImpl& context, bool includeForces, bool includeEnergy) {
497
    if (numAngles == 0)
498
        return 0.0;
499
500
501
502
503
504
505
506
507
508
    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());
    }
509
    cl.executeKernel(kernel, numAngles);
510
511
512
    return 0.0;
}

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
552
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) {
553
554
555
556
    int numContexts = cl.getPlatformData().contexts.size();
    int startIndex = cl.getContextIndex()*force.getNumAngles()/numContexts;
    int endIndex = (cl.getContextIndex()+1)*force.getNumAngles()/numContexts;
    numAngles = endIndex-startIndex;
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
    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;
572
        force.getAngleParameters(startIndex+i, particle1, particle2, particle3, parameters);
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
599
        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;
600
    expressions["float dEdAngle = "] = forceExpression;
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629

    // 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");
}

630
double OpenCLCalcCustomAngleForceKernel::execute(ContextImpl& context, bool includeForces, bool includeEnergy) {
631
    if (numAngles == 0)
632
        return 0.0;
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
    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];
657
            kernel.setArg<cl::Memory>(nextIndex++, buffer.getMemory());
658
659
660
661
662
663
        }
    }
    cl.executeKernel(kernel, numAngles);
    return 0.0;
}

664
665
class OpenCLPeriodicTorsionForceInfo : public OpenCLForceInfo {
public:
666
    OpenCLPeriodicTorsionForceInfo(int requiredBuffers, const PeriodicTorsionForce& force) : OpenCLForceInfo(requiredBuffers), force(force) {
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
    }
    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);
685
        force.getTorsionParameters(group2, particle1, particle2, particle3, particle4, periodicity2, phase2, k2);
686
687
688
689
690
691
692
693
694
695
696
697
698
699
        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) {
700
701
702
703
    int numContexts = cl.getPlatformData().contexts.size();
    int startIndex = cl.getContextIndex()*force.getNumTorsions()/numContexts;
    int endIndex = (cl.getContextIndex()+1)*force.getNumTorsions()/numContexts;
    numTorsions = endIndex-startIndex;
704
705
    if (numTorsions == 0)
        return;
706
707
    params = new OpenCLArray<mm_float4>(cl, numTorsions, "periodicTorsionParams");
    indices = new OpenCLArray<mm_int8>(cl, numTorsions, "periodicTorsionIndices");
708
709
710
711
712
713
    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;
714
        force.getTorsionParameters(startIndex+i, particle1, particle2, particle3, particle4, periodicity, phase, k);
715
716
717
        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]++);
718
719
720
721
722

    }
    params->upload(paramVector);
    indices->upload(indicesVector);
    int maxBuffers = 1;
723
    for (int i = 0; i < (int) forceBufferCounter.size(); i++)
724
        maxBuffers = max(maxBuffers, forceBufferCounter[i]);
725
    cl.addForce(new OpenCLPeriodicTorsionForceInfo(maxBuffers, force));
726
    cl::Program program = cl.createProgram(OpenCLKernelSources::periodicTorsionForce);
727
728
729
    kernel = cl::Kernel(program, "calcPeriodicTorsionForce");
}

730
double OpenCLCalcPeriodicTorsionForceKernel::execute(ContextImpl& context, bool includeForces, bool includeEnergy) {
731
    if (numTorsions == 0)
732
        return 0.0;
733
734
735
736
737
738
739
740
741
742
    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());
    }
743
    cl.executeKernel(kernel, numTorsions);
744
745
746
    return 0.0;
}

747
748
class OpenCLRBTorsionForceInfo : public OpenCLForceInfo {
public:
749
    OpenCLRBTorsionForceInfo(int requiredBuffers, const RBTorsionForce& force) : OpenCLForceInfo(requiredBuffers), force(force) {
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
    }
    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);
768
        force.getTorsionParameters(group2, particle1, particle2, particle3, particle4, c0b, c1b, c2b, c3b, c4b, c5b);
769
770
771
772
773
774
775
776
777
778
779
780
781
782
        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) {
783
784
785
786
    int numContexts = cl.getPlatformData().contexts.size();
    int startIndex = cl.getContextIndex()*force.getNumTorsions()/numContexts;
    int endIndex = (cl.getContextIndex()+1)*force.getNumTorsions()/numContexts;
    numTorsions = endIndex-startIndex;
787
788
    if (numTorsions == 0)
        return;
789
790
    params = new OpenCLArray<mm_float8>(cl, numTorsions, "rbTorsionParams");
    indices = new OpenCLArray<mm_int8>(cl, numTorsions, "rbTorsionIndices");
791
792
793
794
795
796
    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;
797
        force.getTorsionParameters(startIndex+i, particle1, particle2, particle3, particle4, c0, c1, c2, c3, c4, c5);
798
799
800
        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]++);
801
802
803
804
805

    }
    params->upload(paramVector);
    indices->upload(indicesVector);
    int maxBuffers = 1;
806
    for (int i = 0; i < (int) forceBufferCounter.size(); i++)
807
        maxBuffers = max(maxBuffers, forceBufferCounter[i]);
808
    cl.addForce(new OpenCLRBTorsionForceInfo(maxBuffers, force));
809
    cl::Program program = cl.createProgram(OpenCLKernelSources::rbTorsionForce);
810
811
812
    kernel = cl::Kernel(program, "calcRBTorsionForce");
}

813
double OpenCLCalcRBTorsionForceKernel::execute(ContextImpl& context, bool includeForces, bool includeEnergy) {
814
    if (numTorsions == 0)
815
        return 0.0;
816
817
818
819
820
821
822
823
824
825
    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());
    }
826
    cl.executeKernel(kernel, numTorsions);
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
871
    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) {
872
873
874
875
    int numContexts = cl.getPlatformData().contexts.size();
    int startIndex = cl.getContextIndex()*force.getNumTorsions()/numContexts;
    int endIndex = (cl.getContextIndex()+1)*force.getNumTorsions()/numContexts;
    numTorsions = endIndex-startIndex;
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
    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++) {
            coeffVec.push_back(mm_float4(c[j][0], c[j][1], c[j][2], c[j][3]));
            coeffVec.push_back(mm_float4(c[j][4], c[j][5], c[j][6], c[j][7]));
            coeffVec.push_back(mm_float4(c[j][8], c[j][9], c[j][10], c[j][11]));
            coeffVec.push_back(mm_float4(c[j][12], c[j][13], c[j][14], c[j][15]));
        }
    }
    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];
902
        force.getTorsionParameters(startIndex+i, torsionMapsVec[i], ind.s0, ind.s1, ind.s2, ind.s3, ind.s4, ind.s5, ind.s6, ind.s7);
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
        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");
}

928
double OpenCLCalcCMAPTorsionForceKernel::execute(ContextImpl& context, bool includeForces, bool includeEnergy) {
929
    if (numTorsions == 0)
930
        return 0.0;
931
932
933
934
935
936
937
938
939
940
941
942
943
    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);
944
945
946
    return 0.0;
}

947
948
949
950
951
952
953
954
955
956
957
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
958
        particles.resize(4);
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
987
        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) {
988
989
990
991
    int numContexts = cl.getPlatformData().contexts.size();
    int startIndex = cl.getContextIndex()*force.getNumTorsions()/numContexts;
    int endIndex = (cl.getContextIndex()+1)*force.getNumTorsions()/numContexts;
    numTorsions = endIndex-startIndex;
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
    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;
1007
        force.getTorsionParameters(startIndex+i, particle1, particle2, particle3, particle4, parameters);
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
1065
        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");
}

1066
double OpenCLCalcCustomTorsionForceKernel::execute(ContextImpl& context, bool includeForces, bool includeEnergy) {
1067
    if (numTorsions == 0)
1068
        return 0.0;
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
    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];
1093
            kernel.setArg<cl::Memory>(nextIndex++, buffer.getMemory());
1094
1095
1096
1097
1098
1099
        }
    }
    cl.executeKernel(kernel, numTorsions);
    return 0.0;
}

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
1131
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;
};

1132
1133
1134
OpenCLCalcNonbondedForceKernel::~OpenCLCalcNonbondedForceKernel() {
    if (sigmaEpsilon != NULL)
        delete sigmaEpsilon;
1135
1136
1137
1138
    if (exceptionParams != NULL)
        delete exceptionParams;
    if (exceptionIndices != NULL)
        delete exceptionIndices;
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
    if (cosSinSums != NULL)
        delete cosSinSums;
    if (pmeGrid != NULL)
        delete pmeGrid;
    if (pmeBsplineModuliX != NULL)
        delete pmeBsplineModuliX;
    if (pmeBsplineModuliY != NULL)
        delete pmeBsplineModuliY;
    if (pmeBsplineModuliZ != NULL)
        delete pmeBsplineModuliZ;
    if (pmeBsplineTheta != NULL)
        delete pmeBsplineTheta;
    if (pmeBsplineDtheta != NULL)
        delete pmeBsplineDtheta;
    if (pmeAtomRange != NULL)
        delete pmeAtomRange;
    if (pmeAtomGridIndex != NULL)
        delete pmeAtomGridIndex;
1157
1158
    if (erfcTable != NULL)
        delete erfcTable;
1159
1160
1161
1162
    if (sort != NULL)
        delete sort;
    if (fft != NULL)
        delete fft;
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
}

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);
1187
    double sumSquaredCharges = 0.0;
1188
1189
    bool hasCoulomb = false;
    bool hasLJ = false;
1190
1191
1192
1193
    for (int i = 0; i < numParticles; i++) {
        double charge, sigma, epsilon;
        force.getParticleParameters(i, charge, sigma, epsilon);
        posq[i].w = (float) charge;
1194
        sigmaEpsilonVector[i] = mm_float2((float) (0.5*sigma), (float) (2.0*sqrt(epsilon)));
1195
        exclusionList[i].push_back(i);
1196
        sumSquaredCharges += charge*charge;
1197
1198
1199
1200
        if (charge != 0.0)
            hasCoulomb = true;
        if (epsilon != 0.0)
            hasLJ = true;
1201
1202
1203
1204
1205
1206
1207
1208
1209
    }
    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);
1210
    map<string, string> defines;
1211
1212
    defines["HAS_COULOMB"] = (hasCoulomb ? "1" : "0");
    defines["HAS_LENNARD_JONES"] = (hasLJ ? "1" : "0");
1213
    if (useCutoff) {
1214
1215
        // Compute the reaction field constants.

1216
1217
        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);
1218
1219
1220
        defines["REACTION_FIELD_K"] = doubleToString(reactionFieldK);
        defines["REACTION_FIELD_C"] = doubleToString(reactionFieldC);
    }
1221
    if (force.getUseDispersionCorrection() && cl.getContextIndex() == 0)
1222
1223
1224
        dispersionCoefficient = NonbondedForceImpl::calcDispersionCorrection(system, force);
    else
        dispersionCoefficient = 0.0;
1225
    double alpha = 0;
1226
1227
1228
1229
1230
1231
1232
1233
    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";
1234
        ewaldSelfEnergy = (cl.getContextIndex() == 0 ? -ONE_4PI_EPS0*alpha*sumSquaredCharges/std::sqrt(M_PI) : 0.0);
1235
1236
1237
1238
1239
1240
1241
1242
1243

        // 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));
1244
        cl::Program program = cl.createProgram(OpenCLKernelSources::ewald, replacements);
1245
1246
1247
1248
        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");
    }
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
    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";
1260
        ewaldSelfEnergy = (cl.getContextIndex() == 0 ? -ONE_4PI_EPS0*alpha*sumSquaredCharges/std::sqrt(M_PI) : 0.0);
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
        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");
        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");
        pmeBsplineDtheta = new OpenCLArray<mm_float4>(cl, PmeOrder*numParticles, "pmeBsplineDtheta");
        pmeAtomRange = new OpenCLArray<cl_int>(cl, gridSizeX*gridSizeY*gridSizeZ+1, "pmeAtomRange");
1278
1279
        pmeAtomGridIndex = new OpenCLArray<mm_int2>(cl, numParticles, "pmeAtomGridIndex");
        sort = new OpenCLSort<mm_int2>(cl, cl.getNumAtoms(), "int2", "value.y");
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
        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);
        }
    }
1342
1343
    else
        ewaldSelfEnergy = 0.0;
1344
1345
1346
1347
    
    // Tabulate values of erfc().

    if (force.getNonbondedMethod() == NonbondedForce::Ewald || force.getNonbondedMethod() == NonbondedForce::PME) {
1348
1349
1350
1351
1352
1353
1354
        const int tableSize = 2048;
        defines["ERFC_TABLE_SCALE"] = doubleToString((tableSize-1)/(alpha*force.getCutoffDistance()));
        erfcTable = new OpenCLArray<cl_float>(cl, tableSize, "ErfcTable", false, CL_MEM_READ_ONLY);
        vector<cl_float> erfcVector(tableSize);
        for (int i = 0; i < tableSize; ++i)
            erfcVector[i] = (float) erfc(i*(alpha*force.getCutoffDistance())/(tableSize-1));
        erfcTable->upload(erfcVector);
1355
        cl.getNonbondedUtilities().addArgument(OpenCLNonbondedUtilities::ParameterInfo("erfcTable", "float", 1, sizeof(cl_float), erfcTable->getDeviceBuffer()));
1356
    }
1357
1358
1359

    // Add the interaction to the default nonbonded kernel.
    
1360
    string source = cl.replaceStrings(OpenCLKernelSources::coulombLennardJones, defines);
1361
    cl.getNonbondedUtilities().addInteraction(useCutoff, usePeriodic, true, force.getCutoffDistance(), exclusionList, source);
Peter Eastman's avatar
Peter Eastman committed
1362
    if (hasLJ)
1363
        cl.getNonbondedUtilities().addParameter(OpenCLNonbondedUtilities::ParameterInfo("sigmaEpsilon", "float", 2, sizeof(cl_float2), sigmaEpsilon->getDeviceBuffer()));
1364

1365
    // Initialize the exceptions.
1366

1367
1368
1369
1370
    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;
1371
    int maxBuffers = cl.getNonbondedUtilities().getNumForceBuffers();
1372
1373
1374
1375
1376
1377
    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);
1378
        for (int i = 0; i < numExceptions; i++) {
1379
1380
            int particle1, particle2;
            double chargeProd, sigma, epsilon;
1381
            force.getExceptionParameters(exceptions[startIndex+i], particle1, particle2, chargeProd, sigma, epsilon);
1382
1383
            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]++);
1384
        }
1385
1386
        exceptionParams->upload(exceptionParamsVector);
        exceptionIndices->upload(exceptionIndicesVector);
1387
        for (int i = 0; i < (int) forceBufferCounter.size(); i++)
1388
            maxBuffers = max(maxBuffers, forceBufferCounter[i]);
1389
    }
1390
    cl.addForce(new OpenCLNonbondedForceInfo(maxBuffers, force));
1391
    defines.clear();
Peter Eastman's avatar
Bug fix  
Peter Eastman committed
1392
    defines["NUM_ATOMS"] = intToString(cl.getPaddedNumAtoms());
1393
    defines["NUM_EXCEPTIONS"] = intToString(numExceptions);
1394
    cl::Program program = cl.createProgram(OpenCLKernelSources::nonbondedExceptions, defines);
1395
    exceptionsKernel = cl::Kernel(program, "computeNonbondedExceptions");
1396
1397
}

1398
double OpenCLCalcNonbondedForceKernel::execute(ContextImpl& context, bool includeForces, bool includeEnergy) {
1399
    bool deviceIsCpu = (cl.getDevice().getInfo<CL_DEVICE_TYPE>() == CL_DEVICE_TYPE_CPU);
1400
1401
1402
    if (!hasInitializedKernel) {
        hasInitializedKernel = true;
        if (exceptionIndices != NULL) {
1403
1404
1405
1406
1407
            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());
1408
1409
1410
1411
1412
1413
1414
1415
1416
        }
        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());
        }
1417
        if (pmeGrid != NULL) {
1418
1419
            string file = (deviceIsCpu ? OpenCLKernelSources::pme_cpu : OpenCLKernelSources::pme);
            cl::Program program = cl.createProgram(file, pmeDefines);
1420
            pmeUpdateBsplinesKernel = cl::Kernel(program, "updateBsplines");
1421
            pmeAtomRangeKernel = cl::Kernel(program, "findAtomRangeForGrid");
1422
1423
1424
1425
1426
1427
1428
            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());
            pmeUpdateBsplinesKernel.setArg<cl::Buffer>(2, pmeBsplineDtheta->getDeviceBuffer());
            pmeUpdateBsplinesKernel.setArg(3, 2*OpenCLContext::ThreadBlockSize*PmeOrder*sizeof(mm_float4), NULL);
Peter Eastman's avatar
Peter Eastman committed
1429
            pmeUpdateBsplinesKernel.setArg<cl::Buffer>(4, pmeAtomGridIndex->getDeviceBuffer());
1430
1431
1432
            pmeAtomRangeKernel.setArg<cl::Buffer>(0, pmeAtomGridIndex->getDeviceBuffer());
            pmeAtomRangeKernel.setArg<cl::Buffer>(1, pmeAtomRange->getDeviceBuffer());
            pmeAtomRangeKernel.setArg<cl::Buffer>(2, cl.getPosq().getDeviceBuffer());
1433
1434
1435
1436
1437
            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());
1438
1439
1440
1441
1442
1443
1444
1445
1446
1447
1448
            pmeConvolutionKernel.setArg<cl::Buffer>(0, pmeGrid->getDeviceBuffer());
            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());
            pmeInterpolateForceKernel.setArg<cl::Buffer>(2, pmeBsplineTheta->getDeviceBuffer());
            pmeInterpolateForceKernel.setArg<cl::Buffer>(3, pmeBsplineDtheta->getDeviceBuffer());
            pmeInterpolateForceKernel.setArg<cl::Buffer>(4, pmeGrid->getDeviceBuffer());
       }
1449
    }
1450
1451
    if (exceptionIndices != NULL)
        cl.executeKernel(exceptionsKernel, exceptionIndices->getSize());
1452
    if (cosSinSums != NULL && cl.getContextIndex() == 0) {
1453
1454
1455
1456
1457
        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);
        float recipCoefficient = ONE_4PI_EPS0*4*M_PI/(boxSize.x*boxSize.y*boxSize.z);
        ewaldSumsKernel.setArg<mm_float4>(3, recipBoxSize);
        ewaldSumsKernel.setArg<cl_float>(4, recipCoefficient);
1458
        cl.executeKernel(ewaldSumsKernel, cosSinSums->getSize());
1459
1460
        ewaldForcesKernel.setArg<mm_float4>(3, recipBoxSize);
        ewaldForcesKernel.setArg<cl_float>(4, recipCoefficient);
1461
1462
        cl.executeKernel(ewaldForcesKernel, cl.getNumAtoms());
    }
1463
    if (pmeGrid != NULL && cl.getContextIndex() == 0) {
1464
1465
        mm_float4 boxSize = cl.getPeriodicBoxSize();
        mm_float4 invBoxSize = cl.getInvPeriodicBoxSize();
1466
1467
        pmeUpdateBsplinesKernel.setArg<mm_float4>(5, boxSize);
        pmeUpdateBsplinesKernel.setArg<mm_float4>(6, invBoxSize);
1468
        cl.executeKernel(pmeUpdateBsplinesKernel, cl.getNumAtoms());
1469
1470
1471
1472
1473
1474
1475
1476
1477
1478
1479
1480
        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());
            cl.executeKernel(pmeSpreadChargeKernel, cl.getNumAtoms());
        }
1481
        fft->execFFT(*pmeGrid, true);
1482
1483
        pmeConvolutionKernel.setArg<mm_float4>(5, invBoxSize);
        pmeConvolutionKernel.setArg<cl_float>(6, (float) (1.0/(M_PI*boxSize.x*boxSize.y*boxSize.z)));
1484
1485
        cl.executeKernel(pmeConvolutionKernel, cl.getNumAtoms());
        fft->execFFT(*pmeGrid, false);
1486
1487
        pmeInterpolateForceKernel.setArg<mm_float4>(5, boxSize);
        pmeInterpolateForceKernel.setArg<mm_float4>(6, invBoxSize);
1488
1489
        cl.executeKernel(pmeInterpolateForceKernel, cl.getNumAtoms());
    }
1490
1491
1492
1493
1494
1495
    double energy = ewaldSelfEnergy;
    if (dispersionCoefficient != 0.0) {
        mm_float4 boxSize = cl.getPeriodicBoxSize();
        energy += dispersionCoefficient/(boxSize.x*boxSize.y*boxSize.z);
    }
    return energy;
1496
1497
}

1498
1499
1500
1501
1502
1503
1504
1505
1506
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);
1507
        for (int i = 0; i < (int) params1.size(); i++)
1508
1509
1510
1511
1512
            if (params1[i] != params2[i])
                return false;
        return true;
    }
    int getNumParticleGroups() {
1513
        return force.getNumExclusions();
1514
1515
1516
    }
    void getParticlesInGroup(int index, std::vector<int>& particles) {
        int particle1, particle2;
1517
        force.getExclusionParticles(index, particle1, particle2);
1518
1519
1520
1521
1522
1523
1524
1525
1526
1527
1528
1529
1530
1531
1532
1533
        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;
1534
1535
1536
1537
    if (tabulatedFunctionParams != NULL)
        delete tabulatedFunctionParams;
    for (int i = 0; i < (int) tabulatedFunctions.size(); i++)
        delete tabulatedFunctions[i];
1538
1539
1540
1541
1542
1543
1544
1545
1546
1547
1548
}

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();
1549
    params = new OpenCLParameterSet(cl, force.getNumPerParticleParameters(), numParticles, "customNonbondedParameters");
1550
    if (force.getNumGlobalParameters() > 0)
1551
        globals = new OpenCLArray<cl_float>(cl, force.getNumGlobalParameters(), "customNonbondedGlobals", false, CL_MEM_READ_ONLY);
1552
    vector<vector<cl_float> > paramVector(numParticles);
1553
1554
1555
1556
    vector<vector<int> > exclusionList(numParticles);
    for (int i = 0; i < numParticles; i++) {
        vector<double> parameters;
        force.getParticleParameters(i, parameters);
1557
        paramVector[i].resize(parameters.size());
1558
        for (int j = 0; j < (int) parameters.size(); j++)
1559
            paramVector[i][j] = (cl_float) parameters[j];
1560
1561
        exclusionList[i].push_back(i);
    }
1562
1563
1564
1565
1566
    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);
1567
    }
1568
    params->setParameterValues(paramVector);
1569
1570
1571

    // Record the tabulated functions.

1572
    OpenCLExpressionUtilities::FunctionPlaceholder fp;
1573
1574
1575
    map<string, Lepton::CustomFunction*> functions;
    vector<pair<string, string> > functionDefinitions;
    vector<mm_float4> tabulatedFunctionParamsVec(force.getNumFunctions());
1576
1577
1578
1579
    for (int i = 0; i < force.getNumFunctions(); i++) {
        string name;
        vector<double> values;
        double min, max;
1580
        force.getFunctionParameters(i, name, values, min, max);
1581
1582
        string arrayName = prefix+"table"+intToString(i);
        functionDefinitions.push_back(make_pair(name, arrayName));
1583
        functions[name] = &fp;
1584
        tabulatedFunctionParamsVec[i] = mm_float4((float) min, (float) max, (float) ((values.size()-1)/(max-min)), values.size()-2);
1585
        vector<mm_float4> f = OpenCLExpressionUtilities::computeFunctionCoefficients(values, min, max);
1586
1587
        tabulatedFunctions.push_back(new OpenCLArray<mm_float4>(cl, values.size()-1, "TabulatedFunction"));
        tabulatedFunctions[tabulatedFunctions.size()-1]->upload(f);
1588
        cl.getNonbondedUtilities().addArgument(OpenCLNonbondedUtilities::ParameterInfo(arrayName, "float", 4, sizeof(cl_float4), tabulatedFunctions[tabulatedFunctions.size()-1]->getDeviceBuffer()));
1589
1590
1591
1592
    }
    if (force.getNumFunctions() > 0) {
        tabulatedFunctionParams = new OpenCLArray<mm_float4>(cl, tabulatedFunctionParamsVec.size(), "tabulatedFunctionParameters", false, CL_MEM_READ_ONLY);
        tabulatedFunctionParams->upload(tabulatedFunctionParamsVec);
1593
        cl.getNonbondedUtilities().addArgument(OpenCLNonbondedUtilities::ParameterInfo(prefix+"functionParams", "float", 4, sizeof(cl_float4), tabulatedFunctionParams->getDeviceBuffer()));
1594
1595
1596
1597
1598
1599
1600
1601
1602
1603
1604
1605
1606
1607
    }

    // 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);
1608
    Lepton::ParsedExpression energyExpression = Lepton::Parser::parse(force.getEnergyFunction(), functions).optimize();
1609
    Lepton::ParsedExpression forceExpression = energyExpression.differentiate("r").optimize();
1610
1611
1612
    map<string, Lepton::ParsedExpression> forceExpressions;
    forceExpressions["tempEnergy += "] = energyExpression;
    forceExpressions["tempForce -= "] = forceExpression;
1613
1614
1615

    // Create the kernels.

1616
1617
1618
1619
    map<string, string> variables;
    variables["r"] = "r";
    for (int i = 0; i < force.getNumPerParticleParameters(); i++) {
        const string& name = force.getPerParticleParameterName(i);
1620
1621
        variables[name+"1"] = prefix+"params"+params->getParameterSuffix(i, "1");
        variables[name+"2"] = prefix+"params"+params->getParameterSuffix(i, "2");
1622
1623
1624
1625
    }
    for (int i = 0; i < force.getNumGlobalParameters(); i++) {
        const string& name = force.getGlobalParameterName(i);
        string value = "globals["+intToString(i)+"]";
1626
        variables[name] = prefix+value;
1627
    }
1628
    stringstream compute;
1629
    compute << OpenCLExpressionUtilities::createExpressions(forceExpressions, variables, functionDefinitions, prefix+"temp", prefix+"functionParams");
1630
1631
    map<string, string> replacements;
    replacements["COMPUTE_FORCE"] = compute.str();
1632
    string source = cl.replaceStrings(OpenCLKernelSources::customNonbonded, replacements);
1633
    cl.getNonbondedUtilities().addInteraction(useCutoff, usePeriodic, true, force.getCutoffDistance(), exclusionList, source);
1634
1635
    for (int i = 0; i < (int) params->getBuffers().size(); i++) {
        const OpenCLNonbondedUtilities::ParameterInfo& buffer = params->getBuffers()[i];
1636
        cl.getNonbondedUtilities().addParameter(OpenCLNonbondedUtilities::ParameterInfo(prefix+"params"+intToString(i+1), buffer.getComponentType(), buffer.getNumComponents(), buffer.getSize(), buffer.getMemory()));
1637
    }
1638
1639
    if (globals != NULL) {
        globals->upload(globalParamValues);
1640
        cl.getNonbondedUtilities().addArgument(OpenCLNonbondedUtilities::ParameterInfo(prefix+"globals", "float", 1, sizeof(cl_float), globals->getDeviceBuffer()));
1641
    }
1642
    cl.addForce(new OpenCLCustomNonbondedForceInfo(cl.getNonbondedUtilities().getNumForceBuffers(), force));
1643
1644
}

1645
double OpenCLCalcCustomNonbondedForceKernel::execute(ContextImpl& context, bool includeForces, bool includeEnergy) {
1646
1647
    if (globals != NULL) {
        bool changed = false;
1648
        for (int i = 0; i < (int) globalParamNames.size(); i++) {
1649
1650
1651
1652
1653
1654
1655
1656
            cl_float value = (cl_float) context.getParameter(globalParamNames[i]);
            if (value != globalParamValues[i])
                changed = true;
            globalParamValues[i] = value;
        }
        if (changed)
            globals->upload(globalParamValues);
    }
1657
1658
    return 0.0;
}
Peter Eastman's avatar
Peter Eastman committed
1659
1660
1661
1662
1663
1664
1665
1666
1667
1668
1669
1670
1671
1672
1673

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;
};

1674
1675
1676
1677
1678
1679
1680
1681
1682
1683
1684
1685
1686
1687
OpenCLCalcGBSAOBCForceKernel::~OpenCLCalcGBSAOBCForceKernel() {
    if (params != NULL)
        delete params;
    if (bornSum != NULL)
        delete bornSum;
    if (bornRadii != NULL)
        delete bornRadii;
    if (bornForce != NULL)
        delete bornForce;
    if (obcChain != NULL)
        delete obcChain;
}

void OpenCLCalcGBSAOBCForceKernel::initialize(const System& system, const GBSAOBCForce& force) {
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
    bornSum = new OpenCLArray<cl_float>(cl, cl.getPaddedNumAtoms()*nb.getNumForceBuffers(), "bornSum");
    bornForce = new OpenCLArray<cl_float>(cl, cl.getPaddedNumAtoms()*nb.getNumForceBuffers(), "bornForce");
1694
1695
1696
1697
1698
1699
1700
1701
    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;
1702
        paramsVector[i] = mm_float2((float) radius, (float) (scalingFactor*radius));
1703
1704
1705
1706
        posq[i].w = (float) charge;
    }
    posq.upload();
    params->upload(paramsVector);
1707
    prefactor = -ONE_4PI_EPS0*((1.0/force.getSoluteDielectric())-(1.0/force.getSolventDielectric()));
1708
1709
    bool useCutoff = (force.getNonbondedMethod() != GBSAOBCForce::NoCutoff);
    bool usePeriodic = (force.getNonbondedMethod() != GBSAOBCForce::NoCutoff && force.getNonbondedMethod() != GBSAOBCForce::CutoffNonPeriodic);
1710
    string source = OpenCLKernelSources::gbsaObc2;
1711
    nb.addInteraction(useCutoff, usePeriodic, false, force.getCutoffDistance(), vector<vector<int> >(), source);
1712
1713
    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
1714
    cl.addForce(new OpenCLGBSAOBCForceInfo(nb.getNumForceBuffers(), force));
1715
1716
    cl.addAutoclearBuffer(bornSum->getDeviceBuffer(), bornSum->getSize());
    cl.addAutoclearBuffer(bornForce->getDeviceBuffer(), bornForce->getSize());
1717
1718
}

1719
double OpenCLCalcGBSAOBCForceKernel::execute(ContextImpl& context, bool includeForces, bool includeEnergy) {
1720
    OpenCLNonbondedUtilities& nb = cl.getNonbondedUtilities();
1721
    bool deviceIsCpu = (cl.getDevice().getInfo<CL_DEVICE_TYPE>() == CL_DEVICE_TYPE_CPU);
1722
1723
1724
1725
    if (!hasCreatedKernels) {
        // These Kernels cannot be created in initialize(), because the OpenCLNonbondedUtilities has not been initialized yet then.

        hasCreatedKernels = true;
1726
        maxTiles = (nb.getUseCutoff() ? nb.getInteractingTiles().getSize() : 0);
1727
1728
1729
1730
1731
1732
1733
        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";
1734
1735
1736
1737
        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());
1738
        defines["NUM_BLOCKS"] = OpenCLExpressionUtilities::intToString(cl.getNumAtomBlocks());
1739
1740
1741
1742
1743
1744
1745
        string file;
        if (deviceIsCpu)
            file = OpenCLKernelSources::gbsaObc_cpu;
        else if (cl.getSIMDWidth() == 32)
            file = OpenCLKernelSources::gbsaObc_nvidia;
        else
            file = OpenCLKernelSources::gbsaObc_default;
1746
        cl::Program program = cl.createProgram(file, defines);
1747
        int index = 0;
1748
        computeBornSumKernel = cl::Kernel(program, "computeBornSum");
1749
1750
1751
        computeBornSumKernel.setArg<cl::Buffer>(index++, bornSum->getDeviceBuffer());
        computeBornSumKernel.setArg<cl::Buffer>(index++, cl.getPosq().getDeviceBuffer());
        computeBornSumKernel.setArg<cl::Buffer>(index++, params->getDeviceBuffer());
1752
        computeBornSumKernel.setArg(index++, (deviceIsCpu ? OpenCLContext::TileSize : OpenCLContext::ThreadBlockSize)*13*sizeof(cl_float), NULL);
1753
        computeBornSumKernel.setArg(index++, (deviceIsCpu ? 1 : OpenCLContext::ThreadBlockSize)*sizeof(cl_float), NULL);
1754
        if (nb.getUseCutoff()) {
1755
1756
            computeBornSumKernel.setArg<cl::Buffer>(index++, nb.getInteractingTiles().getDeviceBuffer());
            computeBornSumKernel.setArg<cl::Buffer>(index++, nb.getInteractionCount().getDeviceBuffer());
1757
1758
            index += 2; // The periodic box size arguments are set when the kernel is executed.
            computeBornSumKernel.setArg<cl_uint>(index++, maxTiles);
1759
            if (cl.getSIMDWidth() == 32 || deviceIsCpu)
1760
                computeBornSumKernel.setArg<cl::Buffer>(index++, nb.getInteractionFlags().getDeviceBuffer());
1761
        }
1762
1763
        else
            computeBornSumKernel.setArg<cl_uint>(index++, cl.getNumAtomBlocks()*(cl.getNumAtomBlocks()+1)/2);
1764
        force1Kernel = cl::Kernel(program, "computeGBSAForce1");
1765
1766
1767
1768
1769
1770
        index = 0;
        force1Kernel.setArg<cl::Buffer>(index++, cl.getForceBuffers().getDeviceBuffer());
        force1Kernel.setArg<cl::Buffer>(index++, cl.getEnergyBuffer().getDeviceBuffer());
        force1Kernel.setArg<cl::Buffer>(index++, cl.getPosq().getDeviceBuffer());
        force1Kernel.setArg<cl::Buffer>(index++, bornRadii->getDeviceBuffer());
        force1Kernel.setArg<cl::Buffer>(index++, bornForce->getDeviceBuffer());
1771
        force1Kernel.setArg(index++, (deviceIsCpu ? OpenCLContext::TileSize : OpenCLContext::ThreadBlockSize)*13*sizeof(cl_float), NULL);
1772
        force1Kernel.setArg(index++, (deviceIsCpu ? 1 : OpenCLContext::ThreadBlockSize)*sizeof(mm_float4), NULL);
1773
        if (nb.getUseCutoff()) {
1774
1775
            force1Kernel.setArg<cl::Buffer>(index++, nb.getInteractingTiles().getDeviceBuffer());
            force1Kernel.setArg<cl::Buffer>(index++, nb.getInteractionCount().getDeviceBuffer());
1776
1777
            index += 2; // The periodic box size arguments are set when the kernel is executed.
            force1Kernel.setArg<cl_uint>(index++, maxTiles);
1778
            if (cl.getSIMDWidth() == 32 || deviceIsCpu)
1779
                force1Kernel.setArg<cl::Buffer>(index++, nb.getInteractionFlags().getDeviceBuffer());
1780
        }
1781
1782
        else
            force1Kernel.setArg<cl_uint>(index++, cl.getNumAtomBlocks()*(cl.getNumAtomBlocks()+1)/2);
1783
        program = cl.createProgram(OpenCLKernelSources::gbsaObcReductions, defines);
1784
1785
        reduceBornSumKernel = cl::Kernel(program, "reduceBornSum");
        reduceBornSumKernel.setArg<cl_int>(0, cl.getPaddedNumAtoms());
Peter Eastman's avatar
Peter Eastman committed
1786
        reduceBornSumKernel.setArg<cl_int>(1, nb.getNumForceBuffers());
1787
1788
1789
1790
1791
1792
1793
        reduceBornSumKernel.setArg<cl_float>(2, 1.0f);
        reduceBornSumKernel.setArg<cl_float>(3, 0.8f);
        reduceBornSumKernel.setArg<cl_float>(4, 4.85f);
        reduceBornSumKernel.setArg<cl::Buffer>(5, bornSum->getDeviceBuffer());
        reduceBornSumKernel.setArg<cl::Buffer>(6, params->getDeviceBuffer());
        reduceBornSumKernel.setArg<cl::Buffer>(7, bornRadii->getDeviceBuffer());
        reduceBornSumKernel.setArg<cl::Buffer>(8, obcChain->getDeviceBuffer());
1794
        reduceBornForceKernel = cl::Kernel(program, "reduceBornForce");
1795
        reduceBornForceKernel.setArg<cl_int>(0, cl.getPaddedNumAtoms());
Peter Eastman's avatar
Peter Eastman committed
1796
        reduceBornForceKernel.setArg<cl_int>(1, nb.getNumForceBuffers());
1797
1798
1799
1800
1801
        reduceBornForceKernel.setArg<cl::Buffer>(2, bornForce->getDeviceBuffer());
        reduceBornForceKernel.setArg<cl::Buffer>(3, cl.getEnergyBuffer().getDeviceBuffer());
        reduceBornForceKernel.setArg<cl::Buffer>(4, params->getDeviceBuffer());
        reduceBornForceKernel.setArg<cl::Buffer>(5, bornRadii->getDeviceBuffer());
        reduceBornForceKernel.setArg<cl::Buffer>(6, obcChain->getDeviceBuffer());
1802
    }
1803
    if (nb.getUseCutoff()) {
1804
1805
1806
1807
        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());
1808
1809
1810
1811
1812
        if (maxTiles < nb.getInteractingTiles().getSize()) {
            maxTiles = nb.getInteractingTiles().getSize();
            computeBornSumKernel.setArg<cl_uint>(9, maxTiles);
            force1Kernel.setArg<cl_uint>(11, maxTiles);
        }
1813
    }
1814
    int numTiles = cl.getNumAtomBlocks()*(cl.getNumAtomBlocks()+1)/2;
1815
    cl.executeKernel(computeBornSumKernel, numTiles*OpenCLContext::TileSize, deviceIsCpu ? 1 : -1);
1816
    cl.executeKernel(reduceBornSumKernel, cl.getPaddedNumAtoms());
1817
    cl.executeKernel(force1Kernel, numTiles*OpenCLContext::TileSize, deviceIsCpu ? 1 : -1);
1818
    cl.executeKernel(reduceBornForceKernel, cl.getPaddedNumAtoms());
1819
    return 0.0;
1820
}
1821
1822
1823
1824
1825
1826
1827
1828
1829
1830

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);
1831
        for (int i = 0; i < (int) params1.size(); i++)
1832
1833
1834
1835
1836
1837
1838
1839
1840
1841
1842
1843
1844
1845
1846
1847
1848
1849
1850
1851
1852
1853
1854
1855
1856
1857
            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;
1858
1859
    if (energyDerivs != NULL)
        delete energyDerivs;
1860
1861
    if (globals != NULL)
        delete globals;
1862
1863
    if (valueBuffers != NULL)
        delete valueBuffers;
1864
1865
1866
1867
1868
1869
1870
1871
    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) {
    bool useExclusionsForValue = false;
1872
1873
    vector<string> computedValueNames(force.getNumComputedValues());
    vector<string> computedValueExpressions(force.getNumComputedValues());
1874
1875
    if (force.getNumComputedValues() > 0) {
        CustomGBForce::ComputationType type;
1876
        force.getComputedValueParameters(0, computedValueNames[0], computedValueExpressions[0], type);
1877
1878
1879
1880
        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++) {
1881
            force.getComputedValueParameters(i, computedValueNames[i], computedValueExpressions[i], type);
1882
1883
1884
1885
1886
1887
1888
1889
1890
1891
1892
1893
1894
1895
1896
1897
1898
1899
1900
1901
1902
1903
            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());
1904
        for (int j = 0; j < (int) parameters.size(); j++)
1905
1906
1907
1908
1909
1910
1911
1912
1913
1914
1915
1916
1917
1918
1919
1920
1921
            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());
1922
    stringstream tableArgs;
1923
1924
1925
1926
    for (int i = 0; i < force.getNumFunctions(); i++) {
        string name;
        vector<double> values;
        double min, max;
1927
        force.getFunctionParameters(i, name, values, min, max);
1928
1929
1930
        string arrayName = prefix+"table"+intToString(i);
        functionDefinitions.push_back(make_pair(name, arrayName));
        functions[name] = &fp;
1931
        tabulatedFunctionParamsVec[i] = mm_float4((float) min, (float) max, (float) ((values.size()-1)/(max-min)), values.size()-2);
1932
        vector<mm_float4> f = OpenCLExpressionUtilities::computeFunctionCoefficients(values, min, max);
1933
1934
        tabulatedFunctions.push_back(new OpenCLArray<mm_float4>(cl, values.size()-1, "TabulatedFunction"));
        tabulatedFunctions[tabulatedFunctions.size()-1]->upload(f);
1935
        cl.getNonbondedUtilities().addArgument(OpenCLNonbondedUtilities::ParameterInfo(arrayName, "float", 4, sizeof(cl_float4), tabulatedFunctions[tabulatedFunctions.size()-1]->getDeviceBuffer()));
1936
        tableArgs << ", __global float4* " << arrayName;
1937
1938
1939
1940
    }
    if (force.getNumFunctions() > 0) {
        tabulatedFunctionParams = new OpenCLArray<mm_float4>(cl, tabulatedFunctionParamsVec.size(), "tabulatedFunctionParameters", false, CL_MEM_READ_ONLY);
        tabulatedFunctionParams->upload(tabulatedFunctionParamsVec);
1941
        cl.getNonbondedUtilities().addArgument(OpenCLNonbondedUtilities::ParameterInfo(prefix+"functionParams", "float", 4, sizeof(cl_float4), tabulatedFunctionParams->getDeviceBuffer()));
1942
        tableArgs << ", __constant float4* " << prefix << "functionParams";
1943
1944
    }

1945
    // Record the global parameters.
1946
1947
1948
1949
1950
1951
1952
1953
1954

    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);
1955
1956
1957

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

1958
    vector<vector<Lepton::ParsedExpression> > valueGradientExpressions(force.getNumComputedValues());
1959
    vector<vector<Lepton::ParsedExpression> > valueDerivExpressions(force.getNumComputedValues());
Peter Eastman's avatar
Peter Eastman committed
1960
    needParameterGradient = false;
1961
1962
1963
1964
1965
1966
1967
    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;
1968
1969
         for (int j = 0; j < i; j++)
            valueDerivExpressions[i].push_back(ex.differentiate(computedValueNames[j]).optimize());
1970
    }
1971
    vector<vector<Lepton::ParsedExpression> > energyDerivExpressions(force.getNumEnergyTerms());
1972
1973
1974
1975
1976
1977
1978
1979
1980
1981
1982
1983
1984
1985
1986
1987
    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());
            }
        }
    }
    energyDerivs = new OpenCLParameterSet(cl, force.getNumComputedValues(), cl.getPaddedNumAtoms()*cl.getNonbondedUtilities().getNumForceBuffers(), "customGBEnergyDerivatives");

1988
1989
    // Create the kernels.

1990
1991
    bool useCutoff = (force.getNonbondedMethod() != CustomGBForce::NoCutoff);
    bool usePeriodic = (force.getNonbondedMethod() != CustomGBForce::NoCutoff && force.getNonbondedMethod() != CustomGBForce::CutoffNonPeriodic);
1992
    bool deviceIsCpu = (cl.getDevice().getInfo<CL_DEVICE_TYPE>() == CL_DEVICE_TYPE_CPU);
1993
1994
1995
    {
        // Create the N2 value kernel.

1996
1997
1998
        map<string, string> variables;
        map<string, string> rename;
        variables["r"] = "r";
1999
2000
        for (int i = 0; i < force.getNumPerParticleParameters(); i++) {
            const string& name = force.getPerParticleParameterName(i);
2001
2002
2003
2004
            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";
2005
2006
2007
2008
        }
        for (int i = 0; i < force.getNumGlobalParameters(); i++) {
            const string& name = force.getGlobalParameterName(i);
            string value = "globals["+intToString(i)+"]";
2009
            variables[name] = value;
2010
        }
2011
2012
        map<string, Lepton::ParsedExpression> n2ValueExpressions;
        stringstream n2ValueSource;
2013
2014
2015
2016
        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");
2017
2018
2019
2020
2021
2022
2023
        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];
2024
            string paramName = "params"+intToString(i+1);
2025
            extraArgs << ", __global " << buffer.getType() << "* global_" << paramName << ", __local " << buffer.getType() << "* local_" << paramName;
2026
2027
            loadLocal1 << "local_" << paramName << "[localAtomIndex] = " << paramName << "1;\n";
            loadLocal2 << "local_" << paramName << "[localAtomIndex] = global_" << paramName << "[j];\n";
2028
2029
2030
            load1 << buffer.getType() << " " << paramName << "1 = global_" << paramName << "[atom1];\n";
            load2 << buffer.getType() << " " << paramName << "2 = local_" << paramName << "[atom2];\n";
        }
2031
        replacements["PARAMETER_ARGUMENTS"] = extraArgs.str()+tableArgs.str();
2032
2033
2034
2035
2036
2037
2038
2039
2040
2041
2042
2043
2044
2045
2046
2047
        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";
        defines["CUTOFF_SQUARED"] = doubleToString(force.getCutoffDistance()*force.getCutoffDistance());
        defines["NUM_ATOMS"] = intToString(cl.getNumAtoms());
        defines["PADDED_NUM_ATOMS"] = intToString(cl.getPaddedNumAtoms());
2048
        defines["NUM_BLOCKS"] = OpenCLExpressionUtilities::intToString(cl.getNumAtomBlocks());
2049
2050
2051
2052
2053
2054
        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
2055
            file = OpenCLKernelSources::customGBValueN2_default;
2056
        cl::Program program = cl.createProgram(cl.replaceStrings(file, replacements), defines);
2057
2058
2059
2060
2061
2062
2063
2064
2065
2066
        pairValueKernel = cl::Kernel(program, "computeN2Value");
    }
    {
        // 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];
2067
2068
            string paramName = "params"+intToString(i+1);
            extraArgs << ", __global " << buffer.getType() << "* " << paramName;
2069
2070
2071
        }
        for (int i = 0; i < (int) computedValues->getBuffers().size(); i++) {
            const OpenCLNonbondedUtilities::ParameterInfo& buffer = computedValues->getBuffers()[i];
2072
            string valueName = "values"+intToString(i+1);
2073
2074
2075
2076
            extraArgs << ", __global " << buffer.getType() << "* global_" << valueName;
            reductionSource << buffer.getType() << " local_" << valueName << ";\n";
        }
        reductionSource << "local_values" << computedValues->getParameterSuffix(0) << " = sum;\n";
2077
        map<string, string> variables;
2078
2079
2080
        variables["x"] = "pos.x";
        variables["y"] = "pos.y";
        variables["z"] = "pos.z";
2081
2082
2083
2084
2085
2086
2087
2088
2089
2090
        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");
        }
2091
        for (int i = 0; i < (int) computedValues->getBuffers().size(); i++) {
2092
            string valueName = "values"+intToString(i+1);
2093
2094
2095
            reductionSource << "global_" << valueName << "[index] = local_" << valueName << ";\n";
        }
        map<string, string> replacements;
2096
        replacements["PARAMETER_ARGUMENTS"] = extraArgs.str()+tableArgs.str();
2097
2098
2099
        replacements["COMPUTE_VALUES"] = reductionSource.str();
        map<string, string> defines;
        defines["NUM_ATOMS"] = intToString(cl.getNumAtoms());
2100
        cl::Program program = cl.createProgram(cl.replaceStrings(OpenCLKernelSources::customGBValuePerParticle, replacements), defines);
2101
2102
2103
2104
2105
2106
2107
2108
2109
2110
2111
2112
2113
2114
2115
2116
2117
2118
        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)+"]";
2119
        stringstream n2EnergySource;
2120
2121
2122
2123
2124
2125
2126
2127
        bool anyExclusions = false;
        for (int i = 0; i < force.getNumEnergyTerms(); i++) {
            string expression;
            CustomGBForce::ComputationType type;
            force.getEnergyTermParameters(i, expression, type);
            if (type == CustomGBForce::SingleParticle)
                continue;
            bool exclude = (type == CustomGBForce::ParticlePair);
2128
            anyExclusions |= exclude;
2129
            map<string, Lepton::ParsedExpression> n2EnergyExpressions;
2130
2131
2132
2133
2134
2135
2136
2137
2138
2139
2140
            n2EnergyExpressions["tempEnergy += "] = Lepton::Parser::parse(expression, functions).optimize();
            n2EnergyExpressions["dEdR += "] = Lepton::Parser::parse(expression, functions).differentiate("r").optimize();
            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];
            }
            if (exclude)
                n2EnergySource << "if (!isExcluded) {\n";
            n2EnergySource << OpenCLExpressionUtilities::createExpressions(n2EnergyExpressions, variables, functionDefinitions, "temp", prefix+"functionParams");
            if (exclude)
                n2EnergySource << "}\n";
2141
2142
        }
        map<string, string> replacements;
2143
        replacements["COMPUTE_INTERACTION"] = n2EnergySource.str();
2144
        stringstream extraArgs, loadLocal1, loadLocal2, clearLocal, load1, load2, recordDeriv, storeDerivs1, storeDerivs2, declareTemps, setTemps;
2145
2146
2147
2148
2149
2150
        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;
2151
2152
            loadLocal1 << "local_" << paramName << "[localAtomIndex] = " << paramName << "1;\n";
            loadLocal2 << "local_" << paramName << "[localAtomIndex] = global_" << paramName << "[j];\n";
2153
2154
2155
2156
2157
2158
2159
            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;
2160
2161
            loadLocal1 << "local_" << valueName << "[localAtomIndex] = " << valueName << "1;\n";
            loadLocal2 << "local_" << valueName << "[localAtomIndex] = global_" << valueName << "[j];\n";
2162
2163
2164
            load1 << buffer.getType() << " " << valueName << "1 = global_" << valueName << "[atom1];\n";
            load2 << buffer.getType() << " " << valueName << "2 = local_" << valueName << "[atom2];\n";
        }
2165
2166
2167
2168
        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;
2169
            clearLocal << "local_deriv" << index << "[localAtomIndex] = 0.0f;\n";
2170
2171
            load1 << buffer.getType() << " deriv" << index << "_1 = 0.0f;\n";
            load2 << buffer.getType() << " deriv" << index << "_2 = 0.0f;\n";
2172
            recordDeriv << "local_deriv" << index << "[atom2] += deriv" << index << "_2;\n";
2173
2174
2175
2176
            storeDerivs1 << "STORE_DERIVATIVE_1(" << index << ")";
            storeDerivs2 << "STORE_DERIVATIVE_2(" << index << ")";
            declareTemps << "__local " << buffer.getType() << " tempDerivBuffer" << index << "[64];\n";
            setTemps << "tempDerivBuffer" << index << "[get_local_id(0)] = deriv" << index << "_1;\n";
2177
        }
2178
2179
2180
        replacements["PARAMETER_ARGUMENTS"] = extraArgs.str()+tableArgs.str();
        replacements["LOAD_LOCAL_PARAMETERS_FROM_1"] = loadLocal1.str();
        replacements["LOAD_LOCAL_PARAMETERS_FROM_GLOBAL"] = loadLocal2.str();
2181
        replacements["CLEAR_LOCAL_DERIVATIVES"] = clearLocal.str();
2182
2183
        replacements["LOAD_ATOM1_PARAMETERS"] = load1.str();
        replacements["LOAD_ATOM2_PARAMETERS"] = load2.str();
2184
2185
2186
        replacements["RECORD_DERIVATIVE_2"] = recordDeriv.str();
        replacements["STORE_DERIVATIVES_1"] = storeDerivs1.str();
        replacements["STORE_DERIVATIVES_2"] = storeDerivs2.str();
2187
2188
        replacements["DECLARE_TEMP_BUFFERS"] = declareTemps.str();
        replacements["SET_TEMP_BUFFERS"] = setTemps.str();
2189
2190
2191
2192
2193
2194
2195
2196
2197
2198
2199
2200
        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";
        defines["CUTOFF_SQUARED"] = doubleToString(force.getCutoffDistance()*force.getCutoffDistance());
        defines["NUM_ATOMS"] = intToString(cl.getNumAtoms());
        defines["PADDED_NUM_ATOMS"] = intToString(cl.getPaddedNumAtoms());
2201
        defines["NUM_BLOCKS"] = OpenCLExpressionUtilities::intToString(cl.getNumAtomBlocks());
2202
2203
2204
2205
2206
2207
2208
        string file;
        if (deviceIsCpu)
            file = OpenCLKernelSources::customGBEnergyN2_cpu;
        else if (cl.getSIMDWidth() == 32)
            file = OpenCLKernelSources::customGBEnergyN2_nvidia;
        else
            file = OpenCLKernelSources::customGBEnergyN2_default;
2209
        cl::Program program = cl.createProgram(cl.replaceStrings(file, replacements), defines);
2210
2211
2212
2213
2214
        pairEnergyKernel = cl::Kernel(program, "computeN2Energy");
    }
    {
        // Create the kernel to reduce the derivatives and calculate per-particle energy terms.

2215
        stringstream compute, extraArgs, reduce;
2216
2217
2218
2219
2220
2221
2222
2223
2224
2225
2226
2227
        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;
        }
2228
2229
2230
2231
2232
2233
2234
        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;
            reduce << "REDUCE_VALUE(derivBuffers" << index << ", " << buffer.getType() << ")\n";
            compute << buffer.getType() << " deriv" << index << " = derivBuffers" << index << "[index];\n";
        }
2235
        map<string, string> variables;
2236
2237
2238
        variables["x"] = "pos.x";
        variables["y"] = "pos.y";
        variables["z"] = "pos.z";
2239
2240
2241
2242
2243
2244
        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]");
2245
        map<string, Lepton::ParsedExpression> energyExpressions;
2246
2247
2248
2249
2250
2251
        for (int i = 0; i < force.getNumEnergyTerms(); i++) {
            string expression;
            CustomGBForce::ComputationType type;
            force.getEnergyTermParameters(i, expression, type);
            if (type != CustomGBForce::SingleParticle)
                continue;
2252
2253
            Lepton::ParsedExpression parsed = Lepton::Parser::parse(expression, functions).optimize();
            energyExpressions["/*"+intToString(i+1)+"*/ energy += "] = parsed;
2254
2255
            for (int j = 0; j < force.getNumComputedValues(); j++)
                energyExpressions["/*"+intToString(i+1)+"*/ deriv"+energyDerivs->getParameterSuffix(j)+" += "] = energyDerivExpressions[i][j];
2256
2257
2258
2259
2260
2261
2262
2263
2264
            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;
2265
2266
2267
2268
2269
        }
        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";
2270
        }
2271
        compute << "forceBuffers[index] = forceBuffers[index]+force;\n";
2272
2273
        map<string, string> replacements;
        replacements["PARAMETER_ARGUMENTS"] = extraArgs.str()+tableArgs.str();
2274
2275
        replacements["REDUCE_DERIVATIVES"] = reduce.str();
        replacements["COMPUTE_ENERGY"] = compute.str();
2276
2277
        map<string, string> defines;
        defines["NUM_ATOMS"] = intToString(cl.getNumAtoms());
2278
        cl::Program program = cl.createProgram(cl.replaceStrings(OpenCLKernelSources::customGBEnergyPerParticle, replacements), defines);
2279
        perParticleEnergyKernel = cl::Kernel(program, "computePerParticleEnergy");
2280
    }
Peter Eastman's avatar
Peter Eastman committed
2281
2282
2283
2284
2285
2286
2287
2288
2289
2290
2291
2292
2293
2294
2295
2296
2297
2298
2299
2300
2301
2302
2303
2304
2305
2306
2307
2308
2309
2310
2311
2312
2313
    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++) {
2314
2315
2316
2317
2318
2319
2320
2321
2322
2323
2324
2325
            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
2326
            if (!isZeroExpression(valueGradientExpressions[i][0]))
2327
                gradientExpressions["dV"+is+"dR.x += "] = valueGradientExpressions[i][0];
Peter Eastman's avatar
Peter Eastman committed
2328
            if (!isZeroExpression(valueGradientExpressions[i][1]))
2329
                gradientExpressions["dV"+is+"dR.y += "] = valueGradientExpressions[i][1];
Peter Eastman's avatar
Peter Eastman committed
2330
            if (!isZeroExpression(valueGradientExpressions[i][2]))
2331
2332
2333
2334
2335
2336
                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
2337
2338
2339
2340
2341
2342
2343
2344
2345
        }
        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");
    }
2346
    {
Peter Eastman's avatar
Peter Eastman committed
2347
        // Create the code to calculate chain rules terms as part of the default nonbonded kernel.
2348
2349
2350
2351
2352
2353
2354

        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;
        }
2355
2356
2357
        map<string, string> variables = globalVariables;
        map<string, string> rename;
        variables["r"] = "r";
2358
2359
        for (int i = 0; i < force.getNumPerParticleParameters(); i++) {
            const string& name = force.getPerParticleParameterName(i);
2360
2361
2362
2363
            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";
2364
2365
2366
2367
        }
        map<string, Lepton::ParsedExpression> derivExpressions;
        stringstream chainSource;
        Lepton::ParsedExpression dVdR = Lepton::Parser::parse(computedValueExpressions[0], functions).differentiate("r").optimize();
2368
2369
        derivExpressions["float dV0dR1 = "] = dVdR;
        derivExpressions["float dV0dR2 = "] = dVdR.renameVariables(rename);
2370
        chainSource << OpenCLExpressionUtilities::createExpressions(derivExpressions, variables, functionDefinitions, prefix+"temp0_", prefix+"functionParams");
Peter Eastman's avatar
Peter Eastman committed
2371
2372
2373
2374
2375
2376
        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";
2377
2378
2379
        variables = globalVariables;
        map<string, string> rename1;
        map<string, string> rename2;
Peter Eastman's avatar
Peter Eastman committed
2380
2381
2382
2383
2384
2385
2386
2387
2388
2389
2390
2391
        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";
2392
2393
        for (int i = 0; i < force.getNumPerParticleParameters(); i++) {
            const string& name = force.getPerParticleParameterName(i);
2394
2395
2396
2397
            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";
2398
2399
2400
        }
        for (int i = 0; i < force.getNumComputedValues(); i++) {
            const string& name = computedValueNames[i];
2401
2402
2403
2404
            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";
2405
2406
            if (i == 0)
                continue;
2407
            string is = intToString(i);
Peter Eastman's avatar
Peter Eastman committed
2408
2409
2410
2411
2412
            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();
2413
                derivExpressions.clear();
Peter Eastman's avatar
Peter Eastman committed
2414
2415
2416
                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");
2417
            }
Peter Eastman's avatar
Peter Eastman committed
2418
2419
            chainSource << "tempForce -= dV"<< is << "dR1*" << prefix << "dEdV" << energyDerivs->getParameterSuffix(i, "1") << ";\n";
            chainSource << "tempForce -= dV"<< is << "dR2*" << prefix << "dEdV" << energyDerivs->getParameterSuffix(i, "2") << ";\n";
2420
2421
2422
        }
        map<string, string> replacements;
        replacements["COMPUTE_FORCE"] = chainSource.str();
2423
        string source = cl.replaceStrings(OpenCLKernelSources::customGBChainRule, replacements);
2424
2425
        vector<OpenCLNonbondedUtilities::ParameterInfo> parameters;
        vector<OpenCLNonbondedUtilities::ParameterInfo> arguments;
2426
2427
2428
        for (int i = 0; i < (int) params->getBuffers().size(); i++) {
            const OpenCLNonbondedUtilities::ParameterInfo& buffer = params->getBuffers()[i];
            string paramName = prefix+"params"+intToString(i+1);
2429
            parameters.push_back(OpenCLNonbondedUtilities::ParameterInfo(paramName, buffer.getComponentType(), buffer.getNumComponents(), buffer.getSize(), buffer.getMemory()));
2430
2431
2432
2433
        }
        for (int i = 0; i < (int) computedValues->getBuffers().size(); i++) {
            const OpenCLNonbondedUtilities::ParameterInfo& buffer = computedValues->getBuffers()[i];
            string paramName = prefix+"values"+intToString(i+1);
2434
            parameters.push_back(OpenCLNonbondedUtilities::ParameterInfo(paramName, buffer.getComponentType(), buffer.getNumComponents(), buffer.getSize(), buffer.getMemory()));
2435
2436
2437
2438
        }
        for (int i = 0; i < (int) energyDerivs->getBuffers().size(); i++) {
            const OpenCLNonbondedUtilities::ParameterInfo& buffer = energyDerivs->getBuffers()[i];
            string paramName = prefix+"dEdV"+intToString(i+1);
2439
            parameters.push_back(OpenCLNonbondedUtilities::ParameterInfo(paramName, buffer.getComponentType(), buffer.getNumComponents(), buffer.getSize(), buffer.getMemory()));
2440
2441
2442
        }
        if (globals != NULL) {
            globals->upload(globalParamValues);
2443
2444
            arguments.push_back(OpenCLNonbondedUtilities::ParameterInfo(prefix+"globals", "float", 1, sizeof(cl_float), globals->getDeviceBuffer()));
        }
Peter Eastman's avatar
Peter Eastman committed
2445
2446
2447
2448
2449
        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]);
2450
2451
    }
    cl.addForce(new OpenCLCustomGBForceInfo(cl.getNonbondedUtilities().getNumForceBuffers(), force));
2452
2453
2454
2455
    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));
    }
2456
2457
}

2458
double OpenCLCalcCustomGBForceKernel::execute(ContextImpl& context, bool includeForces, bool includeEnergy) {
2459
    bool deviceIsCpu = (cl.getDevice().getInfo<CL_DEVICE_TYPE>() == CL_DEVICE_TYPE_CPU);
2460
2461
2462
    OpenCLNonbondedUtilities& nb = cl.getNonbondedUtilities();
    if (!hasInitializedKernels) {
        hasInitializedKernels = true;
2463
        maxTiles = (nb.getUseCutoff() ? nb.getInteractingTiles().getSize() : 0);
2464
        valueBuffers = new OpenCLArray<cl_float>(cl, cl.getPaddedNumAtoms()*cl.getNumForceBuffers(), "customGBValueBuffers");
2465
2466
        cl.addAutoclearBuffer(valueBuffers->getDeviceBuffer(), valueBuffers->getSize());
        cl.clearBuffer(*valueBuffers);
2467
2468
2469
        int index = 0;
        pairValueKernel.setArg<cl::Buffer>(index++, cl.getPosq().getDeviceBuffer());
        pairValueKernel.setArg(index++, OpenCLContext::ThreadBlockSize*sizeof(cl_float4), NULL);
2470
2471
        pairValueKernel.setArg<cl::Buffer>(index++, cl.getNonbondedUtilities().getExclusions().getDeviceBuffer());
        pairValueKernel.setArg<cl::Buffer>(index++, cl.getNonbondedUtilities().getExclusionIndices().getDeviceBuffer());
2472
        pairValueKernel.setArg<cl::Buffer>(index++, cl.getNonbondedUtilities().getExclusionRowIndices().getDeviceBuffer());
2473
        pairValueKernel.setArg<cl::Buffer>(index++, valueBuffers->getDeviceBuffer());
2474
2475
        pairValueKernel.setArg(index++, (deviceIsCpu ? OpenCLContext::TileSize : OpenCLContext::ThreadBlockSize)*sizeof(cl_float), NULL);
        pairValueKernel.setArg(index++, (deviceIsCpu ? OpenCLContext::TileSize : OpenCLContext::ThreadBlockSize)*sizeof(cl_float), NULL);
2476
2477
2478
        if (nb.getUseCutoff()) {
            pairValueKernel.setArg<cl::Buffer>(index++, nb.getInteractingTiles().getDeviceBuffer());
            pairValueKernel.setArg<cl::Buffer>(index++, nb.getInteractionCount().getDeviceBuffer());
2479
            index += 2; // Periodic box size arguments are set when the kernel is executed.
2480
            pairValueKernel.setArg<cl_uint>(index++, maxTiles);
2481
            if (cl.getSIMDWidth() == 32 || deviceIsCpu)
2482
                pairValueKernel.setArg<cl::Buffer>(index++, nb.getInteractionFlags().getDeviceBuffer());
2483
        }
2484
2485
        else
            pairValueKernel.setArg<cl_uint>(index++, cl.getNumAtomBlocks()*(cl.getNumAtomBlocks()+1)/2);
2486
2487
2488
2489
        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];
2490
            pairValueKernel.setArg<cl::Memory>(index++, buffer.getMemory());
2491
2492
            pairValueKernel.setArg(index++, OpenCLContext::ThreadBlockSize*buffer.getSize(), NULL);
        }
2493
2494
2495
2496
2497
        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());
        }
2498
        index = 0;
2499
2500
2501
        perParticleValueKernel.setArg<cl_int>(index++, cl.getPaddedNumAtoms());
        perParticleValueKernel.setArg<cl_int>(index++, nb.getNumForceBuffers());
        perParticleValueKernel.setArg<cl::Buffer>(index++, valueBuffers->getDeviceBuffer());
2502
        perParticleValueKernel.setArg<cl::Buffer>(index++, cl.getPosq().getDeviceBuffer());
2503
        if (globals != NULL)
2504
            perParticleValueKernel.setArg<cl::Buffer>(index++, globals->getDeviceBuffer());
2505
        for (int i = 0; i < (int) params->getBuffers().size(); i++)
2506
            perParticleValueKernel.setArg<cl::Memory>(index++, params->getBuffers()[i].getMemory());
2507
        for (int i = 0; i < (int) computedValues->getBuffers().size(); i++)
2508
            perParticleValueKernel.setArg<cl::Memory>(index++, computedValues->getBuffers()[i].getMemory());
2509
2510
2511
2512
2513
2514
2515
2516
        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;
        pairEnergyKernel.setArg<cl::Buffer>(index++, cl.getForceBuffers().getDeviceBuffer());
        pairEnergyKernel.setArg<cl::Buffer>(index++, cl.getEnergyBuffer().getDeviceBuffer());
2517
        pairEnergyKernel.setArg(index++, (deviceIsCpu ? OpenCLContext::TileSize : OpenCLContext::ThreadBlockSize)*sizeof(cl_float4), NULL);
2518
        pairEnergyKernel.setArg<cl::Buffer>(index++, cl.getPosq().getDeviceBuffer());
2519
        pairEnergyKernel.setArg(index++, (deviceIsCpu ? OpenCLContext::TileSize : OpenCLContext::ThreadBlockSize)*sizeof(cl_float4), NULL);
2520
2521
        pairEnergyKernel.setArg<cl::Buffer>(index++, cl.getNonbondedUtilities().getExclusions().getDeviceBuffer());
        pairEnergyKernel.setArg<cl::Buffer>(index++, cl.getNonbondedUtilities().getExclusionIndices().getDeviceBuffer());
2522
        pairEnergyKernel.setArg<cl::Buffer>(index++, cl.getNonbondedUtilities().getExclusionRowIndices().getDeviceBuffer());
2523
        pairEnergyKernel.setArg(index++, (deviceIsCpu ? OpenCLContext::TileSize : OpenCLContext::ThreadBlockSize)*sizeof(cl_float4), NULL);
2524
2525
2526
        if (nb.getUseCutoff()) {
            pairEnergyKernel.setArg<cl::Buffer>(index++, nb.getInteractingTiles().getDeviceBuffer());
            pairEnergyKernel.setArg<cl::Buffer>(index++, nb.getInteractionCount().getDeviceBuffer());
2527
            index += 2; // Periodic box size arguments are set when the kernel is executed.
2528
            pairEnergyKernel.setArg<cl_uint>(index++, maxTiles);
2529
            if (cl.getSIMDWidth() == 32 || deviceIsCpu)
2530
                pairEnergyKernel.setArg<cl::Buffer>(index++, nb.getInteractionFlags().getDeviceBuffer());
2531
        }
2532
2533
        else
            pairEnergyKernel.setArg<cl_uint>(index++, cl.getNumAtomBlocks()*(cl.getNumAtomBlocks()+1)/2);
2534
2535
2536
2537
        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];
2538
            pairEnergyKernel.setArg<cl::Memory>(index++, buffer.getMemory());
2539
            pairEnergyKernel.setArg(index++, (deviceIsCpu ? OpenCLContext::TileSize : OpenCLContext::ThreadBlockSize)*buffer.getSize(), NULL);
2540
2541
2542
        }
        for (int i = 0; i < (int) computedValues->getBuffers().size(); i++) {
            const OpenCLNonbondedUtilities::ParameterInfo& buffer = computedValues->getBuffers()[i];
2543
            pairEnergyKernel.setArg<cl::Memory>(index++, buffer.getMemory());
2544
            pairEnergyKernel.setArg(index++, (deviceIsCpu ? OpenCLContext::TileSize : OpenCLContext::ThreadBlockSize)*buffer.getSize(), NULL);
2545
        }
2546
2547
        for (int i = 0; i < (int) energyDerivs->getBuffers().size(); i++) {
            const OpenCLNonbondedUtilities::ParameterInfo& buffer = energyDerivs->getBuffers()[i];
2548
            pairEnergyKernel.setArg<cl::Memory>(index++, buffer.getMemory());
2549
            pairEnergyKernel.setArg(index++, (deviceIsCpu ? OpenCLContext::TileSize : OpenCLContext::ThreadBlockSize)*buffer.getSize(), NULL);
2550
        }
2551
2552
2553
2554
2555
2556
2557
2558
        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());
2559
        perParticleEnergyKernel.setArg<cl::Buffer>(index++, cl.getForceBuffers().getDeviceBuffer());
2560
        perParticleEnergyKernel.setArg<cl::Buffer>(index++, cl.getEnergyBuffer().getDeviceBuffer());
2561
        perParticleEnergyKernel.setArg<cl::Buffer>(index++, cl.getPosq().getDeviceBuffer());
2562
2563
2564
        if (globals != NULL)
            perParticleEnergyKernel.setArg<cl::Buffer>(index++, globals->getDeviceBuffer());
        for (int i = 0; i < (int) params->getBuffers().size(); i++)
2565
            perParticleEnergyKernel.setArg<cl::Memory>(index++, params->getBuffers()[i].getMemory());
2566
        for (int i = 0; i < (int) computedValues->getBuffers().size(); i++)
2567
            perParticleEnergyKernel.setArg<cl::Memory>(index++, computedValues->getBuffers()[i].getMemory());
2568
        for (int i = 0; i < (int) energyDerivs->getBuffers().size(); i++)
2569
            perParticleEnergyKernel.setArg<cl::Memory>(index++, energyDerivs->getBuffers()[i].getMemory());
2570
2571
2572
2573
2574
        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
2575
2576
2577
2578
2579
2580
2581
2582
2583
2584
2585
2586
2587
        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());
        }
2588
2589
2590
    }
    if (globals != NULL) {
        bool changed = false;
2591
        for (int i = 0; i < (int) globalParamNames.size(); i++) {
2592
2593
2594
2595
2596
2597
2598
2599
            cl_float value = (cl_float) context.getParameter(globalParamNames[i]);
            if (value != globalParamValues[i])
                changed = true;
            globalParamValues[i] = value;
        }
        if (changed)
            globals->upload(globalParamValues);
    }
2600
    if (nb.getUseCutoff()) {
2601
2602
2603
2604
        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());
2605
2606
2607
2608
2609
        if (maxTiles < nb.getInteractingTiles().getSize()) {
            maxTiles = nb.getInteractingTiles().getSize();
            pairValueKernel.setArg<cl_uint>(12, maxTiles);
            pairEnergyKernel.setArg<cl_uint>(13, maxTiles);
        }
2610
    }
2611
    int numTiles = cl.getNumAtomBlocks()*(cl.getNumAtomBlocks()+1)/2;
2612
    cl.executeKernel(pairValueKernel, numTiles*OpenCLContext::TileSize, deviceIsCpu ? 1 : -1);
2613
    cl.executeKernel(perParticleValueKernel, cl.getPaddedNumAtoms());
2614
    cl.executeKernel(pairEnergyKernel, numTiles*OpenCLContext::TileSize, deviceIsCpu ? 1 : -1);
2615
    cl.executeKernel(perParticleEnergyKernel, cl.getPaddedNumAtoms());
Peter Eastman's avatar
Peter Eastman committed
2616
2617
    if (needParameterGradient)
        cl.executeKernel(gradientChainRuleKernel, cl.getPaddedNumAtoms());
2618
2619
2620
    return 0.0;
}

2621
2622
2623
2624
2625
2626
2627
2628
2629
2630
2631
2632
2633
2634
2635
2636
2637
2638
2639
2640
2641
2642
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);
2643
        for (int i = 0; i < (int) params1.size(); i++)
2644
2645
2646
2647
2648
2649
2650
2651
2652
2653
2654
2655
2656
2657
2658
2659
2660
2661
2662
2663
            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) {
    numParticles = force.getNumParticles();
2664
    params = new OpenCLParameterSet(cl, force.getNumPerParticleParameters(), numParticles, "customExternalParams");
2665
2666
2667
2668
2669
2670
    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";
    }
2671
    vector<vector<cl_float> > paramVector(numParticles);
2672
2673
2674
2675
    vector<cl_int> indicesVector(numParticles);
    for (int i = 0; i < numParticles; i++) {
        vector<double> parameters;
        force.getParticleParameters(i, indicesVector[i], parameters);
2676
        paramVector[i].resize(parameters.size());
2677
        for (int j = 0; j < (int) parameters.size(); j++)
2678
            paramVector[i][j] = (cl_float) parameters[j];
2679
    }
2680
    params->setParameterValues(paramVector);
2681
2682
2683
2684
2685
2686
2687
2688
2689
2690
2691
2692
2693
2694
2695
2696
2697
2698
2699
2700
2701
2702
2703
2704
2705
2706
2707
2708
2709
2710
2711
    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);
2712
        variables[name] = "particleParams"+params->getParameterSuffix(i);
2713
2714
2715
2716
2717
2718
2719
    }
    for (int i = 0; i < force.getNumGlobalParameters(); i++) {
        const string& name = force.getGlobalParameterName(i);
        string value = "globals["+intToString(i)+"]";
        variables[name] = value;
    }
    stringstream compute;
2720
2721
2722
2723
2724
    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";
    }
2725
2726
    vector<pair<string, string> > functions;
    compute << OpenCLExpressionUtilities::createExpressions(expressions, variables, functions, "temp", "");
2727
    map<string, string> replacements;
2728
2729
    replacements["COMPUTE_FORCE"] = compute.str();
    replacements["EXTRA_ARGUMENTS"] = extraArguments;
2730
    cl::Program program = cl.createProgram(cl.replaceStrings(OpenCLKernelSources::customExternalForce, replacements));
2731
2732
2733
    kernel = cl::Kernel(program, "computeCustomExternalForces");
}

2734
double OpenCLCalcCustomExternalForceKernel::execute(ContextImpl& context, bool includeForces, bool includeEnergy) {
2735
2736
    if (globals != NULL) {
        bool changed = false;
2737
        for (int i = 0; i < (int) globalParamNames.size(); i++) {
2738
2739
2740
2741
2742
2743
2744
2745
2746
2747
2748
2749
2750
2751
            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());
2752
2753
        kernel.setArg<cl::Buffer>(4, indices->getDeviceBuffer());
        int nextIndex = 5;
2754
        if (globals != NULL)
2755
2756
2757
            kernel.setArg<cl::Buffer>(nextIndex++, globals->getDeviceBuffer());
        for (int i = 0; i < (int) params->getBuffers().size(); i++) {
            const OpenCLNonbondedUtilities::ParameterInfo& buffer = params->getBuffers()[i];
2758
            kernel.setArg<cl::Memory>(nextIndex++, buffer.getMemory());
2759
        }
2760
2761
2762
    }
    cl.executeKernel(kernel, numParticles);
    return 0.0;
2763
}
2764

2765
2766
2767
2768
2769
2770
2771
2772
2773
2774
2775
2776
2777
2778
2779
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);
2780
2781
2782
2783
2784
2785
            particles.clear();
            particles.push_back(p1);
            if (p2 > -1)
                particles.push_back(p2);
            if (p3 > -1)
                particles.push_back(p3);
2786
2787
2788
2789
2790
            return;
        }
        index -= force.getNumDonors();
        if (index < force.getNumAcceptors()) {
            force.getAcceptorParameters(index, p1, p2, p3, parameters);
2791
2792
2793
2794
2795
2796
            particles.clear();
            particles.push_back(p1);
            if (p2 > -1)
                particles.push_back(p2);
            if (p3 > -1)
                particles.push_back(p3);
2797
2798
2799
2800
2801
            return;
        }
        index -= force.getNumAcceptors();
        int donor, acceptor;
        force.getExclusionParticles(index, donor, acceptor);
2802
        particles.clear();
2803
        force.getDonorParameters(donor, p1, p2, p3, parameters);
2804
2805
2806
2807
2808
        particles.push_back(p1);
        if (p2 > -1)
            particles.push_back(p2);
        if (p3 > -1)
            particles.push_back(p3);
2809
        force.getAcceptorParameters(acceptor, p1, p2, p3, parameters);
2810
2811
2812
2813
2814
        particles.push_back(p1);
        if (p2 > -1)
            particles.push_back(p2);
        if (p3 > -1)
            particles.push_back(p3);
2815
2816
2817
2818
2819
2820
2821
2822
2823
2824
2825
2826
2827
2828
2829
2830
2831
2832
2833
2834
2835
2836
2837
2838
2839
2840
2841
2842
2843
2844
2845
2846
2847
2848
2849
2850
2851
2852
2853
2854
2855
    }
    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;
2856
2857
2858
2859
    if (donorExclusions != NULL)
        delete donorExclusions;
    if (acceptorExclusions != NULL)
        delete acceptorExclusions;
2860
2861
2862
2863
2864
2865
    if (tabulatedFunctionParams != NULL)
        delete tabulatedFunctionParams;
    for (int i = 0; i < (int) tabulatedFunctions.size(); i++)
        delete tabulatedFunctions[i];
}

2866
2867
2868
2869
2870
2871
2872
2873
2874
2875
2876
2877
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";
}
2878

2879
void OpenCLCalcCustomHbondForceKernel::initialize(const System& system, const CustomHbondForce& force) {
2880
2881
2882
2883
2884
2885
2886
2887
2888
2889
2890
2891
2892
2893
2894
2895
2896
2897
2898
2899
2900
2901
2902
2903
2904
2905
2906
2907
2908
2909
2910
2911
2912
2913
    // Record the lists of donors and acceptors, and the parameters for each one.

    numDonors = force.getNumDonors();
    numAcceptors = force.getNumAcceptors();
    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;
        force.getDonorParameters(i, donorVector[i].x, donorVector[i].y, donorVector[i].z, parameters);
        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);

2914
    // Select an output buffer index for each donor and acceptor.
2915
2916
2917
2918
2919

    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);
2920
    vector<int> donorBufferCounter(numParticles, 0);
2921
    for (int i = 0; i < numDonors; i++)
2922
2923
2924
        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);
2925
    vector<int> acceptorBufferCounter(numParticles, 0);
2926
    for (int i = 0; i < numAcceptors; i++)
2927
2928
2929
        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);
2930
2931
    donorBufferIndices->upload(donorBufferVector);
    acceptorBufferIndices->upload(acceptorBufferVector);
2932
2933
2934
2935
2936
2937
    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));
2938
2939
2940

    // Record exclusions.

2941
2942
    vector<mm_int4> donorExclusionVector(numDonors, mm_int4(-1, -1, -1, -1));
    vector<mm_int4> acceptorExclusionVector(numAcceptors, mm_int4(-1, -1, -1, -1));
2943
2944
2945
    for (int i = 0; i < force.getNumExclusions(); i++) {
        int donor, acceptor;
        force.getExclusionParticles(i, donor, acceptor);
2946
2947
2948
2949
2950
2951
2952
2953
2954
2955
2956
2957
2958
2959
2960
2961
2962
2963
2964
2965
        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");
2966
    }
2967
2968
2969
2970
    donorExclusions = new OpenCLArray<mm_int4>(cl, numDonors, "customHbondDonorExclusions");
    acceptorExclusions = new OpenCLArray<mm_int4>(cl, numDonors, "customHbondAcceptorExclusions");
    donorExclusions->upload(donorExclusionVector);
    acceptorExclusions->upload(acceptorExclusionVector);
2971
2972
2973
2974
2975
2976
2977
2978
2979
2980
2981
2982

    // 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;
2983
        force.getFunctionParameters(i, name, values, min, max);
2984
        string arrayName = "table"+intToString(i);
2985
2986
        functionDefinitions.push_back(make_pair(name, arrayName));
        functions[name] = &fp;
2987
        tabulatedFunctionParamsVec[i] = mm_float4((float) min, (float) max, (float) ((values.size()-1)/(max-min)), values.size()-2);
2988
        vector<mm_float4> f = OpenCLExpressionUtilities::computeFunctionCoefficients(values, min, max);
2989
2990
2991
2992
2993
2994
2995
        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);
2996
        tableArgs << ", __global float4* functionParams";
2997
2998
    }

2999
    // Record information about parameters.
3000
3001
3002
3003
3004
3005
3006
3007
3008
3009
3010
3011
3012
3013
3014
3015
3016
3017
3018
3019
3020
3021

    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)+"]";
    }
3022
3023
3024
3025
3026
3027
3028
3029
3030
3031
3032
3033
3034
3035
3036
3037
3038
3039
3040
3041
3042
3043
3044
3045
3046
3047
3048
3049
3050
3051
3052
3053
3054
3055
3056
3057
3058
3059
3060
3061
3062
3063
3064
3065
3066
3067
3068
3069
3070
3071
3072
3073
3074
3075
3076
3077
3078
3079
3080
3081
3082
3083
3084
3085
3086
3087
3088
3089
3090
3091
3092
3093
3094
3095
3096

    // 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.

3097
    if (force.getNumGlobalParameters() > 0)
3098
        extraArgs << ", __global float* globals";
3099
3100
3101
    for (int i = 0; i < (int) donorParams->getBuffers().size(); i++) {
        const OpenCLNonbondedUtilities::ParameterInfo& buffer = donorParams->getBuffers()[i];
        extraArgs << ", __global "+buffer.getType()+"* donor"+buffer.getName();
3102
        addDonorAndAcceptorCode(computeDonor, computeAcceptor, buffer.getType()+" donorParams"+intToString(i+1)+" = donor"+buffer.getName()+"[index];\n");
3103
3104
3105
3106
    }
    for (int i = 0; i < (int) acceptorParams->getBuffers().size(); i++) {
        const OpenCLNonbondedUtilities::ParameterInfo& buffer = acceptorParams->getBuffers()[i];
        extraArgs << ", __global "+buffer.getType()+"* acceptor"+buffer.getName();
3107
        addDonorAndAcceptorCode(computeDonor, computeAcceptor, buffer.getType()+" acceptorParams"+intToString(i+1)+" = acceptor"+buffer.getName()+"[index];\n");
3108
    }
3109
3110
3111
3112
3113
3114
3115
3116
3117
3118
3119
3120
3121
3122
3123
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

    // 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.

3169
    map<string, string> replacements;
3170
3171
    replacements["COMPUTE_DONOR_FORCE"] = computeDonor.str();
    replacements["COMPUTE_ACCEPTOR_FORCE"] = computeAcceptor.str();
3172
3173
3174
3175
3176
3177
    replacements["PARAMETER_ARGUMENTS"] = extraArgs.str()+tableArgs.str();
    map<string, string> defines;
    defines["PADDED_NUM_ATOMS"] = intToString(cl.getPaddedNumAtoms());
    defines["NUM_DONORS"] = intToString(force.getNumDonors());
    defines["NUM_ACCEPTORS"] = intToString(force.getNumAcceptors());
    defines["M_PI"] = doubleToString(M_PI);
3178
3179
3180
3181
3182
3183
    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";
3184
3185
    if (force.getNumExclusions() > 0)
        defines["USE_EXCLUSIONS"] = "1";
3186
    cl::Program program = cl.createProgram(cl.replaceStrings(OpenCLKernelSources::customHbondForce, replacements), defines);
3187
3188
    donorKernel = cl::Kernel(program, "computeDonorForces");
    acceptorKernel = cl::Kernel(program, "computeAcceptorForces");
3189
3190
}

3191
double OpenCLCalcCustomHbondForceKernel::execute(ContextImpl& context, bool includeForces, bool includeEnergy) {
3192
3193
3194
3195
3196
3197
3198
3199
3200
3201
3202
3203
3204
3205
    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;
3206
3207
3208
        donorKernel.setArg<cl::Buffer>(index++, cl.getForceBuffers().getDeviceBuffer());
        donorKernel.setArg<cl::Buffer>(index++, cl.getEnergyBuffer().getDeviceBuffer());
        donorKernel.setArg<cl::Buffer>(index++, cl.getPosq().getDeviceBuffer());
3209
        donorKernel.setArg<cl::Buffer>(index++, donorExclusions->getDeviceBuffer());
3210
3211
3212
3213
        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);
3214
        index += 2; // Periodic box size arguments are set when the kernel is executed.
3215
        if (globals != NULL)
3216
            donorKernel.setArg<cl::Buffer>(index++, globals->getDeviceBuffer());
3217
3218
        for (int i = 0; i < (int) donorParams->getBuffers().size(); i++) {
            const OpenCLNonbondedUtilities::ParameterInfo& buffer = donorParams->getBuffers()[i];
3219
            donorKernel.setArg<cl::Memory>(index++, buffer.getMemory());
3220
3221
3222
        }
        for (int i = 0; i < (int) acceptorParams->getBuffers().size(); i++) {
            const OpenCLNonbondedUtilities::ParameterInfo& buffer = acceptorParams->getBuffers()[i];
3223
            donorKernel.setArg<cl::Memory>(index++, buffer.getMemory());
3224
3225
3226
3227
3228
3229
3230
3231
3232
3233
        }
        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());
3234
        acceptorKernel.setArg<cl::Buffer>(index++, acceptorExclusions->getDeviceBuffer());
3235
3236
3237
3238
        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);
3239
        index += 2; // Periodic box size arguments are set when the kernel is executed.
3240
3241
3242
3243
        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];
3244
            acceptorKernel.setArg<cl::Memory>(index++, buffer.getMemory());
3245
3246
3247
        }
        for (int i = 0; i < (int) acceptorParams->getBuffers().size(); i++) {
            const OpenCLNonbondedUtilities::ParameterInfo& buffer = acceptorParams->getBuffers()[i];
3248
            acceptorKernel.setArg<cl::Memory>(index++, buffer.getMemory());
3249
3250
3251
3252
3253
        }
        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());
3254
3255
        }
    }
3256
3257
    donorKernel.setArg<mm_float4>(8, cl.getPeriodicBoxSize());
    donorKernel.setArg<mm_float4>(9, cl.getInvPeriodicBoxSize());
3258
    cl.executeKernel(donorKernel, std::max(numDonors, numAcceptors));
3259
3260
    acceptorKernel.setArg<mm_float4>(8, cl.getPeriodicBoxSize());
    acceptorKernel.setArg<mm_float4>(9, cl.getInvPeriodicBoxSize());
3261
    cl.executeKernel(acceptorKernel, std::max(numDonors, numAcceptors));
3262
3263
3264
    return 0.0;
}

3265
3266
3267
3268
OpenCLIntegrateVerletStepKernel::~OpenCLIntegrateVerletStepKernel() {
}

void OpenCLIntegrateVerletStepKernel::initialize(const System& system, const VerletIntegrator& integrator) {
3269
    cl.getPlatformData().initializeContexts(system);
3270
    cl::Program program = cl.createProgram(OpenCLKernelSources::verlet, "");
3271
3272
    kernel1 = cl::Kernel(program, "integrateVerletPart1");
    kernel2 = cl::Kernel(program, "integrateVerletPart2");
3273
    prevStepSize = -1.0;
3274
3275
3276
}

void OpenCLIntegrateVerletStepKernel::execute(ContextImpl& context, const VerletIntegrator& integrator) {
3277
    OpenCLIntegrationUtilities& integration = cl.getIntegrationUtilities();
3278
3279
    int numAtoms = cl.getNumAtoms();
    double dt = integrator.getStepSize();
3280
3281
3282
    if (!hasInitializedKernels) {
        hasInitializedKernels = true;
        kernel1.setArg<cl_int>(0, numAtoms);
3283
        kernel1.setArg<cl::Buffer>(1, cl.getIntegrationUtilities().getStepSize().getDeviceBuffer());
3284
3285
3286
3287
3288
        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);
3289
        kernel2.setArg<cl::Buffer>(1, cl.getIntegrationUtilities().getStepSize().getDeviceBuffer());
3290
3291
3292
3293
        kernel2.setArg<cl::Buffer>(2, cl.getPosq().getDeviceBuffer());
        kernel2.setArg<cl::Buffer>(3, cl.getVelm().getDeviceBuffer());
        kernel2.setArg<cl::Buffer>(4, integration.getPosDelta().getDeviceBuffer());
    }
3294
3295
    if (dt != prevStepSize) {
        vector<mm_float2> stepSizeVec(1);
3296
        stepSizeVec[0] = mm_float2((cl_float) dt, (cl_float) dt);
3297
        cl.getIntegrationUtilities().getStepSize().upload(stepSizeVec);
3298
3299
        prevStepSize = dt;
    }
3300
3301
3302
3303
3304
3305
3306

    // Call the first integration kernel.

    cl.executeKernel(kernel1, numAtoms);

    // Apply constraints.

3307
    integration.applyConstraints(integrator.getConstraintTolerance());
3308
3309
3310
3311
3312
3313
3314
3315
3316

    // 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);
3317
3318
}

3319
3320
3321
3322
3323
3324
OpenCLIntegrateLangevinStepKernel::~OpenCLIntegrateLangevinStepKernel() {
    if (params != NULL)
        delete params;
}

void OpenCLIntegrateLangevinStepKernel::initialize(const System& system, const LangevinIntegrator& integrator) {
3325
    cl.getPlatformData().initializeContexts(system);
3326
3327
3328
3329
    cl.getIntegrationUtilities().initRandomNumberGenerator(integrator.getRandomNumberSeed());
    map<string, string> defines;
    defines["NUM_ATOMS"] = intToString(cl.getNumAtoms());
    defines["PADDED_NUM_ATOMS"] = intToString(cl.getPaddedNumAtoms());
3330
    cl::Program program = cl.createProgram(OpenCLKernelSources::langevin, defines, "");
3331
3332
    kernel1 = cl::Kernel(program, "integrateLangevinPart1");
    kernel2 = cl::Kernel(program, "integrateLangevinPart2");
3333
    params = new OpenCLArray<cl_float>(cl, 3, "langevinParams");
3334
3335
3336
3337
    prevStepSize = -1.0;
}

void OpenCLIntegrateLangevinStepKernel::execute(ContextImpl& context, const LangevinIntegrator& integrator) {
3338
    OpenCLIntegrationUtilities& integration = cl.getIntegrationUtilities();
3339
    int numAtoms = cl.getNumAtoms();
3340
3341
    if (!hasInitializedKernels) {
        hasInitializedKernels = true;
3342
3343
3344
3345
        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());
3346
3347
3348
        kernel1.setArg<cl::Buffer>(4, integration.getStepSize().getDeviceBuffer());
        kernel1.setArg<cl::Buffer>(5, integration.getRandom().getDeviceBuffer());
        kernel2.setArg<cl::Buffer>(0, cl.getPosq().getDeviceBuffer());
3349
        kernel2.setArg<cl::Buffer>(1, integration.getPosDelta().getDeviceBuffer());
3350
3351
        kernel2.setArg<cl::Buffer>(2, cl.getVelm().getDeviceBuffer());
        kernel2.setArg<cl::Buffer>(3, integration.getStepSize().getDeviceBuffer());
3352
    }
3353
3354
3355
3356
3357
3358
3359
3360
    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;
3361
3362
3363
        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);
3364
        vector<cl_float> p(params->getSize());
3365
3366
3367
        p[0] = (cl_float) vscale;
        p[1] = (cl_float) fscale;
        p[2] = (cl_float) noisescale;
3368
        params->upload(p);
3369
3370
        integration.getStepSize()[0].y = stepSize;
        integration.getStepSize().upload();
3371
3372
3373
3374
3375
3376
3377
        prevTemp = temperature;
        prevFriction = friction;
        prevStepSize = stepSize;
    }

    // Call the first integration kernel.

3378
    kernel1.setArg<cl_uint>(6, integration.prepareRandomNumbers(cl.getPaddedNumAtoms()));
3379
3380
3381
3382
    cl.executeKernel(kernel1, numAtoms);

    // Apply constraints.

3383
    integration.applyConstraints(integrator.getConstraintTolerance());
3384
3385
3386
3387
3388
3389
3390
3391
3392
3393

    // 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);
}
3394
3395
3396
3397
3398

OpenCLIntegrateBrownianStepKernel::~OpenCLIntegrateBrownianStepKernel() {
}

void OpenCLIntegrateBrownianStepKernel::initialize(const System& system, const BrownianIntegrator& integrator) {
3399
    cl.getPlatformData().initializeContexts(system);
3400
3401
3402
    cl.getIntegrationUtilities().initRandomNumberGenerator(integrator.getRandomNumberSeed());
    map<string, string> defines;
    defines["NUM_ATOMS"] = intToString(cl.getNumAtoms());
3403
    cl::Program program = cl.createProgram(OpenCLKernelSources::brownian, defines, "");
3404
3405
3406
3407
3408
3409
3410
3411
3412
3413
3414
3415
    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());
3416
3417
        kernel1.setArg<cl::Buffer>(4, cl.getVelm().getDeviceBuffer());
        kernel1.setArg<cl::Buffer>(5, integration.getRandom().getDeviceBuffer());
3418
3419
3420
3421
3422
3423
3424
3425
3426
3427
3428
3429
3430
3431
3432
3433
3434
3435
3436
        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.

3437
    kernel1.setArg<cl_uint>(6, integration.prepareRandomNumbers(cl.getPaddedNumAtoms()));
3438
3439
3440
3441
3442
3443
3444
3445
3446
3447
3448
3449
3450
3451
3452
    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);
}
3453
3454
3455
3456
3457

OpenCLIntegrateVariableVerletStepKernel::~OpenCLIntegrateVariableVerletStepKernel() {
}

void OpenCLIntegrateVariableVerletStepKernel::initialize(const System& system, const VariableVerletIntegrator& integrator) {
3458
    cl.getPlatformData().initializeContexts(system);
3459
    cl::Program program = cl.createProgram(OpenCLKernelSources::verlet, "");
3460
3461
3462
3463
3464
3465
3466
    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) {
3467
    OpenCLIntegrationUtilities& integration = cl.getIntegrationUtilities();
3468
3469
3470
3471
    int numAtoms = cl.getNumAtoms();
    if (!hasInitializedKernels) {
        hasInitializedKernels = true;
        kernel1.setArg<cl_int>(0, numAtoms);
3472
        kernel1.setArg<cl::Buffer>(1, cl.getIntegrationUtilities().getStepSize().getDeviceBuffer());
3473
3474
3475
3476
3477
        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);
3478
        kernel2.setArg<cl::Buffer>(1, cl.getIntegrationUtilities().getStepSize().getDeviceBuffer());
3479
3480
3481
3482
        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);
3483
        selectSizeKernel.setArg<cl::Buffer>(3, cl.getIntegrationUtilities().getStepSize().getDeviceBuffer());
3484
3485
3486
3487
3488
3489
3490
3491
3492
        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);
3493
    selectSizeKernel.setArg<cl_float>(2, (cl_float) integrator.getErrorTolerance());
3494
3495
3496
3497
3498
3499
3500
3501
3502
3503
3504
3505
3506
3507
3508
3509
    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.

3510
3511
    cl.getIntegrationUtilities().getStepSize().download();
    double dt = cl.getIntegrationUtilities().getStepSize()[0].y;
3512
3513
3514
3515
3516
3517
3518
3519
3520
3521
3522
3523
3524
    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) {
3525
    cl.getPlatformData().initializeContexts(system);
3526
3527
3528
3529
    cl.getIntegrationUtilities().initRandomNumberGenerator(integrator.getRandomNumberSeed());
    map<string, string> defines;
    defines["NUM_ATOMS"] = intToString(cl.getNumAtoms());
    defines["PADDED_NUM_ATOMS"] = intToString(cl.getPaddedNumAtoms());
3530
    cl::Program program = cl.createProgram(OpenCLKernelSources::langevin, defines, "");
3531
3532
3533
    kernel1 = cl::Kernel(program, "integrateLangevinPart1");
    kernel2 = cl::Kernel(program, "integrateLangevinPart2");
    selectSizeKernel = cl::Kernel(program, "selectLangevinStepSize");
3534
    params = new OpenCLArray<cl_float>(cl, 3, "langevinParams");
3535
3536
3537
3538
3539
3540
    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) {
3541
    OpenCLIntegrationUtilities& integration = cl.getIntegrationUtilities();
3542
3543
3544
    int numAtoms = cl.getNumAtoms();
    if (!hasInitializedKernels) {
        hasInitializedKernels = true;
3545
3546
3547
3548
        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());
3549
3550
3551
        kernel1.setArg<cl::Buffer>(4, integration.getStepSize().getDeviceBuffer());
        kernel1.setArg<cl::Buffer>(5, integration.getRandom().getDeviceBuffer());
        kernel2.setArg<cl::Buffer>(0, cl.getPosq().getDeviceBuffer());
3552
        kernel2.setArg<cl::Buffer>(1, integration.getPosDelta().getDeviceBuffer());
3553
3554
        kernel2.setArg<cl::Buffer>(2, cl.getVelm().getDeviceBuffer());
        kernel2.setArg<cl::Buffer>(3, integration.getStepSize().getDeviceBuffer());
3555
        selectSizeKernel.setArg<cl::Buffer>(4, integration.getStepSize().getDeviceBuffer());
3556
3557
3558
3559
3560
        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);
3561
3562
3563
3564
3565
    }

    // Select the step size to use.

    float maxStepSize = (float)(maxTime-cl.getTime());
3566
3567
3568
3569
    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()));
3570
3571
3572
3573
    cl.executeKernel(selectSizeKernel, blockSize, blockSize);

    // Call the first integration kernel.

3574
    kernel1.setArg<cl_uint>(6, integration.prepareRandomNumbers(cl.getPaddedNumAtoms()));
3575
3576
3577
3578
3579
3580
3581
3582
3583
3584
3585
3586
    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.

3587
3588
    cl.getIntegrationUtilities().getStepSize().download();
    double dt = cl.getIntegrationUtilities().getStepSize()[0].y;
3589
3590
3591
3592
3593
3594
3595
    double time = cl.getTime()+dt;
    if (dt == maxStepSize)
        time = maxTime; // Avoid round-off error
    cl.setTime(time);
    cl.setStepCount(cl.getStepCount()+1);
}

3596
OpenCLApplyAndersenThermostatKernel::~OpenCLApplyAndersenThermostatKernel() {
3597
3598
    if (atomGroups != NULL)
        delete atomGroups;
3599
3600
3601
3602
3603
3604
}

void OpenCLApplyAndersenThermostatKernel::initialize(const System& system, const AndersenThermostat& thermostat) {
    randomSeed = thermostat.getRandomNumberSeed();
    map<string, string> defines;
    defines["NUM_ATOMS"] = intToString(cl.getNumAtoms());
3605
    cl::Program program = cl.createProgram(OpenCLKernelSources::andersenThermostat, defines);
3606
    kernel = cl::Kernel(program, "applyAndersenThermostat");
3607
3608
3609
3610
3611
3612
3613
3614
3615
3616
3617

    // 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);
3618
3619
3620
3621
3622
3623
3624
3625
3626
}

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());
3627
        kernel.setArg<cl::Buffer>(6, atomGroups->getDeviceBuffer());
3628
3629
3630
3631
3632
3633
3634
    }
    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());
}

3635
3636
3637
3638
3639
3640
3641
3642
3643
3644
3645
3646
3647
3648
3649
3650
3651
3652
3653
3654
3655
3656
3657
3658
3659
3660
3661
3662
3663
3664
3665
3666
3667
3668
3669
3670
3671
3672
3673
3674
3675
3676
3677
3678
3679
3680
3681
3682
3683
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());
3684
3685
    for (int i = 0; i < (int) cl.getPosCellOffsets().size(); i++)
        cl.getPosCellOffsets()[i] = mm_int4(0, 0, 0, 0);
3686
3687
3688
3689
3690
3691
}

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

3692
3693
3694
3695
3696
3697
3698
3699
3700
3701
3702
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.

3703
    OpenCLArray<mm_float4>& velm = cl.getVelm();
3704
    velm.download();
3705
    double energy = 0.0;
3706
    for (size_t i = 0; i < masses.size(); ++i) {
3707
3708
        mm_float4 v = velm[i];
        energy += masses[i]*(v.x*v.x+v.y*v.y+v.z*v.z);
3709
    }
3710
3711
    return 0.5*energy;
}
3712
3713
3714
3715
3716
3717
3718
3719
3720
3721
3722
3723
3724
3725
3726

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);
3727
    cl::Program program = cl.createProgram(OpenCLKernelSources::removeCM, defines);
3728
3729
3730
3731
3732
3733
3734
3735
3736
3737
3738
3739
3740
3741
3742
3743
    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());
}