gpu.cpp 118 KB
Newer Older
Peter Eastman's avatar
Peter Eastman committed
1
2
3
4
5
6
7
8
9
10
11
12
/* -------------------------------------------------------------------------- *
 *                                   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.               *
 *                                                                            *
 * Portions copyright (c) 2009 Stanford University and the Authors.           *
 * Authors: Scott Le Grand, Peter Eastman                                     *
 * Contributors:                                                              *
 *                                                                            *
13
14
15
16
 * 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.                                        *
Peter Eastman's avatar
Peter Eastman committed
17
 *                                                                            *
18
19
20
21
 * 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.                        *
Peter Eastman's avatar
Peter Eastman committed
22
 *                                                                            *
23
24
 * 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/>.      *
Peter Eastman's avatar
Peter Eastman committed
25
26
27
 * -------------------------------------------------------------------------- */

#include <stdio.h>
28
#include <string.h>
Peter Eastman's avatar
Peter Eastman committed
29
30
31
32
33
34
35
36
37
#include <cuda.h>
#include <vector_functions.h>
#include <cstdlib>
#include <string>
#include <iostream>
#include <fstream>
#include <sstream>
#include <cmath>
#include <map>
38
#include <set>
39
#include <algorithm>
Peter Eastman's avatar
Peter Eastman committed
40
#ifdef WIN32
41
42
  #define _USE_MATH_DEFINES /* M_PI */
  #include <math.h>
Peter Eastman's avatar
Peter Eastman committed
43
44
45
46
47
48
49
50
  #include <windows.h>
#else
  #include <stdint.h>
#endif
using namespace std;

#include "gputypes.h"
#include "cudaKernels.h"
51
#include "hilbert.h"
52
#include "openmm/OpenMMException.h"
53
#include "quern.h"
54
#include "Lepton.h"
Peter Eastman's avatar
Peter Eastman committed
55

56
57
58
59
// In case we're using some primitive version of Visual Studio this will
// make sure that erf() and erfc() are defined.
#include "openmm/internal/MSVC_erfc.h"

Peter Eastman's avatar
Peter Eastman committed
60
using OpenMM::OpenMMException;
61
using Lepton::Operation;
Peter Eastman's avatar
Peter Eastman committed
62
63
64
65
66

struct ShakeCluster {
    int centralID;
    int peripheralID[3];
    int size;
67
    bool valid;
Peter Eastman's avatar
Peter Eastman committed
68
69
    float distance;
    float centralInvMass, peripheralInvMass;
70
    ShakeCluster() : valid(true) {
Peter Eastman's avatar
Peter Eastman committed
71
    }
72
    ShakeCluster(int centralID, float invMass) : centralID(centralID), centralInvMass(invMass), size(0), valid(true) {
Peter Eastman's avatar
Peter Eastman committed
73
74
    }
    void addAtom(int id, float dist, float invMass) {
75
76
77
78
79
80
81
        if (size == 3 || (size > 0 && dist != distance) || (size > 0 && invMass != peripheralInvMass))
            valid = false;
        else {
            peripheralID[size++] = id;
            distance = dist;
            peripheralInvMass = invMass;
        }
Peter Eastman's avatar
Peter Eastman committed
82
83
84
    }
};

85
86
87
88
89
90
91
92
struct Constraint
{
    Constraint(int atom1, int atom2, float distance2) : atom1(atom1), atom2(atom2), distance2(distance2) {
    }
    int atom1, atom2;
    float distance2;
};

93
94
95
96
97
98
99
100
101
102
103
104
struct ConstraintOrderer : public binary_function<int, int, bool> {
    const vector<int>& atom1;
    const vector<int>& atom2;
    ConstraintOrderer(const vector<int>& atom1, const vector<int>& atom2) : atom1(atom1), atom2(atom2) {
    }
    bool operator()(int x, int y) {
        if (atom1[x] != atom1[y])
            return atom1[x] < atom1[y];
        return atom2[x] < atom2[y];
    }
};

105
106
107
108
109
110
111
struct Molecule {
    vector<int> atoms;
    vector<int> bonds;
    vector<int> angles;
    vector<int> periodicTorsions;
    vector<int> rbTorsions;
    vector<int> constraints;
112
    vector<int> lj14s;
113
114
};

Peter Eastman's avatar
Peter Eastman committed
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
static const float dielectricOffset         =    0.009f;
static const float probeRadius              =    0.14f;
static const float forceConversionFactor    =    0.4184f;

//static const float surfaceAreaFactor        =   -6.0f * 0.06786f * forceConversionFactor * 1000.0f;  // PI * 4.0f * 0.0049f * 1000.0f;
//static const float surfaceAreaFactor        =   -6.0f * PI * 4.0f * 0.0049f * 1000.0f;
static const float surfaceAreaFactor        = -6.0f*PI*0.0216f*1000.0f*0.4184f;
//static const float surfaceAreaFactor        = -1.7035573959e+001;
//static const float surfaceAreaFactor        = -166.02691f;
//static const float surfaceAreaFactor        = 1.0f;

static const float alphaOBC                 =    1.0f;
static const float betaOBC                  =    0.8f;
static const float gammaOBC                 =    4.85f;
static const float kcalMolTokJNM            =   -0.4184f;
static const float electricConstant         = -166.02691f;
static const float defaultInnerDielectric   =    1.0f;
static const float defaultSolventDielectric =   78.3f;
static const float KILO                     =    1e3;                      // Thousand
static const float BOLTZMANN                =    1.380658e-23f;            // (J/K)    
static const float AVOGADRO                 =    6.0221367e23f;            // ()        
static const float RGAS                     =    BOLTZMANN * AVOGADRO;     // (J/(mol K))
static const float BOLTZ                    =    (RGAS / KILO);            // (kJ/(mol K)) 

#define DUMP_PARAMETERS 0

141
template <int SIZE>
142
static Expression<SIZE> createExpression(gpuContext gpu, const string& expression, const Lepton::ExpressionProgram& program, const vector<string>& variables,
143
        const vector<string>& globalParamNames, unsigned int& maxStackSize) {
144
145
146
147
148
    Expression<SIZE> exp;
    if (program.getNumOperations() > SIZE)
        throw OpenMMException("Expression contains too many operations: "+expression);
    exp.length = program.getNumOperations();
    exp.stackSize = program.getStackSize();
149
150
    if (exp.stackSize > maxStackSize)
        maxStackSize = exp.stackSize;
151
152
153
154
155
    for (int i = 0; i < program.getNumOperations(); i++) {
        const Operation& op = program.getOperation(i);
        switch (op.getId()) {
            case Operation::CONSTANT:
                exp.op[i] = CONSTANT;
156
                exp.arg[i] = dynamic_cast<const Operation::Constant*>(&op)->getValue();
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
                break;
            case Operation::VARIABLE:
                if (variables.size() > 0 && op.getName() == variables[0])
                    exp.op[i] = VARIABLE0;
                else if (variables.size() > 1 && op.getName() == variables[1])
                    exp.op[i] = VARIABLE1;
                else if (variables.size() > 2 && op.getName() == variables[2])
                    exp.op[i] = VARIABLE2;
                else if (variables.size() > 3 && op.getName() == variables[3])
                    exp.op[i] = VARIABLE3;
                else if (variables.size() > 4 && op.getName() == variables[4])
                    exp.op[i] = VARIABLE4;
                else if (variables.size() > 5 && op.getName() == variables[5])
                    exp.op[i] = VARIABLE5;
                else if (variables.size() > 6 && op.getName() == variables[6])
                    exp.op[i] = VARIABLE6;
                else if (variables.size() > 7 && op.getName() == variables[7])
                    exp.op[i] = VARIABLE7;
175
176
177
178
179
180
181
182
                else {
                    int j;
                    for (j = 0; j < globalParamNames.size() && op.getName() != globalParamNames[j]; j++);
                    if (j == globalParamNames.size())
                        throw OpenMMException("Unknown variable '"+op.getName()+"' in expression: "+expression);
                    exp.op[i] = GLOBAL;
                    exp.arg[i] = j;
                }
183
                break;
184
185
186
187
188
189
190
191
            case Operation::CUSTOM:
                exp.op[i] = dynamic_cast<const Operation::Custom*>(&op)->getDerivOrder()[0] == 0 ? CUSTOM : CUSTOM_DERIV;
                for (int j = 0; j < MAX_TABULATED_FUNCTIONS; j++)
                    if (op.getName() == gpu->tabulatedFunctions[j].name) {
                        exp.arg[i] = j;
                        break;
                    }
                break;
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
            case Operation::ADD:
                exp.op[i] = ADD;
                break;
            case Operation::SUBTRACT:
                exp.op[i] = SUBTRACT;
                break;
            case Operation::MULTIPLY:
                exp.op[i] = MULTIPLY;
                break;
            case Operation::DIVIDE:
                exp.op[i] = DIVIDE;
                break;
            case Operation::POWER:
                exp.op[i] = POWER;
                break;
            case Operation::NEGATE:
                exp.op[i] = NEGATE;
                break;
            case Operation::SQRT:
                exp.op[i] = SQRT;
                break;
            case Operation::EXP:
                exp.op[i] = EXP;
                break;
            case Operation::LOG:
                exp.op[i] = LOG;
                break;
            case Operation::SIN:
                exp.op[i] = SIN;
                break;
            case Operation::COS:
                exp.op[i] = COS;
                break;
            case Operation::SEC:
                exp.op[i] = SEC;
                break;
            case Operation::CSC:
                exp.op[i] = CSC;
                break;
            case Operation::TAN:
                exp.op[i] = TAN;
                break;
            case Operation::COT:
                exp.op[i] = COT;
                break;
            case Operation::ASIN:
                exp.op[i] = ASIN;
                break;
            case Operation::ACOS:
                exp.op[i] = ACOS;
                break;
            case Operation::ATAN:
                exp.op[i] = ATAN;
                break;
            case Operation::SQUARE:
                exp.op[i] = SQUARE;
                break;
            case Operation::CUBE:
                exp.op[i] = CUBE;
                break;
            case Operation::RECIPROCAL:
                exp.op[i] = RECIPROCAL;
                break;
255
256
257
            case Operation::ADD_CONSTANT:
                exp.op[i] = ADD_CONSTANT;
                exp.arg[i] = dynamic_cast<const Operation::AddConstant*>(&op)->getValue();
258
                break;
259
260
261
262
263
264
265
            case Operation::MULTIPLY_CONSTANT:
                exp.op[i] = MULTIPLY_CONSTANT;
                exp.arg[i] = dynamic_cast<const Operation::MultiplyConstant*>(&op)->getValue();
                break;
            case Operation::POWER_CONSTANT:
                exp.op[i] = POWER_CONSTANT;
                exp.arg[i] = dynamic_cast<const Operation::PowerConstant*>(&op)->getValue();
266
267
268
269
270
271
                break;
        }
    }
    return exp;
}

Peter Eastman's avatar
Peter Eastman committed
272
273
274
275
276
extern "C"
void gpuSetBondParameters(gpuContext gpu, const vector<int>& atom1, const vector<int>& atom2, const vector<float>& length, const vector<float>& k)
{
    int bonds = atom1.size();
    gpu->sim.bonds                              = bonds;
277
    CUDAStream<int4>* psBondID                  = new CUDAStream<int4>(bonds, 1, "BondID");
Peter Eastman's avatar
Peter Eastman committed
278
279
    gpu->psBondID                               = psBondID;
    gpu->sim.pBondID                            = psBondID->_pDevStream[0];
280
    CUDAStream<float2>* psBondParameter         = new CUDAStream<float2>(bonds, 1, "BondParameter");
Peter Eastman's avatar
Peter Eastman committed
281
282
283
284
    gpu->psBondParameter                        = psBondParameter;
    gpu->sim.pBondParameter                     = psBondParameter->_pDevStream[0];
    for (int i = 0; i < bonds; i++)
    {
285
286
287
288
289
290
        (*psBondID)[i].x = atom1[i];
        (*psBondID)[i].y = atom2[i];
        (*psBondParameter)[i].x = length[i];
        (*psBondParameter)[i].y = k[i];
        psBondID->_pSysData[i].z = gpu->pOutputBufferCounter[psBondID->_pSysData[i].x]++;
        psBondID->_pSysData[i].w = gpu->pOutputBufferCounter[psBondID->_pSysData[i].y]++;
Peter Eastman's avatar
Peter Eastman committed
291
292
293
#if (DUMP_PARAMETERS == 1)                
        cout << 
            i << " " << 
294
295
296
297
298
299
            (*psBondID)[i].x << " " <<
            (*psBondID)[i].y << " " <<
            (*psBondID)[i].z << " " <<
            (*psBondID)[i].w << " " <<
            (*psBondParameter)[i].x << " " <<
            (*psBondParameter)[i].y <<
Peter Eastman's avatar
Peter Eastman committed
300
301
302
303
304
305
306
307
308
309
310
311
312
            endl;
#endif
    }
    psBondID->Upload();
    psBondParameter->Upload();
}

extern "C"
void gpuSetBondAngleParameters(gpuContext gpu, const vector<int>& atom1, const vector<int>& atom2, const vector<int>& atom3,
        const vector<float>& angle, const vector<float>& k)
{
    int bond_angles = atom1.size();
    gpu->sim.bond_angles                        = bond_angles;
313
    CUDAStream<int4>* psBondAngleID1            = new CUDAStream<int4>(bond_angles, 1, "BondAngleID1");
Peter Eastman's avatar
Peter Eastman committed
314
315
    gpu->psBondAngleID1                         = psBondAngleID1;
    gpu->sim.pBondAngleID1                      = psBondAngleID1->_pDevStream[0];
316
    CUDAStream<int2>* psBondAngleID2            = new CUDAStream<int2>(bond_angles, 1, "BondAngleID2");
Peter Eastman's avatar
Peter Eastman committed
317
318
    gpu->psBondAngleID2                         = psBondAngleID2;
    gpu->sim.pBondAngleID2                      = psBondAngleID2->_pDevStream[0];
319
    CUDAStream<float2>* psBondAngleParameter    = new CUDAStream<float2>(bond_angles, 1, "BondAngleParameter");
Peter Eastman's avatar
Peter Eastman committed
320
321
322
323
324
    gpu->psBondAngleParameter                   = psBondAngleParameter;
    gpu->sim.pBondAngleParameter                = psBondAngleParameter->_pDevStream[0];        

    for (int i = 0; i < bond_angles; i++)
    {
325
326
327
328
329
330
331
332
        (*psBondAngleID1)[i].x = atom1[i];
        (*psBondAngleID1)[i].y = atom2[i];
        (*psBondAngleID1)[i].z = atom3[i];
        (*psBondAngleParameter)[i].x = angle[i];
        (*psBondAngleParameter)[i].y = k[i];
        psBondAngleID1->_pSysData[i].w = gpu->pOutputBufferCounter[psBondAngleID1->_pSysData[i].x]++;
        psBondAngleID2->_pSysData[i].x = gpu->pOutputBufferCounter[psBondAngleID1->_pSysData[i].y]++;
        psBondAngleID2->_pSysData[i].y = gpu->pOutputBufferCounter[psBondAngleID1->_pSysData[i].z]++;
Peter Eastman's avatar
Peter Eastman committed
333
334
335
#if (DUMP_PARAMETERS == 1)
         cout << 
            i << " " << 
336
337
338
339
340
341
342
343
            (*psBondAngleID1)[i].x << " " <<
            (*psBondAngleID1)[i].y << " " <<
            (*psBondAngleID1)[i].z << " " <<
            (*psBondAngleID1)[i].w << " " <<
            (*psBondAngleID2)[i].x << " " <<
            (*psBondAngleID2)[i].y << " " <<
            (*psBondAngleParameter)[i].x << " " <<
            (*psBondAngleParameter)[i].y <<
Peter Eastman's avatar
Peter Eastman committed
344
345
346
347
348
349
350
351
352
353
354
355
356
357
            endl;
#endif
    }
    psBondAngleID1->Upload();
    psBondAngleID2->Upload();
    psBondAngleParameter->Upload();
}

extern "C"
void gpuSetDihedralParameters(gpuContext gpu, const vector<int>& atom1, const vector<int>& atom2, const vector<int>& atom3, const vector<int>& atom4,
        const vector<float>& k, const vector<float>& phase, const vector<int>& periodicity)
{
        int dihedrals = atom1.size();
        gpu->sim.dihedrals = dihedrals;
358
        CUDAStream<int4>* psDihedralID1             = new CUDAStream<int4>(dihedrals, 1, "DihedralID1");
Peter Eastman's avatar
Peter Eastman committed
359
360
        gpu->psDihedralID1                          = psDihedralID1;
        gpu->sim.pDihedralID1                       = psDihedralID1->_pDevStream[0];
361
        CUDAStream<int4>* psDihedralID2             = new CUDAStream<int4>(dihedrals, 1, "DihedralID2");
Peter Eastman's avatar
Peter Eastman committed
362
363
        gpu->psDihedralID2                          = psDihedralID2;
        gpu->sim.pDihedralID2                       = psDihedralID2->_pDevStream[0];
364
        CUDAStream<float4>* psDihedralParameter     = new CUDAStream<float4>(dihedrals, 1, "DihedralParameter");
Peter Eastman's avatar
Peter Eastman committed
365
366
367
368
        gpu->psDihedralParameter                    = psDihedralParameter;
        gpu->sim.pDihedralParameter                 = psDihedralParameter->_pDevStream[0];
        for (int i = 0; i < dihedrals; i++)
        {
369
370
371
372
373
374
375
376
377
378
379
            (*psDihedralID1)[i].x = atom1[i];
            (*psDihedralID1)[i].y = atom2[i];
            (*psDihedralID1)[i].z = atom3[i];
            (*psDihedralID1)[i].w = atom4[i];
            (*psDihedralParameter)[i].x = k[i];
            (*psDihedralParameter)[i].y = phase[i];
            (*psDihedralParameter)[i].z = (float) periodicity[i];
            psDihedralID2->_pSysData[i].x = gpu->pOutputBufferCounter[psDihedralID1->_pSysData[i].x]++;
            psDihedralID2->_pSysData[i].y = gpu->pOutputBufferCounter[psDihedralID1->_pSysData[i].y]++;
            psDihedralID2->_pSysData[i].z = gpu->pOutputBufferCounter[psDihedralID1->_pSysData[i].z]++;
            psDihedralID2->_pSysData[i].w = gpu->pOutputBufferCounter[psDihedralID1->_pSysData[i].w]++;
Peter Eastman's avatar
Peter Eastman committed
380
381
382
#if (DUMP_PARAMETERS == 1)
            cout << 
                i << " " << 
383
384
385
386
387
388
389
390
391
392
393
                (*psDihedralID1)[i].x << " " <<
                (*psDihedralID1)[i].y << " " <<
                (*psDihedralID1)[i].z << " " <<
                (*psDihedralID1)[i].w << " " <<
                (*psDihedralID2)[i].x << " " <<
                (*psDihedralID2)[i].y << " " <<
                (*psDihedralID2)[i].z << " " <<
                (*psDihedralID2)[i].w << " " <<
                (*psDihedralParameter)[i].x << " " <<
                (*psDihedralParameter)[i].y << " " <<
                (*psDihedralParameter)[i].z << endl;
Peter Eastman's avatar
Peter Eastman committed
394
395
396
397
398
399
400
401
402
403
404
405
406
#endif
        }
        psDihedralID1->Upload();
        psDihedralID2->Upload();
        psDihedralParameter->Upload();
}

extern "C"
void gpuSetRbDihedralParameters(gpuContext gpu, const vector<int>& atom1, const vector<int>& atom2, const vector<int>& atom3, const vector<int>& atom4,
        const vector<float>& c0, const vector<float>& c1, const vector<float>& c2, const vector<float>& c3, const vector<float>& c4, const vector<float>& c5)
{
    int rb_dihedrals = atom1.size();
    gpu->sim.rb_dihedrals = rb_dihedrals;
407
    CUDAStream<int4>* psRbDihedralID1           = new CUDAStream<int4>(rb_dihedrals, 1, "RbDihedralID1");
Peter Eastman's avatar
Peter Eastman committed
408
409
    gpu->psRbDihedralID1                        = psRbDihedralID1;
    gpu->sim.pRbDihedralID1                     = psRbDihedralID1->_pDevStream[0];
410
    CUDAStream<int4>* psRbDihedralID2           = new CUDAStream<int4>(rb_dihedrals, 1, "RbDihedralID2");
Peter Eastman's avatar
Peter Eastman committed
411
412
    gpu->psRbDihedralID2                        = psRbDihedralID2;
    gpu->sim.pRbDihedralID2                     = psRbDihedralID2->_pDevStream[0];
413
    CUDAStream<float4>* psRbDihedralParameter1  = new CUDAStream<float4>(rb_dihedrals, 1, "RbDihedralParameter1");
Peter Eastman's avatar
Peter Eastman committed
414
415
    gpu->psRbDihedralParameter1                 = psRbDihedralParameter1;
    gpu->sim.pRbDihedralParameter1              = psRbDihedralParameter1->_pDevStream[0];
416
    CUDAStream<float2>* psRbDihedralParameter2  = new CUDAStream<float2>(rb_dihedrals, 1, "RbDihedralParameter2");
Peter Eastman's avatar
Peter Eastman committed
417
418
419
420
421
    gpu->psRbDihedralParameter2                 = psRbDihedralParameter2;
    gpu->sim.pRbDihedralParameter2              = psRbDihedralParameter2->_pDevStream[0];

    for (int i = 0; i < rb_dihedrals; i++)
    {
422
423
424
425
426
427
428
429
430
431
432
433
434
435
        (*psRbDihedralID1)[i].x = atom1[i];
        (*psRbDihedralID1)[i].y = atom2[i];
        (*psRbDihedralID1)[i].z = atom3[i];
        (*psRbDihedralID1)[i].w = atom4[i];
        (*psRbDihedralParameter1)[i].x = c0[i];
        (*psRbDihedralParameter1)[i].y = c1[i];
        (*psRbDihedralParameter1)[i].z = c2[i];
        (*psRbDihedralParameter1)[i].w = c3[i];
        (*psRbDihedralParameter2)[i].x = c4[i];
        (*psRbDihedralParameter2)[i].y = c5[i];
        psRbDihedralID2->_pSysData[i].x = gpu->pOutputBufferCounter[psRbDihedralID1->_pSysData[i].x]++;
        psRbDihedralID2->_pSysData[i].y = gpu->pOutputBufferCounter[psRbDihedralID1->_pSysData[i].y]++;
        psRbDihedralID2->_pSysData[i].z = gpu->pOutputBufferCounter[psRbDihedralID1->_pSysData[i].z]++;
        psRbDihedralID2->_pSysData[i].w = gpu->pOutputBufferCounter[psRbDihedralID1->_pSysData[i].w]++;
Peter Eastman's avatar
Peter Eastman committed
436
437
438
#if (DUMP_PARAMETERS == 1)
        cout << 
            i << " " << 
439
440
441
442
443
444
445
446
447
448
449
450
451
452
            (*psRbDihedralID1)[i].x << " " <<
            (*psRbDihedralID1)[i].y << " " <<
            (*psRbDihedralID1)[i].z << " " <<
            (*psRbDihedralID1)[i].w <<" " <<
            (*psRbDihedralID2)[i].x << " " <<
            (*psRbDihedralID2)[i].y << " " <<
            (*psRbDihedralID2)[i].z << " " <<
            (*psRbDihedralID2)[i].w <<" " <<
            (*psRbDihedralParameter1)[i].x << " " <<
            (*psRbDihedralParameter1)[i].y << " " <<
            (*psRbDihedralParameter1)[i].z << " " <<
            (*psRbDihedralParameter1)[i].w << " " <<
            (*psRbDihedralParameter2)[i].x << " " <<
            (*psRbDihedralParameter2)[i].y <<
Peter Eastman's avatar
Peter Eastman committed
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
            endl;
#endif
    }
    psRbDihedralID1->Upload();
    psRbDihedralID2->Upload();
    psRbDihedralParameter1->Upload();
    psRbDihedralParameter2->Upload();
}

extern "C"
void gpuSetLJ14Parameters(gpuContext gpu, float epsfac, float fudge, const vector<int>& atom1, const vector<int>& atom2,
        const vector<float>& c6, const vector<float>& c12, const vector<float>& q1, const vector<float>& q2)
{
    int LJ14s = atom1.size();
    float scale = epsfac * fudge;

    gpu->sim.LJ14s                              = LJ14s;
470
    CUDAStream<int4>* psLJ14ID                  = new CUDAStream<int4>(LJ14s, 1, "LJ14ID");
Peter Eastman's avatar
Peter Eastman committed
471
472
    gpu->psLJ14ID                               = psLJ14ID;
    gpu->sim.pLJ14ID                            = psLJ14ID->_pDevStream[0];
473
    CUDAStream<float4>* psLJ14Parameter         = new CUDAStream<float4>(LJ14s, 1, "LJ14Parameter");
Peter Eastman's avatar
Peter Eastman committed
474
475
476
477
478
    gpu->psLJ14Parameter                        = psLJ14Parameter;
    gpu->sim.pLJ14Parameter                     = psLJ14Parameter->_pDevStream[0];

    for (int i = 0; i < LJ14s; i++)
    {
479
480
481
482
        (*psLJ14ID)[i].x = atom1[i];
        (*psLJ14ID)[i].y = atom2[i];
        psLJ14ID->_pSysData[i].z = gpu->pOutputBufferCounter[psLJ14ID->_pSysData[i].x]++;
        psLJ14ID->_pSysData[i].w = gpu->pOutputBufferCounter[psLJ14ID->_pSysData[i].y]++;
Peter Eastman's avatar
Peter Eastman committed
483
484
485
486
487
488
489
490
491
492
493
494
        float p0, p1, p2;
        if (c12[i] == 0.0f)
        {
            p0 = 0.0f;
            p1 = 1.0f;
        }
        else
        {
            p0 = c6[i] * c6[i] / c12[i];
            p1 = pow(c12[i] / c6[i], 1.0f / 6.0f);
        }
        p2 = scale * q1[i] * q2[i];
495
496
497
        (*psLJ14Parameter)[i].x = p0;
        (*psLJ14Parameter)[i].y = p1;
        (*psLJ14Parameter)[i].z = p2;
Peter Eastman's avatar
Peter Eastman committed
498
499
500
501
    }
#if (DUMP_PARAMETERS == 1)
        cout << 
            i << " " <<
502
503
504
505
506
507
508
            (*psLJ14ID)[i].x << " " <<
            (*psLJ14ID)[i].y << " " <<
            (*psLJ14ID)[i].z << " " <<
            (*psLJ14ID)[i].w << " " <<
            (*psLJ14Parameter)[i].x << " " <<
            (*psLJ14Parameter)[i].y << " " <<
            (*psLJ14Parameter)[i].z << " " <<
Peter Eastman's avatar
Peter Eastman committed
509
510
511
512
513
514
515
516
517
            p0 << " " << 
            p1 << " " << 
            p2 << " " << 
            endl;
#endif
    psLJ14ID->Upload();
    psLJ14Parameter->Upload();
}

518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
static void setExclusions(gpuContext gpu, const vector<vector<int> >& exclusions) {
    if (gpu->exclusions.size() > 0) {
        bool ok = (exclusions.size() == gpu->exclusions.size());
        for (int i = 0; i < exclusions.size() && ok; i++) {
            if (exclusions[i].size() != gpu->exclusions[i].size())
                ok = false;
            else {
                for (int j = 0; j < exclusions[i].size(); j++)
                    if (find(gpu->exclusions[i].begin(), gpu->exclusions[i].end(), exclusions[i][j]) == gpu->exclusions[i].end())
                        ok = false;
            }
        }
        if (!ok)
            throw OpenMMException("All nonbonded forces must have identical sets of exceptions");
    }
    gpu->exclusions = exclusions;
}

Peter Eastman's avatar
Peter Eastman committed
536
537
extern "C"
void gpuSetCoulombParameters(gpuContext gpu, float epsfac, const vector<int>& atom, const vector<float>& c6, const vector<float>& c12, const vector<float>& q,
538
        const vector<char>& symbol, const vector<vector<int> >& exclusions, CudaNonbondedMethod method)
Peter Eastman's avatar
Peter Eastman committed
539
{
540
    unsigned int coulombs = c6.size();
Peter Eastman's avatar
Peter Eastman committed
541
    gpu->sim.epsfac = epsfac;
542
    gpu->sim.nonbondedMethod = method;
543
544
545
    if (coulombs > 0)
        setExclusions(gpu, exclusions);
    
Peter Eastman's avatar
Peter Eastman committed
546
547
548
549
550
551
552
553
554
555
556
    for (unsigned int i = 0; i < coulombs; i++)
    {
            float p0 = q[i];
            float p1 = 0.5f, p2 = 0.0f;               
            if ((c6[i] > 0.0f) && (c12[i] > 0.0f))
            {
                p1 = 0.5f * pow(c12[i] / c6[i], 1.0f / 6.0f);
                p2 = c6[i] * sqrt(1.0f / c12[i]);
            }
            if (symbol.size() > 0)
                gpu->pAtomSymbol[i] = symbol[i];
557
558
559
            (*gpu->psPosq4)[i].w = p0;
            (*gpu->psSigEps2)[i].x = p1;
            (*gpu->psSigEps2)[i].y = p2;
Peter Eastman's avatar
Peter Eastman committed
560
561
562
563
564
    }

    // Dummy out extra atom data
    for (unsigned int i = coulombs; i < gpu->sim.paddedNumberOfAtoms; i++)
    {
565
566
567
568
569
570
        (*gpu->psPosq4)[i].x       = 100000.0f + i * 10.0f;
        (*gpu->psPosq4)[i].y       = 100000.0f + i * 10.0f;
        (*gpu->psPosq4)[i].z       = 100000.0f + i * 10.0f;
        (*gpu->psPosq4)[i].w       = 0.0f;
        (*gpu->psSigEps2)[i].x     = 0.0f;
        (*gpu->psSigEps2)[i].y     = 0.0f;
Peter Eastman's avatar
Peter Eastman committed
571
572
573
574
    }

    gpu->psPosq4->Upload();
    gpu->psSigEps2->Upload();
575
}
Peter Eastman's avatar
Peter Eastman committed
576

577
578
579
extern "C"
void gpuSetNonbondedCutoff(gpuContext gpu, float cutoffDistance, float solventDielectric)
{
580
581
582
    if (gpu->sim.nonbondedCutoff != 0.0f && gpu->sim.nonbondedCutoff != cutoffDistance)
        throw OpenMMException("All nonbonded forces must use the same cutoff");
    gpu->sim.nonbondedCutoff = cutoffDistance;
583
584
    gpu->sim.nonbondedCutoffSqr = cutoffDistance*cutoffDistance;
    gpu->sim.reactionFieldK = pow(cutoffDistance, -3.0f)*(solventDielectric-1.0f)/(2.0f*solventDielectric+1.0f);
585
    gpu->sim.reactionFieldC = (1.0f / cutoffDistance)*(3.0f*solventDielectric)/(2.0f*solventDielectric+1.0f);
586
}
Peter Eastman's avatar
Peter Eastman committed
587

588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
extern "C"
void gpuSetTabulatedFunction(gpuContext gpu, int index, const string& name, const vector<double>& values, double min, double max, bool interpolating)
{
    if (index < 0 || index >= MAX_TABULATED_FUNCTIONS) {
        stringstream str;
        str << "Only " << MAX_TABULATED_FUNCTIONS << " tabulated functions are supported";
        throw OpenMMException(str.str());
    }
    if (gpu->tabulatedFunctions[index].coefficients != NULL)
        delete gpu->tabulatedFunctions[index].coefficients;
    CUDAStream<float4>* coeff = new CUDAStream<float4>((int) values.size()-1, 1, "TabulatedFunction");
    gpu->tabulatedFunctions[index].coefficients = coeff;
    gpu->sim.pTabulatedFunctionCoefficients[index] = coeff->_pDevData;
    gpu->tabulatedFunctions[index].name = name;
    gpu->tabulatedFunctions[index].min = min;
    gpu->tabulatedFunctions[index].max = max;
604
    gpu->tabulatedFunctionsChanged = true;
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634

    // First create a padded set of function values.

    vector<double> padded(values.size()+2);
    padded[0] = 2*values[0]-values[1];
    for (int i = 0; i < (int) values.size(); i++)
        padded[i+1] = values[i];
    padded[padded.size()-1] = 2*values[values.size()-1]-values[values.size()-2];

    // Now compute the spline coefficients.

    for (int i = 0; i < (int) values.size()-1; i++) {
        float4 c;
        if (interpolating) {
            c.x = padded[i+1];
            c.y = 0.5*(-padded[i]+padded[i+2]);
            c.z = 0.5*(2.0*padded[i]-5.0*padded[i+1]+4.0*padded[i+2]-padded[i+3]);
            c.w = 0.5*(-padded[i]+3.0*padded[i+1]-3.0*padded[i+2]+padded[i+3]);
        }
        else {
            c.x = (padded[i]+4.0*padded[i+1]+padded[i+2])/6.0;
            c.y = (-3.0*padded[i]+3.0*padded[i+2])/6.0;
            c.z = (3.0*padded[i]-6.0*padded[i+1]+3.0*padded[i+2])/6.0;
            c.w = (-padded[i]+3.0*padded[i+1]-3.0*padded[i+2]+padded[i+3])/6.0;
        }
        (*coeff)[i] = c;
    }
    coeff->Upload();
}

635
636
637
extern "C"
void gpuSetCustomNonbondedParameters(gpuContext gpu, const vector<vector<double> >& parameters, const vector<vector<int> >& exclusions,
            const vector<int>& exceptionAtom1, const vector<int>& exceptionAtom2, const vector<vector<double> >& exceptionParams,
638
639
            CudaNonbondedMethod method, float cutoffDistance, const string& energyExp, const vector<string>& combiningRules,
            const vector<string>& paramNames, const vector<string>& globalParamNames)
640
641
642
643
644
{
    if (gpu->sim.nonbondedCutoff != 0.0f && gpu->sim.nonbondedCutoff != cutoffDistance)
        throw OpenMMException("All nonbonded forces must use the same cutoff");
    if (paramNames.size() > 4)
        throw OpenMMException("CudaPlatform only supports four per-atom parameters for custom nonbonded forces");
645
646
    if (globalParamNames.size() > 8)
        throw OpenMMException("CudaPlatform only supports eight global parameters for custom nonbonded forces");
647
648
649
650
    gpu->sim.nonbondedCutoff = cutoffDistance;
    gpu->sim.nonbondedCutoffSqr = cutoffDistance*cutoffDistance;
    gpu->sim.customNonbondedMethod = method;
    gpu->sim.customExceptions = exceptionAtom1.size();
651
    gpu->sim.customParameters = paramNames.size();
652
653
654
655
656
    gpu->sim.custom_exception_threads_per_block = (gpu->sim.customExceptions+gpu->sim.blocks-1)/gpu->sim.blocks;
    if (gpu->sim.custom_exception_threads_per_block < 1)
        gpu->sim.custom_exception_threads_per_block = 1;
    if (gpu->sim.custom_exception_threads_per_block > gpu->sim.max_localForces_threads_per_block)
        gpu->sim.custom_exception_threads_per_block = gpu->sim.max_localForces_threads_per_block;
657
658
659
660
661
662
663
    setExclusions(gpu, exclusions);
    gpu->psCustomParams = new CUDAStream<float4>(gpu->sim.paddedNumberOfAtoms, 1, "CustomParams");
    gpu->sim.pCustomParams = gpu->psCustomParams->_pDevData;
    gpu->psCustomExceptionID = new CUDAStream<int4>(gpu->sim.customExceptions, 1, "CustomExceptionId");
    gpu->sim.pCustomExceptionID = gpu->psCustomExceptionID->_pDevData;
    gpu->psCustomExceptionParams = new CUDAStream<float4>(gpu->sim.customExceptions, 1, "CustomExceptionParams");
    gpu->sim.pCustomExceptionParams = gpu->psCustomExceptionParams->_pDevData;
664
    for (int i = 0; i < (int) parameters.size(); i++) {
665
        if (parameters[i].size() > 0)
666
            (*gpu->psCustomParams)[i].x = (float) parameters[i][0];
667
        if (parameters[i].size() > 1)
668
            (*gpu->psCustomParams)[i].y = (float) parameters[i][1];
669
        if (parameters[i].size() > 2)
670
            (*gpu->psCustomParams)[i].z = (float) parameters[i][2];
671
        if (parameters[i].size() > 3)
672
            (*gpu->psCustomParams)[i].w = (float) parameters[i][3];
673
    }
674
    for (int i = 0; i < (int) exceptionAtom1.size(); i++) {
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
        (*gpu->psCustomExceptionID)[i].x = exceptionAtom1[i];
        (*gpu->psCustomExceptionID)[i].y = exceptionAtom2[i];
        (*gpu->psCustomExceptionID)[i].z = gpu->pOutputBufferCounter[exceptionAtom1[i]]++;
        (*gpu->psCustomExceptionID)[i].w = gpu->pOutputBufferCounter[exceptionAtom2[i]]++;
        if (exceptionParams[i].size() > 0)
            (*gpu->psCustomExceptionParams)[i].x = exceptionParams[i][0];
        if (exceptionParams[i].size() > 1)
            (*gpu->psCustomExceptionParams)[i].y = exceptionParams[i][1];
        if (exceptionParams[i].size() > 2)
            (*gpu->psCustomExceptionParams)[i].z = exceptionParams[i][2];
        if (exceptionParams[i].size() > 3)
            (*gpu->psCustomExceptionParams)[i].w = exceptionParams[i][3];
    }
    gpu->psCustomParams->Upload();
    gpu->psCustomExceptionID->Upload();
    gpu->psCustomExceptionParams->Upload();

692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
    // This class serves as a placeholder for custom functions in expressions.

    class FunctionPlaceholder : public Lepton::CustomFunction {
    public:
        int getNumArguments() const {
            return 1;
        }
        double evaluate(const double* arguments) const {
            return 0.0;
        }
        double evaluateDerivative(const double* arguments, const int* derivOrder) const {
            return 0.0;
        }
        CustomFunction* clone() const {
            return new FunctionPlaceholder();
        }
    };

    // Record the tabulated functions, which were previously set with calls to gpuSetTabulatedFunction().

    FunctionPlaceholder* fp = new FunctionPlaceholder();
    map<string, Lepton::CustomFunction*> functions;
    gpu->psTabulatedFunctionParams = new CUDAStream<float4>(MAX_TABULATED_FUNCTIONS, 1, "TabulatedFunctionRange");
    gpu->sim.pTabulatedFunctionParams = gpu->psTabulatedFunctionParams->_pDevData;
    for (int i = 0; i < MAX_TABULATED_FUNCTIONS; i++) {
        gpuTabulatedFunction& func = gpu->tabulatedFunctions[i];
        if (func.coefficients != NULL) {
719
            (*gpu->psTabulatedFunctionParams)[i] = make_float4((float) func.min, (float) func.max, (float) (func.coefficients->_length/(func.max-func.min)), 0.0f);
720
721
722
723
724
            functions[func.name] = fp;
        }
    }
    gpu->psTabulatedFunctionParams->Upload();

725
726
727
728
    // Create the Expressions.

    vector<string> variables;
    variables.push_back("r");
729
    for (int i = 0; i < (int) paramNames.size(); i++)
730
        variables.push_back(paramNames[i]);
731
    gpu->sim.customExpressionStackSize = 0;
732
733
    SetCustomNonbondedEnergyExpression(createExpression<128>(gpu, energyExp, Lepton::Parser::parse(energyExp, functions).optimize().createProgram(), variables, globalParamNames, gpu->sim.customExpressionStackSize));
    SetCustomNonbondedForceExpression(createExpression<128>(gpu, energyExp, Lepton::Parser::parse(energyExp, functions).differentiate("r").optimize().createProgram(), variables, globalParamNames, gpu->sim.customExpressionStackSize));
734
735
736
    Expression<64> paramExpressions[4];
    vector<string> combiningRuleParams;
    for (int j = 1; j < 3; j++) {
737
        for (int i = 0; i < (int) paramNames.size(); i++) {
738
739
740
741
            stringstream name;
            name << paramNames[i] << j;
            combiningRuleParams.push_back(name.str());
        }
742
743
        for (int i = paramNames.size(); i < 4; i++)
            combiningRuleParams.push_back("");
744
    }
745
    for (int i = 0; i < (int) paramNames.size(); i++)
746
        paramExpressions[i] = createExpression<64>(gpu, combiningRules[i], Lepton::Parser::parse(combiningRules[i], functions).optimize().createProgram(), combiningRuleParams, globalParamNames, gpu->sim.customExpressionStackSize);
747
    SetCustomNonbondedCombiningRules(paramExpressions);
748
    delete fp;
749
750
}

Peter Eastman's avatar
Peter Eastman committed
751
752
753
754
755
756
757
758
759
760
761
762
static void tabulateErfc(gpuContext gpu)
{
    int tableSize = 2048;
    gpu->sim.tabulatedErfcSize = tableSize;
    gpu->sim.tabulatedErfcScale = tableSize/(gpu->sim.alphaEwald*gpu->sim.nonbondedCutoff);
    gpu->psTabulatedErfc = new CUDAStream<float>(tableSize, 1, "TabulatedErfc");
    gpu->sim.pTabulatedErfc = gpu->psTabulatedErfc->_pDevData;
    for (int i = 0; i < tableSize; ++i)
        (*gpu->psTabulatedErfc)[i] = erfc(i*(gpu->sim.alphaEwald*gpu->sim.nonbondedCutoff)/tableSize);
    gpu->psTabulatedErfc->Upload();
}

Rossen Apostolov's avatar
Rossen Apostolov committed
763
extern "C"
764
void gpuSetEwaldParameters(gpuContext gpu, float alpha, int kmaxx, int kmaxy, int kmaxz)
Rossen Apostolov's avatar
Rossen Apostolov committed
765
{
766
767
    gpu->sim.alphaEwald         = alpha;
    gpu->sim.factorEwald        = -1 / (4*alpha*alpha);
768
769
770
771
    gpu->sim.kmaxX              = kmaxx;
    gpu->sim.kmaxY              = kmaxy;
    gpu->sim.kmaxZ              = kmaxz;
    gpu->psEwaldCosSinSum       = new CUDAStream<float2>((gpu->sim.kmaxX*2-1) * (gpu->sim.kmaxY*2-1) * (gpu->sim.kmaxZ*2-1), 1, "EwaldCosSinSum");
772
    gpu->sim.pEwaldCosSinSum    = gpu->psEwaldCosSinSum->_pDevStream[0];
Peter Eastman's avatar
Peter Eastman committed
773
    tabulateErfc(gpu);
Rossen Apostolov's avatar
Rossen Apostolov committed
774
775
}

776
extern "C"
Peter Eastman's avatar
Peter Eastman committed
777
void gpuSetPMEParameters(gpuContext gpu, float alpha, int gridSizeX, int gridSizeY, int gridSizeZ)
778
779
{
    gpu->sim.alphaEwald         = alpha;
Peter Eastman's avatar
Peter Eastman committed
780
    int3 gridSize = make_int3(gridSizeX, gridSizeY, gridSizeZ);
781
    gpu->sim.pmeGridSize = gridSize;
Peter Eastman's avatar
Peter Eastman committed
782
783
784
785
    int3 groupSize = make_int3(2, 4, 4);
    gpu->sim.pmeGroupSize = groupSize;
    const int3 numGroups = make_int3((gridSize.x+groupSize.x-1)/groupSize.x, (gridSize.y+groupSize.y-1)/groupSize.y, (gridSize.z+groupSize.z-1)/groupSize.z);
    const unsigned int totalGroups = numGroups.x*numGroups.y*numGroups.z;
786
787
788
789
790
791
792
793
794
795
796
797
798
    cufftPlan3d(&gpu->fftplan, gridSize.x, gridSize.y, gridSize.z, CUFFT_C2C);
    gpu->psPmeGrid = new CUDAStream<cufftComplex>(gridSize.x*gridSize.y*gridSize.z, 1, "PmeGrid");
    gpu->sim.pPmeGrid = gpu->psPmeGrid->_pDevData;
    gpu->psPmeBsplineModuli[0] = new CUDAStream<float>(gridSize.x, 1, "PmeBsplineModuli0");
    gpu->sim.pPmeBsplineModuli[0] = gpu->psPmeBsplineModuli[0]->_pDevData;
    gpu->psPmeBsplineModuli[1] = new CUDAStream<float>(gridSize.y, 1, "PmeBsplineModuli1");
    gpu->sim.pPmeBsplineModuli[1] = gpu->psPmeBsplineModuli[1]->_pDevData;
    gpu->psPmeBsplineModuli[2] = new CUDAStream<float>(gridSize.z, 1, "PmeBsplineModuli2");
    gpu->sim.pPmeBsplineModuli[2] = gpu->psPmeBsplineModuli[2]->_pDevData;
    gpu->psPmeBsplineTheta = new CUDAStream<float4>(PME_ORDER*gpu->natoms, 1, "PmeBsplineTheta");
    gpu->sim.pPmeBsplineTheta = gpu->psPmeBsplineTheta->_pDevData;
    gpu->psPmeBsplineDtheta = new CUDAStream<float4>(PME_ORDER*gpu->natoms, 1, "PmeBsplineDtheta");
    gpu->sim.pPmeBsplineDtheta = gpu->psPmeBsplineDtheta->_pDevData;
799
800
801
802
    gpu->psPmeAtomRange = new CUDAStream<int>(gridSize.x*gridSize.y*gridSize.z+1, 1, "PmeAtomRange");
    gpu->sim.pPmeAtomRange = gpu->psPmeAtomRange->_pDevData;
    gpu->psPmeAtomGridIndex = new CUDAStream<float2>(gpu->natoms, 1, "PmeAtomGridIndex");
    gpu->sim.pPmeAtomGridIndex = gpu->psPmeAtomGridIndex->_pDevData;
Peter Eastman's avatar
Peter Eastman committed
803
    tabulateErfc(gpu);
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852

    // Initialize the b-spline moduli.

    int maxSize = max(max(gridSize.x, gridSize.y), gridSize.z);
    vector<double> data(PME_ORDER);
    vector<double> ddata(PME_ORDER);
    vector<double> bsplines_data(maxSize);
    data[PME_ORDER-1] = 0.0;
    data[1] = 0.0;
    data[0] = 1.0;
    for (int i = 3; i < PME_ORDER; 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 < PME_ORDER; i++)
        ddata[i] = data[i-1]-data[i];
    double div = 1.0/(PME_ORDER-1);
    data[PME_ORDER-1] = 0.0;
    for (int i = 1; i < (PME_ORDER-1); i++)
        data[PME_ORDER-i-1] = div*(i*data[PME_ORDER-i-2]+(PME_ORDER-i)*data[PME_ORDER-i-1]);
    data[0] = div*data[0];
    for (int i = 0; i < maxSize; i++)
        bsplines_data[i] = 0.0;
    for (int i = 1; i <= PME_ORDER; 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 ? gridSize.x : dim == 1 ? gridSize.y : gridSize.z);
        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);
            }
853
            (*gpu->psPmeBsplineModuli[dim])[i] = (float) (sc*sc+ss*ss);
854
855
856
857
        }
        for (int i = 0; i < ndata; i++)
        {
            if ((*gpu->psPmeBsplineModuli[dim])[i] < 1.0e-7)
858
                (*gpu->psPmeBsplineModuli[dim])[i] = ((*gpu->psPmeBsplineModuli[dim])[i-1]+(*gpu->psPmeBsplineModuli[dim])[i+1])*0.5f;
859
860
861
        }
        gpu->psPmeBsplineModuli[dim]->Upload();
    }
862
863
}

864
865
866
867
868
869
extern "C"
void gpuSetPeriodicBoxSize(gpuContext gpu, float xsize, float ysize, float zsize)
{
    gpu->sim.periodicBoxSizeX = xsize;
    gpu->sim.periodicBoxSizeY = ysize;
    gpu->sim.periodicBoxSizeZ = zsize;
870
871
872
873
    gpu->sim.recipBoxSizeX = 2.0f*PI/gpu->sim.periodicBoxSizeX;
    gpu->sim.recipBoxSizeY = 2.0f*PI/gpu->sim.periodicBoxSizeY;
    gpu->sim.recipBoxSizeZ = 2.0f*PI/gpu->sim.periodicBoxSizeZ;
    gpu->sim.cellVolume = gpu->sim.periodicBoxSizeX*gpu->sim.periodicBoxSizeY*gpu->sim.periodicBoxSizeZ;
Peter Eastman's avatar
Peter Eastman committed
874
875
876
}

extern "C"
877
void gpuSetObcParameters(gpuContext gpu, float innerDielectric, float solventDielectric, const vector<float>& radius, const vector<float>& scale, const vector<float>& charge)
Peter Eastman's avatar
Peter Eastman committed
878
{
879
    unsigned int atoms = radius.size();
880
881

    gpu->bIncludeGBSA = true;
Peter Eastman's avatar
Peter Eastman committed
882
883
    for (unsigned int i = 0; i < atoms; i++)
    {
884
885
            (*gpu->psObcData)[i].x = radius[i] - dielectricOffset;
            (*gpu->psObcData)[i].y = scale[i] * (*gpu->psObcData)[i].x;
886
            (*gpu->psPosq4)[i].w = charge[i];
Peter Eastman's avatar
Peter Eastman committed
887
888
889
890

#if (DUMP_PARAMETERS == 1)
        cout << 
            i << " " << 
891
892
            (*gpu->psObcData)[i].x << " " <<
            (*gpu->psObcData)[i].y;
Peter Eastman's avatar
Peter Eastman committed
893
894
895
896
897
898
#endif
    }

    // Dummy out extra atom data
    for (unsigned int i = atoms; i < gpu->sim.paddedNumberOfAtoms; i++)
    {
899
900
901
        (*gpu->psBornRadii)[i]     = 0.2f;
        (*gpu->psObcData)[i].x     = 0.01f;
        (*gpu->psObcData)[i].y     = 0.01f;
Peter Eastman's avatar
Peter Eastman committed
902
903
904
905
    }

    gpu->psBornRadii->Upload();
    gpu->psObcData->Upload();
906
    gpu->psPosq4->Upload();
Peter Eastman's avatar
Peter Eastman committed
907
908
909
    gpu->sim.preFactor = 2.0f*electricConstant*((1.0f/innerDielectric)-(1.0f/solventDielectric))*gpu->sim.forceConversionFactor;
}

910
static void markShakeClusterInvalid(ShakeCluster& cluster, map<int, ShakeCluster>& allClusters, vector<bool>& invalidForShake)
911
912
913
914
915
916
917
918
919
920
921
{
    cluster.valid = false;
    invalidForShake[cluster.centralID] = true;
    for (int i = 0; i < cluster.size; i++) {
        invalidForShake[cluster.peripheralID[i]] = true;
        map<int, ShakeCluster>::iterator otherCluster = allClusters.find(cluster.peripheralID[i]);
        if (otherCluster != allClusters.end() && otherCluster->second.valid)
            markShakeClusterInvalid(otherCluster->second, allClusters, invalidForShake);
    }
}

Peter Eastman's avatar
Peter Eastman committed
922
extern "C"
923
void gpuSetConstraintParameters(gpuContext gpu, const vector<int>& atom1, const vector<int>& atom2, const vector<float>& distance,
924
        const vector<float>& invMass1, const vector<float>& invMass2, float constraintTolerance)
Peter Eastman's avatar
Peter Eastman committed
925
{
926
927
928
929
    // Create a vector for recording which atoms are handled by SHAKE (or SETTLE).

    vector<bool> isShakeAtom(gpu->natoms, false);

Peter Eastman's avatar
Peter Eastman committed
930
931
932
    // Find how many constraints each atom is involved in.
    
    vector<int> constraintCount(gpu->natoms, 0);
933
    for (int i = 0; i < (int)atom1.size(); i++) {
Peter Eastman's avatar
Peter Eastman committed
934
935
936
        constraintCount[atom1[i]]++;
        constraintCount[atom2[i]]++;
    }
937
938
939
940
941
942

    // Identify clusters of three atoms that can be treated with SETTLE.  First, for every
    // atom that might be part of such a cluster, make a list of the two other atoms it is
    // connected to.

    vector<map<int, float> > settleConstraints(gpu->natoms);
943
    for (int i = 0; i < (int)atom1.size(); i++) {
944
945
946
947
948
949
950
951
952
        if (constraintCount[atom1[i]] == 2 && constraintCount[atom2[i]] == 2) {
            settleConstraints[atom1[i]][atom2[i]] = distance[i];
            settleConstraints[atom2[i]][atom1[i]] = distance[i];
        }
    }

    // Now remove the ones that don't actually form closed loops of three atoms.

    vector<int> settleClusters;
953
    for (int i = 0; i < (int)settleConstraints.size(); i++) {
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
        if (settleConstraints[i].size() == 2) {
            int partner1 = settleConstraints[i].begin()->first;
            int partner2 = (++settleConstraints[i].begin())->first;
            if (settleConstraints[partner1].size() != 2 || settleConstraints[partner2].size() != 2 ||
                    settleConstraints[partner1].find(partner2) == settleConstraints[partner1].end())
                settleConstraints[i].clear();
            else if (i < partner1 && i < partner2)
                settleClusters.push_back(i);
        }
        else
            settleConstraints[i].clear();
    }

    // Record the actual SETTLE clusters.

969
    CUDAStream<int4>* psSettleID          = new CUDAStream<int4>((int) settleClusters.size(), 1, "SettleID");
970
971
    gpu->psSettleID                       = psSettleID;
    gpu->sim.pSettleID                    = psSettleID->_pDevStream[0];
972
    CUDAStream<float2>* psSettleParameter = new CUDAStream<float2>((int) settleClusters.size(), 1, "SettleParameter");
973
974
975
    gpu->psSettleParameter                = psSettleParameter;
    gpu->sim.pSettleParameter             = psSettleParameter->_pDevStream[0];
    gpu->sim.settleConstraints            = settleClusters.size();
976
      for (int i = 0; i < (int)settleClusters.size(); i++) {
977
978
979
980
981
982
983
        int atom1 = settleClusters[i];
        int atom2 = settleConstraints[atom1].begin()->first;
        int atom3 = (++settleConstraints[atom1].begin())->first;
        float dist12 = settleConstraints[atom1].find(atom2)->second;
        float dist13 = settleConstraints[atom1].find(atom3)->second;
        float dist23 = settleConstraints[atom2].find(atom3)->second;
        if (dist12 == dist13) { // atom1 is the central atom
984
985
986
987
988
            (*psSettleID)[i].x = atom1;
            (*psSettleID)[i].y = atom2;
            (*psSettleID)[i].z = atom3;
            (*psSettleParameter)[i].x = dist12;
            (*psSettleParameter)[i].y = dist23;
989
990
        }
        else if (dist12 == dist23) { // atom2 is the central atom
991
992
993
994
995
            (*psSettleID)[i].x = atom2;
            (*psSettleID)[i].y = atom1;
            (*psSettleID)[i].z = atom3;
            (*psSettleParameter)[i].x = dist12;
            (*psSettleParameter)[i].y = dist13;
996
997
        }
        else if (dist13 == dist23) { // atom3 is the central atom
998
999
1000
1001
1002
            (*psSettleID)[i].x = atom3;
            (*psSettleID)[i].y = atom1;
            (*psSettleID)[i].z = atom2;
            (*psSettleParameter)[i].x = dist13;
            (*psSettleParameter)[i].y = dist12;
1003
1004
1005
        }
        else
            throw OpenMMException("Two of the three distances constrained with SETTLE must be the same.");
1006
1007
1008
        isShakeAtom[atom1] = true;
        isShakeAtom[atom2] = true;
        isShakeAtom[atom3] = true;
1009
1010
1011
1012
1013
1014
1015
1016
1017
    }
    psSettleID->Upload();
    psSettleParameter->Upload();
    gpu->sim.settle_threads_per_block     = (gpu->sim.settleConstraints + gpu->sim.blocks - 1) / gpu->sim.blocks;
    if (gpu->sim.settle_threads_per_block > gpu->sim.max_shake_threads_per_block)
        gpu->sim.settle_threads_per_block = gpu->sim.max_shake_threads_per_block;
    if (gpu->sim.settle_threads_per_block < 1)
        gpu->sim.settle_threads_per_block = 1;

1018
1019
1020
1021
    // Find clusters consisting of a central atom with up to three peripheral atoms.

    map<int, ShakeCluster> clusters;
    vector<bool> invalidForShake(gpu->natoms, false);
1022
    for (int i = 0; i < (int)atom1.size(); i++) {
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
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
        if (isShakeAtom[atom1[i]])
            continue; // This is being taken care of with SETTLE.

        // Determine which is the central atom.

        bool firstIsCentral;
        if (constraintCount[atom1[i]] > 1)
            firstIsCentral = true;
        else if (constraintCount[atom2[i]] > 1)
            firstIsCentral = false;
        else if (atom1[i] < atom2[i])
            firstIsCentral = true;
        else
            firstIsCentral = false;
        int centralID, peripheralID;
        float centralInvMass, peripheralInvMass;
        if (firstIsCentral) {
            centralID = atom1[i];
            peripheralID = atom2[i];
            centralInvMass = invMass1[i];
            peripheralInvMass = invMass2[i];
        }
        else {
            centralID = atom2[i];
            peripheralID = atom1[i];
            centralInvMass = invMass2[i];
            peripheralInvMass = invMass1[i];
        }

        // Add it to the cluster.

        if (clusters.find(centralID) == clusters.end()) {
            clusters[centralID] = ShakeCluster(centralID, centralInvMass);
        }
        ShakeCluster& cluster = clusters[centralID];
        cluster.addAtom(peripheralID, distance[i], peripheralInvMass);
        if (constraintCount[peripheralID] != 1 || invalidForShake[atom1[i]] || invalidForShake[atom2[i]]) {
            markShakeClusterInvalid(cluster, clusters, invalidForShake);
            map<int, ShakeCluster>::iterator otherCluster = clusters.find(peripheralID);
            if (otherCluster != clusters.end() && otherCluster->second.valid)
                markShakeClusterInvalid(otherCluster->second, clusters, invalidForShake);
        }
    }
    int validShakeClusters = 0;
    for (map<int, ShakeCluster>::iterator iter = clusters.begin(); iter != clusters.end(); ++iter) {
        ShakeCluster& cluster = iter->second;
        if (cluster.valid) {
            cluster.valid = !invalidForShake[cluster.centralID];
            for (int i = 0; i < cluster.size; i++)
                if (invalidForShake[cluster.peripheralID[i]])
                    cluster.valid = false;
            if (cluster.valid)
                ++validShakeClusters;
        }
    }

    // Fill in the Cuda streams.

1081
    CUDAStream<int4>* psShakeID             = new CUDAStream<int4>(validShakeClusters, 1, "ShakeID");
1082
1083
    gpu->psShakeID                          = psShakeID;
    gpu->sim.pShakeID                       = psShakeID->_pDevStream[0];
1084
    CUDAStream<float4>* psShakeParameter    = new CUDAStream<float4>(validShakeClusters, 1, "ShakeParameter");
1085
1086
1087
1088
1089
1090
1091
1092
    gpu->psShakeParameter                   = psShakeParameter;
    gpu->sim.pShakeParameter                = psShakeParameter->_pDevStream[0];
    gpu->sim.ShakeConstraints               = validShakeClusters;
    int index = 0;
    for (map<int, ShakeCluster>::const_iterator iter = clusters.begin(); iter != clusters.end(); ++iter) {
        const ShakeCluster& cluster = iter->second;
        if (!cluster.valid)
            continue;
1093
1094
1095
1096
1097
1098
1099
1100
        (*psShakeID)[index].x = cluster.centralID;
        (*psShakeID)[index].y = cluster.peripheralID[0];
        (*psShakeID)[index].z = cluster.size > 1 ? cluster.peripheralID[1] : -1;
        (*psShakeID)[index].w = cluster.size > 2 ? cluster.peripheralID[2] : -1;
        (*psShakeParameter)[index].x = cluster.centralInvMass;
        (*psShakeParameter)[index].y = 0.5f/(cluster.centralInvMass+cluster.peripheralInvMass);
        (*psShakeParameter)[index].z = cluster.distance*cluster.distance;
        (*psShakeParameter)[index].w = cluster.peripheralInvMass;
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
        isShakeAtom[cluster.centralID] = true;
        isShakeAtom[cluster.peripheralID[0]] = true;
        if (cluster.size > 1)
            isShakeAtom[cluster.peripheralID[1]] = true;
        if (cluster.size > 2)
            isShakeAtom[cluster.peripheralID[2]] = true;
        ++index;
    }
    psShakeID->Upload();
    psShakeParameter->Upload();
1111
    gpu->sim.shakeTolerance = constraintTolerance;
1112
1113
1114
1115
1116
1117
    gpu->sim.shake_threads_per_block     = (gpu->sim.ShakeConstraints + gpu->sim.blocks - 1) / gpu->sim.blocks;
    if (gpu->sim.shake_threads_per_block > gpu->sim.max_shake_threads_per_block)
        gpu->sim.shake_threads_per_block = gpu->sim.max_shake_threads_per_block;
    if (gpu->sim.shake_threads_per_block < 1)
        gpu->sim.shake_threads_per_block = 1;

1118
    // Find connected constraints for CCMA.
1119

1120
    vector<int> ccmaConstraints;
1121
    for (unsigned i = 0; i < atom1.size(); i++)
1122
        if (!isShakeAtom[atom1[i]])
1123
            ccmaConstraints.push_back(i);
1124
1125
1126

    // Record the connections between constraints.

1127
    int numCCMA = (int) ccmaConstraints.size();
1128
    vector<vector<int> > atomConstraints(gpu->natoms);
1129
1130
1131
    for (int i = 0; i < numCCMA; i++) {
        atomConstraints[atom1[ccmaConstraints[i]]].push_back(i);
        atomConstraints[atom2[ccmaConstraints[i]]].push_back(i);
1132
    }
1133
    vector<vector<int> > linkedConstraints(numCCMA);
1134
1135
1136
    for (unsigned atom = 0; atom < atomConstraints.size(); atom++) {
        for (unsigned i = 0; i < atomConstraints[atom].size(); i++)
            for (unsigned j = 0; j < i; j++) {
1137
1138
1139
1140
1141
1142
                int c1 = atomConstraints[atom][i];
                int c2 = atomConstraints[atom][j];
                linkedConstraints[c1].push_back(c2);
                linkedConstraints[c2].push_back(c1);
            }
    }
1143
    int maxLinks = 0;
1144
    for (unsigned i = 0; i < linkedConstraints.size(); i++)
1145
1146
        maxLinks = max(maxLinks, (int) linkedConstraints[i].size());
    int maxAtomConstraints = 0;
1147
    for (unsigned i = 0; i < atomConstraints.size(); i++)
1148
        maxAtomConstraints = max(maxAtomConstraints, (int) atomConstraints[i].size());
1149

1150
1151
1152
    // Compute the constraint coupling matrix

    vector<vector<int> > atomAngles(gpu->natoms);
1153
    for (int i = 0; i < (int) gpu->sim.bond_angles; i++)
1154
        atomAngles[(*gpu->psBondAngleID1)[i].y].push_back(i);
1155
1156
1157
1158
    vector<vector<pair<int, double> > > matrix(numCCMA);
    if (numCCMA > 0) {
        for (int j = 0; j < numCCMA; j++) {
            for (int k = 0; k < numCCMA; k++) {
1159
1160
1161
1162
1163
                if (j == k) {
                    matrix[j].push_back(pair<int, double>(j, 1.0));
                    continue;
                }
                double scale;
1164
1165
                int cj = ccmaConstraints[j];
                int ck = ccmaConstraints[k];
1166
1167
1168
1169
                int atomj0 = atom1[cj];
                int atomj1 = atom2[cj];
                int atomk0 = atom1[ck];
                int atomk1 = atom2[ck];
1170
1171
1172
1173
1174
                int atoma, atomb, atomc;
                if (atomj0 == atomk0) {
                    atoma = atomj1;
                    atomb = atomj0;
                    atomc = atomk1;
1175
                    scale = invMass1[cj]/(invMass1[cj]+invMass2[cj]);
1176
1177
1178
1179
1180
                }
                else if (atomj1 == atomk1) {
                    atoma = atomj0;
                    atomb = atomj1;
                    atomc = atomk0;
1181
                    scale = invMass2[cj]/(invMass1[cj]+invMass2[cj]);
1182
1183
1184
1185
1186
                }
                else if (atomj0 == atomk1) {
                    atoma = atomj1;
                    atomb = atomj0;
                    atomc = atomk0;
1187
                    scale = invMass1[cj]/(invMass1[cj]+invMass2[cj]);
1188
1189
1190
1191
1192
                }
                else if (atomj1 == atomk0) {
                    atoma = atomj0;
                    atomb = atomj1;
                    atomc = atomk1;
1193
                    scale = invMass2[cj]/(invMass1[cj]+invMass2[cj]);
1194
1195
1196
1197
1198
1199
1200
                }
                else
                    continue; // These constraints are not connected.

                // Look for a third constraint forming a triangle with these two.

                bool foundConstraint = false;
1201
                for (int other = 0; other < numCCMA; other++) {
1202
                    if ((atom1[other] == atoma && atom2[other] == atomc) || (atom1[other] == atomc && atom2[other] == atoma)) {
1203
1204
                        double d1 = distance[cj];
                        double d2 = distance[ck];
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
                        double d3 = distance[other];
                        matrix[j].push_back(pair<int, double>(k, scale*(d1*d1+d2*d2-d3*d3)/(2.0*d1*d2)));
                        foundConstraint = true;
                        break;
                    }
                }
                if (!foundConstraint) {
                    // We didn't find one, so look for an angle force field term.

                    const vector<int>& angleCandidates = atomAngles[atomb];
                    for (vector<int>::const_iterator iter = angleCandidates.begin(); iter != angleCandidates.end(); iter++) {
                        int4 atoms = (*gpu->psBondAngleID1)[*iter];
                        if ((atoms.x == atoma && atoms.z == atomc) || (atoms.z == atoma && atoms.x == atomc)) {
                            double angle = (*gpu->psBondAngleParameter)[*iter].x;
                            matrix[j].push_back(pair<int, double>(k, scale*cos(angle*PI/180.0)));
                            break;
                        }
                    }
                }
            }
        }

        // Invert it using QR.

        vector<int> matrixRowStart;
        vector<int> matrixColIndex;
        vector<double> matrixValue;
1232
        for (int i = 0; i < numCCMA; i++) {
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
            matrixRowStart.push_back(matrixValue.size());
            for (int j = 0; j < (int) matrix[i].size(); j++) {
                pair<int, double> element = matrix[i][j];
                matrixColIndex.push_back(element.first);
                matrixValue.push_back(element.second);
            }
        }
        matrixRowStart.push_back(matrixValue.size());
        int *qRowStart, *qColIndex, *rRowStart, *rColIndex;
        double *qValue, *rValue;
1243
        int result = QUERN_compute_qr(numCCMA, numCCMA, &matrixRowStart[0], &matrixColIndex[0], &matrixValue[0], NULL,
1244
                &qRowStart, &qColIndex, &qValue, &rRowStart, &rColIndex, &rValue);
1245
        vector<double> rhs(numCCMA);
1246
        matrix.clear();
1247
1248
        matrix.resize(numCCMA);
        for (int i = 0; i < numCCMA; i++) {
1249
1250
            // Extract column i of the inverse matrix.

1251
            for (int j = 0; j < numCCMA; j++)
1252
                rhs[j] = (i == j ? 1.0 : 0.0);
1253
1254
1255
1256
            result = QUERN_multiply_with_q_transpose(numCCMA, qRowStart, qColIndex, qValue, &rhs[0]);
            result = QUERN_solve_with_r(numCCMA, rRowStart, rColIndex, rValue, &rhs[0], &rhs[0]);
            for (int j = 0; j < numCCMA; j++) {
                double value = rhs[j]*distance[ccmaConstraints[i]]/distance[ccmaConstraints[j]];
1257
                if (abs(value) > 0.1)
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
                    matrix[j].push_back(pair<int, double>(i, value));
            }
        }
        QUERN_free_result(qRowStart, qColIndex, qValue);
        QUERN_free_result(rRowStart, rColIndex, rValue);
    }
    int maxRowElements = 0;
    for (unsigned i = 0; i < matrix.size(); i++)
        maxRowElements = max(maxRowElements, (int) matrix[i].size());
    maxRowElements++;

1269
    // Sort the constraints.
1270

1271
1272
    vector<int> constraintOrder(numCCMA);
    for (int i = 0; i < numCCMA; ++i)
1273
1274
        constraintOrder[i] = i;
    sort(constraintOrder.begin(), constraintOrder.end(), ConstraintOrderer(atom1, atom2));
1275
1276
    vector<int> inverseOrder(numCCMA);
    for (int i = 0; i < numCCMA; ++i)
1277
        inverseOrder[constraintOrder[i]] = i;
1278
1279
    for (int i = 0; i < (int)matrix.size(); ++i)
        for (int j = 0; j < (int)matrix[i].size(); ++j)
1280
            matrix[i][j].first = inverseOrder[matrix[i][j].first];
1281

1282
1283
    // Fill in the CUDA streams.

1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
    CUDAStream<int2>* psCcmaAtoms = new CUDAStream<int2>(numCCMA, 1, "CcmaAtoms");
    gpu->psCcmaAtoms              = psCcmaAtoms;
    gpu->sim.pCcmaAtoms           = psCcmaAtoms->_pDevData;
    CUDAStream<float4>* psCcmaDistance = new CUDAStream<float4>(numCCMA, 1, "CcmaDistance");
    gpu->psCcmaDistance                = psCcmaDistance;
    gpu->sim.pCcmaDistance             = psCcmaDistance->_pDevData;
    CUDAStream<int>* psCcmaAtomConstraints = new CUDAStream<int>(gpu->natoms*maxAtomConstraints, 1, "CcmaAtomConstraints");
    gpu->psCcmaAtomConstraints             = psCcmaAtomConstraints;
    gpu->sim.pCcmaAtomConstraints          = psCcmaAtomConstraints->_pDevData;
    CUDAStream<int>* psCcmaNumAtomConstraints = new CUDAStream<int>(gpu->natoms, 1, "CcmaAtomConstraintsIndex");
    gpu->psCcmaNumAtomConstraints             = psCcmaNumAtomConstraints;
    gpu->sim.pCcmaNumAtomConstraints          = psCcmaNumAtomConstraints->_pDevData;
    CUDAStream<float>* psCcmaDelta1 = new CUDAStream<float>(numCCMA, 1, "CcmaDelta1");
    gpu->psCcmaDelta1             = psCcmaDelta1;
    gpu->sim.pCcmaDelta1          = psCcmaDelta1->_pDevData;
    CUDAStream<float>* psCcmaDelta2 = new CUDAStream<float>(numCCMA, 1, "CcmaDelta2");
    gpu->psCcmaDelta2             = psCcmaDelta2;
    gpu->sim.pCcmaDelta2          = psCcmaDelta2->_pDevData;
1302
    CUDAStream<short>* psSyncCounter = new CUDAStream<short>(3*gpu->sim.blocks, 1, "SyncCounter");
1303
1304
    gpu->psSyncCounter               = psSyncCounter;
    gpu->sim.pSyncCounter            = psSyncCounter->_pDevData;
1305
1306
1307
    CUDAStream<unsigned int>* psRequiredIterations = new CUDAStream<unsigned int>(1, 1, "RequiredIterations");
    gpu->psRequiredIterations               = psRequiredIterations;
    gpu->sim.pRequiredIterations            = psRequiredIterations->_pDevData;
1308
1309
1310
1311
    CUDAStream<float>* psCcmaReducedMass = new CUDAStream<float>(numCCMA, 1, "CcmaReducedMass");
    gpu->psCcmaReducedMass             = psCcmaReducedMass;
    gpu->sim.pCcmaReducedMass          = psCcmaReducedMass->_pDevData;
    CUDAStream<unsigned int>* psConstraintMatrixColumn = new CUDAStream<unsigned int>(numCCMA*maxRowElements, 1, "ConstraintMatrixColumn");
1312
1313
    gpu->psConstraintMatrixColumn               = psConstraintMatrixColumn;
    gpu->sim.pConstraintMatrixColumn            = psConstraintMatrixColumn->_pDevData;
1314
    CUDAStream<float>* psConstraintMatrixValue = new CUDAStream<float>(numCCMA*maxRowElements, 1, "ConstraintMatrixValue");
1315
1316
    gpu->psConstraintMatrixValue             = psConstraintMatrixValue;
    gpu->sim.pConstraintMatrixValue          = psConstraintMatrixValue->_pDevData;
1317
1318
    gpu->sim.ccmaConstraints = numCCMA;
    for (int i = 0; i < numCCMA; i++) {
1319
        int index = constraintOrder[i];
1320
1321
1322
1323
1324
        int c = ccmaConstraints[index];
        (*psCcmaAtoms)[i].x = atom1[c];
        (*psCcmaAtoms)[i].y = atom2[c];
        (*psCcmaDistance)[i].w = distance[c];
        (*psCcmaReducedMass)[i] = 0.5f/(invMass1[c]+invMass2[c]);
1325
        for (unsigned int j = 0; j < matrix[index].size(); j++) {
1326
            (*psConstraintMatrixColumn)[i+j*numCCMA] = matrix[index][j].first;
1327
            (*psConstraintMatrixValue)[i+j*numCCMA] = (float) matrix[index][j].second;
1328
1329
1330
        }
        (*psConstraintMatrixColumn)[i+matrix[index].size()*numCCMA] = numCCMA;
    }
1331
    for (unsigned int i = 0; i < psSyncCounter->_length; i++)
1332
        (*psSyncCounter)[i] = -1;
1333
    for (unsigned int i = 0; i < atomConstraints.size(); i++) {
1334
        (*psCcmaNumAtomConstraints)[i] = atomConstraints[i].size();
1335
        for (unsigned int j = 0; j < atomConstraints[i].size(); j++) {
1336
1337
            bool forward = (atom1[ccmaConstraints[atomConstraints[i][j]]] == i);
            (*psCcmaAtomConstraints)[i+j*gpu->natoms] = (forward ? inverseOrder[atomConstraints[i][j]]+1 : -inverseOrder[atomConstraints[i][j]]-1);
1338
        }
1339
    }
1340
1341
1342
1343
1344
    psCcmaAtoms->Upload();
    psCcmaDistance->Upload();
    psCcmaReducedMass->Upload();
    psCcmaAtomConstraints->Upload();
    psCcmaNumAtomConstraints->Upload();
1345
    psSyncCounter->Upload();
1346
1347
    psConstraintMatrixColumn->Upload();
    psConstraintMatrixValue->Upload();
1348
1349
1350
1351
1352
    gpu->sim.ccma_threads_per_block = (gpu->sim.ccmaConstraints + gpu->sim.blocks - 1) / gpu->sim.blocks;
    if (gpu->sim.ccma_threads_per_block > gpu->sim.threads_per_block)
        gpu->sim.ccma_threads_per_block = gpu->sim.threads_per_block;
    if (gpu->sim.ccma_threads_per_block < gpu->sim.blocks)
        gpu->sim.ccma_threads_per_block = gpu->sim.blocks;
Peter Eastman's avatar
Peter Eastman committed
1353
1354
1355
1356
1357

    // count number of atoms w/o constraint

    int count = 0;
    for (int i = 0; i < gpu->natoms; i++)
1358
       if (!isShakeAtom[i])
Peter Eastman's avatar
Peter Eastman committed
1359
1360
1361
1362
1363
1364
1365
          count++;

    // Allocate NonShake parameters

    gpu->sim.NonShakeConstraints                  = count;
    if( count || true ){

1366
       CUDAStream<int>* psNonShakeID              = new CUDAStream<int>(count, 1, "NonShakeID");
Peter Eastman's avatar
Peter Eastman committed
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
1378
1379
1380
1381
       gpu->psNonShakeID                          = psNonShakeID;
       gpu->sim.pNonShakeID                       = psNonShakeID->_pDevStream[0];

       gpu->sim.nonshake_threads_per_block        = (count + gpu->sim.blocks - 1) / gpu->sim.blocks;

       if (gpu->sim.nonshake_threads_per_block > gpu->sim.max_shake_threads_per_block)
           gpu->sim.nonshake_threads_per_block = gpu->sim.max_shake_threads_per_block;

       if (gpu->sim.nonshake_threads_per_block < 1)
               gpu->sim.nonshake_threads_per_block = 1;

       // load indices

       count = 0;
       for (int i = 0; i < gpu->natoms; i++){
1382
          if (!isShakeAtom[i]){
1383
             (*psNonShakeID)[count++] = i;
1384
          }
Peter Eastman's avatar
Peter Eastman committed
1385
1386
1387
1388
1389
1390
1391
1392
1393
1394
1395
1396
1397
1398
1399
1400
       }
       psNonShakeID->Upload();

    } else {
       gpu->sim.nonshake_threads_per_block           = 0;
    }
}

extern "C"
int gpuAllocateInitialBuffers(gpuContext gpu)
{
    gpu->sim.atoms                      = gpu->natoms;
    gpu->sim.paddedNumberOfAtoms        = ((gpu->sim.atoms + GRID - 1) >> GRIDBITS) << GRIDBITS;
    gpu->sim.degreesOfFreedom           = 3 * gpu->sim.atoms - 6;
    gpu->gpAtomTable                    = NULL;
    gpu->gAtomTypes                     = 0;
1401
    gpu->psPosq4                        = new CUDAStream<float4>(gpu->sim.paddedNumberOfAtoms, 1, "Posq");
Peter Eastman's avatar
Peter Eastman committed
1402
1403
1404
1405
1406
1407
1408
1409
1410
    gpu->sim.stride                     = gpu->psPosq4->_stride;
    gpu->sim.stride2                    = gpu->sim.stride * 2;
    gpu->sim.stride3                    = gpu->sim.stride * 3;
    gpu->sim.stride4                    = gpu->sim.stride * 4;
    gpu->sim.pPosq                      = gpu->psPosq4->_pDevStream[0];
    gpu->sim.stride                     = gpu->psPosq4->_stride;
    gpu->sim.stride2                    = 2 * gpu->sim.stride;
    gpu->sim.stride3                    = 3 * gpu->sim.stride;
    gpu->sim.stride4                    = 4 * gpu->sim.stride;
1411
    gpu->psPosqP4                       = new CUDAStream<float4>(gpu->sim.paddedNumberOfAtoms, 1, "PosqP");
Peter Eastman's avatar
Peter Eastman committed
1412
    gpu->sim.pPosqP                     = gpu->psPosqP4->_pDevStream[0];
1413
    gpu->psOldPosq4                     = new CUDAStream<float4>(gpu->sim.paddedNumberOfAtoms, 1, "OldPosq");
Peter Eastman's avatar
Peter Eastman committed
1414
    gpu->sim.pOldPosq                   = gpu->psOldPosq4->_pDevStream[0];
1415
    gpu->psVelm4                        = new CUDAStream<float4>(gpu->sim.paddedNumberOfAtoms, 1, "Velm");
Peter Eastman's avatar
Peter Eastman committed
1416
    gpu->sim.pVelm4                     = gpu->psVelm4->_pDevStream[0];
1417
    gpu->psvVector4                     = new CUDAStream<float4>(gpu->sim.paddedNumberOfAtoms, 1, "vVector");
Peter Eastman's avatar
Peter Eastman committed
1418
    gpu->sim.pvVector4                  = gpu->psvVector4->_pDevStream[0];
1419
    gpu->psxVector4                     = new CUDAStream<float4>(gpu->sim.paddedNumberOfAtoms, 1, "xVector");
Peter Eastman's avatar
Peter Eastman committed
1420
    gpu->sim.pxVector4                  = gpu->psxVector4->_pDevStream[0];
1421
    gpu->psBornRadii                    = new CUDAStream<float>(gpu->sim.paddedNumberOfAtoms, 1, "BornRadii");
Peter Eastman's avatar
Peter Eastman committed
1422
    gpu->sim.pBornRadii                 = gpu->psBornRadii->_pDevStream[0];
1423
    gpu->psObcChain                     = new CUDAStream<float>(gpu->sim.paddedNumberOfAtoms, 1, "ObcChain");
Peter Eastman's avatar
Peter Eastman committed
1424
    gpu->sim.pObcChain                  = gpu->psObcChain->_pDevStream[0];
1425
    gpu->psSigEps2                      = new CUDAStream<float2>(gpu->sim.paddedNumberOfAtoms, 1, "SigEps2");
Peter Eastman's avatar
Peter Eastman committed
1426
    gpu->sim.pAttr                      = gpu->psSigEps2->_pDevStream[0];
1427
    gpu->psObcData                      = new CUDAStream<float2>(gpu->sim.paddedNumberOfAtoms, 1, "ObcData");
Peter Eastman's avatar
Peter Eastman committed
1428
    gpu->sim.pObcData                   = gpu->psObcData->_pDevStream[0];
1429
1430
1431
1432
    gpu->psStepSize                     = new CUDAStream<float2>(1, 1, "StepSize");
    gpu->sim.pStepSize                  = gpu->psStepSize->_pDevStream[0];
    (*gpu->psStepSize)[0] = make_float2(0.0f, 0.0f);
    gpu->psStepSize->Upload();
1433
1434
    gpu->psLangevinParameters           = new CUDAStream<float>(11, 1, "LangevinParameters");
    gpu->sim.pLangevinParameters        = gpu->psLangevinParameters->_pDevStream[0];
Peter Eastman's avatar
Peter Eastman committed
1435
    gpu->pAtomSymbol                    = new unsigned char[gpu->natoms];
1436
    gpu->psAtomIndex                    = new CUDAStream<int>(gpu->sim.paddedNumberOfAtoms, 1, "AtomIndex");
1437
1438
    gpu->sim.pAtomIndex                 = gpu->psAtomIndex->_pDevStream[0];
    for (int i = 0; i < (int) gpu->sim.paddedNumberOfAtoms; i++)
1439
        (*gpu->psAtomIndex)[i] = i;
1440
    gpu->psAtomIndex->Upload();
1441
    gpu->posCellOffsets.resize(gpu->natoms, make_int3(0, 0, 0));
Peter Eastman's avatar
Peter Eastman committed
1442
    // Determine randoms
1443
    gpu->seed                           = 1;
1444
    gpu->sim.randomFrames               = 20;
Peter Eastman's avatar
Peter Eastman committed
1445
    gpu->sim.randomIterations           = gpu->sim.randomFrames;
1446
    gpu->sim.randoms                    = gpu->sim.randomFrames * gpu->sim.paddedNumberOfAtoms;
Peter Eastman's avatar
Peter Eastman committed
1447
1448
    gpu->sim.totalRandoms               = gpu->sim.randoms + gpu->sim.paddedNumberOfAtoms;
    gpu->sim.totalRandomsTimesTwo       = gpu->sim.totalRandoms * 2;
1449
1450
1451
1452
    gpu->psRandom4                      = new CUDAStream<float4>(gpu->sim.totalRandomsTimesTwo, 1, "Random4");
    gpu->psRandom2                      = new CUDAStream<float2>(gpu->sim.totalRandomsTimesTwo, 1, "Random2");
    gpu->psRandomPosition               = new CUDAStream<int>(gpu->sim.blocks, 1, "RandomPosition");
    gpu->psRandomSeed                   = new CUDAStream<uint4>(gpu->sim.blocks * gpu->sim.random_threads_per_block, 1, "RandomSeed");
Peter Eastman's avatar
Peter Eastman committed
1453
1454
1455
1456
1457
1458
1459
1460
    gpu->sim.pRandom4a                  = gpu->psRandom4->_pDevStream[0];
    gpu->sim.pRandom2a                  = gpu->psRandom2->_pDevStream[0];
    gpu->sim.pRandom4b                  = gpu->psRandom4->_pDevStream[0] + gpu->sim.totalRandoms;
    gpu->sim.pRandom2b                  = gpu->psRandom2->_pDevStream[0] + gpu->sim.totalRandoms;
    gpu->sim.pRandomPosition            = gpu->psRandomPosition->_pDevStream[0];
    gpu->sim.pRandomSeed                = gpu->psRandomSeed->_pDevStream[0];

    // Allocate and clear linear momentum buffer
1461
    gpu->psLinearMomentum = new CUDAStream<float4>(gpu->sim.blocks, 1, "LinearMomentum");
Peter Eastman's avatar
Peter Eastman committed
1462
1463
1464
    gpu->sim.pLinearMomentum = gpu->psLinearMomentum->_pDevStream[0];
    for (int i = 0; i < (int) gpu->sim.blocks; i++)
    {
1465
1466
1467
1468
        (*gpu->psLinearMomentum)[i].x = 0.0f;
        (*gpu->psLinearMomentum)[i].y = 0.0f;
        (*gpu->psLinearMomentum)[i].z = 0.0f;
        (*gpu->psLinearMomentum)[i].w = 0.0f;
Peter Eastman's avatar
Peter Eastman committed
1469
1470
1471
1472
1473
1474
1475
1476
1477
1478
1479
    }
    gpu->psLinearMomentum->Upload();

    return 1;
}

extern "C"
void gpuSetPositions(gpuContext gpu, const vector<float>& x, const vector<float>& y, const vector<float>& z)
{
    for (int i = 0; i < gpu->natoms; i++)
    {
1480
1481
1482
        (*gpu->psPosq4)[i].x = x[i];
        (*gpu->psPosq4)[i].y = y[i];
        (*gpu->psPosq4)[i].z = z[i];
Peter Eastman's avatar
Peter Eastman committed
1483
1484
1485
1486
1487
1488
1489
1490
1491
1492
1493
1494
1495
    }
    gpu->psPosq4->Upload();

	 // set flag to recalculate Born radii

	 gpu->bRecalculateBornRadii = true;
} 

extern "C"
void gpuSetVelocities(gpuContext gpu, const vector<float>& x, const vector<float>& y, const vector<float>& z)
{
    for (int i = 0; i < gpu->natoms; i++)
    {
1496
1497
1498
        (*gpu->psVelm4)[i].x = x[i];
        (*gpu->psVelm4)[i].y = y[i];
        (*gpu->psVelm4)[i].z = z[i];
Peter Eastman's avatar
Peter Eastman committed
1499
1500
1501
1502
1503
1504
1505
1506
1507
1508
    }
    gpu->psVelm4->Upload();
} 

extern "C"
void gpuSetMass(gpuContext gpu, const vector<float>& mass)
{
    float totalMass = 0.0f;
    for (int i = 0; i < gpu->natoms; i++)
    {
1509
        (*gpu->psVelm4)[i].w = 1.0f/mass[i];
Peter Eastman's avatar
Peter Eastman committed
1510
1511
1512
1513
1514
1515
1516
1517
1518
1519
1520
        totalMass += mass[i];
    }
    gpu->sim.inverseTotalMass = 1.0f / totalMass;
    gpu->psVelm4->Upload();
} 

extern "C"
void gpuInitializeRandoms(gpuContext gpu)
{
    for (int i = 0; i < (int) gpu->sim.blocks; i++)
    {
1521
        (*gpu->psRandomPosition)[i] = 0;
Peter Eastman's avatar
Peter Eastman committed
1522
1523
1524
1525
1526
    }
    int seed = gpu->seed | ((gpu->seed ^ 0xffffffff) << 16);
    srand(seed);
    for (int i = 0; i < (int) (gpu->sim.blocks * gpu->sim.random_threads_per_block); i++)
    {
1527
1528
1529
1530
        (*gpu->psRandomSeed)[i].x = rand();
        (*gpu->psRandomSeed)[i].y = rand();
        (*gpu->psRandomSeed)[i].z = rand();
        (*gpu->psRandomSeed)[i].w = rand();
Peter Eastman's avatar
Peter Eastman committed
1531
1532
1533
1534
1535
1536
1537
1538
1539
1540
1541
1542
1543
1544
1545
1546
1547
    }
    gpu->psRandomPosition->Upload();
    gpu->psRandomSeed->Upload();
    gpuSetConstants(gpu);
    kGenerateRandoms(gpu);
    return;
}

extern "C"
bool gpuIsAvailable()
{
    int deviceCount;
    cudaGetDeviceCount(&deviceCount);
    return (deviceCount > 0);
}

extern "C"
1548
void* gpuInit(int numAtoms, unsigned int device, bool useBlockingSync)
Peter Eastman's avatar
Peter Eastman committed
1549
1550
1551
1552
1553
1554
1555
{
    gpuContext gpu = new _gpuContext;
    int LRFSize = 0;
    int SMCount = 0;
    int SMMajor = 0;
    int SMMinor = 0;

1556
    // Select which device to use
1557
1558
1559
1560
1561
1562
    int currentDevice;
    cudaError_t status = cudaGetDevice(&currentDevice);
    RTERROR(status, "Error getting CUDA device")
    if (device != currentDevice)
        cudaSetDevice(device); // Ignore errors
    status = cudaGetDevice(&gpu->device);
1563
    RTERROR(status, "Error getting CUDA device")
1564
1565
1566
    status = cudaSetDeviceFlags(useBlockingSync ? cudaDeviceBlockingSync : cudaDeviceScheduleAuto);
    RTERROR(status, "Error setting device flags")
    gpu->useBlockingSync = useBlockingSync;
Peter Eastman's avatar
Peter Eastman committed
1567
1568
1569

    // Determine kernel call configuration
    cudaDeviceProp deviceProp;
Peter Eastman's avatar
Peter Eastman committed
1570
    cudaGetDeviceProperties(&deviceProp, currentDevice);
Peter Eastman's avatar
Peter Eastman committed
1571
1572
1573
1574
1575
1576
1577
1578
1579
1580
1581
1582
1583
1584
1585
1586
1587
1588
1589
1590
1591
1592
1593
1594
1595
1596
1597
1598
1599
1600
1601
1602
1603
1604
1605
1606
1607
1608
1609
1610
1611
1612
1613
1614
1615
1616
1617
1618
1619

    // Determine SM version
    if (deviceProp.major == 1)
    {
        switch (deviceProp.minor)
        {
        case 0:
        case 1:
            gpu->sm_version = SM_10;
            gpu->sim.workUnitsPerSM = G8X_NONBOND_WORKUNITS_PER_SM;
            break;

        default:
            gpu->sm_version = SM_12;
            gpu->sim.workUnitsPerSM = GT2XX_NONBOND_WORKUNITS_PER_SM;
            break;
        }
    }

    gpu->sim.nonbond_blocks = deviceProp.multiProcessorCount;
    gpu->sim.bornForce2_blocks = deviceProp.multiProcessorCount;
    gpu->sim.blocks = deviceProp.multiProcessorCount;
    if (deviceProp.regsPerBlock == 8192)
    {
        gpu->sim.nonbond_threads_per_block          = G8X_NONBOND_THREADS_PER_BLOCK;
        gpu->sim.bornForce2_threads_per_block       = G8X_BORNFORCE2_THREADS_PER_BLOCK;
        gpu->sim.max_shake_threads_per_block        = G8X_SHAKE_THREADS_PER_BLOCK;
        gpu->sim.max_update_threads_per_block       = G8X_UPDATE_THREADS_PER_BLOCK;
        gpu->sim.max_localForces_threads_per_block  = G8X_LOCALFORCES_THREADS_PER_BLOCK;
        gpu->sim.threads_per_block                  = G8X_THREADS_PER_BLOCK;
        gpu->sim.random_threads_per_block           = G8X_RANDOM_THREADS_PER_BLOCK;
    }
    else
    {
        gpu->sim.nonbond_threads_per_block          = GT2XX_NONBOND_THREADS_PER_BLOCK;
        gpu->sim.bornForce2_threads_per_block       = GT2XX_BORNFORCE2_THREADS_PER_BLOCK;
        gpu->sim.max_shake_threads_per_block        = GT2XX_SHAKE_THREADS_PER_BLOCK;
        gpu->sim.max_update_threads_per_block       = GT2XX_UPDATE_THREADS_PER_BLOCK;
        gpu->sim.max_localForces_threads_per_block  = GT2XX_LOCALFORCES_THREADS_PER_BLOCK;
        gpu->sim.threads_per_block                  = GT2XX_NONBOND_THREADS_PER_BLOCK;
        gpu->sim.random_threads_per_block           = GT2XX_RANDOM_THREADS_PER_BLOCK;
    }
    gpu->sim.shake_threads_per_block                = gpu->sim.max_shake_threads_per_block;
    gpu->sim.localForces_threads_per_block          = gpu->sim.max_localForces_threads_per_block;

    gpu->natoms = numAtoms;
    gpuAllocateInitialBuffers(gpu);
    for (int i = 0; i < gpu->natoms; i++)
    {
1620
1621
1622
1623
        (*gpu->psxVector4)[i].x = 0.0f;
        (*gpu->psxVector4)[i].y = 0.0f;
        (*gpu->psxVector4)[i].z = 0.0f;
        (*gpu->psxVector4)[i].w = 0.0f;
Peter Eastman's avatar
Peter Eastman committed
1624
1625
1626
1627
1628
1629
1630
    }
    gpu->psxVector4->Upload();

    gpu->iterations = 0;
    gpu->sim.update_threads_per_block               = (gpu->natoms + gpu->sim.blocks - 1) / gpu->sim.blocks;
    if (gpu->sim.update_threads_per_block > gpu->sim.max_update_threads_per_block)
        gpu->sim.update_threads_per_block = gpu->sim.max_update_threads_per_block;
1631
1632
    if (gpu->sim.update_threads_per_block < gpu->psLangevinParameters->_length)
            gpu->sim.update_threads_per_block = gpu->psLangevinParameters->_length;
Peter Eastman's avatar
Peter Eastman committed
1633
1634
1635
1636
1637
1638
1639
1640
1641
1642
1643
1644
    gpu->sim.bf_reduce_threads_per_block = gpu->sim.update_threads_per_block;
    gpu->sim.bsf_reduce_threads_per_block = (gpu->sim.stride4 + gpu->natoms + gpu->sim.blocks - 1) / gpu->sim.blocks;
    gpu->sim.bsf_reduce_threads_per_block = ((gpu->sim.bsf_reduce_threads_per_block + (GRID - 1)) / GRID) * GRID;
    if (gpu->sim.bsf_reduce_threads_per_block > gpu->sim.threads_per_block)
        gpu->sim.bsf_reduce_threads_per_block = gpu->sim.threads_per_block;
    if (gpu->sim.bsf_reduce_threads_per_block < 1)
        gpu->sim.bsf_reduce_threads_per_block = 1;

    // Initialize constants to reasonable values
    gpu->sim.probeRadius            = probeRadius;
    gpu->sim.surfaceAreaFactor      = surfaceAreaFactor;
    gpu->sim.electricConstant       = electricConstant;
1645
    gpu->sim.nonbondedMethod        = NO_CUTOFF;
1646
    gpu->sim.nonbondedCutoff        = 0.0f;
1647
    gpu->sim.nonbondedCutoffSqr     = 0.0f;
Peter Eastman's avatar
Peter Eastman committed
1648
1649
1650
1651
1652
1653
1654
1655

    gpu->sim.bigFloat               = 99999999.0f;
    gpu->sim.forceConversionFactor  = forceConversionFactor;
    gpu->sim.preFactor              = 2.0f*electricConstant*((1.0f/defaultInnerDielectric)-(1.0f/defaultSolventDielectric))*gpu->sim.forceConversionFactor;
    gpu->sim.dielectricOffset       = dielectricOffset;
    gpu->sim.alphaOBC               = alphaOBC;
    gpu->sim.betaOBC                = betaOBC;
    gpu->sim.gammaOBC               = gammaOBC;
1656
    gpuSetLangevinIntegrationParameters(gpu, 1.0f, 2.0e-3f, 300.0f, 0.0f);
Peter Eastman's avatar
Peter Eastman committed
1657
1658
1659
1660
1661
1662
1663
    gpu->sim.maxShakeIterations     = 15;
    gpu->sim.shakeTolerance         = 1.0e-04f * 2.0f;
    gpu->sim.InvMassJ               = 9.920635e-001f;
    gpu->grid                       = GRID;
    gpu->bCalculateCM               = false;
    gpu->bRemoveCM                  = false;
    gpu->bRecalculateBornRadii      = true;
1664
    gpu->bIncludeGBSA               = false;
Peter Eastman's avatar
Peter Eastman committed
1665
1666
1667
1668
1669
    gpuInitializeRandoms(gpu);

    // To be determined later
    gpu->psLJ14ID                   = NULL;
    gpu->psForce4                   = NULL;
1670
    gpu->psEnergy                   = NULL;
Peter Eastman's avatar
Peter Eastman committed
1671
1672
1673
1674
1675
1676
1677
1678
1679
1680
1681
1682
1683
1684
1685
1686
1687
1688
1689
1690
1691
    gpu->sim.pForce4                = NULL;
    gpu->sim.pForce4a               = NULL;
    gpu->sim.pForce4b               = NULL;
    gpu->psBornForce                = NULL;
    gpu->sim.pBornForce             = NULL;
    gpu->psBornSum                  = NULL;
    gpu->sim.pBornSum               = NULL;
    gpu->psBondID                   = NULL;
    gpu->psBondParameter            = NULL;
    gpu->psBondAngleID1             = NULL;
    gpu->psBondAngleID2             = NULL;
    gpu->psBondAngleParameter       = NULL;
    gpu->psDihedralID1              = NULL;
    gpu->psDihedralID2              = NULL;
    gpu->psDihedralParameter        = NULL;
    gpu->psRbDihedralID1            = NULL;
    gpu->psRbDihedralID2            = NULL;
    gpu->psRbDihedralParameter1     = NULL;
    gpu->psRbDihedralParameter2     = NULL;
    gpu->psLJ14ID                   = NULL;
    gpu->psLJ14Parameter            = NULL;
Peter Eastman's avatar
Peter Eastman committed
1692
1693
1694
    gpu->psCustomParams             = NULL;
    gpu->psCustomExceptionID        = NULL;
    gpu->psCustomExceptionParams    = NULL;
1695
    gpu->psEwaldCosSinSum           = NULL;
Peter Eastman's avatar
Peter Eastman committed
1696
    gpu->psTabulatedErfc            = NULL;
1697
1698
1699
1700
1701
1702
    gpu->psPmeGrid                  = NULL;
    gpu->psPmeBsplineModuli[0]      = NULL;
    gpu->psPmeBsplineModuli[1]      = NULL;
    gpu->psPmeBsplineModuli[2]      = NULL;
    gpu->psPmeBsplineTheta          = NULL;
    gpu->psPmeBsplineDtheta         = NULL;
1703
1704
    gpu->psPmeAtomRange             = NULL;
    gpu->psPmeAtomGridIndex         = NULL;
Peter Eastman's avatar
Peter Eastman committed
1705
1706
    gpu->psShakeID                  = NULL;
    gpu->psShakeParameter           = NULL;
1707
1708
    gpu->psSettleID                 = NULL;
    gpu->psSettleParameter          = NULL;
Peter Eastman's avatar
Peter Eastman committed
1709
    gpu->psExclusion                = NULL;
1710
    gpu->psExclusionIndex           = NULL;
Peter Eastman's avatar
Peter Eastman committed
1711
    gpu->psWorkUnit                 = NULL;
1712
1713
1714
1715
1716
    gpu->psInteractingWorkUnit      = NULL;
    gpu->psInteractionFlag          = NULL;
    gpu->psInteractionCount         = NULL;
    gpu->psGridBoundingBox          = NULL;
    gpu->psGridCenter               = NULL;
1717
1718
1719
1720
1721
1722
    gpu->psCcmaAtoms                = NULL;
    gpu->psCcmaDistance             = NULL;
    gpu->psCcmaAtomConstraints      = NULL;
    gpu->psCcmaNumAtomConstraints   = NULL;
    gpu->psCcmaDelta1               = NULL;
    gpu->psCcmaDelta2               = NULL;
1723
    gpu->psSyncCounter              = NULL;
1724
    gpu->psRequiredIterations       = NULL;
1725
    gpu->psCcmaReducedMass          = NULL;
1726
1727
    gpu->psConstraintMatrixColumn   = NULL;
    gpu->psConstraintMatrixValue    = NULL;
1728
1729
1730
    gpu->psTabulatedFunctionParams  = NULL;
    for (int i = 0; i < MAX_TABULATED_FUNCTIONS; i++)
        gpu->tabulatedFunctions[i].coefficients = NULL;
Peter Eastman's avatar
Peter Eastman committed
1731
1732
1733
1734
1735
1736
1737
1738
1739

    // Initialize output buffer before reading parameters
    gpu->pOutputBufferCounter       = new unsigned int[gpu->sim.paddedNumberOfAtoms];
    memset(gpu->pOutputBufferCounter, 0, gpu->sim.paddedNumberOfAtoms * sizeof(unsigned int));

    return (void*)gpu;
}

extern "C"
1740
void gpuSetLangevinIntegrationParameters(gpuContext gpu, float tau, float deltaT, float temperature, float errorTol) {
Peter Eastman's avatar
Peter Eastman committed
1741
1742
    gpu->sim.deltaT                 = deltaT;
    gpu->sim.oneOverDeltaT          = 1.0f/deltaT;
1743
    gpu->sim.errorTol               = errorTol;
Peter Eastman's avatar
Peter Eastman committed
1744
    gpu->sim.tau                    = tau;
1745
1746
    gpu->sim.T                      = temperature;
    gpu->sim.kT                     = BOLTZ * gpu->sim.T;
1747
1748
1749
1750
1751
1752
1753
    double GDT                       = gpu->sim.deltaT / gpu->sim.tau;
    double EPH                       = exp(0.5 * GDT);
    double EMH                       = exp(-0.5 * GDT);
    double EP                        = exp(GDT);
    double EM                        = exp(-GDT);
    double B, C, D;
    if (GDT >= 0.1)
Peter Eastman's avatar
Peter Eastman committed
1754
    {
1755
        double term1 = EPH - 1.0;
Peter Eastman's avatar
Peter Eastman committed
1756
        term1                      *= term1;
1757
1758
1759
        B                           = GDT * (EP - 1.0) - 4.0 * term1;
        C                           = GDT - 3.0 + 4.0 * EMH - EM;
        D                           = 2.0 - EPH - EMH;
Peter Eastman's avatar
Peter Eastman committed
1760
1761
1762
    }
    else
    {
1763
1764
1765
1766
1767
1768
1769
1770
1771
1772
1773
        double term1                 = 0.5 * GDT;
        double term2                 = term1 * term1;
        double term4                 = term2 * term2;

        double third                 = 1.0 / 3.0;
        double o7_9                  = 7.0 / 9.0;
        double o1_12                 = 1.0 / 12.0;
        double o17_90                = 17.0 / 90.0;
        double o7_30                 = 7.0 / 30.0;
        double o31_1260              = 31.0 / 1260.0;
        double o_360                 = 1.0 / 360.0;
Peter Eastman's avatar
Peter Eastman committed
1774

1775
        B                           = term4 * (third + term1 * (third + term1 * (o17_90 + term1 * o7_9)));
1776
1777
1778
1779
1780
1781
1782
1783
        C                           = term2 * term1 * (2.0 * third + term1 * (-0.5 + term1 * (o7_30 + term1 * (-o1_12 + term1 * o31_1260))));
        D                           = term2 * (-1.0 + term2 * (-o1_12 - term2 * o_360));
    }
    double DOverTauC                 = D / (gpu->sim.tau * C);
    double TauOneMinusEM             = gpu->sim.tau * (1.0-EM);
    double TauDOverEMMinusOne        = gpu->sim.tau * D / (EM - 1.0);
    double fix1                      = gpu->sim.tau * (EPH - EMH);
    if (fix1 == 0.0)
1784
        fix1 = deltaT;
1785
1786
1787
1788
1789
    double oneOverFix1               = 1.0 / fix1;
    double V                         = sqrt(gpu->sim.kT * (1.0 - EM));
    double X                         = gpu->sim.tau * sqrt(gpu->sim.kT * C);
    double Yv                        = sqrt(gpu->sim.kT * B / C);
    double Yx                        = gpu->sim.tau * sqrt(gpu->sim.kT * B / (1.0 - EM));
1790
1791
1792
1793
1794
1795
1796
1797
1798
1799
1800
    (*gpu->psLangevinParameters)[0] = (float) EM;
    (*gpu->psLangevinParameters)[1] = (float) EM;
    (*gpu->psLangevinParameters)[2] = (float) DOverTauC;
    (*gpu->psLangevinParameters)[3] = (float) TauOneMinusEM;
    (*gpu->psLangevinParameters)[4] = (float) TauDOverEMMinusOne;
    (*gpu->psLangevinParameters)[5] = (float) V;
    (*gpu->psLangevinParameters)[6] = (float) X;
    (*gpu->psLangevinParameters)[7] = (float) Yv;
    (*gpu->psLangevinParameters)[8] = (float) Yx;
    (*gpu->psLangevinParameters)[9] = (float) fix1;
    (*gpu->psLangevinParameters)[10] = (float) oneOverFix1;
1801
    gpu->psLangevinParameters->Upload();
1802
1803
1804
    gpu->psStepSize->Download();
    (*gpu->psStepSize)[0].y = deltaT;
    gpu->psStepSize->Upload();
Peter Eastman's avatar
Peter Eastman committed
1805
1806
1807
}

extern "C"
1808
void gpuSetVerletIntegrationParameters(gpuContext gpu, float deltaT, float errorTol) {
Peter Eastman's avatar
Peter Eastman committed
1809
1810
    gpu->sim.deltaT                 = deltaT;
    gpu->sim.oneOverDeltaT          = 1.0f/deltaT;
1811
1812
1813
1814
    gpu->sim.errorTol               = errorTol;
    gpu->psStepSize->Download();
    (*gpu->psStepSize)[0].y = deltaT;
    gpu->psStepSize->Upload();
Peter Eastman's avatar
Peter Eastman committed
1815
1816
1817
1818
1819
1820
1821
}

extern "C"
void gpuSetBrownianIntegrationParameters(gpuContext gpu, float tau, float deltaT, float temperature) {
    gpu->sim.deltaT                 = deltaT;
    gpu->sim.oneOverDeltaT          = 1.0f/deltaT;
    gpu->sim.tau                    = tau;
1822
    gpu->sim.tauDeltaT              = gpu->sim.deltaT * gpu->sim.tau;
Peter Eastman's avatar
Peter Eastman committed
1823
1824
    gpu->sim.T                      = temperature;
    gpu->sim.kT                     = BOLTZ * gpu->sim.T;
1825
    gpu->sim.noiseAmplitude         = sqrt(2.0f*gpu->sim.kT*deltaT*tau);
1826
1827
1828
    gpu->psStepSize->Download();
    (*gpu->psStepSize)[0].y = deltaT;
    gpu->psStepSize->Upload();
Peter Eastman's avatar
Peter Eastman committed
1829
1830
1831
}

extern "C"
1832
void gpuSetAndersenThermostatParameters(gpuContext gpu, float temperature, float collisionFrequency) {
Peter Eastman's avatar
Peter Eastman committed
1833
1834
    gpu->sim.T                      = temperature;
    gpu->sim.kT                     = BOLTZ * gpu->sim.T;
1835
    gpu->sim.collisionFrequency     = collisionFrequency;
Peter Eastman's avatar
Peter Eastman committed
1836
1837
1838
1839
1840
1841
1842
1843
1844
1845
1846
1847
1848
1849
1850
1851
}

extern "C"
void gpuShutDown(gpuContext gpu)
{
    // Delete sysmem pointers
    delete[] gpu->pOutputBufferCounter;
    delete[] gpu->gpAtomTable;
    delete[] gpu->pAtomSymbol;

    // Delete device pointers
    delete gpu->psPosq4;
    delete gpu->psPosqP4;
    delete gpu->psOldPosq4;
    delete gpu->psVelm4;
    delete gpu->psForce4;
1852
    delete gpu->psEnergy;
Peter Eastman's avatar
Peter Eastman committed
1853
1854
    delete gpu->psxVector4;
    delete gpu->psvVector4;
1855
    delete gpu->psSigEps2;
1856
1857
1858
1859
1860
    if (gpu->psCustomParams != NULL) {
        delete gpu->psCustomParams;
        delete gpu->psCustomExceptionID;
        delete gpu->psCustomExceptionParams;
    }
1861
    if (gpu->psEwaldCosSinSum != NULL)
1862
        delete gpu->psEwaldCosSinSum;
1863
1864
1865
1866
1867
1868
1869
    if (gpu->psPmeGrid != NULL) {
        delete gpu->psPmeGrid;
        delete gpu->psPmeBsplineModuli[0];
        delete gpu->psPmeBsplineModuli[1];
        delete gpu->psPmeBsplineModuli[2];
        delete gpu->psPmeBsplineTheta;
        delete gpu->psPmeBsplineDtheta;
1870
1871
        delete gpu->psPmeAtomRange;
        delete gpu->psPmeAtomGridIndex;
1872
1873
        cufftDestroy(gpu->fftplan);
    }
Peter Eastman's avatar
Peter Eastman committed
1874
1875
    if (gpu->psTabulatedErfc != NULL)
        delete gpu->psTabulatedErfc;
1876
    delete gpu->psObcData;
Peter Eastman's avatar
Peter Eastman committed
1877
1878
1879
1880
1881
1882
1883
1884
1885
1886
1887
1888
1889
1890
1891
1892
1893
1894
1895
1896
    delete gpu->psObcChain;
    delete gpu->psBornForce;
    delete gpu->psBornRadii;
    delete gpu->psBornSum;
    delete gpu->psBondID;
    delete gpu->psBondParameter;
    delete gpu->psBondAngleID1;
    delete gpu->psBondAngleID2;
    delete gpu->psBondAngleParameter;
    delete gpu->psDihedralID1;
    delete gpu->psDihedralID2;
    delete gpu->psDihedralParameter;
    delete gpu->psRbDihedralID1;
    delete gpu->psRbDihedralID2;
    delete gpu->psRbDihedralParameter1;
    delete gpu->psRbDihedralParameter2;
    delete gpu->psLJ14ID;
    delete gpu->psLJ14Parameter;
    delete gpu->psShakeID;
    delete gpu->psShakeParameter;
1897
1898
    delete gpu->psSettleID;
    delete gpu->psSettleParameter;
Peter Eastman's avatar
Peter Eastman committed
1899
    delete gpu->psNonShakeID;
Peter Eastman's avatar
Peter Eastman committed
1900
    delete gpu->psExclusion;
1901
    delete gpu->psExclusionIndex;
Peter Eastman's avatar
Peter Eastman committed
1902
    delete gpu->psWorkUnit;
1903
1904
1905
    delete gpu->psInteractingWorkUnit;
    delete gpu->psInteractionFlag;
    delete gpu->psInteractionCount;
1906
1907
    delete gpu->psStepSize;
    delete gpu->psLangevinParameters;
Peter Eastman's avatar
Peter Eastman committed
1908
1909
1910
1911
1912
    delete gpu->psRandom4;
    delete gpu->psRandom2;
    delete gpu->psRandomPosition;    
    delete gpu->psRandomSeed;
    delete gpu->psLinearMomentum;
1913
1914
1915
    delete gpu->psAtomIndex;
    delete gpu->psGridBoundingBox;
    delete gpu->psGridCenter;
1916
1917
1918
1919
1920
1921
    delete gpu->psCcmaAtoms;
    delete gpu->psCcmaDistance;
    delete gpu->psCcmaAtomConstraints;
    delete gpu->psCcmaNumAtomConstraints;
    delete gpu->psCcmaDelta1;
    delete gpu->psCcmaDelta2;
1922
    delete gpu->psSyncCounter;
1923
    delete gpu->psRequiredIterations;
1924
    delete gpu->psCcmaReducedMass;
1925
1926
    delete gpu->psConstraintMatrixColumn;
    delete gpu->psConstraintMatrixValue;
1927
1928
1929
1930
    delete gpu->psTabulatedFunctionParams;
    for (int i = 0; i < MAX_TABULATED_FUNCTIONS; i++)
        if (gpu->tabulatedFunctions[i].coefficients != NULL)
            delete gpu->tabulatedFunctions[i].coefficients;
1931
1932
    if (gpu->compactPlan.valid)
        destroyCompactionPlan(gpu->compactPlan);
Peter Eastman's avatar
Peter Eastman committed
1933
1934
1935
1936
1937
1938
1939
1940
1941

    // Wrap up
    delete gpu;
    return;
}

extern "C"
int gpuBuildOutputBuffers(gpuContext gpu)
{
1942
1943
1944
1945
1946
1947
1948
1949
1950
1951
1952
1953
1954
    // Select the number of output buffer to use.
    gpu->bOutputBufferPerWarp           = true;
    gpu->sim.nonbondOutputBuffers       = gpu->sim.nonbond_blocks * gpu->sim.nonbond_threads_per_block / GRID;
    if (gpu->sim.nonbondOutputBuffers >= gpu->sim.paddedNumberOfAtoms/GRID)
    {
        // For small systems, it is more efficient to have one output buffer per block of 32 atoms instead of one per warp.
        gpu->bOutputBufferPerWarp           = false;
        gpu->sim.nonbondOutputBuffers       = gpu->sim.paddedNumberOfAtoms / GRID;
    }
    gpu->sim.totalNonbondOutputBuffers  = (gpu->bIncludeGBSA ? 2 * gpu->sim.nonbondOutputBuffers : gpu->sim.nonbondOutputBuffers);
    gpu->sim.outputBuffers              = gpu->sim.totalNonbondOutputBuffers;


Peter Eastman's avatar
Peter Eastman committed
1955
1956
1957
1958
1959
1960
1961
1962
1963
    unsigned int outputBuffers = gpu->sim.totalNonbondOutputBuffers;
    for (unsigned int i = 0; i < gpu->sim.paddedNumberOfAtoms; i++)
    {
        if (outputBuffers < gpu->pOutputBufferCounter[i])
        {
            outputBuffers = gpu->pOutputBufferCounter[i];
        }
    }    
    gpu->sim.outputBuffers      = outputBuffers;
1964
    gpu->sim.energyOutputBuffers = max(gpu->sim.nonbond_threads_per_block, gpu->sim.localForces_threads_per_block)*gpu->sim.blocks;
1965
    gpu->psForce4               = new CUDAStream<float4>(gpu->sim.paddedNumberOfAtoms, outputBuffers, "Force");
1966
    gpu->psEnergy               = new CUDAStream<float>(gpu->sim.energyOutputBuffers, 1, "Energy");
1967
1968
    gpu->psBornForce            = new CUDAStream<float>(gpu->sim.paddedNumberOfAtoms, gpu->sim.nonbondOutputBuffers, "BornForce");
    gpu->psBornSum              = new CUDAStream<float>(gpu->sim.paddedNumberOfAtoms, gpu->sim.nonbondOutputBuffers, "BornSum");
Peter Eastman's avatar
Peter Eastman committed
1969
1970
1971
    gpu->sim.pForce4            = gpu->psForce4->_pDevStream[0];
    gpu->sim.pForce4a           = gpu->sim.pForce4;
    gpu->sim.pForce4b           = gpu->sim.pForce4 + 1 * gpu->sim.nonbondOutputBuffers * gpu->sim.stride;
1972
    gpu->sim.pEnergy            = gpu->psEnergy->_pDevStream[0];
Peter Eastman's avatar
Peter Eastman committed
1973
1974
1975
1976
1977
1978
1979
1980
1981
1982
1983
1984
1985
1986
1987
1988
1989
1990
1991
    gpu->sim.pBornForce         = gpu->psBornForce->_pDevStream[0];
    gpu->sim.pBornSum           = gpu->psBornSum->_pDevStream[0];

    // Determine local energy paramter offsets for bonded interactions
    gpu->sim.bond_offset        =                                  gpu->psBondParameter->_stride;
    gpu->sim.bond_angle_offset  = gpu->sim.bond_offset           + gpu->psBondAngleParameter->_stride;
    gpu->sim.dihedral_offset    = gpu->sim.bond_angle_offset     + gpu->psDihedralParameter->_stride;
    gpu->sim.rb_dihedral_offset = gpu->sim.dihedral_offset       + gpu->psRbDihedralParameter1->_stride;
    gpu->sim.LJ14_offset        = gpu->sim.rb_dihedral_offset    + gpu->psLJ14Parameter->_stride;
    gpu->sim.localForces_threads_per_block  = (gpu->sim.LJ14_offset / gpu->sim.blocks + 15) & 0xfffffff0;
    if (gpu->sim.localForces_threads_per_block > gpu->sim.max_localForces_threads_per_block)
        gpu->sim.localForces_threads_per_block = gpu->sim.max_localForces_threads_per_block;
    if (gpu->sim.localForces_threads_per_block < 1)
        gpu->sim.localForces_threads_per_block = 1;

    // Flip local force output buffers
    int flip = outputBuffers - 1;
    for (int i = 0; i < (int) gpu->sim.bonds; i++)
    {
1992
1993
        (*gpu->psBondID)[i].z = flip - (*gpu->psBondID)[i].z;
        (*gpu->psBondID)[i].w = flip - (*gpu->psBondID)[i].w;
Peter Eastman's avatar
Peter Eastman committed
1994
1995
1996
    }
    for (int i = 0; i < (int) gpu->sim.bond_angles; i++)
    {
1997
1998
1999
        (*gpu->psBondAngleID1)[i].w = flip - (*gpu->psBondAngleID1)[i].w;
        (*gpu->psBondAngleID2)[i].x = flip - (*gpu->psBondAngleID2)[i].x;
        (*gpu->psBondAngleID2)[i].y = flip - (*gpu->psBondAngleID2)[i].y;
Peter Eastman's avatar
Peter Eastman committed
2000
2001
2002
    }
    for (int i = 0; i < (int) gpu->sim.dihedrals; i++)
    {
2003
2004
2005
2006
        (*gpu->psDihedralID2)[i].x = flip - (*gpu->psDihedralID2)[i].x;
        (*gpu->psDihedralID2)[i].y = flip - (*gpu->psDihedralID2)[i].y;
        (*gpu->psDihedralID2)[i].z = flip - (*gpu->psDihedralID2)[i].z;
        (*gpu->psDihedralID2)[i].w = flip - (*gpu->psDihedralID2)[i].w;
Peter Eastman's avatar
Peter Eastman committed
2007
2008
2009
    }
    for (int i = 0; i < (int) gpu->sim.rb_dihedrals; i++)
    {
2010
2011
2012
2013
        (*gpu->psRbDihedralID2)[i].x = flip - (*gpu->psRbDihedralID2)[i].x;
        (*gpu->psRbDihedralID2)[i].y = flip - (*gpu->psRbDihedralID2)[i].y;
        (*gpu->psRbDihedralID2)[i].z = flip - (*gpu->psRbDihedralID2)[i].z;
        (*gpu->psRbDihedralID2)[i].w = flip - (*gpu->psRbDihedralID2)[i].w;
Peter Eastman's avatar
Peter Eastman committed
2014
2015
2016
    }
    for (int i = 0; i < (int) gpu->sim.LJ14s; i++)
    {
2017
2018
        (*gpu->psLJ14ID)[i].z = flip - (*gpu->psLJ14ID)[i].z;
        (*gpu->psLJ14ID)[i].w = flip - (*gpu->psLJ14ID)[i].w;
Peter Eastman's avatar
Peter Eastman committed
2019
2020
2021
2022
2023
2024
2025
2026
2027
2028
2029
2030
2031
2032
2033
2034
2035
2036
    }
    gpu->psBondID->Upload();
    gpu->psBondAngleID1->Upload();
    gpu->psBondAngleID2->Upload();
    gpu->psDihedralID2->Upload();
    gpu->psRbDihedralID2->Upload();
    gpu->psLJ14ID->Upload();

    return 1;
}

extern "C"
int gpuBuildThreadBlockWorkList(gpuContext gpu)
{
    const unsigned int atoms = gpu->sim.paddedNumberOfAtoms;
    const unsigned int grid = gpu->grid;
    const unsigned int dim = (atoms + (grid - 1)) / grid;
    const unsigned int cells = dim * (dim + 1) / 2;
2037
2038
    CUDAStream<unsigned int>* psWorkUnit = new CUDAStream<unsigned int>(cells, 1u, "WorkUnit");
    unsigned int* pWorkList = psWorkUnit->_pSysData;
Peter Eastman's avatar
Peter Eastman committed
2039
2040
    gpu->psWorkUnit = psWorkUnit;
    gpu->sim.pWorkUnit = psWorkUnit->_pDevStream[0];
2041
    CUDAStream<unsigned int>* psInteractingWorkUnit = new CUDAStream<unsigned int>(cells, 1u, "InteractingWorkUnit");
2042
2043
    gpu->psInteractingWorkUnit = psInteractingWorkUnit;
    gpu->sim.pInteractingWorkUnit = psInteractingWorkUnit->_pDevStream[0];
2044
    CUDAStream<unsigned int>* psInteractionFlag = new CUDAStream<unsigned int>(cells, 1u, "InteractionFlag");
2045
2046
    gpu->psInteractionFlag = psInteractionFlag;
    gpu->sim.pInteractionFlag = psInteractionFlag->_pDevStream[0];
2047
    CUDAStream<size_t>* psInteractionCount = new CUDAStream<size_t>(1, 1u, "InteractionCount");
2048
2049
    gpu->psInteractionCount = psInteractionCount;
    gpu->sim.pInteractionCount = psInteractionCount->_pDevStream[0];
2050
    CUDAStream<float4>* psGridBoundingBox = new CUDAStream<float4>(dim, 1u, "GridBoundingBox");
2051
2052
    gpu->psGridBoundingBox = psGridBoundingBox;
    gpu->sim.pGridBoundingBox = psGridBoundingBox->_pDevStream[0];
2053
    CUDAStream<float4>* psGridCenter = new CUDAStream<float4>(dim, 1u, "GridCenter");
2054
2055
    gpu->psGridCenter = psGridCenter;
    gpu->sim.pGridCenter = psGridCenter->_pDevStream[0];
Peter Eastman's avatar
Peter Eastman committed
2056
2057
2058
2059
    gpu->sim.nonbond_workBlock      = gpu->sim.nonbond_threads_per_block / GRID;
    gpu->sim.bornForce2_workBlock   = gpu->sim.bornForce2_threads_per_block / GRID;
    gpu->sim.workUnits = cells;

2060
2061
    // Initialize the plan for doing stream compaction.
    planCompaction(gpu->compactPlan);
2062
2063
2064
2065
2066
2067
2068
2069
2070
2071
2072
2073

    // Increase block count if necessary for extra large molecules that would
    // otherwise overflow the SM workunit buffers
//    int minimumBlocks = (cells + gpu->sim.workUnitsPerSM - 1) / gpu->sim.workUnitsPerSM;
//    if ((int) gpu->sim.nonbond_blocks < minimumBlocks)
//    {
//        gpu->sim.nonbond_blocks = gpu->sim.nonbond_blocks * ((minimumBlocks + gpu->sim.nonbond_blocks - 1) / gpu->sim.nonbond_blocks);
//    }
//    if ((int) gpu->sim.bornForce2_blocks < minimumBlocks)
//    {
//        gpu->sim.bornForce2_blocks = gpu->sim.bornForce2_blocks * ((minimumBlocks + gpu->sim.bornForce2_blocks - 1) / gpu->sim.bornForce2_blocks);
//    }
Peter Eastman's avatar
Peter Eastman committed
2074
2075
2076
2077
    gpu->sim.nbWorkUnitsPerBlock            = cells / gpu->sim.nonbond_blocks;
    gpu->sim.nbWorkUnitsPerBlockRemainder   = cells - gpu->sim.nonbond_blocks * gpu->sim.nbWorkUnitsPerBlock;
    gpu->sim.bf2WorkUnitsPerBlock           = cells / gpu->sim.bornForce2_blocks;
    gpu->sim.bf2WorkUnitsPerBlockRemainder  = cells - gpu->sim.bornForce2_blocks * gpu->sim.bf2WorkUnitsPerBlock;
2078
2079
    gpu->sim.interaction_threads_per_block = 64;
    gpu->sim.interaction_blocks = (gpu->sim.workUnits + gpu->sim.interaction_threads_per_block - 1) / gpu->sim.interaction_threads_per_block;
2080
2081
    if (gpu->sim.interaction_blocks > 8*gpu->sim.blocks)
        gpu->sim.interaction_blocks = 8*gpu->sim.blocks;
Peter Eastman's avatar
Peter Eastman committed
2082
2083
2084
2085
2086
2087
2088
2089
2090
2091
2092
2093
2094
2095
2096
2097
2098
2099
2100
2101
2102
2103
2104
2105
2106
2107
2108

    // Decrease thread count for extra small molecules to spread computation
    // across entire chip
    int activeWorkUnits = gpu->sim.nonbond_blocks * gpu->sim.nonbond_workBlock;
    if (activeWorkUnits > (int) cells)
    {
        int balancedWorkBlock                   = (cells + gpu->sim.nonbond_blocks - 1) / gpu->sim.nonbond_blocks;
        gpu->sim.nonbond_threads_per_block      = balancedWorkBlock * GRID;
        gpu->sim.nonbond_workBlock              = balancedWorkBlock;
    }
    activeWorkUnits = gpu->sim.bornForce2_blocks * gpu->sim.bornForce2_workBlock;
    if (activeWorkUnits > (int) cells)
    {
        int balancedWorkBlock                   = (cells + gpu->sim.bornForce2_blocks - 1) / gpu->sim.bornForce2_blocks;
        gpu->sim.bornForce2_threads_per_block   = balancedWorkBlock * GRID;
        gpu->sim.bornForce2_workBlock           = balancedWorkBlock;
    }

    unsigned int count = 0;
    for (unsigned int y = 0; y < dim; y++)
    {
        for (unsigned int x = y; x < dim; x++)
        {
            pWorkList[count] = (x << 17) | (y << 2);
            count++;
        }
    }
2109
    (*gpu->psInteractionCount)[0] = gpu->sim.workUnits;
Peter Eastman's avatar
Peter Eastman committed
2110

2111
    gpu->psInteractionCount->Upload();
Peter Eastman's avatar
Peter Eastman committed
2112
2113
2114
2115
2116
2117
    psWorkUnit->Upload();
    gpuSetConstants(gpu);
    return cells;
}

extern "C"
2118
void gpuBuildExclusionList(gpuContext gpu)
Peter Eastman's avatar
Peter Eastman committed
2119
{
2120
2121
    const unsigned int atoms = gpu->sim.paddedNumberOfAtoms;
    const unsigned int grid = gpu->grid;
2122
    const unsigned int dim = atoms/grid;
2123
    unsigned int* pWorkList = gpu->psWorkUnit->_pSysData;
2124

2125
    // Mark which work units have exclusions.
Peter Eastman's avatar
Peter Eastman committed
2126

2127
    for (int atom1 = 0; atom1 < (int)gpu->exclusions.size(); ++atom1)
Peter Eastman's avatar
Peter Eastman committed
2128
    {
2129
        int x = atom1/grid;
2130
        for (int j = 0; j < (int)gpu->exclusions[atom1].size(); ++j)
2131
2132
2133
2134
2135
2136
2137
        {
            int atom2 = gpu->exclusions[atom1][j];
            int y = atom2/grid;
            int cell = (x > y ? x+y*dim-y*(y+1)/2 : y+x*dim-x*(x+1)/2);
            pWorkList[cell] |= 1;
        }
    }
2138
    if ((int)gpu->sim.paddedNumberOfAtoms > gpu->natoms)
2139
2140
    {
        int lastBlock = gpu->natoms/grid;
2141
        for (int i = 0; i < (int)gpu->sim.workUnits; ++i)
2142
2143
2144
2145
2146
2147
2148
2149
2150
2151
        {
            int x = pWorkList[i]>>17;
            int y = (pWorkList[i]>>2)&0x7FFF;
            if (x == lastBlock || y == lastBlock)
                pWorkList[i] |= 1;
        }
    }

    // Build a list of indexes for the work units with exclusions.

2152
    CUDAStream<unsigned int>* psExclusionIndex = new CUDAStream<unsigned int>(gpu->sim.workUnits, 1u, "ExclusionIndex");
2153
2154
2155
2156
    gpu->psExclusionIndex = psExclusionIndex;
    unsigned int* pExclusionIndex = psExclusionIndex->_pSysData;
    gpu->sim.pExclusionIndex = psExclusionIndex->_pDevData;
    int numWithExclusions = 0;
2157
    for (int i = 0; i < (int)psExclusionIndex->_length; ++i)
2158
2159
2160
2161
2162
        if ((pWorkList[i]&1) == 1)
            pExclusionIndex[i] = (numWithExclusions++)*grid;

    // Record the exclusion data.

2163
    CUDAStream<unsigned int>* psExclusion = new CUDAStream<unsigned int>(numWithExclusions*grid, 1u, "Exclusion");
2164
2165
2166
    gpu->psExclusion = psExclusion;
    unsigned int* pExclusion = psExclusion->_pSysData;
    gpu->sim.pExclusion = psExclusion->_pDevData;
2167
    for (int i = 0; i < (int)psExclusion->_length; ++i)
2168
        pExclusion[i] = 0xFFFFFFFF;
2169
    for (int atom1 = 0; atom1 < (int)gpu->exclusions.size(); ++atom1)
2170
2171
2172
    {
        int x = atom1/grid;
        int offset1 = atom1-x*grid;
2173
        for (int j = 0; j < (int)gpu->exclusions[atom1].size(); ++j)
2174
2175
2176
2177
2178
2179
2180
2181
2182
2183
2184
2185
2186
2187
2188
2189
        {
            int atom2 = gpu->exclusions[atom1][j];
            int y = atom2/grid;
            int offset2 = atom2-y*grid;
            if (x > y)
            {
                int cell = x+y*dim-y*(y+1)/2;
                pExclusion[pExclusionIndex[cell]+offset1] &= 0xFFFFFFFF-(1<<offset2);
            }
            else
            {
                int cell = y+x*dim-x*(x+1)/2;
                pExclusion[pExclusionIndex[cell]+offset2] &= 0xFFFFFFFF-(1<<offset1);
            }
        }
    }
2190
2191
2192

    // Mark all interactions that involve a padding atom as being excluded.

2193
    for (int atom1 = gpu->natoms; atom1 < (int)atoms; ++atom1)
2194
2195
2196
    {
        int x = atom1/grid;
        int offset1 = atom1-x*grid;
2197
        for (int atom2 = 0; atom2 < (int)atoms; ++atom2)
2198
2199
2200
        {
            int y = atom2/grid;
            int offset2 = atom2-y*grid;
2201
2202
2203
2204
2205
2206
2207
2208
2209
2210
            if (x >= y)
            {
                int cell = x+y*dim-y*(y+1)/2;
                pExclusion[pExclusionIndex[cell]+offset1] &= 0xFFFFFFFF-(1<<offset2);
            }
            if (y >= x)
            {
                int cell = y+x*dim-x*(x+1)/2;
                pExclusion[pExclusionIndex[cell]+offset2] &= 0xFFFFFFFF-(1<<offset1);
            }
Peter Eastman's avatar
Peter Eastman committed
2211
2212
2213
2214
        }
    }
    
    psExclusion->Upload();
2215
    psExclusionIndex->Upload();
2216
    gpu->psWorkUnit->Upload();
Peter Eastman's avatar
Peter Eastman committed
2217
2218
2219
2220
2221
2222
2223
2224
    gpuSetConstants(gpu);
}

extern "C"
int gpuSetConstants(gpuContext gpu)
{
    SetCalculateCDLJForcesSim(gpu);
    SetCalculateCDLJObcGbsaForces1Sim(gpu);
2225
    SetCalculateCustomNonbondedForcesSim(gpu);
Peter Eastman's avatar
Peter Eastman committed
2226
2227
2228
2229
    SetCalculateLocalForcesSim(gpu);
    SetCalculateObcGbsaBornSumSim(gpu);
    SetCalculateObcGbsaForces2Sim(gpu);
    SetCalculateAndersenThermostatSim(gpu);
2230
    SetCalculatePMESim(gpu);
Peter Eastman's avatar
Peter Eastman committed
2231
    SetForcesSim(gpu);
2232
2233
    SetShakeHSim(gpu);
    SetLangevinUpdateSim(gpu);
Peter Eastman's avatar
Peter Eastman committed
2234
2235
    SetVerletUpdateSim(gpu);
    SetBrownianUpdateSim(gpu);
2236
    SetSettleSim(gpu);
2237
    SetCCMASim(gpu);
Peter Eastman's avatar
Peter Eastman committed
2238
2239
2240
2241
    SetRandomSim(gpu);
    return 1;
}

2242
2243
2244
2245
2246
static void tagAtomsInMolecule(int atom, int molecule, vector<int>& atomMolecule, vector<vector<int> >& atomBonds)
{
    // Recursively tag atoms as belonging to a particular molecule.

    atomMolecule[atom] = molecule;
2247
    for (int i = 0; i < (int)atomBonds[atom].size(); i++)
2248
2249
2250
2251
2252
2253
2254
2255
2256
        if (atomMolecule[atomBonds[atom][i]] == -1)
            tagAtomsInMolecule(atomBonds[atom][i], molecule, atomMolecule, atomBonds);
}

static void findMoleculeGroups(gpuContext gpu)
{
    // First make a list of constraints for future use.

    vector<Constraint> constraints;
2257
    for (int i = 0; i < (int)gpu->sim.ShakeConstraints; i++)
2258
    {
2259
2260
2261
2262
2263
        int atom1 = (*gpu->psShakeID)[i].x;
        int atom2 = (*gpu->psShakeID)[i].y;
        int atom3 = (*gpu->psShakeID)[i].z;
        int atom4 = (*gpu->psShakeID)[i].w;
        float distance2 = (*gpu->psShakeParameter)[i].z;
2264
2265
2266
2267
        constraints.push_back(Constraint(atom1, atom2, distance2));
        if (atom3 != -1)
            constraints.push_back(Constraint(atom1, atom3, distance2));
        if (atom4 != -1)
2268
            constraints.push_back(Constraint(atom1, atom4, distance2));
2269
    }
2270
    for (int i = 0; i < (int)gpu->sim.settleConstraints; i++)
2271
    {
2272
2273
2274
2275
2276
        int atom1 = (*gpu->psSettleID)[i].x;
        int atom2 = (*gpu->psSettleID)[i].y;
        int atom3 = (*gpu->psSettleID)[i].z;
        float distance12 = (*gpu->psSettleParameter)[i].x;
        float distance23 = (*gpu->psSettleParameter)[i].y;
2277
2278
2279
2280
        constraints.push_back(Constraint(atom1, atom2, distance12*distance12));
        constraints.push_back(Constraint(atom1, atom3, distance12*distance12));
        constraints.push_back(Constraint(atom2, atom3, distance23*distance23));
    }
2281
    for (int i = 0; i < (int)gpu->sim.ccmaConstraints; i++)
Peter Eastman's avatar
Peter Eastman committed
2282
    {
2283
2284
2285
        int atom1 = (*gpu->psCcmaAtoms)[i].x;
        int atom2 = (*gpu->psCcmaAtoms)[i].y;
        float distance2 = (*gpu->psCcmaDistance)[i].w;
Peter Eastman's avatar
Peter Eastman committed
2286
2287
        constraints.push_back(Constraint(atom1, atom2, distance2));
    }
2288

2289
    // First make a list of every other atom to which each atom is connect by a bond, constraint, or exclusion.
2290
2291
2292

    int numAtoms = gpu->natoms;
    vector<vector<int> > atomBonds(numAtoms);
2293
    for (int i = 0; i < (int)gpu->sim.bonds; i++)
2294
    {
2295
2296
        int atom1 = (*gpu->psBondID)[i].x;
        int atom2 = (*gpu->psBondID)[i].y;
2297
2298
2299
        atomBonds[atom1].push_back(atom2);
        atomBonds[atom2].push_back(atom1);
    }
2300
    for (int i = 0; i < (int)constraints.size(); i++)
2301
2302
2303
2304
2305
2306
    {
        int atom1 = constraints[i].atom1;
        int atom2 = constraints[i].atom2;
        atomBonds[atom1].push_back(atom2);
        atomBonds[atom2].push_back(atom1);
    }
2307
2308
2309
    for (int i = 0; i < (int)gpu->exclusions.size(); i++)
        for (int j = 0; j < (int)gpu->exclusions[i].size(); j++)
            atomBonds[i].push_back(gpu->exclusions[i][j]);
2310
2311
2312
2313
2314
2315
2316
2317
2318
2319
2320
2321
2322
2323
2324
2325
2326

    // Now tag atoms by which molecule they belong to.

    vector<int> atomMolecule(numAtoms, -1);
    int numMolecules = 0;
    for (int i = 0; i < numAtoms; i++)
        if (atomMolecule[i] == -1)
            tagAtomsInMolecule(i, numMolecules++, atomMolecule, atomBonds);
    vector<vector<int> > atomIndices(numMolecules);
    for (int i = 0; i < numAtoms; i++)
        atomIndices[atomMolecule[i]].push_back(i);

    // Construct a description of each molecule.

    vector<Molecule> molecules(numMolecules);
    for (int i = 0; i < numMolecules; i++)
        molecules[i].atoms = atomIndices[i];
2327
    for (int i = 0; i < (int)gpu->sim.bonds; i++)
2328
    {
2329
        int atom1 = (*gpu->psBondID)[i].x;
2330
2331
        molecules[atomMolecule[atom1]].bonds.push_back(i);
    }
2332
    for (int i = 0; i < (int)gpu->sim.bond_angles; i++)
2333
    {
2334
        int atom1 = (*gpu->psBondAngleID1)[i].x;
2335
2336
        molecules[atomMolecule[atom1]].angles.push_back(i);
    }
2337
    for (int i = 0; i < (int)gpu->sim.dihedrals; i++)
2338
    {
2339
        int atom1 = (*gpu->psDihedralID1)[i].x;
2340
2341
        molecules[atomMolecule[atom1]].periodicTorsions.push_back(i);
    }
2342
    for (int i = 0; i < (int)gpu->sim.rb_dihedrals; i++)
2343
    {
2344
        int atom1 = (*gpu->psRbDihedralID1)[i].x;
2345
2346
        molecules[atomMolecule[atom1]].rbTorsions.push_back(i);
    }
2347
    for (int i = 0; i < (int)constraints.size(); i++)
2348
2349
2350
    {
        molecules[atomMolecule[constraints[i].atom1]].constraints.push_back(i);
    }
2351
2352
2353
2354
2355
    for (int i = 0; i < (int)gpu->sim.LJ14s; i++)
    {
        int atom1 = (*gpu->psLJ14ID)[i].x;
        molecules[atomMolecule[atom1]].lj14s.push_back(i);
    }
2356
2357
2358
2359
2360

    // Sort them into groups of identical molecules.

    vector<Molecule> uniqueMolecules;
    vector<vector<int> > moleculeInstances;
2361
    for (int molIndex = 0; molIndex < (int)molecules.size(); molIndex++)
2362
2363
2364
2365
2366
2367
    {
        Molecule& mol = molecules[molIndex];

        // See if it is identical to another molecule.

        bool isNew = true;
2368
        for (int j = 0; j < (int)uniqueMolecules.size() && isNew; j++)
2369
2370
2371
2372
2373
        {
            Molecule& mol2 = uniqueMolecules[j];
            bool identical = true;
            if (mol.atoms.size() != mol2.atoms.size() || mol.bonds.size() != mol2.bonds.size()
                    || mol.angles.size() != mol2.angles.size() || mol.periodicTorsions.size() != mol2.periodicTorsions.size()
2374
2375
                    || mol.rbTorsions.size() != mol2.rbTorsions.size() || mol.constraints.size() != mol2.constraints.size()
                    || mol.lj14s.size() != mol2.lj14s.size())
2376
2377
2378
2379
2380
                identical = false;
            int atomOffset = mol2.atoms[0]-mol.atoms[0];
            float4* posq = gpu->psPosq4->_pSysData;
            float4* velm = gpu->psVelm4->_pSysData;
            float2* sigeps = gpu->psSigEps2->_pSysData;
2381
            for (int i = 0; i < (int)mol.atoms.size() && identical; i++)
2382
2383
2384
2385
2386
2387
                if (mol.atoms[i] != mol2.atoms[i]-atomOffset || posq[mol.atoms[i]].w != posq[mol2.atoms[i]].w ||
                        velm[mol.atoms[i]].w != velm[mol2.atoms[i]].w || sigeps[mol.atoms[i]].x != sigeps[mol2.atoms[i]].x ||
                        sigeps[mol.atoms[i]].y != sigeps[mol2.atoms[i]].y)
                    identical = false;
            int4* bondID = gpu->psBondID->_pSysData;
            float2* bondParam = gpu->psBondParameter->_pSysData;
2388
            for (int i = 0; i < (int)mol.bonds.size() && identical; i++)
2389
2390
2391
2392
2393
                if (bondID[mol.bonds[i]].x != bondID[mol2.bonds[i]].x-atomOffset || bondID[mol.bonds[i]].y != bondID[mol2.bonds[i]].y-atomOffset ||
                        bondParam[mol.bonds[i]].x != bondParam[mol2.bonds[i]].x || bondParam[mol.bonds[i]].y != bondParam[mol2.bonds[i]].y)
                    identical = false;
            int4* angleID = gpu->psBondAngleID1->_pSysData;
            float2* angleParam = gpu->psBondAngleParameter->_pSysData;
2394
            for (int i = 0; i < (int)mol.angles.size() && identical; i++)
2395
2396
2397
2398
2399
2400
2401
2402
                if (angleID[mol.angles[i]].x != angleID[mol2.angles[i]].x-atomOffset ||
                        angleID[mol.angles[i]].y != angleID[mol2.angles[i]].y-atomOffset ||
                        angleID[mol.angles[i]].z != angleID[mol2.angles[i]].z-atomOffset ||
                        angleParam[mol.angles[i]].x != angleParam[mol2.angles[i]].x ||
                        angleParam[mol.angles[i]].y != angleParam[mol2.angles[i]].y)
                    identical = false;
            int4* periodicID = gpu->psDihedralID1->_pSysData;
            float4* periodicParam = gpu->psDihedralParameter->_pSysData;
2403
            for (int i = 0; i < (int)mol.periodicTorsions.size() && identical; i++)
2404
2405
2406
2407
2408
2409
2410
2411
2412
2413
2414
                if (periodicID[mol.periodicTorsions[i]].x != periodicID[mol2.periodicTorsions[i]].x-atomOffset ||
                        periodicID[mol.periodicTorsions[i]].y != periodicID[mol2.periodicTorsions[i]].y-atomOffset ||
                        periodicID[mol.periodicTorsions[i]].z != periodicID[mol2.periodicTorsions[i]].z-atomOffset ||
                        periodicID[mol.periodicTorsions[i]].w != periodicID[mol2.periodicTorsions[i]].w-atomOffset ||
                        periodicParam[mol.periodicTorsions[i]].x != periodicParam[mol2.periodicTorsions[i]].x ||
                        periodicParam[mol.periodicTorsions[i]].y != periodicParam[mol2.periodicTorsions[i]].y ||
                        periodicParam[mol.periodicTorsions[i]].z != periodicParam[mol2.periodicTorsions[i]].z)
                    identical = false;
            int4* rbID = gpu->psRbDihedralID1->_pSysData;
            float4* rbParam1 = gpu->psRbDihedralParameter1->_pSysData;
            float2* rbParam2 = gpu->psRbDihedralParameter2->_pSysData;
2415
            for (int i = 0; i < (int)mol.rbTorsions.size() && identical; i++)
2416
2417
2418
2419
2420
2421
2422
2423
2424
2425
2426
                if (rbID[mol.rbTorsions[i]].x != rbID[mol2.rbTorsions[i]].x-atomOffset ||
                        rbID[mol.rbTorsions[i]].y != rbID[mol2.rbTorsions[i]].y-atomOffset ||
                        rbID[mol.rbTorsions[i]].z != rbID[mol2.rbTorsions[i]].z-atomOffset ||
                        rbID[mol.rbTorsions[i]].w != rbID[mol2.rbTorsions[i]].w-atomOffset ||
                        rbParam1[mol.rbTorsions[i]].x != rbParam1[mol2.rbTorsions[i]].x ||
                        rbParam1[mol.rbTorsions[i]].y != rbParam1[mol2.rbTorsions[i]].y ||
                        rbParam1[mol.rbTorsions[i]].z != rbParam1[mol2.rbTorsions[i]].z ||
                        rbParam1[mol.rbTorsions[i]].w != rbParam1[mol2.rbTorsions[i]].w ||
                        rbParam2[mol.rbTorsions[i]].x != rbParam2[mol2.rbTorsions[i]].x ||
                        rbParam2[mol.rbTorsions[i]].y != rbParam2[mol2.rbTorsions[i]].y)
                    identical = false;
2427
            for (int i = 0; i < (int)mol.constraints.size() && identical; i++)
2428
2429
2430
2431
                if (constraints[mol.constraints[i]].atom1 != constraints[mol2.constraints[i]].atom1-atomOffset ||
                        constraints[mol.constraints[i]].atom2 != constraints[mol2.constraints[i]].atom2-atomOffset ||
                        constraints[mol.constraints[i]].distance2 != constraints[mol2.constraints[i]].distance2)
                    identical = false;
2432
2433
2434
2435
2436
2437
2438
            int4* lj14ID = gpu->psLJ14ID->_pSysData;
            float4* lj14Param = gpu->psLJ14Parameter->_pSysData;
            for (int i = 0; i < (int)mol.lj14s.size() && identical; i++)
                if (lj14ID[mol.lj14s[i]].x != lj14ID[mol2.lj14s[i]].x-atomOffset || lj14ID[mol.lj14s[i]].y != lj14ID[mol2.lj14s[i]].y-atomOffset ||
                        lj14Param[mol.lj14s[i]].x != lj14Param[mol2.lj14s[i]].x || lj14Param[mol.lj14s[i]].y != lj14Param[mol2.lj14s[i]].y ||
                        lj14Param[mol.lj14s[i]].z != lj14Param[mol2.lj14s[i]].z)
                    identical = false;
2439
2440
2441
2442
2443
2444
2445
2446
2447
2448
2449
2450
2451
2452
            if (identical)
            {
                moleculeInstances[j].push_back(mol.atoms[0]);
                isNew = false;
            }
        }
        if (isNew)
        {
            uniqueMolecules.push_back(mol);
            moleculeInstances.push_back(vector<int>());
            moleculeInstances[moleculeInstances.size()-1].push_back(mol.atoms[0]);
        }
    }
    gpu->moleculeGroups.resize(moleculeInstances.size());
2453
    for (int i = 0; i < (int)moleculeInstances.size(); i++)
2454
2455
2456
2457
    {
        gpu->moleculeGroups[i].instances = moleculeInstances[i];
        vector<int>& atoms = uniqueMolecules[i].atoms;
        gpu->moleculeGroups[i].atoms.resize(atoms.size());
2458
        for (int j = 0; j < (int)atoms.size(); j++)
2459
2460
2461
2462
2463
2464
2465
2466
2467
2468
2469
2470
2471
2472
2473
2474
2475
2476
2477
2478
2479
2480
            gpu->moleculeGroups[i].atoms[j] = atoms[j]-atoms[0];
    }
}

extern "C"
void gpuReorderAtoms(gpuContext gpu)
{
    if (gpu->natoms == 0 || gpu->sim.nonbondedCutoffSqr == 0.0)
        return;
    if (gpu->moleculeGroups.size() == 0)
        findMoleculeGroups(gpu);

    // Find the range of positions and the number of bins along each axis.

    int numAtoms = gpu->natoms;
    gpu->psPosq4->Download();
    gpu->psVelm4->Download();
    float4* posq = gpu->psPosq4->_pSysData;
    float4* velm = gpu->psVelm4->_pSysData;
    float minx = posq[0].x, maxx = posq[0].x;
    float miny = posq[0].y, maxy = posq[0].y;
    float minz = posq[0].z, maxz = posq[0].z;
Peter Eastman's avatar
Bug fix  
Peter Eastman committed
2481
    if (gpu->sim.nonbondedMethod == PERIODIC || gpu->sim.nonbondedMethod == EWALD || gpu->sim.nonbondedMethod == PARTICLE_MESH_EWALD)
2482
2483
2484
2485
2486
2487
2488
2489
2490
2491
2492
2493
2494
2495
2496
2497
2498
2499
2500
2501
2502
2503
2504
2505
    {
        minx = miny = minz = 0.0;
        maxx = gpu->sim.periodicBoxSizeX;
        maxy = gpu->sim.periodicBoxSizeY;
        maxz = gpu->sim.periodicBoxSizeZ;
    }
    else
    {
        for (int i = 1; i < numAtoms; i++)
        {
            minx = min(minx, posq[i].x);
            maxx = max(maxx, posq[i].x);
            miny = min(miny, posq[i].y);
            maxy = max(maxy, posq[i].y);
            minz = min(minz, posq[i].z);
            maxz = max(maxz, posq[i].z);
        }
    }

    // Loop over each group of identical molecules and reorder them.

    vector<int> originalIndex(numAtoms);
    vector<float4> newPosq(numAtoms);
    vector<float4> newVelm(numAtoms);
2506
    vector<int3> newCellOffsets(numAtoms);
2507
    for (int group = 0; group < (int)gpu->moleculeGroups.size(); group++)
2508
2509
2510
2511
2512
2513
2514
2515
2516
2517
2518
2519
    {
        // Find the center of each molecule.

        gpuMoleculeGroup& mol = gpu->moleculeGroups[group];
        int numMolecules = mol.instances.size();
        vector<int>& atoms = mol.atoms;
        vector<float3> molPos(numMolecules);
        for (int i = 0; i < numMolecules; i++)
        {
            molPos[i].x = 0.0f;
            molPos[i].y = 0.0f;
            molPos[i].z = 0.0f;
2520
            for (int j = 0; j < (int)atoms.size(); j++)
2521
2522
2523
2524
2525
2526
2527
2528
2529
2530
            {
                int atom = atoms[j]+mol.instances[i];
                molPos[i].x += posq[atom].x;
                molPos[i].y += posq[atom].y;
                molPos[i].z += posq[atom].z;
            }
            molPos[i].x /= atoms.size();
            molPos[i].y /= atoms.size();
            molPos[i].z /= atoms.size();
        }
Peter Eastman's avatar
Bug fix  
Peter Eastman committed
2531
        if (gpu->sim.nonbondedMethod == PERIODIC || gpu->sim.nonbondedMethod == EWALD || gpu->sim.nonbondedMethod == PARTICLE_MESH_EWALD)
2532
2533
2534
2535
2536
        {
            // Move each molecule position into the same box.

            for (int i = 0; i < numMolecules; i++)
            {
2537
2538
2539
2540
2541
2542
                int xcell = (int) floor(molPos[i].x/gpu->sim.periodicBoxSizeX);
                int ycell = (int) floor(molPos[i].y/gpu->sim.periodicBoxSizeY);
                int zcell = (int) floor(molPos[i].z/gpu->sim.periodicBoxSizeZ);
                float dx = xcell*gpu->sim.periodicBoxSizeX;
                float dy = ycell*gpu->sim.periodicBoxSizeY;
                float dz = zcell*gpu->sim.periodicBoxSizeZ;
2543
2544
2545
2546
2547
                if (dx != 0.0f || dy != 0.0f || dz != 0.0f)
                {
                    molPos[i].x -= dx;
                    molPos[i].y -= dy;
                    molPos[i].z -= dz;
2548
                    for (int j = 0; j < (int)atoms.size(); j++)
2549
2550
2551
2552
2553
                    {
                        int atom = atoms[j]+mol.instances[i];
                        posq[atom].x -= dx;
                        posq[atom].y -= dy;
                        posq[atom].z -= dz;
2554
2555
2556
                        gpu->posCellOffsets[atom].x -= xcell;
                        gpu->posCellOffsets[atom].y -= ycell;
                        gpu->posCellOffsets[atom].z -= zcell;
2557
2558
                    }
                }
2559
2560
2561
2562
2563
            }
        }

        // Select a bin for each molecule, then sort them by bin.

2564
        bool useHilbert = (numMolecules > 5000 || atoms.size() > 8); // For small systems, a simple zigzag curve works better than a Hilbert curve.
2565
2566
        float binWidth;
        if (useHilbert)
2567
            binWidth = (float)(max(max(maxx-minx, maxy-miny), maxz-minz)/255.0);
2568
        else
2569
            binWidth = (float)(0.2*sqrt(gpu->sim.nonbondedCutoffSqr));
2570
2571
        int xbins = 1 + (int) ((maxx-minx)/binWidth);
        int ybins = 1 + (int) ((maxy-miny)/binWidth);
2572
        vector<pair<int, int> > molBins(numMolecules);
2573
        bitmask_t coords[3];
2574
2575
2576
2577
2578
        for (int i = 0; i < numMolecules; i++)
        {
            int x = (int) ((molPos[i].x-minx)/binWidth);
            int y = (int) ((molPos[i].y-miny)/binWidth);
            int z = (int) ((molPos[i].z-minz)/binWidth);
2579
2580
2581
2582
2583
2584
2585
2586
2587
2588
2589
2590
2591
2592
2593
2594
            int bin;
            if (useHilbert)
            {
                coords[0] = x;
                coords[1] = y;
                coords[2] = z;
                bin = (int) hilbert_c2i(3, 8, coords);
            }
            else
            {
                int yodd = y&1;
                int zodd = z&1;
                bin = z*xbins*ybins;
                bin += (zodd ? ybins-y : y)*xbins;
                bin += (yodd ? xbins-x : x);
            }
2595
2596
2597
2598
2599
2600
2601
2602
            molBins[i] = pair<int, int>(bin, i);
        }
        sort(molBins.begin(), molBins.end());

        // Reorder the atoms.

        for (int i = 0; i < numMolecules; i++)
        {
2603
            for (int j = 0; j < (int)atoms.size(); j++)
2604
2605
2606
            {
                int oldIndex = mol.instances[molBins[i].second]+atoms[j];
                int newIndex = mol.instances[i]+atoms[j];
2607
                originalIndex[newIndex] = (*gpu->psAtomIndex)[oldIndex];
2608
2609
                newPosq[newIndex] = posq[oldIndex];
                newVelm[newIndex] = velm[oldIndex];
2610
                newCellOffsets[newIndex] = gpu->posCellOffsets[oldIndex];
2611
2612
2613
2614
2615
2616
            }
        }
    }

    // Update the streams.

2617
    for (int i = 0; i < numAtoms; i++) {
2618
2619
        posq[i] = newPosq[i];
        velm[i] = newVelm[i];
2620
        (*gpu->psAtomIndex)[i] = originalIndex[i];
2621
2622
2623
2624
        gpu->posCellOffsets[i] = newCellOffsets[i];
    }
    gpu->psPosq4->Upload();
    gpu->psVelm4->Upload();
2625
2626
    gpu->psAtomIndex->Upload();
}