CpuPmeKernels.cpp 25.6 KB
Newer Older
peastman's avatar
peastman committed
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
/* -------------------------------------------------------------------------- *
 *                                   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) 2013 Stanford University and the Authors.           *
 * Authors: Peter Eastman                                                     *
 * Contributors:                                                              *
 *                                                                            *
 * Permission is hereby granted, free of charge, to any person obtaining a    *
 * copy of this software and associated documentation files (the "Software"), *
 * to deal in the Software without restriction, including without limitation  *
 * the rights to use, copy, modify, merge, publish, distribute, sublicense,   *
 * and/or sell copies of the Software, and to permit persons to whom the      *
 * Software is furnished to do so, subject to the following conditions:       *
 *                                                                            *
 * The above copyright notice and this permission notice shall be included in *
 * all copies or substantial portions of the Software.                        *
 *                                                                            *
 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR *
 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,   *
 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL    *
 * THE AUTHORS, CONTRIBUTORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM,    *
 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR      *
 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE  *
 * USE OR OTHER DEALINGS IN THE SOFTWARE.                                     *
 * -------------------------------------------------------------------------- */

32
33
34
#ifdef WIN32
  #define _USE_MATH_DEFINES // Needed to get M_PI
#endif
35
#include "CpuPmeKernels.h"
peastman's avatar
peastman committed
36
#include "../src/SimTKUtilities/SimTKOpenMMRealType.h"
37
#include <cmath>
peastman's avatar
peastman committed
38
#include <cstring>
peastman's avatar
peastman committed
39
40
41
42
43
44
45
#include <smmintrin.h>

using namespace OpenMM;
using namespace std;

static const int PME_ORDER = 5;

46
47
bool CpuCalcPmeReciprocalForceKernel::hasInitializedThreads = false;
int CpuCalcPmeReciprocalForceKernel::numThreads = 0;
48

49
static float extractFloat(__m128 v, unsigned int element) {
peastman's avatar
peastman committed
50
51
52
53
54
    float f[4];
    _mm_store_ps(f, v);
    return f[element];
}

55
56
// Define function to get the number of processors.

57
58
59
60
61
62
63
64
65
66
67
#ifdef __APPLE__
   #include <sys/sysctl.h>
   #include <dlfcn.h>
#else
   #ifdef WIN32
      #include <windows.h>
   #else
      #include <dlfcn.h>
      #include <unistd.h>
   #endif
#endif
peastman's avatar
peastman committed
68

69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
static int getNumProcessors() {
#ifdef __APPLE__
    int ncpu;
    size_t len = 4;
    if (sysctlbyname("hw.logicalcpu", &ncpu, &len, NULL, 0) == 0)
       return ncpu;
    else
       return 1;
#else
#ifdef WIN32
    SYSTEM_INFO siSysInfo;
    int ncpu;
    GetSystemInfo(&siSysInfo);
    ncpu = siSysInfo.dwNumberOfProcessors;
    if (ncpu < 1)
        ncpu = 1;
    return ncpu;
#else
    long nProcessorsOnline = sysconf(_SC_NPROCESSORS_ONLN);
    if (nProcessorsOnline == -1)
        return 1;
    else
        return (int) nProcessorsOnline;
#endif
#endif
peastman's avatar
peastman committed
94
95
}

96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
// Define a function to check the CPU's capabilities.

#ifdef _WIN32
#define cpuid __cpuid
#else
static void cpuid(int cpuInfo[4], int infoType){
    __asm__ __volatile__ (
        "cpuid":
        "=a" (cpuInfo[0]),
        "=b" (cpuInfo[1]),
        "=c" (cpuInfo[2]),
        "=d" (cpuInfo[3]) :
        "a" (infoType)
    );
}
#endif

113
static void spreadCharge(int start, int end, float* posq, float* grid, int gridx, int gridy, int gridz, int numParticles, Vec3 periodicBoxSize) {
114
    float temp[4];
peastman's avatar
peastman committed
115
116
117
    __m128 boxSize = _mm_set_ps(0, (float) periodicBoxSize[2], (float) periodicBoxSize[1], (float) periodicBoxSize[0]);
    __m128 invBoxSize = _mm_set_ps(0, (float) (1/periodicBoxSize[2]), (float) (1/periodicBoxSize[1]), (float) (1/periodicBoxSize[0]));
    __m128 gridSize = _mm_set_ps(0, gridz, gridy, gridx);
peastman's avatar
peastman committed
118
    __m128i gridSizeInt = _mm_set_epi32(0, gridz, gridy, gridx);
peastman's avatar
peastman committed
119
120
121
    __m128 one  = _mm_set1_ps(1);
    __m128 scale = _mm_set1_ps(1.0f/(PME_ORDER-1));
    const float epsilonFactor = sqrt(ONE_4PI_EPS0);
122
    memset(grid, 0, sizeof(float)*gridx*gridy*gridz);
123
    for (int i = start; i < end; i++) {
peastman's avatar
peastman committed
124
125
126
127
128
        // Find the position relative to the nearest grid point.
        
        __m128 pos = _mm_load_ps(&posq[4*i]);
        __m128 posInBox = _mm_sub_ps(pos, _mm_mul_ps(boxSize, _mm_floor_ps(_mm_mul_ps(pos, invBoxSize))));
        __m128 t = _mm_mul_ps(_mm_mul_ps(posInBox, invBoxSize), gridSize);
peastman's avatar
peastman committed
129
        __m128i ti = _mm_cvttps_epi32(t);
peastman's avatar
peastman committed
130
        __m128 dr = _mm_sub_ps(t, _mm_cvtepi32_ps(ti));
peastman's avatar
peastman committed
131
        __m128i gridIndex = _mm_sub_epi32(ti, _mm_and_si128(gridSizeInt, _mm_cmpeq_epi32(ti, gridSizeInt)));
peastman's avatar
peastman committed
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
        
        // Compute the B-spline coefficients.
        
        __m128 data[PME_ORDER];
        data[PME_ORDER-1] = _mm_setzero_ps();
        data[1] = dr;
        data[0] = _mm_sub_ps(one, dr);
        for (int j = 3; j < PME_ORDER; j++) {
            __m128 div = _mm_set1_ps(1.0f/(j-1));
            data[j-1] = _mm_mul_ps(_mm_mul_ps(div, dr), data[j-2]);
            for (int k = 1; k < j-1; k++)
                data[j-k-1] = _mm_mul_ps(div, _mm_add_ps(_mm_mul_ps(_mm_add_ps(dr, _mm_set1_ps(k)), data[j-k-2]), _mm_mul_ps(_mm_sub_ps(_mm_set1_ps(j-k), dr), data[j-k-1])));
            data[0] = _mm_mul_ps(_mm_mul_ps(div, _mm_sub_ps(one, dr)), data[0]);
        }
        data[PME_ORDER-1] = _mm_mul_ps(_mm_mul_ps(scale, dr), data[PME_ORDER-2]);
        for (int j = 1; j < (PME_ORDER-1); j++)
            data[PME_ORDER-j-1] = _mm_mul_ps(scale, _mm_add_ps(_mm_mul_ps(_mm_add_ps(dr, _mm_set1_ps(j)), data[PME_ORDER-j-2]), _mm_mul_ps(_mm_sub_ps(_mm_set1_ps(PME_ORDER-j), dr), data[PME_ORDER-j-1])));
        data[0] = _mm_mul_ps(_mm_mul_ps(scale, _mm_sub_ps(one, dr)), data[0]);
        
        // Spread the charges.
        
        int gridIndexX = _mm_extract_epi32(gridIndex, 0);
        int gridIndexY = _mm_extract_epi32(gridIndex, 1);
        int gridIndexZ = _mm_extract_epi32(gridIndex, 2);
        float charge = epsilonFactor*posq[4*i+3];
peastman's avatar
peastman committed
157
        __m128 zdata0to3 = _mm_set_ps(extractFloat(data[3], 2), extractFloat(data[2], 2), extractFloat(data[1], 2), extractFloat(data[0], 2));
158
        float zdata4 = extractFloat(data[4], 2);
peastman's avatar
peastman committed
159
160
161
162
        for (int ix = 0; ix < PME_ORDER; ix++) {
            int xbase = gridIndexX+ix;
            xbase -= (xbase >= gridx ? gridx : 0);
            xbase = xbase*gridy*gridz;
163
            float xdata = extractFloat(data[ix], 0);
peastman's avatar
peastman committed
164
165
166
167
168

            for (int iy = 0; iy < PME_ORDER; iy++) {
                int ybase = gridIndexY+iy;
                ybase -= (ybase >= gridy ? gridy : 0);
                ybase = xbase + ybase*gridz;
169
                float multiplier = charge*xdata*extractFloat(data[iy], 1);
peastman's avatar
peastman committed
170
171
172

                __m128 add0to3 = _mm_mul_ps(zdata0to3, _mm_set1_ps(multiplier));
                if (gridIndexZ+4 < gridz)
173
                    _mm_storeu_ps(&grid[ybase+gridIndexZ], _mm_add_ps(_mm_loadu_ps(&grid[ybase+gridIndexZ]), add0to3));
peastman's avatar
peastman committed
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
                else {
                    _mm_store_ps(temp, add0to3);
                    int zindex = gridIndexZ;
                    grid[ybase+zindex] += temp[0];
                    zindex++;
                    zindex -= (zindex >= gridz ? gridz : 0);
                    grid[ybase+zindex] += temp[1];
                    zindex++;
                    zindex -= (zindex >= gridz ? gridz : 0);
                    grid[ybase+zindex] += temp[2];
                    zindex++;
                    zindex -= (zindex >= gridz ? gridz : 0);
                    grid[ybase+zindex] += temp[3];
                }
                int zindex = gridIndexZ+4;
                zindex -= (zindex >= gridz ? gridz : 0);
                grid[ybase+zindex] += multiplier*zdata4;
            }
        }
    }
}

196
static float reciprocalEnergy(int start, int end, fftwf_complex* grid, int gridx, int gridy, int gridz, double alpha, vector<float>* bsplineModuli, Vec3 periodicBoxSize) {
197
198
199
200
201
202
203
204
205
206
    const unsigned int yzsize = gridy*gridz;
    const unsigned int zsizeHalf = gridz/2+1;
    const unsigned int yzsizeHalf = gridy*zsizeHalf;
    const float scaleFactor = (float) (M_PI*periodicBoxSize[0]*periodicBoxSize[1]*periodicBoxSize[2]);
    const float recipExpFactor = (float) (M_PI*M_PI/(alpha*alpha));
    const float invPeriodicBoxSizeX = (float) (1.0/periodicBoxSize[0]);
    const float invPeriodicBoxSizeY = (float) (1.0/periodicBoxSize[1]);
    const float invPeriodicBoxSizeZ = (float) (1.0/periodicBoxSize[2]);
    float energy = 0.0f;

207
208
    int firstz = (start == 0 ? 1 : 0);
    for (int kx = start; kx < end; kx++) {
209
210
211
212
213
214
        int mx = (kx < (gridx+1)/2) ? kx : kx-gridx;
        float mhx = mx*invPeriodicBoxSizeX;
        float bx = scaleFactor*bsplineModuli[0][kx];
        for (int ky = 0; ky < gridy; ky++) {
            int my = (ky < (gridy+1)/2) ? ky : ky-gridy;
            float mhy = my*invPeriodicBoxSizeY;
215
216
            float mhx2y2 = mhx*mhx + mhy*mhy;
            float bxby = bx*bsplineModuli[1][ky];
217
218
219
220
            for (int kz = firstz; kz < gridz; kz++) {
                int mz = (kz < (gridz+1)/2) ? kz : kz-gridz;
                float mhz = mz*invPeriodicBoxSizeZ;
                float bz = bsplineModuli[2][kz];
221
222
                float m2 = mhx2y2 + mhz*mhz;
                float denom = m2*bxby*bz;
223
224
225
226
227
228
229
230
231
232
233
234
                float eterm = exp(-recipExpFactor*m2)/denom;
                int kx1, ky1, kz1;
                if (kz >= gridz/2+1) {
                    kx1 = (kx == 0 ? kx : gridx-kx);
                    ky1 = (ky == 0 ? ky : gridy-ky);
                    kz1 = gridz-kz;
                }
                else {
                    kx1 = kx;
                    ky1 = ky;
                    kz1 = kz;
                }
235
                int index = kx1*yzsizeHalf + ky1*zsizeHalf + kz1;
236
237
238
239
240
241
242
                float gridReal = grid[index][0];
                float gridImag = grid[index][1];
                energy += eterm*(gridReal*gridReal+gridImag*gridImag);
            }
            firstz = 0;
        }
    }
243
    return 0.5f*energy;
244
245
}

246
static void reciprocalConvolution(int start, int end, fftwf_complex* grid, int gridx, int gridy, int gridz, double alpha, vector<float>* bsplineModuli, Vec3 periodicBoxSize) {
247
248
249
250
251
252
253
254
    const unsigned int zsize = gridz/2+1;
    const unsigned int yzsize = gridy*zsize;
    const float scaleFactor = (float) (M_PI*periodicBoxSize[0]*periodicBoxSize[1]*periodicBoxSize[2]);
    const float recipExpFactor = (float) (M_PI*M_PI/(alpha*alpha));
    const float invPeriodicBoxSizeX = (float) (1.0/periodicBoxSize[0]);
    const float invPeriodicBoxSizeY = (float) (1.0/periodicBoxSize[1]);
    const float invPeriodicBoxSizeZ = (float) (1.0/periodicBoxSize[2]);

255
256
    int firstz = (start == 0 ? 1 : 0);
    for (int kx = start; kx < end; kx++) {
257
258
259
260
261
262
        int mx = (kx < (gridx+1)/2) ? kx : kx-gridx;
        float mhx = mx*invPeriodicBoxSizeX;
        float bx = scaleFactor*bsplineModuli[0][kx];
        for (int ky = 0; ky < gridy; ky++) {
            int my = (ky < (gridy+1)/2) ? ky : ky-gridy;
            float mhy = my*invPeriodicBoxSizeY;
263
264
            float mhx2y2 = mhx*mhx + mhy*mhy;
            float bxby = bx*bsplineModuli[1][ky];
265
266
267
268
269
            for (int kz = firstz; kz < zsize; kz++) {
                int index = kx*yzsize + ky*zsize + kz;
                int mz = (kz < (gridz+1)/2) ? kz : kz-gridz;
                float mhz = mz*invPeriodicBoxSizeZ;
                float bz = bsplineModuli[2][kz];
270
271
                float m2 = mhx2y2 + mhz*mhz;
                float denom = m2*bxby*bz;
272
                float eterm = exp(-recipExpFactor*m2)/denom;
273
274
                grid[index][0] *= eterm;
                grid[index][1] *= eterm;
275
276
277
278
279
280
            }
            firstz = 0;
        }
    }
}

281
static void interpolateForces(int start, int end, float* posq, float* force, float* grid, int gridx, int gridy, int gridz, int numParticles, Vec3 periodicBoxSize) {
282
283
284
    __m128 boxSize = _mm_set_ps(0, (float) periodicBoxSize[2], (float) periodicBoxSize[1], (float) periodicBoxSize[0]);
    __m128 invBoxSize = _mm_set_ps(0, (float) (1/periodicBoxSize[2]), (float) (1/periodicBoxSize[1]), (float) (1/periodicBoxSize[0]));
    __m128 gridSize = _mm_set_ps(0, gridz, gridy, gridx);
peastman's avatar
peastman committed
285
    __m128i gridSizeInt = _mm_set_epi32(0, gridz, gridy, gridx);
286
287
288
    __m128 one  = _mm_set1_ps(1);
    __m128 scale = _mm_set1_ps(1.0f/(PME_ORDER-1));
    const float epsilonFactor = sqrt(ONE_4PI_EPS0);
289
    for (int i = start; i < end; i++) {
290
291
292
293
294
        // Find the position relative to the nearest grid point.
        
        __m128 pos = _mm_load_ps(&posq[4*i]);
        __m128 posInBox = _mm_sub_ps(pos, _mm_mul_ps(boxSize, _mm_floor_ps(_mm_mul_ps(pos, invBoxSize))));
        __m128 t = _mm_mul_ps(_mm_mul_ps(posInBox, invBoxSize), gridSize);
peastman's avatar
peastman committed
295
        __m128i ti = _mm_cvttps_epi32(t);
296
        __m128 dr = _mm_sub_ps(t, _mm_cvtepi32_ps(ti));
peastman's avatar
peastman committed
297
        __m128i gridIndex = _mm_sub_epi32(ti, _mm_and_si128(gridSizeInt, _mm_cmpeq_epi32(ti, gridSizeInt)));
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
        
        // Compute the B-spline coefficients.
        
        __m128 data[PME_ORDER];
        __m128 ddata[PME_ORDER];
        data[PME_ORDER-1] = _mm_setzero_ps();
        data[1] = dr;
        data[0] = _mm_sub_ps(one, dr);
        for (int j = 3; j < PME_ORDER; j++) {
            __m128 div = _mm_set1_ps(1.0f/(j-1));
            data[j-1] = _mm_mul_ps(_mm_mul_ps(div, dr), data[j-2]);
            for (int k = 1; k < j-1; k++)
                data[j-k-1] = _mm_mul_ps(div, _mm_add_ps(_mm_mul_ps(_mm_add_ps(dr, _mm_set1_ps(k)), data[j-k-2]), _mm_mul_ps(_mm_sub_ps(_mm_set1_ps(j-k), dr), data[j-k-1])));
            data[0] = _mm_mul_ps(_mm_mul_ps(div, _mm_sub_ps(one, dr)), data[0]);
        }
        ddata[0] = _mm_sub_ps(_mm_set1_ps(0), data[0]);
        for (int j = 1; j < PME_ORDER; j++)
            ddata[j] = _mm_sub_ps(data[j-1], data[j]);
        data[PME_ORDER-1] = _mm_mul_ps(_mm_mul_ps(scale, dr), data[PME_ORDER-2]);
        for (int j = 1; j < (PME_ORDER-1); j++)
            data[PME_ORDER-j-1] = _mm_mul_ps(scale, _mm_add_ps(_mm_mul_ps(_mm_add_ps(dr, _mm_set1_ps(j)), data[PME_ORDER-j-2]), _mm_mul_ps(_mm_sub_ps(_mm_set1_ps(PME_ORDER-j), dr), data[PME_ORDER-j-1])));
        data[0] = _mm_mul_ps(_mm_mul_ps(scale, _mm_sub_ps(one, dr)), data[0]);
320
                
321
322
323
324
325
        // Compute the force on this atom.
        
        int gridIndexX = _mm_extract_epi32(gridIndex, 0);
        int gridIndexY = _mm_extract_epi32(gridIndex, 1);
        int gridIndexZ = _mm_extract_epi32(gridIndex, 2);
326
327
328
        __m128 zdata[PME_ORDER];
        for (int j = 0; j < PME_ORDER; j++)
            zdata[j] = _mm_set_ps(0, extractFloat(ddata[j], 2), extractFloat(data[j], 2), extractFloat(data[j], 2));
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
        __m128 f = _mm_set1_ps(0);
        for (int ix = 0; ix < PME_ORDER; ix++) {
            int xbase = gridIndexX+ix;
            xbase -= (xbase >= gridx ? gridx : 0);
            xbase = xbase*gridy*gridz;
            float dx = extractFloat(data[ix], 0);
            float ddx = extractFloat(ddata[ix], 0);
            __m128 xdata = _mm_set_ps(0, dx, dx, ddx);

            for (int iy = 0; iy < PME_ORDER; iy++) {
                int ybase = gridIndexY+iy;
                ybase -= (ybase >= gridy ? gridy : 0);
                ybase = xbase + ybase*gridz;
                float dy = extractFloat(data[iy], 1);
                float ddy = extractFloat(ddata[iy], 1);
                __m128 xydata = _mm_mul_ps(xdata, _mm_set_ps(0, dy, ddy, dy));

                for (int iz = 0; iz < PME_ORDER; iz++) {
                    int zindex = gridIndexZ+iz;
                    zindex -= (zindex >= gridz ? gridz : 0);
                    __m128 gridValue = _mm_set1_ps(grid[ybase+zindex]);
350
                    f = _mm_add_ps(f, _mm_mul_ps(xydata, _mm_mul_ps(zdata[iz], gridValue)));
351
352
353
                }
            }
        }
354
        f = _mm_mul_ps(invBoxSize, _mm_mul_ps(gridSize, _mm_mul_ps(f, _mm_set1_ps(-epsilonFactor*posq[4*i+3]))));
355
356
357
358
        _mm_store_ps(&force[4*i], _mm_add_ps(_mm_load_ps(&force[4*i]), f));        
    }
}

359
class CpuCalcPmeReciprocalForceKernel::ThreadData {
360
public:
361
    CpuCalcPmeReciprocalForceKernel& owner;
362
    int index;
363
    float* tempGrid;
364
    ThreadData(CpuCalcPmeReciprocalForceKernel& owner, int index) : owner(owner), index(index), tempGrid(NULL) {
365
366
367
368
    }
};

static void* threadBody(void* args) {
369
    CpuCalcPmeReciprocalForceKernel::ThreadData& data = *reinterpret_cast<CpuCalcPmeReciprocalForceKernel::ThreadData*>(args);
370
    data.owner.runThread(data.index);
371
372
    if (data.tempGrid != NULL)
        fftwf_free(data.tempGrid);
373
374
375
376
    delete &data;
    return 0;
}

377
void CpuCalcPmeReciprocalForceKernel::initialize(int gridx, int gridy, int gridz, int numParticles, double alpha) {
378
379
380
381
382
    if (!hasInitializedThreads) {
        numThreads = getNumProcessors();
        fftwf_init_threads();
        hasInitializedThreads = true;
    }
383
384
385
386
387
388
389
    this->gridx = gridx;
    this->gridy = gridy;
    this->gridz = gridz;
    this->numParticles = numParticles;
    this->alpha = alpha;
    force.resize(4*numParticles);
    
390
391
392
393
    // Initialize threads.
    
    pthread_cond_init(&startCondition, NULL);
    pthread_cond_init(&endCondition, NULL);
394
395
    pthread_cond_init(&mainThreadStartCondition, NULL);
    pthread_cond_init(&mainThreadEndCondition, NULL);
396
397
398
399
400
401
    pthread_mutex_init(&lock, NULL);
    thread.resize(numThreads);
    for (int i = 0; i < numThreads; i++) {
        ThreadData* data = new ThreadData(*this, i);
        threadData.push_back(data);
        pthread_create(&thread[i], NULL, threadBody, data);
402
        data->tempGrid = (float*) fftwf_malloc(sizeof(float)*(gridx*gridy*gridz+3));
403
    }
404
405
    pthread_create(&mainThread, NULL, threadBody, new ThreadData(*this, -1));
    
406
407
    // Initialize FFTW.
    
408
    realGrid = threadData[0]->tempGrid;
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
    complexGrid = (fftwf_complex*) fftwf_malloc(sizeof(fftwf_complex)*gridx*gridy*(gridz/2+1));
    fftwf_plan_with_nthreads(numThreads);
    forwardFFT = fftwf_plan_dft_r2c_3d(gridx, gridy, gridz, realGrid, complexGrid, FFTW_MEASURE);
    backwardFFT = fftwf_plan_dft_c2r_3d(gridx, gridy, gridz, complexGrid, realGrid, FFTW_MEASURE);
    hasCreatedPlan = true;
    
    // Initialize the b-spline moduli.

    int maxSize = max(max(gridx, gridy), gridz);
    vector<double> data(PME_ORDER);
    vector<double> ddata(PME_ORDER);
    vector<double> bsplinesData(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++)
        bsplinesData[i] = 0.0;
    for (int i = 1; i <= PME_ORDER; i++)
        bsplinesData[i] = data[i-1];

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

    bsplineModuli[0].resize(gridx);
    bsplineModuli[1].resize(gridy);
    bsplineModuli[2].resize(gridz);
    for (int dim = 0; dim < 3; dim++) {
        int ndata = bsplineModuli[dim].size();
        vector<float>& moduli = bsplineModuli[dim];
        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 += bsplinesData[j]*cos(arg);
                ss += bsplinesData[j]*sin(arg);
            }
            moduli[i] = (float) (sc*sc+ss*ss);
        }
        for (int i = 0; i < ndata; i++)
            if (moduli[i] < 1.0e-7f)
                moduli[i] = (moduli[i-1]+moduli[i+1])*0.5f;
    }
}

471
CpuCalcPmeReciprocalForceKernel::~CpuCalcPmeReciprocalForceKernel() {
472
473
    isDeleted = true;
    pthread_mutex_lock(&lock);
474
    pthread_cond_broadcast(&startCondition);
475
    pthread_cond_broadcast(&mainThreadStartCondition);
476
    pthread_mutex_unlock(&lock);
477
    for (int i = 0; i < (int) thread.size(); i++)
478
        pthread_join(thread[i], NULL);
479
    pthread_join(mainThread, NULL);
480
481
482
    pthread_mutex_destroy(&lock);
    pthread_cond_destroy(&startCondition);
    pthread_cond_destroy(&endCondition);
483
484
    pthread_cond_destroy(&mainThreadStartCondition);
    pthread_cond_destroy(&mainThreadEndCondition);
485
486
487
488
489
490
    if (complexGrid != NULL)
        fftwf_free(complexGrid);
    if (hasCreatedPlan) {
        fftwf_destroy_plan(forwardFFT);
        fftwf_destroy_plan(backwardFFT);
    }
491
492
}

peastman's avatar
peastman committed
493
494
495
496
497
#include <sys/time.h>
double diff(struct timeval t1, struct timeval t2) {
    return t2.tv_usec-t1.tv_usec+1e6*(t2.tv_sec-t1.tv_sec);
}

498
void CpuCalcPmeReciprocalForceKernel::runThread(int index) {
499
500
501
    if (index == -1) {
        // This is the main thread that coordinates all the other ones.
        
502
        pthread_mutex_lock(&lock);
503
504
505
506
507
        while (true) {
            // Wait for the signal to start.
            
            pthread_cond_wait(&mainThreadStartCondition, &lock);
            if (isDeleted)
508
509
                break;
            posq = io->getPosq();
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
            struct timeval t1, t2, t3, t4, t5, t6, t7;
            gettimeofday(&t1, NULL);
            advanceThreads(); // Signal threads to perform charge spreading.
            advanceThreads(); // Signal threads to sum the charge grids.
            gettimeofday(&t2, NULL);
            fftwf_execute_dft_r2c(forwardFFT, realGrid, complexGrid);
            gettimeofday(&t3, NULL);
            if (includeEnergy)
                advanceThreads(); // Signal threads to compute energy.
            gettimeofday(&t4, NULL);
            advanceThreads(); // Signal threads to perform reciprocal convolution.
            gettimeofday(&t5, NULL);
            fftwf_execute_dft_c2r(backwardFFT, complexGrid, realGrid);
            gettimeofday(&t6, NULL);
            advanceThreads(); // Signal threads to interpolate forces.
            isFinished = true;
            gettimeofday(&t7, NULL);
            printf("time %g %g %g %g %g %g\n", diff(t1, t2), diff(t2, t3), diff(t3, t4), diff(t4, t5), diff(t5, t6), diff(t6, t7));
            pthread_cond_signal(&mainThreadEndCondition);
        }
530
        pthread_mutex_unlock(&lock);
531
532
533
534
535
536
537
538
539
540
541
542
    }
    else {
        // This is a worker thread.
        
        int particleStart = (index*numParticles)/numThreads;
        int particleEnd = ((index+1)*numParticles)/numThreads;
        int gridxStart = (index*gridx)/numThreads;
        int gridxEnd = ((index+1)*gridx)/numThreads;
        int gridSize = (gridx*gridy*gridz+3)/4;
        int gridStart = 4*((index*gridSize)/numThreads);
        int gridEnd = 4*(((index+1)*gridSize)/numThreads);
        while (true) {
543
            threadWait();
544
            if (isDeleted)
545
                break;
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
            spreadCharge(particleStart, particleEnd, posq, threadData[index]->tempGrid, gridx, gridy, gridz, numParticles, periodicBoxSize);
            threadWait();
            int numGrids = threadData.size();
            for (int i = gridStart; i < gridEnd; i += 4) {
                __m128 sum = _mm_load_ps(&realGrid[i]);
                for (int j = 1; j < numGrids; j++)
                    sum = _mm_add_ps(sum, _mm_load_ps(&threadData[j]->tempGrid[i]));
                _mm_store_ps(&realGrid[i], sum);
            }
            threadWait();
            if (includeEnergy) {
                double threadEnergy = reciprocalEnergy(gridxStart, gridxEnd, complexGrid, gridx, gridy, gridz, alpha, bsplineModuli, periodicBoxSize);
                pthread_mutex_lock(&lock);
                energy += threadEnergy;
                pthread_mutex_unlock(&lock);
                threadWait();
            }
            reciprocalConvolution(gridxStart, gridxEnd, complexGrid, gridx, gridy, gridz, alpha, bsplineModuli, periodicBoxSize);
            threadWait();
            interpolateForces(particleStart, particleEnd, posq, &force[0], realGrid, gridx, gridy, gridz, numParticles, periodicBoxSize);
566
567
568
569
        }
    }
}

570
void CpuCalcPmeReciprocalForceKernel::threadWait() {
571
572
573
574
575
576
577
    pthread_mutex_lock(&lock);
    waitCount++;
    pthread_cond_signal(&endCondition);
    pthread_cond_wait(&startCondition, &lock);
    pthread_mutex_unlock(&lock);
}

578
void CpuCalcPmeReciprocalForceKernel::advanceThreads() {
579
580
581
582
583
584
585
    waitCount = 0;
    pthread_cond_broadcast(&startCondition);
    while (waitCount < numThreads) {
        pthread_cond_wait(&endCondition, &lock);
    }
}

586
void CpuCalcPmeReciprocalForceKernel::beginComputation(IO& io, Vec3 periodicBoxSize, bool includeEnergy) {
587
    this->io = &io;
588
589
590
    this->periodicBoxSize = periodicBoxSize;
    this->includeEnergy = includeEnergy;
    energy = 0.0;
591
    pthread_mutex_lock(&lock);
592
    isFinished = false;
593
594
595
596
    pthread_cond_signal(&mainThreadStartCondition);
    pthread_mutex_unlock(&lock);
}

597
double CpuCalcPmeReciprocalForceKernel::finishComputation(IO& io) {
598
599
600
601
602
603
    pthread_mutex_lock(&lock);
    while (!isFinished) {
        pthread_cond_wait(&mainThreadEndCondition, &lock);
    }
    pthread_mutex_unlock(&lock);
    io.setForce(&force[0]);
604
    return energy;
peastman's avatar
peastman committed
605
}
606
607
608
609
610
611
612
613
614
615

bool CpuCalcPmeReciprocalForceKernel::isProcessorSupported() {
    int cpuInfo[4];
    cpuid(cpuInfo, 0);
    if (cpuInfo[0] >= 1) {
        cpuid(cpuInfo, 1);
        return ((cpuInfo[2] & ((int) 1 << 19)) != 0); // Require SSE 4.1
    }
    return false;
}