customGBValueN2_cpu.cc 14.1 KB
Newer Older
1
2
3
/**
 * Compute a value based on pair interactions.
 */
4
KERNEL void computeN2Value(GLOBAL const real4* RESTRICT posq, GLOBAL const unsigned int* RESTRICT exclusions,
5
        GLOBAL const int2* exclusionTiles,
6
#ifdef SUPPORTS_64_BIT_ATOMICS
7
        GLOBAL mm_ulong* RESTRICT global_value,
8
#else
9
        GLOBAL real* RESTRICT global_value,
10
#endif
11
#ifdef USE_CUTOFF
12
13
14
        GLOBAL const int* RESTRICT tiles, GLOBAL const unsigned int* RESTRICT interactionCount, real4 periodicBoxSize, real4 invPeriodicBoxSize,
        real4 periodicBoxVecX, real4 periodicBoxVecY, real4 periodicBoxVecZ, unsigned int maxTiles, GLOBAL const real4* RESTRICT blockCenter,
        GLOBAL const real4* RESTRICT blockSize, GLOBAL const int* RESTRICT interactingAtoms
15
16
17
18
#else
        unsigned int numTiles
#endif
        PARAMETER_ARGUMENTS) {
19
20
21
    LOCAL real3 local_pos[LOCAL_BUFFER_SIZE];
    LOCAL real local_value[LOCAL_BUFFER_SIZE];
    ATOM_PARAMETER_DATA
22
23
24

    // First loop: process tiles that contain exclusions.
    
25
26
    const int firstExclusionTile = FIRST_EXCLUSION_TILE+get_group_id(0)*(LAST_EXCLUSION_TILE-FIRST_EXCLUSION_TILE)/get_num_groups(0);
    const int lastExclusionTile = FIRST_EXCLUSION_TILE+(get_group_id(0)+1)*(LAST_EXCLUSION_TILE-FIRST_EXCLUSION_TILE)/get_num_groups(0);
27
    for (int pos = firstExclusionTile; pos < lastExclusionTile; pos++) {
28
        const int2 tileIndices = exclusionTiles[pos];
29
30
31
32
33
34
35
        const unsigned int x = tileIndices.x;
        const unsigned int y = tileIndices.y;

        // Load the data for this tile.

        for (int localAtomIndex = 0; localAtomIndex < TILE_SIZE; localAtomIndex++) {
            unsigned int j = y*TILE_SIZE + localAtomIndex;
36
            local_pos[localAtomIndex] = trimTo3(posq[j]);
37
38
39
40
41
42
43
44
45
46
47
            LOAD_LOCAL_PARAMETERS_FROM_GLOBAL
        }
        if (x == y) {
            // This tile is on the diagonal.

            for (unsigned int tgx = 0; tgx < TILE_SIZE; tgx++) {
#ifdef USE_EXCLUSIONS
                unsigned int excl = exclusions[pos*TILE_SIZE+tgx];
#endif
                unsigned int atom1 = x*TILE_SIZE+tgx;
                real value = 0;
48
                real3 pos1 = trimTo3(posq[atom1]);
49
50
                LOAD_ATOM1_PARAMETERS
                for (unsigned int j = 0; j < TILE_SIZE; j++) {
51
52
                    real3 pos2 = local_pos[j];
                    real3 delta = make_real3(pos2.x-pos1.x, pos2.y-pos1.y, pos2.z-pos1.z);
53
#ifdef USE_PERIODIC
54
                    APPLY_PERIODIC_TO_DELTA(delta)
55
56
#endif
                    real r2 = dot(delta.xyz, delta.xyz);
57
#ifdef USE_CUTOFF
58
59
60
                    if (r2 < CUTOFF_SQUARED) {
#endif
                        real invR = RSQRT(r2);
peastman's avatar
peastman committed
61
                        real r = r2*invR;
62
63
64
65
66
67
68
69
                        unsigned int atom2 = j;
                        LOAD_ATOM2_PARAMETERS
                        atom2 = y*TILE_SIZE+j;
                        real tempValue1 = 0;
                        real tempValue2 = 0;
#ifdef USE_EXCLUSIONS
                        bool isExcluded = (atom1 >= NUM_ATOMS || atom2 >= NUM_ATOMS || !(excl & 0x1));
                        if (!isExcluded && atom1 != atom2) {
70
#else
71
                        if (atom1 < NUM_ATOMS && atom2 < NUM_ATOMS && atom1 != atom2) {
72
#endif
73
74
75
                            COMPUTE_VALUE
                        }
                        value += tempValue1;
76
                        ADD_TEMP_DERIVS1
77
#ifdef USE_CUTOFF
78
                    }
79
80
#endif
#ifdef USE_EXCLUSIONS
81
                    excl >>= 1;
82
#endif
83
                }
84

85
                // Write results.
86

87
#ifdef SUPPORTS_64_BIT_ATOMICS
peastman's avatar
peastman committed
88
                unsigned int offset1 = atom1;
89
                ATOMIC_ADD(&global_value[offset1], (mm_ulong) realToFixedPoint(value));
90
#else
peastman's avatar
peastman committed
91
92
                unsigned int offset1 = atom1 + get_group_id(0)*PADDED_NUM_ATOMS;
                global_value[offset1] += value;
93
#endif
peastman's avatar
peastman committed
94
                STORE_PARAM_DERIVS1
95
96
            }
        }
97
98
        else {
            // This is an off-diagonal tile.
99

100
101
            for (int tgx = 0; tgx < TILE_SIZE; tgx++)
                local_value[tgx] = 0;
102
103
            for (unsigned int tgx = 0; tgx < TILE_SIZE; tgx++) {
#ifdef USE_EXCLUSIONS
104
                unsigned int excl = exclusions[pos*TILE_SIZE+tgx];
105
106
#endif
                unsigned int atom1 = x*TILE_SIZE+tgx;
107
                real value = 0;
108
                real3 pos1 = trimTo3(posq[atom1]);
109
110
                LOAD_ATOM1_PARAMETERS
                for (unsigned int j = 0; j < TILE_SIZE; j++) {
111
112
                    real3 pos2 = local_pos[j];
                    real3 delta = make_real3(pos2.x-pos1.x, pos2.y-pos1.y, pos2.z-pos1.z);
113
#ifdef USE_PERIODIC
114
                    APPLY_PERIODIC_TO_DELTA(delta)
115
#endif
116
                    real r2 = dot(delta.xyz, delta.xyz);
117
118
119
#ifdef USE_CUTOFF
                    if (r2 < CUTOFF_SQUARED) {
#endif
120
                        real invR = RSQRT(r2);
peastman's avatar
peastman committed
121
                        real r = r2*invR;
122
123
124
125
126
                        unsigned int atom2 = j;
                        LOAD_ATOM2_PARAMETERS
                        atom2 = y*TILE_SIZE+j;
                        real tempValue1 = 0;
                        real tempValue2 = 0;
127
#ifdef USE_EXCLUSIONS
128
129
                        bool isExcluded = (atom1 >= NUM_ATOMS || atom2 >= NUM_ATOMS || !(excl & 0x1));
                        if (!isExcluded) {
130
#else
131
                        if (atom1 < NUM_ATOMS && atom2 < NUM_ATOMS && atom1 != atom2) {
132
#endif
133
134
135
136
                            COMPUTE_VALUE
                        }
                        value += tempValue1;
                        local_value[j] += tempValue2;
137
138
                        ADD_TEMP_DERIVS1
                        ADD_TEMP_DERIVS2
139
140
141
142
143
144
145
146
#ifdef USE_CUTOFF
                    }
#endif
#ifdef USE_EXCLUSIONS
                    excl >>= 1;
#endif
                }

147
                // Write results for atom1.
148

149
#ifdef SUPPORTS_64_BIT_ATOMICS
150
                unsigned int offset1 = atom1;
151
                ATOMIC_ADD(&global_value[offset1], (mm_ulong) realToFixedPoint(value));
152
#else
153
154
                unsigned int offset1 = atom1 + get_group_id(0)*PADDED_NUM_ATOMS;
                global_value[offset1] += value;
155
#endif
156
                STORE_PARAM_DERIVS1
157
158
159
160
161
162
            }

            // Write results.

            for (int tgx = 0; tgx < TILE_SIZE; tgx++) {
#ifdef SUPPORTS_64_BIT_ATOMICS
163
                unsigned int offset2 = y*TILE_SIZE+tgx;
164
                ATOMIC_ADD(&global_value[offset2], (mm_ulong) realToFixedPoint(local_value[tgx]));
165
#else
166
167
                unsigned int offset2 = y*TILE_SIZE+tgx + get_group_id(0)*PADDED_NUM_ATOMS;
                global_value[offset2] += local_value[tgx];
168
#endif
169
                STORE_PARAM_DERIVS2
170
171
            }
        }
172
    }
173

174
175
    // Second loop: tiles without exclusions, either from the neighbor list (with cutoff) or just enumerating all
    // of them (no cutoff).
176

177
178
#ifdef USE_CUTOFF
    const unsigned int numTiles = interactionCount[0];
179
180
    if (numTiles > maxTiles)
        return; // There wasn't enough memory for the neighbor list.
181
182
    int pos = (int) (get_group_id(0)*(numTiles > maxTiles ? NUM_BLOCKS*((mm_long)NUM_BLOCKS+1)/2 : numTiles)/get_num_groups(0));
    int end = (int) ((get_group_id(0)+1)*(numTiles > maxTiles ? NUM_BLOCKS*((mm_long)NUM_BLOCKS+1)/2 : numTiles)/get_num_groups(0));
183
#else
184
185
    int pos = (int) (get_group_id(0)*(mm_long)numTiles/get_num_groups(0));
    int end = (int) ((get_group_id(0)+1)*(mm_long)numTiles/get_num_groups(0));
186
187
188
#endif
    int nextToSkip = -1;
    int currentSkipIndex = 0;
189
    LOCAL int atomIndices[TILE_SIZE];
190
191
192
193
194
195

    while (pos < end) {
        bool includeTile = true;
        
        // Extract the coordinates of this tile.
        
196
        int x, y;
197
198
        bool singlePeriodicCopy = false;
#ifdef USE_CUTOFF
199
200
201
202
203
204
205
206
207
208
        x = tiles[pos];
        real4 blockSizeX = blockSize[x];
        singlePeriodicCopy = (0.5f*periodicBoxSize.x-blockSizeX.x >= CUTOFF &&
                              0.5f*periodicBoxSize.y-blockSizeX.y >= CUTOFF &&
                              0.5f*periodicBoxSize.z-blockSizeX.z >= CUTOFF);
#else
        y = (int) floor(NUM_BLOCKS+0.5f-SQRT((NUM_BLOCKS+0.5f)*(NUM_BLOCKS+0.5f)-2*pos));
        x = (pos-y*NUM_BLOCKS+y*(y+1)/2);
        if (x < y || x >= NUM_BLOCKS) { // Occasionally happens due to roundoff error.
            y += (x < y ? -1 : 1);
209
            x = (pos-y*NUM_BLOCKS+y*(y+1)/2);
210
        }
211

212
        // Skip over tiles that have exclusions, since they were already processed.
213

214
215
        while (nextToSkip < pos) {
            if (currentSkipIndex < NUM_TILES_WITH_EXCLUSIONS) {
216
                int2 tile = exclusionTiles[currentSkipIndex++];
217
                nextToSkip = tile.x + tile.y*NUM_BLOCKS - tile.y*(tile.y+1)/2;
218
            }
219
220
            else
                nextToSkip = end;
221
        }
222
223
        includeTile = (nextToSkip != pos);
#endif
224
225
226
227
228
        if (includeTile) {
            // Load the data for this tile.

            for (int localAtomIndex = 0; localAtomIndex < TILE_SIZE; localAtomIndex++) {
#ifdef USE_CUTOFF
peastman's avatar
peastman committed
229
                unsigned int j = interactingAtoms[pos*TILE_SIZE+localAtomIndex];
230
231
232
233
234
#else
                unsigned int j = y*TILE_SIZE+localAtomIndex;
#endif
                atomIndices[localAtomIndex] = j;
                if (j < PADDED_NUM_ATOMS) {
235
                    local_pos[localAtomIndex] = trimTo3(posq[j]);
236
237
238
239
                    LOAD_LOCAL_PARAMETERS_FROM_GLOBAL
                    local_value[localAtomIndex] = 0;
                }
            }
240
#ifdef USE_PERIODIC
241
242
243
244
245
246
            if (singlePeriodicCopy) {
                // The box is small enough that we can just translate all the atoms into a single periodic
                // box, then skip having to apply periodic boundary conditions later.

                real4 blockCenterX = blockCenter[x];
                for (unsigned int tgx = 0; tgx < TILE_SIZE; tgx++)
247
                    APPLY_PERIODIC_TO_POS_WITH_CENTER(local_pos[tgx], blockCenterX)
248
249
250
                for (unsigned int tgx = 0; tgx < TILE_SIZE; tgx++) {
                    unsigned int atom1 = x*TILE_SIZE+tgx;
                    real value = 0;
251
252
                    real3 pos1 = trimTo3(posq[atom1]);
                    APPLY_PERIODIC_TO_POS_WITH_CENTER(pos1, blockCenterX)
253
254
                    LOAD_ATOM1_PARAMETERS
                    for (unsigned int j = 0; j < TILE_SIZE; j++) {
255
256
                        real3 pos2 = local_pos[j];
                        real3 delta = make_real3(pos2.x-pos1.x, pos2.y-pos1.y, pos2.z-pos1.z);
257
258
259
                        real r2 = dot(delta.xyz, delta.xyz);
                        if (atom1 < NUM_ATOMS && atomIndices[j] < NUM_ATOMS && r2 < CUTOFF_SQUARED) {
                            real invR = RSQRT(r2);
peastman's avatar
peastman committed
260
                            real r = r2*invR;
261
262
263
264
265
266
267
268
                            unsigned int atom2 = j;
                            LOAD_ATOM2_PARAMETERS
                            atom2 = atomIndices[j];
                            real tempValue1 = 0;
                            real tempValue2 = 0;
                            COMPUTE_VALUE
                            value += tempValue1;
                            local_value[j] += tempValue2;
269
270
                            ADD_TEMP_DERIVS1
                            ADD_TEMP_DERIVS2
271
                        }
272
                    }
273

274
                    // Write results for atom1.
275

276
#ifdef SUPPORTS_64_BIT_ATOMICS
peastman's avatar
peastman committed
277
                    unsigned int offset1 = atom1;
278
                    ATOMIC_ADD(&global_value[offset1], (mm_ulong) realToFixedPoint(value));
279
#else
peastman's avatar
peastman committed
280
281
                    unsigned int offset1 = atom1 + get_group_id(0)*PADDED_NUM_ATOMS;
                    global_value[offset1] += value;
282
#endif
peastman's avatar
peastman committed
283
                    STORE_PARAM_DERIVS1
284
285
286
287
288
                }
            }
            else
#endif
            {
289
                // We need to apply periodic boundary conditions separately for each interaction.
290
291
292

                for (unsigned int tgx = 0; tgx < TILE_SIZE; tgx++) {
                    unsigned int atom1 = x*TILE_SIZE+tgx;
293
                    real value = 0;
294
                    real3 pos1 = trimTo3(posq[atom1]);
295
296
                    LOAD_ATOM1_PARAMETERS
                    for (unsigned int j = 0; j < TILE_SIZE; j++) {
297
298
                        real3 pos2 = local_pos[j];
                        real3 delta = make_real3(pos2.x-pos1.x, pos2.y-pos1.y, pos2.z-pos1.z);
299
#ifdef USE_PERIODIC
300
                        APPLY_PERIODIC_TO_DELTA(delta)
301
#endif
302
                        real r2 = dot(delta.xyz, delta.xyz);
303
#ifdef USE_CUTOFF
304
                        if (atom1 < NUM_ATOMS && atomIndices[j] < NUM_ATOMS && r2 < CUTOFF_SQUARED) {
305
#else
306
                        if (atom1 < NUM_ATOMS && atomIndices[j] < NUM_ATOMS) {
307
#endif
308
                            real invR = RSQRT(r2);
peastman's avatar
peastman committed
309
                            real r = r2*invR;
310
311
312
313
314
                            unsigned int atom2 = j;
                            LOAD_ATOM2_PARAMETERS
                            atom2 = atomIndices[j];
                            real tempValue1 = 0;
                            real tempValue2 = 0;
315
                            COMPUTE_VALUE
316
317
                            value += tempValue1;
                            local_value[j] += tempValue2;
318
319
                            ADD_TEMP_DERIVS1
                            ADD_TEMP_DERIVS2
320
321
322
323
324
                        }
                    }

                    // Write results for atom1.

325
#ifdef SUPPORTS_64_BIT_ATOMICS
326
                    unsigned int offset1 = atom1;
327
                    ATOMIC_ADD(&global_value[offset1], (mm_ulong) realToFixedPoint(value));
328
#else
329
330
                    unsigned int offset1 = atom1 + get_group_id(0)*PADDED_NUM_ATOMS;
                    global_value[offset1] += value;
331
#endif
332
                    STORE_PARAM_DERIVS1
333
334
335
                }
            }

336
            // Write results.
337
338

            for (int tgx = 0; tgx < TILE_SIZE; tgx++) {
339
340
341
342
343
344
345
#ifdef USE_CUTOFF
                unsigned int atom2 = atomIndices[tgx];
#else
                unsigned int atom2 = y*TILE_SIZE + tgx;
#endif
                if (atom2 < PADDED_NUM_ATOMS) {
#ifdef SUPPORTS_64_BIT_ATOMICS
346
                    unsigned int offset2 = atom2;
347
                    ATOMIC_ADD(&global_value[offset2], (mm_ulong) realToFixedPoint(local_value[tgx]));
348
#else
349
350
                    unsigned int offset2 = atom2 + get_group_id(0)*PADDED_NUM_ATOMS;
                    global_value[offset2] += local_value[tgx];
351
#endif
352
                    STORE_PARAM_DERIVS2
353
                }
354
355
356
357
358
            }
        }
        pos++;
    }
}