customGBEnergyN2_nvidia.cl 8.17 KB
Newer Older
1
#define TILE_SIZE 32
2
3
#define STORE_DERIVATIVE_1(INDEX) derivBuffers##INDEX[offset1] += deriv##INDEX##_1;
#define STORE_DERIVATIVE_2(INDEX) derivBuffers##INDEX[offset2] += local_deriv##INDEX[get_local_id(0)];
4
5
6
7
8

/**
 * Compute a force based on pair interactions.
 */

9
10
__kernel __attribute__((reqd_work_group_size(WORK_GROUP_SIZE, 1, 1)))
void computeN2Energy(__global float4* forceBuffers, __global float* energyBuffer, __local float4* local_force,
11
	__global float4* posq, __local float4* local_posq, __global unsigned int* exclusions, __global unsigned int* exclusionIndices,
12
        __global unsigned int* exclusionRowIndices, __local float4* tempBuffer,
13
#ifdef USE_CUTOFF
14
        __global ushort2* tiles, __global unsigned int* interactionCount, float4 periodicBoxSize, float4 invPeriodicBoxSize, unsigned int maxTiles, __global unsigned int* interactionFlags
15
16
17
18
19
20
#else
        unsigned int numTiles
#endif
        PARAMETER_ARGUMENTS) {
    unsigned int totalWarps = get_global_size(0)/TILE_SIZE;
    unsigned int warp = get_global_id(0)/TILE_SIZE;
21
22
23
24
25
#ifdef USE_CUTOFF
    unsigned int numTiles = interactionCount[0];
    unsigned int pos = warp*(numTiles > maxTiles ? NUM_BLOCKS*(NUM_BLOCKS+1)/2 : numTiles)/totalWarps;
    unsigned int end = (warp+1)*(numTiles > maxTiles ? NUM_BLOCKS*(NUM_BLOCKS+1)/2 : numTiles)/totalWarps;
#else
26
27
    unsigned int pos = warp*numTiles/totalWarps;
    unsigned int end = (warp+1)*numTiles/totalWarps;
28
#endif
29
30
    float energy = 0.0f;
    unsigned int lasty = 0xFFFFFFFF;
31
32
    __local unsigned int exclusionRange[4];
    __local int exclusionIndex[2];
33
34
35

    while (pos < end) {
        // Extract the coordinates of this tile
36
        unsigned int x, y;
37
#ifdef USE_CUTOFF
38
39
40
41
        if (numTiles <= maxTiles) {
            ushort2 tileIndices = tiles[pos];
            x = tileIndices.x;
            y = tileIndices.y;
42
        }
43
        else
44
#endif
45
46
47
        {
            y = (unsigned 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);
48
49
            if (x < y || x >= NUM_BLOCKS) { // Occasionally happens due to roundoff error.
                y += (x < y ? -1 : 1);
50
51
52
                x = (pos-y*NUM_BLOCKS+y*(y+1)/2);
            }
        }
53
54
        unsigned int tgx = get_local_id(0) & (TILE_SIZE-1);
        unsigned int tbx = get_local_id(0) - tgx;
55
        unsigned int atom1 = x*TILE_SIZE + tgx;
56
57
58
        float4 force = 0.0f;
        float4 posq1 = posq[atom1];
        LOAD_ATOM1_PARAMETERS
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74

        // Locate the exclusion data for this tile.

#ifdef USE_EXCLUSIONS
        int localGroupIndex = get_local_id(0)/TILE_SIZE;
        if (tgx < 2)
            exclusionRange[2*localGroupIndex+tgx] = exclusionRowIndices[x+tgx];
        if (tgx == 0)
            exclusionIndex[localGroupIndex] = -1;
        for (int i = exclusionRange[2*localGroupIndex]+tgx; i < exclusionRange[2*localGroupIndex+1]; i += TILE_SIZE)
            if (exclusionIndices[i] == y)
                exclusionIndex[localGroupIndex] = i*TILE_SIZE;
        bool hasExclusions = (exclusionIndex[localGroupIndex] > -1);
#else
        bool hasExclusions = false;
#endif
75
76
77
        if (x == y) {
            // This tile is on the diagonal.

78
79
            const unsigned int localAtomIndex = get_local_id(0);
            local_posq[localAtomIndex] = posq1;
80
81
            LOAD_LOCAL_PARAMETERS_FROM_1
#ifdef USE_EXCLUSIONS
82
            unsigned int excl = exclusions[exclusionIndex[localGroupIndex]+tgx];
83
84
85
86
87
88
89
90
91
#endif
            for (unsigned int j = 0; j < TILE_SIZE; j++) {
#ifdef USE_EXCLUSIONS
                bool isExcluded = !(excl & 0x1);
#endif
                int atom2 = tbx+j;
                float4 posq2 = local_posq[atom2];
                float4 delta = (float4) (posq2.xyz - posq1.xyz, 0.0f);
#ifdef USE_PERIODIC
92
93
94
                delta.x -= floor(delta.x*invPeriodicBoxSize.x+0.5f)*periodicBoxSize.x;
                delta.y -= floor(delta.y*invPeriodicBoxSize.y+0.5f)*periodicBoxSize.y;
                delta.z -= floor(delta.z*invPeriodicBoxSize.z+0.5f)*periodicBoxSize.z;
95
#endif
96
97
98
99
                float r2 = delta.x*delta.x + delta.y*delta.y + delta.z*delta.z;
#ifdef USE_CUTOFF
                if (r2 < CUTOFF_SQUARED) {
#endif
100
                float r = SQRT(r2);
101
                LOAD_ATOM2_PARAMETERS
102
                atom2 = y*TILE_SIZE+j;
103
104
105
106
                float dEdR = 0.0f;
                float tempEnergy = 0.0f;
                if (atom1 < NUM_ATOMS && atom2 < NUM_ATOMS && atom1 != atom2) {
                    COMPUTE_INTERACTION
107
                    dEdR /= -r;
108
109
110
111
                }
                energy += 0.5f*tempEnergy;
                delta.xyz *= dEdR;
                force.xyz -= delta.xyz;
112
113
114
#ifdef USE_CUTOFF
                }
#endif
115
116
117
118
119
120
121
#ifdef USE_EXCLUSIONS
                excl >>= 1;
#endif
            }

            // Write results
#ifdef USE_OUTPUT_BUFFER_PER_BLOCK
122
            unsigned int offset1 = x*TILE_SIZE + tgx + x*PADDED_NUM_ATOMS;
123
#else
124
            unsigned int offset1 = x*TILE_SIZE + tgx + warp*PADDED_NUM_ATOMS;
125
#endif
126
127
            forceBuffers[offset1].xyz += force.xyz;
            STORE_DERIVATIVES_1
128
129
130
131
        }
        else {
            // This is an off-diagonal tile.

132
            const unsigned int localAtomIndex = get_local_id(0);
133
            if (lasty != y) {
134
                unsigned int j = y*TILE_SIZE + tgx;
135
                local_posq[localAtomIndex] = posq[j];
136
137
                LOAD_LOCAL_PARAMETERS_FROM_GLOBAL
            }
138
            local_force[localAtomIndex] = 0.0f;
139
            CLEAR_LOCAL_DERIVATIVES
140
#ifdef USE_CUTOFF
141
            unsigned int flags = (numTiles <= maxTiles ? interactionFlags[pos] : 0xFFFFFFFF);
142
143
            if (!hasExclusions && flags == 0) {
                // No interactions in this tile.
144
145
146
147
148
149
150
            }
            else
#endif
            {
                // Compute the full set of interactions in this tile.

#ifdef USE_EXCLUSIONS
151
                unsigned int excl = (hasExclusions ? exclusions[exclusionIndex[localGroupIndex]+tgx] : 0xFFFFFFFF);
152
153
154
155
156
157
158
159
160
161
162
                excl = (excl >> tgx) | (excl << (TILE_SIZE - tgx));
#endif
                unsigned int tj = tgx;
                for (unsigned int j = 0; j < TILE_SIZE; j++) {
#ifdef USE_EXCLUSIONS
                    bool isExcluded = !(excl & 0x1);
#endif
                    int atom2 = tbx+tj;
                    float4 posq2 = local_posq[atom2];
                    float4 delta = (float4) (posq2.xyz - posq1.xyz, 0.0f);
#ifdef USE_PERIODIC
163
164
165
                    delta.x -= floor(delta.x*invPeriodicBoxSize.x+0.5f)*periodicBoxSize.x;
                    delta.y -= floor(delta.y*invPeriodicBoxSize.y+0.5f)*periodicBoxSize.y;
                    delta.z -= floor(delta.z*invPeriodicBoxSize.z+0.5f)*periodicBoxSize.z;
166
#endif
167
168
169
170
                    float r2 = delta.x*delta.x + delta.y*delta.y + delta.z*delta.z;
#ifdef USE_CUTOFF
                    if (r2 < CUTOFF_SQUARED) {
#endif
171
                    float r = SQRT(r2);
172
                    LOAD_ATOM2_PARAMETERS
173
                    atom2 = y*TILE_SIZE+tj;
174
175
176
177
                    float dEdR = 0.0f;
                    float tempEnergy = 0.0f;
                    if (atom1 < NUM_ATOMS && atom2 < NUM_ATOMS) {
                        COMPUTE_INTERACTION
178
                        dEdR /= -r;
179
180
181
182
                    }
		    energy += tempEnergy;
                    delta.xyz *= dEdR;
                    force.xyz -= delta.xyz;
183
184
                    atom2 = tbx+tj;
                    local_force[atom2].xyz += delta.xyz;
185
                    RECORD_DERIVATIVE_2
186
187
188
#ifdef USE_CUTOFF
                    }
#endif
189
190
191
192
193
194
195
196
197
#ifdef USE_EXCLUSIONS
                    excl >>= 1;
#endif
                    tj = (tj + 1) & (TILE_SIZE - 1);
                }
            }

            // Write results
#ifdef USE_OUTPUT_BUFFER_PER_BLOCK
198
199
            unsigned int offset1 = x*TILE_SIZE + tgx + y*PADDED_NUM_ATOMS;
            unsigned int offset2 = y*TILE_SIZE + tgx + x*PADDED_NUM_ATOMS;
200
#else
201
202
            unsigned int offset1 = x*TILE_SIZE + tgx + warp*PADDED_NUM_ATOMS;
            unsigned int offset2 = y*TILE_SIZE + tgx + warp*PADDED_NUM_ATOMS;
203
204
205
#endif
            forceBuffers[offset1].xyz += force.xyz;
            forceBuffers[offset2].xyz += local_force[get_local_id(0)].xyz;
206
207
            STORE_DERIVATIVES_1
            STORE_DERIVATIVES_2
208
        }
209
        lasty = y;
210
211
212
213
        pos++;
    }
    energyBuffer[get_global_id(0)] += energy;
}