-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathkernels.cl
292 lines (246 loc) · 10.7 KB
/
kernels.cl
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
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
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
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
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
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
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
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
//#pragma OPENCL EXTENSION cl_khr_fp64 : enable
#define NSPEEDS 9
#define VECSIZE 2
#define GRIDSIZE NX/VECSIZE
#define I(jj,ii,sp) ((sp)*NX*NY+(ii)*NX+(jj))
kernel void accelerate_flow(global float* cells,
global int* obstacles)
{
/* compute weighting factors */
const float w1 = native_divide(DENSITY * ACCEL, 9.0f);
const float w2 = native_divide(DENSITY * ACCEL, 36.0f);
/* modify the 2nd row of the grid */
const int ii = NY - 2;
/* get column index */
int jj = get_global_id(0);
float res1 = cells[I(jj,ii,3)];
float res2 = cells[I(jj,ii,6)];
float res3 = cells[I(jj,ii,7)];
/* if the cell is not occupied and
** we don't send a negative density */
int mask = obstacles[ii*NX + jj]^1;
int mask1 = (res1-w1>0.0f) ? 1 : 0;
int mask2 = (res2-w2>0.0f) ? 1 : 0;
int mask3 = (res3-w2>0.0f) ? 1 : 0;
mask = mask & mask1 & mask2 & mask3;
/* increase 'east-side' densities */
cells[I(jj,ii,1)] = mad( mask, w1,cells[I(jj,ii,1)] );
cells[I(jj,ii,5)] = mad( mask, w2,cells[I(jj,ii,5)] );
cells[I(jj,ii,8)] = mad( mask, w2,cells[I(jj,ii,8)] );
/* decrease 'west-side' densities */
cells[I(jj,ii,3)] = mad( mask,-w1,res1 );
cells[I(jj,ii,6)] = mad( mask,-w2,res2 );
cells[I(jj,ii,7)] = mad( mask,-w2,res3 );
/* increase 'east-side' densities */
//cells[I(jj,ii,1)] = mask * w1 + cells[I(jj,ii,1)] ;
//cells[I(jj,ii,5)] = mask * w2 + cells[I(jj,ii,5)] ;
//cells[I(jj,ii,8)] = mask * w2 + cells[I(jj,ii,8)] ;
/* decrease 'west-side' densities */
//cells[I(jj,ii,3)] = res1 - mask * w1;
//cells[I(jj,ii,6)] = res2 - mask * w2;
//cells[I(jj,ii,7)] = res3 - mask * w2;
}
__kernel void timestep(__global float* restrict cells,
__global float* restrict tmp_cells,
__global int* restrict obstacles,
__local float* local_avgs,
__global float* partial_avgs, int step_mod) //remember to reduce partial_avg in a different kernel
{
//static const float c_sq = 1.0 / 3.0; /* square of speed of sound */
const float ic_sq = 3.0f;
//static const float ic_sq_sq = 9.0;
const float w0 = 0.4444444444444444444444f; /* weighting factor */
const float w1 = 0.1111111111111111111111f; /* weighting factor */
const float w2 = 0.0277777777777777777778f; /* weighting factor */
const unsigned int lookup[9][2] __attribute__((aligned(16))) = {{0,0},{3,1},{4,2},{1,3},{2,4},{7,5},{8,6},{5,7},{6,8}};
/* loop over the cells in the grid
** NB the collision step is called after
** the propagate step and so values of interest
** are in the scratch-space grid */
int ii = get_global_id(1);
int jj = get_global_id(0);
int local_ii = get_local_id(1);
int local_jj = get_local_id(0);
int local_nx = get_local_size(0);
int local_ny = get_local_size(1);
int local_size = mul24(local_nx,local_ny);
int item_id = mul24(local_ii, local_nx) + local_jj;
float tot_u = 0.0f;
//printf("y dimension:%d\n",ii);
float tmp[NSPEEDS*VECSIZE] __attribute__((aligned(16)));
int mask[VECSIZE] __attribute__((aligned(16)));
int y_n = ii + 1;
y_n = (y_n == NY) ? (0) : (y_n);
int y_s = (ii == 0) ? (NY-1) : (ii-1);
int grid,k,p;
for(grid = 0, k = 0;k<VECSIZE;grid+=GRIDSIZE,k++){
int xx = jj + grid;
int x_e = xx + 1;
x_e = (x_e >= NX) ? (x_e -= NX) : (x_e);
int x_w = (xx == 0) ? (NX - 1) : (xx-1);
tmp[k*NSPEEDS+0] = cells[ I(xx ,ii ,0) ];
tmp[k*NSPEEDS+1] = cells[ I(x_w,ii ,1) ];
tmp[k*NSPEEDS+2] = cells[ I(xx ,y_s,2) ];
tmp[k*NSPEEDS+3] = cells[ I(x_e,ii ,3) ];
tmp[k*NSPEEDS+4] = cells[ I(xx ,y_n,4) ];
tmp[k*NSPEEDS+5] = cells[ I(x_w,y_s,5) ];
tmp[k*NSPEEDS+6] = cells[ I(x_e,y_s,6) ];
tmp[k*NSPEEDS+7] = cells[ I(x_e,y_n,7) ];
tmp[k*NSPEEDS+8] = cells[ I(x_w,y_n,8) ];
mask[k] = obstacles[ii*NX+xx]^1;
}
for(grid = 0, p = 0;p<VECSIZE;grid+=GRIDSIZE,p++){
int k = p*NSPEEDS;
float densvec = tmp[k+0];
densvec += tmp[k+1];
densvec += tmp[k+2];
densvec += tmp[k+3];
densvec += tmp[k+4];
densvec += tmp[k+5];
densvec += tmp[k+6];
densvec += tmp[k+7];
densvec += tmp[k+8];
float densinv = native_recip(densvec);
float u_x = tmp[k+1] + tmp[k+5];
u_x += tmp[k+8];
u_x -= tmp[k+3];
u_x -= tmp[k+6];
u_x -= tmp[k+7];
float u_y = tmp[k+2] + tmp[k+5];
u_y += tmp[k+6];
u_y -= tmp[k+4];
u_y -= tmp[k+7];
u_y -= tmp[k+8];
float u_sq = u_x*u_x + u_y*u_y;
float uvec[NSPEEDS]; //try aligning
uvec[1] = u_x;
uvec[2] = u_y;
uvec[3] = - u_x;
uvec[4] = - u_y;
uvec[5] = u_x + u_y;
uvec[6] = - u_x + u_y;
uvec[7] = - u_x - u_y;
uvec[8] = u_x - u_y;
float ic_sqtimesu[NSPEEDS];
ic_sqtimesu[1] = uvec[1]*ic_sq;
ic_sqtimesu[2] = uvec[2]*ic_sq;
ic_sqtimesu[3] = uvec[3]*ic_sq;
ic_sqtimesu[4] = uvec[4]*ic_sq;
ic_sqtimesu[5] = uvec[5]*ic_sq;
ic_sqtimesu[6] = uvec[6]*ic_sq;
ic_sqtimesu[7] = uvec[7]*ic_sq;
ic_sqtimesu[8] = uvec[8]*ic_sq;
float ic_sqtimesu_sq[NSPEEDS];
ic_sqtimesu_sq[1] = ic_sqtimesu[1] * uvec[1];
ic_sqtimesu_sq[2] = ic_sqtimesu[2] * uvec[2];
ic_sqtimesu_sq[3] = ic_sqtimesu[3] * uvec[3];
ic_sqtimesu_sq[4] = ic_sqtimesu[4] * uvec[4];
ic_sqtimesu_sq[5] = ic_sqtimesu[5] * uvec[5];
ic_sqtimesu_sq[6] = ic_sqtimesu[6] * uvec[6];
ic_sqtimesu_sq[7] = ic_sqtimesu[7] * uvec[7];
ic_sqtimesu_sq[8] = ic_sqtimesu[8] * uvec[8];
float d_equ[NSPEEDS];
d_equ[0] = w0 * (densvec - 0.5f*densinv*ic_sq*u_sq);
d_equ[1] = w1 * (densvec + ic_sqtimesu[1] + 0.5f * densinv*ic_sq * (ic_sqtimesu_sq[1]-u_sq) );
d_equ[2] = w1 * (densvec + ic_sqtimesu[2] + 0.5f * densinv*ic_sq * (ic_sqtimesu_sq[2]-u_sq) );
d_equ[3] = w1 * (densvec + ic_sqtimesu[3] + 0.5f * densinv*ic_sq * (ic_sqtimesu_sq[3]-u_sq) );
d_equ[4] = w1 * (densvec + ic_sqtimesu[4] + 0.5f * densinv*ic_sq * (ic_sqtimesu_sq[4]-u_sq) );
d_equ[5] = w2 * (densvec + ic_sqtimesu[5] + 0.5f * densinv*ic_sq * (ic_sqtimesu_sq[5]-u_sq) );
d_equ[6] = w2 * (densvec + ic_sqtimesu[6] + 0.5f * densinv*ic_sq * (ic_sqtimesu_sq[6]-u_sq) );
d_equ[7] = w2 * (densvec + ic_sqtimesu[7] + 0.5f * densinv*ic_sq * (ic_sqtimesu_sq[7]-u_sq) );
d_equ[8] = w2 * (densvec + ic_sqtimesu[8] + 0.5f * densinv*ic_sq * (ic_sqtimesu_sq[8]-u_sq) );
int lmask = mask[p];
tmp_cells[I(jj+grid,ii,lookup[0][lmask])] = tmp[0+k] + lmask*OMEGA*(d_equ[0] - tmp[0+k]);
tmp_cells[I(jj+grid,ii,lookup[1][lmask])] = tmp[1+k] + lmask*OMEGA*(d_equ[1] - tmp[1+k]);
tmp_cells[I(jj+grid,ii,lookup[2][lmask])] = tmp[2+k] + lmask*OMEGA*(d_equ[2] - tmp[2+k]);
tmp_cells[I(jj+grid,ii,lookup[3][lmask])] = tmp[3+k] + lmask*OMEGA*(d_equ[3] - tmp[3+k]);
tmp_cells[I(jj+grid,ii,lookup[4][lmask])] = tmp[4+k] + lmask*OMEGA*(d_equ[4] - tmp[4+k]);
tmp_cells[I(jj+grid,ii,lookup[5][lmask])] = tmp[5+k] + lmask*OMEGA*(d_equ[5] - tmp[5+k]);
tmp_cells[I(jj+grid,ii,lookup[6][lmask])] = tmp[6+k] + lmask*OMEGA*(d_equ[6] - tmp[6+k]);
tmp_cells[I(jj+grid,ii,lookup[7][lmask])] = tmp[7+k] + lmask*OMEGA*(d_equ[7] - tmp[7+k]);
tmp_cells[I(jj+grid,ii,lookup[8][lmask])] = tmp[8+k] + lmask*OMEGA*(d_equ[8] - tmp[8+k]);
tot_u += lmask * native_sqrt(u_sq) * densinv;
}
local_avgs[item_id] = tot_u*FREE_CELLS_INV;
barrier(CLK_LOCAL_MEM_FENCE);
int group_id_X = get_group_id(0);
int group_id_Y = get_group_id(1);
int num_groups_X = get_num_groups(0);
int num_groups_Y = get_num_groups(1);
int groupID = mul24(group_id_Y, num_groups_X) + group_id_X;
if(local_size >= 128){
if (item_id<64) local_avgs[item_id] += local_avgs[item_id + 64];
barrier(CLK_LOCAL_MEM_FENCE);
}
//for(unsigned int s=local_size/2;s>32;s>>=1){
// if(item_id<s){
// local_avgs[item_id] += local_avgs[item_id + s];
// }
// barrier(CLK_LOCAL_MEM_FENCE);
//}
//No need to synchronise in the last warp
if(item_id < 32){
if(local_size>=64) local_avgs[item_id] += local_avgs[item_id + 32];
if(local_size>=32) local_avgs[item_id] += local_avgs[item_id + 16];
if(local_size>=16) local_avgs[item_id] += local_avgs[item_id + 8];
if(local_size>= 8) local_avgs[item_id] += local_avgs[item_id + 4];
if(local_size>= 4) local_avgs[item_id] += local_avgs[item_id + 2];
if(local_size>= 2) local_avgs[item_id] += local_avgs[item_id + 1];
}
if(item_id == 0) partial_avgs[step_mod*REDUCT_WIDTH+groupID] = local_avgs[0];
}
kernel void reduce(global float* partial_avgs,
local float* local_partial_avgs,
global float* avgs, int tt)
{
int group_idX = get_group_id(0);
int group_idY = get_group_id(1);
int local_id = get_local_id(0);
int local_size = get_local_size(0);
int num_groups = get_num_groups(0);
int k = 2*group_idX*local_size + local_id;
//local_partial_avgs[local_id] = 0.0f;
//for(int k=0;k<VECSIZE*GRIDSIZE;k+=GRIDSIZE){
local_partial_avgs[local_id] = partial_avgs[group_idY*REDUCT_WIDTH+k] + partial_avgs[group_idY*REDUCT_WIDTH+k+local_size];//reduce while copying from global to local
//}
barrier(CLK_LOCAL_MEM_FENCE);
if(local_size >= 512){
if(local_id<256) local_partial_avgs[local_id] += local_partial_avgs[local_id + 256];
barrier(CLK_LOCAL_MEM_FENCE);
}
if(local_size >= 256){
if(local_id<128) local_partial_avgs[local_id] += local_partial_avgs[local_id + 128];
barrier(CLK_LOCAL_MEM_FENCE);
}
if(local_size >= 128){
if(local_id<64) local_partial_avgs[local_id] += local_partial_avgs[local_id + 64];
barrier(CLK_LOCAL_MEM_FENCE);
}
//for(unsigned int s=local_size/2;s>0;s>>=1){
// if(local_id<s){
// local_partial_avgs[local_id] += local_partial_avgs[local_id + s];
// }
// barrier(CLK_LOCAL_MEM_FENCE);
//}
//No need to synchronise in the last warp
if(local_id < 32){
if(local_size >= 64) local_partial_avgs[local_id] += local_partial_avgs[local_id + 32];
if(local_size >= 32) local_partial_avgs[local_id] += local_partial_avgs[local_id + 16];
if(local_size >= 16) local_partial_avgs[local_id] += local_partial_avgs[local_id + 8];
if(local_size >= 8) local_partial_avgs[local_id] += local_partial_avgs[local_id + 4];
if(local_size >= 4) local_partial_avgs[local_id] += local_partial_avgs[local_id + 2];
if(local_size >= 2) local_partial_avgs[local_id] += local_partial_avgs[local_id + 1];
}
if(local_id == 0){
if(num_groups == 1)
avgs[tt+group_idY] = local_partial_avgs[0];
else
partial_avgs[group_idY*REDUCT_WIDTH+group_idX] = local_partial_avgs[0];
}
//if(gid == 0) {
// for(int i=0;i<size;i++){
// tmp += partial_avgs[i];
// }
// avgs[tt] = tmp;
//}
}