/
pttSYMSGD.c
312 lines (268 loc) · 9.5 KB
/
pttSYMSGD.c
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
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
#include <stdio.h>
#include <stdlib.h>
#include <math.h>
#include <omp.h>
#include "mkl.h"
#include "pttSymSGD.h"
#define PAGESIZE 4096
int panels, T, nt, panelSz;
float prev_err = 0.0;
// X: a MxD matrix, Y: a M vector, W: a M vector
// W0: a M vector
int main(int argc, char ** argv){
if (argc>1 && argv[1][0]=='h') {
printf ("Usage: parSymSGD M D T C lamda r\n");
printf (" M: number of data points, D: dimensions, T: time iterations, C: cores;\n");
printf (" lamda: learning rate, r: panel size in unit of C.\n");
return 1;
}u
// read in the arguments: M, D, I (time iterations), C (cores), r (each panel contains r*C points)
int M = argc>1?atoi(argv[1]):32;
int D = argc>2?atoi(argv[2]):4;
T = argc>3?atoi(argv[3]):10;
int C = argc>4?atoi(argv[4]):4;
float lamda = argc>5?atof(argv[5]):0.01;
int r = argc>6?atoi(argv[6]):1;
///printf("M=%d, D=%d, T=%d, C=%d, lamda=%8.6f, r=%d\n",M,D,T,C,lamda,r);
int max_threads = mkl_get_max_threads(); // get the max number of threads
int rep;
mkl_set_num_threads(1); // set the number of threads to use by mkl
panelSz = C*r;
panels = M/panelSz;
int i,j,k,p,t;
float *Y, *Wreal, *W, *X;
Y = (float *) mkl_malloc(M*sizeof(float),PAGESIZE);
Wreal = (float *) mkl_malloc(D*sizeof(float),PAGESIZE);
W = (float *) mkl_malloc(D*sizeof(float),PAGESIZE);
X = (float *) mkl_malloc(M*D*sizeof(float),PAGESIZE);
float *Ypred = (float*)mkl_malloc(M*sizeof(float),PAGESIZE);
float *Ytmp = (float*)mkl_malloc(M*sizeof(float),PAGESIZE);
float *I = (float*)mkl_malloc(D*D*sizeof(float),PAGESIZE);
float *Z = (float*)mkl_malloc(M*D*sizeof(float),PAGESIZE);
float *B = (float*)mkl_malloc(panels*D*sizeof(float),PAGESIZE);
if (Y==NULL | Wreal==NULL | W==NULL | X==NULL | Ypred==NULL || Ytmp==NULL || Z==NULL || B==NULL || I== NULL){
printf("Memory allocation error.\n");
return 2;
}
initData(Wreal,W,X,Y, M, D,I);
///printf("panelSz=%d, panels=%d\n", panelSz, panels);
for (nt=1; nt<=max_threads && nt<=panelSz; nt*=2){
omp_set_num_threads(nt);// set the number of openMP threads
for (rep=0; rep<REPEATS; rep++){//repeat measurements
double prepTime, gdTime, sInit;
// preprocessing
sInit=dsecnd();
//preprocessSeq(X, Y, Z, B, panelSz, panels, M, D, lamda);
preprocessPar(X, Y, Z, B, panelSz, panels, M, D, lamda);
prepTime = (dsecnd() - sInit);
///dump2("Z",Z,M,D);
///dump2("B",B,panels,D);
// GD
initW(W,D);
///dump1("W (initial)", W, D);
sInit=dsecnd();
float err;
float fixpoint = 0.0;
for (t=0;t<T;t++){
for (p=0;p<panels;p++){
gd(&(X[p*panelSz*D]),&(Z[p*panelSz*D]), &(B[p*D]), panelSz, D, lamda, W, I);
///printf("(t=%d, p=%d) ",t,p);
///dump1("W", W, D);
///err=calErr(X, Ypred, Ytmp, Y, W, M, D);
printf("finish one panels ............................ \n");
}
}
gdTime = (dsecnd() - sInit);
err=calErr(X, Ypred, Ytmp, Y, W, M, D);
fixpoint = err - prev_err;
// print final err. time is in milliseconds
printf("nt=%d\t ttlTime=%.5f\t prepTime=%.5f\t gdTime=%.5f\t error=%.5f\n", nt, (gdTime+prepTime)*1000, prepTime*1000, gdTime*1000, err);
}
}
if (B) mkl_free(B);
if (Z) mkl_free(Z);
if (Ytmp) mkl_free(Ytmp);
if (Ypred) mkl_free(Ypred);
if (Y) mkl_free(Y);
if (Wreal) mkl_free(Wreal);
if (W) mkl_free(W);
if (X) mkl_free(X);
if (I) mkl_free(I);
return 0;
}
void initData(float*Wreal, float*W, float*X, float*Y, int M, int D, float*I){
int i;
srand(1);
float u1=4;// upper bound of X element values
float u2=0.1; // upper bound of noise in Y
for (i=0;i<D;i++){
Wreal[i] = (float)floor(u1*rand()/RAND_MAX-u1/2);
}
for (i=0;i<M;i++){
Y[i] = u2*rand()/RAND_MAX;
}
for (i=0;i<M*D;i++){
X[i] = u1*rand()/RAND_MAX;
}
float alpha = 1.0;
float beta = 0;
///dump1("Wreal", Wreal, D);
///dump1("Y (initial)", Y, M);
///dump2("X", X, M,D);
// Y= X*Wreal + Y
cblas_sgemv(CblasRowMajor, CblasNoTrans, M, D, alpha, X, D, Wreal, 1, 1, Y, 1);
///dump1("Y (real)", Y, M);
//printf("finish the initial \n");
}
void preprocessPar(float *X, float*Y, float *Z, float *B, int panelSz, int panels, int M, int D, float lamda){
int p,i;
for (i=0;i<panels*D;i++)
B[i] = 0;
// this line will need some changes to be more space efficient (e.g., applying the paralleliszation to inside a panel rather than across panels.)
//printf("nt=%d\n", nt);
float *XX = (float*) malloc(nt*panelSz*panelSz*sizeof(float)); //holder of X[j]X[j] for one panel
if (XX==NULL){
printf("Memory allocation error.\n");
return 5;
}
int chunkSz=panels/nt;
#pragma omp parallel for private(i,p)
for (i=0;i<nt;i++){
for (p=i*chunkSz;p<(i+1)*chunkSz;p++){
///printf("preprocess panel %d...\n", p);
///printf("%f\n", X[p*panelSz*D]);
preprocess1(panelSz, D, &(XX[i*panelSz*panelSz]), &(X[p*panelSz*D]), &(Y[p*panelSz]), lamda, &(Z[p*panelSz*D]), &(B[p*D]));
}
}
if (XX) free(XX);
}
void preprocess1(int panelSz, int D, float *XX, float*X, float*Y, float lamda, float*Z, float*B){
// panelSz: the number of points to process in each round
int i,j;
// step 1: compute all X[i]'*X[j] (i>j)
for (i=panelSz-1;i>0;i--)
for (j=i-1;j>=0;j--){
XX[i*panelSz+j] = cblas_sdot(D, &(X[i*D]), 1, &(X[j*D]), 1);
// printf("XX[%d]=%8.4f, X[%d]=%8.4f, X[%d]=%8.4f\n", i*panelSz+j, XX[i*panelSz+j], i*D, X[i*D], j*D, X[j*D]);
}
// step 2: compute all Z vectors
// Z0=lamda*X[0], B=lamda*X[0]*Y[0]
cblas_scopy(D, X, 1, Z, 1);
cblas_sscal(D, lamda, Z, 1);
float alpha=lamda*Y[0];
cblas_scopy(D, X, 1, B, 1);
cblas_sscal(D, alpha, B, 1);
for (i=1; i<panelSz;i++){
cblas_scopy(D, &(X[i*D]), 1, &(Z[i*D]),1);
// Z[i] = lamda*(X[i] - sum_{j<i} XX[i,j]*Z[j]);
for (j=0;j<i;j++){
cblas_saxpy(D, -1*XX[i*panelSz+j], &(Z[j*D]), 1, &(Z[i*D]), 1);
}
cblas_sscal(D, lamda, &(Z[i*D]), 1);
// B = lamda*(Y[i] - X[i]*B) X[i] + B;
float alpha = lamda*(Y[i]-cblas_sdot(D, &(X[i*D]), 1, B, 1));
cblas_saxpy(D, alpha, &(X[i*D]), 1, B, 1);
}
}
void gd(float *X, float* Z, float*B, int panelSz, int D, float lamda, float * W, float *I){
// each D elements in Z forms a z vector
// W=(I-sum_{j}Z[j]X[j])W[0]+B
int i,j,k,m,n,l;
float temp;
float *Wtmp = (float*)malloc(D*sizeof(float));
memset(Wtmp,NULL,D*sizeof(float));
// every iteration should re-initial I
for(i=0;i<D*D;i++){
I[i]= 0.0;
//I[i*D+i]= 1.0;
}
for(i=0;i<D;i++){
I[i*D+i]= 1.0;
}
// for(i=0;i<D;i++){
// printf("W[%d]= %8.4f \n", i, W[i]);
// }
int chunkSz= D/nt;
float *XZ = (float*) malloc(D*D*sizeof(float)); //holder of matrix XZ
memset(XZ,NULL,D*D*sizeof(float));
int PchunkSz = panelSz/nt;
#pragma omp parallel for schedule(static) private(j,i)
for(k=0;k<nt;k++){
for(i=0;i<D;i++){
for(m=0;m<D;m++){
temp = 0;
for(j=k*PchunkSz;j<(k+1)*PchunkSz;j++){
temp += X[i*panelSz+j]*Z[j*D+m];
}
XZ[i*D+m] = temp;
//printf("XZ[%i*D+%j] = %8.4f \n",i,j,XZ[i*D+j]);
}
}
}
#pragma omp parallel for schedule(static) private(i)
for(k=0;k<nt;k++){
for(i=k*chunkSz;i<(k+1)*chunkSz;i++){
// printf("first,I[%d]=%8.4f \n", i, I[i]);
for(j=0;j<chunkSz*D;j++)
I[i*D+j]+= -XZ[i*D+j];
}
}
cblas_saxpy(D,1,B, 1, Wtmp, 1);
cblas_scopy(D, Wtmp, 1, W, 1);
for(i=0;i<D;i++){
//printf("Wtmp[%d]=%8.4f \n",i,Wtmp[i]);
}
free(XZ);
free(Wtmp);
}
float calErr(float *data, float *Ypred, float *Ytmp, float* Y, float* W, int M, int D){
cblas_sgemv(CblasRowMajor, CblasNoTrans, M, D, 1, data, D, W, 1, 0, Ypred, 1);
cblas_scopy(M,Ypred,1,Ytmp,1);
cblas_saxpy(M,-1,Y,1,Ytmp,1);
return cblas_sdot(M,Ytmp,1,Ytmp,1);
}
void initW(float*W, int D){
// initialize W
int i;
srand(2);
float u1=4;
for (i=0;i<D;i++){
W[i] = (float)floor(u1*rand()/RAND_MAX-u1/2);
}
}
void dump1(char*s, float*V, int D){
printf("%s:\n", s);
int i;
for (i=0;i<D;i++)
printf("%8.4f ", V[i]);
printf("\n-------------\n");
}
void dump2(char*s, float*A, int M, int D){
printf("%s:\n", s);
int i,j;
for (i=0;i<M;i++){
printf("%d:\t", i);
for (j=0;j<D;j++){
float f = A[i*D+j];
printf("%8.4f ", f);
}
printf("\n");
}
printf("--------------\n");
}
void preprocessSeq(float *X, float*Y, float *Z, float *B, int panelSz, int panels, int M, int D, float lamda){
int p,i;
for (i=0;i<panels*D;i++)
B[i] = 0;
float *XX = (float*) malloc(panelSz*panelSz*sizeof(float)); //holder of X[j]X[j] for one panel
if (XX==NULL){
printf("Memory allocation error.\n");
return 5;
}
for (p=0;p<panels;p++){
///printf("preprocess panel %d...\n", p);
///printf("%f\n", X[p*panelSz]);
preprocess1(panelSz, D, XX, &(X[p*panelSz*D]), &(Y[p*panelSz]), lamda, &(Z[p*panelSz*D]), &(B[p*D]));
}
if (XX) free(XX);
}