Actual source code: Index.c

  1: /*$Id: Index.c,v 1.35 2001/08/29 20:59:23 balay Exp $*/

 3:  #include petsc.h
 4:  #include petscsys.h

  6: extern int BlastCache(void);
  7: extern int test1(void);
  8: extern int test2(void);
  9: 
 12: int main(int argc,char **argv)
 13: {

 16:   PetscInitialize(&argc,&argv,0,0);
 17: 
 18:   test1();
 19:   test2();

 21:   PetscFinalize();
 22:   return(0);
 23: }

 27: int test1(void)
 28: {
 29:   PetscLogDouble  t1,t2;
 30:   double      value;
 31:   int         i,ierr,*z,*zi,intval;
 32:   PetscScalar *x,*y;
 33:   PetscRandom r;

 35:   PetscRandomCreate(PETSC_COMM_SELF,RANDOM_DEFAULT,&r);
 36:   PetscMalloc(20000*sizeof(PetscScalar),&x);
 37:   PetscMalloc(20000*sizeof(PetscScalar),&y);

 39:   PetscMalloc(2000*sizeof(int),&z);
 40:   PetscMalloc(2000*sizeof(int),&zi);



 44:   /* Take care of paging effects */
 45:   PetscGetTime(&t1);
 46: 
 47:    /* Form the random set of integers */
 48:   for (i=0; i<2000; i++) {
 49:     PetscRandomGetValue(r,&value);
 50:     intval = (int)(value*20000.0);
 51:     z[i]   = intval;
 52:   }

 54:   for (i=0; i<2000; i++) {
 55:     PetscRandomGetValue(r,&value);
 56:     intval  = (int)(value*20000.0);
 57:     zi[i]   = intval;
 58:   }
 59:   /* fprintf(stdout,"Done setup\n"); */

 61:   BlastCache();

 63:   PetscGetTime(&t1);
 64:   for (i=0; i<2000; i++) {  x[i] = y[i]; }
 65:   PetscGetTime(&t2);
 66:   fprintf(stdout,"%-27s : %e sec\n","x[i] = y[i]",(t2-t1)/2000.0);

 68:   BlastCache();

 70:   PetscGetTime(&t1);
 71:   for (i=0; i<500; i+=4) {
 72:     x[i]   = y[z[i]];
 73:     x[1+i] = y[z[1+i]];
 74:     x[2+i] = y[z[2+i]];
 75:     x[3+i] = y[z[3+i]];
 76:   }
 77:   PetscGetTime(&t2);
 78:   fprintf(stdout,"%-27s : %e sec\n","x[i] = y[idx[i]] - unroll 4",(t2-t1)/2000.0);

 80:   BlastCache();

 82:   PetscGetTime(&t1);CHKERRQ(ierr)
 83:   for (i=0; i<2000; i++) {  x[i] = y[z[i]]; }
 84:   PetscGetTime(&t2);
 85:   fprintf(stdout,"%-27s : %e sec\n","x[i] = y[idx[i]]",(t2-t1)/2000.0);

 87:   BlastCache();

 89:   PetscGetTime(&t1);
 90:   for (i=0; i<1000; i+=2) {  x[i] = y[z[i]];  x[1+i] = y[z[1+i]]; }
 91:   PetscGetTime(&t2);
 92:   fprintf(stdout,"%-27s : %e sec\n","x[i] = y[idx[i]] - unroll 2",(t2-t1)/2000.0);

 94:   BlastCache();

 96:   PetscGetTime(&t1);
 97:   for (i=0; i<2000; i++) {  x[z[i]] = y[i]; }
 98:   PetscGetTime(&t2);
 99:   fprintf(stdout,"%-27s : %e sec\n","x[z[i]] = y[i]",(t2-t1)/2000.0);

101:   BlastCache();

103:   PetscGetTime(&t1);
104:   for (i=0; i<2000; i++) {  x[z[i]] = y[zi[i]]; }
105:   PetscGetTime(&t2);
106:   fprintf(stdout,"%-27s : %e sec\n","x[z[i]] = y[zi[i]]",(t2-t1)/2000.0);
107: 
108:   PetscMemcpy(x,y,10);
109:   PetscMemcpy(z,zi,10);
110:   PetscFree(z);
111:   PetscFree(zi);
112:   PetscFree(x);
113:   PetscFree(y);
114:   PetscRandomDestroy(r);
115:   return(0);
116: }

120: int test2(void)
121: {
122:   PetscLogDouble   t1,t2;
123:   double       value;
124:   int          i,ierr,z[20000],zi[20000],intval,tmp;
125:   PetscScalar  x[20000],y[20000];
126:   PetscRandom  r;

128:   PetscRandomCreate(PETSC_COMM_SELF,RANDOM_DEFAULT,&r);

130:   /* Take care of paging effects */
131:   PetscGetTime(&t1);
132: 
133:   for (i=0; i<20000; i++) {
134:     x[i]  = i;
135:     y[i]  = i;
136:     z[i]  = i;
137:     zi[i] = i;
138:   }

140:    /* Form the random set of integers */
141:   for (i=0; i<20000; i++) {
142:     PetscRandomGetValue(r,&value);
143:     intval = (int)(value*20000.0);
144:     tmp    = z[i];
145:     z[i]   = z[intval];
146:     z[intval] = tmp;
147:   }

149:   for (i=0; i<20000; i++) {
150:     PetscRandomGetValue(r,&value);
151:     intval = (int)(value*20000.0);
152:     tmp    = zi[i];
153:     zi[i]  = zi[intval];
154:     zi[intval] = tmp;
155:   }
156:   /* fprintf(stdout,"Done setup\n"); */

158:   /* BlastCache(); */

160:   PetscGetTime(&t1);
161:   for (i=0; i<2000; i++) {  x[i] = y[i]; }
162:   PetscGetTime(&t2);
163:   fprintf(stdout,"%-27s : %e sec\n","x[i] = y[i]",(t2-t1)/2000.0);

165:   /* BlastCache(); */

167:   PetscGetTime(&t1);
168:   for (i=0; i<2000; i++) {  y[i] = x[z[i]]; }
169:   PetscGetTime(&t2);
170:   fprintf(stdout,"%-27s : %e sec\n","x[i] = y[idx[i]]",(t2-t1)/2000.0);

172:   /* BlastCache(); */

174:   PetscGetTime(&t1);
175:   for (i=0; i<2000; i++) {  x[z[i]] = y[i]; }
176:   PetscGetTime(&t2);
177:   fprintf(stdout,"%-27s : %e sec\n","x[z[i]] = y[i]",(t2-t1)/2000.0);

179:   /* BlastCache(); */

181:   PetscGetTime(&t1);
182:   for (i=0; i<2000; i++) {  y[z[i]] = x[zi[i]]; }
183:   PetscGetTime(&t2);
184:   fprintf(stdout,"%-27s : %e sec\n","x[z[i]] = y[zi[i]]",(t2-t1)/2000.0);


187:   PetscRandomDestroy(r);
188:   return(0);
189: }

193: int BlastCache(void)
194: {
195:   int    i,ierr,n = 1000000;
196:   PetscScalar *x,*y,*z,*a,*b;

198:   PetscMalloc(5*n*sizeof(PetscScalar),&x);
199:   y = x + n;
200:   z = y + n;
201:   a = z + n;
202:   b = a + n;

204:   for (i=0; i<n; i++) {
205:     a[i] = (PetscScalar) i;
206:     y[i] = (PetscScalar) i;
207:     z[i] = (PetscScalar) i;
208:     b[i] = (PetscScalar) i;
209:     x[i] = (PetscScalar) i;
210:   }

212:   for (i=0; i<n; i++) {
213:     a[i] = 3.0*x[i] + 2.0*y[i] + 3.3*z[i] - 25.*b[i];
214:   }
215:   for (i=0; i<n; i++) {
216:     b[i] = 3.0*x[i] + 2.0*y[i] + 3.3*a[i] - 25.*b[i];
217:   }
218:   for (i=0; i<n; i++) {
219:     z[i] = 3.0*x[i] + 2.0*y[i] + 3.3*a[i] - 25.*b[i];
220:   }
221:   PetscFree(x);
222:   return(0);
223: }