2: #include <petscsys.h>
  3: #include <petsctime.h>


  6: extern int BlastCache(void);
  7: extern int test1(void);
  8: extern int test2(void);

 12: int main(int argc,char **argv)
 13: {

 16:   PetscInitialize(&argc,&argv,0,0);

 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,&r);
 36:   PetscRandomSetFromOptions(r);
 37:   PetscMalloc1(20000,&x);
 38:   PetscMalloc1(20000,&y);

 40:   PetscMalloc1(2000,&z);
 41:   PetscMalloc1(2000,&zi);



 45:   /* Take care of paging effects */
 46:   PetscTime(&t1);

 48:   /* Form the random set of integers */
 49:   for (i=0; i<2000; i++) {
 50:     PetscRandomGetValue(r,&value);
 51:     intval = (int)(value*20000.0);
 52:     z[i]   = intval;
 53:   }

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

 62:   BlastCache();

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

 69:   BlastCache();

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

 81:   BlastCache();

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

 88:   BlastCache();

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

 95:   BlastCache();

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

102:   BlastCache();

104:   PetscTime(&t1);
105:   for (i=0; i<2000; i++) x[z[i]] = y[zi[i]];
106:   PetscTime(&t2);
107:   fprintf(stdout,"%-27s : %e sec\n","x[z[i]] = y[zi[i]]",(t2-t1)/2000.0);

109:   PetscMemcpy(x,y,10);
110:   PetscMemcpy(z,zi,10);
111:   PetscFree(z);
112:   PetscFree(zi);
113:   PetscFree(x);
114:   PetscFree(y);
115:   PetscRandomDestroy(&r);
116:   return(0);
117: }

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

129:   PetscRandomCreate(PETSC_COMM_SELF,&r);
130:   PetscRandomSetFromOptions(r);

132:   /* Take care of paging effects */
133:   PetscTime(&t1);

135:   for (i=0; i<20000; i++) {
136:     x[i]  = i;
137:     y[i]  = i;
138:     z[i]  = i;
139:     zi[i] = i;
140:   }

142:   /* Form the random set of integers */
143:   for (i=0; i<20000; i++) {
144:     PetscRandomGetValue(r,&value);
145:     intval    = (int)(value*20000.0);
146:     tmp       = z[i];
147:     z[i]      = z[intval];
148:     z[intval] = tmp;
149:   }

151:   for (i=0; i<20000; i++) {
152:     PetscRandomGetValue(r,&value);
153:     intval     = (int)(value*20000.0);
154:     tmp        = zi[i];
155:     zi[i]      = zi[intval];
156:     zi[intval] = tmp;
157:   }
158:   /* fprintf(stdout,"Done setup\n"); */

160:   /* BlastCache(); */

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

167:   /* BlastCache(); */

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

174:   /* BlastCache(); */

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

181:   /* BlastCache(); */

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


189:   PetscRandomDestroy(&r);
190:   return(0);
191: }

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

200:   PetscMalloc1(5*n,&x);
201:   y    = x + n;
202:   z    = y + n;
203:   a    = z + n;
204:   b    = a + n;

206:   for (i=0; i<n; i++) {
207:     a[i] = (PetscScalar) i;
208:     y[i] = (PetscScalar) i;
209:     z[i] = (PetscScalar) i;
210:     b[i] = (PetscScalar) i;
211:     x[i] = (PetscScalar) i;
212:   }

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