1 2 #include "petsc.h" 3 #include "petscsys.h" 4 5 extern int BlastCache(void); 6 extern int test1(void); 7 extern int test2(void); 8 9 #undef __FUNCT__ 10 #define __FUNCT__ "main" 11 int main(int argc,char **argv) 12 { 13 PetscErrorCode ierr; 14 15 PetscInitialize(&argc,&argv,0,0); 16 17 ierr = test1();CHKERRQ(ierr); 18 ierr = test2();CHKERRQ(ierr); 19 20 ierr = PetscFinalize();CHKERRQ(ierr); 21 PetscFunctionReturn(0); 22 } 23 24 #undef __FUNCT__ 25 #define __FUNCT__ "test1" 26 int test1(void) 27 { 28 PetscLogDouble t1,t2; 29 double value; 30 int i,ierr,*z,*zi,intval; 31 PetscScalar *x,*y; 32 PetscRandom r; 33 34 ierr = PetscRandomCreate(PETSC_COMM_SELF,RANDOM_DEFAULT,&r);CHKERRQ(ierr); 35 ierr = PetscMalloc(20000*sizeof(PetscScalar),&x);CHKERRQ(ierr); 36 ierr = PetscMalloc(20000*sizeof(PetscScalar),&y);CHKERRQ(ierr); 37 38 ierr = PetscMalloc(2000*sizeof(int),&z);CHKERRQ(ierr); 39 ierr = PetscMalloc(2000*sizeof(int),&zi);CHKERRQ(ierr); 40 41 42 43 /* Take care of paging effects */ 44 ierr = PetscGetTime(&t1);CHKERRQ(ierr); 45 46 /* Form the random set of integers */ 47 for (i=0; i<2000; i++) { 48 ierr = PetscRandomGetValue(r,&value);CHKERRQ(ierr); 49 intval = (int)(value*20000.0); 50 z[i] = intval; 51 } 52 53 for (i=0; i<2000; i++) { 54 ierr = PetscRandomGetValue(r,&value);CHKERRQ(ierr); 55 intval = (int)(value*20000.0); 56 zi[i] = intval; 57 } 58 /* fprintf(stdout,"Done setup\n"); */ 59 60 ierr = BlastCache();CHKERRQ(ierr); 61 62 ierr = PetscGetTime(&t1);CHKERRQ(ierr); 63 for (i=0; i<2000; i++) { x[i] = y[i]; } 64 ierr = PetscGetTime(&t2);CHKERRQ(ierr); 65 fprintf(stdout,"%-27s : %e sec\n","x[i] = y[i]",(t2-t1)/2000.0); 66 67 ierr = BlastCache();CHKERRQ(ierr); 68 69 ierr = PetscGetTime(&t1);CHKERRQ(ierr); 70 for (i=0; i<500; i+=4) { 71 x[i] = y[z[i]]; 72 x[1+i] = y[z[1+i]]; 73 x[2+i] = y[z[2+i]]; 74 x[3+i] = y[z[3+i]]; 75 } 76 ierr = PetscGetTime(&t2);CHKERRQ(ierr); 77 fprintf(stdout,"%-27s : %e sec\n","x[i] = y[idx[i]] - unroll 4",(t2-t1)/2000.0); 78 79 ierr = BlastCache();CHKERRQ(ierr); 80 81 ierr = PetscGetTime(&t1);CHKERRQ(ierr) 82 for (i=0; i<2000; i++) { x[i] = y[z[i]]; } 83 ierr = PetscGetTime(&t2);CHKERRQ(ierr); 84 fprintf(stdout,"%-27s : %e sec\n","x[i] = y[idx[i]]",(t2-t1)/2000.0); 85 86 ierr = BlastCache();CHKERRQ(ierr); 87 88 ierr = PetscGetTime(&t1);CHKERRQ(ierr); 89 for (i=0; i<1000; i+=2) { x[i] = y[z[i]]; x[1+i] = y[z[1+i]]; } 90 ierr = PetscGetTime(&t2);CHKERRQ(ierr); 91 fprintf(stdout,"%-27s : %e sec\n","x[i] = y[idx[i]] - unroll 2",(t2-t1)/2000.0); 92 93 ierr = BlastCache();CHKERRQ(ierr); 94 95 ierr = PetscGetTime(&t1);CHKERRQ(ierr); 96 for (i=0; i<2000; i++) { x[z[i]] = y[i]; } 97 ierr = PetscGetTime(&t2);CHKERRQ(ierr); 98 fprintf(stdout,"%-27s : %e sec\n","x[z[i]] = y[i]",(t2-t1)/2000.0); 99 100 ierr = BlastCache();CHKERRQ(ierr); 101 102 ierr = PetscGetTime(&t1);CHKERRQ(ierr); 103 for (i=0; i<2000; i++) { x[z[i]] = y[zi[i]]; } 104 ierr = PetscGetTime(&t2);CHKERRQ(ierr); 105 fprintf(stdout,"%-27s : %e sec\n","x[z[i]] = y[zi[i]]",(t2-t1)/2000.0); 106 107 ierr = PetscMemcpy(x,y,10);CHKERRQ(ierr); 108 ierr = PetscMemcpy(z,zi,10);CHKERRQ(ierr); 109 ierr = PetscFree(z);CHKERRQ(ierr); 110 ierr = PetscFree(zi);CHKERRQ(ierr); 111 ierr = PetscFree(x);CHKERRQ(ierr); 112 ierr = PetscFree(y);CHKERRQ(ierr); 113 PetscRandomDestroy(r); 114 PetscFunctionReturn(0); 115 } 116 117 #undef __FUNCT__ 118 #define __FUNCT__ "test2" 119 int test2(void) 120 { 121 PetscLogDouble t1,t2; 122 double value; 123 int i,ierr,z[20000],zi[20000],intval,tmp; 124 PetscScalar x[20000],y[20000]; 125 PetscRandom r; 126 127 ierr = PetscRandomCreate(PETSC_COMM_SELF,RANDOM_DEFAULT,&r);CHKERRQ(ierr); 128 129 /* Take care of paging effects */ 130 ierr = PetscGetTime(&t1);CHKERRQ(ierr); 131 132 for (i=0; i<20000; i++) { 133 x[i] = i; 134 y[i] = i; 135 z[i] = i; 136 zi[i] = i; 137 } 138 139 /* Form the random set of integers */ 140 for (i=0; i<20000; i++) { 141 ierr = PetscRandomGetValue(r,&value);CHKERRQ(ierr); 142 intval = (int)(value*20000.0); 143 tmp = z[i]; 144 z[i] = z[intval]; 145 z[intval] = tmp; 146 } 147 148 for (i=0; i<20000; i++) { 149 ierr = PetscRandomGetValue(r,&value);CHKERRQ(ierr); 150 intval = (int)(value*20000.0); 151 tmp = zi[i]; 152 zi[i] = zi[intval]; 153 zi[intval] = tmp; 154 } 155 /* fprintf(stdout,"Done setup\n"); */ 156 157 /* ierr = BlastCache();CHKERRQ(ierr); */ 158 159 ierr = PetscGetTime(&t1);CHKERRQ(ierr); 160 for (i=0; i<2000; i++) { x[i] = y[i]; } 161 ierr = PetscGetTime(&t2);CHKERRQ(ierr); 162 fprintf(stdout,"%-27s : %e sec\n","x[i] = y[i]",(t2-t1)/2000.0); 163 164 /* ierr = BlastCache();CHKERRQ(ierr); */ 165 166 ierr = PetscGetTime(&t1);CHKERRQ(ierr); 167 for (i=0; i<2000; i++) { y[i] = x[z[i]]; } 168 ierr = PetscGetTime(&t2);CHKERRQ(ierr); 169 fprintf(stdout,"%-27s : %e sec\n","x[i] = y[idx[i]]",(t2-t1)/2000.0); 170 171 /* ierr = BlastCache();CHKERRQ(ierr); */ 172 173 ierr = PetscGetTime(&t1);CHKERRQ(ierr); 174 for (i=0; i<2000; i++) { x[z[i]] = y[i]; } 175 ierr = PetscGetTime(&t2);CHKERRQ(ierr); 176 fprintf(stdout,"%-27s : %e sec\n","x[z[i]] = y[i]",(t2-t1)/2000.0); 177 178 /* ierr = BlastCache();CHKERRQ(ierr); */ 179 180 ierr = PetscGetTime(&t1);CHKERRQ(ierr); 181 for (i=0; i<2000; i++) { y[z[i]] = x[zi[i]]; } 182 ierr = PetscGetTime(&t2);CHKERRQ(ierr); 183 fprintf(stdout,"%-27s : %e sec\n","x[z[i]] = y[zi[i]]",(t2-t1)/2000.0); 184 185 186 PetscRandomDestroy(r); 187 PetscFunctionReturn(0); 188 } 189 190 #undef __FUNCT__ 191 #define __FUNCT__ "BlastCache" 192 int BlastCache(void) 193 { 194 int i,ierr,n = 1000000; 195 PetscScalar *x,*y,*z,*a,*b; 196 197 ierr = PetscMalloc(5*n*sizeof(PetscScalar),&x);CHKERRQ(ierr); 198 y = x + n; 199 z = y + n; 200 a = z + n; 201 b = a + n; 202 203 for (i=0; i<n; i++) { 204 a[i] = (PetscScalar) i; 205 y[i] = (PetscScalar) i; 206 z[i] = (PetscScalar) i; 207 b[i] = (PetscScalar) i; 208 x[i] = (PetscScalar) i; 209 } 210 211 for (i=0; i<n; i++) { 212 a[i] = 3.0*x[i] + 2.0*y[i] + 3.3*z[i] - 25.*b[i]; 213 } 214 for (i=0; i<n; i++) { 215 b[i] = 3.0*x[i] + 2.0*y[i] + 3.3*a[i] - 25.*b[i]; 216 } 217 for (i=0; i<n; i++) { 218 z[i] = 3.0*x[i] + 2.0*y[i] + 3.3*a[i] - 25.*b[i]; 219 } 220 ierr = PetscFree(x);CHKERRQ(ierr); 221 PetscFunctionReturn(0); 222 } 223