1 2 #include <sys/time.h> 3 /* int gettimeofday(struct timeval *tp, struct timezone *tzp); */ 4 5 double second() 6 { 7 /* struct timeval { long tv_sec; 8 long tv_usec; }; 9 10 struct timezone { int tz_minuteswest; 11 int tz_dsttime; }; */ 12 13 struct timeval tp; 14 struct timezone tzp; 15 int i; 16 17 i = gettimeofday(&tp,&tzp); 18 return ( (double) tp.tv_sec + (double) tp.tv_usec * 1.e-6 ); 19 } 20 # include <stdio.h> 21 # include <math.h> 22 # include <float.h> 23 # include <limits.h> 24 # include <sys/time.h> 25 26 /* 27 * Program: Stream 28 * Programmer: Joe R. Zagar 29 * Revision: 4.0-BETA, October 24, 1995 30 * Original code developed by John D. McCalpin 31 * 32 * This program measures memory transfer rates in MB/s for simple 33 * computational kernels coded in C. These numbers reveal the quality 34 * of code generation for simple uncacheable kernels as well as showing 35 * the cost of floating-point operations relative to memory accesses. 36 * 37 * INSTRUCTIONS: 38 * 39 * 1) Stream requires a good bit of memory to run. Adjust the 40 * value of 'N' (below) to give a 'timing calibration' of 41 * at least 20 clock-ticks. This will provide rate estimates 42 * that should be good to about 5% precision. 43 */ 44 45 # define N 2000000 46 # define NTIMES 50 47 # define OFFSET 0 48 49 /* 50 * 3) Compile the code with full optimization. Many compilers 51 * generate unreasonably bad code before the optimizer tightens 52 * things up. If the results are unreasonably good, on the 53 * other hand, the optimizer might be too smart for me! 54 * 55 * Try compiling with: 56 * cc -O stream_d.c second.c -o stream_d -lm 57 * 58 * This is known to work on Cray, SGI, IBM, and Sun machines. 59 * 60 * 61 * 4) Mail the results to mccalpin@cs.virginia.edu 62 * Be sure to include: 63 * a) computer hardware model number and software revision 64 * b) the compiler flags 65 * c) all of the output from the test case. 66 * Thanks! 67 * 68 */ 69 70 # define HLINE "-------------------------------------------------------------\n" 71 72 # ifndef MIN 73 # define MIN(x,y) ((x)<(y)?(x):(y)) 74 # endif 75 # ifndef MAX 76 # define MAX(x,y) ((x)>(y)?(x):(y)) 77 # endif 78 79 static double a[N+OFFSET], 80 b[N+OFFSET], 81 c[N+OFFSET]; 82 /*double *a,*b,*c;*/ 83 84 static double rmstime[4] = {0}, maxtime[4] = {0}, 85 mintime[4] = {FLT_MAX,FLT_MAX,FLT_MAX,FLT_MAX}; 86 87 static char *label[4] = {"Copy: ", "Scale: ", 88 "Add: ", "Triad: "}; 89 90 static double bytes[4] = { 91 2 * sizeof(double) * N, 92 2 * sizeof(double) * N, 93 3 * sizeof(double) * N, 94 3 * sizeof(double) * N 95 }; 96 97 extern double second(); 98 99 #include <mpi.h> 100 101 int main(int argc,char **args) 102 { 103 int quantum, checktick(); 104 int BytesPerWord; 105 register int j, k; 106 double scalar, t, times[4][NTIMES]; 107 108 MPI_Init(&argc,&args); 109 /* --- SETUP --- determine precision and check timing --- */ 110 111 printf(HLINE); 112 BytesPerWord = sizeof(double); 113 printf("This system uses %d bytes per DOUBLE PRECISION word.\n", 114 BytesPerWord); 115 116 printf(HLINE); 117 printf("Array size = %d, Offset = %d\n" , N, OFFSET); 118 printf("Total memory required = %.1f MB.\n", 119 (3 * N * BytesPerWord) / 1048576.0); 120 printf("Each test is run %d times, but only\n", NTIMES); 121 printf("the *best* time for each is used.\n"); 122 123 /* Get initial value for system clock. */ 124 125 /* a = malloc(N*sizeof(double)); 126 b = malloc(N*sizeof(double)); 127 c = malloc(N*sizeof(double));*/ 128 for (j=0; j<N; j++) { 129 a[j] = 1.0; 130 b[j] = 2.0; 131 c[j] = 0.0; 132 } 133 134 printf(HLINE); 135 136 if ( (quantum = checktick()) >= 1) 137 printf("Your clock granularity/precision appears to be " 138 "%d microseconds.\n", quantum); 139 else 140 printf("Your clock granularity appears to be " 141 "less than one microsecond.\n"); 142 143 t = second(); 144 for (j = 0; j < N; j++) 145 a[j] = 2.0E0 * a[j]; 146 t = 1.0E6 * (second() - t); 147 148 printf("Each test below will take on the order" 149 " of %d microseconds.\n", (int) t ); 150 printf(" (= %d clock ticks)\n", (int) (t/quantum) ); 151 printf("Increase the size of the arrays if this shows that\n"); 152 printf("you are not getting at least 20 clock ticks per test.\n"); 153 154 printf(HLINE); 155 156 157 /* --- MAIN LOOP --- repeat test cases NTIMES times --- */ 158 159 scalar = 3.0; 160 for (k=0; k<NTIMES; k++) 161 { 162 MPI_Barrier(MPI_COMM_WORLD); 163 times[0][k] = second(); 164 MPI_Barrier(MPI_COMM_WORLD); 165 for (j=0; j<N; j++) 166 c[j] = a[j]; 167 MPI_Barrier(MPI_COMM_WORLD); 168 times[0][k] = second() - times[0][k]; 169 170 times[1][k] = second(); 171 MPI_Barrier(MPI_COMM_WORLD); 172 for (j=0; j<N; j++) 173 b[j] = scalar*c[j]; 174 MPI_Barrier(MPI_COMM_WORLD); 175 times[1][k] = second() - times[1][k]; 176 177 times[2][k] = second(); 178 MPI_Barrier(MPI_COMM_WORLD); 179 for (j=0; j<N; j++) 180 c[j] = a[j]+b[j]; 181 MPI_Barrier(MPI_COMM_WORLD); 182 times[2][k] = second() - times[2][k]; 183 184 times[3][k] = second(); 185 MPI_Barrier(MPI_COMM_WORLD); 186 for (j=0; j<N; j++) 187 a[j] = b[j]+scalar*c[j]; 188 MPI_Barrier(MPI_COMM_WORLD); 189 times[3][k] = second() - times[3][k]; 190 } 191 192 /* --- SUMMARY --- */ 193 194 for (k=0; k<NTIMES; k++) 195 { 196 for (j=0; j<4; j++) 197 { 198 rmstime[j] = rmstime[j] + (times[j][k] * times[j][k]); 199 mintime[j] = MIN(mintime[j], times[j][k]); 200 maxtime[j] = MAX(maxtime[j], times[j][k]); 201 } 202 } 203 204 printf("Function Rate (MB/s) RMS time Min time Max time\n"); 205 for (j=0; j<4; j++) { 206 rmstime[j] = sqrt(rmstime[j]/(double)NTIMES); 207 208 printf("%s%11.4f %11.4f %11.4f %11.4f\n", label[j], 209 1.0E-06 * bytes[j]/mintime[j], 210 rmstime[j], 211 mintime[j], 212 maxtime[j]); 213 } 214 MPI_Finalize(); 215 return 0; 216 } 217 218 # define M 20 219 220 int 221 checktick() 222 { 223 int i, minDelta, Delta; 224 double t1, t2, timesfound[M]; 225 226 /* Collect a sequence of M unique time values from the system. */ 227 228 for (i = 0; i < M; i++) { 229 t1 = second(); 230 while( ((t2=second()) - t1) < 1.0E-6 ) 231 ; 232 timesfound[i] = t1 = t2; 233 } 234 235 /* 236 * Determine the minimum difference between these M values. 237 * This result will be our estimate (in microseconds) for the 238 * clock granularity. 239 */ 240 241 minDelta = 1000000; 242 for (i = 1; i < M; i++) { 243 Delta = (int)( 1.0E6 * (timesfound[i]-timesfound[i-1])); 244 minDelta = MIN(minDelta, MAX(Delta,0)); 245 } 246 247 return(minDelta); 248 } 249 250