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