|
| 1 | +/* |
| 2 | + * clock.c - Routines for using the cycle counters on x86, |
| 3 | + * Alpha, and Sparc boxes. |
| 4 | + * |
| 5 | + * Copyright (c) 2002, R. Bryant and D. O'Hallaron, All rights reserved. |
| 6 | + * May not be used, modified, or copied without permission. |
| 7 | + */ |
| 8 | + |
| 9 | +#include <stdio.h> |
| 10 | +#include <stdlib.h> |
| 11 | +#include <unistd.h> |
| 12 | +#include <sys/times.h> |
| 13 | +#include "clock.h" |
| 14 | + |
| 15 | + |
| 16 | +/******************************************************* |
| 17 | + * Machine dependent functions |
| 18 | + * |
| 19 | + * Note: the constants __i386__ and __alpha |
| 20 | + * are set by GCC when it calls the C preprocessor |
| 21 | + * You can verify this for yourself using gcc -v. |
| 22 | + *******************************************************/ |
| 23 | + |
| 24 | +#if defined(__i386__) |
| 25 | +/******************************************************* |
| 26 | + * Pentium versions of start_counter() and get_counter() |
| 27 | + *******************************************************/ |
| 28 | + |
| 29 | + |
| 30 | +/* $begin x86cyclecounter */ |
| 31 | +/* Initialize the cycle counter */ |
| 32 | +static unsigned cyc_hi = 0; |
| 33 | +static unsigned cyc_lo = 0; |
| 34 | + |
| 35 | + |
| 36 | +/* Set *hi and *lo to the high and low order bits of the cycle counter. |
| 37 | + Implementation requires assembly code to use the rdtsc instruction. */ |
| 38 | +void access_counter(unsigned *hi, unsigned *lo) |
| 39 | +{ |
| 40 | + asm("rdtsc; movl %%edx,%0; movl %%eax,%1" /* Read cycle counter */ |
| 41 | + : "=r" (*hi), "=r" (*lo) /* and move results to */ |
| 42 | + : /* No input */ /* the two outputs */ |
| 43 | + : "%edx", "%eax"); |
| 44 | +} |
| 45 | + |
| 46 | +/* Record the current value of the cycle counter. */ |
| 47 | +void start_counter() |
| 48 | +{ |
| 49 | + access_counter(&cyc_hi, &cyc_lo); |
| 50 | +} |
| 51 | + |
| 52 | +/* Return the number of cycles since the last call to start_counter. */ |
| 53 | +double get_counter() |
| 54 | +{ |
| 55 | + unsigned ncyc_hi, ncyc_lo; |
| 56 | + unsigned hi, lo, borrow; |
| 57 | + double result; |
| 58 | + |
| 59 | + /* Get cycle counter */ |
| 60 | + access_counter(&ncyc_hi, &ncyc_lo); |
| 61 | + |
| 62 | + /* Do double precision subtraction */ |
| 63 | + lo = ncyc_lo - cyc_lo; |
| 64 | + borrow = lo > ncyc_lo; |
| 65 | + hi = ncyc_hi - cyc_hi - borrow; |
| 66 | + result = (double) hi * (1 << 30) * 4 + lo; |
| 67 | + if (result < 0) { |
| 68 | + fprintf(stderr, "Error: counter returns neg value: %.0f\n", result); |
| 69 | + } |
| 70 | + return result; |
| 71 | +} |
| 72 | +/* $end x86cyclecounter */ |
| 73 | + |
| 74 | +#elif defined(__alpha) |
| 75 | + |
| 76 | +/**************************************************** |
| 77 | + * Alpha versions of start_counter() and get_counter() |
| 78 | + ***************************************************/ |
| 79 | + |
| 80 | +/* Initialize the cycle counter */ |
| 81 | +static unsigned cyc_hi = 0; |
| 82 | +static unsigned cyc_lo = 0; |
| 83 | + |
| 84 | + |
| 85 | +/* Use Alpha cycle timer to compute cycles. Then use |
| 86 | + measured clock speed to compute seconds |
| 87 | +*/ |
| 88 | + |
| 89 | +/* |
| 90 | + * counterRoutine is an array of Alpha instructions to access |
| 91 | + * the Alpha's processor cycle counter. It uses the rpcc |
| 92 | + * instruction to access the counter. This 64 bit register is |
| 93 | + * divided into two parts. The lower 32 bits are the cycles |
| 94 | + * used by the current process. The upper 32 bits are wall |
| 95 | + * clock cycles. These instructions read the counter, and |
| 96 | + * convert the lower 32 bits into an unsigned int - this is the |
| 97 | + * user space counter value. |
| 98 | + * NOTE: The counter has a very limited time span. With a |
| 99 | + * 450MhZ clock the counter can time things for about 9 |
| 100 | + * seconds. */ |
| 101 | +static unsigned int counterRoutine[] = |
| 102 | +{ |
| 103 | + 0x601fc000u, |
| 104 | + 0x401f0000u, |
| 105 | + 0x6bfa8001u |
| 106 | +}; |
| 107 | + |
| 108 | +/* Cast the above instructions into a function. */ |
| 109 | +static unsigned int (*counter)(void)= (void *)counterRoutine; |
| 110 | + |
| 111 | + |
| 112 | +void start_counter() |
| 113 | +{ |
| 114 | + /* Get cycle counter */ |
| 115 | + cyc_hi = 0; |
| 116 | + cyc_lo = counter(); |
| 117 | +} |
| 118 | + |
| 119 | +double get_counter() |
| 120 | +{ |
| 121 | + unsigned ncyc_hi, ncyc_lo; |
| 122 | + unsigned hi, lo, borrow; |
| 123 | + double result; |
| 124 | + ncyc_lo = counter(); |
| 125 | + ncyc_hi = 0; |
| 126 | + lo = ncyc_lo - cyc_lo; |
| 127 | + borrow = lo > ncyc_lo; |
| 128 | + hi = ncyc_hi - cyc_hi - borrow; |
| 129 | + result = (double) hi * (1 << 30) * 4 + lo; |
| 130 | + if (result < 0) { |
| 131 | + fprintf(stderr, "Error: Cycle counter returning negative value: %.0f\n", result); |
| 132 | + } |
| 133 | + return result; |
| 134 | +} |
| 135 | + |
| 136 | +#else |
| 137 | + |
| 138 | +/**************************************************************** |
| 139 | + * All the other platforms for which we haven't implemented cycle |
| 140 | + * counter routines. Newer models of sparcs (v8plus) have cycle |
| 141 | + * counters that can be accessed from user programs, but since there |
| 142 | + * are still many sparc boxes out there that don't support this, we |
| 143 | + * haven't provided a Sparc version here. |
| 144 | + ***************************************************************/ |
| 145 | + |
| 146 | +void start_counter() |
| 147 | +{ |
| 148 | + printf("ERROR: You are trying to use a start_counter routine in clock.c\n"); |
| 149 | + printf("that has not been implemented yet on this platform.\n"); |
| 150 | + printf("Please choose another timing package in config.h.\n"); |
| 151 | + exit(1); |
| 152 | +} |
| 153 | + |
| 154 | +double get_counter() |
| 155 | +{ |
| 156 | + printf("ERROR: You are trying to use a get_counter routine in clock.c\n"); |
| 157 | + printf("that has not been implemented yet on this platform.\n"); |
| 158 | + printf("Please choose another timing package in config.h.\n"); |
| 159 | + exit(1); |
| 160 | +} |
| 161 | +#endif |
| 162 | + |
| 163 | + |
| 164 | + |
| 165 | + |
| 166 | +/******************************* |
| 167 | + * Machine-independent functions |
| 168 | + ******************************/ |
| 169 | +double ovhd() |
| 170 | +{ |
| 171 | + /* Do it twice to eliminate cache effects */ |
| 172 | + int i; |
| 173 | + double result; |
| 174 | + |
| 175 | + for (i = 0; i < 2; i++) { |
| 176 | + start_counter(); |
| 177 | + result = get_counter(); |
| 178 | + } |
| 179 | + return result; |
| 180 | +} |
| 181 | + |
| 182 | +/* $begin mhz */ |
| 183 | +/* Estimate the clock rate by measuring the cycles that elapse */ |
| 184 | +/* while sleeping for sleeptime seconds */ |
| 185 | +double mhz_full(int verbose, int sleeptime) |
| 186 | +{ |
| 187 | + double rate; |
| 188 | + |
| 189 | + start_counter(); |
| 190 | + sleep(sleeptime); |
| 191 | + rate = get_counter() / (1e6*sleeptime); |
| 192 | + if (verbose) |
| 193 | + printf("Processor clock rate ~= %.1f MHz\n", rate); |
| 194 | + return rate; |
| 195 | +} |
| 196 | +/* $end mhz */ |
| 197 | + |
| 198 | +/* Version using a default sleeptime */ |
| 199 | +double mhz(int verbose) |
| 200 | +{ |
| 201 | + return mhz_full(verbose, 2); |
| 202 | +} |
| 203 | + |
| 204 | +/** Special counters that compensate for timer interrupt overhead */ |
| 205 | + |
| 206 | +static double cyc_per_tick = 0.0; |
| 207 | + |
| 208 | +#define NEVENT 100 |
| 209 | +#define THRESHOLD 1000 |
| 210 | +#define RECORDTHRESH 3000 |
| 211 | + |
| 212 | +/* Attempt to see how much time is used by timer interrupt */ |
| 213 | +static void callibrate(int verbose) |
| 214 | +{ |
| 215 | + double oldt; |
| 216 | + struct tms t; |
| 217 | + clock_t oldc; |
| 218 | + int e = 0; |
| 219 | + |
| 220 | + times(&t); |
| 221 | + oldc = t.tms_utime; |
| 222 | + start_counter(); |
| 223 | + oldt = get_counter(); |
| 224 | + while (e <NEVENT) { |
| 225 | + double newt = get_counter(); |
| 226 | + |
| 227 | + if (newt-oldt >= THRESHOLD) { |
| 228 | + clock_t newc; |
| 229 | + times(&t); |
| 230 | + newc = t.tms_utime; |
| 231 | + if (newc > oldc) { |
| 232 | + double cpt = (newt-oldt)/(newc-oldc); |
| 233 | + if ((cyc_per_tick == 0.0 || cyc_per_tick > cpt) && cpt > RECORDTHRESH) |
| 234 | + cyc_per_tick = cpt; |
| 235 | + /* |
| 236 | + if (verbose) |
| 237 | + printf("Saw event lasting %.0f cycles and %d ticks. Ratio = %f\n", |
| 238 | + newt-oldt, (int) (newc-oldc), cpt); |
| 239 | + */ |
| 240 | + e++; |
| 241 | + oldc = newc; |
| 242 | + } |
| 243 | + oldt = newt; |
| 244 | + } |
| 245 | + } |
| 246 | + if (verbose) |
| 247 | + printf("Setting cyc_per_tick to %f\n", cyc_per_tick); |
| 248 | +} |
| 249 | + |
| 250 | +static clock_t start_tick = 0; |
| 251 | + |
| 252 | +void start_comp_counter() |
| 253 | +{ |
| 254 | + struct tms t; |
| 255 | + |
| 256 | + if (cyc_per_tick == 0.0) |
| 257 | + callibrate(0); |
| 258 | + times(&t); |
| 259 | + start_tick = t.tms_utime; |
| 260 | + start_counter(); |
| 261 | +} |
| 262 | + |
| 263 | +double get_comp_counter() |
| 264 | +{ |
| 265 | + double time = get_counter(); |
| 266 | + double ctime; |
| 267 | + struct tms t; |
| 268 | + clock_t ticks; |
| 269 | + |
| 270 | + times(&t); |
| 271 | + ticks = t.tms_utime - start_tick; |
| 272 | + ctime = time - ticks*cyc_per_tick; |
| 273 | + /* |
| 274 | + printf("Measured %.0f cycles. Ticks = %d. Corrected %.0f cycles\n", |
| 275 | + time, (int) ticks, ctime); |
| 276 | + */ |
| 277 | + return ctime; |
| 278 | +} |
| 279 | + |
0 commit comments