/*----------------------------------------------------------------------------*/ /* * Performance-Monitoring Counters Library, for Intel/AMD Processors and Linux * Author: Don Heller, dheller@scl.ameslab.gov * Last revised: 19 January 2001 */ /*----------------------------------------------------------------------------*/ /* test memory access, print "interesting" events * compile: gcc -DTEST=1 -O `pmc_options` -o mem_events mem_events.c -lpmc -lm * usage: mem_events [min_length] [max_length] [alignment] [trials] * min_length and max_length must be multiples of 8 * min_length must be at least 16 * examples, Pentium Pro/II/III: * mem_events --e 0x62 --bus_agent 0,1 1m * mem_events 1m * examples, Athlon: * mem_events -g 0 1m */ /*----------------------------------------------------------------------------*/ /* default values */ #ifndef TEST #define TEST 2 #endif #define MIN_LENGTH 128 #define MAX_LENGTH 128 /* min and max number of words to copy */ #define ALIGN PMC_CACHE_LINE #define TRIALS 100 #define CUTOFF 0.0001 /* events per word copied, ignore anything less; 1/32 = 0.03125 */ // #define USE_CORRECTION void access(int *src, int *dst, int n); /* n must be a multiple of 8, at least 8 */ /*----------------------------------------------------------------------------*/ #include #include #include /*----------------------------------------------------------------------------*/ int Atoi(char *in) { int n = 0; while (isdigit(*in)) { n = 10*n + (*in - '0'); in++; } if (*in == 'k' || *in == 'K') { n *= 1024; } if (*in == 'm' || *in == 'M') { n *= 1024*1024; } return n; } /*----------------------------------------------------------------------------*/ int main(int argc, char * argv[]) { pmc_control_t Ctl = pmc_control_null; int *src, *dst, incr = 8; int min_length = MIN_LENGTH, max_length = MAX_LENGTH; int align = ALIGN; int i,j,k, m = 0, n, out, trials = TRIALS; pmc_data_t t0, t1; pmc_counter_t c; /* initialize internal data structures, read command-line arguments */ Ctl.clean = 10; Ctl.stats = 1; for (i = 0; i < pmc_event_counters; i++) { Ctl.os[i] = 0; } if (pmc_getargs(stderr, argv[0], &argc, &argv, &Ctl) == FALSE) { exit(RABBIT_FAILURE); } if (argc > 0) { min_length = Atoi(argv[0]); max_length = min_length; } if (argc > 1) { max_length = Atoi(argv[1]); } if (argc > 2) { align = Atoi(argv[2]); } if (argc > 3) { trials = Atoi(argv[3]); } printf("test memory access, #%d, length = %d : %d, alignment = %d, trials = %d\n", TEST, min_length, max_length, align, trials); #ifdef USE_CORRECTION printf(" correction from %d words\n", (m = 16)); #endif if (min_length <= 16 || max_length <= 16 || max_length < min_length || align < 0 || trials <= 0) { printf(" invalid request\n"); exit(RABBIT_FAILURE); } src = (int *) pmc_calloc(max_length, 4, align); dst = (int *) pmc_calloc(max_length, 4, align); i = (int) src; j = (int) dst; printf(" src = 0x%08lx, dst = 0x%08lx, dst-src = 0x%08lx\n", i, j, j-i); if (src == NULL || dst == NULL) { printf(" memory allocation failed\n"); exit(RABBIT_FAILURE); } for (i = 0; i < max_length; i++) { dst[i] = src[i] = 0xbabef00d; } if (pmc_open(0) == FALSE) /* open /dev/pmc */ { exit(RABBIT_FAILURE); } for (n = min_length; n <= max_length; n += incr) { if (n > 1*1024) { incr = 32; } if (n > 4*1024) { incr = 128; } if (n > 16*1024) { incr = 512; } pmc_counter_reset(&c); /* for the number of cycles */ for (out = 0; out < Ctl.num_counters; out++) { pmc_counter_reset(&Ctl.counters[out]); pmc_select(&Ctl.counters[out]); access(src, dst, n); /* first time doesn't count */ for (k = 0; k < trials; k++) { pmc_read(&t0); access(src, dst, n); pmc_read(&t1); pmc_accumulate(&Ctl.counters[out], &t1, &t0); pmc_accumulate(&c, &t1, NULL); } } { int e; double Min, Mean, Max, StdDev, len; pmc_counter_t correction; printf("Copying %d words\n", n); printf("Events per word, cutoff %5.4f Min Mean Max Std.Dev.\n", CUTOFF); #ifdef USE_CORRECTION len = (double) (n-m); for (i = 0, out = 0; i < Ctl.event_pairs; i++) { for (j = 0; j < Ctl.replication; j++, out++) { /* estimate function call overhead */ correction.selector = Ctl.counters[out].selector; correction.clean = 10; correction.stats = 1; pmc_counter_reset(&correction); pmc_select(&correction); access(src, dst, m); /* first time doesn't count */ for (k = 0; k < trials; k++) { pmc_read(&t0); access(src, dst, m); pmc_read(&t1); pmc_accumulate(&correction, &t1, &t0); } for (k = 0; k < pmc_event_counters; k++) { Min = ((double) pmc_min_events(&Ctl.counters[out],k) - (double) pmc_min_events(&correction,k)) / len; Mean = (pmc_mean_events(&Ctl.counters[out],k) - pmc_mean_events(&correction,k)) / len; Max = ((double) pmc_max_events(&Ctl.counters[out],k) - (double) pmc_max_events(&correction,k)) / len; StdDev = (sqrt(pmc_variance_events(&Ctl.counters[out],k)) - sqrt(pmc_variance_events(&correction,k))) / len; if (Mean > CUTOFF) { e = Ctl.events[i][k]; printf("0x%02x %3d %-30s %8.4f %8.4f %8.4f %10.6f\n", e, e, pmc_event_name(e,k), Min, Mean, Max, StdDev); } } } } #else len = (double) n; for (i = 0, out = 0; i < Ctl.event_pairs; i++) { for (j = 0; j < Ctl.replication; j++, out++) { for (k = 0; k < pmc_event_counters; k++) { Min = (double) pmc_min_events(&Ctl.counters[out],k) / len; Mean = pmc_mean_events(&Ctl.counters[out],k) / len; Max = (double) pmc_max_events(&Ctl.counters[out],k) / len; StdDev = sqrt(pmc_variance_events(&Ctl.counters[out],k)) / len; if (Mean > CUTOFF) { e = Ctl.events[i][k]; printf("0x%02x %3d %-30s %8.4f %8.4f %8.4f %10.6f\n", e, e, pmc_event_name(e,k), Min, Mean, Max, StdDev); } } } } #endif printf("Cycles Min Mean Max Std.Dev.\n"); Min = (double) pmc_min_cycles(&c) / len; Mean = pmc_mean_cycles(&c) / len; Max = (double) pmc_max_cycles(&c) / len; StdDev = sqrt(pmc_variance_cycles(&c)) / len; printf("cycles %8.4f %8.4f %8.4f %10.6f\n", Min, Mean, Max, StdDev); fflush(stdout); } } pmc_close(); /* close /dev/pmc */ exit(RABBIT_SUCCESS); } /*----------------------------------------------------------------------------*/ void access(int *src, int *dst, int n) { #if TEST == 0 /* nothing at all */ #endif #if TEST == 1 __asm__ __volatile__ ( "movl %0,%%esi\n\t" /* src */ "movl %1,%%edi\n\t" /* dst */ "movl %2,%%ecx\n\t" /* n */ "cld\n\t" /* rep/movsl will increment esi, edi by 4 */ "rep\n\t" "movsl" : /* output */ : "g" (src), "g" (dst), "g" (n) /* input */ : "esi", "edi", "ecx" /* clobbered */ ); #endif #if TEST == 2 __asm__ __volatile__ ( "movl %0,%%esi\n\t" /* src */ "movl %1,%%edi\n\t" /* dst */ "movl %2,%%ecx\n\t" /* n */ "xorl %%edx,%%edx\n\t" /* int j = 0 */ ".align 4\n" "1:\n\t" "movl (%%esi,%%edx,4),%%eax\n\t" /* eax = src[j] */ "movl %%eax,(%%edi,%%edx,4)\n\t" /* dst[j] = eax */ "incl %%edx\n\t" /* j++ */ "cmpl %%ecx,%%edx\n\t" /* j cmp n */ "jl 1b" : : "g" (src), "g" (dst), "g" (n) : "esi", "edi", "ecx", "eax", "edx" ); #endif #if TEST == 3 __asm__ __volatile__ ( "movl %0,%%esi\n\t" /* src */ "movl %1,%%edi\n\t" /* dst */ "movl %2,%%ecx\n\t" /* n */ "xorl %%edx,%%edx\n\t" /* int j = 0 */ ".align 4\n" "1:\n\t" "movl (%%esi,%%edx,4),%%eax\n\t" /* eax = src[j] */ "movl %%eax,(%%edi,%%edx,4)\n\t" /* dst[j] = eax */ "incl %%edx\n\t" /* j++ */ "movl (%%esi,%%edx,4),%%eax\n\t" /* eax = src[j] */ "movl %%eax,(%%edi,%%edx,4)\n\t" /* dst[j] = eax */ "incl %%edx\n\t" /* j++ */ "cmpl %%ecx,%%edx\n\t" /* j cmp n */ "jl 1b" : : "g" (src), "g" (dst), "g" (n) : "esi", "edi", "ecx", "eax", "edx" ); #endif #if TEST == 4 __asm__ __volatile__ ( "movl %0,%%esi\n\t" /* src */ "movl %1,%%edi\n\t" /* dst */ "movl %2,%%ecx\n\t" /* n */ "xorl %%edx,%%edx\n\t" /* int j = 0 */ ".align 4\n" "1:\n\t" "movl (%%esi,%%edx,4),%%eax\n\t" /* eax = src[j] */ "movl %%eax,(%%edi,%%edx,4)\n\t" /* dst[j] = eax */ "incl %%edx\n\t" /* j++ */ "movl (%%esi,%%edx,4),%%eax\n\t" /* eax = src[j] */ "movl %%eax,(%%edi,%%edx,4)\n\t" /* dst[j] = eax */ "incl %%edx\n\t" /* j++ */ "movl (%%esi,%%edx,4),%%eax\n\t" /* eax = src[j] */ "movl %%eax,(%%edi,%%edx,4)\n\t" /* dst[j] = eax */ "incl %%edx\n\t" /* j++ */ "movl (%%esi,%%edx,4),%%eax\n\t" /* eax = src[j] */ "movl %%eax,(%%edi,%%edx,4)\n\t" /* dst[j] = eax */ "incl %%edx\n\t" /* j++ */ "cmpl %%ecx,%%edx\n\t" /* j cmp n */ "jl 1b" : : "g" (src), "g" (dst), "g" (n) : "esi", "edi", "ecx", "eax", "edx" ); #endif #if TEST == 5 __asm__ __volatile__ ( "movl %0,%%esi\n\t" /* src */ "movl %1,%%edi\n\t" /* dst */ "movl %2,%%ecx\n\t" /* n */ "xorl %%edx,%%edx\n\t" /* int j = 0 */ ".align 4\n" "1:\n\t" "movl (%%esi,%%edx,4),%%eax\n\t" /* eax = src[j] */ "movl %%eax,(%%edi,%%edx,4)\n\t" /* dst[j] = eax */ "movl 4(%%esi,%%edx,4),%%eax\n\t" /* eax = src[j+1] */ "movl %%eax,4(%%edi,%%edx,4)\n\t" /* dst[j+1] = eax */ "movl 8(%%esi,%%edx,4),%%eax\n\t" /* eax = src[j+2] */ "movl %%eax,8(%%edi,%%edx,4)\n\t" /* dst[j+2] = eax */ "movl 12(%%esi,%%edx,4),%%eax\n\t" /* eax = src[j+3] */ "movl %%eax,12(%%edi,%%edx,4)\n\t" /* dst[j+3] = eax */ "addl $4,%%edx\n\t" /* j += 4 */ "cmpl %%ecx,%%edx\n\t" /* j cmp n */ "jl 1b" : : "g" (src), "g" (dst), "g" (n) : "esi", "edi", "ecx", "eax", "edx" ); #endif #if TEST == 6 __asm__ __volatile__ ( "movl %0,%%esi\n\t" /* src */ "movl %1,%%edi\n\t" /* dst */ "movl %2,%%ecx\n\t" /* n */ "xorl %%edx,%%edx\n\t" /* int j = 0 */ ".align 4\n" "1:\n\t" "movl (%%esi,%%edx,4),%%eax\n\t" /* eax = src[j] */ "movl 4(%%esi,%%edx,4),%%ebx\n\t" /* ebx = src[j+1] */ "movl %%eax,(%%edi,%%edx,4)\n\t" /* dst[j] = eax */ "movl %%ebx,4(%%edi,%%edx,4)\n\t" /* dst[j+1] = ebx */ "movl 8(%%esi,%%edx,4),%%eax\n\t" /* eax = src[j+2] */ "movl 12(%%esi,%%edx,4),%%ebx\n\t" /* ebx = src[j+3] */ "movl %%eax,8(%%edi,%%edx,4)\n\t" /* dst[j+2] = eax */ "movl %%ebx,12(%%edi,%%edx,4)\n\t" /* dst[j+3] = ebx */ "addl $4,%%edx\n\t" /* j += 4 */ "cmpl %%ecx,%%edx\n\t" /* j cmp n */ "jl 1b" : : "g" (src), "g" (dst), "g" (n) : "esi", "edi", "ecx", "eax", "ebx", "edx" ); #endif #if TEST == 7 __asm__ __volatile__ ( "movl %0,%%esi\n\t" /* src */ "movl %1,%%edi\n\t" /* dst */ "movl %2,%%ecx\n\t" /* n */ "xorl %%edx,%%edx\n\t" /* int j = 0 */ ".align 4\n" "1:\n\t" "movl (%%esi,%%edx,4),%%eax\n\t" /* eax = src[j] */ "movl 4(%%esi,%%edx,4),%%ebx\n\t" /* ebx = src[j+1] */ "movl %%eax,(%%edi,%%edx,4)\n\t" /* dst[j] = eax */ "movl %%ebx,4(%%edi,%%edx,4)\n\t" /* dst[j+1] = ebx */ "movl 8(%%esi,%%edx,4),%%eax\n\t" /* eax = src[j+2] */ "movl 12(%%esi,%%edx,4),%%ebx\n\t" /* ebx = src[j+3] */ "movl %%eax,8(%%edi,%%edx,4)\n\t" /* dst[j+2] = eax */ "movl %%ebx,12(%%edi,%%edx,4)\n\t" /* dst[j+3] = ebx */ "movl 16(%%esi,%%edx,4),%%eax\n\t" /* eax = src[j+4] */ "movl 20(%%esi,%%edx,4),%%ebx\n\t" /* ebx = src[j+5] */ "movl %%eax,16(%%edi,%%edx,4)\n\t" /* dst[j+4] = eax */ "movl %%ebx,20(%%edi,%%edx,4)\n\t" /* dst[j+5] = ebx */ "movl 24(%%esi,%%edx,4),%%eax\n\t" /* eax = src[j+6] */ "movl 28(%%esi,%%edx,4),%%ebx\n\t" /* ebx = src[j+7] */ "movl %%eax,24(%%edi,%%edx,4)\n\t" /* dst[j+6] = eax */ "movl %%ebx,28(%%edi,%%edx,4)\n\t" /* dst[j+7] = ebx */ "addl $8,%%edx\n\t" /* j += 8 */ "cmpl %%ecx,%%edx\n\t" /* j cmp n */ "jl 1b" : : "g" (src), "g" (dst), "g" (n) : "esi", "edi", "ecx", "eax", "ebx", "edx" ); #endif #if TEST == 8 register int i; for (i = 0; i < n; i++) { dst[i] = src[i]; } #endif #if TEST == 9 register int i; for (i = 0; i < n; i += 2) { dst[i] = src[i]; dst[i+1] = src[i+1]; } #endif #if TEST == 10 register int i; for (i = 0; i < n; i += 4) { dst[i] = src[i]; dst[i+1] = src[i+1]; dst[i+2] = src[i+2]; dst[i+3] = src[i+3]; } #endif #if TEST == 11 register int i; for (i = 0; i < n; i += 8) { dst[i] = src[i]; dst[i+1] = src[i+1]; dst[i+2] = src[i+2]; dst[i+3] = src[i+3]; dst[i+4] = src[i+4]; dst[i+5] = src[i+5]; dst[i+6] = src[i+6]; dst[i+7] = src[i+7]; } #endif return; } /*----------------------------------------------------------------------------*/