/*
   lperfex: A workalike of the IRIX perfex(1) command for Linux systems
   running on Intel P6 core (PPro/PII/PIII/Celeron/Xeon) processors.

   Copyright (C) 1999 Ohio Supercomputer Center.

   This code is licensed under version 2 or later of the GNU GPL; see
   /usr/src/linux/COPYING for details.

   Author: Troy Baer
           Science and Technology Support
	   Ohio Supercomputer Center
	   troy@osc.edu
	   http://www.osc.edu/~troy/

   This code relies on Erik Hendriks' perf patch and library v0.7
   for the Linux kernel.  The patch is available from:

   ftp://beowulf.gsfc.nasa.gov/pub/software/perf-0.7.tar.gz

   To compile:

   gcc -o lperfex lperfex.c -I/usr/local/include -L/usr/local/lib -lperf

   (Assuming your sysadmin has installed libperf.a and perf.h into
   /usr/local/lib and /usr/local/include, respectively.)


   Version History:
   v0.1 released 01 Oct 1999; first public release.
   v0.2 fixed %lu bug and other small bugs

   To do:
   * Support of more of the IRIX perfex command line options,
     especially counter multiplexing and counting inside multithreaded
     code (assuming these are even possible).
   * More and better report statistics (memory->L2 bandwidth is
     currently computed in a *very* stupid way)
*/

#include <stdio.h>
#include <stdlib.h>
#include <sys/resource.h>
#include <sys/time.h>
#include <sys/utsname.h>
#include <unistd.h>
#include <time.h>
#include <math.h>
#include <signal.h>
#include <perf.h>

/* 
   P6 core countable events:

   0: Memory references
   1: L1 data cache lines loaded
   2: L1 data cache lines loaded and modified
   3: L1 data cache lines flushed
   4: Weighed number of cycles spent waiting while a L1 data cache miss
      is resolved
   5: Instruction fetches
   6: L1 instruction cache misses
   7: ITLB misses
   8: Cycles spent waiting for instruction fetches and ITLB misses
   9: Cycles spent waiting on the instruction decoder
  10: L2 cache instruction fetches
  11: L2 cache data loads
  12: L2 cache data stores
  13: L2 cache lines loaded
  14: L2 cache lines flushed
  15: L2 cache lines loaded and modified
  16: L2 cache lines modified and flushed
  17: L2 cache requests
  18: L2 cache address strobes
  19: Cycles spent waiting on the L2 data bus
  20: Cycles spent waiting on data transfer from L2 cache to processor
  21: Cycles spent while DRDY is asserted
  22: Cycles spent while LOCK is asserted
  23: Bus requests outstanding
  24: Burst read transactions
  25: Read-for-ownership transactions
  26: Write-back transactions
  27: Instruction fetch transactions
  28: Invalidate transactions
  29: Partial-write transactions
  30: Partial transactions
  31: I/O transactions
  32: Deferred transactions
  33: Burst transactions
  34: Total number of transactions
  35: Memory transactions
  36: Bus clock cycles spent while the processor is receiving data
  37: Bus clock cycles spent while the processor is driving the BNR pin
  38: Bus clock cycles spent while the processor is driving the HIT pin
  39: Bus clock cycles spent while the processor is driving the HITM pin
  40: Cycles spent while the bus is snoop-stalled
  41: Floating point operations retired (counter 0 only)
  42: Floating point operations executed (counter 0 only)
  43: Floating point exceptions handled by microcode (counter 1 only)
  44: Multiply operations (counter 1 only)
  45: Divide operations (counter 1 only)
  46: Cycles spent doing division (counter 0 only)
  47: Store buffer blocks
  48: Store buffer drain cycles
  49: Misaligned memory references
  50: Instructions retired
  51: uOps retired
  52: Instructions decoded
  53: Hardware interrupts received
  54: Cycles spent while interrupts are disabled
  55: Cycles spent while interrupts and disabled and pending
  56: Branch instructions retired
  57: Mispredicted branches retired
  58: Taken branches retired
  59: Taken mispredicted branches retired
  60: Branch instructions decoded
  61: Branches which miss the BTB
  62: Bogus branches
  63: BACLEAR assertions
  64: Cycles spent during resource related stalls
  65: Cycles spent during partial stalls
  66: Segment register loads
  67: Cycles during which the processor is not halted
*/
int event[68] = {
  PERF_DATA_MEM_REFS,  
  PERF_DCU_LINES_IN,
  PERF_DCU_M_LINES_IN,
  PERF_DCU_M_LINES_OUT,
  PERF_DCU_MISS_STANDING,
  PERF_IFU_IFETCH,
  PERF_IFU_IFETCH_MISS,
  PERF_ITLB_MISS,
  PERF_IFU_MEM_STALL,
  PERF_ILD_STALL,
  PERF_L2_IFETCH,
  PERF_L2_LD,
  PERF_L2_ST,
  PERF_L2_LINES_IN,
  PERF_L2_LINES_OUT,
  PERF_L2_LINES_INM,
  PERF_L2_LINES_OUTM,
  PERF_L2_RQSTS,
  PERF_L2_ADS,
  PERF_L2_DBUS_BUSY,
  PERF_L2_DBUS_BUSY_RD,
  PERF_BUS_DRDY_CLOCKS,
  PERF_BUS_LOCK_CLOCKS,
  PERF_BUS_REQ_OUTSTANDING,
  PERF_BUS_TRAN_BRD,
  PERF_BUS_TRAN_RFO,
  PERF_BUS_TRANS_WB,
  PERF_BUS_TRAN_IFETCH,
  PERF_BUS_TRAN_INVAL,
  PERF_BUS_TRAN_PWR,
  PERF_BUS_TRAN_P,
  PERF_BUS_TRANS_IO,
  PERF_BUS_TRAN_DEF,
  PERF_BUS_TRAN_BURST,
  PERF_BUS_TRAN_ANY,
  PERF_BUS_TRAN_MEM,
  PERF_BUS_DATA_RCV,
  PERF_BUS_BNR_DRV,
  PERF_BUS_HIT_DRV,
  PERF_BUS_HITM_DRV,
  PERF_BUS_SNOOP_STALL,
  PERF_FLOPS,
  PERF_FP_COMP_OPS_EXE,
  PERF_FP_ASSIST,
  PERF_MUL,
  PERF_DIV,
  PERF_CYCLES_DIV_BUSY,
  PERF_LD_BLOCK,
  PERF_SB_DRAINS,
  PERF_MISALIGN_MEM_REF,
  PERF_INST_RETIRED,
  PERF_UOPS_RETIRED,
  PERF_INST_DECODER,
  PERF_HW_INT_RX,
  PERF_CYCLES_INST_MASKED,
  PERF_CYCLES_INT_PENDING_AND_MASKED,
  PERF_BR_INST_RETIRED,
  PERF_BR_MISS_PRED_RETIRED,
  PERF_BR_TAKEN_RETIRED,
  PERF_BR_MISS_PRED_TAKEN_RET,
  PERF_BR_INST_DECODED,
  PERF_BR_BTB_MISSES,
  PERF_BR_BOGUS,
  PERF_BACLEARS,
  PERF_RESOURCE_STALLS,
  PERF_PARTIAL_RAT_STALLS,
  PERF_SEGMENT_REG_LOADS,
  PERF_CPU_CLK_UNHALTED
};
char label[68][81] = {
  "Memory references",
  "L1 data cache lines loaded",
  "L1 data cache lines loaded and modified",
  "L1 data cache lines flushed",
  "Weighed number of cycles spent waiting while a L1 data cache miss is resolved",
  "Instruction fetches",
  "L1 instruction cache misses",
  "ITLB misses",
  "Cycles spent waiting for instruction fetches and ITLB misses",
  "Cycles spent waiting on the instruction decoder",
  "L2 cache instruction fetches",
  "L2 cache data loads",
  "L2 cache data stores",
  "L2 cache lines loaded",
  "L2 cache lines flushed",
  "L2 cache lines loaded and modified",
  "L2 cache lines modified and flushed",
  "L2 cache requests",
  "L2 cache address strobes",
  "Cycles spent waiting on the L2 data bus",
  "Cycles spent waiting on data transfer from L2 cache to processor",
  "Cycles spent while DRDY is asserted",
  "Cycles spent while LOCK is asserted",
  "Bus requests outstanding",
  "Burst read transactions",
  "Read-for-ownership transactions",
  "Write-back transactions",
  "Instruction fetch transactions",
  "Invalidate transactions",
  "Partial-write transactions",
  "Partial transactions",
  "I/O transactions",
  "Deferred transactions",
  "Burst transactions",
  "Total number of transactions",
  "Memory transactions",
  "Bus clock cycles spent while the processor is receiving data",
  "Bus clock cycles spent while the processor is driving the BNR pin",
  "Bus clock cycles spent while the processor is driving the HIT pin",
  "Bus clock cycles spent while the processor is driving the HITM pin",
  "Cycles spent while the bus is snoop-stalled",
  "Floating point operations retired (counter 0 only)",
  "Floating point operations executed (counter 0 only)",
  "Floating point exceptions handled by microcode (counter 1 only)",
  "Multiply operations (counter 1 only)",
  "Divide operations (counter 1 only)",
  "Cycles spent doing division (counter 0 only)",
  "Store buffer blocks",
  "Store buffer drain cycles",
  "Misaligned memory references",
  "Instructions retired",
  "uOps retired",
  "Instructions decoded",
  "Hardware interrupts received",
  "Cycles spent while interrupts are disabled",
  "Cycles spent while interrupts and disabled and pending",
  "Branch instructions retired",
  "Mispredicted branches retired",
  "Taken branches retired",
  "Taken mispredicted branches retired",
  "Branch instructions decoded",
  "Branches which miss the BTB",
  "Bogus branches",
  "BACLEAR assertions",
  "Cycles spent during resource related stalls",
  "Cycles spent during partial stalls",
  "Segment register loads",
  "Cycles during which the processor is not halted"
};

#define TIMEVAL_TO_FLOAT(x) ((float)((x).tv_sec)+0.000001*(float)((x).tv_usec))

int main(int argc, char *argv[], char *environment[])
{
  int event0=-1, event1=-1, mplex=0, mkrpt=0, i, status, rval;
  FILE *extfile=stdout;
  float telapsed;
  pid_t child;
  struct rusage ru;
  unsigned long long counter[PERF_COUNTERS];
  extern char *optarg;
  extern int optind;
  char hostname[80];
  
  /* Parse command line options */
  gethostname(hostname,80);
  extfile=stdout;
  while ((i=getopt(argc,argv,"ae:kmpo:sxy"))!=-1)
    {
      switch(i)
	{
	case 'a':
	  mplex=1;
	  
	  break;
	case 'e':
	  mplex=0;
	  if (event0==-1)
	    {
	      sscanf(optarg,"%d",&event0);
	    }
	  else
	    {
	      sscanf(optarg,"%d",&event1);
	    }
	  break;
	case 'k':
	  fprintf(stderr,"lperfex:  Kernel counting not supported\n");
	  break;
	case 'm':
	  fprintf(stderr,"lperfex:  Multithreaded counting not supported\n");
	  break;
	case 'o':
	  extfile=fopen(optarg,"w");
	  break;
	case 's':
	  fprintf(stderr,"lperfex:  Signalled counting no supported\n");
	  break;
	case 'x':
	  fprintf(stderr,"lperfex:  Exception counting no supported\n");
	  break;
	case 'y':
	  mkrpt=1;
	  break;
	case '?':
	case 'h':
	  printf("Usage:  lperfex [-a | -e event0 [-e event1]] [-mp | -s] [-x] [-k] [-y] [-o file] [-c file] command [args]\n");
	  printf("\nEvent numbers: \n");
	  for (i=0;i<68;i++)
	    printf("\t%2d:  %s\n",i,label[i]);
	  exit(-1);
	}
    }

  if (event0==-1)
    {
      event0=42; /* Default to counting MFLOPS */
    }
  if (event1==-1)
    {
      event1=13; /* Default to count L2 cache line loads */
    }

  /* Set the counters up and run the child program */
  perf_reset();
  if ((child=fork())==0)
    {
      /* Child */
      perf_reset();
      if (mplex==0)
	{
	  perf_set_config(0,event[event0]);
	  perf_set_config(1,event[event1]);
	}
      else
	{
	  fprintf(stderr,"lperfex:  Multiplexing of counters currently not supported\n");
	  exit(-2);
	}
      perf_start();
      execvp(argv[optind],argv+optind);
    }
  else
    {
      /* Parent */
      rval=perf_wait(child,&status,(int)(0),&ru,counter);
      if ( ( rval != child ) || ( status != 0 ) )
	 {
	    fprintf(stderr,"Error waiting for process to terminate; perf_wait returned %d, child=%d, status=%d\n",rval,child,status);
	    kill(child,SIGTERM);
	    exit(-3);
	 }
      perf_stop();
      
      telapsed=TIMEVAL_TO_FLOAT(ru.ru_utime)+TIMEVAL_TO_FLOAT(ru.ru_stime);
      fprintf(extfile,"%f seconds of CPU time elapsed and %f MB of memory on %s\n\n",telapsed,((float)ru.ru_maxrss)/1024.,hostname);
      fprintf(extfile,"Event #\t\t\tEvent\t\t\t\t\t\t\t\t\tEvents Counted\n");
      fprintf(extfile,"-------\t\t\t-----\t\t\t\t\t\t\t\t\t--------------\n");
      if (event0!=-1)
	fprintf(extfile,"   %2d  \t%-80s\t%14llu\n",event0,label[event0],counter[0]);
      if (event1!=-1)
	fprintf(extfile,"   %2d  \t%-80s\t%14llu\n",event1,label[event1],counter[1]);
      if (mkrpt)
	{
	  fprintf(extfile,"\nStatistics:\n");
	  fprintf(extfile,"-----------\n");
	  if (event[event0]==PERF_INST_DECODER || event[event0]==PERF_INST_RETIRED)
	    fprintf(extfile,"MIPS\t\t\t%14lf\n",1.0e-6*((double)counter[0])/(double)telapsed);
	  else if (event[event1]==PERF_INST_DECODER || event[event1]==PERF_INST_RETIRED)
	    fprintf(extfile,"MIPS\t\t\t%14lf\n",1.0e-6*((double)counter[1])/(double)telapsed);
	  if (event[event0]==PERF_FLOPS || event[event0]==PERF_FP_COMP_OPS_EXE)
	    {
	      fprintf(extfile,"MFLOPS\t\t\t%14lf\n",1.0e-6*((double)counter[0])/(double)telapsed);
	      if (event[event1]==PERF_MUL)
		fprintf(extfile,"Multiplications/FP op\t%14lf\n",((double)counter[1])/((double)counter[0]));
	      else if (event[event1]==PERF_DIV)
		fprintf(extfile,"Divisions/FP op\t\t%14lf\n",((double)counter[1])/((double)counter[0]));
	      else if (event[event1]==PERF_INST_DECODER || event[event1]==PERF_INST_RETIRED)
		fprintf(extfile,"Instructions/FP op\t%14lf\n",((double)counter[1])/((double)counter[0]));
	      else if (event[event1]==PERF_CPU_CLK_UNHALTED)
		fprintf(extfile,"Cycles/FP op\t\t%14lf\n",((double)counter[1])/((double)counter[0]));
	    }
	  if (event[event0]==PERF_CYCLES_DIV_BUSY && event[event1]==PERF_DIV)
	    fprintf(extfile,"Cycles/divide op\t\t%14lf\n",((double)counter[0])/((double)counter[1]));
	  if (event[event0]==PERF_DCU_LINES_IN)
	    fprintf(extfile,"L2->L1 memory bandwidth\t%14lf MB/s\n",1.6E-5*((double)counter[0])/telapsed);
	  else if (event[event1]==PERF_DCU_LINES_IN)
	    fprintf(extfile,"L2->L1 memory bandwidth\t%14lf MB/s\n",1.6E-5*((double)counter[1])/telapsed);
	  if (event[event0]==PERF_L2_LINES_IN)
	    fprintf(extfile,"main memory->L2 bandwidth\t%14lf MB/s\n",3.2E-5*((double)counter[0])/telapsed);
	  else if (event[event1]==PERF_L2_LINES_IN)
	    fprintf(extfile,"main memory->L2 bandwidth\t%14lf MB/s\n",3.2E-5*((double)counter[1])/telapsed);
	  if (event[event0]==PERF_DATA_MEM_REFS && event[event1]==PERF_DCU_LINES_IN)
	    fprintf(extfile,"L1 data cache hit rate\t%14lf\n",1.-((double)counter[1])/((double)counter[0]));
	  else if (event[event1]==PERF_DATA_MEM_REFS && event[event0]==PERF_DCU_LINES_IN)
	    fprintf(extfile,"L1 data cache hit rate\t%14lf\n",1.-((double)counter[0])/((double)counter[1]));
	  if (event[event0]==PERF_DATA_MEM_REFS && event[event1]==PERF_L2_LINES_IN)
	    fprintf(extfile,"L2 cache hit rate\t\t%14lf\n",1.-((double)counter[1])/((double)counter[0]));
	  else if (event[event1]==PERF_DATA_MEM_REFS && event[event0]==PERF_L2_LINES_IN)
	    fprintf(extfile,"L2 cache hit rate\t\t%14lf\n",1.-((double)counter[0])/((double)counter[1]));
	  if (event[event0]==PERF_DCU_LINES_IN && event[event1]==PERF_L2_LINES_IN)
	    fprintf(extfile,"L1 data cache miss/L2 cache miss\t%14lf\n",((double)counter[0])/((double)counter[1]));
	  else if (event[event1]==PERF_DCU_LINES_IN && event[event0]==PERF_L2_LINES_IN)
	    fprintf(extfile,"L1 data cache miss/L2 cache miss\t%14lf\n",((double)counter[1])/((double)counter[0]));
	}
    }
  exit(0); 
}