跳转至

Visual Performance Analyzer

Visual Performance Analyzer(VPA)

Download: https://w3-connections.ibm.com/wikis/home?lang=en-us#!/wiki/W5be7727ab90c_4453_b240_01bb7c1df49d/page/Performance%20Tools%20Release

https://ctweb.torolab.ibm.com/wiki/index.php/Prepare_testcase_for_P9_M1_perf_test

https://w3-connections.ibm.com/wikis/home?lang=en-us#!/wiki/Wf6305d6e6a03_468d_9aea_5b6b54446a50/page/Run%20Performance%20Test%20on%20M1

Jviewer

Run Performance Test on M1

The Mambo (Functional simulator, qtrace generator) and M1 (Performance simulator) simulation env have been set up on recycler.

Mambo :

Bash
 P9: /opt/systemsim-p9-release
P10: /opt/systemsim-p10-release

M1:

Bash
 P9: /opt/m1-p9/
P10: /opt/m1-p10/

You can use these simulators directly if you want. But please don't try to run long trace (eg: SPEC!) on recycler!

If you would like to run some simple test, you can use Jinsong's script on recycler.

  1. refer to how to prepare the testcase and how to link it in https://ctweb.torolab.ibm.com/wiki/index.php/Prepare_testcase_for_P9_M1_perf_test first.

  2. source the setup here: source /opt/sim_env.cfg

  3. run run_bare_metal_sim.pl

run_bare_metal_sim.pl -10 -f ./a.out

If you want run it on the real machine, you should not using -8 or -10.

Bash
$ run_bare_metal_sim.pl 
Usage: /home/jji/Mambo_M1/Mambo_M1_barmetal_sim_scripts/run_bare_metal_sim.pl
​        -d <string>, optional, directory to run
​        -f <string>, optional, binary to run
​        -o <string>, optional, output filename, default is timestamp.csv
​        -q , optional, generate qtrace only
​        -m , optional, use the qtrace directly (run M1 only)
​        -s , optional, run scrollpv to ASCII
​        -n , optional, don't generate pipe file
​        -8 , optional, run P8 mambo/M1 instead
​        -10 , optional, run P10 mambo/M1 instead
​        -p9mambo , optional, run P9 mambo for P10, when -10 is used instead
​        -p , optional, PATH to Mambo installation, default is using ENV setting
​        -i , optional, show instruction when simulating with Mambo
​        -r , optional, show register when simulating with Mambo
​        -x , optional, show memory reference when simulating with Mambo
​        -a , optional, analyze the report only
​        -v , optional, verbose print commands

Results

  • You can find the CPI/INST/TIME in output csv file. eg:
Text Only
Directory, Filename, CPI, INST, TIME
lhs/test, lhs_test1a.qt.bz2, 0.47159, 11561, 5452
  • You can get the detail M1 report and pipe file in *.qt.bz2.results in your binary dir.

Prepare testcase for M1 perf test

(Redirected from Prepare testcase for P9 M1 perf test)

Guidelines

  • No system call
  • No getenv/setenv etc
  • No cout/printf etc
  • No malloc/free, unless you write your own library that don't use any system call.
    • Use static array if you can.
  • No runtime dependency
  • No too many iterations.

Simple steps

  • Define special nops for mambo
Python
#define mambo_warm_trace() asm volatile("or 13,13,13")
#define mambo_start_trace() asm volatile("or 14,14,14")
#define mambo_stop_trace() asm volatile("or 15,15,15")
  • Point out where you want to collect the trace with special nops
  • call mambo_warm_trace(); at the beginning of your test. eg: in main .
  • call mambo_start_trace; where the interested part is. eg: at the beginning of Run().
  • call mambo_stop_trace; where the interested part ends. eg: at the end of Run().
  • Static link with nocrt.
  • Add "-e main -qnocrt -static -L/opt/at11.0/lib64/ " to your normal xl link options
  • Add "-Xlinker -emain -nostartfiles -static " to your normal clang link options

cat main.c

C++
#define mambo_warm_trace() asm volatile("or 13,13,13")
#define mambo_start_trace() asm volatile("or 14,14,14")
#define mambo_stop_trace() asm volatile("or 15,15,15")

struct parm {
  int *arr;
  int m;
  int n;
};
void foo(struct parm*);
int main() {
  int a[5000];
  struct parm arg = {a, 1000, 5};pkl     nlkm  mambo_warm_trace();
  mambo_start_trace();
  foo(&arg);
  mambo_stop_trace();
  return 0;
}

cat foo.c

C
struct parm {
  int *arr;
  int m;
  int n;
};
void foo(struct parm *arg) {
//  m--;
//  n--;
  struct parm localArg = *arg;
  int m = localArg.m;
  int *s = localArg.arr;
  int n = localArg.n;
  do{
    int k = n;
//    n--;
//    k--;
    do{
      s[++k] = k++;
      s[k++] = k;
      s[k++] = k;
      s[k] = k;
      s[--k] = k--;
      s[k--] = k;
      s[--k] = k;
//      s[k++] = k--;
//      s[k++] = k--;
//      s[k-1]=k;
    }while(k--);
  } while(m--);

  s[n]=0;
}

cat t.ksh

Bash
mv t.dis t_old.dis
clang -c main.c -O0
#clang -c main.c -O
clang -S foo.c -O -fno-vectorize -mllvm -unroll-count=0
#clang -o t foo.s main.o -Xlinker -emain -nostartfiles -static
clang -o t main.o foo.s -Xlinker -emain -nostartfiles -static
#clang -o t foo.s main.o
#clang -o t main.o foo.s
objdump -dr t > t.dis

After you link the program by using clang -Xlinker -emain -nostartfiles -static, you must run it by using run_bare_metal_sim.pl, or you will get segmentation fault.

Use the VPA to open the pipe file