Record event counts per thread

perf record can be used to record event counts for every thread when using -s option. Check following multi-thread program: 

  1. $ cat sum.c
  2. #include <omp.h>
  4. #define N 100000000
  5. #define THRAED_NUM 8
  7. int values[N];
  9. int main(void)
  10. {
  11.     int sum[THRAED_NUM];
  13.     #pragma omp parallel for
  14.     for (int i = 0; i < THRAED_NUM; i++)
  15.     {
  16.         int local_sum;
  17.         for (int j = 0; j < N; j++)
  18.         {
  19.             local_sum += values[j] >> i;
  20.         }
  21.         sum[i] = local_sum;
  22.     }
  24.     return 0;
  25. }

Build and use perf record to note down event counts for every thread: 

  1. # gcc -fopenmp -g sum.c -o sum
  2. # perf record -s ./sum
  3. [ perf record: Woken up 1 times to write data ]
  4. [ perf record: Captured and wrote 0.404 MB perf.data (8757 samples) ]

Use perf report to analyze it (-T option means displaying per-thread event counts): 

  1. # perf report -T
  2. ......
  3. #  PID   TID      cycles:uppp
  4.   9960  9963        751824252
  5.   9960  9961        750625307
  6.   9960  9965        749742594
  7.   9960  9967        749228142
  8.   9960  9968                0
  9.   9960  9970          9857914
  10.   9960  9974          9150516
  11.   9960  9977          8928628
  12. ......