LI79792: PERFORMANCE IMPROVEMENT FOR __POPCNT8()

XL C/C++ for Blue Gene/Q Fix Pack 15 (October 2017 Update) for 12.1

APAR status

• Closed as program error.

Error description

• When using the __popcnt8() builtin, there is a performance
  concern that needed to be addressed via this APAR.
  
  The following test case demonstrates the issue:
  ===== COMPILE COMMAND:
  /bgsys/drivers/ppcfloor/comm/xl/bin/mpixlcxx_r -O5 -qsmp=omp
  -qthreaded test.c -I/bgsys/drivers/V1R2M2/ppc64
  
  ===== TESTCASE:
  % cat test.c
  #include <stdio.h>
  #include <stdlib.h>
  
  #include <stdio.h>
  #include <inttypes.h>
  
  
  #include "hwi/include/bqc/A2_inlines.h"
  #define  size  (128*1024*1024)
  volatile int64_t value[size];
  
  inline int popcnt8(uint64_t data)
  {
        int ldz;
         asm volatile ("popcntd %0, %1\n\t"
         : "=r" (ldz)
         : "r" (data)
         );
         return ldz;
  }
  
  char BitsSetTable256[256];
  
  int popcount(long long b)
  {
       b = (b & 0x5555555555555555LU) + (b >> 1 &
  0x5555555555555555LU);
       b = (b & 0x3333333333333333LU) + (b >> 2 &
  0x3333333333333333LU);
       b = b + (b >> 4) & 0x0F0F0F0F0F0F0F0FLU;
       b = b + (b >> 8);
       b = b + (b >> 16);
       b = b + (b >> 32) & 0x0000007F;
       return (int) b;
  }
  
  int popcount_table(long long v) {
    int c = BitsSetTable256[v & 0xff] +
      BitsSetTable256[(v >> 8) & 0xff] +
      BitsSetTable256[(v >> 16) & 0xff] +
      BitsSetTable256[v >> 24];
    return c;
  }
  int counttz(long long v) {
    int c;
    if (v)  {
      v = (v ? (v - 1)) >> 1;  // Set v's trailing 0s to 1s and
  zero rest
      for (c = 0; v; c++)
         v >>= 1;
    }
    else
      c = 8 * sizeof(v);
  }
  
  
  
  int main(int argc, char * argv[])
  {
    BitsSetTable256[0] = 0;
  
    for (int i = 0; i < 256; i++)
      BitsSetTable256[i] = (i & 1) + BitsSetTable256[i / 2];
  
  
  
    uint64_t t2, t1, software_cycles, hardware_instruction_cycles;
    uint64_t thr_count[64];
    int nthreads;
    #pragma omp parallel for
    for(int i=0;i<size;i++)
      value[i] = i;
  
    t1 = GetTimeBase();
    #pragma omp parallel
    {
      int i, tid = omp_get_thread_num();
      if(tid ==0) nthreads = omp_get_num_threads();
      uint64_t count = 0;
      thr_count[tid] = 0;
      #pragma omp for
      for(i=0;i<size;i++)
        count += popcount_table(value[i]);
      thr_count[tid] = count;
    }
    t2 = GetTimeBase();
    software_cycles = t2 - t1;
    for(int i = 1;i<nthreads;i++)
      thr_count[0] += thr_count[i];
    printf("Software popcnt cycles: %llu, count
  %llu\n",software_cycles, thr_count[0]);
  
    t1 = GetTimeBase();
    #pragma omp parallel
    {
      int i, tid = omp_get_thread_num();
      uint64_t count = 0;
      thr_count[tid] = 0;
      #pragma omp for
      for(i=0;i<size;i++)
         count += __popcnt8(value[i]);
      thr_count[tid] = count;
    }
    t2 = GetTimeBase();
    hardware_instruction_cycles = t2 - t1;
    for(int i = 1;i<nthreads;i++)
      thr_count[0] += thr_count[i];
    printf("hardware_instruction popcnt cycles: %llu, count
  %llu\n", hardware_instruction_cycles, thr_count[0]);
  
  
  }
  %
  
  
  ===== ACTUAL OUTPUT:
  % /bgsys/drivers/ppcfloor/hlcs/bin/runjob --exe
  a.out --block R00-M0-N04 --np 1 -p 1
  Software popcnt cycles:                   115842638, count
  1811939328
  hardware_instruction popcnt cycles: 242737020, count 1811939328
  %
  

Local fix

• N/A
  

Problem summary

• USERS AFFECTED:
  Users using __popcnt8() builtin in their application and call it
  a number of times within a loop maybe affected by this issue.
  
  PROBLEM DESCRIPTION:
  The popcnt hardware ops are slow on BG/Q.
  

Problem conclusion

• The compiler should not use popcnt hardware ops.
  It has been changed to use a software emulation and shows
  improvement in performance.
  

Temporary fix

Comments

APAR Information

• APAR is sysrouted FROM one or more of the following:

• APAR is sysrouted TO one or more of the following:

Fix information

• Fixed component name

  XL C/C++ FOR BG

• Fixed component ID

  5799AG100

Applicable component levels