//  sum all elements in an array, multicore
//   result  
//           one core 156 clocks
//           dual core 109 clocks
//           quad core 82 clocks

ax[20], s0, s1, s2, s3

suma(i)
  s = 0
  while(i < 20)
    s = s + ax[i]
    i = i + 4
  return s

printa()
  i = 0
  while(i < 20)
    print(ax[i])
    i = i + 1

core0()
  s0 = suma(0)  
  sync()
  s = s0 + s1 + s2 + s3
  print(s)

core1()
  s1 = suma(1)
  sync()

core2()
  s2 = suma(2)
  sync()

core3()
  s3 = suma(3)
  sync()

#data
initdata()
  asm("  1 2 3 4 5 6 7 8 9 10")
  asm("  11 12 13 14 15 16 17 18 19 20")