Mirror of forum.nim-lang.org

8836 :: usage of OpenMP parallel loop iterator for lazybones?

• oyster (orginal) [2022-01-22T06:41:13+01:00] view original

  I have serached this forum, and checked the doc

  however, when I try to compile the following code by nim c -d:release -d:speed --stackTrace:off --opt:speed --passL:-s --passc:-flto --passc:-fopenmp --passL:-fopenmp -d:openmp a_openmp I get

  
  ...............................................................................fatal.nim(53)            sysFatal
  Error: unhandled exception: index out of bounds, the container is empty [IndexDefect]
  Run

  why? Thanks

  import std/math
  import std/times
  
  var
      total: int32 = 1
      beginCount: int32 = 3
      endCount: int32 = 1000_0000
      cacheSize: int32 = 400
      cachePtr: int32 = -1
      primesCache: array[400, int32]
  
  proc isPrime(n:int32): bool =
      var cur = int32(0)
      for i in 0 .. cachePtr:
          cur = primesCache[i]
          if n mod cur == 0:
              return false
      var m = int32(sqrt(float(n)))
      while true:
          cur += 2
          if cur > m:
              break
          if n mod cur == 0:
              return false
      if cachePtr < cacheSize-1:
          cachePtr += 1
          primesCache[cachePtr] = n
      return true
  
  
  let time=cpuTime()
  
  var i = beginCount
  for i in  `||`(beginCount, endCount):
      if (i mod 2) == 0:
          continue;
      if isPrime(i):
          total += 1
  
  let endTime = cpuTime() - time
  echo "Time taken: ", endTime, "s"
  echo total
  Run

---

• mratsim (orginal) [2022-01-22T12:22:05+01:00] view original

  -d:openmp doesn't do anything by itself. if you saw that from Arraymancer it's because I have code like

  when defined(openmp):
    doSomething()
  Run

  Any variable that is shared among threads MUST deal with thread synchronization:

  • total is shared and if 2 threads tries to do +1 at the same time, you might end up with +1 in total instead of +2. So you need atomics here, or have local per-thread count that you sum at the end
  • Same thing for cachePtr and primesCache

  Like here: https://github.com/mratsim/Arraymancer/blob/6cfe743/src/arraymancer/tensor/backend/openmp.nim#L18-L34, you need a per thread-cache or they'll mess each other up.

  when defined(openmp):
    when not defined(cuda): # For cuda, OpenMP flags must be passed
      {.passC: "-fopenmp".} # behind -Xcompiler -fopenmp
      {.passL: "-fopenmp".}
    
    {.pragma: omp, header:"omp.h".}
    
    proc omp_set_num_threads*(x: cint) {.omp.}
    proc omp_get_num_threads*(): cint {.omp.}
    proc omp_get_max_threads*(): cint {.omp.}
    proc omp_get_thread_num*(): cint {.omp.}
  
  else:
    template omp_set_num_threads*(x: cint) = discard
    template omp_get_num_threads*(): cint = 1
    template omp_get_max_threads*(): cint = 1
    template omp_get_thread_num*(): cint = 0
  Run

  What's very possible is that due to lack of synchronization, cachePtr was incremented beyond 400.

---