Mirror of forum.nim-lang.org

4927 :: --opt:speed slow down nimpy code

• oyster (orginal) [2019-06-11T14:42:36+02:00] view original

  Here is the code from a local discussion group

  # foo.nim
  import nimpy
  proc foo(): seq[array[3,int]]{.exportpy.}=
      var a:seq[array[3,int]]
      for i in 1..10000000:
          a.add([i,i*2,0])
      return a
  Run

  and

  # a.py
  import time, foo
  
  n=time.time()
  that=foo.foo()
  print(time.time()-n)
  
  old=time.time()
  def test(num):
      a=[]
      for i in range(num):
          temp=[i,i*2,0]
          a.append(temp)
      return a
  oh=test(10000000)
  print(time.time()-old)
  Run

  when I compile the foo.nim with latest-cloned-and-built nim on Win7 64bits( Nim Compiler Version 0.20.99 [Windows: amd64] Compiled at 2019-06-10) in MSYS2+MINGW with nim c -d:release --threads:on --app:lib --out:foo.pyd foo then

  
  R:\>e:\prg\py\Anaconda3_64\python.exe a.py
  8.30347490310669
  10.290588617324829
  
  R:\>r:\python-2.7.16.amd64\python.exe a.py
  13.0490000248
  21.8800001144
  Run

  shows that nimpy runs faster than pure python which is expected

  however if I compile the python module with nim c -d:release --threads:on --opt:speed --app:lib --out:foo.pyd foo then

  
  R:\>e:\prg\py\Anaconda3_64\python.exe a.py
  17.50800108909607
  17.381994247436523
  
  R:\>r:\python-2.7.16.amd64\python.exe a.py
  17.4839999676
  15.9549999237
  Run

  it is funny that --opt:speed slow down the code! What is the problem?

---

• oyster (orginal) [2019-06-11T14:43:40+02:00] view original

  here is the question two: I suspected that set a as a seq in foo.nim is not a good idea for speed( that is true to set a as a list in python version), so what is a better solution?

---

• oyster (orginal) [2019-06-11T14:43:59+02:00] view original

  question three, as we can find that both python 2 and 3 can use the generated pyd files. So the pyd is not a normal python extension which depends on the python version. How does it come? Will this feature slow down the extension?

  Thanks

---

• juancarlospaco (orginal) [2019-06-11T15:06:54+02:00] view original

  Try newSeqOfCap for speed.

  I dont think you need --threads:on.

  You are not using set on the code.

  Being pure math that can be a func instead of proc but thats really just style.

  Nimpy wont care about Python version, as long as is C API compatible it will work.

  AFAIK.

---