New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Optimize simple for loops #1338

Open
scabug opened this Issue Sep 15, 2008 · 20 comments

Comments

Projects
None yet
2 participants
@scabug
Copy link

scabug commented Sep 15, 2008

I would like to suggest the compiler optimize the common case of for loops, that is,

for (var <- Range [by step])
for (var <- int to int [by step])
for (var <- int until int [by step])

to use while loops under the covers instead of. Currently, nested for loops using ranges/iterators are sometimes an order of magnitude slower than while loops. However, while loop constructs for iterating over arrays are very cumbersome, and the functional style (foreach) is also cumbersome and introduces a lot of function call overhead as well.

The following two matrix multipication implementations

  def matMulUsingIterators (
       a : Array[Array[Double]],
       b : Array[Array[Double]],
       c : Array[Array[Double]]) : Unit = {

    val b_j = new Array[Double](b.length)

    for (j <- 0 until b(0).length) {
        for (k <- 0 until b.length) {
            b_j(k) = b(k)(j)
        }
        for (i <- 0 until a.length) {
            val c_i = c(i)
            val a_i = a(i)
            var s = 0.0d;
            for (k <- 0 until b.length) {
                s += a_i(k) * b_j(k)
            }
            c_i(j) = s
        }
    }
  }

  def matMulUsingRanges (
       a : Array[Array[Double]],
       b : Array[Array[Double]],
       c : Array[Array[Double]]) : Unit = {

    val jRange = 0 until b(0).length;
    val kRange = 0 until b.length;
    val iRange = 0 until a.length;

    val b_j = new Array[Double](b.length)

    for (j <- jRange) {
        for (k <- kRange) {
            b_j(k) = b(k)(j)
        }
        for (i <- iRange) {
            val c_i = c(i);
            val a_i = a(i);
            var s = 0.0d;
            for (k <- kRange) {
                s += a_i(k) * b_j(k)
            }
            c_i(j) = s
        }
    }
  }

are much slower than the same algorithm coded with while loops:

  def matMulUsingWhileLoop (
      a : Array[Array[Double]],
      b : Array[Array[Double]],
      c : Array[Array[Double]]) : Unit = {

    val m = a.length;
    val p = b(0).length;
    val n = b.length;

    val b_j = new Array[Double](b.length);

    var i = 0; var j = 0; var k = 0;
    while (j < p) {
        k = 0
        while (k < n) {
            b_j(k) = b(k)(j);
            k += 1
        }
        i = 0
        while (i < m) {
            val c_i = c(i);
            val a_i = a(i);
            var s = 0.0d;
            k = 0;
            while (k < n) {
                s += a_i(k) * b_j(k);
                k += 1
            }
            c_i(j) = s;
            i += 1
        }
        j += 1;
    }
  }

but the while loop code is more complex and error prone.

(Sorry, Trac appears to remove some line breaks; I
added some explicit semis but might have missed some;
I'll try attaching actual working source code)

Running this while measuring time in nanoseconds:

Iterators   2,807,815,301ns
Ranges      2,789,958,191ns
While Loop  190,778,574ns

MatMul by Iterators is 14 times as slow as with while loops.

It does not appear that the Hotspot runtime profiling and optimization dramatically helps this performance problem
This performance problem can hurt adoption of Scala for many types of uses/applications.

@scabug

This comment has been minimized.

Copy link
Author

scabug commented Sep 15, 2008

@scabug

This comment has been minimized.

Copy link
Author

scabug commented Sep 15, 2008

@DavidBiesack said:

// Scala code to compare performance of nested int loops
object MatMul {


  def matMulUsingIterators (
       a : Array[Array[Double]],
       b : Array[Array[Double]],
       c : Array[Array[Double]]) : Unit = {

    val b_j = new Array[Double](b.length)

    for (j <- 0 until b(0).length) {
        for (k <- 0 until b.length) {
            b_j(k) = b(k)(j)
        }
        for (i <- 0 until a.length) {
            val c_i = c(i)
            val a_i = a(i)
            var s = 0.0d;
            for (k <- 0 until b.length) {
                s += a_i(k) * b_j(k)
            }
            c_i(j) = s
        }
    }
  }

  def matMulUsingRanges (
       a : Array[Array[Double]],
       b : Array[Array[Double]],
       c : Array[Array[Double]]) : Unit = {

    val jRange = 0 until b(0).length // p
    val kRange = 0 until b.length // n
    val iRange = 0 until a.length // m

    val b_j = new Array[Double](b.length)

    for (j <- jRange) {
        for (k <- kRange) {
            b_j(k) = b(k)(j)
        }
        for (i <- iRange) {
            val c_i = c(i)
            val a_i = a(i)
            var s = 0.0d;
            for (k <- kRange) {
                s += a_i(k) * b_j(k)
            }
            c_i(j) = s
        }
    }
  }

  def matMulUsingLimits (
       a : Array[Array[Double]],
       b : Array[Array[Double]],
       c : Array[Array[Double]]) : Unit = {

    val b_j = new Array[Double](b.length)

    val m = a.length
    val p = b(0).length
    val n = b.length

    for (j <- 0 until p) {
        for (k <- 0 until n) {
            b_j(k) = b(k)(j)
        }
        for (i <- 0 until m) {
            val c_i = c(i)
            val a_i = a(i)
            var s = 0.0d;
            for (k <- 0 until n) {
                s += a_i(k) * b_j(k)
            }
            c_i(j) = s
        }
    }
  }

  def matMulUsingWhileLoop (
      a : Array[Array[Double]],
      b : Array[Array[Double]],
      c : Array[Array[Double]]) : Unit = {

    val m = a.length
    val p = b(0).length
    val n = b.length

    val b_j = new Array[Double](b.length)

    var i = 0; var j = 0; var k = 0
    while (j < p) {
        k = 0
        while (k < n) {
            b_j(k) = b(k)(j)
            k += 1
        }
        i = 0
        while (i < m) {
            val c_i = c(i)
            val a_i = a(i)
            var s = 0.0d;
            k = 0
            while (k < n) {
                s += a_i(k) * b_j(k)
                k += 1
            }
            c_i(j) = s
            i += 1
        }
        j += 1
    }
  }

  def time[R](block: => R) : (Long, R) = {
    val start = System.nanoTime()
    val result : R = block
    val time = System.nanoTime() - start
    (time, result)
  }

  val format = new java.text.DecimalFormat("0,000'ns'");
  def report[R](label: String, result: (Long, R)) = {
     println(label + " " + format.format(result._1))
     }

    private val FACTOR = 5
    private val M = 80
    private val N = 70
    private val P = 60

    def main(args : Array[String]) = {
      for (trial <- 3 until 0 by -1) {
      val factor = (if (System.getProperty("factor") != null)
                      Integer.parseInt(System.getProperty("factor"))
                    else
                        FACTOR)
      val multiplier = if (trial == 1) factor else 1;
      val m = M * multiplier
      val n = N * multiplier
      val p = P * multiplier
      val a = new Array[Array[Double]](m,n)
      val b = new Array[Array[Double]](n,p)
      val c = new Array[Array[Double]](m,p)
      println("\nMultiply c[" + m + "," + p + "] = a[" + m + "," + n + "] times b[" + n + "," + p + "]\n");
      val whileTime = time(matMulUsingWhileLoop(a,b,c))
      val iterTime = time(matMulUsingIterators(a,b,c))
      report("Iterators  ", iterTime)
      report("Limits     ", time(matMulUsingLimits(a,b,c)))
      report("Ranges     ", time(matMulUsingRanges(a,b,c)))
      report("While Loop ", whileTime)
      println("MatMul by Iterators is " + iterTime._1 / whileTime._1 + " times as slow as with while loops.")
      }
  }
}
@scabug

This comment has been minimized.

Copy link
Author

scabug commented Apr 10, 2009

Jonathan Shore (jshore) said:
This is a very important performance enhancement for numerical work.

Seems that this could be solved with some pattern matching in the compiler, recognizing a range with no filtering (in the simplest case). Could then remap to a while form for the next stage.

@scabug

This comment has been minimized.

Copy link
Author

scabug commented Apr 10, 2009

@Sciss said:
+1

@scabug

This comment has been minimized.

Copy link
Author

scabug commented Apr 10, 2009

Kipton Barros (kbarros) said:
+1

@scabug

This comment has been minimized.

Copy link
Author

scabug commented May 15, 2009

Erkki Lindpere (villane) said:
+1

It would be nice if for-comprehensions with simple filters could be optimized as well, turning

for (i <- 1 to 10 if shouldProcess(i)) {}

into

var i = 1
while (i < 10) {
  if (shouldProcess(i)) {
  }
  i += 1
}

And extra nice if this would work with random access sequences.
@scabug

This comment has been minimized.

Copy link
Author

scabug commented Oct 22, 2009

Philippe (phdp) said:
+1

This is actually the only thing keeping me from using Scala.

@scabug

This comment has been minimized.

Copy link
Author

scabug commented Nov 30, 2009

@dragos said:
Replying to [comment:12 PhDP]:

+1

This is actually the only thing keeping me from using Scala.
Have you tried '-optimize'? It can help a lot. It's very unlikely this will move from library to compiler-generated loops.

@scabug

This comment has been minimized.

Copy link
Author

scabug commented Dec 20, 2009

Adam Kiezun (akiezun) said:
+1

@scabug

This comment has been minimized.

Copy link
Author

scabug commented Dec 26, 2009

Miguel Garcia (mgarcia) said:

I haven't benchmarked (with and without -optimize) to see whether the current compilation scheme for "simple loops" is good enough.

But in case it isn't, looks like the single place to change in the compiler is method TreeBuilder.makeFor().

According to its comment,

http://lampsvn.epfl.ch/trac/scala/browser/scala/trunk/src/compiler/scala/tools/nsc/ast/parser/TreeBuilder.scala

it performs five transformations. Prepending as special case a transformation for "simple loops" would not change semantics. (Well, assuming that a local definition does not shadow the usual ones: "to" in "1 to 10", "Range", and so on)

Miguel
http://www.sts.tu-harburg.de/people/mi.garcia/

@scabug

This comment has been minimized.

Copy link
Author

scabug commented Jul 5, 2010

Anton Mellit (mellit) said:
I tried to create a script with the following:

def timeit(f : () => Unit) {
    val t1 = System.currentTimeMillis()
    f()
    val t2 = System.currentTimeMillis()

    println(t2-t1)
}    

def repeat(n : Int, f : Int => Unit) : Unit = {
    var i = 0
    while (i<n) {
        f(i)
        i += 1
    }
}

def test0() {
    var i = 0
    var sum = 0
    while (i < 1000000000) {
        sum += i
        i += 1
    }
    println(sum)
}

def test1() {
    var sum = 0
    repeat(1000000000, i => {
        sum += i
    })
    println(sum)
}

def test2() {
    var sum = 0
    for(i <- 0 until 1000000000) {
        sum += i
    }
    println(sum)
}

timeit(test0)
timeit(test1)
timeit(test2)

Result is:

-1243309312
467
-1243309312
504
-1243309312
11899

May be this 'repeat' is a workaround? Warning: works only with 'scala -optimise'. This is not very stable, sometimes some seemingly minor modifications, i.e. moving the code outside of the function, break it and I get 12000 for 'repeat'.

@scabug

This comment has been minimized.

Copy link
Author

scabug commented Jul 6, 2010

@DavidBiesack said:
Replying to [comment:28 mellit]:

I tried to create a script with the following:

{code}
...
def repeat(n : Int, f : Int => Unit) : Unit = {
var i = 0
while (i<n) {
f(i)
i += 1
}
}
}}

Thanks for the suggestion. I hesitate to try or rely on this for several
reasons:

  1. having to remember to not use the standard for loop is problematic. The syntax is different and not based on generators and less general, although one could certainly patch it to be more general; i.e pass in start and end and optional increment values: repeat(start:Int, end:Int, increment:Int) or repeat(range:Range) etc. (and also handle backwards iteration when called for).

  2. This may still involve at least an extra function call per iteration. If the compiler has to inject other synthetic calls into the body to allow access to other lexically scoped variables, this may also affect performance. The goal of this optimization is to achieve Java-level performance of for loops where possible.

  3. most critically, if/when the Scala compiler implements this ticket's optimization, then all code using this repeat control would not get the optimization, requiring code maintenance to undo it.

@scabug

This comment has been minimized.

Copy link
Author

scabug commented Sep 18, 2010

@paulp said:
Please see update on this ticket sent to scala-user, also available here: http://scala-programming-language.1934581.n4.nabble.com/optimizing-simple-fors-td2545502.html#a2545502

I very much agree with mgarcia's comment of nine months ago that TreeBuilder.makeFor already does a whole pile of tree transformations and there is no convincing reason we shouldn't add one which has this kind of impact. Failing agreement on that point, I believe we have a pressing responsibility to clean up the parsing phase and plugin architecture sufficiently that it would be possible to do this transformation with a compiler plugin.

@scabug

This comment has been minimized.

Copy link
Author

scabug commented Oct 3, 2010

Olivier Chafik (olivier.chafik) said:
Hello,

I've just written such a compiler plugin, which you can install straight away on 2.8.0 with sbaz for testing :
http://article.gmane.org/gmane.comp.lang.scala.user/31814

It's probably full of bugs, but it rewrites int-range for loops with no filters and Array[T].foreach, Array[T].map into equivalent while loops.

You can have a look at the auto tests to see the supported cases :
http://code.google.com/p/nativelibs4java/source/browse/branches/OpenCL-BridJ/libraries/OpenCL/ScalaCLPlugin/src/test/scala/scalacl/

Looking forward to seeing something like this mainstream :-)
Cheers

zOlive

@scabug

This comment has been minimized.

Copy link
Author

scabug commented Oct 4, 2010

Olivier Chafik (olivier.chafik) said:
I've adapted the fannkuch and nbody benchmarks that were in the scala-user thread mentioned previously and I had to adapt it a bit (inlining the ranges that were stored as val range = x until y).

Here's the modified code :
http://nativelibs4java.sourceforge.net/sbaz/scalacl/examples/

And to run it (with ScalaCL plugin installed via sbaz: sbaz install scalacl-compiler-plugin) :
DISABLE_SCALACL_PLUGIN=1 scalac fannkuch.scala && scala fannkuch
scalac fannkuch.scala && scala fannkuch

DISABLE_SCALACL_PLUGIN=1 scalac nbody.scala && scala nbody
scalac nbody.scala && scala nbody

With the plugin turned on, the performance of the three variants (While, Limit, Range) is the same (the first while is actually slower, I haven't investigated why).

@scabug

This comment has been minimized.

Copy link
Author

scabug commented Oct 6, 2010

Olivier Chafik (olivier.chafik) said:
(Sorry for spamming you again, this should be the last time)

I've just enhanced the plugin with more conversions to while loops :

  • Array.foldLeft / foldRight
  • Array.reduceLeft / reduceRight
  • Array.scanLeft / scanRight
    (in addition to Array.foreach, Array.map and the int range foreach with no filter)

Also, the conversions should now work on method references and inline lambdas the same way.

Further progress and plans can be tracked at the bottom of this page :
http://code.google.com/p/scalacl/wiki/ScalaCLPlugin

@scabug

This comment has been minimized.

Copy link
Author

scabug commented Nov 15, 2012

Herrmann Khosse (herrkhosse) said:
+1

@scabug

This comment has been minimized.

Copy link
Author

scabug commented Jul 21, 2013

@SethTisue said:
"Spire also provides a loop macro called cfor whose syntax bears a slight resemblance to a traditional for-loop from C or Java. This macro expands to a tail-recursive function, which will inline literal function arguments." https://github.com/non/spire

@scabug

This comment has been minimized.

Copy link
Author

scabug commented Jul 21, 2013

Olivier Chafik (ochafik) said:
Here's another loop macro with an arguably better syntax: Scalaxy/Loops (which reuses code from ScalaCL):

https://github.com/ochafik/Scalaxy/tree/master/Loops

@scabug

This comment has been minimized.

Copy link
Author

scabug commented Mar 28, 2014

@DarkDimius said:
There's a different approach that I've tried in scalaBlitz: dont require users switch from using standard library while they code, but instead give a macro that changes implementation methods, replacing standard library implementation with macro-based one.
Here's small description of it: http://scala-blitz.github.io/home/documentation//optimize.html

The Range example in this ticket will be compiled to while loops and get same performance.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment