Comparing doubles using ULPs (Units in the last place)

问题

I have succeeded in writing an Ulps based function that compares two doubles for equality. According to this page, the comparison can be made using a combination of absolute and relative epsilon or using integers (Ulps).

I have made both epsilon based and Ulps based functions. This is the epsilon based function:

var IsAlmostEqual_Epsilon = function(a, b)
{
  if (a == b) return true;
  var diff = Math.abs(a - b);
  if (diff < 4.94065645841247E-320) return true;
  a = Math.abs(a);
  b = Math.abs(b);
  var smallest = (b < a) ? b : a;
  return diff < smallest * 1e-12;
}

And this is the Ulps based (DoubleToInt64Bits, subtract, negate and lessthan functions are in the below mentioned JSBIN):

var IsAlmostEqual_Ulps = function(A, B)
{
  if (A==B) return true;
  DoubleToInt64Bits(A, aInt);
  if(aInt.hi < 0) aInt = subtract(Int64_MinValue, aInt);
  DoubleToInt64Bits(B, bInt);
  if(bInt.hi < 0) bInt = subtract(Int64_MinValue, bInt);
  var sub = subtract(aInt, bInt);
  if (sub.hi < 0) sub = negate(sub);
  if (lessthan(sub, maxUlps)) return true;
  return false;
}

According to Bruce Dawson the Ulps based is preferred. IsAlmostEqual_Ulps is working ok according to test base of 83 cases, but the function is pretty slow. It takes about 700-900 ms to complete the test base (JSBIN) when executed as a standalone html (outside JSBIN). Epsilon based IsAlmostEqual_Epsilon takes only about 100 ms.

Is there anything that can be done to speedup IsAlmostEqual_Ulps function? You can propose also a completely different solution or some fixings to my code.

I have tested already the inlining everything, but it cuts the time only about 5-10%. I'm hunting something like 50-80% improvement in execution time. 100-500% improvement would be fine, but it may be only a dream.

right_answers in the JSBIN code are got using C# IsAlmostEqual function (see at the top of JSBIN code). Both above functions give the same results in all 83 cases.

EDIT:

C++ version from here:

bool IsAlmostEqual(double A, double B)
{
    //http://www.cygnus-software.com/papers/comparingfloats/comparingfloats.htm
    long long aInt = reinterpret_cast<long long&>(A);
    if (aInt < 0) aInt = -9223372036854775808LL - aInt;
    long long bInt = reinterpret_cast<long long&>(B);
    if (bInt < 0) bInt = -9223372036854775808LL - bInt;
    return (std::abs(aInt - bInt) <= 10000);
}

回答1:

Well, since my original "insight" wasn't too helpful, here's another answer, that simply takes your existing functions and reorganizes them to minimize repetitive object constructor calls:

var ULPS2 = function(){
  var buf2 = new ArrayBuffer(8);
  var dataView2A = new DataView(buf2);
  var dataView2B = new DataView(buf2);
  var aInt=new Int64(0, 0), bInt=new Int64(0, 0);
  var sub;

  this.IsAlmostEqual=function(A,B){
    if (A==B) return true;
    dataView2A.setFloat64(0, A);
    aInt.lo = dataView2A.getInt32(4) | 0;
    aInt.hi = dataView2A.getInt32(0) | 0;
    dataView2B.setFloat64(0, B);
    bInt.lo = dataView2B.getInt32(4) | 0;
    bInt.hi = dataView2B.getInt32(0) | 0;

    if(aInt.hi < 0) aInt = subtract(Int64_MinValue, aInt);

    if(bInt.hi < 0) bInt = subtract(Int64_MinValue, bInt);
    sub = subtract(aInt, bInt);
    if (sub.hi < 0) sub = negate(sub);
    if (lessthan(sub, maxUlps)) return true;
    return false;
  }

}
var Ulps2=new ULPS2();

Based on a test in Chrome and Firefox at http://jsbin.com/IWoyIDO/2/ , it appears to yield a 30%-50% improvement. Not phenomenal, but at least better than the 5-10% improvement you were originally mentioning.

回答2:

There is one way that would be faster for sure (currently as fast as 1.5 times the speed of native code), which would be to use asm.js, a highly optimizable subset of Javascript.

This is the unreal gaming engine running on it, to have an idea of the type of performance (takes a bit to load but runs very smoothly, use firefox or chrome).

The way it works is that you take your code and write in a statically typed language: C or C++. Then compile it into bytecode of the LLVM C++ virtual machine. Then use the emscripten compiler to generate asm.js Javascript code.

If the browser does not detect and optimize asm.js, it runs it as Javascript. If the browser (like latest versions of Chrome/Firefox) detects asm.js, you will have a near native performance.

Check the blog post of John Resig (creator of jQuery) about it. You can't get faster than that in a browser these days, and it keeps getting faster (see Mozilla team blog post here).

回答3:

Here is the basic gist of how to improve the performance using new browser functionality of typed arrays. Note that because no native 64-bit int view is provided, you will have to rework the logic on your own based on 2 32-bit ints, and ensure endianness does not affect the calculation on different systems. I am just trying to address the javascript speed portion of the problem, not be an expert on float encoding. Hopefully this is clear:

var IsAlmostEqual_Ulps = function(A, B)
{
  var smallBufferA = new ArrayBuffer(8); //8 bytes = 64bits
  var viewAsDoubleA = new Float64Array(smallBufferA); //byteOffset=0, length=8 optional
  var viewAsIntA = new Int32Array(smallBufferA); //byteOffset=0, length=8 optional
  viewAsDoubleA.set([A]);
  console.log(viewAsIntA);
  var smallBufferB = new ArrayBuffer(8); //8 bytes = 64bits
  var viewAsDoubleB = new Float64Array(smallBufferB); //byteOffset=0, length=8 optional
  var viewAsIntB = new Int32Array(smallBufferB); //byteOffset=0, length=8 optional
  viewAsDoubleB.set([B]);
  console.log(viewAsIntB);
  //The logic below is just illustrative and may not be right. Please test & adjust
  var A_lo=viewAsIntA[1];
  var B_lo=viewAsIntB[1];
  var A_hi=viewAsIntA[0];
  var B_hi=viewAsIntB[0];
  if(A_hi<0){A_hi=Int32_MinValue-A_hi;}
  if(A_hi<0){A_hi=Int32_MinValue-A_hi;}
  //You'll have to rewrite these slightly, I think you know how
  var sub = subtract(viewAsIntA, viewAsIntB); 
  if (sub.hi < 0) sub = negate(sub);
  if (lessthan(sub, maxUlps)) return true;
}

For example, if you called IsAlmostEqual_Ulps(3.14159,3.14159001), you would get logged in your console:

[-266631570, 1074340345]
[-244113572, 1074340345]

That should get you the biggest performance gains. If you want to make it even faster, you should reuse existing buffers (from the global scope or a closure) between function calls to not incur the constructor overhead and the garbage collection.

来源：https://stackoverflow.com/questions/21340372/comparing-doubles-using-ulps-units-in-the-last-place

标签

javascript

c++

performance

floating-point-precision

emscripten