Sunday, June 7, 2015

Node JS Multiplying Float Matrices

I don't know why I wrote such stupid code for readMatrix.js. I guess when someone uses a new language their brain cells seize up, effective function becomes impossible.

After tests I ran on Friday, I wanted to see how Javascript does with floating point numbers? JS only supports 64 bit floating-point numbers, but there must be special code in their to make integers look like integers ... unless their isn't.

When I went to update my code to support numbers with fractions, my first thought was to flag whether a decimal point had been seen and then keep track of the value by which to divide, multiplying by ten for each fractional digit. I quickly realized I was sinking into the mire of coding stupidity; JS must have a function to convert strings into integers or floats, and indeed it has parseInt() and parseFloat(). But in my research I discovered ( or re-discovered ) that there's no need for the conversion. Like Perl, JS allows arithmetic on the string representation of numbers:

    1  +  2 ⇛ 3
   "1" + "2"⇛ 3

This made me think about how I would parse the file in Perl, using split(). I knew JS had map(), so quite likely it would have split() as well. And it does! So all that manual processing went in the trash, and I simplified readMatrix.js to

   exports.readMatrix = function (err, data) {
      if (err) {
         throw err;
      var rows = data.split(/ *\n/)
                     .filter(function(val) {
                         return val.length > 0;
      var matrix = [], i, len;
      for ( 1 = 0, len = rows.length; i<len; i++ ) {
         matrix[i] = rows[i].split(/ +/);
      return matrix;                  

I think I had the two chunks merged, in an online TryIt editor I was experimenting with,but this is compact enough for the moment.

{Correction - Although the original JS used only floats, including for integer values, Googles V* JS engine, the one used in node.js, has separate 32 bit integers.}

Although integers are represented by 64-bit IEEE floating point numbers, integer multpiplication is still 3 times as fast as floating point multiplication. There are clearly optimizations that are being made use of to improve performance.

Calculating average operations per second shows small matrixes have performance limited by other overhead, but larger matrices are pumping along at 7.8 million multiplications per second. 

No comments: