Skip to main content

BASH Matrix Multiplication

tl;dr Bash is not the language for math-intensive operations.





REPS=$1;
FILE_1=$2;
FILE_2=$3
OUTFILENAME=$4;

readonly COLS=`head -1 $FILE_1 | wc -w`;
readonly ROWS=`cat $FILE_1 | wc -l`;
# echo "rows is $ROWS; cols is $COLS"
if [[ $ROWS != $COLS ]]; then
    echo "Expecting square matrices, " \
         "but rows = $ROWS, cols = $COLS\n";
    exit 1;
fi

# --------------------------------------------------
# SUBROUTINES
#
function outputMatrix()
{
    local matrixname=$1;
    local matrix;
    local elem;
    echo "matrix is '$matrixname'.";
    eval matrix=\( \${${matrixname}[@]} \);
    local i=0;
    for elem in "${matrix[@]}"; do
 echo -n "$elem ";
 if (( ++i == $COLS )); then
     echo '';
     i=0;
 fi
    done  
}

function multiply()
{
    declare -a product;
    local M=$1 N=$2;
    local i j k idx1 idx2 idx3;
    for ((i=0; i < $ROWS; i++ )); do
        for ((j=0; j<$COLS; j++)); do
            local result=0;
            for ((k=0; k<$ROWS; k++)); do
                idx1=$((i*$ROWS+k));
                idx2=$((k*$COLS+j));
                result=$((result+${M}[idx1]*${N}[idx2]));
            done
            idx3=$((i*$COLS+j));
            product[$idx3]=$result;
        done
    done
#    outputMatrix product;
}
# --------------------------------------------------
#

matrix1=($(cat $FILE_1));
matrix2=($(cat $FILE_2));


start=`date  +%s.%N`;
for ((repeat = 0; repeat < $REPS; repeat++)); do
    multiply matrix1 matrix2 
done
end=`date  +%s.%N`;
echo "$end - $start" | bc

My core multiplication routine is similar to the one at RosettaCode. Results are drastically inferior compared to Perl or JavaScript, a factor or 200x ~ 1000x as slow.


Notice that the 100x100 matrix has only a single data point, at 55,000 multiplications per second. That took 547 seconds; I wasn't going to wait around 2 hours for more repetitions.

Part of the problem is from flattening a two-dimensional matrix into a single one-dimensional array. Converting indices into a single array index overwhelms the actual matrix multiplication, resulting in 3N^3 + N^2 multiplications. These all happen in "user space", while Perl & JS have the opportunity to carry out array indexing in "language space", potentially more efficient.

Actually, simplifying the index calculations only improved runtime about 10%, not a whole lot. Possibly variable access overwhelms the difference between addition and multiplication.


    for ((i=0; i <; $ROWS; i++ )); do
        for ((j=0; j < $COLS; j++)); do
            local result=0;
            local idx1=$((i*$ROWS));
            local idx2=$j;

            for ((k=0; k < $ROWS; k++)); do
                result=$((result+${M}[idx1]*${N}[idx2]));
                ((idx1 += 1))
                ((idx2 += COLS))
            done
            product[$idx3]=$result;
            ((idx3 ++))
        done
    done



Comments

Popular posts from this blog

Creating Perl5 Objects with Moxie

Having in the previous article prepared data types for car suits and card ranks, I can now combine them to provide a playing card class, using Stevan Little's Moxie module (version 0.04, so definitely early days.) The goal is to provide an object-oriented paradigm to the Perl 5 programming language which is more sophisticated, more powerful and less verbose than manually bless()-ing hashes. To achieve that goal it needs to be faster and light-weight compared to Moose. Currently, Moxie.pm and and MOP.pm are add-on modules, but eventually, when they are more complete, when the wrinkles have been ironed out, and when they have gained acceptance and a community of users, they might be merged into the Perl core.

One significant feature of Moxie is that it reduces boilerplate code. You don't have to specify warnigns or strict. As well, the features or the perl you are using are enabled, among them say, state, signatures, and post_deref.
A Simple Moxie Class package Card { …

Perl5, Moxie and Enumurated Data Types

Moxie - a new object system for Perl5 Stevan Little created the Moose multiverse to upgrade the Perl 5 programming language's object-oriented system more in line with the wonderfull world of Perl 6. Unfortunately, it's grown into a bloated giant, which has inspired light-weight alternatives Moos, Moo, Mo, and others. Now he's trying to create a modern, efficient OO system that can become built into the language.

I've seen a few of his presentations at YAPC (Yet Another Perl Conference, now known as TPC, The Perl Conference), among them ‎p5 mop final final v5 this is the last one i promise tar gz<. So I was delighted to recently see an announcement of the module Moxie, and decided to try implementing a card game.

While the package provides some POD documentation about the main module, Moxie, it doesn't actually explain the enum package, Moxie::Enum. But delving into the tests directory reveals its secrets.
Creating an Enum package Ranks { use Moxie::Enum; …

Adventures in Autovivification

Having recently started a new job, I was exposed to old code with multi-step tests against autovivification in multi-level hashes. You get used to the code you have seen, but in a new environment it‘s irritating and jarring.
Moose does not generally have the problem, first because class structure is pre-declared, because values are accessed using accessor functions rather than directly, and because responsibility is delegated down to attributes, avoiding long chains. On the other hand, Moose has it's own overhead, so hand-rolled objects, and bare structures still have their use.
If you don‘t protect against autovivification, then mis-spelling a key, or referencing keys which haven‘t been instantiated in this instance, causes those keys to instantly come into existence.
#!/usr/bin/perl use warnings; use strict; use Data::Dump 'dump'; use 5.024; my $var = {key1 => {key2 => {key3 => 'a'}}}; say dump $var; if ( $var->{key1}{key2}{key3b}[13]{foob…