bmantl/autodesktask

A task for Autodesk's recruitment process

Autodesk Task

A story of matrix multiplication.

Goals
Implementation
Compilation
Usage
Benchmark conclusions

Goals:

• Get in touch with JNI
• Implement Matrix multiplication
• Benchmark Java against C++

Implementation

Matrix calsses

Java

• Matrix.java - an implementation using a 2d double array
• Matrix1D.java - an implementation using a 2d double array

Both can multiply using the multiplyBy(other) function.
There is no relation between the classes as the use a different data structure.
There are segments of code that looks almost the same, as the logic is identical betwee the two.
It seemed redundant to create some sort of super class (or interface) as the functions would have to be overriten because of the different data structure.

C++

• Mat.h/.cpp - an implementation using a 2d double array
• Matrix.h/.cpp - an implementation using a 2d double array

Both can multiply using the multiplyBy(other) function.
Design-wise, its identical to the Java variant.

Program classes

• Main.java - for running a basic multiplication with a jni call
• Benchmark.java - for benchmarking over the 4 implementations

Usage in the usage part....

Why two implementations? (2D, 1D)

I started off with a simple 2d array implementation for both c++ and java, using primitive type arrays.
It appears that this kind of implementation is quite slow (especially for the c++ part :\ ).
Writing a 1D implementation of the matrix on c++ (using [i * rows + j] kind of querying) I got much (more than 100%) better results that the 2D implementation.

I did a java one as well to have a good comparison (Java does quite well).

*** All that said testing with a linux sitting on a vm.

Compilation

Jni part

Located in the /jni folder there's a makefile.
It generates .h files using javah and compiles the c++ code.
make sure that the dll/so/jnilib are there.

Warning:

You might want to change the jdk folders in the makefile (or even the library compilation) fo compilation on other OSs.

To make all of the relevant files:

$ make all

Java part

It is already compiled and located in the /bin folder.

To compile again: Using eclipse as a project manager or in the /src folder

$ javac Matrix1D.java
$ javac Matrix.java
$ javac Main.java
$ javac Benchmark.java

Usage

Main.class

$ java -Djava.library.path=../jni Main -i <input-file-location> -o <output-file-location>

-i - an input file. There are files in the /Matrices folder.
-o - Optional. An output file to write the multiplication resullt. Optional, but without it it will multiply and save the result in the defOut.out file in the running folder.

Benchmark.class

Used for benchmarking.
Outputs running times for all implementations per input.
Takes inputs from /Matrices folder.

$ java -Djava.library.path=../jni Benchmark -s <start-size> -e <start-size>

-s - the start size for the benchmarking. Given in 1-20. i.e. 1 is for the 100.in file. 11 is for 1100.in
-e - Optional. The end size for the benchmarking. Given in 1-20. i.e. 1 is for the 100.in file. 11 is for 1100.in
the default will benchmark start + 5 size.

Outputs a file in /Matrices called benchmark_s_e_numOfRunsPerFile.out

Benchmark conclusions

• 1D implementation speed >>>> 2D implementation speed.
• Though it should come out that c++ is faster that java, sometimes Java is faster (might need a better eval).
• If the numOfRuns field in the Benchmark class is > 1, c++ runs worse on average. This might be due to the fact that the call to the jni suffers a huge overhead (which accumilates).

The results are located in the /Matrices in the benchmark...log files.