Moving from Java to CUDA

If you move from Java to CUDA:

• you change the algorithms

• you change the compilers

• you change the execution environment

• you change the hardware architecture

Preparing the Environment: the Java code.

We want to implement our code

• on another environment

• using another compiler

BUT which pieces of code?

Selecting the code

We want to keep part of the code in Java - for interfacing reasons

We want to speed it up in CUDA:

• suited to computationally intensive algorithms

• not suited for I/O- bound algorithms

Isolating the code

• semantic isolation:

  • reduce code dependencies

  • interfacing between classes

• structural isolation:

  • move classes to other packages

  • move constants/fields/methods to other packages

Common code - some pieces of code may have to be kept in both Java and CUDA:

• place them in common package

• make copies of the code

• and dont loose track of them ;-)

Interfacing

Interfacing is important:

• make access to the code independent of implementation (Strategy design pattern :-( )

• implementations can be selected on-the-fly

• results from different implementations can be compared (method contracts are the same)