Introduction to NDSL

The NASA/NOAA Domain Specific Language (NDSL) is a Python based middleware which allows users to accelerate their code and seamlessly switch between CPU and GPU execution. To achieve this, NDSL leverages the functionality of GT4Py and optimization capabilities of DaCe to creating a smooth, unified user experience which allows for easy development of grid-based models.

NDSL has extensive documentation, and this should serve as a first point of reference for questions about the system; however, it is important to note that GT4Py and DaCe also have their own sets of documentation, and these may be useful - particularly for more nuanced applications.

NDSL - like all DSL's - is powerful because it is restrictive. Every degree of freedom added exponentially increases optimization difficulty, and ultimately lowers performance ceilings. While this software has numerous features designed to make the development process smoother, there will inevitably be situations where desired patterns fail to align seamlessly with core concepts of this language. In those moments, it is important to remember that these patterns were "banned" because their implementation would dramatically reduce (or, in some cases, completely eliminate) the create a nimble, efficient executable product.

The requirement to (occasionally) rethink old methods may - and is - a significant request; however, with it comes significant benefit. Below are a few examples of the performance gains achieved with NDSL:

PERFORMANCE GAINS HERE

This user manual introduces relevant concepts and provides knowledge needed to use NDSL. By the end of this guide, you will:

• Learn about the unique storage objects used by NDSL
• Possess a foundational understanding of the core concepts of GT4Py
• Move on to more advanced techniques
• Discuss some common patters and their implementation
• Illustrate the importance of using classes and other object oriented coding practices

This guide is not a FAQ or a replacement for proper documentation. These resources are available [LINK(s) HERE]