AUTO-RPE

Tool to automatically implement the Reduced Precision Emulator into a Fortran code. It includes a workflow manager to launch precision analysis

Documentation Code

Automatic Reduced Precision Emulator for large Fortran-based models

AutoRPE is a tool designed to optimise the numerical precision of computational science models written in Fortran. Originally developed to enable the use of a Reduced Precision Emulator (RPE) (Dawson, A. and Düben, P. D.: 2017) in large-scale codes such as NEMO, AutoRPE has evolved into a framework for streamlining code modifications, managing precision experiments, and facilitating precision sensitivity analysis. By automatically instrumenting models with custom precision-aware data types, it allows researchers to create mixed-precision versions without manually modifying hundreds of thousands of lines of code, and provides a workflow to systematically test and evaluate precision configurations to reduce runtime, memory usage, and computational cost while preserving scientific accuracy.

The rational
“Global warming and climate change pose significant challenges to humanity due to their social, economic, and environmental impacts. Better understanding and improved capacity to forecast climate evolution can support the development of ambitious adaptation policies. Earth system models (ESMs) are crucial for simulating and predicting climate behavior, but their computational demands continue to grow with advancements such as higher resolution and ensemble forecasting.
In this context, efficient use of computational resources has become essential, and mixed-precision approaches are emerging as a way to enhance performance by reducing memory and computational costs. Most modern earth-system models are memory-bound, and using uniform high precision throughout the code often leads to over-engineering, where precision is dictated by the most sensitive parts while other sections consume unnecessary resources. Optimizing precision to match the actual requirements of each variable provides a balance between performance, memory usage, and accuracy.
However, identifying which variables can safely operate at reduced precision in large and complex Fortran codes is challenging. AutoRPE was developed to automate this process by instrumenting models with custom precision-aware data types and systematically exploring mixed-precision configurations, enabling efficient performance improvements in computationally intensive simulations without compromising scientific accuracy.”
Key features
- Automatic Code Manipulation: Modifies Fortran source code to implement reduced precision emulation, reducing manual effort.
- Precision Sensitivity Analysis: Identifies sensitive variables requiring higher precision using algorithms like binary search.
- Workflow Management: Automates precision testing workflows to minimize manual intervention and ensure reproducibility.
- Code Coherency Assurance: Detects and resolves inconsistencies in variable types to ensure code correctness.
Models and platforms support
AutoRPE can be installed via pip and integrated into standard Python-based workflows. It has been used with NEMO 4.2.2, NEMO 5, and is currently being applied to the M7 module of OpenIFS-48r1.