Performance Portable Gradient Computations Using Source Transformation

TL;DR

This paper presents a source transformation approach using Clad to generate portable, efficient gradient computations for Kokkos-based C++ code, enabling effective automatic differentiation across diverse GPU architectures.

Contribution

It introduces modifications to Clad for differentiating Kokkos abstractions, facilitating portable AD for C++ scientific codes on modern GPUs.

Findings

Gradient evaluation took up to 2.17x the input function time on GPUs.

Demonstrated portability of AD across NVIDIA, AMD, and Intel GPUs.

Validated the efficiency of source transformation-based AD for scientific computing.

Abstract

Derivative computation is a key component of optimization, sensitivity analysis, uncertainty quantification, and nonlinear solvers. Automatic differentiation (AD) is a powerful technique for evaluating such derivatives, and in recent years, has been integrated into programming environments such as Jax, PyTorch, and TensorFlow to support derivative computations needed for training of machine learning models, resulting in widespread use of these technologies. The C++ language has become the de facto standard for scientific computing due to numerous factors, yet language complexity has made the adoption of AD technologies for C++ difficult, hampering the incorporation of powerful differentiable programming approaches into C++ scientific simulations. This is exacerbated by the increasing emergence of architectures such as GPUs, which have limited memory capabilities and require massive thread-level concurrency. Portable scientific codes rely on domain specific programming models such as Kokkos making AD for such codes even more complex. In this paper, we will investigate source transformation-based automatic differentiation using Clad to automatically generate portable and efficient gradient computations of Kokkos-based code. We discuss the modifications of Clad required to differentiate Kokkos abstractions. We will illustrate the feasibility of our proposed strategy by comparing the wall-clock time of the generated gradient code with the wall-clock time of the input function on different cutting edge GPU architectures such as NVIDIA H100, AMD MI250x, and Intel Ponte Vecchio GPU. For these three architectures and for the considered example, evaluating up to 10 000 entries of the gradient only took up to 2.17x the wall-clock time of evaluating the input function.