Genetic programming for shader simplification

Abstract

We present a framework based on Genetic Programming (GP) for automatically simplifying procedural shaders. Our approach computes a series of increasingly simplified shaders that expose the inherent trade-off between speed and accuracy. Compared to existing automatic methods for pixel shader simplification [Olano et al. 2003;Pellacini 2005], our approach considers a wider space of code transformations and produces faster and more faithful results. We further demonstrate how our cost function can be rapidly evaluated using graphics hardware, which allows tens of thousands of shader variants to be considered during the optimization process. Our approach is also applicable to multi-pass shaders and perceptualbased error metrics.

Genetic Improvement of Programs

“…Sitthi-amorn et al [33] have shown it is feasible for GP to generate many improvements to GPU shaders. (Shaders are small graphics kernel functions which render three dimensional models into two dimensional coloured pictures to display to the user.)…”

“…(Shaders are small graphics kernel functions which render three dimensional models into two dimensional coloured pictures to display to the user.) In [33] they showed GP generating a Pareto front populated by programs which make different tradeoffs between speed and accuracy. Rinard's group at MIT have also considered speed vs. error tradeoffs [32] but only by omitting non-critical parts of the calculation, whereas the Virginia University group's use of GP [33] allows other changes to the human written code.…”

API-Constrained Genetic Improvement

“…It usually does this by searching for a set of edits that are performed on the software system to be improved, such that the desired functional behaviour of the original is retained, while some functional [10,5] and/or non-functional [15,11] aspects are improved. There has been a recent upsurge of activity in this area, with results demonstrating that genetic improvement is able to improve many different properties of systems, including dynamic memory use [20], speed of execution [9,17] and energy consumption [1,14], as well as augmenting and fixing broken functionality [10,5].…”

Babel Pidgin: SBSE Can Grow and Graft Entirely New Functionality into a Real World System

“…This grow and graft development approach aims to reduce the amount of tedious effort required by human programmer in order to develop and add new functionality into an existing system. Our approach is inspired by the recent trend in Search Based Software Engineering (SBSE) called 'genetic improvement' [2,8,10,11,14,15]. Genetic Improvement (GI) uses existing code as 'genetic material' that helps to automatically improve existing software systems.…”

“…Genetic Improvement (GI) uses existing code as 'genetic material' that helps to automatically improve existing software systems. It has been used to repair broken functionality [10,14], and to achieve dramatic scale-ups for sets of small benchmarks [11,15], and also for a 50k LoC genome matching system [8], for graphic shaders [14] and for a CUDA stereo image processing system [7]. Related work on loop perforation has also reported dramatic speed-ups [13].…”