Bloom Post-Processing Methods on a Mobile Tile-Based GPU: A Comparative Performance and Thermal Study

Abstract

Bloom and glow post-processing effects are perceptually significant components of modern real-time rendering pipelines, yet systematic evaluations of competing bloom implementations on mobile Graphics Processing Unit (GPU) hardware remain scarce. This paper presents a comparative benchmarking study of five bloom techniques: separable Gaussian blur, Kawase dual-filter, mipmap-based bloom, compute shader bloom, and a hybrid Kawase+High Dynamic Range (HDR) pipeline, implemented as an Android Native Development Kit (NDK) application using Open Graphics Library for Embedded Systems version 3.1 (OpenGL ES 3.1). Each technique is evaluated across five synthetic scene variants, with per-pass frame timing measured over 100 captured frames per configuration using Central Processing Unit (CPU)-synchronized timing. The results show that Mipmap-based bloom achieves the lowest mean frame time of 10.29 ms, outperforming Kawase+HDR (11.49 ms) by 10%, Gaussian blur (14.12 ms) by 27%, and the compute shader bloom (21.59 ms) by 52%. Despite leveraging OpenGL ES 3.1 shared-memory tiling, the compute shader implementation was the slowest technique, suggesting that compute dispatch overhead and image2D binding costs outweigh the benefits of parallelism on this architecture. Additionally, thermal monitoring confirmed that no throttling occurred during the benchmark run. These findings provide practical guidance for mobile rendering engineers selecting bloom implementations and introduce a reproducible open-source benchmarking framework for future cross-architecture comparisons.