JVET-AP0219 [AHG11] Additional Information on Stable Float Method for Neural Network Inference on GPU devices [H. Zhang, F. Cricri, N. Le, N. Zou, A. B. Koyunku, L. Murn (Nokia)]

A mechanism known as "stable float" was introduced in JVET-AO0178 where it was reported to facilitate inference on GPU devices for neural network-based coding tools using a floating-point data format. The stable float approach is reported to enable cross-device reproducibility across various computing environments by using integer components within convolution operations while utilizing floating-point numbers for other operations. It is reported that several concerns regarding the stable float method and IEEE 754 standard were raised during the JVET-AO meeting review of the proposal. This document is reported to address these concerns and to provide the proof of the overflow-free property of the proposed method.

V2: editorial changes and more information on IEEE 754 standards on rounding operations.

Just for information - no need for presentation according to proponents