Back to Search Document details
41st Meeting: by teleconference, CH, January 2026 2026-01-14 02:51
AHG14: Decoding time reduction in SADL using SIMD acceleration
Authors: Jiang Han Xidian Univ. Cheolkon Jung Xidian Univ. Qipu Qin Xidian Univ.
Abstract
This contribution presents methods for accelerating several operators in the Small Adhoc Deep Learning Library (SADL) using SIMD acceleration. Experimental results demonstrate that the proposed operators significantly reduce decoding time in SADL.
Meeting Notes
JVET-AO0073 AHG14: Decoding time reduction in SADL using SIMD acceleration [J. Han, C. Jung, Q. Qin (Xidian Univ.)]

This contribution presents methods for accelerating several operators in the Small Adhoc Deep Learning Library (SADL) using SIMD acceleration. Experimental results demonstrate that the proposed operators significantly reduce decoding time in SADL.

Additionally, an alignment of negative clipping range (from -32767 to -32768) is suggested.

It was commented that separate merge requests would be necessary for each element. The clipping issue is also raised in JVET-AO0217.

It was commented that the proposed accelerations for convolutions 1x1 and 3x3 might only work if the number of channels is a power of two.

This was agreed in principle, to be further clarified and left to the discretion of the SADL coordinator which of the elements could be harmonized with speedups proposed in other contributions. See also further notes under JVET-AO0217.

ARCHIVE :
EE1.2.4_acc versus EE1.2.4.xlsm JVET-AO0073-v3.docx JVET-AO0073_slides_final.pdf LOP6_acc versus LOP6.xlsm LOP6_acc versus LOP6_original_saturation_range.xlsm VLOP4_acc versus VLOP4.xlsm
Decisions
This was agreed in principle, to be further clarified and left to the discretion of the SADL coordinator which of the elements could be harmonized with speedups proposed in other contributions. See also further notes under JVET-AO0217.
Citation