JVET-AP0053 AhG14: SIMD Improvements of Operators in SADL Library [X. Chen, J. Zhang, Z. Chen (Wuhan Univ.)]

This contribution presents the SIMD improvements of operators in the Small AdHoc Deep Learning (SADL) library from Wuhan University. The optimized layers include Conv2D, Conv2DTranspose, and GridSample. The implementation utilizes the AVX2 instruction set. Experimental results show that the proposed implementation reduces encoding and decoding time.

According to proponents, the modification gives bit-exact results:

Decision(SW): Adopt JVET-AP0053 (all three improvements)