6.2.5SCE4 related (inter-layer filtering)
JCTVC-M0088 AHG-13, 17: complexity and performance analysis of different length up-sampling filters in SHM1.0 [E. Alshina, A. Alshin (Samsung)]
This contribution contains performance and complexity analysis of SHM1.0 both IntraBL and RefIdx frameworks with up-sampling filters of different length. The complexity assessment methodology developed for SCEs 3 & 4 by AhG 17 was used. By shorten up-sampling filter to 6 taps luma and 2 taps chroma we can reportedly achieve 3% average memory access reduction with reportedly 0.1% (SS×2) and 0.4% (SS×1.5) luma BD-BR performance drop. There is reportedly 0.3% (SS×2) and 0.4% (SS×15) chroma BD-BR gain from this change. There is reportedly no reduction in the memory bandwidth in the worst case, because the worst case is not when the upsampling filter is applied (assuming this is not done on a whole-picture basis). Average numbers are reported for motion compensation test scenarios. Average memory bandwidth reduction benefit is reportedly not very substantial.
It is noted that we currently have the same filter coefficients for upsampling and motion compensation, which is a nice property. Thus, the contributor suggested not to bother with shortening (or otherwise changing) the upsampling filters.
Dostları ilə paylaş: |
