JVET-AO0131 AHG3/7: Reference software extension for instruction counting [J. Sauer, Y. Zhao, Y. Sun, E. Alshina (Huawei)]

This document reports findings on instruction counting using special CPU hardware registers. Similar results can be obtained as shown in JVET-AN0119 or JVET-AN0157, but without a runtime performance penalty. It is shown, that instruction count is more stable than measuring runtime. The contribution proposes that reporting instruction count should be allowed as an alternative to reporting runtimes or even to be used instead. Other hardware counters that may also be of interest to AHG19 are cache-misses and branch-misses.

It was claimed that one advantage compared to runtime measurement might be that the instruction count is independent of homogeneity of cluster, SIMD availability, etc.

When using virtual machines, it may not be possible to access the hardware registers.

It was also commented that admin privileges might be necessary.

Some concern was raised to completely replace runtime by instruction count, or make the choice optional. It might be useful as additional information.

It was asked if comparability would still be given when different compilers (even compiler updates) or different degree of compiler optimization was used. SIMD would also be counted as only one instruction.

It was commented that for completely different algorithms, this would not be a good basis for comparison.

Further study was recommended – using additional data might be useful in the context of future standard development.

Traditional tools for profiling can deliver the same (or even more detailed) information, but those would slow down the runtime.

AHG8: Optimization of encoders and receiving systems for machine analysis of coded video content (1)

Contributions in this area were discussed during 2300–2315 UTC on Thursday 22 Jan. 2026 (chaired by JRO).