While doing experiments with MaxCUSize set to 32, it appeared that this significantly slows down the decoder (by about 2.5x). Further analysis indicates that a lot of time is spent in the following functions related to adaptive scanning: updateScanOrder() and normalizeScanStats() in TComROM.cpp. When removing calls to these functions, decoding times are similar for maximum CU sizes of 32 and 64. Additionally for MaxCUSize set to 64, decoding times are roughly halved when removing calls to these functions. Adaptive scanning should thus be configurable to determine whether its benefits justify the associated increase in decoding time.

While running TE12 tests it is noticed that SIFO is slower compared to other interpolation filters. Basically increase in speed is because of multiple ME for different offsets in SIFO. It can be speedup up by a very small code change in encoder only. (No syntax changes) After speedup SIFO speed will be same as other interpolation method.

The macro to fix this issue is "SIFO_ENC_SPEEDUP"

When Maximum CU size is set to 8 as follows, encoder/decoder mismatch is caused.

MaxCUWidth : 8 MaxCUHeight : 8 MaxPartitionDepth : 1 QuadtreeTULog2MaxSize : 3 QuadtreeTULog2MinSize : 2

Other information,

• mismatch is caused on all intra coding at least
• mismatch is caused with 0.7.3 and 0.8 at least
• mismatch is caused in the first few rows of chroma on some frames