In the past months I have done a lot of benchmarking of dav1d, the AV1 video decoder developed by VideoLAN. Right now most functions are accelerated with hand-writen assembly code, SSSE3, AVX2 but also. I think that there are great marketing oppertunities for both Arm and dav1d, showing of how newer processors improve performance and lower power usage, and how NEON enables performance gains over 100%.

In my past testing: https://medium.com/@ewoutterhoeven/how-arms-neon-enables-efficient-av1-decoding-on-mobile-5fcb3a4f6e7f

More comparisons could be done, showing how more cores can not only increase performance but also lower power consumption. The same is probably through for newer cores. Furthermore comparisons between different compilers with different levels auto-vectorizations and the hand-written NEON-assembly could be done.

Due to the low power usage, low die area and therefore high core counts I believe Arm could be very competative in the video market. Results of tests would be shared and can be used both for marketing and research purposes.

## Hardware ##

Arm64 in-order cores
 - Cortex-A53
 - Cortex-A55

Arm64 out-of-order cores
 - Cortex-A72
 - Cortex-A73
 - Cortex-A75
 - Cortex-A76
 - Cortex-A77

## Tests ##

1. Max. single-threaded performance
dav1d's average frame rate with a single tile and frame thread.

KPI: Frames per second

2. Max. multi-threaded performance
dav1d's average frame rate with multiple tile and frame threads.

KPI: Frames per second

3. NEON speedup
The speed up from optimized C code to optimized C code with NEON assembly enabled.

KPI: Frames per second

4. Power consumption at fixed frame rate:
dav1d will be run at a fixed frame rate. Board power will be measured on the power adapter or over USB.

KPI: Watt (average)

5. Power consumption with different thread counts
dav1d will be run at a fixed frame rate

KPI: Watt (average)

6. Compilers and autovecotrization performance
dav1d will be run at a fixed frame rate compiled with different compilers (at least GCC and Clang) and different levels of autovectorization, and with and without the NEON assembly enabled.

KPI: Frames per second

7. Compilers and autovecotrization power
dav1d will be run at a fixed frame rate compiled with different compilers (at least GCC and Clang) and different levels of autovectorization, and with and without the NEON assembly enabled.

KPI: Watt (average)