benchmark multi-frame textured depth

[bmegli]

Especially check if LattePanda Alpha can handle:

• the new methodology with double HEVC (HEVC Main10 + HEVC Main)
• and forced highest quality effort in VAAPI

Related to HVS#5

[bmegli]

Scenario:

• testing 435eb97a7c2a3b48d9b4ef3e11468e05c01de708 and NHVE bench-multi-frame-time
• with LattePanda Alpha (KabyLake m3-7y30) as sender
• with i7-7820HK laptop as sender
• the laptop is also used to check if result is ok (separate test)
• both on Ubuntu 18.04

848x480@30 fps, 8Mb + 1 Mb bitrate

# on LattePanda Alpha
./realsense-nhve-depth-ir 192.168.0.100 9768 848 480 30 500 /dev/dri/renderD128 8000000 1000000 0.00003125

LattePanda Alpha:

• GPU usage reaches 65% (intel_gpu_top)
• CPU usage reaches 35% of 400% (htop)
• encode/send takes 25-27 ms

Laptop:

• GPU usage reaches 50%
• CPU usage reaches 21% of 800%
• encode/send takes 20-25 ms

LattePanda Alpha again

By using default VAAPI setting (not forcing highest quality)

• GPU usage reaches 35%
• CPU no changes (< 35% of 400%)
• encode/send takes 13-17 ms

By forcing highest quality for depth and using default for IR:

• GPU usage reaches 50%
• CPU no changes (<35% of 400%)
• encode/send takes 20-23 ms

[bmegli]

With highest quality forced in VAAPI LattePanda Alpha is too close to not keeping up with the data for my liking.

Switched back to default quality for texture encoding:

• this brings back GPU usage to around 50%
• and encoding time to 20-23 ms
• which at 30 fps (33.33 ms) leaves safety margin

For depth keep forcing highest quality.

[bmegli]

Comparing to results with #7

• the new methodology will add a few ms to glass-to-glass delay
  • 5-6 ms with default quality
  • non-default quality is not comparable (would need more tests)

Between implementations:

• previous (UV encoded infrared)
• current (Main10 + Main), depth with forced quality
• we add around 10-14 ms delay on encoder side.