Hardware video encoder vs software encoder

Autonomous Machines Jetson & Embedded Systems Jetson AGX Xavier
#1

On my AGX running JP 4.6.1, I have a series of png files (PNG image data, 1296 x 972, 8-bit grayscale, non-interlaced) that I am trying to create rtsp stream. I tried both hardware and software encoder:

test-launch 'multifilesrc location="%07d_0.png" caps="image/png,framerate=13/1" ! pngdec ! nvvidconv ! video/x-raw(memory:NVMM),format=(string)I420 ! nvv4l2h265enc ! rtph265pay name=pay0 pt=96'


test-launch 'multifilesrc location="%07d_0.png" caps="image/png,framerate=13/1" ! pngdec ! videoconvert ! video/x-raw,format=(string)I420 ! x265enc ! rtph265pay name=pay0 pt=96'

On a test client (Debian 11), I tried to capture the rtsp stream and save back to png files, just to see how they compare to the original png files on the AGX.

gst-launch-1.0 -e rtspsrc location=rtsp://agx:8554/test protocols=tcp ! decodebin ! queue ! videoconvert ! video/x-raw,format=GRAY8 ! pngenc ! multifilesink location="images_%07d.png"

When I open the resulting png files and check them visually to compare to the original png files, I can see :

  1. when using hardware encoder, the resulting images seems much lighter gray. I tried adding property “enable-lossless=true” to nvv4l2h265enc, but that doesn’t make a difference.
  2. when using software encoder, the darkness level seems to match very well.

Anything I can tune the hardware encoder so that I can have same level of darkness, like the software encoder?

#3

Hi,
Not sure but probably the PNG file is encoded from GRAY in full range [0, 255], not limited range [16,235]?

#4

Yes, I think they are in full range. Are you saying only software encoder can properly handle full range?

#5

Following the suggestion from here, I took the nvvidconv source from https://developer.nvidia.com/embedded/l4t/r32_release_v7.1/sources/t186/public_sources.tbz2

sed -i .orig 's/NvBufferColorFormat_YUV420/NvBufferColorFormat_YUV420_ER/g' \
  gstnvvconv.c

rebuilt and replaced libgstnvvidconv.so on the camera. After clearing the gstreamer cache and rerunning the pipeline, I don’t see any improvement.

#6

Hi,
Please try to convert GRAY8 to I420 through videoconvert plugin:

test-launch 'multifilesrc location="%07d_0.png" caps="image/png,framerate=13/1" ! pngdec ! videoconvert ! video/x-raw,format=(string)I420 ! nvvidconv ! video/x-raw(memory:NVMM),format=(string)I420 ! nvv4l2h265enc ! rtph265pay name=pay0 pt=96'
#7

That works! The dark level matches now.
But I also notice the cpu usage on the AGX bumped up from 17% to 30-34%.

#8

Hi,
Seems like the GRAY8 is in full range [0,255]. In nvvidconv plguin, it allocates NvBuffer in NvBufferColorFormat_GRAY8, which is in limited range [16,235].

#9

Is it possible to fix nvvidconv plugin to support the full range? Using the software plugin unnecessarily bump up the cpu usage.

I tried your other suggestion as menteioned here to compile a new nvvidconv plugin, but that didn’t help.

#10

Hi,
We would need to check if we can include this use-case in next release. On the existing releases, please use videoconvert plugin for the conversion.

#12

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.