TDA4VEN-Q1: Video Encoding Performance Issue

Daniel Cheng1

Part Number: TDA4VEN-Q1

Hi Expert,

According to the TDA4VEN-Q1 datasheet, the video encoder supports a throughput of up to 400 MPixels/s. Based on this specification, if we utilize four 3MP cameras (totaling 12 megapixels), we expect to achieve a frame rate of at least 30fps ($12 \text{ MP} \times 30 \text{ fps} = 360 \text{ MPixels/s}$, which is less than the 400 MPixels/s limit).

However, during our benchmarking on the EVM (Evaluation Module), we observed the following performance for a 4-channel 3MP setup:

Bitrate 1Mbps: ~24 fps
Bitrate 10Mbps: ~22 fps

There is a significant discrepancy between the theoretical spec and our actual results. Could you please clarify the following:

What is the root cause of this performance drop?
Is there a limitation where the input source cannot exceed 8 megapixels (4K) for a single stream, or are there overheads we are overlooking?

Thanks

Daniel

23 days ago

0 Sarabesh Srinivasan 22 days ago

TI__Mastermind 19750 points

Hello,

Could you share what SDK and HLOS you are on. How are you doing the encoding?

0 John Tu 11 days ago in reply to Sarabesh Srinivasan

Intellectual 280 points

Hi Sarabesh,

1. SDK information:

U-Boot 2024.04-ti-ga970f6e51043 (Nov 13 2024 - 14:26:23 +0000)
Linux version 6.6.44-ti-01478-g541c20281af7-dirty (oe-user@oe-host) (aarch64-oe-linux-gcc (GCC) 13.3.0, GNU ld (GNU Binutils) 2.42.0.20240723) #1 SMP PREEMPT Thu Nov 14 19:20:24 UTC 2024

SDK:j722s-evm-10_01_00_04

2. GStreamer string
appsrc format=GST_FORMAT_TIME is-live=true do-timestamp=true block=false name=myAppSrc0 ! queue
! video/x-raw, width=(int)1920, height=(int)1536, framerate=(fraction)30/1, format=(string)NV12, interlace-mode=(string)progressive, colorimetry=(string)bt601
! v4l2h265enc extra-controls="enc,frame_level_rate_control_enable=1,video_bitrate_mode=1,video_bitrate=10000000,video_gop_size=30"
! h265parse ! video/x-h265, config-interval=0
! appsink name=myAppSink0 max-buffers=5 drop=true sync=false

3. the flow is as the following image.

0 Sarabesh Srinivasan 10 days ago in reply to John Tu

TI__Mastermind 19750 points

Hello,

Is your pipeline based on the app_multi_cam_demo in the Vision-Apps SDK?

Could you share what your appsrc and appsink are doing. Are they camera inputs and display outputs or are you writing/reading files?

Thanks,
Sarabesh S.

0 John Tu 10 days ago in reply to Sarabesh Srinivasan

Intellectual 280 points

Hi Sarabesh,
yes, our pipeline is based on the app_multi_cam_demo in the Vision-Apps SDK.
About camera inputs and display outputs, you can refer our gstreamer string, they are on memory(myAppSrc0 & myAppSink0).
In practice, we use your camera expansion board(Fusion 1) to connect two cameras, and each camera then splits into two image input paths at the LDC node(total 4 ch inputs 1920x1536). You can refer to our flow chart.