PROCESSOR-SDK-AM62X: Gstreamer performance difference b/w SDK8 and SDK9

Part Number: PROCESSOR-SDK-AM62X

Tool/software:

Hi,

My customer asked me that gstreamer performace difference between SDK 8 and SDK9.
Here is detailed information about SDK8 and SDK9.

1. SDK 8_06_00_42
2. SDK 9_02_01_10

When they performed following command, the total time of SDK9 was worse than SDK8.
Is it expected result ?

$ gst-launch-1.0 v4l2src device=/dev/video0 ! video/x-raw, width=4192, height=3120, format=UYVY ! multifilesink location=/dev/shm/gst%d.uyvy max-files=8

Best Regards, 

  • Hi Machida-San,

    On AM62x we don’t have any HW accelerator for video processing. 
    when you say performance is worse compared to SDK 8.0. Is ther performance in terms of DDR BW/Framerate/CPU loading?

    Can you share the performance metrics on SDK 8.0 and 9.0 so we can better support your query?

    Best Regards,

    Suren

  • Hello,

    Thank you for your reply.
    >when you say performance is worse compared to SDK 8.0. Is ther performance in terms of DDR BW/Framerate/CPU loading?
    This is from viewpoint of pipeline processing time.

    >Can you share the performance metrics on SDK 8.0 and 9.0 so we can better support your query?
    Here is what I received from customer.

    Case SDK8 : Approx 300ms / 1 image
    Case SDK9 : Approx 700ms / 1 image

    BR,

  • Hi Machida-san,

    Please share the pipeline so we can replicate the setup and share the performance.

    Best Regards,

    Suren

  • Hi,

    Here is pipeline information.

    $ gst-launch-1.0 v4l2src device=/dev/video0 ! video/x-raw, width=4192, height=3120, format=UYVY ! multifilesink location=/dev/shm/gst%d.uyvy max-files=8

    They perform just "input image from camera" and "save multiple image to memory".

    Best Regards,

  • Hi Machida-san,

    I am on travel, so expect a response by end of this week when I will be able to run these tests on the AM62x board and validate and confirm.

    Apologies for the delay. 

    Best Regards,

    Suren

  • Hello,

    I have chance to check pipeline operation time from src to sink by using TI EVM.
    Here is one of my result using gst_tracer.

    0:00:00.715088250 [333m 1040[00m 0x77a64a0 [37mTRACE [00m [00;34m GST_TRACER :0::[00m latency, src-element-id=(string)0x782c1d0, src-element=(string)v4l2src0, src=(string)src, sink-element-id=(string)0x7830400, sink-element=(string)multifilesink0, sink=(string)sink, time=(guint64)110950, ts=(guint64)715035120;

    I got totally approx 50 result, however all result within 200 us.
    Here is command which I used.

    root@am62xx-evm:~# GST_TRACERS="latency" GST_DEBUG=GST_TRACER:7 GST_DEBUG_FILE="/opt/latency.txt" gst-launch-1.0 v4l2src device=/dev/video0 ! video/x-raw, width=1920, height=1080, format=UYVY ! multifilesink location=/dev/shm/gst%d.uyvy max-files=8

    Since I do not have CSI camera which accept 4K image, so I used 2K image. The difference is just this point.
    It seems that customer's result is too long though I have image size difference.

    I will ask how they measure latency, but if you have any comment about my result could you please let me know ?

    BR,



  • Hi Machida-san,

    Yes using the tracers, you would be able to identify latencies pertaining to each element in the pipeline.

    Could you request the customer to try using the tracers and provide their findings.

    Best regards,

    Suren