• State Resolved
  • Locked Locked
  • Replies 5 replies
  • Answers 2 answers
  • Subscribers 33 subscribers
  • Views 331 views
  • Users 0 members are here
Support feedback
Options
Options
Related

This thread has been locked.

If you have a related question, please click the "Ask a related question" button in the top right corner. The newly created question will be automatically linked to this question.

AWR2944: What's the MSS processing time?

Jesse Wang
Intellectual 2345 points
Part Number: AWR2944


Tool/software:

Hi TI expert,

 My customer has two questions regarding the AWR2944 processing time below:

   1). When does MSS complete its task? According to the picture below, the MSS runs RANSAC, GTrack, and data output from time C in the Nth frame. Should the MSS complete the Nth-frame's task before A, B, and C in the (N+1)th frame? Please mark the timing if it's not A, B, or C.

   

   2). In column B as below picture, the number of point clouds for each frame with the corresponding frame period. We found that the InterFrameProcessingTime is above 30 ms even though there is no point cloud generated. We need to ensure this result also matches TI’s expectations. If so, do you have any suggestions for speeding up the processing or any improvements we can make without compromising performance?

     

  • +1
    TI__Mastermind 22280 points

    Hi Jesse,

    Jesse Wang said:
    When does MSS complete its task? According to the picture below, the MSS runs RANSAC, GTrack, and data output from time C in the Nth frame. Should the MSS complete the Nth-frame's task before A, B, and C in the (N+1)th frame? Please mark the timing if it's not A, B, or C.

    Mostly object data transfer of frame N ends before the start of Doppler processing of frame N+1 i.e before point A.

    Jesse Wang said:
    In column B as below picture, the number of point clouds for each frame with the corresponding frame period. We found that the InterFrameProcessingTime is above 30 ms even though there is no point cloud generated. We need to ensure this result also matches TI’s expectations. If so, do you have any suggestions for speeding up the processing or any improvements we can make without compromising performance?

    The number of detected objects can be known at least after the Doppler processing. So, InterFrameProcessingTime cannot be reduced even if the number of detected objects is 0.

    Regards,

    Samhitha

  • 0 in reply to Samhitha Bobba
    TI__Intellectual 2345 points

    Hi Samhitha,

      Thanks for your answer which solved our questions.

      And a further related question:

      - If we add more features to MSS that take longer processing time, what will happen if MSS processing time exceeds point A? For example, will the system crash?

    Sincerely, Jesse

  • 0 in reply to Jesse Wang
    TI__Mastermind 22280 points

    Hi Jesse,

    Jesse Wang said:
    If we add more features to MSS that take longer processing time, what will happen if MSS processing time exceeds point A? For example, will the system crash?

    You can transmit the data till the next frame's processed data is available. So, the processing on MSS for frame N should end before point C of next frame.

    Note: Ensure that you transmit any required results before the next frame's data is overwritten. For example, if you are transmitting detection matrix over UART, make sure that the detMatrix of frame N is transmitted before the doppler processing of frame N+1.

    Regards,

    Samhitha

  • 0 in reply to Samhitha Bobba
    Prodigy 10 points

    Hi TI expert, May I confirm my understanding is correct or not?

    1. After N+1 frame starts chirping, the rangeproc DPU is automatically triggered in the HWA. However, due to their different processing units, RANSAC and Gtrack can still be processed in parallel on MSS. 

    In our cases, the point clouds are generated at stamp C(processed in DSS). The data will be copied into MSS for further tracking processing.
    We need to ensure the Nth frame tracking processing can be completed before N+1 stamp C. 
    Otherwise, the point cloud will be overwritten. Am I correct?
    2. Regarding the note from TI, depending on the application(or verification purpose), it is possible to copy other information to MSS and transmit it through UART.
    For instance, TI lists an example of transmitting a detection matrix through UART.
    The detection matrix is a range-doppler image processed in doppler proc-DDMA DPU(stamp A ~ stamp B).
    Following the same concept, these data will be copied into MSS for further UART transmission if we want.
    We need to ensure the transmission can be completed before the Doppler processing of frame N+1, stamp A. Otherwise, the detection matrix will be overwritten when data is transmitted. Am I correct?
    3. Thus, I am considering 
    1. The tracking process needs to be completed before the next frame of stamp C.
    2. The detection matrix needs to complete transmission before the next frame of stamp A.
    3. The first FFT(rangeproc DPU) needs to complete transmission before next frame chirping.
      Am I correct?

    4. If the tracking process takes longer, for instance, longer than the N+1 frame of stamp C, expect data to be overwritten. Will the system hang, either?

  • 0 in reply to Milton Wu
    TI__Mastermind 22280 points

    Hi Milton,

    Milton Wu said:
    We need to ensure the Nth frame tracking processing can be completed before N+1 stamp C. 
    Otherwise, the point cloud will be overwritten. Am I correct?

    Tracking algorithm of Nth frame should end before the Range CFAR processing of N+1 frame ends. Till this point you can use the point cloud stored on DSS_L3 as the input for the tracking algorithm. If you have a copy of point cloud stored on MSS_L2 (copy from DSS_L3) then tracking algo execution should be completed before N+1 stamp C.

    Milton Wu said:
    We need to ensure the transmission can be completed before the Doppler processing of frame N+1, stamp A. Otherwise, the detection matrix will be overwritten when data is transmitted. Am I correct?

    Yes, you correct. I have given this example as we have an option to transmit detection matrix in the SDK demo. You don't need to store the detMatrix again in MSS_L2 if you are transmitting before doppler processing on frame N+1. If you are storing a copy of the data on MSS_L2 also, then you need to check at which step will you replace it with the data of next frame.

    For the 3rd point, your understanding is correct. But in the SDK demo, we start to transmit the data over UART after tracking, so please ensure you order each task accordingly. If you are storing a copy of all the data like point cloud, detMatrix in MSS_L2, you can transmit all of them before stamp C. But this will utilize additional memory. Regarding 1D FFT or detMatrix, transmitting over UART will take a longer time. You can also do a SW streaming over LVDS lanes if you are transmitting huge amount of data for debug (instead of transmitting over UART). 

    Milton Wu said:
    If the tracking process takes longer, for instance, longer than the N+1 frame of stamp C, expect data to be overwritten. Will the system hang, either?

    This will lead to an assert as the next frame's processed data is available for tracking and transmission over UART. DSS sends a message to MSS after processing on DSS is completed for a particular frame. By this point of time, processing of previous frame data on MSS should end.

    Regards,

    Samhitha