• State TI Thinks Resolved
  • Locked Locked
  • Replies 8 replies
  • Answers 2 answers
  • Subscribers 99 subscribers
  • Views 560 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.

TDA2: EVE: Buffer transfer between ARP32 kernel and VCP kernel.

Shruti Joshi
Intellectual 400 points
Part Number: TDA2

Hi,

I have an application with three kernels, out of them kernel1 & kernel3 are in the kernel-C running on EVE_VCOP and kernel2 which is fully in natural-C running on EVE_ARP32.  The execution graph is as follows:

Input -> kernel1 -> kernel2 ->kernel3 -> Output.

While executing kernel2, 'BAM_ARP32_computeWrapper()' function is called, wherein the buffers are used in ping pong format. So, the input and output buffer addresses are switched between IBUFHB, IBUFLB and IBUFHA, IBUFLA. The output of this kernel is passed as an input to the kernel3(). But, when I debug I see that kernel3() input address does not change according to the output. BAM has the responsibility to pass these buffer addresses but seems like it retains the addresses it got during kernel3_init() function. 

Is there any way to pass this addresses correctly at runtime? 

Also, I see that the input buffer address also toggles, but the output address of kernel1 is constant. So, will kernel2 take correct data in the alternate iterations? 

Regards,

Shruti

  • 0
    TI__Expert 6335 points

    Hi Shruti,

    For kernels running on EVE_VCOP, such as kernel 1 and kernel 3, it is normal that the addresses don't change. VCOP has a special memory mapping capability called ALIAS view, in which the addresses just need to point to image buffer A. The actual physical location is eventually  determined by the buffer switch. Actually the upper bits are masked out, so what counts in the offset from the base address from image buffer A.

    For EVE_ARP32, ALIAS view is not supported so the processor needs to pass the correct addresses when calling kernel 2. 

    So to me, the behaviour you are describing me is normal.

    regards,

    Victor

  • 0 in reply to Victor Cheng
    Intellectual 400 points

    Hi Victor,

    I think my problem is different.

    I have three Kernels running sequentially one after other.
    Kernel_1 (VCOP)-> O/P Buf1 ->Kernel_2 (ARP32) -> O/P Buf2 -> Kernel_3 (VCOP)-> O/P For Display

    Current Scenario:
    - Kernel_3 (VCOP) while initialization time has input as O/P Buf2 address (say Address 1).
    - ARP32 works in PING-PONG style. So for each block processing it toggles it's O/P buffer O/P Buf2(Say Address1 and Address2).

    Now my problem is
    - As Kernel_3 holds only single address (say Address 1) as input and it can not change to Address 2 at run time.
    So how to change run time input addresses for Kernel_3 to get proper inputs?

  • 0 in reply to Shruti Joshi
    TI__Expert 6335 points

    Hi Shruti,

    Do you mean that all your kernels are called within one BAM node ? 

    You will need to split into 3 BAM nodes. Node 1 executes kernel 1, Node 2 executes kernel 2 and Node 3 executes kernel 3.

    regards,

    Victor

  • 0 in reply to Victor Cheng
    Intellectual 400 points
    Hi Victor,

    No, they are not present in a single node. They are in 3 different nodes.

    Regards,
    Shruti
  • 0 in reply to Shruti Joshi
    TI__Expert 6335 points

    Hi Shruti,

    I am affraid I don't understand the issue. Since kernel 3 is in a separate node, that executes on VCOP only, you don't need to change the address of the input since the buffer switch automatically takes care of which image buffer it accesses to, thanks to alias view.

    regards,

    Victor

  • 0 in reply to Victor Cheng
    TI__Guru 56820 points

    Shruti,

       Can you respond to Victor's queries?

    Regards,

    Anshu

  • 0 in reply to Anshu Jain
    TI__Guru 56820 points
    Shruti,
    Can you update us on this thread?

    Regards,
    anshu
  • 0 in reply to Anshu Jain
    TI__Guru 56820 points

    Shruti,

      I am closing this thread as we haven't heard back from you. Kindly re-open this thread if you have further questions.


    Regards,

    Anshu