• State Resolved
  • Locked Locked
  • Replies 6 replies
  • Answers 1 answer
  • Subscribers 98 subscribers
  • Views 845 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.

Question about TDA4 VPAC MSC pyramid.

jacky.wong
Intellectual 340 points

hi,

    When I use the tivxVpacMscPyramidNode to implement pyramid, the result of the border is not correctly,

    The input data is a 32 x 32 image as blow

line[0] 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
line[1] 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
line[2] 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2,
line[3] 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3,
line[4] 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4,
line[5] 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
line[6] 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6,
line[7] 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7,
line[8] 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8,
line[9] 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9,
line[10] 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10,
line[11] 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11,
line[12] 12, 12, 12, 12, 12, 12, 12, 12, 12, 12, 12, 12, 12, 12, 12, 12, 12, 12, 12, 12, 12, 12, 12, 12, 12, 12, 12, 12, 12, 12, 12, 12,
line[13] 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13,
line[14] 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14,
line[15] 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15, 15,
line[16] 16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16,
line[17] 17, 17, 17, 17, 17, 17, 17, 17, 17, 17, 17, 17, 17, 17, 17, 17, 17, 17, 17, 17, 17, 17, 17, 17, 17, 17, 17, 17, 17, 17, 17, 17,
line[18] 18, 18, 18, 18, 18, 18, 18, 18, 18, 18, 18, 18, 18, 18, 18, 18, 18, 18, 18, 18, 18, 18, 18, 18, 18, 18, 18, 18, 18, 18, 18, 18,
line[19] 19, 19, 19, 19, 19, 19, 19, 19, 19, 19, 19, 19, 19, 19, 19, 19, 19, 19, 19, 19, 19, 19, 19, 19, 19, 19, 19, 19, 19, 19, 19, 19,
line[20] 20, 20, 20, 20, 20, 20, 20, 20, 20, 20, 20, 20, 20, 20, 20, 20, 20, 20, 20, 20, 20, 20, 20, 20, 20, 20, 20, 20, 20, 20, 20, 20,
line[21] 21, 21, 21, 21, 21, 21, 21, 21, 21, 21, 21, 21, 21, 21, 21, 21, 21, 21, 21, 21, 21, 21, 21, 21, 21, 21, 21, 21, 21, 21, 21, 21,
line[22] 22, 22, 22, 22, 22, 22, 22, 22, 22, 22, 22, 22, 22, 22, 22, 22, 22, 22, 22, 22, 22, 22, 22, 22, 22, 22, 22, 22, 22, 22, 22, 22,
line[23] 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23, 23,
line[24] 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24,
line[25] 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25,
line[26] 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26,
line[27] 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27,
line[28] 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28,
line[29] 29, 29, 29, 29, 29, 29, 29, 29, 29, 29, 29, 29, 29, 29, 29, 29, 29, 29, 29, 29, 29, 29, 29, 29, 29, 29, 29, 29, 29, 29, 29, 29,
line[30] 30, 30, 30, 30, 30, 30, 30, 30, 30, 30, 30, 30, 30, 30, 30, 30, 30, 30, 30, 30, 30, 30, 30, 30, 30, 30, 30, 30, 30, 30, 30, 30,
line[31] 31, 31, 31, 31, 31, 31, 31, 31, 31, 31, 31, 31, 31, 31, 31, 31, 31, 31, 31, 31, 31, 31, 31, 31, 31, 31, 31, 31, 31, 31, 31, 31,

The level 1 output data is:

line[0] 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
line[1] 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2,
line[2] 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4,
line[3] 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6,
line[4] 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8,
line[5] 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10,
line[6] 12, 12, 12, 12, 12, 12, 12, 12, 12, 12, 12, 12, 12, 12, 12, 12,
line[7] 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14,
line[8] 16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16,
line[9] 18, 18, 18, 18, 18, 18, 18, 18, 18, 18, 18, 18, 18, 18, 18, 18,
line[10] 20, 20, 20, 20, 20, 20, 20, 20, 20, 20, 20, 20, 20, 20, 20, 20,
line[11] 22, 22, 22, 22, 22, 22, 22, 22, 22, 22, 22, 22, 22, 22, 22, 22,
line[12] 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24,
line[13] 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26, 26,
line[14] 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28,
line[15] 29, 29, 29, 29, 29, 29, 29, 29, 29, 29, 29, 29, 29, 29, 29, 29,

The expected reslut of the last line should be 30, but the output is 29, can you help me debug this issue?

thanks

  • 0
    TI__Expert 5565 points

    Can you please tell the SDK version number, and if you are generating these results on the EVM, or using the PC emulation mode?

    Regards,

    Jesse

  • 0 in reply to Jesse Villarreal
    Intellectual 340 points

    Hi,

        SDK vision number: psdk_rtos_auto_j7_06_01_00_15

        I using the TDA4 EVM generating these results. The node function is tivxVpacMscPyramidNode.

    thank you

  • 0 in reply to jacky.wong
    TI__Expert 5565 points

    I have been able to reproduce your results, and I think I know why this is happening.  I have initiated discussions with the experts to see how to fix it. I just wanted to let you know I am looking into it and give you an update.  I will let you know when I find a solution.

    Regards,

    Jesse

  • 0 in reply to Jesse Villarreal
    Intellectual 340 points

    thank you for your support.

  • 0 in reply to jacky.wong
    TI__Expert 5565 points

    Here is the explanation for what is happening, I hope this helps.

    1. The tivxVpacMscPyramidNode function is programmatically internally using a 5 tap Gaussian filter coefficients before doing a half scale.  The coefficients are as follows:

    16, 64, 96, 64, 16

    These coefficients add to 256, so the down shift value chosen is 8 (fixed point representation of each of the coefficients)

    2. The hardware is using replicated border mode.  This means that if the input pixel it needs for filtering does not exist in the image (because it goes past the border), then it replicates the closest border pixel for the calculation.

    3. For 8 bit input/output mode, the hardware upshifts the input pixels to the native bitdepth of the hardware (12 bits), does the processing and rounding of output in 12 bits, and then finally downshifts the output back to 8 bits (dropping the lower 4 bits).

    So given this information, we can find out what happens on the last output row (in your case):

    Step 1: Upshift each input pixel needed in the 5 tap filter (the 5 rows surrounding the row with the 30; 2 before, and 2 after.  Since there is only 1 after, then the last one is replicated).  This means it needs the following pixels : 28, 29, 30, 31, 31.  Upshift all these by 4 bits, giving 448, 464, 480, 496, 496.

    Step 2: Multiply these pixels by the respective filter coefficients, and sum the results: 

    (448*16)+(464*64)+(480*96)+(496*64)+(496*16) = 122624

    Step 3: Add (1 << shift-1) to the sum, and downshift by shift value (this is do do a rounding divide by the sum of the coefficients, essentially) = (122624+128)>>8 = 479.  This value would be the final answer if we wanted to output 12 bits.  As you see, it is slightly less than the center pixel of 480 (12 bits), due to the replicated last pixel brings down the convolution sum slightly.

    Step 4: However, since we want 8 bits, we simply downshift by 4 bits = 479>>4=29.

    You can see that a floating point divide would be as follows 479/16 = 29.9375, which would have rounded to 30, but since this 8 bit mode is simply downshifting the lower 4 bits, it essentially truncates instead of rounds.

    So this isn't a border issue.  The truncation for 8 bit output mode just manifested itself in the border due to the monotonically increasing values of each row in your test being broken by the border replicate, slightly dropping the filter output down enough for the truncate to go down almost 1 value.

    Please note that if you used 12 bit output then this would not happen since the round happens in the 12 bit precision.  For example, if you wanted to send pyramid results to the DMPAC DOF, it can accept the 12 bit output from this pyramid calculation.

    Also, if you used nearest neighbor configuration for coefficients instead of Gaussian (not added in node yet, but would be easy to add if needed), then there would be no math involved, and no reason to round/truncate.

    Best Regards,

    Jesse

  • 0 in reply to Jesse Villarreal
    Intellectual 340 points

    OK，thank you very much.