• State Resolved
  • Locked Locked
  • Replies 3 replies
  • Answers 1 answer
  • Subscribers 98 subscribers
  • Views 203 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.

AM62A7: Config grouping duplication in DSP streaming engine

Iman Fakhari
Prodigy 10 points
Part Number: AM62A7

Hi,

I have two arrays of input data each float32 and 256 bits DSP units I'm trying to vectorize my calculations to do vectors of 8 (32*8).

My first array,

A = [1,2,3,4,5,6,7,8,...]

has a size of 4N, and the second one,

B=[10,20,30,40,...]

has a size of N. I want to configure my DSP streaming engines to multiply each B element into four A elements. So if C is the result it would be like

C=[1*10+2*10+3*10+4*10,5*20+6*20+7*20+8*20, ...].

I know how to do this operation with float4 however, I have 256 bits available and I was wondering if there might be some grouping configuration that can do this operation for me.

Thanks

  • 0
    TI__Expert 4640 points

    Hello Iman,

     

    It is our pleasure to help. 

    Yes, you can use streaming engine grouping duplication. This is an initial code to set the streaming engine for your use case:

     

    //Initialize template to default value
    __SE_TEMPLATE_v1 seTemplate = __gen_SE_TEMPLATE_v1();

    //Specify element type
    seTemplate.ELETYPE = se_eletype<float_vec>::value;

    //Turn on Group Duplication
    seTemplate.GRPDUP = __SE_GRPDUP_ON;

    //Specify group size using VECLEN field.
    seTemplate.VECLEN = __SE_VECLEN_2ELEMS; // This is specific to your use case: duplicated two elements of the B array to fill an entire vector.

    //Specify the type of transfer
    seTemplate.DIMFMT = __SE_DIMFMT_1D;

    seTemplate.ICNT0 = ARRAY_LENGTH;
    .
    .
    // SE open
    .
    .
    //Fetch input data
    Loop for 4xSize of B = 4N

    Please, let me know if this answers your question.

     

    Best regards,

     

    Qutaiba

  • 0 in reply to Qutaiba Saleh
    Prodigy 10 points

    Thanks for your quick reply.

    The config you mentioned will turn B from [10,20,30,..] into [10,20,10,20,10,20,10,20] however I need it to be [10,10,10,10,20,20,20,20]. In fact I need two elements to be duplicated 4 times instead of the group of 2 duplicated 4 times.

    Do you know how I can address this?

    Best,

  • 0 in reply to Iman Fakhari
    TI__Expert 4640 points

    Hi Iman,

     

    Sorry for the confusion. Please use the Element duplication property instead. See this example:

     

    //Initialize template to default value

    __SE_TEMPLATE_v1 seTemplate = __gen_SE_TEMPLATE_v1();

    //Specify element type

    seTemplate.ELETYPE   = se_eletype<float_vec>::value;

    seTemplate.VECLEN    = se_veclen<float_vec>::value;

    //Duplicate one 32-bit element 4 times to fill 256bit vector

    seTemplate.ELEDUP    = __SE_ELEDUP_4X

    //Specify the type of transfer

    seTemplate.DIMFMT = __SE_DIMFMT_1D;

    seTemplate.ICNT0 = ARRAY_LENGTH;

    .

    .

     

    This should work for your use case.

     

    Best regards,

     

    Qutaiba