• State TI Thinks Resolved
  • Locked Locked
  • Replies 4 replies
  • Answers 1 answer
  • Subscribers 33 subscribers
  • Views 381 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.

D3-3P-TDA3X-SK: How does TDA3XX speed up computation on the C66 module

tang hao
Prodigy 80 points
Part Number: D3-3P-TDA3X-SK

I want to scale and move the video on the TDA3X (based on VISION_SDK) development board, which involves matrix and vector operations. It is inefficient to evaluate the data in the system_buffer directly. I wonder if there is any API I can use?
For example, to move in RGB format, we now write it like this:

for(int i=0;i<width*height;i++)
{
    buffer[i].x++;
    bufer[i].y++;
}

  • 0
    TI__Guru 50045 points

    Hi,

    If you check under ~/ti_components/algorithms/ folder in Vision SDK, you can find dsplib_c66x_3_4_0_0 and vlib_c66x_3_3_2_0.

    Those are the optimized DSP kernels TI provides for C66x. You can find their respective document under its /docs/ folder.

    If it is fix-point operation, you can also consider EVE. Please refer to eve_sw_01_20_01_00 folder for details.

    You can refer to TDA3 Technical Reference Menu for other HW accelerator which may help, like VPE for color conversion and resize. Or use EDMA for data transfer to reduce CPU load.

    Regards,

    Stanley

  • 0 in reply to Stanley Liu
    Prodigy 80 points

    The library I used was D3_VISION_SDK_03_02_103_00_TDA3X_RVP, I did not find dsplib_c66X_3_4_0_0 and vlib_c66X_3_3_2_0 in the directory you mentioned. But I did find vlib_c66X_3_3_0_3 and EVe_SW_01_18_00_00. I'm working on them and hope to solve the problem as soon as possible. Thank you for your answer.

  • 0 in reply to Stanley Liu
    Prodigy 80 points

    By the way, According to the document https://www.ti.com/lit/an/sprabg7/sprabg7.pdf， we write a demo to test matrix operation on DSP.

    Now the question is why the output of the "_cmatmpyr1" API under line 56 is always 0. I have tried many times but I cannot solve this error. What is the problem? See the appendix for the code.

    void MatrixMulTest()
{
    Vps_printf("$$$$$$$$$$$$$$$$$$$$ begin MatrixMul $$$$$$$$$$$$$$$$$$$$$$$\n");
    int32_t input_mat1[16] = {1,1,1,2,2,2,3,3,3,4,4,4,5,5,5,6};
    int32_t input_mat2[16] = {6,6,7,7,7,8,8,8,9,9,9,10,10,10,11,11};
    int32_t MatrixOutput[16] = {1};
    
    int64_t *input_vec1 = (long long *) &input_mat1[0];
    int64_t *input_vec2 = (long long *) &input_mat1[4];
    int64_t *input_vec3 = (long long *) &input_mat1[8];
    int64_t *input_vec4 = (long long *) &input_mat1[12];
    int64_t *inputMatPtr1 = (long long *) &input_mat2[0];
    int64_t *inputMatPtr2 = (long long *) &input_mat2[4];
    int64_t *inputMatPtr3 = (long long *) &input_mat2[8];
    int64_t *inputMatPtr4 = (long long *) &input_mat2[12];

    int64_t llinputV1 = _amem8(input_vec1++);
    Vps_printf("_loll(llinputV1) = %d\n",_loll(llinputV1));
    Vps_printf("_hill(llinputV1) = %d\n",_hill(llinputV1));
    int64_t llinputV2 = _amem8(input_vec2++);
    Vps_printf("_loll(llinputV2) = %d\n",_loll(llinputV2));
    Vps_printf("_hill(llinputV2) = %d\n",_hill(llinputV2));
    int64_t llinputV3 = _amem8(input_vec3++);
    Vps_printf("_loll(llinputV3) = %d\n",_loll(llinputV3));
    Vps_printf("_hill(llinputV3) = %d\n",_hill(llinputV3));
    int64_t llinputV4 = _amem8(input_vec4++);
    Vps_printf("_loll(llinputV4) = %d\n",_loll(llinputV4));
    Vps_printf("_hill(llinputV4) = %d\n",_hill(llinputV4));
    int64_t llinputM1 = _amem8(inputMatPtr1++);
    Vps_printf("_loll(llinputM1) = %d\n",_loll(llinputM1));
    Vps_printf("_hill(llinputM1) = %d\n",_hill(llinputM1));
    int64_t llinputM2 = _amem8(inputMatPtr2++);
    Vps_printf("_loll(llinputM2) = %d\n",_loll(llinputM2));
    Vps_printf("_hill(llinputM2) = %d\n",_hill(llinputM2));

    /* _dadds test */
    int64_t daddTest = _dadd2(llinputM1,llinputM2);
    Vps_printf("_loll(daddTest) = %d\n",_loll(daddTest));
    Vps_printf("_hill(daddTest) = %d\n",_hill(daddTest));

    __x128_t inputMat;
#ifdef _LITTLE_ENDIAN
    inputMat = _llto128(llinputM2,llinputM1);
#else
    inputMat = _llto128(llinputM1,llinputM2);
#endif

    /*print __x128_t inputMat*/
    int32_t *testTmp = (int32_t *)&inputMat;        
    Vps_printf("inputMat.word0 = %d\n",*(testTmp++));
    Vps_printf("inputMat.word1 = %d\n",*(testTmp++));
    Vps_printf("inputMat.word2 = %d\n",*(testTmp++));
    Vps_printf("inputMat.word3 = %d\n",*(testTmp));


    int64_t acc1 = _cmatmpyr1(llinputV1,inputMat);
    Vps_printf("_loll(acc1) = %d\n",_loll(acc1));
    Vps_printf("_hill(acc1) = %d\n",_hill(acc1));
    int64_t acc2 = _cmatmpyr1(llinputV2,inputMat);
    Vps_printf("_loll(acc2) = %d\n",_loll(acc2));
    Vps_printf("_hill(acc2) = %d\n",_hill(acc2));
    int64_t acc3 = _cmatmpyr1(llinputV3,inputMat);
    Vps_printf("_loll(acc3) = %d\n",_loll(acc3));
    Vps_printf("_hill(acc3) = %d\n",_hill(acc3));
    int64_t acc4 = _cmatmpyr1(llinputV4,inputMat);
    Vps_printf("_loll(acc4) = %d\n",_loll(acc4));
    Vps_printf("_hill(acc4) = %d\n",_hill(acc4));

    llinputM1 = _amem8(inputMatPtr1++);
    Vps_printf("_loll(llinputM1) = %d\n",_loll(llinputM1));
    Vps_printf("_hill(llinputM1) = %d\n",_hill(llinputM1));
    llinputM2 = _amem8(inputMatPtr2++);
    Vps_printf("_loll(llinputM2) = %d\n",_loll(llinputM2));
    Vps_printf("_hill(llinputM2) = %d\n",_hill(llinputM2));

#ifdef _LITTLE_ENDIAN
    inputMat = _llto128(llinputM2,llinputM1);
#else
    inputMat = _llto128(llinputM1,llinputM2);
#endif
    int64_t acc5 = _cmatmpyr1(llinputV1,inputMat);
    Vps_printf("_loll(acc5) = %d\n",_loll(acc5));
    Vps_printf("_hill(acc5) = %d\n",_hill(acc5));
    int64_t acc6 = _cmatmpyr1(llinputV2,inputMat);
    Vps_printf("_loll(acc6) = %d\n",_loll(acc6));
    Vps_printf("_hill(acc6) = %d\n",_hill(acc6));
    int64_t acc7 = _cmatmpyr1(llinputV3,inputMat);
    Vps_printf("_loll(acc7) = %d\n",_loll(acc7));
    Vps_printf("_hill(acc7) = %d\n",_hill(acc7));
    int64_t acc8 = _cmatmpyr1(llinputV4,inputMat);
    Vps_printf("_loll(acc8) = %d\n",_loll(acc8));
    Vps_printf("_hill(acc8) = %d\n",_hill(acc8));

    llinputV1 = _amem8(input_vec1++);
    Vps_printf("_loll(llinputV1) = %d\n",_loll(llinputV1));
    Vps_printf("_hill(llinputV1) = %d\n",_hill(llinputV1));
    llinputV2 = _amem8(input_vec2++);
    Vps_printf("_loll(llinputV2) = %d\n",_loll(llinputV2));
    Vps_printf("_hill(llinputV2) = %d\n",_hill(llinputV2));
    llinputV3 = _amem8(input_vec3++);
    Vps_printf("_loll(llinputV3) = %d\n",_loll(llinputV3));
    Vps_printf("_hill(llinputV3) = %d\n",_hill(llinputV3));
    llinputV4 = _amem8(input_vec4++);
    Vps_printf("_loll(llinputV4) = %d\n",_loll(llinputV4));
    Vps_printf("_hill(llinputV4) = %d\n",_hill(llinputV4));
    llinputM1 = _amem8(inputMatPtr3++);
    Vps_printf("_loll(llinputM1) = %d\n",_loll(llinputM1));
    Vps_printf("_hill(llinputM1) = %d\n",_hill(llinputM1));
    llinputM2 = _amem8(inputMatPtr4++);
    Vps_printf("_loll(llinputM2) = %d\n",_loll(llinputM2));
    Vps_printf("_hill(llinputM2) = %d\n",_hill(llinputM2));

#ifdef _LITTLE_ENDIAN
    inputMat = _llto128(llinputM2,llinputM1);
#else
    inputMat = _llto128(llinputM1,llinputM2);
#endif    
    acc1 = _dadd2(_cmatmpyr1(llinputV1,inputMat),acc1);
    Vps_printf("_loll(acc1) = %d\n",_loll(acc1));
    Vps_printf("_hill(acc1) = %d\n",_hill(acc1));
    acc2 = _dadd2(_cmatmpyr1(llinputV2,inputMat),acc2);
    Vps_printf("_loll(acc2) = %d\n",_loll(acc2));
    Vps_printf("_hill(acc2) = %d\n",_hill(acc2));
    acc3 = _dadd2(_cmatmpyr1(llinputV3,inputMat),acc3);
    Vps_printf("_loll(acc3) = %d\n",_loll(acc3));
    Vps_printf("_hill(acc3) = %d\n",_hill(acc3));
    acc4 = _dadd2(_cmatmpyr1(llinputV4,inputMat),acc4);
    Vps_printf("_loll(acc4) = %d\n",_loll(acc4));
    Vps_printf("_hill(acc4) = %d\n",_hill(acc4));
    llinputM1 = _amem8(inputMatPtr3++);
    Vps_printf("_loll(llinputM1) = %d\n",_loll(llinputM1));
    Vps_printf("_hill(llinputM1) = %d\n",_hill(llinputM1));
    llinputM2 = _amem8(inputMatPtr4++);
    Vps_printf("_loll(llinputM2) = %d\n",_loll(llinputM2));
    Vps_printf("_hill(llinputM2) = %d\n",_hill(llinputM2));

#ifdef _LITTLE_ENDIAN
    inputMat = _llto128(llinputM2,llinputM1);
#else
    inputMat = _llto128(llinputM1,llinputM2);
#endif  
    acc5 = _dadd2(_cmatmpyr1(llinputV1,inputMat),acc5);
    Vps_printf("_loll(acc5) = %d\n",_loll(acc5));
    Vps_printf("_hill(acc5) = %d\n",_hill(acc5));
    acc6 = _dadd2(_cmatmpyr1(llinputV2,inputMat),acc6);
    Vps_printf("_loll(acc6) = %d\n",_loll(acc6));
    Vps_printf("_hill(acc6) = %d\n",_hill(acc6));
    acc7 = _dadd2(_cmatmpyr1(llinputV3,inputMat),acc7);
    Vps_printf("_loll(acc7) = %d\n",_loll(acc7));
    Vps_printf("_hill(acc7) = %d\n",_hill(acc7));
    acc8 = _dadd2(_cmatmpyr1(llinputV4,inputMat),acc8);
    Vps_printf("_loll(acc8) = %d\n",_loll(acc8));
    Vps_printf("_hill(acc8) = %d\n",_hill(acc8));

    _amem8(&MatrixOutput[0]) = acc1;
    _amem8(&MatrixOutput[2]) = acc5;
    _amem8(&MatrixOutput[4]) = acc2;
    _amem8(&MatrixOutput[6]) = acc6;
    _amem8(&MatrixOutput[8]) = acc3;
    _amem8(&MatrixOutput[10]) = acc7;
    _amem8(&MatrixOutput[12]) = acc4;
    _amem8(&MatrixOutput[14]) = acc8;

    Vps_printf("MatrixOutput[0] = %d\n",MatrixOutput[0]);
    Vps_printf("MatrixOutput[1] = %d\n",MatrixOutput[1]);
    Vps_printf("MatrixOutput[2] = %d\n",MatrixOutput[2]);
    Vps_printf("MatrixOutput[3] = %d\n",MatrixOutput[3]);
    Vps_printf("MatrixOutput[4] = %d\n",MatrixOutput[4]);
    Vps_printf("MatrixOutput[5] = %d\n",MatrixOutput[5]);
    Vps_printf("MatrixOutput[6] = %d\n",MatrixOutput[6]);
    Vps_printf("MatrixOutput[7] = %d\n",MatrixOutput[7]);
    Vps_printf("MatrixOutput[8] = %d\n",MatrixOutput[8]);
    Vps_printf("MatrixOutput[9] = %d\n",MatrixOutput[9]);
    Vps_printf("MatrixOutput[10] = %d\n",MatrixOutput[10]);
    Vps_printf("MatrixOutput[11] = %d\n",MatrixOutput[11]);
    Vps_printf("MatrixOutput[12] = %d\n",MatrixOutput[12]);
    Vps_printf("MatrixOutput[13] = %d\n",MatrixOutput[13]);
    Vps_printf("MatrixOutput[14] = %d\n",MatrixOutput[14]);
    Vps_printf("MatrixOutput[15] = %d\n",MatrixOutput[15]);

    Vps_printf("$$$$$$$$$$$$$$$$$$$$ end MatrixMul $$$$$$$$$$$$$$$$$$$$$$$\n");
}

    Also, does the sprabg7. PDF document mentioned here have anything to do with the previous question?

  • 0 in reply to tang hao
    TI__Guru 50045 points

    Hi,

    You can download Vision SDK at https://www.ti.com/tool/PROCESSOR-SDK-TDAX.

    It includes DSP Lib. vlib_c66X_3_3_2_0 and vlib_c66X_3_3_0_3  are the same component and the number is release version number.

    tang hao said:
    Now the question is why the output of the "_cmatmpyr1" API under line 56 is always 0.

    Please refer to https://www.ti.com/lit/ug/sprui04c/sprui04c.pdf

    long long _cmatmpyr1 (long long src1, __x128_t src2);

    CMATMPYR1 Multiply a 1x2 complex vector by a 2x2 complex matrix, producing two 32-bit complex results