How to use CMATMPY intrinsic

Hi Alex,

Please check the document sprugh7.pdf -TMS320C66x DSP CPU and Instruction Set paragraph 4.35 CCMATMPY

Thanks,

HR

Hi HR,

I already did check this document, I began with this one.

The thing is that the description and examples are in assembly code, and I'd like to use the intrinsic in order to implement C code in CCS...

The function takes two numbers for input, but we treat a 1x2 array and a 2x2 matrix. I don't understand why we need the conversion from array/matrix to long long/__x128_t. I also checked the example in C code in the document "Optimizing Loops on the C66x DSP" without fully understanding it... I repeat that I'm really new at this, I'm in an internship where I learn day after day...

Best regards,

Alex

Hi again,

I found how to use the function but it doesn't give me the good results...

Here is my code :

void main(){

    int matrix[2][2] = {-3,7,2,1};
    int array[2] = {-5,6};
    long long arr,arr1,arr2;
    __x128_t matrix_128,output;
    long A8,A9,A10,A11;

    arr=_itoll(array[0],array[1]); // conversion of the array in long long
    arr1=_itoll(matrix[0][0],matrix[0][1]); // conversion of the first line of the input matrix into long long
    arr2=_itoll(matrix[1][0],matrix[1][1]); // conversion of the second line of the input matrix into long long
    matrix_128=_llto128(arr1,arr2); // conversion of the input matrix  in __x128_t

    output=_cmatmpy(arr,matrix_128);

    A8=_get32_128 (output, 0);
    A9=_get32_128 (output, 1);
    A10=_get32_128 (output, 2);
    A11=_get32_128 (output, 3);

    printf("A11 = %ld, A10 = %ld, A9 = %ld, A8 = %ld",A11,A10,A9,A8);
}

In the console I get :
A11 = -26, A10 = 8, A9 = 29, A8 = -7    instead of 27, 0, -29, 0

Can someone tell me what's wrong ?

Thanks,

Alex

Alexandre,

It is a complex multiply. With a complex being represented as a 32-bit entity (two 16-bit numbers packed)

In your code I don't see any reference to complex number nor 16-bit numbers packed in a 32-bit one.

Are you trying to do

[A B] x [ C D ] = [ G H ]
            [ E  F ]

with A = -5, B = 6, C = -3, D = 7, E = 2, F = 1 ?
in that case G = 27 H = -29

The code doesn't seem wrong per se, it's just that your understanding of what CMATMPY does is confused.

Hi Clement,

Thank you for your answer.

That's exactly what I want to do, I try the function with some easy input values for understanding it... It doesn't matter if I don't use complex values, right ?

But the output doesn't seem correct. Instead of 27, 0, -29, 0 (for real and imaginary parts) I get -26, 8, 29, -7.

Do you know why ?

Thank you, I think I have understood indeed.

I have one last question though : what do we have in memory for negative numbers, when I do int A = -5 ?

I thought we had something like 00000000 00000000 ; 10000000 00000101 but when I do the computation array/matrix myself, I do not get the values that CCS give me.

Assuming you are in a debug configuration,

use the memory browser view in CCS and/or the variable view too. You can then see your 'int' in memory.

It's a good idea to debug step by step and check in memory how things are stored (when you use itoll for example).

CM

Ok so I saw that we use the 2s compliment to represent a negative number.

For example, -5 is 11111111 11111111 ; 11111111 11111011 in binary.

But when I use the intrinsic, the compiler doesn't understand it as 65535 + i*65531 right ?

Well I don't know

you have to try something like that :

short re, im; //16-bit
int complex; //32-bit

re = -5;
im = 0;

complex = _thegoodintrinsic(re,im);

and see what happens in memory.

or you can use a struct too

pseudo code

typedef struct complex{

short re;
short im; } ;

complex myComplex;

myComplex.re = -5;
myComplex.im = 0;

Hi,

I'll try with your structure today. I saw that you tried to make me understand alone but it's really new for me.

Anyway thanks a lot for your help, I'll come back if necessary but I don't hope so... !

Alexandre

Hi,

I have another problem with the function. Below is the code I wrote for multiplying a 1x2 vector by a 2x2 matrix when all elements are of type int :

typedef struct { 

int re; //32 bit

int im; // 32 bit  

} complex;

complex arr1, arr2, mat1, mat2, mat3, mat4; 

arr1.re = -5;   arr1.im = 0; arr2.re = 6; arr2.im = 0;

 mat1.re = -3; mat1.im = 0; mat2.re = 7; mat2.im = 0; mat3.re = 2; mat3.im = 0; mat4.re = 1; mat4.im = 0;

 int array_1, array_2, matrix_1, matrix_2, matrix_3, matrix_4;

array_1 = _spack2(arr1.re,arr1.im); 

array_2 = _spack2(arr2.re,arr2.im);

matrix_1 = _spack2(mat1.re,mat1.im);

matrix_2 = _spack2(mat2.re,mat2.im); 

matrix_3 = _spack2(mat3.re,mat3.im);

matrix_4 = _spack2(mat4.re,mat4.im);

long long array, line1_matrix, line2_matrix;

array = _itoll(array_1,array_2);  line1_matrix = _itoll(matrix_1,matrix_2);  line2_matrix = _itoll(matrix_3,matrix_4);

__x128_t matrix, product;

matrix = _llto128(line1_matrix,line2_matrix);  product = _cmatmpy(array,matrix);

 long A8,A9,A10,A11;

 A8 = _get32_128(product,0);  A9 = _get32_128(product,1);  A10 = _get32_128(product,2);  A11 = _get32_128(product,3);

 printf("A11 = %ld, A10 = %ld, A9 = %ld, A8 = %ld",A11,A10,A9,A8);

This code works correctly for matrix of type int. I'd like to use the intrinsic for matrix of float. I tried to do things like that but I didn't succeed because of the types of the arguments.

Can you help me to adapt this ? I think that there shouldn't be a lot to modify...

Thanks,

Alex

Hmm I'm wondering if that's possible.

float numbers are coded on 32 bits, and all are used...

With integers it was easier since I just took the 16 LSB of each integer to form the long long.

I can't deprive myself by considering only some bits of a float number... Otherwise the number will not be the same.

But I don't think TI has forgotten the possibility to treat float numbers matrix...

So how can I do this ?

I'm upping the topic...

I recall my problem : I'd like to use the complex matrix multiply when my matrixes are composed by float numbers. I succeedeed for int numbers, but for floats, I really don't see the conversions to do for getting a correct input.

Can someone help me please ?

Hi Alberto,

Thank you for your answer !

That's what I thought... I could have looked for the solution for a long time !

Problem solved !

Bye