TMS320C6748: floating point array and signed int array dot product

Rushina Trivedi

Part Number: TMS320C6748

Hello,

I have studied documents related to floating point operation in TMS320C6748. In 2.4.2.3 of SPRU198K, there is a example of dot product of two floating point array for peak performance. But I want to dot product of one floating point array and other one is signed int array. I could align floating point array for 64 bit (double word). For signed int array, I have fixed constant values , so i initialized it globally in the file. Now how can i align the signed int array to 64 bit. Is there any other intrinsic like _dotp2 which does dot product of different types of array (one is floating point and other one if signed int array with fixed value. I want result value in long int format.

In short ,I want peak performance of a function with below requirement in c6748 .

long int function (float *a, signed int *b)

Any suggestions would be sincerely appreciated!

over 7 years ago

0 Yordan Kovachev over 7 years ago

TI__Guru**** 161600 points

Hi,

I've notified the sw team. They will post their feedback directly here.

Best Regards,
Yordan

0 Rushina Trivedi over 7 years ago in reply to Yordan Kovachev

Intellectual 270 points

Thanks Yordan.

Eagerly waiting for the response from sw team.

0 Rahul Prabhu over 7 years ago

TI__Guru** 116020 points

Rushina,

We currently do not provide a function for dot product of mixed data types. However if you check the code for the DSPF_sp_dotprod, it is mostly optimized using the software pipeling using C6000 Compiler and not hand assembly optimized, you should easily be able to modify this to use for your usecase:

I am providing the source for your reference. you can find it in dsplib_c674x_3_4_0_0\packages\ti\dsplib\src\DSPF_sp_dotprod\c674

#pragma CODE_SECTION(DSPF_sp_dotprod, ".text:optimized");

#include "DSPF_sp_dotprod.h"

float DSPF_sp_dotprod(const float * x, const float * y, const int nx)
{
    int i;
    float sum = 0;

    _nassert(nx > 0);
    _nassert(nx % 8 == 0);
    _nassert((int)x % 8 == 0);
    _nassert((int)y % 8 == 0);
    
    for(i = 0; i < nx; i++)    
      sum += x[i]*y[i];
    
    return (sum);
}

Regards,

Rahul

0 Rushina Trivedi over 7 years ago in reply to Rahul Prabhu

Intellectual 270 points

Thanks Rahul.

I have already tried the library function with my parameters of mixed data types, but it is not optimized with software pipelined. I checked it in .asm file. When I use all the parameters as float, it is optimized with software pipelined.

So I believe this the best library routine to optimize the dot product of mixed data types also, Right?

Regards,

Rushina

Processors

Processors forum

TMS320C6748: floating point array and signed int array dot product