how to enable and use VFPU and SPE2-SIMD for MPC57xx in S32DS project

[enwei]

Hi,

As we know, in most of our MPC57xx parts, such as MPC5775K as below, there are VFPU and SPE2-SIMD, which are very useful when doing DSP and complex math calculation.but how can we enable and use it in a S32DS project?

There is no options for this when create the project, and there are some related options in project properties-->C/C++ Bulid-->Stettings-->Target Processor, I checked Use hardware floating point, and configured Signal Processing: generate SPE v2 SIMD instructions and use ewl_c9x library:

However, after configurations as above, I find the disassemle code has nothing differerent,  no performance improve, no SPE instructions generated for my codes, it seems still call the ewl_c9x library to complete my math operations:

void calc_magnitude_doppler( fft_value_t* inAddr, magnitude_t* outAddr, bool_t bRCSComp, eDistRange distRange){

  fft_value_t real, img;

  uint64_t quad;

  double sqroot;

  uint16_t x, y;

  for(y=0; y<FFT_NUM_SAMPLES; y++){

   // Do it for one chirp

   for(x=0;x<FFT_NUM_CHIRPS;x++){

  // Pointer to values

  real = *(inAddr + (x * FFT_NUM_SAMPLES * 2));

  img = *(inAddr + (x * FFT_NUM_SAMPLES * 2) + 1);

  // Calculate magnitude

  quad = (uint64_t)((int64_t)real*real + (int64_t)img*img);

  sqroot = sqrt((double)quad);

// sqroot = __ev_fssqrt(quad);

  // Calculate RCS compensation

  /*if( par->bRCSComp == bTrue ){

   if( y < (FFT_NUM_SAMPLES/2) ){

  sqroot = sqroot * rcs_comp(y, par->distRange);

   }

   else{

  sqroot = sqroot * rcs_comp((y-(y-128)*2-1), par->distRange);

   }

  }*/

  // Save Value

  *(outAddr) = (magnitude_t)sqroot;

  // Adjust pointer

  outAddr++;

   }

   // Adjust Pointer

   inAddr+=2;

  }

}

So, how can we enable and use the powerful VFPU and SPE2-SIMD in S32DS project?  Is there any addtional header files should be included or any SPE/FPU initialization needed besides the above project settings/configurations?

Thanks~!

Best regard,

Enwei Hu.

[stanish]

Hello Enwei,

Thanks for this report.

This is a know compiler defect - ENGR00382137 - mspe2 options doesn't enable SPE codegen and doesn't switch to SPE libs

Could you possibly add the switch "-mspe" into both compiler and linker flags.

Stan

[enwei]

Hi Stan,

I tried your suggestion and added the switch "-mspe" into both compiler and linker flags, it reported new error:

"Description Resource Path Location Type
C:/Freescale/S32_Power_v1.1/S32DS/e200_ewl2/EWL_C/include/pa/spe.h expected '=', ',', ';', 'asm' or '__attribute__' before 'static' MPC5775K_SPE_demo_prj_Z7_0 line 329, external location: C:\Freescale\S32_Power_v1.1\S32DS\e200_ewl2\EWL_C\include\pa\spe.h C/C++ Problem
"

what's wrong? how can I solve it?

thanks~!

Enwei Hu

[enwei · Accepted solution]

It's OK now, it seems can only work with newlib.

Besides, it's necessary to add the below codes in startup.s to enable SPE before use it:

I also attached my test project in S32DS version for your reference(the spe intrinsics instructions is tested in z7_1 CPU core project):

;#*************************** Enable SPE Bit in MSR ****************************

       mfmsr r6

       se_bseti r6,6

       mtmsr r6

Thanks very much for Stan's great support on this case~!:smileyhappy:

Enwei Hu.

[stanish]

Hello Enwei,

gcc C compiler cannot use these specific SPE2 vector instructions for C code.

The only way is to use these instructions in inline/standalone assembly, or use SPE intrinsics.

Simply include "<S32_Power_v1.x>\S32DS\e200_ewl2\EWL_C\include\pa\spe.h" to your .C module and use the intrinsic functions available in the header file.

Other option is to use a library that implements algorithms using SPE SIMD.

There is a AMMCLIB for MPC577xK available : Automotive Math and Motor Control Library Set|NXP

Two versions are offered -  free and pro. Pro version includes highly optimized asm code (SPE/LSP) routines.

Hope it helps.

Stan

[enwei]

Hi Stan,

Thanks for your reply.

I include "spe.h"(it located in <S32_Power_v1.x>\S32DS\e200_ewl2\EWL_C\include\pa\), and tried to call the intrinsic functions, but got serveral errors, compiler reports :

"C:/Freescale/S32_Power_v1.1/S32DS/e200_ewl2/EWL_C/include/pa/spe.h unknown type name '__ev64_opaque__'"

I searched all the project and found no '__ev64_opaque__' type definition, it's a type  that can accept any data type according to discriptions in "SPE2 Programming Interface Manual, Rev. 1.0-2" .

Could you help to made a simple project to demonstrate how to call and use the  intrinsic functions in spe.h?  the demo code to implement "sqroot = sqrt((double)quad)" is a prefer if possible.

I also tried the free version AMMCLIB for MPC5775K, and it seems has none performance improved, no SPE optimizied. I'd like to try the Pro version.

Thanks again~!

Enwei Hu.