Scalar Golden Reference - 2024.1 English

AI Engine Kernel and Graph Programming Guide (UG1079)

Document ID
UG1079
Release Date
2024-06-05
Version
2024.1 English

The AI Engine contains a scalar processor that can be used to implement scalar math operations, non-linear functions, and other general purpose operations. Sometimes it can be helpful to have a golden scalar reference version of code. But note that the scalar version of code takes much more time to run in simulation and hardware compared to the vectorized version.

The following provides example code for the scalar version of a 32-tap filter:
static cint16 eq_coef[32]={{1,2},{3,4},...};

//keep margin data between different invocations of the graph 
static cint16 delay_line[32];

__attribute__((noinline)) void fir_32tap_scalar(input_stream<cint16> * sig_in,
      output_stream<cint16> * sig_out){
  //For profiling only 
  unsigned cycle_num[2];
  aie::tile tile=aie::tile::current();
  
  for(int i=0;i<SAMPLES;i++){
    cycle_num[0]=tile.cycles();//cycle counter of the AI Engine tile
    cint64 sum={0,0};//larger data to mimic accumulator
    for(int j=0;j<32;j++){
      //auto integer promotion to prevent overflow
      sum.real+=delay_line[j].real*eq_coef[j].real-delay_line[j].imag*eq_coef[j].imag;
      sum.imag+=delay_line[j].real*eq_coef[j].imag+delay_line[j].imag*eq_coef[j].real;
    }
    sum=sum>>SHIFT;
    //produce one sample per loop iteration
    writeincr(sig_out,{(int16)sum.real,(int16)sum.imag});

    for(int j=0;j<32;j++){
      if(j==31){
        delay_line[j]=readincr(sig_in);
      }else{
        delay_line[j]=delay_line[j+1];
      }
    }
    //For profiling only 
    cycle_num[1]=tile.cycles();//cycle counter of the AI Engine tile
    printf("start=%d,end=%d,total=%d\n",cycle_num[0],cycle_num[1],cycle_num[1]-cycle_num[0]);
  }
  
}
void fir_32tap_scalar_init()
{
  //initialize data
  for (int i=0;i<32;i++){
    int tmp=get_ss(0);
    delay_line[i]=*(cint16*)&tmp;
  }
};
Note:
  • Function fir_32tap_scalar_init is used as an initialization function for the kernel, which will only be called once after graph.run().
  • Rounding and saturation modes are not supported in the scalar processor. They can be implemented via standard C operations, like shift.
  • Tile counter is used for profiling the main loop of code.
From the profiling result, you can see that each sample takes 3068 cycles. You can also view similar information under the Profile section in the Vitis IDE if you enable the option --profile during AI Engine simulation.
Figure 1. Profiling Details

For more information about graph construction and different kinds of profiling techniques, see AI Engine Simulation-Based Performance Analysis and Performance Analysis of AI Engine Graph Application on Hardware in AI Engine Tools and Flows User Guide (UG1076).