Modeling Streaming Data Interfaces - 2021.2 English

Vitis High-Level Synthesis User Guide (UG1399)

Document ID
UG1399
Release Date
2021-12-15
Version
2021.2 English

Unlike software, the concurrent nature of hardware systems allows them to take advantage of streaming data. Data is continuously supplied to the design and the design continuously outputs data. An RTL design can accept new data before the design has finished processing the existing data.

As Understanding Volatile Data shows, modeling streaming data in software is non-trivial, especially when writing software to model an existing hardware implementation (where the concurrent/streaming nature already exists and needs to be modeled).

There are several possible approaches:

  • Add the volatile qualifier as shown in the Multi-Access Volatile Pointer Interface example. The test bench does not model unique reads and writes, and RTL simulation using the original C/C++ test bench might fail, but viewing the trace file waveforms shows that the correct reads and writes are being performed.
  • Modify the code to model explicit unique reads and writes. See the following example.
  • Modify the code to using a streaming data type. A streaming data type allows hardware using streaming data to be accurately modeled.

The following code example has been updated to ensure that it reads four unique values from the test bench and write two unique values. Because the pointer accesses are sequential and start at location zero, a streaming interface type can be used during synthesis.

#include "pointer_stream_good.h"

void pointer_stream_good ( volatile dout_t *d_o,  volatile din_t *d_i) {
 din_t acc = 0;

 acc += *d_i;
 acc += *(d_i+1);
 *d_o = acc;
 acc += *(d_i+2);
 acc += *(d_i+3);
 *(d_o+1) = acc;
}

The test bench is updated to model the fact that the function reads four unique values in each transaction. This new test bench models only a single transaction. To model multiple transactions, the input data set must be increased and the function called multiple times.

#include "pointer_stream_good.h"
 
int main () {
 din_t d_i[4];
 dout_t d_o[4];
     int i, retval=0;
     FILE        *fp;

 // Create input data
 for (i=0;i<4;i++) {
     d_i[i] = i;
 }

 // Call the function to operate on the data
 pointer_stream_good(d_o,d_i);

 // Save the results to a file
 fp=fopen(result.dat,w);
 for (i=0;i<4;i++) {
     if (i<2)
 fprintf(fp, %d %d\n, d_i[i], d_o[i]);
     else
 fprintf(fp, %d \n, d_i[i]);
 }
 fclose(fp);

 // Compare the results file with the golden results
 retval = system(diff --brief -w result.dat result.golden.dat);
 if (retval != 0) {
     printf(Test failed  !!!\n); 
     retval=1;
 } else {
     printf(Test passed !\n);
 }

 // Return 0 if the test
 return retval;
}

The test bench validates the algorithm with the following results, showing that:

  • There are two outputs from a single transaction.
  • The outputs are an accumulation of the first two input reads, plus an accumulation of the next two input reads and the previous accumulation.
    Din Dout
    0   1
    1   6
    2 
    3
  • The final issue to be aware of when pointers are accessed multiple time at the function interface is RTL simulation modeling.