当前位置：网站首页>HLS / chisel uses CORDIC hyperbolic system to realize square root calculation

HLS / chisel uses CORDIC hyperbolic system to realize square root calculation

2022-04-23 02:57:00 【AlwaysDayOne】

CORDIC( Coordinate rotation digital algorithm ) It's a computational triangle 、 Numerical algorithms for hyperbolic and other mathematical functions , Each operation produces a result output . This enables us to adjust the accuracy of the algorithm according to the application requirements ; More accurate results can be obtained by increasing the number of operation iterations .CORDIC Is to use only addition 、 Subtraction 、 Shift and look-up table implementation of a simple algorithm , This algorithm is in FPGA High efficiency in , It is often used in hardware algorithm implementation .

This article mainly introduces CORDIC Introduce the calculation of square root based on hyperbolic system .

HLS / Chisel Realization CORDIC Algorithmic hyperbolic system

principle

stay CORDIC In the vector mode of hyperbolic system , The final output is as follows ：

$\begin{aligned} &\left\{\begin{array}{l} x_{n}=K_{n} \sqrt{x_{1}^{2}-y_{1}^{2}} \\ y_{n}=0 \\ z_{n}=z_{1}+\tanh ^{-1}\left(y_{1} / x_{1}\right) \\ K_{n}=\prod_{i=1}^{n-1} \sqrt{1-2^{-2 i}} \end{array}\right.\\ &\sigma_{i}=\left\{\begin{array}{l} +1, y_{i} \leqslant 0 \\ -1, y_{i}>0 \end{array}\right. \end{aligned} \\ -0.784<tanh({z_n}) =\frac {y_1} {x_1}<0.784$

The specific mathematical derivation can be seen in the article cited above .

We can notice that , In this iteration , $x_n$ The square root function appears , So we can make a little change ： set up a Is the square root to be found , set up x The initial value of a+1,y The initial value of a-1,z Any value can , After iteration x The value is ：

$x=K^{*} \sqrt{(a+1)^{2}-(a-1)^{2}}=K^{*} \sqrt{4 a}=2 K^{*} \sqrt{a}$

It is known that

$\begin{gathered} K=1 / P= 0.8297816201389014 \\ P = 1.205136358399695 \end{gathered}$

Then the output of the vector mode of the hyperbolic system $x_n$ ride P, Move one more bit to the right to get a The square root of

$\sqrt{a}=(p*x_n)>>1$

Input range problem

be aware $-0.784<tanh({z_n}) =\frac {y_1} {x_1}<0.784$ , Then there are $\frac {a-1} {a+1}<0.784$ , therefore x The input range of needs to be limited to [0,8] Within the scope of .

Besides , because [0,1] The number in the range is too small , Its error in iterative numerical calculation is too large , The end result is not ideal , So we want the input range to be [1,8].

In order to expand the scope to R+, We can input a Use the shift operation to move N position ,N For the even , Shift to [1,8] Get after range $a ‘$ , Finally through HV The result is obtained after module calculation $x_n$ , In shift N/2 Just recover .

HLS practice

Be careful ： Here only the square operation of the basic class is realized , The input problem is not solved by binary shift , Readers can refer to solve .

#define THETA_TYPE float
#define COS_SIN_TYPE float
#define NUM_ITERATIONS 50

THETA_TYPE cordic_P_mark_n = 1.205136358399695;
void cordic_hv_square_root(COS_SIN_TYPE a, COS_SIN_TYPE &x_out)
{
    
    x = a+1;
    y = a-1;
    
    for(int j = 1;j <= NUM_ITERATIONS; j++){
    
        COS_SIN_TYPE temp_x = x;
        
        if(y < 0){
    
            x = x + y >> j;
            y = y + temp_x >> j;
            theta = theta - cordic_phase_h[j-1];
        } else{
    
            x = x - y >> j;
            y = y - temp_x >> j;
            theta = theta + cordic_phase_h[j-1];
        }
    }
    
    x_out = (x * cordic_P_mark_n) >> 1;
}

Chisel practice

Fixed point number definition

First define the number of fixed points , Pay attention to chisel.experimental There are fixed-point classes in the package

/* Cordic  Global constant definition of  */
trait Cordic_Const_Data extends HasDataConfig{
    
  /*  Number of iterations configuration  */
  val NUM_ITERATIONS = 20
}

trait HasDataConfig {
    
  /*  Bit width definition of fixed-point number  */ 
  val DataWidth = 32
  val BinaryPoint = 20
}

The source code to achieve

Here, on the basis of the quoted article, it is changed cordic Algorithm HV Module code , Note that you only need to calculate x,y that will do , There is no need for... In the calculation of hyperbolic system z

First, implement binary shift , And generate associated objects

class shift_range_1_8 extends Module with HasDataConfig{
    
  /* * @x :  Enter the number of bits to be moved   Fixed point number type   Value range R * @out :  Output the shifted result   Range [1,8] * @cnt :  Number of shifts output , Move left to positive , Shift right to negative  * details:  Combinational logic circuits will x Shift to the range by double left and right [1,8], And output the number of shifts  **/
  val io = IO(new Bundle {
    
    val x: FixedPoint = Input(FixedPoint(DataWidth.W, BinaryPoint.BP))
    val out: FixedPoint = Output(FixedPoint(DataWidth.W, BinaryPoint.BP))
    val cnt: SInt = Output(SInt(log2Ceil(DataWidth).W))
  })

  val temp_x: FixedPoint = io.x

  when(temp_x < FixedPoint.fromDouble(1, DataWidth.W, BinaryPoint.BP)){
    
    /*  Than 0.5 Small , Need to move left  */
    val index = VecInit(Seq.fill(BinaryPoint>>1)(0.B))
    for(i <- 0 until (BinaryPoint>>1)){
    
      when((temp_x << (i<<1)) < FixedPoint.fromDouble(1, DataWidth.W, BinaryPoint.BP)){
    
        index(i) := 0.B
      }.otherwise{
    
        index(i) := 1.B
      }
    }
    /*  The priority encoder first obtains greater than or equal to 1 The location of , namely cnt Value  */
    val temp_cnt = PriorityEncoder(index.asUInt())
    io.cnt := (temp_cnt<<1).asSInt()
    io.out := (temp_x << (temp_cnt<<1))
  }.elsewhen(temp_x > FixedPoint.fromDouble(8, DataWidth.W, BinaryPoint.BP)){
    
    /*  Than 8 Big , Need to move right  */
    val index = VecInit(Seq.fill(DataWidth>>1)(0.B))
    for(i <- 0 until (DataWidth>>1)){
    
      when((temp_x >> (i<<1)) > FixedPoint.fromDouble(8, DataWidth.W, BinaryPoint.BP)){
    
        index(i) := 0.B
      }.otherwise{
    
        index(i) := 1.B
      }
    }
    /*  The priority encoder first obtains less than or equal to 8 The location of , namely cnt Value  */
    val temp_cnt = PriorityEncoder(index.asUInt())
    io.cnt := -((temp_cnt<<1).asSInt())
    io.out := (temp_x >> (temp_cnt<<1))
  }.otherwise{
    
    io.cnt := 0.S
    io.out := temp_x
  }
}


/* *  Define the companion object of this class , A factory method is defined to simplify the instantiation and wiring of modules . **/
object shift_range_1_8 {
    
  def apply(x: FixedPoint): (FixedPoint, SInt) = {
    
    /* * @x :  Enter the number of bits to be moved   Fixed point number type   Value range R * @out :  Output the shifted result   Range [0.5,1] * @cnt :  Number of shifts output , Move left to positive , Shift right to negative  * @flag :  Output x The positive and negative symbols of are 1,0 * details:  take x Shift left and right to range [0.5,1], And output the number of shifts and whether to change the symbol  **/
    val unit = Module(new shift_range_1_8)
    unit.io.x := x
    (unit.io.out, unit.io.cnt)
  }
}

Realization hv Module to calculate the square root source code

class cordic_hv_square_root_origin extends Module with HasDataConfig with Cordic_Const_Data{
    
  /* * @in :  The horizontal axis coordinates of the input vector   Fixed point number type  in∈[0,8] * @out :  Output x^0.5  Fixed point number type  * details:  utilize cordic The scaling of vector mode iteration in hyperbolic coordinate system is obtained (in+1,in-1) Output of coordinates x,  namely 2K(in^0.5), cube 1/K=P, Just move one bit to the right . Be careful HV Mode limits y/x The scope of the , so  in∈[0,8] */
  val io = IO(new Bundle {
    
    val in: FixedPoint = Input(FixedPoint(DataWidth.W, BinaryPoint.BP))
    val out: FixedPoint = Output(FixedPoint(DataWidth.W, BinaryPoint.BP))
  })

  val x: FixedPoint = io.in + (1.F(DataWidth.W, BinaryPoint.BP))
  val y: FixedPoint = io.in - (1.F(DataWidth.W, BinaryPoint.BP))

  /*  When approaching infinity ,K The value becomes constant  */
  val cordic_P_h: FixedPoint = FixedPoint.fromDouble(1.205136358399695, DataWidth.W, BinaryPoint.BP)

  /*  Initializes the calculated register array , formation NUM_ITERATIONS It's a cascade of water  */
  val current_x: Vec[FixedPoint] = RegInit(VecInit(Seq.fill(NUM_ITERATIONS)(FixedPoint.fromDouble(0.0, DataWidth.W, BinaryPoint.BP)))) // cos
  val current_y: Vec[FixedPoint] = RegInit(VecInit(Seq.fill(NUM_ITERATIONS)(FixedPoint.fromDouble(0.0, DataWidth.W, BinaryPoint.BP)))) // sin

  for (i <- 1 to NUM_ITERATIONS) {
    
    /* * x[i] = K`(x[i-1] + sigma * 2^(-i) * y[i-1]) * y[i] = K`(y[i-1] + sigma * 2^(-i) * x[i-1]) * z[i] = z[i-1] - sigma * arctanh(2^(-i)) **/
    if (i == 1) {
    
      /*  The first stage of the pipeline is directly connected to (io.x,io.y) Click to do the operation  */
      when(y < FixedPoint.fromDouble(0.0, DataWidth.W, BinaryPoint.BP)) {
    
        current_x(i - 1) := x + (y >> i)
        current_y(i - 1) := y + (x >> i)
      }.otherwise {
    
        current_x(i - 1) := x - (y >> i)
        current_y(i - 1) := y - (x >> i)
      }
    } else {
    
      when(current_y(i - 2) < FixedPoint.fromDouble(0.0, DataWidth.W, BinaryPoint.BP)) {
    
        current_x(i - 1) := current_x(i - 2) + (current_y(i - 2) >> i) //  Shift instead of multiplication 
        current_y(i - 1) := current_y(i - 2) + (current_x(i - 2) >> i)
      }.otherwise {
    
        current_x(i - 1) := current_x(i - 2) - (current_y(i - 2) >> i)
        current_y(i - 1) := current_y(i - 2) - (current_x(i - 2) >> i)
      }
    }
  }

  io.out := (current_x(NUM_ITERATIONS - 1) * cordic_P_h ) >> 1
}

test

import org.scalatest._
import chisel3._
import chiseltest._
import chisel3.experimental._
import scala.math._

class Cordic_Square_Root_Tester extends FlatSpec with ChiselScalatestTester with Matchers {
    
  behavior of "mytest2"
  it should "do something" in {
    
    test(new cordic_hv_square_root) {
     c =>
      c.io.in.poke(2.F(32.W, 20.BP))
      c.clock.step(20)
      println(s"out ${
      c.io.out.peek}\n")

      c.io.in.poke(4.F(32.W, 20.BP))
      c.clock.step(20)
      println(s"out ${
      c.io.out.peek}\n")

      c.io.in.poke(8.F(32.W, 20.BP))
      c.clock.step(20)
      println(s"out ${
      c.io.out.peek}\n")

      c.io.in.poke(16.F(32.W, 20.BP))
      c.clock.step(20)
      println(s"out ${
      c.io.out.peek}\n")

      c.io.in.poke(64.F(32.W, 20.BP))
      c.clock.step(20)
      println(s"out ${
      c.io.out.peek}\n")

      c.io.in.poke(128.F(32.W, 20.BP))
      c.clock.step(20)
      println(s"out ${
      c.io.out.peek}\n")

      //  Note that there 32 There are not enough seats 
// c.io.in.poke(FixedPoint.fromDouble(65532,32.W, 20.BP))
// c.clock.step(20)
// println(s"out ${c.io.out.peek}\n")

      c.io.in.poke(0.00132.F(32.W, 20.BP))
      c.clock.step(20)
      println(s"out ${
      c.io.out.peek}\n")

      c.io.in.poke(0.6315.F(32.W, 20.BP))
      c.clock.step(20)
      println(s"out ${
      c.io.out.peek}\n")

      c.io.in.poke(0.000632.F(32.W, 20.BP))
      c.clock.step(20)
      println(s"out ${
      c.io.out.peek}\n")
    }
  }
}

//  perform  sbt "testOnly Cordic_Square_Root_Tester"