作者：OpenSLee

1、float IP的創建

搜索float雙擊Floating-point

1 > Operation Selection 我們這里選擇浮點數的加減法驗證。

2 > Precision of Inputs 我們選擇單晶浮點數（Single）,指數位寬Exponent Width 8bit 尾數位寬24 bit

3 > Optimizations默認值

4 > Interface Options Latency選擇1

2、浮點IP加減法的仿真驗證

我們用python 自動生成100000個隨機浮點數a和b以及a和b的相加或相減的結果。

python代碼(float32(a)+float32(b)=float32(c))：

import bitstring, random 
span = 10000000
iteration = 100000

def ieee754(flt):
    b = bitstring.BitArray(float=flt, length=32)
    return b

with open("TestAdd.txt", "w") as f:
    for i in range(iteration):
        a = ieee754(random.uniform(-span, span))
        b = ieee754(random.uniform(-span, span))
        ab = ieee754(a.float  + b.float)
        f.write(a.hex +"_" +  b.hex  +  "_" + ab.hex + "/n")

浮點數加法驗證python結果（部分）：

4a953fc4_ca39838f_49e1f7f2
4b14900e_492290ab_4b1eb919
4aedbfc3_4b146c7e_4b85a630
ca4bb7f6_cb162f1d_cb491d1a
ca1ca77e_4ad19cc6_4a834907
c96e52c3_c9c7778d_ca1f5077
4ab9c1cc_cb187d76_ca6e7240
4b18508e_4a8e5556_4b5f7b39
cb103fa5_ca1765cf_cb361919
4a09db98_ca2feb0d_c9183dd4
ca626910_4a991e54_499fa730
c983aaa6_4b0534fd_4ae97f50
49e9e5f2_cad6005d_ca9b86e0
491b3266_4a3e2d28_4a64f9c2
ca935d66_caae8cbc_cb20f511
4a150544_4a645ebe_4abcb201

3、xilinx float IP的加法驗證

s_axis_a_tdata,s_axis_b_tdata和m_axis_result_tdata分別代表浮點操作的a,b和結果c。

s_axis_operation_tdata的最低位為0時為加法，為1時為減法運算。

m_axis_result_tvalid當次信號為1時，結果有效。

浮點數加減法仿真頂層Float_AddSub_tb：

`timescale 1ns / 1ps
`define N_TESTS 100000
module Float_AddSub_tb();
    
 reg aclk;                           
 reg s_axis_a_tvalid;                
 wire s_axis_a_tready;               
 reg [31 : 0] s_axis_a_tdata;        
 reg s_axis_b_tvalid;                
 wire s_axis_b_tready;               
 reg [31 : 0] s_axis_b_tdata;        
 reg s_axis_operation_tvalid;        
 wire s_axis_operation_tready;       
 reg [7 : 0] s_axis_operation_tdata; 
 wire m_axis_result_tvalid;          
 reg m_axis_result_tready;           
 wire [31 : 0] m_axis_result_tdata;

reg [95:0] testVector [`N_TESTS-1:0];
reg test_stop;
reg [31:0] Expected_result;
reg [31:0] Expected_result_r;
integer mcd;
integer test_n;
integer pass;
integer error;

initial begin
  aclk = 0;
  test_n = 0;
  pass =0;
  error = 0;
  test_stop =0;
  s_axis_a_tvalid = 0;
  s_axis_b_tvalid = 0;
  Expected_result = 0;
  Expected_result_r = 0;
  s_axis_a_tdata = 0;
  s_axis_b_tdata = 0;
  s_axis_operation_tvalid = 1;
  s_axis_operation_tdata =8'b0000_0000;//Add
  //s_axis_operation_tdata =8'b0000_0001;//Sub
  m_axis_result_tready = 1;
  $readmemh("TestAdd.txt", testVector);//Add
  mcd = $fopen("ResultsAdd.txt");//Add
  
  //$readmemh("TestSub.txt", testVector);//Sub
 //mcd = $fopen("ResultsSub.txt");//Sub
  
  repeat(100000) begin
   #10 test_n = test_n + 1'b1;
  end
  
  wait(test_stop==1'b1)begin
    $fclose(mcd);
    $finish;
  end
end


always #(5) aclk = ~aclk;
  
  always @(posedge aclk) begin
    Expected_result_r  %d",test_n);
          pass = pass + 1'b1;
        end


      if ((m_axis_result_tvalid == 1) && (m_axis_result_tdata[31:11] != Expected_result_r[31:11]))
        begin
          $fdisplay (mcd,"Test Failed Expected Result = %h, Obtained s_axis_b_tdata = %h, Test Number -> %d",Expected_result,s_axis_b_tdata,test_n);
          error = error + 1'b1;
        end
      
      if (test_n >= `N_TESTS) 
      begin
        $fdisplay(mcd,"Completed %d tests, %d passed and %d fails.", test_n, pass, error);
        test_stop = 1'b1;
      end
  end    
//----------- Begin Cut here for INSTANTIATION Template ---// INST_TAG
floating_AddSUB your_instance_name (
  .aclk(aclk),                                        // input wire aclk
  .s_axis_a_tvalid(s_axis_a_tvalid),                  // input wire s_axis_a_tvalid
  .s_axis_a_tready(s_axis_a_tready),                  // output wire s_axis_a_tready
  .s_axis_a_tdata(s_axis_a_tdata),                    // input wire [31 : 0] s_axis_a_tdata
  .s_axis_b_tvalid(s_axis_b_tvalid),                  // input wire s_axis_b_tvalid
  .s_axis_b_tready(s_axis_b_tready),                  // output wire s_axis_b_tready
  .s_axis_b_tdata(s_axis_b_tdata),                    // input wire [31 : 0] s_axis_b_tdata
  .s_axis_operation_tvalid(s_axis_operation_tvalid),  // input wire s_axis_operation_tvalid
  .s_axis_operation_tready(s_axis_operation_tready),  // output wire s_axis_operation_tready
  .s_axis_operation_tdata(s_axis_operation_tdata),    // input wire [7 : 0] s_axis_operation_tdata
  .m_axis_result_tvalid(m_axis_result_tvalid),        // output wire m_axis_result_tvalid
  .m_axis_result_tready(m_axis_result_tready),        // input wire m_axis_result_tready
  .m_axis_result_tdata(m_axis_result_tdata)           // output wire [31 : 0] m_axis_result_tdata
);

endmodule

仿真結果：

Completed      100000 tests,       99999 passed and           0 fails.

通過仿真xilinx浮點ip的計算結果與python代碼的輸出結果一致，仿真成功。大家可以按照此方法仿真其他的算法中的計算公式或過程。首先利用C、matlab或者python等高級語言將算法的輸入和輸出一起打印出來，然后再讀入到verilog的算法模型里面，通過打印出計算結果或誤差來分析我們自己的算法的錯誤或者誤差出現在哪里。

編輯：hfy