中文 | English
SystemVerilog 定点数库。
- 可定制整数位宽和小数位宽。
- 运算 : 加、减、乘、除、开方。
- 溢出检测 : 发生溢出时,溢出信号=1,输出结果会设为正最大值(上溢出)或负最小值(下溢出)。
- 舍入控制 : 发生截断时,可选择是否进行四舍五入。
- 与单精度浮点数(IEEE754)互相转换 。
- 所有运算均有 单周期实现 ,组合逻辑延迟长的运算有 流水线实现 。
定点数由若干位整数和若干位小数组成。其值= 该二进制码对应的整数补码 除以 2^小数位数 。
例如,若整数位数为8,小数位数为8,举例如下表:
二进制码 | 整数补码 | 定点数值 (8位整数,8位小数) | 备注 |
---|---|---|---|
0000000000000000 | 0 | 0.0 | 零值 |
0000000100000000 | 256 | 1.0 | |
1111111100000000 | -256 | -1.0 | |
0000000000000001 | 1 | 0.00390625 | 正最小值 |
1111111111111111 | -1 | -0.00390625 | 负最大值 |
0111111111111111 | 32767 | 127.99609375 | 正最大值 |
1000000000000000 | -32768 | -128.0 | 负最小值 |
0001010111000011 | 5571 | 21.76171875 | |
1001010110100110 | -27226 | -106.3515625 |
本库的模块输入输出都可以用参数 (parameter) 来定制定点数位宽,这些参数的命名是统一的:
- WOI 和 WOF 分别是输出的定点数的整数位宽和小数位宽。
- 对于单目运算, WII 和 WIF 分别是输入的定点数的整数位宽和小数位宽。
- 对于双目运算, WIIA 和 WIFA 分别是输入操作数A的定点数的整数位宽和小数位宽;WIIB 和 WIFB 分别是输入操作数B的定点数的整数位宽和小数位宽。
以乘法器为例:
module fxp_mul # ( // 以乘法器为例
parameter WIIA = 8, // 输入(乘数a)的整数位宽,默认=8
parameter WIFA = 8, // 输入(乘数a)的小数位宽,默认=8
parameter WIIB = 8, // 输入(乘数b)的整数位宽,默认=8
parameter WIFB = 8, // 输入(乘数b)的小数位宽,默认=8
parameter WOI = 8, // 输出(积)的整数位宽,默认=8
parameter WOF = 8, // 输出(积)的小数位宽,默认=8
parameter bit ROUND= 1 // 当积的小数截断时,是否四舍五入,默认是
)(
input wire [WIIA+WIFA-1:0] ina, // 乘数a
input wire [WIIB+WIFB-1:0] inb, // 乘数b
output wire [WOI +WOF -1:0] out, // 结果(积) = 乘数a * 乘数b
output wire overflow // 结果是否溢出,若溢出则为 1'b1
// 若为上溢出,则out被置为最大正值
// 若为下溢出,则out被置为最小负值
);
所有可综合的模块实现都在 RTL/fixedpoint.sv 中,各模块名称和功能如下表:
运算 | 单周期版本(组合逻辑) | 流水线版本 | 流水线级数 | 说明 |
---|---|---|---|---|
位宽变换 | fxp_zoom | 不需要 | - | 有溢出、舍入控制 |
加减 | fxp_addsub | 不需要 | - | 具有1bit信号控制加或减 |
加 | fxp_add | 不需要 | - | |
乘法 | fxp_mul | fxp_mul_pipe | 2 | |
除法 | fxp_div | fxp_div_pipe | WOI+WOF+3 | 单周期版时序不易收敛 |
开方(sqrt) | fxp_sqrt | fxp_sqrt_pipe | WII/2+WIF+2 | 单周期版时序不易收敛 |
定点转浮点 | fxp2float | fxp2float_pipe | WII+WIF+2 | 单周期版时序不易收敛 |
浮点转定点 | float2fxp | float2fxp_pipe | WOI+WOF+4 | 单周期版时序不易收敛 |
仿真相关的文件都在 SIM 文件夹中,其中:
代码文件 | 说明 |
---|---|
tb_add_sub_mul_div.sv | 测试单周期版本的加减乘除 |
tb_fxp_mul_div_pipe.sv | 测试流水线版本的乘除 |
tb_fxp_sqrt.sv | 测试开方运算 |
tb_convert_fxp_float.sv | 测试定点数与浮点数互相转化 |
使用 iverilog 进行仿真前,需要安装 iverilog ,见:iverilog_usage
然后双击对应的 .bat 文件就能运行仿真。
SystemVerilog fixed-point library. features:
- Customizable integer bit width and fractional bit width.
- Arithmetic : Addition, Subtraction, Multiplication, Division, Square Root.
- Overflow detection: When an overflow occurs, the overflow signal = 1, and the output result will be set to a positive maximum value (upflow) or a negative minimum value (underflow).
- Round Control : When truncation occurs, you can choose whether to round up or not.
- Supports converting to and from single-precision floating-point numbers (IEEE-754).
- All operations have single-cycle implementation, operations with long combinatorial logic delays have pipeline implementation.
A fixed-point number consists of a number of integers and a number of fractions. Its value = the integer's complement of its binary code divided by 2^fractional bit width .
For example, if the integer bit width is 8 and the fractional bit width is 8, the following table shows some fixed-point values:
binary code | Integer's complement | fixed-point value | remark |
---|---|---|---|
0000000000000000 | 0 | 0.0 | zero |
0000000100000000 | 256 | 1.0 | |
1111111100000000 | -256 | -1.0 | |
0000000000000001 | 1 | 0.00390625 | positive minimum |
1111111111111111 | -1 | -0.00390625 | negative maximum |
0111111111111111 | 32767 | 127.99609375 | positive maximum |
1000000000000000 | -32768 | -128.0 | negative minimum |
0001010111000011 | 5571 | 21.76171875 | |
1001010110100110 | -27226 | -106.3515625 |
All module in this repository use parameters to define the fixed-point bit width. The names of these parameters are unified:
WOI
andWOF
are the integer bit width and fractional bit width of the output fixed-point number, respectively.- For unary operations,
WII
andWIF
are the integer and fractional widths of the input fixed-point number, respectively. - For binocular operations,
WIIA
andWIFA
are the integer and fractional widths of the fixed-point number of input operand A, respectively;WIIB
andWIFB
are the integer and fractional widths of the fixed-point number of input operand B, respectively.
Take the multiplier as an example:
module fxp_mul # (
parameter WIIA = 8,
parameter WIFA = 8,
parameter WIIB = 8,
parameter WIFB = 8,
parameter WOI = 8,
parameter WOF = 8,
parameter bit ROUND= 1
)(
input wire [WIIA+WIFA-1:0] ina,
input wire [WIIB+WIFB-1:0] inb,
output wire [WOI +WOF -1:0] out,
output wire overflow
);
All synthesizable modules are implemented in RTL/fixedpoint.sv . The names and functions of each module are as follows:
Operation | Single-cycle ver. | Pipelined ver. | Pipeline stages |
---|---|---|---|
width conversion | fxp_zoom | unnecessary | - |
Addition/Subtraction | fxp_addsub | unnecessary | - |
Addition | fxp_add | unnecessary | - |
Multiplication | fxp_mul | fxp_mul_pipe | 2 |
Division | fxp_div | fxp_div_pipe | WOI+WOF+3 |
Square Root | fxp_sqrt | fxp_sqrt_pipe | WII/2+WIF+2 |
Fixed-point to Float | fxp2float | fxp2float_pipe | WII+WIF+2 |
Float to Fixed-point | float2fxp | float2fxp_pipe | WOI+WOF+4 |
Simulation related files are in the SIM folder, where:
source file | description |
---|---|
tb_add_sub_mul_div.sv | Test single-cycle version of addition, subtraction, multiplication and division |
tb_fxp_mul_div_pipe.sv | Test the pipelined version of multiply and divide. |
tb_fxp_sqrt.sv | Test Square Root. |
tb_convert_fxp_float.sv | Test the conversion between fixed-point and floating-point. |
Before using iverilog for simulation, you need to install iverilog , see: iverilog_usage
Then double-click the corresponding .bat file to run the simulation.