试图理解 Booth 的乘法 radix-4 实现

Trying to understand a Booth's multiplication radix-4 implementation

我正在尝试理解一些描述 Booth 乘法与基数 4 实现的 VHDL 代码。我知道算法是如何工作的,但我似乎无法理解代码的某些部分具体做了什么。
这是整个实现:

library ieee;
use ieee.std_logic_1164.all;
use ieee.std_logic_arith.all;
use ieee.std_logic_signed.all;

entity booth_mult is
  port(
    clk             : in std_logic;
    start           : in std_logic;
    n_reset : in std_logic;
    mcand       : in std_logic_vector(15 downto 0);
    mplier      : in std_logic_vector(15 downto 0);
    done        : out std_logic;
      product   : out std_logic_vector(31 downto 0)
    );
end booth_mult;

architecture arch of booth_mult is
    type state_type is(IDLE, BUSY);
    attribute ENUM_ENCODING : string;                                           -- used for explicit state machine encoding
    attribute ENUM_ENCODING of state_type   : type is "01 10";
    signal state_reg, state_next                                : state_type;

    signal q_reg, q_next                                                : unsigned(6 downto 0);
    signal mcand_reg                                                    : std_logic_vector(15 downto 0);                -- registers for the multiplicand
    signal prod_reg, prod_next                              : std_logic_vector(32 downto 0);
    signal result_reg, result_next                          : std_logic_vector(32 downto 0);            -- this holds the result before shift
    signal q_add, q_reset                                           : std_logic;

begin
    -- increment sequential logic on rising clock edge process
    process(clk, n_reset)   
    begin
        if rising_edge(clk) then
            if n_reset = '0' then
                state_reg <= IDLE;
                q_reg <= (others => '0');
                prod_reg <= (others => '0');
            else
                q_reg <= q_next;
                state_reg <= state_next;
                prod_reg <= prod_next(32) & prod_next(32 downto 1);  -- shift prod register each time
                result_reg <= prod_next;
            end if; 
        end if;
    end process;

  -- control unit process
    process(state_reg, q_reg, result_reg, start, prod_reg, mplier, mcand )                  
        begin
            -- initialize signals and no register update
            q_add <= '0';
            q_reset <= '0';
            done <= '0';
            state_next <= state_reg;
            prod_next <= prod_reg;
            result_next <= result_reg;  

            case state_reg is

                when IDLE => 
                    if (start = '1') then   -- load numbers to multiply
                        mcand_reg <= mcand;
                        prod_next(32 downto 17) <= (others => '0');  -- prod_next reg = [0000...0000(mplier)0]
                        prod_next(16 downto 1) <= mplier;
                        prod_next(0) <= '0';
                        state_next <= BUSY;
                    end if;

                when BUSY =>
                    q_add <= '1';
                    if (q_reg = '0' & conv_unsigned(16, 7)(6 downto 1)  and start /= '1') then  -- after 8 clock cycles multiplication is done
                        product <= prod_next(32) & prod_next(32  downto 2);
                        done <= '1' ;
                        q_add <= '0';
                        q_reset <= '1';
                        state_next <= IDLE;
                    end if;
                    -- radix-4 decoding
                    case result_reg(2 downto 0) is     
                        when "001" | "010" =>   -- + mcand
                            prod_next <= ((prod_reg(32) & prod_reg(32 downto 17)) + (mcand_reg(16 - 1) & mcand_reg)) & prod_reg(16 downto 1);
                        when "011" =>                       -- + 2*mcand
                            prod_next <= ((prod_reg(32) & prod_reg(32 downto 17)) + (mcand_reg & '0' )) & prod_reg(16 downto 1);
                        when "100" =>                       -- - 2*mcand
                            prod_next <= ((prod_reg(32) & prod_reg(32 downto 17)) - (mcand_reg & '0' )) & prod_reg(16 downto 1);
                        when "101" | "110" =>   -- - mcand
                            prod_next <= ((prod_reg(32) & prod_reg(32 downto 17)) - (mcand_reg(16 - 1) & mcand_reg)) & prod_reg(16 downto 1); -- 2*mcand
                        when others =>                  -- shift only
                            prod_next <= prod_reg(32) & prod_reg(32 downto 1);
                    end case;

                end case;

    end process;


    -- timer/counter for timed logic
  q_next <= (others => '0') when q_reset = '1' else         -- reset q_next to bottom if q_reset is 1
                   q_reg + 1 when q_add = '1' else                            -- increment q_reg by 1 if q_add is 1
                   q_reg;   


end arch;   

我不明白的是:

  1. if (q_reg = '0' & conv_unsigned(16, 7)(6 downto 1) and start /= '1') then -- after 8 clock cycles multiplication is done.
    该评论使其显而易见,但它到底做了什么? conv_unsigned 的文档说它应该 return 值 16 作为大小为 7(我猜它可能是)的无符号值(它是)。 conv_usigned0 编辑的 return 不会 &ing 只是让整个事情变成 0 吗?
  2. prod_next <= prod_reg(32) & prod_reg(32 downto 1);
    再次参考评论,这应该只是一个转变。我所理解的实际发生的是 prod_reg 寄存器的第 32 位与第 32 位和第 1 位之间的同一寄存器的每一位进行 &ed,然后分配给 prod_next。这究竟是怎样一种转变?

代码在测试时有效,所以这 100% 是我缺乏 VHDL 知识的问题,所以如果问题很愚蠢,请原谅我。

回答问题 1:'&' 是连接运算符,因此 '0' & seven_bit_number 构成一个 8 位数。

所以整个表达式 '0' & conv_unsigned(16, 7)(6 downto 1)B"001000" 或 8 的冗长说法。我认为...

这是否来自混淆 VHDL 竞赛? 与类型系统进行如此费力的斗争通常意味着您缺少某些东西,或者设计存在严重错误。

我会删除这些非标准库:

use ieee.std_logic_arith.all;
use ieee.std_logic_signed.all;

赞成标准, 使用 ieee.numeric_std.all; 并简单地写

if q_reg = 8 and start /= '1' then

请注意,现在难以理解的评论不仅有道理,而且是多余的。

(可能还有其他非标准库依赖需要清理,所以这是否真的值得做取决于你)。

关于 Q2 的一个小提示:该操作不是一个班次:它似乎是一个带符号扩展的班次,所以那里可能有带符号的数字。同样,如果合适的话,将 prod_next 设为 numeric_std.signed 将使真正发生的事情变得清晰,还清一点技术债务。