I'm trying to implement a floating point adder/subtractor. I've already achieved the code to work. When I run the simulation it works as expected.

The thing is when I try to synthesize it in Vivado I get this errors: "constant expression required width mismatch in assignment" and "target has 279 bits, source has 86 bits" on a line that says:

my_p_nt_a <= to_unsigned(0, to_integer(abs_exp_diff)) & my_p & to_unsigned(0, 2**NE - 1- to_integer(abs_exp_diff));

This corresponds to the mantissa's alignment.

Some more info:

signal my_p_nt_a:      unsigned(NF + 2**NE-1 downto 0);
signal abs_exp_diff:   unsigned(NE downto 0);

NE is the number of bits the exponent has.

I don't know how to solve it. I've tried different ways but I keep getting problems when synthesizing.

I'm trying to implement a floating point adder/subtractor. I've already achieved the code to work. When I run the simulation it works as expected.

The thing is when I try to synthesize it in Vivado I get this errors: "constant expression required width mismatch in assignment" and "target has 279 bits, source has 86 bits" on a line that says:

my_p_nt_a <= to_unsigned(0, to_integer(abs_exp_diff)) & my_p & to_unsigned(0, 2**NE - 1- to_integer(abs_exp_diff));

This corresponds to the mantissa's alignment.

Some more info:

signal my_p_nt_a:      unsigned(NF + 2**NE-1 downto 0);
signal abs_exp_diff:   unsigned(NE downto 0);

NE is the number of bits the exponent has.

I don't know how to solve it. I've tried different ways but I keep getting problems when synthesizing.

• floating-point
• vhdl

• We probably can't say very much without knowing the size of all of those signals. If you really are concatenating 279 bits and trying to store it in an 86 bit destination, that certainly is a problem. – Tim Roberts Commented Nov 20, 2024 at 0:35
• 1 Please edit your question to add a minimal reproducible example. – the busybee Commented Nov 20, 2024 at 6:51

1 Answer 1

The meaning of the error message is, in Vivado synthesis you can't extend your operand by a number of 0's where the number depends on the value at abs_exp_diff. This works only, if abs_exp_diff would be a constant. I expect most of the synthesis tools will not be able to do this.

Instead you should use this solution:

entity test_shift is
end entity test_shift;
library ieee;
use ieee.std_logic_1164.all;
use ieee.numeric_std.all;
architecture struct of test_shift is
    constant NF : natural := 10; -- example value
    constant NE : natural := 12; -- example value
    signal abs_exp_diff : unsigned(NE downto 0);
    signal my_p         : unsigned(NF-1 downto 0);
    signal my_p_nt_a    : unsigned(NF+2**NE-1 downto 0);
begin
    process(my_p, abs_exp_diff)
        variable my_p_nt_a_v : unsigned(my_p_nt_a'range);
    begin
        my_p_nt_a_v := (others => '0');
        my_p_nt_a_v(my_p'range) := my_p;
        my_p_nt_a <= shift_left(my_p_nt_a_v, 2*NE-1-to_integer(abs_exp_diff));
    end process;
end architecture;

Edit: You could also use a function:

function insert_my_p (my_p         : in unsigned(NF-1 downto 0);
                      abs_exp_diff : in unsigned(NE   downto 0))
return unsigned is
    variable my_p_nt_a_v : unsigned(NF+2**NE-1 downto 0);
begin
    my_p_nt_a_v := (others => '0');
    my_p_nt_a_v(my_p'range) := my_p;
    return shift_left(my_p_nt_a_v, 2*NE-1-to_integer(abs_exp_diff));
end function;