constants of sfixed in architecture body

[Paul]

Hi,

in the example attached I did try to set some filter coeffs constants
for a fir filter (direct form I second order structure). Modelsim
complains on

Type mark "coeefficient_type" cannot be actual parameter.
Illegal name (subtype declaration "coeefficient_type") in expression.

Is there a limitation or simple a syntax error. Is there a way to do
this - how to get it?

Thanks,
Olaf


library ieee;
use ieee.std_logic_1164.all;
use ieee.numeric_std.all;

library floatfixlib;
use floatfixlib.math_utility_pkg.all; -- ieee_proposed for VHDL-93 version
use floatfixlib.fixed_pkg.all; -- ieee_proposed for compatibility version

entity df1tsos is

port (
clk : in std_ulogic;
clk_en : in std_ulogic;
reset : in std_ulogic;
input : in sfixed(16 downto -14); -- [-2 2)
output : out sfixed(16 downto -14) -- [-2 2)
);

end entity df1tsos;

architecture rtl of df1tsos is

subtype numerator_type is sfixed(16 downto -14); -- [-2 2)
subtype denominator_type is sfixed(16 downto -14); -- [-2 2)
subtype numerator_state_type is sfixed(16 downto -12); -- [-8 8)
subtype denominator_state_type is sfixed(16 downto -12); -- [-8 8)
subtype coeefficient_type is sfixed(16 downto -14); -- [-2 2)

signal numerator_s : numerator_type;
signal denominator_s : denominator_type;

constant b0_coeff : coeefficient_type := to_sfixed(
0.000000000000000, coeefficient_type);
constant b1_coeff : coeefficient_type := to_sfixed(
1.000000000000000, coeefficient_type);
constant b2_coeff : coeefficient_type := to_sfixed(
0.843811035156250, coeefficient_type);
constant a0_coeff : coeefficient_type := to_sfixed(
1.000000000000000, coeefficient_type);
constant a1_coeff : coeefficient_type :=
to_sfixed(-1.538757324218750, coeefficient_type);
constant a2_coeff : coeefficient_type := to_sfixed(
0.601013183593750, coeefficient_type);

begin


end architecture rtl;

 

[Paul]

architecture rtl of df1tsos is

subtype numerator_type is sfixed(16 downto -14); -- [-2 2)
subtype denominator_type is sfixed(16 downto -14); -- [-2 2)
subtype numerator_state_type is sfixed(16 downto -12); -- [-8 8)
subtype denominator_state_type is sfixed(16 downto -12); -- [-8 8)
subtype coeefficient_type is sfixed(16 downto -14); -- [-2 2)

signal numerator_s : numerator_type;
signal denominator_s : denominator_type;

constant b0_coeff : coeefficient_type := to_sfixed(
0.000000000000000, coeefficient_type);
constant b1_coeff : coeefficient_type := to_sfixed(
1.000000000000000, coeefficient_type);
constant b2_coeff : coeefficient_type := to_sfixed(
0.843811035156250, coeefficient_type);
constant a0_coeff : coeefficient_type := to_sfixed(
1.000000000000000, coeefficient_type);
constant a1_coeff : coeefficient_type :=
to_sfixed(-1.538757324218750, coeefficient_type);
constant a2_coeff : coeefficient_type := to_sfixed(
0.601013183593750, coeefficient_type);

constant b0_coeff : coeefficient_type := to_sfixed(
0.000000000000000, coeefficient_type'high, coeefficient_type'low);
constant b1_coeff : coeefficient_type := to_sfixed(
1.000000000000000, coeefficient_type'high, coeefficient_type'low);
constant b2_coeff : coeefficient_type := to_sfixed(
0.843811035156250, coeefficient_type'high, coeefficient_type'low);
constant a0_coeff : coeefficient_type := to_sfixed(
1.000000000000000, coeefficient_type'high, coeefficient_type'low);
constant a1_coeff : coeefficient_type :=
to_sfixed(-1.538757324218750, coeefficient_type'high,
coeefficient_type'low);
constant a2_coeff : coeefficient_type := to_sfixed(
0.601013183593750, coeefficient_type'high, coeefficient_type'low);

must it really be complicated?

Thanks,
Olaf

 

[Tricky]

architecture rtl of df1tsos is

  subtype numerator_type is sfixed(16 downto -14);  -- [-2 2)
  subtype denominator_type is sfixed(16 downto -14);  -- [-2 2)
  subtype numerator_state_type is sfixed(16 downto -12); -- [-8 8)
  subtype denominator_state_type is sfixed(16 downto -12); -- [-8 8)
  subtype coeefficient_type is sfixed(16 downto -14); -- [-2 2)

  signal numerator_s            : numerator_type;
  signal denominator_s          : denominator_type;

  constant b0_coeff             : coeefficient_type := to_sfixed(
0.000000000000000, coeefficient_type);
  constant b1_coeff             : coeefficient_type := to_sfixed(
1.000000000000000, coeefficient_type);
  constant b2_coeff             : coeefficient_type := to_sfixed(
0.843811035156250, coeefficient_type);
  constant a0_coeff             : coeefficient_type := to_sfixed(
1.000000000000000, coeefficient_type);
  constant a1_coeff             : coeefficient_type :=
to_sfixed(-1.538757324218750, coeefficient_type);
  constant a2_coeff             : coeefficient_type := to_sfixed(
0.601013183593750, coeefficient_type);

  constant b0_coeff             : coeefficient_type := to_sfixed(
0.000000000000000, coeefficient_type'high, coeefficient_type'low);
  constant b1_coeff             : coeefficient_type := to_sfixed(
1.000000000000000, coeefficient_type'high, coeefficient_type'low);
  constant b2_coeff             : coeefficient_type := to_sfixed(
0.843811035156250, coeefficient_type'high, coeefficient_type'low);
  constant a0_coeff             : coeefficient_type := to_sfixed(
1.000000000000000, coeefficient_type'high, coeefficient_type'low);
  constant a1_coeff             : coeefficient_type :=
to_sfixed(-1.538757324218750, coeefficient_type'high,
coeefficient_type'low);
  constant a2_coeff             : coeefficient_type := to_sfixed(
0.601013183593750, coeefficient_type'high, coeefficient_type'low);

must it really be complicated?

Thanks,
Olaf

The function you are trying to call expects a declared object (so a
signal, variable or constant). you cannot pass types into functions.

you could do it with something like:

constant COEFF_SIZER : coeeffiecient_type;

constant a2_coeff : coeefficient_type := to_sfixed(0.724545,
COEFF_SIZER);

Just a quick question: are you sure you want such large coefficients
and data types, and ones that dont divide by 18? Altera multipliers
(and I think Xilinx too) work on 18 bit inputs (giving 36 bits
results). 2x 36bit inputs eats up 4 multipliers.

 

[Paul]

Thanks for your help,

constant COEFF_SIZER : coeeffiecient_type;

got unfortunately the rror

Deferred constants are allowed only in a package declaration.

by modelsim. If I write

subtype coeefficient_type is sfixed(16 downto -14); -- [-2 2)

signal coeeffiecient_sizer : coeefficient_type;

constant b0_coeff : coeefficient_type := to_sfixed(
0.000000000000000, coeeffiecient_sizer);

Warning: value of "b0_coeff" depends on value of signal
"coeeffiecient_sizer". What are the consequences on this, is it fixable?



Thanks,
Olaf

 

[Tricky]

Thanks for your help,

Ooops, my bad. It needs an actual value. try this instead:
constant COEFF_SIZER : coeeffiecient_type := (others => '0');

got unfortunately the rror

Deferred constants are allowed only in a package declaration.

by modelsim. If I write

  subtype coeefficient_type is sfixed(16 downto -14); -- [-2 2)

  signal coeeffiecient_sizer  : coeefficient_type;

  constant b0_coeff             : coeefficient_type := to_sfixed(
0.000000000000000, coeeffiecient_sizer);

Warning: value of "b0_coeff" depends on value of signal
"coeeffiecient_sizer". What are the consequences on this, is it fixable?

Thanks,
Olaf

Nothing for this function. This function only uses the sizing
attributes from the signal, not the actual value stored in the signal.

 

[Paul]

Just a quick question: are you sure you want such large coefficients

and data types, and ones that dont divide by 18? Altera multipliers
(and I think Xilinx too) work on 18 bit inputs (giving 36 bits
results). 2x 36bit inputs eats up 4 multipliers.

why is it better to use data types which divide by 18? I want to use
Xilinx DSP48 slices on virtex4. My Coeffs have a width of 16bit (18 I
could use). Input is 16bit also, multiplying get 32 bit, DSP48's adder
has an input width of 48bit - there is room for sum results bigger than
1. I did scale my input to [1 1), my coeffs are [2 -2). Maybe I miss
something? I did wonder why they have 18bits.

BTW, my first implementation of fixed point arithmetic / FIR/IIR in FPGA

Thanks,
Olaf