> Hi Symon,
> for (most) of the given examples case and if-elsif will give the same
> result.
> Why?
> Because in both cases :-) your selector is fully covered (uses all of
> the bits of a vector or whatever you use as a selector). Your if-elsif
> collapses into a parallel structure, because there is no priority of one
> value over the other possible.
>
> but how about this:
>
> Selector <= A&B&C; -- I MUST do this for a case statement !
> case Selector is
> when "001" => Output <= Input1;
> when "010" => Output <= Input2;
> when "100" => Output <= Input2;
> when others => Output <= (others => 'Z');
> end case;
>
> vs.
>
> If C = '1' then
> Output <= Input1;
> elsif B= '1' then
> Output <= Input2;
> elsif C= '1' then
> Output <= Input3;
> else
> Output <= (others => 'Z');
> end if;
>
> NOW the case produces a parallel multiplexer structure that is sensitive
> for the given code of Selector.
> The if-elsif does something different. Whenever C becomes '1' (No matter
> how unlikely or unneccesary this might be in your particular design) it
> switches Input1 to the output. Here we have the always cited priority
> encoder.
>
> To get the same functionality with a case you have to write a different
> code:
>
> Selector <= A&B&C; -- I MUST do this for a case statement !
> case Selector is
> when "--1" => Output <= Input1;
> when "-10" => Output <= Input2;
> when "100" => Output <= Input2;
> when others => Output <= (others => 'Z');
> end case;
>
> Now the first >when< hits whenever C becomes '1'.
> This code might produce something more "parallel" than the if-elsif, but
> who knows about the tricks of modern synthesis tools. :-)
>
> Have a nice synthesis
>
> Eilert
Just a few errors :
1. you can in fact qualify the expression and not use a signal.
A variable is usually preferable if you want one (otherwise,
you must add Selector in your sensitivity list, and this slows down
the simulation without usefulness).
case SLV3'(A&B&C) is -- with subtype SLV3 is std_logic_vector (3 downto 0);
2.
output <= 'Z' infers a tristate, nothing to do with a don't care!
A don't care '-' does eliminate C from the result.
3. case ... is when "--1" is wrong !
Comparing anything (but a '-') to '-' produces a false.
Comparing with '-' (to ignore the comparison) requires the use
of std_match (not usable in case statement).
VHDL is not always intuitive...
Bert Cuzeau
Reply by info_●May 4, 20052005-05-04
Neo wrote:
> Info,
> The code above given by you for onehot did sysnthesize differently. the
> "case" version systhesized to 3 LUT's involving only OR gates and didnt
> infer priority structure infact it optimized as you have mentioned to a
> series of OR gates. But the "if" version systhesiszed to 6 LUT's and
> inferred a priority structure. Leonardo was used for the systhesis.
>
Yes my point exactly. Not all tools do this correctly. Leoanrdo is right.
Just try Precision Synthesis (or other tools) if you can and compare
the results.
The if .. elsif is a priority encoder which has a well defined behavior
for the overlapping cases (more than one 1 in the vector), and this
requires more logic than the "pure one hot". This edscription is
more predictible acroos tools.
Bert
Reply by backhus●May 2, 20052005-05-02
Hi Symon,
for (most) of the given examples case and if-elsif will give the same result.
Why?
Because in both cases :-) your selector is fully covered (uses all of the bits of a vector or whatever you use as a
selector). Your if-elsif collapses into a parallel structure, because there is no priority of one value over the other
possible.
but how about this:
Selector <= A&B&C; -- I MUST do this for a case statement !
case Selector is
when "001" => Output <= Input1;
when "010" => Output <= Input2;
when "100" => Output <= Input2;
when others => Output <= (others => 'Z');
end case;
vs.
If C = '1' then
Output <= Input1;
elsif B= '1' then
Output <= Input2;
elsif C= '1' then
Output <= Input3;
else
Output <= (others => 'Z');
end if;
NOW the case produces a parallel multiplexer structure that is sensitive for the given code of Selector.
The if-elsif does something different. Whenever C becomes '1' (No matter how unlikely or unneccesary this might be in
your particular design) it switches Input1 to the output. Here we have the always cited priority encoder.
To get the same functionality with a case you have to write a different code:
Selector <= A&B&C; -- I MUST do this for a case statement !
case Selector is
when "--1" => Output <= Input1;
when "-10" => Output <= Input2;
when "100" => Output <= Input2;
when others => Output <= (others => 'Z');
end case;
Now the first >when< hits whenever C becomes '1'.
This code might produce something more "parallel" than the if-elsif, but who knows about the tricks of modern synthesis
tools. :-)
Have a nice synthesis
Eilert
Reply by Neo●May 2, 20052005-05-02
Info,
The code above given by you for onehot did sysnthesize differently. the
"case" version systhesized to 3 LUT's involving only OR gates and didnt
infer priority structure infact it optimized as you have mentioned to a
series of OR gates. But the "if" version systhesiszed to 6 LUT's and
inferred a priority structure. Leonardo was used for the systhesis.
Reply by info_●May 1, 20052005-05-01
I need to add some more information to be absolutely accurate :
> -- -------------------------------------
> Architecture RTL_if of HOTDECOD is
> -- -------------------------------------
> -- and how many LUTS for this one ?
> -- Isn't this description easier to rewrite under
> -- the form of a (size-independant) function ?
> begin
>
> process (A)
> begin
> if A(7) = '1' then Q <= "111";
> elsif A(6) = '1' then Q <= "110";
> elsif A(5) = '1' then Q <= "101";
> elsif A(4) = '1' then Q <= "100";
> elsif A(3) = '1' then Q <= "011";
> elsif A(2) = '1' then Q <= "010";
> elsif A(1) = '1' then Q <= "001";
> elsif A(0) = '1' then Q <= "000";
> else Q <= "---"; -- don't care
> end if;
> end process;
> -- do you see a "priority" in the synthesized result ???
There IS in fact a priority in the description above which is
indeed respected in the synthesized result ! This description
works as a one hot decoder, but it is also a priority encoder
when it handles the overlapping cases.
This is why this solution is still not optimal (as a one-hot decoder)!
(5 LUTs instead of 3).
A really optimized one hot decoder (with all synthesis tools) can
be built with the following function (provided by S. Weir in an
old post) :
----------------------------------------------------------------------------
-- Encode a vector of discrete bits into a binary vector representation,
-- WITHOUT guarding against multiple bits on.
--
-- For a single bit on, the encoded value is the offset from the
-- low index of src of the asserting bit, regardless of src's
-- endianness.
----------------------------------------------------------------------------
function fnEncBin ( src : std_logic_vector ) return std_logic_vector is
variable rslt : std_logic_vector( fnLog2M( src'length ) - 1 downto 0 ) ;
begin
rslt := ( rslt'range => '0' ) ;
for rslt_idx in rslt'range loop
for src_idx in src'length - 1 downto 0 loop
if( ( ( src_idx / ( 2**rslt_idx ) ) MOD 2 ) = 1 ) then
rslt( rslt_idx ) := rslt( rslt_idx ) OR src( src_idx + src'low ) ;
end if ;
end loop ;
end loop ;
return( rslt ) ;
end ;
This function simply infers the OR gates directly !
In the case of this post, the one hot decoder ends up in
three (4 inputs) LUTs, one logic layer instead of 2 (or more).
I don't think there is any other description that infers the
minimal result with all synthesis tools.
Bert Cuzeau
Reply by Mike Treseler●May 1, 20052005-05-01
Hi Bert,
info_ wrote:
> if A(7) = '1' then Q <= "111";
> elsif A(6) = '1' then Q <= "110";
> etc...
> do overlap, but even then, it does not necessarily
> require or imply a priority-based implementation.
I agree. Synthesis is free to make
any netlist that sims like the code.
I have found that coding style has a
negligible effect on utilization
for equivalent descriptions.
My point was that the logical idea
of priority does not apply to all
problems. Some are pieces of pie
and some are Olympic rings.
> I still won't use it, nor RTL procedures, but this is more
> a matter of habits and personal taste. I once suspected that
> you were using both as a hidden signature to copyright your code
> ;-)
Yes. It's a little like changing the
spelling of my name on magazine
subscriptions to see where it goes.
> Cheers
Hey howdy,
-- Mike Treseler
Reply by info_●April 30, 20052005-04-30
Hi Mike,
Mike Treseler wrote:
> "if then else elsif elsif ... "
> only implies priority if cases overlap.
if A(7) = '1' then Q <= "111";
elsif A(6) = '1' then Q <= "110";
etc...
do overlap, but even then, it does not necessarily
require or imply a priority-based implementation.
> I agree that std_ulogic_vector is usually
> more trouble than it's worth.
> However std_ulogic has no such downside.
> I use std_ulogic as my default bit type.
> It port maps directly to std_logic without conversion.
Seems I was a bit too quick to burry std_ulogic...
I still won't use it, nor RTL procedures, but this is more
a matter of habits and personal taste. I once suspected that
you were using both as a hidden signature to copyright your code
;-)
Cheers,
Bert Cuzeau
Reply by Mike Treseler●April 30, 20052005-04-30
info_ wrote:
> They don't quite say this (or they'd lie).
> "If ... elseif" has an implied priority (first test true -> next tests
> are not taken)... but that makes no sense when you describe a truth table !
> (as in my first encoder example)
"case" and "if then else"
have no overlapping cases and therefore no priority.
"if then else elsif elsif ... "
only implies priority if cases overlap.
> I think this is the same with std_ulogic, which use was supposed
> to help detect multiple drivers situations (it did). But there
> are other ways to check this and so many other bad things are
> to be tested at synthesis that everybody has dropped now this
> unresolved type and obsolete style.
I agree that std_ulogic_vector is usually
more trouble than it's worth.
However std_ulogic has no such downside.
I use std_ulogic as my default bit type.
It port maps directly to std_logic without conversion.
-- Mike Treseler
Reply by info_●April 30, 20052005-04-30
Just a few extra comments :
Mohammed A khader wrote:
> All say that case infers a parallel logic and if-elsif-else infers
> a proirity encoder structure ...
They don't quite say this (or they'd lie).
"If ... elseif" has an implied priority (first test true -> next tests
are not taken)... but that makes no sense when you describe a truth table !
(as in my first encoder example)
Whatever the means to describe behaviorally the truth table, it
will end up the same, and by way of logic reduction it will lead
to implementations that may or may not be similar depending
on synthesis tools optimization goals, constraints, internal
logic reduction algorithm, etc...
Things may become different when complex operators are inferred.
The BIG difference is that, from a synthesis perspective,
describing a truth table isn't the same as describing higher
level structures !
What I mean is be wary of general rules about synthesis. There are more
than one tool, and tens of years of research behind them...
I think it's pretty dangerous to say "this does infer that".
Or you have to be damn accurate : given code snippet, given tool,
given version, given technology, given constraints, etc etc
> problem of qulifying expression has been addressed by vhdl committee
> and it is going be fixed soon.
1. I didn't mean it was a problem ! You just need to know the language.
Qualified expressions are just often unknown to newbies, but they
are extremely useful. write (L,"hello"); for example.
There's no problem in writing "case A&B" (a qualified expr does it).
It's also possible to get directly the two MSBs of A + B (vectors).
One just needs to know a bit more than the basics of VHDL.
2. What do you mean by "has been addressed..."? I don't want to start
another controversy, but VHDL200X isn't out of the woods.
We all just hope it will happen (out of IEEE and then into our tools)
before the EDA community majority has switched to SystemVerilog.
But there is absolutely no real need of a better VHDL for simple designs.
> completness of case is an advantage to find the bugs.Lets take an
> example....
It's MUCH easier to follow good coding rules which ensure that the
bad situation you mention will never happen in your design that
trying to test every signal again 'U' !
If you want to do this, be consistent and write a test for every
numeric vector, making sure it doesn't include any 'U'. Not cool.
Simple gross mistakes are easier to prevent than to detect.
I think this is the same with std_ulogic, which use was supposed
to help detect multiple drivers situations (it did). But there
are other ways to check this and so many other bad things are
to be tested at synthesis that everybody has dropped now this
unresolved type and obsolete style.
> I dont know partiuclarly about one hot minimal decoder but
> Synthesis tools say case statments are good for area opptimizations.
It's called hearsay I think.
Then, why doesn't it show up in the simple examples I gave ?
(XST 6.3.03i creates in fact a larger design with the "case" version)
I have reproduced below an old example that I sometimes used
in my courses. I kept it simple minded (no fancy function).
Why don't all synthesis tools give the simple OR gates ?
- The case version "looks" nicer, but check the gates count (with your tool).
- In the "if" solution : do you see a priority after synthesis ??
Chances are your synthesis tool will produce three times larger solution
with the case than with the if.
Don't believe ppt slides (nor me) !
> Dont under estimate the number of gates as it is know
> causing more static power.
0. How do you know you have "saved" some gates and achieved
and an optimal solution ? What do you compare against ?
In the example below, was 15 LUTs acceptable or not ? (without
the comments saying it was not).
1. The case doesn't always produce less gates, sometimes
the contrary as proved here.
I try to not underestimate anything ! I just check by myself
as much as I can and I encourage you, if you're sensitive to gate
count and QOR, to always verify and never simply "assume".
2. If design could be made significantly smaller by avoiding "ifs",
then don't you think the Synthesis tools would automatically do
the translation internally ? (they do, this and many other tricks)
3. There is one or two orders of magnitude higher potential gain
by smart implementation and good design know-how :
being a smart architect pays !
4. Most synthesis tools are smarter and smarter, so a good
understanding of your specific tool is also a good investment.
5. FPGAs and ASICs are two different worlds. What applies to one not
necessarily applies to the other. Synthesizing to LUTs and to GATEs
isn't the same. Don't forget Synopsys is an ASIC company.
Driving DC ins't the same as doing FPGA synthesis.
6. Significant Power consumption reduction requires other techniques
than using case instead of if (supposing that there is any gain
at all doing so).
7. Gate count is definitely less and less an issue, even in the
ASIC field ! Challenges have moved, and old books don't reflect this.
My experience is : Synthesis tools are often smarter than the designer
thinks, but there are also sometimes doing some apparently "stupid"
things (at least looking like such for unexperienced designers).
Just try Q <= A + B - A - B; -- unsigned vectors.
And please do *NOT* avoid the case statement when it is appropriate !!!
Again, HDL code should be as simple, clear and expressive as possible,
easy to understand and maintain, and not error-prone. I assign a higher
priority to these criteria in our designs.
After a good ground learning, learn by trying. This is fun anyway.
You must end up "thinking" like your synthesizer, then you'll be
real efficient.
Bert Cuzeau
"Let's save the poor IF from damnation !" Commitee
12, Gate(s) Lane
Mux Island
:-)
(donations accepted, FFs* and LUTs only)
* FlipFlops, not French Francs
----------------------------
-- HOTDECOD.VHD
-- -------------------------------------
-- One-Hot Decoder
-- -------------------------------------
-- ALSE. http://www.alse-fr.com/english
-- This design must be synthesized as 3 x OR gates (4-inputs)
-- Any extra logic is unnecessary.
--
library IEEE;
use IEEE.std_logic_1164.all;
-- -------------------------------------
Entity HOTDECOD is
-- -------------------------------------
port ( A : in std_logic_vector(0 to 7);
Q : out std_logic_vector(2 downto 0)
);
end;
-- -------------------------------------
Architecture RTL_case of HOTDECOD is
-- -------------------------------------
-- check how many LUTS for this one...
-- How do you code this as a generic size decoder ?
begin
process (A)
begin
case A is
when "00000001" => Q <= "111";
when "00000010" => Q <= "110";
when "00000100" => Q <= "101";
when "00001000" => Q <= "100";
when "00010000" => Q <= "011";
when "00100000" => Q <= "010";
when "01000000" => Q <= "001";
when "10000000" => Q <= "000";
when others => Q <= "---"; -- don't care
end case;
end process;
-- Note : the "case" above is auto-full and auto-parallel in Verilog sense.
-- Quite obviously, one solution is :
-- Q2 = A0 or A1 or A2 or A3
-- Q1 = A2 or A3 or A6 or A7
-- Q0 = A1 or A3 or A5 or A7
end RTL_case;
-- -------------------------------------
Architecture RTL_if of HOTDECOD is
-- -------------------------------------
-- and how many LUTS for this one ?
-- Isn't this description easier to rewrite under
-- the form of a (size-independant) function ?
begin
process (A)
begin
if A(7) = '1' then Q <= "111";
elsif A(6) = '1' then Q <= "110";
elsif A(5) = '1' then Q <= "101";
elsif A(4) = '1' then Q <= "100";
elsif A(3) = '1' then Q <= "011";
elsif A(2) = '1' then Q <= "010";
elsif A(1) = '1' then Q <= "001";
elsif A(0) = '1' then Q <= "000";
else Q <= "---"; -- don't care
end if;
end process;
-- do you see a "priority" in the synthesized result ???
end RTL_if;
Reply by ●April 30, 20052005-04-30
Hi,
it would be interesting to know, if the if-style gets worse compared to
the case-style, when we have a very long if-chain (10x elsif or so).
regards,
Benjamin