Confusion between Behavioural and Dataflow model Programs in VHDL

Question

I'm using the textbook "VHDL: Programming By Example" by Douglas L Perry, Fourth Edition. He gave an example of the Dataflow programming model in page 4:

Code I:

ENTITY mux IS
PORT ( a, b, c, d : IN BIT;
s0, s1 : IN BIT;
x, : OUT BIT);
END mux;
ARCHITECTURE dataflow OF mux IS
SIGNAL select : INTEGER;
BEGIN
select <= 0 WHEN s0 = ‘0’ AND s1 = ‘0’ ELSE
          1 WHEN s0 = ‘1’ AND s1 = ‘0’ ELSE
          2 WHEN s0 = ‘0’ AND s1 = ‘1’ ELSE
          3;
x <= a AFTER 0.5 NS WHEN select = 0 ELSE
     b AFTER 0.5 NS WHEN select = 1 ELSE
     c AFTER 0.5 NS WHEN select = 2 ELSE
     d AFTER 0.5 NS;
END dataflow;

Now in page 17, Code II

LIBRARY IEEE;
USE IEEE.std_logic_1164.ALL;
ENTITY mux4 IS
PORT ( i0, i1, i2, i3, a, b : IN std_logic;
PORT ( i0, i1, i2, i3, a, q : OUT std_logic);
END mux4;
ARCHITECTURE mux4 OF mux4 IS
SIGNAL sel: INTEGER;
BEGIN
WITH sel SELECT
q <= i0 AFTER 10 ns WHEN 0,
q <= i1 AFTER 10 ns WHEN 1,
q <= i2 AFTER 10 ns WHEN 2,
q <= i3 AFTER 10 ns WHEN 3,
q <= ‘X’ AFTER 10 ns WHEN OTHERS;
sel <= 0 WHEN a = ‘0’ AND b = ‘0’ ELSE
       1 WHEN a = ‘1’ AND b = ‘0’ ELSE
       2 WHEN a = ‘0’ AND b = ‘1’ ELSE
       3 WHEN a = ‘1’ AND b = ‘1’ ELSE
       4;
END mux4;

This is supposed to be a behavioural model, as per the same textbook. Aside from the differences in variable name, the only major difference I see here is that there is an extra statement

WITH sel SELECT

in the second case, and small syntax differences. This Code II is concurrent. But from other sources in the internet(listed below), I've seen that a behavioural model is supposed to be sequential. Which one should I believe?

Now from some other sources of the internet, the definition of these models are as follows:

Behavioral – Circuit is described as an i/o relationship using sequential statements inside a process.

Dataflow – Circuit is described using concurrent statements

-San Jose State university

Behavioral – describes how the output is derived from the inputs using structured statements.

Dataflow – describes how the data flows.

-University of Akron College of Engineering

Here I do not understand what structured statements mean.

in Behaviour level, process keyword is present

in dataflow level , concurrent statement (<=) is present

---

This was seen in an online forum.

Is process statement compulsory for Behavioural model?

What is the actual difference between codes I and II? According to the author, they have different models, dataflow and behavioural. I cannot see how this is possible. What should I believe?

Lastly, in Perry D L, Page 45, 46:

LIBRARY IEEE;
USE IEEE.std_logic_1164ALL;
ENTITY mux IS
PORT (i0, i1, i2, i3, a, b : IN std_logic;
PORT (q : OUT std_logic);
END mux;
ARCHITECTURE better OF mux IS
BEGIN
PROCESS ( i0, i1, i2, i3, a, b )
VARIABLE muxval : INTEGER;
BEGIN
muxval := 0;
IF (a = ‘1’) THEN
muxval := muxval + 1;
END IF;
IF (b = ‘1’) THEN
muxval := muxval + 2;
END IF;
CASE muxval IS
WHEN 0 =>
q <= I0 AFTER 10 ns;
WHEN 1 =>
q <= I1 AFTER 10 ns;
WHEN 2 =>
q <= I2 AFTER 10 ns;
WHEN 3 =>
q <= I3 AFTER 10 ns;
WHEN OTHERS =>
NULL;
END CASE;
END PROCESS;
END better;

This is a sequential version of MUX. According to the other definitions, this is supposed to be behavioural, but the author does not state so. Could you clear up my confusion regarding these models?

Answer 1

Don't look for a mathematically rigorous description of these terms; they are a lot vaguer than that, loose classifications that can overlap.

"Dataflow" I think is fairly clear here; it DOES describe the flow of data, and it describes it in terms of concurrent statements. But I would add that each concurrent statement is woken by changes on its inputs and delivers its outputs; therefore (the important bit:) there is no correspondence between the order of things happening and the order of elements in the source code. In that respect it has a lot in common with functional programming. And both the first two models are dataflow; in (I) the elements are in logical order while (II) is not.

"Behavioural" SHOULD be fairly clear too - it simply describes a circuit in terms of its behaviour.

But it is not in general opposed to dataflow - though your San Jose quote is somewhat correct - behavioural descriptions are commonly sequential simply because the sequential paradigm (inside a VHDL process) is common and familiar to programmers. Even so, the behaviour of several such processes interacting with each other is ... dataflow.

Behavioral then is NOT correctly opposed to dataflow. It is more correctly opposed to RTL (Register Transfer Level) and structural which have fairly clear meanings.

A structural description consists of a number of building block (gates, multiplexers, entire CPUs) and the signals interconnecting them : a textual block diagram (perhaps auto-generated from a graphical one). As such it can be either the lowest level (see frequent questions here about making an adder out of gates!) or the highest level (connecting CPU to memory, peripherals, etc).

An RTL description is fairly low level; it describes the transfer and operations on data between storage elements (registers) and is common inside a process; it is rather like an assembly language listing from a (behavioural) C program.

Lastly - too many descriptions and too many extraneous details get in the way of doing a proper design job. Look at the task in hand, extract its essence, and implement that.

A multiplexer selects one of a collection of input elements according to the index of the element you want. The most natural form of index is usually an integer type, rarely including negative indices, and the most natural form of collection in VHDL is ... an array.

So why not write

ENTITY mux IS
  PORT ( a, b, c, d : in BIT;
         sel        : in natural range 0 to 3;
         x          : out BIT);
END mux;
ARCHITECTURE simple OF mux IS
SIGNAL values : array (0 to 3) of BIT;
BEGIN
   values <= a & b & c & d;
   x      <= values(sel);   -- after 0.5 ns; if you need to model timing!
END simple;

or better, make "values" an input port...

Answer 2

There's a distinction between behavioral and structural implementations that isn't well explained in this particular book. Dataflow can be a poorly applied description to hardware based on who is trying to convey what.

I'd suspect this: Switching & Logic Laboratory Spring 2008 jeg 1 Lab 2 – Behavioral and Dataflow VHDL (PDF, 66KB, 12 pages) was provided to overcome confusion of those readers of the book like yourself:

Within VHDL we can describe the logic in three different manners. These three different architectures are:

• Behavioral – describes how the output is derived from the inputs using structured statements.

• Dataflow – describes how the data flows from the inputs to the output most often using NOT, AND and OR operations.

• Structural – describes how gates are interconnected similar to schematic approach.

Dataflow can imply concurrency, while neither behavioral or structural preclude either sequential or concurrent descriptions. All concurrent descriptions are converted to sequential processes for VHDL simulation, counting on delta simulation cycles to synchronize signal assignments emulating concurrency.

There's this concept of a design network in the simulator that reflects the design hierarchy like a schematic can be represented by a flat net list.

There are good reasons to use what's called Dataflow here, in that describing logic behavior with say 9 level logic (MVL9, used by package std_logic_1164) results in the propagation of unknowns ('X's) and uniintialized values ('U's). It brings closure between behavioral and structural models without hand waving around visible values you can't reconcile between the levels of abstraction, other than by exception. The stimulus and expected results are the same between Dataflow behavioral description and a structural implementation.

Speaking from experience, it's a lot easier up front to get the different levels of abstraction to match than to hear back from a foundry wanting to delay first silicon waiting on you to approve exceptions to the stimulus you provided, plus at least historically, the number of exceptions you could make on an IC tester was rather limited. You could think of real silicon on an IC tester as another level of abstraction.

For simulation different levels of abstraction in VHDL you basically write a behavioral model that behaves more closely to what a structural model would. A

I found the above PDF by googling for 'VHDL dataflow', which reveals many more sources of information.

Answer 3

Behavioral – describes how the output is derived from the inputs using structured statements.

Dataflow – describes how the data flows.

-University of Akron College of Engineering

A dataflow model requires that you have a clear understanding of the dataflow(i.e. the physical circuit). However, when using a behavioral model, you only need to pay attention to the main behavior of the design. So a behavioral model is easier to understand and maintain. For example, to implement a parallel multiplier

dataflow model

You will need components like registers, FAUs, multiplexers, etc. And you are supposed to implement them all by yourself. That's terrible especilally when the input numbers are 16-bit or more.

behavioral model

All you need is to write down a statement like this:

p <= a * b;

Whether a model is a dataflow one or a behavioral one is determined by how you model the design. Sequential statements or concurrent statements? That doesn't matter.