Week 13

advertisement
ECE 551
Digital System Design &
Synthesis
Lecture 13:
Life After Verilog
Picking a Language


New lecture – Alternatives to Verilog
Things used to be so simple
 Just VHDL vs Verilog


These days the market for digital design tools and
languages is getting more diverse
Competitors to the entrenched HDLs are becoming
more attractive
 Hardware Verification Languages (HVLs)
 Proprietary Modeling Languages
 High-Level Model-to-Gates Converters
2
Oh, Verilog

Now that you’ve had a some time to get used to
Verilog-2001 you’ve probably noticed a few flaws.

What do you dislike about Verilog?
3
Personality Problems [1]

It’s not always clear what will/won’t synthesize
 While loops, divide/modulo operators, implicit state
machines, complex for loops, etc.

Sometimes it would be nice if you could just say
what you intended, instead of having to imply it.
 Make this always block combinational
 Make this infer a flip-flop / make this infer a latch
 Make this case statement parallel**
 Make this case statement have priority**
** Without relying on “hidden” synthesis pragmas!
4
Personality Problems [2]

It’s easy to get into trouble if you’re not careful
 Inferring latches
 Mixing blocking & non-blocking statements
 Mixing synchronous & asynchronous triggers in sensitivity
lists

Some things are just misleading
 I can call something a “reg” but it’s not necessarily a
register?
5
Personality Problems [3]

Some things are inconvenient
 Why can’t I use multi-dimensional arrays of wires as
inputs & outputs?

Wouldn’t it be nice if it had…
 Language features like TYPEDEF, STRUCT, ENUM
 Better synchonization features for parallel operation in


testbenches, like semaphores and FIFOs
Built-in object-oriented programming for testbenches
Easier integration with C programs
6
Where do you go from here?

“Verilog just needs a tune-up”
 SystemVerilog

“Verilog needs stricter rules”
 VHDL

“Actually, we need something even more abstract”
 BlueSpec

“I need something that is easier to plug in to my
high-level system model”
 SystemC
7
SystemVerilog – The ‘new’ Verilog


New IEEE Standard introduced in 2005
“Successor” to Verilog – merged into Verilog-2009


Mostly the same syntax/semantics
Backwards compatible with most Verilog code


A lot of new features for Verification (testbenches)
Tweaks to existing Verilog constructs to make them
easier to use
New datatypes

8
SystemVerilog – Datatypes [1]

New variable types: LOGIC, BIT

Logic can be used
 As left side in Behavioral block
 As left side in continuous assignment
 As output of a structural module



No more worrying about declaring wire vs reg!
Less confusing – the name doesn’t trick people into
thinking it is always a register, like reg does.
bit – Like logic, but cannot have x or z values
9
SystemVerilog – Datatypes [2]

TYPEDEF – C-style defined data types
typedef pixel logic[23:0];
pixel a_pixel;
pixel[1920:1][1080:1] my_display;

Why not just do this with a macro?


Typedefs do not have the global scope problem of
macros.
Typedefs can be evaluated during Syntax Check rather
than during Elaboration, so they are easier to debug.
10
SystemVerilog – Datatypes [3]

ENUM – Enumerated types
typedef enum logic[2:0] {RESET, COMPUTE1,
COMPUTE2, SEND, ERROR} states_t;
states_t current_state, next_state;
…
if(current_state == COMPUTE1)
next_state = COMPUTE2;
…
11
SystemVerilog – Datatypes [4]

Structures & Unions
typedef struct packed {
logic [2:0] packet_type;
logic [3:0] dest_addr;
logic [47:0] data_payload;
} router_packet;
case(my_packet.packet_type)
DATA3:
if(my_packet.dest_addr == MY_ADDR)
…
…
12
SystemVerilog – Always Constructs [1]

Always Blocks – can specify what you’re trying to
do!

Combinational Block:
always_comb begin //Can omit sensitivity list!
a = b;
//Tool adds it like @(*)
c = ~d;
end
13
SystemVerilog – Always Constructs [2]
always_ff@(posedge clk) //still need sensitivity list
state <= next_state; //for flip-flops. Why?
always_latch
if(en)
d_out <= d_in;
This can’t “force” the synthesis tool to synthesize flip
flops if you describe latch-like behavior
However, it can warn you that what you described
didn’t match what you said you wanted!
14
SystemVerilog – Control Constructs[1]





Case and If/Else have priority by default in Verilog.
To match this behavior in synthesis, we need a
cascade of muxes.
Designers commonly use synopsys parallel_case to
force the synthesizer to make a single mux instead.
Synthesizer pragmas give different information to
the synthesis tool than to the simulator.
This is fundamentally bad. We want synthesis tool
and simulator to have the same information!
15
SystemVerilog – Control Constructs[2]
 unique keyword modifier
unique case (sel)
CASE1: …
CASE2: …
CASE3: …
endcase


unique tells the synthesizer and simulator that one,
and only one, case will be selected – ignore priority
so you can synthesis a single parallel mux!
Also works with if: unique if(…) …
16
SystemVerilog – Control Constructs[3]
 priority keyword modifier
priority case (sel)
CASE1: …
CASE2: …
endcase



priority tells the synthesizer and simulator that at
least one of the cases will always match. If this
doesn’t happen in simulation it will warn you.
Also works with if: priority if(…) …
Easy way to avoid accidental latches!
17
SystemVerilog – Interfaces

Can use multi-dimensional arrays in I/O
module(input [1:0] a[9:0][3:0], output b);

Can define interfaces separately from modules
 Allow an interface to be re-used in multiple modules
interface intf;
logic a, b;
modport in (input a, output b);
modport out (input b, output a);
endinterface
18
SystemVerilog – Verification



Most of SystemVerilog’s new features are in nonsynthesizable constructs for testbenches
SystemVerilog is sometimes referred to as an “HVL”
– Hardware Verification Language.
New verification features include
 Dynamic & associative arrays, Classes
 FIFOs, semaphores, mailboxes
 Assertions and time sequencing
 Built-in support for computing test coverage
 Extended support for easy generation of random test
vectors
19
SystemVerilog – Current State




SystemVerilog is already widely used in industry for
designing testbenches
So why, you ask, did we learn Verilog-2001 instead
of SystemVerilog/Verilog-2009?
Unfortunately, many companies are using not yet
adopting SystemVerilog for synthesis
Main problem: Tool Support is Lacking
 Ex: Xilinx ISE still does not support SystemVerilog at all.
 Most companies cannot migrate until their full toolchain
supports the language
20
VHDL – Introduction

VHDL was introduced in the 1980s as a true
Hardware Description Language
 Meant to describe existing circuits
 Not originally meant as a design tool

As a result, VHDL is very verbose and stronglytyped; this is both its greatest strength and
weakness

Because it has been around so long, VHDL still
enjoys great support from design tools.
21
VHDL – Compared to Verilog

VHDL and Verilog are not radically different.

Both have 3 levels of abstraction
 Structural / Gate
 RTL
 Behavioral



Both have built-in parallelism and the concept of
sensitivity lists
Both have some constructs that synthesize and
others that are meant for verification
Both have If/Else, Case, For, Generate
22
VHDL – Sample Code
library ieee;
use ieee.std_logic_1164.all;
ENTITY compare8 IS
PORT( x, y: ;IN std_logic_vector(7 DOWNTO 0) ;
res: ;OUT std_logic );
END compare8;
ARCHITECTURE struct OF compare8 IS BEGIN
res <= '1' WHEN (x = y) ELSE '0';
END struct;
23
Verilog – Same Functionality
module compare8(input [7:0] x, y, output res);
assign res = (x == y);
endmodule



VHDL is a lot more verbose than Verilog in general.
Some people like this, because it forces them to
more explicitly describe each circuit
Like Verilog-2009, VHDL allows (actually requires)
you to separately define your interfaces.
VHDL also allows you to form packages and
libraries to help design re-use.
24
VHDL – Strict Typing


Variables in VHDL are strictly typed
Verilog is more loosely typed

For example, in Verilog you can
 Assign a value from a variable to an array of wires
 Assign the value of a 4-bit wire to a 6-bit wire (Verilog


will automatically extend the value)
Assign the value of a 6-bit wire to a 4-bit wire (Verilog
will automatically truncate the value).
VHDL requires you to explicitly convert variables to
the same type or it will give an error.
25
VHDL – Conclusion



Second most popular HDL after Verilog
Very good tool support (maybe better than Verilog)
Stylistically, some people may prefer a more
verbose, strongly-typed language

Most comparisons have shown that people are
more productive with Verilog.
Verilog is better for verification, since it is easier to
simulate and has better verification constructs.
VHDL is still trying to catch up to some of the
additions to Verilog-2009/SystemVerilog


26
Bluespec – Introduction

Extension to SystemVerilog – Bluespec
SystemVerilog (BSV) – but uses a much different
design paradigm than Verilog

Bluespec uses its own compiler to convert from
Bluespec SV to an RTL Verilog description
Goal: Create increased abstraction between design
description and implementation
Key Concept: Let the Bluespec compiler do the
heavy lifting rather than the synthesis tool!
What are the pros/cons of a proprietary language
and compiler?



27
Bluespec – Proprietary Compiler



Have very smart people design a good compiler
that automates common HDL design tasks
Let average** people take advantage of the
automation and save time
This can be better than including it as part of the
Verilog synthesis standard because
 Don’t have to wait for tool vendor support (years!)
 Don’t have to wait for IEEE standardization (years!)
 Can be responsive to current trends in design
**Engineering average
28
Bluespec – Major Concepts


Originally created as an internal language to make
router/switch design more efficient.
Idea: Often we spend a long time coding
interfaces.
 Particularly true as more designs need to support USB,




eSATA, BaseX Ethernet, etc.
Add features geared towards making it easier to design
interfaces
Add in the concept of atomic operations to handle
resource contention -> No race conditions
Automate FIFO queues, ready/enable signals
Force compliance so the designer can’t screw up.
29
Bluespec – Sample Code [1]


Describing a simple Counter in Bluespec
Bluespec uses a functional interface to modules.
 Call module functions rather than connecting to ports
 Keep in mind that the Bluespec compiler will convert this
all to standard Verilog in the end
interface Counter;
method Bit#(8) read();
method Action load(Bit#(8) newval);
method Action increment();
endinterface
30
Bluespec – Sample Code [2]
(* synthesize *)
module mkCounter(Counter);
Reg#(Bit#(8)) value <- mkReg(0);
method Bit#(8) read();
return value;
endmethod
method Action load(Bit#(8) newval);
value <= newval;
endmethod
method Action increment();
value <= value + 1;
endmethod
endmodule //What’s missing that you would expect?
31
Bluespec – Abstraction




Things like clock and reset signals are implicit in
Bluespec
Although no clock was ever mentioned, creating a
register implies a clock. The compiler takes care of
making sure that the counter only increments on
clock cycles.
Muxes – such as the one needed to do Load vs
Increment – are also hidden from the user.
Bluespec methods are designed to always be
atomic; you can try to load and increment in the
same clock cycle. The compiler will set up the
necessary priority logic.
32
Bluespec – Future




Is Bluespec a challenger to Verilog? A supplement?
I’m not really sure yet.
Some pro designers have told me that Bluespec is
great. Others have different opinions…
It provides a much different type of abstraction
that may be a lot easier for programmers without
hardware experience to use
One big hurdle: It isn’t free. You must license the
Bluespec compiler if you want to use it at all.
33
SystemC – Introduction







SystemC is not really a proper HDL
It is a framework for the C language that allows
you to model hardware-like behavior
Advantage – Makes it easy to model a full system;
can simulate other elements in C.
Can test software algorithms in C side-by-side with
their hardware models
Can be used to verify HDL code efficiently
Disadvantage – Not designed to be synthesized.
Some companies are working on synthesis of
SystemC, but not ‘mainstream’ quite yet.
34
SystemC – Implementation

Fundamentally, SystemC is just a library you can
use in C.

How does a software implementation in C need to
be extended to properly model cycle-accurate
hardware?
 Concurrency, race conditions, delays, timing


requirements, event-driven updates/edges
Resolving multiple drivers, 4-value logic
It must be possible since this is what the simulators
are doing internally!
35
SystemC – Sample Code [1]
#include "systemc.h"
SC_MODULE (first_counter) {
//counter design example
sc_in_clk clock ;
//module I/O interface
sc_in<bool> reset , enable;
sc_out<sc_uint<4>> counter_out;
sc_uint<4> count;
void incr_count () {
//method to increment count
if (reset.read() == 0) { count = 0; counter_out.write(count);}
else if (enable.read() == 1) {
count = count + 1;
counter_out.write(count);
cout<<"@" << sc_time_stamp() <<" :: Incremented Counter "
<<counter_out.read()<<endl;
}
36
SystemC – Sample Code [2]
//counter example continued
SC_CTOR(first_counter) {
//constructor for counter class
cout<<"Executing new"<<endl;
SC_METHOD(incr_count);
sensitive << reset.neg();
//add to sensitivity list
sensitive << clock.pos();
}
}
37
SystemC – Current State

SystemC is getting pretty popular as an
intermediate hardware modeling step.

It provides a bridge between algorithm
development (in C) and implementation (in HDLs)

SystemC synthesis tools are beginning to become
mainstream.
 MentorGraphics shipped Catapult C in 2010.
 Smaller companies like Forte, Bluespec, SystemCrafter
 OSCI is working on a synthesis standard
38
Other Projects


The idea of skipping HDL authoring altogether is
gaining steam among some
Many recent products that attempt to convert from
C, MatLab, and Dataflow Graphs

The jury is still out on whether these tools can
achieve similar levels of performance, power, and
area efficiency as direct HDL implementation.

But in some cases, improved productivity is more
important than an efficient hardware design.
39
Wrapping Up
40
Digital Design & Synthesis

The goal of this course is to give you most of the
skills needed to take a design from an idea to a
product

At this point you should be capable of handling
many different phases of the design project while
working solo and in a team

These are real job skills! Put them on your CV and
talk about them in job interviews.
41
42
Planning

What platform should I use for this project?
 Standard Cell, Custom ASIC, MPGA
 FPGA, PLD
 ROM
 Software?

What HDL / tools should I use?
 Verilog, VHDL, SystemVerilog
 Bluespec, SystemC
43
Design Entry

How should I architect this design?
 Hierarchy
 Abstraction level (structural, RTL, behavioral)

How do I describe specific types of logic?
 Datapaths
 Control logic
 State Machines

How can I design for re-use?
 Parameterization
 Generated Instantiation
44
Functional Test

Architecting Testbenches
 File-based testing
 Self-checking / Automated testbenches

Designing Test Patterns
 Exhaustive Testing vs Non-Exhaustive Testing
 Selecting tests and writing a test plan
 Identifying corner cases

Debugging a Design
 Reading waveforms
 Identifying problems
45
Synthesis

Setting up Design Rules and Constraints to meet
my design goals
 Timing, Area

How do I read synthesis reports?
 Identifying when constraints are met / un-met
 Identifying the critical path, finding slack time

How do I control synthesis optimizations?
 Top-down vs Bottom-up
 Logic restructuring, ungrouping, effort levels
46
Post-Synthesis Validation

What sort of potential problems should we
anticipate?
 Inferring Latches
 Multiple Drivers
 Timing differences
 How do we fix these problems?

What should we look for on a post-synthesis
waveform?
 X’s, Z’s
 Timing delays
47
Download