Design Notation VHDL or Verilog?

Glen,

The point of optimizing HW arithmetic to the task at hand is that
avoiding excess precision saves resources (= $).

In SW, running on a majority of platforms that support double
precision floating point (already paid for whether it is used or not)
there is no incentive to optimize, so few computer programming
languages support it. Many older programming languages, conceived
before HW floating point support was ubiquitous, supported fixed point
arithmetic because the SW could be optimized to provide only the
precision required, thus consuming only the CPU cycles required.

CPUs have to be optimized for the universe of applications that may be
run on them, so the flexibility of floating point is worth the cost,
especially because only one or two floating point ALUs are typically
required by a CPU core.

Most ASIC/FPGA designs are much more task specific, and do not need
the flexibility of floating point, especially when considering how
many arithmetic units (multiply/add) may be employed by a typical
design, and thus how many times one would have to pay for the
resources consumed by that unneeded flexibility.

Sure, you can design fixed point arithmetic HW in verilog, but you
have to keep track of the binary point yourself, with no help from the
compiler. These VHDL packages allow you to define the width and binary
point at critical stages, and the compiler takes care of the rest.

Most synthesis tools support retiming, which can be used to pipeline
arithmetic operators over a limited number of clock cycles by simply
inserting empty clock cycles in RTL before or after the operator, and
letting the retiming spread out the operator logic into those clock
cycles. With most modern FPGA families, multiplies and multiply/adds
are implemented using available DSP blocks anyway, so intra-operator
pipelining is not often needed uness you have a really wide data path.
And of course, just like verilog, you are in charge of pipelining
multiple operations by the way you describe the behavior (number of
clock cycles it takes to get from input to output). Note that I use
the term clock cycles as opposed to registers. It's an abstraction...

And yes, both the floating and fixed point packages are completely
synthesizable.

Andy

Andy <jonesandy@comcast.net> writes:

> there is negligible advantage to using VHDL over verilog. It is at
> higher levels of abstraction, where design productivity is maximized,
> that VHDL shines.

It depends of the Verilog version in question and the type of design.
SystemVerilog has a much higher level of abstraction when it comes to
verification and testbench code. Especially when using libraries like
UVM etc.

//Petter

-- 
.sig removed by request. 

On Jan 30, 6:16=A0pm, glen herrmannsfeldt <g...@ugcs.caltech.edu> wrote:
> rickman <gnu...@gmail.com> wrote:
>
> (snip, someone wrote)
>
> >> > With the 2008 fixed point package, VHDL allows synthesizable fixed
> >> > point arithmetic of arbitrary precision and binary point location.
>
> (snip, I wrote)
>
> >> An interesting idea, but note that, as far as I know, no computer
> >> hardware has support for this, so it seems strange that an HDL
> >> would need it.
> > Whaaaatttt??? =A0What does your computer capabilities have to do with
> > your ASIC/FPGA designs? =A0Mixed precision/radix point arithmetic can b=
e
> > very useful in an HDL design. =A0If you are just passing data and
> > controlling timing then maybe not. =A0But signal processing often has t=
o
> > deal with the results of varying precision in operations.
>
> If it is useful in ASIC/FPGA designs, wouldn't it also be even more
> useful in software running on commonly (or not so commonly) available
> processors? Yet none have bothered to implement it.
>
> It could be done in software, but none of the popular languages
> have any support for fixed point non-integer arithmetic.
>
> It was one of the fun things I did in PL/I so many years ago, though.

I don't really follow your reasoning here.  The idea of the fixed
point package is to allow a user to work with fixed point values in
VHDL without having to manually deal with all the details.  You seem
to be saying that fixed point arithmetic is of no value because it is
not implemented in software.  I don't see how software has anything to
do with it.  Software is implemented on programmable hardware and for
the most part is designed for the most common platforms and does not
do a good job of utilizing unusual hardware effectively.  I don't see
how fixed point arithmetic would be of any value on conventional
integer oriented hardware.


> >> Yes you can do fixed point with different positions of the radix
> >> point, PL/I and a small number of other languages do, but on hardware
> >> that doesn't keep track of the radix point.
> > You seem to be thinking only of CPU designs. =A0HDL is used for a lot
> > more than CPU designs.
>
> Well, more than CPU designs, I work on systolic arrays, but even
> there so, it doesn't seem likely to help much. Will it generate
> pipelined arithmetic units? Of all the things that I have to think
> about in logic design, the position of the binary point is one of
> the smallest.

Asking about pipelining with fixed point numbers is like asking if
boxes help your car.  Sure, if you want to carry oranges in a car you
can use boxes, but you don't have to.  If your systolic arrays use
fixed point arithmetic then the fixed point package will help your
systolic arrays.  If you don't care about the precision of your
calculations then don't bother using the fixed point package.


> >> For multiply, you do need to keep all the product digits, so that
> >> you can select the right ones based on the position of the product
> >> radix point.
> > Actually you need to keep the bits that are important to your
> > application. =A0That will vary.
>
> And I don't expect the processor to help me figure out which
> ones those are.

What processor?  You have lost me here.  A multiply provides a result
with as many bits as the sum of the bits in the inputs.  But this many
bits are seldom needed.  The fixed point package makes it a little
easier to work with the variety of formats that typically are used in
a case like this.  No, the package does not figure out what you want
to do.  You have to know that, but it does help hide some of the
details.


> >> > Similarly, VHDL also has a floating point package that handles
> >> > arbitrary sizes of data and exponent.
>
> >> And that is synthesizable by anyone?
> > I haven't looked at the floating point package, but if it is described
> > in terms of basic operators in VHDL it should be synthesizable in
> > every tool. =A0That is the point of overloading. =A0You can define a
> > multiply operator for real data in terms of the basic building blocks
> > and this is now a part of the language. =A0VHDL is extensible that way.
>
> Again, does it synthesize pipelined arithmetic units? If not, then
> they aren't much use to anything I would work on. If I do it in
> an FPGA that is because normal processors aren't fast enough.
>
> But the big problem with floating point is that it takes too much logic.
> The barrel shifter for pre/post normalization for an adder is huge.
> (The actual addition, not so huge.)

Nope, the fixed point package does not design pipelines, it isn't a
pipeline package.  Yup, the barrel shifter is as large as a
multiplier.  So what is your point?  The purpose of the package is to
allow you to design floating point arithmetic without designing all
the details of the logic.  It isn't going to reduce the size of the
logic.


> >> I suppose higher level of abstraction is nice, but does it really
> >> help design real systems?
> > Not if you refuse to consider it.
>
> (snip)
>
> > You seem to be very CPU focused. =A0For every Pentium type design done
> > in HDL there are what, thousands of other designs?
>
> Well, it was supposed to be an example....
>
> -- glen

An example of what?  The fact that you can't find the utility of the
fixed point package does not mean it is not useful.  I really don't
see your points.

Rick

On Jan 31, 2:00=A0pm, Petter Gustad <newsmailco...@gustad.com> wrote:
> Andy <jonesa...@comcast.net> writes:
> > there is negligible advantage to using VHDL over verilog. It is at
> > higher levels of abstraction, where design productivity is maximized,
> > that VHDL shines.
>
> It depends of the Verilog version in question and the type of design.
> SystemVerilog has a much higher level of abstraction when it comes to
> verification and testbench code. Especially when using libraries like
> UVM etc.
>
> //Petter
>
> --
> .sig removed by request.

AFAIK, the synthesizable subset of SV is pretty much plane old
verilog, with all its warts.

For verification, SV is indeed really nice. However, a new open source
VHDL Verification Methodology (VVM) library of VHDL packages has
emerged that provides VHDL with some of the capabilities of SV with
OVM/UVM.

Andy

rickman <gnuarm@gmail.com> wrote:

(snip, I wrote)
>> If it is useful in ASIC/FPGA designs, wouldn't it also be even more
>> useful in software running on commonly (or not so commonly) available
>> processors? Yet none have bothered to implement it.

>> It could be done in software, but none of the popular languages
>> have any support for fixed point non-integer arithmetic.

>> It was one of the fun things I did in PL/I so many years ago, though.

> I don't really follow your reasoning here.  The idea of the fixed
> point package is to allow a user to work with fixed point values in
> VHDL without having to manually deal with all the details.  You seem
> to be saying that fixed point arithmetic is of no value because it is
> not implemented in software.  I don't see how software has anything to
> do with it.  

The suggestion, which could be wrong, was that if scaled fixed point
is useful in an HDL, it should also be useful in non-HDL applications.

The designers of PL/I believed that it would be, but designers of
other languages don't seem to have believed in it enough to include.
Personally, I liked PL/I 40 years ago, and believe that it should
have done better than it did, and scaled fixed point was a feature
that I liked to use.
 
> Software is implemented on programmable hardware and for
> the most part is designed for the most common platforms and does not
> do a good job of utilizing unusual hardware effectively.  I don't see
> how fixed point arithmetic would be of any value on conventional
> integer oriented hardware.

Well, the designers of conventional hardware might argue that they
support scaled fixed point as long as you keep track of the radix
point yourself. It is, then, up to software to make it easier for
programmers by helping them keep track of the radix point.

I believe that in addition to PL/I that there are some other less
commonly used languages, maybe ADA, that support it.

It can be done in languages like Pascal and C, TeX and Metafont do it.
Multiplication is complicated, though, because most HLL's don't supply
the high half of the product in multiply, or allow a double length
dividend in division. Knuth wrote the Pascal routines for TeX and MF
to do it, and suggests that implementations rewrite them in assembler
for higher performance. 

>> >> Yes you can do fixed point with different positions of the radix
>> >> point, PL/I and a small number of other languages do, but on hardware
>> >> that doesn't keep track of the radix point.
>> > You seem to be thinking only of CPU designs. &#4294967295;HDL is used for a lot
>> > more than CPU designs.

>> Well, more than CPU designs, I work on systolic arrays, but even
>> there so, it doesn't seem likely to help much. Will it generate
>> pipelined arithmetic units? Of all the things that I have to think
>> about in logic design, the position of the binary point is one of
>> the smallest.

> Asking about pipelining with fixed point numbers is like asking if
> boxes help your car.  Sure, if you want to carry oranges in a car you
> can use boxes, but you don't have to.  If your systolic arrays use
> fixed point arithmetic then the fixed point package will help your
> systolic arrays.  If you don't care about the precision of your
> calculations then don't bother using the fixed point package.

In the past, the tools would not synthesize division with a
non-constant divisor. If you generate the division logic yourself,
you can add in any pipelining needed. (I did one once, and not for
a systolic array.) With the hardware block multipliers in many FPGAs,
it may not be necessary to generate pipelined multipliers, but many
will want pipelined divide.

(snip)
>> And I don't expect the processor to help me figure out which
>> ones those are.

> What processor?  You have lost me here.  A multiply provides a result
> with as many bits as the sum of the bits in the inputs.  But this many
> bits are seldom needed.  The fixed point package makes it a little
> easier to work with the variety of formats that typically are used in
> a case like this.  No, the package does not figure out what you want
> to do.  You have to know that, but it does help hide some of the
> details.

Yes, multiply generates a wide product, and most processors with
a multiply instruction supply all the bits. But most HLLs don't
provide a way to get those bits. For scaled fixed point, you 
shift as appropriate to get the needed product bits out.

>> >> > Similarly, VHDL also has a floating point package that handles
>> >> > arbitrary sizes of data and exponent.

>> >> And that is synthesizable by anyone?

>> > I haven't looked at the floating point package, but if it is described
>> > in terms of basic operators in VHDL it should be synthesizable in
>> > every tool. &#4294967295;That is the point of overloading. &#4294967295;You can define a
>> > multiply operator for real data in terms of the basic building blocks
>> > and this is now a part of the language. &#4294967295;VHDL is extensible that way.

>> Again, does it synthesize pipelined arithmetic units? If not, then
>> they aren't much use to anything I would work on. If I do it in
>> an FPGA that is because normal processors aren't fast enough.

>> But the big problem with floating point is that it takes too much logic.
>> The barrel shifter for pre/post normalization for an adder is huge.
>> (The actual addition, not so huge.)

> Nope, the fixed point package does not design pipelines, it isn't a
> pipeline package.  Yup, the barrel shifter is as large as a
> multiplier.  So what is your point?  The purpose of the package is to
> allow you to design floating point arithmetic without designing all
> the details of the logic.  It isn't going to reduce the size of the
> logic.

The point I didn't make very well was that floating point is still
not very usable in FPGA designs, as it is still too big. If a 
design can be pipelined, then it has a chance to be fast enough
to be useful.

Historically, FPGA based hardware to accelerate scientific
programming has not done very well in the marketplace. People
keep trying, though, and I would like to see someone succeed.

-- glen

In article <5a603ea6-876d-4f20-a35f-79dcfe6ebd1e@h3g2000yqe.googlegroups.com>,
Andy  <jonesandy@comcast.net> wrote:
>On Jan 31, 2:00&#4294967295;pm, Petter Gustad <newsmailco...@gustad.com> wrote:
>> Andy <jonesa...@comcast.net> writes:
>> > there is negligible advantage to using VHDL over verilog. It is at
>> > higher levels of abstraction, where design productivity is maximized,
>> > that VHDL shines.
>>
>> It depends of the Verilog version in question and the type of design.
>> SystemVerilog has a much higher level of abstraction when it comes to
>> verification and testbench code. Especially when using libraries like
>> UVM etc.
>>
>> //Petter
>>
>> --
>> .sig removed by request.
>
>AFAIK, the synthesizable subset of SV is pretty much plane old
>verilog, with all its warts.

Nope.  SV synthesis is quite powerful.  Arrays, and structs are first class 
citizens in SV and can be members of ports.  Further most tools handle 
interfaces now pretty well in synthesis.  There's quite a benefit to using
these in your RTL.  

Trying to resist making this a language war, but I see little advantage 
one or the other with regard to a "higher level of abstraction".  You can 
probably achieve the same with either language.

--Mark

On 1.2.2012 17:20, glen herrmannsfeldt wrote:

> The suggestion, which could be wrong, was that if scaled fixed point
> is useful in an HDL, it should also be useful in non-HDL applications.

It is very useful in some applications, for example in DSP processing.
There are DSP processors that support fixed point arithmetic. Also many
"fixed" FPGA based DSP algorithms use different resolution fixed point
numbers all around. Standard package to help that kind of design is
very welcome. Previously they have been proprietary packages.

> The point I didn't make very well was that floating point is still
> not very usable in FPGA designs, as it is still too big. If a
> design can be pipelined, then it has a chance to be fast enough
> to be useful.

Floating point might be too heavy in many applications, but on the
other hand if the number range and operators are limited to what is
needed it is not that big anymore, when comparing to the leading edge
FPGA sizes. For example leading edge FPGA can support ~2000 single
precision floating point multipliers.

> Historically, FPGA based hardware to accelerate scientific
> programming has not done very well in the marketplace. People
> keep trying, though, and I would like to see someone succeed.

You are fixed to quite narrow band of FPGA applications. FPGAs
are almost everywhere, for example they are very popular in mobile
network basestations. And also telecom equipment use some forms of
floating/fixed point arithmetic often in policing and shaping
for example.

--Kim

On Feb 1, 10:20=A0am, glen herrmannsfeldt <g...@ugcs.caltech.edu> wrote:
> rickman <gnu...@gmail.com> wrote:
>
> (snip, I wrote)
>
> >> If it is useful in ASIC/FPGA designs, wouldn't it also be even more
> >> useful in software running on commonly (or not so commonly) available
> >> processors? Yet none have bothered to implement it.
> >> It could be done in software, but none of the popular languages
> >> have any support for fixed point non-integer arithmetic.
> >> It was one of the fun things I did in PL/I so many years ago, though.
> > I don't really follow your reasoning here. =A0The idea of the fixed
> > point package is to allow a user to work with fixed point values in
> > VHDL without having to manually deal with all the details. =A0You seem
> > to be saying that fixed point arithmetic is of no value because it is
> > not implemented in software. =A0I don't see how software has anything t=
o
> > do with it.
>
> The suggestion, which could be wrong, was that if scaled fixed point
> is useful in an HDL, it should also be useful in non-HDL applications.
>
> The designers of PL/I believed that it would be, but designers of
> other languages don't seem to have believed in it enough to include.
> Personally, I liked PL/I 40 years ago, and believe that it should
> have done better than it did, and scaled fixed point was a feature
> that I liked to use.

I have no idea what bearing a forty year old computer language has to
do with this issue.


> > Software is implemented on programmable hardware and for
> > the most part is designed for the most common platforms and does not
> > do a good job of utilizing unusual hardware effectively. =A0I don't see
> > how fixed point arithmetic would be of any value on conventional
> > integer oriented hardware.
>
> Well, the designers of conventional hardware might argue that they
> support scaled fixed point as long as you keep track of the radix
> point yourself. It is, then, up to software to make it easier for
> programmers by helping them keep track of the radix point.
>
> I believe that in addition to PL/I that there are some other less
> commonly used languages, maybe ADA, that support it.
>
> It can be done in languages like Pascal and C, TeX and Metafont do it.
> Multiplication is complicated, though, because most HLL's don't supply
> the high half of the product in multiply, or allow a double length
> dividend in division. Knuth wrote the Pascal routines for TeX and MF
> to do it, and suggests that implementations rewrite them in assembler
> for higher performance.

Again, I don't know why you are dragging high level languages into
this.  What HLLs do has nothing to do with the utility of features of
HDLs.


> >> >> Yes you can do fixed point with different positions of the radix
> >> >> point, PL/I and a small number of other languages do, but on hardwa=
re
> >> >> that doesn't keep track of the radix point.
> >> > You seem to be thinking only of CPU designs. =A0HDL is used for a lo=
t
> >> > more than CPU designs.
> >> Well, more than CPU designs, I work on systolic arrays, but even
> >> there so, it doesn't seem likely to help much. Will it generate
> >> pipelined arithmetic units? Of all the things that I have to think
> >> about in logic design, the position of the binary point is one of
> >> the smallest.
> > Asking about pipelining with fixed point numbers is like asking if
> > boxes help your car. =A0Sure, if you want to carry oranges in a car you
> > can use boxes, but you don't have to. =A0If your systolic arrays use
> > fixed point arithmetic then the fixed point package will help your
> > systolic arrays. =A0If you don't care about the precision of your
> > calculations then don't bother using the fixed point package.
>
> In the past, the tools would not synthesize division with a
> non-constant divisor. If you generate the division logic yourself,
> you can add in any pipelining needed. (I did one once, and not for
> a systolic array.) With the hardware block multipliers in many FPGAs,
> it may not be necessary to generate pipelined multipliers, but many
> will want pipelined divide.

Oh, I see why you are talking about pipelining now.  I don't think the
package provides pipelining.  But that does not eliminate the utility
of the package.  It may not be useful to you if it isn't pipelined,
but many other apps can use non-pipelined arithmetic just fine.


> >> And I don't expect the processor to help me figure out which
> >> ones those are.
> > What processor? =A0You have lost me here. =A0A multiply provides a resu=
lt
> > with as many bits as the sum of the bits in the inputs. =A0But this man=
y
> > bits are seldom needed. =A0The fixed point package makes it a little
> > easier to work with the variety of formats that typically are used in
> > a case like this. =A0No, the package does not figure out what you want
> > to do. =A0You have to know that, but it does help hide some of the
> > details.
>
> Yes, multiply generates a wide product, and most processors with
> a multiply instruction supply all the bits. But most HLLs don't
> provide a way to get those bits. For scaled fixed point, you
> shift as appropriate to get the needed product bits out.

I have no interest in what HLLs do.  I use HDLs to design hardware and
I determine how the HDL shifts or rounds or whatever it is that I
need.


> >> >> > Similarly, VHDL also has a floating point package that handles
> >> >> > arbitrary sizes of data and exponent.
> >> >> And that is synthesizable by anyone?
> >> > I haven't looked at the floating point package, but if it is describ=
ed
> >> > in terms of basic operators in VHDL it should be synthesizable in
> >> > every tool. =A0That is the point of overloading. =A0You can define a
> >> > multiply operator for real data in terms of the basic building block=
s
> >> > and this is now a part of the language. =A0VHDL is extensible that w=
ay.
> >> Again, does it synthesize pipelined arithmetic units? If not, then
> >> they aren't much use to anything I would work on. If I do it in
> >> an FPGA that is because normal processors aren't fast enough.
> >> But the big problem with floating point is that it takes too much logi=
c.
> >> The barrel shifter for pre/post normalization for an adder is huge.
> >> (The actual addition, not so huge.)
> > Nope, the fixed point package does not design pipelines, it isn't a
> > pipeline package. =A0Yup, the barrel shifter is as large as a
> > multiplier. =A0So what is your point? =A0The purpose of the package is =
to
> > allow you to design floating point arithmetic without designing all
> > the details of the logic. =A0It isn't going to reduce the size of the
> > logic.
>
> The point I didn't make very well was that floating point is still
> not very usable in FPGA designs, as it is still too big. If a
> design can be pipelined, then it has a chance to be fast enough
> to be useful.

Whether floating point is too big depends on your app.  Pipelining
does not reduce the size of a design and can only be used in apps that
can tolerate the pipeline latency.  You speak as if your needs are the
same as everyone else's.


> Historically, FPGA based hardware to accelerate scientific
> programming has not done very well in the marketplace. People
> keep trying, though, and I would like to see someone succeed.
>
> -- glen

Scientific programming is not the only app for FPGAs and fixed or
floating point arithmetic.  Fixed point arithmetic is widely used in
signal processing apps and floating point is often used when fixed
point is too limiting.

Rick

On Feb 1, 11:38=A0am, gtw...@sonic.net (Mark Curry) wrote:>
> Nope. =A0SV synthesis is quite powerful. =A0Arrays, and structs are first=
 class
> citizens in SV and can be members of ports. =A0Further most tools handle
> interfaces now pretty well in synthesis. =A0There's quite a benefit to us=
ing
> these in your RTL.
>
> Trying to resist making this a language war, but I see little advantage
> one or the other with regard to a "higher level of abstraction". =A0You c=
an
> probably achieve the same with either language.
>
> --Mark- Hide quoted text -
>
> - Show quoted text -

Mark,

Sounds like SV synthesis has improved quite a bit, which is good to
know, and good for the industry. Which tools support these features (I
assume at least Synplify)? As soon as someone standardizes a library
to do fixed point in SV, then it will be able to do what VHDL does :)

I'm not that familiar with SV or SV interfaces in particular. There's
one thing I wish VHDL would do, and that is allow an interface (a
record port with different directionality associated with each element
of the record) on a component/entity/subprogram.

From the abstraction point of view, I think you are right, SV and VHDL
synthesis are pretty close, with specific features being edges in
each's favor.

However, I (and I believe the OP) was not considering SV as just a
version of verilog though. Perhaps that is something that is beginning
to change (the realization that there is another viable choice besides
vanilla verilog and vhdl for synthesis.)

Andy

In article <315bd1eb-26b4-4164-b503-9e2146ea7499@l14g2000vbe.googlegroups.com>,
Andy  <jonesandy@comcast.net> wrote:
>On Feb 1, 11:38&#4294967295;am, gtw...@sonic.net (Mark Curry) wrote:>
>> Nope. &#4294967295;SV synthesis is quite powerful. &#4294967295;Arrays, and structs are first class
>> citizens in SV and can be members of ports. &#4294967295;Further most tools handle
>> interfaces now pretty well in synthesis. &#4294967295;There's quite a benefit to using
>> these in your RTL.
>>
>> Trying to resist making this a language war, but I see little advantage
>> one or the other with regard to a "higher level of abstraction". &#4294967295;You can
>> probably achieve the same with either language.
>>
>> --Mark- Hide quoted text -
>>
>> - Show quoted text -
>
>Mark,
>
>Sounds like SV synthesis has improved quite a bit, which is good to
>know, and good for the industry. Which tools support these features (I
>assume at least Synplify)? As soon as someone standardizes a library
>to do fixed point in SV, then it will be able to do what VHDL does :)

I've used Mentor Precision (for FPGAs), and am currently eval'ing Synplify.
Synopsys has supported SV for a while now for ASICs.

I've followed this thread some with some confusion as to what this 
"fixed point library" is and/or does.  I do all kinds of DSP math
apps in verilog, with variable precision.  Yes, you need
to pay attention scaling along the data path.  I'm unsure how
a library can really help this.  

Although I tend to want to be very precise - drawing out
my data path with all truncation/rounding/etc specifically
called out.  I suppose if some library somehow standardized
this... Hmm, need to think about it.  Can't see how it'd offer
much benefit.

--Mark