Ask Slashdot: How Many (Electronics) Gates Is That Software Algorithm?

dryriver writes "We have developed a graphics algorithm that got an electronics manufacturer interested in turning it into hardware. Here comes the problematic bit... The electronics manufacturer asked us to describe how complex the algorithm is. More specifically, we were asked 'How many (logic) gates would be needed to turn your software algorithm into hardware?' This threw us a bit, since none of us have done electronics design before. So here is the question: Is there a piece of software or another tool that can analyze an algorithm written in C/C++ and estimate how many gates would be needed to turn it into hardware? Or, perhaps, there is a more manual method of converting code lines to gates? Maybe an operation like 'Add' would require 3 gates while an operation like 'Divide' would need 6 gates? Something along those lines, anyway. To state the question one more time: How do we get from a software algorithm that is N lines long and executes X number of total operations overall, to a rough estimate of how many gates this algorithm would use when translated into electronic hardware?"

Comments

Holy crap

[CajunArson]

Either implement it as shaders for a GPU (or a DSP) or hire somebody who actually knows about hardware design if you are hell-bent on implementing an ASIC.

Slashdot: Where *not* to go to get specific advice about specific technical issues.

Re:Holy crap

[Megane]

And to think, they rejected my Ask Slashdot submission on how to find a cheat code on my bank's web site for unlimited moneys

Just walk up to any ATM and press: up up down down left right left right B A start.

Re:Holy crap

[Joce640k]

Just 'fess up and say "We don't know, we're software people, not hardware people".

If it's really important they might offer some help.

Re:Holy crap

[Goaway]

This is the only sane answer. They probably only asked to find out if you happened to know.

Say you don't know, and let them look at the code to figure it out.

Re:Holy crap

[Austerity Empowers]

To give a more helpful, unhelpful answer, it's an ill-formed question. "How many gates" depends on the target on which you synthesize the hardware: a PCB, an FPGA, actual silicon (which fab? Which process? whose std cell library? what clock frequency?).

If somehow the above could be narrowed down by asking the customer, then the next thing I'd advise is contracting someone who can write RTL using an HDL (verilog is most popular). The synthesizeable subset of HDL is tricky to learn for non-HW people, so unless you understand digital logic well I'd suggest finding someone else to do it for you. They can then synthesize it to the targeted device/platform. If you can do this, you should charge quite a lot of money since this form of IP is expensive, and they know it. If they're ok with that, you may also want to have this contractor also write the design verification suite, since this company will certainly want that to integrate into their own testing. Lots of contractors are out there for this due to the cyclic nature of this job, make sure you also have some support feature in place if you need them to fix/update the code later.

Even simple software algorithms can be very big in HW, but some surpisingly complex SW algorithms are next to 1 liners in HW (like any form of bit masking or bit swizzling is free!). But generally if there are a lot of sequential steps, and those steps are different...it gets big. Also assume that for every 1 SW guy that wrote the code, you will need 1 RTL designer. If you take the verification step, it may be 1-2 verification engineers for 1 RTL, depending on your timeline.

Re:Holy crap

[fractoid]

It's also ill-formed (to the point of being almost meaningless) in the sense that the smallest number of gates for a given algorithm is probably going to be to implement some kind of low-end processor which then runs the algorithm as code.

What they really wanted to ask was "what's the best price/performance option for executing this algorithm, given the following expected parameters and an initial production run size of X".

Re:Holy crap

[Goaway]

I work with plenty of people with that kind of degree or higher, and I doubt any of them could. Very few CS educations would teach you that. That is highly specialist knowledge, in an usual field.

I really don't know why you would ever think that would be a common skill.

Verilog

[tepples]

If you learn to program in Verilog, you could try synthesizing for some FPGA and see how much space it takes up on the FPGA. But then programming for an FPGA differs from programming for a serial computer in that each line of code runs essentially as a separate thread, usually triggered on another signal (such as a clock) having a positive or negative edge.

Re:Verilog

[Anonymous Coward]

if you only need a estimation, use something like bamboo from PandA to convert your C Code to Verilog. Then synthesize this code for a FPGA. In the summery you should find how many logic cells would be used as well as how many digital gates in an asics are necessary. This value is only a estimation, but for your question, this should work.

Re:Verilog

[Andy Dodd]

While there are some compilers that ATTEMPT to convert C/C++ into a hardware representation - These will usually fail unless you understand the target hardware.

http://www.drdobbs.com/embedded-systems/c-for-fpgas/230800194

One thing is: Even if you can successfully compile from C to Verilog or VHDL, there is no guarantee that the Verilog or VHDL will successfully synthesize on your target hardware.

Even if it successfully synthesizes, there is no guarantee that it will be in any way an optimal implementation.

Some C algorithms may never transfer well into a hardware implementation.

Re:Verilog

[ranulf]

The number of slices or logic cells or whatever else a particular synthesis program for a particular chip generates doesn't exactly correspond to a number of gates either. For instance, a single 4-in 1-out LUT on a Xilinx can be used for 1 gate or 6.

I wouldn't have much confidence in automatic C to HDL conversion either. Good HDL design is about understanding the problem in terms of gates and parallelism. FPGAs and ASICs in general aren't particularly good at things that CPUs are good for, and inversely CPUs aren't especially good for things that FPGAs and ASICs can do well.

The OP shows such a lack of understanding of hardware design that it's not funny! "Add = 3 gates, Divide = 6 gates" is quite comical to anyone who actually knows these things. A more ball park is that an n-bit add can be done with 2n LUTs, in terms of gates it's about 5n gates, but really that depends what gates you have available. A multiplier is massively more, dividing is even more complicated still. Fortunately, many FPGAs come with a few dedicator multipliers... Unless your algorithm requires only as many multipliers as you have available, you're probably best building a state machine and multiplexing a single multiplier unit, in much the same way as a CPU multiplexes the ALU at its core.

The whole thing is massively dependent on algorithm and experience of the person doing the porting. The best advice is to say "I don't know" or to hire someone who does or suggest them running the algorithm on an embedded CPU.

Re:Verilog

[harrkev]

Seriously???? Asking a C++ programmer to begin to use Verilog is simply not practical. There is a VERY STEEP learning curve in trying to target real hardware. There is even a very different frame of mind that has to be learned in order to target gates.

I speak from experience. I program Verilog and SystemVerilog for a living doing ASIC design.

Now, to answer the OP:

The answer is very strongly: it depends. The most optimistic answer is a couple hundred thousand. Implement an 8-bit CPU and write the thing in under 32K of code.

On the other end of the spectrum is "many billions." Design your own x86 multi-core CPU, throw a couple of gigs of SRAM on the ASIC, tons of flash for a solid-state disc drive, and you will have a complete high-end PC on a chip. Then add your software.

Of course, these are both ridiculous extremes. Everything depends on the TYPE of operations being done. In a CPU a simple 32-bit multiply can be done with one character ("*"). In gates, if you need the answer in a single clock cycle, it can take an EXTREME amount of logic. However, if you are willing to wait 32 clock cycles for the answer, the amount of logic is reduced to a very manageable level. This is why C++ is a bad choice of input. How time-sensitive is it? Hardware is also very parallel in nature. Different parts of the chip can indeed be working on different things at the same time. You can go for a strictly pipelined architecture where each block does one little bit of the job and passes it off to the next block. High throughput, but lots of gates. Or you could design a general-purpose block and have it to everything slowly (the most extreme example of this approach is a common CPU).

While I have heard of magic "C to gates" compilers, after almost 15 years in the business, I have never actually seen one. The closest that I have seen are tools that can turn Matlab code into (messy-looking) gates. If your algorithm is DSP in nature, this is a very viable alternative. Otherwise, the only advice that I can give you is to consult somebody who does hardware design for a living (like me).

Otherwise, you really need to look at where the input comes from, where the output goes, and how fast you need to do the work.

Re:Verilog

[SecurityTheatre]

"Add = 3 gates, Divide = 6 gates" is quite comical to anyone who actually knows these things.

Looking at an old reference I have, a 16-bit ripple-carry style adder requires 576 transistors, and a 16-bit carry-lookahead style adder (faster) requires 784 transistors.

This is not including ANY control circuitry, nor a subtract feature.

A pure-hardware 16-bit integer DIVIDE is between 15-30 times more complicated. To do it in pure hardware, would require on the order of 23,000 transistors.

Unless you need your division to happen wicked fast with low latency and you don't care about transistor count, it's better to build add/shift hardware and simply perform a division operation using those bits of hardware repeatedly.

Also, we're only doing 16-bit. If you need 64-bit, multiple all of those numbers by about 50 (spitballing).

And converting from C into VHDL is probably not going to be the best way to go about this. Hire a decent hardware engineer.

Re:Verilog

[harrkev]

Oh, one more thing about "C to Gates" compilers. In the industry I have not seen one in actual use, but they do supposedly exist. However, they would only work in a limited domain.

For example, if you have C++ that does simple control or DSP-type stuff, then it might work (cannot vouch for the quality of the results). On the other hand, if you get one of these compilers and try feeding it the source code for the Apache web server or the Quake engine source code, you are completely screwed.

If your application is, say, a novel type of network filter that inspects and does something to Ethernet packets, you have to figure out how to interface your design with a real Ethernet SerDes .. which is a *LOT* different than opening up something in the "/dev/" directory. If your application is robotics, then you also need to get data into and out of the chip. How exactly is this done? How fast does the logic need to run? Is it speech processing? If so, then this will involve a lot of straight-forward DSP. If you constrain the design to tell it how fast the data needs to flow through, you should be able to get a reasonable estimate. Does your application need a lot of memory? If so, you might need some type of RAM controller. DRAM controllers can be hairy to work with, and you also have to consider latency and throughput.

In theory, C to gates can work quite well, ***for a limited subset of applications***.

HOWEVER: as others have pointed out, anybody who needs to know the answer to this question should be qualified to answer it for themselves.

Re: Verilog

[harrkev]

I still must disagree. Yes, the syntax is somewhat like C. However, WHAT you are coding is completely different. In particular, things that C and do with a simple "if" statement are not allowed at all in proper gate design. It is not hard to imagine a software guy coding latches all over the place, assigning the same signals from withing different always blocks, etc. Even "always @(posedge clock)" may be a fundamental paradigm shift for a software guy. And not to mention the rather arbitrary way that Verilog treats wire vs. reg.

wire a = b & c;

wire a;
assign a = b & c;

reg a
always @(*) a = b & c;

These three constructs do the same thing. Why is one "wire" and one "reg"?

What is the difference between the two blocks (they are NOT the same - blocking vs. non-blocking)?

always @(posedge clk) begin
    a = b;
    c = a & b;
end

always @(posedge clk) begin
    a = b;
    c = a & b;
end

What about race conditions? Glitches on combinatorial logic? Proper coding of state machines? Need memory? How do you drop in an encrypted 3rd party DDR controller and PHY? Interface with AHB bus? In a given process, how many levels are logic are reasonable for a given clock speed? What exactly are hold violations?

I am not saying that any of these are insurmountable. What I am saying is that a good digital designer is worth paying for, and a software guy may have a very steep learning curve indeed.

Re: Verilog

[harrkev]

Gaaa. On the blocking vs. non-blocking, Slashdot swallowed the "less than" sign. Apologies.

Re:Verilog

[harrkev]

Actually, that depends on what the 24,000 transistors are doing. Let's assume that you stupidly did a divide using Verilog "/". This implies a one-cycle divide which might well take that many transistors. The problem is that you would not likely be able to get this to work in real life. With so many levels of logic, your timing would be pure crap. Plus you might have fanout and congestion issues that would further limit your timing. So you could get a divide in one clock cycle, but limit yourself to a clock speed of 10 MHz, for example.

Once you get past about 10 or 12 levels of logic (in my opinion), it is time to re-code, no matter what your clock speed is. If you can't get the job done in 12 level, it is time to re-think your approach. Register re-timing can certainly be useful, but it is much better to do the job right in RTL, the way God intended. Register re-timing can make later steps more complicated (including formal verification).

Why don't they know?

[Anonymous Coward]

You'd think the "electronics manufacturer" would have some idea how to estimate this.

Re:Why don't they know?

[janeuner]

They do have a way. They asked if it had already been determined.

The correct response is, "We don't know."

Re:Why don't they know?

[Sarten-X]

Because they're robots with no AI functionality?

Re:Why don't they know?

[Megane]

A more accurate car analogy would be GM wanting to build a car using your technology and asking you how many assembly line workers it would take.

Just like any other software project

[mbadolato]

Make up a number, then when they complain that it was way off, blame it on their management changing scope a hundred times throughout the life of the project!

Re:Just like any other software project

[Imrik]

Made me think of this.

Try Stackoverflow perhaps?

[Anonymous Coward]

I think you may have a better chance of getting an answer if you ask this question on Stackoverflow (or one of its related sites).

Unfortunately, I think asking on Slashdot is only likely to get you some tired and outdated memes / jokes...

They shouldn't be asking you.

[pavon]

If they plan on implementing this in hardware, then they should have people who are capable of answering that question. If instead, they are just a manufacturer and aren't capable of doing the actual hardware design, then you have bigger problems than answering this question. That is something you should find out about ASAP.

Minecraft

[nbetcher]

Develop out the algorithm in Minecraft using ProjectRed (Integration module, specifically) and then you can easily count the gates! :-)

Completely stupid question

[dskoll]

The question "How many gates does it take to implement this algorithm?" is stupid. It's like asking "How long is a piece of string?"

There will always be a time/space tradeoff, even with translating an algorithm to hardware. You can save time by throwing more gates at the problem to increase parallelism, or you can save space by reusing gates in sequential operations.

You need a C to VHDL translator

[Animats]

You need a C to VHDL translator. Here's a tutorial for one.

Only the parts of the algorithm that have to go really fast need to be fully translated into hardware. Control, startup, debugging, and rarely used functions can be done in some minimal CPU on or off the chip. So, for sizing purposes, extract the core part of the code that uses most of the time and work only on that.

Re:You need a C to VHDL translator

[Trepidity]

One caveat to going this route: if the algorithm contains well-known operations as building blocks, you probably don't want to synthesize your own VHDL versions of those standard operations, since they already have highly optimized hardware implementations. For example, if one step of the algorithm is "compute an FFT", you probably want to use an existing FFT IP core to implements it, rather than translating some FFT C code to new VHDL.

At one extreme, where the algorithm is nothing but a chain of such cores (common in DSP applications), you could get a rough estimate just by looking up the gate counts for each operation and adding them up.

VHDL

[Anonymous Coward]

Implement the algorithm in VHDL and test it on an FPGA. I would imagine you'll need to pay someone $$$$$ to do that for you...

Break down your algorithm

[neutrino38]

Hello,

It is probable that you can break down your algorithm -(I do not mean code) into a pipeline of elementary processing and find implementations (IP) for each of them.

to give out an estimate:
- subdivise your algorithm into simpler pieces
- find for each simple piece how it can or could be implemented in hardware and the complexity of each piece.
- do the sum.

  Indeed an hardware designer or consultant would be of a great help here.

HLS

[orledrat]

What you want to do is called high-level synthesis (going from C to hardware description language (HDL) to generating gate-lists from that HDL) and there's plenty of software to do that with. A neat open-source package for HLS is LegUp (http://legup.eecg.utoronto.ca/), check it out to get an idea of what the process consists of.

It doesn't work like that

[cjonslashdot]

It's about more than gates. It is about registers, ALUs, gates, and how they are all connected. There are many different possible architectures, so it depends on the design: some designs are faster but take more real estate. There are algorithm-to-silicon compilers (I know: I wrote one for a product company during the '80s and it is apparently still in use today) but each compiler will assume a certain architecture. I would recommend one but I have been out of that field for decades.

Easy calculation

[Anonymous Coward]

Here is a proven method for calculation.

If your code is:
a) C: divide the number of lines with 7
b) C++: divide the number of lines with 5
c) Ruby/Python/Java: divide the number of lines with 3
d) Perl: multiply the number of lines with 42
e) C#: resign.

Sounds like a joke

[Cryacin]

How many Gates will it take to implement your software project?

One. His name is Bill, and here is yours.

It's an optimization problem

[swm]

You already have your algorithm running in electronic hardware, right?
Your current gate count is the sum of
  * the gate count of your CPU
  * the gate count of your RAM
  * the gate count of your program ROM

So that's an upper bound on the gate count.
If that number is too big for your manufacturing partner,
then you have an optimization problem.

Optimization is a hard problem...

Accurate answer

[Sarten-X]

Write out the truth table for each output as a Karnaugh map incorporating every input. Count the number of gates needed to solve the map, and that's your answer for that output bit. Repeat for every other output bit. Add all those numbers together, and that's a fair estimate of how many gates you'll need.

Of course, this method requires that your number of input bits must be fairly small. Don't forget that memory counts as both input (when read) and output (when written). For nontrivial applications, you'll find that the number of gates quickly approaches "a lot".

The key to success.

[ttucker]

Do not ask a computer scientist to be an electrical engineer.

Re:The key to success.

[geoskd]

Except ... Wow. An early course in my computer science curriculum was: 201. Computer Logic Design I (3) Prerequisite: MATH 113 or equivalent all with a grade of "C" or better. Basic topics in combinational and sequential switching circuits with applications to the design of digital devices. Introduction to Electronic Design Automation (EDA) tools. Laboratory projects with Field Programmable Gate Arrays (FPGA). (Lecture 2 hours, lab 3 hours) Letter grade only (A-F). (We used Verilog and a Xilinx FPGA board.) I'm surprised a reputable CS degree wouldn't require at least a basic course in digital logic; Cal State Long Beach is a great school, but it's certainly not a standards bearer...

There is a world of difference between an entry level college course on ASIC/FPGA design, and actually being able to do the job. Just because you can design and synthesize a projct with a few hundred gates in it does not mean you are even remotely prepared to know where to begin a project with 10^6+ gates in it. More impotantly, high level software languages allow for indescriminant serial loops which are massively difficult to deal with in pure hardware. In short, the design methodology is completely different if you are trying to build for a software path, or a hardware path. You need someone with a hardware mindset to take your algorithm back to scratch and start over. Even knowing the HDLs is not good enough, as it is relatively trivial to write "valid" VHDL or Verlilog code that cant be synthesized...

I've done this before

[Asmodae]

There's been several people who suggested using a high-level synthesis tool to convert your software (c/c++) directly to HDL (verilog/VHDL) of some kind. This can work and I've been on this task and seen it's output before. The catch is; unless that software was expressly and purpose written to describe hardware (by someone who understands that hardware and it's limitations and how that particular converter works), it almost always makes awful and extraordinarily inefficient hardware.

Case in point - we had one algorithm developed in Simulink/Matlab that needed to end up in an FPGA. After 'pushing the button' and letting the tool generate the HDL, it consumed not just 1 but about 4 FPGAs worth of logic gates, RAMs, and registers. Needless to say the hardware platform only had one FPGA and a good portion of it was already dedicated to platform tasks so only about 20% was available for the algorithm. We got it working after basically re-implementing the algorithm with the goal of hardware in mind. The generation tool's output was 20 times worse than what was even feasible. If you're doing an ASIC you can just throw a crap-load of extra silicon at it, but that gets expensive very quickly. Plus closing timing on that will be a nightmare.

My job recently has been to go through and take algorithms written by very smart people (but oriented to software) and re-implement them so they can fit on reasonably sized FPGAs. It can be a long task sometimes and there's no push-button solution for getting something good, fast, cheap. Techies usually say you can pick two during the design process, but when converting from software to hardware you usually only get one.

Granted this all varies a lot and depends heavily on the specifics of the algorithm in question. But the most likely way to get a reasonable estimate is going to be to explain the algorithm in detail to an ASIC/FPGA engineer and let them work up a prelim architecture and estimate. The high-level synthesis push-button tools will give you a number but it probably won't be something people actually want to build/sell or buy.

Knowing The Algorithm Is NOT Enough

[SplawnDarts]

Knowing what algorithm you want to run in hardware in not even close to enough to estimate gates. You need to know the algorithm, and the required performance, and have a sketched out HW design that meets those goals. THEN you can estimate gate count.

For a simple example of why this is, consider processors. A 386 and a Sandy Bridge i7 implement very similar "algorithms" - it's just fetch->decode->execute->writeback all day long. If you implemented them in software emulation, it would be very similar software with some additional bits for the newer ISA features on the i7. But a 386 is about 280 THOUSAND gates, and the i7 is about 350 MILLION gates/core - three orders of magnitude different. Of course, there's at least a 2 order of magnitude performance difference too - it's not like those gates are going to waste.

Point is, knowing the algorithm isn't enough to get even a finger in the wind guess at gate count. If you need an answer to this question, you need to get competent HW design people looking at it.