# Equivalent. But better.

Ever since primary (elementary) school, we’ve known that multiplying an integer by 10 is easy. No need for the long written calculations we churn through when doing multiplication by hand. Just stick an extra zero on the end, and we’re done. Multiplication is (relatively) hard, concatenation of digits is easy. And yet, in this case, they’re equivalent in terms of the operation they perform.

Similar equivalences abound in the design of digital hardware for arithmetic computation, and my PhD student Sam Coward (jointly supervised by Theo Drane from Intel) has been devising ways to automatically take advantage of such equivalences to make Intel hardware smaller and more efficient. He will be presenting our work on this topic at the main computer arithmetic conference, ARITH, next week. The conference will be online, and registration is free: https://arith2022.arithsymposium.org.

Let’s explore this example from our early school years a bit more. I’ll use Verilog notation $\{\cdot,\cdot\}$ to denote a function taking two bit vectors and concatenating them together. Of course the ‘multiplication by 10 is easy’ becomes ‘multiplication by 2 is easy’ in binary. Putting this together, we can write $2*x \simeq \{x,0\}$, meaning that multiplication by two is the same as concatenation with a zero. But what does ‘the same as’ actually mean here? Clearly they are not the same expression syntactically and one is cheap to compute whereas one is expensive. What we mean is that no matter which value of $x$ I choose, the value computed on the left hand side is the same as the value computed on the right hand side. This is why I’ve chosen to write $\simeq$ rather than $=$. $\simeq$ clearly defines a relation on the set of expressions. This is a special kind of relation called a congruence: it’s an equivalence relation, i.e. it is symmetric, transitive, and reflexive, but it also ‘plays well’ with function application: if $x \simeq y$ then it necessarily follows that $f(x) \simeq f(y)$ for every function symbol $f$. Like any equivalence relation on a set, $\simeq$ partitions the set into a set of equivalence classes: in our setting a class corresponds to expressions that can be freely interchanged without changing the functionality of our hardware, even if it changes the performance, area or energy consumption of the resulting design.

Our colleagues Willsey, Nandi, Wang, Flat, Tatlock and Panchekha recently published egg, a wonderful open source library for building and exploring data structures known as ‘e-graphs’, specifically designed to capture these relations on expressions. Sam, Theo and I have developed a set of ‘rewrites’ capturing some of the important intuition that Intel designers apply manually, and encoded these for use within egg. To give you a flavour of these rewrites, here’s the table from Sam’s paper; you can see the example we started with is hiding in there by the name ‘Mult by Two’. The subscripts are used to indicate how many digits we’re dealing with; not all these rules are true for arbitrarily-sized integers, and Sam has gone to some lengths to discover simple rules – listed here as ‘sufficient condition’ – for when they can be applied. This is really important in hardware, where we can use as few or as many bits as the job requires.

You can imagine that, when you have this many equivalences, they all interact and you can very quickly build up a very large set of possible equivalent expressions. e-graphs help us to compactly represent this large set.

Once our tool has spent enough time building such a representation of equivalences, we need to extract an efficient implementation as a hardware implementation. This is actually a hard problem itself, because common subexpressions change the hardware cost. For example if I’m calculating $(x+1)*(x+1)$ then I wouldn’t bother to calculate $x+1$ twice. We describe in our paper how we address this problem via an optimisation formulation. Our tool solves this optimisation and produces synthesisable Verilog code for the resulting circuit.

So, does it generate good circuits? It certainly does! The graph below shows the possible circuit area and performance achievable before (blue) and after (orange) the application of our tool flow before standard logic synthesis tools. For this example, silicon area can be reduced by up to around 70% – a very significant saving.

I’ve really enjoyed working on this topic with Sam and Theo. Lots more exciting content to follow. In the meantime, please tune in to hear Sam talk about it next week.

# Keeping the Pipelines Full

On the 16th May, my PhD student Jianyi Cheng (jointly advised with John Wickerson) will present his most recent paper “Dynamic C-Slow Pipelining for HLS” at FCCM 2022 in New York, the FPGA community’s first in-person conference since the pandemic hit.

Readers of this blog may remember that Jianyi has been working on high-level synthesis techniques that combine the best of dynamic scheduling with the best of static scheduling [FPGA 2020,FCCM 2021]. The general principle underlying his work is to make the most of what information we have at compile time to develop highly efficient custom architectures, while leaving what we don’t know at compile time to influence execution at run-time.

A very common design pattern in hardware acceleration is the idea of C-slow pipelining. Pipelining tends to be taught early in undergraduate programmes, but C-slow pipelining rarely gets mentioned. The idea arises in circuits with feedback loops. The basic approach to pipelining doesn’t really work in this setting: although we can throw multiple registers into the circuit, potentially improving clock frequency at the cost of latency, just like with feed-forward circuits, we can’t then overlap computation to achieve improved throughput, unlike the feed-forward case, because of the data dependency corresponding to the feedback loop.

C-slow pipelining essentially says “OK, but you can use the spare capacity induced by the pipeline registers to overlap computation of independent streams of data, if you happen to have them available.”

Our new paper introduces a dynamic HLS flow for C-slow pipelining. This is particularly valuable in the context of a globally dynamic environment but where certain components exhibit static control flow and can be efficiently pipelined, for example some deep but predictable computation that must be repeated many times but with the arrival times and sources for this computation may change dynamically at runtime, a perfect fit for our prior work.

Jianyi presents a way to leverage the Boogie language and tool flow from Microsoft Research to automatically prove sufficient conditions for C-slowing to be correct. He is then able to introduce a new hardware component within the Dynamatic HLS tool that allows the schedule to “run ahead” to implement certain bounded out-of-order executions corresponding to C-slowing at the circuit level.

At the cost of a small area overhead in the region of 10%, this combined software analysis and hardware transformation is able to reduce wall-clock execution time by more than half compared to the vanilla dynamic scheduling approach.

If you’ll be in NYC in mid-May, go along and hear Jianyi’s talk!

# Nonlinearity is Your Friend

My former PhD student Erwei Wang and I recently teamed up with some collaborators at UCL: Dovydas Joksas, Nikolaos Barmpatsalos, Wing Ng, Tony Kenyon and Adnan Mehonic and our paper has just been published by Advanced Science (open access).

Our goal was to start to answer the question of how specific circuit and device features can be accounted for in the training of neural networks built from analogue memristive components. This is a step outside my comfort zone of digital computation, but naturally fits with the broader picture I’ve been pursuing under the auspices of the Center for Spatial Computational Learning on bringing circuit-level features into the neural network design process.

One of the really interesting aspects of deep neural networks is that the basic functional building blocks can be quite diverse and still result in excellent classification accuracy, both in theory and in practice. Typically these building blocks include linear operations and a type of nonlinear function known as an activation function; the latter being essential to the expressive power of ‘depth’ in deep neural networks. This linear / nonlinear split is something Erwei and I, together with our coauthors James Davis and Peter Cheung, challenged for FPGA-based design, where we showed that the nonlinear expressive power of Boolean lookup tables provides considerable advantages. Could we apply the a similar kind of reasoning to analogue computation with memristors?

Memristive computation for the linear part of the computation performed in neural networks has been proposed for some time. Computation essentially comes naturally, using Ohm’s law to perform scalar multiplication and Kirchhoff’s Current Law to perform addition, resulting in potentially energy-efficient analogue dot product computation in a physical structure known as a ‘crossbar array’. To get really high energy efficiency, though, devices should have high resistance. But high resistance brings nonlinearity in practice. So do we back away from high resistance devices so we can be more like our mathematical abstractions used in our training algorithms? We argue not. Instead, we argue that we should make our mathematical abstractions more like our devices! After all, we need nonlinearity in deep neural networks. Why not embrace the nonlinearity we have, rather than compromise energy efficiency to minimise it in linear components, only to reintroduce it later in activation functions?

I think our first experiments in this area are a great success. We have been able to not only capture a variety of behaviours traditionally considered ‘non-ideal’ and harness them for computation, but also show very significant energy efficiency savings as a result. You can see an example of this in the figure above (refer to the paper for more detail). In high power consumption regimes, you can see little impact of our alternative training flow (green & blue) compared to the standard approach (orange) but when you try to reduce power consumption, a very significant gap opens up between the two precisely because our approach is aware of the impact this has on devices, and the training process learns to adapt the network accordingly.

We’ve only scratched the surface of what’s possible – I’m looking forward to lots more to come! I’m also very pleased that Dovydas has open-sourced our training code and provided a script to reproduce the results in the paper: please do experiment with it.

# Pruning Circuits

On Tuesday, my former PhD student Erwei Wang (now at AMD) will present our recent paper “Logic Shrinkage: Learned FPGA Netlist Sparsity for Efficient Neural Network Inference” at the ACM International Symposium on FPGAs. This is joint work with our collaborator Mohamed Abdelfattah from Cornell Tech as well as James Davis, George-Ilias Stavrou and Peter Y.K. Cheung at Imperial College.

In 2019, I published a paper in Phil. Trans. Royal Soc. A, suggesting that it would be fruitful to explore the possibilities opened up when considering the graph of a Boolean circuit – known as a netlist – as a neural network topology. The same year, in a paper at FCCM 2019 (and then with a follow-up article in IEEE Transactions on Computers), Erwei, James, Peter and I showed how to put some of these ideas into practice by learning the content of the Boolean lookup tables that form the reconfigurable fabric of an FPGA, for a fixed circuit topology.

Our new paper takes these ideas significantly further and actually learns the topology itself, by pruning circuit elements. Those working on deep neural networks will be very familiar with the idea of pruning – removing certain components of a network to achieve a more efficient implementation. Our new paper shows how to apply these ideas to prune circuits made of lookup tables, leaving a simplified circuit capable of highly-efficient inference. Such pruning can consist of reducing the number of inputs of each Boolean LUT and, in the limit, removing the LUT completely from the netlist. We show that very significant savings are possible compared to binarising a neural network, pruning that network, and only then mapping it into reconfigurable logic – the standard approach to date.

We have open-sourced our flow, and our paper has been given a number of ACM reproducibility badges. Please try it out at https://github.com/awai54st/Logic-Shrinkage and let us know what you think. And, if you’re attending FPGA next week, reach out and say hello.

# Islands of Certainty

Readers of this blog may remember that my PhD student Jianyi Cheng (jointly supervised by John Wickerson) has been working on high-level synthesis, combining dynamic scheduling with static scheduling. His latest contribution, to be presented on 28th February at the ACM FPGA conference, is about finding islands of static control flow in a sea of dynamic behaviour.

Here’s the story so far:

So now, two years later, we are back at the same conference to present a method to do just that. We now have an automated flow to select parts of a program to statically schedule, resulting in a 4x reduction in area combined with a 13% boost in performance compared to a fully dynamic circuit, a result that is close to the best achievable — as shown by exhaustively enumerating different parts of the program to schedule statically.

The basic idea of the paper is to develop the concept of a static island — a part of a dataflow graph where making decisions on scheduling of operations once, at compile time, is likely to have minimal impact on performance (or may even improve it) while opening the door to static resource sharing. We can throw a boundary around these islands, synthesise them efficiently with commercial HLS tools (we use Xilinx Vitis HLS), and integrate the result into the overall dynamic circuit using our previous open-source compiler flow.

So what makes a good static island? Unsurprisingly, these islands should exhibit static control flow or control flow with balanced path timing, e.g. in a conditional statement the if and else branch should take the same time, and loops should have constant dependence distances (or none at all). Jianyi also shows that there is an advantage to having these islands consume their inputs at offset-times, e.g. for a two-input island we may wish the static scheduler to be aware that second input is available — on average — two cycles after the first. He shows precisely how to generate ‘wrapper’ circuits for these components, allowing them to communicate with a dynamically scheduled environment.

The overall design flow, shown below, is now fully automated – freeing the user from writing the pragmas we required two years ago.

# What’s the Rush?

At FPL 2021, my PhD student Jianyi Cheng (jointly supervised by John Wickerson) will present our short paper “Exploiting the Correlation between Dependence Distance and Latency in Loop Pipelining for HLS”. In this post, I explain the simple idea behind this paper and how it can significantly accelerate certain neglected corner cases in high-level synthesis (HLS).

By far the most significant way to extract high performance from a hardware accelerator in high-level synthesis is to use loop pipelining. Loop pipelining is the idea of starting the next iteration of a loop before the previous one finishes, allowing multiple iterations to be executing simultaneously. However, some loop iterations may need a result produced by earlier loop iterations, limiting the extent to which this can be done. HLS tools generally determine a ‘safe’ initiation interval – the number of clock cycles between starting two adjacent loop iterations – and then schedule the iterations statically at multiples of this interval.

This limit on initiation interval of the loop essentially derives from two properties. Firstly, if it takes a long time for the computation of a loop iteration to execute, then any iterations waiting on its result must be delayed. But secondly if an iteration’s result is only needed many iterations later, it can afford to take a long time to compute: what’s the rush? These two factors – latency and dependence distance – together determine the safe initiation interval.

The simple observation of our paper is that typically HLS tools will generally independently over-approximate latency and under-approximate dependence distance. However, there are some examples of programs where there is a correlation between dependence distance and latency. Jianyi gives this nice motivating example in the paper:

double f( double a ) {
return (((((a+0.64)*a+0.7)*a+0.21)*a+0.33)*a+0.25)*a+0.125;
}

void example( double vec[M] ) {

for (int i = 0; i < N; i++) {
double e = vec[i];
if (e > 0) vec[i+63] = f(e);
else vec[i*i+9] = e * e;
}

}


In this code snippet, you can see two control paths in the loop. The if branch has a long latency (it computes the Horner scheme polynomial f) but also writes to elements of vec that only get read many iterations later. Meanwhile the else branch has a short latency but can write – in the early stages of the loop at least – to values read in nearby iterations.

The end result is that the commercial tools Jianyi tried don’t cope very well with scheduling this loop. However, Jianyi has developed an approach that uses the formal verification tool Boogie to show that this loop can actually be scheduled very efficiently by exploiting this correlation.

He has developed an LLVM pass called iiProver that proves that it is safe to use a certain II with the commercial Vitis HLS tool from Xilinx. iiProver and our benchmarks are available – please take a look: https://github.com/JianyiCheng/iiProver. And you can hear Jianyi talking about his work on Youtube here: https://www.youtube.com/watch?v=SdQeBBc85jc.

# It Probably Works!

Followers of my research will know that I’ve long been interested in rounding errors and how they can be controlled to best achieve efficient hardware designs. Going back 20 years, I published a book on this topic based on my PhD dissertation, where I addressed the question of how to choose the precision / word-length (often called ‘bit width’ in the literature) of fixed point variables in a digital signal processing algorithm, in order to achieve a controlled tradeoff between signal-to-noise ratio and implementation cost.

Fast forward several years, and my wonderful collaborators Fredrik Dahlqvist, Rocco Salvia, Zvonimir Rakamarić and I have a new paper out on this topic, to be presented by Rocco and Fredrik at CAV 2021 next week. In this post, I summarise what’s new here – after all, the topic has been studied extensively since Turing!

I would characterise the key elements of this work as: (i) probabilistic, i.e. we’re interested in showing that computation probably achieves its goal, (ii) floating point (especially of the low custom-precision variety), and (iii) small-scale computation on straight-line code, i.e. we’re interested in deep analysis of small kernels rather than very large code, code with complex control structures, or code operating on very large data structures.

Why would one be interested in showing that something probably works, rather than definitely works? In short because worst-case behaviour is often very far from average case behaviour of numerical algorithms, a point discussed with depth in Higham and Mary’s SIAM paper. Often, ‘probably works’ is good enough, as we’ve seen recently with the huge growth of machine learning techniques predicated on this assumption.

In recent work targeting large-scale computation, Higham and Mary and, independently, Ipsen, have considered models of rounding error that are largely / partially independent of the statistical distribution of the error induced by a specific rounding operation. Fredrik was keen to take a fresh look at the kind of distributions one might see in practice, and in our paper has derived a ‘typical distribution’ that holds under fairly common assumptions.

Rocco and Fredrik then decided that a great way to approximate the probabilistic behaviour of the program is to sandwich whatever distribution is of interest between two other easy to compute distributions, utilising the prior idea of a p-box.

One of the core problems of automated analysis of numerical programs has always been that of ‘dependence’. Imagine adding together two variables each in the range $[-1,1]$. Clearly their sum is in the range $[-2,2]$. But what if we knew, a priori, that these two variables were related somehow? For example in the expression $X + (-X)$, which is clearly always zero. Ideally, an automated system should be able to produce a tighter result that $[-2,2]$ for this! Over the years, many approaches to dealing with this issue have arisen, from very the very simple approach of affine arithmetic to the more complex semialgebraic techniques Magron, Donaldson and myself developed using sequences of semidefinite relaxations. In our CAV paper, we take the practical step of cutting-out regions of the resulting probability space with zero probability using modern SMT solver technology. Another interesting approach used in our paper is in the decision of which nonlinear dependences to keep and which to throw away for scalability reasons. Similar to my work with Magron, we keep first-order dependence on small rounding error variables but higher-order dependence on input program variables.

I am really excited by the end result: not only a wonderful blend of ideas from numerical analysis, programming languages, automated reasoning and hardware, but also a practical open-source tool people can use: https://github.com/soarlab/paf. Please give it a try!

Readers interested in learning more about the deeply fascinating topic of numerical properties of floating point would be well advised to read Higham’s outstanding book on the topic. Readers interested in the proofs of the theorems presented in our CAV paper should take a look at the extended version we have on arXiv. Those interested in some of the issues arising (in the worst case setting) when moving beyond straight-line code could consult this paper with Boland. Those interested in the history of this profoundly beautiful topic, especially in its links to linear algebra, would do well to read Wilkinson.

# Scheduling with Probabilities

Readers of this blog may remember that Jianyi Cheng, my PhD student jointly supervised by John Wickerson, has been investigating ways to combine dynamic and static scheduling in high-level synthesis (HLS). The basic premise has been that static scheduling, when it works well due to static control, works very well indeed. Meanwhile, for programs exhibiting highly dynamic control flow, static scheduling can be very conservative, a problem addressed by our colleagues Lana Josipović, Radhika Ghosal and Paolo Ienne at EPFL. Together with Lana and Paolo, we developed a scheme to combine the best of both worlds, which we published at FPGA 2020 (and recently extended in IEEE Transactions on CAD). I blogged about this work previously here. We provided a tool flow allowing us to stitch large efficient statically-scheduled components into a dynamic circuit.

However, when scheduling a circuit statically, there are many design choices that can be made, typically to trade off time (throughput, latency) against area. So while our previous work was useful to stitch pre-existing statically-scheduled components into a dynamically-scheduled environment, we had no way of automatically designing those components to optimally fit the dynamic environment.

Enter Jianyi’s latest contribution – to be presented at FCCM 2021 next week.

In his paper “Probabilistic Scheduling in High-Level Synthesis”, Jianyi tackles this problem. He demonstrates that the dynamic environment, including data-dependent decisions and even load-store queues, can be adequately modelled using a Petri net formalism, and uses the PRISM model checker from Kwiatowska et al. to extract an appropriate initiation interval for each statically-scheduled component.

The initiation intervals inferred by Jianyi’s tool can then be given to a commercial HLS tool – in our case Vitis HLS – to schedule each component. The components – together with any remaining dynamically-scheduled code – is then integrated using our previously published framework, producing the complete FPGA-ready design. The whole process provides a quality of result very close to an exhaustive search of possible initiation intervals, without having to perform multiple scheduling runs, and so in a fraction of the time.

# Watch Where You’re Pointing That!

This week Nadesh Ramanathan, a member of research staff in my group, will be presenting a paper at the virtual FPL 2020 conference entitled “Precise Pointer Analysis in High Level Synthesis” (jointly with John Wickerson and myself). This blog post is intended as an accessible summary of the key message of the paper.

People are now aiming to generate hardware accelerators for more complex algorithms than things like classical CNNs, low-level image processing tasks, and other bread-and-butter hardware acceleration tasks. Inevitably, this is a difficult task to get right, and the prevalence of C/C++-based high-level synthesis (HLS) tools offers a great opportunity to experiment with the design space. Sophisticated algorithms written in C/C++ often incorporate pointers, which have long been difficult for HLS tools. Previously, I proposed a relatively sophisticated analysis using separation logic, together with my PhD student Felix Winterstein, which is an intensive analysis specialised to certain data structures. Nadesh’s most recent work can, in some sense, be viewed as the opposite. He is trying to make more simple, but more generally applicable pointer analyses more widely understood and used within HLS, while trying to quantify how much this might bring to hardware accelerator design.

The basic idea is that since FPGA compile times are long, we can afford to spend a bit more time being precise about which variables can point to which other variables. The question is, what are the benefits of being more precise in the context of HLS? Nadesh has studied two different types of ‘sensitivity’ of pointer analyses – to flow and to context. Flow-sensitive analyses consider the ordering of memory operations, context sensitive analyses consider the calling context of functions. The most common form of analysis in HLS is Andersen analysis, which is neither flow- nor context-sensitive. So how much do we gain by utilising more precise analyses?

Nadesh studies this question by modifying the LegUp source code, showing that over the PTABen benchmark set, area utilisation can be halved and performance doubled by using these analyses. This suggests that as we move towards greater diversity in hardware accelerators, HLS tool developers should think carefully about their pointer analyses.

# When are Digits Correct?

Often, we compute with iterative algorithms. Start with some value, often an initial guess to be refined, and keep iterating until some stopping criterion is met. If we actually think about what goes on in a modern digital computer when we execute these algorithms, we soon see that – often – the same digits end up being computed time and again. As we converge to a value, it’s reasonable to expect that most of the time the most significant digits become stable. But we still compute them, time and again at each iteration, wasting computational resource.

In general, in standard binary representations, this re-computation may not be avoidable – most-significant digits might be stable for 1000 iterations and then flip, e.g. from 0.99999 to 1.00000. As a child, I used to play with such iterations using my HP32S calculator – a gift from fred harris – it provided endless entertainment.

There is, however, a class of number representations in which these digit flips can be avoided: redundant number representations. These representations have a long history – indeed, as my friend and colleague Miloš Ercegovac has identified, they can be traced back as far as a 1727 publication in Phil. Trans. Royal Soc by John Colson FRS. Miloš developed these ideas to a mature state between the 1970s and today, in the guise of Online Arithmetic.

Together with my PhD students He Li (now research staff at Cambridge) and Ian McInerney and collaborator James Davis, I have recently done some work on methods to detect and establish exactly when digits become stable using such schemes and what the implications might be for hardware that makes use of this stability. In our recent IEEE Transactions on Computers paper, we adapt standard forward error analyses of stationary iterative methods to this setting. We mathematically derive some conditions that can be checked at run-time to determine when you don’t need to compute certain digits in any future iteration, and also present a toy hardware implementation taking advantage of this approach using a non-standard arithmetic processor design.

We hope that – in the future – only what needs to be computed will be computed.