POLITECNICO DI TORINO

Master Degree In Computer Engineering

Master Thesis

Interfacing a Neuromorphic
Coprocessor with a RISC-V
Architecture

Supervisors
PhD Gianvito Urgese
Dr. Evelina Forno
Candidate
Andrea Spitale

Academic Year 2020-2021

”Intelligence is the ability to adapt to change”
— Stephen Hawking

3
Summary

The concept of neural network is nowadays spread everywhere. Commonly meant

as a mean to provide what is called artificial intelligence, neural network constitute
one of few state of the art technologies which are constantly growing in complex-
ity and efficiency to overcome new challenges, providing fascinating services and
features, ranging from image classification and manipulation, up to speech recogni-
tion and many more that are still being explored. A neural network is a computing
system which takes inspiration from the human brain, exploiting its parallel in-
terconnections to solve complex data problems, modifying its internal parameters
(training phase) in order to recognize unknown input data with higher accuracy
(inference phase). Spiking Neural Networks (SNN) represent an emerging class of
neural networks, coming from neuroscience research and aiming at accurately re-
producing both static and dynamic behaviours of human brain neurons. Initially
conceived to help neuroscientists enrich their knowledge on the human brain struc-
ture and working principles, SNNs have gathered attention in the computer science
community, for several reasons: power efficiency, continuous learning, natural sup-
port for spatio-temporal input.
Internet of Things (IoT) oriented applications, running on smart devices capable
of analyzing data in real time with limited power resources, are predicted to be
those that will benefit a lot from the adoption of SNN based solutions. Indeed,
since spikes are sparse in time and space, the network is characterized by very low
activity, thus the overall power consumption is greatly reduced if compared to other
solutions, such as Convolutional Neural Networks.
To enhance the computing capability of an IoT device during the execution of SNN
based algorithms, the thesis describes the design of an interfacing solution for con-
trolling a reconfigurable SNN-accelerated coprocessor, named ODIN, by means of
a RISC-V based System on Chip (SoC), without any remote controller from the
cloud. The designed architecture exploits the serial peripheral interface (SPI) to
let the RISC-V core configure the accelerator parameters, thus offloading the SNN
task on ODIN.
The capability of the system to work as a standalone device has been validated by
configuring a synfire chain simulation without the intervention of an host computer.
A synfire chain is a particular arrangement of spiking neurons, which are connected
4
in a loop, firing in series once a proper input stimuli is provided. The synfire chain
has been used to benchmark the proposed architecture because of its predictable
behaviour, which can be monitored through Register Transfer Level (RTL) simu-
lation. The network behaviour is tested on a few case scenarios, where a number
of parameters are changed, and the results are matched against the expected ones.
Thus, demonstrating that the proposed solution could avoid the usage of a host
computer to control and configure the SNN based accelerator.

5
Contents

1 Introduction 9

2 Technical Background 17
2.1 Machine Learning and Applications . . . . . . . . . . . . . . . . . . 17
2.1.1 Machine Learning . . . . . . . . . . . . . . . . . . . . . . . . 17
2.1.2 Neural Networks . . . . . . . . . . . . . . . . . . . . . . . . 17
2.2 The RISC-V Instruction Set Architecture (ISA) . . . . . . . . . . . 21
2.2.1 RV32I - Base Integer ISA . . . . . . . . . . . . . . . . . . . 24
2.2.2 Choices and Consequences . . . . . . . . . . . . . . . . . . . 25
2.3 ODIN : A Spiking Neural Network Coprocessor . . . . . . . . . . . 28
2.3.1 Supported Neurons Models . . . . . . . . . . . . . . . . . . . 30
2.3.2 SPI Slave . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
2.3.3 Controller . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
2.3.4 AER Output . . . . . . . . . . . . . . . . . . . . . . . . . . 38
2.3.5 Scheduler . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
2.3.6 Neuron Core . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
2.3.7 Synaptic Core . . . . . . . . . . . . . . . . . . . . . . . . . . 41
2.4 Chipyard Hardware Design Framework . . . . . . . . . . . . . . . . 45
2.4.1 Rocket Core . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
2.4.2 Tools & Toolchains . . . . . . . . . . . . . . . . . . . . . . . 48
2.4.3 Simulators . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
2.5 Parallel Ultra Low Platform (PULP) . . . . . . . . . . . . . . . . . 50
2.5.1 PULPino . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
2.5.2 PULPissimo . . . . . . . . . . . . . . . . . . . . . . . . . . . 52

3 Methods 55
3.1 Setting up ODIN and Chipyard Environments . . . . . . . . . . . . 56
3.2 ODIN Integration Inside Chipyard . . . . . . . . . . . . . . . . . . . 57
3.3 ODIN Parameters Definition . . . . . . . . . . . . . . . . . . . . . . 67
6
4 Results & Discussion 71
4.1 RTL Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
4.2 Synthesis Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
4.2.1 Area . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83

5 Conclusions 85

Bibliography 88

7
Chapter 1

Introduction

Neuromorphic engineering, of which you might have heard as ”neurmorphic com-

puting”, represents one of the most interesting yet challenging interdisciplinary field
for all computer science enthusiasts, both from academic and industrial communi-
ties. Conceived in the 80s by Carver Andress Mead, an American engineer that
is considered as one of the pioneers of modern microelectronics and that originally
coined Gordon Moore’s prediction as ”Moore’s Law”, it involves the usage of very
large scale integrated (VLSI) circuits to assemble systems that strive to reproduce
the behaviour of the human brain, which is capable of processing and transmitting
information at a really high speed, due to dense connectivity and high degree of
parallelism, while consuming a very low amount of power (≈ 25 W [22]). The key
aspect to be taken in consideration when building such a system is the way neurons
and synapses are interconnected, since this strongly influences the computations
the system will be able to perform, how robust the architecture is with respect to
external and undesired inputs and conditions, and whether and how the intercon-
nections will adapt to changes. Neuromorphic engineering goes hand in hand with
neural networks, those widespread objects that are involved in the most fascinat-
ing electronic systems we deal with everyday, ranging from AI driven cameras on
smartphones up to circuits which are provided to robots so that they can react and
improve their responses to external events.
According to [50], there are 10 main reasons that make neuromorphic hardware
so interesting in the upcoming years.
1. Faster. The Von Neumann architecture had been devised, together with Har-
vard one, to have machines process information at a way faster rate with
respect to the average one with which humans process it. However, if one
compares the computational speed of such an architecture with that of the
human brain, it becomes evident there are limits and defects that have to
be overcome if one wants to reach the efficiency of that biological system. In
particular, researchers working in the 80s stated that neural networks compu-
tation could have benefited a lot from the development of custom, dedicated
9
 Introduction

Parallelism Scalability Footprint Faster Neuroscience

Real-Time Von Neumann Fault Online

Performance Bottleneck Low Power Tolerance Learning

2015 (150)

2014 (123)

2013 (127)

2012 (102)

2010-2011 (104)

2006-2009 (112)

1998-2005 (108)

1988-1997 (103)

0 10 20 30 40 50 60 70 80 90 100
Percentage of papers in each field

Figure 1.1: A graph showing 10 main motivations for which neuromorphic com-
puting research was conducted. On the bottom, the percentage of papers citing a
particular motivation is shown. On the left side the year to which each istogram
refers is reported. Finally the right side gives, within round parenthesis, the number
of papers on neuromorphic computing published during the years that row refers
to. Picture taken from [50].

chips, foreseeing the arrival of an era in which neuromorphic accelerator would

have become a fundamental device for machine learning tasks.
2. Parallelism. Biological brain performs lots of computations in parallel, achiev-
ing a really high throughput, so neuromorphic hardware architectures are
tailored to have an high degree of parallelism, providing simple yet efficient
computing units, called neurons, with condensed interconnections between
them, named synapses, that exploit some encoding protocol to process infor-
mation and make it possible for humans to give it a useful meaning. These
characteristics make neuromorphic hardware unique among all architectures
scientists are aware of.
3. Von Neumann Bottleneck. Modern microprocessor based systems can process
data way faster with respect to accessing it from main memory. This is known
10
 Introduction

as the Von Neumann Bottleneck. Overall throughput is reduced and the CPU
spends most of the operating time in idle state, wasting resources and power.
The problem is quite evident nowadays, with CPU operating frequencies and
memory sizes that have increased so faster with respect to the speed with
which memory and CPU communicate.

4. Real Time Performance. Neuromorphic devices find large interests in all

those domains for which real time tasks are to be handled. Real time implies
reacting within precise deadlines, that must be satisfied in order to prevent
undesired consequences from happening. Consequences may be acceptable in
case of soft real time, for which the system performance may degrade if those
deadlines are not respected, or they may be catastrophic in case of hard real
time, ultimately leading to a total system failure. Applications requiring such
performance may be related to digital image reconstruction or autonomous
robot control.

5. Low Power. This stands firm as the main reason behind neuromorphic hard-
ware development. Nowadays Internet of Things (IoT) and many embedded
systems related devices are needed to perform complex computations, yet the
power resources are limited. This is strictly related to Moore’s law apparently
fading away and Dennard scaling. Dennard observed that power density of a
given device stays the same as the transistors reduced in size, because current
and operating voltage reduce consequently. Thus number of transistors on a
given area could be doubled. This allowed the semiconductor industry to have
microprocessors which operated at always increasing frequencies during the
years, while maintaining acceptable power consumption and density. How-
ever two issues rose in the past 20 years, them being dark silicon and leakage
current impact on the more refined dynamic power consumption equation
Pdyn = αCf V 2 + V Ileak , which ultimately lead to modern high performance
processors stop at around 4 GHZ operating frequency, and making designers
move to a multi core approach.

6. Scalability. As it happens for classic Von Neumann computers, which bene-

fit a lot from having multiple cores scattered across the entire architecture,
neuromorphic devices are devised with an high degree of scalability, making
it possible to have multiple devices running in parallel and communicating
with each other, sharing information and computational resources.

7. Footprint. Depending on the application domain, a given design may be too

large to fit within given area constraints. Neuromorphic hardware helps in
keeping area cost as low as possible.

8. Fault Tolerance. Neuromorphic architectures are intrinsically fault tolerant.

11
 Introduction

This is due to the exploited information encoding and the self healing ca-
pability of such devices. Moreover they may help in reducing the impact of
process variation in the fabrication of devices, that often lead to degraded
performance or unacceptable errors.

9. Online Learning. As explained in 2.1.2, is to be intended as the ability of a

given neuromorphic device to adapt to input stimuli, refining the efficiency of
the tasks they perform, without any external intervention. Given the recent
demands for systems that may classify and process data in an unsupervised
manner, online learning algorithms have to be developed and successfully
implemented in such systems.

10. Neuroscience. Neuromorphic computing was firstly conceived to help neuro-

scientists, that is people who devote their career to the study of of nervous
system, trying to delineate its behaviour and working principles. Indeed,
simulating neuronal behaviour on classical supercomputers is simply not fea-
sibly, due to scale, speed, and associated power consumption. Neuromorphic
architectures are fundamental to perform neuroscience simulations within ac-
ceptable processing times.

Neuromorphic computing and devices are widely adopted to perform tasks which
belong to various domains, as illustrated in Figure 1.2. A few hints are given in
Table 1.1 and Table 1.2.
There is the need for technologies and algorithms which may enable the develop-
ment of more powerful computing edge devices, while keeping power consumption
as low as possible. To this extent, since neuromorphic computing seems to represent
a key factor in enabling the transition from a cloud oriented computing environment
to one which exploits small devices to perform intensive tasks, this thesis focuses
on the development and validation of an architecture that puts a RISC-V based
System on Chip (SOC) with a Spiking Neural Network (SNN) oriented hardware
accelerator.

12
Applica- Category Description
tion
Domain
Imaging Edge identify edges in a digital image, by identifying points at
Detection which brightness rapidly changes. An example is described
in [24].
Imaging Compression minimizing image size in bytes, possibly without reducing
image overall quality. An example is presented in [19].
Imaging Filtering changing an image properties to highlight certain features
or hide some others. An example is presented in [35].
Imaging Segmenta- typically used in medical imaging, to search for tumors or
tion similar illnesses, it aims at changing the image
representation according to new schemes, which make it
easier to analyze. An example is analyzed in [11].
Imaging Feature algorithms used to reduce the number of variables or
Extraction characteristics associated to an input dataset, in order to
speed up the processing phase without losing meaningful
information [14].
Imaging Classifica- probably the most widespread application, it consists in
tion Or analyzing an image, to either detect a certain class of
Detection objects and/or classify the image and associate it to a
precise category, according to the objects being detected.
An example is described in [40].
Speech Word technologies and algorithms devoted to analyze spoken
Recognition language to recognize words and translate them into
machine readable format and text. Some recent
advancements have been made, like algorithms that can
generate images according to a given text based
description, as reported in [41].
Data Classifica- the scope of this kind of this category of algorithms is to
Analysis tion analyze input data and associate them to a specific class,
through the assignment of a label. An example could
consist in having a system that determines whether a given
email is purely spam or not. An example is given in [49].
Imaging, Anomaly algorithms to identify events or data that deviate from a
Control, Detection given dataset normal behaviour. An example is presented
Security in [9].
Neuroscience Simulation neuromorphic architectures are way better suited to
Research provide scientists with meaningful simulations that can be
conducted in a reasonable amount of time. An example is
reported in [10].
Visual architectures tailored to run algorithms that improve or
Systems repair tissues associated to sense of sight. An example is
presented in [3].
Auditory architectures tailored to run algorithms that improve or
Systems repair tissues associated to sense of hearing. An example is
described in [34].
Olfactory architectures tailored to run algorithms that improve or
Systems repair tissues associated to sense of smell. An example is
given in [27].
Somatosen- architectures tailored to run algorithms that improve or
sory Systems repair tissues associated to sense of touch. An example is
analyzed in [33].

Table 1.1: Machine Learning Applications. Taken from [50]

 Introduction

Applica- Category Description

tion
Domain
Medical as discussed in [53], machine learning is changing the
Treatment healthcare sector around the world. Indeed, sophisticated
technologies are being used to create new drugs to cure
patients more quicky, provide wearable devices that
manage blood sugar levels in people affected by diabetes,
increase speed and accuracy of breast cancer diagnosis.
Brain- A Brain Machine Interface (BCI) between a wired brain
Machine and a custom device, which enables the user to translate
Interface its brain activity into messages that can be recognized by
the custom device and used to execute some sort of task
The aim of such system is to enhance or assist human
cognitive or motor functions [57].
Robotics Imitation like toddlers, also robots and alike can imitate some
Learning behaviours of a given target, by replicating the observed
patter. These patterns are typically known as Bayesian
patterns [12].
Central Central Pattern Generators are biological circuits that
Pattern produce basic rhythmic patterns without any external
Generation rhythmic input. They constitute the fundamentals by
(CPG) which walking, slimming and other basic motor behaviours
are possible. Such circuits are often built by exploiting
neuromorphic hardware. An example is described in [15].
Control Machine subset of machine learning methods to improve control
Learning systems, making them optimal. Examples of such
Control techniques include altitude control of satellites and
(MLC) feedback turbulence control in aerodynamics. An example
is presented in [16].

Table 1.2: Machine Learning Applications - Part 2. Taken from [50]

14
 Introduction

Figure 1.2: A graph showing application domains for which neuromorphic devices
have been developed. The size of boxes is proportional to the number of papers
that have been published for that domain. Picture taken from [50].

15
Chapter 2

Technical Background

This chapter covers all concepts and aspects that are needed to understand the the-
sis workflow, from what machine learning is, how and why it is exploited nowadays,
up to employed Instruction Set Architecture and the digital hardware architectures
involved.

2.1 Machine Learning and Applications

2.1.1 Machine Learning
The term Machine Learning refers to the possibility of building up systems that
are able to learn and adapt their parameters, fine tuning their behaviour in order
to fulfill a given scope, without being programmed to do so. Then there is deep
learning, which is a subset of tools taken from the machine learning domain [54],
such as neural networks, organized in layers, and brain inspired algorithms, to sam-
ple and elaborate information with as many details as possible as the information
proceeds from a layer to another [58].

2.1.2 Neural Networks

An Artificial Neural Network (ANN) is a biologically inspired computational model
that is used to solve complex tasks. The term neural network was coined back in
1944, when Warren McCullough and Walter Pitts, former researchers belonging to
the University of Chicago, proposed the first ever computational model of a human
brain neuron [37].
As depicted in Figure 2.1, a neuron structure computes a linear combination of
inputs, all of which are either 0 or 1, called g, which value is then fed to function
17
 Technical Background

Figure 2.1: McCulloch & Pitts Neuron Model. Picture taken from [8].

Figure 2.2: Biological Neuron Structure Picture taken from [8].

f, that will determine the outcome y, according to the equations 2.1 and 2.2
n
X
g(x) = (x1 , x2 , ..., xn ) = xi (2.1)
i=1
(
f (g(x)) = 1, if g(x) ≥ θ
(2.2)
0, if g(x) < θ
The aforementioned inputs are to be intended as parameters that function f
needs to determine whether the neuron has to ”fire”, where firing is to be nowa-
days intended as informing all neurons connected to y output that something has
18
 Technical Background

Figure 2.3: Generic Neural Network

happened, typically passing the value of function g as input to them, by means of

the synapses. This process takes inspiration from real biological neurons, which in-
ternal architecture is briefly sketched in Figure 2.2. As it can be seen, the dendrites
represent the input parameters, processed by the soma, resembling the g function
which output is transmitted by means of the axon to a given number of neurons
through synapses, that act as interconnection points. As said in Section 2.1.1, a
common neural network is organized in layers, composed of a multitude of neurons,
which change in number according to the selected layer and are connected according
to a given scheme through synapses. As illustrated in Figure 2.3, each layer has the
scope of processing a certain aspect of the input information, identify some feature
or pattern, and inform the subsequent layer with refined data, up to the point in
which the system is able to correctly compute output features from neurons in the
rightmost layer. In this regard, it is necessary to introduce the following concepts
related to the network learning process

• parameterized : a learning algorithm that is characterized by a predetermined

number of parameters, which can only change in value during the learning
process, so they are independent from the data being fed in.

• unparameterized : a learning algorithm that is characterized by a non fixed

number of parameters, which can increase or decrease as the learning process
goes on, so they are determined by data being fed in.

• supervised : direct imitation of a pattern between two datasets, so that the

19
 Technical Background

system feeds in data that is known in order to let the neural network output
data the system would like to predict, transforming one dataset into another.
• unsupervised : transforms one dataset into another, as if happens in the
aforementioned supervised learning, but in this case input data or its charac-
teristics are unknown. This implies that an unsupervised learning algorithm
typically clusters input data into groups, according to some specific features
it finds during the processing phase.
• offline : the input examples set is fixed in size, so the neural network is
fed with one example at a time and synapses weights are changed according
to a cost function. Once the global cost function is minimized, learning is
interrupted and the neural network is deployed. It will just perform infer-
ence on the incoming input data, leaving synapses weight fixed for the whole
operational time of the device.
• online : the neural network model learns and adjust its interconnection after
one input data is processed, so the model continuously learn and adapt. Thus,
changes made to synaptic weights solely depend on the actual input data and
possibly on the current model state.
Let’s make an example ([54]). Let’s imagine having a set of objects to be fed into
geometrically shaped holes (e.g. squares, triangles). babies would probably take an
object, and try to put it inside any hole, until it perfectly fits; this is an example of
parameterized learning. Teenagers, instead, would probably count the number of
sides of each object and look for an hole with that number of sides, before trying
to feed the object it; this is unparameterized learning. In most cases, one could say
that parameterized learning is all about trial and error, whereas unparameterized
one is about counting features. Finally let’s have a little digression on the evolution
history of Artificial Neural Networks (ANNs).

Spiking Neural Networks

While first and second generation of ANNs have inputs to each neuron consisting
of the sum of the incident values multiplied by some weight, third generation ANNs
inputs consist of pulse spikes arriving randomly, which indeed are values of arriving
spikes multiplied by the weights of the preceptors, hence the name spiking neural
networks. First emerged during research to model human brain behaviour, they
are still being explored and are not widespread, due to
• lack of efficient supervised learning algorithms
• a so called ”killer” application is yet to be identified. A killer application is
a task for which SNNs would outperform any other kind of neural network,
not only from a power consumption point of view.
20
 Technical Background

Still, there are a few advantages that make them stand out among the possible
neural network models

• power consumption. The fact that computation in SNNs consists of spikes

which are sparse in time and space and that converting Convolutional Neural
Networks (CNNs), which represent state of the art architectures in the com-
puter vision domain, into SNNs provides similar computational capabilities at
lower power cost, constitute another reason for which the neuromorphic sci-
entists push towards the widespread of SNNs in domains like that of Internet
of Things (IoT), where power capabilities are tightly constrained [52].

• encoding capabilities. As reported in 2.3.7, SNNs allow for the implemen-

tation of biologically inspired learning algorithms, which allow for the so
called online learning, meaning that synaptic interconnections are reinforced
or weakened according to the cause-effect relationship that incurs between
presynaptic and postsynaptic neurons firing events in the network. More-
over, some algorithms have been developed to provide unsupervised online
learning, which gave the possibility for reaching state of the art recognition
rates when dealing with MNIST database [36]. That said, another advantage
comes from the possibility of encoding inputs and outputs either in time or
rate of the spikes that characterize neurons.

2.2 The RISC-V Instruction Set Architecture (ISA)

Originally conceived at Berkeley EECS department, RISC-V represents the most
discussed Instruction Set Architecture (ISA) among both academic and industry
communities, and there are several valid reasons for it to be so:

1. completely open, meaning that specification are publicly available and no roy-
alty fees are due to the RISC-V Foundation, which is a non profit foundation
which shall keep the ISA stable, preventing it from being abandoned, as it
happened for other closed source ones [42].

2. avoids any dependence on a particular microarchitecture or specific fabrica-

tion technology

3. easily extensible to support variants that are more suitable to the desired do-
main of application, starting from a base yet complete integer ISA (RV32I/RV64I),
which is usable as standalone for customized accelerators or educational pur-
poses. This means it should suit all kind of hardware, be it a small micro-
controller or a powerful supercomputer.
21
 Technical Background

4. support for highly parallel multicore implementations, such as heterogeneous

multiprocessors, optional variable length instructions to both provide a denser
instruction encoding and expand available instruction encoding space
5. provides support for hypervisor and privileged instructions, for application
domains in which these are required
The name RISC originates from the acronym of the Reduced Instruction Set Com-
puter project, led in the same university by David A. Patterson and Carlo H. Sequin,
which aim was exploring an alternative to the general trend toward computers with
increasingly complex instruction sets, commonly identified as CISC, which further
evolution were accompanied by an always increasing hardware complexity. The
RISC approach reduces the instruction set and the available addressing modes,
causing larger code size but also reducing design time, number of design errors,
and execution time of individual instructions in terms of required clock cycles, be-
ing overall advantageous with respect to CISC. The latter family are nowadays
maintained as compatibility shell of modern processors, which instead internally
rely on a RISC one to improve performance. The V in RISC-V underlines the
main goal of this ISA, being variations, apart from being the fifth instalment in
the RISC family developed at Berkeley, whose predecessor were RISC-I, RISC-II,
SOAR and SPUR.

A core, that is any component including an independent instruction fetch unit

(IFU), serves one or more threads, identified as harts within RISC-V, by means
of multithreading of some kind. A core might provide support for additional spe-
cialized instructions, either directly or through a so called coprocessor, which is
in strict communication with the RISC-V core, performing some domain specific
tasks. There are 4 base instruction sets families, each characterized by the width
of the integer registers and the corresponding size of the address space and the
number of integer registers. Integer registers width is indicated as XLEN. Each one
uses two’s complement representation for signed integer values. Every ISA version
has a RV prefix, followed by the width of employed data and a sequence of letters,
among
• I → integer instructions
• E → integer instructions, but registers are restricted to be only 16 (x0-x15).
Used for low end implementations for which register files occupy a significant
amount of chip area.
• M → integer multiply and divide instructions
• A → 32 bit address space and integer instructions
• F → single precision (32 bit) floating point instructions
22
 Technical Background

• D → double precision (64 bit) floating point instructions

• Q → quad precision (128 bit) floating point instructions. Requires both F

and D extensions implementation.

• C → 16 bit compressed integer instructions

Treating the 4 base ISAs as distinct ones has advantages and disadvantages : in-
deed, it leads more complicate hardware needed to emulate one ISA on another
(e.g. RV32I on RV64I), but it allows for an higher degree of optimization for each
specific ISA, without the need for each one to support all operations of the one on
which it is built upon.
In order to support customization, through the addition of any extension to the
chosen basic ISA, three instruction set categories have been outlined: standard,
reserved, custom. The former are those defined by the RISC-V Foundation, so in-
structions which encodings are not interfering with each other; reserved instructions
are those that will be used in the future, still undefined. The latter are those that
use only custom encodings and extensions, or non conforming extensions, meaning
that it uses any standard or reserved encoding space, and are generally available for
vendor specific extensions. Any RISC-V implementation must include at least one
base integer subset (RV32I, RV32E, RV64I) and any of the above listed extensions.
Instructions in the base ISA have the lowest 2 bits set to 11. Compressed C in-
struction sets have their lowest 2 bits set to 00,01,10, as depicted in Figure 2.4.
Instruction sets using more than 32 bits have additional low order bits set to 1 (48
bit r → 011111, 64 bit → 0111111, 80-176 bits → xnnnxxxxx0111111, reserved for
at least 192 bits → x111xxxxx1111111).

Figure 2.4: RISC-V Instruction Length Encoding Schemes. Picture taken from
[47].

Base ISAs use either little endian or big endian memory systems, with the
23
 Technical Background

privileged architecture further defining bi-endian operations, but the IRAM is filled
by 16 bit little endian parcels, regardless of memory system endianess, to ensure
that length encoding bits always appear first in halfword address order, allowing
the length of a variable length instruction to be quickly determined by a IFU that
examines only the first few bits of the first 16 bit instruction parcel.

2.2.1 RV32I - Base Integer ISA

Figure 2.5: RV32I Instruction Formats. Picture taken from [47].

RV32I is the most basic instruction set that is provided and that must be im-
plemented in any case. It was designed to reduce the hardware needed by the most
basic application and to support modern operating systems. It handles 32 32-bit
registers, from x0, that always contains 010 and can be used to discard an instruc-
tion result, to x31, plus the program counter pc, all of which are XLEN wide, with
XLEN representing the data width in a given ISA version, so 32 bit in this case.
Although there is no precise indication on which register must hold the return ad-
dress of a given subroutine or the actual stack pointer, that is the address of the
stack holding the passed subroutine variables from top down, the standard software
calling convention uses x1 as link/return register and x5 as alternative one, plus
x2 as stack pointer. The instruction formats were designed so to ease the decoding
phase as much as possible by:

• keeping source rs1, rs2 and destination rd registers numbers positions fixed
between an instruction class and another.

• having immediate MSB always at position 31 (leftmost) to speed up the

sign extension operation which involves immediates of all instructions, all
located to the left end of the instruction array. The sign extension is not
performed on Control Status Register (CSR) instructions. This organization
has been chosen because instruction decoding is often part of critical paths
in microprocessors, so it helps in reducing them at the expense of moving
immediate bits from an instruction format to another, a property already
24
 Technical Background

seen in RISC-IV, also known as SPUR. Zero extension for immediate values
is not available as the authors did not find any real advantage in providing it
in nowadays applications.

There are a total of four main formats R, I, S, U, plus two variants names SB
and UJ, which differ from S and U for the immediate encoding scheme, respectively.

2.2.2 Choices and Consequences

In this part, technical analysis is conducted on the consequences originating from a
series of choices that characterize the RISC-V ISA, which ultimately led to choosing
it as the target ISA for this work.

• Cost : history showed that most companies chose to make their Instruction
Set Architectures grow over time, following an approach that is known as
incremental ISA. It provides for having an instruction set that increases its
instruction count over time, always keeping instruction that had been pre-
viously developed. The reason is that they are eager to maintain binary
compatibility, that is the possibility of running very old software on modern
processors, no matter the consequences and the strain that derive from such
choice. As [42] reports, the consequence for Intel has been to have a tremen-
dous amount of assembly instructions (about 3000), which include a few to
support the old fashioned and long abandoned ones to support Binary Coded
Decimal (BCD) arithmetic, at the expense of power and occupied area, the
latter influencing the die cost quadratically. Moreover one should consider
that yield, that is the number of dies per wafer that are working, decreases as
the area of the production wafer increases. On the contrary, the RISC-V ar-
chitects decided to have a single and basic integer ISA, yet complete for most
applications, called RV32I, which is guaranteed to be stable, so it won’t ever
change in the future. To make the ISA suitable for other applications, be it
power constrained ones or requiring very high computing capacities, they pro-
vide other modular extensions, that can be attached on top of RV32I, keeping
them optional, and labeling the target ISA according to the extensions being
used (e.g. RV32IM indicates the support for multiply instructions), if any.
This choice makes it possible for compilers to produce efficient code that bet-
ter suits the hardware it runs on and guarantees that new instructions will
be added by the RISC-V foundation only if there are valid technical reasons
that justify the introduction of such instructions, after they are discussed by
a dedicated commission.

• Simplicity : coming up with complicate instructions, trying to have one

macro instruction doing multiple smaller ones , doesn’t really pay. Indeed,
25
 Technical Background

most compilers are optimized so to exploit simpler yet effective instructions

whenever possible, so they could translate a given program portion in ways
one could not expect, as if they neglect specialized instructions that wouldn’t
carry any advantage with respect to using multiple simpler instructions that
serve the same scope.
• Performance : although one could think that fewer instructions could imply
lower execution time, truth is this is seldom true. Indeed, it is true that a
program compiled on a processor supporting a simple, RISC oriented, ISA
often leads to more instructions being necessary, but it is also true that ei-
ther instructions will be executed at a faster rate, because of the higher clock
frequency achievable by the underlying processor, or the machine is able to
execute more simpler instructions per clock cycle, so it has Cycles Per In-
struction (CPI) that is lower with respect to that of the machine exploiting
a complex ISA. Moreover one should consider that in the end simple instruc-
tions are the most used ones, thus making simplicity a key factor in the choice
of an ISA for an application.
• Isolation of Architecture from Implementation : an hardware archi-
tect should not add features that are particularly helpful in a precise context
rather than in another, nor should neglect the consequences that those fea-
tures might have on other kind of application domains. An example coming
from [42] concerns the delay slot or delayed branch that originates from ar-
chitects of MIPS-32 ISA. First, it is important to state what jump or branch
instructions are, why they are needed and what is the main issue concerning
their execution. A jump instruction makes the processor alter the execu-
tion flow, bypassing the instruction which comes just after the jump one,
and ”jumping” to the one which is located at address indicated by the jump
instruction. A branch instruction works similarly to a jumping one, but it
runs only if a given condition, specified in the instruction, together with the
target address, holds true when the branch instruction is fetched and pro-
cessed. However, the condition outcome can be determined only in a phase
that comes after the fetching stage, and it depends on whether the variables
(i.e. registers in this case) involved in the condition check are already filled
with the needed value or not. This situation may lead to a stall, meaning that
the instructions being fetched after the jump or branch one may be stopped
for a number of clock cycles that cannot be determined at compile time and
strongly depends on the factors discussed above. A delayed branch serves the
purpose of having an useful instruction placed right after a branch instruc-
tion (i.e. instructions that alter the program flow, so either a branch or a
jump), so that the stall that would originate from the execution of the jump
or branch instruction is compensated by running an instruction which would
be executed in any case, regardless of the branch condition outcome. As the
26
 Technical Background

authors note, this feature helps when there is the need for loading multiple
values from the memory in common single instruction fetch processors, but
could impact on the performance of superscalar ones, that rely on scheduling
of such load instructions in parallel to improve the throughput of the system.

• Program Size : a program size, in terms of occupied bytes, is one of major

concerns of system architects. Indeed, smaller programs require smaller area
on a chip and lower power, as on chip SRAM access needs lower energy with
respect to off-chip DRAM, and it often leads to fewer chances of cache misses.
As previously discussed, RISC-V standard dictates that instructions are 32
bits wide, although 64 bit and 128 bit extensions are provided, the latter being
in ongoing development. [42] states that RISC-V typically leads to programs
that are about 10% larger with respect to those written in x86-32 ISA when
all instructions are 32 bits long, but the latter proves to be inefficient when
dealing with compressed instructions, giving programs that are about 26%
larger than RISC-V ones.

• Ease of Programming : variables and temporary data are stored into reg-
isters, as they are faster in access with respect to common memory. To this
extent, RISC-V provides 32 integer registers, whereas other proprietary ones
such as ARM-32 or x86-32 have 16 and 8, respectively. Having 32 registers
has been proved to be enough for modern applications, although one must
note that the cost associated with having an higher number of registers back
in the days in which MIPS and ARM were born was too high to make 32
registers affordable. This makes it easier for programmers and compilers to
handle complex programs. Moreover, RISC-V instructions typically take one
clock cycle, if one ignores cache misses and some other amenities, whereas
complex instructions from Intel x86-32 may take way more cycles, and this
rises an issue when programming embedded devices, as programmers may
want to have more or less precise timing in such applications. Least but not
last, modern applications benefit a lot from support for Position Indepen-
dent Code (PIC). PIC means having a block of code that can be correctly
executed regardless of the absolute address, that is the compiler assigned a
specific and explicit starting address to the block of code. This is generally
possible thanks to Program Counter (PC) relative addressing, and it is often
used to support shared libraries.

• Room for Growth : back in the 70s, Gorgon Moore, co-founder of Intel
and Fairchild Semiconductor, predicted that the number of transistors being
integrated on a single chip would double every year, and this remained true
until very recently. Back in those days, designers were concerned with squeez-
ing the number of instructions required per program, to reduce the execution
time of that program, according to the equation instructions
program
· CP I · fck , with fck
27
 Technical Background

being the processor clock frequency. Today, Moore is slowly fading, mainly
due to technological issues that hinder the die manufacturing process. Given
this, RISC-V architects chose to make the ISA as modular as possible, having
enough room for optional and custom extensions that would help the imple-
mentation and usage of domain specific accelerators, such as the one involved
in this work. In this regard, a large part of opcode space for base RV32I ISA
has been reserved for the integration of specialized coprocessors, that may
not need any extension but the basic RV32I and a few specialized instruction
to perform the tasks they were designed for [46].

2.3 ODIN : A Spiking Neural Network Copro-

cessor

Figure 2.6: ODIN Architecture. Picture taken from [7].

ODIN stands for Online-learning DIgital spiking Neuromorphic processor. It is

a neurosynaptic core which supports up to 256 neurons with the possibility o The
concept of neural network is nowadays spread everywhere. Commonly meant as
a mean to provide what is called artificial intelligence, neural network constitute
28
 Technical Background

one of few state of the art technologies which are constantly growing in complex-
ity and efficiency to overcome new challenges, providing fascinating services and
features, ranging from image classification and manipulation, up to speech recogni-
tion and many more that are still being explored. A neural network is a computing
system which takes inspiration from the human brain, exploiting its parallel in-
terconnections to solve complex data problems, modifying its internal parameters
(training phase) in order to recognize unknown input data with higher accuracy
(inference phase). Spiking Neural Networks (SNN) represent an emerging class
of neural networks, coming from neuroscience research and aiming at accurately
reproducing both static and dynamic behaviours of human brain neurons. Ini-
tially conceived to help neuroscientists enrich their knowledge on the human brain
structure and working principles, SNNs have gathered attention in the computer
science f all-to-all synaptic interconnections, for a total of 28 ∗ 28 = 64k synapses,
thus emulating a crossbar which allows every neuron to be connected to every other.

Figure 2.7: AER general working scheme

The communication between ODIN and other external modules happens by

means of an input and an output interface that implements the Address Event
Representation (AER) protocol. AER stands for Asynchronous Event Represen-
tation and it is a protocol originally conceived for neuromorphic devices to allow
exchanging information on asynchronous events between computing devices run-
ning operations in parallel. As depicted in Figure 2.7, an AER transaction consists
of

• ADDR: an N bit wide bus providing an event ”address”, which is made up of

control and data bits.

• REQ: a single line that is asserted once ADDR is ready and stable, in order
to request attention for that event to the module it is connected to. When
the ACK signal goes to logic 1, then the REQ line is lowered to logic 0.

• ACK : a single line that is asserted once the requested event has been received
and completely processed. As soon as the REQ signal is de-asserted, this line
can be driven to logic 0 as well.
29
 Technical Background

The lines can be an input or an output, depending on whether they are used to
receive data from external devices or to provide it to other modules, as shown in
Figure 2.6.
Although the standard [55] dictates that every involved unit must be identified
through a unique address, in this work ODIN is present as the only AER compatible
device, so there is not really need to define such address. An address event consists
in packing up information of a given spiking event and transmitting it over to
ODIN. The message packet contains various information, depending on the event
being represented and trasmitted, as will be detailed in subsections concerning
AER OUTPUT and CONTROLLER modules. The coprocessor architecture is
illustrated in 2.6 and consists of the following modules

2.3.1 Supported Neurons Models

ODIN supports two neuron models, the former being well known and simple,
whereas the latter is a custom model that is able to reproduce main Izhikevich
behaviours while being way more computationally efficient with respect to stan-
dard Izhikevich or Hodgkin & Huxley ones, due to the fact that the neuron is
event driven, and the state is updated only when presynaptic neurons fire or time
references trigger it [21].

Leaky Integrate & Fire Model

Originally described by Louis Lapicque, this model is an extension of the classical
Integrate & Fire (IF), which adds a term VRmm(t) to Cm dVdt
m (t)
= Is (t)+Iinj (t)− Vm R
(t)−V0
m
equation [6], with Vm being the membrane voltage, Rm is the membrane resistance,
Is (t) indicates the membrane input synaptic current, Iinj (t) is a current injected into
the neuron by an intracellular electrode, and V0 is the membrane resting potential.
That leaky term takes into account the leaky phenomenon which characterize the
neuron membrane, that is not a perfect insulator, leading to its potential decaying
over time. The neuron increases its potential from time to time, according to the
stimuli given by Is , until it reaches a given threshold Vth , after which the membrane
potential is reset to V0 and the neuron ”fires”, that is an impulse is emitted at the
output (i.e. through the axon) and transmitted to all neurons that are connected
to it, according to the synaptic weights. Since the models has just one equation, it
cannot exhibit dynamics such as bursting or phasic spikes, plus it only fires spikes
with a fixed latency, given by the Vth threshold ([31]). Section 3.3.2 at [7] contains
a complete description of parameters to be set up for this neuron model.

Custom Phenomenological Izhikevich Inspired Model

A custom neuron model has been designed in order to bridge the area gap between
digital and analog implementations of neuron models [21]. It allows for reproducing
30
 Technical Background

20 Izhikevich behaviours [31], and it is event driven, reducing the computational

effort and resources needed to handle the Hodgkin-Huxley and Izhikevich neuron
models equations. At the time being, details on its parameters and working princi-
ples are not yet openly published by the author of ODIN, but some effort has been
put into extracting some information from the verilog files. Although the docu-
mentation inside [7] is sufficient to the extent of the present work, a table listing all
parameters of the custom phenomenological Izhikevich based neuron model, which
is reported in Tab 2.1 and Tab 2.2 for convenience and future usage.

2.3.2 SPI Slave

The Serial Peripheral Interface (SPI) slave unit is responsible for properly setting
ODIN configuration registers, which are listed in Table 2.3, writing to and reading
from either neurons or synapses memories. The communication protocol strictly
adheres to SPI standard, according to [59], and described on page 29. Both CPOL
and CPHA are set to 0, so that transmitted data can change only at falling edge
of the clock, whereas it is sampled at the rising edge. In order to have a working
transmission and ensure that the system operations have sufficient time to elaborate
sampled data, ODIN requires the SCK clock frequency to be at least four times
lower than that of the ODIN main clock source, since some operations, such as
that needed to handle neurons and synapses memory operation, lasts 4 clock cycles
overall (e.g. WAIT → W NEUR (memory addressing phase) → W NEUR (memory
write phase) → W SPIDN).

Serial Peripheral Interface

The Serial Peripheral Interface, commonly indicated as SPI, is a serial communi-
cation protocol originally conceived by Motorola Inc. in 1972. This protocol,
as well as I 2 C, can be applied only to short interconnection (e.g. same board, or
between different boards that are very close together), because of the skew that
limits the maximum performance. Also known as 3-wire serial interface, it is a full
duplex transmission, i.e. it allows for a concurrent communication between master
and slave and between slave and master allowing for only one master and multiple
slaves, typically up to 5, because of costs. There 4 types of wire:

• Master Out Slave In (MOSI), that is the wire for data from master to slaves

• Master In Slave Out (MISO), that is the wire for data from slaves to master.
There is only one MISO, so there must be a way to avoid conflicts because of
multiple wires trying to drive it.

• SPICLK, the clock wire, always driven by the master

31
Bits Parame- Name Width Description
ter/State
Variable
0 Parameter lif izh sel 1 bit Selects the neuron model. 0 for
Custom Izhikevich, 1 for LIF.
7-1 Parameter leak str 7 bits Defines the amount by which the
membrane potential is decreased in the
case of time reference l leakage event.
The membrane potential cannot go
negative.
8 Parameter leak en 1 bit Enables the leakage mechanism.
11-9 Parameter fi sel 3 bits accumulator depth parameter for
fan-in configuration.
14-12 Parameter spk ref 3 bits number of spikes per burst.
17-15 Parameter isi ref 3 bits inter-spike-interval in burst.
18 Parameter reson sharp 1 bit inter-spike-interval in burst.
en
21-19 Parameter thr 3 bits neuron firing threshold.
24-22 Parameter rfr 3 bits neuron refractory period.
27-25 Parameter dapdel 3 bits delay for spike latency or for DAP
parameter.
28 Parameter spklat en 1 bit spike latency enable.
29 Parameter dap en 1 bit DAP enable.
32-30 Parameter stim thr 3 bits simulation threshold
(phasic,mixed,etc).
33 Parameter phasic en 1 bit phasic behaviour enable.
34 Parameter mixed en 1 bit mixed mode behaviour enable.
35 Parameter class2 en 1 bit class 2 excitability enable.
36 Parameter neg en 1 bit negative state enable enable.
37 Parameter rebound en 1 bit rebound behaviour enable.
38 Parameter inhin en 1 bit inhibition-induced behaviour enable.
39 Parameter bist en 1 bit bistability behaviour enable.
40 Parameter reson en 1 bit resonant behaviour enable.
41 Parameter thrvar en 1 bit threshold variability and spike
frequency adaption behaviour enable.
42 Parameter thr sel of 1 bit selection between O (0) and F (1)
behaviors parameter (according to
Izhikevich behavior numbering).
46-43 Parameter thrleak 3 bits threshold leakage strength.
47 Parameter acc en 1 bit accommodation behaviour enable
(requires threshold variability
enabled).
48 Parameter ca en 1 bit Calcium concentration enable.
51-49 Parameter thetamem 3 bits Defines the SDSP threshold on the
membrane potential.
54-52 Parameter ca theta1 3 bits Defines the first SDSP threshold on
the Calcium variable.
57-55 Parameter ca theta2 3 bits Defines the second SDSP threshold on
the Calcium variable.

Table 2.1: Custom Izhikevich Neuron Parameters - Part 1

Bits Parame- Name Width Description
ter/State
Variable
60-58 Parameter ca theta3 3 bits Defines the third SDSP threshold on
the Calcium variable.
65-61 Parameter ca leak 5 bits Defines the Calcium variable leakage
time constant: a number of time
reference events must be
accumulated for the Calcium
variable to be decremented. If set to
0, disables Calcium leakage.
66 Parameter burst incr 1 bit amount by which effective threshold
and calcium are incremented at
every burst event.
69-67 Parameter reson sharp 3 bits sharp resonant behaviour time
amt constant.
80-70 State inacc 11 bits input accumulator state.
81 State refrac 1 bit tells whether neuron is in refractory
period or not.
85-82 State core 4 bits contains the current value of the
membrane potential.
88-86 State dapdel cnt 3 bits dapdel counter state.
92-89 State stim thr 4 bits stimulation strength state.
96-93 State stim str tmp 4 bits temporary stimulation strength
state.
97 State phasic lock 1 bit phasic lock state.
98 State phasic lock 1 bit mixed lock state.
99 State spkout done 1 bit tells whether in current time
reference interval a spike was fired or
not.
101-100 State stim0 prev 2 bits zero stimulation monitoring state.
103-102 State inhexc prev 2 bits inhibitory/excitatory stimulation
monitoring state.
104 State bist lock 1 bit bistability lock state.
105 State inhin lock 1 bit inhibition-induced lock state.
107-106 State reson sign 2 bits resonant sign state.
111-108 State thrmod 4 bits threshold modificator state.
115-112 State thrleak cnt 4 bits threshold leakage state.
118-116 State calcium 3 bits Contains the current value of the
Calcium variable.
123-119 State caleak cnt 5 bits Contains the current value of the
Calcium variable leakage counter
(cfr. ca leak parameter).
124 State burst lock 1 bit burst lock state.
127 Parameter neur disable 1 bit Disables the neuron.

Table 2.2: Custom Izhikevich Neuron Parameters - Part 2

Register Address Width Description
Name <15:0>
SPI GATE 0 1 bit Gates the network activity and allows the SPI to
ACTIVITY access the neuron and synapse memories for
programming and readback.
SPI OPEN 1 1 bit Prevents spike events generated locally by the
LOOP neuron array from entering the scheduler, they will
thus not be processed by the controller and the
scheduler only handles events received from the
input AER interface. Locally-generated spike
events can still be transmitted by the output AER
interface if the SPI AER SRC CTRL nNEUR
configuration register is de-asserted.
SPI SYN 2..17 256 bits Configures each of the 256 ODIN neurons as either
SIGN inhibitory (1) or excitatory (0), i.e. all of their
downstream synapses either take a negative (1) or
positive (0) sign.
SPI 18 20 bits Defines the number of clock cycles required to
BURST increment the scheduler inter-spike-interval
TIMEREF counter. Useful only if the neuron bursting
behavior is used (i.e. Izhikevich model only),
otherwise this register should be set to 0 to save
power in the scheduler.
SPI AER 19 1 bit Defines the source of the AER output events when
SRC CTRL a neuron spikes: either directly from the neuron
nNEUR when the event is generated (0) or from the
controller when the event is processed (1). This
distinction is of importance especially if SPI
OPEN LOOP is asserted.
SPI OUT 20 1 bit Enables automatic neuron and synapse state
AER monitoring through the AER output link (Refer to
MONITOR 2.3.4).
EN
SPI 21 8 bits Neuron address to be monitored if SPI OUT AER
MONITOR MONITOR EN is asserted.
NEUR
ADDR
SPI 22 8 bits Synapse address of the post-synaptic neuron SPI
MONITOR MONITOR NEUR ADDR to be monitored if SPI
SYN ADDR OUT AER MONITOR EN is asserted.
SPI 23 1 bit Allows SDSP online learning to be carried out in
UPDATE all synaptic weights, even in synapses whose
UN- mapping table bit is disabled (Refer to Section
MAPPED SDSP ONLINE LEARNING).
SYN
SPI PROP- 24 1 bit Allows all the synaptic weights to be propagated to
AGATE their post-synaptic neuron, independently of the
UN- mapping table bit value.
MAPPED
SYN
SPI SDSP 25 1 bit Enables SDSP online learning for synapse events
ON SYN (Refer to Section 2.3.3).
STIM

Table 2.3: SPI Module Configuration Registers

 Technical Background

• SPISS, that is the slave select wire, connected to the Chip Select input of
slave and driven by the master. Of course there might be multiple SPISS,one
per each slave, or they may be absent in case of daisy chain connections. A
daisy chain network has a master and all slaves are connected in series, so
that the MOSI of one slave becomes the MISO of the following one, up to the
point in which the last slave MISO is connected to the master. An example
of daisy chain configuration is depicted in Figure 2.8.

Figure 2.8: SPI with daisy chain configuration

Note that master and slaves are always both transmitters and receivers; indeed, a
master is just a block that initiates the transmission, but it doesn’t mean it cannot
act as a receiver. Furthermore, if one data direction wire for a block is not needed,
either MOSI or MISO can be removed, but internal registers will need to have input
fixed either to high or low logic value. The slave select of the target slave is active
during the whole transmission, while others must have their SPISS deactivated and
their MISO in high impedance, ignoring whatever happens on SPICLK and MOSI.
The master will generate n clock cycles, transmitting data starting from MSB; data
is generated on a precise clock edge and is sampled on the opposite edge, according
to the following parameter

• CPHA=0 → data change on trailing (second) edge, and is sampled on the

leading (first) one

• CPHA=1 → data change on leading (first) edge, and is sampled on the trailing
(second) one

Once the LSB is received, the slave SPISS is deactivated, and SPICLK returns to
idle level.
Master and slaves must agree on the number n of bits to be transferred, which is
typically equal to 8. One bit is transferred per each clock cycle, exploiting NRZ-L
encoding. The SPICLK is idle while there is nothing to be transmitted, according
to the value of parameter CPOL

• CPOL=0 → clock is low while idle

35
 Technical Background

Figure 2.9: CPOL and CPHA define three parameters of the connection

• CPOL=1 → clock is high while idle

So in the end, before setting up a SPI interconnection, master and slaves shall
agree on number n of transmitted bits, clock frequency, CPOL and CPHA. CPOL
and CPHA define the idle clock level, the active edge for data output and the
active edge for data sampling, according to the table in Figure 2.9. SPI doesn’t
need arbitration mechanism, as there is only one master; addressing is obtained
through chip select signals and bit synchronization is achieved since the clock is
only generated by the master. However, a slow slave cannot stop the master, so the
SPICLK must be set to a lower frequency in order to make things work, and there is
no error checking protocol. Furthermore, the only shared wire is MISO, which can
be driven by multiple slaves, for which conflict should be avoided through careful
addressing, making sure that only one SPISS is active at a given time. Finally,
performance is in the order of tens of Mbit/s, and it can be implemented only on
interconnections up to tens of centimeters (e.g. SD card memory).

2.3.3 Controller
The whole system is administered by a controller. It is a Moore Finite State
Machine (FSM), as its outputs solely depends on the current state, that moves
from a state to another, depending on the ongoing operations. States are listed
hereby

• WAIT : this state makes ODIN waiting until the AEROUT bus is freed, as it
is transmitting a number of AER transactions.

• W NEUR : a write operation has been requested on neurons memory. It lasts

two clock cycles, as in the first one the memory is enabled through the Chip
Select (CS), then the write request is granted by asserting the Write Enable
(WE) signal of that memory.

• R NEUR : a read operation has been requested on neurons memory. It lasts

two clock cycles, as in the first one the memory is enabled through the Chip
Select (CS), then the memory latches the and gives back the word at that
address in the subsequent clock cycle.
36
 Technical Background

• W SYN : a write operation has been requested on synapses memory. It lasts

two clock cycles, as in the first one the memory is enabled through the Chip
Select (CS), then the write request is granted by asserting the Write Enable
(WE) signal of that memory.
• R SYN : a read operation has been requested on synapses memory. It lasts
two clock cycles, as in the first one the memory is enabled through the Chip
Select (CS), then the memory latches the address and gives back the word at
that address in the subsequent clock cycle.
• TREF : a ”time reference” event has been detected coming through input
AER bus, and will take two clock cycles, plus the one in which the state
changes to TREF, in order to be completed.
• BIST : a ”bistability” event has been detected coming through input AER
bus, and will take two clock cycles, plus the one in which the state changes
to BIST, in order to be completed.
• SYNAPSE : a ”single synapse” event has been detected coming through input
AER bus, and will take two clock cycles, plus the one in which the state
changes to SYNAPSE, in order to be completed.
• PUSH : the controller has detected an incoming AER event, that is either
”virtual” or ”neuron”. In case any of these events must be handled, their in-
formation is pushed onto the scheduler in one clock cycle and will be processed
according to the scheduler algorithm, so not necessarily in the subsequent cy-
cle.
• POP NEUR : this state is reached either when the scheduler has at least
one event yet to be processes, or it cannot hold any more events due to the
limited FIFOs memory available. If the FSM transitions to POP NEUR, then
a ”neuron spike” event has to be processed and removed from the scheduler
FIFOs once it has been successfully handled. It takes N*2 clock cycles, for
the same reasons detailed for states W NEUR or W SYN, where N is the
number of ODIN supported neurons (i.e. N = 256 by default).
• POP VIRT : this state is reached either when the scheduler has at least
one event yet to be processes, or it cannot hold any more events due to the
limited FIFOs memory available. If the FSM transitions to POP NEUR, then
a ”neuron spike” event has to be processed and removed from the scheduler
FIFOs once it has been successfully handled. It takes two clock cycles, for
the same reasons detailed in states W NEUR or W SYN.
• WAIT SPIDN : this is the state in which the controller moves when any
memory operation related to neurons or synapses has been taken care of.
37
 Technical Background

Indeed, each of these operations lasts 40 SPI clock cycles, so the controller
leaves the state only as soon as the last bit of data on SPI bus has been
correctly loaded.
• WAIT REQDN : this is the state in which the controller moves after an input
AER event has been either loaded into the scheduler (i.e. previous state is
PUSH) or it has just ended (e.g. BIST event), so the controller shall wait for
input AER REQ signal to be driven low before moving on.
The controller also handles AER requests coming from the input AER signals,
according to events listed in Table 2.4 and Table 2.5.
Ad- Ad- Ad- Event Num- Description
dress dress dress Type ber of
<16> <15:8> <7:0> Cycles
1 pre neur pre neur Single 2 Stimulates neuron at address
<7:0 > <7:0 > Synapse post neur <7:0 >with the
synaptic weight associated to
pre-synaptic neuron address
pre neur <7:0 >. Ignores the
value of the mapping table bit.
0 neur 0xFF Single 2 Activates a time reference event
<7:0 > Neuron for neuron neur <7:0>only.
Time
Refer-
ence
0 Don’t 0x7F All 2 * 256 Activates a time reference event
Care Neurons for all neurons.
Time
Refer-
ence
0 neur 0x80 Single 128 Activates a bistability event for
<7:0 > Neuron all synapses in the dendritic
Bistabil- tree of neuron neur <7:0>.
ity
0 Don’t 0x00 All 128 * Activates a bistability event for
Care Neurons 256 = all synapses in the crossbar
Bistabil- 32k array.
ity

Table 2.4: AER Input Supported Events - Part 1

2.3.4 AER Output

This module handles all outgoing AER transactions, in order to communicate with
another neuromorphic processor or let external modules receive information on
ODIN neurons and synapses status.
38
 Technical Background

Ad- Ad- Ad- Event Num- Description

dress dress dress Type ber of
<16> <15:8> <7:0> Cycles
0 pre neur 0x00 Neuron 1+ (2 * Stimulates all neurons with the
<7:0 > Spike 512) synaptic weight associated to
pre-synaptic neuron pre neur
<7:0>. Takes the mapping
table bit into account. Neuron
spike events go through the
scheduler (1 cycle for a push to
the scheduler, 512 cycles for
processing when the event is
popped from the scheduler).
0 neur {w<2:0>, Virtual 1+2 Stimulates a specific neuron
<7:0 > s,l,001} Spike with weight w<2:0>and sign
bit s (1: inhibitory, 0:
excitatory), without activating
a physical synapse. If the leak
bit l is asserted, weight
information is ignored and a
time reference l leakage event is
triggered in the neuron instead.
Virtual events go through the
scheduler (1 cycle for a push to
the scheduler, 2 cycles for
processing when the event is
popped from the scheduler).

Table 2.5: AER Input Supported Events - Part 2

39
 Technical Background

2.3.5 Scheduler

Figure 2.10: ODIN Scheduler. Picture taken from [21].

The scheduler represents one of two crucial ODIN components, the other one
being the FSM based controller, and is sketched in Figure 2.10. It is inspired to
priority based ordered First In First Out memory structures [39]. The scheduler
is responsible for handling spiking and bursting (i.e. multi ple spikes in sequence)
events, distinguishing whether they come from ODIN neurons or from other neuro-
morphic devices through the input AER interface. Every spiking event is encoded
into a 14 bit wide packet, consisting of
• spiking neuron address

• number of spikes - 1

• inter spike interval (ISI)

All single spike events are stored into the main FIFO, which can accommodate 32
events, each one storing the address of the neuron that fires. Moreover the main
FIFO has the highest priority, so its events are served as soon as they are available,
even if there are some bursting events. The latter are stocked into 57 secondary
40
 Technical Background

FIFOs, each one having space for 4 events, each one storing the corresponding
neuron address. For an exhaustive description, please refer to Section [Link] of
[21].

2.3.6 Neuron Core

Figure 2.11: ODIN Neurons SRAM Organization. Picture taken from [7].

The neuron core handles events related to neurons state change. In particular
it provides information taken from the neurons state SRAM to the neurons update
logic blocks, so that it can determine whether the neuron has to fire or not, and
updates the SRAM accordingly. The neurons state SRAM is made of 28 128 bits
wide words, for a total of 4 kBytes of memory. The memory layout is depicted
in Figure 2.11. The memory input address corresponds to the 8 bit wide neuron
address, whereas the byte to be written or read is selected according to a 4 bits
wide selection address which is give through SPI when an operation on neurons
memory is requested.

2.3.7 Synaptic Core

The synaptic core is strongly bound to information taken from and given to neuron
core. Indeed it provides proper synaptic weight to the latter and receives data
needed to determine whether the synapse associated to the event being handled in
the system is going to grow, as it made the neuron membrane potential increase, or
shrink, as it made the potential decrease. The memory is organized as illustrated
in Figure 2.12. The address is formed by intersecting the 8 bits presynaptic neu-
ron address PRE NEUR[7:0] with the 8 bit postsynaptic one POST NEUR[7:0],
taking all bits from the presynaptic one and concatenating them with bits [7:3]
of postsynaptic one, forming a 13 bit wide address; bits [2:1] of postsynaptic neu-
ron address select one of the 4 available word bytes, and the least significant bit
41
 Technical Background

Bitwise Parame- Name Width Description

location in ter/State
word
0 Parameter lif izh sel 1 bit Selects the neuron model. 0 for
Custom Izhikevich, 1 for LIF.
7-1 Parameter leak str 7 bits Defines the amount by which the
membrane potential is decreased
in the case of time reference l
leakage event. The membrane
potential cannot go negative.
8 Parameter leak en 1 bit Enables the leakage mechanism.
16-9 Parameter thr 8 bits Defines the firing threshold: a
spike is issued and the neuron is
reset when the membrane
potential reaches the thr value.
17 Parameter ca en 1 bit Enables the Calcium variable
and SDSP online learning in the
dendritic tree of the current
neuron.
25-18 Parameter theta mem 8 bits Defines the SDSP threshold on
the membrane potential.
28-26 Parameter ca theta 3 bits Defines the first SDSP threshold
on the Calcium variable.
31-29 Parameter ca theta2 3 bits Defines the second SDSP
threshold on the Calcium
variable.
34-32 Parameter ca theta3 3 bits Defines the third SDSP
threshold on the Calcium
variable.
39-35 Parameter ca leak 5 bits Defines the Calcium variable
leakage time constant: a number
ca leak of time reference events
must be accumulated for the
Calcium variable to be
decremented. If set to 0, disables
Calcium leakage.
77-70 State core 8 bits Contains the current value of the
membrane potential.
80-78 State calcium 3 bits Contains the current value of the
Calcium variable.
85-81 State caleak cnt 5 bits Contains the current value of the
Calcium variable leakage
counter.
127 Parameter neur disable 1 bit Disables the neuron if set to 1.

Table 2.6: ODIN Neuron State

42
 Technical Background

Figure 2.12: ODIN Synapses SRAM Organization. Picture taken from [7].

determines whether the upper or lower half of the selected byte is to be written
or read, according to the requested memory operation. However it seems this is
not true in the implemented architecture, as the mechanism to choose either the
upper half or lower part of the selected byte is based on a byte of masking bits, as
described in 2.3.2. In order to benefit from the usage of high density SRAMs, each
word consists of 32 bits, storing a total of 8 synapses data blocks. Indeed, each
synapse is characterized by a 3 bit wide weight value and a so called mapping table
bit, which serves the purpose of allowing a given synapse to exploit online SDSP
learning mechanism or not, thus being static in value.

Synaptic Plasticity and other Synaptic Features

Figure 2.13: STDP vs SDSP Learning. Picture taken from [21].

In the neuroscience domain, the term plasticity refers to the possibility for
biological systems to modify synaptic strengths from time to time, according to
43
 Technical Background

neurons activity. Spike Timing Dependent Plasticity (STDP) is a popular synaptic

learning algorithm that improves or weakens synaptic interconnections between
neurons according to firing events. As illustrated in Figure 2.13, the learning rule
increases the synaptic weight if the postsynaptic neuron fires after the presynaptic
one has fired, or decreases it if the posysynaptic one fires before the presynaptic one
does. Thus, the process tends to favor communications between couples of neurons
where the presynaptic one contribute to the possibility of making the postsynaptic
one fire. Equations 2.3 describe the STDP behaviour in updating synaptic weights
when needed
− ∆T
(
w → w + A+ e τ+ , if ∆T > 0
− ∆T
(2.3)
w → w + A− e τ− , if ∆T < 0
with A+ , A− being amplitudes that can be modified to scale the effects of the
learning phase, ∆T is tpostsynaptic − tpresynaptic and τ+ , τ− are time constants for
which details can be found in [5]. The Spike Driven Synaptic Plasticity (SDSP) is
a simplification of STDP. It consists in evaluating the new state of the postsynaptic
neuron at the time in which the presynaptic one fires, according to equations 2.4

(
w → w + A+ , if Vmem (tpresynaptic ) ≥ θm , θ1 ≤ Ca(tpresynaptic ) < θ3
(2.4)
w → w + A− , if Vmem (tpresynaptic ) < θm , θ1 ≤ Ca(tpresynaptic ) < θ2

The equations concerning the Ca variable refer to a rule to limit the learning pro-
cess to run under well established conditions, as described in [5]. Indeed, the Ca
variable indicates the calcium concentration in the neuron and gives some insights
into recent firing activity of the neuron. If it exceeds the ranges delimited by
θ1 , θ2 , θ3 , then the learning process could lead to a phenomenon called overfitting,
which means that the employed neural network state evolved so that is it absolutely
able to correctly predict when fed with input samples on which it was trained dur-
ing the learning process, but the prediction accuracy lowers a lot when the neural
network handles previous unseen data samples. This implies that the model actu-
ally memorized the input examples, rather than the correlation that exists between
inputs and outputs. The rule formulated by [5] belongs to the early stopping meth-
ods, a modern approach to prevent overfitting from happening. Last but not least,
the author implemented a set of rules to handle bistable synapses in such net-
works. The solution is originally described in [29] and furthermore modified in [5].
Indeed, Complementary Metal Oxide Semiconductor (CMOS) technology is ubiq-
uitous when dealing with digital designs, but is not really suited for storing analog
values for a relatively long amount of time, due to parasitic capacitances and leaky
currents that affect such structure. Thus, stored values may substantially change
over time and this is highly unacceptable. The adopted solution provides for a com-
paring circuit which matches the actual synaptic weight against a fixed threshold
and either increase or decreases the weight if the actual value is above or below
44
 Technical Background

that threshold, respectively. Please note that bistability and membrane potential
leakage mechanisms (the latter being commonly indicated as leakage mechanism
in [7]) are not automatically handled from any synaptic related logic, but specific
events have to be generated and sent to ODIN AER input interface, according to
Tab 2.4 and Tab 2.5.

2.4 Chipyard Hardware Design Framework

Figure 2.14: The Chipyard project logo

Figure 2.15: An example of two Rocket tiles, one with Berkeley Out of Order
Machine (BOOM) core and the other with a standard Rocket core, inside a complete
SoC. [2]

45
 Technical Background

Chipyard is the name people at Berkeley Architecture Research gave to an hard-

ware design framework which enables for composing a RISC-V based architecture
as complex as you may think of. Starting from the previously conceived RocketChip
generator, which allows for producing a 5 stage in order RISC-V core with a great
degree of customization, chipyard makes a clear statement: ”As the world around
us keeps growing its demand for new features, devices responsible for providing
those ones are too complex to be designed at the RTL level. There is the need for
something which eases this task, among others, and lets companies and research
groups alike build at an higher abstraction level, delegating the task of writing
clunky HDL code to an autonomous system, of which chipyard just represents one
of the many more to come”.

2.4.1 Rocket Core

Figure 2.16: The Rocket core pipeline [2]

Rocket is an in order microprocessor that supports RV64GC, with G meaning

that it supports base integer instruction with 32 registers all 32 bits wide(32I),
integer multiply and divide (M), atomic ones (A), and both single and double pre-
cision floating point operations (F and D, respectively). A Rocket tile, which is
the basic microprocessor being instantiated in Chipyard, is composed of both L1
and L2 data and instruction caches. Moreover, it consists of a configurable number
of Translation Lookaside Buffers (TLBs), pipelined floating point units, a Mem-
ory Management Unit (MMU) with support for virtual page based memory and
branch prediction mechanisms, the latter being provided through Branch Target
Buffer (BTB), Branch History Table (BTH) and Return Address Stack(RAS). The
whole core pipeline is shown in Figure 2.16and a complete overview of rocket tiles
connection inside the generated SoC is represented in Figure 2.15.

TileLink Coherent Bus Protocol

TileLink [51] is the name of the open source RISC-V oriented interconnect standard
used in Chipyard. It provides support for other ISAs, multiple masters, meaning
that multiple units can initiate a transfer operation at the same time, and is con-
ceived with the purpose of using it with specialized accelerators, not general purpose
46
 Technical Background

Figure 2.17: MESI Cache Coherence Protocol - State Diagram. Picture taken from
[32].

multicore processors only. It provides a physically addressed, shared memory sys-

tem, with coherent access to both memories and slave device. Memory coherence
indeed refers to cache coherency. Indeed, as long as the cores do not have data
caches, if one writes some data to the shared memory, the others will read the
most updated value, so coherency is guaranteed. However, if one or more cores
have some cache, this might not be the case. In fact, if one core updates a given
chunk of data inside its cache, there is no guarantee that it will update the same
data address into the shared memory, so other cores might read that address and
get an old value [48]. This is known as cache incoherency. There are a number of
protocols used to tackle with this issue, many based on the ”Invalidate” pattern,
that is when a given core updates the value of a piece of data, all other copies of
that data in other cores caches are marked as ”non valid”. This solution reduces
the buses data traffic, and is enhanced with the usage of MESI protocol, the one
upon which TileLink developed its own. MESI is an acronym deriving from the
four states into which a given core cache might reside.

• Modified : cache line present only in current cache, and it differs from the
value stored into main shared memory, so it is a so called dirty value. The
value will have to be stored back into the main memory before or then, and
the cache line originally holding it will be labeled as Exclusive once that
happens.
47
 Technical Background

• Exclusive : cache line present only in current cache, the stored value matches
that hold by shared main memory, thus it is a so called clean value. The
state of the cache line may switch to Modified is the core holding that cache
modifies the line, or to Shared if it detects that any other core requests that
data.

• Shared : cache line has a clean value and might be actually stored in other
cores caches as well, so data can be also retrieved elsewhere, and matches the
value stored in main shared memory. The line may switch its state to Invalid,
once the same data is modified in any other core.

• Invalid : line value not valid anymore, as some other core modified the value
elsewhere.

A state diagram illustrating the MESI states is depicted in Figure 2.17 to give a
clearer overview.

2.4.2 Tools & Toolchains

The frameworks consists of tools and toolchains [25] that all together make it pos-
sible to customize the System on Chip you may want to build. In the following
I’ll list all of them, giving a deeper description for those that are actually deeply
involved in my thesis. Please note that only RISCV-Tools and ESP-Tools refer to
toolchains, whereas the others are simpler tools.

• Chisel: Chisel is an hardware design language, which aims to ease hardware

modules and interconnections description while giving the user powerful fea-
tures other Hardware Description Languages (HDL) like Verilog do not offer.

• FIRRTL: Chisel teams up with another language, named Flexible Intermedi-

ate Representation for RTL (FIRRTL), which role consists in taking what has
been produced by the Chisel compiler and optimize it (e.g. removing unused
signals) before writing an equivalent yet more efficient verilog description of
the user design.

• Barstools: a collection of common FIRRTL transformations used to manipu-

late a digital circuit without changing the generator source RTL.

• Dsptools: a Chisel library for writing custom signal processing hardware, as

well as integrating custom signal processing hardware into an SoC (especially
a Rocket-based SoC).

• Dromajo: a RV64GC emulator primarily used for co-simulation and was orig-
inally developed by Esperanto Technologies
48
 Technical Background

• RISCV-Tools: a collection of software used to develop, compile and run pro-

grams based on RISC-V Instruction Set Architecture (ISA). It inclused a
functional ISA simulator, named Spike, the berkeley boot loader (BBL) to
have linux OS running on a chipyard generated SoC, and other amenities. In
this work it is used to have the RISC-V Front End Server (FESVR) running
on the chosen software emulator to power up the SoC and feed it with proper
commands.

• ESP-Tools: a fork of riscv-tools, designed to work with the Hwacha non-

standard RISC-V extension. This fork can also be used as an example demon-
strating how to add additional RoCC accelerators to the ISA-level simulation
(Spike) and the higher-level software toolchain (GNU binutils, riscv-opcodes,
etc.)

2.4.3 Simulators
The chipyard environment can take advantage of three different choices when it
comes to simulators.

Verilator

Verilator is a cycle-accurate verilog and systemverilog compiler, not really a simu-

lator. It is a software package that can convert all synthesizable, and even some be-
havioural, verilog and systemverilog constructs into either C++ or SystemC based
models. Once the model is ready, a user written wrapper must be written; it should
include a main function, instantiate the top level ”verilated” model and describe
the operations that should be taken during simulation. With the wrapper ready, a
C++ compiler such as gcc can be used to create the simulation executable, accord-
ing to a series of parameters, enabling for Value Change Dump (VCD) waveform
traces to be produced at user request, similarly to what one would pretend from a
commercial hardware language simulator. Finally note that this software is open
source, so no fees are due to use it. Moreover the authors state that is often
outperforms some commercial competitors.

VCS

VCS is a functional simulator made by Synopsys ®. Provided that the user has a
valid product license, chipyard allows them to have wrappers to build VCS based
simulators from the given Scala and Chisel files, together will all features that
are mandatory for such tools, including faster compilation times and VCD trace
support.
49
 Technical Background

FireSim
FireSIM is an open source hardware simulation platform that runs on EC2 F1
FPGAs hosted on the Amazon Web Services (AWS). Its primary target is allowing
anyone to evaluate the performances of an hardware design at FPGA specific speeds,
among other possibilities, such as simulating datacenters and having profiling tools.

2.5 Parallel Ultra Low Platform (PULP)

PULP is an acronym for Parallel Ultra Low Power, and represents a family of pro-
cessors being developed as joint effort by ETH Zurich and Università di Bologna,
which aims at supporting a wide variety of IoT (Internet of Things) related appli-
cations, striving to achieve performance comparable to ARM Cortex-M family but
with lower power cost. The idea is to exploit parallel calculus as much as possible,
organizing multiple smaller cores as clusters, up to a maximum of 16 cores per clus-
ter. A cluster contains a shared instruction cache and a shared scratchpad memory,
which provides recent data. The project started from Open RISC, then has gradu-
ally moved to RISC-V ISA, adding instructions that serve this scope and possibly
accelerators, working with a ”near threshold” technique, which aims at using the
core with the lowest possible operating voltage which still meet the performance
they expect. Many designs were taped out [60]

• Main : PULPv1,PULPv2,PULPv3,PULPv4

• PULP Experiments : Artemis, Hecate, Imperio, Fulmine

• RISC-V Based: Honey Bunny

• Mixed-signal : VivoSoC,EdgeSoc

• Approximate computing : Diego, Manny, Sid

Since the design space they are exploring is pretty huge, the open source list started
with a simple yet effective core named PULPino.

2.5.1 PULPino
Small microcontroller with no caches, no memory hierarchy and no DMA. All IPs
from PULP projects, working on FPGA. As one can see from Figure 2.18, the
core is connected to two separated single port instruction and data RAMs, each
one accessible in a single clock cycle, with no wait states, which all contribute
to provide small power consumption. An AXI4 interconnect is there to provide
connection to the two RAMs and between other peripherals, through an APB
bridge, which allows for high flexibility as long as one uses components suited for
50
 Technical Background

Figure 2.18: PULPino Architecture

it or AXI. The peripherals on the bottom left part, namely GPIOs, UART, I 2 C
and SPI master are used to communicate with external devices and are fine grained
clock gated, meaning that they can be shut down whenever not needed. The core
is in a low power state whenever there is nothing to be done, and a simple event
unit, which waits for an interrupt or an event caused by a peripheral, inhibits the
clock gating on the core and wakes it up. Furthermore a SPI slave is provided to
allow external devices to access the entire memory map of pulpino and an advanced
debug unit, accessible via JTAG, allows for debugging. Finally a BOOT ROM has
been integrated in order to allow user to use PULPino as standalone system, simply
loading a bootloader into the core through an external SPI flash memory. Moving
from OPEN RISC to RISC-V was justified by the need for easily extensible ISA,
the possibility of having less and eventually compressed instructions, which overall
helps in power consumption reduction. A couple of extensions which are worth
mentioning are

1. Hardware Loops : since a loop is typically a performance impact for a simple

core, mostly because of the branch evaluation at the end of every iteration,
hardware loops allow to set up the number of overall iterations before the
loop starts, removing that cost.

2. Post Incrementing Load and Store : load and store are often part of pat-
terns where the target addresses are repeatedly incremented, but this is done
51
 Technical Background

through a separate instruction. It can be avoided by specifying that the ad-

dress has to be incremented, so that in a single execution cycle memory gets
updated or accessed for a read and the target address is updated.

2.5.2 PULPissimo
PULPissimo is the most recent and advanced single core platform of the PULP
project. Although it hosts only one core, it is used to as the main controller for
applications in which there are multiple cores, thereby providing

• Autonomous I/O Subsystems, called uDMA (micro Direct Memory Access)

[43]. Once the main core configures it, the uDMA takes care of the peripherals
transfers.

• new memory subsystem devised to improve both performance and power con-
sumption

• support for Hardware Processing Engines (HWPEs), that is custom acceler-

ators/coprocessors

• new interrupt controller (INTC): up to 32 requests from peripherals, with

possibility of masking events, have pending interrupts, and much more.

• new peripherals (QuadSPI, Camera Interface, I 2 C, I 2 S, JTAG, GPIOs, BootROM)

• new Software Development Kit (SDK): contains tools and runtime support
for all PULP microcontrollers, with procedurs for setting up cores and pe-
ripherals, so that application developers can exploit their full potential.

RI5CY
It is an in order single instruction core, made up of 4 pipelined stages, with full
support for RV32I, RV32C and RV32M RISC-V instructions. It can be configured
to support single precision floating point instructions, that is RV32F extension.
Moreover it provides a number of specific features

• Interrupts

• Events, allowing the core to sleep and wait for an event (as seen for PULPino)

• Exceptions

• performance counters, telling the number of compressed instructions and so

on

• Debug through software breakpoints and access to all registers

52
 Technical Background

Figure 2.19: RI5CY Datapath Pipeline Stages

• Hardware loops and post incrementing load/store, as seen for PULPino

Further details provided in [18].

ZERO-RISCY

Figure 2.20: ZERO RISCY Datapath

A block diagram is shown in Figure 2.20. Derived from RI5CY, it is a 2 stage in

order core, designed to be small, efficient, and configurable through 2 parameters
to support one or multiple of the following extensions
• RV32I: base integer ISA, must be always implemented.
53
 Technical Background

• RV32E: reduced version of RV32I, targeting embedded systems.

• RV32C: 16 bit instructions.

• RV32M: integer multiply and divide instructions.

It has been designed to target application domains for which area and power are
strongly constrained. Further details are available at [13]. RI5CY has recently
being taken over by lowRISC foundation, which provides full documentation at
[17].

54
Chapter 3

Methods

This chapter provides a deep dive into the thesis workflow, giving details on each
and every step needed to have a working simulation environment and a properly
set up architecture consisting of ODIN and RocketChip interfaced by means of
SPI. Chipyard has been selected as RISC-V SoC for the thesis, because it felt eas-
ier to use and offered a very high degree of configuration choices. First, ODIN and
Chipyard will be properly downloaded and configured, then the accelerator will be
integrated inside the Chipyard compilation flow so that it can be instantiated in the
final architecture. Finally, the configuration procedure of ODIN will be explained
and applied to a specific neural network. The steps of the workflow are summarized
in the diagram of Figure 3.1. All steps up to ODIN Configuration Setup will be
outlined in this chapter, whereas RTL Simulation and Synthesis will be carried out
in Chapter 4.

55
 Methods

Figure 3.1: Thesis Workflow

3.1 Setting up ODIN and Chipyard Environments

The ODIN files and documentation are publicly available at [20]. The source files
of the architecture, listed on the right part of Figure 2.6, are located inside src
folder. In order to use ODIN, one should create a folder and type git init, followed
by git remote add origin <odin url>, where odin url is the one at [20]. Finally
execute command git pull origin master to let the Git system download everything
that’s on the master branch of the repository. No further work is needed. The files of
Chipyard project are available at [45], whereas the documentation is hosted at [44].
Considering that the thesis workflow has been tested on a Linux distribution, most
of the software dependencies listed in that documentation are already satisfied, so
that section may be skipped, and can go on with setting the chipyard installation.
First, one should create a folder and type git init, followed by git remote add
origin <chipyard url>, where chipyard url is again [45]. Finally execute com-
mand git pull origin master to let the Git system download everything that’s on
the master branch of the repository. In order to make chipyard fully operational,
56
 Methods

one should then execute the [Link] script,which will

initialize all project submodules (e.g. rocketchip, sifive-blocks etc), then the build-
[Link] one; the latter will take a while, as it builds all required toolchains
that are needed to compile and run RISC-V programs on any architecture you may
compose with Chipyard. To speed up the process, it is possible to type export
MAKEFLAGS=-j<cores>, where cores is an integer indicating the number
of cores of the CPU on which the host system is running, before executing the
toolchain building script.

3.2 ODIN Integration Inside Chipyard

Companies and researchers are used to exploit designs which someone else build
and provides, either freely or by paying for a license, in order to reach their projects
goal, often within a strict deadline. Members of Berkeley Architecture Research
(BAR) group are aware of this necessary routine, so Chipyard offers the possibility
of integrating those design, provided they are written in Verilog or SystemVerilog,
or describe their architectures using the Chisel language, if necessary. Given a com-
plete, syntactically correct and synthesizable design, the first step to take consists
in copying all .sv or .v source files of ODIN coprocessor inside generators/chip-
yard/src/main/resources/vsrc. The top level entity, that is ODIN.v in this
specific case, should be renamed to ODINBlackBox.v before proceeding.
The next move is to define a Scala file named [Link] in any folder inside
generators/chipyard/src/main/scala; In this work, it has been placed inside
the example folder, as it was done for the Greatest Common Divider (GCD) ex-
ample of Verilog design integration illustrated in the chipyard documentation. The
file [Link] serves the purpose of declaring a Chisel specified ODIN blackbox
module, that acts as a wrapper for the underlying verilog design. Every chisel
blackbox consists of
• an io field, which bundles all input and output signal of ODIN top level entity
• a constructor parameter
• a listing of all verilog resources that must be loaded to make the design work
Chisel offers the possibility of having parameterized blocks, and this is absolutely
fundamental when one or more blocks have to be adaptable to different domains of
application.
1 case class ODINParams(
2 address: BigInt = 0x2000,
3 N: Int = 256,
4 M: Int = 8)
Listing 3.1: ODIN Parameters

57
 Methods

Let’s analyze the code contained in [Link]. In Listing 3.1, the parameters to
build ODIN are declared. In this simple case, there are only three parameters; N
and M indicate the number of neurons and number of bits needed to represent the
neuron address, respectively, whereas address is the address that is to be associated
to ODIN in the memory map of the SoC.
1 case object ODINKey extends Field[Option[ODINParams]](None)
Listing 3.2: ODINKey
In Listing 3.2, a Key is declared as Option type; a key is used by Chisel to specify
which parameters are needed to customize the functionalities of a given module.
They are to be declared as Option, with default value being None, meaning that
in case the user doesn’t explicitly call ODINKey with parameters, the default values
specified in ODINParams are to be selected.
1 class ODINIO(val N: Int, val M: Int) extends Bundle {
2 val CLK = Input(Clock())
3 val RST = Input(Bool())
4 val SCK = Input(Clock())
5 val MOSI = Input(UInt(1.W))
6 val MISO = Output(UInt(1.W))
7 val AERIN ADDR = Input(UInt((2∗M+1).W))
8 val AERIN REQ = Input(UInt(1.W))
9 val AERIN ACK = Output(UInt(1.W))
10 val AEROUT ADDR = Output(UInt(M.W))
11 val AEROUT REQ = Output(UInt(1.W))
12 val AEROUT ACK = Input(UInt(1.W))
13 }
Listing 3.3: ODINIO
In Listing 3.3, the top level signals of ODIN are declared, with one bit signals being
of Bool type and multi bit ones being declared as unsigned integers, which width is
parameterized according to the value of M. The ”.W” expression serves the purpose
of converting the Scala integer value inside the parenthesis to Chisel Width type.
1 trait HasODINIO extends BaseModule {
2 val N: Int
3 val M: Int
4 val io = IO(new ODINIO(N,M))
5 }
Listing 3.4: ODINIO
In Listing 3.4 the IO bundle required by the blackbox integration is declared. No-
tice that top level parameters are indicated as well, as they are needed for the
input/output definition.
1 class ODINBlackBox(val N: Int, val M: Int) extends BlackBox(Map(”N” −> IntParam(N), ”
M” −> IntParam(M))) with HasBlackBoxResource
2 with HasODINIO

58
 Methods

3 {
4 addResource(”/vsrc/ODINBlackBox.v”)
5 addResource(”/vsrc/aer out.v”)
6 addResource(”/vsrc/controller.v”)
7 addResource(”/vsrc/fifo.v”)
8 addResource(”/vsrc/izh calcium.v”)
9 addResource(”/vsrc/izh effective threshold.v”)
10 addResource(”/vsrc/izh input accumulator.v”)
11 addResource(”/vsrc/izh neuron state.v”)
12 addResource(”/vsrc/izh neuron.v”)
13 addResource(”/vsrc/izh stimulation strength.v”)
14 addResource(”/vsrc/lif calcium.v”)
15 addResource(”/vsrc/lif neuron state.v”)
16 addResource(”/vsrc/lif neuron.v”)
17 addResource(”/vsrc/neuron core.v”)
18 addResource(”/vsrc/sdsp update.v”)
19 addResource(”/vsrc/spi slave.v”)
20 addResource(”/vsrc/synaptic core.v”)
21 addResource(”/vsrc/scheduler.v”)
22 }
Listing 3.5: ODINBlackBox

In Listing 3.5, the blackbox class is declared. Notice that the parameters are
mapped as integers, and that traits HasBlackBoxResource and HasODINIO are
appended to the class declaration. The body of the class just lists all the verilog
files needed for the design description.
1 trait ODINModule extends HasRegMap{
2
3 val io: ODINTopIO
4 implicit val p: Parameters
5 def params: ODINParams
6 val clock: Clock
7 val reset: Reset
8
9 val impl = Module(new ODINBlackBox(params.N,params.M))
10
11 val aerin ack = RegInit(0.U(1.W))
12 val aerout ack = RegInit(0.U(1.W))
13 val aerin req = RegInit(0.U(1.W))
14 val aerout req = RegInit(0.U(1.W))
15 val aerin addr = RegInit(0.U((2∗params.M+1).W))
16 val aerout addr = RegInit(0.U((params.M).W))
17 [Link] := clock
18 [Link] := [Link]
19 [Link] := clock
20
21 [Link] ACK := aerout ack
22 aerout addr := [Link] ADDR
23 aerout req := [Link] REQ

59
 Methods

24 [Link] REQ := aerin req

25 [Link] ADDR := aerin addr
26 aerin ack := [Link] ACK
27
28
29 [Link] busy := [Link]
30
31 regmap(
32 0x00 −> Seq(
33 RegField.r(1, aerout req)), // a read−only register capturing aerout req
34 0x04 −> Seq(
35 RegField.r(params.M, aerout addr)),
36 0x08 −> Seq(
37 RegField.w(1, aerout ack)),
38 0x0C −> Seq(
39 RegField.w(1, aerin req)),
40 0x10 −> Seq(
41 RegField.w(2∗params.M+1, aerin addr)),
42 0x14 −> Seq(
43 RegField.r(1, aerin ack))) // read−only, aerout addr
44
45 }
Listing 3.6: ODIN Interconnections

In Listing 3.6 a trait is needed to describe the interconnections for ODIN top
level signals. Let’s analyze this trait. An impl val is declared to instantiate the
ODIN module, then all required input and output signals are declared as well,
making sure that those related to input AER and output AER are declared as
registers. Every impl signal is configured so that it appears on the left side of the
:= assignment operator if the signal is an input to ODIN, whereas it appears to the
left side of the operator if it is a signal coming out from ODIN module. The AER
specific signals must be declared as registers because ODIN will be integrated as
a MMIO Peripheral. A MMIO Peripheral is a device which makes it possible for
the processor to communicate with it through memory mapped registers, exploiting
TileLink interconnects.
Once ODIN signals are properly interconnected, it is necessary to inform Chipyard
toolchain on how those registers are to be inserted into the SoC memory map.
In particular, aerout req will be a read-only register, 1 bit wide, placed at offset
0x00, with respect to base address of ODIN module, which is indicated by address
parameter in ODINParams class, whereas aerout ack will be a write-only register,
1 bit wide, placed at offset 0x0C. Similar considerations apply for other ODIN
signals, as illustrated into Listing 3.6 at lines 31-43.
1 class ODINTL(params: ODINParams, beatBytes: Int)(implicit p: Parameters)
2 extends TLRegisterRouter(
3 [Link], ”odin”, Seq(”ucvlouvain,odin”),
4 beatBytes = beatBytes)(

60
 Methods

5 new TLRegBundle(params, ) with ODINTopIO)(

6 new TLRegModule(params, , ) with ODINModule)
Listing 3.7: ODIN TileLink
In Listing 3.7 a class named ODINTL is declared and filled in order to properly set
ODIN to take advantage of TileLink interconnection. This is done by having the
class extend the TLRegisterRouter; the first argument ([Link], ”gcd”,
Seq(”ucbbar,gcd”), beatBytes = beatBytes) determines in which place of the
global memory map ODIN will be placed and what information will be provided
in the device tree, that is a data structure used by an operating system kernel to
detect and manage attached peripherals; the second one indicates which signals will
be seen by TileLink and the last one provides the proper constructor to have ODIN
properly connected to TileLink bus. Now it is time to make Chipyard recognize
ODIN when needed.
1 trait CanHavePeripheryODIN { this: BaseSubsystem =>
2 private val portName = ”odin”
3

4 // Only build if we are using the TL (nonAXI4) version

5 val odin = p(ODINKey) match {
6 case Some(params) => {
7 val odin = LazyModule(new ODINTL(params, [Link])(p))
8 [Link](Some(portName)) { [Link] }
9 Some(odin)
10 }
11 case None => None
12 }
13
14 }
Listing 3.8: ODIN Periphery
In Listing 3.8 a LazyModule trait is shown, containing indications on what must be
initialized before all hardware is elaborated and instantiated. In particular, since
this is a memory mapped peripheral, it is sufficient to have ODIN TileLink specific
node connected to the MMIO crossbar. Every RegisterRouter has a tilelink node
named ”node”, that can be connected to the main core peripheral bus, in order
to add the address to the memory map and all device tree info that have been
provided in Listing 3.7.
1 trait CanHavePeripheryODINModuleImp extends LazyModuleImp {
2 val outer: CanHavePeripheryODIN
3 val odin busy = [Link] match {
4 case Some(odin) => {
5 val busy = IO(Output(Bool()))
6 busy := [Link] busy
7 Some(busy)
8 }
9 case None => None

61
 Methods

10 }
11 }
Listing 3.9: ODIN Top level implementation
Listing 3.9 shows the top level trait needed to have top level ports visible. A few
things still need to be addressed before compiling the chipyard project and have
the synthesizable verilog ready.
1 class WithODIN extends Config((site, here, up) => {
2 case ODINKey => Some(ODINParams(N = 256, M = 8))
3 })
Listing 3.10: ODIN Config Fragment
Listing 3.10 describes the config fragment needed to put ODIN in whatever config-
uration class. A more configurable version would be
1 class WithODIN(N: Int, M: Int) extends Config((site, here, up) => {
2 case ODINKey => Some(ODINParams(N = N, M = M))
3 })
but the one shown in 3.10 is modified in order to have ODIN instantiated with
default values suggested by its author, as the note on ODIN documentation states
that the crossbar scales according to N and M, but other ODIN modules do not
automatically scale and proper modifications are needed if one changes either M,N
or both.
1 class DigitalTop(implicit p: Parameters) extends System
2 with [Link] // Enables optionally adding trace IO
3 with [Link] // Enables optionally adding a backing
scratchpad
4 with [Link] // Enables optionally adding the block
device
5 with [Link] // Enables optionally adding the TSI serial−
adapter and port
6 with [Link] // Enables optionally adding the sifive
UART
7 with [Link] // Enables optionally adding the sifive
GPIOs
8 with [Link] // Enables optionally adding the sifive
SPI flash controller
9 with [Link] // Enables optionally adding the IceNIC for FireSim
10 with [Link] // Enables optionally adding the initzero
example widget
11 with [Link] // Enables optionally adding the GCD
example widget
12 with [Link] // Enables optionally adding the ODIN
example widget
13 with [Link] // Enables optionally adding the
DSPTools FIR example widget
14 with [Link] // Enables optionally
adding the DSPTools streaming−passthrough example widget

62
 Methods

15 with [Link] // Enables optionally having an NVDLA

16 {
17 override lazy val module = new DigitalTopModule(this)
18 }
Listing 3.11: Chipyard DigitalTop

There are still two steps to take before ODIN can be utilized in Chipyard framework.
The first one consists in locating the file [Link] in generators/chip-
yard/src/main/scala and modify the class DigitalTop as shown in Listing 3.11,
thus adding the string with [Link].
1 class DigitalTopModule[+L <: DigitalTop](l: L) extends SystemModule(l)
2 with [Link]
3 with [Link]
4 with [Link]
5 with [Link]
6 with [Link]
7 with [Link]
8 with [Link]
9 with [Link]
10 with [Link]
11 with [Link]
Listing 3.12: Chipyard DigitalTop part 2

The last one consists in modifying the very same file, this time adding the string
with [Link] to the class
DigitalTopModule, as depicted in Listing 3.12.
1 class ODINRocketConfig extends Config(
2 new [Link] ++ // display UART with a
SimUARTAdapter
3 new [Link] ++ // tie off top−level
interrupts
4 new [Link] ++ // drive the master AXI4
memory with a blackbox DRAMSim model
5 // new [Link] ++
6 new [Link] ++ // tie off debug (since we
are using SimSerial for testing)
7 //new [Link] ++
8 new [Link] ++ // drive TSI with SimSerial for
testing
9 new [Link] ++ // use testchipip serial offchip link
10 new [Link] ++
11 new [Link] ++ // use default bootrom
12 new [Link] ++ // add a UART
13 new [Link].WithL2TLBs(1024) ++ // use L2 TLBs
14

15 new [Link] ++ // no top−level MMIO

master port (overrides default set in rocketchip)

63
 Methods

16 new [Link] ++ // no top−level MMIO

slave port (overrides default set in rocketchip)
17 /∗ new [Link] ++ // use Sifive L2 cache∗/
18 new [Link](0) ++ // no external
interrupts
19 new [Link](1) ++ // single rocket−core
20 new [Link] ++ // hierarchical buses
including mbus+l2
21 new [Link]) // ”base” rocketchip system
Listing 3.13: Chipyard RocketConfigs
Now it is time to create a new configuration inside file [Link] located
in generators/chipyard/main/scala/config, simply by having a new class, that
extends the Config base class, with all components one wants to put. Listing 3.13
shows such modification. As the Chipyard documentation suggests, the design can
be compiled by executing the make CONFIG=SmallSPIFlashODINRocketConfig
BINARY=../../tests/[Link] verilog -j4 command on a Terminal, mak-
ing sure that the working directory is sims/verilator, located inside the Chipyard
folder that has been created in 3.1. Once compilation starts, the software will deter-
mine whether they are compatible with each other or not. As soon as the process
ends, the architecture verilog files are available in sims/verilator/generated-
src/[Link] folder.
1 module ODINTL( // @[:[Link]@
284084.2]
2 input clock, // @[:[Link]@
284085.4]
3 input reset, // @[:[Link]@
284086.4]
4 output auto in a ready, // @[:[Link]
.fir@284087.4]
5 input auto in a valid, // @[:[Link].
fir@284087.4]
6 input [2:0] auto in a bits opcode, // @[:[Link].
[Link]@284087.4]
7 input [2:0] auto in a bits param, // @[:[Link].
[Link]@284087.4]
8 input [1:0] auto in a bits size, // @[:[Link].
[Link]@284087.4]
9 input [6:0] auto in a bits source, // @[:[Link].
[Link]@284087.4]
10 input [13:0] auto in a bits address, // @[:[Link].
[Link]@284087.4]
11 input [7:0] auto in a bits mask, // @[:[Link].
[Link]@284087.4]
12 input auto in a bits corrupt, // @[:[Link].
[Link]@284087.4]
13 input auto in d ready, // @[:[Link].
fir@284087.4]

64
 Methods

14 input spi slave sck,

15 input spi slave mosi,
16 output auto in d valid, // @[:[Link].
fir@284087.4]
17 output [2:0] auto in d bits opcode, // @[:[Link].
[Link]@284087.4]
18 output [1:0] auto in d bits size, // @[:[Link].
[Link]@284087.4]
19 output [6:0] auto in d bits source, // @[:[Link].
[Link]@284087.4]
20 output [63:0] auto in d bits data // @[:[Link].
[Link]@284087.4]
21 );
Listing 3.14: ODINTL changes

1 assign impl SCK = spi slave sck; // @[[Link] 115:15:[Link].

[Link]@284141.4]
2 assign impl MOSI = spi slave mosi;
Listing 3.15: ODINTL SCK and MOSI assignments

1 ODINTL odin ( // @[[Link] 128:30:[Link].

[Link]@284671.4]
2 .clock(odin clock),
3 .reset(odin reset),
4 .auto in a ready(odin auto in a ready),
5 .auto in a valid(odin auto in a valid),
6 .auto in a bits opcode(odin auto in a bits opcode),
7 .auto in a bits param(odin auto in a bits param),
8 .auto in a bits size(odin auto in a bits size),
9 .auto in a bits source(odin auto in a bits source),
10 .auto in a bits address(odin auto in a bits address),
11 .auto in a bits mask(odin auto in a bits mask),
12 .auto in a bits corrupt(odin auto in a bits corrupt),
13 .auto in d ready(odin auto in d ready),
14 .auto in d valid(odin auto in d valid),
15 .auto in d bits opcode(odin auto in d bits opcode),
16 .auto in d bits size(odin auto in d bits size),
17 .auto in d bits source(odin auto in d bits source),
18 .auto in d bits data(odin auto in d bits data),
19 .spi slave sck(qspi 0 0 sck),
20 .spi slave mosi(qspi 0 0 dq 1 i)
21
22 );
Listing 3.16: ODINTL instantiated block changes
There are some of modifications that should be performed on the file
1 sims/verilator/generated−src/[Link]/
[Link].v

65
 Methods

Indeed, one should modify the module ODINTL as illustrated in Listing 3.14,
then look for impl SCK and impl MOSI, which lie inside the ODINTL module,
and connect them according to Listing 3.15. Yet another modification should
be applied to the instantiated ODIN module; look for ”ODINTL odin” inside
the file, and modify it according to Listing 3.16. These changes are necessary
to properly interface the SPIFlash Master with ODIN. Finally, run make CON-
FIG=SmallSPIFlashODINRocketConfig BINARY=../../tests/[Link] run-
binary-debug -j4 again on the terminal, still with sims/verilator as working direc-
tory, in order to recompile the whole architecture and produce the simulation ex-
ecutable. You’ll notice that an error has been detected, as shown in the following
code snippet
1 [0] \% Error: plusarg\ file\ [Link]: Assertion failed in [Link]\ mem\ 0.
memory
This error arises due to the fact that the simulator doesn’t know where to get the
file that SPIFlash module should read to feed ODIN with configuration patterns.
To fix the error, one should issue the following command
1 /home/andrea/Documents/SNN Gianvito/Papers/RocketChip/chipyard/sims/verilator/
simulator−chipyard−SmallSPIFlashODINRocketConfig−debug +permissive +dramsim
+spiflash0=/home/andrea/Documents/SNN Gianvito/Papers/RocketChip/chipyard/
tests/spiflash [Link] +verbose −v [Link].
[Link] +permissive−off ../../tests/[Link] </
dev/null 2> >(spike−dasm > [Link].
[Link]) | tee [Link].
[Link]
Listing 3.17: Command to run ODIN + ROCKET simulation.
You’ll notice this is almost the same command issued before the error occurred.
Indeed the binary file to be read by SPIFlash master module has been specified
with a path given through variable spiflash0, as reported in Listing 3.17. If you
launch the command shown in that code portion, you’ll launch the software RTL
simulation and, after a given amount of time, the process will end, providing you
with three files
• [Link] : this
is a long listing of all RISC-V assembly instructions being executed, starting
from the base address of the bootrom, up to the end of spiflashread program.
An excerpt is shown in Listing 3.18. That string tells the instruction that is
being executed (i.e. auipc, 0x0), that it refers to the 19th simulation cycle
and core hart #0 (i.e. C0), that r10 is going to be filled with a new value and
that, a few cycles before, when the instruction was located into decode stage,
it made the processor read two registers, in this case being both r0. Indeed
each string in this file refers to the write-back stage of the 5 stage pipeline of
rocketchip, that is the last stage of execution of a given instruction, during
which the result is written back into the register file, if any.
66
 Methods

1 C0: 19 [1] pc=[0000000000010040] W[r10=0000000000010040][1] R[r 0=

0000000000000000] R[r 0=0000000000000000] inst=[00000517] auipc a0, 0x0
Listing 3.18: First line of .out file given by ODIN + ROCKET simulation, running
[Link] program.

• [Link] : this
is a log of the entire simulation, listing output of the modules involved in the
simulation. In this work it just lists debug information coming from the pro-
gram being run for simulation, that is [Link], consisting in a few
printf functions being called in specific points, so to inform the user about
the status of the simulation.

• [Link] : this
is the Value Change Dump (VCD) file, that is a dump of the waveforms in-
volved in the simulation, which can be read by any VCD reader, such as the
open source GTKWave.

3.3 ODIN Parameters Definition

As one can imagine, manually configuring ODIN can be challenging, both because
it is quite a complex piece of hardware and it is configurable through standard Serial
Peripheral Interface (SPI) protocol, which is not complex, but can lead to a long
and error prone configuration phase, due to its characteristics. To ease this phase,
a C program has been developed, allowing for simple yet effective configuration
of ODIN SPI internal registers, and loading neuron and synapse Static Random
Access Memories (SRAMs). The program is called odin configurator and can be
executed by simply calling it by a terminal on Linux. Once executed, a menu will
be provided to the user, as shown in Figure 3.5. The proposed choices are

0. Exit. Simply terminates the program.

1. Set SPI Configuration Registers. Once selected, a list of possible registers

will be shown, as depicted in Figure 3.6. In example of Figure 3.7, the value
4 is given for the least significant bits of SPI SYN SIGN registers, which
determines whether synapses of a given neuron are inhibitory or excitatory,
according to Table 2.3.

2. Add Synapse. Once selected, the routine asks for presynaptic and postsynap-
tic neuron numbers, that is the neurons one wants to connect through the
synapse that is going to be set up. Then, it asks whether to set up the map-
ping table bit and which weight value to assign to the established synapse.
This process is shown in Figure 3.3 and Figure 3.4.
67
 Methods

Figure 3.2: An example of neuron configuration using odin configurator application

Figure 3.3: An example of synapse configuration using odin configurator applica-

tion. Note that since the postsynaptic neuron number is odd, 4 bit configuration
data is moved left by 4 positions.

3. Add Neuron. When this option is selected, all parameters listed in Figure
3.2 have to be tuned. First it ask for the neuron number one wants to cus-
tomize, then all LIF specific parameters can be specified. Once determined,
the program gives back the values that were specified, together with a brief
indication on the parameters to which each byte refers.
68
 Methods

Figure 3.4: An example of synapse configuration using odin configurator applica-

tion.

Figure 3.5: ODIN configurator menu.

Figure 3.6: List of SPI configuration registers that can be set up through
odin configurator application.

One may notice that some pictures concerning odin configurator application report
the string ”little-endian” on the terminal. This is due to the fact that the program
can handle either little or big endian systems, even if most desktop systems run
69
 Methods

Figure 3.7: Configuation of SPI SYN SIGN for neurons in range [15,0]. Value 4
means that neuron number 2 will have inhibitory synapses, whereas all others in
that range will have excitatory synapses.

little endian processors. Please note that little and big endian system support is
available for synapses and neurons configurations only at this time being.

70
Chapter 4

Results & Discussion

4.1 RTL Simulation

The final designed architecture, comprising ODIN and RocketChip interfaced through
a SPIFlash Master module is depicted in Figure 4.1.

Figure 4.1: Rocket and ODIN SPI Interconnection

A Register Transfer Level (RTL) simulation has been performed to determine

71
 Results & Discussion

whether the synfire chain is correctly stimulated and working. A synfire chain is
a feed-forward (i.e. the neural network neurons are arranged so that there are no
cycles) consisting of many layers of neurons. All synapes are excitatory, meaning
that the postsynaptic neuron membrane potential increases whenever one or more
presynaptic neurons fire. Once the first layer of neurons is characterized by some
firing activity, all the subsequent layers are excited and fire, giving birth to a volley
of spikes synchronously propagating from a layer to another [1]. The synfire chain
has been chosen as a benchmark to validate the proposed architecture because of its
simple and predictable behaviour. An example of synfire chain comprising 8 neu-
rons, which constitute the network that is going to be configured in the simulation,
is shown in Figure 4.3.

Figure 4.2: ODIN SPI Slave - Transmission Details. Every write or read operation
consists of 40 clock cycles, first 20 devoted to operation address transmission, the
subsequent 20 are instead needed to transfer data that should be written some-
where, according to Tab 4.2, or it consists of data that someone requested from the
outside.

Figure 4.3: Synfire Chain Network composed of 8 neurons. This is the network
that serves as validating example for the final architecture comprising ODIN and
RocketChip.

Before running the simulation, issuing command in Listing 3.17, a C program,

shown in Listing 4.1, has been written to correctly initialize SPIFlash master device,
72
 Results & Discussion

let it configure ODIN, and gather the results from output AER interface of the
ODIN coprocessor. The original program was developed by the authors of Chipyard
framework and retained its original filename. Initially conceived to configure the
SPI master and tinker with its parameters to read data from a specific area of
memory, it has been modified to suit the needs of ODIN.
1 #include <stdlib.h>
2 #include <stdio.h>
3
4 #include ”mmio.h”
5 #include ”spiflash.h”
6 #define ODIN AEROUT REQ 0x2000
7 #define ODIN AEROUT ADDR 0x2004
8 #define ODIN AEROUT ACK 0x2008
9 #define ODIN AERIN REQ 0x200C
10 #define ODIN AERIN ADDR 0x2010
11 #define ODIN AERIN ACK 0x2014
12 #define SYNFIRECHAIN NEURONS 8
13 int main(void)
14 {
15 int i;
16
17 spiflash ffmt ffmt;
18 uint8 t neurons[SYNFIRECHAIN NEURONS];
19
20 [Link] en = 1;
21 [Link] len = 4; // Valid options are 3 or 4 for our model
22 [Link] cnt = 0; // Our SPI flash model assumes 8 dummy cycles for fast reads, 0
for slow
23 [Link] proto = SPIFLASH PROTO SINGLE; // Our SPI flash model only
supports single−bit commands
24 [Link] proto = SPIFLASH PROTO SINGLE; // We support both single and
quad
25 [Link] proto = SPIFLASH PROTO SINGLE; // We support both single and
quad
26 [Link] code = 0x13; // Slow read 4 byte
27 [Link] code = 0x00; // Not used by our model
28
29
30 printf(”Initiating ODIN configuration...\n”);
31 configure spiflash(ffmt);
32 test spiflash(0x0,0x2c6,0);
33 //test spiflash(0x0,0xabe, 0); //32 neuroni
34 //test spiflash(0x0,0x551e,0); 256
35 //test spiflash(0x0,0x36B,0); // 10 neuroni
36 /∗∗ ADATTARE IN BASE A QUANTI NEURONI E SINAPSI CONFIGURO∗/
37 printf(”ALL NEURONS CONFIGURED!\n”);
38 reg write32(ODIN AERIN ADDR,0x0021);
39
40 reg write32(ODIN AERIN REQ,1);

73
 Results & Discussion

41 while ((reg read16(ODIN AERIN ACK)) == 0);

42 reg write32(ODIN AERIN REQ,0);
43
44
45 for (i = 0; i < SYNFIRECHAIN NEURONS; i++)
46 {
47 while ((reg read16(ODIN AEROUT REQ) & 0x1) == 0);
48 neurons[i] = reg read16(ODIN AEROUT ADDR);
49 reg write16(ODIN AEROUT ACK,1);
50 while ((reg read16(ODIN AEROUT REQ) & 0x1) == 1);
51 reg write16(ODIN AEROUT ACK,0);
52 }
53 printf(”ALL NEURONS FIRED!\n”);
54

55 /∗for (i = 0; i < 10; i++) printf(”Ordine %d : %d.\n”, i, neurons[i]);∗/

56 for (i = 0; i < SYNFIRECHAIN NEURONS; i++)
57 {
58 reg write32(0x89000000+(i∗4),neurons[i]);
59
60 }
61
62
63 return 0;
64
65 }
Listing 4.1: Example program spiflashread.c.
Here are the changes that have been applied
1. observing lines 6-11, one can notice that a few preprocessor directives have
been added to define memory offsets to be used to access to ODIN memory
mapped registers, which store AER related signals. The offsets are those
defined in 3.10.
2. line 12 shows a constant that indicates the number of neurons that consti-
tute the configured synfire chain, so it depends on the number of neurons
that have been provided to odin configurator program. It should be changed
accordingly before any simulation, if needed.
3. line 20-27 show the configuration of parameters for the SPIFlash module.
Line 20 enables the command, line 21 determines the number of bytes that
make up the address field, line 22 refers to the number of dummy (i.e. no op-
erations) SPI cycles that have to be performed between a complete operation
and another. Lines 23 to 25 show configuration of protocols to be employed
for commands, addresses, and data transmission, making it possible to use
standard mono bit SPI or the more recent one, called quad SPI, which sends
4 bits per transmission on 4 serial channels. Please note that command trans-
mission doesn’t support quad SPI mode, but standard one only. Finally, line
74
 Results & Discussion

27 can be ignored, as it is a parameter not currently supported, whereas line

26 specifies the command to be sent to the SPI flash master. A list of SPI
model commands is shown is Table 4.1.

4. line 31 calls the function devoted to set internal spi flash master module pa-
rameters according to the values in ffmt structure, whereas line 32 effectively
starts the SPI transmission, for all bytes in range 0x0-0x2c6.

5. line 38-42 sets memory mapped input AER registers to values needed to
request a VIRTUAL SYNAPSE event targeting neuron 0. Note that the
AERIN REQ signal is lowered only after AERIN ACK is asserted.

6. line 45-52 handles output AER events as soon as neuron i fires. In particular,
the software continuously (i.e. polling technique, since no interrupts are avail-
able for this peripheral) reads the memory mapped ODIN AEROUT REQ
register, until its value becomes logic 1. This means that a neuron fired, and
the neuron address is stored into neurons[i]. Then the software sends an ac-
knowledge through register ODIN AEROUT ACK and waits for
ODIN AEROUT REQ to be lowered in response to the acknowledge. Fi-
nally the acknowledge register is cleared so that ODIN can handle subse-
quent events. These operations are repeated for all neurons in range [0,SYN-
FIRECHAIN NEURONS].

7. lines 56-60 properly store addresses of neurons that fired into external DRAM,
which covers the 0x80000000-0X8FFFFFFF range in the Soc memory map,
as stated into sims/verilator/generated-src
/[Link]
/[Link] device tree source
file. Thus, the defalt configuration for external DRAM make it able to store
up to 4096 Mebibyte (0x10000000 possible addresses) of data, which corre-
spond to a little more than 4 Gigabytes (GB).

Let’s proceed with a step by step description of RTL simulation of a synfire

chain composed of 8 neurons, namely neurons with addresses ranging from 0 to 7,
that start with a zeroed membrane potential and their threshold firing voltage set
to 1. Membrane potential leakage and calcium leakage are disabled. All neurons
but those that are part of the neural network are disabled as well.
Figure 4.4 shows the first step. In this picture

1 mem req valid indicates whether the SPI transition is valid or not. Every
low to high transition signals that the transmission is being executed and data
is being sent to ODIN.
75
 Results & Discussion

Command Further Modifications Description

cmd code = pad cnt = 0 and addr len=4 and addr proto= Slow read 4 byte
0x13 SPIFLASH PROTO SINGLE and data proto= address
SPIFLASH PROTO SINGLE
cmd code = pad cnt = 0 and addr len=3 and addr proto= Slow read 3 byte
0x03 SPIFLASH PROTO SINGLE and data proto= address
SPIFLASH PROTO SINGLE
cmd code = pad cnt = 8 and addr len=3 and addr proto= Fast read 3 byte
0x0B SPIFLASH PROTO SINGLE and data proto= address.
SPIFLASH PROTO SINGLE
cmd code = pad cnt = 8 and addr len=4 and addr proto= Fast read 4 byte
0x0C SPIFLASH PROTO SINGLE and data proto= address.
SPIFLASH PROTO SINGLE
cmd code = pad cnt = 8 and addr len=3 and addr proto= Fast read 3 byte
0x6B SPIFLASH PROTO SINGLE and data proto= address, with quad
SPIFLASH PROTO QUAD data.
cmd code = pad cnt = 8 and addr len=4 and addr proto= Fast read 4 byte
0x6C SPIFLASH PROTO SINGLE and data proto= address, with quad
SPIFLASH PROTO QUAD data.
cmd code = pad cnt = 8 and addr len=3 and addr proto= Fast read 3 byte
0xEB SPIFLASH PROTO QUAD and data proto= address, with quad
SPIFLASH PROTO QUAD data and quad address.
cmd code = pad cnt = 8 and addr len=4 and addr proto= Fast read 4 byte
0xEC SPIFLASH PROTO QUAD and data proto= address, with quad
SPIFLASH PROTO QUAD data and quad address.

Table 4.1: SPI Flash Module Commands. Note that cmd proto is always set to
SPIFLASH PROTO SINGLE and pad code is always set to 0, as it is unused.

76
 Results & Discussion

Figure 4.4: Synfire chain with 8 neurons: setup of ODIN SPI slave configuration
registers.

Figure 4.5: Synfire chain with 8 neurons: configuration of a synaptic interconnec-

tion by performing a write operation into the synapses SRAM.

2 mem req addr indicates the byte, coming from the configuration file,
that spi master is reading and sending to ODIN.

3 spi addr consists of first 20 bits sent through SPI master. The address
determines the operation that is to be executed by ODIN SPI slave module,
as summarized in Table 4.2.

4 spi cnt counts the internal clock cycles, in order to distinguish incoming
address bits from data related ones. Indeed, the first 20 SPI clock cycles are
devoted to the transmission of the operation address, whereas the remaining
20 are dedicated to data transmission, as depicted in Figure 4.2.
77
 Results & Discussion

Figure 4.6: Synfire chain with 8 neurons: configuration of least significant byte of
neuron 0, by performing a write operation into the neurons SRAM.

Figure 4.7: Synfire chain with 8 neurons: configuration phase ends as soon as
SPI GATE ACTIVITY is lowered. This lets ODIN run the synfire chain.

Figure 4.8: Synfire chain with 8 neurons: neuron 0 is stimulated by a virtual

synapse event (points 1-3), then every neuron of the synfire chain fires in sequence
(points 4-7).

78
 Results & Discussion

5 SPI GATE ACTIVITY is one of the configuration registers that is prop-

erly set during the simulation. Since it enables write and read operations on
neurons and synapses SRAMs, it must be set to logic 1 before providing data
to be written in those memories.

Figure 4.5 shows the second step. In this picture signals from A[12:0] down to WE
refer to neurons SRAM.

1 A[12:0] is the address referring to the synapses SRAM. Indeed, spi addr[19:0]
contains the value 64060, which means that synapse memory must be written.
In particular, synapse between neuron 3 and neuron 4 must be modified, as
the byte address is 010.

2 once synapses SRAM Chip Select (CS) signal is asserted, the target
address is latched into A[12:0]. In the subsequent clock cycle the Write Enable
(WE) is asserted as well and data provided through D[31:0], that is ready once
spi cnt is equal to 27hexadecimal = 39decimal , is written into the memory.

Figure 4.6 shows the third step. In this picture signals from A[7:0] down to WE
refer to neurons SRAM.

1 A[7:0] is the address referring to the neurons SRAM. Indeed, spi addr[19:0]
contains the value 50000, which means that neurons memory must be writ-
ten. In particular, byte 0 of neuron 0 must be modified. Once neurons SRAM
Chip Select (CS) signal is asserted, the target address is latched into A[7:0].

2 In the subsequent clock cycle the Write Enable (WE) is asserted as well
and data provided through D[127:0], that is ready once spi cnt is equal to
27hexadecimal = 39decimal , is written into the memory.

Figure 4.7 shows the fourth step. In this picture signals from A[7:0] down to WE
refer to neurons SRAM.

1 A[7:0] is the address referring to the neurons SRAM. Indeed, spi addr[19:0]
contains the value 50000, which means that neurons memory must be writ-
ten. In particular, byte 0 of neuron 0 must be modified. Once neurons SRAM
Chip Select (CS) signal is asserted, the target address is latched into A[7:0].

2 In the subsequent clock cycle the Write Enable (WE) is asserted as well
and data provided through D[127:0], that is ready once spi cnt is equal to
27hexadecimal = 39decimal , is written into the memory.

Figure 4.8 shows the last step. In this picture

79
 Results & Discussion

1 AERIN REQ is asserted to trigger the ODIN controller and let it acco-
modate the incoming AER event request if AEROUT CTRL BUSY is low.

2 AERIN ADDR indicates the type of AER event that is being requested
and information about it. In this case a VIRTUAL SYNAPSE EVENT, for
which details can be found in 2.5, is being handled and is targeted to neuron
0.

3 AERIN ACK is asserted once the virtual synapse event has been cor-
rectly handled and operations are complete.

4 Once the virtual synapse event processing comes to an end, neuron 0

is excited so that its membrane potential matches the configured threshold,
it fires and the event is sent over output AER interface, first asserting the
AEROUT REQ and AEROUT CTRL BUSY signals.

5 AEROUT ADDR is filled with the neuron number that has just fired
and generated the event.

6 AEROUT ACK is a software programmed acknowledge which informs

ODIN that the latest firing event has been correctly received from the RISC-V
CPU.

7 AEROUT CTRL BUSY indicates that the accelerator is currently un-

able to process other requests, either coming from internal modules or through
input AER interface, and is lowered only when an external AEROUT ACK
is received.

In the end a all neurons in the range {0,7} fire in sequence, and their output
AER events are correctly received and an acklowedge is sent to ODIN for each one
of them. As soon as neuron 7 fires and its event is correctly handled, the firing
sequence starts again from neuron 0, since neuron 7 is connected to it, and the
simulation is stopped. Other simulation experiments, using larger synfire chains,
have been conducted, but the one presented here was selected due to its simplicity
and shorter simulation time. Simulation of larger networks, such as 256 neurons
wide synfire chains, work similarly.

80
 Results & Discussion

Read Write Command spi addr[15:0] Description

(spi addr[19]) (spi addr[18]) (spi addr
[17:16])
- - 00 con- Configuration
fig reg addr[15:0] register write
at address
config reg addr.
1 0 01 {-, Read to the
byte addr[3:0], neuron memory
word addr[7:0]} (256 128-bit
words). Byte
byte addr[3:0]
from word
word addr[7:0]
is retrieved.
0 1 01 {-, Write to the
byte addr[3:0], neuron memory
word addr[7:0]} (256 128-bit
words). Byte
byte addr[3:0]
from word
word addr[7:0]
is written.
1 0 10 {-, Read to the
byte addr[1:0], synapse
word addr[12:0]} memory (8192
32-bit words).
Byte
byte addr[1:0]
from word
word addr[12:0]
is retrieved.
0 1 10 {-, Write to the
byte addr[1:0], synapse
word addr[12:0]} memory (8192
32-bit words).
Byte
byte addr[1:0]
from word
word addr[12:0]
is written.

Table 4.2: SPI Slave Configuration Commands

81
 Results & Discussion

4.2 Synthesis Results

This section serves the purpose of conducting a feasibility study on the entire
design synthesis on Field Programmable Gate Array (FPGA), namely Xilinx PYNQ
Z2, comparing timing information and occupied area on that FPGA against those
obtained by ECE department of University of Virginia, for which one can find more
details here, on a Xilinx Zedboard. Before proceeding, the design comprising ODIN
and RocketChip must be evicted from all unnecessary components and signals which
make it impossible to synthesize it on PYNQ Z2 board, due to the fact that it
requires too many I/O pins (i.e. the outermost input and output wires). As shown
in Listing 4.2, the following modules have been removed or modified

• no UART adapter, simply removing the [Link]

config fragment

• removed SiFive L2 cache. InclusiveCache config fragment has been removed,

so the design exploits TileLink proprietary broadcast system. To perform this
modification, just remove [Link].

• SPI Flash master has been changed to read-only peripheral, meaning that
it can only read from a given file. This is done by giving the true ar-
gument to [Link](true) config fragment.
The [Link](0x100000) fragment should be inserted as
well, with the parameter between parenthesis being the size of addressable
space for the SPI controller.

• external DRAM removed (no need to save ODIN results outside, as in Verila-
tor simulation). To do so, remove the [Link]
fragment.

• TileLink (TL) monitors removed. Simply add

[Link].

• smallest RISC-V RocketChip core available has been put in place of standard
core. To achieve this result, remove [Link]
and [Link] config fragments, then add
the [Link] one, which will instantiate the
smallest core available, use an incoherent bus interconnect rather than a co-
herent one, and remove all interconnects needed for external memories.

1 class TinyRDOnlySPIFlashODINRocketConfig extends Config(

2 // no uart
3 // no sifive l2 cache
4 // read only spi flash
5 // small core

82
 Results & Discussion

6 // tiny config
7 // no tlmonitors
8 // no external DRAM
9 // bufferless broadcast
10 new [Link] ++
11 //new [Link] ++
12 //new [Link] ++
13 //new [Link] ++
14 new [Link](true) ++ // add the SPI flash model in
the harness (read−only)
15 new [Link] ++
16 //new [Link] ++
17
18 new [Link] ++
19 new [Link](0x100000) ++ // add the SPI flash controller
(1 MiB)
20 new [Link].WithL2TLBs(1024) ++
21 //new [Link] ++
22 new [Link] ++
23 new [Link] ++
24 new [Link] ++
25 new [Link](0) ++
26 //new [Link](1) ++
27 //new [Link] ++
28 new [Link])
Listing 4.2: First line of .out file given by ODIN + ROCKET simulation, running
[Link] program.

4.2.1 Area
The synthesis result are shown in Tab 4.3 and Tab 4.4. As one can see, the resource
utilization on PYNQ Z2 board is really low, accounting for 15.99 % of LUTs slices
and 11.07 % of Block RAMs (BRAMs), the latter being instantiated for ODIN
neurons and synapses states, RocketChip data and instruction caches. The number
of I/O pins is 8
1. clock: main clock source
2. reset: global synchronous reset
3. SCK: SPIFlash Master and ODIN SPI Slave clock source
4. CS: SPIFlash Master Chip Select
5. 4 in/out pins for quad spi data transmission
and is the minimum needed to have a working design. The PYNQ Z2 provides up
to 125 user programmable I/O pins (known as IOBs), so there more than enough
to integrate other peripherals or systems.
83
 Results & Discussion

Site Type Scope Type Used Available Utilization %

Slice LUTs 8506 53200 15.99
Logic 7928 53200 14.90
Memory 578 53200 1.09
Distributed RAM 578
Shift Register 0
Slice Registers 4317 106400 4.06
Flip Flop 4317 4.06
Latch 0
F7 Muxes 179 26600 0.67
F8 Muxes 34 13300 0.26

Table 4.3: ODIN + RocketCore Synthesis - Slices

Site Type Scope Type Used Available Utilization %

BRAM Tile 15.5 4140 11.07
RAMB36/FIFO 15 140 10.71
RAMB36E1 15
RAMB18 1 280 0.36
RAMB18E1 1

Table 4.4: ODIN + RocketCore Synthesis - RAM

84
Chapter 5

Conclusions

This work was intended to establish a first contribution towards seamless integration
of neuromorphic technologies with state of the art processors, which cannot really
handle operations deep learning algorithms require to achieve the results they were
conceived for. The architecture should take the best from both domains; on one side
there are advantages coming from using a ”standard” SoC system, as the principles
behind programming such devices are known and it is for sure easier than directly
interacting with a neuromorphic architecture like that on which ODIN is built, plus
it exploits an open source ISA that is gaining more and more attention due to its
modularity, as the base ISA is stable, minimal, and won’t be discontinued in terms
of support, and simplicity, plus all positive points cited in 2.2; on another side
it allows for efficient execution of above mentioned algorithms due to the nature
of networks involved, which further pushes down power consumption, and this is
mandatory when dealing with IoT devices. All in all there is so much room for
improvements, a few of which are provided in the following
1. make RocketChip able to reconfigure ODIN at run time. This means that
proper hardware and sotware (APIs) support should be implemented and
provided to the end user.
2. FPGA porting. Chipyard leverages a boot ROM, containing the instructions
to run when the SoC is powered on, together with all details concerning the
SoC modules through the Device Tree Binary structure. The SoC runs those
instructions, then run a wait-for-interrupt (WFI) instruction, waiting for the
RISC-V Frontend Server (FESVR) to load the program and wake up the core
through an external interrupt. If one wants to deploy and run the system on
a FPGA, the booting process should be changed, removing the need for an
external interrupt and let the user program run as soon as the boot loader
completely set up the SoC modules.
3. exploit ROcket Custom Coprocessor interface instead of MMIO/SPI, first
evaluating pros and cons deriving from the possibility of exploiting custom
85
 Conclusions

coprocessor specific instructions. A RoCC accelerator exploits a custom com-

munication protocol and reserved non-standard ISA instructions, having a
specific format that is detailed in [25]. This change might help in reaching
the functionality of point 1.

4. change SPIFlash module to allow the user to specify separated input and
output data files/streams, as at the moment it can only read and write,
depending on the configuration parameters, from and to one file only.

5. change SPI slave interface of ODIN to support quad spi, thus accelerating
the configuration phase. This would imply not only modifying the external
interface of the SPI slave module, but also change the internal behaviour so
that the controller can process 4 bits at a time, rather than one.

6. have a cluster of ODIN modules talking one to each other, so to increase the
computing capabilities of such systems.

7. modify ODIN so to handle bistability and membrane potential leakage mech-

anism automatically, removing the dependency from external events.

86
Acknowledgements

This work established the first step towards the start of my professional career.
Lots of difficulties showed up and I was surrounded by a multitude of people that
supported me without even asking for. I’d start with a few words on my parents,
Marco and Milena. They truly believed in my desire to pursue a career in engi-
neering, although my dad initially thought I’d be better with studying law, and
provided me with everything a son might ask for, love above all, and I felt encour-
aged in leaving home to move to Turin. This city changed my way of being, and
acting, and fortified my personality while providing me with opportunities to grow
up. It is also the city that gave me friendships I’d never live without, and they are
countless, yet I’d like to list a few of them : Vito, Sofia, Cosimo, Samuel, Marco. A
special mention goes to Francesco, who spent hundreds of hours in the past months
listening to my concerns, providing hints on how I could face certain problems. I
could not forget to mention my best friend Aurelio, who has always been present,
no matter the distance we were at or the time that passes since the last time we
had to hear each other. I’d like to thank my uncle Carlo, the man that stands as
my lighthouse in my engineering path. Least but not last, I don’t know how to
express the gratitude I have towards Giorgia, my girlfriend since two years, who
truly understood the malaise I went through and always had kind words to relieve
it.

87
Bibliography

[1] M. Abeles. Synfire chains. Scholarpedia, 4(7):1441, 2009. doi: 10.4249/

scholarpedia.1441. revision #150018.

[2] K. Asanović, R. Avizienis, J. Bachrach, S. Beamer, D. Biancolin, C. Celio,

H. Cook, D. Dabbelt, J. Hauser, A. Izraelevitz, S. Karandikar, B. Keller,
D. Kim, J. Koenig, Y. Lee, E. Love, M. Maas, A. Magyar, H. Mao, M. Moreto,
A. Ou, D. A. Patterson, B. Richards, C. Schmidt, S. Twigg, H. Vo, and A. Wa-
terman. The rocket chip generator. Technical Report UCB/EECS-2016-17,
EECS Department, University of California, Berkeley, Apr 2016. URL http:
//[Link]/Pubs/TechRpts/2016/[Link].

[3] S. Barzegarjalali and A. C. Parker. Neuromorphic circuit modeling directional

selectivity in the visual cortex. In 2016 38th Annual International Conference
of the IEEE Engineering in Medicine and Biology Society (EMBC), pages
6130–6133, 2016. doi: 10.1109/EMBC.2016.7592127.

[4] N. Boldrini. Deep learning, cos’è l’apprendimento profondo, come funziona e

quali sono i casi di applicazione, 2019. URL [Link]
intelligenza-artificiale/deep-learning/deep-learning-cose/.

[5] F. S. Brader JM, Senn W. Learning real-world stimuli in a neural network

with spike-driven synaptic dynamics.. (English). Neural Computation, 19(11),
2007. doi: [Link]

[6] Burkitt A.N. A review of the integrate-and-fire neuron model: I. ho-

mogeneous synaptic input. Biological Cybernetics, 95(1):1–19, 2006. doi:
10.1007/s00422-006-0068-6.

[7] J.-D. L. [Link], [Link] and D. Bol. A 0.086-mm² 12.7-pJ/SOP 64k-

Synapse 256-Neuron Online-Learning Digital Spiking Neuromorphic Proces-
sor in 28-nm CMOS. (English). IEEE Transactions on Biomedical Circuits
and Systems, 13(1):145–158, 2019. doi: [Link]
2880425.
88
 BIBLIOGRAPHY

[8] A. L. Chandra. Mcculloch-pitts neuron — mankind’s first mathematical

model of a biological neuron, 2018. URL [Link]
mcculloch-pitts-model-5fdf65ac5dd1.

[9] Q. Chen, Q. Qiu, H. Li, and Q. Wu. A neuromorphic architecture for anomaly
detection in autonomous large-area traffic monitoring. In 2013 IEEE/ACM
International Conference on Computer-Aided Design (ICCAD), pages 202–
205, 2013. doi: 10.1109/ICCAD.2013.6691119.

[10] K. Cheung, S. R. Schultz, and W. Luk. Neuroflow: A general purpose

spiking neural network simulation platform using customizable processors.
Frontiers in Neuroscience, 9:516, 2016. ISSN 1662-453X. doi: 10.3389/
fnins.2015.00516. URL [Link]
fnins.2015.00516.

[11] G. Crebbin. Image segmentation using neuromorphic integrate-and-fire cells.

In 2005 5th International Conference on Information Communications Signal
Processing, pages 305–309, 2005. doi: 10.1109/ICICS.2005.1689056.

[12] Daniel Faggella. Machine learning in robotics – 5 modern appli-

cations, 2020. URL [Link]
machine-learning-in-robotics/.

[13] P. Davide Schiavone, F. Conti, D. Rossi, M. Gautschi, A. Pullini, E. Flamand,

and L. Benini. Slow and steady wins the race? a comparison of ultra-low-power
risc-v cores for internet-of-things applications. In 2017 27th International Sym-
posium on Power and Timing Modeling, Optimization and Simulation (PAT-
MOS), pages 1–8, 2017. doi: 10.1109/PATMOS.2017.8106976.

[14] DeepAI. Feature extraction. URL [Link]

machine-learning-glossary-and-terms/feature-extraction.

[15] E. Donati, F. Corradi, C. Stefanini, and G. Indiveri. A spiking implemen-

tation of the lamprey’s central pattern generator in neuromorphic vlsi. In
2014 IEEE Biomedical Circuits and Systems Conference (BioCAS) Proceed-
ings, pages 512–515, 2014. doi: 10.1109/BioCAS.2014.6981775.

[16] J. Dungen and J. . Brault. Simulated control of a tracking mobile robot by

four avlsi integrate-and-fire neurons paired into maps. In Proceedings. 2005
IEEE International Joint Conference on Neural Networks, 2005., volume 2,
pages 695–699 vol. 2, 2005. doi: 10.1109/IJCNN.2005.1555936.

[17] l. ETH Zurich, University of Bologna. Ibex documentation, 2021. URL https:
//[Link]/en/latest/[Link].
89
 BIBLIOGRAPHY

[18] o. G. ETH Zurich, University of Bologna. Cv32e40p documentation, 2021.

URL [Link]

[19] W. . C. Fang, B. J. Sheu, O. T. . C. Chen, and J. Choi. A vlsi neural processor

for image data compression using self-organization networks. IEEE Transac-
tions on Neural Networks, 3(3):506–518, 1992. doi: 10.1109/72.129423.

[20] C. Frenkel. Odin spiking neural network (snn) processor. URL https://
[Link]/ChFrenkel/ODIN.

[21] C. Frenkel. Bottom-up and top-down neuromorphic processor design: Unveil-

ing roads to embedded cognition, 2020. URL [Link]
1/226494.

[22] S. B. Furber, D. R. Lester, L. A. Plana, J. D. Garside, E. Painkras, S. Temple,

and A. D. Brown. Overview of the spinnaker system architecture. IEEE
Transactions on Computers, 62(12):2454–2467, 2013. doi: 10.1109/TC.2012.
142.

[23] W. Gerstner and W. Kistler. Spiking Neuron Models: Single Neurons, Popula-
tions, Plasticity. Cambridge University Press, 2002. URL [Link]
ch/~gerstner/SPNM/[Link].

[24] B. Glackin, J. Harkin, T. M. McGinnity, L. P. Maguire, and Q. Wu. Emu-

lating spiking neural networks for edge detection on fpga hardware. In 2009
International Conference on Field Programmable Logic and Applications, pages
670–673, 2009. doi: 10.1109/FPL.2009.5272339.

[25] U. B. A. R. Group. Chipyard’s documentation, 2021. URL https://

[Link]/en/latest/[Link].

[26] L. Hardesty. Explained: Neural networks, 2017. URL [Link]

edu/2017/explained-neural-networks-deep-learning-0414.

[27] N. Imam, T. Cleland, R. Manohar, P. Merolla, J. Arthur, F. Akopyan, and

D. Modha. Implementation of olfactory bulb glomerular-layer computations
in a digital neurosynaptic core. Frontiers in Neuroscience, 6:83, 2012. ISSN
1662-453X. doi: 10.3389/fnins.2012.00083. URL [Link]
org/article/10.3389/fnins.2012.00083.

[28] G. Indiveri and S. C. Liu. Memory and information processing in neuromorphic

systems. Proceedings of the IEEE, 103(8):1379–1397, Aug 2015. ISSN 0018-
9219. doi: 10.1109/JPROC.2015.2444094.
90
 BIBLIOGRAPHY

[29] G. Indiveri, E. Chicca, and R. Douglas. A vlsi array of low-power spiking

neurons and bistable synapses with spike-timing dependent plasticity. IEEE
Transactions on Neural Networks, 17(1):211–221, 2006. doi: 10.1109/TNN.
2005.860850.
[30] G. Indiveri, F. Stefanini, and E. Chicca. Spike-based learning with a gener-
alized integrate and fire silicon neuron. In Proceedings of 2010 IEEE Inter-
national Symposium on Circuits and Systems, pages 1951–1954, 2010. doi:
10.1109/ISCAS.2010.5536980.
[31] E. M. Izhikevich. Which model to use for cortical spiking neurons? IEEE
Transactions on Neural Networks, 15(5):1063–1070, 2004. doi: 10.1109/TNN.
2004.832719.
[32] J. Jones. Mesi cache coherency protocol, 2017. URL [Link]
ie/[Link]/VivioJS/caches/[Link].
[33] J. Joshi, A. C. Parker, and T. Celikel. Neuromorphic network implementation
of the somatosensory cortex. In 2013 6th International IEEE/EMBS Confer-
ence on Neural Engineering (NER), pages 907–910, 2013. doi: 10.1109/NER.
2013.6696082.
[34] J. Lazzaro, J. Wawrzynek, M. Mahowald, M. Sivilotti, and D. Gillespie. Silicon
auditory processors as computer peripherals. IEEE Transactions on Neural
Networks, 4(3):523–528, 1993. doi: 10.1109/72.217193.
[35] J. . Lee and B. J. Sheu. Parallel digital image restoration using adaptive
vlsi neural chips. In Proceedings., 1990 IEEE International Conference on
Computer Design: VLSI in Computers and Processors, pages 126–129, 1990.
doi: 10.1109/ICCD.1990.130181.
[36] D. Marković. Spiking neural networks : What makes these biologi-
cally realistic neurons interesting for computing?, 2020. URL https://
[Link]/spiking-neural-networks-558dc4479903.
[37] McCulloch and Pitts. A logical calculus of the ideas immanent in nervous
activity. Bulletin of Mathematical Biophysics, (5):115–133, Dec 1943. doi:
10.1007/BF02478259.
[38] C. Mead. Neuromorphic electronic systems. Proceedings of the IEEE, 78(10):
1629–1636, Oct 1990. ISSN 0018-9219. doi: 10.1109/5.58356.
[39] S. . Moon, K. G. Shin, and J. Rexford. Scalable hardware priority queue
architectures for high-speed packet switches. In Proceedings Third IEEE Real-
Time Technology and Applications Symposium, pages 203–212, 1997. doi: 10.
1109/RTTAS.1997.601359.
91
 BIBLIOGRAPHY

[40] W. Murphy, M. Renz, and Q. Wu. Binary image classification using a

neurosynaptic processor: A trade-off analysis. In 2016 IEEE International
Symposium on Circuits and Systems (ISCAS), pages 1342–1345, 2016. doi:
10.1109/ISCAS.2016.7527497.

[41] NVIDIA Developer. Top 5 ai speech applications using nvidia’s

gpus for inference, 2019. URL [Link]
top-5-ai-speech-applications-using-nvidias-gpus-for-inference/.

[42] D. Patterson and A. Waterman. The RISC-V Reader: An Open Architecture

Atlas. Strawberry Canyon, 1st edition, 2017. ISBN 0999249118.

[43] A. Pullini, D. Rossi, G. Haugou, and L. Benini. udma: An autonomous i/o

subsystem for iot end-nodes. In 2017 27th International Symposium on Power
and Timing Modeling, Optimization and Simulation (PATMOS), pages 1–8,
2017. doi: 10.1109/PATMOS.2017.8106971.

[44] U. B. A. Research. Chipyard documentation, . URL [Link]

[Link]/en/latest/.

[45] U. B. A. Research. Chipyard framework, . URL [Link]

ucb-bar/chipyard.

[46] RISC-V Foundation. Extending risc-v, 2019. URL [Link]

com/riscv-isa-manual/latest/[Link].

[47] RISC-V Foundation. The risc-v instruction set manual - volume i : Unprivi-
leged isa, 2019. URL [Link]

[48] Roger Shepherd. What is coherent interconnect in

arm soc design?, 2017. URL [Link]
What-is-Coherent-Interconnect-in-ARM-SoC-Design.

[49] M. Schmuker, T. Pfeil, and M. P. Nawrot. A neuromorphic network for

generic multivariate data classification. Proceedings of the National Academy
of Sciences, 111(6):2081–2086, 2014. ISSN 0027-8424. doi: 10.1073/pnas.
1303053111. URL [Link]

[50] C. D. Schuman, T. E. Potok, R. M. Patton, J. D. Birdwell, M. E. Dean,

G. S. Rose, and J. S. Plank. A survey of neuromorphic computing and neural
networks in hardware. CoRR, abs/1705.06963, 2017. URL [Link]
org/abs/1705.06963.

[51] SiFive Inc. Sifive tilelink specification, 2019. URL [Link]

documentation.
92
 BIBLIOGRAPHY

[52] M. Sorbaro, Q. Liu, M. Bortone, and S. Sheik. Optimizing the energy con-
sumption of spiking neural networks for neuromorphic applications, 2020.

[53] M. Thomas. 15 examples of machine learning in healthcare that

are revolutionizing medicine, 2020. URL [Link]
artificial-intelligence/machine-learning-healthcare.

[54] A. W. Trask. Grokking Deep Learning. Manning, 2019. URL [Link]

[Link]/books/grokking-deep-learning.

[55] S. University. Extended address event representation draft standard

v0.4, 2002. URL [Link]
documents/[Link].

[56] Wikipedia Community. Artificial neural network — Wikipedia, the free

encyclopedia, 2021. URL [Link]
neural_network.

[57] Wikipedia Community. Brain-computer interface — Wikipedia, the free

encyclopedia, 2021. URL [Link]
textendashcomputer_interface.

[58] Wikipedia Community. Approfondimento profondo — Wikipedia, the free en-

cyclopedia, 2021. URL [Link]
profondo.

[59] Wikipedia Community. Serial peripheral interface — Wikipedia, the

free encyclopedia, 2021. URL [Link]
Peripheral_Interface.

[60] E. Zurich and U. of Bologna. Pulp silicon proven designs, 2021. URL https:
//[Link]/[Link].

93