% JMF : revoir l'abstract : on y avait mis le Zynq7010 de la redpitaya en montrant
  %   comment optimiser les perfs a surface finie. Ici aussi on tombait dans le cas ou`
  %   la solution a 1 seul FIR n'etait simplement pas synthetisable => fusionner les deux
  %   contributions pour le papier TUFFC

  \documentclass[a4paper,conference]{IEEEtran/IEEEtran}
  \usepackage{graphicx,color,hyperref}
  \usepackage{amsfonts}

  \usepackage{amsthm}
  \usepackage{amssymb}
  \usepackage{amsmath}

  \usepackage{algorithm2e}

  \usepackage{url,balance}

  \usepackage[normalem]{ulem}

  % correct bad hyphenation here
  \hyphenation{op-tical net-works semi-conduc-tor}
  \textheight=26cm
  \setlength{\footskip}{30pt}
  \pagenumbering{gobble}
  \begin{document}
  \title{Filter optimization for real time digital processing of radiofrequency signals: application
  to oscillator metrology}

  \author{\IEEEauthorblockN{A. Hugeat\IEEEauthorrefmark{1}\IEEEauthorrefmark{2}, J. Bernard\IEEEauthorrefmark{2},
  G. Goavec-M\'erou\IEEEauthorrefmark{1},

  P.-Y. Bourgeois\IEEEauthorrefmark{1}, J.-M. Friedt\IEEEauthorrefmark{1}}

  \IEEEauthorblockA{\IEEEauthorrefmark{1}FEMTO-ST, Time \& Frequency department, Besan\c con, France }
  \IEEEauthorblockA{\IEEEauthorrefmark{2}FEMTO-ST, Computer Science department DISC, Besan\c con, France \\
  Email: \{pyb2,jmfriedt\}@femto-st.fr}
  }
  \maketitle
  \thispagestyle{plain}
  \pagestyle{plain}

  
  ewtheorem{definition}{Definition}

  
  \begin{abstract}
  Software Defined Radio (SDR) provides stability, flexibility and reconfigurability to
  radiofrequency signal processing. Applied to oscillator characterization in the context
  of ultrastable clocks, stringent filtering requirements are defined by spurious signal or
  noise rejection needs. Since real time radiofrequency processing must be performed in a
  Field Programmable Array to meet timing constraints, we investigate optimization strategies
  to design filters meeting rejection characteristics while limiting the hardware resources
  required and keeping timing constraints within the targeted measurement bandwidths.
  \end{abstract}
  
  \begin{IEEEkeywords}
  Software Defined Radio, Mixed-Integer Linear Programming, Finite Impulse Response filter
  \end{IEEEkeywords}
  
  \section{Digital signal processing of ultrastable clock signals}
  
  Analog oscillator phase noise characteristics are classically performed by downconverting

  the radiofrequency signal using a saturated mixer to bring the radiofrequency signal to baseband,

  followed by a Fourier analysis of the beat signal to analyze phase fluctuations close to carrier. In
  a fully digital approach, the radiofrequency signal is digitized and numerically downconverted by

  multiplying the samples with a local numerically controlled oscillator (Fig. \ref{schema}) \cite{rsi}.

  
  \begin{figure}[h!tb]
  \begin{center}
  \includegraphics[width=.8\linewidth]{images/schema}
  \end{center}
  \caption{Fully digital oscillator phase noise characterization: the Device Under Test
  (DUT) signal is sampled by the radiofrequency grade Analog to Digital Converter (ADC) and
  downconverted by mixing with a Numerically Controlled Oscillator (NCO). Unwanted signals
  and noise aliases are rejected by a Low Pass Filter (LPF) implemented as a cascade of Finite
  Impulse Response (FIR) filters. The signal is then decimated before a Fourier analysis displays
  the spectral characteristics of the phase fluctuations.}
  \label{schema}
  \end{figure}
  
  As with the analog mixer,
  the non-linear behavior of the downconverter introduces noise or spurious signal aliasing as
  well as the generation of the frequency sum signal in addition to the frequency difference.
  These unwanted spectral characteristics must be rejected before decimating the data stream

  for the phase noise spectral characterization \cite{andrich2018high}. The characteristics introduced between the

  downconverter

  and the decimation processing blocks are core characteristics of an oscillator characterization
  system, and must reject out-of-band signals below the targeted phase noise -- typically in the
  sub -170~dBc/Hz for ultrastable oscillator we aim at characterizing. The filter blocks will
  use most resources of the Field Programmable Gate Array (FPGA) used to process the radiofrequency
  datastream: optimizing the performance of the filter while reducing the needed resources is
  hence tackled in a systematic approach using optimization techniques. Most significantly, we
  tackle the issue by attempting to cascade multiple Finite Impulse Response (FIR) filters with
  tunable number of coefficients and tunable number of bits representing the coefficients and the
  data being processed.

  \section{Finite impulse response filter}
  
  We select FIR filter for their unconditional stability and ease of design. A FIR filter is defined

  by a set of weights $b_k$ applied to the inputs $x_k$ through a convolution to generate the

  outputs $y_k$

  $$y_n=\sum_{k=0}^N b_k x_{n-k}$$

  As opposed to an implementation on a general purpose processor in which word size is defined by the

  processor architecture, implementing such a filter on an FPGA offer more degrees of freedom since

  not only the coefficient values and number of taps must be defined, but also the number of bits

  defining the coefficients and the sample size. For this reason, and because we consider pipeline

  processing (as opposed to First-In, First-Out FIFO memory batch processing) of radiofrequency

  signals, High Level Synthesis (HLS) languages \cite{kasbah2008multigrid} are not considered but

  the problem is tackled at the Very-high-speed-integrated-circuit Hardware Description Language (VHDL) level.

  Since latency is not an issue in a openloop phase noise characterization instrument, the large

  numbre of taps in the FIR, as opposed to the shorter Infinite Impulse Response (IIR) filter,

  is not considered as an issue as would be in a closed loop system.

  The coefficients are classically expressed as floating point values. However, this binary
  number representation is not efficient for fast arithmetic computation by an FPGA. Instead,
  we select to quantify these floating point values into integer values. This quantization

  will result in some precision loss.

  
  %As illustrated in Fig. \ref{float_vs_int}, we see that we aren't
  %need too coefficients or too sample size. If we have lot of coefficients but a small sample size,
  %the first and last are equal to zero. But if we have too sample size for few coefficients that not improve the quality.

  % JMF je ne comprends pas la derniere phrase ci-dessus ni la figure ci dessous

  % AH en gros je voulais dire que prendre trop peu de bit avec trop de coeff, ça induit ta figure (bien mieux faite que moi)
  %    et que l'inverse trop de bit sur pas assez de coeff on ne gagne rien, je vais essayer de la reformuler

  %\begin{figure}[h!tb]
  %\includegraphics[width=\linewidth]{images/float-vs-integer.pdf}
  %\caption{Impact of the quantization resolution of the coefficients}

  %\label{float_vs_int}

  %\end{figure}

  \begin{figure}[h!tb]
  \includegraphics[width=\linewidth]{images/demo_filtre}
  \caption{Impact of the quantization resolution of the coefficients: the quantization is

  set to 6~bits -- with the horizontal black lines indicating $\pm$1 least significant bit -- setting

  the 30~first and 30~last coefficients out of the initial 128~band-pass
  filter coefficients to 0 (red dots).}

  \label{float_vs_int}
  \end{figure}

  The tradeoff between quantization resolution and number of coefficients when considering
  integer operations is not trivial. As an illustration of the issue related to the
  relation between number of fiter taps and quantization, Fig. \ref{float_vs_int} exhibits
  a 128-coefficient FIR bandpass filter designed using floating point numbers (blue). Upon
  quantization on 6~bit integers, 60 of the 128~coefficients in the beginning and end of the
  taps become null, making the large number of coefficients irrelevant and allowing to save
  processing resource by shrinking the filter length. This tradeoff aimed at minimizing resources
  to reach a given rejection level, or maximizing out of band rejection for a given computational
  resource, will drive the investigation on cascading filters designed with varying tap resolution

  and tap length, as will be shown in the next section. Indeed, our development strategy closely
  follows the skeleton approach \cite{crookes1998environment, crookes2000design, benkrid2002towards}
  in which basic blocks are defined and characterized before being assembled \cite{hide}
  in a complete processing chain. In our case, assembling the filter blocks is a simpler block

  combination process since we assume a single value to be processed and a single value to be

  generated at each clock cycle. The FIR filters will not be considered to decimate in the
  current implementation: the decimation is assumed to be located after the FIR cascade at the
  moment.

  \section{Filter optimization}
  
  A basic approach for implementing the FIR filter is to compute the transfer function of
  a monolithic filter: this single filter defines all coefficients with the same resolution
  (number of bits) and processes data represented with their own resolution. Meeting the
  filter shape requires a large number of coefficients, limited by resources of the FPGA since
  this filter must process data stream at the radiofrequency sampling rate after the mixer.
  
  An optimization problem \cite{leung2004handbook} aims at improving one or many
  performance criteria within a constrained resource environment. Amongst the tools
  developed to meet this aim, Mixed-Integer Linear Programming (MILP) provides the framework to

  formally define the stated problem and search for an optimal use of available

  resources \cite{yu2007design, kodek1980design}.

  First we need to ensure that our problem is a real optimization problem. When

  designing a processing function in the FPGA, we aim at meeting some requirement such as
  the throughput, the computation time or the noise rejection noise. However, due to limited

  resources to design the process like BRAM (high performance RAM), DSP (Digital Signal Processor)

  or LUT (Look Up Table), a tradeoff must be generally searched between performance and available
  computational resources: optimizing some criteria within finite, limited
  resources indeed matches the definition of a classical optimization problem.

  
  Specifically the degrees of freedom when addressing the problem of replacing the single monolithic

  FIR with a cascade of optimized filters are the number of coefficients $N_i$ of each filter $i$,

  the number of bits $C_i$ representing the coefficients and the number of bits $D_i$ needed to represent
  the data $x_k$ fed to each filter as provided by the acquisition or previous processing stage. 
  Because each FIR in the chain is fed the output of the previous stage,

  the optimization of the complete processing chain within a constrained resource environment is not

  trivial. The resource occupation of a FIR filter is considered as $C_i \times N_i$ which aims

  at approximating the number of bits needed in a worst case condition to represent the output of the

  FIR. Indeed, the number of bits generated by the $i$th FIR is $(C_i+D_i)\times\log_2(N_i)$, but the

  $\log$ function is avoided for its incompatibility with a linear programming description, and
  the simple product is approximated as the number of gates needed to perform the calculation. Such an

  occupied area estimate assumes that the number of gates scales as the number of bits and the number
  of coefficients, but does not account for the detailed implementation of the hardware. Indeed,
  various FPGA implementations will provide different hardware functionalities, and we shall consider
  at the end of the design a synthesis step using vendor software to assess the validity of the solution
  found. As an example of the limitation linked to the lack of detailed hardware consideration, Block Random
  Access Memory (BRAM) used to store filter coefficients are not shared amongst filters, and multiplications

  are most efficiently implemented by using DSP blocks whose input word

  size is finite. DSPs are a scarce resource to be saved in a practical implementation. Keeping a high
  abstraction on the resource occupation is nevertheless selected in the following discussion in order
  to leave enough degrees of freedom in the problem to try and find original solutions: too many
  constraints in the initial statement of the problem leave little room for finding an optimal solution.

  \begin{figure}[h!tb]
  \begin{center}
  \includegraphics[width=.5\linewidth]{schema2}

  \caption{Shape of the filter transmitted power $P$ as a function of frequency:

  the bandpass BP is considered to occupy the initial
  40\% of the Nyquist frequency range, the stopband the last 40\%, allowing 20\% transition

  width.}
  \label{rejection-shape}
  \end{center}
  \end{figure}
  
  Following these considerations, the model is expressed as:

  \begin{align}
    \begin{cases}
      \mathcal{R}_i &= \mathcal{F}(N_i, C_i)\\

      \mathcal{A}_i &= N_i \times C_i\\

      \Delta_i &= \Delta _{i-1} + \mathcal{P}_i
    \end{cases}
    \label{model-FIR}
  \end{align}

  To explain the system \ref{model-FIR}, $\mathcal{R}_i$ represents the stopband rejection dependence with $N_i$ and $C_i$, $\mathcal{A}_i$

  is a theoretical area occupation of the processing block on the FPGA as discussed earlier, and $\Delta_i$ is the total rejection for the current stage $i$.

  Since the function $\mathcal{F}$ cannot be explictly expressed, we run simulations to determine the rejection depending

  on $N_i$ and $C_i$. However, selecting the right filter requires a clear definition of the rejection criterion. Selecting an
  incorrect criterion will lead the linear program solver to produce a solution which might not meet the user requirements.

  Hence, amongst various criteria including the mean or median value of the FIR response in the stopband as will
  be illustrated lated (section \ref{median}), we have designed

  a criterion aimed at avoiding ripples in the passband and considering the maximum of the FIR spectral response in the stopband

  (Fig. \ref{rejection-shape}). The bandpass criterion is defined as the sum of the absolute values of the spectral response
  in the bandpass, reminiscent of a standard deviation of the spectral response: this criterion must be minimized to avoid
  ripples in the passband. The stopband transfer function maximum must also be minimized in order to improve the filter
  rejection capability. Weighing these two criteria allows designing the linear program to be solved.

  \begin{figure}[h!tb]
  \includegraphics[width=\linewidth]{images/noise-rejection.pdf}
  \caption{Rejection as a function of number of coefficients and number of bits}
  \label{noise-rejection}
  \end{figure}

  The objective function maximizes the noise rejection ($\max(\Delta_{i_{\max}})$) while keeping resource
  occupation below a user-defined threshold, or as will be discussed here, aims at minimizing the area
  needed to reach a given rejection ($\min(S_q)$ in the forthcoming discussion, Eqs. \ref{cstr_size}

  and \ref{cstr_rejection}). The MILP solver is allowed to choose the number of successive

  filters, within an upper bound. The last problem is to model the noise rejection. Since filter

  noise rejection capability is not modeled with linear equations, a look-up-table is generated

  for multiple filter configurations in which the $C_i$, $D_i$ and $N_i$ parameters are varied: for each

  one of these conditions, the low-pass filter rejection is stored as computed by the frequency response
  of the digital filter (Fig. \ref{noise-rejection}). Various rejection criteria have been investigated,
  including mean value of the stopband response, median value of the stopband response, or as finally

  selected, maximum value in the stopband. An intuitive analysis of the chart of Fig. \ref{noise-rejection}

  hints at an optimum

  set of tap length and number of bit for representing the coefficients along the line of the pyramidal
  shaped rejection capability function.

  
  Linear program formalism for solving the problem is well documented: an objective function is
  defined which is linearly dependent on the parameters to be optimized. Constraints are expressed

  as linear equations and solved using one of the available solvers, in our case GLPK\cite{glpk}.
  With the notations used in the description of system \ref{model-FIR}, we have defined the linear problem as:

  \paragraph{Variables}
  \begin{align*}

  x_{i,j} \in \lbrace 0,1 \rbrace & \text{ $i$ is a given filter} \\

  & \text{ $j$ is the stage} \\
  & \text{ If $x_{i,j}$ is equal to 1, the filter is selected} \\
  \end{align*}
  \paragraph{Constants}
  \begin{align*}
  \mathcal{F} = \lbrace F_1 ... F_p \rbrace & \text{ All possible filters}\\
  & \text{ $p$ is the number of different filters} \\

  % N(i) & \text{ % Constant to let the
  % number of coefficients %} \\ & \text{
  % for filter $i$}\\
  % C(i) & \text{ % Constant to let the
  % number of bits of %}\\ & \text{
  % each coefficient for filter $i$}\\
  \mathcal{S}_{\max} & \text{ Total space available inside the FPGA}

  \end{align*}
  \paragraph{Constraints}

  \begin{align}
  1 \leq i \leq p & 
  onumber\\
  1 \leq j \leq q & \text{ $q$ is the max of filter stage} 
  onumber \\
  \forall j, \mathlarger{\sum_{i}} x_{i,j} = 1 & \text{ At most one filter by stage} 
  onumber\\
  \mathcal{S}_0 = 0 & \text{ initial occupation} 
  onumber\\

  \forall j, \mathcal{S}_j = \mathcal{S}_{j-1} + \mathlarger{\sum_i (x_{i,j} \times \mathcal{A}_i)} \label{cstr_size} \\

  \mathcal{S}_j \leq \mathcal{S}_{\max}
  onumber \\

  \mathcal{N}_0 = 0 & \text{ initial rejection}
  onumber\\

  \forall j, \mathcal{N}_j = \mathcal{N}_{j-1} + \mathlarger{\sum_i (x_{i,j} \times \mathcal{R}_i)} \label{cstr_rejection} \\
  \mathcal{N}_q \geqslant 160 & \text{ an user defined bound}
  onumber\\

  & \text{ (e.g. 160~dB here)}
  onumber\\
  onumber
  \end{align}

  \paragraph{Goal}
  \begin{align*}
  \min \mathcal{S}_q
  \end{align*}

  The constraint \ref{cstr_size} means the occupation for the current stage $j$ depends on
  the previous occupation and the occupation of current selected filter (it is possible
  that no filter is selected for this stage). And the second one \ref{cstr_rejection}
  means the same thing but for the rejection, the rejection depends the previous rejection
  plus the rejection of selected filter.

  \subsection{Low bandpass ripple and maximum rejection criteria}

  The MILP solver provides a solution to the problem by selecting a series of small FIR with
  increasing number of bits representing data and coefficients as well as an increasing number

  of coefficients, instead of a single monolithic filter.

  
  \begin{figure}[h!tb]
  % \includegraphics[width=\linewidth]{images/compare-fir.pdf}

  \includegraphics[width=\linewidth]{images/fir-mono-vs-fir-series-noise-fixe-jmf-light.pdf}

  \caption{Comparison of the rejection capability between a series of FIR and a monolithic FIR
  with a cutoff frequency set at half the Nyquist frequency.}
  \label{compare-fir}
  \end{figure}

  Fig. \ref{compare-fir} exhibits the
  performance comparison between one solution and a monolithic FIR when selecting a cutoff
  frequency of half the Nyquist frequency: a series of 5 FIR and a series of 10 FIR with the
  same space usage are provided as selected by the MILP solver. The FIR cascade provides improved
  rejection than the monolithic FIR at the expense of a lower cutoff frequency which remains to
  be tuned or compensated for.

  The resource occupation when synthesizing such FIR on a Xilinx FPGA is summarized as Tab. \ref{t1}.

  We have considered a set of resources representative of the hardware platform we work on,

  Avnet's Zedboard featuring a Xilinx XC7Z020-CLG484-1 Zynq System on Chip (SoC). The results reported in

  Tab. \ref{t1} emphasize that implementing the monolithic single FIR is impossible due to
  the insufficient hardware resources (exhausted LUT resources), while the FIR cascading 5 or 10

  filters fit in the available resources. However, in all cases the DSP resources are fully

  used: while the design can be synthesized using Xilinx proprietary Vivado 2016.2 software,
  implementing the design fails due to the excessive resource usage preventing routing the signals
  on the FPGA. Such results emphasize on the one hand the improvement prospect of the optimization
  procedure by finding non-trivial solutions matching resource constraints, but on the other
  hand also illustrates the limitation of a model with an abstraction layer that does not account
  for the detailed architecture of the hardware.

  
  \begin{table}[h!tb]
  \caption{Resource occupation on a Xilinx Zynq-7000 series FPGA when synthesizing the FIR cascade
  identified as optimal by the MILP solver within a finite resource criterion. The last line refers

  to available resources on a Zynq-7020 as found on the Zedboard.}

  \begin{center}

  \begin{tabular}{|c|cccc|}\hline
  FIR & BlockRAM & LookUpTables & DSP & rejection (dB)\\\hline\hline

  1 (monolithic) & 1 & 76183 & 220 & -162 \\
  5 & 5 & 18597 & 220 & -160 \\
  10 & 8 & 24729 & 220 & -161 \\\hline\hline
  \textbf{Zynq 7020} & \textbf{420} & \textbf{53200} & \textbf{220} &  \\\hline

  %\begin{tabular}{|c|ccccc|}\hline
  %FIR & BRAM36 & BRAM18 & LUT & DSP & rejection (dB)\\\hline\hline
  %1 (monolithic) & 1 & 0 & {\color{Red}76183} & 220 & -162 \\
  %5 & 0 & 5 & {\color{Green}18597} & 220 & -160 \\
  %10 & 0 & 8 & {\color{Green}24729} & 220 & -161 \\\hline\hline
  %\textbf{Zynq 7020} & \textbf{140} & \textbf{280} & \textbf{53200} & \textbf{220} &  \\\hline

  \end{tabular}
  \end{center}
  %\vspace{-0.7cm}
  \label{t1}
  \end{table}

  \subsection{Alternate criteria}\label{median}
  
  Fig. \ref{compare-fir} provides FIR solutions matching well the targeted transfer
  function, namely low ripple in the bandpass defined as the first 40\% of the frequency
  range and maximum rejection of 160~dB in the last 40\% stopband. We illustrate now, for
  demonstrating the need to properly select the optimization criterion, two cases of poor
  filter shapes obtained by selecting the mean value and median value of the rejection,
  with no consideration for the ripples in the bandpass. The results of the optimizations,
  in these cases, are shown in Figs. \ref{compare-mean} and \ref{compare-median}.
  
  \begin{figure}[h!tb]

  \includegraphics[width=\linewidth]{images/fir-mono-vs-fir-series-noise-fixe-mean-light.pdf}

  \caption{Comparison of the rejection capability between a series of FIR and a monolithic FIR
  with a cutoff frequency set at half the Nyquist frequency.}
  \label{compare-mean}
  \end{figure}
  
  In the case of the mean value criterion (Fig. \ref{compare-mean}), the solution is not
  acceptable since the notch at the end of the transition band compensates for some unacceptable
  rise in the rejection close to the Nyquist frequency. Applying such a filter might yield excessive
  high frequency spurious components to be aliased at low frequency when decimating the signal.
  Similarly, the lack of criterion on the bandpass shape induces a shape with poor flatness and
  and slowly decaying transfer function starting to attenuate spectral components well before the
  transition band starts. Such issues are partly aleviated by replacing a mean rejection value with
  a median rejection value (Fig. \ref{compare-median}) but solutions remain unacceptable for
  the reasons stated previously and much poorer than those found with the maximum rejection criterion
  selected earlier (Fig. \ref{compare-fir}).
  
  \begin{figure}[h!tb]

  \includegraphics[width=\linewidth]{images/fir-mono-vs-fir-series-noise-fixe-median-light.pdf}

  \caption{Comparison of the rejection capability between a series of FIR and a monolithic FIR
  with a cutoff frequency set at half the Nyquist frequency.}
  \label{compare-median}
  \end{figure}

  \section{Filter coefficient selection}
  
  The coefficients of a single monolithic filter are computed as the impulse response
  of the filter transfer function, and practically approximated by a multitude of methods
  including least square optimization (Matlab's {\tt firls} function), Hamming or Kaiser windowing

  (Matlab's {\tt fir1} function).

  
  \begin{figure}[h!tb]
  \includegraphics[width=\linewidth]{images/fir1-vs-firls}
  \caption{Evolution of the rejection capability of least-square optimized filters and Hamming
  FIR filters as a function of the number of coefficients, for floating point numbers and 8-bit
  encoded integers.}
  \label{2}
  \end{figure}

  Cascading filters opens a new optimization opportunity by
  selecting various coefficient sets depending on the number of coefficients. Fig. \ref{2}
  illustrates that for a number of coefficients ranging from 8 to 47, {\tt fir1} provides a better
  rejection than {\tt firls}: since the linear solver increases the number of coefficients along
  the processing chain, the type of selected filter also changes depending on the number of coefficients
  and evolves along the processing chain.

  \section{Conclusion}
  
  We address the optimization problem of designing a low-pass filter chain in a Field Programmable Gate
  Array for improved noise rejection within constrained resource occupation, as needed for
  real time processing of radiofrequency signal when characterizing spectral phase noise
  characteristics of stable oscillators. The flexibility of the digital approach makes the result
  best suited for closing the loop and using the measurement output in a feedback loop for
  controlling clocks, e.g. in a quartz-stabilized high performance clock whose long term behavior

  is controlled by non-piezoelectric resonator (sapphire resonator, microwave or optical

  atomic transition).
  
  \section*{Acknowledgement}

  This work is supported by the ANR Programme d'Investissement d'Avenir in
  progress at the Time and Frequency Departments of the FEMTO-ST Institute
  (Oscillator IMP, First-TF and Refimeve+), and by R\'egion de Franche-Comt\'e.

  The authors would like to thank E. Rubiola, F. Vernotte, and G. Cabodevila 
  for support and fruitful discussions.

  \bibliographystyle{IEEEtran}

  \balance

  \bibliography{references,biblio}
  \end{document}

  	\section{Contexte d'ordonnancement}

  	Dans cette partie, nous donnerons des d\'efinitions de termes rattach\'es au domaine de l'ordonnancement

  	et nous verrons que le sujet trait\'e se rapproche beaucoup d'un problème d'ordonnancement. De ce fait

  	nous pourrons aller plus loin que les travaux vus pr\'ec\'edemment et nous tenterons des approches d'ordonnancement

  	et d'optimisation.

  	\subsection{D\'efinition du vocabulaire}

  	Avant tout, il faut d\'efinir ce qu'est un problème d'optimisation. Il y a deux d\'efinitions

  	importantes à donner. La première est propos\'ee par Legrand et Robert dans leur livre \cite{def1-ordo} :
  	\begin{definition}
  		\label{def-ordo1}

  		Un ordonnancement d'un système de t\^aches $G\ =\ (V,\ E,\ w)$ est une fonction $\sigma$ :

  		$V \rightarrow \mathbb{N}$ telle que $\sigma(u) + w(u) \leq \sigma(v)$ pour toute arête $(u,\ v) \in E$.
  	\end{definition}

  
  	Dit plus simplement, l'ensemble $V$ repr\'esente les t\^aches à ex\'ecuter, l'ensemble $E$ repr\'esente les d\'ependances
  	des t\^aches et $w$ les temps d'ex\'ecution de la t\^ache. La fonction $\sigma$ donne donc l'heure de d\'ebut de
  	chacune des t\^aches. La d\'efinition dit que si une t\^ache $v$ d\'epend d'une t\^ache $u$ alors
  	la date de d\'ebut de $v$ sera plus grande ou \'egale au d\'ebut de l'ex\'ecution de la t\^ache $u$ plus son

  	temps d'ex\'ecution.

  	Une autre d\'efinition importante qui est propos\'ee par Leung et al. \cite{def2-ordo} est :
  	\begin{definition}
  		\label{def-ordo2}
  		L'ordonnancement traite de l'allocation de ressources rares à des activit\'es avec
  		l'objectif d'optimiser un ou plusieurs critères de performance.
  	\end{definition}

  	Cette d\'efinition est plus g\'en\'erique mais elle nous int\'eresse d'avantage que la d\'efinition \ref{def-ordo1}.
  	En effet, la partie qui nous int\'eresse dans cette première d\'efinition est le respect de la pr\'ec\'edance des t\^aches.
  	Dans les faits les dates de d\'ebut ne nous int\'eressent pas r\'eellement.

  	En revanche la d\'efinition \ref{def-ordo2} sera au c\oe{}ur du projet. Pour se convaincre de cela,

  	il nous faut d'abord d\'efinir quel est le type de problème d'ordonnancement qu'on traite et quelles

  	sont les m\'ethodes qu'on peut appliquer.

  	Les problèmes d'ordonnancement peuvent être class\'es en diff\'erentes cat\'egories :
  	\begin{itemize}

  		\item T\^aches ind\'ependantes : dans cette cat\'egorie de problèmes, les t\^aches sont complètement ind\'ependantes

  		les unes des autres. Dans notre cas, ce n'est pas le plus adapt\'e.
  		\item Graphe de t\^aches : la d\'efinition \ref{def-ordo1} d\'ecrit cette cat\'egorie. La plupart du temps,

  		les t\^aches sont repr\'esent\'ees par une DAG. Cette cat\'egorie est très proche de notre cas puisque nous devons \'egalement ex\'ecuter
  		des t\^aches qui ont un certain nombre de d\'ependances. On pourra même dire que dans certain cas,
  		on a des anti-arbres, c'est à dire que nous avons une multitude de t\^aches d'entr\'ees qui convergent vers une
  		t\^ache de fin.

  		\item Workflow : cette cat\'egorie est une sous cat\'egorie des graphes de t\^aches dans le sens où
  		il s'agit d'un graphe de t\^aches r\'ep\'et\'e de nombreuses de fois. C'est exactement ce type de problème
  		que nous traitons ici.
  	\end{itemize}

  	Bien entendu, cette liste n'est pas exhaustive et il existe de nombreuses autres classifications et sous-classifications
  	de ces problèmes. Nous n'avons parl\'e ici que des cat\'egories les plus communes.

  
  	Un autre point à d\'efinir, est le critère d'optimisation. Il y a là encore un grand nombre de

  	critères possibles. Nous allons donc parler des principaux :
  	\begin{itemize}
  		\item Temps de compl\'etion total (ou Makespan en anglais) : ce critère est l'un des critères d'optimisation

  		les plus courant. Il s'agit donc de minimiser la date de fin de la dernière t\^ache de l'ensemble des

  		t\^aches à ex\'ecuter. L'enjeu de cette optimisation est donc de trouver l'ordonnancement optimal permettant
  		la fin d'ex\'ecution au plus tôt.
  		\item Somme des temps d'ex\'ecution (Flowtime en anglais) : il s'agit de faire la somme des temps d'ex\'ecution de toutes les t\^aches
  		et d'optimiser ce r\'esultat.
  		\item Le d\'ebit : ce critère quant à lui, vise à augmenter au maximum le d\'ebit de traitement des donn\'ees.
  	\end{itemize}

  	En plus de cela, on peut avoir besoin de plusieurs critères d'optimisation. Il s'agit dans ce cas d'une optimisation
  	multi-critères. Bien entendu, cela complexifie d'autant plus le problème car la solution la plus optimale pour un
  	des critères peut être très mauvaise pour un autre critère. De ce cas, il s'agira de trouver une solution qui permet
  	de faire le meilleur compromis entre tous les critères.

  	\subsection{Formalisation du problème}
  	\label{formalisation}

  	Maintenant que nous avons donn\'e le vocabulaire li\'e à l'ordonnancement, nous allons pouvoir essayer caract\'eriser

  	formellement notre problème. En effet, nous allons reprendre les contraintes \'enonc\'ees dans la sections \ref{def-contraintes}
  	et nous essayerons de les formaliser le plus finement possible.

  	Comme nous l'avons dit, une t\^ache est un bloc de traitement. Chaque t\^ache $i$ dispose d'un ensemble de paramètres
  	que nous nommerons $\mathcal{P}_{i}$. Cet ensemble $\mathcal{P}_i$ est propre à chaque t\^ache et il variera d'une
  	t\^ache à l'autre. Nous reviendrons plus tard sur les paramètres qui peuvent composer cet ensemble.

  	Outre cet ensemble $\mathcal{P}_i$, chaque t\^ache dispose de paramètres communs :
  	\begin{itemize}
  		\item Dur\'ee de la t\^ache : Comme nous l'avons dit auparavant, dans le cadre d'un FPGA le temps est compt\'e en nombre de coup d'horloge.
  		En outre, les blocs sont toujours sollicit\'es, certains même sont capables de lire et de renvoyer une r\'esultat à chaque coups d'horloge.

  		Donc la dur\'ee d'une t\^ache ne peut être le laps de temps entre l'entr\'ee d'une donn\'ee et la sortie d'une autre. Nous d\'efinirons la
  		dur\'ee comme le temps de traitement d'une donn\'ee, c'est à dire la diff\'erence de temps entre la date de sortie d'une donn\'ee

  		et de sa date d'entr\'ee. Nous nommerons cette dur\'ee $\delta_i$. % Je devrais la nomm\'ee w comme dans la def2

  		\item La pr\'ecision : La pr\'ecision d'une donn\'ee est le nombre de bits significatifs qu'elle compte. En effet, au fil des traitements

  		les pr\'ecisions peuvent varier. On nomme donc la pr\'ecision d'entr\'ee d'une t\^ache $i$ comme $\pi_i^-$ et la pr\'ecision en sortie $\pi_i^+$.

  		\item La fr\'equence du flux en entr\'ee (ou sortie) : Cette fr\'equence repr\'esente la fr\'equence des donn\'ees qui arrivent (resp. sortent).

  		Selon les t\^aches, les fr\'equences varieront. En effet, certains blocs ralentissent le flux c'est pourquoi on distingue la fr\'equence du
  		flux en entr\'ee et la fr\'equence en sortie. Nous nommerons donc la fr\'equence du flux en entr\'ee $f_i^-$ et la fr\'equence en sortie $f_i^+$.
  		\item La quantit\'e de donn\'ees en entr\'ee (ou en sortie) : Il s'agit de la quantit\'e de donn\'ees que le bloc s'attend à traiter (resp.
  		est capable de produire). Les t\^aches peuvent avoir à traiter des gros volumes de donn\'ees et n'en ressortir qu'une partie. Cette

  		fois encore, il nous faut donc diff\'erencier l'entr\'ee et la sortie. Nous nommerons donc la quantit\'e de donn\'ees entrantes $q_i^-$

  		et la quantit\'e de donn\'ees sortantes $q_i^+$ pour une t\^ache $i$.

  		\item Le d\'ebit d'entr\'ee (ou de sortie) : Ce paramètre correspond au d\'ebit de donn\'ees que la t\^ache est capable de traiter ou qu'elle
  		fournit en sortie. Il s'agit simplement de l'expression des deux pr\'ec\'edents paramètres. Nous d\'efinirons donc la d\'ebit entrant de la

  		t\^ache $i$ comme $d_i^-\ =\ q_i^-\ *\ f_i^-$ et le d\'ebit sortant comme $d_i^+\ =\ q_i^+\ *\ f_i^+$.
  		\item La taille de la t\^ache : La taille dans les FPGA \'etant limit\'ee, ce paramètre exprime donc la place qu'occupe la t\^ache au sein du bloc.
  		Nous nommerons $\mathcal{A}_i$ cette taille.

  		\item Les pr\'ed\'ecesseurs et successeurs d'une t\^ache : cela nous permet de connaître les t\^aches requises pour pouvoir traiter

  		la t\^ache $i$ ainsi que les t\^aches qui en d\'ependent. Ces ensemble sont not\'es $\Gamma _i ^-$ et $ \Gamma _i ^+$ \\
  		%TODO Est-ce vraiment un paramètre ?
  	\end{itemize}

  
  	Ces diff\'erents paramètres communs sont fortement li\'es aux \'el\'ements de $\mathcal{P}_i$. Voici quelques exemples de relations
  	que nous avons identifi\'ees :

  	\begin{itemize}
  		\item $ \delta _i ^+ \ = \ \mathcal{F}_{\delta}(\pi_i^-,\ \pi_i^+,\ d_i^-,\ d_i^+,\ \mathcal{P}_i) $ donne le temps d'ex\'ecution
  		de la t\^ache en fonction de la pr\'ecision voulue, du d\'ebit et des paramètres internes.
  		\item $ \pi _i ^+ \ = \ \mathcal{F}_{p}(\pi_i^-,\ \mathcal{P}_i) $, la fonction $F_p$ donne la pr\'ecision en sortie selon la pr\'ecision de d\'epart
  		et les paramètres internes de la t\^ache.
  		\item $d_i^+\ =\ \mathcal{F}_d(d_i^-, \mathcal{P}_i)$, la fonction $F_d$ donne le d\'ebit sortant de la t\^ache en fonction du d\'ebit
  		sortant et des variables internes de la t\^ache.
  		\item $A_i^+\ =\ \mathcal{F}_A(\pi_i^-,\ \pi_i^+,\ d_i^-,\ d_i^+, \mathcal{P}_i)$

  	\end{itemize}
  	Pour le moment, nous ne sommes pas capables de donner une d\'efinition g\'en\'erale de ces fonctions. Mais en revanche,
  	sur quelques exemples simples (cf. \ref{def-contraintes}), nous parvenons à donner une \'evaluation de ces fonctions.

  	Maintenant que nous avons donn\'e toutes les notations utiles, nous allons \'enoncer des contraintes relatives à notre problème. Soit
  	un DGA $G(V,\ E)$, on a pour toutes arêtes $(i, j)\ \in\ E$ les in\'equations suivantes :

  	\paragraph{Contrainte de pr\'ecision :}

  	Cette in\'equation traduit la contrainte de pr\'ecision d'une t\^ache à l'autre :

  	\begin{align*}
  		\pi _i ^+ \geq \pi _j ^-
  	\end{align*}

  	\paragraph{Contrainte de d\'ebit :}

  	Cette in\'equation traduit la contrainte de d\'ebit d'une t\^ache à l'autre :

  	\begin{align*}
  		d _i ^+ = q _j ^- * (f_i + (1 / s_j) ) & \text{ où } s_j \text{ est une valeur positive de temporisation de la t\^ache}
  	\end{align*}

  	\paragraph{Contrainte de synchronisation :}

  	Il s'agit de la contrainte qui impose que si à un moment du traitement, le DAG se s\'epare en plusieurs branches parallèles

  	et qu'elles se rejoignent plus tard, la somme des latences sur chacune des branches soit la même.
  	Plus formellement, s'il existe plusieurs chemins disjoints, partant de la t\^ache $s$ et allant à la t\^ache de $f$ alors :
  	\begin{align*}

  		\forall \text{ chemin } \mathcal{C}1(s, .., f),
  			\forall \text{ chemin } \mathcal{C}2(s, .., f)
  				\text{ tel que } \mathcal{C}1 
  eq \mathcal{C}2
  		\Rightarrow

  			\sum _{i} ^{i \in \mathcal{C}1} \delta_i = \sum _{i} ^{i \in \mathcal{C}2} \delta_i
  	\end{align*}

  	\paragraph{Contrainte de place :}

  	Cette in\'equation traduit la contrainte de place dans le FPGA. La taille max de la puce FPGA est nomm\'e $\mathcal{A}_{FPGA}$ :

  	\begin{align*}
  		\sum ^{\text{t\^ache } i} \mathcal{A}_i \leq \mathcal{A}_{FPGA}
  	\end{align*}

  	\subsection{Exemples de mod\'elisation}
  	\label{exemples-modeles}

  	Nous allons maintenant prendre quelques blocs de traitement simples afin d'illustrer au mieux notre modèle.

  	Pour tous nos exemple, nous prendrons un d\'ebit en entr\'ee de 200 Mo/s avec une pr\'ecision de 16 bit.

  	Prenons tout d'abord l'exemple d'un bloc de d\'ecimation. Le but de ce bloc est de ralentir le flux en ne gardant
  	que certaines donn\'ees à intervalle r\'egulier. Cet intervalle est appel\'e le facteur de d\'ecimation, on le notera $N$.

  	Donc d'après notre mod\'elisation :
  	\begin{itemize}
  		\item $N \in \mathcal{P}_i$
  		%TODO N ou 1 ?
  		\item $\delta _i = N\ c.h.$ (coup d'horloge)
  		\item $\pi _i ^+ = \pi _i ^- = 16 bits$
  		\item $f _i ^+ = f _i ^-$
  		\item $q _i ^+ = q _i ^- / N$
  		\item $d _i ^+ = q _i ^- / N / f _i ^-$
  		\item $\Gamma _i ^+ = \Gamma _i ^- = 1$\\
  		%TODO Je ne sais pas trouver la taille...
  	\end{itemize}

  
  	Un autre exemple int\'eressant que l'on peut donner, c'est le cas des spliters. Il s'agit la aussi d'un bloc très
  	simple qui permet de dupliquer un flux. On peut donc donner un nombre de sorties à cr\'eer, on note ce paramètre

  	%TODO pas très inspir\'e...
  	$X$. Voici ce que donne notre mod\'elisation :
  	\begin{itemize}
  		\item $X \in \mathcal{P}_i$
  		\item $\delta _i = 1\ c.h.$
  		\item $\pi _i ^+ = \pi _i ^- = 16 bits$
  		\item $f _i ^+ = f _i ^-$
  		\item $q _i ^+ = q _i ^-$
  		\item $d _i ^+ = d _i ^-$
  		\item $\Gamma _i ^- = 1$
  		\item $\Gamma _i ^+ = X$\\
  	\end{itemize}

  	L'exemple suivant traite du cas du shifter. Il s'agit d'un bloc qui a pour but de diminuer le nombre de bits des

  	donn\'ees afin d'acc\'el\'erer les traitement sur les blocs suivants. On peut donc donner le nombre de bits à shifter,

  	on note ce paramètre $S$. Voici ce que donne notre mod\'elisation :
  	\begin{itemize}
  		\item $S \in \mathcal{P}_i$
  		\item $\delta _i = 1\ c.h.$
  		\item $\pi _i ^+ = \pi _i ^- - S$
  		\item $f _i ^+ = f _i ^-$
  		\item $q _i ^+ = q _i ^-$
  		\item $d _i ^+ = d _i ^-$
  		\item $\Gamma _i ^+ = \Gamma _i ^- = 1$\\
  	\end{itemize}

  
  	Nous allons traiter un dernier exemple un peu plus complexe, le cas d'un filtre d\'ecimateur (ou FIR). Ce bloc

  	est compos\'e de beaucoup de paramètres internes. On peut d\'efinir un nombre d'\'etages $E$, qui repr\'esente le nombre

  	d'it\'erations à faire avant d'arrêter le traitement. Afin d'effectuer son filtrage, on doit donner au bloc un ensemble

  	de coefficients $C$ et par cons\'equent ces coefficients ont leur propre pr\'ecision $\pi _C$. Pour finir, le dernier
  	paramètre à donner est le facteur de d\'ecimation $N$. Si on applique notre mod\'elisation, on peut obtenir cela :
  	\begin{itemize}
  		\item $E \in \mathcal{P}_i$
  		\item $C \in \mathcal{P}_i$
  		\item $\pi _C \in \mathcal{P}_i$
  		\item $N \in \mathcal{P}_i$
  		\item $\delta _i = E * |C| * q_i^-\ c.h.$ %Trop simpliste
  		\item $\pi _i ^+ = \pi _i ^- * \pi _C$
  		\item $f _i ^+ = f _i ^-$
  		\item $q _i ^+ = q _i ^- / N$
  		\item $d _i ^+ = q _i ^- / N / f _i ^-$
  		\item $\Gamma _i ^+ = \Gamma _i ^- = 1$\\
  	\end{itemize}

  
  	Ces exemples ne sont que des modèles provisoires; pour s'assurer de leur performance, il faudra les

  	confronter à des simulations.

  
  
  Bien que les articles sur les skeletons, \cite{gwen-cogen}, \cite{skeleton} et \cite{hide}, nous aient donn\'e des indices sur une possible

  	mod\'elisation, ils \'etaient encore trop focalis\'es sur l'optimisation spatiale des blocs. Nous nous sommes donc inspir\'es de ces travaux
  	pour proposer notre modèle, en faisant abstraction des optimisations bas niveau.