%%
%% $Id$
%%
%% Copyright 1989-2016 MINES ParisTech
%%
%% This file is part of PIPS.
%%
%% PIPS is free software: you can redistribute it and/or modify it
%% under the terms of the GNU General Public License as published by
%% the Free Software Foundation, either version 3 of the License, or
%% any later version.
%%
%% PIPS is distributed in the hope that it will be useful, but WITHOUT ANY
%% WARRANTY; without even the implied warranty of MERCHANTABILITY or
%% FITNESS FOR A PARTICULAR PURPOSE.
%%
%% See the GNU General Public License for more details.
%%
%% You should have received a copy of the GNU General Public License
%% along with PIPS.  If not, see <http://www.gnu.org/licenses/>.
%%
\documentclass[12pt,compress]{beamer}
%\documentclass[12pt,handout,compress]{beamer}
%\documentclass[handout,compress,10pt]{beamer}
%\usepackage{pgfpages}
%\pgfpagesuselayout{4 on 1}[a4paper,landscape,border shrink=5mm]
%\documentclass[11pt,a4]{powersem}
\usepackage[latin9]{inputenc}
\usepackage[T1]{fontenc}
\usepackage{amsmath,soul}
\usepackage[hyper,procnames]{listings}
\usepackage{alltt}%verbatim,
%\usepackage[linkbordercolor={1 1 1},bookmarks=true,colorlinks=false,urlbordercolor={1 1 1},runbordercolor={1 1 1},citebordercolor={1 1 1},bookmarksnumbered=true,hypertexnames=false]{hyperref}
\usepackage{tabularx}
\usepackage{tikz}
\usepackage{beamer_rk,perso2e,bbold2e,abbrev_reactive,getVersion,exemple}
\usepackage[francais]{babel}
\frenchbsetup{og=«, fg=», ThinColonSpace=true,
  % ShowOptions, 
  % Évite de casser les références bibliographiques du style harvard avec
  % des ":" dedans :
  %AutoSpacePunctuation=false
}
% S'arrange pour avoir une virgule qui ne rajoute pas de blanc dans les
% nombres à virgule :
%\iffalse
\DecimalMathComma
% Et pour avoir \nombre qui marche :
\usepackage[autolanguage]{numprint}
%\usepackage{/usr/share/doc/texlive-doc-de/german/latex-tipps-und-tricks/fraktur}

\makeindex

\lstset{extendedchars=true, language=C, basicstyle=\footnotesize\ttfamily, numbers=left,
  numberstyle=\tiny, stepnumber=2, numberfirstline=true, showspaces=true,
  showstringspaces=false, showtabs=true,
  tabsize=8, tab=\rightarrowfill, keywordstyle=\color{orange}\bf,
  stringstyle=\rmfamily, commentstyle=\rmfamily\itshape,
  index=[1][keywords],indexprocnames=true}
%\lstloadlanguages{Basic}

%\usepackage{beamerthemesplit}
%{\usebeamercolor{palette primary}}
%\setbeamertemplate{sidebar canvas}[vertical shading][top=palette primary.bg,middle=white,bottom=palette primary.bg]

\getRCSDollarsVersion$Revision: 1.22 $

\usepackage[dvips]{preview}
\PreviewEnvironment{trans}
\PreviewEnvironment{frame}
%\fi

% Le plan est petit :
\renewcommand{\OutlineColumnNumbers}{1}

\title[PIPS @ TB]{PIPS-related projects at TÉLÉCOM Bretagne}

%\subtitle{SympA'2008, Fribourg, Suisse}

%\subject{Partie \Atb}

\author[R. \textsc{Keryell}]{Ronan \bsc{Keryell} \texttt{rk@enstb.org}}

\institute[T\rlap{\red \'E}EL\rlap{\red \'E}ECOM Bretagne --- Département
Informatique]{Institut T\rlap{\red \'E}EL\rlap{\red \'E}ECOM, T\rlap{\red
    \'E}EL\rlap{\red \'E}ECOM Bretagne, Département Informatique,
  Lab-STICC/CAC/HPCAS\\
  Plouzané, France}

\date{22 février 2008}


\begin{document}

\begin{frame}[plain]
  \titlepage
\end{frame}

%\newcommand{\Idee}{{\red\Handwash}\xspace}
%\Stopsign \Laserbeam \Estatically \Attention \Idee
%\FilledRainCloud\Coffeecup

\section{CryptoPage}
\label{sec:cryptopage}

\subsection{Introduction}
\label{sec:introduction}


\begin{trans}{Introduction}
  Computer Science department at TÉLÉCOM Bretagne involved in the SAFESCALE
  project:
  \begin{itemize}
  \item Refining secure high-performance processor: CryptoPage
  \item Simulating this processor
  \item Writing a parallelizing tool to transform classical applications
    for the SAFESCALE architecture based on Kaapi execution model
  \end{itemize}
\end{trans}

\subsection{Compilation}
\label{sec:compilation}


\begin{trans}{SAFESCALE execution model}
  \begin{itemize}
  \item Parallel applications describes as a task graph
  \item Communicating tasks with a data-flow model with consumer-producer
    paradigm
  \item SAFESCALE adds secure execution through different mechanisms (from
    more to less efficient)
    \begin{itemize}
    \item Use inherently fault tolerant algorithm if possible (fix point,
      iterative solvers...)
      \begin{itemize}
      \item Open question: how many applications are in this category?
      \end{itemize}
    \item Secure execution platform (if available)
      \begin{itemize}
      \item Trust the microprocessor factory...
      \item Right now, only ``smartcard supercomputers'' (LaBRI) or
        military computing platforms (SAGEM Fox...) \frownie
      \end{itemize}
    \item Use fault tolerant property to component based-applications
    \item Replay few tasks on some more (``verifiers'') or less secure
      nodes to verify they produce the same results
      \begin{itemize}
      \item Use a probabilistic approach to grant a probabilistic security
        level
      \item Use the dependence graph to optimize task verification
        selection
      \end{itemize}
    \item Use the BCP (\emph{Best Current Practices}) to secure execution
      in \emph{the} real world
    \end{itemize}
  \end{itemize}
\end{trans}


\begin{trans}{Writing SAFESCALE applications}
  \begin{itemize}
  \item Write application as a task DAG
  \item Task DAG $\approx$ SSA for imperative programs
    \begin{itemize}
    \item If loops with task calls in them, need to unroll them...
    \item May not be quite convenient for average programmers \smiley
    \end{itemize}
  \item Need to study more deeply Kaapi model and API
  \item Need to define a SAFESCALE API (Kaapi is a good starting point
    \smiley)
  \item A tool to help writing SAFESCALE applications could be nice...
  \end{itemize}
\end{trans}


\begin{trans}{Automatic compilation for SAFESCALE}
  \begin{itemize}
  \item Real programmers love big sequential machines \smiley
  \item Need to split a big application in several tasks...
  \item ...if tasks are parallel, it is better \smiley
  \item Need to figure out
    \begin{itemize}
    \item Data needed by a task to run
    \item Data produced by a task used later by other tasks to run
    \end{itemize}
  \item Add some Kaapi/SAFESCALE glue to
    \begin{itemize}
    \item Define and calling the tasks
    \item Map the task inputs to actual variables used in each task
    \item Map the task variable productions to task output
    \end{itemize}
  \item Orchestrate all these tasks to preserve the initial program
    semantics
  \end{itemize}
  Idea : use the PIPS source-to-source compiler from ENSMP to automate
  this process
\end{trans}


\begin{trans}{Code distribution}
  \begin{itemize}
  \item General problem already studied and going on at ENSMP, ENSTB and
    in many other places
    \begin{itemize}
    \item Compilation for SoC with hardware accelerators : CoMap project
      with UBO \& ENSTB
      \begin{itemize}
      \item Extract pieces of code in an application
      \item Compile them in hardware accelerators
      \item Add some glue to interface to the sequential program (bus, bus
        API, DMA generation...)
      \end{itemize}
      PhD thesis at UBO and MSc of Matthieu \textsc{Godet}
    \item Poor man OpenMP : INT with ENSTB
      \begin{itemize}
      \item Offer distributed shared memory computing without... shared
        memory ! \smiley
      \item Compute more or less the memory zones used as input and output
      \item Mimic the shared memory semantics with different API (remote
        write, MPI...)
      \end{itemize}
    \item Mobython at INT, CNAM and ENSTB
      \begin{itemize}
      \item Build a grid on cellular phones \smiley
      \item Transform an application into distributed tasks
      \end{itemize}
    \item HPF compilation : Jean-Louis \textsc{Pazat} \smiley, ENSMP
    \item SAFESCALE !
    \end{itemize}
  \item Idea \vavers factorizing these transformations in PIPS
  \item MSc of The Nhan \textsc{Luong} and 1-year post-doc to be found for
    SAFESCALE... Done ! \smiley
  \end{itemize}
\end{trans}


\begin{trans}{Distribution concepts in PIPS}
  \begin{itemize}
  \item Many semantics analysis and code transformations available in PIPS
  \item Parallelization transformations usable here to extract parallel
    tasks
  \item Interprocedural dependence graph usable to feed the Kaapi
    scheduler
  \item PIPS compute ``Regions'' that define storage areas used by a piece
    of code : use them to generate communications
  \item Parallelization is in the general case not decidable : need to
    think about SAFESCALE directives to help the compiler too
  \end{itemize}
\end{trans}


\newcommand{\RegionRE}{\green\textsl{-r-exact}}
\newcommand{\RegionWE}{\green\textsl{-w-exact}}
\newcommand{\RegionRI}{\magenta\textsl{-r-inexact}}
\newcommand{\RegionWI}{\magenta\textsl{-w-inexact}}
\newcommand{\RegionIE}{\green\textsl{-in-exact}}
\newcommand{\RegionOE}{\green\textsl{-out-exact}}
\newcommand{\RegionII}{\magenta\textsl{-in-inexact}}
\newcommand{\RegionOI}{\magenta\textsl{-out-inexact}}

\begin{trans}{PIPS read and write array regions}
  \begin{itemize}
  \item Array elements described as integer polyhedra
  \item Integer polyhedra: compromise between expressivity and easy
    mathematical management
  \item Not all the memory accessed can be sum up with polyhedra
  \item \vavers Array regions can be exact (the elements accessed are
    exactly described) or inexact (with more points than really accessed,
    to be conservative)
  \item Regions are built up bottom-up through hierarchical statements
  \end{itemize}
%\begin{multicols}{2}
\begin{lstlisting}[escapechar=\%]
// %N\RegionRE\{\}%
double b[N], a[N]; int i; ...

// %i\RegionWE\{\}%
// %i\RegionRE\{\}%
// %N\RegionRE\{\}%
// %a[$\varphi_1$]\RegionRE\{$0 \leq \varphi_1 \bigwedge \varphi_1 \leq \mathrm{N}$\}%
// %b[$\varphi_1$]\RegionRI\{$0 \leq \varphi_1 \bigwedge \varphi_1 \leq \mathrm{N}$\}%
// Code to execute somewhere else: 
{ 
  // %s\RegionWE\{\}%
  double s = 0;
  // %i\RegionWE\{\}%
  // %i\RegionRE\{\}%
  // %N\RegionRE\{\}%
  // %a[$\varphi_1$]\RegionRE\{$0 \leq \varphi_1 \bigwedge \varphi_1 \leq \mathrm{N}$\}%
  // %b[$\varphi_1$]\RegionRI\{$0 \leq \varphi_1 \bigwedge \varphi_1 \leq \mathrm{N}$\}%
  for(i = 0; i < N; i++) {
    // %i\RegionRE\{\}%
    // %a[$\varphi_1$]\RegionRE\{$\varphi_1 == \mathrm{i}$\}%
    // %s\RegionRE\{\}%
    // %s\RegionWE\{\}%
    s += a[i];
    // %s\RegionRE\{\}%
    // %i\RegionRI\{\}%
    // %a[$\varphi_1$]\RegionRI\{$\varphi_1 == \mathrm{i}$\}%
    // %b[$\varphi_1$]\RegionWI\{$\varphi_1 == \mathrm{i}$\}%
    if (s > 0)
      // %i\RegionRE\{\}%
      // %a[$\varphi_1$]\RegionRE\{$\varphi_1 == \mathrm{i}$\}%
      // %b[$\varphi_1$]\RegionWE\{$\varphi_1 == \mathrm{i}$\}%
      b[i] = 2*a[i];
  }
}
// b is used later
...
\end{lstlisting}
%\end{multicols}
Read and write regions are overkill for us because some statements may
write elements that are not used later\ldots
\end{trans}


\begin{trans}{PIPS in and out array regions}
  \begin{itemize}
  \item Use the dependence graph to compute elements that are
    \emph{really} used by a statement (\emph{in} regions) and that are
    written and will \emph{really} be needed by a future statement
    (\emph{out} regions)
  \end{itemize}
\begin{lstlisting}[escapechar=\%]
// %N\RegionIE\{\}%
double b[N], a[N]; int i; ...

// %N\RegionIE\{\}%
// %a[$\varphi_1$]\RegionIE\{$0 \leq \varphi_1 \bigwedge \varphi_1 \leq \mathrm{N}$\}%
// %b[$\varphi_1$]\RegionOI\{$0 \leq \varphi_1 \bigwedge \varphi_1 \leq \mathrm{N}$\}%
// Code to execute somewhere else: 
{ 
  // %s\RegionOE\{\}%
  double s = 0;
  // %N\RegionIE\{\}%
  // %a[$\varphi_1$]\RegionIE\{$0 \leq \varphi_1 \bigwedge \varphi_1 \leq \mathrm{N}$\}%
  // %b[$\varphi_1$]\RegionOI\{$0 \leq \varphi_1 \bigwedge \varphi_1 \leq \mathrm{N}$\}%
  // %s\RegionIE\{\}%
  for(i = 0; i < N; i++) {
    // %i\RegionIE\{\}%
    // %a[$\varphi_1$]\RegionIE\{$\varphi_1 == \mathrm{i}$\}%
    // %s\RegionIE\{\}%
    // %s\RegionOE\{\}%
    s += a[i];
    // %s\RegionIE\{\}%
    // %i\RegionII\{\}%
    // %a[$\varphi_1$]\RegionII\{$\varphi_1 == \mathrm{i}$\}%
    // %b[$\varphi_1$]\RegionOI\{$\varphi_1 == \mathrm{i}$\}%
    if (s > 0)
      // %i\RegionIE\{\}%
      // %a[$\varphi_1$]\RegionIE\{$\varphi_1 == \mathrm{i}$\}%
      // %b[$\varphi_1$]\RegionOE\{$\varphi_1 == \mathrm{i}$\}%
      b[i] = 2*a[i];
  }
}
// b is used later
...
\end{lstlisting}
  
\end{trans}


\begin{trans}{SAFESCALE generation blueprint}
  To distribute a statement $\mathcal{S}$ on a node $\mathcal{N}$:
  \begin{itemize}
  \item $\mathtt{receive}_{\mathcal{N}}(e\in \mathrm{InExact}(\mathcal{S}) \bigcup
    \mathrm{InInexact}(\mathcal{S}))$
  \item $\mathtt{receive}_{\mathcal{N}}(e\in \mathrm{OutInexact}(\mathcal{S}))$
  \item $\mathtt{executeTask}_{\mathcal{N}}(\mathcal{S})$
  \item $\mathtt{send}_{\mathcal{N}}(e\in \mathrm{OutExact}(\mathcal{S}) \bigcup
    \mathrm{OutInexact}(\mathcal{S}))$
  \end{itemize}
\end{trans}


\begin{trans}{Optimized SAFESCALE generation}
  \begin{itemize}
  \item For inexact out regions, may use combining write instead of
    read-then-write
    \begin{itemize}
    \item Need to track modified elements with a run-time resolution
    \item May use inspector-executors
    \item Need to detect loop invariant region patterns
    \end{itemize}
  \item In the generation sketch up, guess that variables are allocated in
    the same way on each task using them
    \begin{itemize}
    \item Quite inefficient if a task use only few element
    \item Use a more compact allocation
    \end{itemize}
  \item What about the general pointers ?
  \item Some science fiction: reorganize globally variables for more
    efficient communications and data access (spitting red-black
    relaxation data into red and black in 2 different arrays\ldots)
  \item Big collaboration needed here in SAFESCALE : everything still to
    do \smiley
  \end{itemize}
\end{trans}



\subsection{Conclusion}
\label{sec:conclusion}



\begin{trans}{Conclusion}
  \begin{itemize}
  \item CryptoPage architecture developing at good pace with PhD and MSc
    students
  \item Good simulation results : only 3\,\% slowdown on average by
    combining HIDE, fast cipher at page level and speculative insecure
    execution (presentation at ACSAC 12/2006)
  \item Interesting side effects: need SAFESCALE to simulate
    \st{SAFESCALE}CryptoPage \smiley
  \item Still looking for a (good \smiley) 1-year post-doc to work on
    compilation aspect with \vavers cold start for this part of the
    project with IMAG team \frownie
  \item Factorizing compilation aspects for other heterogeneous parallel
    machines 
  \end{itemize}
\end{trans}


\section{Phases développées à TÉLÉCOM Bretagne}
\label{sec:phases-dveloppes-}

\begin{frame}{Projet PHRASE}
  Générer code pour acélérateurs configurables
  \begin{itemize}
  \item Détourage de code marqué par directives qui est déporté sur
    accélérateur
  \item Atomiseur de code 3 adresses
  \item Génération bus logiciel/matériel
  \item Transformation graphe de contrôle en FSM hiérarchique ou à plat
  \item Prettyprinter SmallTalk pour outil de synthèse Madeo du
    Lab-SSTIC/CAC/AS
  \end{itemize}
\end{frame}


\begin{frame}{Projet SAC}
  SIMD Architecture Compiler
  \begin{itemize}
  \item Vectoriseur par déroulage et rassemblement d'instructions « en
    vrac »
  \item Calcul taille opérandes à partir des précondition
  \item Génération d'intrinsèques à partir fichier configuration
    architecture
  \item Gère réduction
  \end{itemize}
\end{frame}


\begin{frame}{Projet CoMap}
  \begin{itemize}
  \item Détourage de code marqué par directives
  \item Génération de descripteur DMA et des synchronisations
  \item Génération code de contrôle
  \item Génération du code des opérateurs
  \end{itemize}
\end{frame}


\section{Moult travaux futurs}
\label{sec:moult-travaux-futurs}



\begin{frame}{NewGen}
  Échanger données avec autres outils ?
  \begin{itemize}
  \item Rajouter un backend \Axml
  \item Regarder STEP/EXPRESS ISO 10303 pour échanger données ?
  \end{itemize}
  \Attention La syntaxe ne fait pas tout... Quid de la sémantique \frownie
\end{frame}


\begin{frame}{Interfaces graphiques}
  {\red \Estatically} Remettre en ordre...
  \begin{itemize}
  \item \texttt{Jpips} (au dessus de \texttt{tpips}) devrait encore
    marcher ?
  \item \texttt{wpips} utilise \texttt{XView} plus maintenu \frownie
  \item \texttt{epips} utilise GNU/Emacs mais au dessus de \texttt{wpips}
  \end{itemize}
  \belleboite{{\red \Laserbeam} Epips dans Eclipse ?}

  Rajouter snapshot dans PIPSdbm pour permettre de défaire des actions
\end{frame}


\begin{frame}{PIPS \& checkpoint}
  Problème de robustesse des machines à moult processeurs\\
  \begin{example}
    BlueGene EPFL : 1 panne/2,17 jours
  \end{example}
  \begin{itemize}
  \item Contact avec le LRI sur snapshot/checkpoint
  \item Collaboration avec TÉLÉCOM SudParis qui travaille dans le domaine
    ?

  \end{itemize}
\end{frame}


\begin{trans}{Relations avec IEF}
  \begin{itemize}
  \item IEF QUAFF génère code à partir de templates C++ pour programmes
    basés sur squelettes algorithmiques
  \item Utilisation de PIPS
    \begin{itemize}
    \item Front-end : génère programmes QUAFF
    \item Back-end/Middle-end : optimise/transforme sortie de QUAFF
    \end{itemize}
  \item Question : peut-on récupérer C++ détemplatisé sans génération de
    code ? Comment générer du C++ (compilable par la suite y compris avec
    des templates) plutôt que de l'objet ?

    Si oui, permettrait d'utiliser QUAFF comme middle-end moult simplement
    pour moult cibles
    \begin{itemize}
    \item Langages StreamC, RapidMind, CUDA, HMPP (CAPS Entreprise)
    \item Bibliothèques AMD FrameWave, Intel Threading Building Blocks,
      KAAPI (LIG)
    \item MPI (\textsc{Skell} BE)
    \item OpenMP, \textsc{Forest}GOMP (LaBRI)
    \item Implémentations de C++ STL PaSTeL (multi-c½urs, LIG), MCSTL
      (Karlsruhe), STXXL, STAPL...
    \item FREIA : templates pour C++ ou SystemC synthétisable
    \item PIPS en post-production : rajouter des annotations
      éventuellement pour raffiner avec PIPS en post-production ?
    \end{itemize}
  \end{itemize}
\end{trans}


\begin{trans}{À rajouter/faire dans PIPS}
  \begin{itemize}
  \item Terminer passage en 64 bits
  \item Refaire un site WWW moderne
    \begin{itemize}
    \item Utiliser un CMS collaboratif, style bibliographie interactive
      \LienPDF{hpcas.enstb.org/resources-related-projects}
    \end{itemize}
  \item Généralisation de la transformation de programmes
  \item Utiliser langage de réécriture de graphes ? Pattern matching dans
    PIPS ?
  \item Tiling généralisé et paramétrable
  \item Pipeline logiciel
  \item Généralisation factorisation sous-expressions communes
  \item Faire état des lieux de tout ce qui a été fait et refactoriser
  \end{itemize}  
\end{trans}


\begin{frame}{Conclusion}
  \begin{itemize}
  \item Moult choses disponibles dans PIPS
  \item Moult choses encore à faire
  \item Projet toujours vivant au bout de 20 ans ! {\yellow \Sun}
  \item {\FilledSunCloud} Gros travail de public relation à faire...
  \item {\red \Coffeecup} Organiser fiesta pour les 20 ans de PIPS !
    {\yellow \smiley}
  \end{itemize}
\end{frame}


\section*{Table des matières}

% Do not use frame environment because it does not split multicols
\begin{multicols}{2}
  \tiny
  \tableofcontents[frametitles]
  \textbf{Vous êtes ici      !}\hfill\expandafter\the\csname c@page\endcsname
\end{multicols}


\end{document}

%%% Local Variables: 
%%% mode: latex
%%% TeX-master: t
%%% End: