From e2f7ea69b2321fbf77291f35360751e460a99f44 Mon Sep 17 00:00:00 2001 From: couturie Date: Tue, 23 Jul 2013 21:57:49 +0200 Subject: [PATCH] new --- BookGPU/Chapters/chapter12/biblio12.bib | 53 +- BookGPU/Chapters/chapter12/ch12.tex | 279 ++-- BookGPU/Chapters/chapter12/ch12.tex~ | 1232 ----------------- .../Chapters/chapter12/figures/cluster.eps | 4 +- .../Chapters/chapter12/figures/cluster.fig | 2 +- .../chapter12/figures/cluster.fig.bak | 9 +- .../Chapters/chapter12/figures/cluster.pdf | Bin 14158 -> 14213 bytes .../Chapters/chapter12/figures/compress.eps | 278 ++-- .../Chapters/chapter12/figures/compress.fig | 222 +-- .../chapter12/figures/compress.fig.bak | 10 +- .../Chapters/chapter12/figures/compress.pdf | Bin 13541 -> 13606 bytes .../Chapters/chapter12/figures/partition.eps | 36 +- .../Chapters/chapter12/figures/partition.fig | 200 +-- .../chapter12/figures/partition.fig.bak | 205 +-- .../Chapters/chapter12/figures/partition.pdf | Bin 10695 -> 11064 bytes .../Chapters/chapter12/figures/reorder.eps | 236 ++-- .../Chapters/chapter12/figures/reorder.fig | 184 +-- .../chapter12/figures/reorder.fig.bak | 24 +- .../Chapters/chapter12/figures/reorder.pdf | Bin 9341 -> 9228 bytes BookGPU/Chapters/chapter13/biblio13.bib | 84 +- BookGPU/Chapters/chapter13/ch12.tex~ | 1080 --------------- BookGPU/Chapters/chapter13/ch13.tex | 303 ++-- BookGPU/Chapters/chapter13/ch13.tex~ | 710 ---------- BookGPU/Chapters/chapter13/ex1.cu | 2 +- .../Chapters/chapter13/figures/splitGPU.eps | 17 +- .../Chapters/chapter13/figures/splitGPU.pdf | Bin 9410 -> 9341 bytes 26 files changed, 1049 insertions(+), 4121 deletions(-) delete mode 100755 BookGPU/Chapters/chapter12/ch12.tex~ delete mode 100755 BookGPU/Chapters/chapter13/ch12.tex~ delete mode 100755 BookGPU/Chapters/chapter13/ch13.tex~ diff --git a/BookGPU/Chapters/chapter12/biblio12.bib b/BookGPU/Chapters/chapter12/biblio12.bib index f7dd392..803d354 100644 --- a/BookGPU/Chapters/chapter12/biblio12.bib +++ b/BookGPU/Chapters/chapter12/biblio12.bib @@ -1,36 +1,35 @@ -@article{ch12:ref1, -title = {Iterative methods for sparse linear systems}, + +@book{ch12:ref1, author = {Saad, Y.}, -journal = {Society for Industrial and Applied Mathematics, 2nd edition}, -volume = {}, -number = {}, -pages = {}, +title = {Iterative Methods for Sparse Linear Systems}, +publisher = {Society for Industrial and Applied Mathematics}, year = {2003}, +edition = {Second} } @article{ch12:ref2, title = {Methods of conjugate gradients for solving linear systems}, -author = {Hestenes, M. R. and Stiefel, E.}, +author = {Hestenes, M.R. and Stiefel, E.}, journal = {Journal of Research of the National Bureau of Standards}, volume = {49}, number = {6}, pages = {409--436}, -year = {1952}, +year = {1952} } @article{ch12:ref3, title = {{GMRES}: a generalized minimal residual algorithm for solving nonsymmetric linear systems}, -author = {Saad, Y. and Schultz, M. H.}, +author = {Saad, Y. and Schultz, M.H.}, journal = {SIAM Journal on Scientific and Statistical Computing}, volume = {7}, number = {3}, pages = {856--869}, -year = {1986}, +year = {1986} } @article{ch12:ref4, title = {Solution of sparse indefinite systems of linear equations}, -author = {Paige, C. C. and Saunders, M. A.}, +author = {Paige, C.C. and Saunders, M.A.}, journal = {SIAM Journal on Numerical Analysis}, volume = {12}, number = {4}, @@ -40,7 +39,7 @@ year = {1975}, @article{ch12:ref5, title = {The principle of minimized iteration in the solution of the matrix eigenvalue problem}, -author = {Arnoldi, W. E.}, +author = {Arnoldi, W.E.}, journal = {Quarterly of Applied Mathematics}, volume = {9}, number = {17}, @@ -49,8 +48,8 @@ year = {1951}, } @article{ch12:ref6, -title = {{CUDA} Toolkit 4.2 {CUBLAS} Library}, -author = {NVIDIA Corporation}, +title = {{CUDA} {T}oolkit 4.2 {CUBLAS} {L}ibrary}, +author = {NVIDIA {C}orporation}, journal = {}, volume = {}, number = {}, @@ -59,14 +58,14 @@ note = {\url{http://developer.download.nvidia.com/compute/DevZone/docs/html/CUDA year = {2012}, } -@article{ch12:ref7, -title = {Efficient sparse matrix-vector multiplication on {CUDA}}, -author = {Bell, N. and Garland, M.}, -journal = {NVIDIA Technical Report NVR-2008-004, NVIDIA Corporation}, -volume = {}, -number = {}, -pages = {}, -year = {2008}, +@techreport{ch12:ref7, + author = {Bell, N. and Garland, M.}, + title = {Efficient Sparse Matrix-Vector Multiplication on {CUDA}}, + month = dec, + year = 2008, + institution = {NVIDIA Corporation}, + type = {NVIDIA Technical Report}, + number = {NVR-2008-004}, } @article{ch12:ref8, @@ -81,7 +80,7 @@ year = {}, } @article{ch12:ref9, -title = {{NVIDIA} {CUDA} {C} programming guide}, +title = {{NVIDIA} {CUDA} {C} {P}rogramming {G}uide}, author = {NVIDIA Corporation}, journal = {}, volume = {}, @@ -92,7 +91,7 @@ year = {2012}, } @article{ch12:ref10, -title = {The university of {F}lorida sparse matrix collection}, +title = {The {U}niversity of {F}lorida {S}parse {M}atrix {C}ollection}, author = {Davis, T. and Hu, Y.}, journal = {}, volume = {}, @@ -115,7 +114,7 @@ year = {1999}, @article{ch12:ref12, title = {{hMETIS}: A hypergraph partitioning package}, -author = {Karypis, George and Kumar, Vipin}, +author = {Karypis, G. and Kumar, V.}, journal = {}, volume = {}, number = {}, @@ -126,7 +125,7 @@ year = {1998}, @article{ch12:ref13, title = {{PaToH}: partitioning tool for hypergraphs}, -author = {Catalyurek, Umit V. and Aykanat, Cevdet}, +author = {Catalyurek, U.V. and Aykanat, C.}, journal = {}, volume = {}, number = {}, @@ -137,7 +136,7 @@ year = {1999}, @article{ch12:ref14, title = {Parallel hypergraph partitioning for scientific computing}, -author = {Devine, Karen D. and Boman, Erik G. and Heaphy, Robert T. and Bisseling, Rob H. and Catalyurek, Umit V.}, +author = {Devine, K.D. and Boman, E.G. and Heaphy, R.T. and Bisseling, R.H. and Catalyurek, U.V.}, journal = {In Proceedings of the 20th international conference on Parallel and distributed processing, IPDPS’06}, volume = {}, number = {}, diff --git a/BookGPU/Chapters/chapter12/ch12.tex b/BookGPU/Chapters/chapter12/ch12.tex index 0b743eb..3843597 100755 --- a/BookGPU/Chapters/chapter12/ch12.tex +++ b/BookGPU/Chapters/chapter12/ch12.tex @@ -5,7 +5,7 @@ %%%%%%%%%%%%%%%%%%%%%%%%%%%%%% %\chapterauthor{}{} -\chapterauthor{Lilia Ziane Khodja, Raphaël Couturier and Jacques Bahi}{Femto-ST Institute, University of Franche-Comte, France} +\chapterauthor{Lilia Ziane Khodja, Raphaël Couturier, and Jacques Bahi}{Femto-ST Institute, University of Franche-Comte, France} %\chapterauthor{Raphaël Couturier}{Femto-ST Institute, University of Franche-Comte, France} %\chapterauthor{Jacques Bahi}{Femto-ST Institute, University of Franche-Comte, France} @@ -20,15 +20,15 @@ Sparse linear systems are used to model many scientific and industrial problems, such as the environmental simulations or the industrial processing of the complex or non-Newtonian fluids. Moreover, the resolution of these problems often involves the -solving of such linear systems which are considered as the most expensive process in +solving of such linear systems that are considered the most expensive process in terms of execution time and memory space. Therefore, solving sparse linear systems must be as efficient as possible in order to deal with problems of ever increasing size. There are, in the jargon of numerical analysis, different methods of solving sparse -linear systems that can be classified in two classes: the direct and iterative methods. -However, the iterative methods are often more suitable than their counterpart, direct -methods, to solve these systems. Indeed, they are less memory consuming and easier +linear systems that can be classified in two classes: direct and iterative methods. +However, the iterative methods are often more suitable than their counterparts, direct +methods, to solve these systems. Indeed, they are less memory-consuming and easier to parallelize on parallel computers than direct methods. Different computing platforms, sequential and parallel computers, are used to solve sparse linear systems with iterative solutions. Nowadays, graphics processing units (GPUs) have become attractive to solve @@ -38,8 +38,8 @@ traditional CPUs. In Section~\ref{ch12:sec:02}, we describe the general principle of two well-known iterative methods: the conjugate gradient method and the generalized minimal residual method. In Section~\ref{ch12:sec:03}, we give the main key points of the parallel implementation of both methods on a cluster of -GPUs. Finally, in Section~\ref{ch12:sec:04}, we present the experimental results obtained on a -CPU cluster and on a GPU cluster, to solve large sparse linear systems. +GPUs. Finally, in Section~\ref{ch12:sec:04}, we present the experimental results, obtained on a +CPU cluster and on a GPU cluster of solving large sparse linear systems. %%--------------------------%% @@ -54,12 +54,12 @@ Ax=b, \label{ch12:eq:01} \end{equation} where $A\in\mathbb{R}^{n\times n}$ is a sparse nonsingular square matrix, $x\in\mathbb{R}^{n}$ -is the solution vector, $b\in\mathbb{R}^{n}$ is the right-hand side and $n\in\mathbb{N}$ is a +is the solution vector, $b\in\mathbb{R}^{n}$ is the right-hand side, and $n\in\mathbb{N}$ is a large integer number. The iterative methods\index{Iterative~method} for solving the large sparse linear system~(\ref{ch12:eq:01}) proceed by successive iterations of a same block of elementary operations, during which an -infinite number of approximate solutions $\{x_k\}_{k\geq 0}$ are computed. Indeed, from an +infinite number of approximate solutions $\{x_k\}_{k\geq 0}$ is computed. Indeed, from an initial guess $x_0$, an iterative method determines at each iteration $k>0$ an approximate solution $x_k$ which, gradually, converges to the exact solution $x^{*}$ as follows: \begin{equation} @@ -78,9 +78,9 @@ where $\varepsilon<1$ is the required convergence tolerance threshold\index{Conv Some of the most iterative methods that have proven their efficiency for solving large sparse linear systems are those called \textit{Krylov subspace methods}~\cite{ch12:ref1}\index{Iterative~method!Krylov~subspace}. -In the present chapter, we describe two Krylov methods which are widely used: the conjugate -gradient method (CG) and the generalized minimal residual method (GMRES). In practice, the -Krylov subspace methods are usually used with preconditioners that allow to improve their +In the present chapter, we describe two Krylov methods which are widely used: the CG method (conjugate +gradient method) and the GMRES method (generalized minimal residual method). In practice, the +Krylov subspace methods are usually used with preconditioners that allow the improvement of their convergence. So, in what follows, the CG and GMRES methods are used to solve the left-preconditioned\index{Sparse~linear~system!Preconditioned} sparse linear system: \begin{equation} @@ -95,7 +95,7 @@ where $M$ is the preconditioning matrix. \subsection{CG method} \label{ch12:sec:02.01} The conjugate gradient method was initially developed by Hestenes and Stiefel in 1952~\cite{ch12:ref2}. -It is one of the well known iterative method to solve large sparse linear systems. In addition, it +It is one of the well-known iterative methods to solve large sparse linear systems. In addition, it can be adapted to solve nonlinear equations and optimization problems. However, it can only be applied to problems with positive definite symmetric matrices. @@ -111,7 +111,7 @@ such that the Galerkin condition\index{Galerkin~condition} must be satisfied: r_k \bot \mathcal{K}_k(A,r_0), \label{ch12:eq:05} \end{equation} -where $x_0$ is the initial guess, $r_k=b-Ax_k$ is the residual of the computed solution $x_k$ and $\mathcal{K}_k$ +where $x_0$ is the initial guess, $r_k=b-Ax_k$ is the residual of the computed solution $x_k$, and $\mathcal{K}_k$ the Krylov subspace of order $k$: \[\mathcal{K}_k(A,r_0) \equiv\text{span}\{r_0, Ar_0, A^2r_0,\ldots, A^{k-1}r_0\}.\] In fact, CG is based on the construction of a sequence $\{p_k\}_{k\in\mathbb{N}}$ of direction vectors in $\mathcal{K}_k$ which are pairwise $A$-conjugate ($A$-orthogonal): @@ -142,9 +142,9 @@ p_0=r_0, & p_k=r_k+\beta_k p_{k-1}, & \beta_k\in\mathbb{R}. \label{ch12:eq:09} \end{equation} Moreover, the scalars $\{\alpha_k\}_{k>0}$ are chosen so as to minimize the $A$-norm error $\|x^{*}-x_k\|_A$ -over the Krylov subspace $\mathcal{K}_{k}$ and the scalars $\{\beta_k\}_{k>0}$ are chosen so as to ensure +over the Krylov subspace $\mathcal{K}_{k}$, and the scalars $\{\beta_k\}_{k>0}$ are chosen so as to ensure that the direction vectors are pairwise $A$-conjugate. So, the assumption that matrix $A$ is symmetric and -the recurrences~(\ref{ch12:eq:08}) and~(\ref{ch12:eq:09}) allow to deduce that: +the recurrences~(\ref{ch12:eq:08}) and~(\ref{ch12:eq:09}) allow the deduction that: \begin{equation} \begin{array}{ll} \alpha_{k}=\frac{r^{T}_{k-1}r_{k-1}}{p_{k}^{T}Ap_{k}}, & \beta_{k}=\frac{r_{k}^{T}r_{k}}{r_{k-1}^{T}r_{k-1}}. @@ -176,21 +176,21 @@ the recurrences~(\ref{ch12:eq:08}) and~(\ref{ch12:eq:09}) allow to deduce that: $k = k + 1$\; } } -\caption{Left-preconditioned CG method} +\caption{left-preconditioned CG method} \label{ch12:alg:01} \end{algorithm} Algorithm~\ref{ch12:alg:01} shows the main key points of the preconditioned CG method. It allows -to solve the left-preconditioned\index{Sparse~linear~system!Preconditioned} sparse linear system~(\ref{ch12:eq:11}). +the solving the left-preconditioned\index{Sparse~linear~system!Preconditioned} sparse linear system~(\ref{ch12:eq:11}). In this algorithm, $\varepsilon$ is the convergence tolerance threshold, $maxiter$ is the maximum -number of iterations and $(\cdot,\cdot)$ defines the dot product between two vectors in $\mathbb{R}^{n}$. +number of iterations, and $(\cdot,\cdot)$ defines the dot product between two vectors in $\mathbb{R}^{n}$. At every iteration, a direction vector $p_k$ is determined, so that it is orthogonal to the preconditioned residual $z_k$ and to the direction vectors $\{p_i\}_{i0$ an approximate -solution $x_k$ which, gradually, converges to the exact solution $x^{*}$ as follows: -\begin{equation} -x^{*}=\lim\limits_{k\to\infty}x_{k}=A^{-1}b. -\label{ch12:eq:02} -\end{equation} -The number of iterations necessary to reach the exact solution $x^{*}$ is not known beforehand -and can be infinite. In practice, an iterative method often finds an approximate solution $\tilde{x}$ -after a fixed number of iterations and/or when a given convergence criterion\index{Convergence} -is satisfied as follows: -\begin{equation} -\|b-A\tilde{x}\| < \varepsilon, -\label{ch12:eq:03} -\end{equation} -where $\varepsilon<1$ is the required convergence tolerance threshold\index{Convergence!Tolerance~threshold}. - -Some of the most iterative methods that have proven their efficiency for solving large sparse -linear systems are those called \textit{Krylov subspace methods}~\cite{ch12:ref1}\index{Iterative~method!Krylov~subspace}. -In the present chapter, we describe two Krylov methods which are widely used: the conjugate -gradient method (CG) and the generalized minimal residual method (GMRES). In practice, the -Krylov subspace methods are usually used with preconditioners that allow to improve their -convergence. So, in what follows, the CG and GMRES methods are used for solving the left-preconditioned\index{Sparse~linear~system!Preconditioned} -sparse linear system: -\begin{equation} -M^{-1}Ax=M^{-1}b, -\label{ch12:eq:11} -\end{equation} -where $M$ is the preconditioning matrix. - - -%%****************%% -%%****************%% -\subsection{CG method} -\label{ch12:sec:02.01} -The conjugate gradient method is initially developed by Hestenes and Stiefel in 1952~\cite{ch12:ref2}. -It is one of the well known iterative method for solving large sparse linear systems. In addition, it -can be adapted for solving nonlinear equations and optimization problems. However, it can only be applied -to problems with positive definite symmetric matrices. - -The main idea of the CG method\index{Iterative~method!CG} is the computation of a sequence of approximate -solutions $\{x_k\}_{k\geq 0}$ in a Krylov subspace\index{Iterative~method!Krylov~subspace} of order $k$ as -follows: -\begin{equation} -x_k \in x_0 + \mathcal{K}_k(A,r_0), -\label{ch12:eq:04} -\end{equation} -such that the Galerkin condition\index{Galerkin~condition} must be satisfied: -\begin{equation} -r_k \bot \mathcal{K}_k(A,r_0), -\label{ch12:eq:05} -\end{equation} -where $x_0$ is the initial guess, $r_k=b-Ax_k$ is the residual of the computed solution $x_k$ and $\mathcal{K}_k$ -the Krylov subspace of order $k$: \[\mathcal{K}_k(A,r_0) \equiv\text{span}\{r_0, Ar_0, A^2r_0,\ldots, A^{k-1}r_0\}.\] -In fact, CG is based on the construction of a sequence $\{p_k\}_{k\in\mathbb{N}}$ of direction vectors in $\mathcal{K}_k$ -which are pairwise $A$-conjugate ($A$-orthogonal): -\begin{equation} -\begin{array}{ll} -p_i^T A p_j = 0, & i\neq j. -\end{array} -\label{ch12:eq:06} -\end{equation} -At each iteration $k$, an approximate solution $x_k$ is computed by recurrence as follows: -\begin{equation} -\begin{array}{ll} -x_k = x_{k-1} + \alpha_k p_k, & \alpha_k\in\mathbb{R}. -\end{array} -\label{ch12:eq:07} -\end{equation} -Consequently, the residuals $r_k$ are computed in the same way: -\begin{equation} -r_k = r_{k-1} - \alpha_k A p_k. -\label{ch12:eq:08} -\end{equation} -In the case where all residuals are nonzero, the direction vectors $p_k$ can be determined so that -the following recurrence holds: -\begin{equation} -\begin{array}{lll} -p_0=r_0, & p_k=r_k+\beta_k p_{k-1}, & \beta_k\in\mathbb{R}. -\end{array} -\label{ch12:eq:09} -\end{equation} -Moreover, the scalars $\{\alpha_k\}_{k>0}$ are chosen so as to minimize the $A$-norm error $\|x^{*}-x_k\|_A$ -over the Krylov subspace $\mathcal{K}_{k}$ and the scalars $\{\beta_k\}_{k>0}$ are chosen so as to ensure -that the direction vectors are pairwise $A$-conjugate. So, the assumption that matrix $A$ is symmetric and -the recurrences~(\ref{ch12:eq:08}) and~(\ref{ch12:eq:09}) allow to deduce that: -\begin{equation} -\begin{array}{ll} -\alpha_{k}=\frac{r^{T}_{k-1}r_{k-1}}{p_{k}^{T}Ap_{k}}, & \beta_{k}=\frac{r_{k}^{T}r_{k}}{r_{k-1}^{T}r_{k-1}}. -\end{array} -\label{ch12:eq:10} -\end{equation} - -\begin{algorithm}[!t] - Choose an initial guess $x_0$\; - $r_{0} = b - A x_{0}$\; - $convergence$ = false\; - $k = 1$\; - \Repeat{convergence}{ - $z_{k} = M^{-1} r_{k-1}$\; - $\rho_{k} = (r_{k-1},z_{k})$\; - \eIf{$k = 1$}{ - $p_{k} = z_{k}$\; - }{ - $\beta_{k} = \rho_{k} / \rho_{k-1}$\; - $p_{k} = z_{k} + \beta_{k} \times p_{k-1}$\; - } - $q_{k} = A \times p_{k}$\; - $\alpha_{k} = \rho_{k} / (p_{k},q_{k})$\; - $x_{k} = x_{k-1} + \alpha_{k} \times p_{k}$\; - $r_{k} = r_{k-1} - \alpha_{k} \times q_{k}$\; - \eIf{$(\rho_{k} < \varepsilon)$ {\bf or} $(k \geq maxiter)$}{ - $convergence$ = true\; - }{ - $k = k + 1$\; - } - } -\caption{Left-preconditioned CG method} -\label{ch12:alg:01} -\end{algorithm} - -Algorithm~\ref{ch12:alg:01} shows the main key points of the preconditioned CG method. It allows -to solve the left-preconditioned\index{Sparse~linear~system!Preconditioned} sparse linear system~(\ref{ch12:eq:11}). -In this algorithm, $\varepsilon$ is the convergence tolerance threshold, $maxiter$ is the maximum -number of iterations and $(\cdot,\cdot)$ defines the dot product between two vectors in $\mathbb{R}^{n}$. -At every iteration, a direction vector $p_k$ is determined, so that it is orthogonal to the preconditioned -residual $z_k$ and to the direction vectors $\{p_i\}_{i0}$ in -a Krylov subspace\index{Iterative~method!Krylov~subspace} $\mathcal{K}_k$ as follows: -\begin{equation} -\begin{array}{ll} -x_k \in x_0 + \mathcal{K}_k(A, v_1),& v_1=\frac{r_0}{\|r_0\|_2}, -\end{array} -\label{ch12:eq:12} -\end{equation} -so that the Petrov-Galerkin condition\index{Petrov-Galerkin~condition} is satisfied: -\begin{equation} -\begin{array}{ll} -r_k \bot A \mathcal{K}_k(A, v_1). -\end{array} -\label{ch12:eq:13} -\end{equation} -GMRES uses the Arnoldi process~\cite{ch12:ref5}\index{Iterative~method!Arnoldi~process} to construct an -orthonormal basis $V_k$ for the Krylov subspace $\mathcal{K}_k$ and an upper Hessenberg matrix\index{Hessenberg~matrix} -$\bar{H}_k$ of order $(k+1)\times k$: -\begin{equation} -\begin{array}{ll} -V_k = \{v_1, v_2,\ldots,v_k\}, & \forall k>1, v_k=A^{k-1}v_1, -\end{array} -\label{ch12:eq:14} -\end{equation} -and -\begin{equation} -V_k A = V_{k+1} \bar{H}_k. -\label{ch12:eq:15} -\end{equation} - -Then, at each iteration $k$, an approximate solution $x_k$ is computed in the Krylov subspace $\mathcal{K}_k$ -spanned by $V_k$ as follows: -\begin{equation} -\begin{array}{ll} -x_k = x_0 + V_k y, & y\in\mathbb{R}^{k}. -\end{array} -\label{ch12:eq:16} -\end{equation} -From both formulas~(\ref{ch12:eq:15}) and~(\ref{ch12:eq:16}) and $r_k=b-Ax_k$, we can deduce that: -\begin{equation} -\begin{array}{lll} - r_{k} & = & b - A (x_{0} + V_{k}y) \\ - & = & r_{0} - AV_{k}y \\ - & = & \beta v_{1} - V_{k+1}\bar{H}_{k}y \\ - & = & V_{k+1}(\beta e_{1} - \bar{H}_{k}y), -\end{array} -\label{ch12:eq:17} -\end{equation} -such that $\beta=\|r_0\|_2$ and $e_1=(1,0,\cdots,0)$ is the first vector of the canonical basis of -$\mathbb{R}^k$. So, the vector $y$ is chosen in $\mathbb{R}^k$ so as to minimize at best the Euclidean -norm of the residual $r_k$. Consequently, a linear least-squares problem of size $k$ is solved: -\begin{equation} -\underset{y\in\mathbb{R}^{k}}{min}\|r_{k}\|_{2}=\underset{y\in\mathbb{R}^{k}}{min}\|\beta e_{1}-\bar{H}_{k}y\|_{2}. -\label{ch12:eq:18} -\end{equation} -The QR factorization of matrix $\bar{H}_k$ is used to compute the solution of this problem by using -Givens rotations~\cite{ch12:ref1,ch12:ref3}, such that: -\begin{equation} -\begin{array}{lll} -\bar{H}_{k}=Q_{k}R_{k}, & Q_{k}\in\mathbb{R}^{(k+1)\times (k+1)}, & R_{k}\in\mathbb{R}^{(k+1)\times k}, -\end{array} -\label{ch12:eq:19} -\end{equation} -where $Q_kQ_k^T=I_k$ and $R_k$ is an upper triangular matrix. - -The GMRES method computes an approximate solution with a sufficient precision after, at most, $n$ -iterations ($n$ is the size of the sparse linear system to be solved). However, the GMRES algorithm -must construct and store in the memory an orthonormal basis $V_k$ whose size is proportional to the -number of iterations required to achieve the convergence. Then, to avoid a huge memory storage, the -GMRES method must be restarted at each $m$ iterations, such that $m$ is very small ($m\ll n$), and -with $x_m$ as the initial guess to the next iteration. This allows to limit the size of the basis -$V$ to $m$ orthogonal vectors. - -\begin{algorithm}[!t] - Choose an initial guess $x_0$\; - $convergence$ = false\; - $k = 1$\; - $r_{0} = M^{-1}(b-Ax_{0})$\; - $\beta = \|r_{0}\|_{2}$\; - \While{$\neg convergence$}{ - $v_{1} = r_{0}/\beta$\; - \For{$j=1$ \KwTo $m$}{ - $w_{j} = M^{-1}Av_{j}$\; - \For{$i=1$ \KwTo $j$}{ - $h_{i,j} = (w_{j},v_{i})$\; - $w_{j} = w_{j}-h_{i,j}v_{i}$\; - } - $h_{j+1,j} = \|w_{j}\|_{2}$\; - $v_{j+1} = w_{j}/h_{j+1,j}$\; - } - Set $V_{m}=\{v_{j}\}_{1\leq j \leq m}$ and $\bar{H}_{m}=(h_{i,j})$ a $(m+1)\times m$ upper Hessenberg matrix\; - Solve a least-squares problem of size $m$: $min_{y\in\mathrm{I\!R}^{m}}\|\beta e_{1}-\bar{H}_{m}y\|_{2}$\; - $x_{m} = x_{0}+V_{m}y_{m}$\; - $r_{m} = M^{-1}(b-Ax_{m})$\; - $\beta = \|r_{m}\|_{2}$\; - \eIf{ $(\beta<\varepsilon)$ {\bf or} $(k\geq maxiter)$}{ - $convergence$ = true\; - }{ - $x_{0} = x_{m}$\; - $r_{0} = r_{m}$\; - $k = k + 1$\; - } - } -\caption{Left-preconditioned GMRES method with restarts} -\label{ch12:alg:02} -\end{algorithm} - -Algorithm~\ref{ch12:alg:02} shows the main key points of the GMRES method with restarts. -It solves the left-preconditioned\index{Sparse~linear~system!Preconditioned} sparse linear -system~(\ref{ch12:eq:11}), such that $M$ is the preconditioning matrix. At each iteration -$k$, GMRES uses the Arnoldi process\index{Iterative~method!Arnoldi~process} (defined from -line~$7$ to line~$17$) to construct a basis $V_m$ of $m$ orthogonal vectors and an upper -Hessenberg matrix\index{Hessenberg~matrix} $\bar{H}_m$ of size $(m+1)\times m$. Then, it -solves the linear least-squares problem of size $m$ to find the vector $y\in\mathbb{R}^{m}$ -which minimizes at best the residual norm (line~$18$). Finally, it computes an approximate -solution $x_m$ in the Krylov subspace spanned by $V_m$ (line~$19$). The GMRES algorithm is -stopped when the residual norm is sufficiently small ($\|r_m\|_2<\varepsilon$) and/or the -maximum number of iterations\index{Convergence!Maximum~number~of~iterations} ($maxiter$) -is reached. - - -%%--------------------------%% -%% SECTION 3 %% -%%--------------------------%% -\section{Parallel implementation on a GPU cluster} -\label{ch12:sec:03} -In this section, we present the parallel algorithms of both iterative CG\index{Iterative~method!CG} -and GMRES\index{Iterative~method!GMRES} methods for GPU clusters. The implementation is performed on -a GPU cluster composed of different computing nodes, such that each node is a CPU core managed by a -MPI process and equipped with a GPU card. The parallelization of these algorithms is carried out by -using the MPI communication routines between the GPU computing nodes\index{Computing~node} and the -CUDA programming environment inside each node. In what follows, the algorithms of the iterative methods -are called iterative solvers. - - -%%****************%% -%%****************%% -\subsection{Data partitioning} -\label{ch12:sec:03.01} -The parallel solving of the large sparse linear system~(\ref{ch12:eq:11}) requires a data partitioning -between the computing nodes of the GPU cluster. Let $p$ denotes the number of the computing nodes on the -GPU cluster. The partitioning operation consists in the decomposition of the vectors and matrices, involved -in the iterative solver, in $p$ portions. Indeed, this operation allows to assign to each computing node -$i$: -\begin{itemize} -\item a portion of size $\frac{n}{p}$ elements of each vector, -\item a sparse rectangular sub-matrix $A_i$ of size $(\frac{n}{p},n)$ and, -\item a square preconditioning sub-matrix $M_i$ of size $(\frac{n}{p},\frac{n}{p})$, -\end{itemize} -where $n$ is the size of the sparse linear system to be solved. In the first instance, we perform a naive -row-wise partitioning (decomposition row-by-row) on the data of the sparse linear systems to be solved. -Figure~\ref{ch12:fig:01} shows an example of a row-wise data partitioning between four computing nodes -of a sparse linear system (sparse matrix $A$, solution vector $x$ and right-hand side $b$) of size $16$ -unknown values. - -\begin{figure} -\centerline{\includegraphics[scale=0.35]{Chapters/chapter12/figures/partition}} -\caption{A data partitioning of the sparse matrix $A$, the solution vector $x$ and the right-hand side $b$ into four portions.} -\label{ch12:fig:01} -\end{figure} - - -%%****************%% -%%****************%% -\subsection{GPU computing} -\label{ch12:sec:03.02} -After the partitioning operation, all the data involved from this operation must be -transferred from the CPU memories to the GPU memories, in order to be processed by -GPUs. We use two functions of the CUBLAS\index{CUBLAS} library (CUDA Basic Linear -Algebra Subroutines), developed by Nvidia~\cite{ch12:ref6}: \verb+cublasAlloc()+ -for the memory allocations on GPUs and \verb+cublasSetVector()+ for the memory -copies from the CPUs to the GPUs. - -An efficient implementation of CG and GMRES solvers on a GPU cluster requires to -determine all parts of their codes that can be executed in parallel and, thus, take -advantage of the GPU acceleration. As many Krylov subspace methods, the CG and GMRES -methods are mainly based on arithmetic operations dealing with vectors or matrices: -sparse matrix-vector multiplications, scalar-vector multiplications, dot products, -Euclidean norms, AXPY operations ($y\leftarrow ax+y$ where $x$ and $y$ are vectors -and $a$ is a scalar) and so on. These vector operations are often easy to parallelize -and they are more efficient on parallel computers when they work on large vectors. -Therefore, all the vector operations used in CG and GMRES solvers must be executed -by the GPUs as kernels. - -We use the kernels of the CUBLAS library to compute some vector operations of CG and -GMRES solvers. The following kernels of CUBLAS (dealing with double floating point) -are used: \verb+cublasDdot()+ for the dot products, \verb+cublasDnrm2()+ for the -Euclidean norms and \verb+cublasDaxpy()+ for the AXPY operations. For the rest of -the data-parallel operations, we code their kernels in CUDA. In the CG solver, we -develop a kernel for the XPAY operation ($y\leftarrow x+ay$) used at line~$12$ in -Algorithm~\ref{ch12:alg:01}. In the GMRES solver, we program a kernel for the scalar-vector -multiplication (lines~$7$ and~$15$ in Algorithm~\ref{ch12:alg:02}), a kernel for -solving the least-squares problem and a kernel for the elements updates of the solution -vector $x$. - -The least-squares problem in the GMRES method is solved by performing a QR factorization -on the Hessenberg matrix\index{Hessenberg~matrix} $\bar{H}_m$ with plane rotations and, -then, solving the triangular system by backward substitutions to compute $y$. Consequently, -solving the least-squares problem on the GPU is not interesting. Indeed, the triangular -solves are not easy to parallelize and inefficient on GPUs. However, the least-squares -problem to solve in the GMRES method with restarts has, generally, a very small size $m$. -Therefore, we develop an inexpensive kernel which must be executed in sequential by a -single CUDA thread. - -The most important operation in CG\index{Iterative~method!CG} and GMRES\index{Iterative~method!GMRES} -methods is the sparse matrix-vector multiplication (SpMV)\index{SpMV~multiplication}, -because it is often an expensive operation in terms of execution time and memory space. -Moreover, it requires to take care of the storage format of the sparse matrix in the -memory. Indeed, the naive storage, row-by-row or column-by-column, of a sparse matrix -can cause a significant waste of memory space and execution time. In addition, the sparsity -nature of the matrix often leads to irregular memory accesses to read the matrix nonzero -values. So, the computation of the SpMV multiplication on GPUs can involve non coalesced -accesses to the global memory, which slows down even more its performances. One of the -most efficient compressed storage formats\index{Compressed~storage~format} of sparse -matrices on GPUs is HYB\index{Compressed~storage~format!HYB} format~\cite{ch12:ref7}. -It is a combination of ELLpack (ELL) and Coordinate (COO) formats. Indeed, it stores -a typical number of nonzero values per row in ELL\index{Compressed~storage~format!ELL} -format and remaining entries of exceptional rows in COO format. It combines the efficiency -of ELL due to the regularity of its memory accesses and the flexibility of COO\index{Compressed~storage~format!COO} -which is insensitive to the matrix structure. Consequently, we use the HYB kernel~\cite{ch12:ref8} -developed by Nvidia to implement the SpMV multiplication of CG and GMRES methods on GPUs. -Moreover, to avoid the non coalesced accesses to the high-latency global memory, we fill -the elements of the iterate vector $x$ in the cached texture memory. - - -%%****************%% -%%****************%% -\subsection{Data communications} -\label{ch12:sec:03.03} -All the computing nodes of the GPU cluster execute in parallel the same iterative solver -(Algorithm~\ref{ch12:alg:01} or Algorithm~\ref{ch12:alg:02}) adapted to GPUs, but on their -own portions of the sparse linear system\index{Sparse~linear~system}: $M^{-1}_iA_ix_i=M^{-1}_ib_i$, -$0\leq iC9qzzloDo)li|VmD5lNwUhms9TnBQMk&YjT@ zds?tbbTuP&2z2|wEU)z=imHN&bs|}P?BwI#`w)TQ-qRV z03b;3xM-lyCiXKcSZ$miLm#7tS{}WGH1R$r?)L(=*af%w5H6@Cp)= zILHMsH0QsP|JvB3G0Lel(tF9B(-`*)Fn3f)8VI4#XYlJdnQKD+-a56 z+46cizrW~+s7@*Fs@!K;P_uFmbM>TozRWk#(=`%Ic5Ms0LO~_7xX6MLMUj}rw^-u7 zjQ6I($ViISZR750q!!H-X{=PUNiz#2DP1f|vLQkgaH*|yt?y%JYb=1J)}`DDSNVbT zcQ_)Tl{|vy*8>C|o3775BBOeDFhqlB1~u8Isg5sy^UEE|xZ1k*b_*G7q4(Io=-QJ5 zo(N07)}eC$wHAdh?Q%!LdqmnbQ0mBW8C66M!q+sDZ(5=nqjrZSy2aA}{&n$+Y}3pX zn`<}*?baUkX}5|cYONxke%Y_%LBLusOfSnF2;rJy7q&8rp{_{rOmJ()6%vX}`bMPi zK}Bdnph;m9*@Yp}Nro7^4MC<(eCK3fq6r$BuBs@6cYj`LuaSo{GGm1Q;?=Up9U9%CPxkD3H(qdZ9C)B|3 zW+-08!yU0!{>|j|Hbvyf1OT1jcysRRij-a-Z_Jv$JRt%?v9%lGxq(Oye^GDR0S2$r z7!`(U{ny**!ti(a?S_oshTQ`LWVc5i7dj@|HyyuS{O&bO{U9wn6w9iJr?ep$&0iTZ zx>qp7&z5GT9hPs)`0cCY?HI}pzDyc-YgSnC%zhD+^k=qNk=dWN0$|X1hxUnrx^+=0 zZ<6?PMS7h;W^P;I!xW;Ui8;PNL$1f~nG5Wc`)*B`x1Gq@6St-02oxj$iWFU3I@hD( z^T(cus}>tUX*Oh$ijbailil#oE5j_S-YzaU#(j(Fq`mNgmT@H&#b|a=n}>`yErPu!QO1HlQ)n|sH}jU+#JnruNA)I63uy=PJ3~xzJP^=bD}FUBRQ7^P(?bp{VUb^L=B&c&}~$g ztI95uc&)2YEB<&18tB}^*t239cGrizy4Z%5G(a+_)2cIwn~ewO8f!svE+rrRG*7oR zjmOE4it&CC=611kC2?)jv%9Jx=QDlrCW+-(XDxq9Mnk)jB<$$W_e&GPY7)3vHq}Yq z?MXIGoCEsAKHL;J{gisUl$uW^)8-?=WtMFDTQvHj^O0(f1;>(MQF;Rj2aU$qW$RzK zx1gk@RakoasW~79cTtrYIV8D}-1k}WG8_cEN&A!Bs*Wj7tI{{*#y0+1RXB!nL4xDy zHf*l?*&;LV^3^+zveg)GPE*-o%NYjwcCs{s7*oPn5ci^g!}Q1Upqx&43?=`dNVmi*3jXJheECSn)=F4 zgvNRiw*0@cZ4mW_Ql^u>D0=jpZwt24hI@Y5j5?Z*he_JKp<}z!(`WTk@0!i0lkb8>>R{CUls@sB z)Dblb1#caL5N7DKhTdL5o21|Gun#G!6Z~z!CNv1iL|47UB=Vq{N^|1 zgrzQvgs|7jFy=2xCVigwrpzSBqboi}Q5e{_q>6f~uqkQ%uIiZaqpfh(r^Zy3v-p~V zo;zE0zlSq)B7AG{jk6MfZS8t8F-|Gu6*^|Rc-0B`E2fD2jA^jZpx0LIo#ssf#mJl7 zNX3_OvVIpoHlYia_DXvG?3F2%&RIj`Wp0=|qciMACsXVcg$mJ}b`I~qMpe`J45$Z6 zF`0!I{q@w5uSp%yvPx1+Fb6vv&QD`~W-qyBWQ!#BzOCM|Jb07`R!gq4>X6QT?G9&^ z<8zG}>_cekva}&D)UT}SGwzqv9#lnsaAbk1b<)35u?|~p0BTvt&W-|7t zI=X}oifo%P<(8ff!bS;MafqF_B3khzqisRqB>4z zQ~(&A&g*ZGrbXRjPqCUwNw8TZod2R_DC+FjmNK;zcZ`6=BT*HGgCCN?bcx2-g5nc` z5Ksp<4`O-LPiSVbET(ZOa*KdSn7#(jk7!0+$|6=&NGcP#g1o$<)W%+Dxc($ZhbvE_ zp&=T$U!IM4c6dhOW0+a7F@cn8Rhf^70AA#r)(u77oGF%Iyu>pO`1}OdJCM#pj;RX% zgp-sZvl6lA;$qzr10Zq>{X1Ha?cNO=(hZGg8)4~aFg-}A(-J7d0-_z*@|MpLw zKRj-uIdkX5dTPamKKdTdW1myw0fd)BXiLLA5s3Q431<5}TDs5oMWCN`Dfb8bc6Tdu zQ(2C`8pHJBRj4GKm154TE#!zL#1K|TW!5V$j(oSu&LbdV*(9GvoXC;OX-_F5p9==D35@LO!tbE5Fj~=%Tf!7D8z!7W2EA=w6 z_MV%Kx!|fd6o*XEaO2ZuOPPx8&%+zMhZ1|kAG!<0fb+{?0>^Znr~Rv0Md@BPe&;u* zUj%0R`wY+GcnY-()U+B<^-t|eFVjaQu|dPiLu~33Xjn?x-$Y+S+FSC*15VH%pfIm&V3 zuWRbLUOaWQIKSXLCj$`>t}*>}q3W&rLw{+` z?Px_#=kw2>cz8db&U8LVo9EgGr`yv5WH@{)+@Dux)9ugWp@L;Loi=Qp6ui=qk;BVoFsb!9nNe7f3ZAiEQ2=0pZjC7y3fW&dJ(?OT)#<+?hw##o1F* z(#6-13!X~I5P>S0n#=$raOd&h@v!sd@o+TvuuYv|a0Ub+U@ZZloCRzdTh%bmtO+G#$bJ~%DAk^$Xs(`Okx zrN#S0gbJl@b~(7Co(Z)1Dv**-G|oe+xOoasg|9!}oYOqq&)N_X5oKbJKGt^l#%NlU zkQyzeFhConb0~RKUI#6VrXFRywwUdbRwRpo036&K$;m%n=jy-ekfz;j64xQg&0^ZuQoum+GpQf2u<}X~iif$pxd?VoVE7`xpxBDulWPxR4;t_}}uKbJ>P*z}ao1#fr z+83loL5~m&VEH5$2;&^%;Sc_uQCr{pvuPr}I6>ommE|J!0joWiSKe8LNx6zzy;{8n z8&K7o#zwB>r(U-xW>4B$=w3pohE-L&$n-JFz4U5PL*W%%o%X=*;E6p!Xhu_sRo-?U z`^@kYR6LuG?#@D_%$ILl$3Lu3`%1O{po1`OsF*bS=Yq{whfXN`xTuYfT017b&-Q(P zSZ2*ud`q_^)bpbOiAbLIG5d|kHh!SrodoQ*_@-;EP07-zIOD1r7fcZN6dTMa_f6WN ze(y*|vn0%7ny4?Jdm_HO?v`FeR%1D=STcI}MeRb*#@jeP;fEVztP+zBkl@M!qL(8jGHT>>$DDU9C}Gx zeocbm9O3wd3e=@8tv27%Z1@#Umk6j&TXI7i%+cjw5yz?na=J;q0j+btye zR=wsTS28tC)HVN}jo$sNBnQiidEa-3&NuVg?az)eQ;aU()B|+DP7dEE!2%t_fMs)H zk5=DCM`%&9OhTW+h(;?I*Z-Bz1o2~#+(;dWC!{VPS$ES5>>H*TpXsC-D{)P9`h(;&D<%Vmm>OlwdGQ(_CZFU+QOBHvE(qm}_fa>bNR z6>n*m<`#1ud$u#r3kT9-a@b(8x_*1M9pd-A{r5oOX`YmDRSzZ3juU{&`l2c=K;F~D zy^Tn)h)VKjzRt|o)8OCmGTvQPhe(_whE%=Zn|ZS^8YbL6l!{rr2iD8&XOG+l7{FKc0JTlU64vu9X+pI_xZ0u)XApa?rE;zK{L-(v0F0pcc zcalja8tV+g?bCo>OqDu!TB~jDVye<}Z{GQ_k0}=JrhPt7YB+2^G7-9eMV=k#u0FaR z%%rCkl=Lzy1+mv^w+-;`hSpsqPt2@vDMz@fSEF8BqN)fK>t|K~tDTI0x=}J;>>=(B z?(FmR4Kr|uDer`$f6MYiiAH8-!~!Hc>9S4j&o}AfzEfdUit$+p{+4#oUyM+ zueJ(n8$LWt;Su_daZl#uM~0)oK`ROl?LrO#VvT(ksPu2AsIMk?4^{1z7Fb7aVrfNg zO4^ejikq5*YiRjnUWLjibP2E3gt}L9%x&43jD4OoktB!nUGp^zQ7!3G@laCo5^9A# zpq1EI5j1^32PR`Z@cqN#gj>(qLN5_zdkfZd44G8I1Y04NIr9A+-ulg(Bg=c+ zQ7U(nHM46NRMj;`tX3N5e8bGr#dNdBYWx-FZ}k%-Y&HwbPP82SH9f*x8%rBn=GZwa z*nbmM@M9g8CEkC2`0=pS>aZ(gmUWI-i1`q;ov1rCzu&>Kg%g&ny$#>1A z&Le}HA_RiCLQ;%sU%rrG{rr)U^qHtQ3nx(8d41WYiXMxc#24}i3{Z1|X$eVLSNq8JaB?Wug~8PfYdF$rcom@7 zvHSZRyIrX~-~8X8)pkCYtUJQg^wG9%9{lq5K@`YX9<=M6wKkP@4oeD}=Lk-66TV|b z>+kWoG_U6nuGY}X(BZDtg}vHcl2u$7>6J?jJfa&61cm(6KcM4Zy3iZ=uAx2MuNSavb<^_hlpABc z^ct|po1pmZ+K~6axt#Xo68gIytwWa8XO_uZpW;+UUY7JO+k6>S3o8EcJ;DixI;!=k z*jojOuTZ#CTtSU$ATaa(N$e0>7E!jX@#pC1`y|*%(mTrO4g6gtH4`n&Nu~u|L=4DW zfbpkFjgX?y6W39ZJ7$xcFAlbY(*0h4diDTt)l=chtcKum#DPot_h z;B%xE620izrtJ{7G*5B7WKI+L#Ss;Wv3{E#M4_o{tt{VPThp{vB;Nmgwx+e3vv`9T zQylX*n$FB{_tqa!%{)P##ZDqh@Oq(4H_54T4N=T;MiY`9K@)01*p1qWC+Qu@N(;OV z2G=|@*O;$A;piAFqG|d0Oj*v|C-7wdZ7LYMKqJ~xRfff1f(R4u#7S!X!))ENgA|bX zD;f-KqM6HYAp&O2i+^&QXy5yPc|6|EV-vmRlJ=Mv-&uWfC1?_}yuLuw{jKggw!Ei5 z=1$m34P+2wG`C+!LVw-QbJo|}>mv$;es#u3VE5mwd*b61;R_Kd|6UlZk(-yCmqShJ zAxn&+I?Q_xF=msSxmO}X&2s|$?b130VM|x^Ru5TWhh6Hv!n;pX#&X+Id=p@69E!St zRQz*&Snd~vCovFAqaJGoWeD++>3n>T)6F$#bQB?Cim z^1F8u?MdiYQsQ+TW7k&H*Pz|pdD6kMx)CjH+5YX~Rf- zT-5$Z(|}Dq-;eVPk}w>1f=2y`uFdS3n6@Dv4(Nc!IDY!!ko0ditI}QDAg>V<3(P+2 z9D9pLC8?(Vm6qwFmm0{H>3g8liz0L-$t2sAUaqLgx?L&%v#8Uz+6+ZkGQz6c6rqvu zpS3hNRKm6o*6=GO-E4y10%dd`j)>9xT{!1K5>g~ zI#JQn@*#fAiJ_A1p#^ye{1DG^A%3KPC2%k{V@gOYD<~hoV=bvdnFnw_+vSj#(0ZC1 z%UR^{n|B_8F{hc@yM|dhhEJGr!0A-MH1GyJSt%(ihclhOyv+y3=iO4l&zi*C06KX; z2IpNEM_z0gM86WC`sjq#+se!<9WjX_g3zGnVWCf*W-QXvf5HRZLkk(*PM5V*Tu}S= zWE@{$EJxda*W|mt9(+K>((ZC0)EXW4%jX{}nuN{m#^7)$|yE9y_Lixpk;A zH$EqIJFbfdbh}8|_C*xgg6oqp z@MzP&Zd4Yxl{pk+`LRfB9DKSI#P)DmgJF$pVi6=|2Lw8N&-9_p|KNi5Fnc+l+YAf7 zl|o%wc%4{q>Ii9R_I{T4d#xnpgVB6fea_7`4iu2E=`x^S6p(Z|O`QB)%#(+lgfhM&*hA^QHZKgvfT`^O^U1 z{Y5pTwi__GB!(Zc#d(;Q#Nrjg^2dq##Y1D^Wuud1rNJ((A$(djX&>h))kLv4{%myY zFH!C>lpm2G5C7Al`OP4Nu4ehVwxHH>fcZgh@Yzym2AF-7;GSF<=XWZ`Lgqvu$kMHH zvdsz6U-FvFgUaoVG~vDu|3K_rt>{1DZA%^ir!Vg(U^ zTY)WYxS)a-yj&27wKW&`ufk=`XKe|F!(f6C8^P3Uu+x7AazZ3P0$dQd76dK?5fp;) zzW_s110XN{DYikP|1(PqsO0niBuF-%R7NQMf8^PW&pU$3EdYW4RR#YQBO4gZ#(~SD z>|!Nt?rF^~Ed&Kac)|aZNPU2Li9#9wJ{vHjGb8T5IuKqy$iHh~UO@=&zj0t*ett-5 z10Ndz`Ns&HSK!|@7#IfshX()WN-#b!-@kJLf&%}T%L^9}{4dVr|Ah{Qfx-VR{(p@R z3i*ejAQb$M?Z9A=f1eA1g5iRxE&L)T{8BQqvS2Amu!N+btSrAYKOY|$A|MEZ^GZv} n2*M#^BL6F4j4LLF>)~nc?&<4pZG#Kp<%h#@Sy^RNWpV!xLj4sE delta 8106 zcma)=Wl$W?mAcq?lwRmc!1y@EV#S7CAeE~2pTLfIGgw0 zt=+Bv`(bZaS6A1spFVZE>zuAr9p*IQ7;gzPt6umLV$71_O7kMB%}}oQb6h|Nge1j< zVjB+e@R!j^&>qjpC9WLHzc79AdvNx?6lSj$rR-2#Sgm<{Q0q3$5fdUcY>4u?00|O} zUm_6dYQ?;36(hB&_i1gLyF4e2?>ED0@VcdJ^;cqKR+qT2L5oeWjH)T)qBQ?H&!U4nEC_c)}O-b!&0 zyH~5hQ^w!>1(5pYEeUaSwDpOFa8E7Fr*=6wyP>07=<#+ikNM0Ev(08Pf)Y)04ui?5 z3J&1ND=RGFuooN!6Q*a;Q2zWh>Y>bC%~1=K5hceK6-Nk56L~YxiZ?ZLUxi-mF;a38Kd4;6cz|`;j&bT4JKg|hqs{Qw>j__-fesqa_kZrkE5%O7u)(ZY{rZ!RK zGPYvowuu8#l|!cFLUi-Ss;>FF+=y(hv)j?3&AaS7EG6Tf;vfQv#k~^B8ql4Evi@zq zeO&%&5ErAj}$X!wRR4UM;5^|4H)9Synmxe+S zZV&0uxfmJIvBv^!%pD!l)!XK>zBB@ywV2Q+BhY>?5U(_oO}^^x!FjDc;!*sOF!*OU zGdu*zMSs>`K0A*Ax19tw;1`np{f-FNE->KsisZ*?zbTC*kwd``Q}Za3&s1pVa#a5L zv*Gr{8h1S^IQ0Fe#PaYO%8eX#fr}Tg_qBrpO&6Qu42Lo--4N%$~x6wmHM>dvP z@-n{+7u0h?Jc>t)@0O+BLTytjQEJ~>V$&XyCt*1M2ZV-x!+@~p<~)x%hy^u`KyqWN zrS=RdCXKV14DEy6=>5$}o__0p(!^+oHy*$eX~JM(o)&WOmM)981sZ3$x4lu znUksB5^lyKahTCcVt)KF9M5Dnx@W}-I%DaZz$*RuXuwW0wi=RE%~uJ^ePsE=Nywd3W-a9F7s42uhv7SNOzV2*kxq0qO* zJGchGbOy4c!}d8}EeH(Fh2cTnY2O_#DKOp1>ZrQY5@y=8$8XXS*A)0VkVvw4H(|Dp zq1YWE>fD#S83;FuOT3!)L|sgPU&Yx;GDj_b=(T>!!9zuMs4tV6POWX>fFhcZ^0WjWJr1MLX(Kx>Z)?*x{y@Vc+uaCm6 zbA8+h!fZvY%WE<_e`2X-;%esd$oNoMyqhbhK-vr*S{ymX+i}oyW9SCPvBnN@4jJ#X zyrM1tQcP=T%DjXyoMLReaU04f*4GwRR**v$vnX3KwUdGN5J4tmG~oVEv#m3~j4{fv};Z&G{p`0m2h{>m-*V35}2$NL(~t-9P$r4oNc? znht6Yd{_Nl{wxg`6UremP)N+gljcFna7nbt|gNdy{O$* zC6cB+u@v9q%QH3NS$Q;24MCA@!fl0u2%PWrhPJ|HTGYra?S&$^S8%?$pA;tg)0Z=G ziWz9-bU*eWOD)H>JRZPCy`C?3=w8!I1xf719SOxWA*z14N(JgVQ~9vIq;(abUMyR0 z5&TF{&*{8v8T*q){yK#kP{vS3@y%JrE-TolmItsX#1l1GOpvd*vQl=FcQi!hOt_(j zzC`o)sgoRp#|z3u#>>=2Y^?aYUr9i|+lr1Tn1}xCD!WU&hY`TwwxZvTn=;Q%GlbuA zyH_wq)rcqIqirf`1j^@V<@cY6Cl*T8aGCpEM{T?=- z+`lq2+VZtB#t-*W`oLmifduGhs$zzhJ?wyb!_%B?yp*4!8-8LgH&}e>{Aq_3NjS75 z5<1#A-7DuxZ2ulcx;>O?=j}IwBdYa$Ba-A0+ZlaD+OV@zC+hAU+f`HJR!erg@<2N8 zbu}Py-iCzn?5VkSGjNwgSz+V(sSS6g^MU6RwpyKGuHD7NZRzPP$RaNJ$j zduf`VU*h7>Vfbj1kXPr{!p-6F)Q>{h@@!k!aeb<$pVnMbX(e39Dn~%Dqe?u)iS-Jj zRWo%s2uigCr1lhd_AVbWbercXFa0HdZ(hQ7L*ga?xnG4(KGX zXqWnQmq->R7g+9d@+^WImt ziT8f4w*1Y^04WxSZtoC42%k=s%VR6;yMLJb71pepg^YKJjMCR)6k&{`|KxWsdBWHp z!sAnE^aos5BBj{j_EM)2(6G3F@07|AUng7L7MOVBD<;vd;IBnu+aE5Qo26$OAg*V z=Ok71`Id-*Mp~SQ-qGH5mSZzu+zLDb-o)$t`O0T~e9{vq?g=#Vpj%V;E#NoI5tzbh zN(svm4##6IOWIB$5u;w=ydJX|w~DQpBU?%KDn>vxOpn))@Q=lG&Vlw8GZ-yPkf* zIh}Gs^7an-HfUE9e<6^4_CT=1w^FidmWTaNefjfPK1jnyIkQZ^b0#z&_?-DXdclO7 z+f+po|IrTYjrd-2-u|SXGqZUrUNVQR$waCBey`rC>zLSl-nor6H5|Dx%KnGrE40K& zu{Q&f^pN@T<%Zjj1NSR+uDx$`oT@d>&WGhz*!aL(Bb!35$RJ!pgo?7S3uHuH8+b5o zf5^&#h7ZgSUd`9+^`rl1e>?m|g2nPYb_AS+1evlr@2`6ZOVo`iC-*q|%T{vtKe zej!H%5_9>&IKRdzIzI*}O$kgd{p2-q%>-E0x+>SYx-F0|>Gdq#53JHEcS`DIf&737 zjW(vyM$wwUJ{vt;>;2Qkb~9Y6IuX0aA@!n*F*n!47!Pizzl4^sy2ePki?>|{1O1>QcN{un&65qo_47z77Azi_621YHuDaVQZ5>Sb{3 zkYCBFNodP!CpO?aBXT7&5rF=Vh6G_q5U%=70z53L|6XR*viksaW7lx{VD88%>*VMz zDe2^G$d!l#3{6Y|@?bb~x^lX4y4iUr&H)`^e7qo0Q8#y2sQCwgch(18AFag2rWXtO ziDJ4Bk*yI@eX~tXI1!b^4vhwgl|=Y>@d{*MB<@g9=sPPbK|C51bi|Oc0<3%*!lEet zInnY)gVWFo!_~6RfTGjmJMVD#f>15I3O;;nps#=DIwAUS(g}NLh5@5p^~7uS*oI-7 z1BB}PLo_am4rNb1$n_zE6-GT37E)P)@YOYyiMQXYgF8_K>+%ZvOOrV{JY`?JF+940 z{%~qI5G&c+UlSMaSu+eVS+#Z&6S&`AUmx>5ZirP{L!wdW+B!GaVG|LYVEDR`NU1MXOF)1h zu^8psR3uBNj`baML3ehCyrDIVd3ces8z9MsG^NUj+3Y0fh>E}G198rhZQWG(v+F1K1^NI7bfMhOM65}B z)T||Y&$)9sxz`8+3s#05tgI|J^k5!}3(|$Wbpx9-6SbIl7$bGx~8q-&0 z!5(4q_?0YJCbT7DuaET(yY_ZUNVSnCcb`#d8bKWV*a;unWzt3K`Y6^r%!2!o^Q<#g z|H+%E(1j1K56R$A2PzJzveP+7j}GS#&zD zbL@sPHWPOq4J(&x7gRu_2IK4F{1KTo^;gXW`T52{I?f7^48rToLmfw2gZ}ih%}d3? z&79u+(5^rJ2ti$ngN_8*aYXiKY$(TfAfKg4Z8K^4FG{_5O6BVnL5HQzzcq zK}nl)+OtWEBQ1Po8Cjg`;;X+~&UwG8PO}Fh8jPyDMMpKMtYih1@-ocm}a34qV#L~k~$ReJL#r=gi{u$o#t8l2$HM$ghqVa&*DREj0*q{Y6; zTtk@2`*aTa1(L!|^}yH+@aru!D1Cir!eC*dmRVlNs;2F{ILjQN9ROUtl>x1$sn2Vw zAA>QA#&)qGVuZukGnrc` zWFOp|W$*Bom4lNPBYJEwc}lt&4#tjcF8+l6i0COsGdmbj=av+C&M=O7MG@I0>UuLb z;DI>l+R`E6>N#lVUG}$DVcSyPxHN~kcIkXcEhT)<7_WktOHG`Dua$KwrzCsALNnKL zYq@Z0mvrUDyTaf6G1rv5Tz*!j)JW5VQlNyH<>MQ9-1NrUzB|GA$Aev6+hsO`poOm_ zo}&84KvG}=J-P=!&w|*+c^}S6=yJkCGnMGd+S{60rz5ai_O7nOHA!_cS5(Ow>zDM zVqb2Gge-D_Wyils8k(0@+N)+%UK)3EBP?UNu zf#u)uMDLlP_8K$v99|XaO(eNGhbBEJcQNAr{`PAT@wQIm2ukDW!+Y)sORpe%mgISl z_%;)`Gv|7Wni1qNmXD+Jfb~7yx*!+q-{2XdiRKy|ZAM$T<1t?#x&MZCKwdw@x_e9; z=dIFPGsZ%+u?&9N8-)Cvh9k>RmLQ@1&3W&`CwAN3U9+7fyRE<(P8_xB$=MP&H*Rog z{0-n4&>6(P1Mj-E6@IG)llL|FS<)zX>=1V|jxQ8S2FbO1b#5>pG_GN&ZOCiuR4&Xy z%2vT%V^5L1@9iBl6xS=9o8Wq`!s11`aP%qv;0`&6&EsP1;^ zMMA6ybt+PrX5a?Zi%|N_mMx>Jlw6=n0HMF~f!At1OhlovpP6#}RJDz8x)XnUkW0Lg z=~oS}R7yf}a=bGb#z#mUEJs1mh8vgj7-mbrzj`s-3c4RV_nw8yzX+f%-h5rcS*R|I z;X7f!K%(7XF@$fH8YtU;{L0ZOtBHGPZ7Agib6@R6&a;!72|Ju_yWRFW_@J<{X=v19 znEmTRsuMA59PF1Vg6w(7w#;#>5!&yl)JzcT(BrZK!+3#I5pfI%>D^(PU-BBB?Gfb(9A~%~A7;qU^mt z$p#>#=;)-ss~qpY4v43=Q=}=9M?8-=ptADHfuc$LKzNOima7LGWrX!R; zXT&F)1E!|R8(B%y2`b`L_)AKcppn)q|_@T#qvM5Rg2IMlsxmeVH z0~W=m&(%+DV%d`T`(^1pCx^m5VEE=M!VQGP$=wbIz$cMrT9E+85w~G{uSvd8Z(N++UC@uy;xD3)&lx;z z4@C!6m}k5z5nX$xVdFCFD-_E%A)C}IJO>})Gp_I0nNV1;xf8R05h*J4Ip1WXnKorp z!CrsP{*W2cyeV_1XYZww*zjgWmf^=~?Lx59kM4D`Ru*A{6hcmU1rilt1u+`t4a zO*>Fw(&2nECH`k{jKv!tXAurpHl}8%Cb7kcwGf%=HYbrV7lhpllcCVDMR~5aLGE$g zDH*fmPRQ~!Rp?)@M1)YZ8b?#J|1NWP_n+Hr3C({d@k=h=uIJ+PM`Yl(w1NuoK)Kk> zEkG9RAQ04wo$GI7=eHE#=C$B4w=joTBu;UC_@4>QK@xb}>|6p`+}wiPAVCm_l?#+u z2V(s$ZY|C)e!z!1LwU|c->Al`p5US4qGF9;J1{I3-O zFxP)FFc*aTUrYeR^DhSB;{Q*^$It&SdprXCp#RpH{2$l3z+9mJ4gW8LaR1wNevrVw zjtB;G|L0y1Hy4;E@t9ZGMB?w2%1TRf3-IwtL1blRK?1yB83_qquz-Y&EUy%|sPO++ fVGIxz1-QAJySjV3LahNH9$qjQz{n`0Dhv2uH5}kZ diff --git a/BookGPU/Chapters/chapter12/figures/compress.eps b/BookGPU/Chapters/chapter12/figures/compress.eps index 8d4d84e..62a4f5b 100644 --- a/BookGPU/Chapters/chapter12/figures/compress.eps +++ b/BookGPU/Chapters/chapter12/figures/compress.eps @@ -1,8 +1,8 @@ %!PS-Adobe-2.0 EPSF-2.0 %%Title: compress.fig %%Creator: fig2dev Version 3.2 Patchlevel 5c -%%CreationDate: Fri Feb 8 16:15:28 2013 -%%BoundingBox: 0 0 964 614 +%%CreationDate: Thu Jul 18 18:50:59 2013 +%%BoundingBox: 0 0 961 614 %Magnification: 1.0000 %%EndComments %%BeginProlog @@ -77,52 +77,12 @@ end bind def /shd {dup dup currentrgbcolor 4 -2 roll mul 4 -2 roll mul 4 -2 roll mul srgb} bind def -/reencdict 12 dict def /ReEncode { reencdict begin -/newcodesandnames exch def /newfontname exch def /basefontname exch def -/basefontdict basefontname findfont def /newfont basefontdict maxlength dict def -basefontdict { exch dup /FID ne { dup /Encoding eq -{ exch dup length array copy newfont 3 1 roll put } -{ exch newfont 3 1 roll put } ifelse } { pop pop } ifelse } forall -newfont /FontName newfontname put newcodesandnames aload pop -128 1 255 { newfont /Encoding get exch /.notdef put } for -newcodesandnames length 2 idiv { newfont /Encoding get 3 1 roll put } repeat -newfontname newfont definefont pop end } def -/isovec [ -8#055 /minus 8#200 /grave 8#201 /acute 8#202 /circumflex 8#203 /tilde -8#204 /macron 8#205 /breve 8#206 /dotaccent 8#207 /dieresis -8#210 /ring 8#211 /cedilla 8#212 /hungarumlaut 8#213 /ogonek 8#214 /caron -8#220 /dotlessi 8#230 /oe 8#231 /OE -8#240 /space 8#241 /exclamdown 8#242 /cent 8#243 /sterling -8#244 /currency 8#245 /yen 8#246 /brokenbar 8#247 /section 8#250 /dieresis -8#251 /copyright 8#252 /ordfeminine 8#253 /guillemotleft 8#254 /logicalnot -8#255 /hyphen 8#256 /registered 8#257 /macron 8#260 /degree 8#261 /plusminus -8#262 /twosuperior 8#263 /threesuperior 8#264 /acute 8#265 /mu 8#266 /paragraph -8#267 /periodcentered 8#270 /cedilla 8#271 /onesuperior 8#272 /ordmasculine -8#273 /guillemotright 8#274 /onequarter 8#275 /onehalf -8#276 /threequarters 8#277 /questiondown 8#300 /Agrave 8#301 /Aacute -8#302 /Acircumflex 8#303 /Atilde 8#304 /Adieresis 8#305 /Aring -8#306 /AE 8#307 /Ccedilla 8#310 /Egrave 8#311 /Eacute -8#312 /Ecircumflex 8#313 /Edieresis 8#314 /Igrave 8#315 /Iacute -8#316 /Icircumflex 8#317 /Idieresis 8#320 /Eth 8#321 /Ntilde 8#322 /Ograve -8#323 /Oacute 8#324 /Ocircumflex 8#325 /Otilde 8#326 /Odieresis 8#327 /multiply -8#330 /Oslash 8#331 /Ugrave 8#332 /Uacute 8#333 /Ucircumflex -8#334 /Udieresis 8#335 /Yacute 8#336 /Thorn 8#337 /germandbls 8#340 /agrave -8#341 /aacute 8#342 /acircumflex 8#343 /atilde 8#344 /adieresis 8#345 /aring -8#346 /ae 8#347 /ccedilla 8#350 /egrave 8#351 /eacute -8#352 /ecircumflex 8#353 /edieresis 8#354 /igrave 8#355 /iacute -8#356 /icircumflex 8#357 /idieresis 8#360 /eth 8#361 /ntilde 8#362 /ograve -8#363 /oacute 8#364 /ocircumflex 8#365 /otilde 8#366 /odieresis 8#367 /divide -8#370 /oslash 8#371 /ugrave 8#372 /uacute 8#373 /ucircumflex -8#374 /udieresis 8#375 /yacute 8#376 /thorn 8#377 /ydieresis] def -/Times-Roman /Times-Roman-iso isovec ReEncode -/AvantGarde-Book /AvantGarde-Book-iso isovec ReEncode -/Times-Italic /Times-Italic-iso isovec ReEncode /$F2psBegin {$F2psDict begin /$F2psEnteredState save def} def /$F2psEnd {$F2psEnteredState restore end} def /pageheader { save -newpath 0 614 moveto 0 0 lineto 964 0 lineto 964 614 lineto closepath clip newpath +newpath 0 614 moveto 0 0 lineto 961 0 lineto 961 614 lineto closepath clip newpath 171.0 1185.8 translate 1 -1 scale $F2psBegin @@ -299,339 +259,339 @@ n 7065 13455 m % arrowhead 15.000 slw n 9280 12372 m 9386 12250 l 9225 12265 l col0 s -/AvantGarde-Book-iso ff 396.88 scf sf +/AvantGarde-Book ff 396.88 scf sf 2790 10755 m gs 1 -1 sc (X) col0 sh gr -/AvantGarde-Book-iso ff 396.88 scf sf +/AvantGarde-Book ff 396.88 scf sf 4140 9855 m gs 1 -1 sc (X) col0 sh gr -/AvantGarde-Book-iso ff 396.88 scf sf +/AvantGarde-Book ff 396.88 scf sf 4140 10755 m gs 1 -1 sc (X) col0 sh gr -/AvantGarde-Book-iso ff 396.88 scf sf +/AvantGarde-Book ff 396.88 scf sf 4590 10305 m gs 1 -1 sc (X) col0 sh gr -/AvantGarde-Book-iso ff 396.88 scf sf +/AvantGarde-Book ff 396.88 scf sf 5040 9855 m gs 1 -1 sc (X) col0 sh gr -/AvantGarde-Book-iso ff 396.88 scf sf +/AvantGarde-Book ff 396.88 scf sf 5040 10755 m gs 1 -1 sc (X) col0 sh gr -/AvantGarde-Book-iso ff 396.88 scf sf +/AvantGarde-Book ff 396.88 scf sf 5490 10305 m gs 1 -1 sc (X) col0 sh gr -/AvantGarde-Book-iso ff 396.88 scf sf +/AvantGarde-Book ff 396.88 scf sf 5490 11205 m gs 1 -1 sc (X) col0 sh gr -/AvantGarde-Book-iso ff 396.88 scf sf +/AvantGarde-Book ff 396.88 scf sf 5940 11205 m gs 1 -1 sc (X) col0 sh gr -/AvantGarde-Book-iso ff 396.88 scf sf +/AvantGarde-Book ff 396.88 scf sf 6390 9855 m gs 1 -1 sc (X) col0 sh gr -/AvantGarde-Book-iso ff 396.88 scf sf +/AvantGarde-Book ff 396.88 scf sf 6390 10305 m gs 1 -1 sc (X) col0 sh gr -/AvantGarde-Book-iso ff 396.88 scf sf +/AvantGarde-Book ff 396.88 scf sf 6390 10755 m gs 1 -1 sc (X) col0 sh gr -/AvantGarde-Book-iso ff 396.88 scf sf +/AvantGarde-Book ff 396.88 scf sf 8235 10305 m gs 1 -1 sc (X) col0 sh gr -/AvantGarde-Book-iso ff 396.88 scf sf +/AvantGarde-Book ff 396.88 scf sf 8235 11205 m gs 1 -1 sc (X) col0 sh gr -/AvantGarde-Book-iso ff 396.88 scf sf +/AvantGarde-Book ff 396.88 scf sf 8685 10305 m gs 1 -1 sc (X) col0 sh gr -/Times-Roman-iso ff 381.00 scf sf +/Times-Roman ff 381.00 scf sf 4905 14310 m gs 1 -1 sc (1) col0 sh gr -/Times-Roman-iso ff 381.00 scf sf +/Times-Roman ff 381.00 scf sf 5355 14310 m gs 1 -1 sc (8) col0 sh gr -/Times-Roman-iso ff 381.00 scf sf +/Times-Roman ff 381.00 scf sf 5805 14310 m gs 1 -1 sc (9) col0 sh gr -/Times-Roman-iso ff 381.00 scf sf +/Times-Roman ff 381.00 scf sf 6120 14310 m gs 1 -1 sc (13) col0 sh gr -/Times-Roman-iso ff 381.00 scf sf +/Times-Roman ff 381.00 scf sf 6615 14310 m gs 1 -1 sc (14) col0 sh gr -/AvantGarde-Book-iso ff 396.88 scf sf +/AvantGarde-Book ff 396.88 scf sf 4860 13905 m gs 1 -1 sc (X) col0 sh gr -/AvantGarde-Book-iso ff 396.88 scf sf +/AvantGarde-Book ff 396.88 scf sf 5310 13905 m gs 1 -1 sc (X) col0 sh gr -/AvantGarde-Book-iso ff 396.88 scf sf +/AvantGarde-Book ff 396.88 scf sf 5760 13905 m gs 1 -1 sc (X) col0 sh gr -/AvantGarde-Book-iso ff 396.88 scf sf +/AvantGarde-Book ff 396.88 scf sf 6210 13905 m gs 1 -1 sc (X) col0 sh gr -/AvantGarde-Book-iso ff 396.88 scf sf +/AvantGarde-Book ff 396.88 scf sf 6705 13905 m gs 1 -1 sc (X) col0 sh gr -/AvantGarde-Book-iso ff 396.88 scf sf +/AvantGarde-Book ff 396.88 scf sf 3015 16155 m gs 1 -1 sc (X) col0 sh gr -/Times-Roman-iso ff 381.00 scf sf +/Times-Roman ff 381.00 scf sf 2385 17910 m gs 1 -1 sc (0) col0 sh gr -/Times-Roman-iso ff 381.00 scf sf +/Times-Roman ff 381.00 scf sf 2835 17910 m gs 1 -1 sc (1) col0 sh gr -/Times-Roman-iso ff 381.00 scf sf +/Times-Roman ff 381.00 scf sf 3285 17910 m gs 1 -1 sc (2) col0 sh gr -/Times-Roman-iso ff 381.00 scf sf +/Times-Roman ff 381.00 scf sf 3735 17910 m gs 1 -1 sc (3) col0 sh gr -/AvantGarde-Book-iso ff 396.88 scf sf +/AvantGarde-Book ff 396.88 scf sf 2340 18405 m gs 1 -1 sc (X) col0 sh gr -/AvantGarde-Book-iso ff 396.88 scf sf +/AvantGarde-Book ff 396.88 scf sf 3240 18405 m gs 1 -1 sc (X) col0 sh gr -/AvantGarde-Book-iso ff 396.88 scf sf +/AvantGarde-Book ff 396.88 scf sf 2790 18405 m gs 1 -1 sc (X) col0 sh gr -/AvantGarde-Book-iso ff 396.88 scf sf +/AvantGarde-Book ff 396.88 scf sf 3735 18405 m gs 1 -1 sc (X) col0 sh gr -/Times-Roman-iso ff 381.00 scf sf +/Times-Roman ff 381.00 scf sf 5040 17910 m gs 1 -1 sc (8) col0 sh gr -/Times-Roman-iso ff 381.00 scf sf +/Times-Roman ff 381.00 scf sf 5535 17910 m gs 1 -1 sc (9) col0 sh gr -/Times-Roman-iso ff 381.00 scf sf +/Times-Roman ff 381.00 scf sf 5895 17910 m gs 1 -1 sc (10) col0 sh gr -/Times-Roman-iso ff 381.00 scf sf +/Times-Roman ff 381.00 scf sf 6390 17910 m gs 1 -1 sc (11) col0 sh gr -/AvantGarde-Book-iso ff 396.88 scf sf +/AvantGarde-Book ff 396.88 scf sf 5040 18405 m gs 1 -1 sc (X) col0 sh gr -/AvantGarde-Book-iso ff 396.88 scf sf +/AvantGarde-Book ff 396.88 scf sf 5490 18405 m gs 1 -1 sc (X) col0 sh gr -/AvantGarde-Book-iso ff 396.88 scf sf +/AvantGarde-Book ff 396.88 scf sf 5985 18405 m gs 1 -1 sc (X) col0 sh gr -/AvantGarde-Book-iso ff 396.88 scf sf +/AvantGarde-Book ff 396.88 scf sf 6435 18405 m gs 1 -1 sc (X) col0 sh gr -/Times-Roman-iso ff 381.00 scf sf +/Times-Roman ff 381.00 scf sf 7650 17910 m gs 1 -1 sc (12) col0 sh gr -/Times-Roman-iso ff 381.00 scf sf +/Times-Roman ff 381.00 scf sf 8145 17910 m gs 1 -1 sc (13) col0 sh gr -/Times-Roman-iso ff 381.00 scf sf +/Times-Roman ff 381.00 scf sf 8595 17910 m gs 1 -1 sc (14) col0 sh gr -/Times-Roman-iso ff 381.00 scf sf +/Times-Roman ff 381.00 scf sf 9045 17910 m gs 1 -1 sc (15) col0 sh gr -/AvantGarde-Book-iso ff 396.88 scf sf +/AvantGarde-Book ff 396.88 scf sf 8235 18405 m gs 1 -1 sc (X) col0 sh gr -/AvantGarde-Book-iso ff 396.88 scf sf +/AvantGarde-Book ff 396.88 scf sf 7785 18405 m gs 1 -1 sc (X) col0 sh gr -/AvantGarde-Book-iso ff 396.88 scf sf +/AvantGarde-Book ff 396.88 scf sf 8685 18405 m gs 1 -1 sc (X) col0 sh gr -/AvantGarde-Book-iso ff 396.88 scf sf +/AvantGarde-Book ff 396.88 scf sf 9090 18405 m gs 1 -1 sc (X) col0 sh gr -/Times-Roman-iso ff 381.00 scf sf +/Times-Roman ff 381.00 scf sf 2385 11610 m gs 1 -1 sc (0) col0 sh gr -/Times-Roman-iso ff 381.00 scf sf +/Times-Roman ff 381.00 scf sf 6435 11610 m gs 1 -1 sc (9) col0 sh gr -/Times-Roman-iso ff 381.00 scf sf +/Times-Roman ff 381.00 scf sf 5985 11610 m gs 1 -1 sc (8) col0 sh gr -/Times-Roman-iso ff 381.00 scf sf +/Times-Roman ff 381.00 scf sf 5535 11610 m gs 1 -1 sc (7) col0 sh gr -/Times-Roman-iso ff 381.00 scf sf +/Times-Roman ff 381.00 scf sf 5085 11610 m gs 1 -1 sc (6) col0 sh gr -/Times-Roman-iso ff 381.00 scf sf +/Times-Roman ff 381.00 scf sf 4635 11610 m gs 1 -1 sc (5) col0 sh gr -/Times-Roman-iso ff 381.00 scf sf +/Times-Roman ff 381.00 scf sf 4185 11610 m gs 1 -1 sc (4) col0 sh gr -/Times-Roman-iso ff 381.00 scf sf +/Times-Roman ff 381.00 scf sf 3735 11610 m gs 1 -1 sc (3) col0 sh gr -/Times-Roman-iso ff 381.00 scf sf +/Times-Roman ff 381.00 scf sf 3285 11610 m gs 1 -1 sc (2) col0 sh gr -/Times-Roman-iso ff 381.00 scf sf +/Times-Roman ff 381.00 scf sf 2790 11610 m gs 1 -1 sc (1) col0 sh gr -/Times-Roman-iso ff 381.00 scf sf +/Times-Roman ff 381.00 scf sf 9045 11610 m gs 1 -1 sc (15) col0 sh gr -/Times-Roman-iso ff 381.00 scf sf +/Times-Roman ff 381.00 scf sf 8595 11610 m gs 1 -1 sc (14) col0 sh gr -/Times-Roman-iso ff 381.00 scf sf +/Times-Roman ff 381.00 scf sf 8145 11610 m gs 1 -1 sc (13) col0 sh gr -/Times-Roman-iso ff 381.00 scf sf +/Times-Roman ff 381.00 scf sf 7695 11610 m gs 1 -1 sc (12) col0 sh gr -/Times-Roman-iso ff 381.00 scf sf +/Times-Roman ff 381.00 scf sf 7290 11610 m gs 1 -1 sc (11) col0 sh gr -/Times-Roman-iso ff 381.00 scf sf +/Times-Roman ff 381.00 scf sf 3060 16560 m gs 1 -1 sc (1) col0 sh gr -/Times-Roman-iso ff 381.00 scf sf +/Times-Roman ff 381.00 scf sf 5535 16560 m gs 1 -1 sc (8) col0 sh gr -/Times-Roman-iso ff 381.00 scf sf +/Times-Roman ff 381.00 scf sf 5985 16560 m gs 1 -1 sc (9) col0 sh gr -/Times-Roman-iso ff 381.00 scf sf +/Times-Roman ff 381.00 scf sf 8100 16560 m gs 1 -1 sc (13) col0 sh gr -/Times-Roman-iso ff 381.00 scf sf +/Times-Roman ff 381.00 scf sf 8595 16560 m gs 1 -1 sc (14) col0 sh gr -/Times-Roman-iso ff 381.00 scf sf +/Times-Roman ff 381.00 scf sf 6750 11610 m gs 1 -1 sc (10) col0 sh gr -/AvantGarde-Book-iso ff 396.88 scf sf +/AvantGarde-Book ff 396.88 scf sf 5985 16155 m gs 1 -1 sc (X) col0 sh gr -/AvantGarde-Book-iso ff 396.88 scf sf +/AvantGarde-Book ff 396.88 scf sf 8190 16155 m gs 1 -1 sc (X) col0 sh gr -/AvantGarde-Book-iso ff 396.88 scf sf +/AvantGarde-Book ff 396.88 scf sf 8685 16155 m gs 1 -1 sc (X) col0 sh gr -/AvantGarde-Book-iso ff 396.88 scf sf +/AvantGarde-Book ff 396.88 scf sf 5490 16155 m gs 1 -1 sc (X) col0 sh gr -/Times-Italic-iso ff 381.00 scf sf +/Times-Italic ff 381.00 scf sf -2250 11700 m gs 1 -1 sc (Node 1) col0 sh gr -/AvantGarde-Book-iso ff 396.88 scf sf +/AvantGarde-Book ff 396.88 scf sf 4140 12105 m gs 1 -1 sc (X) col0 sh gr -/AvantGarde-Book-iso ff 396.88 scf sf +/AvantGarde-Book ff 396.88 scf sf 4635 12105 m gs 1 -1 sc (X) col0 sh gr -/AvantGarde-Book-iso ff 396.88 scf sf +/AvantGarde-Book ff 396.88 scf sf 5040 12105 m gs 1 -1 sc (X) col0 sh gr -/AvantGarde-Book-iso ff 396.88 scf sf +/AvantGarde-Book ff 396.88 scf sf 5490 12105 m gs 1 -1 sc (X) col0 sh gr -/AvantGarde-Book-iso ff 396.88 scf sf +/AvantGarde-Book ff 396.88 scf sf 2790 12105 m gs 1 -1 sc (X) col0 sh gr -/AvantGarde-Book-iso ff 396.88 scf sf +/AvantGarde-Book ff 396.88 scf sf 5985 12105 m gs 1 -1 sc (X) col0 sh gr -/AvantGarde-Book-iso ff 396.88 scf sf +/AvantGarde-Book ff 396.88 scf sf 2340 12105 m gs 1 -1 sc (X) col32 sh gr -/AvantGarde-Book-iso ff 396.88 scf sf +/AvantGarde-Book ff 396.88 scf sf 3240 12105 m gs 1 -1 sc (X) col32 sh gr -/AvantGarde-Book-iso ff 396.88 scf sf +/AvantGarde-Book ff 396.88 scf sf 3690 12105 m gs 1 -1 sc (X) col32 sh gr -/AvantGarde-Book-iso ff 396.88 scf sf +/AvantGarde-Book ff 396.88 scf sf 7740 12105 m gs 1 -1 sc (X) col32 sh gr -/AvantGarde-Book-iso ff 396.88 scf sf +/AvantGarde-Book ff 396.88 scf sf 9135 12105 m gs 1 -1 sc (X) col32 sh gr -/AvantGarde-Book-iso ff 396.88 scf sf +/AvantGarde-Book ff 396.88 scf sf 7335 12105 m gs 1 -1 sc (X) col32 sh gr -/AvantGarde-Book-iso ff 396.88 scf sf +/AvantGarde-Book ff 396.88 scf sf 6435 12105 m gs 1 -1 sc (X) col0 sh gr -/AvantGarde-Book-iso ff 396.88 scf sf +/AvantGarde-Book ff 396.88 scf sf 8685 12105 m gs 1 -1 sc (X) col0 sh gr -/AvantGarde-Book-iso ff 396.88 scf sf +/AvantGarde-Book ff 396.88 scf sf 8235 12105 m gs 1 -1 sc (X) col0 sh gr -/AvantGarde-Book-iso ff 396.88 scf sf +/AvantGarde-Book ff 396.88 scf sf 6840 12105 m gs 1 -1 sc (X) col32 sh gr -/Times-Roman-iso ff 317.50 scf sf +/Times-Roman ff 317.50 scf sf 4680 12420 m gs 1 -1 sc (local) col0 sh gr -/Times-Roman-iso ff 317.50 scf sf +/Times-Roman ff 317.50 scf sf 4275 12645 m -gs 1 -1 sc (Sub-vector ) col0 sh gr -/Times-Roman-iso ff 317.50 scf sf +gs 1 -1 sc (Subvector ) col0 sh gr +/Times-Roman ff 317.50 scf sf -135 10350 m -gs 1 -1 sc (Local sub-matrix) col0 sh gr -/Times-Roman-iso ff 317.50 scf sf +gs 1 -1 sc (Local submatrix) col0 sh gr +/Times-Roman ff 317.50 scf sf 270 12015 m gs 1 -1 sc (Global vector ) col0 sh gr -/Times-Roman-iso ff 317.50 scf sf +/Times-Roman ff 317.50 scf sf 2205 13815 m -gs 1 -1 sc (Shared sub-vector) col0 sh gr -/Times-Roman-iso ff 317.50 scf sf +gs 1 -1 sc (Shared subvector) col0 sh gr +/Times-Roman ff 317.50 scf sf 10575 12015 m gs 1 -1 sc (Sparse storage) col4 sh gr -/Times-Roman-iso ff 317.50 scf sf +/Times-Roman ff 317.50 scf sf 990 14580 m gs 1 -1 sc (Send vector elements ) col4 sh gr -/Times-Roman-iso ff 317.50 scf sf +/Times-Roman ff 317.50 scf sf 1800 14940 m gs 1 -1 sc (to Node 1) col4 sh gr -/Times-Roman-iso ff 317.50 scf sf +/Times-Roman ff 317.50 scf sf 8190 13815 m gs 1 -1 sc (Compressed storage ) col4 sh gr -/Times-Roman-iso ff 317.50 scf sf +/Times-Roman ff 317.50 scf sf 11115 12330 m gs 1 -1 sc (format) col4 sh gr -/Times-Roman-iso ff 317.50 scf sf +/Times-Roman ff 317.50 scf sf 9135 14130 m gs 1 -1 sc (format) col4 sh gr -/Times-Roman-iso ff 317.50 scf sf +/Times-Roman ff 317.50 scf sf -675 17010 m gs 1 -1 sc (Determine vector elements) col4 sh gr -/Times-Roman-iso ff 317.50 scf sf +/Times-Roman ff 317.50 scf sf 0 17415 m gs 1 -1 sc (needed by Node 1) col4 sh gr -/Times-Roman-iso ff 317.50 scf sf +/Times-Roman ff 317.50 scf sf -225 18315 m -gs 1 -1 sc (Local sub-vectors) col0 sh gr -/Times-Italic-iso ff 317.50 scf sf +gs 1 -1 sc (Local subvectors) col0 sh gr +/Times-Italic ff 317.50 scf sf 2655 18810 m gs 1 -1 sc (Node 0) col0 sh gr -/Times-Italic-iso ff 317.50 scf sf +/Times-Italic ff 317.50 scf sf 5355 18810 m gs 1 -1 sc (Node 2) col0 sh gr -/Times-Italic-iso ff 317.50 scf sf +/Times-Italic ff 317.50 scf sf 8055 18810 m gs 1 -1 sc (Node 3) col0 sh gr -/Times-Italic-iso ff 381.00 scf sf +/Times-Italic ff 381.00 scf sf -2700 17055 m gs 1 -1 sc (Neighbors ) col0 sh gr -/Times-Italic-iso ff 381.00 scf sf +/Times-Italic ff 381.00 scf sf -2475 17505 m -gs 1 -1 sc (0, 2 et 3) col0 sh gr +gs 1 -1 sc (0, 2, and 3) col0 sh gr % here ends figure; pagefooter showpage diff --git a/BookGPU/Chapters/chapter12/figures/compress.fig b/BookGPU/Chapters/chapter12/figures/compress.fig index ac628da..5cc0187 100644 --- a/BookGPU/Chapters/chapter12/figures/compress.fig +++ b/BookGPU/Chapters/chapter12/figures/compress.fig @@ -17,76 +17,76 @@ Single 5850 9450 5850 11250 2 1 0 2 0 7 50 -1 -1 0.000 0 0 -1 0 0 2 7650 9450 7650 11250 -4 0 0 50 -1 4 25 0.0000 4 315 255 2790 10755 X\001 -4 0 0 50 -1 4 25 0.0000 4 315 255 4140 9855 X\001 -4 0 0 50 -1 4 25 0.0000 4 315 255 4140 10755 X\001 -4 0 0 50 -1 4 25 0.0000 4 315 255 4590 10305 X\001 -4 0 0 50 -1 4 25 0.0000 4 315 255 5040 9855 X\001 -4 0 0 50 -1 4 25 0.0000 4 315 255 5040 10755 X\001 -4 0 0 50 -1 4 25 0.0000 4 315 255 5490 10305 X\001 -4 0 0 50 -1 4 25 0.0000 4 315 255 5490 11205 X\001 -4 0 0 50 -1 4 25 0.0000 4 315 255 5940 11205 X\001 -4 0 0 50 -1 4 25 0.0000 4 315 255 6390 9855 X\001 -4 0 0 50 -1 4 25 0.0000 4 315 255 6390 10305 X\001 -4 0 0 50 -1 4 25 0.0000 4 315 255 6390 10755 X\001 -4 0 0 50 -1 4 25 0.0000 4 315 255 8235 10305 X\001 -4 0 0 50 -1 4 25 0.0000 4 315 255 8235 11205 X\001 -4 0 0 50 -1 4 25 0.0000 4 315 255 8685 10305 X\001 +4 0 0 50 -1 4 25 0.0000 4 300 255 2790 10755 X\001 +4 0 0 50 -1 4 25 0.0000 4 300 255 4140 9855 X\001 +4 0 0 50 -1 4 25 0.0000 4 300 255 4140 10755 X\001 +4 0 0 50 -1 4 25 0.0000 4 300 255 4590 10305 X\001 +4 0 0 50 -1 4 25 0.0000 4 300 255 5040 9855 X\001 +4 0 0 50 -1 4 25 0.0000 4 300 255 5040 10755 X\001 +4 0 0 50 -1 4 25 0.0000 4 300 255 5490 10305 X\001 +4 0 0 50 -1 4 25 0.0000 4 300 255 5490 11205 X\001 +4 0 0 50 -1 4 25 0.0000 4 300 255 5940 11205 X\001 +4 0 0 50 -1 4 25 0.0000 4 300 255 6390 9855 X\001 +4 0 0 50 -1 4 25 0.0000 4 300 255 6390 10305 X\001 +4 0 0 50 -1 4 25 0.0000 4 300 255 6390 10755 X\001 +4 0 0 50 -1 4 25 0.0000 4 300 255 8235 10305 X\001 +4 0 0 50 -1 4 25 0.0000 4 300 255 8235 11205 X\001 +4 0 0 50 -1 4 25 0.0000 4 300 255 8685 10305 X\001 -6 6 4725 13455 7065 14355 2 2 0 2 0 7 50 -1 -1 0.000 0 0 -1 0 0 5 4770 13500 7020 13500 7020 13950 4770 13950 4770 13500 -4 0 0 50 -1 0 24 0.0000 4 270 210 4905 14310 1\001 -4 0 0 50 -1 0 24 0.0000 4 270 210 5355 14310 8\001 -4 0 0 50 -1 0 24 0.0000 4 285 210 5805 14310 9\001 -4 0 0 50 -1 0 24 0.0000 4 270 420 6120 14310 13\001 -4 0 0 50 -1 0 24 0.0000 4 270 420 6615 14310 14\001 -4 0 0 50 -1 4 25 0.0000 4 315 255 4860 13905 X\001 -4 0 0 50 -1 4 25 0.0000 4 315 255 5310 13905 X\001 -4 0 0 50 -1 4 25 0.0000 4 315 255 5760 13905 X\001 -4 0 0 50 -1 4 25 0.0000 4 315 255 6210 13905 X\001 -4 0 0 50 -1 4 25 0.0000 4 315 255 6705 13905 X\001 +4 0 0 50 -1 0 24 0.0000 4 270 195 4905 14310 1\001 +4 0 0 50 -1 0 24 0.0000 4 270 195 5355 14310 8\001 +4 0 0 50 -1 0 24 0.0000 4 270 195 5805 14310 9\001 +4 0 0 50 -1 0 24 0.0000 4 270 390 6120 14310 13\001 +4 0 0 50 -1 0 24 0.0000 4 270 390 6615 14310 14\001 +4 0 0 50 -1 4 25 0.0000 4 300 255 4860 13905 X\001 +4 0 0 50 -1 4 25 0.0000 4 300 255 5310 13905 X\001 +4 0 0 50 -1 4 25 0.0000 4 300 255 5760 13905 X\001 +4 0 0 50 -1 4 25 0.0000 4 300 255 6210 13905 X\001 +4 0 0 50 -1 4 25 0.0000 4 300 255 6705 13905 X\001 -6 6 2880 15705 3420 16245 2 2 0 2 0 7 50 -1 -1 0.000 0 0 -1 0 0 5 2925 15750 3375 15750 3375 16200 2925 16200 2925 15750 -4 0 0 50 -1 4 25 0.0000 4 315 255 3015 16155 X\001 +4 0 0 50 -1 4 25 0.0000 4 300 255 3015 16155 X\001 -6 6 2205 17640 4095 18495 2 2 0 2 0 7 50 -1 -1 0.000 0 0 -1 0 0 5 2250 18000 4050 18000 4050 18450 2250 18450 2250 18000 -4 0 0 50 -1 0 24 0.0000 4 270 210 2385 17910 0\001 -4 0 0 50 -1 0 24 0.0000 4 270 210 2835 17910 1\001 -4 0 0 50 -1 0 24 0.0000 4 270 210 3285 17910 2\001 -4 0 0 50 -1 0 24 0.0000 4 270 210 3735 17910 3\001 -4 0 0 50 -1 4 25 0.0000 4 315 255 2340 18405 X\001 -4 0 0 50 -1 4 25 0.0000 4 315 255 3240 18405 X\001 -4 0 0 50 -1 4 25 0.0000 4 315 255 2790 18405 X\001 -4 0 0 50 -1 4 25 0.0000 4 315 255 3735 18405 X\001 +4 0 0 50 -1 0 24 0.0000 4 270 195 2385 17910 0\001 +4 0 0 50 -1 0 24 0.0000 4 270 195 2835 17910 1\001 +4 0 0 50 -1 0 24 0.0000 4 270 195 3285 17910 2\001 +4 0 0 50 -1 0 24 0.0000 4 270 195 3735 17910 3\001 +4 0 0 50 -1 4 25 0.0000 4 300 255 2340 18405 X\001 +4 0 0 50 -1 4 25 0.0000 4 300 255 3240 18405 X\001 +4 0 0 50 -1 4 25 0.0000 4 300 255 2790 18405 X\001 +4 0 0 50 -1 4 25 0.0000 4 300 255 3735 18405 X\001 -6 6 4905 17640 6840 18495 2 2 0 2 0 7 50 -1 -1 0.000 0 0 -1 0 0 5 4950 18000 6750 18000 6750 18450 4950 18450 4950 18000 -4 0 0 50 -1 0 24 0.0000 4 270 210 5040 17910 8\001 -4 0 0 50 -1 0 24 0.0000 4 285 210 5535 17910 9\001 -4 0 0 50 -1 0 24 0.0000 4 270 420 5895 17910 10\001 -4 0 0 50 -1 0 24 0.0000 4 270 420 6390 17910 11\001 -4 0 0 50 -1 4 25 0.0000 4 315 255 5040 18405 X\001 -4 0 0 50 -1 4 25 0.0000 4 315 255 5490 18405 X\001 -4 0 0 50 -1 4 25 0.0000 4 315 255 5985 18405 X\001 -4 0 0 50 -1 4 25 0.0000 4 315 255 6435 18405 X\001 +4 0 0 50 -1 0 24 0.0000 4 270 195 5040 17910 8\001 +4 0 0 50 -1 0 24 0.0000 4 270 195 5535 17910 9\001 +4 0 0 50 -1 0 24 0.0000 4 270 390 5895 17910 10\001 +4 0 0 50 -1 0 24 0.0000 4 270 390 6390 17910 11\001 +4 0 0 50 -1 4 25 0.0000 4 300 255 5040 18405 X\001 +4 0 0 50 -1 4 25 0.0000 4 300 255 5490 18405 X\001 +4 0 0 50 -1 4 25 0.0000 4 300 255 5985 18405 X\001 +4 0 0 50 -1 4 25 0.0000 4 300 255 6435 18405 X\001 -6 6 7605 17595 9495 18495 2 2 0 2 0 7 50 -1 -1 0.000 0 0 -1 0 0 5 7650 18000 9450 18000 9450 18450 7650 18450 7650 18000 -4 0 0 50 -1 0 24 0.0000 4 270 420 7650 17910 12\001 -4 0 0 50 -1 0 24 0.0000 4 270 420 8145 17910 13\001 -4 0 0 50 -1 0 24 0.0000 4 270 420 8595 17910 14\001 -4 0 0 50 -1 0 24 0.0000 4 285 420 9045 17910 15\001 -4 0 0 50 -1 4 25 0.0000 4 315 255 8235 18405 X\001 -4 0 0 50 -1 4 25 0.0000 4 315 255 7785 18405 X\001 -4 0 0 50 -1 4 25 0.0000 4 315 255 8685 18405 X\001 -4 0 0 50 -1 4 25 0.0000 4 315 255 9090 18405 X\001 +4 0 0 50 -1 0 24 0.0000 4 270 390 7650 17910 12\001 +4 0 0 50 -1 0 24 0.0000 4 270 390 8145 17910 13\001 +4 0 0 50 -1 0 24 0.0000 4 270 390 8595 17910 14\001 +4 0 0 50 -1 0 24 0.0000 4 270 390 9045 17910 15\001 +4 0 0 50 -1 4 25 0.0000 4 300 255 8235 18405 X\001 +4 0 0 50 -1 4 25 0.0000 4 300 255 7785 18405 X\001 +4 0 0 50 -1 4 25 0.0000 4 300 255 8685 18405 X\001 +4 0 0 50 -1 4 25 0.0000 4 300 255 9090 18405 X\001 -6 2 2 0 2 0 7 50 -1 -1 0.000 0 0 -1 0 0 5 5400 15750 6300 15750 6300 16200 5400 16200 5400 15750 @@ -136,64 +136,64 @@ Single 2 1 1 1 0 32 50 -1 -1 4.000 0 0 -1 1 0 2 0 0 2.00 120.00 150.00 7065 13455 9405 12240 -4 0 0 50 -1 0 24 0.0000 4 270 210 2385 11610 0\001 -4 0 0 50 -1 0 24 0.0000 4 285 210 6435 11610 9\001 -4 0 0 50 -1 0 24 0.0000 4 270 210 5985 11610 8\001 -4 0 0 50 -1 0 24 0.0000 4 270 210 5535 11610 7\001 -4 0 0 50 -1 0 24 0.0000 4 270 210 5085 11610 6\001 -4 0 0 50 -1 0 24 0.0000 4 285 210 4635 11610 5\001 -4 0 0 50 -1 0 24 0.0000 4 270 210 4185 11610 4\001 -4 0 0 50 -1 0 24 0.0000 4 270 210 3735 11610 3\001 -4 0 0 50 -1 0 24 0.0000 4 270 210 3285 11610 2\001 -4 0 0 50 -1 0 24 0.0000 4 270 210 2790 11610 1\001 -4 0 0 50 -1 0 24 0.0000 4 285 420 9045 11610 15\001 -4 0 0 50 -1 0 24 0.0000 4 270 420 8595 11610 14\001 -4 0 0 50 -1 0 24 0.0000 4 270 420 8145 11610 13\001 -4 0 0 50 -1 0 24 0.0000 4 270 420 7695 11610 12\001 -4 0 0 50 -1 0 24 0.0000 4 270 420 7290 11610 11\001 -4 0 0 50 -1 0 24 0.0000 4 270 210 3060 16560 1\001 -4 0 0 50 -1 0 24 0.0000 4 270 210 5535 16560 8\001 -4 0 0 50 -1 0 24 0.0000 4 285 210 5985 16560 9\001 -4 0 0 50 -1 0 24 0.0000 4 270 420 8100 16560 13\001 -4 0 0 50 -1 0 24 0.0000 4 270 420 8595 16560 14\001 -4 0 0 50 -1 0 24 0.0000 4 270 420 6750 11610 10\001 -4 0 0 50 -1 4 25 0.0000 4 315 255 5985 16155 X\001 -4 0 0 50 -1 4 25 0.0000 4 315 255 8190 16155 X\001 -4 0 0 50 -1 4 25 0.0000 4 315 255 8685 16155 X\001 -4 0 0 50 -1 4 25 0.0000 4 315 255 5490 16155 X\001 -4 0 0 50 -1 1 24 0.0000 4 285 1185 -2250 11700 Node 1\001 -4 0 0 50 -1 4 25 0.0000 4 315 255 4140 12105 X\001 -4 0 0 50 -1 4 25 0.0000 4 315 255 4635 12105 X\001 -4 0 0 50 -1 4 25 0.0000 4 315 255 5040 12105 X\001 -4 0 0 50 -1 4 25 0.0000 4 315 255 5490 12105 X\001 -4 0 0 50 -1 4 25 0.0000 4 315 255 2790 12105 X\001 -4 0 0 50 -1 4 25 0.0000 4 315 255 5985 12105 X\001 -4 0 32 50 -1 4 25 0.0000 4 315 255 2340 12105 X\001 -4 0 32 50 -1 4 25 0.0000 4 315 255 3240 12105 X\001 -4 0 32 50 -1 4 25 0.0000 4 315 255 3690 12105 X\001 -4 0 32 50 -1 4 25 0.0000 4 315 255 7740 12105 X\001 -4 0 32 50 -1 4 25 0.0000 4 315 255 9135 12105 X\001 -4 0 32 50 -1 4 25 0.0000 4 315 255 7335 12105 X\001 -4 0 0 50 -1 4 25 0.0000 4 315 255 6435 12105 X\001 -4 0 0 50 -1 4 25 0.0000 4 315 255 8685 12105 X\001 -4 0 0 50 -1 4 25 0.0000 4 315 255 8235 12105 X\001 -4 0 32 50 -1 4 25 0.0000 4 315 255 6840 12105 X\001 -4 0 0 50 -1 0 20 0.0000 4 225 645 4680 12420 local\001 -4 0 0 50 -1 0 20 0.0000 4 225 1530 4275 12645 Sub-vector \001 -4 0 0 50 -1 0 20 0.0000 4 225 2280 -135 10350 Local sub-matrix\001 -4 0 0 50 -1 0 20 0.0000 4 225 1905 270 12015 Global vector \001 -4 0 0 50 -1 0 20 0.0000 4 225 2400 2205 13815 Shared sub-vector\001 -4 0 4 50 -1 0 20 0.0000 4 300 1935 10575 12015 Sparse storage\001 -4 0 4 50 -1 0 20 0.0000 4 225 2940 990 14580 Send vector elements \001 -4 0 4 50 -1 0 20 0.0000 4 225 1320 1800 14940 to Node 1\001 -4 0 4 50 -1 0 20 0.0000 4 300 2790 8190 13815 Compressed storage \001 -4 0 4 50 -1 0 20 0.0000 4 225 870 11115 12330 format\001 -4 0 4 50 -1 0 20 0.0000 4 225 870 9135 14130 format\001 -4 0 4 50 -1 0 20 0.0000 4 225 3585 -675 17010 Determine vector elements\001 -4 0 4 50 -1 0 20 0.0000 4 300 2430 0 17415 needed by Node 1\001 -4 0 0 50 -1 0 20 0.0000 4 225 2385 -225 18315 Local sub-vectors\001 -4 0 0 50 -1 1 20 0.0000 4 225 960 2655 18810 Node 0\001 -4 0 0 50 -1 1 20 0.0000 4 225 960 5355 18810 Node 2\001 -4 0 0 50 -1 1 20 0.0000 4 225 960 8055 18810 Node 3\001 -4 0 0 50 -1 1 24 0.0000 4 375 1845 -2700 17055 Neighbors \001 -4 0 0 50 -1 1 24 0.0000 4 315 1350 -2475 17505 0, 2 et 3\001 +4 0 0 50 -1 0 24 0.0000 4 270 195 2385 11610 0\001 +4 0 0 50 -1 0 24 0.0000 4 270 195 6435 11610 9\001 +4 0 0 50 -1 0 24 0.0000 4 270 195 5985 11610 8\001 +4 0 0 50 -1 0 24 0.0000 4 270 195 5535 11610 7\001 +4 0 0 50 -1 0 24 0.0000 4 270 195 5085 11610 6\001 +4 0 0 50 -1 0 24 0.0000 4 270 195 4635 11610 5\001 +4 0 0 50 -1 0 24 0.0000 4 270 195 4185 11610 4\001 +4 0 0 50 -1 0 24 0.0000 4 270 195 3735 11610 3\001 +4 0 0 50 -1 0 24 0.0000 4 270 195 3285 11610 2\001 +4 0 0 50 -1 0 24 0.0000 4 270 195 2790 11610 1\001 +4 0 0 50 -1 0 24 0.0000 4 270 390 9045 11610 15\001 +4 0 0 50 -1 0 24 0.0000 4 270 390 8595 11610 14\001 +4 0 0 50 -1 0 24 0.0000 4 270 390 8145 11610 13\001 +4 0 0 50 -1 0 24 0.0000 4 270 390 7695 11610 12\001 +4 0 0 50 -1 0 24 0.0000 4 270 390 7290 11610 11\001 +4 0 0 50 -1 0 24 0.0000 4 270 195 3060 16560 1\001 +4 0 0 50 -1 0 24 0.0000 4 270 195 5535 16560 8\001 +4 0 0 50 -1 0 24 0.0000 4 270 195 5985 16560 9\001 +4 0 0 50 -1 0 24 0.0000 4 270 390 8100 16560 13\001 +4 0 0 50 -1 0 24 0.0000 4 270 390 8595 16560 14\001 +4 0 0 50 -1 0 24 0.0000 4 270 390 6750 11610 10\001 +4 0 0 50 -1 4 25 0.0000 4 300 255 5985 16155 X\001 +4 0 0 50 -1 4 25 0.0000 4 300 255 8190 16155 X\001 +4 0 0 50 -1 4 25 0.0000 4 300 255 8685 16155 X\001 +4 0 0 50 -1 4 25 0.0000 4 300 255 5490 16155 X\001 +4 0 0 50 -1 1 24 0.0000 4 285 1110 -2250 11700 Node 1\001 +4 0 0 50 -1 4 25 0.0000 4 300 255 4140 12105 X\001 +4 0 0 50 -1 4 25 0.0000 4 300 255 4635 12105 X\001 +4 0 0 50 -1 4 25 0.0000 4 300 255 5040 12105 X\001 +4 0 0 50 -1 4 25 0.0000 4 300 255 5490 12105 X\001 +4 0 0 50 -1 4 25 0.0000 4 300 255 2790 12105 X\001 +4 0 0 50 -1 4 25 0.0000 4 300 255 5985 12105 X\001 +4 0 32 50 -1 4 25 0.0000 4 300 255 2340 12105 X\001 +4 0 32 50 -1 4 25 0.0000 4 300 255 3240 12105 X\001 +4 0 32 50 -1 4 25 0.0000 4 300 255 3690 12105 X\001 +4 0 32 50 -1 4 25 0.0000 4 300 255 7740 12105 X\001 +4 0 32 50 -1 4 25 0.0000 4 300 255 9135 12105 X\001 +4 0 32 50 -1 4 25 0.0000 4 300 255 7335 12105 X\001 +4 0 0 50 -1 4 25 0.0000 4 300 255 6435 12105 X\001 +4 0 0 50 -1 4 25 0.0000 4 300 255 8685 12105 X\001 +4 0 0 50 -1 4 25 0.0000 4 300 255 8235 12105 X\001 +4 0 32 50 -1 4 25 0.0000 4 300 255 6840 12105 X\001 +4 0 0 50 -1 0 20 0.0000 4 225 615 4680 12420 local\001 +4 0 0 50 -1 0 20 0.0000 4 225 1395 4275 12645 Subvector \001 +4 0 0 50 -1 0 20 0.0000 4 225 2130 -135 10350 Local submatrix\001 +4 0 0 50 -1 0 20 0.0000 4 225 1845 270 12015 Global vector \001 +4 0 0 50 -1 0 20 0.0000 4 225 2235 2205 13815 Shared subvector\001 +4 0 4 50 -1 0 20 0.0000 4 300 1890 10575 12015 Sparse storage\001 +4 0 4 50 -1 0 20 0.0000 4 225 2850 990 14580 Send vector elements \001 +4 0 4 50 -1 0 20 0.0000 4 225 1260 1800 14940 to Node 1\001 +4 0 4 50 -1 0 20 0.0000 4 300 2670 8190 13815 Compressed storage \001 +4 0 4 50 -1 0 20 0.0000 4 225 885 11115 12330 format\001 +4 0 4 50 -1 0 20 0.0000 4 225 885 9135 14130 format\001 +4 0 4 50 -1 0 20 0.0000 4 225 3495 -675 17010 Determine vector elements\001 +4 0 4 50 -1 0 20 0.0000 4 300 2265 0 17415 needed by Node 1\001 +4 0 0 50 -1 0 20 0.0000 4 225 2175 -225 18315 Local subvectors\001 +4 0 0 50 -1 1 20 0.0000 4 240 885 2655 18810 Node 0\001 +4 0 0 50 -1 1 20 0.0000 4 240 885 5355 18810 Node 2\001 +4 0 0 50 -1 1 20 0.0000 4 240 885 8055 18810 Node 3\001 +4 0 0 50 -1 1 24 0.0000 4 390 1725 -2700 17055 Neighbors \001 +4 0 0 50 -1 1 24 0.0000 4 345 1665 -2475 17505 0, 2, and 3\001 diff --git a/BookGPU/Chapters/chapter12/figures/compress.fig.bak b/BookGPU/Chapters/chapter12/figures/compress.fig.bak index b8ecc58..ac628da 100644 --- a/BookGPU/Chapters/chapter12/figures/compress.fig.bak +++ b/BookGPU/Chapters/chapter12/figures/compress.fig.bak @@ -126,14 +126,16 @@ Single 2 1 0 1 4 7 50 -1 -1 0.000 0 0 -1 0 1 2 1 1 2.00 90.00 150.00 9540 11925 10530 11925 -2 1 1 1 0 32 50 -1 -1 4.000 0 0 -1 0 0 2 - 4680 13455 2295 12240 -2 1 1 1 0 32 50 -1 -1 4.000 0 0 -1 0 0 2 - 7065 13455 9405 12240 2 1 0 1 0 32 50 -1 -1 0.000 0 0 -1 0 0 1 9405 12240 2 1 0 1 0 32 50 -1 -1 0.000 0 0 -1 0 0 1 9405 12240 +2 1 1 1 0 32 50 -1 -1 4.000 0 0 -1 1 0 2 + 0 0 2.00 120.00 150.00 + 4680 13455 2295 12240 +2 1 1 1 0 32 50 -1 -1 4.000 0 0 -1 1 0 2 + 0 0 2.00 120.00 150.00 + 7065 13455 9405 12240 4 0 0 50 -1 0 24 0.0000 4 270 210 2385 11610 0\001 4 0 0 50 -1 0 24 0.0000 4 285 210 6435 11610 9\001 4 0 0 50 -1 0 24 0.0000 4 270 210 5985 11610 8\001 diff --git a/BookGPU/Chapters/chapter12/figures/compress.pdf b/BookGPU/Chapters/chapter12/figures/compress.pdf index 3163620e6dd76ce8d65bd05bd1373092aef87a03..e53930c198fb6692e66e1c348b19dcff3869d9c5 100644 GIT binary patch delta 9535 zcma)hbx>W;x+N}$;1-+&_rt*s?gWCn1rprdb)&(89tiFh+}+(R5Hz@l;2vyp-@K`r zcYlA(?W&eqwZ2tfOZDEhde^GRI!+eQchKj<^IxcfI+P(1j=c$MT4;~}eA(x8`Lr^vlO5b&|_S{}$8U)GR#*y`!t*KK=UBuNb%R5whm>BpR%}}VmpWq&8;9(0SwEl); zv%oKtqN!jevGKCV1GyJy(wVSf^~$X9Cwcjk56rD3CJ&o&K-^r3F%6qHK2Ts($+)(9pPO?gq) zw`Xq>Oqb-S<4g+YLQcy|yNeF!%?na1*(CV?XhmXpXZ&t$N0aO0)=SgEs-JrFGV`{M zSp5K0=;B@sFhFc`t%bXAhLG#$BSin^lUor85@dwfB7sLH)8BQl@wSq$Wr#}6OyCA4}2i~OlN@% zA53)-%=V!6dCy)C?TcE8bLa?fv^EvrOl`bB@8VddBk~`wE{t+QG~-W>Pj(7xiq=vX zK^vTNUK#XG8UjuB*hi3wTfC*kv3f(HYtlrtsBV}ukXRt5d!Z#jt$yWP43%}Hho3~+ za{KyvObtahUtd5kISxCaPE0!Jr25RQ(jgZamC0>h#x?jAx~BGtdOqf}fa%^ZN}|sF zyH09U(+qK=0X;2F^FfOC+3@%Pth^wq8lnvG^G0ONN_ZB$*il>05SNo9)^}{aLSAgV zr21&6z}HELnT;L3{Ju*-gH&#T|C)-|iKqy98x^|~u}r*#o7{+7GS!1Zu2StAFuN%! zLtkT?kbUEezuWx-y+C0_Qv=1vwX2IiPtm?FLxaP$z7L7Yf(M%7nwA_{AKmZ&P*A#K zspMmN`U3Y=zv{$5E>yXJBS3iE{!Vhy6m_-(aQ4u&F`;8Lil#1TqDtu~Bp4jkgs!Bq zZg=2#V;wBmA}zvr(z)Z*Nf=&2UktH@ZW94t57bRN35_{#O#gJ>JHdD2~- zYX#N){+$5Nax5%99~OmJEdV{nwj}!sh&B34PgNY8M`jt!aOr9_Rq0E_%Qv4bjiyVZ zo=xCX=oa%_YW!ZP{}rkcV%Q20TsY<1XBqVJD7RTL3mzkF=`oYd88jaF{)p^1o@};? z$*ETFr$r5s{SMuMGJFN=m|un@eH}D+p8q2M<|N2@HQt^jvP4fKjgMY|7ez`f4-04WCI}{UJN~Ksz7+vf&qodPDBz+5PQme1*ea^YyH40$S6k z@jFY+O9H05Xd(Kr^!zb38qrqz+2FPePGmni;Rfz=)k;?D&8%mV%Ki|Iw zZm6;-l?-OkUuWf=qS38g*F4Fvy2b*AR8?6@+0UbyU8%i_OrObrtYl+jFQ|~G(gd## z{yYVpE^V*=umWgqBF$I)+JvVUjwvam{JrBr8Kw%|dmPWr0>US4@0weH*42FqZ%6zL zv7w~tdxR%r08kG+#eUd+em>+p74uK#Y8E*KBY303&-i;EeDQ65*AHr)LzPx$=ebcrzV|$|xZU0(~vb`&G3uR7w1HarDsBuPgJA9M+Qy z2uxj&>D=4@@?c*-`ovu}MHKjHU<7Pt2P^@t9c?Hf&Uk((zX#PzpDVf6-L%7z1mErw z0WYX8fA35)ds7!zXEP%^(BF}Ru{DU78chAK9s&mQfXraLXe{uYFk`fz2%P*dF?2g@ zFo+ZUm!ig@#t8v&LSXIaHLwK?LS$(rB{^}}IfgCl69x**?9~BG3M7p3URqgNQw|me zdI!$~n*yaG@PJ{K81#5NU@-OH3i_867nJ{Rg+W2wb$yuX2(a%YO#fRgfs_@23j$*! zB|+fhg~^f%!NzgW;dx{w=85nGN>^$y8iI?genU10MWcYeQc zoE3}92ahWZgVs3D*Es!{e6%4!5VOeHs8sRV_QC4LbBpFDFU65)K{I*Gcb(I>u^Q50 zvcV>qL=(xT2&uftVtdkCfR6(sM!Y( zpGGo&?o29GDrMnWWsHmddFje!v=$HfN$0HZIzG7w4ef$-Yq~+E74c{>vCb1htMod| zk6WiQ<{cajRl&pyx z(w8MG3U5^H6NLwg%~cS7nT!T*sl+74rOEk8v6*rf*)k%#rPc*;%f7;K<%}Rj&#>eTkUR&LWf830aL{k ze>q6*MhT}@@E2+7kLZk|~`S9>@?*_OP6X1dGrT6Xc>Mdf3WCkPg zmyZS#dpIlGf_$hbKdu@3FEG$=H$?uJ9Sr)#ul9qE;V$1{7vs{LAPtZc6nqM;2(nP2 zr5%t2SvvC|hmHAR#0GKM#vNa-bt>1A57X+bwV$-_HwcT$`y<+7bYU94P?j(^*{)D@ z!iQ9DnDNi-3InV@=qXWWC)#(2LcmVYmTBxao^jTy<={oKxHFJjAt>xfkEG4!CU(&2 zrwgjT9xm5Al|;Lcw_?T7t@&FXJ>2B)mE8HLpAmSJ^6!UO4kwl(=+1b0%MNa)1q`UD zh3+kEgZH$HHDUs6Ll^qWz>1zvq?5;Y4JVxB^ zW_q9(2am%`S$jP$z5i3v$%x!oy}jOuuJ<0~lfC{Pji1HaImd$n2Q<7BaGgzc9dRbB zhjC!4rV5lTVu@PZUlfL3dS=*|nuA82qiW^he-P8%MaZ#Zc6vXp(SAks=IY?mw)>15 zvVAcB?1G6bZPkjW+S1tn2|>f6($ZF4N8*vjxQT-3MWJ|?{0oe}l||PX&5cKn$h?KF zX7KqD^SAF*U>Cu9PZRQ?XQLg2gNsW#J1GfE0ueZHPPoMF$^)~rb6C&IC{qzv&BLZ* z!rw`*Y_k}P(O+S}9zuo{c;3h7i4^iWA<@Mp(Pz~_Il_cF)*H}9AvR0h8m~#ozOz_F zuab9ouou|@VJRwnySr)LmN+-{+*mQ(Vda$#TNTGf&#Qe_144s`+9x3#&rJaYrJheD zg#bxm$zEcR4m5Ds!Zk&|07JJ6=dya`XfPY)vovBKAL&Uq;cG2-`Sji0_-p%bUy~yX z8=ZuL)e1EOu<)EU2eqa02>X?kEA{2Q^$KFhc_`J2`y zR@_uen?$R~=GJ8$=*97V+gOhl&2hUTFTe@H7taCR7L<6Afw;T{xprl8IoWrQ79-xl zB6jqqrWSkVZ=YAAcJLhhE+fkN4BUv9r6TAeEgRUeLd4TPR}D0$?}Gxz1vWhVa+?11y5X+7GIpuW2p(~3BsY50R1E`>RLIow z@~X50mb*+Kv7U?Br!67g(hdsGmM9&T_O_K|6kk84lm3z% zJr){v6^3fKSRQEruYt}YN zJK;`y0`gQXNg{PRe-evn3oJ+=S6c%G5!21xDqrn`@V*L}n*dlCDgj?bk__WNd>a%> z*iA{d#&Wn~OAeAjO9OPY-I{rsl}eBi3w91Jl4RVn(&*%TeHAQiT-T55w9ne*joUMH zt*zOZD#r-*My>;D?6&4&ol;vo3pa+{t-dhiB>8mpT~Q^U0_#Tav6zX7`w2%`G=ESF zX{u}bm_1DY=@U8s4p&o^?aN>*wX?I%eEjK3m_6Kkq*bKKjz9Xa5=Rz(z zInebB$#WBV#aS;Y37sUMpv0CJ!l#QaoQE}ktciht(EA&K5ou1h00Re>RA0bTyF**X zcI9j=zn4G~JHI8T-O4IiJm6zBUIX7DB9~kq89Ofxx4iL0kriVXMkUe-^}WhT`O#r2 z|NcjaVv_LQE3ml@O1WGgEk=qSX(geDR_bWv*hTBbA{-slVda-vSC>d>ecDjJ@b-dr z^w))KO9eaElDgJYs#o?@6Ssk0db=oyYFh=&f69d1*dAk2mblmZn>)yA8}dAf-_;wp z2KTWJkRiR}&4jViF>rK~;LZfQr9AJ>XewMs4CA+Js9fy0?z=T(xjE13U3nbqEk-_( z@dMQ=Yt;}l0ay?nooT1oq>*UciqK+O#*VR@06o`nM*0VYSe(x2v5T|qpDFD$NKSYyUx5$b)T#@;WWjPT_e7rrm#EeMsN z9h4Med=K;9ZlZu;(38N+!|3Q)0f0|^y69yR{OX{%QL1F9NX)oxFyAU?-tw0q!{ksE zk~Y~2s=iW5-BFoROXRP>k9IiN1Z1F1_TF~~U5^}oCnr>OFDYoT(^f5S4A z=iWK|Ee|7ZTfPF%g4#wAngoy;wegFfci z*>`evcU-iiwFfLrO!n3tiurE`<5wq-bH<(l;G+94*ji|BO-yz#(iR7cmeYy`7njj$ zC|p!ma4n2v9l{5mpIypYt>xh%)~ zCR_Erj>W3gTg}JZ=+D5csQ+1~H2G=Ne#{^2ZaUYNs-wYFk-~Zd*dzl8z6!R)zy#O` z2=B+~<*norEh}o&EKam^dx^vtR8ZLy+Wb;%3Dr^uN*Oo9QwVexrajGSY^2f;gZb0& z4rH1$KWFqB=ht*LcI8n_iz$N147Gwg5@5m=(}QVNQ?`o8KgFQg_9qTym*(W$g#>PsxwRQ**OWD7NM6BO`$4MhikIR#r zF~8HFS~o4?GiM;^)9rt>o)GsOpA*^H$j)*h;m@Lx9ml>(pgn27R_u4_T!2~mSAAx+<>oQ~ErPbU&lk!|$>po_oRoMax%%Ipgsexm@V!<1K zpQ2+Mv8KtpmiIj04=kNLn?KZX?K$`a5U-i8HVKPd?v!z&KC{St#z(p&!4{=#d7A8V z=MYis%(?H1-Q-xi7_Ll5$lEs(_BW}BT~;#RVM0~@xxKYI?t$>Tj|DZXJ>!GP0tO73 znFa+{Z>#Zs493pP48Y1NisyMDyulf>RgP=&?A!b`x?-LJ*t=R(5`Mc%o?_;`%b$6k zIt~1jjqR(Sro*NHqlYdM{Vs<3YzYo0LcR?;+c2q3&0-vM>eZnA{SU+MOx1E$Zg@}) zw+Z(#$k@!C+`Qy*$Mgg<#!gbKjXM2?jj$A0iQ2c%e~$D<0&~xTNkwDS{hLD0cdINb z`aQ*9;}vFK`q-T7RN!!yqL&?i>Y6a}9MjZ36;%2HSm@v_U~U6!Iq{!6p&ES~@^ zDJ6b6WfE2|K*;S~bbBQDJ^KW&cf6pw{se>XBdtDiEMVaRIm>a z>5LIY62dUrC%i*^ORcWOL0uQV8a&16)>74eo5vVX&hUYumK(HL<|zSpe#kV!bdi!H z#gwD?j&hTUW_O`fIf8}tnocp`*~y$@iLr6*ZZ6~rU{YEov-R#9kl|fz@A*hU@OzW) z4R;)NfaCHos-$72|GtULrqGqPYkjQYiE+A|5T?;f(p^`0FeA--u>a?4e7v{kE_zLq zr|VE0)a;(0`^Iga7%tO69c=A;BU44(#7dce%AJ34NB;Skidky#FHhrtxEnA69OnPy zNl4-3a2RnIbC@{T*%@(|a+q;ga9DC!aoBS>a5&hTaX53haJV=cnV507S~{EkZ9N=d z52V@v1j7G69EU_d71%q+R zRA>0Hb&4?YxKS{F-LK?H5mMx;cDOzuX!nwnH>$AeC~{1j1Z4W1JlfLCc`GnkyDH{+ zT~@)?ww!VHZ0L(+loT6`S(+s!M(gdAHWTgp33^rXI=x9+>V#3uQgWbyQIhp*xJB=c z#W0;$o*kq`p>T8|Qfsh-Rj5ad6^1FvSb}HQvC>&cePrm;+(0-IkR8#Xn;n6>FGpJ!Tx41B&lTt(6g)yivN zH1D&#*gn!09IT96sJa$`^Eh+3jqYHtw~lyj{M|p%0?4h1o1Z2V6d*oUKVA*tAIGw< zy`74=al1nFrDOx(-gtY3Bk!}5e|A{7#r}CrX?IkPn$)em$8tk_FE0t_P~~v&doJvf zhZ*+kb!uvS(x-8V{C#q6(1mSdVl%Hkvh+`U5sq&Np6YfrJ+A`Jg`H_yaYlYSpp*2v zy4HyhAnP`vp0IBvWq)QtqwwZ|44t$ZF!2EDFpg}oo?Ppgj4G-|g*X7JV- z$$5|6%kw_9=jS1D+EKhcRpHnMiq>(6&)zt~O~($_N#(NWXXJ_x-NUILA4F+$eipK$QnC5%MY3*NpD@X6S~uek z_N1hTeQtCRj*EEPi1s3PDes3WUod&>4o{E|o>@<86BUt8tWF>QT3|!6yIZt9u}vX* ztr)j|Y5={Q`Eo&!KmP72NJV4g&CiL0yy80QblEjK7A6GGnvOr72E-W)m^8!l44YLP zJ$;2Dz_Wk)T+3_(v6HB~rm7eQt9>y)9cD*Q(=ADmaX^0Pfm>pLmffLZm;1Kw!W5sZ ze&iSuH?&Gi(*e4~=X-Kf9hV%m7Qh#Qr7ebC0FCrV?`YF1vKM|h#_6wGCunTA_ z^P06$%(W8gH-0>tLVC~RJow%wDv&_OKr3t#P;40s@VT$M_oXcwdpm@Ch!cKK%Rqeb63byT=~92tpY#8UA)gsgzBDBH4tvR>uWpi?Bj$gjbcB$47ZmOlpL zf!ovdR*#b?n7+;1~ zZ#RGa{rpO|1GCP5%#5+a5DF=M3QONK$@B)84z)jx2GS2KazfsW&V7ZUc38Eu$|DLB~N0;#FZe&IOaV3234!4#@pnmP`v_sDD?7E zv0WU!1em|S)iVqegE~xQpa1s&(3tmj_Vi`tq^k5BrIb`|<66t>u{lit#0%+EC5zCT zg0ARwcISA2k?*n+NoK1?eUih^{X|5HEQ&d|S68kI(dZ3V??T=etM2evujr?xuPpPh zm68xqgB8I;>>&p5H2ZOlMJ26yJwU?Mop93bnllwurxYp+LYDOgzD9F9K;Q`9q<>e$ zG`NA_o;*PZb{Z0z*o9}AMh}?P{9rBCYr)&eqp$y1WU_Z1&~_EE!YZ&O$lLKMA>>ca ze8PA;Oo|GXp3{ssW0eP~663kk%-A5NVDTGH(N3A^XmfS&x=TfuANfX;43_N=HSgk3 z4F;-J!Y)EH8q6*M*;G;-pW=3vO z5!1*kvsSAPbPI5xo`*c=q|aGNqu}AhfS7f1+K;qyntNR_Y;K-Ndb35HLR8PrnGe+b z#T`7nS3M*C&d#4KLB80XYFM_X{T(0n5?g|cTOC+CJ3LHh8&`D&0T&YV4u$JCK`Er< zuaWaT9mFjvYgLMwOY3=v&=JkEDF_h%x)9VBD3SSgX)A-j(n7Xjws(D3WJs3``SFyx zxG%rTM6Yj=1s;a88BThl;n@2&|AbI!qf^}X%MQgWS~?L$Dt=IUzr|(t4n8Y`?`q=^ ztU%!}k=5zlcqz0U%i4<5L5CC@0-cyz1v*>2C&^zuU@IK?o&V!!3&hD^XU=K~&tbyD z#|`E;HDiPDneno5a+;a3f&UgZZf;(v84sr!j}bo~1GT4}t-XtYqp3M8hMfq8&#plG zpTxm}*~M}A*f^mYoKOKCumBG~3m5{MV`up%h|LlHpGmfHNc?BSH(=`jh*)O^9)M@( z2Iwz_@sMA^*@2 h|1u0fA|fCcS0iUvPiHf8(BB>8fq>}fq?Khr{|ou*nVbLs delta 9501 zcma)hWmH^S)-CSt?i!>hrEr3~1Phwr65PEABtb)f!ks{n;O@bK6Ck(*cY;H3dAIxP z_r}+Kf4p8}>@m)nYcARQ&zWbhLWeE;I3>iQGlMW$P)_A;BO|>SK{UVFEtDDqCkCY& z!wQw1xMn(;VMM9Ap#9+?n=Qj?MInyR(p-8{$FxjTyo|m-N36Vf^33J_-gP193g#*N zsnraqIQtNgV&>z5ST+PbPaAi`DQMf9Gb~a&h$z$h=}FvkWYOP;oCw(cydof|Q@xK1 z&o*94PxZpZ&-SDdi)gLhHE3ta{!oHhoLQm70$?aDwifIJ} zV{uMP^n-;#yn*3C9I|Y(!s+0osd8+>yJ9T*e{d(nw){dp_DU19b?IplPN#`6gjH|2 z`B(tlVT@6c@)cMVq}-SDZ<68QyvC^Sl49X4JMrN}@73dg0Yz*eFb2ifDgn3U&1bps zy@sGGy+PmE`COT?C1z<~Q_GcI5C@7$#C4FstjZQ(omP`l~_|< z=8zE0#TWOxhTqN`hv5RkS7mh{}AYeRd-2HCzp6-rDdZ3^FZpYF3rJan6oiCJsYQ%|0G_t{AsJ*&$ z3kkKMAzj2bUCBX$vZ}&R46YL0rf#c6O@kcopEdEpGY4NZnou9wRw!rCFo5GG2zq2= zEuD54i!Xr}Q;{-5l1BMrPm9`lz8vflR8Fngj>Xx{c2dBA`12x{-;a+mj@Q3ghk+9w zfI5n??v+3@`~u4`-&k}ZIgU7?R$4yvsOikB!X*zAo7Zbb!7~g4SI^)`2TAxi1M=`7 z=37w>d=-Zd8qKY;5GbV~-ymAH?^cq=*6z}Li9ed_yBPCqI)pG%(HbV+)qLTVoKGj2 z-9|;9B=8O$p=z5}TVEepf6m$Y;md& z4L8&IwP(9;2j=24WtUxdiFn0H2zMGo2fr(FP|w(H!L3qS9D9>Y*AzO0)g8uqh)<$H zirPB8h)=4)7)9*gx@(<}(LV&668YeQj@dlHrMYkpP`T>}e;Vmk0XM@W)qun15B?e| zh+wzVdz{FD^qq(SwZM%s-k&y`Ws@EjlTH@h??cV~hKwaEWg`vSySCm=j`7w7oS2~SA5SJ?m*=*{eqArZNNmSEgOI)qmK}52J^hfqG!9nQZ>W4D6aU+7 zt@gceG|_4)>q3o&@iy-wz#fXni{I|~#Iw1fU@FndG%J3%U2;%I9${SKsO2;N`J9rE zESh&?TO_nwgA~a?!lrCF>~>_>{W|LhSl@vzW8M%f_`V5$$V;)MoafOWA>mnv`W7na zh-g}1PSz`wM-Y}NyL0e=!LD=tz7(0rTkbu4!o{G$;4yB9qcX_-cmFp<4fZAOxSqP()6)QBkGBI}?ql*9y8_Q5Mm_&(0;Lj{cF@&v+fm&tAi?j|a=y}M$AT*< zBF&zjy~(92xtv^7+Q8AFu}5c`innIJ9}wID*IcKHG~1s?h!o;K`lUN{A=e~k5=5aw zN4Ao~9G<&*qqX-$$sHMIS?5xf7mm zrgb7hOJ|<4|5e00TZ8gYO>(dE3B&s0Yjj2i?KeIcrmns6$vja*mG4NDSY(vYSTW9` zQ+{z^pyA=hg~(jjN3ikmuysf2l}tyWc~c<{>}HQ*vKmH%u2VIb4?S-ccb1zyc76pM zHD7rgT?xfu1|x97x?VB>lJU-Aio-e_>vevzIL+zLD^A&^Y@d}(&6mxu+#+HbDqc0w z@LPn+!J>b`M-?TvwbHcmH;e?2bXQmjP ztWcU22x;;UZI>5uoUd3)K({S3}VVI3jcc)a@FQ8w=7rZET88awF#suiC*Nu zGE^*)1B3_#j{HTy+%Xe4#xMA-FX~p#u~gT>U`sEZO*?S}TO=ok_Bv;yFRcK$%I^bi z(Ev@%f)b{VamjcU>Eq%lwp0Gf#s+bKS*YzUv7(}ndp%y+SPE04)cLG(_{oK3_b;oE z>;#h*@Xy=rZSwmeB^P66>6AEzRn}QQ8rRb()+R({Mdg#qa`rSKn2QMxbI?$9>xr?3;|t#QWF-|j%`GiC}SO5iKbHg)a>HEkH1uGG8l^? zlJNPg3b<<}UTooFNaXj!l2wImU>qcS9-u^L~;Q2Ze)rOtcD7Y{d*jZJ5p z=JiH|aMAqpHp-Xf-uHxoTT(sgV|wJx7hGXh9ab_O^GS229+;?SZ)3;l&VOLlc3FVv z=cdv{Y7vGZfiFbSgu9l6&^|~?iqLqi`aavD9()iRgo$YE%JuO`VB_%sW@Z!o_(c(K zsp-Rt@R5|BAWr$6^uC4QqpZ$Kb+e8@OUX10!aJ8lPYE`BYwQ$2;i(T@ zQ0{#}&mNG#EZH)1x<+c2@LCYf9a4-?&^K_c_SxN^WOdd;tH+n@TNn|fB+1hqpq@%G z?)4Ke2&Q?rdu>@!S>M`atB^9{EO0jRnF*`>>6N+e=X9vUS6)oET<9A@6?}7E^kgp& z)%nB_{_~>lsj{~q7q|A1TSvhdfr8}U^=|&UgqkhrWJ7v8A;eR25|m~GY(BLJ^R-_g z2nqZO&ga74hxm?W-IL?0IcEpA)MGAZwV^=NJkUJO{~MIY}<_ zFiJ%C7gF*(XX5d8poesY@7~y7tX`*eba$q;RvvSk;WrPci~7tEniQ9oH22w{Q;E&m zSJdmGLUuP6a~>gET;dAEk^^|Ra_TQ!fP7Vkj#RB*(p@toZM=8J`%ga=A`TZl_Rqrb zDZP%coYhTTiaeHSv5=+JM<3dBXcst=XvQ>?HltCWja(77(FdK96ul~-4|!_p4KXM> z)KPnZLv3+iS;y%RvQMC=H+^*HCF}e9?gOqX>S2664w1Zq&hG}-H8#-lCftZ$DA011 z0*XCulsm0u?3PDj7t)tliWqKqiSrSdV*4hir2JV^l-#F z2+E}Sg)#@C6q4-Gpt0m$TDtXTuESbP+Jb4|O2}E#vfE^!`>clGb|Rs>`|}G%=|BFq z*7Zwz^C9T1?GxK_PN@2e^w%*}pUZi)ELs2}Cl>iQZLGY>?y;dc)X&nMQ}OUyLIG4Eutn*5R?WWptNfz`c7 zT{FuU_J~59KwMSvDCh!^c6p_;Q^X^uJnyUG?f&zUEwK;7fgkbXi{1EeK-B1UdufOg zU}AG{Qe&q8l`_simA+583X_x%`M1|uv`W38vDPk3_v$u z_c#?@X_ZbCpEP+j35n_jV&&eKb)_(ezg`~~yPpW-JDlCUExw40XYyU>ak%ZBvQ^iM zreYoFK)KSRpTM(rw)&UI`0L5V!T}Em{7dk-Ab${Z{7V~M67w@a@Bq&}EUleASzvHD z9zfRIRms}k*6yFLNGTF=WCV#A=0BAWvYSL1X~Hg#fPldw|69Nq*z4--v@8ST9;Q(R zfbZcPzh+8Cy^1FmL|3&auXs-T^sSQPFTGhiM8)J_rr+pR+`O;kv(HQq~Lj zS2Q$SUd%M?1ZMm>-8h=_nt5=W9U2-MeeBx_THXmVBeDj+Uw4I1w##;N0{IZ)TWgP` zGdo)&CPZKtJa7&^hsEbWi`K)yq51?d^g<8w7WF^kcb^VUH}9@5)XH+?-9#9^R4>aj zl|&K6WA%#)h^+eUZdeZ|KdngRz{SMGFpD40`|Mjexa8>8>LMQU;s!S3*AgP;#PGGTjjY#3Y zM?&AsmOo9|_bBX8mhw7CZD|PPiWk36q~1~&$RpCiE|Hyr4W+o}>1!dvOAe=q94PCN z3-UqohlQ6~zl(zs>8&+SOaA&2+=jU4!V}p!3%0cgr|zc6sBDHlywFM!p%;075|1TW7i zAAGA15J!rEEw?*tPHhm%X#`?i=p!SU!vhDN$mqD->CiUnwF7INd(Tcy!5tu1Zf@f@&Gdaansnt6nO_RIj-G;l+-9yUXtd%^?wrLz z$TWt}?A@ZQl}MS>w~>r3oVu5!yt+V@I;C+M5Xr^SxIpORpH2vbZu2pQ5r+$t%Qb~J zHZlA;u(LU4#Lyek*!C-`8G0ORBp~)+xSmH?#G;rakO>#Dep8DSvyWM=Y@j!AHa!f) z?dUE&;a*PhS>ntj0x^@6Y@P1#?iJmBrkERwMJ#gJ@$w!7x$&`XmhBF{UH&7@!=i%@ z3-q4i?nbcNGv*?d<;yL^?nAeRsi{Uc+$!SJ*m;Hph6FKnUnQ<`b2US*_x}VSUo!3f zxTr9xr&S<{%{RUqgHlj@*;A3*S|AI<)yiuAT&-@7d-^G2Dfs0*c~-KAkVg7$h1DSI z)HH`-YEtrdy1R4Xs&fhKNB2PN6{>;_zlOoKkr4z%kiNbs4q@)F@~Gl{RCiNTTXA!S zc#uh(k*1^+8xx_^r^*P_v~hsi5jh)pnnynBHgVxQX`-}#hHu|`8^dK=;4B-@aG5o_ zj=7zqQe~!?x5veDb85o_qlL^;V#=HSyytV16L9zhLrvjKRAWQjCawg23;{~mHEJk{qrGy{hLi}_z-+sDcS9Dg@7%%Vx zK9%{%*0<^M*MHogO?+04<{1}P3Stsw<`W;q+Tyv1;OF9&be?nXH4O~dp~uZe1E6%Q zEKkN(F%uf7s^{Mf5H;=!3QZaxs>44X(EA`(rTX9VU)}gnn}6pU9fC$&@P43Z*H{l) zSCSE3m(rY}zp(dUVHX1nlb;@&aisXtr)!!p9(`>c=ThOP^r4_Rh2ZeN~tFH|3Ouq%NDMJP(>lP53 z18~C^@Z8R-$smhh%wQEN`@LFPYdZJ`CCT=^va!ajr7c3j&yT^2A0aDHzQxoQJe3&KihPZ>JrF34-Ig|lGFja3koXr1^xp5&Qo%wh9i_g4^4gFiCB>jCPB^)Z@;qCqu_PgfqW3_^VfNnlUZn%9)cP5M=hjd+I`idk3+=1gFZy} z;Fv{Ww0an?Iy#U3D;P^(xLs4+OhHNXZ@BgrkzMahPepa|^+Jn-s=AqeaI$u$VWJ+6 z!qm9YB={GJoAjmvGU;xrgqy?dbb$m%K9G49wc|H>`!sb$^mU=^O;Xfya}gr1rLH|X zNM8Se1;x-Pij7x+Pg5HOFFEIkdZ$|Z@URn%Rn~$=WmBFOW-Qs+KO#>1J!Kec!0}+? z8a>-=dfa3=T>yfF(zS@x&Y>77sYFF5f1yp#jXX7q&TrH7qEpkd0l{o*=Q5`#e*VzP zo^!vjt?7(EiSY0Z(O^zzokFa>w=v32ba^I}u1 zjrv`>hYv&llQ4B<>a|Q<@=?7%zYK#sjoMQB80g1;BRvPdUfY>sBzZQeAB#s>*~dEZONI{r z0vqpv>?WlI@9V?{tJ;KFV(StK#h;s!>Bs2!bweMq3cf4rU8MN!8(lL3&+WcNx@9ob z)PMG+R;;sV3G3nSr^i%;-1@1f z_9t?Hyv;=GY}0ectBJ_VeNyzo=g{a^)C0W;qDpJI3S`ZN^rPowlT;ivUqLR4&n0JW zT6l?afe$k?Owc}&eEWE~>qEfM^+o{MNl&Fvuwf`h<`To_8N(kt@(HcZy>!@N7Hd#A z&*F>97g)i21&%&haUK<<|3)Ko6v2BI&y&a%z^ClY{Z^KJhZ4K@f=zC9E^m+w4E%_k zOm@3TDKG*0ZMA^*ezi#4T*V(dx?*Vh?P%or>6@2nv9lTvR%R`P`9AKc6%C;n0WVX@L10?iCC_+L z&G|f9|B>jAU(Q|NWpf)>Lv=|*VF|DSCga#|GUfft_*qcv+m`apLcmN;H_Lx@!R@h>% zdYNNyi90zxn=60!rAdCNGsT}G)2Grqirx1+u0W16@*>7O{YP2rN1KgsxuI8 z&n*~Qjii1(o0-aR$1g8@)Ya}L1@cx^_q! z-=@EYwUuE7aB>G2<$Z)sX3wYNHNAC}T>7=5FM?xN0CLYyRhs8wg-*+k5hY~o6=$Ag zeplJ0T1TMJX%e_?4v;3!x_xqfTX`jb+^;trLWbCKBY9Rhq{F&~7F9p1LO zjtV2%ES+vJYJqfc002c3FXIgeid&S(xw0?5utK0PR(a&-A19639fN8vLmE6EY9nr@ z)5A_uOM(Wvq`s|!+{ItTr7~}(rMfe`<46h@UPBxLUinCU;O>69aXer@mUZI}lQU;c z7y5I#59!ZFc-JjsLL}Fl*P<=?gv&^gS>JnLk!hBudW&86zCW0xe4qKPc6LM?tTJO5 zxq&d%L z&Tj|>!+nEZ|C(0WKUpc;fsI!~&t6G+rRq{Jl|)7@x@hO^@Ri=~Z23_C@Xa(h#lpGjVZ4mGqv~Ph zz-fhaK{GFPRxA}uw@E6zj zELZMX*FSZ=91u!7#Veu0`6i*ujR~mRxjvQ12vT@*t1Wc#RDy_ey#zmFe23t(yHGox z^|^iU_8XV@>Bzq-!M)STY!RNj_z?Zk@zF1=LzHK( zM^(m-he>3FVJFM?-H_MaCOAfeh; zemhhSYYQ0M!W;w@FgFKU3jl%ERsuqQnE)IrVgVKgia@P}ERl+WPX8ToPVhMiNI(d# z3xtURp`su-pAZnq3*h@ZXa$J)?^2Qo$#DHQO>zT}qX4G=DraqrP%o+=NC*yu0iplw zky}WJn->qD?qVfp?rF^0)vI7X}0g3;#6$`-lBMOb7&m{v87f!I7`QTnPAorVs*w;o!fefC|C>69Yp*f8&9L zz<-V8{=Ga5CiM66FxdapMf|TmAwZz;e+GjfpueVk|DF&8f&L8x3q$@^7y=RcYwGy# zu`rQ;=P~@l4TgaKb9X=>7zRPcKp0Sg$Px&%8AL$@DlaS}3x@!~BC=3fI20xf6c&a- sMdan-@^Gl6`2Uw-iYFvMuf!AP@)|ENC~-O*am~T^lDj!KKlV&`27G0F5R{a0qT8xCaOj+}&M6 zaC!UOch7kD?)~x3S|6*{tXVZ?)lZF4y*l4vaf%4TsOQ3D0mrYi>$^jf*^*+XtT33# z5~P|#DDeCYj#s@>Pk%HKUJMTxPU(#=3Q90_R4(e6rtFod1n?(L$mAn7{k&+*nu*qi z*2(hgP!x=z*EPP^>tvWp6pD~_#YQCq1@%u#sS&J8JzYa?K&hB$Wi4xdQdb=3^c2

zHYD4vV{_&N^*3TK?z~ZHRtSY2O<)Rt}w- z3rT7nQ8hksf7O0@=xV>!Tmu&pK8fdDu8deUIXp+g=*xy@M;WxKbK;e2;@(oF*FHe< z{+K7VynxL@Et5)aG%1cLBs|-{YLm>2jekAIGSBJ=05;=l^U-3Xl$<1ewGV6_Y8j7k z3d@qMvCiJ8v*sl-e_rCUS$;_yMuZ>D?iFiUn-!(t$jdS8_i3vsD2{Zjmp$dlhgw8! zkiA19`*hwfEh50!!owJ>`tcTiTF&9ceZ_*xu}I%g<_le>c?`YeC<71kfpY@{jE3g< zrS(Md+r<)%V(pdF8sxkPsujX!&EnIdJ-t!px{2->*mgWcaMAv1_G30Jl*kvm`8efV zRu;)UXajy9O6V>$&n@;87J+O)V6tMZ`G3(6+@F1t?qU58(rw< zhM9iw#z2UemkJnz)#3z&s-V>i*iZa7_qyz%p-_E9R#Z}n_Jv>n z&&8tZOZyrVnpRe9r6#62n|P*>8)n3W6=DA69@k&}>vI#jtI~s^mmt&2%DWm7h*DgqOF{+m~nG zMZ%ALW)9;6iy~+;W)>02Vx8r-8PVtr=6hVMLq~vjx4kus%U!8^a#ns?mOkQl_-i{% zlJTw;zwR`24;Mca&pj&=#A<%1yS1kw4lqc!K7aE_E{9$JTgy#I+#-yY_=t20`GAe; zv~M?V9YL_`;5nutzg1?Jwer-GTSuPE{W2<;olVY$Z$nq5XT}h)!%}N`tC6V|st=Co zF>krc#IlRi44AyM*4Au#!u?iW=;VWp8Zx@W4Wk|wB){ZZaC5IVgP<7Z-qt~A71BJ(kRvm8 zC8M-R+ARO{Xj}HG;C`u7q)@?sH>|zW9Q4)^YD4NHF_EhA4BYM#X7MJsV^~-{f$AXzBDxI6e2gyTJvBfaJNHg`#YvG8^?a(TzWjJS8Ag_Gad9QvDFcOJ^^86H z13uG?OJ;iM9z_%VQE>5?Uidav(NM1BQwv89!>{i`~ zD7p7U_9J-D3`li8kPL;p$j^D7^xMB*QZ$O!ioWm4hFi*E$`M@xn8&k1JeT zNrKk9IXgPcaD4D6aQu@sh&MHz-cg}#JMV-Gb+aM9%o_9{3X*r|818G&6--|KNbhe* zYvvOa@QHtZd3-jZ`90xkT@#W1Iar~GZYP`5ecASPCD7jU$l%=Ch!u-L%cAy4g7DQX z|C5w=xHc}0d8UNEq^OOz6KcF)@B!KNKDg4LYEP8RtDZ zj*6UtfxHYV4_6+wjjxXh1_Dvi1k@-^0&J879vcP-fC|CmC-^J+Usp&-2sKL}h}y)f z$5EG+QPh`1<>05IUg6Q83ikM?A!GCPSysnc z({w5Ju2@{=T%DoYd-3P*pLmyrv~mC#V(IvL;FGjg878pY++Ow= zk)$7wMjhZYa`nnUNj}60*ga9Jr zTgN(>T~VT9&E-+)G<3KoLD|A3^@6+nRd{V` zi4!ZPhIc^{mJ_<$v*uM&pB51DHA@k>>?lsfq$4)?i6U(oKNJt=?HEjvhP;z!1{f)F z^(=j0&_mMdw3W@*CIXkytry8j=&aDBb@5P)SgSZ*Jj}|>K6-zC@#+M3O`|R2q2aqB z|MLhX!q?v`5AR$#cV8-*LYaPXhAw5b-E3a*cv*3cN@9c)#X3dO8oJn3FgG%+!L2TZ zt>1>Nz$`H5dXFhTh=c@rpNQSZuX|-*O|sj;Nc}Iu_;qV`%d1t08COh zf79+}ChDFA*vA1IwlZa7;!5&AJmbRN2;Pen(jGNun`@#jw6Q%r({qf9X^r>zjHfAB z&rE{+a00}UaARhTH#H*kQ??n^vz9pQ=7*KN3(OiL>lL?QenhhJ)S>cC(qtb@4u3^m zfPL`D4)XB^TBnM|`Qu0Z60h%nXbfr*Gwla0yv^xYU~vEuhjq0AnD|4!{2A*nLA=S} z#4WSe1k(%!dZI-wKXGypymbx)q^c5E$frF7MXicb=)O)rR2e+a_0&>O!ll6NTKV}| z@HKhKuS*Y&==Jm+X3_#4SF(T=l13@<%EnLhe_Wd8kK7v%S!(@yhFq1#c{vMLNPN+@ zIuC|_vd1z&O(V2Pq8@u9(u1^11`+NP@|_Od1PV8N=^`*0>gR=d2_EmC5F>a0EY069 zvu_oTB7H&;^^x*H*YrolX75%A-+b>4k(4q?HU46rFekG+V%<%R-&1VBJ!sk1M#>P6 zznP?6i*1wpJkiViY!kPJP+0OoDq7dA{-mt;H(}fBL%LiY8fyYJwz7Dd%Y8&d-gZ&s z=F|0&k&s+xMmT@KPr8jatNOw3xre96sUx}2W}O@fZ(9NofeyrPjiuwvV@%(QTTS%i zf^t7f-ii*ky?l87>y()f@5ePBg$Z(Y+O807l-zsHl%ak=sm@?g-JGMAyrA$h<|uQ2 z%)rpl>;74qf#ISXtRZA31FI8(%`(u*De)i_tw1B4*4o0cFh}30dM(rXW%`&ne5J<3 z&XIN3A<)U|*0h~ijNz17=i4Z;lS|EKp?tO5w=MCuBT^V)^J_vZr;&yeJd?&Gv--SP zmkq+(ZhV{MALlV%W`EON)+NHzr6Ft2*)YJI?&fG0KXVP+zMN3xiPknjt+q_B?eU7xt z4PIgtzjAklh|_mH5O1Kf8P`~l$6)+>5}s3hM#WPmBLKqxV`4eO{B?Eo8zq6P{?jZO z%mq(I(^c&Xy!Hs4MICx$%dXxi<1n*V$Q+>X(`<^bena zX#dR4DMl?55`}fzaQ!6vflZ`bQUYlA<+C<}Chx_^#~SF)*gt!Ea-QPWGYD{%^0Y80YUFKu7h zVot2*p!Ip~iH<JhL7BW@(CNQuy+wFoPXMH0%0LA|u%0bt~Mvo6{sO zJ#Q}}8PQ5!7;_63PKM9W@xB_|V*Yi?U-hWM=VIZ3+}QnIC#!1tc*hg_V|4jW_~|+o zT^sv=eXcSD;l*pH=IeU+{*R*cqcv68r?Q3!myQa%^5rKr-@>a^rV(X5$^&>5e*(@q zejD7sL4H~4Dt-F)fNw+EZO{g)$o~{Q(E`jd`(?$Dz8p8WzOkdk+e-8XZB1>(+P!U) zC)js4%MI%R`Y?LLpQ(|QT&S|^-@c_-!6i_zD|>aVkL*h69bRsz43X*_3IIKr|GcV% z5J*~{`m=R6hz@%Jxe>^RrypA|JJ2;^64`mtYz)h8ReAWP&}Qay2){tXC(W(*GCIH- zE%~Mk_HXL1$qq<8iDO{M4Zr-Prs`eEptrr~UfDpUu3)z6^G`B^p>b8hH;ene^Fvg@ zURRe_KAkwtk#<|cph=_8pXuWILU~Bw+?1XTlMdUXGlwtJP7%c~!Wg$k^dmK3K6*?@ zMWdXbiQMm6Xx%5JoU`0Xrs5!s%bQ*z1_*3?h7H#}tJU3VVq&mo@6zqb&mkW!`Q_(@ z_PQCS$e}nwe+?542wzrdWmOIe#J?DoFwwXet{@ZZcG(KlbXzwhx?b9P`gz94 zv82w(+P(n_B1UMQj5yU9?(65s`|GGsP%ij>G}&-{_N(QsMGnGS*U71ZYua!^QTGeu zUEq(xpq177yCNjUtMFNdH9(sk?|w(>q@$cjUcFxG066|EmG`^04+HZCMfQ^SbTlDO zxGt$W$TRiTE?bJUTOPvatkc}S5kekEeCy@aAmCLe9UGa=Bxqsu*7-(t_yAPZo-#UT>c#FI{pWE5 z%U>28z$Y7dGsmTal-4swn#VghMS5d+{uzXrfr-1Y=`GFh+)g42PqWcp2YnT#8n=f2 zioSlXfP)(*s&J5lHd~6@rDjGaz=TpVEibB1t3~{9c*QC21rrT@*LMsjo;R!{*%vpb z@uJmueiwPi6ME7QwrTA#THqX8&%zD-IkD#m$9~6rgC4L4jE6Xr&{apoP0eX8P+olC z{mCJ@vXI(4iP%{CLs)^gPw#hA-54UTWlvv$Q#+f{hg``k1b;L;z={m($y!T2K6!Gd zbBCCSt;keSj-a%R(Z9P47@{p4soF17}cN zc-GI-L~LeQ22Le^4}V6$CyND56&&(JAbL%&*x)+fQ?W;2D0hGHiIq{NnOTRE8scyz zjGg3RpF(0BB@!kDYWloT_Y=i$p^ss2v+xf;L=9t}3x?R1#7qiecT{9FN}TgrXJXDY zK)Qba?m`>NQMcVCvpXR#&)i;XGx%`TU}(lM0v67e$0FRH8y+7FbCT>Er82zOMaY8M zon2qKU}^aATnj~DR>xOgbm2Xz3_9r(w7%xxUt>$*kP);xmk5NC_c?$J2j})YR9*%f zk4yAuSS%q$eU_WtPSA#6sZd?ISg^?RL&9+j0YO{laomyHB^~Lk<$WJ@$n1$KEnsxV z6>lPxC{w1EF@e7$CvNnlkAxV3Ah_OVC++OHmx{YS&N0Qtl3(?h<%~pV{`;sP6m#=I)6VVkeZnCU(R;cY{|y*ZXje12 zdNM1nIPrKdHG)a$3OvR~kSeR7HIEu?%hfagkZQMHc0Wpvo2|Fb;xYJLsH^l-)`WhP}7=PBc%YB1*Aao6^p zS0u<;RT84zz}&wP{wx!bY)>0ioIEbA@sNvA-_?!zIWKP&!35;*?d$#7k_-3 zjD!&0QfE>ph!H<(BdqG56>G-S4d$^9^LkkI8Qp!qT5pt~B+LG?EV|~Fe*PSqg6DO-|fu1hc%w<$SY035qdRieBvkymQA{(n(LoUK~G)K!3mSXfzPz zU{qDm&J%D_3(Tq{)G7!y4L{FeOufx~goTG}e*EtHX^7uXJk}14zU)xFFn4}(^PFX7 zeKcic6{c&VZ>WPPsQ?LWA@3y_(B=<8X6In*2aL{wBZ2omkgL-de7T(^$RAU|lvanv zin4oGzM@Nida^W@?~oSWwbxidXE*>=Si*W_3>O*{#?OV^(Lk#Ui%|V--}=$1{rGy{5CSID`}G| zJ(z0nJm8$5`x;p=A$%+j@wN@kgu)L9Zs_0E+ZLS5dQ}v>qP6x3rW3?0fw7@B9v%R$jzy&odDK0-n>GP?Xho1<9|6V)M=p%4n z*IK=UezW))uFB=N%eE_dLT=(o0%si6nq=OR!Yf2xb*V5)Ko4T0-_6iBMCUwEKbbkA zsF7!Wht6sZ6@&iwZ-E6pAmgtNRtWk z0-?G^o;IAaA_4%QAmD#8P--0B5>P>=e>DfdD{=pN13j6~T00BUMj|G8%!hgqr zP{{w4`Qra+2M9o*0{i0zfFi zA)p*c_?4_26burOmzR~7myv}+08k-mkTg_IRzghv|Hm*TkdPpNds=vS`g+*d5&!`p OAmFd9hMa~x!T$ljHXLXG delta 6775 zcma)AbyU^O)|M0wt#pGNI?mw)97+#e(%s#M?)o9Eq#jy2r5mJ?5D5tdL6neIT9Fj| zym#HZ)?M%W$M^0vvu5Vm&pfmDtXcbyd1)}Bo2m?aa{r8^44XGRs>R%ba>dadHxfL- zRWL35_^0ng+1BqRF07cE7X`+Ia!qN=uCT-l`~hx{`Xz&o$>REg+p?et#!wr9hL$IA#VC9>jVT^=o@Rql1AW zZF10LPv@zKY0pn26-&dl_NMV@t>#;G$o%H-URhEtEE z2E&Cb{OQet+j5^JdG}+tTfAsM+x!e}W--7(hs#N}0khE5%uI$3UGY=Z=Coj%+?E^h zGH}DdGNr&CbK`)1C~wSJBn4SRaT*@-tc6z^XF*^=Y)LxRdxkel-R5c%Ya4ew*W0z9 z+h+2N>P1Oos+LZyn*_~RuQoiAlsJ*!FWIJr?YWYx(Cw+9s;y7SS`_a__%lJ#EiJ$! z=c+V*!?H_|grY3NHb>Q=yO@z+u4K?jWSmq{qDvLRKPDa#Ne}z>S)-|DG(<(OjQnLYM!9xI z5TSZpu41pC7D@aj!M1HXgFnYlIS{BQMUW1uXFaIN>)ZP7Z2Pg`t!lESLxuQx@U0|K zUo&>aLGqRs(W*2PcKvUHVDga|WIN(jyrO3gjXy!H%YR}|Y!_v80Nkr_M`R>?FTU^Eafc^r4oEE$OgaU^IQ81di*c#J__M4ltXz z8-q~DkWfQgGFnLbmh$!T`R$9KIhA#or3$x=a@q88__K?se>wT zJi_IEXRC7y+B+-`JZbaJa=!P*?n_x-nR;upu-x_<>I>TFy#cT%XcuhjSsNliRN#gD z*4p$vTJ7x+X=TD4HVcE3tlKi8`X=lA?NK{WSFii+bT4NbXe1A-ziZ!1XRM4Q>1)(f z8<3Iqs_F0PuMnvrO0=^J61bUMszX_8(3+!FVrVt#q%UYlw>mWpts8saEK+*TS(%#K z)xMMr9!5APw9Wv}%c@LZ36$eZ!&zjWLv=0DRFWPL0mui+*ZHqgZ9QJ+7|Auf5hT%6 zIwChn;}RJQywc?i)d%XWrPGz9Ti6>QMoC==rH}8V3Fb7lV`&;(A*2|yqjkE9SN@fm z_K>rdU-|JU)A6p-c8`EPPqaSAkN*dEEzV&yLjXkvY2?)D>8FGhTCA4pt+T-`nlXrsB z29~jWy-hZNcY+jhD6LcuQC$7Blf7HKVTwt`v`H1^q)6H~dPg=}L*PM~X{Jq;W@Nru zbl<*k!2#>gRb~3Uwj3jqpdVAigXc;x=s@M(UTx)YMkE0qcM5Z%V}(}_=rSEJQVBV*Zp zK|x)X(WZ9^63X?KjCXDXQTu5hqp87SH6Es7{@RG-yOm{vJh<{X6b~GU{ea%0B)_$LDRi)t}tjG5wX(;&d`D$zulh}1! z{4*@}@!c&zdPukfT{e5i)i^4x+AlcC%vGI3=c5xs5~oP-`OgbN*_=$Cspp!K?E2#X zNjyB=q-=gy*(U58At#gr-IA?udR7hRc|sa(RPog1@pY#iXB6dlx`K-{$1z30@7wG| zP0EgtF`D8bE>0Y@=u<)Y$Wo4wRC+F2l-%yN8Md`5gGvp_}c zkLWg~_otxy*X3-|_wlCQdai~AvlToB96!KJ>1RIdJ}VeelvHwGwQZLHZiV-UUs`Pl zS))b|DR_q@HR~jAgzLoFMJ5OqwEdYhkMvW)3^2Z<#xXKvV1<>67A!!m$s5V+3Zx6gTuT z8iX?_)^U9+}V|0x|@wpn)8pxNJmX zyo$qGI|UF3{IXctua8Njb^ih&C4&ytA}*%16kAm)Kde3p?FbTh+~@|-XBXs?%j!yw z)pR{P>=-1T!T(Jx@^!O~bImS;nrKXj>>+FBh;H2chb|Yi^p;rrGcJ2%mF(qs^U7~x z*kyFU7ks+n?o0MebRb>f&f2+OM4o$9+hQJh2*N^l^97;7#A?~1q>Qs`-wPr;$88!G zeGnpO7?qQ>K$}s;61WYV%lIB4=udRI-I_XjD%)z01rOYP=ewlpI>pZ=mUMRg#2!hY zYX@y_ldt-M0@C6q$gRBe#PPvT?;Vh5wlhH@>)K z_gKAm5R}_Y2P5me&A8;O4dzv%0-VJ4m|Oq)xFP+Lr5{1BJ`u_&((>(m zFb~By0b;-khT~;c6lk&FWtCd(y!ugn^k&ZTD1T@iQpeP6+T~Mr=WJ(- z?uPoVZ+2PUe#pc~HcVtc8wO0|{?M2Vxj5)q>FK-bB1s@JmTfSqnFhEXtD80Qp@ql! zh{HWf?-`%ufKN=2LrpP#!h2gaV4Kbb0V`gn*jQ%MyO(F4zQ0vwukl5H$6GZUsVc3o z231y>>1Y}81tU=$KOLEnZodxin+6rJMv^1@LcYpk_sgcj9 zenA;_*uiVit4n=Pm6DFP6t(#~M3z#1qZ<7E-S;PB-W*e_JNVqke~xDBh~5LU2`x3AjN4C% zxSHK;d`};z@s(SUl_4!`A-){J`5^nG04V*qBix;UIk`~>qiVGY9m4oi?$YsAST3^NLQj`WKtdGTL!i7Bt^7O`n6NYeV462 zAp<4gB+gX6kCF|NNu4S7h@(?3#kI6I5Bk}l#}fA7oH7f*Fz`R-s8sC#?8IEyPS&6i z`l#EoK^vC$c0^MF(=W?>u^CLAJkWO`d8b(4wD+|DUtaG001Tc*0!oRz`&s6Sq=mw$ zql^2iPaE^^#oS34OpQKN9_Vmnd&QH7i2Yhem1bEESNqJkUIR>pBY^q|_JEZ&|WyK44R_k8B; zv|lpx`!=IYrdDlEAK|@qVao+S#U*{%yR`DH3f4vpNVZ^TDY!zy^6&=raa1%^HkV1i zofS@I6!T6G^D#8wzjhCcSZ1E~i|g~j+wX3D(oPi?A3j*GpqJcS(jnh;*@c5xo|0Y5){q8ao^*2xOATrT<=bsc$(#oKng4gM6I^+J=; zac8>7lcH+y_O`(>D+f>B!1j_^igI07*^y(xK@0oKv)g! zA>|OWM*O5@rd~lq{(6-%6dWf+Rsqj#7txoV8HBlY9%A^bXaIi|ix3wwDsG z0S5jXzT#asAGL@M+~^sv*B_&+edRsQ)2F7J3+mSLN^F6*hOR&T7?lljdUUB(2|UXW zqRS1TIr}M58|8*D8-Bbw>7EhXrPJ~Z?%Q)aF5a@YKi;2}4lnBCGNvav+KO_~Z8OrY zP4DM?mPrds_X~4bTbS_Un8Gf;(??LDfPr(sB@s1XDzh%r9ql@+M3jkGHYwR_ft z7l^T;uFc|%B48;jIBT39t(dlEGXzyjmg)wARTIuhV=k|5OUBmVvMrb6L?yYC%*ysy zXbaai@91~BVAy5D4G_4})4N`zCJYV9E@oDg{yQ41m!RKVxIV`k#&t7hv+UDAVZmjD z*~*%;-?XLxwmBwO&Mlkyj?=U|P;1PpTnX{oUHB`}KRaE3T* ze}|LF*?ZDq16HgD)46wRdavAoV3sFqjt2rSlq2l)PI!oZZGPX^RFaMj6`QN_9P82> zDO?R~@7o^h{kqTKJV(%>7o73y)@Oya3RZW;J)d_ADGChzIodca|LjPMdm-qB?&k*@Al-4vEiSfd_ktj z#F}_|QOTa|j8|}gb7X$z+&8j&x;V0@9vBR!cRl)W<$`Boh$E%ECeqJYVBme^4i2~z z{BS+)&2Usi=IQTZbG$%SiHyf$SxwJh*bzO_rtP8HUv)f~6n-2_3_z~rM{8Urq&^4c z$cNCKW21#TdaJd+*}aBrG+wtKNz&26juzYIfK(Jj-s>J>l*$mp1pg z>10{K$fjO*kWst?8Zd8`w{%ynp-LmXaBfG!Y9HnL<}_6f+uS-cc3FRz*gJcB%`|C0 zQRY2SwrS>-2Ep``!vVI?vperee#f5TD=lVlCC82$biK9pL!%GR@X&axtR>cfrxs(_ zY7#A$X_7Ktm12R{|sUQGt~WjB^GFo`88aUNX^_WeSB#bfq* zZW4fS1X*y300k^vUI#Tzt7%7ejqNq}lK0lb*<^N8eB6HHK3^HbG`E>DxRZCZP7Dd* zL5!;tDs62L=MNIi4`VY`E(Nqa)btaQuI3^SFWMU-jorwFrA9_jst22~!gy(ej6RhX zu?5Cf$UNcJNtUGrf^vz4#H4h@I=f||()#KAK0zHePm>15+q0uP`V_&NWDLySO zyEyD!X6M%AS-+hs%32*fg_$N*d}R3SlHBDY%J-S^PyrVNYJFMasXeCLP*A1C2iiMWXJG692Jyjd_^(Q zv(v+3#W5iC9&_W%=bt$VhU-z9hwnE!Bm#-z6nd>?HXf)9Wi9w+!0^%MX>Tdbk%=Rc+n@8_je ziRXw2W;i{W3rU$%Nc`>PyTi@3?QplPFEsGSDTaDEWgQh9>^EKIpBDbhUA*&$R;VM* zA@^dm%(`WibW%^N7`5Y_)o|~SfqDy}TK|(R5_@x(xEzKUMvVTK(zOrQaU(qCm0usP z8-4*iRwWVd!93Mn>HEgv4m>YWO97YfzG;@!yQxg+(9sdum*t2KKkvPaf1EN&)^~f9 zV$yz?+6go^&X=uThthgQZR#_fzE1LG#JFdLeM-!2!>U0w67}uZz_~n>kWwTj3z3+q z!IM83T(QOZ35VHPV-2}5|BSf9g{Bl0f*|yuP>DYcAu0fYVo_j$zk|Nu_aRQm1PR4T@N@n7M-vtQs}m{?|CO&B{I3`ShWHy# z{Li!b!^0NwOPGtx!bRbtXs99_Eh-CBl!M60Ar z?tRX=&pprHKkokito6L>UEliF`uDBd7hx|xE24f2kk)T&A5^+wIKgQ@7!ptrP}D=g zgP5UWX5^g;#+^?@ZZ9sqGi4K^V|Q?8C;=wHUacPd$L2X~>CE`CcUKEv4UX@IL;FgC zhvVP28!69I#xXnwCT5-a62TP=se81x7=z*wgUiMCg^7{A^(R9?s-RlCu#@Qxm#~QO zm#9a^h;LzPI{~rCdCE0;IhJRdIi;^N3dV)vxioZg)q-BgX9;PM1okD&NH+q1?|n-r zI~8myQh{@dasv#=`2mim{3Vgz$N-ACUdmTp1;m@Njo>oBfv01!1JNA-s97zhGSSPh zN0Bw7b(pF6ZoguQ1G3np0u%y385+KynAmTySn#iFAXYZ&%W>Js znVF`NwHyU}R&421RfJy!TV}>`SQ?@nf;ZZo3lrJe%(~cRKlfIK|9&O5^25%V1XdjN zcylqcO@-kW$iah!h4}Su)6zbxD$!umH5=c=Y07$;>+=3oT9o+G#S$9v><6K`C8c=&d?lvrOL8pnylsqD##mW`Ze>)R8H7N$GP5{$R+>+*j~!h`fF|MU z*R*Kw8K1(~uY2$v%OrrZg{Jl(*h-IV{OeDIO&vIzGP6`!uVoM5Hub(J7K-{HmDVn+ zXbpmbxNYq!M=3Jh>iL(Q9SDZ44m&li8LYs=k7gw^qm*iel8!S@@ku2x$0%?~ac*!0MAUF(xe?sx%2ttBBlNn+oT z;P;f$+h@*OiKSbr`J<^Ktl>=9@2Hc7zi4aLPpr#l|1y>aCsDU)xafD~>eM(4%J%kx zZScJv=+<;-O8ENL?p%eNkLN*rc}`O>>JGAl-(52*1JrugoAR8UHz?vzd(!@RifR6H zAZqmv-k9BQUd6&C^Poj?nN&;O(7Q&QcEY4=4%ILpA~7L2a4pjSX0n!Ie`Dq~l>6}F zqH}%j^mu7$Kbvk!i1#e-ho@WBKhoZKFfjn>o865yp1FKV}Xl z)<6M95aXXC7z~A%En*V@00Lz@I067%Ej1Yxspn{RJX+HaOt28Tk?=8! z;v3A^Hh=t4g;q(WwUix~+?iZE23bpDG0W%A*wU0=NHG0*{80KM@WceOw z<@*pcBG;*9(xs|T>h9^j{ao$t{rt|+!GU9F^zDAo*EmKV6?Wvz3-2f(S$mfPrrH&r zeah%gTU)iWeT%FP(l$}g%%z-hx`A5=qV{>wjhh$D3)kUxo!Pl8oUwO5H<&@?b{eKu zXR<3K&KO1gybKbXt(&{SkybD<+q!3H2HE<(NJ>g{6E$E8?+p{C;t6`#an*^#%(^>W zzg}Oz|8sX{e<{M44Q$j)LePX+()WgxGVIe^22*8${?cdFu;66IETe215 zvvwI3_YZ77WXfRCG|M|wJILvm?`_I&5GJSIQ>{qHB{!zQL}TOY-yyhkBdIMS86qOA zXkw1qMs%v0?3Uw#FK6_vQG78PY(;U+%&t_MCt1h!a_l-t<4vw#r#m4`&vGziBh|bE z&LNa|PvP(eVc!Zh;!~k*Td@z-FEnu)m=<~K1RLbU2p?1*S!`e|_!QCI8o$JlfCUZc zHMzD+4ZqP$OTwz6HFxhl313JS^?x8M;W~L7hBCsNz~@!F@U~2!Lm6d0>8jXyEF-4| zW7Op;y%05Pl8V>zWcu17XVVe+WW^lxQWAe220`%T59a4rS9EhGDeE-<$GvvQGWyS2~J&~SpG??C#kzZGjqMJ zx)Ex`(f5xtU8Ptz&$ztdgmK&DQD$_EG=#F`vhO2{{a99AZyWwUWq&cCS_Gg-} zpienmHIO*d$I$-Bf?>2zx5L87fTx|O?40FKzN4XvjxgwF&3T^6S6iIYgoMQR)z&;D z>+P2Ii^%UojjdXb`8EZO=y+gnD@160A1lJSl*0+Y^mRREs%@K{c<6*EZA8?5&CiR@ z?j9p+3RZj5xZDg#I$d$kPOfmiqdbyKF73EHV-l55e*HXKx#YC>ASus@-YdfGcY!@z zJk{t2t}%9ZhtHcLjWy<%WMycKhslHdy{2Q$ftDF$sC%YHc|Z%kMhL?V)5**+R)rRq zyP1CpgM~*Ng_6!y9`ER6kF3>G5?xC(U@UN%Bi`WB9QF%)3)3~_J%65?H=fpp^^uJv z4&JCXjVF;m%_q)-JIsvtH+CCAxCWuFxId)SS-D1Hd?0V32?MRLgF>6@aENK3k0X=! zSlOL3D$^$PxYxA$QiN)pdh?h_udsKV3nQxV2hD0TDXr4>%p;!Yh9F!cNm~hIx5!-d z(Co0K6(EB|)SWTZl{L}}R!-3=1uLsf>^>>uT|QV04RL zgNoeFBFAmLj#ewQ^A{Z|VN7mY&dVNjPJ_c235-UvyM1EZWH#-J}aX)RtJf&ZvE{ zYLn~*$WW_@k4z%gpmo_8yfjRs;sNP37DKc4Q`4ip+fcNHug6)ry|fxt+qLD2Ky|A3 z)hquE>;t@qyWezl8Y_P;asrAVFr)k#^`FsKEl*@5sfC`Aw5_~u-rsdrp8usQ%##|K*Cd8%7^qTh zpe0-87>QF`-g4pd`bnP~fjQ3`I~Pcb3xdSU-vZ;)(^hs4cjZDQUS&XA#4}XN-n>?` z;9;Fof>Pl~XQ&1ZBi4)jvOq|kbo6g2!iP5>?|bU;>{>@=@A_+b2T9fj2pPbPZV5Jy zPS!X%IA8V*Z9L@CO9n9>1C>4@7DcI{?z%ti)(k;7)2DsiyqH!-g$Q>WhD}z$p?a0# zhCBOX@YzL0q`(v9R`t5q7cnUB_HmKQAV=HlvyBxW=2)hy zLI(p|Y^l8Z>FM`#o?x_TRdlPKS54w+Zgn>+W2SW6ht;_>2|{+wgsA$L z-^=K`<*_n`qHWi`6`}w!m?oL}r{Mcz+wG21U+9LPh!f*XXW6@HA5!0vm>_bZM5}TLGc*R@{T!nebaC4W;m8B_;pMZ zF2s}(#0N}<^zF?bT+fW@qVPo?A4Do@vN{HSG`>&*_sVJ0IP#$%T^&Jd+A1|;Nm?e> znJNPno_$J~J&1WTH;#$=sr2fy2y(8wxJA`ZVpykodLWVheGaulXvWTdq)qiTgcIZY zD&MKrVv~BZb8kOM1QkNua5zY{rA1}731s2nANSPx|(R(QU9`(s~O?z6|nv&@E6XODKcUNiS=pj%4&L; z>7)|1K@_72rRm<)D*3^*4JC=4I)8;t@!B&p@o>>Z?h!TPTR!5D9+;RtYxWtgHPg!K>3hhx!e4Afz(Hmw(i^cHnDC*fca})_2$d3+f+98dX^yjVB+Vw zo|386{4wubtuz%Boa##oGs#H9ABAAg_kd{IY=+B_k7~MaUU=T@n%}%t(_*>sWkglD zV>zbUHoWwcK=BMIfb|EoVmmJKkpq~-!=NRpjnmcQ^TJiFJ*J7%IBFqqN#tyIE#352WR!AY#eUw-Gt|{E&AbPD_3xHi{HoT@9j>&R zG!}P5$RbI^!|NLfk?162+u$YEv7T_GV$35&tDT?~)Obs4Wjg+R+p8JcYGPh9&8sUUJ=IODfNP04rzAwQFL@$ z$ZsoDDEF*O_)&4c<4*H)0^Wk}W=y>%jWSn<2^&{0B|7jY{luGe8u^+?3_(L|6O%7` z{BX;Miqx3e9KKlW+w-=1?4X~Z`v(xarC{#F;E_tZb519{auQOgg;~(MHQ#}Z6aI4yYA~y zj~Z|ng)!gn_!lt>=vH_>(+y18qy0F!x-^|0I?G;8GBh_>Nr7qT(_{NND()+qOC?kd zAKVa5Px}q_0)B`>dRyIizl$oxgHR)C(B2>5Nv`r`PF+3@Uw*GAC|J~f9qaM(dDbeZx zo+04;f>3iaa~_BwLVyPhMj&`Ve*}+^sTl+e6*7hkg3!O{?EjIcLoZD#$O9JA01FAj zz``&%7YKsRr|0@RT%!;AM;W6SWLW<}g&;OQ^eYDXf0D9Hgs~mKD+m?>3&Ftus*nu? zV&eevDLa_S8oMCaWrZOiFhA(O4Ky#4mcAhH&v*HEFbcq6{4hpy#{V!dQ~=8OuO&wN z|7AijsL)>-7z%>@jX_{g^db`*>TemqFbMo_wg43TR|bKA{`jQ%Gyd;%un>s4VSKmtWY8Tr($-Xj>nf@lCU zg8@GjB)~5Rhk!vaenAKrEW;0zkp+Wfg`l#sAR(}Txaj|D7y!k^fzB?*NSAj=ggH=1 OkRJwQVUbgn2mTLw6WGT9 delta 6052 zcma)=WmMH$x5fz(q(izkt<(-;Z@Q5Zq#GnQY#ISc;ZK8rfP{1if`D{MmxR(FE#1;B z&GkL^eaF4`_~D-UVU0C@^BK>Y^V^y;O5sK!RtfllAa35)JMkQpY00c9Tm~CkNG>w!F84o3Q+o z?K~i&9H!p=$bC*p$O+FiHz)3`cR}Uo6aA7GkYCS`c`?$mFwU-uU(QqM^*kp&X0xW+ zS~<6b=L3(QgoLqWUv;XW%}%N8)IL>BT8^8>;7l23YCiI?sy&|0YuKpzP=Wa}WiOyR zl&zZ~rf{qYBMYatFOe`bk3%X{QC`-LCvOFyaa~A`k&o80(oMh%{pRttZ>S5ap#aFlp`=gyYd@Jg7UrYAtB;c+%V zAn0>6E-PH4RMb<}=EN=|l>BFYm{#AKHOa^Yd3=lYF_zv6 zP#Za&y`0#+va^gGy&(}=E4H%E3xd<;jiXNO^VeUpf>ZjV)&(BhbHcwL8}Rk6?)0yI ziEtg)J8W_nY!eyk`01bgKEye6@RhkiQgF?>qY153)CcN9d(EVBL;sA)c{?`XO%@3X z`ZkURp8y}vFpWn5tI&8m>U{P`l$*hIxxy|Bk9T*rt#P@kp`z|H3tb1; ztg5aD3P0LkeNAu3FTVWFS+hZ?n2UqXaQ`FW+W4pWZC4h}RY(q%D!w_VaOvt;@vI<< z*(yJIJh%_r=wB(YoTG)4rH!LCSlY}L8SjWgO(APzWrcJ> zI$9uI4RQGezz#N!FI|nR5^zE={_{E`9ydCF{8Kz?3;RMgr^IdI9H$#tQTcBBTg zNlKIAC{Bq_l!ufD3+#ZdG>FN|H4G<<3QE2@HK2B#{APZhG^P*nSv#Ker%0@)r~AZR zIqYA(8oCl!qe>pYjcR{Bbt%__7Z224DlC8LoLDe?#Bu-h4A+*P40Px+P$^0Lzc4cglK>M)bh8&7 zdK)YwFGe&9S-ZS)e_#)a7#q3U=P@>l<*s}|Dh`5KJkAp8gnbZKIC7PB<(+pDfo4;U z=CApoKT(z3Y}`E8D!~wqhihcJw3MwD8>K`Zz-p6q*|rWPzDTr}*?cm1JeWHR&+7)(q`q3?(M^jN6M(UTtD}JF+ zP+D%R#;?xS{AE+0%`@vm^n0$I((X^N>j-W$i3E$ic0II~Wj)*x7@EATap~)2)OFCy zt3*ySRNi=>=8qfdVwd@$Jo@(%<{l>6Jz2vbt2tv{X=ipbwVZgi55(rlkz*zLqquZ4 z7jrS_&Mk`qU2xcvUBH1E1taTv+9&ljhU43 zLEW9Sf{29e@90aGfKP0z%XtGRdh-l%ZwPax7bFCnqS}h?NZP8_zpy&u^ZwjGE`zSy zuHjLm)6PD(^mH61M`OCdlXsLni&cx4Xelsss|7XteX@al)-M9Sl7pgoIVncl-kHXfuteKxBo@qp0$_)NXvfnt@`+ zY6&=zu@k}>sBN=5;7q|>utHWIA62?tTnPDp!cOhzCk@mzWREB3KSv*9r_)%Y&7ho zq9|9kV1RySOhvQS_)I?eO^bosqG%ve3+-Ed3j^s+#5J>+T2kK7s7Dq*1GAl}oGmfE} zSGL;YXK>0TlFI}GPMe5+t1wHQ=;I~LYO)PCF5uu|mXiR7{n>p&)@e@dd@YJ35~A?% zsZ2BRj}-GywzbbE>&KOdrL#$2-AzhKq|+Bv=y|fR6gMM!auH6wv$|fTGY0IOd^HRG z*`%9V#HQc&dIfAGX@*yAT=;3#6IZUvjZ8BP33}_ZOGk$dWs{%9lI{=jp}H4>YEi=# z&j8a8T%+%Lm22{{A5ul-S2+_j?N0fXj18;HPN|!u>Sqb{R!sY?GtHjoVp>NXU1w?) zz$z2{6(+P}j?6%W`Bgs>2P#{Jy}^?9Vs0HQ)9Eu~v(Fv52f^Rq2gkj~ef|OEZUTg= z2A#K}YIF`)O798Cdo{F2K%tBGyrOVQ2}f(AIqqUWxp@_oW&shgE0=o zE5gObqR94530*efgLSZ}%EHz&Mk6sLJ)iiL8KimJ;OFPzuj#Qh?3U$6XsA}tiDxDZ zg>4&f>)5L`Uwf=-ZwBC6T7D&6sXV22d0EuE^N9N0FEo|{K^kFPl>OkNzC*%ixqG6( z>AmKS@Tso?;zqfV-_|&nuGo#@)XljClB29{ea|n;zSirR7GQhx#%{Zq9~HkHzCn(E zj@pidoGzp|Zy#}3x|=0y6p7`nN7NL%HC1O*QK3QgGReCE^`IWH2ro4HtS{sT#^;v- z9JHd%avU07=z^GJzTbc4i%YqOdmSYMc8fprDKbwP9My~6<@+$=y|Ei&Ifx?TUZE|y z?4WnvHF+NQh+9}$Vzp}NuouR|(H!k8e82ZEy%ysL)R0@cejc#;Xi6>n^PW>se?@!h zpm=6pO6dA%U-CyOf||-VX8M7Jkt*k5p~1BI`T}*L5=5&Q(jW3(D;vLW98i=?0M?TO z)1s5e69Qq9={%Fc$8Yg?cj9z1-}ezOgz9b5SFaC-Id`eFdgX)x_+j|8+24^U*jnLh z8Yt-OQp~?hW@2@>af3q*>!8HOIBc0W{1>!<#~JK(Bw27yb%6eUcUj-7wNzvuh7PW( zS+JQQYh1pProoVcAlNA7vv8Au225RyL&x3F^L-CqDf{HGy>=eUwdTyKq;}*mH`Gs| zX{(X8;kh2|ikk3eFT1x*s!3NyzSe+Djl5CMC&hx%F^rtTmeNACVejitXf_HM5>naZ zYk*2!_ib=#WxI6>wWH2BEY*23r(elA%wP9Cy<;NzA@?jYl#mBvmHh#DGvbb!57wJW z*1}&?Abb_^5d*1SB&D8oS_omyy8Qx){bu|J;PHD?R3rRC`p#t>vJriYpuNJ0OoUsk+W94>z z2x_a^?Q@EE`$N=_qsB{sMB$e?&vc1TcGkVlvN_wgzXEVNXV&J#XDlIvYKO{;7DpO} z`<~XHm|H@1DYeNlm~l*BXnj2_-_JZUv1$H})lC#R9N%q^{Yr^2#$${S&Adp?EjdPY zaj`&tw*+82yCuJbIcNBD9JY4w?i}MP3o~>CRe8FryUFKFz%#gjnxmGwY@7r%^A9St zy#PsM(3eYYC!S{U!C;>x+oH`=_r4!eVy7F^)!tVdL+@!|bO^jY+AT#hb=#0S?2s=S zJ)gRxiyW18?QS5+Xe?V*4|Y4=H|;~p zmcW#Et0b;I;O;0a)_t`Y3gubPu z;2O+4?23Q|emAD4)~DEhhV;OiDE_^`m?ysP-%Ro>xp_vG8yUx8#Dxz9$%y~> z`GRBy>o{(j{!zsE*LBY^N-gcQbWSM;)#m7V#c7EGdbhyyp%~Njiw&e zNgC1ZWiisKlAQua!WN&%Y@5O<&iYQM?~E)bzuh+=*C;dRECd1qhq zT?hD=7`mv*7n!BR+Ffxm$)yKKuqGYrGrJw<_$_Qmy!9mdy=ORSB72KkQ_2*iBiq-G)XPy5fxNglBnUvw2<}uU3sv z03bBdaTB6t!d+OAeQ9!TVU}_StzErI`y}A6X$Aa%i5LSd9 z3D~4aC6*q(h9JCwOx*B*T4>2U!xM5kW&K_?$$IhsO#IqGpo=21BkI}nZCGscQt4`U z1CD%q4cp{(O+Blk+Zj%vgI=|05`Kg7S*A>HqX}LfJnl>VqHlX9j>)_5yaQZ-&5Cb6 z>A&2;EOR#$t$!aG!7dUaulQ9S|HFGZ4#w3NRgtcJ4fkJeF)?(q3;|ezq-kPAesM(* z{_bP*v^XvlBS;6yipvqV$30(}p0*ql0|@j7HZJPScpfA?AN%;A#J*g<#5UKyn_4W! z>=3NdYkdKFJw>x$JrqcMR!C`%SxxZ@H7!6o3$BahSM#}ozw`Ter(Jxp^CTEg!%do~ z1>ou$IYkJ%pJ!EymuVVSM`;GRj(JGeb-m5njoYY;jkSTBLY5C zC)^#eA7Q2Kh6s+?J}Dii24nC>S?7|08LC<=z7(u^m@JRh+D}?Zxc)kDcvQdVk+kRK z9Nh8k4zVDDUrHiT?^}fqJ@ef=LN=e5?TQW#;kYJt(Hl4ZyjA}iSc|uC_icsSKV(y} z`jS%mmf=F?c@@q%scEM8&00WqJpdTl&}F7AKIM%S*^pg%9-2%`dwDh!z8D^}S{)Oc zlz}IyT1w7bK>gl+hKjxw3jU^mvS`u+9F8RRO+uZ}Hf1FzcMk+)bF@V!SvV{`>=-+F z2`BM0HbQd6!hc59vJxMV8nH{+s_Iz1V|}c_gV7lf@3B}ko$5qma8a}Vt`~5Zytc51 zv$`<`8j1#Jau99>ajnKpg!*sk>D%Ds5ZFOHPPGzR=YDkRivyIise$ujiLP5H8? z=b7Z`%Bh$eZ3+&i`?Fns3TSny{${)f3$d7Bcz_XIZgUv%1FXSD54_t@5{NB6Ea}|6 z(2BDS|E7N@Ix$A)Wy17~IAc+j@f)){%)0eCnt#px{~pEvvjKotEz+XUS@_J%p-3Tq z2#=W=On?UpMOyMeeg}`0AkxeXfv_?cgvHy@JN_fxHN7-3j0b|yg2II%5Me$97X%uw z%E0v}e9aL4j~w|KW&V*am67S6=&I%!yU_7r5Cjwsf&6!QWlR+R7mP6Jpd;cRF*At$ z-k|&|K#m~Xzb+7jAV2i)49X9I|G{7ag7HDj9KatAPyxO_9R&IL{_X&SArOCOPy|Ha z@9eh>_ICyqfd88T|5G|t5b}o*Ch(^rAM8&CgF*f%8V-Z~ueRM>%xvtDF2A=)Ah4#5 zHxdLzfWR6~@d+%n#xQA^th6j#Ru;l9Cnzn?CnEzz$V$sdNg?=PFgZC{anb)*VT}8G bbLHx0=Hlk*g0#Z@-7^B0jZIEX9{0ZhaL)K4 diff --git a/BookGPU/Chapters/chapter13/biblio13.bib b/BookGPU/Chapters/chapter13/biblio13.bib index 0903bfe..3d6e56c 100644 --- a/BookGPU/Chapters/chapter13/biblio13.bib +++ b/BookGPU/Chapters/chapter13/biblio13.bib @@ -1,27 +1,23 @@ @article{ch13:ref1, title = {Asynchronous grid computation for {A}merican options derivatives}, -author = {Chau, M. and Couturier, R. and Bahi, J. M. and Spiteri, P.}, +author = {Chau, M. and Couturier, R. and Bahi, J.M. and Spiteri, P.}, journal = {Advances in Engineering Software}, -volume = {***-***}, -number = {***-***}, -pages = {***-***}, -note = {Online version first}, -year = {2012}, +volume = {60-61}, +pages={136--144}, +year = {2012} } -@article{ch13:ref2, -title = {Matrix iterative analysis}, -author = {Varga, R. S.}, -journal = {Prentice Hall}, -volume = {}, -number = {}, -pages = {}, -note = {}, -year = {}, +@book{ch13:ref2, + author = "Varga, R.S.", + title = "Matrix Iterative Analysis", + publisher = "Springer", + address = "Dordrecht", + series = "Springer Series in Computational Mathematics", + year = "2009", } @article{ch13:ref3, - author = {Baudet, G. M.}, + author = {Baudet, G.M.}, title = {Asynchronous iterative methods for multiprocessors}, journal = {Journal Assoc. Comput. Mach.}, volume = {25}, @@ -30,27 +26,27 @@ year = {}, year = {1978}, } -@article{ch13:ref4, - author = {Bertsekas, D. P. and Tsitsiklis, J. N.}, - title = {Parallel and distributed computation, numerical methods}, - journal = {Prentice Hall Englewood Cliffs N. J. (1989)}, - volume = {}, - number = {}, - pages = {}, +@book{ch13:ref4, + author = {Bertsekas, D.P. and Tsitsiklis, J.N.}, + title = {Parallel and Distributed Computation: Numerical Methods}, year = {1989}, -} + publisher = {Prentice-Hall, Inc.}, + address = {Upper Saddle River, NJ, USA}, +} @book{ch13:ref5, - author = {Bahi, J. M. and Contassot-Vivier, S. and Couturier, R.}, - title = {Parallel iterative algorithms: from sequential to grid computing}, - journal = {Chapman \& Hall/CRC, Numerical Analysis \& Scientific Computating 1 (2007)}, - year = {2007}, + title = {{Parallel Iterative Algorithms: from Sequential to Grid Computing}}, + author = {Bahi, J.M. and Contassot-Vivier, S. and Couturier, R.}, + publisher = {Chapman \& Hall/CRC}, + pages = {240}, + series = {Numerical Analysis \& Scientific Computing Series }, + year = {2007}, } @article{ch13:ref6, author = {Miellou, J.-C. and Spiteri, P.}, title = {Two criteria for the convergence of asynchronous iterations}, - journal = {in Computers and computing, P. Chenin et al. ed., Wiley Masson}, + journal = {in P. Chenin et al. ed., Computers and computing, Wiley Masson}, volume = {}, number = {}, pages = {91--95}, @@ -68,7 +64,7 @@ year = {}, } @article{ch13:ref8, -title = {{CUDA} Toolkit 4.2 {CUBLAS} Library}, +title = {{CUDA} {T}oolkit 4.2 {CUBLAS} {L}ibrary}, author = {NVIDIA Corporation}, journal = {}, volume = {}, @@ -78,18 +74,20 @@ note = {\url{http://developer.download.nvidia.com/compute/DevZone/docs/html/CUDA year = {2012}, } -@article{ch13:ref9, +@inproceedings{ch13:ref9, author = {Micikevicius, P.}, title = {{3D} finite difference computation on {GPUs} using {CUDA}}, - journal = {Proceedings of 2nd Workshop on General Purpose Processing on Graphics Processing Units}, - volume = {}, - number = {}, - pages = {79--84}, + booktitle = {Proceedings of 2nd Workshop on General Purpose Processing on Graphics Processing Units}, + series = {GPGPU-2}, year = {2009}, -} + pages = {79--84}, + numpages = {6}, + publisher = {ACM}, + address = {New York, NY, USA} +} @article{ch13:ref10, - author = {Leist, A. and Playne, D. P. and Hawick, K. A.}, + author = {Leist, A. and Playne, D.P. and Hawick, K.A.}, title = {Exploiting graphical processing units for data-parallel scientific applications}, journal = {Concurrency and Computation: Practice and Experience}, volume = {21}, @@ -99,7 +97,7 @@ year = {2012}, } @article{ch13:ref11, - author = {Chau, M. and Couturier, R. and Bahi, J. M. and Spiteri, P.}, + author = {Chau, M. and Couturier, R. and Bahi, J.M. and Spiteri, P.}, title = {Parallel solution of the obstacle problem in grid environments}, journal = {International Journal of High Performance Computing Applications}, volume = {25}, @@ -109,9 +107,9 @@ year = {2012}, } @article{ch13:ref12, - author = {Nvidia}, + author = {{NVIDIA}}, title = {{NVIDIA} {CUDA} {C} {P}rogramming {G}uide}, - journal = {Version 4.2 (2012)}, + journal = {Version 4.2}, volume = {}, number = {}, pages = {}, @@ -119,7 +117,7 @@ year = {2012}, } @article{ch13:ref13, - author = {Evans, D. J.}, + author = {Evans, D.J.}, title = {Parallel {S}.{O}.{R}. iterative methods}, journal = {Parallel Computing}, volume = {1}, @@ -133,7 +131,7 @@ year = {2012}, title = {Block red-black ordering method for parallel processing of {ICCG} solver}, journal = {High Performance Computing}, volume = {2327}, - number = {}, + number = {0}, pages = {}, year = {2006}, } @@ -141,7 +139,7 @@ year = {2012}, @article{ch13:ref15, title = {Iterative methods for sparse linear systems}, author = {Saad, Y.}, -journal = {Society for Industrial and Applied Mathematics, 2nd edition}, +journal = {Society for Industrial and Applied Mathematics, second edition}, volume = {}, number = {}, pages = {}, diff --git a/BookGPU/Chapters/chapter13/ch12.tex~ b/BookGPU/Chapters/chapter13/ch12.tex~ deleted file mode 100755 index 370c6c7..0000000 --- a/BookGPU/Chapters/chapter13/ch12.tex~ +++ /dev/null @@ -1,1080 +0,0 @@ -%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% -%% %% -%% CHAPTER 12 %% -%% %% -%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% - -\chapterauthor{}{} -\chapter{Solving sparse linear systems with GMRES and CG methods on GPU clusters} - -%%--------------------------%% -%% SECTION 1 %% -%%--------------------------%% -\section{Introduction} -\label{sec:01} -The sparse linear systems are used to model many scientific and industrial problems, such as the environmental simulations or -the industrial processing of the complex or non-Newtonian fluids. Moreover, the resolution of these problems often involves the -solving of such linear systems which is considered as the most expensive process in terms of time execution and memory space. -Therefore, solving sparse linear systems must be as efficient as possible in order to deal with problems of ever increasing size. - -There are, in the jargon of numerical analysis, different methods of solving sparse linear systems that we can classify in two -classes: the direct and iterative methods. However, the iterative methods are often more suitable than their counterpart, direct -methods, for solving large sparse linear systems. Indeed, they are less memory consuming and easier to parallelize on parallel -computers than direct methods. Different computing platforms, sequential and parallel computers, are used for solving sparse -linear systems with iterative solutions. Nowadays, graphics processing units (GPUs) have become attractive for solving these -linear systems, due to their computing power and their ability to compute faster than traditional CPUs. - -In Section~\ref{sec:02}, we describe the general principle of two well-known iterative methods: the conjugate gradient method and -the generalized minimal residual method. In Section~\ref{sec:03}, we give the main key points of the parallel implementation of both -methods on a cluster of GPUs. Then, in Section~\ref{sec:04}, we present the experimental results obtained on a CPU cluster and on -a GPU cluster, for solving sparse linear systems associated to matrices of different structures. Finally, in Section~\ref{sec:05}, -we apply the hypergraph partitioning technique to reduce the total communication volume between the computing nodes and, thus, to -improve the execution times of the parallel algorithms of both iterative methods. - - -%%--------------------------%% -%% SECTION 2 %% -%%--------------------------%% -\section{Krylov iterative methods} -\label{sec:02} -Let us consider the following system of $n$ linear equations in $\mathbb{R}$: -\begin{equation} -Ax=b, -\label{eq:01} -\end{equation} -where $A\in\mathbb{R}^{n\times n}$ is a sparse nonsingular square matrix, $x\in\mathbb{R}^{n}$ is the solution vector, -$b\in\mathbb{R}^{n}$ is the right-hand side and $n\in\mathbb{N}$ is a large integer number. - -The iterative methods for solving the large sparse linear system~(\ref{eq:01}) proceed by successive iterations of a same -block of elementary operations, during which an infinite number of approximate solutions $\{x_k\}_{k\geq 0}$ are computed. -Indeed, from an initial guess $x_0$, an iterative method determines at each iteration $k>0$ an approximate solution $x_k$ -which, gradually, converges to the exact solution $x^{*}$ as follows: -\begin{equation} -x^{*}=\lim\limits_{k\to\infty}x_{k}=A^{-1}b. -\label{eq:02} -\end{equation} -The number of iterations necessary to reach the exact solution $x^{*}$ is not known beforehand and can be infinite. In -practice, an iterative method often finds an approximate solution $\tilde{x}$ after a fixed number of iterations and/or -when a given convergence criterion is satisfied as follows: -\begin{equation} -\|b-A\tilde{x}\| < \varepsilon, -\label{eq:03} -\end{equation} -where $\varepsilon<1$ is the required convergence tolerance threshold. - -Some of the most iterative methods that have proven their efficiency for solving large sparse linear systems are those -called \textit{Krylov sub-space methods}~\cite{ref1}. In the present chapter, we describe two Krylov methods which are -widely used: the conjugate gradient method (CG) and the generalized minimal residual method (GMRES). In practice, the -Krylov sub-space methods are usually used with preconditioners that allow to improve their convergence. So, in what -follows, the CG and GMRES methods are used for solving the left-preconditioned sparse linear system: -\begin{equation} -M^{-1}Ax=M^{-1}b, -\label{eq:11} -\end{equation} -where $M$ is the preconditioning matrix. - -%%****************%% -%%****************%% -\subsection{CG method} -\label{sec:02.01} -The conjugate gradient method is initially developed by Hestenes and Stiefel in 1952~\cite{ref2}. It is one of the well -known iterative method for solving large sparse linear systems. In addition, it can be adapted for solving nonlinear -equations and optimization problems. However, it can only be applied to problems with positive definite symmetric matrices. - -The main idea of the CG method is the computation of a sequence of approximate solutions $\{x_k\}_{k\geq 0}$ in a Krylov -sub-space of order $k$ as follows: -\begin{equation} -x_k \in x_0 + \mathcal{K}_k(A,r_0), -\label{eq:04} -\end{equation} -such that the Galerkin condition must be satisfied: -\begin{equation} -r_k \bot \mathcal{K}_k(A,r_0), -\label{eq:05} -\end{equation} -where $x_0$ is the initial guess, $r_k=b-Ax_k$ is the residual of the computed solution $x_k$ and $\mathcal{K}_k$ the Krylov -sub-space of order $k$: \[\mathcal{K}_k(A,r_0) \equiv\text{span}\{r_0, Ar_0, A^2r_0,\ldots, A^{k-1}r_0\}.\] -In fact, CG is based on the construction of a sequence $\{p_k\}_{k\in\mathbb{N}}$ of direction vectors in $\mathcal{K}_k$ -which are pairwise $A$-conjugate ($A$-orthogonal): -\begin{equation} -\begin{array}{ll} -p_i^T A p_j = 0, & i\neq j. -\end{array} -\label{eq:06} -\end{equation} -At each iteration $k$, an approximate solution $x_k$ is computed by recurrence as follows: -\begin{equation} -\begin{array}{ll} -x_k = x_{k-1} + \alpha_k p_k, & \alpha_k\in\mathbb{R}. -\end{array} -\label{eq:07} -\end{equation} -Consequently, the residuals $r_k$ are computed in the same way: -\begin{equation} -r_k = r_{k-1} - \alpha_k A p_k. -\label{eq:08} -\end{equation} -In the case where all residuals are nonzero, the direction vectors $p_k$ can be determined so that the following recurrence -holds: -\begin{equation} -\begin{array}{lll} -p_0=r_0, & p_k=r_k+\beta_k p_{k-1}, & \beta_k\in\mathbb{R}. -\end{array} -\label{eq:09} -\end{equation} -Moreover, the scalars $\{\alpha_k\}_{k>0}$ are chosen so as to minimize the $A$-norm error $\|x^{*}-x_k\|_A$ over the Krylov -sub-space $\mathcal{K}_{k}$ and the scalars $\{\beta_k\}_{k>0}$ are chosen so as to ensure that the direction vectors are -pairwise $A$-conjugate. So, the assumption that matrix $A$ is symmetric and the recurrences~(\ref{eq:08}) and~(\ref{eq:09}) -allow to deduce that: -\begin{equation} -\begin{array}{ll} -\alpha_{k}=\frac{r^{T}_{k-1}r_{k-1}}{p_{k}^{T}Ap_{k}}, & \beta_{k}=\frac{r_{k}^{T}r_{k}}{r_{k-1}^{T}r_{k-1}}. -\end{array} -\label{eq:10} -\end{equation} - -Algorithm~\ref{alg:01} shows the main key points of the preconditioned CG method. It allows to solve the left-preconditioned -sparse linear system~(\ref{eq:11}). In this algorithm, $\varepsilon$ is the convergence tolerance threshold, $maxiter$ is the maximum -number of iterations and $(\cdot,\cdot)$ defines the dot product between two vectors in $\mathbb{R}^{n}$. At every iteration, a direction -vector $p_k$ is determined, so that it is orthogonal to the preconditioned residual $z_k$ and to the direction vectors $\{p_i\}_{i0}$ in a Krylov sub-space $\mathcal{K}_k$ as follows: -\begin{equation} -\begin{array}{ll} -x_k \in x_0 + \mathcal{K}_k(A, v_1),& v_1=\frac{r_0}{\|r_0\|_2}, -\end{array} -\label{eq:12} -\end{equation} -so that the Petrov-Galerkin condition is satisfied: -\begin{equation} -\begin{array}{ll} -r_k \bot A \mathcal{K}_k(A, v_1). -\end{array} -\label{eq:13} -\end{equation} -GMRES uses the Arnoldi process~\cite{ref5} to construct an orthonormal basis $V_k$ for the Krylov sub-space $\mathcal{K}_k$ -and an upper Hessenberg matrix $\bar{H}_k$ of order $(k+1)\times k$: -\begin{equation} -\begin{array}{ll} -V_k = \{v_1, v_2,\ldots,v_k\}, & \forall k>1, v_k=A^{k-1}v_1, -\end{array} -\label{eq:14} -\end{equation} -and -\begin{equation} -V_k A = V_{k+1} \bar{H}_k. -\label{eq:15} -\end{equation} - -Then, at each iteration $k$, an approximate solution $x_k$ is computed in the Krylov sub-space $\mathcal{K}_k$ spanned by $V_k$ -as follows: -\begin{equation} -\begin{array}{ll} -x_k = x_0 + V_k y, & y\in\mathbb{R}^{k}. -\end{array} -\label{eq:16} -\end{equation} -From both formulas~(\ref{eq:15}) and~(\ref{eq:16}) and $r_k=b-Ax_k$, we can deduce that: -\begin{equation} -\begin{array}{lll} - r_{k} & = & b - A (x_{0} + V_{k}y) \\ - & = & r_{0} - AV_{k}y \\ - & = & \beta v_{1} - V_{k+1}\bar{H}_{k}y \\ - & = & V_{k+1}(\beta e_{1} - \bar{H}_{k}y), -\end{array} -\label{eq:17} -\end{equation} -such that $\beta=\|r_0\|_2$ and $e_1=(1,0,\cdots,0)$ is the first vector of the canonical basis of $\mathbb{R}^k$. So, -the vector $y$ is chosen in $\mathbb{R}^k$ so as to minimize at best the Euclidean norm of the residual $r_k$. Consequently, -a linear least-squares problem of size $k$ is solved: -\begin{equation} -\underset{y\in\mathbb{R}^{k}}{min}\|r_{k}\|_{2}=\underset{y\in\mathbb{R}^{k}}{min}\|\beta e_{1}-\bar{H}_{k}y\|_{2}. -\label{eq:18} -\end{equation} -The QR factorization of matrix $\bar{H}_k$ is used to compute the solution of this problem by using Givens rotations~\cite{ref1,ref3}, -such that: -\begin{equation} -\begin{array}{lll} -\bar{H}_{k}=Q_{k}R_{k}, & Q_{k}\in\mathbb{R}^{(k+1)\times (k+1)}, & R_{k}\in\mathbb{R}^{(k+1)\times k}, -\end{array} -\label{eq:19} -\end{equation} -where $Q_kQ_k^T=I_k$ and $R_k$ is an upper triangular matrix. - -The GMRES method computes an approximate solution with a sufficient precision after, at most, $n$ iterations ($n$ is the size of the -sparse linear system to be solved). However, the GMRES algorithm must construct and store in the memory an orthonormal basis $V_k$ whose -size is proportional to the number of iterations required to achieve the convergence. Then, to avoid a huge memory storage, the GMRES -method must be restarted at each $m$ iterations, such that $m$ is very small ($m\ll n$), and with $x_m$ as the initial guess to the -next iteration. This allows to limit the size of the basis $V$ to $m$ orthogonal vectors. - -Algorithm~\ref{alg:02} shows the main key points of the GMRES method with restarts. It solves the left-preconditioned sparse linear -system~(\ref{eq:11}), such that $M$ is the preconditioning matrix. At each iteration $k$, GMRES uses the Arnoldi process (defined -from line~$7$ to line~$17$) to construct a basis $V_m$ of $m$ orthogonal vectors and an upper Hessenberg matrix $\bar{H}_m$ of size -$(m+1)\times m$. Then, it solves the linear least-squares problem of size $m$ to find the vector $y\in\mathbb{R}^{m}$ which minimizes -at best the residual norm (line~$18$). Finally, it computes an approximate solution $x_m$ in the Krylov sub-space spanned by $V_m$ -(line~$19$). The GMRES algorithm is stopped when the residual norm is sufficiently small ($\|r_m\|_2<\varepsilon$) and/or the maximum -number of iterations ($maxiter$) is reached. - -\begin{algorithm} - \SetLine - \linesnumbered - Choose an initial guess $x_0$\; - $convergence$ = false\; - $k = 1$\; - $r_{0} = M^{-1}(b-Ax_{0})$\; - $\beta = \|r_{0}\|_{2}$\; - \While{$\neg convergence$}{ - $v_{1} = r_{0}/\beta$\; - \For{$j=1$ \KwTo $m$}{ - $w_{j} = M^{-1}Av_{j}$\; - \For{$i=1$ \KwTo $j$}{ - $h_{i,j} = (w_{j},v_{i})$\; - $w_{j} = w_{j}-h_{i,j}v_{i}$\; - } - $h_{j+1,j} = \|w_{j}\|_{2}$\; - $v_{j+1} = w_{j}/h_{j+1,j}$\; - } - Set $V_{m}=\{v_{j}\}_{1\leq j \leq m}$ and $\bar{H}_{m}=(h_{i,j})$ a $(m+1)\times m$ upper Hessenberg matrix\; - Solve a least-squares problem of size $m$: $min_{y\in\mathrm{I\!R}^{m}}\|\beta e_{1}-\bar{H}_{m}y\|_{2}$\; - $x_{m} = x_{0}+V_{m}y_{m}$\; - $r_{m} = M^{-1}(b-Ax_{m})$\; - $\beta = \|r_{m}\|_{2}$\; - \eIf{ $(\beta<\varepsilon)$ {\bf or} $(k\geq maxiter)$}{ - $convergence$ = true\; - }{ - $x_{0} = x_{m}$\; - $r_{0} = r_{m}$\; - $k = k + 1$\; - } - } -\caption{Left-preconditioned GMRES method with restarts} -\label{alg:02} -\end{algorithm} - - -%%--------------------------%% -%% SECTION 3 %% -%%--------------------------%% -\section{Parallel implementation on a GPU cluster} -\label{sec:03} -In this section, we present the parallel algorithms of both iterative CG and GMRES methods for GPU clusters. -The implementation is performed on a GPU cluster composed of different computing nodes, such that each node -is a CPU core managed by a MPI process and equipped with a GPU card. The parallelization of these algorithms -is carried out by using the MPI communication routines between the GPU computing nodes and the CUDA programming -environment inside each node. In what follows, the algorithms of the iterative methods are called iterative -solvers. - -%%****************%% -%%****************%% -\subsection{Data partitioning} -\label{sec:03.01} -The parallel solving of the large sparse linear system~(\ref{eq:11}) requires a data partitioning between the computing -nodes of the GPU cluster. Let $p$ denotes the number of the computing nodes on the GPU cluster. The partitioning operation -consists in the decomposition of the vectors and matrices, involved in the iterative solver, in $p$ portions. Indeed, this -operation allows to assign to each computing node $i$: -\begin{itemize*} -\item a portion of size $\frac{n}{p}$ elements of each vector, -\item a sparse rectangular sub-matrix $A_i$ of size $(n,\frac{n}{p})$ and, -\item a square preconditioning sub-matrix $M_i$ of size $(\frac{n}{p},\frac{n}{p})$, -\end{itemize*} -where $n$ is the size of the sparse linear system to be solved. In the first instance, we perform a naive row-wise partitioning -(decomposition row-by-row) on the data of the sparse linear systems to be solved. Figure~\ref{fig:01} shows an example of a row-wise -data partitioning between four computing nodes of a sparse linear system (sparse matrix $A$, solution vector $x$ and right-hand -side $b$) of size $16$ unknown values. - -\begin{figure} -\centerline{\includegraphics[scale=0.35]{Chapters/chapter12/figures/partition}} -\caption{A data partitioning of the sparse matrix $A$, the solution vector $x$ and the right-hand side $b$ into four portions.} -\label{fig:01} -\end{figure} - -%%****************%% -%%****************%% -\subsection{GPU computing} -\label{sec:03.02} -After the partitioning operation, all the data involved from this operation must be transferred from the CPU memories to the GPU -memories, in order to be processed by GPUs. We use two functions of the CUBLAS library (CUDA Basic Linear Algebra Subroutines), -developed by Nvidia~\cite{ref6}: \verb+cublasAlloc()+ for the memory allocations on GPUs and \verb+cublasSetVector()+ for the -memory copies from the CPUs to the GPUs. - -An efficient implementation of CG and GMRES solvers on a GPU cluster requires to determine all parts of their codes that can be -executed in parallel and, thus, take advantage of the GPU acceleration. As many Krylov sub-space methods, the CG and GMRES methods -are mainly based on arithmetic operations dealing with vectors or matrices: sparse matrix-vector multiplications, scalar-vector -multiplications, dot products, Euclidean norms, AXPY operations ($y\leftarrow ax+y$ where $x$ and $y$ are vectors and $a$ is a -scalar) and so on. These vector operations are often easy to parallelize and they are more efficient on parallel computers when -they work on large vectors. Therefore, all the vector operations used in CG and GMRES solvers must be executed by the GPUs as kernels. - -We use the kernels of the CUBLAS library to compute some vector operations of CG and GMRES solvers. The following kernels of CUBLAS -(dealing with double floating point) are used: \verb+cublasDdot()+ for the dot products, \verb+cublasDnrm2()+ for the Euclidean -norms and \verb+cublasDaxpy()+ for the AXPY operations. For the rest of the data-parallel operations, we code their kernels in CUDA. -In the CG solver, we develop a kernel for the XPAY operation ($y\leftarrow x+ay$) used at line~$12$ in Algorithm~\ref{alg:01}. In the -GMRES solver, we program a kernel for the scalar-vector multiplication (lines~$7$ and~$15$ in Algorithm~\ref{alg:02}), a kernel for -solving the least-squares problem and a kernel for the elements updates of the solution vector $x$. - -The least-squares problem in the GMRES method is solved by performing a QR factorization on the Hessenberg matrix $\bar{H}_m$ with -plane rotations and, then, solving the triangular system by backward substitutions to compute $y$. Consequently, solving the least-squares -problem on the GPU is not interesting. Indeed, the triangular solves are not easy to parallelize and inefficient on GPUs. However, -the least-squares problem to solve in the GMRES method with restarts has, generally, a very small size $m$. Therefore, we develop -an inexpensive kernel which must be executed in sequential by a single CUDA thread. - -The most important operation in CG and GMRES methods is the sparse matrix-vector multiplication (SpMV), because it is often an -expensive operation in terms of execution time and memory space. Moreover, it requires to take care of the storage format of the -sparse matrix in the memory. Indeed, the naive storage, row-by-row or column-by-column, of a sparse matrix can cause a significant -waste of memory space and execution time. In addition, the sparsity nature of the matrix often leads to irregular memory accesses -to read the matrix nonzero values. So, the computation of the SpMV multiplication on GPUs can involve non coalesced accesses to -the global memory, which slows down even more its performances. One of the most efficient compressed storage formats of sparse -matrices on GPUs is HYB format~\cite{ref7}. It is a combination of ELLpack (ELL) and Coordinate (COO) formats. Indeed, it stores -a typical number of nonzero values per row in ELL format and remaining entries of exceptional rows in COO format. It combines -the efficiency of ELL due to the regularity of its memory accesses and the flexibility of COO which is insensitive to the matrix -structure. Consequently, we use the HYB kernel~\cite{ref8} developed by Nvidia to implement the SpMV multiplication of CG and -GMRES methods on GPUs. Moreover, to avoid the non coalesced accesses to the high-latency global memory, we fill the elements of -the iterate vector $x$ in the cached texture memory. - -%%****************%% -%%****************%% -\subsection{Data communications} -\label{sec:03.03} -All the computing nodes of the GPU cluster execute in parallel the same iterative solver (Algorithm~\ref{alg:01} or Algorithm~\ref{alg:02}) -adapted to GPUs, but on their own portions of the sparse linear system: $M^{-1}_iA_ix_i=M^{-1}_ib_i$, $0\leq i0$, the fixed point mapping $F_{\gamma}$ of the projected Richardson method\index{Iterative~method!Projected~Richardson} is defined as follows: \begin{equation} @@ -251,8 +251,8 @@ F_{\gamma}(U) & = & (F_{1,\gamma}(U),\ldots,F_{\alpha,\gamma}(U)). \\ \end{array} \label{ch13:eq:11} \end{equation} -Assume that the convex set $K=\displaystyle\prod_{i=1}^{\alpha}K_{i}$, such that $\forall i\in\{1,\ldots,\alpha\},K_i\subset E_i$ -and $K_i$ is a closed convex set. Let also $G=(G_1,\ldots,G_{\alpha})\in E$ and, for any +Assume that the convex set $K=\displaystyle\prod_{i=1}^{\alpha}K_{i}$, such that $\forall i\in\{1,\ldots,\alpha\},K_i\subset E_i$, +and $K_i$ is a closed convex set. Let also $G=(G_1,\ldots,G_{\alpha})\in E$; for any $U\in E$, $P_K(U)=(P_{K_1}(U_1),\ldots,P_{K_{\alpha}}(U_{\alpha}))$ is the projection of $U$ on $K$ where $\forall i\in\{1,\ldots,\alpha\},P_{K_i}$ is the projector from $E_i$ onto $K_i$. So, the fixed point mapping of the projected Richardson method~(\ref{ch13:eq:10})\index{Iterative~method!Projected~Richardson} @@ -267,7 +267,7 @@ $\mathcal{A}_{i,j}$ denote block matrices of $\mathcal{A}$. The parallel asynchronous iterations of the projected Richardson method for solving the obstacle problem~(\ref{ch13:eq:08}) are defined as follows: let $U^0\in E,U^0=(U^0_1,\ldots,U^0_\alpha)$ be the initial solution, then for all $p\in\mathbb{N}$, the iterate $U^{p+1}=(U^{p+1}_1,\ldots,U^{p+1}_{\alpha})$ -is recursively defined by: +is recursively defined by \begin{equation} U_i^{p+1} = \left\{ @@ -283,7 +283,7 @@ where \left\{ \begin{array}{l} \forall p\in\mathbb{N}, s(p)\subset\{1,\ldots,\alpha\}\mbox{~and~} s(p)\ne\emptyset, \\ -\forall i\in\{1,\ldots,\alpha\},\{p \ | \ i \in s(p)\}\mbox{~is denombrable}, +\forall i\in\{1,\ldots,\alpha\},\{p \ | \ i \in s(p)\}\mbox{~is enumerable}, \end{array} \right. \label{ch13:eq:14} @@ -300,27 +300,27 @@ and $\forall j\in\{1,\ldots,\alpha\}$, \end{equation} The previous asynchronous scheme\index{Asynchronous} of the projected Richardson -method models computations that are carried out in parallel without order nor +method models computations that are carried out in parallel without order or synchronization (according to the behavior of the parallel iterative method) and describes a subdomain method without overlapping. It is a general model that takes into account all possible situations of parallel computations and nonblocking message -passing. So, the synchronous iterative scheme\index{Synchronous} is defined by: +passing. So, the synchronous iterative scheme\index{Synchronous} is defined by \begin{equation} \forall j\in\{1,\ldots,\alpha\} \mbox{,~} \forall p\in\mathbb{N} \mbox{,~} \rho_j(p)=p. \label{ch13:eq:16} \end{equation} The values of $s(p)$ and $\rho_j(p)$ are defined dynamically and not explicitly by the parallel asynchronous or synchronous execution of the algorithm. Particularly, -it enables one to consider distributed computations whereby processors compute at +They allow us to consider distributed computations whereby processors compute at their own pace according to their intrinsic characteristics and computational load. The parallelism between the processors is well described by the set $s(p)$ which contains at each step $p$ the index of the components relaxed by each processor on a parallel way while the use of delayed components in~(\ref{ch13:eq:13}) permits one to model nondeterministic behavior and does not imply inefficiency of the considered distributed scheme of computation. Note that, according to~\cite{ch13:ref7}, theoretically, -each component of the vector must be relaxed an infinity of time. The choice of the -relaxed components to be used in the computational process may be guided by any criterion -and, in particular, a natural criterion is to pick-up the most recently available +each component of the vector must be relaxed an infinite number of times. The choice of the +relaxed components to be used in the computational process may be guided by any criterion, +and in particular, a natural criterion is to pickup the most recently available values of the components computed by the processors. Furthermore, the asynchronous iterations are implemented by means of nonblocking MPI communication subroutines\index{MPI~subroutines!Nonblocking} (asynchronous communications). @@ -329,7 +329,7 @@ The important property ensuring the convergence of the parallel projected Richar method, both synchronous and asynchronous algorithms, is the fact that $\mathcal{A}$ is an M-matrix. Moreover, the convergence\index{Convergence} proceeds from a result of~\cite{ch13:ref6}. Indeed, there exists a value $\gamma_0>0$, such that $\forall\gamma\in ]0,\gamma_0[$, -the parallel iterations~(\ref{ch13:eq:13}), (\ref{ch13:eq:14}) and~(\ref{ch13:eq:15}), +the parallel iterations~(\ref{ch13:eq:13}), (\ref{ch13:eq:14}), and~(\ref{ch13:eq:15}), associated to the fixed point mapping $F_\gamma$~(\ref{ch13:eq:12}), converge to the unique solution $U^{*}$ of the discretized problem. @@ -343,64 +343,64 @@ In this section, we give the main key points of the parallel implementation of t projected Richardson method, both synchronous and asynchronous versions, on a GPU cluster, for solving the nonlinear systems derived from the discretization of large obstacle problems. More precisely, each nonlinear system is solved iteratively using -the whole cluster. We use a heterogeneous CUDA/MPI programming. Indeed, the communication +the whole cluster. We use a heterogeneous CUDA and MPI programming. Indeed, the communication of data, at each iteration between the GPU computing nodes, can be either synchronous or asynchronous using the MPI communication subroutines, whereas inside each GPU node, a CUDA parallelization is performed. -\begin{figure}[!h] -\centerline{\includegraphics[scale=0.30]{Chapters/chapter13/figures/splitCPU}} -\caption{Data partitioning of a problem to be solved among $S=3\times 4$ computing nodes.} -\label{ch13:fig:01} -\end{figure} - Let $S$ denote the number of computing nodes\index{Computing~node} on the GPU cluster, where a computing node is composed of CPU core holding one MPI process and a GPU card. So, before starting computations, the obstacle problem of size $(NX\times NY\times NZ)$ -is split into $S$ parallelepipedic sub-problems, each for a node (MPI process, GPU), as +is split into $S$ parallelepipedic subproblems, each for a node (MPI process, GPU), as is shown in Figure~\ref{ch13:fig:01}. Indeed, the $NY$ and $NZ$ dimensions (according -to the $y$ and $z$ axises) of the three-dimensional problem are, respectively, split +to the $y$ and $z$ axises) of the three-dimensional problem are split, respectively, into $Sy$ and $Sz$ parts, such that $S=Sy\times Sz$. In this case, each computing node -has at most four neighboring nodes. This kind of the data partitioning reduces the data +has at most four neighboring nodes. This kind of data partitioning reduces the data exchanges at subdomain boundaries compared to a naive $z$-axis-wise partitioning. -\begin{algorithm}[!t] -Initialization of the parameters of the sub-problem\; -Allocate and fill the data in the global memory GPU\; -\For{$i=1$ {\bf to} $NbSteps$}{ - $G = \frac{1}{k}.U^0 + F$\; - Solve($A$, $U^0$, $G$, $U$, $\varepsilon$, $MaxRelax$)\; - $U^0 = U$\; -} -Copy the solution $U$ back from GPU memory\; -\caption{Parallel solving of the obstacle problem on a GPU cluster} -\label{ch13:alg:01} -\end{algorithm} +\begin{figure} +\centerline{\includegraphics[scale=0.30]{Chapters/chapter13/figures/splitCPU}} +\caption{Data partitioning of a problem to be solved among $S=3\times 4$ computing nodes.} +\label{ch13:fig:01} +\end{figure} -All the computing nodes of the GPU cluster execute in parallel the same Algorithm~\ref{ch13:alg:01} -but on different three-dimensional sub-problems of size $(NX\times ny\times nz)$. +All the computing nodes of the GPU cluster execute in parallel the Algorithm~\ref{ch13:alg:01} +on a three-dimensional subproblems of size $(NX\times ny\times nz)$. This algorithm gives the main key points for solving an obstacle problem\index{Obstacle~problem} defined in a three-dimensional domain, where $A$ is the discretization matrix, $G$ -is the right-hand side and $U$ is the solution vector. After the initialization step, +is the right-hand side, and $U$ is the solution vector. After the initialization step, all the data generated from the partitioning operation are copied from the CPU memories -to the GPU global memories, to be processed on the GPUs. Next, the algorithm uses $NbSteps$ +to the GPU global memories to be processed on the GPUs. Next, the algorithm uses $NbSteps$ time steps to solve the global obstacle problem. In fact, it uses a parallel algorithm -adapted to GPUs of the projected Richardson iterative method for solving the nonlinear +adapted to GPUs from the projected Richardson iterative method for solving the nonlinear systems\index{Nonlinear} of the obstacle problem. This function is defined by {\it Solve()} in Algorithm~\ref{ch13:alg:01}. At every time step, the initial guess $U^0$ for the iterative algorithm is set to the solution found at the previous time step. Moreover, the right-hand side $G$ is computed as follows: \[G = \frac{1}{k}.U^{prev} + F\] where $k$ is the time step, -$U^{prev}$ is the solution computed in the previous time step and each element $f(x, y, z)$ +$U^{prev}$ is the solution computed in the previous time step, and each element $f(x, y, z)$ of the vector $F$ is computed as follows: \begin{equation} f(x,y,z)=\cos(2\pi x)\cdot\cos(4\pi y)\cdot\cos(6\pi z). \label{ch13:eq:18} \end{equation} Finally, the solution $U$ of the obstacle problem is copied back from the GPU global -memories to the CPU memories. We use the communication subroutines of the CUBLAS library~\cite{ch13:ref8}\index{CUBLAS} -(CUDA Basic Linear Algebra Subroutines) for the memory allocations in the GPU (\verb+cublasAlloc()+) +memories to the CPU memories. We use the communication subroutines of the CUBLAS +(CUDA Basic Linear Algebra Subroutines) library~\cite{ch13:ref8}\index{CUBLAS} for the memory allocations in the GPU (\verb+cublasAlloc()+) and the data transfers between the CPU and its GPU: \verb+cublasSetVector()+ and \verb+cublasGetVector()+. +\begin{algorithm}[t] +Initialization of the parameters of the subproblem\; +Allocate and fill the data in the global memory GPU\; +\For{$i=1$ {\bf to} $NbSteps$}{ + $G = \frac{1}{k}.U^0 + F$\; + Solve($A$, $U^0$, $G$, $U$, $\varepsilon$, $MaxRelax$)\; + $U^0 = U$\; +} +Copy the solution $U$ back from GPU memory\; +\caption{parallel solving of the obstacle problem on a GPU cluster} +\label{ch13:alg:01} +\end{algorithm} + \begin{algorithm}[!t] $p = 0$\; $conv = false$\; @@ -414,17 +414,17 @@ and the data transfers between the CPU and its GPU: \verb+cublasSetVector()+ and $p = p + 1$\; $conv$ = Convergence($error$, $p$, $\varepsilon$, $MaxRelax$)\; } -\caption{Parallel iterative solving of the nonlinear systems on a GPU cluster ($Solve()$ function)} +\caption{parallel iterative solving of the nonlinear systems on a GPU cluster ($Solve()$ function)} \label{ch13:alg:02} \end{algorithm} -As many other iterative methods, the algorithm of the projected Richardson +As are many other iterative methods, the algorithm of the projected Richardson method\index{Iterative~method!Projected~Richardson} is based on algebraic functions operating on vectors and/or matrices, which are more efficient on parallel computers when they work on large vectors. Its parallel implementation on the GPU cluster is carried out so that the GPUs execute the vector operations as kernels and the CPUs execute the serial codes, supervise the kernel executions -and the data exchanges with the neighboring nodes\index{Neighboring~node} and +and the data exchanges with the neighboring nodes\index{Neighboring~node}, and supply the GPUs with data. Algorithm~\ref{ch13:alg:02} shows the main key points of the parallel iterative algorithm (function $Solve()$ in Algorithm~\ref{ch13:alg:01}). All the vector operations inside the main loop ({\bf repeat} ... {\bf until}) @@ -432,7 +432,7 @@ are executed by the GPU. We use the following functions of the CUBLAS library\in \begin{itemize} \item \verb+cublasDaxpy()+ to compute the difference between the solution vectors $U^{p}$ and $U^{p+1}$ computed in two successive relaxations $p$ and $p+1$ (line~$7$ in Algorithm~\ref{ch13:alg:02}), -\item \verb+cublasDnrm2()+ to perform the Euclidean norm (line~$8$) and, +\item \verb+cublasDnrm2()+ to perform the Euclidean norm (line~$8$), and \item \verb+cublasDcpy()+ for the data copy of a vector to another one in the GPU memory (lines~$3$ and~$9$). \end{itemize} @@ -441,41 +441,41 @@ depend on the resources of the GPU multiprocessor and the resource requirements of the kernel. So, if $block$ defines the size of a thread block, which must not exceed the maximum size of a thread block, then the number of thread blocks in the grid, denoted by $grid$, can be computed according to the size of the -local sub-problem as follows: \[grid = \frac{(NX\times ny\times nz)+block-1}{block}.\] +local subproblem as follows: \[grid = \frac{(NX\times ny\times nz)+block-1}{block}.\] However, when solving very large problems, the size of the thread grid can exceed -the maximum number of thread blocks that can be executed on the GPUs (up-to $65.535$ -thread blocks) and, thus, the kernel will fail to launch. Therefore, for each kernel, -we decompose the three-dimensional sub-problem into $nz$ two-dimensional slices of size +the maximum number of thread blocks that can be executed on the GPUs (upto $65.535$ +thread blocks), and thus, the kernel will fail to launch. Therefore, for each kernel, +we decompose the three-dimensional subproblem into $nz$ two-dimensional slices of size ($NX\times ny$), as is shown in Figure~\ref{ch13:fig:02}. All slices of the same kernel -are executed using {\bf for} loop by $NX\times ny$ parallel threads organized in a +are executed using a {\bf for} loop by $NX\times ny$ parallel threads organized in a two-dimensional grid of two-dimensional thread blocks, as is shown in Listing~\ref{ch13:list:01}. Each thread is in charge of $nz$ discretization points (one from each slice), accessed in the GPU memory with a constant stride $(NX\times ny)$. \begin{figure} \centerline{\includegraphics[scale=0.30]{Chapters/chapter13/figures/splitGPU}} -\caption{Decomposition of a sub-problem in a GPU into $nz$ slices.} +\caption{Decomposition of a subproblem in a GPU into $nz$ slices.} \label{ch13:fig:02} \end{figure} \begin{center} -\lstinputlisting[label=ch13:list:01,caption=Skeleton codes of a GPU kernel and a CPU function]{Chapters/chapter13/ex1.cu} +\lstinputlisting[label=ch13:list:01,caption=skeleton codes of a GPU kernel and a CPU function]{Chapters/chapter13/ex1.cu} \end{center} The function $Determine\_Bordering\_Vector\_Elements()$ (line~$5$ in Algorithm~\ref{ch13:alg:02}) determines the values of the vector elements shared at the boundaries with neighboring computing -nodes. Its main operations are defined as follows: +nodes. Its main operations are as follows: \begin{enumerate} \item define the values associated to the bordering points needed by the neighbors, \item copy the values associated to the bordering points from the GPU to the CPU, -\item exchange the values associated to the bordering points between the neighboring CPUs, -\item copy the received values associated to the bordering points from the CPU to the GPU, +\item exchange the values associated to the bordering points between the neighboring CPUs, and +\item copy the received values associated to the bordering points from the CPU to the GPU. \end{enumerate} The first operation of this function is implemented as kernels to be performed by the GPU: \begin{itemize} -\item a kernel executed by $NX\times nz$ threads to define the values associated to the bordering vector elements along $y$-axis and, -\item a kernel executed by $NX\times ny$ threads to define the values associated to the bordering vector elements along $z$-axis. +\item a kernel executed by $NX\times nz$ threads to define the values associated to the bordering vector elements along the $y$-axis, and +\item a kernel executed by $NX\times ny$ threads to define the values associated to the bordering vector elements along the $z$-axis. \end{itemize} -As mentioned before, we develop the \emph{synchronous} and \emph{asynchronous} +As mentioned previously, we develop the \emph{synchronous} and \emph{asynchronous} algorithms of the projected Richardson method. Obviously, in this scope, the synchronous\index{Synchronous} or asynchronous\index{Asynchronous} communications refer to the communications between the CPU cores (MPI processes) on the GPU cluster, @@ -487,11 +487,11 @@ and \verb+cublasGetVectorAsync()+ in the asynchronous algorithm. Moreover, we use the communication routines of the MPI library to carry out the data exchanges between the neighboring nodes. We use the following communication routines: \verb+MPI_Isend()+ and \verb+MPI_Irecv()+ to perform nonblocking\index{MPI~subroutines!Nonblocking} -sends and receptions, respectively. For the synchronous algorithm, we use the MPI +sends and receives, respectively. For the synchronous algorithm, we use the MPI routine \verb+MPI_Waitall()+ which puts the MPI process of a computing node in -blocking status until all data exchanges with neighboring nodes (sends and receptions) +blocking status until all data exchanges with neighboring nodes (sends and receives) are completed. In contrast, for the asynchronous algorithms, we use the MPI routine -\verb+MPI_Test()+ which tests the completion of a data exchange (send or reception) +\verb+MPI_Test()+ which tests the completion of a data exchange (send or receives) without putting the MPI process in blocking status\index{MPI~subroutines!Blocking}. The function $Compute\_New\_Vector\_Elements()$ (line~$6$ in Algorithm~\ref{ch13:alg:02}) @@ -510,7 +510,7 @@ u^{p+1}(x,y,z) =& \frac{1}{Center}(g(x,y,z) - (Center\cdot u^{p}(x,y,z) + \\ \end{equation} where $u^{p}(x,y,z)$ is an element of the iterate vector $U$ computed at the iteration $p$ and $g(x,y,z)$ is a vector element of the right-hand side $G$. -The scalars $Center$, $West$, $East$, $South$, $North$, $Rear$ and $Front$ +The scalars $Center$, $West$, $East$, $South$, $North$, $Rear$, and $Front$ define constant coefficients of the block matrix $A$. Figure~\ref{ch13:fig:03} shows the positions of these coefficients in a three-dimensional domain. @@ -533,6 +533,7 @@ of the projected Richardson method, which are: the matrix-vector multiplication (\verb+MV_Multiplication()+) and the vector elements updates (\verb+Vector_Updates()+). The codes of these kernels are based on that presented in Listing~\ref{ch13:list:01}. +\pagebreak \lstinputlisting[label=ch13:list:02,caption=GPU kernels of the projected Richardson method]{Chapters/chapter13/ex2.cu} \begin{figure} @@ -544,37 +545,38 @@ The codes of these kernels are based on that presented in Listing~\ref{ch13:list Each kernel is executed by $NX\times ny$ GPU threads so that $nz$ slices of $(NX\times ny)$ vector elements are computed in a {\bf for} loop. In this case, each thread is in charge of one vector element from each slice -(in total $nz$ vector elements along $z$-axis). We can notice from the +(in total $nz$ vector elements along the $z$-axis). We can notice from the formula~(\ref{ch13:eq:17}) that the computation of a vector element $u^{p+1}(x,y,z)$, by a thread at iteration $p+1$, requires seven vector elements computed at the previous iteration $p$: two vector elements in each dimension plus -the vector element at the intersection of the three axises $x$, $y$ and $z$ +the vector element at the intersection of the three axes $x$, $y$, and $z$ (see Figure~\ref{ch13:fig:04}). So, to reduce the memory accesses to the high-latency global memory, the vector elements of the current slice can be stored in the low-latency shared memories of thread blocks, as is described in~\cite{ch13:ref9}. Nevertheless, the fact that the computation of a vector -element requires only two elements in each dimension does not allow to maximize +element requires only two elements in each dimension does not allow us to maximize the data reuse from the shared memories. The computation of a slice involves in total $(bx+2)\times(by+2)$ accesses to the global memory per thread block, to fill the required vector elements in the shared memory where $bx$ and $by$ are the dimensions of a thread block. Then, in order to optimize the memory accesses on GPUs, the elements of the iterate vector $U$ are filled in the -cache texture memory (see~\cite{ch13:ref10}). In new GPU generations as Fermi +cache texture memory (see~\cite{ch13:ref10}). In new GPU hardware and software as Fermi or Kepler, the global memory accesses are always cached in L1 and L2 caches. -For example, for a given kernel, we can favour the use of the L1 cache to that +For example, for a given kernel, we can favor the use of the L1 cache to that of the shared memory by using the function \verb+cudaFuncSetCacheConfig(Kernel,cudaFuncCachePreferL1)+. So, the initial access to the global memory loads the vector elements required by the threads of a block into the cache memory (texture or L1/L2 caches). Then, all the following memory accesses read from this cache memory. In Listing~\ref{ch13:list:02}, the function \verb+fetch_double(v,i)+ is used to read from the texture memory -the $i^{th}$ element of the double-precision vector \verb+v+ (see Listing~\ref{ch13:list:03}). +the $ith$ element of the double-precision vector \verb+v+ (see Listing~\ref{ch13:list:03}). Moreover, the seven constant coefficients of matrix $A$ can be stored in the constant memory but, since they are reused $nz$ times by each thread, it is more -interesting to fill them on the low-latency registers of each thread. +efficient to fill them on the low-latency registers of each thread. -\lstinputlisting[label=ch13:list:03,caption=Memory access to the cache texture memory]{Chapters/chapter13/ex3.cu} +\pagebreak +\lstinputlisting[label=ch13:list:03,caption=memory access to the cache texture memory]{Chapters/chapter13/ex3.cu} -The function $Convergence()$ (line~$11$ in Algorithm~\ref{ch13:alg:02}) allows +The function $Convergence()$ (line~$11$ in Algorithm~\ref{ch13:alg:02}) allows us to detect the convergence of the parallel iterative algorithm and is based on the tolerance threshold\index{Convergence!Tolerance~threshold} $\varepsilon$ and the maximum number of relaxations\index{Convergence!Maximum~number~of~relaxations} @@ -598,13 +600,13 @@ conv \leftarrow true; $$ where the function $AllReduce()$ uses the MPI global reduction subroutine\index{MPI~subroutines!Global} \verb+MPI_Allreduce()+ to compute the maximal value, $maxerror$, among the local -absolute errors, $error$, of all computing nodes and $p$ (in Algorithm~\ref{ch13:alg:02}) +absolute errors, $error$, of all computing nodes, and $p$ (in Algorithm~\ref{ch13:alg:02}) is used as a counter of the local relaxations carried out by a computing node. In the asynchronous\index{Asynchronous} algorithms, the global convergence is detected when all computing nodes locally converge. For this, we use a token ring architecture around which a boolean token travels, in one direction, from a computing node to another. Starting from node $0$, the boolean token is set to $true$ by node $i$ if the local -convergence is reached or to $false$ otherwise and, then, it is sent to node $i+1$. +convergence is reached or to $false$ otherwise, and then, it is sent to node $i+1$. Finally, the global convergence is detected when node $0$ receives from its neighbor node $S-1$, in the ring architecture, a token set to $true$. In this case, node $0$ sends a stop message (end of parallel solving) to all computing nodes in the cluster. @@ -615,54 +617,48 @@ sends a stop message (end of parallel solving) to all computing nodes in the clu %%--------------------------%% \section{Experimental tests on a GPU cluster} \label{ch13:sec:05} -The GPU cluster\index{GPU~cluster} of tests, that we used in this chapter, is an $20Gbps$ +The GPU cluster\index{GPU~cluster} of tests that we used in this chapter is an $20GB/s$ Infiniband network of six machines. Each machine is a Quad-Core Xeon E5530 CPU running at $2.4$GHz. It provides a RAM memory of $12$GB with a memory bandwidth of $25.6$GB/s and it -is equipped with two Nvidia Tesla C1060 GPUs. A Tesla GPU contains in total $240$ cores +is equipped with two NVIDIA Tesla C1060 GPUs. A Tesla GPU contains in total $240$ cores running at $1.3$GHz. It provides $4$GB of global memory with a memory bandwidth of $102$GB/s, accessible by all its cores and also by the CPU through the PCI-Express 16x Gen 2.0 interface with a throughput of $8$GB/s. Hence, the memory copy operations between the GPU and the CPU are about $12$ times slower than those of the Tesla GPU memory. We have performed our simulations -on a cluster of $24$ CPU cores and on a cluster of $12$ GPUs. Figure~\ref{ch13:fig:05} describes +on a cluster of $24$ CPU cores and on a cluster of $12$ GPUs. Figure~\ref{ch12:fig:04} describes the components of the GPU cluster of tests. Linux cluster version 2.6.39 OS is installed on CPUs. C programming language is used for coding the parallel algorithms of the methods on both GPU cluster and CPU cluster. CUDA version 4.0~\cite{ch13:ref12} is used for programming GPUs, using CUBLAS library~\cite{ch13:ref8} -to deal with vector operations in GPUs and, finally, MPI functions of OpenMPI 1.3.3 are +to deal with vector operations in GPUs, and finally, MPI functions of OpenMPI 1.3.3 are used to carry out the synchronous and asynchronous communications between CPU cores. Indeed, -in our experiments, a computing node is managed by a MPI process and it is composed of +in our experiments, a computing node is managed by one MPI process and it is composed of one CPU core and one GPU card. All experimental results of the parallel projected Richardson algorithms are obtained from simulations made in double precision data. The obstacle problems to be solved are defined in constant three-dimensional domain $\Omega\subset\mathbb{R}^{3}$. The numerical -values of the parameters of the obstacle problems are: $\eta=0.2$, $c=1.1$, $f$ is computed -by formula~(\ref{ch13:eq:18}) and final time $T=0.02$. Moreover, three time steps ($NbSteps=3$) +values of the parameters of the obstacle problems are $\eta=0.2$, $c=1.1$, $f$ is computed +by formula~(\ref{ch13:eq:18}), and final time $T=0.02$. Moreover, three time steps ($NbSteps=3$) are computed with $k=0.0066$. As the discretization matrix is constant along the time steps, the convergence properties of the iterative algorithms do not change. Thus, the performance characteristics obtained with three time steps will still be valid for more time steps. The initial function $u(0,x,y,z)$ of the obstacle problem~(\ref{ch13:eq:01}) is set to $0$, with a constraint $u\geq\phi=0$. The relaxation parameter $\gamma$ used by the projected Richardson method is computed automatically thanks to the diagonal entries -of the discretization matrix. The formula and its proof can be found in~\cite{ch13:ref11}, -Section~2.3. The convergence tolerance threshold $\varepsilon$ is set to $1e$-$04$ and the +of the discretization matrix. The formula and its proof can be found in~\cite{ch13:ref11}. +The convergence tolerance threshold $\varepsilon$ is set to $1e$-$04$ and the maximum number of relaxations is limited to $10^{6}$ relaxations. Finally, the number of threads per block is set to $256$ threads, which gives, in general, good performances for most GPU applications. We have performed some tests for the execution configurations and -we have noticed that the best configuration of the $256$ threads per block is an organization +have noticed that the best configuration of the $256$ threads per block is an organization into two dimensions of sizes $(64,4)$. -\begin{figure} -\centerline{\includegraphics[scale=0.25]{Chapters/chapter13/figures/cluster}} -\caption{GPU cluster of tests composed of 12 computing nodes (six machines, each with two GPUs.} -\label{ch13:fig:05} -\end{figure} - The performance measures that we took into account are the execution times and the number of relaxations performed by the parallel iterative algorithms, both synchronous and asynchronous versions, on the GPU and CPU clusters. These algorithms are used for solving nonlinear systems -derived from the discretization of obstacle problems of sizes $256^{3}$, $512^{3}$, $768^{3}$ +derived from the discretization of obstacle problems of sizes $256^{3}$, $512^{3}$, $768^{3}$, and $800^{3}$. In Table~\ref{ch13:tab:01} and Table~\ref{ch13:tab:02}, we show the performances of the parallel synchronous and asynchronous algorithms of the projected Richardson method implemented, respectively, on a cluster of $24$ CPU cores and on a cluster of $12$ GPUs. In @@ -670,14 +666,14 @@ these tables, the execution time defines the time spent by the slowest computing number of relaxations is computed as the summation of those carried out by all computing nodes. In the sixth column of Table~\ref{ch13:tab:01} and in the eighth column of Table~\ref{ch13:tab:02}, -we give the gains in $\%$ obtained by using an asynchronous algorithm compared to a synchronous +we give the gains in percentage obtained by using an asynchronous algorithm compared to a synchronous one. We can notice that the asynchronous version on CPU and GPU clusters is slightly faster than the synchronous one for both methods. Indeed, the cluster of tests is composed of local and homogeneous nodes communicating via low-latency connections. So, in the case of distant and/or heterogeneous -nodes (or even with geographically distant clusters) the asynchronous version would be faster than +nodes (or even with geographically distant clusters), the asynchronous version would be faster than the synchronous one. However, the gains obtained on the GPU cluster are better than those obtained on the CPU cluster. In fact, the computation times are reduced by accelerating the computations on -GPUs while the communication times still unchanged. +GPUs while the communication times remain unchanged. \begin{table} \centering @@ -685,7 +681,7 @@ GPUs while the communication times still unchanged. \hline \multirow{2}{*}{\bf Pb. size} & \multicolumn{2}{c|}{\bf Synchronous} & \multicolumn{2}{c|}{\bf Asynchronous} & \multirow{2}{*}{\bf Gain\%} \\ \cline{2-5} - & $\mathbf{T_{cpu}}$ & {\bf \#relax.} & $\mathbf{T_{cpu}}$ & {\bf \#relax.} & \\ \hline \hline + & $\mathbf{T_{cpu}}$ & {\bf \# Relax.} & $\mathbf{T_{cpu}}$ & {\bf \# Relax.} & \\ \hline \hline $256^{3}$ & $575.22$ & $198,288$ & $539.25$ & $198,613$ & $6.25$ \\ \hline \hline @@ -706,7 +702,7 @@ $800^{3}$ & $222,108.09$ & $1,769,232$ & $188,790 \hline \multirow{2}{*}{\bf Pb. size} & \multicolumn{3}{c|}{\bf Synchronous} & \multicolumn{3}{c|}{\bf Asynchronous} & \multirow{2}{*}{\bf Gain\%} \\ \cline{2-7} - & $\mathbf{T_{gpu}}$ & {\bf \#relax.} & $\mathbf{\tau}$ & $\mathbf{T_{gpu}}$ & {\bf \#relax.} & $\mathbf{\tau}$ & \\ \hline \hline + & $\mathbf{T_{gpu}}$ & {\bf \# Relax.} & $\mathbf{\tau}$ & $\mathbf{T_{gpu}}$ & {\bf \# Relax.} & $\mathbf{\tau}$ & \\ \hline \hline $256^{3}$ & $29.67$ & $100,692$ & $19.39$ & $18.00$ & $94,215$ & $29.96$ & $39.33$ \\\hline \hline @@ -728,11 +724,11 @@ $\tau$ as a ratio between the execution time $T_{cpu}$ spent on the CPU cluster that $T_{gpu}$ spent on the GPU cluster: \[\tau=\frac{T_{cpu}}{T_{gpu}}.\] We can see from these ratios that solving large obstacle problems is faster on the GPU cluster than on the CPU cluster. Indeed, the GPUs are more efficient than their counterpart -CPUs to execute large data-parallel operations. In addition, the projected Richardson -method is implemented as a fixed point-based iteration and uses the Jacobi vector updates -that allow a well thread-parallelization on GPUs, such that each GPU thread is in charge +CPUs at executing large data-parallel operations. In addition, the projected Richardson +method is implemented as a fixed point based iteration and uses the Jacobi vector updates +that allow a well-suited thread-parallelization on GPUs, such that each GPU thread is in charge of one vector component at a time without being dependent on other vector components -computed by other threads. Then, this allow to exploit at best the high performance +computed by other threads. Then, this allows us to exploit at best the high performance computing of the GPUs by using all the GPU resources and avoiding the idle cores. Finally, the number of relaxations performed by the parallel synchronous algorithm @@ -747,11 +743,11 @@ consequently it also depends on the number of computing nodes. %%--------------------------%% %% SECTION 6 %% %%--------------------------%% -\section{Red-Black ordering technique} +\section{Red-black ordering technique} \label{ch13:sec:06} -As is well-known, the Jacobi method\index{Iterative~method!Jacobi} is characterized +As is wellknown, the Jacobi method\index{Iterative~method!Jacobi} is characterized by a slow convergence\index{Convergence} rate compared to some iterative methods\index{Iterative~method} -(for example Gauss-Seidel method\index{Iterative~method!Gauss-Seidel}). So, in this +(for example, Gauss-Seidel method\index{Iterative~method!Gauss-Seidel}). So, in this section, we present some solutions to reduce the execution time and the number of relaxations and, more specifically, to speed up the convergence of the parallel projected Richardson method on the GPU cluster. We propose to use the point red-black @@ -762,30 +758,30 @@ apply it to the projected Richardson method as a compromise between the Jacobi and Gauss-Seidel iterative methods. The general principle of the red-black technique is as follows. Let $t$ be the -summation of the integer $x$-, $y$- and $z$-coordinates of a vector element $u(x,y,z)$ +summation of the integer $x$-, $y$-, and $z$-coordinates of a vector element $u(x,y,z)$ on a three-dimensional domain: $t=x+y+z$. As is shown in Figure~\ref{ch13:fig:06.01}, -the red-black ordering technique consists in the parallel computing of the red -vector elements having even value $t$ by using the values of the black ones then, +the red-black ordering technique consists of the parallel computing of the red +vector elements having even value $t$ by using the values of the black ones, then the parallel computing of the black vector elements having odd values $t$ by using the new values of the red ones. This technique can be implemented on the GPU in two different manners: \begin{itemize} -\item among all launched threads ($NX\times ny$ threads), only one thread out of two computes its red or black vector element at a time or, -\item all launched threads (on average half of $NX\times ny$ threads) compute the red vector elements first and, then, the black ones. +\item among all launched threads ($NX\times ny$ threads), only one thread out of two computes its red or black vector element at a time or +\item all launched threads (on average half of $NX\times ny$ threads) compute the red vector elements first, and then the black ones. \end{itemize} However, in both solutions, for each memory transaction, only half of the memory segment addressed by a half-warp is used. So, the computation of the red and black -vector elements leads to use twice the initial number of memory transactions. Then, +vector elements leads to using twice the initial number of memory transactions. Then, we apply the point red-black ordering\index{Iterative~method!Red-Black~ordering} accordingly to the $y$-coordinate, as is shown in Figure~\ref{ch13:fig:06.02}. In this case, the vector elements having even $y$-coordinate are computed in parallel -using the values of those having odd $y$-coordinate and then vice-versa. Moreover, +using the values of those having odd $y$-coordinate and then viceversa. Moreover, in the GPU implementation of the parallel projected Richardson method (Section~\ref{ch13:sec:04}), -we have shown that a sub-problem of size $(NX\times ny\times nz)$ is decomposed into +we have shown that a subproblem of size $(NX\times ny\times nz)$ is decomposed into $nz$ grids of size $(NX\times ny)$. Then, each kernel is executed in parallel by $NX\times ny$ GPU threads, so that each thread is in charge of $nz$ vector elements -along $z$-axis (one vector element in each grid of the sub-problem). So, we propose +along the $z$-axis (one vector element in each grid of the subproblem). So, we propose to use the new values of the vector elements computed in grid $i$ to compute those of the vector elements in grid $i+1$. Listing~\ref{ch13:list:04} describes the kernel of the matrix-vector multiplication and the kernel of the vector elements updates of @@ -793,11 +789,12 @@ the parallel projected Richardson method using the red-black ordering technique. \begin{figure} \centering - \mbox{\subfigure[Red-Black ordering on x, y and z axises]{\includegraphics[width=2.3in]{Chapters/chapter13/figures/rouge-noir}\label{ch13:fig:06.01}}\quad - \subfigure[Red-Black ordering on y axis]{\includegraphics[width=2.3in]{Chapters/chapter13/figures/rouge-noir-y}\label{ch13:fig:06.02}}} -\caption{Red-Black ordering for computing the iterate vector elements in a three-dimensional space.} + \mbox{\subfigure[Red-black ordering on x, y, and z axises]{\includegraphics[width=2.3in]{Chapters/chapter13/figures/rouge-noir}\label{ch13:fig:06.01}}\quad + \subfigure[Red-black ordering on y axis]{\includegraphics[width=2.3in]{Chapters/chapter13/figures/rouge-noir-y}\label{ch13:fig:06.02}}} +\caption{Red-black ordering for computing the iterate vector elements in a three-dimensional space.} \end{figure} +\pagebreak \lstinputlisting[label=ch13:list:04,caption=GPU kernels of the projected Richardson method using the red-black technique]{Chapters/chapter13/ex4.cu} Finally, we exploit the concurrent executions between the host functions and the GPU @@ -814,8 +811,8 @@ neighboring CPUs and this in both synchronous and asynchronous cases. In Table~\ref{ch13:tab:03}, we report the execution times and the number of relaxations performed on a cluster of $12$ GPUs by the parallel projected Richardson algorithms; it -can be noted that the performances of the projected Richardson are improved by using the -point read-black ordering. We compare the performances of the parallel projected Richardson +can be noted that the performances of the projected Richardson algorithm are improved by using the +point red-black ordering. We compare the performances of the parallel projected Richardson method with and without this later ordering (Tables~\ref{ch13:tab:02} and~\ref{ch13:tab:03}). We can notice that both parallel synchronous and asynchronous algorithms are faster when they use the red-black ordering. Indeed, we can see in Table~\ref{ch13:tab:03} that the @@ -828,7 +825,7 @@ shown in Table~\ref{ch13:tab:02}. \hline \multirow{2}{*}{\bf Pb. size} & \multicolumn{2}{c|}{\bf Synchronous} & \multicolumn{2}{c|}{\bf Asynchronous} & \multirow{2}{*}{\bf Gain\%} \\ \cline{2-5} - & $\mathbf{T_{gpu}}$ & {\bf \#relax.} & $\mathbf{T_{gpu}}$ & {\bf \#relax.} & \\ \hline \hline + & $\mathbf{T_{gpu}}$ & {\bf \# Relax.} & $\mathbf{T_{gpu}}$ & {\bf \# Relax.} & \\ \hline \hline $256^{3}$ & $18.37$ & $71,988$ & $12.58$ & $67,638$ & $31.52$ \\ \hline \hline @@ -839,7 +836,7 @@ $768^{3}$ & $2,773.65$ & $590,652$ & $2,222.2 $800^{3}$ & $2,748.23$ & $638,916$ & $2,502.61$ & $592,525$ & $8.92$ \\ \hline \end{tabular} \vspace{0.5cm} -\caption{Execution times in seconds of the parallel projected Richardson method using read-black ordering technique implemented on a cluster of 12 GPUs.} +\caption{Execution times in seconds of the parallel projected Richardson method using red-black ordering technique implemented on a cluster of 12 GPUs.} \label{ch13:tab:03} \end{table} @@ -854,15 +851,15 @@ the communication time of the parallel projected Richardson algorithms on a GPU cluster. The experimental tests are carried out on a cluster composed of one to ten Tesla GPUs. We have focused on the weak scaling of both parallel, synchronous and asynchronous, algorithms using the red-black ordering technique. For this, we -have fixed the size of a sub-problem to $256^{3}$ per computing node (a CPU core +have fixed the size of a subproblem to $256^{3}$ per computing node (a CPU core and a GPU). Then, Figure~\ref{ch13:fig:07} shows the number of relaxations performed, on average, per second by a computing node. We can see from this figure that the efficiency of the asynchronous algorithm is almost stable, while that of the synchronous -algorithm decreases (down to $81\%$ in this example) with the increasing of the +algorithm decreases (down to $81\%$ in this example) with the increase in the number of computing nodes on the cluster. This is due to the fact that the ratio between the time of the computation over that of the communication is reduced when the computations are performed on GPUs. Indeed, GPUs compute faster than CPUs and -communications are more time consuming. In this context, asynchronous algorithms +communications are more time-consuming. In this context, asynchronous algorithms are more scalable than synchronous ones. So, with large scale GPU clusters, synchronous\index{Synchronous} algorithms might be more penalized by communications, as can be deduced from Figure~\ref{ch13:fig:07}. That is why we think that asynchronous\index{Asynchronous} iterative algorithms @@ -880,7 +877,7 @@ spatial discretization of three-dimensional obstacle problems. For this, we have both synchronous and asynchronous algorithms of the Richardson iterative method using a projection on a convex set. Indeed, this method uses point-based iterations of the Jacobi method that are very easy to parallelize on parallel computers. We have shown that its adapted parallel -algorithms to GPU architectures allows to exploit at best the computing power of the GPUs and +algorithms to GPU architectures allow us to exploit at best the computing power of the GPUs and to accelerate the resolution of large nonlinear systems. Consequently, the experimental results have shown that solving nonlinear systems of large obstacle problems with this method is about fifty times faster on a cluster of $12$ GPUs than on a cluster of $24$ CPU cores. Moreover, @@ -890,7 +887,7 @@ performed on the cluster of $12$ GPUs are reduced on average of $32\%$. Afterwards, the experiments have shown that the asynchronous version is slightly more efficient than the synchronous one. In fact, the computations are accelerated by using GPUs while the communication -times still unchanged. In addition, we have studied the weak-scaling in the synchronous and asynchronous +times are still unchanged. In addition, we have studied the weak-scaling in the synchronous and asynchronous cases, which has confirmed that the ratio between the computations and the communications are reduced when using a cluster of GPUs. We highlight that asynchronous iterative algorithms are more scalable than synchronous ones. Therefore, we can conclude that asynchronous iterations are well suited to diff --git a/BookGPU/Chapters/chapter13/ch13.tex~ b/BookGPU/Chapters/chapter13/ch13.tex~ deleted file mode 100755 index 65a9980..0000000 --- a/BookGPU/Chapters/chapter13/ch13.tex~ +++ /dev/null @@ -1,710 +0,0 @@ -%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% -%% %% -%% CHAPTER 13 %% -%% %% -%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% - -\chapterauthor{}{} -\newcommand{\scalprod}[2]% -{\ensuremath{\langle #1 \, , #2 \rangle}} -\chapter{Solving sparse nonlinear systems of obstacle problems on GPU clusters} - -%%--------------------------%% -%% SECTION 1 %% -%%--------------------------%% -\section{Introduction} -\label{sec:01} -The obstacle problem is one kind of free boundary problems. It allows to model, for example, an elastic membrane covering a solid obstacle. -In this case, the objective is to find an equilibrium position of this membrane constrained to be above the obstacle and which tends to minimize -its surface and/or its energy. The study of such problems occurs in many applications, for example: fluid mechanics, bio-mathematics (tumour growth -process) or financial mathematics (American or European option pricing). - -In this chapter, we focus on solutions of large obstacle problems defined in a three-dimensional domain. Particularly, the present study consists -in solving large nonlinear systems derived from the spatial discretization of these problems. Owing to the great size of such systems, in order to -reduce computation times, we proceed by solving them by parallel synchronous or asynchronous iterative algorithms. Moreover, we aim at harnessing -the computing power of GPUs to accelerate computations of these parallel algorithms. For this, we use an iterative method involving a projection -on a convex set, which is: the projected Richardson method. We choose this method among other iterative methods because it is easy to implement on -parallel computers and easy to adapt to GPU architectures. - -In Section~\ref{sec:02}, we present the mathematical model of obstacle problems then, in Section~\ref{sec:03}, we describe the general principle of -the parallel projected Richardson method. Next, in Section~\ref{sec:04}, we give the main key points of the parallel implementation of both synchronous -and asynchronous algorithms of the projected Richardson method on a GPU cluster. In Section~\ref{sec:05}, we present the performances of both parallel -algorithms obtained from simulations carried out on a CPU and GPU clusters. Finally, in Section~\ref{sec:06}, we use the read-black ordering technique -to improve the convergence and, thus, the execution times of the parallel projected Richardson algorithms on the GPU cluster. - - -%%--------------------------%% -%% SECTION 2 %% -%%--------------------------%% -\section{Obstacle problems} -\label{sec:02} -In this section, we present the mathematical model of obstacle problems defined in a three-dimensional domain. -This model is based on that presented in~\cite{ref1}. - -%%******************* -%%******************* -\subsection{Mathematical model} -\label{sec:02.01} -An obstacle problem, arising for example in mechanics or financial derivatives, consists in solving a time dependent -nonlinear equation: -\begin{equation} -\left\{ -\begin{array}{l} -\frac{\partial u}{\partial t}+b^t.\nabla u-\eta.\Delta u+c.u-f\geq 0\mbox{,~}u\geq\phi\mbox{,~a.e.w. in~}\lbrack 0,T\rbrack\times\Omega\mbox{,~}\eta>0,\\ -(\frac{\partial u}{\partial t}+b^t.\nabla u-\eta.\Delta u+c.u-f)(u-\phi)=0\mbox{,~a.e.w. in~}\lbrack 0,T\rbrack\times\Omega,\\ -u(0,x,y,z)=u_0(x,y,z),\\ -\mbox{B.C. on~}u(t,x,y,z)\mbox{~defined on~}\partial\Omega, -\end{array} -\right. -\label{eq:01} -\end{equation} -where $u_0$ is the initial condition, $c\geq 0$, $b$ and $\eta$ are physical parameters, $T$ is the final time, $u=u(t,x,y,z)$ -is an element of the solution vector $U$ to compute, $f$ is the right-hand right that could represent, for example, the external -forces, B.C. describes the boundary conditions on the boundary $\partial\Omega$ of the domain $\Omega$, $\phi$ models a constraint -imposed to $u$, $\Delta$ is the Laplacian operator, $\nabla$ is the gradient operator, a.e.w. means almost every where and ``.'' -defines the products between two scalars, a scalar and a vector or a matrix and a vector. In practice the boundary condition, -generally considered, is the Dirichlet condition (where $u$ is fixed on $\partial\Omega$) or the Neumann condition (where the -normal derivative of $u$ is fixed on $\partial\Omega$). - -The time dependent equation~(\ref{eq:01}) is numerically solved by considering an implicit or a semi-implicit time marching, -where at each time step $k$ a stationary nonlinear problem is solved: -\begin{equation} -\left\{ -\begin{array}{l} -b^t.\nabla u-\eta.\Delta u+(c+\delta).u-g\geq 0\mbox{,~}u\geq\phi\mbox{,~a.e.w. in~}\lbrack 0,T\rbrack\times\Omega\mbox{,~}\eta>0, \\ -(b^t.\nabla u-\eta.\Delta u+(c+\delta).u- g)(u-\phi)=0\mbox{,~a.e.w. in~}\lbrack 0,T\rbrack\times\Omega, \\ -\mbox{B.C. on~}u(t,x,y,z)\mbox{~defined on~}\partial\Omega, -\end{array} -\right. -\label{eq:02} -\end{equation} -where $\delta=\frac{1}{k}$ is the inverse of the time step $k$, $g=f+\delta u^{prev}$ and $u^{prev}$ is the solution computed at the -previous time step. - - -%%******************* -%%******************* -\subsection{Discretization} -\label{sec:02.02} -First, we note that the spatial discretization of the previous stationary problem~(\ref{eq:02}) does not provide a symmetric matrix, -because the convection-diffusion operator is not self-adjoint. Moreover, the fact that the operator is self-adjoint or not plays an -important role in the choice of the appropriate algorithm for solving nonlinear systems derived from the discretization of the obstacle -problem. Nevertheless, since the convection coefficients arising in the operator~(\ref{eq:02}) are constant, we can formulate the same -problem by an self-adjoint operator by performing a classical change of variables. Then, we can replace the stationary convection-diffusion -problem: -\begin{equation} -b^{t}.\nabla v-\eta.\Delta v+(c+\delta).v=g\mbox{,~a.e.w. in~}\lbrack 0,T\rbrack\times\Omega\mbox{,~}c\geq 0\mbox{,~}\delta\geq~0, -\label{eq:03} -\end{equation} -by the following stationary diffusion operator: -\begin{equation} --\eta.\Delta u+(\frac{\|b\|^{2}_{2}}{4\eta}+c+\delta).u=e^{-a}g=f, -\label{eq:04} -\end{equation} -where $b=\{b_{1},b_{2},b_{3}\}$, $\|b\|_{2}$ denotes the Euclidean norm of $b$ and $v=e^{-a}.u$ represents the general change of variables -such that $a=\frac{b^{t}(x,y,z)}{2\eta}$. Consequently, the numerical resolution of the diffusion problem (the self-adjoint operator~(\ref{eq:04})) -is done by optimization algorithms, in contrast, that of the convection-diffusion problem (non self-adjoint operator~(\ref{eq:03})) is -done by relaxation algorithms. In the case of our studied algorithm, the convergence is ensured by M-matrix property then, the performance -is linked to the magnitude of the spectral radius of the iteration matrix, which is independent of the condition number. - -Next, the three-dimensional domain $\Omega\subset\mathbb{R}^{3}$ is set to $\Omega=\lbrack 0,1\rbrack^{3}$ and discretized with an uniform -Cartesian mesh constituted by $M=m^3$ discretization points, where $m$ related to the spatial discretization step by $h=\frac{1}{m+1}$. This -is carried out by using a classical order 2 finite difference approximation of the Laplacian. So, the complete discretization of both stationary -boundary value problems~(\ref{eq:03}) and~(\ref{eq:04}) leads to the solution of a large discrete complementary problem of the following -form, when both Dirichlet or Neumann boundary conditions are used: -\begin{equation} -\left\{ -\begin{array}{l} -\mbox{Find~}U^{*}\in\mathbb{R}^{M}\mbox{~such~that} \\ -(A+\delta I)U^{*}-G\geq 0\mbox{,~}U^{*}\geq\bar{\Phi},\\ -((A+\delta I)U^{*}-G)^{T}(U^{*}-\bar{\Phi})=0,\\ -\end{array} -\right. -\label{eq:05} -\end{equation} -where $A$ is a matrix obtained after the spatial discretization by a finite difference method, $G$ is derived from the Euler first order implicit time -marching scheme and from the discretized right-hand side of the obstacle problem, $\delta$ is the inverse of the time step $k$ and $I$ is the identity -matrix. The matrix $A$ is symmetric when the self-adjoint operator is considered and nonsymmetric otherwise. - -According to the chosen discretization scheme of the Laplacian, $A$ is an M-matrix (irreducibly diagonal dominant, see~\cite{ref2}) and, consequently, -the matrix $(A+\delta I)$ is also an M-matrix. This property is important to the convergence of iterative methods. - - -%%--------------------------%% -%% SECTION 3 %% -%%--------------------------%% -\section{Parallel iterative method} -\label{sec:03} -Owing to the large size of the previous discrete complementary problem~(\ref{eq:05}), we will solve it by parallel synchronous or asynchronous iterative -algorithms (see~\cite{ref3,ref4,ref5}). In this chapter, we aim at harnessing the computing power of GPU clusters for solving these large nonlinear systems. -Then, we choose to use the projected Richardson iterative method for solving the diffusion problem~(\ref{eq:04}). Indeed, this method is based on the iterations -of the Jacobi method which are easy to parallelize on parallel computers and easy to adapt to GPU architectures. Then, according to the boundary value problem -formulation with a self-adjoint operator~(\ref{eq:04}), we can consider here the equivalent optimization problem and the fixed point mapping associated to -its solution. - -Assume that $E=\mathbb{R}^{M}$ is a Hilbert space, in which $\scalprod{.}{.}$ is the scalar product and $\|.\|$ its associated norm. So, the general fixed -point problem to be solved is defined as follows: -\begin{equation} -\left\{ -\begin{array}{l} -\mbox{Find~} U^{*} \in E \mbox{~such that} \\ -U^{*} = F(U^{*}), \\ -\end{array} -\right. -\label{eq:06} -\end{equation} -where $U\mapsto F(U)$ is an application from $E$ to $E$. - -Let $K$ be a closed convex set defined by: -\begin{equation} -K = \{U | U \geq \Phi \mbox{~everywhere in~} E\}, -\label{eq:07} -\end{equation} -where $\Phi$ is the discrete obstacle function. In fact, the obstacle problem~(\ref{eq:05}) is formulated as the following constrained optimization problem: -\begin{equation} -\left\{ -\begin{array}{l} -\mbox{Find~} U^{*} \in K \mbox{~such that} \\ -\forall V \in K, J(U^{*}) \leq J(V), \\ -\end{array} -\right. -\label{eq:08} -\end{equation} -where the cost function is given by: -\begin{equation} -J(U) = \frac{1}{2}\scalprod{\mathcal{A}.U}{U} - \scalprod{G}{U}, -\label{eq:09} -\end{equation} -in which $\scalprod{.}{.}$ denotes the scalar product in $E$, $\mathcal{A}=A+\delta I$ is a symmetric positive definite, $A$ is the discretization matrix -associated with the self-adjoint operator~(\ref{eq:04}) after change of variables. - -For any $U\in E$, let $P_K(U)$ be the projection of $U$ on $K$. For any $\gamma\in\mathbb{R}$, $\gamma>0$, the fixed point mapping $F_{\gamma}$ of the projected -Richardson method is defined as follows: -\begin{equation} -U^{*} = F_{\gamma}(U^{*}) = P_K(U^{*} - \gamma(\mathcal{A}.U^{*} - G)). -\label{eq:10} -\end{equation} -In order to reduce the computation time, the large optimization problem is solved in a numerical way by using a parallel asynchronous algorithm of the projected -Richardson method on the convex set $K$. Particularly, we will consider an asynchronous parallel adaptation of the projected Richardson method~\cite{ref6}. - -Let $\alpha\in\mathbb{N}$ be a positive integer. We consider that the space $E=\displaystyle\prod_{i=1}^{\alpha} E_i$ is a product of $\alpha$ subspaces $E_i$ -where $i\in\{1,\ldots,\alpha\}$. Note that $E_i=\mathbb{R}^{m_i}$, where $\displaystyle\sum_{i=1}^{\alpha} m_{i}=M$, is also a Hilbert space in which $\scalprod{.}{.}_i$ -denotes the scalar product and $|.|_i$ the associated norm, for all $i\in\{1,\ldots,\alpha\}$. Then, for all $u,v\in E$, $\scalprod{u}{v}=\displaystyle\sum_{i=1}^{\alpha}\scalprod{u_i}{v_i}_i$ -is the scalar product on $E$. - -Let $U\in E$, we consider the following decomposition of $U$ and the corresponding decomposition of $F_\gamma$ into $\alpha$ blocks: -\begin{equation} -\begin{array}{rcl} -U & = & (U_1,\ldots,U_{\alpha}), \\ -F_{\gamma}(U) & = & (F_{1,\gamma}(U),\ldots,F_{\alpha,\gamma}(U)). \\ -\end{array} -\label{eq:11} -\end{equation} -Assume that the convex set $K=\displaystyle\prod_{i=1}^{\alpha}K_{i}$, such that $\forall i\in\{1,\ldots,\alpha\},K_i\subset E_i$ and $K_i$ is a closed convex set. -Let also $G=(G_1,\ldots,G_{\alpha})\in E$ and, for any $U\in E$, $P_K(U)=(P_{K_1}(U_1),\ldots,P_{K_{\alpha}}(U_{\alpha}))$ is the projection of $U$ on $K$ where $\forall i\in\{1,\ldots,\alpha\},P_{K_i}$ -is the projector from $E_i$ onto $K_i$. So, the fixed point mapping of the projected Richardson method~(\ref{eq:10}) can be written in the following way: -\begin{equation} -\forall U\in E\mbox{,~}\forall i\in\{1,\ldots,\alpha\}\mbox{,~}F_{i,\gamma}(U) = P_{K_i}(U_i - \gamma(\mathcal{A}_i.U - G_i)). -\label{eq:12} -\end{equation} -Note that $\displaystyle\mathcal{A}_i.U= \sum_{j=1}^{\alpha}\mathcal{A}_{i,j}.U_j$, where $\mathcal{A}_{i,j}$ denote block matrices of $\mathcal{A}$. - -The parallel asynchronous iterations of the projected Richardson method for solving the obstacle problem~(\ref{eq:08}) are defined as follows: let $U^0\in E,U^0=(U^0_1,\ldots,U^0_\alpha)$ be -the initial solution, then for all $p\in\mathbb{N}$, the iterate $U^{p+1}=(U^{p+1}_1,\ldots,U^{p+1}_{\alpha})$ is recursively defined by: -\begin{equation} -U_i^{p+1} = -\left\{ -\begin{array}{l} -F_{i,\gamma}(U_1^{\rho_1(p)}, \ldots, U_{\alpha}^{\rho_{\alpha}(p)}) \mbox{~if~} i\in s(p), \\ -U_i^p \mbox{~otherwise}, \\ -\end{array} -\right. -\label{eq:13} -\end{equation} -where -\begin{equation} -\left\{ -\begin{array}{l} -\forall p\in\mathbb{N}, s(p)\subset\{1,\ldots,\alpha\}\mbox{~and~} s(p)\ne\emptyset, \\ -\forall i\in\{1,\ldots,\alpha\},\{p \ | \ i \in s(p)\}\mbox{~is denombrable}, -\end{array} -\right. -\label{eq:14} -\end{equation} -and $\forall j\in\{1,\ldots,\alpha\}$, -\begin{equation} -\left\{ -\begin{array}{l} -\forall p\in\mathbb{N}, \rho_j(p)\in\mathbb{N}, 0\leq\rho_j(p)\leq p\mbox{~and~}\rho_j(p)=p\mbox{~if~} j\in s(p),\\ -\displaystyle\lim_{p\to\infty}\rho_j(p) = +\infty.\\ -\end{array} -\right. -\label{eq:15} -\end{equation} - -The previous asynchronous scheme of the projected Richardson method models computations that are carried out in parallel -without order nor synchronization (according to the behavior of the parallel iterative method) and describes a subdomain -method without overlapping. It is a general model that takes into account all possible situations of parallel computations -and non-blocking message passing. So, the synchronous iterative scheme is defined by: -\begin{equation} -\forall j\in\{1,\ldots,\alpha\} \mbox{,~} \forall p\in\mathbb{N} \mbox{,~} \rho_j(p)=p. -\label{eq:16} -\end{equation} -The values of $s(p)$ and $\rho_j(p)$ are defined dynamically and not explicitly by the parallel asynchronous or synchronous -execution of the algorithm. Particularly, it enables one to consider distributed computations whereby processors compute at -their own pace according to their intrinsic characteristics and computational load. The parallelism between the processors is -well described by the set $s(p)$ which contains at each step $p$ the index of the components relaxed by each processor on a -parallel way while the use of delayed components in~(\ref{eq:13}) permits one to model nondeterministic behavior and does not -imply inefficiency of the considered distributed scheme of computation. Note that, according to~\cite{ref7}, theoretically, -each component of the vector must be relaxed an infinity of time. The choice of the relaxed components to be used in the -computational process may be guided by any criterion and, in particular, a natural criterion is to pick-up the most recently -available values of the components computed by the processors. Furthermore, the asynchronous iterations are implemented by -means of non-blocking MPI communication subroutines (asynchronous communications). - -The important property ensuring the convergence of the parallel projected Richardson method, both synchronous and asynchronous -algorithms, is the fact that $\mathcal{A}$ is an M-matrix. Moreover, the convergence proceeds from a result of~\cite{ref6}. -Indeed, there exists a value $\gamma_0>0$, such that $\forall\gamma\in ]0,\gamma_0[$, the parallel iterations~(\ref{eq:13}), -(\ref{eq:14}) and~(\ref{eq:15}), associated to the fixed point mapping $F_\gamma$~(\ref{eq:12}), converge to the unique solution -$U^{*}$ of the discretized problem. - - -%%--------------------------%% -%% SECTION 4 %% -%%--------------------------%% -\section{Parallel implementation on a GPU cluster} -\label{sec:04} -In this section, we give the main key points of the parallel implementation of the projected Richardson method, both synchronous -and asynchronous versions, on a GPU cluster, for solving the nonlinear systems derived from the discretization of large obstacle -problems. More precisely, each nonlinear system is solved iteratively using the whole cluster. We use a heteregeneous CUDA/MPI -programming. Indeed, the communication of data, at each iteration between the GPU computing nodes, can be either synchronous -or asynchronous using the MPI communication subroutines, whereas inside each GPU node, a CUDA parallelization is performed. - -\begin{figure}[!h] -\centerline{\includegraphics[scale=0.30]{Chapters/chapter13/figures/splitCPU}} -\caption{Data partitioning of a problem to be solved among $S=3\times 4$ computing nodes.} -\label{fig:01} -\end{figure} - -Let $S$ denote the number of computing nodes on the GPU cluster, where a computing node is composed of CPU core holding one MPI -process and a GPU card. So, before starting computations, the obstacle problem of size $(NX\times NY\times NZ)$ is split into $S$ -parallelepipedic sub-problems, each for a node (MPI process, GPU), as is shown in Figure~\ref{fig:01}. Indeed, the $NY$ and $NZ$ -dimensions (according to the $y$ and $z$ axises) of the three-dimensional problem are, respectively, split into $Sy$ and $Sz$ parts, -such that $S=Sy\times Sz$. In this case, each computing node has at most four neighboring nodes. This kind of the data partitioning -reduces the data exchanges at subdomain boundaries compared to a naive $z$-axis-wise partitioning. - -\begin{algorithm}[!t] -\SetLine -\linesnumbered -Initialization of the parameters of the sub-problem\; -Allocate and fill the data in the global memory GPU\; -\For{$i=1$ {\bf to} $NbSteps$}{ - $G = \frac{1}{k}.U^0 + F$\; - Solve($A$, $U^0$, $G$, $U$, $\varepsilon$, $MaxRelax$)\; - $U^0 = U$\; -} -Copy the solution $U$ back from GPU memory\; -\caption{Parallel solving of the obstacle problem on a GPU cluster} -\label{alg:01} -\end{algorithm} - -All the computing nodes of the GPU cluster execute in parallel the same Algorithm~\ref{alg:01} but on different three-dimensional -sub-problems of size $(NX\times ny\times nz)$. This algorithm gives the main key points for solving an obstacle problem defined in -a three-dimensional domain, where $A$ is the discretization matrix, $G$ is the right-hand side and $U$ is the solution vector. After -the initialization step, all the data generated from the partitioning operation are copied from the CPU memories to the GPU global -memories, to be processed on the GPUs. Next, the algorithm uses $NbSteps$ time steps to solve the global obstacle problem. In fact, -it uses a parallel algorithm adapted to GPUs of the projected Richardson iterative method for solving the nonlinear systems of the -obstacle problem. This function is defined by {\it Solve()} in Algorithm~\ref{alg:01}. At every time step, the initial guess $U^0$ -for the iterative algorithm is set to the solution found at the previous time step. Moreover, the right-hand side $G$ is computed -as follows: \[G = \frac{1}{k}.U^{prev} + F\] where $k$ is the time step, $U^{prev}$ is the solution computed in the previous time -step and each element $f(x, y, z)$ of the vector $F$ is computed as follows: -\begin{equation} -f(x,y,z)=\cos(2\pi x)\cdot\cos(4\pi y)\cdot\cos(6\pi z). -\label{eq:18} -\end{equation} -Finally, the solution $U$ of the obstacle problem is copied back from the GPU global memories to the CPU memories. We use the -communication subroutines of the CUBLAS library~\cite{ref8} (CUDA Basic Linear Algebra Subroutines) for the memory allocations in -the GPU (\verb+cublasAlloc()+) and the data transfers between the CPU and its GPU: \verb+cublasSetVector()+ and \verb+cublasGetVector()+. - -\begin{algorithm}[!t] - \SetLine - \linesnumbered - $p = 0$\; - $conv = false$\; - $U = U^{0}$\; - \Repeat{$(conv=true)$}{ - Determine\_Bordering\_Vector\_Elements($U$)\; - Compute\_New\_Vector\_Elements($A$, $G$, $U$)\; - $tmp = U^{0} - U$\; - $error = \|tmp\|_{2}$\; - $U^{0} = U$\; - $p = p + 1$\; - $conv$ = Convergence($error$, $p$, $\varepsilon$, $MaxRelax$)\; - } -\caption{Parallel iterative solving of the nonlinear systems on a GPU cluster ($Solve()$ function)} -\label{alg:02} -\end{algorithm} - -As many other iterative methods, the algorithm of the projected Richardson method is based on algebraic functions operating on vectors -and/or matrices, which are more efficient on parallel computers when they work on large vectors. Its parallel implementation on the GPU -cluster is carried out so that the GPUs execute the vector operations as kernels and the CPUs execute the serial codes, supervise the -kernel executions and the data exchanges with the neighboring nodes and supply the GPUs with data. Algorithm~\ref{alg:02} shows the -main key points of the parallel iterative algorithm (function $Solve()$ in Algorithm~\ref{alg:01}). All the vector operations inside -the main loop ({\bf repeat} ... {\bf until}) are executed by the GPU. We use the following functions of the CUBLAS library: -\begin{itemize*} -\item \verb+cublasDaxpy()+ to compute the difference between the solution vectors $U^{p}$ and $U^{p+1}$ computed in two successive relaxations -$p$ and $p+1$ (line~$7$ in Algorithm~\ref{alg:02}), -\item \verb+cublasDnrm2()+ to perform the Euclidean norm (line~$8$) and, -\item \verb+cublasDcpy()+ for the data copy of a vector to another one in the GPU memory (lines~$3$ and~$9$). -\end{itemize*} - -The dimensions of the grid and blocks of threads that execute a given kernel depend on the resources of the GPU multiprocessor and the -resource requirements of the kernel. So, if $block$ defines the size of a thread block, which must not exceed the maximum size of a thread -block, then the number of thread blocks in the grid, denoted by $grid$, can be computed according to the size of the local sub-problem -as follows: \[grid = \frac{(NX\times ny\times nz)+block-1}{block}.\] However, when solving very large problems, the size of the thread -grid can exceed the maximum number of thread blocks that can be executed on the GPUs (up-to $65.535$ thread blocks) and, thus, the kernel -will fail to launch. Therefore, for each kernel, we decompose the three-dimensional sub-problem into $nz$ two-dimensional slices of size -($NX\times ny$), as is shown in Figure~\ref{fig:02}. All slices of the same kernel are executed using {\bf for} loop by $NX\times ny$ parallel -threads organized in a two-dimensional grid of two-dimensional thread blocks, as is shown in Listing~\ref{list:01}. Each thread is in charge -of $nz$ discretization points (one from each slice), accessed in the GPU memory with a constant stride $(NX\times ny)$. - -\begin{figure} -\centerline{\includegraphics[scale=0.30]{Chapters/chapter13/figures/splitGPU}} -\caption{Decomposition of a sub-problem in a GPU into $nz$ slices.} -\label{fig:02} -\end{figure} - -\begin{center} -\lstinputlisting[label=list:01,caption=Skeleton codes of a GPU kernel and a CPU function]{Chapters/chapter13/ex1.cu} -\end{center} -The function $Determine\_Bordering\_Vector\_Elements()$ (line~$5$ in Algorithm~\ref{alg:02}) determines the values of the vector -elements shared at the boundaries with neighboring computing nodes. Its main operations are defined as follows: -\begin{enumerate*} -\item define the values associated to the bordering points needed by the neighbors, -\item copy the values associated to the bordering points from the GPU to the CPU, -\item exchange the values associated to the bordering points between the neighboring CPUs, -\item copy the received values associated to the bordering points from the CPU to the GPU, -\end{enumerate*} -The first operation of this function is implemented as kernels to be performed by the GPU: -\begin{itemize*} -\item a kernel executed by $NX\times nz$ threads to define the values associated to the bordering vector elements along $y$-axis and, -\item a kernel executed by $NX\times ny$ threads to define the values associated to the bordering vector elements along $z$-axis. -\end{itemize*} -As mentioned before, we develop the \emph{synchronous} and \emph{asynchronous} algorithms of the projected Richardson method. Obviously, -in this scope, the synchronous or asynchronous communications refer to the communications between the CPU cores (MPI processes) on the -GPU cluster, in order to exchange the vector elements associated to subdomain boundaries. For the memory copies between a CPU core and -its GPU, we use the synchronous communication routines of the CUBLAS library: \verb+cublasSetVector()+ and \verb+cublasGetVector()+ -in the synchronous algorithm and the asynchronous ones: \verb+cublasSetVectorAsync()+ and \verb+cublasGetVectorAsync()+ in the -asynchronous algorithm. Moreover, we use the communication routines of the MPI library to carry out the data exchanges between the neighboring -nodes. We use the following communication routines: \verb+MPI_Isend()+ and \verb+MPI_Irecv()+ to perform non-blocking sends and receptions, -respectively. For the synchronous algorithm, we use the MPI routine \verb+MPI_Waitall()+ which puts the MPI process of a computing node -in blocking status until all data exchanges with neighboring nodes (sends and receptions) are completed. In contrast, for the asynchronous -algorithms, we use the MPI routine \verb+MPI_Test()+ which tests the completion of a data exchange (send or reception) without putting the -MPI process in blocking status. - -The function $Compute\_New\_Vector\_Elements()$ (line~$6$ in Algorithm~\ref{alg:02}) computes, at each iteration, the new elements -of the iterate vector $U$. Its general code is presented in Listing~\ref{list:01} (CPU function). The iterations of the projected -Richardson method, based on those of the Jacobi method, are defined as follows: -\begin{equation} -\begin{array}{ll} -u^{p+1}(x,y,z) =& \frac{1}{Center}(g(x,y,z) - (Center\cdot u^{p}(x,y,z) + \\ -& West\cdot u^{p}(x-h,y,z) + East\cdot u^{p}(x+h,y,z) + \\ -& South\cdot u^{p}(x,y-h,z) + North\cdot u^{p}(x,y+h,z) + \\ -& Rear\cdot u^{p}(x,y,z-h) + Front\cdot u^{p}(x,y,z+h))), -\end{array} -\label{eq:17} -\end{equation} -where $u^{p}(x,y,z)$ is an element of the iterate vector $U$ computed at the iteration $p$ and $g(x,y,z)$ is a vector element of the -right-hand side $G$. The scalars $Center$, $West$, $East$, $South$, $North$, $Rear$ and $Front$ define constant coefficients of the -block matrix $A$. Figure~\ref{fig:03} shows the positions of these coefficients in a three-dimensional domain. - -\begin{figure} -\centerline{\includegraphics[scale=0.35]{Chapters/chapter13/figures/matrix}} -\caption{Matrix constant coefficients in a three-dimensional domain.} -\label{fig:03} -\end{figure} - -The kernel implementations of the projected Richardson method on GPUs uses a perfect fine-grain multithreading parallelism. Since the -projected Richardson algorithm is implemented as a fixed point method, each kernel is executed by a large number of GPU threads such -that each thread is in charge of the computation of one element of the iterate vector $U$. Moreover, this method uses the vector elements -updates of the Jacobi method, which means that each thread $i$ computes the new value of its element $u_{i}^{p+1}$ independently of the -new values $u_{j}^{p+1}$, where $j\neq i$, of those computed in parallel by other threads at the same iteration $p+1$. Listing~\ref{list:02} -shows the GPU implementations of the main kernels of the projected Richardson method, which are: the matrix-vector multiplication -(\verb+MV_Multiplication()+) and the vector elements updates (\verb+Vector_Updates()+). The codes of these kernels are based on -that presented in Listing~\ref{list:01}. - -\lstinputlisting[label=list:02,caption=GPU kernels of the projected Richardson method]{Chapters/chapter13/ex2.cu} - -\begin{figure} -\centerline{\includegraphics[scale=0.3]{Chapters/chapter13/figures/points3D}} -\caption{Computation of a vector element with the projected Richardson method.} -\label{fig:04} -\end{figure} - -Each kernel is executed by $NX\times ny$ GPU threads so that $nz$ slices of $(NX\times ny)$ vector elements are computed in -a {\bf for} loop. In this case, each thread is in charge of one vector element from each slice (in total $nz$ vector elements -along $z$-axis). We can notice from the formula~(\ref{eq:17}) that the computation of a vector element $u^{p+1}(x,y,z)$, by -a thread at iteration $p+1$, requires seven vector elements computed at the previous iteration $p$: two vector elements in -each dimension plus the vector element at the intersection of the three axises $x$, $y$ and $z$ (see Figure~\ref{fig:04}). -So, to reduce the memory accesses to the high-latency global memory, the vector elements of the current slice can be stored -in the low-latency shared memories of thread blocks, as is described in~\cite{ref9}. Nevertheless, the fact that the computation -of a vector element requires only two elements in each dimension does not allow to maximize the data reuse from the shared memories. -The computation of a slice involves in total $(bx+2)\times(by+2)$ accesses to the global memory per thread block, to fill the -required vector elements in the shared memory where $bx$ and $by$ are the dimensions of a thread block. Then, in order to optimize -the memory accesses on GPUs, the elements of the iterate vector $U$ are filled in the cache texture memory (see~\cite{ref10}). -In new GPU generations as Fermi or Kepler, the global memory accesses are always cached in L1 and L2 caches. For example, for -a given kernel, we can favour the use of the L1 cache to that of the shared memory by using the function \verb+cudaFuncSetCacheConfig(Kernel,cudaFuncCachePreferL1)+. -So, the initial access to the global memory loads the vector elements required by the threads of a block into the cache memory -(texture or L1/L2 caches). Then, all the following memory accesses read from this cache memory. In Listing~\ref{list:02}, the -function \verb+fetch_double(v,i)+ is used to read from the texture memory the $i^{th}$ element of the double-precision vector -\verb+v+ (see Listing~\ref{list:03}). Moreover, the seven constant coefficients of matrix $A$ can be stored in the constant memory -but, since they are reused $nz$ times by each thread, it is more interesting to fill them on the low-latency registers of each thread. - -\lstinputlisting[label=list:03,caption=Memory access to the cache texture memory]{Chapters/chapter13/ex3.cu} - -The function $Convergence()$ (line~$11$ in Algorithm~\ref{alg:02}) allows to detect the convergence of the parallel iterative algorithm -and is based on the tolerance threshold $\varepsilon$ and the maximum number of relaxations $MaxRelax$. We take into account the number -of relaxations since that of iterations cannot be computed in the asynchronous case. Indeed, a relaxation is the update~(\ref{eq:13}) of -a local iterate vector $U_i$ according to $F_i$. Then, counting the number of relaxations is possible in both synchronous and asynchronous -cases. On the other hand, an iteration is the update of at least all vector components with $F_i$. - -In the synchronous algorithm, the global convergence is detected when the maximal value of the absolute error, $error$, is sufficiently small -and/or the maximum number of relaxations, $MaxRelax$, is reached, as follows: -$$ -\begin{array}{l} -error=\|U^{p}-U^{p+1}\|_{2}; \\ -AllReduce(error,\hspace{0.1cm}maxerror,\hspace{0.1cm}MAX); \\ -\text{if}((maxerror<\varepsilon)\hspace{0.2cm}\text{or}\hspace{0.2cm}(p\geq MaxRelax)) \\ -conv \leftarrow true; -\end{array} -$$ -where the function $AllReduce()$ uses the MPI reduction subroutine \verb+MPI_Allreduce()+ to compute the maximal value, $maxerror$, among the -local absolute errors, $error$, of all computing nodes and $p$ (in Algorithm~\ref{alg:02}) is used as a counter of the local relaxations carried -out by a computing node. In the asynchronous algorithms, the global convergence is detected when all computing nodes locally converge. For this, -we use a token ring architecture around which a boolean token travels, in one direction, from a computing node to another. Starting from node $0$, -the boolean token is set to $true$ by node $i$ if the local convergence is reached or to $false$ otherwise and, then, it is sent to node $i+1$. -Finally, the global convergence is detected when node $0$ receives from its neighbor node $S-1$, in the ring architecture, a token set to $true$. -In this case, node $0$ sends a stop message (end of parallel solving) to all computing nodes in the cluster. - - -%%--------------------------%% -%% SECTION 5 %% -%%--------------------------%% -\section{Experimental tests on a GPU cluster} -\label{sec:05} -The GPU cluster of tests, that we used in this chapter, is an $20Gbps$ Infiniband network of six machines. Each machine is a Quad-Core Xeon -E5530 CPU running at $2.4$GHz. It provides a RAM memory of $12$GB with a memory bandwidth of $25.6$GB/s and it is equipped with two Nvidia -Tesla C1060 GPUs. A Tesla GPU contains in total $240$ cores running at $1.3$GHz. It provides $4$GB of global memory with a memory bandwidth -of $102$GB/s, accessible by all its cores and also by the CPU through the PCI-Express 16x Gen 2.0 interface with a throughput of $8$GB/s. -Hence, the memory copy operations between the GPU and the CPU are about $12$ times slower than those of the Tesla GPU memory. We have performed -our simulations on a cluster of $24$ CPU cores and on a cluster of $12$ GPUs. Figure~\ref{fig:05} describes the components of the GPU cluster -of tests. - -\begin{figure} -\centerline{\includegraphics[scale=0.25]{Chapters/chapter13/figures/cluster}} -\caption{GPU cluster of tests composed of 12 computing nodes (six machines, each with two GPUs.} -\label{fig:05} -\end{figure} - -Linux cluster version 2.6.39 OS is installed on CPUs. C programming language is used for coding the parallel algorithms of the methods on both -GPU cluster and CPU cluster. CUDA version 4.0~\cite{ref12} is used for programming GPUs, using CUBLAS library~\cite{ref8} to deal with vector -operations in GPUs and, finally, MPI functions of OpenMPI 1.3.3 are used to carry out the synchronous and asynchronous communications between -CPU cores. Indeed, in our experiments, a computing node is managed by a MPI process and it is composed of one CPU core and one GPU card. - -All experimental results of the parallel projected Richardson algorithms are obtained from simulations made in double precision data. The obstacle -problems to be solved are defined in constant three-dimensional domain $\Omega\subset\mathbb{R}^{3}$. The numerical values of the parameters of the -obstacle problems are: $\eta=0.2$, $c=1.1$, $f$ is computed by formula~(\ref{eq:18}) and final time $T=0.02$. Moreover, three time steps ($NbSteps=3$) -are computed with $k=0.0066$. As the discretization matrix is constant along the time steps, the convergence properties of the iterative algorithms -do not change. Thus, the performance characteristics obtained with three time steps will still be valid for more time steps. The initial function -$u(0,x,y,z)$ of the obstacle problem~(\ref{eq:01}) is set to $0$, with a constraint $u\geq\phi=0$. The relaxation parameter $\gamma$ used by the -projected Richardson method is computed automatically thanks to the diagonal entries of the discretization matrix. The formula and its proof can be -found in~\cite{ref11}, Section~2.3. The convergence tolerance threshold $\varepsilon$ is set to $1e$-$04$ and the maximum number of relaxations is -limited to $10^{6}$ relaxations. Finally, the number of threads per block is set to $256$ threads, which gives, in general, good performances for -most GPU applications. We have performed some tests for the execution configurations and we have noticed that the best configuration of the $256$ -threads per block is an organization into two dimensions of sizes $(64,4)$. - -\begin{table}[!h] -\centering -\begin{tabular}{|c|c|c|c|c|c|} -\hline -\multirow{2}{*}{\bf Pb. size} & \multicolumn{2}{c|}{\bf Synchronous} & \multicolumn{2}{c|}{\bf Asynchronous} & \multirow{2}{*}{\bf Gain\%} \\ \cline{2-5} - - & $\mathbf{T_{cpu}}$ & {\bf \#relax.} & $\mathbf{T_{cpu}}$ & {\bf \#relax.} & \\ \hline \hline - -$256^{3}$ & $575.22$ & $198,288$ & $539.25$ & $198,613$ & $6.25$ \\ \hline \hline - -$512^{3}$ & $19,250.25$ & $750,912$ & $18,237.14$ & $769,611$ & $5.26$ \\ \hline \hline - -$768^{3}$ & $206,159.44$ & $1,635,264$ & $183,582.60$ & $1,577,004$ & $10.95$ \\ \hline \hline - -$800^{3}$ & $222,108.09$ & $1,769,232$ & $188,790.04$ & $1,701,735$ & $15.00$ \\ \hline -\end{tabular} -\vspace{0.5cm} -\caption{Execution times in seconds of the parallel projected Richardson method implemented on a cluster of 24 CPU cores.} -\label{tab:01} -\end{table} - -\begin{table}[!h] -\centering -\begin{tabular}{|c|c|c|c|c|c|c|c|} -\hline -\multirow{2}{*}{\bf Pb. size} & \multicolumn{3}{c|}{\bf Synchronous} & \multicolumn{3}{c|}{\bf Asynchronous} & \multirow{2}{*}{\bf Gain\%} \\ \cline{2-7} - - & $\mathbf{T_{gpu}}$ & {\bf \#relax.} & $\mathbf{\tau}$ & $\mathbf{T_{gpu}}$ & {\bf \#relax.} & $\mathbf{\tau}$ & \\ \hline \hline - -$256^{3}$ & $29.67$ & $100,692$ & $19.39$ & $18.00$ & $94,215$ & $29.96$ & $39.33$ \\ \hline \hline - -$512^{3}$ & $521.83$ & $381,300$ & $36.89$ & $425.15$ & $347,279$ & $42.89$ & $18.53$ \\ \hline \hline - -$768^{3}$ & $4,112.68$ & $831,144$ & $50.13$ & $3,313.87$ & $750,232$ & $55.40$ & $19.42$ \\ \hline \hline - -$800^{3}$ & $3,950.87$ & $899,088$ & $56.22$ & $3,636.57$ & $834,900$ & $51.91$ & $7.95$ \\ \hline -\end{tabular} -\vspace{0.5cm} -\caption{Execution times in seconds of the parallel projected Richardson method implemented on a cluster of 12 GPUs.} -\label{tab:02} -\end{table} - -The performance measures that we took into account are the execution times and the number of relaxations performed by the parallel iterative algorithms, -both synchronous and asynchronous versions, on the GPU and CPU clusters. These algorithms are used for solving nonlinear systems derived from the discretization -of obstacle problems of sizes $256^{3}$, $512^{3}$, $768^{3}$ and $800^{3}$. In Table~\ref{tab:01} and Table~\ref{tab:02}, we show the performances -of the parallel synchronous and asynchronous algorithms of the projected Richardson method implemented, respectively, on a cluster of $24$ CPU cores -and on a cluster of $12$ GPUs. In these tables, the execution time defines the time spent by the slowest computing node and the number of relaxations -is computed as the summation of those carried out by all computing nodes. - -In the sixth column of Table~\ref{tab:01} and in the eighth column of Table~\ref{tab:02}, we give the gains in $\%$ obtained by using an -asynchronous algorithm compared to a synchronous one. We can notice that the asynchronous version on CPU and GPU clusters is slightly faster -than the synchronous one for both methods. Indeed, the cluster of tests is composed of local and homogeneous nodes communicating via low-latency -connections. So, in the case of distant and/or heterogeneous nodes (or even with geographically distant clusters) the asynchronous version -would be faster than the synchronous one. However, the gains obtained on the GPU cluster are better than those obtained on the CPU cluster. -In fact, the computation times are reduced by accelerating the computations on GPUs while the communication times still unchanged. - -The fourth and seventh columns of Table~\ref{tab:02} show the relative gains obtained by executing the parallel algorithms on the cluster -of $12$ GPUs instead on the cluster of $24$ CPU cores. We compute the relative gain $\tau$ as a ratio between the execution time $T_{cpu}$ -spent on the CPU cluster over that $T_{gpu}$ spent on the GPU cluster: \[\tau=\frac{T_{cpu}}{T_{gpu}}.\] We can see from these ratios that -solving large obstacle problems is faster on the GPU cluster than on the CPU cluster. Indeed, the GPUs are more efficient than their -counterpart CPUs to execute large data-parallel operations. In addition, the projected Richardson method is implemented as a fixed point-based -iteration and uses the Jacobi vector updates that allow a well thread-parallelization on GPUs, such that each GPU thread is in charge -of one vector component at a time without being dependent on other vector components computed by other threads. Then, this allow to exploit -at best the high performance computing of the GPUs by using all the GPU resources and avoiding the idle cores. - -Finally, the number of relaxations performed by the parallel synchronous algorithm is different in the CPU and GPU versions, because the number -of computing nodes involved in the GPU cluster and in the CPU cluster is different. In the CPU case, $24$ computing nodes ($24$ CPU cores) are -considered, whereas in the GPU case, $12$ computing nodes ($12$ GPUs) are considered. As the number of relaxations depends on the domain decomposition, -consequently it also depends on the number of computing nodes. - - -%%--------------------------%% -%% SECTION 6 %% -%%--------------------------%% -\section{Red-Black ordering technique} -\label{sec:06} -As is well-known, the Jacobi method is characterized by a slow convergence rate compared to some iterative methods (for example Gauss-Seidel method). -So, in this section, we present some solutions to reduce the execution time and the number of relaxations and, more specifically, to speed up the -convergence of the parallel projected Richardson method on the GPU cluster. We propose to use the point red-black ordering technique to accelerate -the convergence. This technique is often used to increase the parallelism of iterative methods for solving linear systems~\cite{ref13,ref14,ref15}. -We apply it to the projected Richardson method as a compromise between the Jacobi and Gauss-Seidel iterative methods. - -The general principle of the red-black technique is as follows. Let $t$ be the summation of the integer $x$-, $y$- and $z$-coordinates of a vector -element $u(x,y,z)$ on a three-dimensional domain: $t=x+y+z$. As is shown in Figure~\ref{fig:06.01}, the red-black ordering technique consists in the -parallel computing of the red vector elements having even value $t$ by using the values of the black ones then, the parallel computing of the black -vector elements having odd values $t$ by using the new values of the red ones. - -\begin{figure} -\centering - \mbox{\subfigure[Red-black ordering on x, y and z axises]{\includegraphics[width=2.3in]{Chapters/chapter13/figures/rouge-noir}\label{fig:06.01}}\quad - \subfigure[Red-black ordering on y axis]{\includegraphics[width=2.3in]{Chapters/chapter13/figures/rouge-noir-y}\label{fig:06.02}}} -\caption{Red-black ordering for computing the iterate vector elements in a three-dimensional space.} -\end{figure} - -This technique can be implemented on the GPU in two different manners: -\begin{itemize*} -\item among all launched threads ($NX\times ny$ threads), only one thread out of two computes its red or black vector element at a time or, -\item all launched threads (on average half of $NX\times ny$ threads) compute the red vector elements first and, then, the black ones. -\end{itemize*} -However, in both solutions, for each memory transaction, only half of the memory segment addressed by a half-warp is used. So, the computation of the -red and black vector elements leads to use twice the initial number of memory transactions. Then, we apply the point red-black ordering accordingly to -the $y$-coordinate, as is shown in Figure~\ref{fig:06.02}. In this case, the vector elements having even $y$-coordinate are computed in parallel using -the values of those having odd $y$-coordinate and then vice-versa. Moreover, in the GPU implementation of the parallel projected Richardson method (Section~\ref{sec:04}), -we have shown that a sub-problem of size $(NX\times ny\times nz)$ is decomposed into $nz$ grids of size $(NX\times ny)$. Then, each kernel is executed -in parallel by $NX\times ny$ GPU threads, so that each thread is in charge of $nz$ vector elements along $z$-axis (one vector element in each grid of -the sub-problem). So, we propose to use the new values of the vector elements computed in grid $i$ to compute those of the vector elements in grid $i+1$. -Listing~\ref{list:04} describes the kernel of the matrix-vector multiplication and the kernel of the vector elements updates of the parallel projected -Richardson method using the red-black ordering technique. - -\lstinputlisting[label=list:04,caption=GPU kernels of the projected Richardson method using the red-black technique]{Chapters/chapter13/ex4.cu} - -Finally, we exploit the concurrent executions between the host functions and the GPU kernels provided by the GPU hardware and software. In fact, the kernel -launches are asynchronous (when this environment variable is not disabled on the GPUs), such that the control is returned to the host (MPI process) before -the GPU has completed the requested task (kernel)~\cite{ref12}. Therefore, all the kernels necessary to update the local vector elements, $u(x,y,z)$ where -$068+r8%aSx zl#&MF^{sQ>=XrkT{GRz|?)l7j=H5GV&YXKYRDY?&so*SJ7$j+)E?QN_KBNDVvbX*y zKkB{Or|5CD{OfSaz#}r$Tm72W7W<~!spIXSh;cHW8h%})-RBOqVVDx!*3j1Z z<{CPEM^<~%?p4aZzvOA$2jf&W+EVDE_5y&J1>+{0lK%C}YEwViLb@#+RiV<`$CH_Y zc)dZUK8Sk^-7K#C+3@Yhb3vB%mswY-fQr-oi3B1aNjAPAAdLfN$i{!vXM(F;IoFJh zkI4c~s9wfRw>$=vA7L4KPB-_sG>!7vHRNG2BlQgp>aLPyF4dV2*?U4_@d7j}2^NUSFI8jku{q`!f+#d*S zjO_4?)a_}ZL(~Jh9h~)6Zpi+ypGA5a?I+St6hRRV@q8fDixcC z;q`F_{r4uS_KLvTj2h@j4I6n@JBMUC6Gum2)jK<8JIcha3SbqyUqHQc0$o(0WRN;k ztmb)3v2P)DN%_sqb&(ipUI1TTuO4-K9e7_~lvvyIi=IKIm}z_$I#HY*U6ibwoXlJz zB4eavVh-fX(@l0{Mmw}Mh+La`tL|><>438g21j-|sItG^kGvKdSfVod4Zmgf-n1G! zAtG$#v`&Z4gqp|m@s%{A;|ff*@=U61bqWgl1vG$$e*GYg8eoTqvF7U-N1B2=p4JUh zpn=;b9D}84I7RQ46!YF%onfmyqY+0FM$v^4m`WQgoK?j(0;9X5z40ris>_y*5Enyt ztMx@|Qn;GkjfjwBRp#@TACp=6=VJ**;zl!Ai(!`WHx^{Dl_#Tyl(0ehLwwg3GLI+p z{0m=l_tj?Ch4oJWfCzApBgCT&cU4E_MQ2BBCj(f?fEfyEL9j*miG9RqRg4JXC)K$4 zQsxBp%Qqt6J9HjGeI%oBK5ixy4&7(?^G2wu4odWgSugp#t}&h09x{4D)jh7Ga?p4E zcCIVYSrHj|w5$_`=o~sW}Fsb6o-Pz6g z-0`R&JGuE@sqO--7ldBvpc8}iq5{yfm~dgkFt z>oCQ`+SJ^qg&32m8Uw592L;r!yEIqQhC|6pkKCJg-r=8%E7v|~Wwj&4PII3=UoBmt zW-}5#ri|7g3IPom-kjuc7I$fB-_l)YQXW%U8Hp)E8`dNc8e#`b<>SL!3Fy6`cfnp^h^#Irtfv9ZzBm?$t(yP!+ zIa5lA7B-ZhMRm>KUVn`L`DArvW6T>L=_YG>L>n0S@RY0iJD~@okj;G)?}@w_-{-fL zcPOqO>Wy`>QRJ>2pM5zCxZHMp;u|@fja)zd_T!zuP%hGXVKA>_lZj2bAuF@_i4P8j z^S^U-y1jh#>C?sB-j_p;tSqZvvK9n!6Qlleud(kqE#@_b@HjHbL!m$vaivm21Yxzv ztu9haE0W+hVrRR1Uz1hlbv;DIc-^*J%^z!0m-nYy-NZw&D%R{VJLN9xQ$wqV?(^Xg z8RU2i6|46a;r&iH2m8~yGjPU<&Y|34Pd<(8;j57zG1=R<<$feug zaUOvIS_)M|XSlr~ZgP(yj(|yilf_DZuXf$GG)C;DWpF7ffKKwvk zUk3jVla{`%@9T=btz&JEzMW7(MneXMf#48ibvKy{5t))HN?S)U;SME=2qKXXNcpwe z6Oc*zU(8IioNz&7h?A6%kd^(9OP_SKfuBB|J!2Lp4@|#H6p|FdSF~*?(=zNBT3iEA z6Q%W~B(kBnCC|gN98bU**yIyH!%j%FtmzXB3JQu?p<;Sm6LDyvSQ}e5kvW;^+-e;z zV^2OWbKY|av&U?E?;JNCUvBTbiO3}mdb82-K($E#=b(@+Z7Gm;WvN}4^8rh4LwYUl z9Mg&M9y0=6n9g!l#E~&dNr+g*eJ9#!rea}K7|Wu$6;+Td;n&CD5P8GxG1TSLa@%mt z)@0e$V%De3%*>{ZcUPu93vYNLjl0jEJ<*NK!zwnZBsT@GD`q6C@%H%IIHYn(MDk0! z$$Qu=;6zaZ3|)Ii>UlRp!*D(zW$ckDy4}4 zx~s(?P%kt^8q&vaa`Wj$hyb7A4f@~TY}1;;Lh7|g^R%lpE3qV;?Y*~F!BtFq@=fti zdjmd@BcJ)lJXWmb^j`(71cBpuDWVXZWxaQLag7ly26sQu>T(*59FK64bA)T%%9`Ac zKwzX~b9b_jA&f9B*kx#vwfF9t57(vkg_V|HMo_v?1}4(no{MQZ0&O(HKQ`=_YN%Oi zm-M=yN{C>=uz+&?CFxCyafjH|rCT>2b;i&;lx8!97gkUyL{B6JBxQ5hBZ)? zy~m%Y;#o(3^jy!T^6)*q-W#Yc`3ohsjOjHQlbnnfRyl)YX={I&J~uEweR#$O4!wd) zL%Q6;VDx7eha15Trh=6(4A5f26+}(-I1!6B-cL6vU@87EL47`!ueOGBZJw+`m3Vh1 zGxd*EhYabCPKt)SZ9i&LYn=_6Q3C2m)a#dQ{bE4fDAPfZ80*Hm9n`bYojCw zP1T9@>;a@M)z~F$LrnZDd+^%&I)NACFw?gvH%zo*k%*e>9J9{{;kR2SKPk4SaebMI zsXQPw)yk?e&*;Z){g~pJ;Y2q!ZT@*-=KM9qP21?hDjMnQJPQIsH0Fy?cCg`HE|U07KyazOa=%oCx^2!Z3#6 z%V34xT{V&K*rJe`)%MN&f%Reekc!s{9QV?7GhGH(UIR_Y5-jOi8R9d8aj>i{4U>ZJ zXwZAmP{7+Sp7 zt4BjisV>RTaEewlks1k@8S$>2-UNs#%;o{S(Zbf~NNM7f*uOK=H30imEulH5rBa|3 z(VCcMlfKNpR3!dtlgz8+A4GU7ex$ zk{1AO>~|(WSHHBTtGnbsgU(6a6!O_seUQ{@t$6!My0OMd&c z>PF+4Q2FQ%A9%1RY;>&hWe}GRVD}Ol9CybkS~V-npUA>WfxHfY6I=LYQ)BfEAstG$ z;)AL4J{ak;$b+dq-#rA%OpXtv3|)%JxhSV`j7B64XE0Q1r_)oodN1&a z@EYqJIm%V8`H0|qa(6?)qw}!$;T)>8%I=*cSdI!h{zbng+R?DHk~@CLvZIE{7#0_B zY>?whhQvEhvI31U+>)>i6`noc$kyT17sRw=W838`wnr*XOZ&~ZxjZ!^#sr}DjC856 z|6s&?l#R|)P$))Ezg^;`2bQ;f_0n*|wX8hggno)C_$@a$1$x-oN`YHpVyIz|`OOKbD=CIT_EBhB|Gy z=sVNV^Y?Yp^Y(m`I=@wXP-Uv+8DwBZyj##PTVN^IpA*a()l1?W)Fv>v80gn6!|MWT zv$QCfYQh`fX2IIkc%sie%T60Fnv*o&3vWz~4aWjnm|N^JWEOSZ$(ESXy3XtD5M=%b z%!DQ`o(Lw)>ZTZU8gEKuceY2ZO(sZ@6r%&&b!W7t_ZFij zKCc22Dt)wa>?Befw; zxkZX>GGfao^OwFSPguONr;Vw%S7>ZBRw(o<>HGSyzP4C1KRsJNGi<_(Z#`(I@~TV_ zInSY(9!%$a>;6|P*Tc!+e4ACCEe}UEwsA#As=@F0sL09>^@@anXL%xqseB|k4|*QZ zcpS9igl|PVGbW+wOYYscG%;RX5WOfFM!rZG*Z|pXMg#d;@olDP_Ywkeq$i)y%YyR7 z=^e7l9rmju_=)OAp5*ybP8@P z^}OOM7P1?!wH_Z`*pVYOP_1G{+0sqf&j7?h@7bEtZ9bc0Hrb^+`;6dzei7nqKg~I- zPtHboSsTnQaUfVDg>eumb|zgT*by6wSQ;z>)1;N#Xnr24z zM*`krgv8-B%omRVNrPc~&!PS82I)R>r{TqlXTKbe=F-KCG$(R(pbEjk`a8WtpEpeh zOwD!m^^MI;^-ycmIt7tm21Biv@*p_md^Y5vd+4eDfEP<`HK>`@ye-nmh2rwr!-nt< z4$Q+4R{R4reGdudW(=W(gI!z3@OdN0kTgN;bgTb`tAXp=eIFpd)|*H@KS{gBqpQzz z;*`6aTY+=liq4-~Le;nVi8oR{RGq$^wOk2rZD?zZRrG!ZQqx%R8 zd~ml#X7{H|RN{RpRlJP%`IW_07(z!NhL;lI$2$ zfrQRah2l=Hxk1iBk{ym*F6_WPF^VxwHR}2X{>hl^oz@?j)PJJ zeP91_mw6H>G)7vi;m5T?XT#Nud*6*G%cBFO6ix-4zY2_CTzi;aGwa!-=^MKeFzPf< zyi0Y`yzj7unnNk?%tNA3Uv%84c%^cJJRjkg2#&Jef^EVkmA7r@8;3`*XBXp(PXY*s zDcg7<<+A>4PdRfS?}TCVhm@`2wLkRE?HL-|9V>)76_)wed!9EFeJv`&Ii+|y$o;r& zz1Htr;8zh8sKZ9bLc_yVDM>KmJnDehc(-(a^kyB&;Kaev3yUAYh-c4T1w6uc64a6P zl%<=TYDtQDDs6MMn!DuEgtp-zr?9Q```RJnol8%YdhQG{c^ET_`CB%+(gt0yZx4^FUXMIe_BaX2aM~F*cM4<>98&L=ZjTQy}9-=Jga6DR2=U)N$sctP$=_M?a`WL3(7!Ap zB@usSP#F9#3=Wq3%U&oL`4{(4#P7c6Pws!`h5Q$j{GZsNVEDh`|1jA9!Qe1Scmg30 z9mDVL>yKGULZRXbf;`d|;_^s13IT>7kWe@jhLVtk%1a=jisBH2k|G3-gvd(&zX}V0 btSrFY!y4=1gGFNi2m}-X;N?}+Rs#G3_QF3H delta 5238 zcma)&byO5wyT$1kl^8k%L>L+-U|1 zS`Htfr7B@4Gg?nzt@}Vhtp91Vrx3?bs!_f8vR%wbi7T;FGaL@7dd{7l$imc2de{j& zvylMuh^R>boI2~TpV&W=P_<)&HGO-VILzh^Az-6)@gI?mNX_x#vou^!^Hai_ZhXw} zyrfK}BFXQWk-IXx2WM3nh!~P54D2p;jc13xq8qn#d1O+Rr%I<eim3L78 zC~S>3x(YPl;ZXD@cM6z07(4fNN=U-e-QB;f&1u_}KDNej+tn?-D2sbmd6bp(Ftehy z$uKk2u%!cZ!Y8U;IK)?}+Ap;2^ZFVmM`%G!*+4b7P3yPtj{qs@(t4j;I+=%UZDZ!H zxqju9Lqm4TpWUket=Zsf*W$v*?asAcO{OYR+CioZk^=UnA_~0 zPF;I9WN6IqnZ%6h7<4QPpC@i%0CU!OsjHV=l*3s!o`ug5-%fP1e>2!t73y%eJU7Rp z_WB0LIiuXpuq~vpk(Q%HifN1^x63tCHv#lzi#%f>NI!MGCkbTmWg_PSA3ZdSeMYf( zMK!Yb%(&NBfDkE^tgAiGW4MssKvibKZ)diw&+viEtSqrhLmFs1Ebjc8Wu^<*7@#@A zzPk_ao%wyMaro>Fn=Boe?7`9$@$_QVvx!F}Uop#&>DIV;98#^R;sa5;h{m!VuiNMC z;3VWR?4-=wbUmX`%;M6IrzcKp*l1CydC*$9fMehD@(6k*9`-mndyzt7BbO|#IPXqT z&;zb3B6dYJeq(|B;^AHPUW?)!1EoPD>$I;P13@ z##;^&PZFCA@mjNU#_k+L{1C?cHP0+30kLaYM{ek&93gv=Z;9yZJ>ernte@@Y;-+Xz zCq?;$tLz`%3uS}BF(PXoC=BCIYB!XX9AQv#qY3*zlboq<2 z)U7*$Xlqd`k%;ON0Fa%@=PW(>S2eM8hwSW0>o?g-m5@ey(Xd0Uf^zd17Xp%31n~l_V%TNGCXHc+AG64JKwZvo(J;VAH^J%u5it zw-U-Qf^wB-i565h*y1mKFO5Mw4w)`ZJk;>Lt<1H7d+gBp(H|RsZrfV~cTSyo(j@J{ zD|qN!n>*X72Anp=z8kIf(kU1ckwP^JbC{vu`06Epf9Gq%%=tD8-2i7Qd^Zk6*XGH} zcLWcHC4Du=x`%dB+sQ^W{I)6_-j`08<@l)$cfGW_erEr(I(ay@!(%v4id|}Gxpw;N z%A0|gI|tKe2NR5pi2a6_pY|};7YFO>>-z)!EH4y2FsTEISfxB$8%7yvdAWDuc-5=T zq3FR#YxbCx$ekoRgBg_{#SLDG##i-bR}6qJPY3KJ9CkPPlA2zw3MOWNQ-b&#J%>AF z_MY=?ebJFqL(7x}PuXUNAI!`+P1#FYFi3sto10eHo7(Yi956|6D>sUO#8Z95Pge7X zUX|bO?j@tVgB~;3#t2fEVl;MRPcLSDqpBJ~hmJ1-(!|}rClnv$*WJ#rb-4{8E3-T4 zc<}u~*moG;adO5#DgeO}qG)(1kZ?i~jT8+8EFcSkx4xzEqM%XMHc&EDC#ch_5|Zcu zgbezR3i?0kGc6O2n9Q+A7#M@`^Y?YOzYhp5a5ix>H*#U`g$Vsv zi8gW&XVx<;tp3?q5)f5W&z@l^*fJs>|M;4!GBAsbUh9?e9EMIOQgM!6o=#9SQS_lh zziE=IwMW`El(_J6!RM!S(P+xktuffmjfsf8o=-vJArq+m5JqYr^T^wx8AD1d!S>jx z3NnjsMfC3;=AiN=367=Osz@Z52Yx1YpY26$i8OgqwOzBy#cQ>J@vqBaP4;4A=a-br zZ%=9GDLDH&KQU*`AF~dU6>z$UzryqThf8fj0xiPw^b<=%HdsqC-S?)Dtzj%MSHVV9 z@8g>zAuk=F6&9;BRk0Mk6@Y#bYf%9fafw|s)@L!s0cS3qUG6;_(y;02=lA{@E`0YZ z+e;0J^iHodzX)eOD?WsxsG%km?l(4CAhFgsD`%}xG7(dP7E8vl(b5cQwR@f=c6m;R zBB)T&C)c7hnOYc755wbw1#L8`e92IW!PgT~Wm3PqVu(%PRWD$sX}Me?fY-!cYGDdt zw9v|8@=-KQHF!;%(NF#H61qykJ)d%AX&bD47Fki=Q$joW8@H#zZ#N3Y1djVOu3|I@ zj!#z!abaJ1MHFb~zOt2Oh)hp< zE*q7Es^xQ8pt#PbRD%SshG{s`wkVP>C<_LE;k9)CF5F6M`mLYemE4e!`ABDqpA^>2 z!&U8~ta0=WVHCr(*PZ|bv)XBo?uH5=t!ezdS-j$SrZk`a@T z6gYqemp_Jfyj4YmiS16aotSg?-SP6WhU+(YDNeOb%RMQi>U!~{OSA8X-gD3Ew?gtT z=Mxxhk+Rjr?3sD+i3p)dbbHgU4rN;U=61p5CtwE=?oEjwLV?)&IRCA)?xW2p`{TZ^ zs-uV{Zccr93HNN1k^Jyx^=HfvT>;O;^s2SULsLCXvv@nyEb)}dNcFmFB%s00on!0o ztBaTuWjc;XF0xNMdppAd{M3j#Hb=D_mfgver6l=62|IS5;84^mSKCs*?(kXyQ)%{k zuHLS0%t&fA2MtzDt<1$S+2xt8(OnWCl~C+4KEz2dORPP#{$hZL5BqZ6`riEz?I+re zWObYYqk0SqNzkGT{h${YhqIK!-W*-uiTlgX_OoC#KA-OBW`eHkTxnVxdri)8S*c&s zG=HIC3s-YQIahJ9+`aZ}Mm;RTTE*VRz=mF=Lw15fyD1Ah9^bpnzHQJhqkRuk2UVwM ztLBCv;~>f*l;;Ogzka6qz)J0O(YN^|?PccMGwwNFdh4;G?Itd~owpSbFC z+HPDqwi`T4L8yGTvo+CR>>9FpKlLDHaF=%Bfb9z_HeAHtS^UzGeT9Gh!*T=Ml@g~P z^Zs*@jg|9>X08m)sirD2Jua;g^x#(&EJVK@H}`OQbn!7=)eDYxL(98%pJmy4XJE?< z%$965W24k4G8V2;8Sk%+uSy=?9{L<%T<1TNZhKjhOx~y_WrNAHs>#UaL3DD0C&+cs z-C_RX3)Z`z`0-iEG%r7^A+O@yyYV)ax_UD@%rq8wHB=bAda|2!&U!*x-ZlBf?k6K6!o+R}Okh&cKYP@aiFbI@AwM$;y?q1I93K@$|pL zrd&{IX$a1LMdvxQy9jYpPFs8n!Lx1TD>3(t-$^zoc#34@1#_K0D(e}H|H-jaUv>jk z=QN|%zK8ByBYrQPl^+R-puDju7p%jir*}MTc{|M(V1C@eU85dVX%pnl*ZVP^PaK*R z)x{YrO34@}rW@khqDx9$Tp-O&lJXX@qQYlBYSX1v1Goo1h2Li`&z$!bGy@q60$eI4 zTL}SZHZgFGawfH-zsc~R`HEF3nQeeJw}b&ek#6n=6W!JGJ%GlO{Mhn^C#GwEuOUlCj+BY4h3s$2_0fF~H+gcyqyX^YKZKb6A4(2|TTZRr{EpC7~ z$;_JH`t`8HY)+|u*u{6gjupRIhhqiOxgUHQH?SM=5SPqV%ke{r&Q(u8b+LEzZel+Y zoiUCKmg8s6-ZM|t0*>j= z=+`$=seUw$&~c~E#_56Q%Ilp!lLq79mqty^EfibE9Iii+J!}nGpI4+zrg;R4SxnZK z?v&5778pzAzL>T+P+xe@ZuB6(f|Q=~aC`lf6z;o#9+iEG^rII=*1Y|OZP7GF^SqWA z6QwNW{8iA#Xh45NYf=}4vaT+uDOEhU%*K$~a~%^@>UP`1AjIBYNlz=_TCQV6H5w1& zh>tcJUh|Jxcs##A15bDi5VB0HvLfL(A2XGz||8R6kyrNwcr09WUFpo`R z_@UUyk42r4DLC}}JXy}eK4u!iaRwHcU#perRVC>sV^2A$maPUAX8j|Q1&U^Gwb6xTdx*A<_rvfRHLdU zO-BR?A9t^e54?PHp;A!tYXjoVW2MiZ9#{r1vp6NL>F4pfqx}M~LqV9?_SgdxAHl$G ztHQ@fhta_8=ZLug7CT$UZy|z2QAipKzl>3sNLq{&Xx;v}F(4z!DOWn(d4Zp~n2BYZ zcq33sfS0@<+4?Kt`H_aKg!)RIiFB9O>u2jf2%bB9k!DMJL(^F}67IOD%TWk7TyoA% zy&vhjQUCN@Webb#ev!!KF5?q<`eo>N1IAJD>*EDLQG+adsQY{?;BY85K6O0VKU<_b zYmO9gJTZEc+M}PlzH8BJIEf?V=Y)~bw8GA!Zu9d&vBLDfbRZoAbK62SBP*M8KE0g;o*XaB?Ag2S+EE9K=xy<>-Kt zfseJ ze2M?eGL>H$^iN8P$Po1T`Tmh-D_UTPTp9v}Lgc`RzdMS6!6IS+89gs2ReOJDQB^b) z41t0FT|tl%48VY)0{>bLSin;N@b?3ZM1WEM$skBs*dGjzf)JJjL0AOjk1rG)@t+I| zgZ;r||LXdO4vhFC9*P9R|I~qze>2;E>kb9O{u}-qgMsD#m{At?e}o_)e*`0-C|SbV zbp>k_TosN)$tlS}l)-XpC^!rur>r8Uf<(a7)lg6+6h`6yRagTs7=WL@y{~_;uk&31 P5{W_pgoV{^sRRBCiKjK} -- 2.39.5