From: Jean-Paul Pelteret Date: Mon, 1 Jun 2020 12:51:51 +0000 (+0200) Subject: Add section on SD::BatchOptimizer X-Git-Url: https://gitweb.dealii.org/cgi-bin/gitweb.cgi?a=commitdiff_plain;h=dfb71e5af7ea34c843bb70276b3a05706dc7c267;p=release-papers.git Add section on SD::BatchOptimizer --- diff --git a/9.2/paper.tex b/9.2/paper.tex index 45c8c77..4278337 100644 --- a/9.2/paper.tex +++ b/9.2/paper.tex @@ -645,16 +645,88 @@ To visualize the motion of particles, the \texttt{Particles::DataOut} class was \subsection{Improved performance of the symbolic differentiation framework} \label{subsec:symbdiff} -\todo[inline]{J-P's section} +In the previous release we added support for symbolic expressions, leveraging the +SymEngine library \cite{symengine-web-page}. +Although effective, evaluating lengthy expressions could be a bottle-neck as this +was performed using dictionary-based substitution. +We have improved on this by implementing a \texttt{BatchOptimizer} class in the +namespace \texttt{Differentiation::SD} that collects several \texttt{Expression}s +and transforms them in such a way that the equivalent result is returned through +a quicker code path. +This may be done by simply using common subexpression elimination (CSE) for the +dictionary-based expressions, by transformation to a set of nested +\texttt{std::function}s (the equivalent to \texttt{SymPy}'s "lambdify", with or +without using CSE), or by offloading to these expressions to the \texttt{LLVM} +just-in-time (JIT) compiler. +Although, each of these features is implemented and tested in the SymEngine +library itself, the \texttt{BatchOptimizer} class provides both uniform +interface to their classes and a convenient interface for scalar expressions, +as well as tensorial expressions formed using the \texttt{deal.II} tensor and +symmetric tensor classes. +It, like the \texttt{Expression} class, is also serializable. + +The way that the batch optimizer may be employed within a user's code is shown +in the pseudo-code below. +As per usual, one would first define some independent variables, and +subsequently compute some symbolic expressions that are dependent on these +independent variables. +These expression could be, for example, scalar expressions or tensors of +expressions. +Instead of evaluating these expressions directly, the user would now create an +optimizer to evaluate the dependent functions. +In this example, the selected arithmetic type numerical result will be of type +\texttt{double}, and the \texttt{LLVM} JIT optimizer will be invoked. +It will employ common subexpression elimination and aggressive optimizations +during compilation. +The user then informs the optimizer of all of the independent variables and the +dependent expressions, and invokes the optimization process. +This is an expensive call, as it determines an equivalent code path to evaluate +all of the dependent functions at once. +However, in many cases each evaluation has significantly less computational cost +than evaluating the symbolic expressions directly. +Evaluation is performed when the user constructs a substitution map, giving each +independent variable a numerical representation, and passes those to the +optimizer. +After this step, the numerical equivalent of the individual dependent expressions +may finally be retreived from the optimizer. -BatchOptimizer for symbolic expressions -\begin{itemize} - \item CSE for dictionary-based expressions - \item ``lambdify'' (via SymEngine) - \item Offloading to LLVM JIT compiler (via SymEngine) - \item Serialization, so complex expressions can be compiled offline -\end{itemize} +\begin{c++} +using namespace Differentiation::SD; + +const Expression x("x"); +const Expression y("y"); +... +const auto f = calculate_f(x, y, ...); // User function +const auto g = calculate_g(x, y, ...); // User function +... + +BatchOptimizer optimizer (OptimizerType::llvm, + OptimizationFlags::optimize_all); +optimizer.register_symbols(x, y, ...); +optimizer.register_functions(f, g, ...); +optimizer.optimize(); + +const auto substitution_map + = make_substitution_map({x, ...}, {y, ...}, ...); +optimizer.substitute(substitution_map); + +const auto result_f = optimizer.evaluate(f); +const auto result_g = optimizer.evaluate(g); +\end{c++} +This expense of invoking the optimizer may be offset not only by +the number of evaluations performed, but also by the amount of reuse each +instance of a \texttt{BatchOptimizer} has. +In certain circumstances, this can be maximized by generalizing the way in which +the dependent expressions are formulated. +For example, in the context of constitutive modelling the material coefficients +may be made symbolic rather than encoding these into the dependent expressions +as numerical values. +The optimizer may then be used to evaluate an entire family of constitutive laws, +and not a specific one that describes the response of a single material. +Thereafter, serializing the optimizer instance and reloading the contents during +subsequent simulations permits the user to skip the optimization process entirely. +Serialization also enables these complex expressions to be compiled offline. %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% \subsection{Advances of the SIMD capabilities and the matrix-free infrastructure}