From: Martin Kronbichler Date: Thu, 19 Sep 2024 07:34:42 +0000 (+0200) Subject: A few improvements to the text X-Git-Url: https://gitweb.dealii.org/cgi-bin/gitweb.cgi?a=commitdiff_plain;h=163bd9b1ea94c0d694fdc8eec5c53a57c97561e3;p=release-papers.git A few improvements to the text --- diff --git a/9.6/paper.tex b/9.6/paper.tex index d8fa044..be7af24 100644 --- a/9.6/paper.tex +++ b/9.6/paper.tex @@ -362,28 +362,36 @@ can be found We made different updates to the matrix-free infrastructure in \dealii. These include: \begin{itemize} -\item Our own implementation of \texttt{std::simd} (\texttt{VectorizedArray}) now - also supports Arm Neon. Arm Neon is an architecture extension for - the Arm Cortex-A and Arm Cortex-R series of processors, commonly used in Apple - products. With these instructions 2 doubles or 4 floats can be processed in one - go. Since matrix-free infrastructure works directly with \texttt{VectorizedArray} - data structures, it automatically benefits from this implementation extension. +\item Our own implementation of \texttt{std::simd}, called + \texttt{VectorizedArray}, now also supports Arm Neon. Arm Neon is an + architecture extension of the ARMv7, ARMv8 and ARMv9 architecture families, + used e.g.~for the Arm Cortex-A and Arm Cortex-R series of processors or in + the Apple 64-bit silicon series (A7--A18, M1--M4 as of 2024). + With these instructions, 2 doubles or 4 floats can be processed in one + go. Since the matrix-free infrastructure works directly with the \texttt{VectorizedArray} + abstraction as data structure, it automatically benefits from this new implementation. -\item We made some improvement for evaluation of values and gradients for - $H$(div)-conforming Raviart--Thomas elements on non-Cartesian elements. +\item The application of the Piola transformation for values and gradients of + $H$(div)-conforming Raviart--Thomas elements on non-Cartesian elements has + been rewritten for better performance. Together with additional changes in + the sum-factorization algorithms that are now more similar to the kernels + described in~\cite{KronbichlerKormann2019}, the matrix-free operator + evaluation is now three to five times faster in these cases. -\item We furthermore improved the internal data structures of the - tensor-product evaluators as well as the evaluators for simplex elements, - which speed up the evaluation in several scenarios, especially multi-component systems. +\item Furthermore, improved the internal data structures of the tensor-product + evaluators as well as the evaluators for simplex elements. This speeds up + the operator-evaluation in several scenarios, especially for simplices + (around two times higher throughput for operator evaluation) and + multi-component systems. \end{itemize} Furthermore, we performed substantial improvements to the global-coarsening multigrid infrastructure: \texttt{MGTransferMF} (previously: \texttt{MGTransferGlobalCoarsening}) and \texttt{MGTwoLevelTransfer}. They now allow to perform local smoothing, which is a key step towards unifying -all transfer operators in \dealii. \texttt{MGTwoLevelTransfer} now also -can be set up with an existing \texttt{MatrixFree} object in the case -of $p$-multigrid, allowing to reduce the setup costs significantly. +all transfer operators in \dealii. Finally, \texttt{MGTwoLevelTransfer} can now +be set up with an existing \texttt{MatrixFree} object in the case +of $p$-multigrid, allowing to reduce the setup costs and memory consumption significantly. %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% @@ -392,18 +400,23 @@ of $p$-multigrid, allowing to reduce the setup costs significantly. %\todo[inline]{All: If you contributed to this functionality, please % help flesh out this section.} -We made different advances to the non-matching infrastructure in \dealii. These include: +For the non-matching infrastructure in \dealii, the following improvements were made: \begin{itemize} \item The performance of the non-nested multigrid infrastructure (\texttt{MGTwoLevelTransferNon\-Nested}) has been improved significantly by avoiding -redundant copy operations. Furthermore, the support for simplex-shaped cells and -multiple-component elements have been added. +redundant copy operations. Furthermore, support for simplex-shaped cells and +multiple-component elements has been added. +\item Several minor performance improvements for \dealii's evaluator class on + sets of unstructured points, \texttt{FEPointEvaluation}, have been + made. This is particularly useful for evaluating several quantities on the + same set of points, which is enabled by \texttt{NonMatching::MappingInfo}. \item The new class \texttt{FERemoteEvaluation} has been added. This is a class to access data in a distributed matrix-free loop for non-matching discretizations. -Interfaces are named with \texttt{FEEvaluation} in mind. Key component is the -underlying MPI communication, which is done via \texttt{RemotePointEvaluation}. +Interfaces are named in analogy to \texttt{FEEvaluation}, in order to seamlessly switch +local evaluator functionality at quadrature points of a computation. The key component is the +underlying MPI communication infrastructure, which is done via \texttt{RemotePointEvaluation}. Tutorial \texttt{step-89} has been added to present its usage in the context of an application to acoustic conservation equations~\cite{heinz2023high}. \item FECouplingValues @@ -432,8 +445,8 @@ time integration and discretization. Historically, \trilinos{} has implemented \petra{} objects -- such as parallel vector, sparsity pattern, and sparse matrix classes -- in its \epetra{} package that uses \MPI{} as its only source of -parallelism. The \dealii{} interfaces to \trilinos{} are therefore all -built using \epetra{}. +parallelism. The \dealii{} interfaces to \trilinos{} have therefore traditionally all +been built using \epetra{}. However, several years ago, \trilinos also introduced the newer \tpetra{} (``templated'' Petra) package that provides additional shared memory parallelism and GPU capabilities by building on \kokkos{}. \tpetra{} @@ -534,7 +547,7 @@ programs: checkpoint/restart functionality in \dealii{}-based programs, using the BOOST serialization functionality as a foundation. step-83 was written by Pasquale Africa, Wolfgang - Bangerth, and Bruno Blais and uses step-19 as its basis. + Bangerth, and Bruno Blais using step-19 as its basis. \item \texttt{step-86} is a program that solves the heat equation using PETSc's TS (time @@ -823,9 +836,7 @@ German Ministry of Education and Research, project ``PDExa: Optimized software methods for solving partial differential equations on exascale supercomputers'' and the Bayerisches Kompetenznetzwerk f\"ur Technisch-Wissen\-schaft\-li\-ches Hoch- und H\"ochstleistungsrechnen -(KONWIHR), projects ``High-order matrix-free finite -element implementations with hybrid parallelization and improved data -locality'' and ``Fast and scalable finite element algorithms for coupled +(KONWIHR), project ``Fast and scalable finite element algorithms for coupled multiphysics problems and non-matching grids''. M.~Maier was partially supported by NSF Award DMS-2045636 and and by the