Skip to content

Latest commit

 

History

History
539 lines (518 loc) · 41.4 KB

CHANGELOG.md

File metadata and controls

539 lines (518 loc) · 41.4 KB

Changelog

All notable changes to this project will be documented in this file. The format is based on Keep a Changelog.

We mark updates with a (pre-)release tag on github.

Creating a release on github will create an associated archived version on zenodo.

As of v8.0 we try to follow Semantic versioning i.e. major version number changes break backwards compatibility and minor version number changes indicate new features. Before that we kind of made up our own version numbers. A new major version number is often associated with a journal publication, but other than that there is no defined mapping from the amount or kind of change to a version number.

Only changes in code are reported here, we do not track changes in the doxygen documentation, READMEs or tex writeups. As of v7.0 we also stop reporting changes in test and benchmark programs.

[v8.2.2] MSVC + OpenMP

Fixed

  • Fix non-standard use of uint
  • Fix OpenMP 2 compliance on Windows
  • Fix OpenMP version detection using _OPENMP Macro instead of by-compiler detection
  • Fix error in Probes module where only first coordinate is written to file
  • Fix a bug in Feltor's restart routine

Added

  • msvc-omp cmake preset for Windows

Changed

  • Optimize implementation of dg::geo::solovev and dg::geo::polynomial flux functions. Speed up evaluation and fieldline integration.
  • Change cmake location of test and benchmark executables to follow original folder structure exactly (i.e. build/tests and build/benchmarks replaced with build/inc)
  • Fully integrate geometries and matrix into Catch2 tests. Changed some file names from *_t.cpp to *_b.cpp to strictly enforce all ``*_t.cppfiles as Catch2 test files. For exampleds_t.cpp` is now called `ds_b.cpp`

[v8.2.1] Fix Clang compilation

Changed

  • Update cccl dependency version from 2.8.0 to 2.8.4

Fixed

  • Fix all warnings and errors from clang-19
  • Fix some remaining warnings from msvc
  • FP_FAST_FMA macro no longer used for fma support test except for GCC

Removed

  • Revmove DG_BENCHMARK macro everywhere

[v8.2] CMake integration

Added

  • Complete cmake build system for all programs and libraries
  • Macro VCL_NO_INCLUDE_PREFIX can be used to include vectorclass.h directly

Changed

  • matrix now depends on lapack directly, not lapacke
  • The timestamp in the file history now follows ISO8601 formatting in UTC
  • probes_params uses std::vector instead of thrust::host_vector
  • All remaining *.cu files are renamed to *.cpp
  • Moved all diag programs to their respective src project

Fixed

  • everything compiles without warnings from -Wextra -Wpedantic and \W4 on Windows
  • Remove the use of variable length arrays (VLAs)
  • Change uint to unsigned in mpi_kron.h
  • Change iterator type in dg::View to given pointer type (else undefined construction of unkown iterator type from pointer)
  • dg::is_divisable ignored eps parameter
  • Avoid zero sized array warnings in aRealMPITopology for Nd=0
  • filesystem paths are given in wchar on windows; fix ability to read wchar
  • Fixed MPI file write error when get does not set value type
  • Fix probes parser when using nlohmann-json

Removed

  • diag/compare.cpp Removed because it can be trivially implemented in python
  • Boost dependency of geometries header completely removed using C++17 special functions, in particular in taylor.h

[v8.1.1] Hotfix aliases

Fixed

  • Aliases in dg::blas1::pointwiseDot and dg::blas1::pointwiseDivide lead to wrong results

[v8.1] Remove cusp dependency

Added

  • New dg::SparseMatrix class to replace our previous cusp::csr_matrix and cusp::coo_matrix
  • The SPARSELIB Makefile variables expands to -lcusparse when compiling with cuda
  • resize method in dg::TriDiagonal
  • Writeable data method in dg::SquareMatrix

Changed

  • The LIBS Makefile variable includes -lcusparse dependency when compiling with cuda. This is rather trivial because cuda ships with cusparse.
  • Default build should now use https://github.com/nvidia/cccl repository
  • Default build should now use https://github.com/vectorclass/version2 (instead of version1)
  • All dg functions/classes/typedefs returning or depending on any cusp coo or csr matrix now use dg::SparseMatrix
  • All dg functions/classes returning or depending on cusp::dia_matrix now use dg::TriDiagonal
  • All dg functions/classes returning or depending on cusp::array1d now use thrust::host_vector
  • All dg functions/classes returning or depending on cusp::array2d now use dg::SquareMatrix
  • The inverse type of tridiagonal matrices in dg/matrix/matrix.h is dg::SquareMatrix

Deprecated

  • dg::InverseTriDiagonal because untested and unused over dg::mat::compute_Tinv_y

Removed

  • All cusp header files are removed from dg/algorithm.h, dg/geometries/geometries.h and dg/matrix/matrix.h (The relevant tensor traits for cusp can still manually be included though)

Fixed

  • Thrust library can now be used at newest version

[v8.0] Improved foundations

Added

  • A novel C++17 style dg::file::NcFile class together with its MPI counterpart and supporting infrastructure that greatly simplifies netCDF-4 output
  • Add modular dg::file::Probes class with default dg::file::parse_probes that can be added to any project
  • Simplify output design with dg::file::Record
  • Add possibility to use nlohmann::json instead of jsoncpp in our WrappedJsonValue using the macro DG_USE_JSONHPP
  • New blas1 function dg::blas1::kronecker for Nd dimensional product space evaluations
  • New blas1 function dg::blas1::vdot that can handle general and in particular complex scalar products
  • Novel N-dimensional Grid class dg::aRealTopology and dg::RealGrid (and its MPI counterparts) that generalize the previous 1d, 2d and 3d grids. Previous grid classes still fully function using typedefs.
  • Use Catch2 as a unit test framework for dg, dg/file and dg/topology. We introduce tests/run-all-tests script
  • A novel MPI communication backend allowing for consistent overlapping communication and computation in all sparse matrices
  • dg::mpi_cart_xxx with xxx in create, kron, split for more consistent and easier Cartesian communicator handling in MPI grids
  • dg::mpi_read* functions for more consistent and easier MPI initialisation
  • Improved handling of cuda-aware MPI detection for non-OpenMPI libraries
  • Add default constructors to most classes notably Grid classes
  • Add optional weights parameter to dg::least_squares
  • Add experimental extrapolate_least_squares method to dg::Extrapolate (unfortunately not very successful)
  • Template version of dg::is_same
  • Add more verbose output to SeparatrixOrthogonal
  • Add dg::mat::BesselJ and dg::mat::LaguerreL and dg::mat::GyroLagK functors
  • Add dg::exblas::fpedot_cpu, dg::exblas::fpedot_gpu and dg::exblas::fpedot_omp
  • Add dg::exblas::ufloat
  • Probes module in feltor 3d code
  • More general interfaces working with new N dimensional grid class to all topology functions like evaluate, interpolate, derivatives etc.
  • dg::cHVec, dg::cDVec, dg::cMHVec and dg::cMDVec typedefs for complex vector types
  • Expose dg::is_scalar_v, dg::is_vector_v, dg::has_policy_v and dg::is_matrix_v type predicates
  • Extended Tag system on scalar types in particular complex types (std::complex and thrust::complex)
  • Add dg::geo::SOLRegion and dg::geo::ClosedFieldlineRegion as predicates
  • Add predicate parameter to dg::geo::WallFieldlineDistance and dg::geo::WallFieldlineCoordinate, which may accelerate sheath generation by a factor 4
  • Add double X-point switch in dg::geo::createModifiedField as "sol_pfr_2X"

Changed

  • std=c++17 Change C++ standard to C++-17
  • dg::Average has one additional template parameter (the interpolation matrix). Changed in all occurences in src programs.
  • dg::EllSparseBlockMat and dg::CooSparseBlockMat got one additional template parameter (but does not matter if used through dg::HMatrix typedef)
  • Bump Doxyfile(s) to 1.9.8
  • dg::mpi_init now also sets GPU
  • dg/geometries/geometries.h now always includes dg/file/json_utilities.h
  • The main documentation page is changed to topics.html from modules.html
  • Complete rework of MPI sparse matrix communication backend
  • Most *.cu file endings were changed to *.cpp (Nicer for C++ automatic syntax highlighting)
  • Most inc/*_mpi*.cpp programs were merged into its corresponding inc/*.cpp program (reduces future maintenance load)
  • Changed static inline function declarations to inline
  • MPI grid classes can handle arbitrary parallelization (even those that do not evenly divide cell numbers)
  • dg::geo::DS class no longer adds jump terms in its symv method. Thus removed one template parameter from dg::geo::DS class. Adapted in all src prorams.
  • SeparatrixOrthogonal class only integrates with 1e-11 default accuracy (prior 1e-13)
  • dg/geometries/geometries.h now automatically incurs json dependency
  • The tridiag method of dg::mat::UniversalLanczos now returns the tridagonal matrix T
  • Restructure feltor, feltorSH, feltorSHp, feltorSesol, feltorShw, toefl, lamb-dipole, geometry-diag programs using new I/O design
  • Use WrappedJsonValue in poet, esol, hasegawa and heat code
  • Use macro WITH_MPI in most executable programs for easy compiler switch
  • Move ping_mpit.cu program to src/ping/mpi-ping.cpp
  • Changed x,y,z coordinate to R,Z,P in geometry-diag
  • Add capability for 2nd X-point in geometry-diag
  • Add ClosedFieldlineRegion and SOL output in geometry-diag
  • Major code refactoring of 3d feltor code in preparation of thermal code
  • Member functions of dg::DLT are now static
  • dg::Operator is now named dg::SquareMatrix and can be used in dg::blas2::symv functions together with std::vector (dg::Operator typedef still exists for backward compatibility)
  • dg::lu_solve is now exposed in main dg namespace
  • Allow general value types in all blas1 functions
  • dg::zero and dg::one have more general template interface
  • dg::blas1::dot and dg::blas2::dot allow complex input vectors (in fact they allow any addable type)
  • dg::MPI_Vector::set_communicator member only needs one MPI_Comm parameter

Deprecated

  • All previous netcdf utilitiy functions are now deprecated
  • dg::mpi_init1d, dg::mpi_init2d and dg::mpi_init3d and all overloads

Removed

  • dg::EllSparseBlockMatDevice and dg::CooSparseBlockMatDevice now merged into dg::EllSparseBlockMat and dg::CooSparseBlockMat
  • dg::sainv_precond as it is not used
  • device=cpp is no longer supported. Use device=cpu.
  • Last remains of DG_DEBUG and SILENT macros completely removed
  • doxygen-awesome css and js files. We now use doxygen-awesome as external library instead of keeping copies in feltor/doc
  • dg::average and dg::mpi_average (now replaced by dg::create::reduction)
  • dg::extend_line and dg::extend_column (now replaced by dg::create::prolongation`)
  • diag/probes.h is now replaced by dg/file/probes.h
  • Some low level operator generators for dg derivatives
  • fehlberg-4-2-3, billington-3-3-2, trbdf2-3-3-2 Butcher tableaus as they performed very badly and should not be used
  • dg::transpose
  • communicator_mod and communicator_mod_reduce members of dg::MPI_Vector
  • All previous MPI communication objects
  • dg::Buffer memory class

Fixed

  • Fix unnecessary copy in Adaptive and SingleStep timeloop
  • Fix Broadcase in mpi_init3d goes to comm not to world
  • Fix value init in default constructor of dg::MPI_Vector
  • Fix bug in double version of dg::is_same
  • Fix sign in SafetyFactor
  • Fix grid generation close to O-point
  • Fix verbose output in SeparatrixOrthogonal
  • Performance of serial sparseblock matrix now equals that of OpenMP with 1 thread

[v7.0] Three-dimensional

Added

  • A left looking sparse inverse preconditioner dg::sainv_precond (sadly did not yield any benefits over dg::nested_iterations in our tests so we did not implement an MPI version)
  • dg::blas2::parallel_for for general parallelization tasks on shared memory containers (very close in functionality to kokkos::parallel_for
  • dg::blas2::stencil A generalization of dg::blas2::parallel_for that works for MPI parallelization at the cost of having to encode the communication stencil in a sparse matrix and only one vector argument
  • dg::create::window_stencil and dg::create::limiter_stencil overloaded for various topologies for use as communication stencils in dg::blas2::stencil
  • Functors dg::CSRSlopeLimiter, dg::CSRMedianFilter, dg::CSRSWMFilter, dg::CSRAverageFilter, dg::CSRSymvFilter for use as functors in dg::blas2::stencil
  • isLocalBijective method in dg::SurjectiveComm
  • write access to inner_matrix and outer_matrix in dg::MPI_Matrix
  • MPI_Comm parameter in constructor of dg::NearestNeighborComm
  • set_right_size and set_left_size methods in dg::EllSarseBlockMat
  • dg::Elliptic1d (since we have 2d and 3d versions)
  • dg::Sign a 1d functor that returns the sign of its argument
  • dg::GeneralHelmholtz generalizes all the Helmholtz classes into one class (and specializes to dg::Helmholtz1d, dg::Helmholtz, dg::Helmholtz2d, dg::Helmholtz3d
  • Experimental dg::FilteredERKStep class (sadly does not work so well)
  • n method to access number of polynomial coefficients in dg::DLT
  • dg::create::fast_transform for creating block diagonal matrices on various topologies
  • overloads for dg::create::fast_interpolation and dg::create::fast_projection to select directions
  • new fem functionality: dg::create::fem_linear2const, dg::create::inv_fem_mass dg::create::inv_fem_linear2const, dg::create::fem_mass2d, dg::create::inv_fem_mass2d, dg::create::fem_linear2const2d, dg::create::inv_fem_linear2const2d
  • dg::create::interpolation and dg::create::projection functions now all have a method parameter (that defaults to "dg")
  • dg::invert overload acting on dg::Operator as an alias to dg::create::inverse
  • dg::create::backproject and dg::create::inv_backproject for projecting a dG grid vector to an equidistant grid and back
  • dg::create::inv_backscatter for inverting an interpolation to an equidistant grid
  • dg::convertGlobal2LocalRows and dg::convertLocal2GlobalCols helper functions to enable mpi conversion of column distributed matrices
  • access_string method in dg::file::WrappedJsonValue to access creation history of an object
  • Restructure of dg::geo::Fieldaligned to allow for different interpolation and backprojection from dG to fine grid and back (to a possibly equidistant grid and with possibly cubic, linear or nearest methods)
  • in dg::geo::DS the derivative methods now have the general zero transform
  • dg::geo::createCircularField has now two radii parameter to allow for the creation of an ellipsis and changed the sign
  • dg::geo::SeparatrixOrthogonalAdaptor to create Separatrix grids without XGrid topology
  • Utilities for exponential integration: dg::mat::ExponentialStep, dg::mat::ExponentialERKStep, dg::mat::phi1, dg::mat::phi2, dg::mat::phi3, dg::mat::phi4, dg::mat::FuncionalButcherTableau, dg::mat::ConvertsToFunctionalButcherTableau with the four methods "Euler", "Midpoint-2-2", "Runge-Kutta-4-4", and "Hochbruck-3-3-4"
  • Utility function dg::mat::MatrixFunction for easy general purpose use
  • Test of filters and Limiters and implicit timesteppers in the src/lamb_dipole project
  • Test of new schemes in src/navier_stokes project
  • File src/toefl/input/default_hpc.json to be consistent with Quick Start guide

Changed

  • explicit use of "default" constructors in dg::Adaptive, dg::Advection, dg::AndersonAcceleration, dg::ArakawaX, dg::BijectiveComm, dg::GeneralComm, dg::BICGSTABl, dg::ChebyshevIteration, dg::EVE, dg::LGMRES, dg::MultigridCG2d, dg::PCG, dg::Poisson, dg::ERKStep, dg::SinglestepTimeloop, dg::ButcherTableau, dg::ShuOSherTableau, dg::RealGrid1d, dg::Operator
  • dg::create::modal_filter no longer returns a dg::MultiMatrix directly. Instead, a dg::Operator is returned that can be converted to dg::MultiMatrix using dg::create::fast_transform
  • rename getLocalGatherMap to getLocalIndexMap in dg::SurjectiveComm and dg::GeneralComm
  • rename getPidGatherMap to getPidIndexMap in dg::SurjectiveComm and dg::GeneralComm
  • rename getSortedGatherMap to getSortedIndexMap in dg::SurjectiveComm
  • Constructors of dg::Helmholtz, and dg::Helmholtz3d change the orders of parameter alpha and reverts to constructor of respective dg::Elliptic class
  • 1d overloads for dg::create::fast_interpolation and dg::create::fast_projection return dg::HMatrix_t instead of dg::MultiMatrix
  • dg::create::fem_mass now returns dg::TriDiagonal instead of cusp::coo_matrix
  • dg::create::precond overload for dg::DLT renamed to dg::create::inv_weights
  • parameters to dg::tensorproduct changed from cusp::coo_matrix to cusp::csr_matrix
  • all appearances of guenther or Guenther in dg::geo are corrected to guenter respectively Guenter (because that is how her name is spelled)
  • clone and other virtual methods in dg::geo::RealCurvilinearGridX2d, dg::geo::RealCurvilinearProductGridX3d, dg::geo::DSPGenerator, dg::geo::FluxGenerator, dg::geo::Hector, dg::geo::RealCurvilinearProductMPIGrid3d, dg::geo::PolarGenerator, dg::geo::LogPolarGenerator, dg::geo::Ribeiro, dg::geo::RibeiroX, dg::geo::SimpleOrthogonalX, dg::geo::SeparatrixOrthogonal, dg::geo::SimpleOrthogonal have override final keywords
  • default parameters changed in dg::geo::Fieldaligned constructor
  • restructure of the src/feltor project (including renaming *.cu files to *.cpp)

Deprecated

  • dg::geo::createModifiedSolovevField is superseded by dg::geo::make_field

Removed

  • dg::ModalFilter is removed in favor of generating and using dg::MultiMatrix directly
  • dg::transpose_dispatch (replaced by dg::blas2::parallel_for)
  • dg::extend_line (replaced by dg::blas2::parallel_for)
  • dg::extend_column (replaced by dg::blas2::parallel_for)
  • dg::create::interpolationT (is not used anywhere)
  • file src/toefl/toefl_hpc.cu (was previously merged into toefl.cpp
  • dg::integrate replaced by dg::aTimeloop functionality
  • dg::integrateERK replaced by dg::aTimeloop functionality
  • dg::stepperRK replaced by dg::aTimeloop functionality

Fixed

  • Fix left shift in exblas::AccumulateWord by -1 by explicit conversion to unsigned
  • Fix isCommunicating in dg::BijectiveComm
  • Fix global_gather and global_scatter_reduce in dg::aCommunicator in the case that a map is not communicating but still needs to locally gather values. This makes the two functions truly the transpose of each other.
  • Fix dg::ShuOsher copyable method segmentation fault in case of default constructed
  • Fix size method in dg::Operator for integer value types
  • Fix convergence criterion in fieldline integrator for dg::geo::SimpleOrthogonal
  • Fix constructor of dg::mat::PolCharge and dg::mat::PolChargeN

[v6.0] More modular

Added

  • Add ability to use lambdas and functors without dg::TensorTraits in dg::apply, dg::blas2::symv and dg::blas2::gemv (Extremely useful!)
  • Add --extended-lambda flag in default NVCCFLAGS
  • New class dg::file::WrappedJsonValue plus test file; allows easy bug-tracing on jsoncpp file input
  • Allow arbitrary nx, ny, nz polynomial number in all grids; adapt all weights, derivatives, etc. creation accordingly
  • New sub-project dg/matrix/matrix.h written by Markus and optimized by Matthias. Allows matrix-function computations in particular square roots
  • CONTRIBUTING.md file (moved from homepage)
  • LAPACKLIB variable in Makefile configuration
  • new device=cpu configuration (enable to compile without parallelization)
  • new file config/version.mk provides git hash and branch information for file output
  • new wall and sheath generation capabilities in inc/geometries
  • new dg::aTimeloop family of classes to help create time-stepping
  • new explicit timesteppers Cavaglieri-3-1-3 (explicit and implicit), Fehlberg-3-2-3, Fehlberg-4-2-3, Cavaglieri-4-2-3 (explicit and implicit), Tsitouras09/11-7-4-5, Verner-9-5-6, Verner-10-6-7, Dormand-Prince-13-7-8,
  • new implicit timesteppers SDIRK-4-2-3, Sanchez-3-3, Sanchez-3-4, Sanchez-6-5, Sanchez-7-6
  • Add vector support for std::map in blas1 functions
  • Add UnaryOp parameter to dg::blas1::reduce
  • Add DenseMatrix-Vector multiplication suppoert in dg::blas2::symv
  • Add append method to dg::Message
  • Add dg::abort_program function
  • Add more cuda Error management in backend functions
  • Add convenience dg::mpi_init function
  • Matrix conversion from dg::EllSparseBlockMat to cusp::coo_matrix and associated dg::tensorproduct and dg::create::diagonal function to re-create elliptic operators
  • Add convenience mpi Macro DG_RANK0 helping to write shared/mpi programs more easily
  • Add new multigrid implementation dg::nested_iterations for maximum modularity, sovlers can be separately chosen at each stage
  • Add experimental and preliminary FEM support
  • Add dg::least_squares function
  • project esol Edge-SOL turbulence in Full-F models
  • project poet Full-F Full-K blob simulations
  • project navier_stokes 3d field-aligned Navier-Stokes equations

Changed

  • Redesign implicit and semi-implicit timestepper interfaces for more modularity and the ability to use lambdas
  • Redesign dg::Adaptive
  • Unify stopping criterions for Matrix inversion classes and use of weighted scalar product
  • dg::apply, dg::blas2::symv and dg::blas2::gemv are now formally the same
  • dg::PCG works with self-adjoint matrices in weights instead of symmetric ones
  • Renamed dg::lr_pivot to dg::lu_pivot and make it use binary reproducible dot functions, same for dg::lu_solve
  • Rename dg::create::invert to dg::creat::inverse
  • Redesign project src/toefl; follows standard design on user-guide
  • Redesign project src/reco2D; follows standard design on user-guide
  • Redesign project src/lamb_dipole; adapt to dg::file::WrappedJsonValue
  • Adapt all src projects to new semi-implicit timestepper interface
  • Redesign DS and all related projects
  • Rename inc/geometries/magnetic_field_t.cu to inc/geometries/solovev_t.cu

Deprecated

  • dg::integrateERK replaced by dg::aTimeloop functionality
  • dg::stepperRK replaced by dg::aTimeloop functionality

Removed

  • dg::FilteredImplicitMultistep; barely any theory behind that
  • dg::Karniadakis was marked deprecated before
  • "Verner-8-5-6" Timestepper tableau
  • dg::norm ; Everything is normed now; adapt all codes to it; specifically also makes inv_weights and inv_volume functions unnecessary; matrix inversion works with weighted scalar products
  • class dg::Invert; replaced by lambda functionality
  • file diag/impRdiag.cu; replaced by python diagnostics
  • file diag/reco2Ddiag.cu; replaced by python diagnostics
  • file diag/toeflEPdiag.cu; replaced by python diagnostics
  • file diag/toeflRdiag.cu; replaced by python diagnostics
  • file diag/vmaxnc.cu; replaced by python diagnostics
  • file diag/window_params.txt; Useless
  • project src/ep; merged and oursourced with impurities project
  • project src/polar; Useless
  • project src/impurities; outsourced to https://github.com/mwiesenberger/impurities
  • nc_enddef and nc_redef calls; they are unnecessary according to netcdf documentation
  • all files inc/geometries/[*magnetic-geometry.js*] These are now part of a separate magnetic field repository
  • hector_t.cu, ribeiro_t.cu, simple_orthogonal_t.cu merged into flux_t.cu

Fixed

  • fix bug from linking -lhdf5 by linking -lhdf5_serial in default.mk
  • fix bug nvcc did not compile .cpp programs now it does
  • fix bugs in dg::AndersonAcceleration
  • fix bugs in dg::LGMRES
  • fix bugs in dg::BICGSTABl
  • fix bug in exblas/accumulate.cuh and exblas/accumulate.h
  • fix bug in dg::ClonePtr copy-constructor
  • fix bug symv call to CooMatrix in mpi_matrix.h
  • fix NaN bugs in dg::blas2::symv even though beta == 0
  • fix nvcc compiler warnings when using host lambdas in dg::blas1 functions
  • fix tableau Tsitouras11-7-4-5

[v5.2] More Multistep

Added

  • M100 config file
  • json utility functions dg::file::get, dg::file::get_idx in dg/file/json_utilities.h which adds a small abstraction layer that gives a user more control over what happens if a variable is not found
  • json utility functions dg::file::file2Json, and dg::file::string2Json in json_utilities.h which adds a small abstraction layer that gives a user more control over what happens if an error happens during the parsing of a file
  • "easy output" netcdf utility functions that are particularly useful for MPI output: either write data in parallel or funnel through the master thread
  • new include files dg/file/file.h, dg/file/json_utilities.h and dg/exblas/exblas.h
  • new class dg::Gradient for gradient
  • new class dg::Advection for the upwind advection scheme
  • new dg::blas1::reduce function for custom reductions
  • new "exchangeable" dg::x::DVec, dg::x::HVec, ..., dg::x::CartesianGrid2d, ..., dg::x::IHMatrix, ... typedefs. The idea is that these resolve to either shared memory or mpi distributed memory versions depending on the MPI_VERSION macro. This helps merging shared and mpi programs into single ones.
  • added "simple" mode to dg::Average computation, which is beneficial for GPU computing
  • add dg::integrate that computes an indefinite integral of a function (essentially the opposite of the derivative)
  • add dg::ModalFilter and tests thereof
  • new dg::compose function that concatenates two or more functors to one
  • add dg::cooRZP2X coordinate transformation functions to easily transform between Cylindrical and Cartesian coordinates
  • interpolate function has an additional dg::space parameter to indicate nodal or modal values
  • Grid classes now have host_vector and host_grid member typedefs
  • new tensor functions dg::tensor::scalar_product2d and dg::scalar_product3d that can compute uE2 in one go
  • new extended tensor functions dg::tensor::multiply2d and dg::multiply3d that can compute uE2 in one go
  • new single step timestepper dg::ShuOsher including new dg::ShuOsherTableau and dg::ConvertsToShuOsherTableau classes to hold corresponding coefficients
  • new ShuOsher tableaus SSPRK
  • new Runge Kutta embedded tableaus tsitouras, the default timesteppers in Julia
  • new implicit RK tableau trapezoidal, implicit midpoint and sdirk-2-1-2
  • new class Simpsons that implements Simpsons rule for (time) integration
  • new implicit timesteppers dg::DIRKStep and dg::ImplicitRungeKutta
  • Redesign of multistep time steppers consistent with Runge-Kutta ones in terms of dg::MultistepTableau and dg::ConvertsToMultistepTableau
  • a host of new explicit, implicit and semi-implicit multistep tableaus
  • experimental "filtered" multistep time-steppers that allow modal filtering (first tests are not promising though)
  • new experimental multigrid solvers involving Chebyshev iterations as smoother (but none are better than nested iterations so they remain experimental)
  • new class dg::EVE that computes the largest Eigenvalue of a matrix (courtesy of Eduard Reiter)
  • new class dg::ChebyshevIterations and dg::ChebyshevPreconditioner (for chebyshev iterations)
  • new solvers dg::LGMRES, dg::BICGSTABL, and dg::AndersonAcceleration (courtesy of Aslak Poulsen)
  • new dg::FixedPointSolver and dg::AndersonSolver for nonlinear problems in time
  • a host of new functors for the evaluate and pullback functions
  • dg::geo::FluxSurfaceIntegral, dg::geo::FluxVolumeIntegral and dg::geo::SafetyFactorAverage classes
  • new implementation: dg::geo::ds_centered_bc_along_field and dg::geo::dss_centered_bc_along_field that implement boundary condition "Stegmeir" style along the magnetic field lines
  • new Fieldaligned member functions integrate_between_coarse_grid and interpolate_from_coarse_grid that allow field-aligned interpolations
  • dg::geo::Periodify class and dg::geo::periodify function to extend flux-functions periodically beyond grid boundaries
  • new dg::geo::findCriticalPoint function that generalizes X-point and O-point identification
  • new classes dg::geo::SquareNorm and dg::geo::ScalarProduct that work on cylindrical vector fields
  • new set utility functors dg::geo::SetUnion, dg::geo::SetIntersection, and dg::geo::SetNot that help construct damping regions
  • dg::geo::createMagneticField and dg::geo::createModifiedField with associated utility functions and classes that generalize the creation of magnetic flux functions and wall and sheath regions
  • new polynomial expansion and associated dg::Horner2d functor for magnetic flux functions that can in particular approximate any experimental equilibrium
  • new equilibrium, modifier and description fields for tokamak magnetic fields
  • Sign reversal of magnetic field and associated flux functions is now possible
  • new utility dg::geo::createAlignmentTensor and dg::geo::createProjectionTensor with respect to the magnetic unit vector

Changed

  • namespace file changed to dg::file and exblas changed to dg::exblas (for consistency reasons, everything should go into the dg namespace, which in particular reduces the chance for name-clashes to just one, namely 'dg')
  • Moved variation member function into dg::Elliptic (previously in ArakawaX and Poisson)
  • std=c++14 We use the C++-14 standard now (previously 11)
  • vectorclass dependency changed to vectorclass/version1 (previously we used a custom upload on feltor-dev repository)
  • default cuda compute capability bumped to sm-61 (previously sm-35)
  • marconi config now uses jsoncpp module (previously manually installed)
  • dg::blas1::dot and dg::blas2::dot and corresponding exblas functions now detect NaN and Inf errors
  • dg::blas1::dot and dg::blas2::dot now both do not accumulate rest of multiplication (inconsistent before)
  • All blas1 functions that do not read or alias their result vector now overwrite NaN and Inf
  • all our mpi communications on GPUs now fall-back to host2host communication for cuda-unaware mpi-installations
  • swapped input and output parameters in dg::blas1::evaluate first subroutine
  • the fast_interpolation and fast_projection functions now can also double / divide the polynomial coefficient consistent with the grids
  • change shift_topologic() shared RealTopology member functions to shift() and have an additional negative parameter that indicates sign swaps
  • clarify and unify the behaviour of the interpolation functions when points lie outside the grid boundaries
  • split and join functions have an additional real_type template parameter
  • improved stopping criterion for dg::bisection1d function
  • implicit time steppers give access to solver
  • multistep time-stepper now initialize with Runge-Kutta timesteppers of corresponding order
  • Multigrid nested iteration algorithm now allows accuracies for each stage separately (which can give a significant speed-up)
  • dg::inverse( bc) function is now a free-standing function to invert a boundary condition
  • dg::Elliptic classes now have jump_weighting member function
  • dg::CG operator now has a test-frequency parameter to control the number of times the error condition is evaluated
  • dg::Extrapolation class now has a derive member function to interpolate the derivative of the interpolating polynomial
  • Adapt all src and diag projects to changed file and json utilities and the moved variation member
  • Rename all input files with correct json file-ending
  • Complete redesign of src/feltor and src/lamb_dipole
  • Merge toefl_hpc with old toefl_mpi program
  • bump Doxygen version to 1.8.17
  • DS forward, backward, centered and dss functions are now free-standing, only requiring a fielaligned object, plus, and minus applications (this allows to reduce the number of times the plus and minus interpolation has to be applied)
  • changed Fieldaligned members hp_inv to hbp
  • changed name dg::forward_transform function (previously dg::create::forward_transform)
  • new dg::geo::MagneticFieldParameters struct to unify the representation of Meta-data in the dg::geo::TokamakMagneticField class (simplifies construction)

Deprecated

  • dg::Karniadakis time-stepper is now superceded by the dg::ImExMultistep class

Removed

  • remove diag/feltordiag.cu
  • remove dg::MemoryTraits and associated dimensionality and memory_category traits in favor of direct host_vector and host_grid typedefs in topology classes
  • old txt input files
  • dg::geo::DeltaFunction and dg::geo::Alpha for the computation of flux-surface averages no longer needed
  • dg::blas1::transfer (previously marked deprecated)

Fixed

  • Fix bug: race condition in dg::blas1::dot and dg::blas2::dot on GPUs that led to hard to reproduce and seemingly unreasonable crashes
  • Fix bug: std namespace in diag/probes.h
  • Fix bug: const in exblas::cpu::get_element
  • Fix bug: correct indices in exblas::cpu::make_vcl_vec8d
  • Fix bug: infinite creation of MPI communicators in exblas::mpi_reduce_communicator . Lead to MPI crashes due to memory overflow.
  • Fix bug: correct capture of cuda-aware mpi in configuration
  • Fix bug: test for no-communication in mpi_communicator.h (indicated false positives)
  • Fix bug: coefficient and initialization in dg::Extrpolate
  • Fix bug: Fpsi safety-factor in case nan is encountered still works
  • Fix bug: Fpsi safety-factor works up to the O-point
  • Fix bug: dg::pushForwardPerp on functors computed wrong result (only affects dg::geo::Hector)
  • Fix bug(s): several bugs in dg::geo::Hector which computed wrong grid (happened probably when we changed the grid design to polymorphic)
  • Fix bug: in perpendicular grid of MPI Curvlinear grid
  • Fix bug: missing direction initialization in mpi fieldaligned class
  • Fix bug: host mpi code compiled with nvcc
  • Fix bug: non-zero parallel boundary condition in mpi fieldaligned
  • Fix bug: GPU symv on X-point grids

[v5.1] Adaptive Timesteppers

Added

  • dg::Elliptic3d: a three-dimensional version of dg::Elliptic
  • Add 4 parameter symv member to dg::Elliptic class
  • dg::Helmholtz3d: a three-dimensional version of dg::Helmholtz
  • a 2nd set_chi function in Elliptic allowing a SparseTensor argument
  • new member functions for DS
  • Adaptive timesteppers through the dg::Adaptive and dg::ERKStep and dg::ARKStep classes in inc/dg/adaptive.h and tests in inc/dg/adaptive_t.cu
  • dg::ButcherTableau and dg::ConvertsToButcherTableau in inc/dg/tableau.h
  • Add interface to Extrapolation class so that it can take a time variable
  • dg::View vector view class in inc/dg/view.h usable in dg::blas1 and dg::blas2 functions (tested in inc/dg/view_t.cu)
  • introduce dg::assign and dg::construct that replace the dg::transfer functions and have increased flexibility through an additional template parameter pack
  • dg::tensor::inv_multiply3d and dg::tensor::inv_multiply2d functions
  • Add unified typedefs for geometry, matrix, container and value types in dg library classes
  • Add more ContainerType template parameters to ArakawaX, CG, Invert, Elliptic and Helmholtz class members
  • Add set_chi member to ArakawaX to allow for an optional prefactor (like the magnetic field)
  • Additional constructors in BijectiveComm consistent with SujectiveComm
  • Add consistency test to mpi_collective_mpit.cu
  • copyable member function in CG, Invert, RungeKutta, AdamsBashforth and Karniadakis
  • float overloads of dg::one and dg::zero functions
  • Add trivial 2d and 3d operator() for most 1d functors in functors.h
  • New subroutines in subroutines.h
  • Add shift_topologic member to RealTopology overriding boundary conditions

Changed

  • Redesign: dg::AB now named dg::AdamsBashforth and stages not a template parameter any more (changed src/lamb_dipole/shu_time.cu accordingly)
  • Redesign: dg::RK now named dg::RungeKutta and stages not a template parameter any more
  • stepperRK does not need a stage template parameter any more
  • Redesign dg::SparseTensor and all associated dg::tensor functions; (Not allowing unset values significantly reduces complexity and especially allows easy and fast implementation of dg::tensor::multiply2d and dg::tensor::multiply3d in terms of dg::blas1::subroutine)
  • Change construction of Metric and Jacobian in grid geometries and implementation of dg::pushForward transformation functions
  • Change grids() method interface of Multigrid class to grid(unsigned) and adapt this in all programs using it (now we completely avoid exposing ClonePtr to users)
  • Reimplement and merge src/feltor with src/asela
  • Reimplement src/heat project
  • experimental use of adaptive timestepper in src/toeflR.cu and src/toeflR.cuh
  • change folder name from inc/dg/geometry to inc/dg/topology (mainly to avoid clashes with the new inc/dg/geometries) and adapt include directives accordingly
  • dg::split now creates vector views instead of hard copies
  • default constructor of MPI_Vector constructs empty communicator instead of MPI_COMM_WORLD
  • set_communicator in MPI_Vector takes three arguments now to avoid group creation
  • cuda-aware mpi no longer a requirement, fall-back to traditional mpi implemented
  • rewrite feltordiag.cu merging ncdiag and filamentdiag
  • Remove container argument from dg::geo::SafetyFactor constructor (Since it always takes Alpha)
  • More general interface for geometries/init.h functors including Nprofile and ZonalFlow (Old Psi functors are now regular functors)
  • Rename BinaryFunctors to CylindricalFunctors (they are not only binary after all) and change files in inc/geometries accordingly
  • Rewrite dg::ClonePtr as a true pointer class and adapt memory_t.cu
  • dg::pushForwardPerp for tensors uses SparseTensor as output argument
  • interface for DS
  • boundary condition implementation of Fieldaligned (reflected in parallel.pdf)
  • boundary conditions in interpolation functions due to Fieldaligned
  • better test programs ds_t.cu, ds_curv_t.cu and ds_guenther_t.cu including corresponding *mpit.cu programs
  • FieldR, FieldZ and FieldP are now named BFieldR, BFieldZ and BFieldP
  • change default optimization to -O2 due to gcc-7 bug
  • replace use of dg::transfer with either dg::assign or dg::construct throughout the dg library
  • improve 3d benchmark functionality in cluster_mpib.cu
  • add hasXpoint() and default parameters to solovev::parameter
  • add map_index to NearestNeighborComm to signify responsibility of mapping matrix indices to buffer indices
  • Optimization: implement fast EllSparseBlockMat kernel for z derivative
  • Optimization: change buffer layout in dg::NearestNeighborComm and CooSparseBlockMat kernels to avoid slow scatter/gather operations in mpi matrix-vector multiplication
  • Optimization: implement faster kernels for CooSparseBlockMat symv kernel to accelerate mpi symv with low computation to communication ratio
  • separate modification of fluxfunctions into mod namespace that works on flux functions in general (previously only solovev)

Deprecated

  • dg::blas1::transfer (replaced by the more general dg::assign and dg::construct)
  • the header geometries/geometries.h (now dg/geometries/geometries.h is preferred for unified access and easier recognition as a dg header file)
  • the header file/nc_utilities.h ( now dg/file/nc_utilities.h is preferred for unified access and easier recognition as a dg header file)
  • the Helmholtz2 class (it's much faster to use Helmholtz twice)

Removed

  • diag/filamentdiag.cu and diag/ncdiag.cpp (merged into feltordiag.cu)
  • src/asela and src/asela2d (now merged into src/feltor and src/feltor2d)
  • src/feltor2D (now merged in src/feltor project as an input parameter)
  • src/feltor/feltor.cu (not really useful any more with our adatpive Timesteppers as we don't need to find suitable timesteps anymore)
  • src/feltor/feltor_mpi.cu (now merged into feltor_hpc.cu)
  • BinaryFunctorAdapter and aCloneableBinaryFunctor ( we now use the better CylindricalFunctor based on std::function to do the type erasure)
  • TensorElliptic, GeneralElliptic and GeneralEllipticSym classes ( superceded by Elliptic3d)
  • inc/geometries/ds_geom_t.cu and ds_guenther_b.cu (superceded by better test programs)
  • dg::SparseElement (it's more intuitive and practical to use a container directly and the performance gain is negligeable)
  • src/heat/init.h
  • dg::join (no longer necessary because split works on views)

Fixed

  • fix bug in diag/reco2Ddiag.cu regarding DVec in interpolation
  • fix bug in src/reco2D/reconnection_mpi.cu regarding output
  • fix bug in src/polar/polar.cu regarding DVec in interpolation function
  • fix bug of too many constructor parameters in BathRZ constructor in src/feltorShw/feltur.cu, feltorShw/feltor_hpc.cu, feltorShw/feltor_mpi.cu, inc/dg/bathRZ_t.cu
  • fix bug in src/feltorSHp/feltor_hpc.cu regarding step function in timestepper
  • fix bug in inc/dg/backend/exblas/mylibm.hpp regarding msvc preprocessor
  • fix bug in dg::Histogramm: remove DG_DEVICE from operator()
  • change parameter type in members of aCommunicator and NearesNeighborComm to pointer types to allow different container types in blas2 functions for mpi vectors too
  • fix bug in GeneralComm global_scatter_reduce member initializing values to zero and adding local_size to constructor
  • geometries functors can now be used as std::function (by removing the clone method)
  • geometries/geometries.h and file/nc_utilities.h headers are now also includable by multiple translation units