Tasmanian

--------------

* added AMD HIP/ROCm backend
* all CUDA capabilities have an AMD HIP/ROCm equivalent
* this includes MAGMA-HIP capabilities
* capabilities are accessed with accel_gpu_cuda and accel_gpu_cublas
* removed the XSDK names for options
* the XSDK requirement was alleviated due to spack
* CMake now builds only one type of library shared or static
* matlab interface works well with shared libs thanks to rpaths
* magma search is simpler looking for either shared or static libs
* ::shared and ::static targets are deprecated

--------------

* the Python pip installer can now enable CUDA, MAGMA, MPI, etc.
* requires setting environment variables

* the Python pip installer now handles dependencies automatically
* requires an up-to-date version of pip

* added the API to manually set the CUDA handles
* manual handles can be set for cuBlas, cuSparse and cuSolverDn

* new addon method for constructing a surrogate from unstructured data
* utilizes QR factorization from cuSolver and MAGMA
* using out-of-core implementation within MAGMA

* an LAPACK implementation is required for the BLAS option
* BLAS and LAPACK are packaged together in all tested libraries

* added Swig generated Fortran 2003 interface
* naming conventions and objects mimic C++, similar to python

* split the Tasmanian::Tasmanian and Tasmanian::Fortran targets
* also added Tasmanian::Fortran::shared and Tasmanian::Fortran::static

* added mixed-precision templates for
* `evaluateBatchGPU()` and `evaluateHierarchicalFunctionsGPU()`
* `evaluateBatch()` when the acceleration mode is cuda or magma

--------------

* improved MPI capability, see the updated documentation
* fully automated distributed adaptive sparse grid construction

* new module "Tasmanian Addons" consisting of miscellaneous templates
* algorithms for automated sampling from a lambda model
* use through the master target and header, cannot be used independently

* new Tasmanian CMake master targets that handle all modules
* now available Tasmanian::shared and Tasmanian::static
* also Tasmanian::Tasmanian will always point to an available target
* added Tasmanian::tasgrid imported executable target

* new Tasmanian master Python module: `import Tasmanian`
* the python changes are backwards compatible, the old module is still there
* the sparse grid class can be called with `grid = Tasmanian.SparseGrid()`
* the DREAM binding is available through `Tasmanian.DREAM`

* new Tasmanian master header to add all modules: Tasmanian.hpp

* new dynamic construction algorithms have been added
* removed the need for blocking between refinement iterations
* Tasmanian can accept data one-point at a time
* new candidate points can be requested at any time
* candidate points for dynamic construction are weighted by "importance"
* the dynamic construction process is available through C++ and Python interfaces

* modernized C++ compatibility
* see the updated DREAM api notes
* TasmanianSparseGrid has move and copy constructors and `operator=` overloads
* Tasmanian C++ API no-longer returns raw-pointers, only STL containers
* **broke backward api** affects mostly get points and weights
* externally allocated raw-pointers are still accepted and used by C/Python/Fortran APIs

* updated the Fortran interface, the library no longer has a global state
* replaced the integer grid-ids with a derived type(TasmanianSparseGrid)
* grid variables now require tsgAllocateGrid(grid) and tsgDeallocateGrid(grid)
* tsgAllocateGrid() is now thread safe (when called on different grid variables)
* no other changes appear on the front-end, see the Fortran examples

* improved the `add_subdirectory()` capability
* can specify the export name used by the Tasmanian install commands
* `set(Tasmanian_export_name <name> CACHE INTERNAL "")`
* `add_subdirectory(<path-to-Tasmanian-source> <work-folder>)`
* `install(EXPORT ${Tasmanian_export_name} ...)`
* the export name functionality is required to import the transitive dependencies
* in addition, when using `add_subdirectory()`:
* will not enable testing, tests are still set if enabled by the master project
* will not install package-config, that is the job of the master project
* will not install example CMakeLists.txt or post-install tests
* (`make test_install` and examples require Tasmanian package-config)

* updated the DREAM interface:
* excessive polymorphism is replaced by lambdas
* sampling is done by a template
* can specify arbitrary domain (using lambdas)
* the new API is also available though Python
* Note: the old API is completely obsolete and incompatible

* added Doxygen documentation (CMake option, extra pages, etc.)

* improved CUDA support
* added support for CUDA 10
* all grids now benefit from all acceleration modes
* added `evaluateBatchGPU()` where both `x` and `y` sit on the GPU

* removed the deprecated MS Windows build system with batch scripts

* require cmake 3.10 or newer
* removed the clumsy work-around for OpenMP on old cmake systems
* removed other small legacy cmake fixes
* added cuda as a cmake lang, added multiple hacks to avoid feature regression

--------------

* as always, a new version includes numerous bug fixes and performance enhancements

* CXX standard 2011 is now required and enabled by default

* required cmake 3.5 or newer (as opposed to 2.8 in version 5.1)

* merged Tasmanian_ENABLE_CUBLAS option into Tasmanian_ENABLE_CUDA
* the option is OFF by default
* two distinct acceleration modes use this variable: `gpu-cublas` and `gpu-cuda`

* added new acceleration mode `gpu-magma` that uses UTK MAGMA library (requires CUDA)

* removed Tasmanian_STRICT_OPTIONS, now all options are considered strict by default

* new option Tasmanian_ENABLE_RECOMMENDED
* searches for OpenMP, BLAS, and Python, and enables if found
* set the `-O3` flag for Debug and Release
* adjusted the install script, see `./install --help`

* modified install folder structure now everything sits in four places
* `<prefix>/bin` takes the executable files
* `<prefix>/lib` takes the libraries and Fortran `.mod` files
* `<prefix>/lib/Tasmanian` takes the cmake package-config
* `<prefix>/lib/PythonX.Y` takes the python module
* `<prefix>/include` takes the headers
* `<prefix>/share/Tasmanian` takes everything else

* added package-config file, now it is possible to use the command

find_package(Tasmanian 6.0 PATHS "<Tasmanian install prefix>")
PATHS "<Tasmanian install prefix>" not needed if the install folder
is included in CMAKE_PREFIX_PATH


* added new grids and rules
* `GridFourier` that uses trigonometric basis functions (see Manual)
* `localpb` rule to local polynomial grids that favors the boundary

* API change for error handling (not backward compatible)
* errors in C++ now throw exceptions
* `std::runtime_error` is thrown on wrong file format
* `std::runtime_error` is thrown when requesting level too high
* `std::invalid_argument` is thrown when calling function for wrong grid
* `logstream` is no longer needed and has been removed
* the error handling in the interface calls remains the same as before

* updated C++ API with overloaded functions
* using `std::vector<int/double>` as opposed to arrays
* error checking is provided based on the vector size
* massive portion of the internal API has also changed

* acceleration with custom CUDA kernels is now available
* Sequence grids, i.e., build with `makeSequenceGrid()`
* Fourier grids, i.e., build with `makeFourierGrid()`

* the library is more robust when working with large data sets
* number of points, outputs, or batch size is still `int`
* in most cases the product (i.e., matrix size) can exceed `MAX_INT`

* improved MS Windows support using cmake
* recommended optimization flags are properly set (Debug and Release)
* shared symbols are exported by cmake
* CUDA is now supported under Windows
* `getGPUName()` under Python currently does not work
* the old `WindowsMake.bat` script is deprecated

--------------

* visit us on github.com/ORNL/TASMANIAN

* added functionality to access hierarchical functions and coefficients
which allows constructing model approximation from an arbitrary
cloud of samples, as opposed to requiring the use of the specific
points in the hypercube

* added custom CUDA kernels for evaluations for LocalPolynomial grids,
both sparse/dense basis of hierarchical functions and batch evaluations
can be performed with the CUDA kernels for extra speed.
evaluateFast for LocalPolynomial also benefits from acceleration

* grids can be written/read in either binary or ascii format,
MATLAB defaults to binary format now

* CUDA accelerated evaluations can be triggered from MATLAB,
just specify `lGrid.gpuDevice = X,` where X is a valid CUDA device
run `tsgCoreTests()` to see a list of the available devices

* added tsgCoreTests() script to MATLAB

* added tweaks and simplifications to the build system,
the main options remain the same, but check the manual for changes
to the additional options

* added Fortran 90/95 interface (still very EXPERIMENTAL)

* significantly expanded the automated testing procedures and code
coverage

* numerous bug fixes and performance enhancements