1# Changes/Release Notes 2 3On this page we provide a summary of the main API changes, new features and examples for each release of libCEED. 4 5(main)= 6 7## Current `main` branch 8 9### New features 10- Added {c:func}`CeedOperatorGetFieldByName` to access a specific `CeedOperatorField` by its name 11 12(v0-11)= 13 14## v0.11 (Dec 24, 2022) 15 16### Interface changes 17 18- Added {c:func}`CeedOperatorSetName` for more readable {c:func}`CeedOperatorView` output. 19- Added {c:func}`CeedBasisCreateProjection` to facilitate interpolation between nodes for separate `CeedBases`. 20- Rename and move {c:func}`CeedCompositeOperatorGetNumSub` and {c:func}`CeedCompositeOperatorGetSubList` to public interface. 21 22### New features 23 24- Update `/cpu/self/memcheck/*` backends to help verify `CeedQFunctionContext` data sizes provided by user. 25- Improved support for $H(\text{div})$ bases. 26- Added `CeedInt_FMT` to support potential future use of larger integer sizes. 27- Added `CEED_QFUNCTION_ATTR` for setting compiler attributes/pragmas to `CEED_QFUNCTION_HELPER` and `CEED_QFUNCTION`. 28- OCCA backend updated to latest OCCA release; DPC++ and OMP OCCA modes enabled. 29Due to a limitation of the OCCA parser, typedefs are required to use pointers to arrays in QFunctions with the OCCA backend. 30This issue will be fixed in a future OCCA release. 31 32### Bugfix 33 34- Fix bug in setting device id for GPU backends. 35- Fix storing of indices for `CeedElemRestriction` on the host with GPU backends. 36- Fix `CeedElemRestriction` sizing for {c:func}`CeedOperatorAssemblePointBlockDiagonal`. 37- Fix bugs in CPU implementation of {c:func}`CeedOperatorLinearAssemble` when there are different number of active input modes and active output modes. 38 39### Examples 40 41#### {ref}`example-petsc-navier-stokes` 42 43- Various performance enhancements, analytic matrix-free and assembled Jacobian, and PETSc solver configurations for GPUs. 44- Refactored to improve code reuse and modularity. 45- Support for primitive variables for more accurate boundary layers and all-speed flow. 46- Added $YZ\beta$ shock capturing scheme and Shock Tube example. 47- Added Channel example, with comparison to analytic solutions. 48- Added Flat Plate with boundary layer mesh and compressible Blasius inflow condition based on Chebyshev collocation solution of the Blasius equations. 49- Added strong and weak synthetic turbulence generation (STG) inflow boundary conditions. 50- Added "freestream" boundary conditions based on HLLC Riemann solver. 51- Automated stabilization coefficients for different basis degree. 52 53#### {ref}`example-petsc-bps` 54 55- Support for convergence studies. 56 57### Maintainability 58 59- Refactored `/gpu/cuda/shared` and `/gpu/cuda/gen` as well as `/gpu/hip/shared` and `/gpu/hip/gen` backend to improve maintainablity and reduce duplicated code. 60- Enabled support for `p > 8` for `/gpu/*/shared` backends. 61- Switch to `clang-format` over `astyle` for automatic formatting; Makefile command changed to `make format` from `make style`. 62- Improved test harness. 63 64(v0-10-1)= 65 66## v0.10.1 (Apr 11, 2022) 67 68### Interface changes 69 70- Added {c:func}`CeedQFunctionSetUserFlopsEstimate` and {c:func}`CeedOperatorGetFlopsEstimate` to facilitate estimating FLOPs in operator application. 71 72### New features 73 74- Switched MAGMA backends to use runtime compilation for tensor basis kernels (and element restriction kernels, in non-deterministic `/gpu/*/magma` backends). 75This reduces time to compile the library and increases the range of parameters for which the MAGMA tensor basis kernels will work. 76 77### Bugfix 78 79- Install JiT source files in install directory to fix GPU functionality for installed libCEED. 80 81(v0-10)= 82 83## v0.10 (Mar 21, 2022) 84 85### Interface changes 86 87- Update {c:func}`CeedQFunctionGetFields` and {c:func}`CeedOperatorGetFields` to include number of fields. 88- Promote to the public API: QFunction and Operator field objects, `CeedQFunctionField` and `CeedOperatorField`, and associated getters, {c:func}`CeedQFunctionGetFields`; {c:func}`CeedQFunctionFieldGetName`; {c:func}`CeedQFunctionFieldGetSize`; {c:func}`CeedQFunctionFieldGetEvalMode`; {c:func}`CeedOperatorGetFields`; {c:func}`CeedOperatorFieldGetElemRestriction`; {c:func}`CeedOperatorFieldGetBasis`; and {c:func}`CeedOperatorFieldGetVector`. 89- Clarify and document conditions where `CeedQFunction` and `CeedOperator` become immutable and no further fields or suboperators can be added. 90- Add {c:func}`CeedOperatorLinearAssembleQFunctionBuildOrUpdate` to reduce object creation overhead in assembly of CeedOperator preconditioning ingredients. 91- Promote {c:func}`CeedOperatorCheckReady`to the public API to facilitate interactive interfaces. 92- Warning added when compiling OCCA backend to alert users that this backend is experimental. 93- `ceed-backend.h`, `ceed-hash.h`, and `ceed-khash.h` removed. Users should use `ceed/backend.h`, `ceed/hash.h`, and `ceed/khash.h`. 94- Added {c:func}`CeedQFunctionGetKernelName`; refactored {c:func}`CeedQFunctionGetSourcePath` to exclude function kernel name. 95- Clarify documentation for {c:func}`CeedVectorTakeArray`; this function will error if {c:func}`CeedVectorSetArray` with `copy_mode == CEED_USE_POINTER` was not previously called for the corresponding `CeedMemType`. 96- Added {c:func}`CeedVectorGetArrayWrite` that allows access to uninitalized arrays; require initalized data for {c:func}`CeedVectorGetArray`. 97- Added {c:func}`CeedQFunctionContextRegisterDouble` and {c:func}`CeedQFunctionContextRegisterInt32` with {c:func}`CeedQFunctionContextSetDouble` and {c:func}`CeedQFunctionContextSetInt32` to facilitate easy updating of {c:struct}`CeedQFunctionContext` data by user defined field names. 98- Added {c:func}`CeedQFunctionContextGetFieldDescriptions` to retreive user defined descriptions of fields that are registered with `CeedQFunctionContextRegister*`. 99- Renamed `CeedElemTopology` entries for clearer namespacing between libCEED enums. 100- Added type `CeedSize` equivalent to `ptrdiff_t` for array sizes in {c:func}`CeedVectorCreate`, {c:func}`CeedVectorGetLength`, `CeedElemRestrictionCreate*`, {c:func}`CeedElemRestrictionGetLVectorSize`, and {c:func}`CeedOperatorLinearAssembleSymbolic`. This is a breaking change. 101- Added {c:func}`CeedOperatorSetQFunctionUpdated` to facilitate QFunction data re-use between operators sharing the same quadrature space, such as in a multigrid hierarchy. 102- Added {c:func}`CeedOperatorGetActiveVectorLengths` to get shape of CeedOperator. 103 104### New features 105 106- `CeedScalar` can now be set as `float` or `double` at compile time. 107- Added JiT utilities in `ceed/jit-tools.h` to reduce duplicated code in GPU backends. 108- Added support for JiT of QFunctions with `#include "relative/path/local-file.h"` statements for additional local files. Note that files included with `""` are searched relative to the current file first, then by compiler paths (as with `<>` includes). To use this feature, one should adhere to relative paths only, not compiler flags like `-I`, which the JiT will not be aware of. 109- Remove need to guard library headers in QFunction source for code generation backends. 110- `CeedDebugEnv()` macro created to provide debugging outputs when Ceed context is not present. 111- Added {c:func}`CeedStringAllocCopy` to reduce repeated code for copying strings internally. 112- Added {c:func}`CeedPathConcatenate` to facilitate loading kernel source files with a path relative to the current file. 113- Added support for non-tensor H(div) elements, to include CPU backend implementations and {c:func}`CeedBasisCreateHdiv` convenience constructor. 114- Added {c:func}`CeedQFunctionSetContextWritable` and read-only access to `CeedQFunctionContext` data as an optional feature to improve GPU performance. By default, calling the `CeedQFunctionUser` during {c:func}`CeedQFunctionApply` is assumed to write into the `CeedQFunctionContext` data, consistent with the previous behavior. Note that if a user asserts that their `CeedQFunctionUser` does not write into the `CeedQFunctionContext` data, they are responsible for the validity of this assertion. 115- Added support for element matrix assembly in GPU backends. 116 117### Maintainability 118 119- Refactored preconditioner support internally to facilitate future development and improve GPU completeness/test coverage. 120- `Include-what-you-use` makefile target added as `make iwyu`. 121- Create backend constant `CEED_FIELD_MAX` to reduce magic numbers in codebase. 122- Put GPU JiTed kernel source code into separate files. 123- Dropped legacy version support in PETSc based examples to better utilize PETSc DMPlex and Mat updates to support libCEED; current minimum PETSc version for the examples is v3.17. 124 125(v0-9)= 126 127## v0.9 (Jul 6, 2021) 128 129### Interface changes 130 131- Minor modification in error handling macro to silence pedantic warnings when compiling with Clang, but no functional impact. 132 133### New features 134 135- Add {c:func}`CeedVectorAXPY` and {c:func}`CeedVectorPointwiseMult` as a convenience for stand-alone testing and internal use. 136- Add `CEED_QFUNCTION_HELPER` macro to properly annotate QFunction helper functions for code generation backends. 137- Add `CeedPragmaOptimizeOff` macro for code that is sensitive to floating point errors from fast math optimizations. 138- Rust support: split `libceed-sys` crate out of `libceed` and [publish both on crates.io](https://crates.io/crates/libceed). 139 140### Performance improvements 141 142### Examples 143 144- Solid mechanics mini-app updated to explore the performance impacts of various formulations in the initial and current configurations. 145- Fluid mechanics example adds GPU support and improves modularity. 146 147### Deprecated backends 148 149- The `/cpu/self/tmpl` and `/cpu/self/tmpl/sub` backends have been removed. These backends were intially added to test the backend inheritance mechanism, but this mechanism is now widely used and tested in multiple backends. 150 151(v0-8)= 152 153## v0.8 (Mar 31, 2021) 154 155### Interface changes 156 157- Error handling improved to include enumerated error codes for C interface return values. 158- Installed headers that will follow semantic versioning were moved to {code}`include/ceed` directory. These headers have been renamed from {code}`ceed-*.h` to {code}`ceed/*.h`. Placeholder headers with the old naming schema are currently provided, but these headers will be removed in the libCEED v0.9 release. 159 160### New features 161 162- Julia and Rust interfaces added, providing a nearly 1-1 correspondence with the C interface, plus some convenience features. 163- Static libraries can be built with `make STATIC=1` and the pkg-config file is installed accordingly. 164- Add {c:func}`CeedOperatorLinearAssembleSymbolic` and {c:func}`CeedOperatorLinearAssemble` to support full assembly of libCEED operators. 165 166### Performance improvements 167 168- New HIP MAGMA backends for hipMAGMA library users: `/gpu/hip/magma` and `/gpu/hip/magma/det`. 169- New HIP backends for improved tensor basis performance: `/gpu/hip/shared` and `/gpu/hip/gen`. 170 171### Examples 172 173- {ref}`example-petsc-elasticity` example updated with traction boundary conditions and improved Dirichlet boundary conditions. 174- {ref}`example-petsc-elasticity` example updated with Neo-Hookean hyperelasticity in current configuration as well as improved Neo-Hookean hyperelasticity exploring storage vs computation tradeoffs. 175- {ref}`example-petsc-navier-stokes` example updated with isentropic traveling vortex test case, an analytical solution to the Euler equations that is useful for testing boundary conditions, discretization stability, and order of accuracy. 176- {ref}`example-petsc-navier-stokes` example updated with support for performing convergence study and plotting order of convergence by polynomial degree. 177 178(v0-7)= 179 180## v0.7 (Sep 29, 2020) 181 182### Interface changes 183 184- Replace limited {code}`CeedInterlaceMode` with more flexible component stride {code}`compstride` in {code}`CeedElemRestriction` constructors. 185 As a result, the {code}`indices` parameter has been replaced with {code}`offsets` and the {code}`nnodes` parameter has been replaced with {code}`lsize`. 186 These changes improve support for mixed finite element methods. 187- Replace various uses of {code}`Ceed*Get*Status` with {code}`Ceed*Is*` in the backend API to match common nomenclature. 188- Replace {code}`CeedOperatorAssembleLinearDiagonal` with {c:func}`CeedOperatorLinearAssembleDiagonal` for clarity. 189- Linear Operators can be assembled as point-block diagonal matrices with {c:func}`CeedOperatorLinearAssemblePointBlockDiagonal`, provided in row-major form in a {code}`ncomp` by {code}`ncomp` block per node. 190- Diagonal assemble interface changed to accept a {ref}`CeedVector` instead of a pointer to a {ref}`CeedVector` to reduce memory movement when interfacing with calling code. 191- Added {c:func}`CeedOperatorLinearAssembleAddDiagonal` and {c:func}`CeedOperatorLinearAssembleAddPointBlockDiagonal` for improved future integration with codes such as MFEM that compose the action of {ref}`CeedOperator`s external to libCEED. 192- Added {c:func}`CeedVectorTakeAray` to sync and remove libCEED read/write access to an allocated array and pass ownership of the array to the caller. 193 This function is recommended over {c:func}`CeedVectorSyncArray` when the {code}`CeedVector` has an array owned by the caller that was set by {c:func}`CeedVectorSetArray`. 194- Added {code}`CeedQFunctionContext` object to manage user QFunction context data and reduce copies between device and host memory. 195- Added {c:func}`CeedOperatorMultigridLevelCreate`, {c:func}`CeedOperatorMultigridLevelCreateTensorH1`, and {c:func}`CeedOperatorMultigridLevelCreateH1` to facilitate creation of multigrid prolongation, restriction, and coarse grid operators using a common quadrature space. 196 197### New features 198 199- New HIP backend: `/gpu/hip/ref`. 200- CeedQFunction support for user `CUfunction`s in some backends 201 202### Performance improvements 203 204- OCCA backend rebuilt to facilitate future performance enhancements. 205- Petsc BPs suite improved to reduce noise due to multiple calls to {code}`mpiexec`. 206 207### Examples 208 209- {ref}`example-petsc-elasticity` example updated with strain energy computation and more flexible boundary conditions. 210 211### Deprecated backends 212 213- The `/gpu/cuda/reg` backend has been removed, with its core features moved into `/gpu/cuda/ref` and `/gpu/cuda/shared`. 214 215(v0-6)= 216 217## v0.6 (Mar 29, 2020) 218 219libCEED v0.6 contains numerous new features and examples, as well as expanded 220documentation in [this new website](https://libceed.org). 221 222### New features 223 224- New Python interface using [CFFI](https://cffi.readthedocs.io/) provides a nearly 225 1-1 correspondence with the C interface, plus some convenience features. For instance, 226 data stored in the {cpp:type}`CeedVector` structure are available without copy as 227 {py:class}`numpy.ndarray`. Short tutorials are provided in 228 [Binder](https://mybinder.org/v2/gh/CEED/libCEED/main?urlpath=lab/tree/examples/tutorials/). 229- Linear QFunctions can be assembled as block-diagonal matrices (per quadrature point, 230 {c:func}`CeedOperatorAssembleLinearQFunction`) or to evaluate the diagonal 231 ({c:func}`CeedOperatorAssembleLinearDiagonal`). These operations are useful for 232 preconditioning ingredients and are used in the libCEED's multigrid examples. 233- The inverse of separable operators can be obtained using 234 {c:func}`CeedOperatorCreateFDMElementInverse` and applied with 235 {c:func}`CeedOperatorApply`. This is a useful preconditioning ingredient, 236 especially for Laplacians and related operators. 237- New functions: {c:func}`CeedVectorNorm`, {c:func}`CeedOperatorApplyAdd`, 238 {c:func}`CeedQFunctionView`, {c:func}`CeedOperatorView`. 239- Make public accessors for various attributes to facilitate writing composable code. 240- New backend: `/cpu/self/memcheck/serial`. 241- QFunctions using variable-length array (VLA) pointer constructs can be used with CUDA 242 backends. (Single source is coming soon for OCCA backends.) 243- Fix some missing edge cases in CUDA backend. 244 245### Performance Improvements 246 247- MAGMA backend performance optimization and non-tensor bases. 248- No-copy optimization in {c:func}`CeedOperatorApply`. 249 250### Interface changes 251 252- Replace {code}`CeedElemRestrictionCreateIdentity` and 253 {code}`CeedElemRestrictionCreateBlocked` with more flexible 254 {c:func}`CeedElemRestrictionCreateStrided` and 255 {c:func}`CeedElemRestrictionCreateBlockedStrided`. 256- Add arguments to {c:func}`CeedQFunctionCreateIdentity`. 257- Replace ambiguous uses of {cpp:enum}`CeedTransposeMode` for L-vector identification 258 with {cpp:enum}`CeedInterlaceMode`. This is now an attribute of the 259 {cpp:type}`CeedElemRestriction` (see {c:func}`CeedElemRestrictionCreate`) and no 260 longer passed as `lmode` arguments to {c:func}`CeedOperatorSetField` and 261 {c:func}`CeedElemRestrictionApply`. 262 263### Examples 264 265libCEED-0.6 contains greatly expanded examples with {ref}`new documentation <Examples>`. 266Notable additions include: 267 268- Standalone {ref}`ex2-surface` ({file}`examples/ceed/ex2-surface`): compute the area of 269 a domain in 1, 2, and 3 dimensions by applying a Laplacian. 270 271- PETSc {ref}`example-petsc-area` ({file}`examples/petsc/area.c`): computes surface area 272 of domains (like the cube and sphere) by direct integration on a surface mesh; 273 demonstrates geometric dimension different from topological dimension. 274 275- PETSc {ref}`example-petsc-bps`: 276 277 - {file}`examples/petsc/bpsraw.c` (formerly `bps.c`): transparent CUDA support. 278 - {file}`examples/petsc/bps.c` (formerly `bpsdmplex.c`): performance improvements 279 and transparent CUDA support. 280 - {ref}`example-petsc-bps-sphere` ({file}`examples/petsc/bpssphere.c`): 281 generalizations of all CEED BPs to the surface of the sphere; demonstrates geometric 282 dimension different from topological dimension. 283 284- {ref}`example-petsc-multigrid` ({file}`examples/petsc/multigrid.c`): new p-multigrid 285 solver with algebraic multigrid coarse solve. 286 287- {ref}`example-petsc-navier-stokes` ({file}`examples/fluids/navierstokes.c`; formerly 288 `examples/navier-stokes`): unstructured grid support (using PETSc's `DMPlex`), 289 implicit time integration, SU/SUPG stabilization, free-slip boundary conditions, and 290 quasi-2D computational domain support. 291 292- {ref}`example-petsc-elasticity` ({file}`examples/solids/elasticity.c`): new solver for 293 linear elasticity, small-strain hyperelasticity, and globalized finite-strain 294 hyperelasticity using p-multigrid with algebraic multigrid coarse solve. 295 296(v0-5)= 297 298## v0.5 (Sep 18, 2019) 299 300For this release, several improvements were made. Two new CUDA backends were added to 301the family of backends, of which, the new `cuda-gen` backend achieves state-of-the-art 302performance using single-source {ref}`CeedQFunction`. From this release, users 303can define Q-Functions in a single source code independently of the targeted backend 304with the aid of a new macro `CEED QFUNCTION` to support JIT (Just-In-Time) and CPU 305compilation of the user provided {ref}`CeedQFunction` code. To allow a unified 306declaration, the {ref}`CeedQFunction` API has undergone a slight change: 307the `QFunctionField` parameter `ncomp` has been changed to `size`. This change 308requires setting the previous value of `ncomp` to `ncomp*dim` when adding a 309`QFunctionField` with eval mode `CEED EVAL GRAD`. 310 311Additionally, new CPU backends 312were included in this release, such as the `/cpu/self/opt/*` backends (which are 313written in pure C and use partial **E-vectors** to improve performance) and the 314`/cpu/self/ref/memcheck` backend (which relies upon the 315[Valgrind](http://valgrind.org/) Memcheck tool to help verify that user 316{ref}`CeedQFunction` have no undefined values). 317This release also included various performance improvements, bug fixes, new examples, 318and improved tests. Among these improvements, vectorized instructions for 319{ref}`CeedQFunction` code compiled for CPU were enhanced by using `CeedPragmaSIMD` 320instead of `CeedPragmaOMP`, implementation of a {ref}`CeedQFunction` gallery and 321identity Q-Functions were introduced, and the PETSc benchmark problems were expanded 322to include unstructured meshes handling were. For this expansion, the prior version of 323the PETSc BPs, which only included data associated with structured geometries, were 324renamed `bpsraw`, and the new version of the BPs, which can handle data associated 325with any unstructured geometry, were called `bps`. Additionally, other benchmark 326problems, namely BP2 and BP4 (the vector-valued versions of BP1 and BP3, respectively), 327and BP5 and BP6 (the collocated versions---for which the quadrature points are the same 328as the Gauss Lobatto nodes---of BP3 and BP4 respectively) were added to the PETSc 329examples. Furthermoew, another standalone libCEED example, called `ex2`, which 330computes the surface area of a given mesh was added to this release. 331 332Backends available in this release: 333 334| CEED resource (`-ceed`) | Backend | 335|--------------------------|-----------------------------------------------------| 336| `/cpu/self/ref/serial` | Serial reference implementation | 337| `/cpu/self/ref/blocked` | Blocked reference implementation | 338| `/cpu/self/ref/memcheck` | Memcheck backend, undefined value checks | 339| `/cpu/self/opt/serial` | Serial optimized C implementation | 340| `/cpu/self/opt/blocked` | Blocked optimized C implementation | 341| `/cpu/self/avx/serial` | Serial AVX implementation | 342| `/cpu/self/avx/blocked` | Blocked AVX implementation | 343| `/cpu/self/xsmm/serial` | Serial LIBXSMM implementation | 344| `/cpu/self/xsmm/blocked` | Blocked LIBXSMM implementation | 345| `/cpu/occa` | Serial OCCA kernels | 346| `/gpu/occa` | CUDA OCCA kernels | 347| `/omp/occa` | OpenMP OCCA kernels | 348| `/ocl/occa` | OpenCL OCCA kernels | 349| `/gpu/cuda/ref` | Reference pure CUDA kernels | 350| `/gpu/cuda/reg` | Pure CUDA kernels using one thread per element | 351| `/gpu/cuda/shared` | Optimized pure CUDA kernels using shared memory | 352| `/gpu/cuda/gen` | Optimized pure CUDA kernels using code generation | 353| `/gpu/magma` | CUDA MAGMA kernels | 354 355Examples available in this release: 356 357:::{list-table} 358:header-rows: 1 359:widths: auto 360* - User code 361 - Example 362* - `ceed` 363 - * ex1 (volume) 364 * ex2 (surface) 365* - `mfem` 366 - * BP1 (scalar mass operator) 367 * BP3 (scalar Laplace operator) 368* - `petsc` 369 - * BP1 (scalar mass operator) 370 * BP2 (vector mass operator) 371 * BP3 (scalar Laplace operator) 372 * BP4 (vector Laplace operator) 373 * BP5 (collocated scalar Laplace operator) 374 * BP6 (collocated vector Laplace operator) 375 * Navier-Stokes 376* - `nek5000` 377 - * BP1 (scalar mass operator) 378 * BP3 (scalar Laplace operator) 379::: 380 381(v0-4)= 382 383## v0.4 (Apr 1, 2019) 384 385libCEED v0.4 was made again publicly available in the second full CEED software 386distribution, release CEED 2.0. This release contained notable features, such as 387four new CPU backends, two new GPU backends, CPU backend optimizations, initial 388support for operator composition, performance benchmarking, and a Navier-Stokes demo. 389The new CPU backends in this release came in two families. The `/cpu/self/*/serial` 390backends process one element at a time and are intended for meshes with a smaller number 391of high order elements. The `/cpu/self/*/blocked` backends process blocked batches of 392eight interlaced elements and are intended for meshes with higher numbers of elements. 393The `/cpu/self/avx/*` backends rely upon AVX instructions to provide vectorized CPU 394performance. The `/cpu/self/xsmm/*` backends rely upon the 395[LIBXSMM](http://github.com/hfp/libxsmm) package to provide vectorized CPU 396performance. The `/gpu/cuda/*` backends provide GPU performance strictly using CUDA. 397The `/gpu/cuda/ref` backend is a reference CUDA backend, providing reasonable 398performance for most problem configurations. The `/gpu/cuda/reg` backend uses a simple 399parallelization approach, where each thread treats a finite element. Using just in time 400compilation, provided by nvrtc (NVidia Runtime Compiler), and runtime parameters, this 401backend unroll loops and map memory address to registers. The `/gpu/cuda/reg` backend 402achieve good peak performance for 1D, 2D, and low order 3D problems, but performance 403deteriorates very quickly when threads run out of registers. 404 405A new explicit time-stepping Navier-Stokes solver was added to the family of libCEED 406examples in the `examples/petsc` directory (see {ref}`example-petsc-navier-stokes`). 407This example solves the time-dependent Navier-Stokes equations of compressible gas 408dynamics in a static Eulerian three-dimensional frame, using structured high-order 409finite/spectral element spatial discretizations and explicit high-order time-stepping 410(available in PETSc). Moreover, the Navier-Stokes example was developed using PETSc, 411so that the pointwise physics (defined at quadrature points) is separated from the 412parallelization and meshing concerns. 413 414Backends available in this release: 415 416| CEED resource (`-ceed`) | Backend | 417|--------------------------|-----------------------------------------------------| 418| `/cpu/self/ref/serial` | Serial reference implementation | 419| `/cpu/self/ref/blocked` | Blocked reference implementation | 420| `/cpu/self/tmpl` | Backend template, defaults to `/cpu/self/blocked` | 421| `/cpu/self/avx/serial` | Serial AVX implementation | 422| `/cpu/self/avx/blocked` | Blocked AVX implementation | 423| `/cpu/self/xsmm/serial` | Serial LIBXSMM implementation | 424| `/cpu/self/xsmm/blocked` | Blocked LIBXSMM implementation | 425| `/cpu/occa` | Serial OCCA kernels | 426| `/gpu/occa` | CUDA OCCA kernels | 427| `/omp/occa` | OpenMP OCCA kernels | 428| `/ocl/occa` | OpenCL OCCA kernels | 429| `/gpu/cuda/ref` | Reference pure CUDA kernels | 430| `/gpu/cuda/reg` | Pure CUDA kernels using one thread per element | 431| `/gpu/magma` | CUDA MAGMA kernels | 432 433Examples available in this release: 434 435:::{list-table} 436:header-rows: 1 437:widths: auto 438* - User code 439 - Example 440* - `ceed` 441 - * ex1 (volume) 442* - `mfem` 443 - * BP1 (scalar mass operator) 444 * BP3 (scalar Laplace operator) 445* - `petsc` 446 - * BP1 (scalar mass operator) 447 * BP3 (scalar Laplace operator) 448 * Navier-Stokes 449* - `nek5000` 450 - * BP1 (scalar mass operator) 451 * BP3 (scalar Laplace operator) 452::: 453 454(v0-3)= 455 456## v0.3 (Sep 30, 2018) 457 458Notable features in this release include active/passive field interface, support for 459non-tensor bases, backend optimization, and improved Fortran interface. This release 460also focused on providing improved continuous integration, and many new tests with code 461coverage reports of about 90%. This release also provided a significant change to the 462public interface: a {ref}`CeedQFunction` can take any number of named input and output 463arguments while {ref}`CeedOperator` connects them to the actual data, which may be 464supplied explicitly to `CeedOperatorApply()` (active) or separately via 465`CeedOperatorSetField()` (passive). This interface change enables reusable libraries 466of CeedQFunctions and composition of block solvers constructed using 467{ref}`CeedOperator`. A concept of blocked restriction was added to this release and 468used in an optimized CPU backend. Although this is typically not visible to the user, 469it enables effective use of arbitrary-length SIMD while maintaining cache locality. 470This CPU backend also implements an algebraic factorization of tensor product gradients 471to perform fewer operations than standard application of interpolation and 472differentiation from nodes to quadrature points. This algebraic formulation 473automatically supports non-polynomial and non-interpolatory bases, thus is more general 474than the more common derivation in terms of Lagrange polynomials on the quadrature points. 475 476Backends available in this release: 477 478| CEED resource (`-ceed`) | Backend | 479|-------------------------|-----------------------------------------------------| 480| `/cpu/self/blocked` | Blocked reference implementation | 481| `/cpu/self/ref` | Serial reference implementation | 482| `/cpu/self/tmpl` | Backend template, defaults to `/cpu/self/blocked` | 483| `/cpu/occa` | Serial OCCA kernels | 484| `/gpu/occa` | CUDA OCCA kernels | 485| `/omp/occa` | OpenMP OCCA kernels | 486| `/ocl/occa` | OpenCL OCCA kernels | 487| `/gpu/magma` | CUDA MAGMA kernels | 488 489Examples available in this release: 490 491:::{list-table} 492:header-rows: 1 493:widths: auto 494* - User code 495 - Example 496* - `ceed` 497 - * ex1 (volume) 498* - `mfem` 499 - * BP1 (scalar mass operator) 500 * BP3 (scalar Laplace operator) 501* - `petsc` 502 - * BP1 (scalar mass operator) 503 * BP3 (scalar Laplace operator) 504* - `nek5000` 505 - * BP1 (scalar mass operator) 506 * BP3 (scalar Laplace operator) 507::: 508 509(v0-21)= 510 511## v0.21 (Sep 30, 2018) 512 513A MAGMA backend (which relies upon the 514[MAGMA](https://bitbucket.org/icl/magma) package) was integrated in libCEED for this 515release. This initial integration set up the framework of using MAGMA and provided the 516libCEED functionality through MAGMA kernels as one of libCEED’s computational backends. 517As any other backend, the MAGMA backend provides extended basic data structures for 518{ref}`CeedVector`, {ref}`CeedElemRestriction`, and {ref}`CeedOperator`, and implements 519the fundamental CEED building blocks to work with the new data structures. 520In general, the MAGMA-specific data structures keep the libCEED pointers to CPU data 521but also add corresponding device (e.g., GPU) pointers to the data. Coherency is handled 522internally, and thus seamlessly to the user, through the functions/methods that are 523provided to support them. 524 525Backends available in this release: 526 527| CEED resource (`-ceed`) | Backend | 528|-------------------------|---------------------------------| 529| `/cpu/self` | Serial reference implementation | 530| `/cpu/occa` | Serial OCCA kernels | 531| `/gpu/occa` | CUDA OCCA kernels | 532| `/omp/occa` | OpenMP OCCA kernels | 533| `/ocl/occa` | OpenCL OCCA kernels | 534| `/gpu/magma` | CUDA MAGMA kernels | 535 536Examples available in this release: 537 538:::{list-table} 539:header-rows: 1 540:widths: auto 541* - User code 542 - Example 543* - `ceed` 544 - * ex1 (volume) 545* - `mfem` 546 - * BP1 (scalar mass operator) 547 * BP3 (scalar Laplace operator) 548* - `petsc` 549 - * BP1 (scalar mass operator) 550* - `nek5000` 551 - * BP1 (scalar mass operator) 552::: 553 554(v0-2)= 555 556## v0.2 (Mar 30, 2018) 557 558libCEED was made publicly available the first full CEED software distribution, release 559CEED 1.0. The distribution was made available using the Spack package manager to provide 560a common, easy-to-use build environment, where the user can build the CEED distribution 561with all dependencies. This release included a new Fortran interface for the library. 562This release also contained major improvements in the OCCA backend (including a new 563`/ocl/occa` backend) and new examples. The standalone libCEED example was modified to 564compute the volume volume of a given mesh (in 1D, 2D, or 3D) and placed in an 565`examples/ceed` subfolder. A new `mfem` example to perform BP3 (with the application 566of the Laplace operator) was also added to this release. 567 568Backends available in this release: 569 570| CEED resource (`-ceed`) | Backend | 571|-------------------------|---------------------------------| 572| `/cpu/self` | Serial reference implementation | 573| `/cpu/occa` | Serial OCCA kernels | 574| `/gpu/occa` | CUDA OCCA kernels | 575| `/omp/occa` | OpenMP OCCA kernels | 576| `/ocl/occa` | OpenCL OCCA kernels | 577 578Examples available in this release: 579 580:::{list-table} 581:header-rows: 1 582:widths: auto 583* - User code 584 - Example 585* - `ceed` 586 - * ex1 (volume) 587* - `mfem` 588 - * BP1 (scalar mass operator) 589 * BP3 (scalar Laplace operator) 590* - `petsc` 591 - * BP1 (scalar mass operator) 592* - `nek5000` 593 - * BP1 (scalar mass operator) 594::: 595 596(v0-1)= 597 598## v0.1 (Jan 3, 2018) 599 600Initial low-level API of the CEED project. The low-level API provides a set of Finite 601Elements kernels and components for writing new low-level kernels. Examples include: 602vector and sparse linear algebra, element matrix assembly over a batch of elements, 603partial assembly and action for efficient high-order operators like mass, diffusion, 604advection, etc. The main goal of the low-level API is to establish the basis for the 605high-level API. Also, identifying such low-level kernels and providing a reference 606implementation for them serves as the basis for specialized backend implementations. 607This release contained several backends: `/cpu/self`, and backends which rely upon the 608[OCCA](http://github.com/libocca/occa) package, such as `/cpu/occa`, 609`/gpu/occa`, and `/omp/occa`. 610It also included several examples, in the `examples` folder: 611A standalone code that shows the usage of libCEED (with no external 612dependencies) to apply the Laplace operator, `ex1`; an `mfem` example to perform BP1 613(with the application of the mass operator); and a `petsc` example to perform BP1 614(with the application of the mass operator). 615 616Backends available in this release: 617 618| CEED resource (`-ceed`) | Backend | 619|-------------------------|---------------------------------| 620| `/cpu/self` | Serial reference implementation | 621| `/cpu/occa` | Serial OCCA kernels | 622| `/gpu/occa` | CUDA OCCA kernels | 623| `/omp/occa` | OpenMP OCCA kernels | 624 625Examples available in this release: 626 627| User code | Example | 628|-----------------------|-----------------------------------| 629| `ceed` | ex1 (scalar Laplace operator) | 630| `mfem` | BP1 (scalar mass operator) | 631| `petsc` | BP1 (scalar mass operator) | 632``` 633