NOLINT for OCCA tensor contract false positive

OCCA Backend clang-tidy fixes

Update Ceed Delegate refrencing

Add delegates for specific objects

VecCreate -> VectorCreate

Use blocking in optimized serial backends

Include full evec blocked backend

Add restriction by block to /cpu/self/*/blocked

Add ElemRestrictionApplyBlock

make style

Remove useless function declaration.

Add a reference non-tensor BasisApply for cuda backends.

Cuda backend using shared memory (#247)

Add a GPU backend based on Cuda using shared memory.

* Draft of a shared memory backend

* New basis apply passes all tests.

* Add the possibility to

Cuda backend using shared memory (#247)

Add a GPU backend based on Cuda using shared memory.

* Draft of a shared memory backend

* New basis apply passes all tests.

* Add the possibility to treat several elements in one block of threads.

* Fix an error in 2D and 3D gradient.

* Put the cuda-shared backend in its own folder.

* Minor cleaning.

* Replace <ceed-impl.h> with <ceed-backend.h>

* make style

* Add a few CeedChk_Cu

show more ...

Fix include statements

Set QFunction outputs undefined before apply in new memcheck backend

Force Context existence with cudaFree(0).

Modify the device initialization

Fix CeedChk with CeedChk_Cu in the Cuda backend.

Add atomicAdd in /cuda/ref backend for compute capability < 6.0

Add CeedGetPreferredMemType

Add error message if XSMM kernel fails to build

Add kernel caching to XSMM backend

Make style and comment updates

XSMM tensor ind logic fix

Logic cleanup

Add Basis argument to TensorContractCreate

Small clang-tidy fixes (#215)

Simple Cuda backend using one thread per element (#195)

Thanks-to: Jeremy Thompson

* Take into account the compute capability of the GPU

* Add the cuda/reg backend and rename cuda to cuda/ref.

Simple Cuda backend using one thread per element (#195)

Thanks-to: Jeremy Thompson

* Take into account the compute capability of the GPU

* Add the cuda/reg backend and rename cuda to cuda/ref.

- cuda/reg uses a simple approach where each element is
processed by one thread. This approach is expected to be
efficient for 1D and 2D problems, but very ineficient
as soon as the kernels start to spill, which should arise
around Q1D=4 for 3D problems.

* Compilation takes into account the deviceId

* Make style

* Remove dead code in cuda qFunctions.

* Cuda-reg specialized Restriction.

* Split the Prolongation operator into Identity/not Identity.

* Remove "#pragma unroll" until further perf investigation.

* README update

* Add a description of cuda/reg.

* Add CompositeOperator msg to CUDA backends

show more ...