CUDA - add small leak fix in QFunction JiT load

CUDA - clean up includes

CUDA - clean up minor memory leak

XSMM - switch to LIBXSMM 143409c, reverted redefiniton of _Bool

comment out the hiprtc-related calls causing memory errors

XSMM - use v1.16

fix size of options list for hiprtcCompile

Op - add AssembleAdd version of diagonal assembly functions, will be helpful for MFEM integration

Op - change LinearAssemble* to accept CeedVector instead of pointer to CeedVector, allows for less memory movement and interfaces with parent code better

Experimental HIP port of cuda-ref, round 2, with ROCm3.5

make style

style - rename AssembleLinear* to LinearAssemble*

style - change Get*Status to Is*

Composite Operator support for AssembleLinearDiag/PBDiag (#552)

* ref - add composite operator support for building diagonal/pb diagonal

* ref - add error for non-composite mult-field operator di

Composite Operator support for AssembleLinearDiag/PBDiag (#552)

* ref - add composite operator support for building diagonal/pb diagonal

* ref - add error for non-composite mult-field operator diagonal/pb diagonal assembly

* tap - add t538 exclusion because OCCA does not support galleries

* tests - adjust test cases for ceed examples for test coverage

* Op - fix documentation

show more ...

CUDA/MAGMA - add AssembleLinearQFunction (#553)

* CUDA - refactor operator apply for CUDA assemble linear QF impl

* CUDA/MAGMA - add AssembleLinearQFunction

* make style

* CUDA - clean up a

CUDA/MAGMA - add AssembleLinearQFunction (#553)

* CUDA - refactor operator apply for CUDA assemble linear QF impl

* CUDA/MAGMA - add AssembleLinearQFunction

* make style

* CUDA - clean up assembleLinearQF after q/e layout refactor

* CUDA - fallback operator for cuda/gen to cuda/ref

* CUDA - use delegation for cuda/gen prefered memtype

show more ...

Convert CUDA ref/reg/shared E-Layout (#554)

* tests - update tests for multiple e-layouts

* CUDA - convert ref and reg restrictions to Q-layout

* CUDA - ref/reg/shared use gen/magma E-Layout f

Convert CUDA ref/reg/shared E-Layout (#554)

* tests - update tests for multiple e-layouts

* CUDA - convert ref and reg restrictions to Q-layout

* CUDA - ref/reg/shared use gen/magma E-Layout for multi elememnt basis apply and operator apply

* CUDA/MAGMA - drop eandqdiffer and separate MAGMA operator code

* CUDA - update operator comment

* reg - clarify read/write dofs/quads

* CUDA - drop dead code

show more ...

CPU Impl of AssemblePointBlockDiagonal (#503)

* Operator - add AssemblePointBlockDiagonal for CPU backends, with test

* CUDA - add point block diagonal not supported message

* make style

*

CPU Impl of AssemblePointBlockDiagonal (#503)

* Operator - add AssemblePointBlockDiagonal for CPU backends, with test

* CUDA - add point block diagonal not supported message

* make style

* Operator - improve point block description

* OCCA - explicitly remove OCCA fallback to CPU assembly functions, will update after new OCCA backend

* Op - remove gap removal in point block diagonal

* Op - update diagonal assembly documentation

* Update backends/ref/ceed-ref-operator.c

Co-authored-by: Jed Brown <jed@jedbrown.org>

* style - fix extra space in * with nopad

Co-authored-by: Jed Brown <jed@jedbrown.org>

show more ...

OCCA - small fixes highlighed by icc

CUDA - fix empty restriction support, must use GetStridedStatus over buggy indices pointer check

Icl/magma queue (#524)

Update the MAGMA backend:

* add new specialized tensor basis kernels

* add batched DGEMM wrapper for non-tensor basis kernels

* switch backend to use MAGMA v2 inte

Icl/magma queue (#524)

Update the MAGMA backend:

* add new specialized tensor basis kernels

* add batched DGEMM wrapper for non-tensor basis kernels

* switch backend to use MAGMA v2 interface

Co-authored-by: nbeams <246972+nbeams@users.noreply.github.com>
Co-authored-by: Stan Tomov <tomov@eecs.utk.edu>

show more ...

cuda-gen: avoid gcc-10 warning on use of strncat

XSMM - fix q=1 index computation (#531)

* XSMM - use hash table for kernel index lookup, add khash to libCEED

* Hash - add CeedHashIJKLM to simplify xsmm tensor hash table code

* simplify use

XSMM - fix q=1 index computation (#531)

* XSMM - use hash table for kernel index lookup, add khash to libCEED

* Hash - add CeedHashIJKLM to simplify xsmm tensor hash table code

* simplify use of hash table, use kernels as values

* XSMM: more idiomatic use of khash

* make style

Co-authored-by: Jed Brown <jed@jedbrown.org>

show more ...

CEED_STRIDES_BACKEND optimization for cuda-ref operator apply (#528)

* add check for backend stride status for input vectors

* add backend strides check for output vectors

* replace output cop

CEED_STRIDES_BACKEND optimization for cuda-ref operator apply (#528)

* add check for backend stride status for input vectors

* add backend strides check for output vectors

* replace output copy with elem restriction for none emode

* move input skip_restrict check to setup and never allocate E-vec if not needed

* add boolean variable for E/Q vector layout for
further optimization of output and add wrapper function in magma backend
to create a cuda-ref operator and change this state variable

* Add missing CeedChks

* style changes to better match cuda backends

* missed style change for evec check

* add CeedChk from PR #525 (merge conflict)

* make style changes

* adjust size of nqpts for non-tensor basis

show more ...

MAGMA - support empty restrictions

CUDA - support empty restrictions