Fix regression but from shared-memory non-tensor basis transpose

Remove need to zero out V vector before applying basis transpose for magma backends

Initial commit to optimize magma backend transpose basis application

Use shared-memory for transpose non-tensor basis kernel to accelerate load of A matrix

H(div) and H(curl) basis support for magma backend

Formatting consistency for magma backend with cuda-ref and hip-ref

Includes JiT upgrades for Magma non-tensor basis to only compile for N values which are used at runtime. Adds JiT for Magma non-ten

Formatting consistency for magma backend with cuda-ref and hip-ref

Includes JiT upgrades for Magma non-tensor basis to only compile for N values which are used at runtime. Adds JiT for Magma non-tensor basis CEED_EVAL_WEIGHT mode.

show more ...

Formatting updates for MAGMA JiT kernels

Rename files in magma backend for consistency with other libCEED backends

Enforce consistent pointer alignment with clang-format

Fix some missing consistency issues from #1315

style - fixes for CUDA backends

style - fix header guards

Whitespace, style, and formatting updates for consistency between CUDA and HIP backends

Adds include guards in JiT header files, even if not strictly necessary, to match the precedent set in cuda-sh

Whitespace, style, and formatting updates for consistency between CUDA and HIP backends

Adds include guards in JiT header files, even if not strictly necessary, to match the precedent set in cuda-shared and hip-shared as well as sycl.

show more ...

Remove two unused files from cuda-shared and hip-shared backends

Minor formatting consistency for SYCL backend files

Unify magma backend restriction with cuda/hip-ref, keeping runtime option for deterministic and non-deterministic using magma/det

This also opens the opportunity to make cuda/hip-ref non-determinist

Unify magma backend restriction with cuda/hip-ref, keeping runtime option for deterministic and non-deterministic using magma/det

This also opens the opportunity to make cuda/hip-ref non-deterministic by default and add cuda/hip-ref/det variants.

show more ...

Add sycl/gen backend and other sycl changes (#1258)

---------

Co-authored-by: Kris Rowe <kris.rowe@anl.gov>
Co-authored-by: Kris Rowe <krowe@anl.gov>
Co-authored-by: Jed Brown <jed@jedbrown.org

Add sycl/gen backend and other sycl changes (#1258)

---------

Co-authored-by: Kris Rowe <kris.rowe@anl.gov>
Co-authored-by: Kris Rowe <krowe@anl.gov>
Co-authored-by: Jed Brown <jed@jedbrown.org>
Co-authored-by: Varsha Madananth <vmadananth@uan-0002.head.cm.americas.sgi.com>
Co-authored-by: James Wright <jrwrigh.iii@gmail.com>

show more ...

CUDA: improvements for handling large CeedVectors requiring CeedSize for length

HIP: Improve support for CeedVectors that are longer than the max size of 32 bit integers

Very minor backend style and whitespace fixes

Add sycl/ref and sycl/shared backends (#1229)

* Merge sycl_backend from ALCF fork

---------
Co-authored-by: Umesh Unnikrishnan <unnikrishnan@anl.gov>
Co-authored-by: Kris Rowe <krowe@anl.gov>

Add sycl/ref and sycl/shared backends (#1229)

* Merge sycl_backend from ALCF fork

---------
Co-authored-by: Umesh Unnikrishnan <unnikrishnan@anl.gov>
Co-authored-by: Kris Rowe <krowe@anl.gov>
Co-authored-by: Jeremy L Thompson <jeremy@jeremylt.org>

show more ...

Fix file endings inconsistency

include - move enums to ceed/types.h

magma: non-tensor rtc (#1141)

* some refactoring in magma's jit src

* fix path

* fix loading src

* refactor magma nontensor backend

* refactor magma nontensor backend

* [WIP]: new non

magma: non-tensor rtc (#1141)

* some refactoring in magma's jit src

* fix path

* fix loading src

* refactor magma nontensor backend

* refactor magma nontensor backend

* [WIP]: new nontensor basis kernels

* [WIP]: new nontensor basis kernels

* [WIP]: new nontensor basis kernels

* call the new nontensor kernels for low order problems

* multiple compilation for the same kernels but with different tuning parmaters

* magma: allow different nb's for different non-tensor kernels

* tuning data for the non-tensor rtc kernels

* remove no-longer used functions, add new one for tuning the nontensor kernels

* constants for tuning

* tuning functions

* use the tuning functions in compiling/running the new kernels

* bug fix

* fixes

* fixes

* minor

* switch tuning data

* fix name

* fix name

* add function to run cuda kernels with opt-in shared memory feature

* minor fix

* minor fix

* fix calls to batch api

* allow more kernel instances

* temporary timing function

* temporary timing function

* tuning data based on hiprtc

* rollback tuning parameters

* fixes

* fixes

* fix inconsistency in the parameters passed to nvrtc/hiprtc

* minor

* a fix to the nb selector

* cleanup

* merge the opt-in feature in CeedRunKernelDimSharedOptinCuda into CeedRunKernelDimSharedCuda

* fix paths for hip-magma backends

* style

* fixes

* running make format

* undo changes from the last commit

* change HIP_DIR to ROCM_DIR and adjust the paths for magma accordingly

* replace HIP_DIR with ROCM_DIR

show more ...

minor - assorted formatting fixes