backends/cuda-shared: limit 1D thread counts

We need to avoid this error:

CUDA_ERROR_LAUNCH_OUT_OF_RESOURCES: max_threads_per_block 512 on block size (24,1,32), shared_size 0, num_regs 106

A pro

backends/cuda-shared: limit 1D thread counts

We need to avoid this error:

CUDA_ERROR_LAUNCH_OUT_OF_RESOURCES: max_threads_per_block 512 on block size (24,1,32), shared_size 0, num_regs 106

A proper solution is to use cuOccupancyMaxPotentialBlockSize to place a
number of elements per block that stays within resource limits. This
would involve a bit more refactoring to do cleanly.

show more ...

backends/cuda: more informative error reporting

xsmm: support for 1.17 headers

Add launch bounds to HIP QFunction kernels

add launch bounds to magma kernels;
add macro definition for y-dim of magma basis kernel threadblocks

Co-authored-by: Ahmad Abdelfattah <ahmad@icl.utk.edu>

Add flag to use atomic adds on supported AMD GPU hardware

updated ceed-ref-restriction.c

Added CeedElemRestrictionIsOriented function

backends/ref: Added oriented element restriction.

H(div) basis constructor (#878)

* backends/interface: added H(div) basis constructor

* tests: CEED_EVAL_INTERP/DIV tested for H(div) basis.

* Update interface/ceed-basis.c

Co-authored-by:

H(div) basis constructor (#878)

* backends/interface: added H(div) basis constructor

* tests: CEED_EVAL_INTERP/DIV tested for H(div) basis.

* Update interface/ceed-basis.c

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

* Update interface/ceed-basis.c

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

* Update interface/ceed-basis.c

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

* Update interface/ceed-basis.c

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

* Update interface/ceed-basis.c

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

* interface/ceed-basis.c: delete [-1,1] for non-tensor basis constructor

* Update interface/ceed-basis.c

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

* include/ceed: added CEEDFESpace enum and used it at ceed-basis.c

* tests: added t330-basis.out and changed xhat-->x_hat

* Changed name convention for CeedElemTopology and CeedFESpace enum.

* make style

* interface/ceed-basis.c: Added FE space/topology of the basis to CeedBasisView function

* tests/output: update ...basis.out file

* Changed CEED_FE_TOPO_ ==> CEED_TOPOLOGY_

* backends/ref: delete CeedBasisApplyHdiv_Ref

* Update interface/ceed-basis.c

Co-authored-by: Jeremy L Thompson <jeremy@jeremylt.org>

* Update interface/ceed-basis.c

Co-authored-by: Jeremy L Thompson <jeremy@jeremylt.org>

* Update include/ceed-impl.h

Co-authored-by: Jeremy L Thompson <jeremy@jeremylt.org>

* interface/ceed-bassi.c: update CeedBasisView

* deleted L2 and H(curl) FE spaces

* minor - update output file

* minor - update output file

* style - consistency with main

Co-authored-by: Jed Brown <jed@jedbrown.org>
Co-authored-by: Jeremy L Thompson <jeremy@jeremylt.org>

show more ...

Add launch bounds to hip-gen operator kernel

xsmm - add compatibility for xsmm > 1.17

Slight modifications for hiprtc usage in ROCm 4.5 (#850)

cpu - drop some empty TensorContract destructors

vec/ctx - make backend restore function optional

gpu - style consistency

gpu - minor style

gpu - improved human-readability of debugging output

gpu - pull quoted kernels into separate files

style - consistent nameing and style for gpu backends

gpu - drop unused argument in init

gpu - drop unused [HIP, CUDA]_MAX_PATH constant

gpu - split common cuda/hip data into separate folder

cuda - separate compile functionality into new header

cpu - fall through now on INTERP/GRAD cases