Current NOT_IMPLEMENTED kernels

[thoasm]

We should discuss what kernels we should implement before the release. Since I lost the overview, I looked in the code and searched for kernels which are marked as GKO_NOT_IMPLEMENTED.

Interestingly, we do not support Coo::transpose and Coo::conj_transpose anywhere, which should be fairly trivial to implement (swapping col_idxs_ and row_idxs_ with std::move, sorting, followed by complex conjugation for conj_transpose).

The following are lists of GKO_NOT_IMPLEMENTED kernels (which are actually not implemented and not just have a case which is not implemented):

CUDA

• [ ] index_set
• [x] csr
  • [x] convert_to_fbcsr
• [x] dense
  • [x] convert_to_fbcsr
  • [x] count_nonzero_blocks_per_row
  • [x] convert_to_sparsity_csr #904
• [ ] fbcsr
  • [x] fill_in_dense
  • [ ] sort_by_column_index
  • [ ] extract_diagonal
• [ ] sparsity_csr
• [ ] jacobi
  • [ ] convert_to_dense

DPC++

• [ ] index_set
• [ ] partition #1034
• [x] factorization #928
• [ ] ic
• [ ] ilu
• [x] par_ic #928
• [x] par_ilu #928
• [x] par_ict #928
• [x] par_ilut #928
• [ ] csr
  • [ ] convert_to_fbcsr
• [ ] dense
  • [ ] convert_to_fbcsr
  • [ ] count_nonzero_blocks_per_row
  • [x] convert_to_hybrid #904
  • [x] convert_to_sparsity_csr #904
• [ ] fbcsr
• [ ] fft
• [x] hybrid #904
• [ ] sparsity_csr
• [x] amgx_pgm #933
• [x] isai #924
• [ ] jacobi #929
• [ ] lower_trs
• [ ] upper_trs
• [x] multigrid

HIP

• [ ] index_set
• [x] csr
  • [x] convert_to_fbcsr
• [x] dense
  • [x] convert_to_fbcsr
  • [x] count_nonzero_blocks_per_row
  • [x] convert_to_sparsity_csr #904
• [x] fbcsr
  • [x] spmv
  • [x] advanced_spmv
  • [x] fill_in_dense
  • [ ] transpose
  • [ ] conj_transpose
  • [ ] sort_by_column_index
  • [x] is_sorted_by_column_index
  • [ ] extract_diagonal
• [ ] sparsity_csr
• [ ] jacobi
  • [ ] convert_to_dense

OpenMP

• [ ] ic
• [ ] ilu
• [x] fbcsr
  • [x] convert_to_csr

Current conversions supported (maybe not fully implemented)

Everything -> Csr Everything -> Dense Csr -> Everything Dense -> Everything

Effective: 14.10.2019

Updated: 01.02.2022

[thoasm]

I will definitively take care of the move_to_XXX replacements, but for the others, we should discuss what is necessary. I can also take care of the COO::transpose, it should also be pretty straight forward.

[hartwiganzt]

I think the reference transpose and conjugate transpose should be easy. No need for the other executors. I don't have a strong opinion on the others - I think they are all optional.

[thoasm]

transpose in general is easy, I don't even have to implement any kernels (since I just swap the col and row index arrays). Additionally, I think we should also implement at least the OMP kernels, so we actually support everything with OMP that we do with CUDA.

[tcojean]

I tried to check everything that is "taken care of". Overall, what is left I think is:

• CUDA Dense <-> Hybrid conversions (hard to implement AFAIK)
• OpenMP CSR <-> Ell, Sellp
• Everywhere CSR <-> Hybrid

I believe none are that much required for the release.

[thoasm]

That is true, I removed the dependency to the Release. We also said previously that we do not require the OpenMP kernels since we have the reference version.

[tcojean]

Note: this was updated to the last status.