IntelPython / dpnp / 41674999499

84%

push

github

Refactor libtensor and backend into libtensor (#3131)

This PR proposes a refactor merging the dpnp backend pybind11 extensions
into libtensor, moving libtensor to the dpnp directory (out from
dpnp/tensor) and merging kernels from the dpnp backend into libtensor

This makes libtensor the sole home of all kernels and Python bindings,
such that users of the libtensor library have access to the full suite
of functionality from Python or C++

Slips in some small additional changes:
* vm: the 2-D binary contig dispatch check now looks up both type ids
(the row lookup could never be `nullptr`; the check still can't fire,
since the earlier checks already guarantee a match) and fixes the
`fmax`/`fmin` docstrings copied from `fmod`
* lapack: cast the `syevd` eigenvalue output to `RealT` (a no-op for the
supported `T == RealT` types); fix the "outpuy" typo and the gesvd error
message naming `gesv()`
* fix the "c-contiguos" typo (visible in the `dpnp.interp` error for
non-contiguous input)
* remove unused helpers: `DispatchTable2` getters, local-memory-size
overloads, and `FourZeroOffsets_Indexer`, `FixedDimStridedIndexer` and
`ThreeOffsets_FixedDimStridedIndexer` from `utils/offset_utils.hpp`
* CMake: remove dead branches and options, unused targets and variables,
and no-op properties
* coverage: the dpnp sources are listed first in the merged modules so
coverage builds keep their instrumented copies of shared inline
functions, and the shared static library is instrumented too
* Fixes unnecessary copies of dependent event vectors

1792 of 2998 branches covered (59.77%)

Branch coverage included in aggregate %.

166 of 173 new or added lines in 79 files covered. (95.95%)

29539 of 34140 relevant lines covered (86.52%)

39098.48 hits per line