[amontoison]

Follow-up to #581 and #583, from a library-by-library audit of the bindings against the ROCm 10.1 headers. Regenerated with rocm-fortran 3b65cdb.

More typed forms

These scalar pointers still needed c_loc(). They now also get a _typed form:

• rocBLAS / hipBLAS:
  • the dot/dotu/dotc/nrm2/asum/iamax/iamin result;
  • the rotg and rotmg scalars;
  • c/s of hipBLAS rot.
• rocSPARSE / hipSPARSE:
  • the doti/dotci results;
  • the nnz counts of the *_nnz/*_nnzb/nnz_compress/csr2bsrNnz/csr2gebsrNnz routines;
  • the pivot positions of csritsv_zero_pivot and csric0/csrilu0_singular_pivot;
  • ell_width of csr2ell_width.
• hipBLAS: info of getrs/geqrf/gels, which the header documents as host memory.
• HIP:
  • hipChooseDevice, whose metadata key named the short alias instead of the hipChooseDeviceR0600 symbol, so the entry never applied;
  • the failIdx and nbGroups outputs.

The array overloads take the BLAS results and the hipBLAS info typed. New _rank_N_devptr variants keep the type(c_ptr) these arguments always had there, so existing c_loc() calls still compile.

Fixes

• rocSOLVER larf/larfg: alpha and tau are device scalars, since the library forces the device pointer mode. They lose the _typed form introduced by Bind every scalar pointer as type(c_ptr), add its typed form as _typed #581, which would have handed the kernel a host address.
• hipsolverRfBatchZeroPivot and rocblas_set_solution_fitness_query lose theirs too. The first one's position is an array of batchSize ints. The second one's fitness is a pointer the library keeps.
• csrcolor (rocSPARSE and hipSPARSE): coloring and reordering are arrays again in the array overloads. The color count and fraction_to_color are host scalars.
• rocSPARSE: the nullable error arguments (spgeam, spmv_set_input, v2_spmv, ...) are optional.
• hipMalloc(..., source=) and its hipMallocManaged/hipHostMalloc siblings: the copy used to overwrite the allocation's status. They now copy only after a successful allocation, so an out-of-memory is reported.
• USE_CUDA_NAMES: the compat hipsolver?gesvd_bufferSize/??gels_bufferSize were bound to cusolverDn routines with different arguments. These were amdflang's 8 -Wexternal-interface-mismatch warnings, and the build now has none. They are ROCm-only now.
• hipChooseDevice is a generic over hipChooseDeviceR0600's own specifics, instead of a second interface on the same binding label (F2018 19.2).

Also in this PR

• three tests that take a result through the new typed forms: rocBLAS ddot, hipBLAS snrm2, and rocSPARSE ddoti in host pointer mode;
• the regenerated symbol tests;
• the supported-API tables, with only the generated diff applied;
• the changelog: the Document and test the typed and type(c_ptr) forms of scalar arguments #583 docs section and tests, plus everything above.

Validation

ROCm 10.1, gfortran and amdflang.

Check	Result
Modules compile, with no macro, USE_ASSUMED_SHAPE, USE_ASSUMED_RANK and USE_ASSUMED_SHAPE + USE_CUDA_NAMES	all four configurations, no amdflang warning (8 on develop with USE_CUDA_NAMES)
Current tests build	1248/1248
hipfort 0.9.0 test programs build	728/807, the same set as on develop
ctest with -DHIPFORT_EXTENDED_TESTS=ON	1258/1261: the three known WSL failures (hip_mem_advise ×2, hip_module_kernel)
Shared-link tests	every symbol resolves against ROCm 10.1, and against CUDA 13.3 without libcuda
gfx1101 GPU, 0.9.0 test programs that build	same results as against the 0.9.0 library

Not in this PR

• NVIDIA backend (enum values not translated for CUDA, and a few CUDA mappings whose signature differs): this will be a separate PR, verified by compilation only.
• New functionality (hipfftXt, hipFFTW array overloads, more .native routines): after the release.