Skip to content

rocm-10.2.0

Latest

Choose a tag to compare

@amontoison amontoison released this 19 Nov 20:13
· 1 commit to develop since this release

hipfort 0.11.0 for ROCm 10.2.0

Added

  • Regenerated the bindings against the ROCm 10.2 API, adding 77 routines in
    HIP, hipSOLVER, rocSOLVER and rocSPARSE.
  • The hipFFT Xt API, in the new hipfort_hipfftxt module (hipfort::hipfftxt
    CMake component). use hipfort_hipfftxt gives the whole hipFFT API.
    hipfort_hipfft keeps the hipfft.h API only. The multi-process routines
    exist only in a hipFFT built with MPI, which the ROCm packages are not.
  • Array overloads for every routine with a data buffer, including the ILP64
    (_64) routines of rocBLAS, hipBLAS and rocSOLVER, with _devptr twins that
    take a type(c_ptr) wherever an argument may be a device pointer.
  • X_strided_rank array overloads for the strided-batched routines of rocBLAS,
    hipBLAS, rocSOLVER and hipSOLVER, which take the batch as one more dimension,
    as in A(lda, n, batch_count).
  • The optional arguments of hipStreamGetCaptureInfo_spt and
    hipStreamGetCaptureInfo_v2_spt.

Changed

  • Breaking. Arguments are now classified against the C documentation.

    • The single values that rocSOLVER writes on the GPU (info, nev,
      n_sweeps, residual, rcond, ...) and the info of
      hipsolverDnXpotrf and hipsolverDnXpotrs are taken as one value on the
      device, not as an array.
    • Host values (the info of hipblasXgeqrf and hipblasXgetrs, and a few
      rocSPARSE and hipSPARSE arguments) lose their _devptr forms. Pass a
      Fortran variable.
    • The array overloads of the hipSOLVER batched routines take devInfo as an
      array of batch_count entries.

    Calls through the generic names with a type(c_ptr) compile as before, but a call
    to a specific procedure by name may need its _devptr form.

  • Breaking. The array overloads of hipMalloc, hipFree, hipMemcpy and
    related routines are now part of the Fortran 2008 interfaces, so a build with
    -DHIPFORT_USE_FPOINTER_INTERFACES=OFF no longer has them.

  • Breaking. hipfftXtSetWorkArea moves from hipfort_hipfft to the new
    hipfort_hipfftxt module, with the rest of the hipFFT Xt API. Code that calls
    it writes use hipfort_hipfftxt.

  • Breaking for calls by keyword. The hipblasSetVector/GetVector/
    SetMatrix/GetMatrix families name their arguments as in the hipBLAS
    headers (elemSize, AP, BP).

  • hipblasXrotBatched and hipblasXrotStridedBatched take c and s as
    scalars read per the pointer mode, as rocBLAS does.

  • The archives are compiled with one section per procedure. Link with
    -Wl,--gc-sections to keep only the wrappers you call and to link against an
    older ROCm. A shared library also needs
    -Wl,--exclude-libs,libhipfort-amdgcn.a.

  • The assumed-rank overloads (-DHIPFORT_ASSUMED_RANK=ON) no longer declare
    their arrays contiguous, like the per-rank ones, so a section is passed as
    it is, without a copy. A submatrix works when the leading dimension is that of
    the whole array, and a strided vector when the increment matches the stride.

  • The modules compile 2 to 3 times faster with gfortran and amdflang, and much
    faster with the Cray compiler.

Fixed

  • On the CUDA backend, the GPUDirect RDMA device attributes had their HIP
    values instead of their CUDA values.
  • A failed hipMalloc, hipMallocManaged or hipHostMalloc array overload
    now leaves its pointer unassociated.
  • -DHIPFORT_ASSUMED_RANK=ON with nvfortran now falls back to the per-rank
    interfaces, since its c_loc of an assumed-rank argument is wrong.
  • Warnings with the Cray compiler.
  • The find_package test on platforms that install to lib64.