Skip to content

Correct the 0.10.0 changelog after a claim-by-claim check #588

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
amontoison merged 4 commits into from
Oct 9, 2026
Merged

Correct the 0.10.0 changelog after a claim-by-claim check #588

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
7ea52c6
Correct the 0.10.0 changelog after a claim-by-claim check
amontoison Oct 9, 2026
46a53b0
Merge branch 'develop' into changelog_final
amontoison Oct 9, 2026
77cf7ee
Reflow the rocTX comment in lib/CMakeLists.txt
amontoison Oct 9, 2026
ee37b16
Changelog: ROCTx string form, and direct calls to the bind(C) specifics
amontoison Oct 9, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
128 changes: 77 additions & 51 deletions CHANGELOG.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -28,10 +28,13 @@
pass the variable and when its address, the device pointer mode, and
hipSOLVER's `devInfo`.
* Test coverage for:
* the BLAS `nrm2`, `asum`, `iamax`/`iamin`, `ger`, `syrk`/`herk`,
`symm`/`hemm`, `trmm`, `geam` and `gemm_ex` routines in rocBLAS and hipBLAS;
* the BLAS `nrm2`, `asum`, `iamax`, `syrk`, `symm`, `geam` and `gemm_ex`
routines in rocBLAS and hipBLAS, `iamin`, `ger`, `herk` and `hemm` in
rocBLAS, and `trmm` in hipBLAS;
* every `N`/`T`/`C` combination of `gemm` and `gemv`, on rectangular padded
matrices with complex `alpha`/`beta` and non-unit increments;
matrices with complex `alpha`/`beta`, and non-unit increments for `gemv`;
* the HIP cache configuration, memory pool and stream capture queries,
`hipMalloc` with `lbounds=`, and the `<lib>Check` messages;
* the device pointer mode of the BLAS and sparse libraries;
* the version, pointer-mode, handle-state and descriptor getters of nine
libraries;
Expand All @@ -44,12 +47,11 @@

### Changed

* Every pointer-mode scalar and host output now comes in two forms under the
same generic.
The `bind(C)` specific takes it as `type(c_ptr), value`, the C API's own
spelling, and a `<routine>_typed` module procedure takes the Fortran variable
(`integer`, `real`, `complex` or the enum kind) and passes its address, so
both calls work:
* A pointer-mode scalar or host output now comes in two forms under the same
generic. The `bind(C)` specific takes it as `type(c_ptr), value`, the C API's
own spelling, and a `<routine>_typed` module procedure takes the Fortran
variable (`integer`, `real`, `complex` or the enum kind) and passes its
address, so both calls work:

```fortran
st = rocblas_daxpy(handle, n, alpha, dx, incx, dy, incy)
Expand All @@ -58,26 +60,31 @@

This covers:
* the scalars a library reads or writes per its handle's pointer mode:
`alpha` and `beta` throughout rocBLAS, hipBLAS, rocSPARSE and hipSPARSE,
`alpha` and `beta` throughout rocBLAS, hipBLAS, rocSPARSE and hipSPARSE;
the `dot`, `nrm2`, `asum` and `iamax`/`iamin` results and the `rot`, `rotg`
and `rotmg` scalars of rocBLAS and hipBLAS, the sparse `doti`/`dotci`
and `rotmg` scalars of rocBLAS and hipBLAS; the sparse `doti`/`dotci`
results, `nnzTotalDevHostPtr`, the nnz counts of the `*_nnz` routines and
the pivot positions. The `type(c_ptr)` form makes
`rocblas_pointer_mode_device` and its equivalents usable from Fortran;
the pivot positions. The half-precision routines, the `_ex`/`Ex` and
generic sparse routines, whose data type is chosen at run time, and the
hipBLAS batched `rot` keep the `type(c_ptr)` form only. The `type(c_ptr)`
form makes `rocblas_pointer_mode_device` and its equivalents usable from
Fortran;
* the host outputs: the HIP runtime queries (`hipStreamGetId`,
`hipStreamGetCaptureInfo`, `hipFuncGetAttribute`,
`hipDeviceGetP2PAttribute`, the `hipOccupancy*` and `hipGraph*` counts,
...), the version, pointer-mode, math-mode and handle-state getters of the
BLAS, sparse and solver libraries, the hipSPARSE and rocSPARSE descriptor
getters, `hipblasGetProperty`, `hipfftGetProperty`, the hipBLAS
`getrs`/`geqrf`/`gels` `info`, the hipSOLVER buffer sizes, `*gels`
`niters` and `*{sy,he}evdx`/`*{sy,he}gvdx` `nev`, the `csrcolor` color
count, `hipChooseDevice`, and `rocfft_cache_serialize`. An output the C
API lets the caller skip is `optional` in the `_typed` form; pass
`hipDeviceGetP2PAttribute`, the `hipOccupancy*` and `hipGraph*` counts, ...,
but not the `_spt` variants), the version, pointer-mode, math-mode and
handle-state getters of the BLAS, sparse and solver libraries, the hipSPARSE
and rocSPARSE descriptor getters, `hipblasGetProperty`, `hipfftGetProperty`,
the hipBLAS `getrs`/`geqrf`/`gels` `info`, the hipSOLVER buffer sizes,
`*gels` `niters` and `*{sy,he}evdx`/`*{sy,he}gvdx` `nev`, the `csrcolor`
color count, `hipChooseDevice`, and `rocfft_cache_serialize`. An output the
C API lets the caller skip is `optional` in the `_typed` form; pass
`c_null_ptr` to the `type(c_ptr)` one.

Scalars that only ever live on the device, such as rocSOLVER `info` and
hipSOLVER `devInfo`, keep the `type(c_ptr)` form only.
Scalars that only ever live on the device, such as rocSOLVER `info`, the
`alpha` and `tau` of `rocsolver_?larfg`, the `alpha` of `rocsolver_?larf`, and
hipSOLVER `devInfo`, have no `_typed` form. The array overloads take them like
the device arrays next to them.

The 28 rocSPARSE and hipSPARSE routines with both kinds (`alpha`/`beta` beside
a buffer size) also get `<routine>_devptr`, for the device pointer mode: the
Expand All @@ -95,40 +102,52 @@
`type(fftw_iodim)` / `type(fftw_iodim64)` arrays in every precision. The
`hipfftGetProperty` rank overloads are removed, as its value is a plain
integer.
* **Breaking.** Apart from ROCTx (below), six patterns of code written against
* **Breaking.** Apart from ROCTx, the preprocessor macros and direct calls to
the `bind(C)` specifics (below), eight patterns of code written against
hipfort 0.9.0 no longer compile:
* hipSOLVER, an `integer` `devInfo` passed with `type(c_ptr)` buffers: pass
`c_loc(devInfo)`, or the `type(c_ptr)` from `hipMalloc`;
* hipSOLVER, an `integer` `devInfo` passed with `type(c_ptr)` buffers: pass a
device pointer, such as the `type(c_ptr)` from `hipMalloc`;
* hipSOLVER array overloads, `tau` passed as an element, `dTau(1)`: pass the
array;
* hipSOLVER `*syevdx_bufferSize`, `*heevdx_bufferSize`, `*sygvdx_bufferSize`
and `*hegvdx_bufferSize`, `nev` passed as a `type(c_ptr)` while `lwork` is
typed: pass both variables, or both addresses;
* `hipsolverDsyevdx_bufferSize`, the host `nev` passed as a `type(c_ptr)`
while `lwork` is typed (the other precisions took an integer): pass both
variables, or both addresses;
* `rocsolver_chegvdx` and `rocsolver_zhegvdx`, `nev` passed as an integer: it
is a device pointer, as for the rest of the `*gvdx` family, so pass
`c_loc(nev)`;
* the hipFFTW guru planners, a scalar `dims`: pass an array;
is a device pointer, as for the rest of the `*gvdx` family, so pass a
device pointer, such as the `type(c_ptr)` from `hipMalloc`;
* `rocsolver_?larfg` (`alpha`, `tau`), `rocsolver_?larf` (`alpha`) and their
`_64` variants, a variable: pass a device pointer;
* the double-precision hipFFTW guru planners, a scalar or an array element
as `dims` or `howmany_dims`: pass an array;
* the single-precision hipFFTW guru planners (`fftwf_plan_guru*`),
`c_loc(dims)`: pass the `type(fftw_iodim)` array itself.
`c_loc(dims)`: pass the `type(fftw_iodim)` array itself;
* `hipfftGetProperty`, a rank-1 array as `value`: pass an integer.

One more change compiles but behaves differently: the array of host pointers
of `hipsolverRfBatchSetupHost`, `hipsolverRfBatchResetValues` and
`hipsolverRfBatchSolve` is now `type(c_ptr), value`. Pass `c_loc` of the
array; code that passed its first element, `arr(1)`, now passes the first
pointer instead of the address of the array.
* **Breaking.** `hipfort_roctx` is generated from
* **Breaking.** Calling a `bind(C)` specific directly, by its `<routine>_`
name, with a Fortran variable where the C API takes a pointer no longer
compiles for the 223 specifics whose argument became `type(c_ptr), value`,
such as `hipGetDevice_`, `hipMemGetInfo_` and `hipfftGetSize_`. Call the
generic, `hipGetDevice(dev)`, which picks the `_typed` form, or pass `c_loc`
to the specific.
* **Breaking.** `hipfort_roctx` is now generated from
`rocprofiler-sdk-roctx/roctx.h`, which is the header behind the library that
`hipfort::roctx` links, rather than the legacy `roctracer/roctx.h`. Its
`const char*` arguments are now `type(c_ptr)`, like every other `char*`
argument in hipfort. Pass `c_loc` of a NUL-terminated
`character(kind=c_char)` array instead of a Fortran string:
`hipfort::roctx` links. The hand-written 0.9.0 module followed the legacy
`roctracer/roctx.h`. Its `const char*` arguments are bound as `type(c_ptr)`,
and each of these routines also accepts a `character(kind=c_char)` string,
passed by address without a copy. As in 0.9.0, the caller supplies the
terminator:

```fortran
character(kind=c_char), dimension(5), target :: msg = &
[c_char_"z", c_char_"o", c_char_"n", c_char_"e", c_null_char]
ret = roctxRangePush(c_loc(msg))
ret = roctxRangePush("zone"//c_null_char)
```

A `character(kind=c_char)` array is no longer accepted: pass `c_loc` of it.

`hipfort_roctx` is now AMD-only and is no longer part of
`libhipfort-nvptx.a`.
* **Breaking, for anyone compiling the `.F90` files by hand.** The preprocessor
Expand All @@ -141,14 +160,17 @@
unchanged.
* hipfort now follows the platform install layout. Libraries and modules go to
`CMAKE_INSTALL_LIBDIR`/`CMAKE_INSTALL_INCLUDEDIR` as set by `GNUInstallDirs`
(`lib64` on Fedora, RHEL and SUSE, `lib/<triplet>` on Debian multiarch)
instead of a literal `lib`. A user-supplied `-DCMAKE_INSTALL_LIBDIR` no longer
drops the `fortran/<compiler>` subdirectory. The package config shim and the
version file are installed beside the config they point to.
(`lib64` on Fedora, RHEL and SUSE, `lib/<triplet>` on Debian multiarch when
installing to `/usr`) instead of a literal `lib`. A user-supplied
`-DCMAKE_INSTALL_LIBDIR` no longer drops the `fortran/<compiler>`
subdirectory. The package config shim and its version file go to
`<libdir>/cmake/hipfort` instead of a literal `lib/cmake/hipfort`, so
`find_package(hipfort)` finds them on these platforms.
* The package now requires `hip-runtime-amd >= 6.0.0`, the first release that
exports the `hipGetDevicePropertiesR0600` symbol that hipfort binds.
* Every source and build file carries the same MIT license header with an
`SPDX-License-Identifier` line.
* Every hipfort source and build file carries the same MIT license header with
an `SPDX-License-Identifier` line. The two vendored `cmake/Modules/Set*.cmake`
files keep their upstream form.

### Removed

Expand All @@ -171,16 +193,20 @@
arrays of device pointers by reference, so rocBLAS received a host address.
* `rocsolver_chegvdx` and `rocsolver_zhegvdx` take `nev` as a device pointer,
like the rest of the `*gvdx` family (see *Changed*).
* `rocsolver_?larfg`, `rocsolver_?larf` and their `_64` variants take `alpha`
(and `tau` for `?larfg`) as a device pointer, `type(c_ptr), value`. They took
`real`/`complex` variables, whose host address rocSOLVER dereferenced on the
device (see *Changed*).
* The eight-byte integer overloads of the BLAS `Set`/`Get` `Vector`/`Matrix`
routines and their `Async` variants declared their arrays `integer(c_long)`,
which is four bytes on LLP64 targets. They are `integer(c_int64_t)` now.
* `hipfort::hipblas` was silently skipped when ROCm was installed outside the
default CMake search prefixes. `ROCM_PATH` is now added to
`CMAKE_PREFIX_PATH`, so the dependencies of the hipBLAS package resolve as
well.
* Release builds passed wrong flags to two compilers: `-ta=host` instead of
`-tp=host` to NVHPC/PGI, and the removed `-vec-report0` instead of `-vec` to
Intel.
* Wrong compiler flags: with `-DBUILD_NATIVE=ON`, NVHPC/PGI got `-ta=host`
instead of `-tp=host`, and Release builds passed Intel the removed
`-vec-report0` instead of `-vec`.
* Tests can now fail. 413 failure branches ended in a bare `call exit`, which
returns 0 under gfortran. Fixing this uncovered 27 tests that had been failing
silently. Several other tests that could not fail, tolerances that did not
Expand All @@ -199,7 +225,7 @@
* Under `USE_CUDA_NAMES`, the compatibility-API `hipsolver?gesvd_bufferSize`
and `hipsolver??gels_bufferSize` were bound to the `cusolverDn` routines of
the same name, whose arguments differ. They are now ROCm-only.
* The modules compile without warnings under gfortran `-Wall`. The bind(C)
* The modules compile without warnings under gfortran `-Wall`. The `bind(C)`
interfaces declared enum-valued arguments and results as
`integer(kind(<enumerator>))`, which gfortran reported about 12,000 times as
possibly not C interoperable (`-Wc-binding-type`). They are declared
Expand Down
9 changes: 5 additions & 4 deletions lib/CMakeLists.txt
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -459,10 +459,11 @@ if(HIPFORT_EXTENDED_TESTS)
set(_amd_dir "${ROCM_PATH}/lib")
endif()
# NB: the HIP runtime is libamdhip64, not "libhip".
# rocTX ships under two names and they are NOT interchangeable -- librocprofiler-sdk-roctx exports
# the plain symbols and is what hipfort::roctx links, while the legacy
# libroctx64 exports them versioned (roctxMarkA@@ROCTX_4.1) -- so resolve the
# same one the component does, falling back to the legacy name.
# rocTX ships under two names and they are NOT interchangeable --
# librocprofiler-sdk-roctx exports the plain symbols and is what
# hipfort::roctx links, while the legacy libroctx64 exports them versioned
# (roctxMarkA@@ROCTX_4.1) -- so resolve the same one the component does,
# falling back to the legacy name.
set(_amd_libs hipfftw hipfft rocfft hipsolver rocsolver hipblas rocblas
rocrand hiprand hipsparse rocsparse roctx amdhip64)
set(_amd_names_roctx rocprofiler-sdk-roctx roctx64)
Expand Down