[amontoison]

A sandbox preview, not for merge. It shows what the Fortran bindings look like once they move into the bindings/fortran/ tree of rocm-systems and rocm-libraries, apart from the libraries they wrap, so the shape can be tried before anything lands. lib/rocm-systems/bindings/fortran/ and lib/rocm-libraries/bindings/fortran/ mirror those trees, with one project per library (the monorepos build HIP with ROCTx, and hipFFT with hipFFTW, from one project). The bindings are generated from the ROCm 10.2 headers with rocm-fortran-generator main, and are byte-identical to those of the per-library pull requests below. The HIP module of rocm-systems#11923 is generated from the headers of rocm-systems develop (HIP 7.18), so it also has hipExtMemcpyBatchAsync and hipKernelSetCacheConfig, which ROCm 10.2 does not export; the sandbox keeps the ROCm 10.2 one.

The real pull requests are one per library. This branch is where you can build the whole set at once.

Library	Pull request
HIP	rocm-systems#11923
ROCTx	rocm-systems#14024
rocBLAS	rocm-libraries#12533
hipBLAS	rocm-libraries#12370
rocSOLVER	rocm-libraries#12582
hipSOLVER	rocm-libraries#12581
rocSPARSE	rocm-libraries#12514
hipSPARSE	rocm-libraries#12368
rocFFT	rocm-libraries#12366
hipFFT, hipFFTW	rocm-libraries#12367
rocRAND	rocm-libraries#12505
hipRAND	rocm-libraries#12506

Status (2026-11-07)

All the binding pull requests are up to date and green, locally and in TheRock CI (rocm-libraries#15008, a test-only PR that merges the ten rocm-libraries bindings with ROCm/TheRock#9446: 1271 binding tests pass on MI300). The ten rocm-libraries PRs are still drafts until the RFC is approved.

Pull request	What	State
ROCm/TheRock#9433	RFC0015 update	up to date, to be approved
rocm-systems#11923, #14024	HIP, ROCTx bindings	up to date, in review; #14024 contains #11923 and merges after it
the ten rocm-libraries PRs above	math-library bindings	up to date, draft
ROCm/TheRock#9446	TheRock builds and tests the bindings (fortran-bindings stage, one GPU test job per binding)	up to date, in review
ROCm/rocm-cmake#303	find_package(<lib>) also loads <lib>-fortran when installed	up to date, in review; the bindings work without it
hipfort#548	migration guides (hipFORT, in-tree modules, CUDA Fortran) and hipFORT deprecation	draft until the RFC is approved

Each binding installs its CMake package in <libdir>/cmake/<lib>-fortran/ and is ready for the rocm-cmake hook (hipfft-fortran also loads hipfftw-fortran).

Merge order: rocm-systems#11923, then #14024, then the ten rocm-libraries PRs (each appends its entry to .github/repos-config.json, so they rebase in turn, regenerating the bindings), then ROCm/TheRock#9446.

After the merge:

• packaging in TheRock: one amdrocm-<family>-fortran-devel package per family and the amdrocm-fortran metapackage, which also removes the *-fortran entries from KNOWN_UNCOVERED_COMPONENTS;
• deprecation of the hand-written modules: rocRAND and hipRAND (their legacy <lib>-fortran-config.cmake, whose removal enables the rocm-cmake hook for them), hipSOLVER (EXPORT_FORTRAN_BINDINGS OFF by default), rocSPARSE docs;
• before the bindings ship: let the HIP arguments that C requires to be NULL take c_null_ptr (hipStreamBeginCaptureToGraph, hipModuleLaunchKernel/hipDrvLaunchKernelEx, hipGraphAddMemFreeNode);
• in TheRock, when the pins include the bindings: regenerate the consumer graph so a C library change runs its binding's tests.

What changes

lib/hipfort, a flat directory of 35 files, becomes 13 directories, one per library, laid out as in the repository each will be committed to:

lib/rocm-systems/bindings/fortran/{hip,roctx}
lib/rocm-libraries/bindings/fortran/{rocblas,hipblas,rocsolver,hipsolver,rocsparse,hipsparse,rocfft,hipfft,hipfftw,rocrand,hiprand}

Each is one self-contained module named after its library, with the enums, types, handles and the <lib>Check helper folded in. use hipfort_rocblas becomes use rocblas. hipFFT is the one library with two modules, see below: 14 modules in all.

hipFFT: hipfft and hipfftxt

As CUDA Fortran's cufft and cufftXt, the hipFFT binding has two modules, built into the same libhipfft_fortran.a (hip::hipfft_fortran, find_package(hipfft-fortran)):

Module	File	Lines	Contents
hipfft	lib/rocm-libraries/bindings/fortran/hipfft/hipfft.F90	1683	the plans, hipfftExec*, the work area and stream routines
hipfftxt	lib/rocm-libraries/bindings/fortran/hipfft/hipfftxt.F90	651	every hipfftXt* and hipfftMp* function, hipfftXtSetWorkArea included, with the hipXtDesc/hipLibXtDesc types and the Xt enums; it uses and re-exports hipfft

A program calling the Xt or multi-process API writes use hipfftxt (the Xt tests do); any other keeps use hipfft. rocm_add_fortran_binding takes several SOURCE files for this and installs one .mod per source. The files are byte-identical to rocm-libraries#12367. hipFORT itself has the same split (hipfort_hipfftxt).

Each also builds its own archive and config package, the way its binding project will:

target	roc::rocblas_fortran, hip::hipfft_fortran
archive	librocblas_fortran.a
module	rocblas.mod (hipFFT: hipfft.mod, hipfftxt.mod)
package	find_package(rocblas-fortran)

The .mod and the archive install under include/fortran/<compiler> and lib/fortran/<compiler>, because a .mod is not portable across compilers.

find_package(hipfort) and the single libhipfort-amdgcn.a are gone, as they are after the split. So is the CUDA backend.

Try it

cmake -S . -B build -DCMAKE_Fortran_COMPILER=amdflang
cmake --build build -j$(nproc)
ctest --test-dir build

1296 tests, on by default, the same set as develop (fewer with FORTRAN_ARRAY_INTERFACES=none or assumed-rank, which skip the f2008 tests). They need a GPU to pass.

Build one binding on its own, against the installed library, as TheRock and a user with another compiler do:

cmake -S lib/rocm-libraries/bindings/fortran/rocblas -B build-rocblas -DCMAKE_Fortran_COMPILER=amdflang -DCMAKE_PREFIX_PATH=/opt/rocm
cmake --build build-rocblas

Install and consume it, which is what a user does:

cmake --install build --prefix /tmp/rocm-fortran

find_package(rocblas-fortran REQUIRED)   # find_dependency()s rocblas itself
find_package(hip-fortran REQUIRED)
target_link_libraries(app PRIVATE roc::rocblas_fortran hip::hip_fortran)

cmake -S your-app -B build-app -DCMAKE_PREFIX_PATH="/tmp/rocm-fortran;/opt/rocm"

Without CMake:

amdflang app.f90 -I/tmp/rocm-fortran/include/fortran/amdflang -L/tmp/rocm-fortran/lib/fortran/amdflang -lrocblas_fortran -L/opt/rocm/lib -lrocblas -lamdhip64 -o app

Options

Option	Default	Notes
BUILD_FORTRAN_BINDINGS	ON	skipped, not an error, when no Fortran compiler is found
BUILD_FORTRAN_CLIENTS	ON	each binding's test/; BUILD_TESTING is honoured once, with a warning
FORTRAN_ARRAY_INTERFACES	assumed-shape	none / assumed-shape / assumed-rank
BUILD_FORTRAN_CLIENT_SAMPLES	OFF	the libraries' own samples, see below

The libraries' own samples

test/rocm-libraries/ holds the 18 Fortran samples rocBLAS and rocSPARSE ship in clients/samples/, copied verbatim and compiled unedited. They open with use rocblas and use rocsparse, the names the generated bindings provide, so building them is what shows the generated module is a drop-in for the hand-written one it replaces.

cmake -S . -B build -DBUILD_FORTRAN_CLIENTS=ON -DBUILD_FORTRAN_CLIENT_SAMPLES=ON
cmake --build build --target hipfort_sample_rocblas_example_fortran_axpy

OFF by default because 16 of the 18 still need one line each: a use <lib>_enums to delete for two of them, and a c_loc() around the character buffer passed to the revision helpers for the fourteen rocSPARSE ones, each of which opens by printing the revision. Details in test/rocm-libraries/README.md.

[github-advanced-security]

You are seeing this message because GitHub Code Scanning has recently been set up for this repository, or this pull request contains the workflow file for the Code Scanning tool.

What Enabling Code Scanning Means:

• The 'Security' tab will display more code scanning analysis results (e.g., for the default branch).
• Depending on your configuration and choice of analysis tool, future pull requests will be annotated with code scanning analysis results.
• You will be able to see the analysis results for the pull request's branch on this overview once the scans have completed and the checks have passed.

For more information about GitHub Code Scanning, check out the documentation.

[github-actions]

Security scan found no critical or high findings on lines changed by this pull request

2 medium, low, or unrated finding(s) are not listed; the scan run has them.

View the security scan run for complete reports.

“New” means the finding is located in a changed file on an added or modified line; it is not a base-versus-head findings delta.