Enable flang and openmp use throughout ROCm. by stellaraccident · Pull Request #3928 · ROCm/TheRock

stellaraccident · 2026-03-12T03:49:01Z

Summary

Splits the ROCm LLVM build into a small bootstrap compiler plus optional Fortran and OpenMP/offload layers, then adjusts TheRock's build topology so downstream projects can consume the right compiler view without forcing every build to pay for Flang/offload.

The main goals are:

Build a base LLVM compiler that contains Clang/LLD/backends/device-libs, but not Flang, MLIR, OpenMP/offload, or flang-rt.
Keep Fortran and offload available for full Linux builds, while allowing minimal HIP/runtime builds to disable Fortran/offload completely.
Make downstream Fortran use explicit via TheRock subproject metadata and the ROCM_BUILD_FORTRAN_LIBS flag.
Avoid advertising LLVM's installed OpenMP CMake config as the global provider for find_package(OpenMP).
Collapse the old foundation stage into compiler-runtime so CI does not add a lightweight root stage or require a second LLVM checkout.
Move hip-tests and rocrtst to a separate runtime-tests stage so runtime tests do not rebuild on downstream stages that only need HIP/core runtime artifacts.

LLVM Layering

The LLVM-related artifacts are now layered as follows:

amd-llvm-base: bootstrap compiler layer. Builds the LLVM project with clang, lld, clang-tools-extra, X86/AMDGPU targets, compiler-rt/device-libs, and no Flang/MLIR/OpenMP/offload. This is the base compiler used by later layers.
amd-llvm: compiler infrastructure layer. Builds amd-comgr and hipcc on top of amd-llvm-base.
amd-llvm-flang: optional Linux-only Flang frontend layer. Builds Flang/MLIR as implementation details needed by the frontend, but does not install MLIR into the final ROCm dist.
amd-llvm-offload: optional Linux-only runtime layer. Builds OpenMP/offload runtimes and, when Flang is enabled, flang runtime pieces. Depends on amd-llvm, amd-llvm-flang when enabled, and core-runtime.

hip-clr uses the amd-llvm-offload compiler view when that target exists. That keeps HIP, Flang, OpenMP headers/libraries, and offload runtime visibility together in the amd-hip toolchain view without making the base compiler advertise OpenMP globally.

Build Topology

The old foundation stage has been removed. Its former artifact groups are now part of compiler-runtime:

third-party-sysdeps
base
compiler
core/runtime/profiler groups

This avoids a root stage that mostly adds critical path and avoids splitting LLVM-related source checkout/build behavior across stages.

A new runtime-tests stage contains:

core-hiptests
rocrtst

runtime-tests depends on the runtime/compiler artifacts it needs and runs parallel to math-libs. Linux builds both hip-tests and rocrtst; Windows has the same stage so core-hiptests artifacts are still produced, while rocrtst remains disabled by platform metadata.

Fortran Controls

This adds or wires:

THEROCK_ENABLE_FLANG
THEROCK_ENABLE_AMD_LLVM_OFFLOAD
THEROCK_FLAG_BUILD_FORTRAN_LIBS
FORTRAN_OPTIONAL / FORTRAN_REQUIRED subproject declaration keywords
ROCM_BUILD_FORTRAN_LIBS=ON/OFF propagation to downstream subprojects

Known Fortran consumers can now declare whether Fortran is optional or required. When Fortran libs are disabled, optional users can configure without forcing Flang/offload into the build.

OpenMP Handling

HIP-facing projects get OpenMP visibility through the amd-hip toolchain view when hip-clr is built on top of amd-llvm-offload.
rocprofiler-systems gets a targeted pre-hook that passes libomptarget explicitly and seeds module-mode FindOpenMP cache entries where needed.
GCC-configured subprojects use GCC's own libgomp path, avoiding accidental selection of the offload package's libgomp.so shim.

Downstream Wiring

The change updates downstream projects that need Fortran/OpenMP awareness, including RAND/BLAS-family projects and rocprofiler-systems.

Notable fixes:

Math library Fortran wrappers are gated by ROCM_BUILD_FORTRAN_LIBS.
rocprofiler-systems receives the offload library directory explicitly for libomptarget.
rocprofiler-systems examples are disabled in TheRock integration builds to avoid using deep example builds as the first test of the Fortran/offload toolchain.

Verification

Local verification was done on a gfx1100 machine.

Static checks:

python build_tools/topology_to_cmake.py --validate-only
python build_tools/configure_stage.py --stage compiler-runtime --print
python build_tools/configure_stage.py --stage runtime-tests --print
python build_tools/configure_stage.py --stage runtime-tests --platform windows --print
python -m unittest discover -s build_tools/github_actions/tests -p configure_multi_arch_ci_test.py
python -m unittest build_tools.tests.build_topology_test build_tools.github_actions.tests.post_stage_upload_test
git diff --check

No-Fortran HIP/RAND build:

Configured with THEROCK_ENABLE_ALL=OFF, THEROCK_ENABLE_HIP_RUNTIME=ON, THEROCK_ENABLE_RAND=ON, THEROCK_ENABLE_FLANG=OFF, THEROCK_ENABLE_AMD_LLVM_OFFLOAD=OFF, and THEROCK_FLAG_BUILD_FORTRAN_LIBS=OFF.
Built dist-rocm.
Verified the dist contains HIP/RAND pieces and does not contain Flang/OpenMP/offload outputs such as amdflang, flang, libomp.so, libomptarget.so, omp_lib.mod, or libflang_rt*.
Built and ran a tiny HIP smoke test with hipcc --offload-arch=gfx1100.

Fortran/offload toolchain build:

Configured with compiler/core runtime/HIP/RAND/BLAS enabled plus THEROCK_ENABLE_FLANG=ON, THEROCK_ENABLE_AMD_LLVM_OFFLOAD=ON, and THEROCK_FLAG_BUILD_FORTRAN_LIBS=ON.
Built and verified amd-llvm-base, amd-llvm, amd-llvm-flang, amd-llvm-offload, HIP runtime, RAND, and BLAS-family artifacts.
Verified dist/rocm contains amdflang, flang, omp_lib.mod, flang runtime libraries, libomp.so, libomptarget.so, AMDGCN device runtime pieces, BLAS/RAND libraries, and no installed MLIR CMake package in the final ROCm dist.
Passed small Fortran, Fortran OpenMP, C OpenMP offload, Fortran OpenMP/offload, and HIP smoke tests.

BLAS/RAND integration:

Expunged cached downstreams after toolchain changes.
Built/refreshed mxDataGenerator, hipBLASLt, rocBLAS, hipBLAS, rocRAND, and hipRAND.
Confirmed Fortran wrappers are enabled in the Fortran build and absent in the no-Fortran build.

marbre · 2026-03-12T13:27:14Z

Builds are failing:

[rocprofiler-systems configure] -- Checking out 'https://github.com/OpenMP-Validation-and-Verification/OpenMP_VV.git' @ 'master'...
[rocprofiler-systems configure] Cloning into 'external/ompvv'...
[rocprofiler-systems configure] CMake Error at cmake/MacroUtilities.cmake:20 (message):
[rocprofiler-systems configure]   [rocprofiler-systems] Could not find amdflang.  This is required for the
[rocprofiler-systems configure]   OMPVV tests.
[rocprofiler-systems configure] Call Stack (most recent call first):
[rocprofiler-systems configure]   examples/openmp/external/CMakeLists.txt:184 (rocprofiler_systems_message)
[rocprofiler-systems configure] 
[rocprofiler-systems configure] 
[rocprofiler-systems configure] -- Configuring incomplete, errors occurred!

ScottTodd

Overall looks good once the CI passes

stellaraccident · 2026-03-13T05:49:45Z

Try 2.

marbre · 2026-03-18T10:30:29Z

Try 2.

[rocRoller configure] CMake Error at /__w/TheRock/TheRock/rocm-libraries/shared/mxdatagenerator/CMakeLists.txt:14 (target_link_libraries):
[rocRoller configure]   The link interface of target "mxDataGenerator" contains:
[rocRoller configure] 
[rocRoller configure]     OpenMP::OpenMP_CXX
[rocRoller configure] 
[rocRoller configure]   but the target was not found.  Possible reasons include:
[rocRoller configure] 
[rocRoller configure]     * There is a typo in the target name.
[rocRoller configure]     * A find_package call is missing for an IMPORTED target.
[rocRoller configure]     * An ALIAS target is missing.
[rocRoller configure] 
[rocRoller configure] 
[rocRoller configure] 
[rocRoller configure] CMake Error at test/common/CMakeLists.txt:7 (target_link_libraries):
[rocRoller configure]   Target "common-test-utilities" links to:
[rocRoller configure] 
[rocRoller configure]     OpenMP::OpenMP_CXX
[rocRoller configure] 
[rocRoller configure]   but the target was not found.  Possible reasons include:
[rocRoller configure] 
[rocRoller configure]     * There is a typo in the target name.
[rocRoller configure]     * A find_package call is missing for an IMPORTED target.
[rocRoller configure]     * An ALIAS target is missing.

😢

ronlieb · 2026-04-01T01:26:58Z

what are the valid combinations of the 3 amd-llvm*

can amd-llvm + amd-llvm-offload be built ?

same for amd-llvm + amd-llvm-flang ?

if i specify amd-llvm-flang does it imply amd-llvm ? or do i need to explicitly as for amd-llvm?

LLVM's openmp-config.cmake exports OpenMP::omp, not OpenMP::OpenMP_CXX. When provide_package redirected find_package(OpenMP) to the LLVM config, downstream consumers (rocRoller, mxdatagenerator) that link against OpenMP::OpenMP_CXX failed at configure time. Removing the provide_package lets CMake's built-in FindOpenMP module handle detection, which correctly creates the OpenMP::OpenMP_CXX target. Fixes: #3928 (marbre report on "Try 2") Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

stellaraccident · 2026-04-01T01:35:43Z

what are the valid combinations of the 3 amd-llvm*

can amd-llvm + amd-llvm-offload be built ?

same for amd-llvm + amd-llvm-flang ?

if i specify amd-llvm-flang does it imply amd-llvm ? or do i need to explicitly as for amd-llvm?

My agent organized my answer a bit:

The three sub-projects form a dependency chain:

amd-llvm-base → always built. Minimal compiler: clang, lld, compiler-rt, device-libs, openmp host runtime.
amd-llvm → depends on amd-llvm-base. Adds comgr, hipify, etc. This is what most of ROCm needs.
amd-llvm-flang → depends on amd-llvm. Adds the flang compiler frontend (+ MLIR). Linux-only, gated by THEROCK_ENABLE_FLANG.
amd-llvm-offload → depends on amd-llvm, and optionally on amd-llvm-flang. Adds libomptarget, device RTLs, and (if flang is enabled) flang-rt device runtime. Linux-only.

Valid combinations:

amd-llvm + amd-llvm-offload (no flang): Yes. You get libomptarget + device RTLs but no flang-rt. C/C++ OpenMP offload works, Fortran doesn't.
amd-llvm + amd-llvm-flang (no offload): Technically valid but not useful — you get the flang compiler binary but no Fortran runtime (flang-rt lives in amd-llvm-offload), so you can't link anything.
All three: The full stack. flang frontend + flang-rt + offload runtimes.

Dependencies are enforced via artifact_deps in the topology, so enabling amd-llvm-flang automatically pulls in amd-llvm — you don't need to request it explicitly.

marbre · 2026-04-01T10:54:06Z

[rocprofiler-systems] error: Semantic errors in /__w/TheRock/TheRock/build/profiler/rocprofiler-systems/build/examples/openmp/external/ompvv/tests/5.0/teams/test_team_default_shared.F90
[rocprofiler-systems] /__w/TheRock/TheRock/build/profiler/rocprofiler-systems/build/examples/openmp/external/ompvv/tests/5.0/teams/test_team_default_shared.F90:28:7: error: Cannot parse module file for module 'iso_fortran_env': Source file 'iso_fortran_env.mod' was not found
[rocprofiler-systems]     USE iso_fortran_env
[rocprofiler-systems]         ^^^^^^^^^^^^^^^
[rocprofiler-systems] /__w/TheRock/TheRock/build/profiler/rocprofiler-

In the latest run rocprofiler-systems is unhappy about the Fortran related changes.

stellaraccident · 2026-05-02T05:50:05Z

A month later - needs more cleanup before landing but pushing a WIP branch that passes my local build to see what CI says.

Changes: - Add amd-llvm-base as a foundation compiler artifact without Flang/OpenMP/offload. - Add optional amd-llvm-flang and amd-llvm-offload layers with flang-rt and OpenMP runtime packaging. - Wire Fortran optional subproject behavior and initial ROCm consumers. - Add tactical LLVM patches for split OpenMP/offload builds. Generated with Claude Code. Co-Authored-By: Claude <noreply@anthropic.com>

Changes: - Stop advertising LLVM OpenMP config as the global OpenMP provider. - Build hip-clr against amd-llvm-offload when that layer is enabled so the normal amd-hip toolchain carries HIP, Flang, and OpenMP together. - Keep rocprofiler-systems' LLVM OpenMP config shim local and explicit. Generated with Claude Code Co-authored-by: Claude <noreply@anthropic.com>

Changes: - Remove the foundation stage and fold sysdeps/base into compiler-runtime. - Add a runtime-tests stage for hip-tests and rocrtst in multi-stage CI. - Keep rocprofiler-systems on module-mode OpenMP hints instead of LLVM OpenMP config mode. Generated with Claude Code

Changes: - keep kpacked test host binaries paired with per-ISA artifacts instead of publishing them through shared generic test artifacts - keep setuptools and wheel in setup_venv-created venvs and disable build isolation for explicit artifact-only installs - cover the no-index setup_venv command shape in unit tests Claude Code

Changes: - normalize hipblas-test's Windows process status to 0 when GTest reports success - keep quick_exit on Windows so post-main DLL cleanup still does not run Claude Code

Partial landing split from #3928. Changes: - Replace install(FILES) for the root llvm and amdgcn directory symlinks with install(CODE) that creates the symlinks at install time. - Create the install prefix inside the install code before creating root-level symlinks; local validation showed the closed #4301 form failed when the prefix did not exist yet. - Leave the existing amdflang symlink behavior unchanged; conditional Fortran symlink handling remains with the split compiler work. Rationale: - CMake install(FILES) rejects directory symlinks, which breaks aux-overlay installation for the SDK layout symlinks tracked by #3714. Validation: - cmake -S base/aux-overlay -B /tmp/therock-aux-overlay-test-build -GNinja -DTHEROCK_SOURCE_DIR=/srv/vm-shared/projects/therock-workspace1/worktrees/TheRock -DTHEROCK_PACKAGE_VERSION=0.0.0 -DGPU_TARGETS=gfx1100 - cmake --build /tmp/therock-aux-overlay-test-build - cmake --install /tmp/therock-aux-overlay-test-build --prefix /tmp/therock-aux-overlay-test-install - verified llvm, amdgcn, and bin/amdclang are symlinks in /tmp/therock-aux-overlay-test-install - git diff --cached --check Generated with Codex

Partial landing split from #3928. Changes: - When a stage depends on an artifact group, include not only the artifacts produced by that group but also each artifact's transitive artifact dependencies. - Add topology regression coverage for a group dependency where a pulled-in artifact depends on additional toolchain artifacts from the same producer group. Rationale: - Multi-stage CI fetches stage inputs from get_inbound_artifacts(). Without this closure, a downstream stage can fetch a higher-level artifact while missing lower-level artifacts that it needs at configure/build time. - This is a standalone topology correctness fix that derisks the split compiler artifacts from #3928. Validation: - python -m unittest build_tools.tests.build_topology_test - python build_tools/topology_to_cmake.py --validate-only - git diff --cached --check Generated with Codex

ronlieb · 2026-05-11T17:55:24Z

+
+When OpenMP is built standalone against an installed LLVM tree,
+LIBOMP_HEADERS_INTDIR can point at the installed LLVM resource include
+directory instead of the local runtime build directory. That makes GCC


shouldnt this build of openmp files always use clang rather than gcc ?

lumachad · 2026-05-11T18:26:07Z

I wonder if we need to wire something in the cmake/therock_sanitizers.cmake to handle passing sanitizer options to flang.

Meinersbur · 2026-05-13T23:16:01Z

+if(TARGET FortranSupport AND TARGET MLIRBuiltinTypeInterfacesIncGen)
+  add_dependencies(FortranSupport MLIRBuiltinTypeInterfacesIncGen)
+endif()
+if(TARGET FIRSupport AND TARGET MLIRLinalgOpsIncGen)
+  add_dependencies(FIRSupport MLIRLinalgOpsIncGen)
+endif()
+if(TARGET FortranSupport)
+  target_include_directories(FortranSupport PRIVATE
+    ${CMAKE_BINARY_DIR}/flang/lib/Support)
+endif()


Followup from today's conversation:

I think duplicating the CMakeLists.txt from upstream is highly undesirable. If changes upstream, it is hard to debug what changes and replicated it in this copy of CMakeLists.txt. FLANG_INCLUDE_QUADMATH_H is a great example for this: Build will not fail initially, but REAL(16) will be silently unavailable and some poor soul has to find out why. I am planning larger changes to this file upstream (llvm/llvm-project#171515) and would rather not have to do the same a second time for TheRock.

Upstream LLVM already has a means to build components separately in standalone builds. That is: LLVM+Flang -> MLIR -> Flang. You mentioned the disadvantage is that the MLIR build includes a lot of other stuff that is actually not needed for Flang, and a common build dir just handles dependency resolution.

This is an argument for MLIR, Flang which are build through LLVM_ENABLE_PROJECTS, but not for LLVM_ENABLE_RUNTIMES (i.e. amd-llvm-offload). A runtimes build can build multiple runtimes via passing them all to LLVM_ENABLE_RUNTIMES, e.g -DLLVM_ENABLE_RUNTIMES=openmp;offload.

For MLIR, it would be less work to only build the components actually needed, either by ninja MLIRLinalgOpsIncGen ..., creating a new target in our downstream llvm-project (such as

add_custom_target(mlir-flang-deps) add_dependencies(mlir-flang-deps MLIRLinalgOpsIncGen ...)

) that depend only on needed targets, or straight up commenting-out upstream components that we do not need. Dependencies are already managed in this CMakeLists copy and a missing dependency due to an upstream change is much more obvious than hunting down changes to configuration parameters.

If we have to manually manage out top-level dependencies anyway, why not keep everything in a single build dir and create targets that only build/install components that are needed (replacing ninja all and ninja install)? From AI:

# Install base_lib into the "base" component install(TARGETS base_lib DESTINATION lib COMPONENT base) # Create the custom target for 'ninja install-base' add_custom_target(install-base # 1. The command to run the component-specific install COMMAND ${CMAKE_COMMAND} --install ${CMAKE_BINARY_DIR} --component base # 2. The targets that MUST be built before installing DEPENDS base_lib COMMENT "Building and installing the 'base' component..." )

(upstream LLVM already uses COMPONENT in install commands)

or reconfigure the same build dir with LLVM_ENABLE_PROJECTS=${LLVM_ENABLE_PROJECTS};flang a second build to incrementally build flang on top of the previous build.

Partial landing split from #3928. Changes: - Remove the foundation stage and fold sysdeps/base into compiler-runtime. - Add a runtime-tests stage for hip-tests and rocrtst so runtime test builds no longer sit on the compiler-runtime critical path. - Update multi-stage workflow dependencies, helper text, docs, and CI unit-test expectations for the new stage layout. - Keep rocprofiler-systems OpenMP/example changes out of this split; those remain separate from the topology refactor. Evidence: - python build_tools/topology_to_cmake.py --validate-only - python -m pytest build_tools/tests/build_topology_test.py build_tools/github_actions/tests/configure_multi_arch_ci_test.py - python build_tools/configure_stage.py --stage compiler-runtime --print - python build_tools/configure_stage.py --stage runtime-tests --print Generated with Claude Code

Partial landing split from #3928. Changes: - Remove the foundation stage and fold sysdeps/base into compiler-runtime. - Add a runtime-tests stage for hip-tests and rocrtst so runtime test builds no longer sit on the compiler-runtime critical path. - Update multi-stage workflow dependencies, helper text, docs, and CI unit-test expectations for the new stage layout.

Partial landing split from #3928. Changes: - Keep setuptools and wheel available when setup_venv updates a venv so ROCm sdists can build without resolving build dependencies from the package index under test. - Treat an explicit empty --index-url as an artifact-only install request and pass pip --no-index --no-build-isolation while preserving --find-links. - Add unit coverage for the empty-index command shape. Validation: - python -m unittest build_tools.tests.setup_venv_test - git diff --cached --check Generated with Codex Co-authored-by: Scott Todd <Scott.Todd@amd.com>

Partial upstream split from ROCm/TheRock#3928. Changes: - Rename the local HIP C binding function result from c_int to hip_status in the rocBLAS Fortran samples. - Declare the renamed result explicitly as integer(c_int) under implicit none. Rationale: - The samples imported iso_c_binding::c_int and also used result(c_int), which makes the function result name collide with the imported kind name. - Flang diagnoses the collision and missing explicit result type; gfortran also accepts the explicit non-conflicting result name. Validation: - Applied on current rocm-libraries develop after git pull --ff-only origin develop. - gfortran -fsyntax-only on example_fortran_{axpy,gemv,scal}.f90 with temporary stub rocblas modules. - git diff --check. Generated with Codex --------- Co-authored-by: Michael Klemm <michael.klemm@amd.com> Co-authored-by: Copilot Autofix powered by AI <175728472+Copilot@users.noreply.github.com> Co-authored-by: Torre Zuk <Torre.Zuk@amd.com>

rocBLAS: Fix Fortran sample HIP result names Partial upstream split from ROCm/TheRock#3928. Changes: - Rename the local HIP C binding function result from c_int to hip_status in the rocBLAS Fortran samples. - Declare the renamed result explicitly as integer(c_int) under implicit none. Rationale: - The samples imported iso_c_binding::c_int and also used result(c_int), which makes the function result name collide with the imported kind name. - Flang diagnoses the collision and missing explicit result type; gfortran also accepts the explicit non-conflicting result name. Validation: - Applied on current rocm-libraries develop after git pull --ff-only origin develop. - gfortran -fsyntax-only on example_fortran_{axpy,gemv,scal}.f90 with temporary stub rocblas modules. - git diff --check. Generated with Codex

Partial upstream split from ROCm/TheRock#3928. Changes: - Rename the local HIP C binding function result from c_int to hip_status in the rocBLAS Fortran samples. - Declare the renamed result explicitly as integer(c_int) under implicit none. Rationale: - The samples imported iso_c_binding::c_int and also used result(c_int), which makes the function result name collide with the imported kind name. - Flang diagnoses the collision and missing explicit result type; gfortran also accepts the explicit non-conflicting result name. Validation: - Applied on current rocm-libraries develop after git pull --ff-only origin develop. - gfortran -fsyntax-only on example_fortran_{axpy,gemv,scal}.f90 with temporary stub rocblas modules. - git diff --check. Generated with Codex --------- Co-authored-by: Michael Klemm <michael.klemm@amd.com> Co-authored-by: Copilot Autofix powered by AI <175728472+Copilot@users.noreply.github.com> Co-authored-by: Torre Zuk <Torre.Zuk@amd.com>

github-project-automation Bot added this to TheRock Triage Mar 12, 2026

github-project-automation Bot moved this to TODO in TheRock Triage Mar 12, 2026

stellaraccident mentioned this pull request Mar 12, 2026

[Issue]: amd-llvm project isn't marked as exporting openmp #3839

Open

stellaraccident changed the title ~~POC: First class fortran/openmp/offload plumbing (take 2)~~ Split amd-llvm into base/flang/offload sub-projects and re-layer build order Mar 12, 2026

stellaraccident marked this pull request as ready for review March 12, 2026 06:37

stellaraccident requested review from ScottTodd and marbre as code owners March 12, 2026 06:37

stellaraccident requested a review from ronlieb March 12, 2026 06:40

ScottTodd reviewed Mar 12, 2026

View reviewed changes

Comment thread compiler/artifact-core-runtime.toml Resolved

Comment thread base/artifact-amd-llvm-base.toml Resolved

Comment thread BUILD_TOPOLOGY.toml Resolved

Comment thread cmake/therock_subproject.cmake Resolved

ScottTodd mentioned this pull request Mar 12, 2026

Split LLVM build: extract Flang and offload runtimes from critical path #3219

Draft

stellaraccident requested a review from skganesan008 as a code owner April 1, 2026 01:15

ScottTodd mentioned this pull request Apr 1, 2026

[docs] Adding an overarching CI doc page #4275

Merged

ScottTodd self-requested a review April 1, 2026 18:30

stellaraccident force-pushed the users/stellaraccident/fortran-first-class-take2 branch from 5c0a394 to 9879624 Compare May 2, 2026 05:38

stellaraccident commented May 2, 2026

View reviewed changes

stellaraccident mentioned this pull request May 2, 2026

[rocprofiler-systems] Examples need to be moved out of the main build ROCm/rocm-systems#5729

Closed

stellaraccident changed the title ~~Split amd-llvm into base/flang/offload sub-projects and re-layer build order~~ Enable flang and openmp use throughout ROCm. May 2, 2026

stellaraccident mentioned this pull request May 2, 2026

[aux-overlay] Use install(CODE) for directory symlinks #4301

Closed

stellaraccident and others added 4 commits May 3, 2026 21:23

Revert local debugging hack

429da3c

stellaraccident force-pushed the users/stellaraccident/fortran-first-class-take2 branch from baff225 to c83335c Compare May 4, 2026 04:23

stellaraccident added 2 commits May 4, 2026 02:58

Fix Windows hipBLAS success exit

23b97ff

Changes: - normalize hipblas-test's Windows process status to 0 when GTest reports success - keep quick_exit on Windows so post-main DLL cleanup still does not run Claude Code

stellaraccident mentioned this pull request May 11, 2026

Add CMake option to enable or disable Flang support #1589

Open

ronlieb reviewed May 11, 2026

View reviewed changes

lumachad mentioned this pull request May 11, 2026

[cmake] Handle sanitizers for autotools projects #5182

Closed

ScottTodd mentioned this pull request May 13, 2026

[amd-llvm] Provide OpenMP through super-project dependency provider #4708

Closed

4 tasks done

Meinersbur reviewed May 13, 2026

View reviewed changes

ScottTodd mentioned this pull request May 19, 2026

Add a variable to enable/disable flang build #5336

Closed

1 task done

This was referenced Jul 16, 2026

Fallback libhipcxx tests to AMDGPU_TARGETS #6597

Closed

Fallback libhipcxx tests to AMDGPU_TARGETS #6599

Closed

ScottTodd mentioned this pull request Aug 24, 2026

Is Fortran meant to be usable in the Windows ROCm SDK? flang.exe there is byte-identical to clang.exe #7596

Open

Conversation

stellaraccident commented Mar 12, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

LLVM Layering

Build Topology

Fortran Controls

OpenMP Handling

Downstream Wiring

Verification

Uh oh!

marbre commented Mar 12, 2026

Uh oh!

ScottTodd left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

stellaraccident commented Mar 13, 2026

Uh oh!

marbre commented Mar 18, 2026

Uh oh!

ronlieb commented Apr 1, 2026

Uh oh!

stellaraccident commented Apr 1, 2026

Uh oh!

marbre commented Apr 1, 2026

Uh oh!

stellaraccident commented May 2, 2026

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

ronlieb May 11, 2026

Choose a reason for hiding this comment

Uh oh!

lumachad commented May 11, 2026

Uh oh!

Meinersbur May 13, 2026

Choose a reason for hiding this comment

Uh oh!

This branch has not been deployed

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

6 participants

stellaraccident commented Mar 12, 2026 •

edited

Loading