GitLab

b6783cd5 doc(README): fix URL + repository name <> Thu May 15 09:03:54 2025 -0700 <> Philipp Samfass <philipp.samfass@amd.com>

717112f7 docs/index.rst: fix trademark symbol <> Mon May 5 15:12:16 2025 +0000 <> domcharrier <docharri@amd.com>

dc177cbc README.md: remove GitHub org move note <> Mon May 5 10:55:07 2025 +0000 <> domcharrier <docharri@amd.com>

20af9489 docs/install/INSTALL.md: add PyPI install instructions <> Fri May 2 18:07:37 2025 +0000 <> domcharrier <docharri@amd.com>

d1b3a22b doc changes: <> Mon Apr 28 16:12:46 2025 -0400 <> Alex Xu <alex.xu@amd.com>
add label settings

add rocm_docs to extension

replace deprecated doxygen setup with new variables

comment out unused extensions

remove unused extensions

change version to 25.05

revert version change

30e5aaf7 versions.json: use rocmds-cmake defaults <> Thu Apr 24 19:56:03 2025 +0000 <> domcharrier <docharri@amd.com>

d897d85d rapids_config.*: configurable ROCm-DS.cmake repo <> Thu Apr 24 19:31:47 2025 +0000 <> domcharrier <docharri@amd.com>
Make the repository from which to obtain the ROCm-DS.cmake
script configurable via the environment variable
`ROCm-DS_CMAKE_SCRIPT_REPO`.

f60fe68e fix(conda/*/all_rocm*.yaml): update versions <> Thu Apr 24 16:50:41 2025 +0000 <> domcharrier <docharri@amd.com>

9f732a56 fix(docs/install/INSTALL.md): use scikit-buil-core <> Thu Apr 24 13:37:59 2025 +0000 <> domcharrier <docharri@amd.com>

a02e313d docs: bump version {1=>2}.0.0b1 <> Thu Apr 24 13:17:32 2025 +0000 <> docharri@amd.com <domcharrier>

bf46c436 Update install with comments from Dominic re PR #34 <> Fri Apr 11 15:23:32 2025 -0700 <> randyh62 <randy.hartgrove@amd.com>
edit metadata content, edit index

fix spacing in metadata

add TradeMark to ROCm-DS

edit for Leo's comments

update version

78489ba1 doc(INSTALL.md): revise C++-only instructions for librmm, remove conda requirement <> Wed Apr 9 01:11:47 2025 -0700 <> Philipp Samfass <philipp.samfass@amd.com>

2f6497cb .readthedocs.yaml: add file <> Wed Mar 19 15:30:36 2025 -0400 <> Alex Xu <alex.xu@amd.com>

9542cee3 docs/*/requirements*: bump rocm-docs-core, deps <> Thu Apr 3 21:44:49 2025 -0400 <> Alex Xu <alex.xu@amd.com>

ed6605d5 docs/*/INSTALL.md: fix indent, whitespace errors <> Mon Apr 7 10:09:26 2025 +0000 <> domcharrier <docharri@amd.com>

a775adf7 docs/*/INSTALL.md: update install instructions <> Wed Mar 26 15:13:04 2025 -0700 <> randyh62 <randy.hartgrove@amd.com>

6a98eca4 docs/*.{rst,txt,md,..}: kick off docs/ subdir <> Tue Mar 4 15:45:10 2025 -0800 <> randyh62 <randy.hartgrove@amd.com>

6b48ddb3 fix(CMakeLists.txt): init archs before finding HIP <> Tue Apr 22 15:40:13 2025 +0000 <> domcharrier <docharri@amd.com>
In contrast to the CUDA implementation, we need to initialize the HIP architectures
already before calling rapids_find_package(HIP...). Otherwise, GPU_TARGETS will be unset
on systems without GPUs and find_package(HIP) will warn that it
cannot autotect an architecture.

0e241dbb fix(conda/*/all_rocm*.yaml): install cmake<4 <> Thu Apr 24 10:02:56 2025 +0000 <> domcharrier <docharri@amd.com>

9b5b4997 rapids_config.*: ROCm-DS_CMAKE{_FROM=>_}MODULE_PATH <> Thu Apr 3 06:01:55 2025 -0700 <> Philipp Samfass <philipp.samfass@amd.com>
Remove the ROCm-DS_CMAKE_FROM_MODULE_PATH option in favor
of a new cached filepath varaible ROCm-DS_CMAKE_MODULE_PATH.

This allows to specify the path to the rapids-cmake modules
via an external environment variable `ROCm-DS_CMAKE_MODULE_PATH`
instead of passing it directly via a cmake command.

Squash commit constructed from the following 4 individual commits:

* feat: allow to specify CMAKE_MODULE_PATH as ROCm-DS_CMAKE_MODULE_PATH
environment variable
* tmpfix: make sure that rapids-cmake-src directory is set when using
local rocm-ds-cmake
* feat(fetch_rapids.cmake): deprecate option ROCm-DS_CMAKE_FROM_MODULE_PATH
* fix: make ROCm-DS_CMAKE_MODULE_PATH a variable

8c928daf feat(CMake): Add basic support for packaging hipMM <> Wed Apr 2 03:52:15 2025 -0700 <> Philipp Samfass <philipp.samfass@amd.com>
Details:

This commit allows to build a hipmm package
of the C++/HIP source code files.
Presently, in some environments, the packaging
will also include fmt and spdlog into the package. This might have to be
revisited later.

77ecae9b fix(rapids_conf*):make ROCmDS-CMake 1.0.0 default <> Tue Mar 18 17:03:25 2025 +0000 <> domcharrier <docharri@amd.com>
Set the ROCm-DS_CMAKE_SCRIPT_BRANCH default to
`branch-1.0.0*`.

d331bb5e feat: +option to find ROCmDS-CMake on module path <> Tue Mar 18 16:53:07 2025 +0000 <> domcharrier <docharri@amd.com>
Adds an option ROCm-DS_CMAKE_FROM_MODULE_PATH [default=OFF]
that disables fetching of ROCm-DS CMake script and repo
assuming that ROCm-DS CMake modules can be found via the
CMAKE_MODULE_PATH.

a0d7b777 fix(rapids_config.cmake): fix regex to accept semantic VERSION <> Wed Mar 12 10:19:22 2025 -0700 <> Philipp Samfass <philipp.samfass@amd.com>

a9e2f1ba add beta label to VERSION <> Wed Mar 12 10:17:33 2025 -0700 <> Philipp Samfass <philipp.samfass@amd.com>

4100cf85 refactor(python/librmm):add some metadata for amd-hipmm package <> Wed Mar 12 06:23:19 2025 -0700 <> Philipp Samfass <philipp.samfass@amd.com>

6dcfdf6e change librmm Python package name to amd-libhipmm <> Wed Mar 12 06:22:13 2025 -0700 <> Philipp Samfass <philipp.samfass@amd.com>

151c3f91 fix(allocators/numba.py): use numba.hip <> Thu Mar 13 16:33:56 2025 +0000 <> domcharrier <docharri@amd.com>

5023d6f4 fix(stream.pyx): use numba.hip <> Thu Mar 13 16:30:31 2025 +0000 <> domcharrier <docharri@amd.com>

0b063ce2 bump hip-python dependencies to >=6.4 <> Thu Feb 27 01:44:58 2025 -0800 <> Philipp Samfass <philipp.samfass@amd.com>

30ef89b2 doc(pyproject.toml): modify authors <> Thu Feb 27 01:41:37 2025 -0800 <> Philipp Samfass <philipp.samfass@amd.com>

55788211 fix: do not require git for installing numba-hip <> Fri Mar 14 08:56:18 2025 -0700 <> Philipp Samfass <philipp.samfass@amd.com>

c2637a3e fix: do not require git for installing numba-hip <> Wed Feb 26 06:16:16 2025 -0800 <> Philipp Samfass <philipp.samfass@amd.com>

8fa4be2b doc: revise LICENSE notices in pyproject.toml <> Wed Feb 26 03:47:13 2025 -0800 <> Philipp Samfass <philipp.samfass@amd.com>

07dc876f doc(README.md): add compatible ROCm-DS RMM version <> Mon Feb 24 08:46:34 2025 -0800 <> Philipp Samfass <philipp.samfass@amd.com>

81af0567 rename python project to amd-hipmm <> Fri Feb 21 07:45:24 2025 -0800 <> Philipp Samfass <philipp.samfass@amd.com>

a2163037 refactor: change version to 2.0.0 <> Fri Feb 21 07:41:41 2025 -0800 <> Philipp Samfass <philipp.samfass@amd.com>

0bb038d0 tmpfix(README.md): comment 3rd party section <> Fri Feb 14 16:20:04 2025 +0000 <> domcharrier <docharri@amd.com>

d419ee3e README.md): RMM* -> hipMM <> Fri Feb 14 16:18:53 2025 +0000 <> domcharrier <docharri@amd.com>

b6c6c3af README.md: clarify python part is wip <> Thu Feb 13 14:10:31 2025 +0000 <> domcharrier <docharri@amd.com>

b166cce5 hipify(test_rmm.py): + import numba hip as cuda <> Thu Jan 16 14:34:14 2025 +0000 <> domcharrier <docharri@amd.com>

e368c5eb tmpfix(CMakeLists.txt): - nvtx3 incl + link dep <> Thu Jan 16 11:12:45 2025 +0000 <> domcharrier <docharri@amd.com>
NOTE: This is marked as tmpfix as we might replace nvtx3
      by something equivalent later on.

cde18f4b tmpfix(CMakeList.txt):handler for unsupported opts <> Wed Jan 15 13:29:15 2025 +0000 <> domcharrier <docharri@amd.com>
* Handle user specifying USE_NVTX=ON.
* Handle user specifying CUDA_STATIC_RUNTIME=ON.

NOTE: This is marked as tmpfix as we might reenable these
      options later on again.

9dafc6f4 pyproject.toml: + rocm-X-Y-Z opt deps keys <> Thu Dec 12 13:51:33 2024 +0000 <> domcharrier <docharri@amd.com>
Add additional optional dependency keys.

ccebf458 pyproject.toml: rename project <> Wed Dec 11 16:20:22 2024 +0000 <> domcharrier <docharri@amd.com>

b2a1994c hipify(print_env.sh): use rocm tools <> Wed Dec 11 16:13:57 2024 +0000 <> domcharrier <docharri@amd.com>

ae07d677 refactor(._upstream): mv low prio code here <> Wed Dec 11 16:13:34 2024 +0000 <> domcharrier <docharri@amd.com>

51a64cfd README.md: clearer hipMM vs RMM, info boxes <> Tue Dec 10 10:39:21 2024 +0100 <> Philipp Samfass <philippsamfass@googlemail.com>
* Distinguish better between hipMM and RMM.
* Use github-markdown info boxes instead of
  other highlighting mechanisms.

775a1dfa tmpfix(**/nvtx/ranges): disable nvtx ranges <> Tue Dec 3 21:56:35 2024 +0000 <> domcharrier <docharri@amd.com>

7a831d03 fix: include rmm/cuda_runtime_api.h <> Thu Nov 28 19:09:55 2024 +0000 <> domcharrier <docharri@amd.com>

7285ef61 fix(rmm/cuda_runt*api.h): + 2 CUDA->HIP maps <> Thu Nov 28 19:09:20 2024 +0000 <> domcharrier <docharri@amd.com>

a8d7ddaa clean(build.sh): pip already builds+installs wheel <> Thu Nov 28 19:08:07 2024 +0000 <> domcharrier <docharri@amd.com>

f16566c0 fix(all_rocm*yaml): use scikit-build-core <> Wed Apr 17 08:55:32 2024 +0000 <> domcharrier <docharri@amd.com>

564d8cf6 feat(rapids_config.cmake): environ var controls <> Tue Apr 16 16:19:30 2024 +0000 <> domcharrier <docharri@amd.com>
Makes ROCmDS-CMake script branch as well as ROCmDS-CMake
repo URL and branch configurable via environment variables.

---

This commit is the result of squashing the following 5 original commits:

1. {fetch_rapids->rapids_config}.cm*: +ENV for branch

  * Introduce environent variable ROCm-DS_CMAKE_BRANCH for
    specifying the ROCm-DS_CMAKE_BRANCH, defaults to `hipdf-dev`.

2. fix(rapids_config.cmake): check download status

  * Check download status and abort with error if download was not
    successful. Also delete `${CMAKE_CURRENT_BINARY_DIR}/RMM_ROCm-DS.cmake`
    in this case.

3. fix(rapids_config.cmake): dl + check errors once

4. Set default rapids-cmake branch to include missing libhipcxx headers by default.

5. fetch_rapids.cmake: redef ENV{ROCm-DS_CMAKE_BRANCH}

  * Introduces new variable ROCm-DS_CMAKE_SCRIPT_BRANCH (env + cmake var) for specifying
    where to fetch the ROCm-DS.cmake script from.
  * Repurpose ROCm-DS_CMAKE_BRANCH environment variable for specfiying what
    ROCm/rapids-cmake branch shall be checked out by the ROCm-DS.cmake
    script. This var overrides CMake var `rapids-cmake-branch`.

67bd66ca include/rmm/cuda_device.hpp: fix incomplete license <> Wed Nov 6 12:17:12 2024 +0000 <> domcharrier <docharri@amd.com>

ab49218c fix(__init__.py): fix problems arising with map... <> Tue Nov 5 14:06:20 2024 +0000 <> Philipp Samfass <philipp.samfass@amd.com>
device_id_to_resource not being globally unique across all
translation units when RMM is built with hipcc/clang as host compiler.

f84a05b3 feat(dynamic*runtime*): enable rocm>=6.0.2 feature <> Wed Sep 18 11:52:20 2024 +0000 <> domcharrier <docharri@amd.com>
Device attribute 'hipDeviceAttributeMemoryPoolSupportedHandleTypes'
available with ROCm>=6.0.2.

0fd61717 fix(**cuda_run*_api.h):hipDev{Attr->iceAttribute} <> Wed Sep 18 11:49:27 2024 +0000 <> domcharrier <docharri@amd.com>
Fix enum constant name.

b31217f4 feat: + versions.json <> Tue Sep 17 14:15:00 2024 +0000 <> domcharrier <docharri@amd.com>
* Allows to fix third-party dependencies to branch-X.Y compatible
  versions.
* Allows to control dependencies from CI process.

NOTE ("first to record wins"):

  "If a project is listed in multiple override files, the first file values will be used,
  and all later calls for that packaged will be ignored.  This "first to record, wins"
  approach is used to match FetchContent, and allows parent projects to override child
  projects."

  Source: Docs of `rapids_cpm_package_override` in `rapids-cmake/cpm/package_override.cmake`.

09176a74 **cuda_runtime_api.h: only macros, use ifndef <> Mon Aug 26 09:24:16 2024 +0000 <> domcharrier <docharri@amd.com>
We only use macros and use `#ifndef` to prevent
redefinition warnings if this project is used
by another project with a similar header file.

03dc3d83 feat(py*/pyproject*): + rocm-6-2-0 opt deps key <> Wed Aug 14 12:17:56 2024 +0000 <> domcharrier <docharri@amd.com>

c3d5b264 README.md: Make user aware of rocm-* keys <> Fri Aug 9 13:39:53 2024 +0000 <> domcharrier <docharri@amd.com>
Remove ROCm requirements file that installed RMM dependencies.
Instead, explain in README.md how to install dependencies
correctly for a given ROCm installation.

2224740a pyproject.toml: configurable rocm version of deps <> Fri Aug 9 13:33:23 2024 +0000 <> domcharrier <docharri@amd.com>
Dependencies compatible with particular ROCm versions
can now be selected via optional dependency keys
`rocm-<major>-<minor>-<patch>` (example: `rocm-6-1-2`).

When you specify an optional dependency key `rocm_12_12_12` (note the underscores),
pip (tested with v24.2) might report a warning where the key's
underscores have all been replaced by hyphens:

```text
WARNING: numba-hip 0.1 does not provide the extra 'rocm-12-12-12'
```

This pip behavior has a high potential to confuse users, so we try to prevent
such issues from the start on by using hyphens for all optional dependency keys.

64c98236 fix(cuda_runtime_api.h): constexpr -> #define to prevent redef errors <> Mon Jun 3 16:19:41 2024 +0200 <> domcharrier <docharri@amd.com>
Use #define for cudaStreamLegacy instead of constexpr statement to avoid redefinition errors if this file is used (indirectly) by a dependency that makes the same definition in its reverse header file.

dc1fa00a dynamic_load_runtime: add todo <> Wed Aug 7 09:13:19 2024 +0000 <> domcharrier <docharri@amd.com>

3e4c4ac2 cuda_async_mem*_res*: add hip version check <> Wed Aug 7 09:05:50 2024 +0000 <> domcharrier <docharri@amd.com>
Remark: From the standing of the build system and the majority
of the ode can only be compiled for AMD GPUs. So some
of the checks need to be revisited simplified.

3c4abc34 fix(tests,benchmarks): + old+new HIP PTDS compile def <> Wed Aug 7 08:55:20 2024 +0000 <> domcharrier <docharri@amd.com>

9a43a6bf build(conda/**/all_rocm*): readd gcc_linux-64 <> Thu Dec 14 12:42:31 2023 +0000 <> domcharrier <docharri@amd.com>
This package installs activation and deactivation
scripts into the conda environment:

* ${CONDA_PREFIX}/etc/conda/activate.d/activate-gcc_linux-64.sh
* ${CONDA_PREFIX}/etc/conda/deactivate.d/deactivate-gcc_linux-64.sh

Among other things, the activation script set the current conda
environment's lib folder as RPATH for compiled binaries.

6bff2a65 REMOVE: build.sh: 'build.sh rmm' creates binary wheel now <> Wed Nov 29 12:52:29 2023 +0000 <> domcharrier <docharri@amd.com>
The script creates a binary Python package wheel if
called like `build.sh rmm` now.

This is a change to the upstream behavior where
the package is installed directly into the current Python
environment.

959aeeaf docs(README): revise two notes <> Wed Nov 29 12:30:51 2023 +0000 <> domcharrier <docharri@amd.com>

9e4495e2 docs(all_rocm*.yaml): more detailed comments <> Wed Nov 29 12:30:20 2023 +0000 <> domcharrier <docharri@amd.com>

fde0c2d6 hotfix(dynamic_load_runtime.hpp): back to old behavior <> Thu Nov 23 17:57:25 2023 +0000 <> domcharrier <docharri@amd.com>
* Reverting to behavior similar to CUDA.
* Presence of hipMemPool* API in runtime does not imply support of this
  feature by the respective GPU. Apparently some specific HW support is
  required (which is not present on MI200 according to hipGetAttributes).

f6e85361 fix (all_rocm*yaml): readd sysroot_linux-64==2.17 <> Thu Nov 23 12:18:40 2023 +0000 <> domcharrier <docharri@amd.com>
Otherwise, linker errors caused by libfmts.o are observed.

c1cfac95 + modifications licenses <> Thu Nov 23 09:57:54 2023 +0000 <> domcharrier <docharri@amd.com>

04b2008d docs(README): revise <> Thu Nov 23 09:14:58 2023 +0000 <> domcharrier <docharri@amd.com>

d27ab904 fix (python/rmm/_cuda/gpu.py): use cudart.cudaRuntimeGetVersion() <> Thu Nov 23 08:59:40 2023 +0000 <> domcharrier <docharri@amd.com>
It is available in `hip-python` and `hip-python-as-cuda`.

11df9c9c docs(README.md): add section on py wheel creation <> Wed Nov 22 21:12:11 2023 +0000 <> domcharrier <docharri@amd.com>

daa174b1 fix(python/pyproject.toml): rename project <> Wed Nov 22 21:11:09 2023 +0000 <> domcharrier <docharri@amd.com>
`rmm-rocm` for now.

0a038a06 fix (README.md): clarify installation steps <> Wed Nov 22 19:40:04 2023 +0000 <> domcharrier <docharri@amd.com>

37e86031 fix: enable benchmarks <> Wed Nov 22 19:39:42 2023 +0000 <> domcharrier <docharri@amd.com>

74deeff8 fix (tests/CMakeLists.txt):set __HIP_API_PER_THREAD_DEFAULT_STREAM__ <> Wed Nov 22 19:34:53 2023 +0000 <> domcharrier <docharri@amd.com>

dd1717ac fix (tests/CMakeLists.txt): coverage analysis also for Clang CXX compiler <> Wed Nov 22 19:31:52 2023 +0000 <> domcharrier <docharri@amd.com>

1c350885 fix (tests/CMakeLists.txt): Set HIP C++ standard to 17 <> Wed Nov 22 19:30:30 2023 +0000 <> domcharrier <docharri@amd.com>

f0e3356b fix(cuda_async_view_memory_resource.hpp): exclude unnecessary code for HIP/AMD build <> Wed Nov 22 18:51:10 2023 +0000 <> domcharrier <docharri@amd.com>
hip(Malloc|Free)Async and hipMemPool* APIs have been introduced with
ROCm HIP SDK 5.1. Therefore the presence of the former APIs implies
the presence of the hipMemPool* APIs.

acc52f86 fix (random_allocations.cpp): if -> else if <> Wed Nov 22 18:50:06 2023 +0000 <> domcharrier <docharri@amd.com>

cfb28ec1 fix (benchmarks/CMakeLists.txt): compile synchronization.cpp with HIP too <> Wed Nov 22 18:49:17 2023 +0000 <> domcharrier <docharri@amd.com>

db792596 fix(dynamic_load_runtime.hpp): async memory + function name fix <> Wed Nov 22 17:48:24 2023 +0000 <> domcharrier <docharri@amd.com>

4bc0dd27 fix (rmm/cuda_runtime_api.h): support benchmarks + tests <> Wed Nov 22 17:42:11 2023 +0000 <> domcharrier <docharri@amd.com>

17f31890 fix (exec_policy.hpp): thrust::{cuda,cuda_cub}->thrust::{hip,hip_rocprim} <> Wed Nov 22 17:36:28 2023 +0000 <> domcharrier <docharri@amd.com>

2bb06789 fix (mr_multithreadedtests.cpp): no 3rd arg default value with hipStreamWaitEvent <> Wed Nov 22 17:28:55 2023 +0000 <> domcharrier <docharri@amd.com>

dd736fdf hipify (README.md) <> Fri Nov 10 21:27:22 2023 +0000 <> domcharrier <docharri@amd.com>

2791b044 hipify (CMakeLists.txt) <> Fri Nov 10 21:10:21 2023 +0000 <> domcharrier <docharri@amd.com>
* Change to CUDA to HIP dependencies.
* Disable CUDA static library built option for the time being.

334e0793 hipify (pyproject.toml): switch to hip-python dependencies, change metadata <> Fri Nov 10 21:08:11 2023 +0000 <> domcharrier <docharri@amd.com>

37e0b5d2 hipify(get_cccl.cmake): get rocthrust, libhipcxx <> Fri Nov 10 21:05:41 2023 +0000 <> domcharrier <docharri@amd.com>
There is no CCCL equivalent in the AMD ecosystem.
Therefore, we get rocthrust and libhipcxx individually
via CPM.

aff5f7e4 hipify (benchmarks,tests): tag sources as hip files <> Fri Nov 10 21:04:41 2023 +0000 <> domcharrier <docharri@amd.com>

2470dabb fix (stack_trace.hpp): use n::free -> sinstfree <> Fri Nov 10 20:58:12 2023 +0000 <> domcharrier <docharri@amd.com>

aa1edfe4 fix (error.hpp): void-cast cuda/hipGetLastError() <> Fri Nov 10 20:56:12 2023 +0000 <> domcharrier <docharri@amd.com>
Prevents warnings about ignored no discard attribute.

26ee1010 hipify (dynamic_load_runtime.hpp): dlopen libamdhip64, limited HIP capabilities <> Fri Nov 10 20:53:52 2023 +0000 <> domcharrier <docharri@amd.com>
* Look for libamdhip64 and dlopen it.
* Limited HIP cababilities prevent that we can use all
  features for AMD GPUs.

d56de4d1 refactor (device_buffer.pyx): slight differences between cuda-python and hip-python-as-cuda <> Fri Nov 10 20:49:40 2023 +0000 <> domcharrier <docharri@amd.com>
hip-python-as-cuda's extension modules are compiled with a C compiler.

aabb0bab refactor(benchmarks,test,include): employ inverse hipify header <> Fri Nov 10 20:40:24 2023 +0000 <> domcharrier <docharri@amd.com>
Co-authored-by: domcharrier <docharri@amd.com>
Co-authored-by: psamfass <psamfass@amd.com>
Co-authored-by: monoatamd <mnorouzi@amd.com>
Co-authored-by: reger-men <adil.lashab@amd.com>
Co-authored-by: dkbhaskaran <dineshkumarb@gmail.com>
Co-authored-by: obersteiner <moberste@amd.com>
Co-authored-by: alexxu-amd <alex.xu@amd.com>
Co-authored-by: randyh62 <randy.hartgrove@amd.com>