CPU and GPU identification data by Adrien-Tab · Pull Request #6 · CExA-project/CExA-Experimental

Adrien-Tab · 2026-03-04T13:53:23Z

No description provided.

Signed-off-by: Adrien Taberner <adrien.taberner@cea.fr>

Adrien-Tab · 2026-03-06T08:27:56Z

https://developer.nvidia.com/management-library-nvml
https://rocm.docs.amd.com/projects/amdsmi/en/latest/doxygen/docBin/html/topics.html

- add a CEXA_ARCHINFO_ENABLE_TESTS option - try to find gtest before calling FetchContent - link to gtest instead of gtest_main - Use the PRE_TEST test discovery mode

- Use macros to choose between the platform specific functions instead of doing it at configure time

- Remove the compilInfo.hpp

- Properly install the library - Do not install gtest alongside the library - Add an install test subdirectory

- Only ubuntu and macos for the moment

tretre91

I added all the missing functionality, namely:

windows support
macos support
sycl support
openacc support
getting the cuda driver version

I also added a basic CI which runs the tests, installs the library and tries to build a cmake project using it

tretre91 · 2026-03-11T23:25:28Z

I kept the tests as is (just removed the prints in the main function), we might want to do something more specific than checking if the string is non-empty. For example checking that if there is no GPU backend, the GPU functions return "N/A".

tretre91 · 2026-03-11T23:34:02Z

+std::string get_gpu_driver_version() {
+  if (NVML_SUCCESS != nvmlInit_v2()) {
+    std::cerr << "get_gpu_driver_version: failed to initialize nvml\n";
+    return "ERROR";
+  }
+  char buffer[NVML_SYSTEM_NVML_VERSION_BUFFER_SIZE];
+  nvmlSystemGetDriverVersion(buffer, NVML_SYSTEM_NVML_VERSION_BUFFER_SIZE);
+  nvmlShutdown();
+  return buffer;
+}


According to the documentation, nvml can be initialized/shutdown multiple times (they do a noop if the library was already initialized) so there shouldn't be any conflicts with a user code using nvml and our library.

For the error checking, nvmlSystemGetDriverVersion and nvmlShutdown cannot fail in our use case (nvml is initialized, buffer is not NULL, buffer size is sufficient), that's why I only check for errors in the call to nvmlInit_v2.
Also, I don't know if we should keep the std::cerr with the error message, as it is the only place in the code where we actually print an error message to the console.

I am in favor of removing the cerr and simply returning "N/A".

I would still check the return value of nvmlSystemGetDriverVersion. I trust you that with current specification we do not expect it to fail but:

we can't guarantee that new condition for failure won't appear in the future

we have no guarantee that the implementation contains no bug and doesn't throw error in other cases too

Co-authored-by: Paul Zehner <paul.zehner@cea.fr>

PaulGannay · 2026-03-16T08:47:11Z

+std::string get_gpu_driver_version() {
+  if (NVML_SUCCESS != nvmlInit_v2()) {
+    std::cerr << "get_gpu_driver_version: failed to initialize nvml\n";
+    return "ERROR";
+  }
+  char buffer[NVML_SYSTEM_NVML_VERSION_BUFFER_SIZE];
+  nvmlSystemGetDriverVersion(buffer, NVML_SYSTEM_NVML_VERSION_BUFFER_SIZE);
+  nvmlShutdown();
+  return buffer;
+}


I would still check the return value of nvmlSystemGetDriverVersion. I trust you that with current specification we do not expect it to fail but:

we can't guarantee that new condition for failure won't appear in the future

we have no guarantee that the implementation contains no bug and doesn't throw error in other cases too

PaulGannay · 2026-03-16T08:50:48Z

+std::string get_gpu_name() {
+  hipDeviceProp_t device_prop;
+  int device_id = Kokkos::device_id();
+  KOKKOS_IMPL_HIP_SAFE_CALL(hipGetDeviceProperties(&device_prop, device_id));
+  return std::string{device_prop.name};
+}
+
+std::string get_gpu_arch() {
+  hipDeviceProp_t device_prop;
+  int device_id = Kokkos::device_id();
+  KOKKOS_IMPL_HIP_SAFE_CALL(hipGetDeviceProperties(&device_prop, device_id));
+  return std::string{device_prop.gcnArchName};
+}


I would consider using a macro here to avoid code duplication (this is only a suggestion).

I think it's fine since it's only used twice

pzehner

Some remarks about the CMake files.

Co-authored-by: PaulGannay <paul.gannay@cea.fr>

Co-authored-by: Paul Zehner <paul.zehner@alumni.enseeiht.fr>

Signed-off-by: Adrien Taberner <adrien.taberner@cea.fr>

tretre91 · 2026-04-16T10:45:07Z

+std::string get_gpu_name() {
+  hipDeviceProp_t device_prop;
+  int device_id = Kokkos::device_id();
+  KOKKOS_IMPL_HIP_SAFE_CALL(hipGetDeviceProperties(&device_prop, device_id));
+  return std::string{device_prop.name};
+}
+
+std::string get_gpu_arch() {
+  hipDeviceProp_t device_prop;
+  int device_id = Kokkos::device_id();
+  KOKKOS_IMPL_HIP_SAFE_CALL(hipGetDeviceProperties(&device_prop, device_id));
+  return std::string{device_prop.gcnArchName};
+}


I think it's fine since it's only used twice

Signed-off-by: Adrien Taberner <adrien.taberner@cea.fr>

Initial version for CPU + OS for Unix

c589581

Signed-off-by: Adrien Taberner <adrien.taberner@cea.fr>

Adrien-Tab changed the title ~~Initial version for CPU + OS for Unix~~ CPU and GPU identification data Mar 4, 2026

Adrien-Tab self-assigned this Mar 4, 2026

PaulGannay requested changes Mar 5, 2026

View reviewed changes

Comment thread archInfo/src/cexa_ArchInfo.cpp Outdated

Comment thread archInfo/src/cexa_ArchInfo.cpp Outdated

Comment thread archInfo/unit_test/TestArchInfo.cpp Outdated

Comment thread archInfo/src/cexa_ArchInfo.hpp Outdated

Comment thread archInfo/src/cexa_ArchInfo.hpp Outdated

Hip Gpu info + fallback values from cmake

f574ac3

Signed-off-by: Adrien Taberner <adrien.taberner@cea.fr>

Adrien-Tab marked this pull request as draft March 5, 2026 14:45

CUDA GPU

fb95cf3

Signed-off-by: Adrien Taberner <adrien.taberner@cea.fr>

tretre91 added 17 commits March 10, 2026 10:40

Add a cmake option to enable the tests

42f9b15

- add a CEXA_ARCHINFO_ENABLE_TESTS option - try to find gtest before calling FetchContent - link to gtest instead of gtest_main - Use the PRE_TEST test discovery mode

Dispatch to the platform specific impls at compile time

0ae4c5e

- Use macros to choose between the platform specific functions instead of doing it at configure time

Add windows specific functions

1972d61

Use nvml to get the driver version

26d2581

Implement sycl functions

4bd1165

Make the platform specific functions non-inline

45c6682

Add macos specific functions

78797ab

- Remove the compilInfo.hpp

Small changes in the gpu functions

144620a

Add a proper installation mechanism

2c847f0

- Properly install the library - Do not install gtest alongside the library - Add an install test subdirectory

Add a CI

d95722a

- Only ubuntu and macos for the moment

ci: install kokkos outside of the github workspace

0dabf4f

ci: checkout the code before installing kokkos

76ad351

Remove the call to the print functions in the tests

38a8773

Return "ERROR" if the file open failed for macos get_sys_name

34568b0

Rename the functions

3bb1ba0

Update the README

4ceaeab

Remove unnecessary backslashes in the README

00233f5

tretre91 reviewed Mar 11, 2026

View reviewed changes

tretre91 self-assigned this Mar 11, 2026

tretre91 marked this pull request as ready for review March 11, 2026 23:35

tretre91 marked this pull request as draft March 12, 2026 10:57

tretre91 reviewed Mar 12, 2026

View reviewed changes

Comment thread archInfo/src/cexa_unixArchInfo.hpp Outdated

tretre91 and others added 9 commits March 13, 2026 16:26

ci: Simplify the declaration of paths triggering the ci

760e4f5

Add comments as advised in code review

87cd6d2

Co-authored-by: Paul Zehner <paul.zehner@cea.fr>

Use std::size_t instead of size_t

0458c15

Co-authored-by: Paul Zehner <paul.zehner@cea.fr>

Simplify some if conditions

af8fac8

Co-authored-by: Paul Zehner <paul.zehner@cea.fr>

Use a more modern CMake installation mechanism

25b3a1e

Co-authored-by: Paul Zehner <paul.zehner@cea.fr>

Add check_required_components to the cmake Config file

3945223

Add OpenACC support

2b8bf69

Rename the include guard for cexa_ArchInfo.hpp

e12245f

Read the cpu name from the registry on windows

940d42b

Adrien-Tab requested a review from PaulGannay March 13, 2026 22:27

tretre91 and others added 4 commits March 13, 2026 23:31

Add a windows arm ci

8926f0c

ci: set the extra flags for the windows arm runner

b708cde

Remove a useless target_include_directories

56c7f86

Add the copyright notice

31279ef

PaulGannay requested changes Mar 16, 2026

View reviewed changes

pzehner reviewed Mar 16, 2026

View reviewed changes

Comment thread archInfo/CMakeLists.txt Outdated

Comment thread archInfo/cmake/CexaArchInfoConfig.cmake.in

Adrien-Tab and others added 5 commits April 2, 2026 11:08

Update archInfo/src/cexa_ArchInfo.cpp

159c649

Co-authored-by: PaulGannay <paul.gannay@cea.fr>

Update archInfo/CMakeLists.txt

11c08ab

Co-authored-by: Paul Zehner <paul.zehner@alumni.enseeiht.fr>

CPU Name detection on GH200, rename 'Error' to 'Unknown'

b53a630

Signed-off-by: Adrien Taberner <adrien.taberner@cea.fr>

RAII: FILE*

85b3104

Signed-off-by: Adrien Taberner <adrien.taberner@cea.fr>

Use ubuntu 24

164a576

Signed-off-by: Adrien Taberner <adrien.taberner@cea.fr>

Adrien-Tab force-pushed the cpu-gpu-arch branch from dc675c0 to 164a576 Compare April 2, 2026 14:25

tretre91 self-requested a review April 15, 2026 09:10

tretre91 added 2 commits April 16, 2026 12:40

Use source files instead of headers for the cpu info functions

f20af8a

Correctly call find_dependency

6c823ee

tretre91 reviewed Apr 16, 2026

View reviewed changes

Adrien-Tab added 3 commits May 5, 2026 14:21

move functions from cexa::impl to cexa::

ac4548d

Signed-off-by: Adrien Taberner <adrien.taberner@cea.fr>

Fix Tests

3d3db37

Signed-off-by: Adrien Taberner <adrien.taberner@cea.fr>

Add check for nvml call

9c76fd6

Signed-off-by: Adrien Taberner <adrien.taberner@cea.fr>

Adrien-Tab merged commit ab346c3 into main May 11, 2026
6 checks passed

Conversation

Adrien-Tab commented Mar 4, 2026

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Adrien-Tab commented Mar 6, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

tretre91 left a comment • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

pzehner left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

Adrien-Tab commented Mar 6, 2026 •

edited

Loading

tretre91 left a comment •

edited

Loading