[OpenMP] Change CMake Configuration to Build for Highest CUDA Architecture by Default

[jhuber6]

Summary

This patch changes the CMake files for Clang and Libomptarget to query the system for its supported CUDA architecture. This will simplify the experience of building LLVM with OpenMP Offloading support by removing the need to manually specify the most optimal architecture for each system. Libomptarget will also build support for sm_35 as a fallback. This uses the find_cuda methods from CMake to detect the architecture which is deprecated in Cmake 3.18.

[ye-luo]

FindCUDA has been deprecated.
Please explore the following feature without directly calling FindCUDA.
https://gitlab.kitware.com/cmake/cmake/-/merge_requests/1856

[jhuber6]

In D88929#2315378, @ye-luo wrote:

FindCUDA has been deprecated.
Please explore the following feature without directly calling FindCUDA.
https://gitlab.kitware.com/cmake/cmake/-/merge_requests/1856

Finding architectures using the CUDA language support requires CMake 3.18 as far as I know. LLVM's minimum CMake requirement is 3.13.4 so this method is probably the best I could see. Libomptarget already uses FindCUDA inside openmp/libomptarget/cmake/Modules/LibomptargetGetDependencies.cmake so I figured it was okay. Alternatively, we could do the detection manually. All this amount to is compiling and running some CUDA code that prints the major and minor version. We could do that manually and try to use the original language support if you think that's a better option, I'm not a CMake guru or anything.

[ye-luo]

The link I posted indicated that independent feature is merged since 3.12. Better to avoid deprecated stuff when introducing new cmake lines even though some existing lines may still rely on deprecated cmake.

[ye-luo]

3.18 introduces CMAKE_CUDA_ARCHITECTURES. Does 3.18 supports detection? If we know a new way works since 3.18, I think putting both with if-else makes sense.

Regarding doing it manually. I think it is more risky than using existing schemes shipped with cmake.

[jhuber6]

In D88929#2315402, @ye-luo wrote:

The link I posted indicated that independent feature is merged since 3.12. Better to avoid deprecated stuff when introducing new cmake lines even though some existing lines may still rely on deprecated cmake.

This requires adding CUDA as a language, is that alright inside of Clang? Nothing is using CUDA, we're just checking the architecture.

In D88929#2315404, @ye-luo wrote:

3.18 introduces CMAKE_CUDA_ARCHITECTURES. Does 3.18 supports detection? If we know a new way works since 3.18, I think putting both with if-else makes sense.

Regarding doing it manually. I think it is more risky than using existing schemes shipped with cmake.

I'm not sure how that would affect the other Libomptarget stuff that uses FindCUDA since language enabling is done at the highest level, does that conflict?

[ye-luo]

I just realized that this patch affects clang and libomptarget.
I cannot comment on clang. Regarding libomptarget, Could you explain why the detection is not put together with other cuda stuff in openmp/libomptarget/cmake/Modules/LibomptargetGetDependencies.cmake

[jhuber6]

In D88929#2315451, @ye-luo wrote:

I just realized that this patch affects clang and libomptarget.
I cannot comment on clang. Regarding libomptarget, Could you explain why the detection is not put together with other cuda stuff in openmp/libomptarget/cmake/Modules/LibomptargetGetDependencies.cmake

If we're sticking with using FindCUDA it's definitely redundant here since it was already called by the time we get here. The support for CUDA language would use the same method but have enable_language(CUDA) somewhere instead of find_package(CUDA)

[ye-luo]

In D88929#2315513, @jhuber6 wrote:

In D88929#2315451, @ye-luo wrote:

I just realized that this patch affects clang and libomptarget.
I cannot comment on clang. Regarding libomptarget, Could you explain why the detection is not put together with other cuda stuff in openmp/libomptarget/cmake/Modules/LibomptargetGetDependencies.cmake

If we're sticking with using FindCUDA it's definitely redundant here since it was already called by the time we get here. The support for CUDA language would use the same method but have enable_language(CUDA) somewhere instead of find_package(CUDA)

Probably not messing with enable_language(CUDA) at the moment, just add cuda_select_nvcc_arch_flags(CUDA_ARCH_FLAGS) to `openmp/libomptarget/cmake/Modules/LibomptargetGetDependencies.cmake?

[jhuber6]

In D88929#2315519, @ye-luo wrote:

Probably not messing with enable_language(CUDA) at the moment, just add cuda_select_nvcc_arch_flags(CUDA_ARCH_FLAGS) to `openmp/libomptarget/cmake/Modules/LibomptargetGetDependencies.cmake?

That only controls loading the library, since this is where we set all the CUDA options I think it's fine to call it here.

[ye-luo]

In D88929#2315538, @jhuber6 wrote:

In D88929#2315519, @ye-luo wrote:

Probably not messing with enable_language(CUDA) at the moment, just add cuda_select_nvcc_arch_flags(CUDA_ARCH_FLAGS) to `openmp/libomptarget/cmake/Modules/LibomptargetGetDependencies.cmake?

That only controls loading the library, since this is where we set all the CUDA options I think it's fine to call it here.

Yes. I'm reducing my review to libomptarget only. I could not comment on clang/CMakeLists.txt. Regarding libomptarget, keep all the CUDA detection inside LibomptargetGetDependencies.cmake

[JonChesterfield]

An alternative approach is to build the deviceRTL for multiple cuda versions and then pick whichever one is the best fit when compiling application code. That has advantages when building the deviceRTL libraries on a different machine to the one that intends to use it.

Cmake isn't my thing, but I see that my trunk build only has libomptarget-nvptx-sm_35.bc when the local card is a sm_50. The downstream amd toolchain builds lots of this library, my install dir has fifteen of them (including sm_50).

[ye-luo]

In D88929#2315640, @JonChesterfield wrote:

An alternative approach is to build the deviceRTL for multiple cuda versions and then pick whichever one is the best fit when compiling application code. That has advantages when building the deviceRTL libraries on a different machine to the one that intends to use it.

Cmake isn't my thing, but I see that my trunk build only has libomptarget-nvptx-sm_35.bc when the local card is a sm_50. The downstream amd toolchain builds lots of this library, my install dir has fifteen of them (including sm_50).

You can build multiple deviceRTL today with LIBOMPTARGET_NVPTX_COMPUTE_CAPABILITIES=50,61,70. This patch tries to add the high arch automatically.

[jhuber6]

Removing redundant call to find_package.

[ye-luo]

openmp/libomptarget/deviceRTLs/nvptx/CMakeLists.txt

80

1. Doesn't work right now. Missing comma ",${CMAKE_MATCH_1}"
2. using CUDA_ARCH as "string(REGEX MATCH" output causes problems.
3. "append" needs to protect redundant 35.
4. I think it is better to move this part of logic to openmp/libomptarget/cmake/Modules/LibomptargetGetDependencies.cmake

after find_package(CUDA QUIET).

[jhuber6]

openmp/libomptarget/deviceRTLs/nvptx/CMakeLists.txt

80

1. I noticed the comma problem, it's because the compute capabilities uses commas instead of semicolons like the rest of CMake to delimit the values. There's another weird error I'm getting while trying to build with this where it will add sm_70 as an argument causing an nvcc error. like in nvcc -o out file.cpp sm_70.
2. What's the problem here? Pretty much the output gives some string that you would pass to nvcc like --arch=sm_70 and I'm regexing out the sm_70 if there was no match or the architecture is too small it doesn't add anything.
3. I'm thinking it would just be easiest to change it do something like set(default_capabilities "35,${CMAKE_MATCH_1}")
4. My feeling is that there's not enough complexity here to justify moving it since I'd need to add just as much code to generate the output and then even more to use it here. Since the LibomptargetGetDependencies.cmake doesn't bother checking whether or not the find_package(CUDA) was successful I feel like there's no need to add logic that requires checking if it was there when we already have it here.

[ye-luo]

openmp/libomptarget/deviceRTLs/nvptx/CMakeLists.txt

95

my point 2 refers to here CUDA_ARCH which gets into the compile line, your point 1 issue. rename your output variable ot CUDA_ARCH_MATCH_OUTPUT should solve the problem.

I still think it is better to move default_capabilities. Very natural to have cuda_select_nvcc_arch_flags next to find_package(cuda) in one place.

[ye-luo]

openmp/libomptarget/deviceRTLs/nvptx/CMakeLists.txt

80

3. better than append but still need a check to avoid "35,35"
4. https://github.com/ye-luo/llvm-project/commit/ac5f20f9770e894ff48467a9317ec0649f5c7562

libomptarget part should be fulling working.

[jhuber6]

Implementing Ye's code. I changed it to putput the architecture as a dependency and then set the value where we define the default architecture. I did a full build from scratch using this and got the sm_70 libraries for my machine without needing to specify it in my build script.

[jdoerfert]

@ye-luo good?

[ye-luo]

LGTM

[JonChesterfield]

Nice patch. Thanks!

[jhuber6]

Screwed up and put the wrong revision in the commit message. Closed in rGd564409946a5a13cb6391fc0fec54dcbd6f6d249

clang/CMakeLists.txt

set(CLANG_DEFAULT_OBJCOPY "objcopy" CACHE STRING
   "Default objcopy executable to use.")
 
 set(CLANG_DEFAULT_OPENMP_RUNTIME "libomp" CACHE STRING
   "Default OpenMP runtime used by -fopenmp.")
 
 # OpenMP offloading requires at least sm_35 because we use shuffle instructions
 # to generate efficient code for reductions and the atomicMax instruction on
 # 64-bit integers in the implementation of conditional lastprivate.
-set(CLANG_OPENMP_NVPTX_DEFAULT_ARCH "sm_35" CACHE STRING
-  "Default architecture for OpenMP offloading to Nvidia GPUs.")
-string(REGEX MATCH "^sm_([0-9]+)$" MATCHED_ARCH "${CLANG_OPENMP_NVPTX_DEFAULT_ARCH}")
-if (NOT DEFINED MATCHED_ARCH OR "${CMAKE_MATCH_1}" LESS 35)
-  message(WARNING "Resetting default architecture for OpenMP offloading to Nvidia GPUs to sm_35")
+set(CUDA_ARCH_FLAGS "sm_35")
+
+# Try to find the highest architecture the host supports
+if (NOT DEFINED CLANG_OPENMP_NVPTX_DEFAULT_ARCH)
+  find_package(CUDA QUIET)
+  if (CUDA_FOUND)
+    cuda_select_nvcc_arch_flags(CUDA_ARCH_FLAGS)
+  endif()
+else()
+  set(CUDA_ARCH_FLAGS ${CLANG_OPENMP_NVPTX_DEFAULT_ARCH})
+endif()
+
+string(REGEX MATCH "sm_([0-9]+)" CUDA_ARCH ${CUDA_ARCH_FLAGS})
+if (NOT DEFINED CUDA_ARCH OR "${CMAKE_MATCH_1}" LESS 35)
   set(CLANG_OPENMP_NVPTX_DEFAULT_ARCH "sm_35" CACHE STRING
     "Default architecture for OpenMP offloading to Nvidia GPUs." FORCE)
+  message(WARNING "Resetting default architecture for OpenMP offloading to Nvidia GPUs to sm_35")
+else()
+  set(CLANG_OPENMP_NVPTX_DEFAULT_ARCH ${CUDA_ARCH} CACHE STRING
+    "Default architecture for OpenMP offloading to Nvidia GPUs.")
 endif()
 
 set(CLANG_SYSTEMZ_DEFAULT_ARCH "z10" CACHE STRING "SystemZ Default Arch")
 
 set(CLANG_VENDOR ${PACKAGE_VENDOR} CACHE STRING
   "Vendor-specific text for showing with version information.")
 
 if( CLANG_VENDOR )

openmp/libomptarget/deviceRTLs/nvptx/CMakeLists.txt

set(cuda_src_files
       ${devicertl_common_directory}/src/sync.cu
       ${devicertl_common_directory}/src/task.cu
       src/target_impl.cu
   )
 
   set(omp_data_objects ${devicertl_common_directory}/src/omp_data.cu)
 
   # Get the compute capability the user requested or use SM_35 by default.
-  # SM_35 is what clang uses by default.
-  set(default_capabilities 35)
+  set(compute_capabilities 35)
+  if (NOT DEFINED LIBOMPTARGET_NVPTX_COMPUTE_CAPABILITIES)
+    find_package(CUDA QUIET)
+    if (CUDA_FOUND)
+      cuda_select_nvcc_arch_flags(CUDA_ARCH_FLAGS)
+      string(REGEX MATCH "sm_([0-9]+)" CUDA_ARCH ${CUDA_ARCH_FLAGS})
+      if (NOT DEFINED CUDA_ARCH OR "${CMAKE_MATCH_1}" LESS 35)
+          message(WARNING "Setting default architecture for OpenMP target library to sm_35")
+      else()
+        list(APPEND compute_capabilities ${CMAKE_MATCH_1})
+      endif()
+    endif()
+  endif()
+
+
   if (DEFINED LIBOMPTARGET_NVPTX_COMPUTE_CAPABILITY)
     set(default_capabilities ${LIBOMPTARGET_NVPTX_COMPUTE_CAPABILITY})
     libomptarget_warning_say("LIBOMPTARGET_NVPTX_COMPUTE_CAPABILITY is deprecated, please use LIBOMPTARGET_NVPTX_COMPUTE_CAPABILITIES")
   endif()
-  set(LIBOMPTARGET_NVPTX_COMPUTE_CAPABILITIES ${default_capabilities} CACHE STRING
+  set(LIBOMPTARGET_NVPTX_COMPUTE_CAPABILITIES ${compute_capabilities} CACHE STRING
     "List of CUDA Compute Capabilities to be used to compile the NVPTX device RTL.")
   string(REPLACE "," ";" nvptx_sm_list ${LIBOMPTARGET_NVPTX_COMPUTE_CAPABILITIES})
 
   foreach(sm ${nvptx_sm_list})
     set(CUDA_ARCH ${CUDA_ARCH} -gencode arch=compute_${sm},code=sm_${sm})
   endforeach()
 
   # Override default MAX_SM in src/target_impl.h if requested
   if (DEFINED LIBOMPTARGET_NVPTX_MAX_SM)
     set(MAX_SM_DEFINITION "-DMAX_SM=${LIBOMPTARGET_NVPTX_MAX_SM}")
   endif()
 
   # Activate RTL message dumps if requested by the user.