Enable OpenMP back-ends for clang-cuda

cmake/alpakaCommon.cmake

@@ -386,7 +386,9 @@ if(alpaka_ACC_GPU_CUDA_ENABLE)
             endif()
 
             if(alpaka_ACC_CPU_B_OMP2_T_SEQ_ENABLE OR alpaka_ACC_CPU_B_SEQ_T_OMP2_ENABLE)
-                message(FATAL_ERROR "Clang as a CUDA compiler does not support OpenMP 2!")
+                # FindOpenMP.cmake guards the OpenMP::OpenMP_CXX target with a $<COMPILE_LANGUAGE:CXX> generator expression.
+                # We have to duplicate the flags here.
+                alpaka_set_compiler_options(DEVICE target alpaka "$<$<COMPILE_LANGUAGE:CUDA>:SHELL:${OpenMP_CXX_FLAGS}>")
             endif()
 
             target_compile_options(alpaka INTERFACE $<$<COMPILE_LANGUAGE:CUDA>:-Wno-unknown-cuda-version>)

include/alpaka/acc/AccCpuOmp2Blocks.hpp

@@ -1,4 +1,4 @@
-/* Copyright 2022 Axel Huebl, Benjamin Worpitz, René Widera, Jan Stephan, Bernhard Manfred Gruber
+/* Copyright 2023 Axel Hübl, Benjamin Worpitz, René Widera, Jan Stephan, Bernhard Manfred Gruber
  * SPDX-License-Identifier: MPL-2.0
  */
 
@@ -40,7 +40,12 @@
 
 #ifdef ALPAKA_ACC_CPU_B_OMP2_T_SEQ_ENABLED
 
-#    if _OPENMP < 200203
+/* clang doesn't define the _OPENMP macro when compiling in CUDA mode. We need to turn off -Wundef so clang doesn't
+   complain about our preprocessor guards.*/
+#    pragma clang diagnostic push
+#    pragma clang diagnostic ignored "-Wundef"
+
+#    if(_OPENMP < 200203) && !defined(BOOST_COMP_CLANG_CUDA)
 #        error If ALPAKA_ACC_CPU_B_OMP2_T_SEQ_ENABLED is set, the compiler has to support OpenMP 2.0 or higher!
 #    endif
 
@@ -223,4 +228,6 @@ namespace alpaka
     } // namespace trait
 } // namespace alpaka
 
+#    pragma clang diagnostic pop
+
 #endif

include/alpaka/acc/AccCpuOmp2Threads.hpp

@@ -1,4 +1,4 @@
-/* Copyright 2022 Axel Huebl, Benjamin Worpitz, René Widera, Jan Stephan, Bernhard Manfred Gruber
+/* Copyright 2023 Axel Hübl, Benjamin Worpitz, René Widera, Jan Stephan, Bernhard Manfred Gruber
  * SPDX-License-Identifier: MPL-2.0
  */
 
@@ -40,7 +40,12 @@
 
 #ifdef ALPAKA_ACC_CPU_B_SEQ_T_OMP2_ENABLED
 
-#    if _OPENMP < 200203
+/* clang doesn't define the _OPENMP macro when compiling in CUDA mode. We need to turn off -Wundef so clang doesn't
+   complain about our preprocessor guards.*/
+#    pragma clang diagnostic push
+#    pragma clang diagnostic ignored "-Wundef"
+
+#    if _OPENMP < 200203 && !defined(BOOST_COMP_CLANG_CUDA)
 #        error If ALPAKA_ACC_CPU_B_SEQ_T_OMP2_ENABLED is set, the compiler has to support OpenMP 2.0 or higher!
 #    endif
 
@@ -233,4 +238,6 @@ namespace alpaka
     } // namespace trait
 } // namespace alpaka
 
+#    pragma clang diagnostic pop
+
 #endif

include/alpaka/atomic/AtomicOmpBuiltIn.hpp

@@ -1,4 +1,4 @@
-/* Copyright 2022 René Widera, Bernhard Manfred Gruber
+/* Copyright 2023 René Widera, Bernhard Manfred Gruber, Jan Stephan
  * SPDX-License-Identifier: MPL-2.0
  */
 
@@ -8,7 +8,16 @@
 #include "alpaka/atomic/Traits.hpp"
 #include "alpaka/core/BoostPredef.hpp"
 
-#ifdef _OPENMP
+#if defined(ALPAKA_ACC_CPU_B_OMP2_T_SEQ_ENABLED) || defined(ALPAKA_ACC_CPU_B_SEQ_T_OMP2_ENABLED)
+
+/* clang doesn't define the _OPENMP macro when compiling in CUDA mode. We need to turn off -Wundef so clang doesn't
+   complain about our preprocessor guards.*/
+#    pragma clang diagnostic push
+#    pragma clang diagnostic ignored "-Wundef"
+
+#    if(_OPENMP < 200203) && !defined(BOOST_COMP_CLANG_CUDA)
+#        error If ALPAKA_ACC_CPU_B_OMP2_T_SEQ_ENABLED is set, the compiler has to support OpenMP 2.0 or higher!
+#    endif
 
 namespace alpaka
 {
@@ -24,7 +33,7 @@ namespace alpaka
     {
 // check for OpenMP 3.1+
 // "omp atomic capture" is not supported before OpenMP 3.1
-#    if _OPENMP >= 201107
+#    if _OPENMP >= 201107 || BOOST_COMP_CLANG_CUDA
 
         //! The OpenMP accelerators atomic operation: ADD
         template<typename T, typename THierarchy>
@@ -169,11 +178,11 @@ namespace alpaka
             }
         };
 
-#    endif // _OPENMP >= 201107
+#    endif // _OPENMP >= 201107 || BOOST_COMP_CLANG_CUDA
 
 // check for OpenMP 5.1+
-// "omp atomic compare" was introduced with OpenMP 5.1
-#    if _OPENMP >= 202011
+// "omp atomic compare" was introduced with OpenMP 5.1 (supported since clang-15)
+#    if _OPENMP >= 202011 || (BOOST_COMP_CLANG_CUDA >= BOOST_VERSION_NUMBER(15, 0, 0))
 
         //! The OpenMP accelerators atomic operation: Min
         template<typename T, typename THierarchy>
@@ -319,9 +328,11 @@ namespace alpaka
             }
         };
 
-#    endif // _OPENMP >= 202011
+#    endif // _OPENMP >= 202011 || (BOOST_COMP_CLANG_CUDA >= BOOST_VERSION_NUMBER(15, 0, 0))
 
     } // namespace trait
 } // namespace alpaka
 
+#    pragma clang diagnostic pop
+
 #endif

include/alpaka/block/sync/BlockSyncBarrierOmp.hpp

@@ -9,7 +9,16 @@
 
 #include <cstdint>
 
-#ifdef _OPENMP
+#if defined(ALPAKA_ACC_CPU_B_OMP2_T_SEQ_ENABLED) || defined(ALPAKA_ACC_CPU_B_SEQ_T_OMP2_ENABLED)
+
+/* clang doesn't define the _OPENMP macro when compiling in CUDA mode. We need to turn off -Wundef so clang doesn't
+   complain about our preprocessor guards.*/
+#    pragma clang diagnostic push
+#    pragma clang diagnostic ignored "-Wundef"
+
+#    if(_OPENMP < 200203) && !defined(BOOST_COMP_CLANG_CUDA)
+#        error If ALPAKA_ACC_CPU_B_OMP2_T_SEQ_ENABLED is set, the compiler has to support OpenMP 2.0 or higher!
+#    endif
 
 namespace alpaka
 {
@@ -106,4 +115,6 @@ namespace alpaka
     } // namespace trait
 } // namespace alpaka
 
+#    pragma clang diagnostic pop
+
 #endif

include/alpaka/core/OmpSchedule.hpp

@@ -1,17 +1,23 @@
-/* Copyright 2022 Sergei Bastrakov, Bernhard Manfred Gruber
+/* Copyright 2023 Sergei Bastrakov, Bernhard Manfred Gruber, Jan Stephan
  * SPDX-License-Identifier: MPL-2.0
  */
 
 #pragma once
 
+#include "alpaka/core/BoostPredef.hpp"
 #include "alpaka/core/Common.hpp"
 
-#ifdef _OPENMP
+#if defined(_OPENMP)
 #    include <omp.h>
 #endif
 
 #include <cstdint>
 
+/* clang doesn't define the _OPENMP macro when compiling in CUDA mode. We need to turn off -Wundef so clang doesn't
+   complain about our preprocessor guards.*/
+#pragma clang diagnostic push
+#pragma clang diagnostic ignored "-Wundef"
+
 namespace alpaka::omp
 {
     //! Representation of OpenMP schedule information: kind and chunk size. This class can be used regardless of
@@ -30,7 +36,7 @@ namespace alpaka::omp
             Dynamic = 2u,
             Guided = 3u,
             // Auto supported since OpenMP 3.0
-#if defined _OPENMP && _OPENMP >= 200805
+#if(defined _OPENMP && _OPENMP >= 200805) || BOOST_COMP_CLANG_CUDA
             Auto = 4u,
 #endif
             Runtime = 5u
@@ -86,3 +92,5 @@ namespace alpaka::omp
         }
     }
 } // namespace alpaka::omp
+
+#pragma clang diagnostic pop

include/alpaka/idx/bt/IdxBtOmp.hpp

@@ -1,4 +1,4 @@
-/* Copyright 2022 Axel Huebl, Benjamin Worpitz, Matthias Werner, Jan Stephan, Bernhard Manfred Gruber
+/* Copyright 2023 Axel Hübl, Benjamin Worpitz, Matthias Werner, Jan Stephan, Bernhard Manfred Gruber, Jan Stephan
  * SPDX-License-Identifier: MPL-2.0
  */
 
@@ -12,7 +12,16 @@
 #include "alpaka/vec/Vec.hpp"
 #include "alpaka/workdiv/Traits.hpp"
 
-#ifdef _OPENMP
+#ifdef ALPAKA_ACC_CPU_B_SEQ_T_OMP2_ENABLED
+
+/* clang doesn't define the _OPENMP macro when compiling in CUDA mode. We need to turn off -Wundef so clang doesn't
+   complain about our preprocessor guards.*/
+#    pragma clang diagnostic push
+#    pragma clang diagnostic ignored "-Wundef"
+
+#    if(_OPENMP < 200203) && !defined(BOOST_COMP_CLANG_CUDA)
+#        error If ALPAKA_ACC_CPU_B_OMP2_T_SEQ_ENABLED is set, the compiler has to support OpenMP 2.0 or higher!
+#    endif
 
 #    include <omp.h>
 
@@ -74,4 +83,6 @@ namespace alpaka
     } // namespace trait
 } // namespace alpaka
 
+#    pragma clang diagnostic pop
+
 #endif

include/alpaka/kernel/TaskKernelCpuOmp2Blocks.hpp

@@ -1,4 +1,4 @@
-/* Copyright 2022 Benjamin Worpitz, Bert Wesarg, René Widera, Sergei Bastrakov, Bernhard Manfred Gruber
+/* Copyright 2023 Benjamin Worpitz, Bert Wesarg, René Widera, Sergei Bastrakov, Bernhard Manfred Gruber, Jan Stephan
  * SPDX-License-Identifier: MPL-2.0
  */
 
@@ -31,7 +31,12 @@
 
 #ifdef ALPAKA_ACC_CPU_B_OMP2_T_SEQ_ENABLED
 
-#    if _OPENMP < 200203
+/* clang doesn't define the _OPENMP macro when compiling in CUDA mode. We need to turn off -Wundef so clang doesn't
+   complain about our preprocessor guards.*/
+#    pragma clang diagnostic push
+#    pragma clang diagnostic ignored "-Wundef"
+
+#    if _OPENMP < 200203 && !defined(BOOST_COMP_CLANG_CUDA)
 #        error If ALPAKA_ACC_CPU_B_OMP2_T_SEQ_ENABLED is set, the compiler has to support OpenMP 2.0 or higher!
 #    endif
 
@@ -515,7 +520,7 @@ namespace alpaka
         {
         };
 
-#    if _OPENMP >= 200805
+#    if _OPENMP >= 200805 || BOOST_COMP_CLANG_CUDA
         //! Executor of parallel OpenMP loop with auto schedule set
         //!
         //! Does not use chunk size.
@@ -682,7 +687,7 @@ namespace alpaka
                         numIterations,
                         schedule);
                     break;
-#    if _OPENMP >= 200805
+#    if _OPENMP >= 200805 || BOOST_COMP_CLANG_CUDA
                 case omp::Schedule::Auto:
                     ParallelForImpl<TKernel, omp::Schedule, omp::Schedule::Auto>{}(
                         kernel,
@@ -949,4 +954,6 @@ namespace alpaka
     } // namespace trait
 } // namespace alpaka
 
+#    pragma clang diagnostic pop
+
 #endif

include/alpaka/kernel/TaskKernelCpuOmp2Threads.hpp

@@ -1,4 +1,4 @@
-/* Copyright 2022 Axel Huebl, Benjamin Worpitz, Bert Wesarg, René Widera, Jan Stephan, Bernhard Manfred Gruber
+/* Copyright 2023 Axel Hübl, Benjamin Worpitz, Bert Wesarg, René Widera, Jan Stephan, Bernhard Manfred Gruber
  * SPDX-License-Identifier: MPL-2.0
  */
 
@@ -29,7 +29,12 @@
 
 #ifdef ALPAKA_ACC_CPU_B_SEQ_T_OMP2_ENABLED
 
-#    if _OPENMP < 200203
+/* clang doesn't define the _OPENMP macro when compiling in CUDA mode. We need to turn off -Wundef so clang doesn't
+   complain about our preprocessor guards.*/
+#    pragma clang diagnostic push
+#    pragma clang diagnostic ignored "-Wundef"
+
+#    if _OPENMP < 200203 && !defined(BOOST_COMP_CLANG_CUDA)
 #        error If ALPAKA_ACC_CPU_B_SEQ_T_OMP2_ENABLED is set, the compiler has to support OpenMP 2.0 or higher!
 #    endif
 
@@ -195,4 +200,6 @@ namespace alpaka
     } // namespace trait
 } // namespace alpaka
 
+#    pragma clang diagnostic pop
+
 #endif

include/alpaka/mem/fence/MemFenceOmp2Blocks.hpp

@@ -1,4 +1,4 @@
-/* Copyright 2022 Jan Stephan, Bernhard Manfred Gruber, Andrea Bocci
+/* Copyright 2023 Jan Stephan, Bernhard Manfred Gruber, Andrea Bocci
  * SPDX-License-Identifier: MPL-2.0
  */
 
@@ -9,7 +9,12 @@
 
 #ifdef ALPAKA_ACC_CPU_B_OMP2_T_SEQ_ENABLED
 
-#    if _OPENMP < 200203
+/* clang doesn't define the _OPENMP macro when compiling in CUDA mode. We need to turn off -Wundef so clang doesn't
+   complain about our preprocessor guards.*/
+#    pragma clang diagnostic push
+#    pragma clang diagnostic ignored "-Wundef"
+
+#    if _OPENMP < 200203 && !defined(BOOST_COMP_CLANG_CUDA)
 #        error If ALPAKA_ACC_CPU_B_OMP2_T_SEQ_ENABLED is set, the compiler has to support OpenMP 2.0 or higher!
 #    endif
 
@@ -51,4 +56,6 @@ namespace alpaka
     } // namespace trait
 } // namespace alpaka
 
+#    pragma clang diagnostic pop
+
 #endif

include/alpaka/mem/fence/MemFenceOmp2Threads.hpp

@@ -1,4 +1,4 @@
-/* Copyright 2022 Jan Stephan, Bernhard Manfred Gruber
+/* Copyright 2023 Jan Stephan, Bernhard Manfred Gruber
  * SPDX-License-Identifier: MPL-2.0
  */
 
@@ -9,7 +9,12 @@
 
 #ifdef ALPAKA_ACC_CPU_B_SEQ_T_OMP2_ENABLED
 
-#    if _OPENMP < 200203
+/* clang doesn't define the _OPENMP macro when compiling in CUDA mode. We need to turn off -Wundef so clang doesn't
+   complain about our preprocessor guards.*/
+#    pragma clang diagnostic push
+#    pragma clang diagnostic ignored "-Wundef"
+
+#    if _OPENMP < 200203 && !defined(BOOST_COMP_CLANG_CUDA)
 #        error If ALPAKA_ACC_CPU_B_SEQ_T_OMP2_ENABLED is set, the compiler has to support OpenMP 2.0 or higher!
 #    endif
 
@@ -65,4 +70,7 @@ namespace alpaka
         };
     } // namespace trait
 } // namespace alpaka
+
+#    pragma clang diagnostic pop
+
 #endif

include/alpaka/test/queue/QueueCpuOmp2Collective.hpp

@@ -1,4 +1,4 @@
-/* Copyright 2022 Axel Huebl, Benjamin Worpitz, Matthias Werner, Jan Stephan, Bernhard Manfred Gruber
+/* Copyright 2023 Axel Hübl, Benjamin Worpitz, Matthias Werner, Jan Stephan, Bernhard Manfred Gruber
  * SPDX-License-Identifier: MPL-2.0
  */
 
@@ -21,7 +21,12 @@
 
 #ifdef ALPAKA_ACC_CPU_B_OMP2_T_SEQ_ENABLED
 
-#    if _OPENMP < 200203
+/* clang doesn't define the _OPENMP macro when compiling in CUDA mode. We need to turn off -Wundef so clang doesn't
+   complain about our preprocessor guards.*/
+#    pragma clang diagnostic push
+#    pragma clang diagnostic ignored "-Wundef"
+
+#    if _OPENMP < 200203 && !defined(BOOST_COMP_CLANG_CUDA)
 #        error If ALPAKA_ACC_CPU_B_OMP2_T_SEQ_ENABLED is set, the compiler has to support OpenMP 2.0 or higher!
 #    endif
 
@@ -292,6 +297,8 @@ namespace alpaka
     };
 } // namespace alpaka
 
+#    pragma clang diagnostic pop
+
 #    include "alpaka/event/EventCpu.hpp"
 
 #endif

script/job_generator/alpaka_filter.py

@@ -45,14 +45,4 @@ def alpaka_post_filter(row: List) -> bool:
     ):
         return False
 
-    # OpenMP is not supported for clang as cuda compiler
-    # https://github.com/alpaka-group/alpaka/issues/639
-    if row_check_name(row, DEVICE_COMPILER, "==", CLANG_CUDA) and (
-        row_check_backend_version(row, ALPAKA_ACC_CPU_B_OMP2_T_SEQ_ENABLE, "==", ON_VER)
-        or row_check_backend_version(
-            row, ALPAKA_ACC_CPU_B_SEQ_T_OMP2_ENABLE, "==", ON_VER
-        )
-    ):
-        return False
-
     return True