Merge pull request #6106 from crtrott/fix-nvhpc-compilation

Fix nvhpc 23.x Compilation
sandialabs · May 10, 2023 · fb0c1b8 · fb0c1b8
2 parents f15b5ab + 531b01d
commit fb0c1b8
Show file tree

Hide file tree

Showing 4 changed files with 35 additions and 7 deletions.
diff --git a/core/unit_test/TestJoinBackwardCompatibility.hpp b/core/unit_test/TestJoinBackwardCompatibility.hpp
@@ -21,13 +21,16 @@
                                // unimplemented reduction features
 namespace {
 
+// FIXME_NVHPC 23.3 errors out when using enums here
+// NVC++-F-0000-Internal compiler error. process_acc_put_dinit: unexpected
+// datatype    5339
+#ifndef KOKKOS_COMPILER_NVHPC
 enum MyErrorCode {
   no_error                           = 0b000,
   error_operator_plus_equal          = 0b001,
   error_operator_plus_equal_volatile = 0b010,
   error_join_volatile                = 0b100,
   expected_join_volatile             = 0b1000
-
 };
 
 KOKKOS_FUNCTION constexpr MyErrorCode operator|(MyErrorCode lhs,
@@ -36,6 +39,17 @@ KOKKOS_FUNCTION constexpr MyErrorCode operator|(MyErrorCode lhs,
                                   static_cast<int>(rhs));
 }
 
+#else
+
+using MyErrorCode                                        = unsigned;
+constexpr MyErrorCode no_error                           = 0b000;
+constexpr MyErrorCode error_operator_plus_equal          = 0b001;
+constexpr MyErrorCode error_operator_plus_equal_volatile = 0b010;
+constexpr MyErrorCode error_join_volatile                = 0b100;
+constexpr MyErrorCode expected_join_volatile             = 0b1000;
+
+#endif
+
 static_assert((no_error | error_operator_plus_equal_volatile) ==
                   error_operator_plus_equal_volatile,
               "");
@@ -130,8 +144,9 @@ void test_join_backward_compatibility() {
 }
 
 TEST(TEST_CATEGORY, join_backward_compatibility) {
-#if defined(KOKKOS_ENABLE_CUDA) && \
-    defined(KOKKOS_COMPILER_NVHPC)  // FIXME_NVHPC
+#if defined(KOKKOS_ENABLE_CUDA) && defined(KOKKOS_COMPILER_NVHPC) && \
+    KOKKOS_COMPILER_NVHPC <                                          \
+        230300  // FIXME_NVHPC test passes with workaround in 23.3
   GTEST_SKIP() << "FIXME wrong result";
 #endif
   test_join_backward_compatibility();

diff --git a/core/unit_test/TestMathematicalConstants.hpp b/core/unit_test/TestMathematicalConstants.hpp
@@ -63,7 +63,7 @@ struct TestMathematicalConstants {
 
   KOKKOS_FUNCTION void use_on_device() const {
 #if defined(KOKKOS_COMPILER_NVCC) || defined(KOKKOS_ENABLE_OPENMPTARGET) || \
-    defined(KOKKOS_ENABLE_OPENACC)
+    defined(KOKKOS_ENABLE_OPENACC) || defined(KOKKOS_COMPILER_NVHPC)
     take_by_value(Trait::value);
 #else
     (void)take_address_of(Trait::value);

diff --git a/core/unit_test/TestNumericTraits.hpp b/core/unit_test/TestNumericTraits.hpp
@@ -40,7 +40,14 @@ struct extrema {
 
   DEFINE_EXTREMA(float, -FLT_MAX, FLT_MAX);
   DEFINE_EXTREMA(double, -DBL_MAX, DBL_MAX);
+
+// FIXME_NVHPC: with 23.3 using long double in KOKKOS_FUNCTION is hard error
+#if !defined(KOKKOS_ENABLE_CUDA) || !defined(KOKKOS_COMPILER_NVHPC)
   DEFINE_EXTREMA(long double, -LDBL_MAX, LDBL_MAX);
+#else
+  static long double min(long double) { return -LDBL_MAX; }
+  static long double max(long double) { return LDBL_MAX; }
+#endif
 
 #undef DEFINE_EXTREMA
 };
@@ -163,7 +170,8 @@ struct TestNumericTraits {
   }
 
   KOKKOS_FUNCTION void use_on_device() const {
-#if defined(KOKKOS_COMPILER_NVCC) || defined(KOKKOS_ENABLE_OPENMPTARGET)
+#if defined(KOKKOS_COMPILER_NVCC) || defined(KOKKOS_COMPILER_NVHPC) || \
+    defined(KOKKOS_ENABLE_OPENMPTARGET) || defined(KOKKOS_ENABLE_OPENACC)
     take_by_value(trait<T>::value);
 #else
     (void)take_address_of(trait<T>::value);

diff --git a/tpls/desul/include/desul/atomics/cuda/CUDA_asm.hpp b/tpls/desul/include/desul/atomics/cuda/CUDA_asm.hpp
@@ -2,10 +2,15 @@
 namespace desul {
 namespace Impl {
 // Choose the variant of atomics we are using later
+// The __isGlobal intrinsic was only introduced in CUDA 11.2
+// It also stopped working in NVC++ 23.1 - it works in 22.11
+// this is a bug in NVHPC, not treating CUDA intrinsics correctly
+// FIXME_NVHPC
 #if !defined(DESUL_IMPL_ATOMIC_CUDA_PTX_PREDICATE) && \
     !defined(DESUL_IMPL_ATOMIC_CUDA_PTX_ISGLOBAL)
-#if (__CUDACC_VER_MAJOR__ > 11) || \
-    ((__CUDACC_VER_MAJOR__ == 11) && (__CUDACC_VER_MINOR__ > 1))
+#if ((__CUDACC_VER_MAJOR__ > 11) ||                                   \
+     ((__CUDACC_VER_MAJOR__ == 11) && (__CUDACC_VER_MINOR__ > 1))) && \
+    (!defined(__NVCOMPILER) || __NVCOMPILER_MAJOR__ < 23)
 #define DESUL_IMPL_ATOMIC_CUDA_PTX_ISGLOBAL
 #else
 #define DESUL_IMPL_ATOMIC_CUDA_PTX_PREDICATE