[CUDA][HIP] add test for uses of std::array on device code

External/CUDA/CMakeLists.txt

@@ -63,6 +63,7 @@ macro(create_local_cuda_tests VariantSuffix)
   list(APPEND CUDA_LOCAL_TESTS assert)
   list(APPEND CUDA_LOCAL_TESTS axpy)
   list(APPEND CUDA_LOCAL_TESTS algorithm)
+  list(APPEND CUDA_LOCAL_TESTS array)
   list(APPEND CUDA_LOCAL_TESTS cmath)
   list(APPEND CUDA_LOCAL_TESTS complex)
   list(APPEND CUDA_LOCAL_TESTS math_h)

External/CUDA/array.cu

@@ -0,0 +1,71 @@
+// Check that we can use std::array on device code
+//
+// After libstdc++ 15, some internal asserts rely on function that are neither
+// constexpr nor device. This can trigger errors when using std::array members
+// on device code.
+//
+// This workaround is implemented in bits/c++config.h
+
+#include <stdio.h>
+
+#if __cplusplus >= 201103L
+
+#include <array>
+#include <assert.h>
+
+#if __cplusplus >= 201402L && STDLIB_VERSION >= 2014
+constexpr
+#endif
+    __host__ __device__ size_t
+    test_array() {
+  std::array<int, 4> A = {0, 1, 2, 3};
+  size_t N = A.size();
+  assert(N == 4);
+
+#if __cplusplus >= 201402L && STDLIB_VERSION >= 2014
+  const int fst = A[0];
+  assert(fst == 0);
+#endif
+
+#if __cplusplus >= 201703L && STDLIB_VERSION >= 2017
+  A[0] = 4;
+  int snd = A[0];
+  assert(snd == 4);
+#endif
+  return N;
+}
+
+__host__ __device__ void do_test() {
+  // call the function in a constexpr and a non-constexpr context
+  size_t M = test_array();
+  (void)(M);
+#if __cplusplus >= 201402L && STDLIB_VERSION >= 2014
+  constexpr size_t N = test_array();
+  (void)(N);
+#endif
+}
+
+__global__ void kernel() { do_test(); }
+
+int main() {
+  kernel<<<32, 32>>>();
+  cudaError_t err = cudaDeviceSynchronize();
+  if (err != cudaSuccess) {
+    printf("CUDA error %d\n", (int)err);
+    return 1;
+  }
+
+  do_test();
+
+  printf("Success!\n");
+  return 0;
+}
+
+#else
+
+int main() {
+  printf("Success!\n");
+  return 0;
+}
+
+#endif

External/CUDA/array.reference_output

@@ -0,0 +1,2 @@
+Success!
+exit 0

External/HIP/CMakeLists.txt

@@ -22,6 +22,7 @@ macro(create_local_hip_tests VariantSuffix)
   list(APPEND HIP_LOCAL_TESTS saxpy)
   list(APPEND HIP_LOCAL_TESTS memmove)
   list(APPEND HIP_LOCAL_TESTS split-kernel-args)
+  list(APPEND HIP_LOCAL_TESTS array)
 
   # TODO: Re-enable InOneWeekend after it is fixed
   #list(APPEND HIP_LOCAL_TESTS InOneWeekend)

External/HIP/array.hip

@@ -0,0 +1,2 @@
+#include "cuda2hip.h"
+#include "../CUDA/array.cu"

External/HIP/array.reference_output

@@ -0,0 +1,2 @@
+Success!
+exit 0

External/HIP/cuda2hip.h

@@ -0,0 +1,10 @@
+#ifndef CUDA_2_HIP
+#define CUDA_2_HIP
+
+#include "hip/hip_runtime.h"
+
+#define cudaError_t hipError_t
+#define cudaSuccess hipSuccess
+#define cudaDeviceSynchronize hipDeviceSynchronize
+
+#endif