[SYCL] Use std::array as storage for syc::vec on device

sycl/include/sycl/detail/generic_type_traits.hpp

@@ -328,11 +328,11 @@ template <typename T> auto convertToOpenCLType(T &&x) {
  // TODO: for some mysterious reasons on NonUniformGroups E2E tests fail if
  // we use the "else" version only. I suspect that's an issues with
  // non-uniform groups implementation.
- if constexpr (std::is_same_v<MatchingVec, no_ref>)
- return static_cast<typename MatchingVec::vector_t>(x);
- else
- return static_cast<typename MatchingVec::vector_t>(
-  x.template as<MatchingVec>());
+ if constexpr (std::is_same_v<MatchingVec, no_ref>) {
+ return bit_cast<typename no_ref::vector_t>(x);
+ } else {
+ return bit_cast<typename MatchingVec::vector_t>(x);
+ }
 #else
  return x.template as<MatchingVec>();
 #endif

sycl/include/sycl/ext/oneapi/bfloat16.hpp

@@ -38,6 +38,12 @@ using Vec3StorageT = Bfloat16StorageT __attribute__((ext_vector_type(3)));
 using Vec4StorageT = Bfloat16StorageT __attribute__((ext_vector_type(4)));
 using Vec8StorageT = Bfloat16StorageT __attribute__((ext_vector_type(8)));
 using Vec16StorageT = Bfloat16StorageT __attribute__((ext_vector_type(16)));
+
+using Vec2DeviceStorageT = std::array<Bfloat16StorageT, 2>;
+using Vec3DeviceStorageT = std::array<Bfloat16StorageT, 3>;
+using Vec4DeviceStorageT = std::array<Bfloat16StorageT, 4>;
+using Vec8DeviceStorageT = std::array<Bfloat16StorageT, 8>;
+using Vec16DeviceStorageT = std::array<Bfloat16StorageT, 16>;
 #else
 using Vec2StorageT = std::array<Bfloat16StorageT, 2>;
 using Vec3StorageT = std::array<Bfloat16StorageT, 3>;

sycl/include/sycl/half_type.hpp

@@ -249,6 +249,12 @@ using Vec3StorageT = StorageT __attribute__((ext_vector_type(3)));
 using Vec4StorageT = StorageT __attribute__((ext_vector_type(4)));
 using Vec8StorageT = StorageT __attribute__((ext_vector_type(8)));
 using Vec16StorageT = StorageT __attribute__((ext_vector_type(16)));
+
+using Vec2DeviceStorageT = std::array<StorageT, 2>;
+using Vec3DeviceStorageT = std::array<StorageT, 4>;
+using Vec4DeviceStorageT = std::array<StorageT, 4>;
+using Vec8DeviceStorageT = std::array<StorageT, 8>;
+using Vec16DeviceStorageT = std::array<StorageT, 16>;
 #else // SYCL_DEVICE_ONLY
 using StorageT = detail::host_half_impl::half;
 // No need to extract underlying data type for built-in functions operating on