[SYCL][ESIMD] Make lsc_block_load/store try to use 64 bit blocks by default

sycl/include/sycl/ext/intel/esimd/detail/simd_obj_impl.hpp

@@ -99,6 +99,15 @@ static inline constexpr bool is_simd_flag_type_v = is_simd_flag_type<T>::value;
 namespace detail {
 
 /// @cond ESIMD_DETAIL
+/// \c dqword_element_aligned_tag type. Flag of this type should be used in load
+/// and store operations when memory address is aligned by simd object's element
+/// type or dword whatever is greater.
+struct dqword_element_aligned_tag {
+  template <typename VT, typename ET = detail::element_type_t<VT>>
+  static constexpr unsigned alignment = alignof(ET) > 4 ? alignof(ET) : 4;
+};
+
+inline constexpr dqword_element_aligned_tag dqword_element_aligned = {};
 
 // Functions to support efficient simd constructors - avoiding internal loop
 // over elements.
@@ -881,6 +890,9 @@ class simd_obj_impl {
 
 } // namespace detail
 
+template <>
+struct is_simd_flag_type<detail::dqword_element_aligned_tag> : std::true_type {
+};
 } // namespace ext::intel::esimd
 } // __SYCL_INLINE_VER_NAMESPACE(_V1)
 } // namespace sycl

sycl/include/sycl/ext/intel/experimental/esimd/common.hpp

@@ -109,6 +109,18 @@ template <typename T, lsc_data_size DS> constexpr void check_lsc_data_size() {
   static_assert(DS != lsc_data_size::default_size || sizeof(T) == 1 ||
                     sizeof(T) == 2 || sizeof(T) == 4 || sizeof(T) == 8,
                 "Unsupported data type");
+  static_assert(
+      DS == lsc_data_size::default_size ||
+          (sizeof(T) == 1 &&
+           (DS == lsc_data_size::u8 || DS == lsc_data_size::u8u32)) ||
+          (sizeof(T) == 2 &&
+           (DS == lsc_data_size::u16 || DS == lsc_data_size::u16u32 ||
+            DS == lsc_data_size::u16u32h)) ||
+          (sizeof(T) == 4 &&
+           (DS == lsc_data_size::u32 || DS == lsc_data_size::u8u32 ||
+            DS == lsc_data_size::u16u32 || DS == lsc_data_size::u16u32h)) ||
+          (sizeof(T) == 8 && DS == lsc_data_size::u64),
+      "Data type does not match data size");
 }
 
 template <lsc_vector_size VS> constexpr uint8_t to_int() {