[NativeCPU] Avoid using vector_t in NativeCPU

[steffenlarsen]

This commit avoid the use of the to-be-removed sycl::vec::vector_t in the native-cpu utilities.

[steffenlarsen]

@aelovikov-intel - There are no other nativecpu reviewers, but since you worked on removing the vector_t alias in times past, I was hoping you could help review this. I am open to other solutions if you can think of any.

[aelovikov-intel]

Would

llvm/sycl/include/sycl/detail/generic_type_traits.hpp

Lines 117 to 194 in 62397e8

	template <typename T> auto convertToOpenCLType(T &&x) {
	using no_ref = std::remove_reference_t<T>;
	if constexpr (is_multi_ptr_v<no_ref>) {
	return convertToOpenCLType(x.get_decorated());
	} else if constexpr (std::is_pointer_v<no_ref>) {
	// TODO: Below ignores volatile, but we didn't have a need for it yet.
	using elem_type = remove_decoration_t<std::remove_pointer_t<no_ref>>;
	using converted_elem_type_no_cv = decltype(convertToOpenCLType(
	std::declval<std::remove_const_t<elem_type>>()));
	using converted_elem_type =
	std::conditional_t<std::is_const_v<elem_type>,
	const converted_elem_type_no_cv,
	converted_elem_type_no_cv>;
	#ifdef __SYCL_DEVICE_ONLY__
	using result_type =
	typename DecoratedType<converted_elem_type,
	deduce_AS<no_ref>::value>::type *;
	#else
	using result_type = converted_elem_type *;
	#endif
	return reinterpret_cast<result_type>(x);
	} else if constexpr (is_vec_v<no_ref>) {
	using ElemTy = typename no_ref::element_type;
	using ConvertedElemTy =
	decltype(convertToOpenCLType(std::declval<ElemTy>()));
	static constexpr int NumElements = no_ref::size();
	#ifdef __SYCL_DEVICE_ONLY__
	using vector_t =
	std::conditional_t<NumElements == 1, ConvertedElemTy,
	ConvertedElemTy
	__attribute__((ext_vector_type(NumElements)))>;
	return sycl::bit_cast<vector_t>(x);
	#else
	// sycl::half may convert to _Float16, and we would try to
	// instantiate vec class with _Float16 DataType, which is not
	// expected there. As such, leave vector<half, N> as-is.
	using MatchingVec = vec<std::conditional_t<std::is_same_v<ElemTy, half>,
	ElemTy, ConvertedElemTy>,
	NumElements>;
	return x.template as<MatchingVec>();
	#endif
	#if (!defined(_HAS_STD_BYTE) || _HAS_STD_BYTE != 0)
	} else if constexpr (std::is_same_v<no_ref, std::byte>) {
	return static_cast<uint8_t>(x);
	#endif
	} else if constexpr (std::is_integral_v<no_ref>) {
	using OpenCLType = std::conditional_t<std::is_signed_v<no_ref>,
	fixed_width_signed<sizeof(no_ref)>,
	fixed_width_unsigned<sizeof(no_ref)>>;
	static_assert(sizeof(OpenCLType) == sizeof(T));
	return static_cast<OpenCLType>(x);
	} else if constexpr (std::is_same_v<no_ref, half>) {
	// Make it template-param-dependent to compile with incomplete `half`:
	using OpenCLType =
	std::enable_if_t<std::is_same_v<no_ref, half>,
	sycl::detail::half_impl::BIsRepresentationT>;
	static_assert(sizeof(OpenCLType) == sizeof(T));
	return static_cast<OpenCLType>(x);
	} else if constexpr (std::is_same_v<no_ref, ext::oneapi::bfloat16>) {
	// On host, don't interpret BF16 as uint16.
	#ifdef __SYCL_DEVICE_ONLY__
	return sycl::bit_cast<uint16_t>(x);
	#else
	return std::forward<T>(x);
	#endif
	} else if constexpr (std::is_floating_point_v<no_ref>) {
	static_assert(std::is_same_v<no_ref, float> ||
	std::is_same_v<no_ref, double>,
	"Other FP types are not expected/supported (yet?)");
	static_assert(std::is_same_v<float, sycl::opencl::cl_float> &&
	std::is_same_v<double, sycl::opencl::cl_double>);
	return std::forward<T>(x);
	} else {
	using OpenCLType = select_cl_scalar_complex_or_T_t<no_ref>;
	static_assert(sizeof(OpenCLType) == sizeof(T));
	return static_cast<OpenCLType>(x);
	}
	}

help here?

[steffenlarsen]

Would

llvm/sycl/include/sycl/detail/generic_type_traits.hpp

Lines 117 to 194 in 62397e8

	template <typename T> auto convertToOpenCLType(T &&x) {
	using no_ref = std::remove_reference_t<T>;
	if constexpr (is_multi_ptr_v<no_ref>) {
	return convertToOpenCLType(x.get_decorated());
	} else if constexpr (std::is_pointer_v<no_ref>) {
	// TODO: Below ignores volatile, but we didn't have a need for it yet.
	using elem_type = remove_decoration_t<std::remove_pointer_t<no_ref>>;
	using converted_elem_type_no_cv = decltype(convertToOpenCLType(
	std::declval<std::remove_const_t<elem_type>>()));
	using converted_elem_type =
	std::conditional_t<std::is_const_v<elem_type>,
	const converted_elem_type_no_cv,
	converted_elem_type_no_cv>;
	#ifdef __SYCL_DEVICE_ONLY__
	using result_type =
	typename DecoratedType<converted_elem_type,
	deduce_AS<no_ref>::value>::type *;
	#else
	using result_type = converted_elem_type *;
	#endif
	return reinterpret_cast<result_type>(x);
	} else if constexpr (is_vec_v<no_ref>) {
	using ElemTy = typename no_ref::element_type;
	using ConvertedElemTy =
	decltype(convertToOpenCLType(std::declval<ElemTy>()));
	static constexpr int NumElements = no_ref::size();
	#ifdef __SYCL_DEVICE_ONLY__
	using vector_t =
	std::conditional_t<NumElements == 1, ConvertedElemTy,
	ConvertedElemTy
	__attribute__((ext_vector_type(NumElements)))>;
	return sycl::bit_cast<vector_t>(x);
	#else
	// sycl::half may convert to _Float16, and we would try to
	// instantiate vec class with _Float16 DataType, which is not
	// expected there. As such, leave vector<half, N> as-is.
	using MatchingVec = vec<std::conditional_t<std::is_same_v<ElemTy, half>,
	ElemTy, ConvertedElemTy>,
	NumElements>;
	return x.template as<MatchingVec>();
	#endif
	#if (!defined(_HAS_STD_BYTE) || _HAS_STD_BYTE != 0)
	} else if constexpr (std::is_same_v<no_ref, std::byte>) {
	return static_cast<uint8_t>(x);
	#endif
	} else if constexpr (std::is_integral_v<no_ref>) {
	using OpenCLType = std::conditional_t<std::is_signed_v<no_ref>,
	fixed_width_signed<sizeof(no_ref)>,
	fixed_width_unsigned<sizeof(no_ref)>>;
	static_assert(sizeof(OpenCLType) == sizeof(T));
	return static_cast<OpenCLType>(x);
	} else if constexpr (std::is_same_v<no_ref, half>) {
	// Make it template-param-dependent to compile with incomplete `half`:
	using OpenCLType =
	std::enable_if_t<std::is_same_v<no_ref, half>,
	sycl::detail::half_impl::BIsRepresentationT>;
	static_assert(sizeof(OpenCLType) == sizeof(T));
	return static_cast<OpenCLType>(x);
	} else if constexpr (std::is_same_v<no_ref, ext::oneapi::bfloat16>) {
	// On host, don't interpret BF16 as uint16.
	#ifdef __SYCL_DEVICE_ONLY__
	return sycl::bit_cast<uint16_t>(x);
	#else
	return std::forward<T>(x);
	#endif
	} else if constexpr (std::is_floating_point_v<no_ref>) {
	static_assert(std::is_same_v<no_ref, float> ||
	std::is_same_v<no_ref, double>,
	"Other FP types are not expected/supported (yet?)");
	static_assert(std::is_same_v<float, sycl::opencl::cl_float> &&
	std::is_same_v<double, sycl::opencl::cl_double>);
	return std::forward<T>(x);
	} else {
	using OpenCLType = select_cl_scalar_complex_or_T_t<no_ref>;
	static_assert(sizeof(OpenCLType) == sizeof(T));
	return static_cast<OpenCLType>(x);
	}
	}

help here?

Interesting idea! Let's try!