[SYCL][CUDA] add bf16 builtins

Author: t4c1

sycl/include/CL/__spirv/spirv_ops.hpp

@@ -755,6 +755,28 @@ __spirv_ocl_printf(const __attribute__((opencl_constant)) char *Format, ...);
 extern SYCL_EXTERNAL int __spirv_ocl_printf(const char *Format, ...);
 #endif
 
+extern SYCL_EXTERNAL __SYCL_EXPORT uint16_t __clc_fabs(uint16_t) noexcept;
+
+#define __CLC_BF16(...) \
+extern SYCL_EXTERNAL __SYCL_EXPORT __VA_ARGS__ __clc_fabs(__VA_ARGS__) noexcept; \
+extern SYCL_EXTERNAL __SYCL_EXPORT __VA_ARGS__ __clc_fmin(__VA_ARGS__, __VA_ARGS__) noexcept; \
+extern SYCL_EXTERNAL __SYCL_EXPORT __VA_ARGS__ __clc_fmax(__VA_ARGS__, __VA_ARGS__) noexcept; \
+extern SYCL_EXTERNAL __SYCL_EXPORT __VA_ARGS__ __clc_fma(__VA_ARGS__, __VA_ARGS__, __VA_ARGS__) noexcept;
+
+#define __CLC_BF16_SCAL_VEC(TYPE) \
+__CLC_BF16(TYPE) \
+__CLC_BF16(__ocl_vec_t<TYPE, 2>) \
+__CLC_BF16(__ocl_vec_t<TYPE, 3>) \
+__CLC_BF16(__ocl_vec_t<TYPE, 4>) \
+__CLC_BF16(__ocl_vec_t<TYPE, 8>) \
+__CLC_BF16(__ocl_vec_t<TYPE, 16>)
+
+__CLC_BF16_SCAL_VEC(uint16_t)
+__CLC_BF16_SCAL_VEC(uint32_t)
+
+#undef __CLC_BF16_SCAL_VEC
+#undef __CLC_BF16
+
 #else // if !__SYCL_DEVICE_ONLY__
 
 template <typename dataT>

sycl/include/CL/sycl.hpp

@@ -60,6 +60,7 @@
 #if SYCL_EXT_ONEAPI_BACKEND_LEVEL_ZERO
 #include <sycl/ext/oneapi/backend/level_zero.hpp>
 #endif
+#include <sycl/ext/oneapi/bf16_storage_builtins.hpp>
 #include <sycl/ext/oneapi/experimental/builtins.hpp>
 #include <sycl/ext/oneapi/filter_selector.hpp>
 #include <sycl/ext/oneapi/group_algorithm.hpp>

sycl/include/sycl/ext/oneapi/bf16_storage_builtins.hpp

@@ -0,0 +1,79 @@
+#pragma once
+
+#include <CL/__spirv/spirv_ops.hpp>
+#include <CL/sycl/builtins.hpp>
+#include <CL/sycl/detail/builtins.hpp>
+#include <CL/sycl/detail/generic_type_lists.hpp>
+#include <CL/sycl/detail/generic_type_traits.hpp>
+#include <CL/sycl/detail/type_traits.hpp>
+
+__SYCL_INLINE_NAMESPACE(cl) {
+namespace sycl {
+namespace ext {
+namespace oneapi {
+
+namespace detail {
+
+template <typename T> struct is_bf16_storage_type {
+  static constexpr int value = false;
+};
+
+template <> struct is_bf16_storage_type<uint16_t> {
+  static constexpr int value = true;
+};
+
+template <> struct is_bf16_storage_type<uint32_t> {
+  static constexpr int value = true;
+};
+
+template <int N> struct is_bf16_storage_type<vec<uint16_t, N>> {
+  static constexpr int value = true;
+};
+
+template <int N> struct is_bf16_storage_type<vec<uint32_t, N>> {
+  static constexpr int value = true;
+};
+
+} // namespace detail
+
+template <typename T>
+std::enable_if_t<detail::is_bf16_storage_type<T>::value, T> fabs(T x) {
+#ifdef __SYCL_DEVICE_ONLY__
+  return __clc_fabs(x);
+#else
+  throw runtime_error("bf16 is not supported on host device.",
+                      PI_INVALID_DEVICE);
+#endif
+}
+template <typename T>
+std::enable_if_t<detail::is_bf16_storage_type<T>::value, T> fmin(T x, T y) {
+#ifdef __SYCL_DEVICE_ONLY__
+  return __clc_fmin(x, y);
+#else
+  throw runtime_error("bf16 is not supported on host device.",
+                      PI_INVALID_DEVICE);
+#endif
+}
+template <typename T>
+std::enable_if_t<detail::is_bf16_storage_type<T>::value, T> fmax(T x, T y) {
+#ifdef __SYCL_DEVICE_ONLY__
+  return __clc_fmax(x, y);
+#else
+  throw runtime_error("bf16 is not supported on host device.",
+                      PI_INVALID_DEVICE);
+#endif
+}
+template <typename T>
+std::enable_if_t<detail::is_bf16_storage_type<T>::value, T> fma(T x, T y, T z) {
+#ifdef __SYCL_DEVICE_ONLY__
+  return __clc_fma(x, y, z);
+#else
+  throw runtime_error("bf16 is not supported on host device.",
+                      PI_INVALID_DEVICE);
+#endif
+}
+
+} // namespace oneapi
+} // namespace ext
+} // namespace sycl
+} // __SYCL_INLINE_NAMESPACE(cl)