[ESIMD] Enable some of intrinsic math ops on wrapper element types.

- Enable
  * inv, log, exp, sqrt, sqrt_ieee, rsqrt, sin, cos, pow, saturation
  on wrapper element types (sycl::half at the moment)
- Fix host implementation to support wrapper types
- TODO some intrinsics like rounding, inverse trigonometric functions
  are still not supported - will cause compile-time error

Signed-off-by: Konstantin S Bobrovsky <[email protected]>

Author: kbobrovs

sycl/include/CL/sycl/builtins_esimd.hpp

@@ -28,7 +28,7 @@ cos(__ESIMD_NS::simd<float, SZ> x) __NOEXC {
 #ifdef __SYCL_DEVICE_ONLY__
   return __ESIMD_NS::detail::ocl_cos<SZ>(x.data());
 #else
-  return __esimd_cos<SZ>(x.data());
+  return __esimd_cos<float, SZ>(x.data());
 #endif // __SYCL_DEVICE_ONLY__
 }
 
@@ -39,7 +39,7 @@ sin(__ESIMD_NS::simd<float, SZ> x) __NOEXC {
 #ifdef __SYCL_DEVICE_ONLY__
   return __ESIMD_NS::detail::ocl_sin<SZ>(x.data());
 #else
-  return __esimd_sin<SZ>(x.data());
+  return __esimd_sin<float, SZ>(x.data());
 #endif // __SYCL_DEVICE_ONLY__
 }
 
@@ -50,7 +50,7 @@ exp(__ESIMD_NS::simd<float, SZ> x) __NOEXC {
 #ifdef __SYCL_DEVICE_ONLY__
   return __ESIMD_NS::detail::ocl_exp<SZ>(x.data());
 #else
-  return __esimd_exp<SZ>(x.data());
+  return __esimd_exp<float, SZ>(x.data());
 #endif // __SYCL_DEVICE_ONLY__
 }
 
@@ -61,7 +61,7 @@ log(__ESIMD_NS::simd<float, SZ> x) __NOEXC {
 #ifdef __SYCL_DEVICE_ONLY__
   return __ESIMD_NS::detail::ocl_log<SZ>(x.data());
 #else
-  return __esimd_log<SZ>(x.data());
+  return __esimd_log<float, SZ>(x.data());
 #endif // __SYCL_DEVICE_ONLY__
 }
 

sycl/include/sycl/ext/intel/experimental/esimd/detail/elem_type_traits.hpp

@@ -133,6 +133,10 @@ template <class T, class SFINAE> struct element_type_traits {
   // Whether a value or clang vector value the raw element type can be used
   // directly as operand to std C++ operations.
   static inline constexpr bool use_native_cpp_ops = true;
+  // W/A for MSVC compiler problems which thinks
+  // std::is_floating_point_v<_Float16> is false; so require new element types
+  // implementations to state "is floating point" trait explicitly
+  static inline constexpr bool is_floating_point = false;
 };
 
 // Element type traits specialization for C++ standard element type.
@@ -141,8 +145,19 @@ struct element_type_traits<T, std::enable_if_t<is_vectorizable_v<T>>> {
   using RawT = T;
   using EnclosingCppT = T;
   static inline constexpr bool use_native_cpp_ops = true;
+  static inline constexpr bool is_floating_point = std::is_floating_point_v<T>;
 };
 
+#ifdef __SYCL_DEVICE_ONLY__
+template <> struct element_type_traits<_Float16, void> {
+  using RawT = _Float16;
+  using EnclosingCppT = _Float16;
+  __SYCL_DEPRECATED("use sycl::half as element type")
+  static inline constexpr bool use_native_cpp_ops = true;
+  static inline constexpr bool is_floating_point = true;
+};
+#endif
+
 // --- Type conversions
 
 // Low-level conversion functions to and from a wrapper element type.
@@ -563,7 +578,7 @@ class WrapperElementTypeProxy {
 // the wrapper floating-point types such as sycl::half.
 template <typename T>
 static inline constexpr bool is_generic_floating_point_v =
-    std::is_floating_point_v<typename element_type_traits<T>::EnclosingCppT>;
+    element_type_traits<T>::is_floating_point;
 
 // @{
 // Get computation type of a binary operator given its operand types:
@@ -664,6 +679,8 @@ struct element_type_traits<T, std::enable_if_t<std::is_same_v<T, sycl::half>>> {
   // operations on half type.
   static inline constexpr bool use_native_cpp_ops = false;
 #endif // __SYCL_DEVICE_ONLY__
+
+  static inline constexpr bool is_floating_point = true;
 };
 
 using half_raw = __raw_t<sycl::half>;

sycl/include/sycl/ext/intel/experimental/esimd/detail/host_util.hpp

@@ -12,6 +12,7 @@
 
 #ifndef __SYCL_DEVICE_ONLY__
 
+#include <assert.h>
 #include <limits>
 
 #define SIMDCF_ELEMENT_SKIP(i)
@@ -46,7 +47,12 @@ static long long abs(long long a) {
   }
 }
 
-template <typename RT> struct satur {
+template <typename RT, class SFINAE = void> struct satur;
+
+template <typename RT>
+struct satur<RT, std::enable_if_t<std::is_integral_v<RT>>> {
+  static_assert(!__SEIEED::is_wrapper_elem_type_v<RT>);
+
   template <typename T> static RT saturate(const T val, const int flags) {
     if ((flags & sat_is_on) == 0) {
       return (RT)val;
@@ -72,35 +78,29 @@ template <typename RT> struct satur {
   }
 };
 
-template <> struct satur<float> {
-  template <typename T> static float saturate(const T val, const int flags) {
-    if ((flags & sat_is_on) == 0) {
-      return (float)val;
+// Host implemenation of saturation for FP types, including non-standarad
+// wrapper types such as sycl::half. Template parameters are defined in terms
+// of user-level types (sycl::half), function parameter and return types -
+// in terms of raw bit representation type(_Float16 for half on device).
+template <class Tdst>
+struct satur<Tdst,
+             std::enable_if_t<__SEIEED::is_generic_floating_point_v<Tdst>>> {
+  template <typename Tsrc>
+  static __SEIEED::__raw_t<Tdst> saturate(const __SEIEED::__raw_t<Tsrc> raw_src,
+                                          const int flags) {
+    Tsrc src = __SEIEED::bitcast_to_wrapper_type<Tsrc>(raw_src);
+
+    // perform comparison on user type!
+    if ((flags & sat_is_on) == 0 || (src >= 0 && src <= 1)) {
+      // convert_scalar accepts/returns user types - need to bitcast
+      Tdst dst = __SEIEED::convert_scalar<Tdst, Tsrc>(src);
+      return __SEIEED::bitcast_to_raw_type<Tdst>(dst);
     }
-
-    if (val < 0.) {
-      return 0;
-    } else if (val > 1.) {
-      return 1.;
-    } else {
-      return (float)val;
-    }
-  }
-};
-
-template <> struct satur<double> {
-  template <typename T> static double saturate(const T val, const int flags) {
-    if ((flags & sat_is_on) == 0) {
-      return (double)val;
-    }
-
-    if (val < 0.) {
-      return 0;
-    } else if (val > 1.) {
-      return 1.;
-    } else {
-      return (double)val;
+    if (src < 0) {
+      return __SEIEED::bitcast_to_raw_type<Tdst>(Tdst{0});
     }
+    assert(src > 1);
+    return __SEIEED::bitcast_to_raw_type<Tdst>(Tdst{1});
   }
 };
 
@@ -116,6 +116,10 @@ template <> struct SetSatur<double, true> {
   static unsigned int set() { return sat_is_on; }
 };
 
+// TODO replace restype_ex with detail::computation_type_t and represent half
+// as sycl::half rather than 'using half = sycl::detail::half_impl::half;'
+// above
+
 // used for intermediate type in dp4a emulation
 template <typename T1, typename T2> struct restype_ex {
 private:
@@ -430,36 +434,6 @@ template <> struct fptype<float> { static const bool value = true; };
 template <typename T> struct dftype { static const bool value = false; };
 template <> struct dftype<double> { static const bool value = true; };
 
-template <typename T> struct esimdtype;
-template <> struct esimdtype<char> { static const bool value = true; };
-
-template <> struct esimdtype<signed char> { static const bool value = true; };
-
-template <> struct esimdtype<unsigned char> { static const bool value = true; };
-
-template <> struct esimdtype<short> { static const bool value = true; };
-
-template <> struct esimdtype<unsigned short> {
-  static const bool value = true;
-};
-template <> struct esimdtype<int> { static const bool value = true; };
-
-template <> struct esimdtype<unsigned int> { static const bool value = true; };
-
-template <> struct esimdtype<unsigned long> { static const bool value = true; };
-
-template <> struct esimdtype<half> { static const bool value = true; };
-
-template <> struct esimdtype<float> { static const bool value = true; };
-
-template <> struct esimdtype<double> { static const bool value = true; };
-
-template <> struct esimdtype<long long> { static const bool value = true; };
-
-template <> struct esimdtype<unsigned long long> {
-  static const bool value = true;
-};
-
 template <typename T> struct bytetype;
 template <> struct bytetype<char> { static const bool value = true; };
 template <> struct bytetype<unsigned char> { static const bool value = true; };