Adds sycl::vec overloads to abs, cos, expm1, log, log1p, and sqrt

ndgrigorian · ndgrigorian · commit a0959d0d267c · 2023-11-24T21:08:08.000-08:00
diff --git a/dpctl/tensor/libtensor/include/kernels/elementwise_functions/abs.hpp b/dpctl/tensor/libtensor/include/kernels/elementwise_functions/abs.hpp
@@ -52,13 +52,15 @@ namespace py = pybind11;
 namespace td_ns = dpctl::tensor::type_dispatch;
 
 using dpctl::tensor::type_utils::is_complex;
+using dpctl::tensor::type_utils::vec_cast;
 
 template <typename argT, typename resT> struct AbsFunctor
 {
 
     using is_constant = typename std::false_type;
     // constexpr resT constant_value = resT{};
-    using supports_vec = typename std::false_type;
+    using supports_vec = typename std::negation<
+        std::disjunction<is_complex<resT>, is_complex<argT>>>;
     using supports_sg_loadstore = typename std::negation<
         std::disjunction<is_complex<resT>, is_complex<argT>>>;
 
@@ -127,6 +129,40 @@ template <typename argT, typename resT> struct AbsFunctor
 #endif
         }
     }
+
+    template <int vec_sz>
+    sycl::vec<resT, vec_sz> operator()(const sycl::vec<argT, vec_sz> &in)
+    {
+        if constexpr (std::is_integral<argT>::value) {
+            if constexpr (std::is_same_v<argT, bool> ||
+                          std::is_unsigned<argT>::value) {
+                return in;
+            }
+            else {
+                auto const &res_vec = sycl::abs(in);
+                using deducedT = typename std::remove_cv_t<
+                    std::remove_reference_t<decltype(res_vec)>>::element_type;
+                if constexpr (std::is_same_v<resT, deducedT>) {
+                    return res_vec;
+                }
+                else {
+
+                    return vec_cast<resT, deducedT, vec_sz>(res_vec);
+                }
+            }
+        }
+        else {
+            auto const &res_vec = sycl::fabs(in);
+            using deducedT = typename std::remove_cv_t<
+                std::remove_reference_t<decltype(res_vec)>>::element_type;
+            if constexpr (std::is_same_v<resT, deducedT>) {
+                return res_vec;
+            }
+            else {
+                return vec_cast<resT, deducedT, vec_sz>(res_vec);
+            }
+        }
+    }
 };
 
 template <typename argT,
diff --git a/dpctl/tensor/libtensor/include/kernels/elementwise_functions/cos.hpp b/dpctl/tensor/libtensor/include/kernels/elementwise_functions/cos.hpp
@@ -50,6 +50,7 @@ namespace py = pybind11;
 namespace td_ns = dpctl::tensor::type_dispatch;
 
 using dpctl::tensor::type_utils::is_complex;
+using dpctl::tensor::type_utils::vec_cast;
 
 template <typename argT, typename resT> struct CosFunctor
 {
@@ -59,7 +60,8 @@ template <typename argT, typename resT> struct CosFunctor
     // constant value, if constant
     // constexpr resT constant_value = resT{};
     // is function defined for sycl::vec
-    using supports_vec = typename std::false_type;
+    using supports_vec = typename std::negation<
+        std::disjunction<is_complex<resT>, is_complex<argT>>>;
     // do both argTy and resTy support sugroup store/load operation
     using supports_sg_loadstore = typename std::negation<
         std::disjunction<is_complex<resT>, is_complex<argT>>>;
@@ -165,6 +167,20 @@ template <typename argT, typename resT> struct CosFunctor
             return std::cos(in);
         }
     }
+
+    template <int vec_sz>
+    sycl::vec<resT, vec_sz> operator()(const sycl::vec<argT, vec_sz> &in)
+    {
+        auto const &res_vec = sycl::cos(in);
+        using deducedT = typename std::remove_cv_t<
+            std::remove_reference_t<decltype(res_vec)>>::element_type;
+        if constexpr (std::is_same_v<resT, deducedT>) {
+            return res_vec;
+        }
+        else {
+            return vec_cast<resT, deducedT, vec_sz>(res_vec);
+        }
+    }
 };
 
 template <typename argTy,
diff --git a/dpctl/tensor/libtensor/include/kernels/elementwise_functions/expm1.hpp b/dpctl/tensor/libtensor/include/kernels/elementwise_functions/expm1.hpp
@@ -51,6 +51,7 @@ namespace py = pybind11;
 namespace td_ns = dpctl::tensor::type_dispatch;
 
 using dpctl::tensor::type_utils::is_complex;
+using dpctl::tensor::type_utils::vec_cast;
 
 template <typename argT, typename resT> struct Expm1Functor
 {
@@ -60,7 +61,8 @@ template <typename argT, typename resT> struct Expm1Functor
     // constant value, if constant
     // constexpr resT constant_value = resT{};
     // is function defined for sycl::vec
-    using supports_vec = typename std::false_type;
+    using supports_vec = typename std::negation<
+        std::disjunction<is_complex<resT>, is_complex<argT>>>;
     // do both argTy and resTy support sugroup store/load operation
     using supports_sg_loadstore = typename std::negation<
         std::disjunction<is_complex<resT>, is_complex<argT>>>;
@@ -132,6 +134,20 @@ template <typename argT, typename resT> struct Expm1Functor
             return std::expm1(in);
         }
     }
+
+    template <int vec_sz>
+    sycl::vec<resT, vec_sz> operator()(const sycl::vec<argT, vec_sz> &in)
+    {
+        auto const &res_vec = sycl::expm1(in);
+        using deducedT = typename std::remove_cv_t<
+            std::remove_reference_t<decltype(res_vec)>>::element_type;
+        if constexpr (std::is_same_v<resT, deducedT>) {
+            return res_vec;
+        }
+        else {
+            return vec_cast<resT, deducedT, vec_sz>(res_vec);
+        }
+    }
 };
 
 template <typename argTy,
diff --git a/dpctl/tensor/libtensor/include/kernels/elementwise_functions/log.hpp b/dpctl/tensor/libtensor/include/kernels/elementwise_functions/log.hpp
@@ -51,6 +51,7 @@ namespace py = pybind11;
 namespace td_ns = dpctl::tensor::type_dispatch;
 
 using dpctl::tensor::type_utils::is_complex;
+using dpctl::tensor::type_utils::vec_cast;
 
 template <typename argT, typename resT> struct LogFunctor
 {
@@ -60,7 +61,8 @@ template <typename argT, typename resT> struct LogFunctor
     // constant value, if constant
     // constexpr resT constant_value = resT{};
     // is function defined for sycl::vec
-    using supports_vec = typename std::false_type;
+    using supports_vec = typename std::negation<
+        std::disjunction<is_complex<resT>, is_complex<argT>>>;
     // do both argTy and resTy support sugroup store/load operation
     using supports_sg_loadstore = typename std::negation<
         std::disjunction<is_complex<resT>, is_complex<argT>>>;
@@ -79,6 +81,20 @@ template <typename argT, typename resT> struct LogFunctor
             return std::log(in);
         }
     }
+
+    template <int vec_sz>
+    sycl::vec<resT, vec_sz> operator()(const sycl::vec<argT, vec_sz> &in)
+    {
+        auto const &res_vec = sycl::log(in);
+        using deducedT = typename std::remove_cv_t<
+            std::remove_reference_t<decltype(res_vec)>>::element_type;
+        if constexpr (std::is_same_v<resT, deducedT>) {
+            return res_vec;
+        }
+        else {
+            return vec_cast<resT, deducedT, vec_sz>(res_vec);
+        }
+    }
 };
 
 template <typename argTy,
diff --git a/dpctl/tensor/libtensor/include/kernels/elementwise_functions/log1p.hpp b/dpctl/tensor/libtensor/include/kernels/elementwise_functions/log1p.hpp
@@ -50,6 +50,7 @@ namespace py = pybind11;
 namespace td_ns = dpctl::tensor::type_dispatch;
 
 using dpctl::tensor::type_utils::is_complex;
+using dpctl::tensor::type_utils::vec_cast;
 
 // TODO: evaluate precision against alternatives
 template <typename argT, typename resT> struct Log1pFunctor
@@ -60,7 +61,8 @@ template <typename argT, typename resT> struct Log1pFunctor
     // constant value, if constant
     // constexpr resT constant_value = resT{};
     // is function defined for sycl::vec
-    using supports_vec = typename std::false_type;
+    using supports_vec = typename std::negation<
+        std::disjunction<is_complex<resT>, is_complex<argT>>>;
     // do both argTy and resTy support sugroup store/load operation
     using supports_sg_loadstore = typename std::negation<
         std::disjunction<is_complex<resT>, is_complex<argT>>>;
@@ -99,6 +101,20 @@ template <typename argT, typename resT> struct Log1pFunctor
             return std::log1p(in);
         }
     }
+
+    template <int vec_sz>
+    sycl::vec<resT, vec_sz> operator()(const sycl::vec<argT, vec_sz> &in)
+    {
+        auto const &res_vec = sycl::log1p(in);
+        using deducedT = typename std::remove_cv_t<
+            std::remove_reference_t<decltype(res_vec)>>::element_type;
+        if constexpr (std::is_same_v<resT, deducedT>) {
+            return res_vec;
+        }
+        else {
+            return vec_cast<resT, deducedT, vec_sz>(res_vec);
+        }
+    }
 };
 
 template <typename argTy,
diff --git a/dpctl/tensor/libtensor/include/kernels/elementwise_functions/sqrt.hpp b/dpctl/tensor/libtensor/include/kernels/elementwise_functions/sqrt.hpp
@@ -53,6 +53,7 @@ namespace py = pybind11;
 namespace td_ns = dpctl::tensor::type_dispatch;
 
 using dpctl::tensor::type_utils::is_complex;
+using dpctl::tensor::type_utils::vec_cast;
 
 template <typename argT, typename resT> struct SqrtFunctor
 {
@@ -62,7 +63,8 @@ template <typename argT, typename resT> struct SqrtFunctor
     // constant value, if constant
     // constexpr resT constant_value = resT{};
     // is function defined for sycl::vec
-    using supports_vec = typename std::false_type;
+    using supports_vec = typename std::negation<
+        std::disjunction<is_complex<resT>, is_complex<argT>>>;
     // do both argTy and resTy support sugroup store/load operation
     using supports_sg_loadstore = typename std::negation<
         std::disjunction<is_complex<resT>, is_complex<argT>>>;
@@ -263,6 +265,20 @@ template <typename argT, typename resT> struct SqrtFunctor
                    ? csqrt_finite_unscaled(x, y)
                    : csqrt_finite_scaled(x, y);
     }
+
+    template <int vec_sz>
+    sycl::vec<resT, vec_sz> operator()(const sycl::vec<argT, vec_sz> &in)
+    {
+        auto const &res_vec = sycl::sqrt(in);
+        using deducedT = typename std::remove_cv_t<
+            std::remove_reference_t<decltype(res_vec)>>::element_type;
+        if constexpr (std::is_same_v<resT, deducedT>) {
+            return res_vec;
+        }
+        else {
+            return vec_cast<resT, deducedT, vec_sz>(res_vec);
+        }
+    }
 };
 
 template <typename argTy,
