[SYCL][COMPAT] Add vectorized_ternary and vectorized_with_pred. (#15550)

Add vectorized_ternary and vectorized_with_pred.
Update relu and vectorized_binary.

Signed-off-by: Tang, Jiajun jiajun.tang@intel.com

---------

Signed-off-by: Tang, Jiajun jiajun.tang@intel.com

Author: tangjj11

sycl/doc/syclcompat/README.md

@@ -1906,17 +1906,11 @@ template <typename ValueT, typename ValueU>
 inline typename std::enable_if_t<!std::is_floating_point_v<ValueT>, double>
 pow(const ValueT a, const ValueU b);
 
-template <typename ValueT>
-inline std::enable_if_t<std::is_floating_point_v<ValueT> ||
-                            std::is_same_v<sycl::half, ValueT>,
-                        ValueT>
-relu(const ValueT a);
+template <typename ValueT> inline ValueT relu(const ValueT a);
 
-template <class ValueT>
-inline std::enable_if_t<std::is_floating_point_v<ValueT> ||
-                            std::is_same_v<sycl::half, ValueT>,
-                        sycl::vec<ValueT, 2>>
-relu(const sycl::vec<ValueT, 2> a);
+template <class ValueT, int NumElements>
+inline sycl::vec<ValueT, NumElements>
+relu(const sycl::vec<ValueT, NumElements> a);
 
 template <class ValueT>
 inline std::enable_if_t<std::is_floating_point_v<ValueT> ||
@@ -2019,9 +2013,12 @@ inline dot_product_acc_t<T1, T2> dp4a(T1 a, T2 b,
 
 `vectorized_binary` computes the `BinaryOperation` for two operands,
 with each value treated as a vector type. `vectorized_unary` offers the same
-interface for operations with a single operand.
+interface for operations with a single operand. `vectorized_ternary` offers the
+interface for three operands with two `BinaryOperation`.
 The implemented `BinaryOperation`s are `abs_diff`, `add_sat`, `rhadd`, `hadd`,
 `maximum`, `minimum`, and `sub_sat`.
+And the `vectorized_with_pred` offers the `BinaryOperation` for two operands,
+meanwihle provides the pred of high/low halfword operation.
 
 ```cpp
 namespace syclcompat {
@@ -2036,7 +2033,19 @@ struct abs {
 
 template <typename VecT, class BinaryOperation>
 inline unsigned vectorized_binary(unsigned a, unsigned b,
-                                  const BinaryOperation binary_op);
+                                  const BinaryOperation binary_op,
+                                  bool need_relu = false);
+
+template <typename VecT, typename BinaryOperation1, typename BinaryOperation2>
+inline unsigned vectorized_ternary(unsigned a, unsigned b, unsigned c,
+                                   const BinaryOperation1 binary_op1,
+                                   const BinaryOperation2 binary_op2,
+                                   bool need_relu = false);
+
+template <typename ValueT, typename BinaryOperation>
+inline unsigned vectorized_with_pred(unsigned a, unsigned b,
+                                     const BinaryOperation binary_op,
+                                     bool *pred_hi, bool *pred_lo);
 
 // A sycl::abs_diff wrapper functor.
 struct abs_diff {
@@ -2062,11 +2071,15 @@ struct hadd {
 struct maximum {
   template <typename ValueT>
   auto operator()(const ValueT x, const ValueT y) const;
+  template <typename ValueT>
+  auto operator()(const ValueT x, const ValueT y, bool *pred) const;
 };
 // A sycl::min wrapper functor.
 struct minimum {
   template <typename ValueT>
   auto operator()(const ValueT x, const ValueT y) const;
+  template <typename ValueT>
+  auto operator()(const ValueT x, const ValueT y, bool *pred) const;
 };
 // A sycl::sub_sat wrapper functor.
 struct sub_sat {

sycl/include/syclcompat/math.hpp

@@ -856,23 +856,24 @@ pow(const ValueT a, const ValueU b) {
 /// Performs relu saturation.
 /// \param [in] a The input value
 /// \returns the relu saturation result
-template <typename ValueT>
-inline std::enable_if_t<syclcompat::is_floating_point_v<ValueT>, ValueT>
-relu(const ValueT a) {
-  if (!detail::isnan(a) && a < ValueT(0))
+template <typename ValueT> inline ValueT relu(const ValueT a) {
+  if constexpr (syclcompat::is_floating_point_v<ValueT>)
+    if (detail::isnan(a))
+      return a;
+  if (a < ValueT(0))
     return ValueT(0);
   return a;
 }
-template <class ValueT>
-inline std::enable_if_t<syclcompat::is_floating_point_v<ValueT>,
-                        sycl::vec<ValueT, 2>>
-relu(const sycl::vec<ValueT, 2> a) {
-  return {relu(a[0]), relu(a[1])};
+template <class ValueT, int NumElements>
+inline sycl::vec<ValueT, NumElements>
+relu(const sycl::vec<ValueT, NumElements> a) {
+  sycl::vec<ValueT, NumElements> ret;
+  for (int i = 0; i < NumElements; ++i)
+    ret[i] = relu(a[i]);
+  return ret;
 }
 template <class ValueT>
-inline std::enable_if_t<syclcompat::is_floating_point_v<ValueT>,
-                        sycl::marray<ValueT, 2>>
-relu(const sycl::marray<ValueT, 2> a) {
+inline sycl::marray<ValueT, 2> relu(const sycl::marray<ValueT, 2> a) {
   return {relu(a[0]), relu(a[1])};
 }
 
@@ -990,6 +991,10 @@ struct maximum {
   auto operator()(const ValueT x, const ValueT y) const {
     return sycl::max(x, y);
   }
+  template <typename ValueT>
+  auto operator()(const ValueT x, const ValueT y, bool *pred) const {
+    return (x >= y) ? ((*pred = true), x) : ((*pred = false), y);
+  }
 };
 
 /// A sycl::min wrapper functors.
@@ -998,6 +1003,10 @@ struct minimum {
   auto operator()(const ValueT x, const ValueT y) const {
     return sycl::min(x, y);
   }
+  template <typename ValueT>
+  auto operator()(const ValueT x, const ValueT y, bool *pred) const {
+    return (x <= y) ? ((*pred = true), x) : ((*pred = false), y);
+  }
 };
 
 /// A sycl::sub_sat wrapper functors.
@@ -1037,19 +1046,76 @@ struct average {
 /// \tparam [in] BinaryOperation The binary operation class
 /// \param [in] a The first value
 /// \param [in] b The second value
+/// \param [in] binary_op The operation to do with the two values
+/// \param [in] need_relu Whether the result need relu saturation
 /// \returns The vectorized binary operation value of the two values
 template <typename VecT, class BinaryOperation>
 inline unsigned vectorized_binary(unsigned a, unsigned b,
-                                  const BinaryOperation binary_op) {
+                                  const BinaryOperation binary_op,
+                                  bool need_relu = false) {
   sycl::vec<unsigned, 1> v0{a}, v1{b};
   auto v2 = v0.as<VecT>();
   auto v3 = v1.as<VecT>();
   auto v4 =
       detail::vectorized_binary<VecT, BinaryOperation>()(v2, v3, binary_op);
+  if (need_relu)
+    v4 = relu(v4);
   v0 = v4.template as<sycl::vec<unsigned, 1>>();
   return v0;
 }
 
+/// Compute two vectorized binary operation value with pred for three values,
+/// with each value treated as a 2 \p T type elements vector type.
+///
+/// \tparam [in] VecT The type of the vector
+/// \tparam [in] BinaryOperation1 The first binary operation class
+/// \tparam [in] BinaryOperation2 The second binary operation class
+/// \param [in] a The first value
+/// \param [in] b The second value
+/// \param [in] c The third value
+/// \param [in] binary_op1 The first operation to do with the first two values
+/// \param [in] binary_op2 The second operation to do with the third values
+/// \param [in] need_relu Whether the result need relu saturation
+/// \returns The two vectorized binary operation value of the three values
+template <typename VecT, typename BinaryOperation1, typename BinaryOperation2>
+inline unsigned vectorized_ternary(unsigned a, unsigned b, unsigned c,
+                                   const BinaryOperation1 binary_op1,
+                                   const BinaryOperation2 binary_op2,
+                                   bool need_relu = false) {
+  const auto v1 = sycl::vec<unsigned, 1>(a).as<VecT>();
+  const auto v2 = sycl::vec<unsigned, 1>(b).as<VecT>();
+  const auto v3 = sycl::vec<unsigned, 1>(c).as<VecT>();
+  auto v4 =
+      detail::vectorized_binary<VecT, BinaryOperation1>()(v1, v2, binary_op1);
+  v4 = detail::vectorized_binary<VecT, BinaryOperation2>()(v4, v3, binary_op2);
+  if (need_relu)
+    v4 = relu(v4);
+  return v4.template as<sycl::vec<unsigned, 1>>();
+}
+
+/// Compute vectorized binary operation value with pred for two values, with
+/// each value treated as a 2 \p T type elements vector type.
+///
+/// \tparam [in] VecT The type of the vector
+/// \tparam [in] BinaryOperation The binary operation class
+/// \param [in] a The first value
+/// \param [in] b The second value
+/// \param [in] binary_op The operation with pred to do with the two values
+/// \param [out] pred_hi The pred pointer that pass into high halfword operation
+/// \param [out] pred_lo The pred pointer that pass into low halfword operation
+/// \returns The vectorized binary operation value of the two values
+template <typename VecT, typename BinaryOperation>
+inline unsigned vectorized_binary_with_pred(unsigned a, unsigned b,
+                                            const BinaryOperation binary_op,
+                                            bool *pred_hi, bool *pred_lo) {
+  auto v1 = sycl::vec<unsigned, 1>(a).as<VecT>();
+  auto v2 = sycl::vec<unsigned, 1>(b).as<VecT>();
+  VecT ret;
+  ret[0] = binary_op(v1[0], v2[0], pred_lo);
+  ret[1] = binary_op(v1[1], v2[1], pred_hi);
+  return ret.template as<sycl::vec<unsigned, 1>>();
+}
+
 template <typename T1, typename T2>
 using dot_product_acc_t =
     std::conditional_t<std::is_unsigned_v<T1> && std::is_unsigned_v<T2>,

sycl/test-e2e/syclcompat/math/math_fixt.hpp

@@ -134,6 +134,8 @@ class BinaryOpTestLauncher : OpTestLauncher {
   ValueT *op1_;
   ValueU *op2_;
   ResultT res_h_, *res_;
+  bool *res_hi_;
+  bool *res_lo_;
 
 public:
   BinaryOpTestLauncher(const syclcompat::dim3 &grid,
@@ -147,6 +149,8 @@ class BinaryOpTestLauncher : OpTestLauncher {
     op1_ = syclcompat::malloc<ValueT>(data_size);
     op2_ = syclcompat::malloc<ValueU>(data_size);
     res_ = syclcompat::malloc<ResultT>(data_size);
+    res_hi_ = syclcompat::malloc<bool>(1);
+    res_lo_ = syclcompat::malloc<bool>(1);
   };
 
   virtual ~BinaryOpTestLauncher() {
@@ -155,6 +159,8 @@ class BinaryOpTestLauncher : OpTestLauncher {
     syclcompat::free(op1_);
     syclcompat::free(op2_);
     syclcompat::free(res_);
+    syclcompat::free(res_hi_);
+    syclcompat::free(res_lo_);
   }
 
   template <auto Kernel>
@@ -169,6 +175,37 @@ class BinaryOpTestLauncher : OpTestLauncher {
 
     CHECK(ResultT, res_h_, expected);
   };
+  template <auto Kernel>
+  void launch_test(ValueT op1, ValueU op2, ResultT expected, bool need_relu) {
+    if (skip_)
+      return;
+    syclcompat::memcpy<ValueT>(op1_, &op1, data_size_);
+    syclcompat::memcpy<ValueU>(op2_, &op2, data_size_);
+    syclcompat::launch<Kernel>(grid_, threads_, op1_, op2_, res_, need_relu);
+    syclcompat::wait();
+    syclcompat::memcpy<ResultT>(&res_h_, res_, data_size_);
+
+    CHECK(ResultT, res_h_, expected);
+  };
+  template <auto Kernel>
+  void launch_test(ValueT op1, ValueU op2, ResultT expected, bool expected_hi,
+                   bool expected_lo) {
+    if (skip_)
+      return;
+    syclcompat::memcpy<ValueT>(op1_, &op1, data_size_);
+    syclcompat::memcpy<ValueU>(op2_, &op2, data_size_);
+    syclcompat::launch<Kernel>(grid_, threads_, op1_, op2_, res_, res_hi_,
+                               res_lo_);
+    syclcompat::wait();
+    syclcompat::memcpy<ResultT>(&res_h_, res_, data_size_);
+    bool res_hi_h_, res_lo_h_;
+    syclcompat::memcpy<bool>(&res_hi_h_, res_hi_, 1);
+    syclcompat::memcpy<bool>(&res_lo_h_, res_lo_, 1);
+
+    CHECK(ResultT, res_h_, expected);
+    assert(res_hi_h_ == expected_hi);
+    assert(res_lo_h_ == expected_lo);
+  };
 };
 
 template <typename ValueT, typename ResultT = ValueT>
@@ -208,3 +245,54 @@ class UnaryOpTestLauncher : OpTestLauncher {
     CHECK(ResultT, res_h_, expected);
   }
 };
+
+// Templated ResultT to support both arithmetic and boolean operators
+template <typename ValueT, typename ValueU, typename ValueV,
+          typename ResultT = std::common_type_t<ValueT, ValueU, ValueV>>
+class TernaryOpTestLauncher : OpTestLauncher {
+protected:
+  ValueT *op1_;
+  ValueU *op2_;
+  ValueV *op3_;
+  ResultT res_h_, *res_;
+
+public:
+  TernaryOpTestLauncher(const syclcompat::dim3 &grid,
+                        const syclcompat::dim3 &threads,
+                        const size_t data_size = 1)
+      : OpTestLauncher{grid, threads, data_size,
+                       should_skip<ValueT, ValueU, ValueV, ResultT>()(
+                           syclcompat::get_current_device())} {
+    if (skip_)
+      return;
+    op1_ = syclcompat::malloc<ValueT>(data_size);
+    op2_ = syclcompat::malloc<ValueU>(data_size);
+    op3_ = syclcompat::malloc<ValueV>(data_size);
+    res_ = syclcompat::malloc<ResultT>(data_size);
+  };
+
+  virtual ~TernaryOpTestLauncher() {
+    if (skip_)
+      return;
+    syclcompat::free(op1_);
+    syclcompat::free(op2_);
+    syclcompat::free(op3_);
+    syclcompat::free(res_);
+  }
+
+  template <auto Kernel>
+  void launch_test(ValueT op1, ValueU op2, ValueV op3, ResultT expected,
+                   bool need_relu = false) {
+    if (skip_)
+      return;
+    syclcompat::memcpy<ValueT>(op1_, &op1, data_size_);
+    syclcompat::memcpy<ValueU>(op2_, &op2, data_size_);
+    syclcompat::memcpy<ValueV>(op3_, &op3, data_size_);
+    syclcompat::launch<Kernel>(grid_, threads_, op1_, op2_, op3_, res_,
+                               need_relu);
+    syclcompat::wait();
+    syclcompat::memcpy<ResultT>(&res_h_, res_, data_size_);
+
+    CHECK(ResultT, res_h_, expected);
+  };
+};

sycl/test-e2e/syclcompat/math/math_ops.cpp

@@ -226,8 +226,10 @@ template <typename ValueT> void test_syclcompat_relu() {
   UnaryOpTestLauncher<ValueT>(grid, threads)
       .template launch_test<relu_kernel<ValueT>>(op1, res1);
 
-  const ValueT op2 = static_cast<ValueT>(-3);
-  const ValueT res2 = static_cast<ValueT>(0);
+  const ValueT op2 = std::is_signed_v<ValueT> ? static_cast<ValueT>(-3)
+                                              : static_cast<ValueT>(2);
+  const ValueT res2 = std::is_signed_v<ValueT> ? static_cast<ValueT>(0)
+                                               : static_cast<ValueT>(2);
   UnaryOpTestLauncher<ValueT>(grid, threads)
       .template launch_test<relu_kernel<ValueT>>(op2, res2);
 
@@ -374,7 +376,7 @@ int main() {
   test_syclcompat_pow<float, int>();
   test_syclcompat_pow<double, int>();
 
-  INSTANTIATE_ALL_TYPES(fp_type_list, test_syclcompat_relu);
+  INSTANTIATE_ALL_TYPES(value_type_list, test_syclcompat_relu);
   INSTANTIATE_ALL_TYPES(fp_type_list_no_bfloat16, test_syclcompat_cbrt);
 
   INSTANTIATE_ALL_CONTAINER_TYPES(fp_type_list, sycl::vec, test_isnan);