COMPMID-2269: Implement POW operator for NEON

Change-Id: I7135f665d89da3c24c9bbe00e991a64713a41d0e
Signed-off-by: Usama Arif <usama.arif@arm.com>
Reviewed-on: https://review.mlplatform.org/c/1128
Reviewed-by: Michalis Spyrou <michalis.spyrou@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>

8 files changed, 341 insertions, 30 deletions

diff --git a/arm_compute/core/NEON/kernels/NEElementwiseOperationKernel.h b/arm_compute/core/NEON/kernels/NEElementwiseOperationKernel.h
index 351b346c5a..195fd8951d 100644
--- a/arm_compute/core/NEON/kernels/NEElementwiseOperationKernel.h
+++ b/arm_compute/core/NEON/kernels/NEElementwiseOperationKernel.h
@@ -151,6 +151,35 @@ protected:
     static Status validate_arguments(const ITensorInfo &input1, const ITensorInfo &input2, const ITensorInfo &output);
 };
 
+class NEPowerOperationKernel : public NEArithmeticOperationKernel
+{
+public:
+    /** Default constructor */
+    NEPowerOperationKernel() = default;
+
+    /** Static function to check if given info will lead to a valid configuration of @ref NEArithmeticOperationKernel
+     *
+     * @param[in]  input1 First tensor input. Data types supported: F16/F32.
+     * @param[in]  input2 Second tensor input. Data types supported: Same as @p input1.
+     * @param[out] output Output tensor. Data types supported: Same as @p input1.
+     */
+    void configure(const ITensor *input1, const ITensor *input2, ITensor *output);
+
+    /** Static function to check if given info will lead to a valid configuration of @ref NEArithmeticOperationKernel
+     *
+     * @param[in]  input1 First tensor input info. Data types supported: F16/F32.
+     * @param[in]  input2 Second tensor input info. Data types supported: Same as @p input1.
+     * @param[out] output Output tensor info. Data types supported: Same as @p input1.
+     *
+     * @return a Status
+     */
+    static Status validate(const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output);
+
+protected:
+    // Inherited methods overridden:
+    static Status validate_arguments(const ITensorInfo &input1, const ITensorInfo &input2, const ITensorInfo &output);
+};
+
 class NEComparisonOperationKernel : public NEElementwiseOperationKernel
 {
 public:
diff --git a/arm_compute/core/Types.h b/arm_compute/core/Types.h
index 9ecd7ff2e1..99ec44d5e0 100644
--- a/arm_compute/core/Types.h
+++ b/arm_compute/core/Types.h
@@ -571,6 +571,7 @@ enum class ArithmeticOperation
     MIN,          /**< Min(x, y) */
     MAX,          /**< Max(x, y) */
     SQUARED_DIFF, /**< (x - y)^2 */
+    POWER,        /**< x ^ y */
 };
 
 /** Available element wise unary operations */
diff --git a/arm_compute/runtime/NEON/functions/NEElementwiseOperations.h b/arm_compute/runtime/NEON/functions/NEElementwiseOperations.h
index ca3717a709..586546c8bc 100644
--- a/arm_compute/runtime/NEON/functions/NEElementwiseOperations.h
+++ b/arm_compute/runtime/NEON/functions/NEElementwiseOperations.h
@@ -135,6 +135,33 @@ public:
     static Status validate(const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output);
 };
 
+/** Basic function to run @ref NEArithmeticOperationKernel for power
+ *
+ * @note The tensor data type for the inputs must be F16/F32.
+ * @note The function performs a elementwise power of in1 to in2 (i.e., out[i] = in1[i] ^ in2[i])
+ * @note For an exponent that is a float, this function will only work with a positive base.
+ */
+class NEElementwisePower : public INESimpleFunction
+{
+public:
+    /** Initialise the kernel's inputs, output and conversion policy.
+     *
+     * @param[in, out] input1 First tensor input. Data types supported: F16/F32.
+     * @param[in, out] input2 Second tensor input. Data types supported: Same as @p input1.
+     * @param[out]     output Output tensor. Data types supported: Same as @p input1.
+     */
+    void configure(ITensor *input1, ITensor *input2, ITensor *output);
+    /** Static function to check if given info will lead to a valid configuration of @ref NEArithmeticOperationKernel for power
+     *
+     * @param[in] input1 First tensor input info. Data types supported: F16/F32.
+     * @param[in] input2 Second tensor input info. Data types supported: Same as @p input1.
+     * @param[in] output Output tensor info. Data types supported: Same as @p input1.
+     *
+     * @return a status
+     */
+    static Status validate(const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output);
+};
+
 /** Basic function to run @ref NEComparisonOperationKernel.
  *
  * @note The tensor data type for the inputs must be QASYMM8/S16/F16/S32/F32.
diff --git a/src/core/NEON/kernels/NEElementwiseOperationKernel.cpp b/src/core/NEON/kernels/NEElementwiseOperationKernel.cpp
index 6b87ea017b..33457e1fca 100644
--- a/src/core/NEON/kernels/NEElementwiseOperationKernel.cpp
+++ b/src/core/NEON/kernels/NEElementwiseOperationKernel.cpp
@@ -130,6 +130,11 @@ inline ScalarType elementwise_arithm_op_scalar(const ScalarType &a, const Scalar
             res = a / b;
             break;
         }
+        case ArithmeticOperation::POWER:
+        {
+            res = std::pow(a, b);
+            break;
+        }
         default:
             ARM_COMPUTE_ERROR("NOT_SUPPORTED!");
     }
@@ -174,12 +179,24 @@ inline float32x4_t elementwise_arithm_op<ArithmeticOperation::DIV, float32x4_t>(
     return wrapper::vdiv(a, b);
 }
 
+template <>
+inline float32x4_t elementwise_arithm_op<ArithmeticOperation::POWER, float32x4_t>(const float32x4_t &a, const float32x4_t &b)
+{
+    return wrapper::vpow(a, b);
+}
+
 #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
 template <>
 inline float16x8_t elementwise_arithm_op<ArithmeticOperation::DIV, float16x8_t>(const float16x8_t &a, const float16x8_t &b)
 {
     return wrapper::vdiv(a, b);
 }
+
+template <>
+inline float16x8_t elementwise_arithm_op<ArithmeticOperation::POWER, float16x8_t>(const float16x8_t &a, const float16x8_t &b)
+{
+    return wrapper::vpow(a, b);
+}
 #endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
 
 template <ArithmeticOperation op>
@@ -879,6 +896,27 @@ Status NEDivisionOperationKernel::validate(const ITensorInfo *input1, const ITen
     return Status{};
 }
 
+/** The power operator */
+void NEPowerOperationKernel::configure(const ITensor *input1, const ITensor *input2, ITensor *output)
+{
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(*input1->info(), *input2->info(), *output->info()));
+    configure_common(input1, input2, output);
+    _function = configure_arithm_func<ArithmeticOperation::POWER>(input1, input2, output);
+}
+
+Status NEPowerOperationKernel::validate_arguments(const ITensorInfo &input1, const ITensorInfo &input2, const ITensorInfo &output)
+{
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&input1, 1, DataType::F16, DataType::F32);
+    return NEArithmeticOperationKernel::validate_arguments(input1, input2, output);
+}
+
+Status NEPowerOperationKernel::validate(const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output)
+{
+    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input1, input2, output);
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(*input1, *input2, *output));
+    return Status{};
+}
+
 /** Comparison operators (equal, not equal, less than, greater than, less than or equal, greater than or equal) */
 
 void NEComparisonOperationKernel::configure(ComparisonOperation op, const ITensor *input1, const ITensor *input2, ITensor *output)
diff --git a/src/runtime/NEON/functions/NEElementwiseOperators.cpp b/src/runtime/NEON/functions/NEElementwiseOperators.cpp
index 74c195764b..699363111d 100644
--- a/src/runtime/NEON/functions/NEElementwiseOperators.cpp
+++ b/src/runtime/NEON/functions/NEElementwiseOperators.cpp
@@ -79,6 +79,18 @@ Status NEElementwiseDivision::validate(const ITensorInfo *input1, const ITensorI
     return NEDivisionOperationKernel::validate(input1, input2, output);
 }
 
+void NEElementwisePower::configure(ITensor *input1, ITensor *input2, ITensor *output)
+{
+    auto k = arm_compute::support::cpp14::make_unique<NEPowerOperationKernel>();
+    k->configure(input1, input2, output);
+    _kernel = std::move(k);
+}
+
+Status NEElementwisePower::validate(const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output)
+{
+    return NEPowerOperationKernel::validate(input1, input2, output);
+}
+
 template <ComparisonOperation COP>
 void NEElementwiseComparisonStatic<COP>::configure(ITensor *input1, ITensor *input2, ITensor *output)
 {
diff --git a/tests/validation/NEON/ElementwisePower.cpp b/tests/validation/NEON/ElementwisePower.cpp
new file mode 100644
index 0000000000..3ca39e840a
--- /dev/null
+++ b/tests/validation/NEON/ElementwisePower.cpp
@@ -0,0 +1,155 @@
+/*
+ * Copyright (c) 2019 ARM Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#include "arm_compute/core/Types.h"
+#include "arm_compute/runtime/NEON/functions/NEElementwiseOperations.h"
+#include "arm_compute/runtime/Tensor.h"
+#include "arm_compute/runtime/TensorAllocator.h"
+#include "tests/NEON/Accessor.h"
+#include "tests/PaddingCalculator.h"
+#include "tests/datasets/ShapeDatasets.h"
+#include "tests/framework/Asserts.h"
+#include "tests/framework/Macros.h"
+#include "tests/framework/datasets/Datasets.h"
+#include "tests/validation/Validation.h"
+#include "tests/validation/fixtures/ElementwiseOperationsFixture.h"
+
+namespace arm_compute
+{
+namespace test
+{
+namespace validation
+{
+namespace
+{
+RelativeTolerance<float> tolerance_fp32(0.001f);
+/** Input data sets **/
+#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+RelativeTolerance<half> tolerance_fp16(static_cast<half>(0.01f));
+const auto              ElementwisePowerFP16Dataset = combine(combine(framework::dataset::make("DataType", DataType::F16), framework::dataset::make("DataType", DataType::F16)),
+                                                              framework::dataset::make("DataType", DataType::F16));
+#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */
+const auto ElementwisePowerFP32Dataset = combine(combine(framework::dataset::make("DataType", DataType::F32), framework::dataset::make("DataType", DataType::F32)),
+                                                 framework::dataset::make("DataType", DataType::F32));
+} // namespace
+
+TEST_SUITE(NEON)
+TEST_SUITE(ElementwisePower)
+
+template <typename T>
+using NEElementwisePowerFixture = ElementwisePowerValidationFixture<Tensor, Accessor, NEElementwisePower, T>;
+
+// *INDENT-OFF*
+// clang-format off
+DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(
+               framework::dataset::make("Input1Info", { TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32),
+                                                        TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32),
+                                                        TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32),
+                                                        TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32),     // Invalid data type combination
+                                                        TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32),     // Mismatching shapes
+                                                      }),
+               framework::dataset::make("Input2Info",{ TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32),
+                                                       TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32),
+                                                       TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32),
+                                                       TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::S32),
+                                                       TensorInfo(TensorShape(48U, 11U, 2U), 1, DataType::F32),
+                                                     })),
+               framework::dataset::make("OutputInfo",{ TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32),
+                                                       TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32),
+                                                       TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32),
+                                                       TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32),
+                                                       TensorInfo(TensorShape(48U, 11U, 2U), 1, DataType::F32),
+                                                     })),
+               framework::dataset::make("Expected", { true, true, true, false, false})),
+               input1_info, input2_info, output_info, expected)
+{
+    ARM_COMPUTE_EXPECT(bool(NEElementwisePower::validate(&input1_info.clone()->set_is_resizable(false), &input2_info.clone()->set_is_resizable(false), &output_info.clone()->set_is_resizable(false))) == expected, framework::LogLevel::ERRORS);
+}
+// clang-format on
+// *INDENT-ON*
+
+TEST_SUITE(Float)
+#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+TEST_SUITE(F16)
+FIXTURE_DATA_TEST_CASE(RunSmall, NEElementwisePowerFixture<half>, framework::DatasetMode::ALL, combine(datasets::SmallShapes(), ElementwisePowerFP16Dataset))
+{
+    // Validate output
+    validate(Accessor(_target), _reference, tolerance_fp16, 0.01);
+}
+TEST_SUITE_END() // F16
+#endif           /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */
+
+TEST_SUITE(F32)
+DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, framework::dataset::concat(datasets::SmallShapes(), datasets::LargeShapes()),
+               shape)
+{
+    // Create tensors
+    Tensor ref_src1 = create_tensor<Tensor>(shape, DataType::F32);
+    Tensor ref_src2 = create_tensor<Tensor>(shape, DataType::F32);
+    Tensor dst      = create_tensor<Tensor>(shape, DataType::F32);
+
+    // Create and Configure function
+    NEElementwisePower power;
+    power.configure(&ref_src1, &ref_src2, &dst);
+
+    // Validate valid region
+    const ValidRegion valid_region = shape_to_valid_region(shape);
+    validate(dst.info()->valid_region(), valid_region);
+}
+
+FIXTURE_DATA_TEST_CASE(RunSmall, NEElementwisePowerFixture<float>, framework::DatasetMode::ALL, combine(datasets::SmallShapes(), ElementwisePowerFP32Dataset))
+{
+    // Validate output
+    validate(Accessor(_target), _reference, tolerance_fp32, 0.01);
+}
+
+FIXTURE_DATA_TEST_CASE(RunLarge, NEElementwisePowerFixture<float>, framework::DatasetMode::NIGHTLY, combine(datasets::LargeShapes(), ElementwisePowerFP32Dataset))
+{
+    // Validate output
+    validate(Accessor(_target), _reference, tolerance_fp32, 0.01);
+}
+
+template <typename T>
+using NEElementwisePowerBroadcastFixture = ElementwisePowerBroadcastValidationFixture<Tensor, Accessor, NEElementwisePower, T>;
+
+FIXTURE_DATA_TEST_CASE(RunSmallBroadcast, NEElementwisePowerBroadcastFixture<float>, framework::DatasetMode::ALL, combine(datasets::SmallShapesBroadcast(),
+                       ElementwisePowerFP32Dataset))
+{
+    // Validate output
+    validate(Accessor(_target), _reference, tolerance_fp32, 0.01);
+}
+
+FIXTURE_DATA_TEST_CASE(RunLargeBroadcast, NEElementwisePowerBroadcastFixture<float>, framework::DatasetMode::NIGHTLY, combine(datasets::LargeShapesBroadcast(),
+                       ElementwisePowerFP32Dataset))
+{
+    // Validate output
+    validate(Accessor(_target), _reference, tolerance_fp32, 0.01);
+}
+TEST_SUITE_END() // F32
+TEST_SUITE_END() // Float
+
+TEST_SUITE_END() // ElementwisePower
+TEST_SUITE_END() // NEON
+} // namespace validation
+} // namespace test
+} // namespace arm_compute
diff --git a/tests/validation/fixtures/ElementwiseOperationsFixture.h b/tests/validation/fixtures/ElementwiseOperationsFixture.h
index 53030842e7..e86e7a0f20 100644
--- a/tests/validation/fixtures/ElementwiseOperationsFixture.h
+++ b/tests/validation/fixtures/ElementwiseOperationsFixture.h
@@ -58,7 +58,17 @@ protected:
     template <typename U>
     void fill(U &&tensor, int i)
     {
-        (_op == ArithmeticOperation::DIV) ? library->fill_tensor_uniform_ranged(tensor, i, { std::pair<float, float>(-0.001f, 0.001f) }) : library->fill_tensor_uniform(tensor, i);
+        switch(_op)
+        {
+            case ArithmeticOperation::DIV:
+                library->fill_tensor_uniform_ranged(tensor, i, { std::pair<float, float>(-0.001f, 0.001f) });
+                break;
+            case ArithmeticOperation::POWER:
+                library->fill_tensor_uniform(tensor, i, 0.0f, 5.0f);
+                break;
+            default:
+                library->fill_tensor_uniform(tensor, i);
+        }
     }
 
     TensorType compute_target(const TensorShape &shape0, const TensorShape &shape1, DataType data_type0, DataType data_type1, DataType output_data_type,
@@ -382,6 +392,32 @@ public:
     }
 };
 
+template <typename TensorType, typename AccessorType, typename FunctionType, typename T>
+class ElementwisePowerBroadcastValidationFixture : public ArithmeticOperationsGenericFixture<TensorType, AccessorType, FunctionType, T>
+{
+public:
+    template <typename...>
+    void setup(const TensorShape &shape0, const TensorShape &shape1, DataType data_type0, DataType data_type1, DataType output_data_type)
+    {
+        ArithmeticOperationsGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(ArithmeticOperation::POWER, shape0, shape1,
+                                                                                             data_type0, data_type1, output_data_type,
+                                                                                             QuantizationInfo(), QuantizationInfo(), QuantizationInfo());
+    }
+};
+
+template <typename TensorType, typename AccessorType, typename FunctionType, typename T>
+class ElementwisePowerValidationFixture : public ArithmeticOperationsGenericFixture<TensorType, AccessorType, FunctionType, T>
+{
+public:
+    template <typename...>
+    void setup(const TensorShape &shape, DataType data_type0, DataType data_type1, DataType output_data_type)
+    {
+        ArithmeticOperationsGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(ArithmeticOperation::POWER, shape, shape,
+                                                                                             data_type0, data_type1, output_data_type,
+                                                                                             QuantizationInfo(), QuantizationInfo(), QuantizationInfo());
+    }
+};
+
 } // namespace validation
 } // namespace test
 } // namespace arm_compute
diff --git a/tests/validation/reference/ElementwiseOperations.cpp b/tests/validation/reference/ElementwiseOperations.cpp
index 2ffb0faa75..82f42a0c21 100644
--- a/tests/validation/reference/ElementwiseOperations.cpp
+++ b/tests/validation/reference/ElementwiseOperations.cpp
@@ -43,38 +43,51 @@ T arithm_op(ArithmeticOperation op, T src1, T src2, ConvertPolicy convert_policy
 
     intermediate_type val;
 
-    if(op == ArithmeticOperation::ADD)
+    switch(op)
     {
-        val = static_cast<intermediate_type>(src1) + static_cast<intermediate_type>(src2);
-    }
-    else if(op == ArithmeticOperation::SUB)
-    {
-        val = static_cast<intermediate_type>(src1) - static_cast<intermediate_type>(src2);
-    }
-    else if(op == ArithmeticOperation::MIN)
-    {
-        val = std::min(static_cast<intermediate_type>(src1), static_cast<intermediate_type>(src2));
-    }
-    else if(op == ArithmeticOperation::MAX)
-    {
-        val = std::max(static_cast<intermediate_type>(src1), static_cast<intermediate_type>(src2));
-    }
-    else if(op == ArithmeticOperation::SQUARED_DIFF)
-    {
-        intermediate_type tmp = (static_cast<intermediate_type>(src1) - static_cast<intermediate_type>(src2));
-        val                   = tmp * tmp;
-    }
-    else if(op == ArithmeticOperation::DIV)
-    {
-        val = (static_cast<intermediate_type>(src1) / static_cast<intermediate_type>(src2));
-    }
-    else
-    {
-        ARM_COMPUTE_ERROR("Not handled");
+        case ArithmeticOperation::ADD:
+        {
+            val = static_cast<intermediate_type>(src1) + static_cast<intermediate_type>(src2);
+            break;
+        }
+        case ArithmeticOperation::SUB:
+        {
+            val = static_cast<intermediate_type>(src1) - static_cast<intermediate_type>(src2);
+            break;
+        }
+        case ArithmeticOperation::MIN:
+        {
+            val = std::min(static_cast<intermediate_type>(src1), static_cast<intermediate_type>(src2));
+            break;
+        }
+        case ArithmeticOperation::MAX:
+        {
+            val = std::max(static_cast<intermediate_type>(src1), static_cast<intermediate_type>(src2));
+            break;
+        }
+        case ArithmeticOperation::SQUARED_DIFF:
+        {
+            intermediate_type tmp = (static_cast<intermediate_type>(src1) - static_cast<intermediate_type>(src2));
+            val                   = tmp * tmp;
+            break;
+        }
+        case ArithmeticOperation::DIV:
+        {
+            val = (static_cast<intermediate_type>(src1) / static_cast<intermediate_type>(src2));
+            break;
+        }
+        case ArithmeticOperation::POWER:
+        {
+            val = std::pow(static_cast<intermediate_type>(src1), static_cast<intermediate_type>(src2));
+            break;
+        }
+        default:
+        {
+            ARM_COMPUTE_ERROR("Not handled");
+        }
     }
-
     T result;
-    if(op == ArithmeticOperation::ADD || op == ArithmeticOperation::SUB || op == ArithmeticOperation::DIV)
+    if(op == ArithmeticOperation::ADD || op == ArithmeticOperation::SUB || op == ArithmeticOperation::DIV || op == ArithmeticOperation::POWER)
     {
         result = (convert_policy == ConvertPolicy::SATURATE) ? saturate_cast<T>(val) : static_cast<T>(val);
     }