TheRock/rocm-libraries/projects/hipdnn/test_sdk/tests/utilities/TestCpuReferencePointwise.cpp

// Copyright © Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT

#include <cmath>
#include <gtest/gtest.h>
#include <hipdnn_data_sdk/types.hpp>
#include <hipdnn_data_sdk/utilities/Tensor.hpp>
#include <hipdnn_test_sdk/utilities/CpuFpReferenceValidation.hpp>
#include <hipdnn_test_sdk/utilities/DynamicTolerances.hpp>
#include <hipdnn_test_sdk/utilities/FlatbufferGraphTestUtils.hpp>
#include <hipdnn_test_sdk/utilities/pointwise/CpuReferencePointwise.hpp>
#include <hipdnn_test_sdk/utilities/pointwise/PointwiseErrorClassification.hpp>

using namespace hipdnn_test_sdk::utilities;
using namespace hipdnn_data_sdk::utilities;
using namespace hipdnn_flatbuffers_sdk::data_objects;
using namespace hipdnn_data_sdk::types;
using hipdnn_test_sdk::detail::safeTestTypeCast;

namespace
{

// Mathematical constants
constexpr float PI = 3.14159265f;
constexpr float E = 2.71828183f;
constexpr float SQRT_2 = 1.41421356f;
constexpr float SQRT_3 = 1.73205081f;
constexpr float SQRT_5 = 2.23606798f;
constexpr float GOLDEN_RATIO = 1.61803399f; // φ
constexpr float LN_2 = 0.69314718f;

// Common test values
constexpr float TEST_VALUE_1 = 1.0f;
constexpr float TEST_VALUE_2 = 2.0f;
constexpr float TEST_VALUE_3 = 3.0f;
constexpr float TEST_VALUE_5 = 5.0f;
constexpr float TEST_VALUE_2_5 = 2.5f;
constexpr float TEST_VALUE_1_5 = 1.5f;
constexpr float TEST_VALUE_4 = 4.0f;

// Precision test values
constexpr float PRECISION_TEST_A = 0.123456789f;
constexpr float PRECISION_TEST_B = 0.987654321f;

// Broadcasting test values
constexpr float BROADCAST_MULTIPLIER_10 = 10.0f;

} // namespace

template <typename Input1Type,
          typename Input2Type = Input1Type,
          typename OutputType = Input1Type,
          typename ComputeType = double>
class CpuReferencePointwiseFixture : public ::testing::Test
{
protected:
    /// Calculates dynamic tolerance based on the operation and input data range.
    /// Takes the max of ComputeType-based and float-based tolerance because test
    /// expected values are computed from float constants (PI, E, TEST_VALUE_*),
    /// so precision is bounded by float even when ComputeType is double.
    float getDynamicTolerance(PointwiseMode mode, float scale) const
    {
        auto errorClass = pointwise::classifyPointwiseOp(mode);
        auto effectiveScale = static_cast<double>(pointwise::isBoundedOutput(mode) ? 1.0f : scale);
        auto tolerance
            = pointwise::calculatePointwiseTolerance<OutputType, Input1Type, ComputeType>(
                effectiveScale, errorClass);
        auto floatFloor = pointwise::calculatePointwiseTolerance<OutputType, Input1Type, float>(
            effectiveScale, errorClass);
        return std::max(tolerance, floatFloor);
    }

    // ======================= BINARY OPERATIONS =======================

    void testBinaryAddOperation()
    {
        Tensor<Input1Type> input1({1, 3, 2, 2});
        Tensor<Input2Type> input2({1, 3, 2, 2});
        Tensor<OutputType> output({1, 3, 2, 2});

        input1.fillWithValue(safeTestTypeCast<Input1Type>(TEST_VALUE_1));
        input2.fillWithValue(safeTestTypeCast<Input2Type>(TEST_VALUE_2));

        CpuReferencePointwiseImpl<OutputType, Input1Type, Input2Type>::pointwiseCompute(
            PointwiseMode::ADD, output, input1, input2);

        Tensor<OutputType> expected({1, 3, 2, 2});
        expected.fillWithValue(safeTestTypeCast<OutputType>(TEST_VALUE_3));

        auto tolerance = getDynamicTolerance(PointwiseMode::ADD, TEST_VALUE_2);
        auto validator = createAllCloseValidator<OutputType>(tolerance, tolerance);
        EXPECT_TRUE(validator->allClose(expected, output));
    }

    void testBinarySubtractOperation()
    {
        Tensor<Input1Type> input1({1, 3, 2, 2});
        Tensor<Input2Type> input2({1, 3, 2, 2});
        Tensor<OutputType> output({1, 3, 2, 2});

        input1.fillWithValue(safeTestTypeCast<Input1Type>(TEST_VALUE_5));
        input2.fillWithValue(safeTestTypeCast<Input2Type>(TEST_VALUE_2));

        CpuReferencePointwiseImpl<OutputType, Input1Type, Input2Type>::pointwiseCompute(
            PointwiseMode::SUB, output, input1, input2);

        Tensor<OutputType> expected({1, 3, 2, 2});
        expected.fillWithValue(safeTestTypeCast<OutputType>(TEST_VALUE_3));

        auto tolerance = getDynamicTolerance(PointwiseMode::SUB, TEST_VALUE_5);
        auto validator = createAllCloseValidator<OutputType>(tolerance, tolerance);
        EXPECT_TRUE(validator->allClose(expected, output));
    }

    void testBinaryAddOperationSanityValidation()
    {
        Tensor<Input1Type> input1({1, 1, 2, 2});
        Tensor<Input2Type> input2({1, 1, 2, 2});
        Tensor<OutputType> output({1, 1, 2, 2});

        // Create expected tensor with computed results
        Tensor<OutputType> expected({1, 1, 2, 2});

        if constexpr(std::is_integral_v<Input1Type> || std::is_integral_v<Input2Type>)
        {
            input1.setHostValue(safeTestTypeCast<Input1Type>(TEST_VALUE_1), 0, 0, 0, 0);
            input1.setHostValue(safeTestTypeCast<Input1Type>(TEST_VALUE_2), 0, 0, 0, 1);
            input1.setHostValue(safeTestTypeCast<Input1Type>(TEST_VALUE_3), 0, 0, 1, 0);
            input1.setHostValue(safeTestTypeCast<Input1Type>(TEST_VALUE_4), 0, 0, 1, 1);

            input2.setHostValue(safeTestTypeCast<Input2Type>(-TEST_VALUE_4), 0, 0, 0, 0);
            input2.setHostValue(safeTestTypeCast<Input2Type>(-TEST_VALUE_3), 0, 0, 0, 1);
            input2.setHostValue(safeTestTypeCast<Input2Type>(-TEST_VALUE_2), 0, 0, 1, 0);
            input2.setHostValue(safeTestTypeCast<Input2Type>(-TEST_VALUE_1), 0, 0, 1, 1);

            expected.setHostValue(
                safeTestTypeCast<OutputType>(TEST_VALUE_1 + (-TEST_VALUE_4)), 0, 0, 0, 0);
            expected.setHostValue(
                safeTestTypeCast<OutputType>(TEST_VALUE_2 + (-TEST_VALUE_3)), 0, 0, 0, 1);
            expected.setHostValue(
                safeTestTypeCast<OutputType>(TEST_VALUE_3 + (-TEST_VALUE_2)), 0, 0, 1, 0);
            expected.setHostValue(
                safeTestTypeCast<OutputType>(TEST_VALUE_4 + (-TEST_VALUE_1)), 0, 0, 1, 1);
        }
        else
        {
            input1.setHostValue(safeTestTypeCast<Input1Type>(PI), 0, 0, 0, 0); // π
            input1.setHostValue(safeTestTypeCast<Input1Type>(E), 0, 0, 0, 1); // e
            input1.setHostValue(safeTestTypeCast<Input1Type>(SQRT_2), 0, 0, 1, 0); // √2
            input1.setHostValue(
                safeTestTypeCast<Input1Type>(GOLDEN_RATIO), 0, 0, 1, 1); // φ (golden ratio)

            input2.setHostValue(safeTestTypeCast<Input2Type>(LN_2), 0, 0, 0, 0); // ln(2)
            input2.setHostValue(
                safeTestTypeCast<Input2Type>(hipdnn_data_sdk::types::sin(1.0f)), 0, 0, 0, 1);
            input2.setHostValue(
                safeTestTypeCast<Input2Type>(hipdnn_data_sdk::types::cos(1.0f)), 0, 0, 1, 0);
            input2.setHostValue(
                safeTestTypeCast<Input2Type>(hipdnn_data_sdk::types::tan(1.0f)), 0, 0, 1, 1);

            expected.setHostValue(safeTestTypeCast<OutputType>(PI + LN_2), 0, 0, 0, 0); // π + ln(2)
            expected.setHostValue(
                safeTestTypeCast<OutputType>(E + hipdnn_data_sdk::types::sin(1.0f)),
                0,
                0,
                0,
                1); // e + sin(1)
            expected.setHostValue(
                safeTestTypeCast<OutputType>(SQRT_2 + hipdnn_data_sdk::types::cos(1.0f)),
                0,
                0,
                1,
                0); // √2 + cos(1)
            expected.setHostValue(
                safeTestTypeCast<OutputType>(GOLDEN_RATIO + hipdnn_data_sdk::types::tan(1.0f)),
                0,
                0,
                1,
                1); // φ + tan(1)
        }

        CpuReferencePointwiseImpl<OutputType, Input1Type, Input2Type>::pointwiseCompute(
            PointwiseMode::ADD, output, input1, input2);

        auto tolerance = getDynamicTolerance(PointwiseMode::ADD, PI);
        auto validator = createAllCloseValidator<OutputType>(tolerance, tolerance);
        EXPECT_TRUE(validator->allClose(expected, output));
    }

    void testBinarySubtractOperationSanityValidation()
    {
        Tensor<Input1Type> input1({1, 1, 2, 2});
        Tensor<Input2Type> input2({1, 1, 2, 2});
        Tensor<OutputType> output({1, 1, 2, 2});

        // Create expected tensor with computed results
        Tensor<OutputType> expected({1, 1, 2, 2});

        if constexpr(std::is_integral_v<Input1Type> || std::is_integral_v<Input2Type>)
        {
            input1.setHostValue(safeTestTypeCast<Input1Type>(TEST_VALUE_1), 0, 0, 0, 0);
            input1.setHostValue(safeTestTypeCast<Input1Type>(TEST_VALUE_2), 0, 0, 0, 1);
            input1.setHostValue(safeTestTypeCast<Input1Type>(TEST_VALUE_3), 0, 0, 1, 0);
            input1.setHostValue(safeTestTypeCast<Input1Type>(TEST_VALUE_4), 0, 0, 1, 1);

            input2.setHostValue(safeTestTypeCast<Input2Type>(-TEST_VALUE_4), 0, 0, 0, 0);
            input2.setHostValue(safeTestTypeCast<Input2Type>(-TEST_VALUE_3), 0, 0, 0, 1);
            input2.setHostValue(safeTestTypeCast<Input2Type>(-TEST_VALUE_2), 0, 0, 1, 0);
            input2.setHostValue(safeTestTypeCast<Input2Type>(-TEST_VALUE_1), 0, 0, 1, 1);

            expected.setHostValue(
                safeTestTypeCast<OutputType>(TEST_VALUE_1 - (-TEST_VALUE_4)), 0, 0, 0, 0);
            expected.setHostValue(
                safeTestTypeCast<OutputType>(TEST_VALUE_2 - (-TEST_VALUE_3)), 0, 0, 0, 1);
            expected.setHostValue(
                safeTestTypeCast<OutputType>(TEST_VALUE_3 - (-TEST_VALUE_2)), 0, 0, 1, 0);
            expected.setHostValue(
                safeTestTypeCast<OutputType>(TEST_VALUE_4 - (-TEST_VALUE_1)), 0, 0, 1, 1);
        }
        else
        {
            input1.setHostValue(safeTestTypeCast<Input1Type>(TEST_VALUE_2 * PI), 0, 0, 0, 0); // 2π
            input1.setHostValue(safeTestTypeCast<Input1Type>(E * E), 0, 0, 0, 1); // e²
            input1.setHostValue(safeTestTypeCast<Input1Type>(SQRT_5), 0, 0, 1, 0); // √5
            input1.setHostValue(
                safeTestTypeCast<Input1Type>(TEST_VALUE_2), 0, 0, 1, 1); // Simple test value

            input2.setHostValue(safeTestTypeCast<Input2Type>(PI / TEST_VALUE_2), 0, 0, 0, 0); // π/2
            input2.setHostValue(safeTestTypeCast<Input2Type>(E), 0, 0, 0, 1); // e
            input2.setHostValue(safeTestTypeCast<Input2Type>(SQRT_3), 0, 0, 1, 0); // √3
            input2.setHostValue(
                safeTestTypeCast<Input2Type>(TEST_VALUE_1), 0, 0, 1, 1); // Simple test value

            expected.setHostValue(
                safeTestTypeCast<OutputType>((TEST_VALUE_2 * PI) - (PI / TEST_VALUE_2)),
                0,
                0,
                0,
                0); // 2π - π/2 = 3π/2
            expected.setHostValue(safeTestTypeCast<OutputType>((E * E) - E), 0, 0, 0,
                                  1); // e² - e
            expected.setHostValue(
                safeTestTypeCast<OutputType>(SQRT_5 - SQRT_3), 0, 0, 1, 0); // √5 - √3
            expected.setHostValue(
                safeTestTypeCast<OutputType>(TEST_VALUE_2 - TEST_VALUE_1), 0, 0, 1, 1); // 2 - 1
        }

        CpuReferencePointwiseImpl<OutputType, Input1Type, Input2Type>::pointwiseCompute(
            PointwiseMode::SUB, output, input1, input2);

        auto tolerance = getDynamicTolerance(PointwiseMode::SUB, E * E);
        auto validator = createAllCloseValidator<OutputType>(tolerance, tolerance);
        EXPECT_TRUE(validator->allClose(expected, output));
    }

    void testBinaryAddOperation3D()
    {
        Tensor<Input1Type> input1({2, 3, 10});
        Tensor<Input2Type> input2({2, 3, 10});
        Tensor<OutputType> output({2, 3, 10});

        input1.fillWithValue(safeTestTypeCast<Input1Type>(TEST_VALUE_2_5));
        input2.fillWithValue(safeTestTypeCast<Input2Type>(TEST_VALUE_1_5));

        CpuReferencePointwiseImpl<OutputType, Input1Type, Input2Type>::pointwiseCompute(
            PointwiseMode::ADD, output, input1, input2);

        Tensor<OutputType> expected({2, 3, 10});
        expected.fillWithValue(safeTestTypeCast<OutputType>(TEST_VALUE_4));

        auto tolerance = getDynamicTolerance(PointwiseMode::ADD, TEST_VALUE_2_5);
        auto validator = createAllCloseValidator<OutputType>(tolerance, tolerance);
        EXPECT_TRUE(validator->allClose(expected, output));
    }

    void testBinarySingleElementTensors()
    {
        Tensor<Input1Type> input1({1, 1, 1, 1});
        Tensor<Input2Type> input2({1, 1, 1, 1});
        Tensor<OutputType> output({1, 1, 1, 1});

        Tensor<OutputType> expected({1, 1, 1, 1});

        if constexpr(std::is_integral_v<Input1Type> || std::is_integral_v<Input2Type>)
        {
            input1.setHostValue(safeTestTypeCast<Input1Type>(TEST_VALUE_2), 0, 0, 0, 0);
            input2.setHostValue(safeTestTypeCast<Input2Type>(TEST_VALUE_5), 0, 0, 0, 0);

            expected.setHostValue(
                safeTestTypeCast<OutputType>(TEST_VALUE_2 - TEST_VALUE_5), 0, 0, 0, 0);
        }
        else
        {
            input1.setHostValue(safeTestTypeCast<Input1Type>(E * E), 0, 0, 0, 0); // e²
            input2.setHostValue(safeTestTypeCast<Input2Type>(E), 0, 0, 0, 0); // e

            expected.setHostValue(safeTestTypeCast<OutputType>((E * E) - E), 0, 0, 0,
                                  0); // e² - e
        }

        CpuReferencePointwiseImpl<OutputType, Input1Type, Input2Type>::pointwiseCompute(
            PointwiseMode::SUB, output, input1, input2);

        auto tolerance = getDynamicTolerance(PointwiseMode::SUB, E * E);
        auto validator = createAllCloseValidator<OutputType>(tolerance, tolerance);
        EXPECT_TRUE(validator->allClose(expected, output));
    }

    void testBinaryNumericalPrecision()
    {
        Tensor<Input1Type> input1({1, 1, 1, 1});
        Tensor<Input2Type> input2({1, 1, 1, 1});
        Tensor<OutputType> output({1, 1, 1, 1});

        input1.setHostValue(safeTestTypeCast<Input1Type>(PRECISION_TEST_A), 0, 0, 0, 0);
        input2.setHostValue(safeTestTypeCast<Input2Type>(PRECISION_TEST_B), 0, 0, 0, 0);

        CpuReferencePointwiseImpl<OutputType, Input1Type, Input2Type>::pointwiseCompute(
            PointwiseMode::ADD, output, input1, input2);

        Tensor<OutputType> expected({1, 1, 1, 1});
        expected.setHostValue(
            safeTestTypeCast<OutputType>(PRECISION_TEST_A + PRECISION_TEST_B), 0, 0, 0, 0);

        auto tolerance = getDynamicTolerance(PointwiseMode::ADD, PRECISION_TEST_B);
        auto validator = createAllCloseValidator<OutputType>(tolerance, tolerance);
        EXPECT_TRUE(validator->allClose(expected, output));
    }

    void testElementwise1D()
    {
        Tensor<Input1Type> input1({5});
        Tensor<Input2Type> input2({5});
        Tensor<OutputType> output({5});

        for(int i = 0; i < 5; ++i)
        {
            input1.setHostValue(safeTestTypeCast<Input1Type>(static_cast<float>(i + 1)), i);
            input2.setHostValue(safeTestTypeCast<Input2Type>(static_cast<float>(i * 2)), i);
        }

        CpuReferencePointwiseImpl<OutputType, Input1Type, Input2Type>::pointwiseCompute(
            PointwiseMode::ADD, output, input1, input2);

        // Create expected tensor: [1,2,3,4,5] + [0,2,4,6,8] = [1,4,7,10,13]
        Tensor<OutputType> expected({5});
        expected.setHostValue(safeTestTypeCast<OutputType>(static_cast<float>(1)), 0);
        expected.setHostValue(safeTestTypeCast<OutputType>(static_cast<float>(4)), 1);
        expected.setHostValue(safeTestTypeCast<OutputType>(static_cast<float>(7)), 2);
        expected.setHostValue(safeTestTypeCast<OutputType>(static_cast<float>(10)), 3);
        expected.setHostValue(safeTestTypeCast<OutputType>(static_cast<float>(13)), 4);

        auto tolerance = getDynamicTolerance(PointwiseMode::ADD, 8.0f);
        auto validator = createAllCloseValidator<OutputType>(tolerance, tolerance);
        EXPECT_TRUE(validator->allClose(expected, output));
    }

    void testBroadcast2Dx1D()
    {
        Tensor<Input1Type> input1({3, 4}); // [M,N] = [3,4]
        Tensor<Input2Type> input2({4}); // [N] = [4]
        Tensor<OutputType> output({3, 4}); // Output: [3,4]

        // Fill input1 with pattern: row*10 + col
        for(int m = 0; m < 3; ++m)
        {
            for(int n = 0; n < 4; ++n)
            {
                input1.setHostValue(
                    safeTestTypeCast<Input1Type>(static_cast<float>((m * 10) + n)), m, n);
            }
        }

        // Fill input2 with pattern: [10, 20, 30, 40]
        for(int n = 0; n < 4; ++n)
        {
            input2.setHostValue(safeTestTypeCast<Input2Type>(static_cast<float>((n + 1) * 10)), n);
        }

        CpuReferencePointwiseImpl<OutputType, Input1Type, Input2Type>::pointwiseCompute(
            PointwiseMode::ADD, output, input1, input2);

        // Create expected tensor: broadcasting input2[n] to all input1[m,n]
        Tensor<OutputType> expected({3, 4});
        for(int m = 0; m < 3; ++m)
        {
            for(int n = 0; n < 4; ++n)
            {
                auto input1Val = static_cast<float>((m * 10) + n);
                auto input2Val = static_cast<float>((n + 1) * 10);
                expected.setHostValue(safeTestTypeCast<OutputType>(input1Val + input2Val), m, n);
            }
        }

        auto tolerance = getDynamicTolerance(PointwiseMode::ADD, 40.0f);
        auto validator = createAllCloseValidator<OutputType>(tolerance, tolerance);
        EXPECT_TRUE(validator->allClose(expected, output));
    }

    void testBroadcast3D()
    {
        Tensor<Input1Type> input1({2, 3, 4}); // [2,3,4]
        Tensor<Input2Type> input2({1, 3, 1}); // [1,3,1] - broadcasts to [2,3,4]
        Tensor<OutputType> output({2, 3, 4}); // Output: [2,3,4]

        input1.fillWithValue(safeTestTypeCast<Input1Type>(TEST_VALUE_5));

        input2.setHostValue(safeTestTypeCast<Input2Type>(TEST_VALUE_1), 0, 0, 0); // Channel 0
        input2.setHostValue(safeTestTypeCast<Input2Type>(TEST_VALUE_2), 0, 1, 0); // Channel 1
        input2.setHostValue(safeTestTypeCast<Input2Type>(TEST_VALUE_3), 0, 2, 0); // Channel 2

        CpuReferencePointwiseImpl<OutputType, Input1Type, Input2Type>::pointwiseCompute(
            PointwiseMode::SUB, output, input1, input2);

        // Create expected tensor: broadcasting subtraction
        Tensor<OutputType> expected({2, 3, 4});
        for(int n = 0; n < 2; ++n)
        {
            for(int c = 0; c < 3; ++c)
            {
                for(int h = 0; h < 4; ++h)
                {
                    const float input1Val = TEST_VALUE_5;
                    auto input2Val = static_cast<float>(c + 1); // Channel values: 1.0, 2.0, 3.0
                    expected.setHostValue(
                        safeTestTypeCast<OutputType>(input1Val - input2Val), n, c, h);
                }
            }
        }

        auto tolerance = getDynamicTolerance(PointwiseMode::SUB, TEST_VALUE_5);
        auto validator = createAllCloseValidator<OutputType>(tolerance, tolerance);
        EXPECT_TRUE(validator->allClose(expected, output));
    }

    void testBroadcast3DImplicitLeading()
    {
        Tensor<Input1Type> input1({2, 3, 4}); // [2,3,4]
        Tensor<Input2Type> input2(
            {3, 1}); // [3,1] - broadcasts to [2,3,4] (implicit leading dimension)
        Tensor<OutputType> output({2, 3, 4}); // Output: [2,3,4]

        input1.fillWithValue(safeTestTypeCast<Input1Type>(TEST_VALUE_5));

        input2.setHostValue(safeTestTypeCast<Input2Type>(TEST_VALUE_1), 0, 0); // Channel 0
        input2.setHostValue(safeTestTypeCast<Input2Type>(TEST_VALUE_2), 1, 0); // Channel 1
        input2.setHostValue(safeTestTypeCast<Input2Type>(TEST_VALUE_3), 2, 0); // Channel 2

        CpuReferencePointwiseImpl<OutputType, Input1Type, Input2Type>::pointwiseCompute(
            PointwiseMode::SUB, output, input1, input2);

        // Create expected tensor: broadcasting subtraction with implicit leading dimension
        Tensor<OutputType> expected({2, 3, 4});
        for(int n = 0; n < 2; ++n)
        {
            for(int c = 0; c < 3; ++c)
            {
                for(int h = 0; h < 4; ++h)
                {
                    const float input1Val = TEST_VALUE_5;
                    auto input2Val = static_cast<float>(c + 1); // Channel values: 1.0, 2.0, 3.0
                    expected.setHostValue(
                        safeTestTypeCast<OutputType>(input1Val - input2Val), n, c, h);
                }
            }
        }

        auto tolerance = getDynamicTolerance(PointwiseMode::SUB, TEST_VALUE_5);
        auto validator = createAllCloseValidator<OutputType>(tolerance, tolerance);
        EXPECT_TRUE(validator->allClose(expected, output));
    }

    // Test case: 4D × 4D: [N,C,H,W] + [1,C,1,1] → broadcast to [N,C,H,W]
    void testBroadcast4Dx4D()
    {
        Tensor<Input1Type> input1({2, 3, 2, 2}); // [N,C,H,W] = [2,3,2,2]
        Tensor<Input2Type> input2({1, 3, 1, 1}); // [1,C,1,1] = [1,3,1,1]
        Tensor<OutputType> output({2, 3, 2, 2}); // Output: [2,3,2,2]

        input1.fillWithValue(safeTestTypeCast<Input1Type>(TEST_VALUE_1));

        input2.setHostValue(
            safeTestTypeCast<Input2Type>(BROADCAST_MULTIPLIER_10), 0, 0, 0, 0); // Channel 0
        input2.setHostValue(safeTestTypeCast<Input2Type>(TEST_VALUE_2 * BROADCAST_MULTIPLIER_10),
                            0,
                            1,
                            0,
                            0); // Channel 1
        input2.setHostValue(safeTestTypeCast<Input2Type>(TEST_VALUE_3 * BROADCAST_MULTIPLIER_10),
                            0,
                            2,
                            0,
                            0); // Channel 2

        CpuReferencePointwiseImpl<OutputType, Input1Type, Input2Type>::pointwiseCompute(
            PointwiseMode::ADD, output, input1, input2);

        // Create expected tensor: broadcasting addition
        Tensor<OutputType> expected({2, 3, 2, 2});
        for(int n = 0; n < 2; ++n)
        {
            for(int c = 0; c < 3; ++c)
            {
                for(int h = 0; h < 2; ++h)
                {
                    for(int w = 0; w < 2; ++w)
                    {
                        const float input1Val = TEST_VALUE_1;
                        auto input2Val = static_cast<float>(
                            (static_cast<float>(c) + 1.0f)
                            * BROADCAST_MULTIPLIER_10); // Channel values: 10.0, 20.0, 30.0
                        expected.setHostValue(
                            safeTestTypeCast<OutputType>(input1Val + input2Val), n, c, h, w);
                    }
                }
            }
        }

        auto tolerance = getDynamicTolerance(PointwiseMode::ADD, 30.0f);
        auto validator = createAllCloseValidator<OutputType>(tolerance, tolerance);
        EXPECT_TRUE(validator->allClose(expected, output));
    }

    // Test case: Complex N-D broadcasting: [2,1,3,1] + [1,2,1,4] → [2,2,3,4]
    void testBroadcastComplexND()
    {
        Tensor<Input1Type> input1({2, 1, 3, 1});
        Tensor<Input2Type> input2({1, 2, 1, 4});
        Tensor<OutputType> output({2, 2, 3, 4});

        for(int n = 0; n < 2; ++n)
        {
            for(int h = 0; h < 3; ++h)
            {
                input1.setHostValue(
                    safeTestTypeCast<Input1Type>(static_cast<float>((n * 10) + h)), n, 0, h, 0);
            }
        }

        for(int c = 0; c < 2; ++c)
        {
            for(int w = 0; w < 4; ++w)
            {
                input2.setHostValue(
                    safeTestTypeCast<Input2Type>(static_cast<float>((c * 100) + w)), 0, c, 0, w);
            }
        }

        CpuReferencePointwiseImpl<OutputType, Input1Type, Input2Type>::pointwiseCompute(
            PointwiseMode::ADD, output, input1, input2);

        // Create expected tensor: broadcasting addition
        Tensor<OutputType> expected({2, 2, 3, 4});
        for(int n = 0; n < 2; ++n)
        {
            for(int c = 0; c < 2; ++c)
            {
                for(int h = 0; h < 3; ++h)
                {
                    for(int w = 0; w < 4; ++w)
                    {
                        // input1[n,0,h,0] broadcasts to input1[n,c,h,w]
                        auto input1Val = static_cast<float>((n * 10) + h);
                        // input2[0,c,0,w] broadcasts to input2[n,c,h,w]
                        auto input2Val = static_cast<float>((c * 100) + w);
                        expected.setHostValue(
                            safeTestTypeCast<OutputType>(input1Val + input2Val), n, c, h, w);
                    }
                }
            }
        }

        auto tolerance = getDynamicTolerance(PointwiseMode::ADD, 103.0f);
        auto validator = createAllCloseValidator<OutputType>(tolerance, tolerance);
        EXPECT_TRUE(validator->allClose(expected, output));
    }

    void testBroadcast5D()
    {
        const std::vector<int64_t> dims1 = {2, 3, 2, 2, 2}; // [N,C,D,H,W] = [2,3,2,2,2]
        const std::vector<int64_t> strides1 = {24, 8, 4, 2, 1}; // Row-major strides for [2,3,2,2,2]
        Tensor<Input1Type> input1(dims1, strides1);

        const std::vector<int64_t> dims2 = {1, 3, 1, 1, 1}; // [1,C,1,1,1] = [1,3,1,1,1]
        const std::vector<int64_t> strides2 = {3, 1, 1, 1, 1}; // Row-major strides for [1,3,1,1,1]
        Tensor<Input2Type> input2(dims2, strides2);

        const std::vector<int64_t> outputDims = {2, 3, 2, 2, 2}; // Output: [2,3,2,2,2]
        const std::vector<int64_t> outputStrides
            = {24, 8, 4, 2, 1}; // Row-major strides for [2,3,2,2,2]
        Tensor<OutputType> output(outputDims, outputStrides);

        input1.fillWithValue(safeTestTypeCast<Input1Type>(TEST_VALUE_2));

        // Set channel-specific values in input2
        input2.setHostValue(
            safeTestTypeCast<Input2Type>(BROADCAST_MULTIPLIER_10), 0, 0, 0, 0, 0); // Channel 0
        input2.setHostValue(safeTestTypeCast<Input2Type>(TEST_VALUE_2 * BROADCAST_MULTIPLIER_10),
                            0,
                            1,
                            0,
                            0,
                            0); // Channel 1
        input2.setHostValue(safeTestTypeCast<Input2Type>(TEST_VALUE_3 * BROADCAST_MULTIPLIER_10),
                            0,
                            2,
                            0,
                            0,
                            0); // Channel 2

        CpuReferencePointwiseImpl<OutputType, Input1Type, Input2Type>::pointwiseCompute(
            PointwiseMode::ADD, output, input1, input2);

        // Create expected tensor: broadcasting addition
        Tensor<OutputType> expected(outputDims, outputStrides);
        for(int n = 0; n < 2; ++n)
        {
            for(int c = 0; c < 3; ++c)
            {
                for(int d = 0; d < 2; ++d)
                {
                    for(int h = 0; h < 2; ++h)
                    {
                        for(int w = 0; w < 2; ++w)
                        {
                            auto input1Val = TEST_VALUE_2;
                            auto input2Val = static_cast<float>(
                                (static_cast<float>(c) + 1.0f)
                                * BROADCAST_MULTIPLIER_10); // Channel values: 10.0, 20.0, 30.0
                            expected.setHostValue(
                                safeTestTypeCast<OutputType>(input1Val + input2Val), n, c, d, h, w);
                        }
                    }
                }
            }
        }

        auto tolerance = getDynamicTolerance(PointwiseMode::ADD, 30.0f);
        auto validator = createAllCloseValidator<OutputType>(tolerance, tolerance);
        EXPECT_TRUE(validator->allClose(expected, output));
    }

    // ======================= UNARY OPERATIONS =======================

    void testReluForwardOperation()
    {
        Tensor<Input1Type> input({1, 3, 2, 2});
        Tensor<OutputType> output({1, 3, 2, 2});

        // Fill with mix of positive and negative values
        input.setHostValue(safeTestTypeCast<Input1Type>(TEST_VALUE_1), 0, 0, 0, 0); // 1.0
        input.setHostValue(safeTestTypeCast<Input1Type>(-TEST_VALUE_2), 0, 0, 0, 1); // -2.0
        input.setHostValue(safeTestTypeCast<Input1Type>(TEST_VALUE_3), 0, 0, 1, 0); // 3.0
        input.setHostValue(safeTestTypeCast<Input1Type>(-TEST_VALUE_1), 0, 0, 1, 1); // -1.0
        input.setHostValue(safeTestTypeCast<Input1Type>(0.0f), 0, 1, 0, 0); // 0.0
        input.setHostValue(safeTestTypeCast<Input1Type>(TEST_VALUE_5), 0, 1, 0, 1); // 5.0
        input.setHostValue(safeTestTypeCast<Input1Type>(-TEST_VALUE_5), 0, 1, 1, 0); // -5.0
        input.setHostValue(safeTestTypeCast<Input1Type>(TEST_VALUE_2_5), 0, 1, 1, 1); // 2.5
        input.setHostValue(safeTestTypeCast<Input1Type>(-TEST_VALUE_1_5), 0, 2, 0, 0); // -1.5
        input.setHostValue(safeTestTypeCast<Input1Type>(TEST_VALUE_4), 0, 2, 0, 1); // 4.0
        input.setHostValue(safeTestTypeCast<Input1Type>(-TEST_VALUE_3), 0, 2, 1, 0); // -3.0
        input.setHostValue(safeTestTypeCast<Input1Type>(TEST_VALUE_1_5), 0, 2, 1, 1); // 1.5

        CpuReferencePointwiseImpl<OutputType, Input1Type>::pointwiseCompute(
            PointwiseMode::RELU_FWD, output, input);

        // Create expected tensor: ReLU(x) = max(0, x)
        Tensor<OutputType> expected({1, 3, 2, 2});
        expected.setHostValue(
            safeTestTypeCast<OutputType>(TEST_VALUE_1), 0, 0, 0, 0); // max(0, 1) = 1
        expected.setHostValue(safeTestTypeCast<OutputType>(0.0f), 0, 0, 0, 1); // max(0, -2) = 0
        expected.setHostValue(
            safeTestTypeCast<OutputType>(TEST_VALUE_3), 0, 0, 1, 0); // max(0, 3) = 3
        expected.setHostValue(safeTestTypeCast<OutputType>(0.0f), 0, 0, 1, 1); // max(0, -1) = 0
        expected.setHostValue(safeTestTypeCast<OutputType>(0.0f), 0, 1, 0, 0); // max(0, 0) = 0
        expected.setHostValue(
            safeTestTypeCast<OutputType>(TEST_VALUE_5), 0, 1, 0, 1); // max(0, 5) = 5
        expected.setHostValue(safeTestTypeCast<OutputType>(0.0f), 0, 1, 1, 0); // max(0, -5) = 0
        expected.setHostValue(
            safeTestTypeCast<OutputType>(TEST_VALUE_2_5), 0, 1, 1, 1); // max(0, 2.5) = 2.5
        expected.setHostValue(safeTestTypeCast<OutputType>(0.0f), 0, 2, 0, 0); // max(0, -1.5) = 0
        expected.setHostValue(
            safeTestTypeCast<OutputType>(TEST_VALUE_4), 0, 2, 0, 1); // max(0, 4) = 4
        expected.setHostValue(safeTestTypeCast<OutputType>(0.0f), 0, 2, 1, 0); // max(0, -3) = 0
        expected.setHostValue(
            safeTestTypeCast<OutputType>(TEST_VALUE_1_5), 0, 2, 1, 1); // max(0, 1.5) = 1.5

        auto tolerance = getDynamicTolerance(PointwiseMode::RELU_FWD, TEST_VALUE_5);
        auto validator = createAllCloseValidator<OutputType>(tolerance, tolerance);
        EXPECT_TRUE(validator->allClose(expected, output));
    }

    void testReluBackwardOperation()
    {
        Tensor<Input1Type> input({1, 2, 2, 2}); // Forward input (x)
        Tensor<Input2Type> upstreamGrad({1, 2, 2, 2}); // Upstream gradient (dy)
        Tensor<OutputType> output({1, 2, 2, 2}); // Downstream gradient (dx)

        // Fill input with mix of positive and negative values
        input.setHostValue(safeTestTypeCast<Input1Type>(TEST_VALUE_1), 0, 0, 0, 0); // 1.0
        input.setHostValue(safeTestTypeCast<Input1Type>(-TEST_VALUE_2), 0, 0, 0, 1); // -2.0
        input.setHostValue(safeTestTypeCast<Input1Type>(0.0f), 0, 0, 1, 0); // 0.0
        input.setHostValue(safeTestTypeCast<Input1Type>(TEST_VALUE_3), 0, 0, 1, 1); // 3.0
        input.setHostValue(safeTestTypeCast<Input1Type>(-TEST_VALUE_1), 0, 1, 0, 0); // -1.0
        input.setHostValue(safeTestTypeCast<Input1Type>(TEST_VALUE_2_5), 0, 1, 0, 1); // 2.5
        input.setHostValue(safeTestTypeCast<Input1Type>(-TEST_VALUE_5), 0, 1, 1, 0); // -5.0
        input.setHostValue(safeTestTypeCast<Input1Type>(TEST_VALUE_1_5), 0, 1, 1, 1); // 1.5

        // Fill upstream gradient with test values
        upstreamGrad.setHostValue(safeTestTypeCast<Input2Type>(TEST_VALUE_2), 0, 0, 0, 0); // 2.0
        upstreamGrad.setHostValue(safeTestTypeCast<Input2Type>(TEST_VALUE_1_5), 0, 0, 0, 1); // 1.5
        upstreamGrad.setHostValue(safeTestTypeCast<Input2Type>(TEST_VALUE_3), 0, 0, 1, 0); // 3.0
        upstreamGrad.setHostValue(safeTestTypeCast<Input2Type>(TEST_VALUE_1), 0, 0, 1, 1); // 1.0
        upstreamGrad.setHostValue(safeTestTypeCast<Input2Type>(TEST_VALUE_2_5), 0, 1, 0, 0); // 2.5
        upstreamGrad.setHostValue(safeTestTypeCast<Input2Type>(TEST_VALUE_4), 0, 1, 0, 1); // 4.0
        upstreamGrad.setHostValue(safeTestTypeCast<Input2Type>(TEST_VALUE_1_5), 0, 1, 1, 0); // 1.5
        upstreamGrad.setHostValue(safeTestTypeCast<Input2Type>(TEST_VALUE_3), 0, 1, 1, 1); // 3.0

        // ReLU backward expects dy first, then x.
        CpuReferencePointwiseImpl<OutputType, Input2Type, Input1Type>::pointwiseCompute(
            PointwiseMode::RELU_BWD, output, upstreamGrad, input);

        // Create expected tensor: dx = dy * (x > 0 ? 1 : 0)
        Tensor<OutputType> expected({1, 2, 2, 2});
        expected.setHostValue(safeTestTypeCast<OutputType>(TEST_VALUE_2 * 1.0f),
                              0,
                              0,
                              0,
                              0); // dy=2.0, x=1.0>0: dx=2.0*1=2.0
        expected.setHostValue(safeTestTypeCast<OutputType>(TEST_VALUE_1_5 * 0.0f),
                              0,
                              0,
                              0,
                              1); // dy=1.5, x=-2.0<0: dx=1.5*0=0.0
        expected.setHostValue(safeTestTypeCast<OutputType>(TEST_VALUE_3 * 0.0f),
                              0,
                              0,
                              1,
                              0); // dy=3.0, x=0.0<=0: dx=3.0*0=0.0
        expected.setHostValue(safeTestTypeCast<OutputType>(TEST_VALUE_1 * 1.0f),
                              0,
                              0,
                              1,
                              1); // dy=1.0, x=3.0>0: dx=1.0*1=1.0
        expected.setHostValue(safeTestTypeCast<OutputType>(TEST_VALUE_2_5 * 0.0f),
                              0,
                              1,
                              0,
                              0); // dy=2.5, x=-1.0<0: dx=2.5*0=0.0
        expected.setHostValue(safeTestTypeCast<OutputType>(TEST_VALUE_4 * 1.0f),
                              0,
                              1,
                              0,
                              1); // dy=4.0, x=2.5>0: dx=4.0*1=4.0
        expected.setHostValue(safeTestTypeCast<OutputType>(TEST_VALUE_1_5 * 0.0f),
                              0,
                              1,
                              1,
                              0); // dy=1.5, x=-5.0<0: dx=1.5*0=0.0
        expected.setHostValue(safeTestTypeCast<OutputType>(TEST_VALUE_3 * 1.0f),
                              0,
                              1,
                              1,
                              1); // dy=3.0, x=1.5>0: dx=3.0*1=3.0

        auto tolerance = getDynamicTolerance(PointwiseMode::RELU_BWD, TEST_VALUE_5);
        auto validator = createAllCloseValidator<OutputType>(tolerance, tolerance);
        EXPECT_TRUE(validator->allClose(expected, output));
    }

    void testParameterizedReluForwardOperation()
    {
        Tensor<Input1Type> input({1, 2, 2, 2});
        Tensor<OutputType> output({1, 2, 2, 2});

        // Fill with values that will test all three regions: below lower_clip, between clips, above upper_clip
        input.setHostValue(
            safeTestTypeCast<Input1Type>(-TEST_VALUE_5), 0, 0, 0, 0); // -5.0 (below lower_clip)
        input.setHostValue(
            safeTestTypeCast<Input1Type>(-TEST_VALUE_1), 0, 0, 0, 1); // -1.0 (between clips)
        input.setHostValue(
            safeTestTypeCast<Input1Type>(TEST_VALUE_2), 0, 0, 1, 0); // 2.0 (between clips)
        input.setHostValue(
            safeTestTypeCast<Input1Type>(TEST_VALUE_5), 0, 0, 1, 1); // 5.0 (above upper_clip)
        input.setHostValue(
            safeTestTypeCast<Input1Type>(-TEST_VALUE_2), 0, 1, 0, 0); // -2.0 (at lower_clip)
        input.setHostValue(safeTestTypeCast<Input1Type>(0.0f), 0, 1, 0, 1); // 0.0 (between clips)
        input.setHostValue(
            safeTestTypeCast<Input1Type>(TEST_VALUE_4), 0, 1, 1, 0); // 4.0 (at upper_clip)
        input.setHostValue(
            safeTestTypeCast<Input1Type>(TEST_VALUE_1), 0, 1, 1, 1); // 1.0 (between clips)

        // Parameters: lower_clip = -2.0, upper_clip = 4.0, lower_slope = 0.1
        const float lowerClip = -TEST_VALUE_2; // -2.0
        const float upperClip = TEST_VALUE_4; // 4.0
        auto lowerSlope = static_cast<float>(0.1);

        CpuReferencePointwiseImpl<OutputType, Input1Type>::pointwiseCompute(
            PointwiseMode::RELU_FWD, output, input, lowerClip, upperClip, lowerSlope);

        // Create expected tensor: Generalized ReLU with parameters
        Tensor<OutputType> expected({1, 2, 2, 2});
        // For x < lower_clip: output = lower_clip + lower_slope * (x - lower_clip)
        // For lower_clip <= x <= upper_clip: output = x
        // For x > upper_clip: output = upper_clip
        expected.setHostValue(
            safeTestTypeCast<OutputType>(lowerClip + (lowerSlope * (-TEST_VALUE_5 - lowerClip))),
            0,
            0,
            0,
            0); // -2 + 0.1*(-5-(-2)) = -2.3
        expected.setHostValue(
            safeTestTypeCast<OutputType>(-TEST_VALUE_1), 0, 0, 0, 1); // -1.0 (in range)
        expected.setHostValue(
            safeTestTypeCast<OutputType>(TEST_VALUE_2), 0, 0, 1, 0); // 2.0 (in range)
        expected.setHostValue(safeTestTypeCast<OutputType>(upperClip), 0, 0, 1, 1); // 4.0 (clamped)
        expected.setHostValue(
            safeTestTypeCast<OutputType>(lowerClip), 0, 1, 0, 0); // -2.0 (at boundary)
        expected.setHostValue(safeTestTypeCast<OutputType>(0.0f), 0, 1, 0, 1); // 0.0 (in range)
        expected.setHostValue(
            safeTestTypeCast<OutputType>(upperClip), 0, 1, 1, 0); // 4.0 (at boundary)
        expected.setHostValue(
            safeTestTypeCast<OutputType>(TEST_VALUE_1), 0, 1, 1, 1); // 1.0 (in range)

        auto tolerance = getDynamicTolerance(PointwiseMode::RELU_FWD, TEST_VALUE_5);
        auto validator = createAllCloseValidator<OutputType>(tolerance, tolerance);
        EXPECT_TRUE(validator->allClose(expected, output));
    }

    void testParameterizedReluBackwardOperation()
    {
        Tensor<Input1Type> input({1, 2, 2, 2}); // Forward input (x)
        Tensor<Input2Type> upstreamGrad({1, 2, 2, 2}); // Upstream gradient (dy)
        Tensor<OutputType> output({1, 2, 2, 2}); // Downstream gradient (dx)

        // Fill forward input with values that will test all three regions: below lower_clip, between clips, above upper_clip
        input.setHostValue(
            safeTestTypeCast<Input1Type>(-TEST_VALUE_5), 0, 0, 0, 0); // -5.0 (below lower_clip)
        input.setHostValue(
            safeTestTypeCast<Input1Type>(-TEST_VALUE_1), 0, 0, 0, 1); // -1.0 (between clips)
        input.setHostValue(
            safeTestTypeCast<Input1Type>(TEST_VALUE_2), 0, 0, 1, 0); // 2.0 (between clips)
        input.setHostValue(
            safeTestTypeCast<Input1Type>(TEST_VALUE_5), 0, 0, 1, 1); // 5.0 (above upper_clip)
        input.setHostValue(
            safeTestTypeCast<Input1Type>(-TEST_VALUE_2), 0, 1, 0, 0); // -2.0 (at lower_clip)
        input.setHostValue(safeTestTypeCast<Input1Type>(0.0f), 0, 1, 0, 1); // 0.0 (between clips)
        input.setHostValue(
            safeTestTypeCast<Input1Type>(TEST_VALUE_4), 0, 1, 1, 0); // 4.0 (at upper_clip)
        input.setHostValue(
            safeTestTypeCast<Input1Type>(TEST_VALUE_1), 0, 1, 1, 1); // 1.0 (between clips)

        // Fill upstream gradient with test values
        upstreamGrad.setHostValue(safeTestTypeCast<Input2Type>(TEST_VALUE_2), 0, 0, 0, 0); // 2.0
        upstreamGrad.setHostValue(safeTestTypeCast<Input2Type>(TEST_VALUE_1_5), 0, 0, 0, 1); // 1.5
        upstreamGrad.setHostValue(safeTestTypeCast<Input2Type>(TEST_VALUE_3), 0, 0, 1, 0); // 3.0
        upstreamGrad.setHostValue(safeTestTypeCast<Input2Type>(TEST_VALUE_1), 0, 0, 1, 1); // 1.0
        upstreamGrad.setHostValue(safeTestTypeCast<Input2Type>(TEST_VALUE_2_5), 0, 1, 0, 0); // 2.5
        upstreamGrad.setHostValue(safeTestTypeCast<Input2Type>(TEST_VALUE_4), 0, 1, 0, 1); // 4.0
        upstreamGrad.setHostValue(safeTestTypeCast<Input2Type>(TEST_VALUE_1_5), 0, 1, 1, 0); // 1.5
        upstreamGrad.setHostValue(safeTestTypeCast<Input2Type>(TEST_VALUE_3), 0, 1, 1, 1); // 3.0

        // Parameters: lower_clip = -2.0, upper_clip = 4.0, lower_slope = 0.1
        const float lowerClip = -TEST_VALUE_2; // -2.0
        const float upperClip = TEST_VALUE_4; // 4.0
        auto lowerSlope = static_cast<float>(0.1);

        // ReLU backward expects dy first, then x.
        CpuReferencePointwiseImpl<OutputType, Input2Type, Input1Type>::pointwiseCompute(
            PointwiseMode::RELU_BWD, output, upstreamGrad, input, lowerClip, upperClip, lowerSlope);

        // Create expected tensor: dx = dy * local_gradient
        Tensor<OutputType> expected({1, 2, 2, 2});
        for(int n = 0; n < 1; ++n)
        {
            for(int c = 0; c < 2; ++c)
            {
                for(int h = 0; h < 2; ++h)
                {
                    for(int w = 0; w < 2; ++w)
                    {
                        auto inputVal = static_cast<float>(input.getHostValue(n, c, h, w));
                        auto upstreamGradVal
                            = static_cast<float>(upstreamGrad.getHostValue(n, c, h, w));

                        // For x <= lower_clip: local_gradient = lower_slope
                        // For x > lower_clip && x <= upper_clip: local_gradient = 1.0
                        // For x > upper_clip: local_gradient = 0.0
                        float localGradient;
                        if(inputVal > lowerClip && inputVal <= upperClip)
                        {
                            localGradient = 1.0f;
                        }
                        else if(inputVal <= lowerClip)
                        {
                            localGradient = lowerSlope;
                        }
                        else
                        {
                            localGradient = 0.0f;
                        }

                        auto downstreamGrad = upstreamGradVal * localGradient;
                        expected.setHostValue(
                            safeTestTypeCast<OutputType>(downstreamGrad), n, c, h, w);
                    }
                }
            }
        }

        auto tolerance = getDynamicTolerance(PointwiseMode::RELU_BWD, TEST_VALUE_5);
        auto validator = createAllCloseValidator<OutputType>(tolerance, tolerance);
        EXPECT_TRUE(validator->allClose(expected, output));
    }

    void testSigmoidForwardOperation()
    {
        Tensor<Input1Type> input({1, 2, 2, 2});
        Tensor<OutputType> output({1, 2, 2, 2});

        // Fill with values that will test sigmoid function
        input.setHostValue(safeTestTypeCast<Input1Type>(0.0f), 0, 0, 0, 0); // 0.0
        input.setHostValue(safeTestTypeCast<Input1Type>(TEST_VALUE_1), 0, 0, 0, 1); // 1.0
        input.setHostValue(safeTestTypeCast<Input1Type>(-TEST_VALUE_1), 0, 0, 1, 0); // -1.0
        input.setHostValue(safeTestTypeCast<Input1Type>(TEST_VALUE_2), 0, 0, 1, 1); // 2.0
        input.setHostValue(safeTestTypeCast<Input1Type>(-TEST_VALUE_2), 0, 1, 0, 0); // -2.0
        input.setHostValue(safeTestTypeCast<Input1Type>(TEST_VALUE_5), 0, 1, 0, 1); // 5.0
        input.setHostValue(safeTestTypeCast<Input1Type>(-TEST_VALUE_5), 0, 1, 1, 0); // -5.0
        input.setHostValue(safeTestTypeCast<Input1Type>(TEST_VALUE_1_5), 0, 1, 1, 1); // 1.5

        CpuReferencePointwiseImpl<OutputType, Input1Type>::pointwiseCompute(
            PointwiseMode::SIGMOID_FWD, output, input);

        // Create expected tensor: sigmoid(x) = 1 / (1 + exp(-x))
        Tensor<OutputType> expected({1, 2, 2, 2});
        expected.setHostValue(
            safeTestTypeCast<OutputType>(1.0f / (1.0f + hipdnn_data_sdk::types::exp(-0.0f))),
            0,
            0,
            0,
            0); // sigmoid(0) = 0.5
        expected.setHostValue(safeTestTypeCast<OutputType>(
                                  1.0f / (1.0f + hipdnn_data_sdk::types::exp(-TEST_VALUE_1))),
                              0,
                              0,
                              0,
                              1); // sigmoid(1)
        expected.setHostValue(
            safeTestTypeCast<OutputType>(1.0f / (1.0f + hipdnn_data_sdk::types::exp(TEST_VALUE_1))),
            0,
            0,
            1,
            0); // sigmoid(-1)
        expected.setHostValue(safeTestTypeCast<OutputType>(
                                  1.0f / (1.0f + hipdnn_data_sdk::types::exp(-TEST_VALUE_2))),
                              0,
                              0,
                              1,
                              1); // sigmoid(2)
        expected.setHostValue(
            safeTestTypeCast<OutputType>(1.0f / (1.0f + hipdnn_data_sdk::types::exp(TEST_VALUE_2))),
            0,
            1,
            0,
            0); // sigmoid(-2)
        expected.setHostValue(safeTestTypeCast<OutputType>(
                                  1.0f / (1.0f + hipdnn_data_sdk::types::exp(-TEST_VALUE_5))),
                              0,
                              1,
                              0,
                              1); // sigmoid(5)
        expected.setHostValue(
            safeTestTypeCast<OutputType>(1.0f / (1.0f + hipdnn_data_sdk::types::exp(TEST_VALUE_5))),
            0,
            1,
            1,
            0); // sigmoid(-5)
        expected.setHostValue(safeTestTypeCast<OutputType>(
                                  1.0f / (1.0f + hipdnn_data_sdk::types::exp(-TEST_VALUE_1_5))),
                              0,
                              1,
                              1,
                              1); // sigmoid(1.5)

        auto tolerance = getDynamicTolerance(PointwiseMode::SIGMOID_FWD, TEST_VALUE_5);
        auto validator = createAllCloseValidator<OutputType>(tolerance, tolerance);
        EXPECT_TRUE(validator->allClose(expected, output));
    }

    void testSigmoidBackwardOperation()
    {
        Tensor<Input1Type> input({1, 2, 2, 2}); // Forward input (x)
        Tensor<Input2Type> upstreamGrad({1, 2, 2, 2}); // Upstream gradient (dy)
        Tensor<OutputType> output({1, 2, 2, 2}); // Downstream gradient (dx)

        // Fill forward input with values that will test sigmoid backward function
        input.setHostValue(safeTestTypeCast<Input1Type>(0.0f), 0, 0, 0, 0); // 0.0
        input.setHostValue(safeTestTypeCast<Input1Type>(TEST_VALUE_1), 0, 0, 0, 1); // 1.0
        input.setHostValue(safeTestTypeCast<Input1Type>(-TEST_VALUE_1), 0, 0, 1, 0); // -1.0
        input.setHostValue(safeTestTypeCast<Input1Type>(TEST_VALUE_2), 0, 0, 1, 1); // 2.0
        input.setHostValue(safeTestTypeCast<Input1Type>(-TEST_VALUE_2), 0, 1, 0, 0); // -2.0
        input.setHostValue(safeTestTypeCast<Input1Type>(TEST_VALUE_3), 0, 1, 0, 1); // 3.0
        input.setHostValue(safeTestTypeCast<Input1Type>(-TEST_VALUE_3), 0, 1, 1, 0); // -3.0
        input.setHostValue(safeTestTypeCast<Input1Type>(TEST_VALUE_1_5), 0, 1, 1, 1); // 1.5

        // Fill upstream gradient with test values
        upstreamGrad.setHostValue(safeTestTypeCast<Input2Type>(TEST_VALUE_2), 0, 0, 0, 0); // 2.0
        upstreamGrad.setHostValue(safeTestTypeCast<Input2Type>(TEST_VALUE_1_5), 0, 0, 0, 1); // 1.5
        upstreamGrad.setHostValue(safeTestTypeCast<Input2Type>(TEST_VALUE_3), 0, 0, 1, 0); // 3.0
        upstreamGrad.setHostValue(safeTestTypeCast<Input2Type>(TEST_VALUE_1), 0, 0, 1, 1); // 1.0
        upstreamGrad.setHostValue(safeTestTypeCast<Input2Type>(TEST_VALUE_2_5), 0, 1, 0, 0); // 2.5
        upstreamGrad.setHostValue(safeTestTypeCast<Input2Type>(TEST_VALUE_4), 0, 1, 0, 1); // 4.0
        upstreamGrad.setHostValue(safeTestTypeCast<Input2Type>(TEST_VALUE_1_5), 0, 1, 1, 0); // 1.5
        upstreamGrad.setHostValue(safeTestTypeCast<Input2Type>(TEST_VALUE_3), 0, 1, 1, 1); // 3.0

        // Backward activations expect dy first, then x.
        CpuReferencePointwiseImpl<OutputType, Input2Type, Input1Type>::pointwiseCompute(
            PointwiseMode::SIGMOID_BWD, output, upstreamGrad, input);

        // Create expected tensor: dx = dy * sigmoid(x) * (1 - sigmoid(x))
        Tensor<OutputType> expected({1, 2, 2, 2});
        for(int n = 0; n < 1; ++n)
        {
            for(int c = 0; c < 2; ++c)
            {
                for(int h = 0; h < 2; ++h)
                {
                    for(int w = 0; w < 2; ++w)
                    {
                        auto inputVal = input.getHostValue(n, c, h, w);
                        auto upstreamGradVal = upstreamGrad.getHostValue(n, c, h, w);
                        // Compute using the template ComputeType
                        auto xCompute = static_cast<ComputeType>(inputVal);
                        auto dyCompute = static_cast<ComputeType>(upstreamGradVal);
                        auto sigmoid = ComputeType{1}
                                       / (ComputeType{1} + hipdnn_data_sdk::types::exp(-xCompute));
                        auto localGradient = sigmoid * (ComputeType{1} - sigmoid);
                        auto downstreamGrad = dyCompute * localGradient;
                        expected.setHostValue(
                            safeTestTypeCast<OutputType>(downstreamGrad), n, c, h, w);
                    }
                }
            }
        }

        auto tolerance = getDynamicTolerance(PointwiseMode::SIGMOID_BWD, TEST_VALUE_4);
        auto validator = createAllCloseValidator<OutputType>(tolerance, tolerance);
        EXPECT_TRUE(validator->allClose(expected, output));
    }

    void testTanhForwardOperation()
    {
        Tensor<Input1Type> input({1, 2, 2, 2});
        Tensor<OutputType> output({1, 2, 2, 2});

        // Fill with values that will test tanh function
        input.setHostValue(safeTestTypeCast<Input1Type>(0.0f), 0, 0, 0, 0); // 0.0
        input.setHostValue(safeTestTypeCast<Input1Type>(TEST_VALUE_1), 0, 0, 0, 1); // 1.0
        input.setHostValue(safeTestTypeCast<Input1Type>(-TEST_VALUE_1), 0, 0, 1, 0); // -1.0
        input.setHostValue(safeTestTypeCast<Input1Type>(TEST_VALUE_2), 0, 0, 1, 1); // 2.0
        input.setHostValue(safeTestTypeCast<Input1Type>(-TEST_VALUE_2), 0, 1, 0, 0); // -2.0
        input.setHostValue(safeTestTypeCast<Input1Type>(TEST_VALUE_3), 0, 1, 0, 1); // 3.0
        input.setHostValue(safeTestTypeCast<Input1Type>(-TEST_VALUE_3), 0, 1, 1, 0); // -3.0
        input.setHostValue(safeTestTypeCast<Input1Type>(TEST_VALUE_1_5), 0, 1, 1, 1); // 1.5

        CpuReferencePointwiseImpl<OutputType, Input1Type>::pointwiseCompute(
            PointwiseMode::TANH_FWD, output, input);

        // Create expected tensor: tanh(x)
        Tensor<OutputType> expected({1, 2, 2, 2});
        expected.setHostValue(
            safeTestTypeCast<OutputType>(std::tanh(0.0f)), 0, 0, 0, 0); // tanh(0) = 0
        expected.setHostValue(
            safeTestTypeCast<OutputType>(std::tanh(TEST_VALUE_1)), 0, 0, 0, 1); // tanh(1)
        expected.setHostValue(
            safeTestTypeCast<OutputType>(std::tanh(-TEST_VALUE_1)), 0, 0, 1, 0); // tanh(-1)
        expected.setHostValue(
            safeTestTypeCast<OutputType>(std::tanh(TEST_VALUE_2)), 0, 0, 1, 1); // tanh(2)
        expected.setHostValue(
            safeTestTypeCast<OutputType>(std::tanh(-TEST_VALUE_2)), 0, 1, 0, 0); // tanh(-2)
        expected.setHostValue(
            safeTestTypeCast<OutputType>(std::tanh(TEST_VALUE_3)), 0, 1, 0, 1); // tanh(3)
        expected.setHostValue(
            safeTestTypeCast<OutputType>(std::tanh(-TEST_VALUE_3)), 0, 1, 1, 0); // tanh(-3)
        expected.setHostValue(
            safeTestTypeCast<OutputType>(std::tanh(TEST_VALUE_1_5)), 0, 1, 1, 1); // tanh(1.5)

        auto tolerance = getDynamicTolerance(PointwiseMode::TANH_FWD, TEST_VALUE_3);
        auto validator = createAllCloseValidator<OutputType>(tolerance, tolerance);
        EXPECT_TRUE(validator->allClose(expected, output));
    }

    void testTanhBackwardOperation()
    {
        Tensor<Input1Type> input({1, 2, 2, 2}); // Forward input (x)
        Tensor<Input2Type> upstreamGrad({1, 2, 2, 2}); // Upstream gradient (dy)
        Tensor<OutputType> output({1, 2, 2, 2}); // Downstream gradient (dx)

        // Fill forward input with values that will test tanh backward function
        input.setHostValue(safeTestTypeCast<Input1Type>(0.0f), 0, 0, 0, 0); // 0.0
        input.setHostValue(safeTestTypeCast<Input1Type>(TEST_VALUE_1), 0, 0, 0, 1); // 1.0
        input.setHostValue(safeTestTypeCast<Input1Type>(-TEST_VALUE_1), 0, 0, 1, 0); // -1.0
        input.setHostValue(safeTestTypeCast<Input1Type>(TEST_VALUE_2), 0, 0, 1, 1); // 2.0
        input.setHostValue(safeTestTypeCast<Input1Type>(-TEST_VALUE_2), 0, 1, 0, 0); // -2.0
        input.setHostValue(safeTestTypeCast<Input1Type>(TEST_VALUE_3), 0, 1, 0, 1); // 3.0
        input.setHostValue(safeTestTypeCast<Input1Type>(-TEST_VALUE_3), 0, 1, 1, 0); // -3.0
        input.setHostValue(safeTestTypeCast<Input1Type>(TEST_VALUE_1_5), 0, 1, 1, 1); // 1.5

        // Fill upstream gradient with test values
        upstreamGrad.setHostValue(safeTestTypeCast<Input2Type>(TEST_VALUE_2), 0, 0, 0, 0); // 2.0
        upstreamGrad.setHostValue(safeTestTypeCast<Input2Type>(TEST_VALUE_1_5), 0, 0, 0, 1); // 1.5
        upstreamGrad.setHostValue(safeTestTypeCast<Input2Type>(TEST_VALUE_3), 0, 0, 1, 0); // 3.0
        upstreamGrad.setHostValue(safeTestTypeCast<Input2Type>(TEST_VALUE_1), 0, 0, 1, 1); // 1.0
        upstreamGrad.setHostValue(safeTestTypeCast<Input2Type>(TEST_VALUE_2_5), 0, 1, 0, 0); // 2.5
        upstreamGrad.setHostValue(safeTestTypeCast<Input2Type>(TEST_VALUE_4), 0, 1, 0, 1); // 4.0
        upstreamGrad.setHostValue(safeTestTypeCast<Input2Type>(TEST_VALUE_1_5), 0, 1, 1, 0); // 1.5
        upstreamGrad.setHostValue(safeTestTypeCast<Input2Type>(TEST_VALUE_3), 0, 1, 1, 1); // 3.0

        // Backward activations expect dy first, then x.
        CpuReferencePointwiseImpl<OutputType, Input2Type, Input1Type>::pointwiseCompute(
            PointwiseMode::TANH_BWD, output, upstreamGrad, input);

        // Create expected tensor: dx = dy * (1 - tanh²(x))
        // The functor computes in double (default ComputeType), so we must match that
        Tensor<OutputType> expected({1, 2, 2, 2});
        for(int n = 0; n < 1; ++n)
        {
            for(int c = 0; c < 2; ++c)
            {
                for(int h = 0; h < 2; ++h)
                {
                    for(int w = 0; w < 2; ++w)
                    {
                        auto inputVal = input.getHostValue(n, c, h, w);
                        auto upstreamGradVal = upstreamGrad.getHostValue(n, c, h, w);
                        // Compute using the template ComputeType
                        auto xCompute = static_cast<ComputeType>(inputVal);
                        auto dyCompute = static_cast<ComputeType>(upstreamGradVal);
                        auto tanhVal = std::tanh(xCompute);
                        auto localGradient = ComputeType{1} - (tanhVal * tanhVal);
                        auto downstreamGrad = dyCompute * localGradient;
                        expected.setHostValue(
                            safeTestTypeCast<OutputType>(downstreamGrad), n, c, h, w);
                    }
                }
            }
        }

        auto tolerance = getDynamicTolerance(PointwiseMode::TANH_BWD, TEST_VALUE_4);
        auto validator = createAllCloseValidator<OutputType>(tolerance, tolerance);
        EXPECT_TRUE(validator->allClose(expected, output));
    }

    void testAbsoluteValueOperation()
    {
        Tensor<Input1Type> input({1, 2, 2, 2});
        Tensor<OutputType> output({1, 2, 2, 2});

        // Fill with mix of positive, negative, and zero values
        input.setHostValue(safeTestTypeCast<Input1Type>(TEST_VALUE_1), 0, 0, 0, 0); // 1.0
        input.setHostValue(safeTestTypeCast<Input1Type>(-TEST_VALUE_2), 0, 0, 0, 1); // -2.0
        input.setHostValue(safeTestTypeCast<Input1Type>(0.0f), 0, 0, 1, 0); // 0.0
        input.setHostValue(safeTestTypeCast<Input1Type>(-TEST_VALUE_3), 0, 0, 1, 1); // -3.0
        input.setHostValue(safeTestTypeCast<Input1Type>(TEST_VALUE_5), 0, 1, 0, 0); // 5.0
        input.setHostValue(safeTestTypeCast<Input1Type>(-TEST_VALUE_1_5), 0, 1, 0, 1); // -1.5
        input.setHostValue(safeTestTypeCast<Input1Type>(TEST_VALUE_2_5), 0, 1, 1, 0); // 2.5
        input.setHostValue(safeTestTypeCast<Input1Type>(-TEST_VALUE_4), 0, 1, 1, 1); // -4.0

        CpuReferencePointwiseImpl<OutputType, Input1Type>::pointwiseCompute(
            PointwiseMode::ABS, output, input);

        // Create expected tensor: abs(x)
        Tensor<OutputType> expected({1, 2, 2, 2});
        expected.setHostValue(
            safeTestTypeCast<OutputType>(hipdnn_data_sdk::types::abs(TEST_VALUE_1)),
            0,
            0,
            0,
            0); // |1| = 1
        expected.setHostValue(
            safeTestTypeCast<OutputType>(hipdnn_data_sdk::types::abs(-TEST_VALUE_2)),
            0,
            0,
            0,
            1); // |-2| = 2
        expected.setHostValue(
            safeTestTypeCast<OutputType>(hipdnn_data_sdk::types::abs(0.0f)), 0, 0, 1, 0); // |0| = 0
        expected.setHostValue(
            safeTestTypeCast<OutputType>(hipdnn_data_sdk::types::abs(-TEST_VALUE_3)),
            0,
            0,
            1,
            1); // |-3| = 3
        expected.setHostValue(
            safeTestTypeCast<OutputType>(hipdnn_data_sdk::types::abs(TEST_VALUE_5)),
            0,
            1,
            0,
            0); // |5| = 5
        expected.setHostValue(
            safeTestTypeCast<OutputType>(hipdnn_data_sdk::types::abs(-TEST_VALUE_1_5)),
            0,
            1,
            0,
            1); // |-1.5| = 1.5
        expected.setHostValue(
            safeTestTypeCast<OutputType>(hipdnn_data_sdk::types::abs(TEST_VALUE_2_5)),
            0,
            1,
            1,
            0); // |2.5| = 2.5
        expected.setHostValue(
            safeTestTypeCast<OutputType>(hipdnn_data_sdk::types::abs(-TEST_VALUE_4)),
            0,
            1,
            1,
            1); // |-4| = 4

        auto tolerance = getDynamicTolerance(PointwiseMode::ABS, TEST_VALUE_5);
        auto validator = createAllCloseValidator<OutputType>(tolerance, tolerance);
        EXPECT_TRUE(validator->allClose(expected, output));
    }

    void testNegationOperation()
    {
        Tensor<Input1Type> input({1, 2, 2, 2});
        Tensor<OutputType> output({1, 2, 2, 2});

        // Fill with mix of positive, negative, and zero values
        input.setHostValue(safeTestTypeCast<Input1Type>(TEST_VALUE_1), 0, 0, 0, 0); // 1.0
        input.setHostValue(safeTestTypeCast<Input1Type>(-TEST_VALUE_2), 0, 0, 0, 1); // -2.0
        input.setHostValue(safeTestTypeCast<Input1Type>(0.0f), 0, 0, 1, 0); // 0.0
        input.setHostValue(safeTestTypeCast<Input1Type>(-TEST_VALUE_3), 0, 0, 1, 1); // -3.0
        input.setHostValue(safeTestTypeCast<Input1Type>(TEST_VALUE_5), 0, 1, 0, 0); // 5.0
        input.setHostValue(safeTestTypeCast<Input1Type>(-TEST_VALUE_1_5), 0, 1, 0, 1); // -1.5
        input.setHostValue(safeTestTypeCast<Input1Type>(TEST_VALUE_2_5), 0, 1, 1, 0); // 2.5
        input.setHostValue(safeTestTypeCast<Input1Type>(-TEST_VALUE_4), 0, 1, 1, 1); // -4.0

        CpuReferencePointwiseImpl<OutputType, Input1Type>::pointwiseCompute(
            PointwiseMode::NEG, output, input);

        // Create expected tensor: -x
        Tensor<OutputType> expected({1, 2, 2, 2});
        expected.setHostValue(safeTestTypeCast<OutputType>(-TEST_VALUE_1), 0, 0, 0, 0); // -1
        expected.setHostValue(safeTestTypeCast<OutputType>(TEST_VALUE_2), 0, 0, 0, 1); // -(-2) = 2
        expected.setHostValue(safeTestTypeCast<OutputType>(-0.0f), 0, 0, 1, 0); // -0 = 0
        expected.setHostValue(safeTestTypeCast<OutputType>(TEST_VALUE_3), 0, 0, 1, 1); // -(-3) = 3
        expected.setHostValue(safeTestTypeCast<OutputType>(-TEST_VALUE_5), 0, 1, 0, 0); // -5
        expected.setHostValue(
            safeTestTypeCast<OutputType>(TEST_VALUE_1_5), 0, 1, 0, 1); // -(-1.5) = 1.5
        expected.setHostValue(safeTestTypeCast<OutputType>(-TEST_VALUE_2_5), 0, 1, 1, 0); // -2.5
        expected.setHostValue(safeTestTypeCast<OutputType>(TEST_VALUE_4), 0, 1, 1, 1); // -(-4) = 4

        auto tolerance = getDynamicTolerance(PointwiseMode::NEG, TEST_VALUE_5);
        auto validator = createAllCloseValidator<OutputType>(tolerance, tolerance);
        EXPECT_TRUE(validator->allClose(expected, output));
    }

    void testGeluForwardOperation()
    {
        Tensor<Input1Type> input({1, 2, 2, 2});
        Tensor<OutputType> output({1, 2, 2, 2});

        // Fill with values that will test gelu function
        input.setHostValue(safeTestTypeCast<Input1Type>(0.0f), 0, 0, 0, 0); // 0.0
        input.setHostValue(safeTestTypeCast<Input1Type>(TEST_VALUE_1), 0, 0, 0, 1); // 1.0
        input.setHostValue(safeTestTypeCast<Input1Type>(-TEST_VALUE_1), 0, 0, 1, 0); // -1.0
        input.setHostValue(safeTestTypeCast<Input1Type>(TEST_VALUE_2), 0, 0, 1, 1); // 2.0
        input.setHostValue(safeTestTypeCast<Input1Type>(-TEST_VALUE_2), 0, 1, 0, 0); // -2.0
        input.setHostValue(safeTestTypeCast<Input1Type>(TEST_VALUE_5), 0, 1, 0, 1); // 5.0
        input.setHostValue(safeTestTypeCast<Input1Type>(-TEST_VALUE_5), 0, 1, 1, 0); // -5.0
        input.setHostValue(safeTestTypeCast<Input1Type>(TEST_VALUE_1_5), 0, 1, 1, 1); // 1.5

        CpuReferencePointwiseImpl<OutputType, Input1Type>::pointwiseCompute(
            PointwiseMode::GELU_FWD, output, input);

        // Create expected tensor: gelu_erf(x) = 0.5 * x * (1 + erf(x / sqrt(2)))
        auto geluErf = [](float x) {
            auto xCompute = static_cast<ComputeType>(safeTestTypeCast<Input1Type>(x));
            ComputeType sqrt2 = hipdnn_data_sdk::types::sqrt(ComputeType{2});
            return ComputeType{0.5} * xCompute
                   * (ComputeType{1.0} + hipdnn_data_sdk::types::erf(xCompute / sqrt2));
        };
        Tensor<OutputType> expected({1, 2, 2, 2});
        expected.setHostValue(safeTestTypeCast<OutputType>(geluErf(0.0f)),
                              0,
                              0,
                              0,
                              0); // gelu_erf(0) = 0.5
        expected.setHostValue(safeTestTypeCast<OutputType>(geluErf(TEST_VALUE_1)),
                              0,
                              0,
                              0,
                              1); // gelu_erf(1)
        expected.setHostValue(safeTestTypeCast<OutputType>(geluErf(-TEST_VALUE_1)),
                              0,
                              0,
                              1,
                              0); // gelu_erf(-1)
        expected.setHostValue(safeTestTypeCast<OutputType>(geluErf(TEST_VALUE_2)),
                              0,
                              0,
                              1,
                              1); // gelu_erf(2)
        expected.setHostValue(safeTestTypeCast<OutputType>(geluErf(-TEST_VALUE_2)),
                              0,
                              1,
                              0,
                              0); // gelu_erf(-2)
        expected.setHostValue(safeTestTypeCast<OutputType>(geluErf(TEST_VALUE_5)),
                              0,
                              1,
                              0,
                              1); // gelu_erf(5)
        expected.setHostValue(safeTestTypeCast<OutputType>(geluErf(-TEST_VALUE_5)),
                              0,
                              1,
                              1,
                              0); // gelu_erf(-5)
        expected.setHostValue(safeTestTypeCast<OutputType>(geluErf(TEST_VALUE_1_5)),
                              0,
                              1,
                              1,
                              1); // gelu_erf(1.5)

        auto tolerance = getDynamicTolerance(PointwiseMode::GELU_FWD, TEST_VALUE_5);
        auto validator = createAllCloseValidator<OutputType>(tolerance, tolerance);
        EXPECT_TRUE(validator->allClose(expected, output));
    }

    void testGeluApproxTanhForwardOperation()
    {
        Tensor<Input1Type> input({1, 2, 2, 2});
        Tensor<OutputType> output({1, 2, 2, 2});

        input.setHostValue(safeTestTypeCast<Input1Type>(0.0f), 0, 0, 0, 0);
        input.setHostValue(safeTestTypeCast<Input1Type>(TEST_VALUE_1), 0, 0, 0, 1);
        input.setHostValue(safeTestTypeCast<Input1Type>(-TEST_VALUE_1), 0, 0, 1, 0);
        input.setHostValue(safeTestTypeCast<Input1Type>(TEST_VALUE_2), 0, 0, 1, 1);
        input.setHostValue(safeTestTypeCast<Input1Type>(-TEST_VALUE_2), 0, 1, 0, 0);
        input.setHostValue(safeTestTypeCast<Input1Type>(TEST_VALUE_5), 0, 1, 0, 1);
        input.setHostValue(safeTestTypeCast<Input1Type>(-TEST_VALUE_5), 0, 1, 1, 0);
        input.setHostValue(safeTestTypeCast<Input1Type>(TEST_VALUE_1_5), 0, 1, 1, 1);

        CpuReferencePointwiseImpl<OutputType, Input1Type>::pointwiseCompute(
            PointwiseMode::GELU_APPROX_TANH_FWD, output, input);

        // gelu_tanh(x) ≈ 0.5 * x * (1 + tanh(sqrt(2/π) * (x + 0.044715 * x³)))
        auto geluTanh = [](float x) {
            auto xCompute = static_cast<ComputeType>(safeTestTypeCast<Input1Type>(x));
            const auto sqrt2OverPi = static_cast<ComputeType>(
                hipdnn_data_sdk::types::sqrt(ComputeType{2} / static_cast<ComputeType>(PI)));
            static const auto s_kCoeff = ComputeType{0.044715};
            auto inner = sqrt2OverPi * (xCompute + s_kCoeff * xCompute * xCompute * xCompute);
            return ComputeType{0.5} * xCompute
                   * (ComputeType{1.0} + hipdnn_data_sdk::types::tanh(inner));
        };
        Tensor<OutputType> expected({1, 2, 2, 2});
        expected.setHostValue(safeTestTypeCast<OutputType>(geluTanh(0.0f)), 0, 0, 0, 0);
        expected.setHostValue(safeTestTypeCast<OutputType>(geluTanh(TEST_VALUE_1)), 0, 0, 0, 1);
        expected.setHostValue(safeTestTypeCast<OutputType>(geluTanh(-TEST_VALUE_1)), 0, 0, 1, 0);
        expected.setHostValue(safeTestTypeCast<OutputType>(geluTanh(TEST_VALUE_2)), 0, 0, 1, 1);
        expected.setHostValue(safeTestTypeCast<OutputType>(geluTanh(-TEST_VALUE_2)), 0, 1, 0, 0);
        expected.setHostValue(safeTestTypeCast<OutputType>(geluTanh(TEST_VALUE_5)), 0, 1, 0, 1);
        expected.setHostValue(safeTestTypeCast<OutputType>(geluTanh(-TEST_VALUE_5)), 0, 1, 1, 0);
        expected.setHostValue(safeTestTypeCast<OutputType>(geluTanh(TEST_VALUE_1_5)), 0, 1, 1, 1);

        auto tolerance = getDynamicTolerance(PointwiseMode::GELU_APPROX_TANH_FWD, TEST_VALUE_5);
        auto validator = createAllCloseValidator<OutputType>(tolerance, tolerance);
        EXPECT_TRUE(validator->allClose(expected, output));
    }

    void testSwishForwardOperation()
    {
        Tensor<Input1Type> input({1, 2, 2, 2});
        Tensor<OutputType> output({1, 2, 2, 2});

        input.setHostValue(safeTestTypeCast<Input1Type>(0.0f), 0, 0, 0, 0);
        input.setHostValue(safeTestTypeCast<Input1Type>(TEST_VALUE_1), 0, 0, 0, 1);
        input.setHostValue(safeTestTypeCast<Input1Type>(-TEST_VALUE_1), 0, 0, 1, 0);
        input.setHostValue(safeTestTypeCast<Input1Type>(TEST_VALUE_2), 0, 0, 1, 1);
        input.setHostValue(safeTestTypeCast<Input1Type>(-TEST_VALUE_2), 0, 1, 0, 0);
        input.setHostValue(safeTestTypeCast<Input1Type>(TEST_VALUE_5), 0, 1, 0, 1);
        input.setHostValue(safeTestTypeCast<Input1Type>(-TEST_VALUE_5), 0, 1, 1, 0);
        input.setHostValue(safeTestTypeCast<Input1Type>(TEST_VALUE_1_5), 0, 1, 1, 1);

        CpuReferencePointwiseImpl<OutputType, Input1Type>::pointwiseCompute(
            PointwiseMode::SWISH_FWD, output, input);

        // swish(x) = x * sigmoid(x) = x / (1 + exp(-x))
        auto swish = [](float x) {
            auto xCompute = static_cast<ComputeType>(safeTestTypeCast<Input1Type>(x));
            auto exp = hipdnn_data_sdk::types::exp(-xCompute);
            return xCompute / (ComputeType{1} + exp);
        };
        Tensor<OutputType> expected({1, 2, 2, 2});
        expected.setHostValue(safeTestTypeCast<OutputType>(swish(0.0f)), 0, 0, 0, 0);
        expected.setHostValue(safeTestTypeCast<OutputType>(swish(TEST_VALUE_1)), 0, 0, 0, 1);
        expected.setHostValue(safeTestTypeCast<OutputType>(swish(-TEST_VALUE_1)), 0, 0, 1, 0);
        expected.setHostValue(safeTestTypeCast<OutputType>(swish(TEST_VALUE_2)), 0, 0, 1, 1);
        expected.setHostValue(safeTestTypeCast<OutputType>(swish(-TEST_VALUE_2)), 0, 1, 0, 0);
        expected.setHostValue(safeTestTypeCast<OutputType>(swish(TEST_VALUE_5)), 0, 1, 0, 1);
        expected.setHostValue(safeTestTypeCast<OutputType>(swish(-TEST_VALUE_5)), 0, 1, 1, 0);
        expected.setHostValue(safeTestTypeCast<OutputType>(swish(TEST_VALUE_1_5)), 0, 1, 1, 1);

        auto tolerance = getDynamicTolerance(PointwiseMode::SWISH_FWD, TEST_VALUE_5);
        auto validator = createAllCloseValidator<OutputType>(tolerance, tolerance);
        EXPECT_TRUE(validator->allClose(expected, output));
    }

    void testUnary1DOperation()
    {
        Tensor<Input1Type> input({5});
        Tensor<OutputType> output({5});

        // Fill with simple values for 1D testing
        for(int i = 0; i < 5; ++i)
        {
            input.setHostValue(safeTestTypeCast<Input1Type>(static_cast<float>(i - 2)),
                               i); // [-2, -1, 0, 1, 2]
        }

        CpuReferencePointwiseImpl<OutputType, Input1Type>::pointwiseCompute(
            PointwiseMode::RELU_FWD, output, input);

        // Create expected tensor: ReLU applied to [-2, -1, 0, 1, 2] = [0, 0, 0, 1, 2]
        Tensor<OutputType> expected({5});
        expected.setHostValue(safeTestTypeCast<OutputType>(0.0f), 0); // max(0, -2) = 0
        expected.setHostValue(safeTestTypeCast<OutputType>(0.0f), 1); // max(0, -1) = 0
        expected.setHostValue(safeTestTypeCast<OutputType>(0.0f), 2); // max(0, 0) = 0
        expected.setHostValue(safeTestTypeCast<OutputType>(1.0f), 3); // max(0, 1) = 1
        expected.setHostValue(safeTestTypeCast<OutputType>(2.0f), 4); // max(0, 2) = 2

        auto tolerance = getDynamicTolerance(PointwiseMode::RELU_FWD, 2.0f);
        auto validator = createAllCloseValidator<OutputType>(tolerance, tolerance);
        EXPECT_TRUE(validator->allClose(expected, output));
    }

    void testUnary2DOperation()
    {
        Tensor<Input1Type> input({3, 4});
        Tensor<OutputType> output({3, 4});

        // Fill with pattern for 2D testing
        for(int m = 0; m < 3; ++m)
        {
            for(int n = 0; n < 4; ++n)
            {
                input.setHostValue(
                    safeTestTypeCast<Input1Type>(static_cast<float>((m - 1) + (n - 2))), m, n);
            }
        }

        CpuReferencePointwiseImpl<OutputType, Input1Type>::pointwiseCompute(
            PointwiseMode::ABS, output, input);

        // Create expected tensor: abs applied to the pattern
        Tensor<OutputType> expected({3, 4});
        for(int m = 0; m < 3; ++m)
        {
            for(int n = 0; n < 4; ++n)
            {
                auto val = static_cast<float>((m - 1) + (n - 2));
                expected.setHostValue(
                    safeTestTypeCast<OutputType>(hipdnn_data_sdk::types::abs(val)), m, n);
            }
        }

        auto tolerance = getDynamicTolerance(PointwiseMode::ABS, 3.0f);
        auto validator = createAllCloseValidator<OutputType>(tolerance, tolerance);
        EXPECT_TRUE(validator->allClose(expected, output));
    }

    void testUnary3DOperation()
    {
        Tensor<Input1Type> input({2, 3, 4});
        Tensor<OutputType> output({2, 3, 4});

        input.fillWithValue(safeTestTypeCast<Input1Type>(-TEST_VALUE_2_5));

        CpuReferencePointwiseImpl<OutputType, Input1Type>::pointwiseCompute(
            PointwiseMode::NEG, output, input);

        // Create expected tensor: negation of -2.5 = 2.5
        Tensor<OutputType> expected({2, 3, 4});
        expected.fillWithValue(safeTestTypeCast<OutputType>(TEST_VALUE_2_5));

        auto tolerance = getDynamicTolerance(PointwiseMode::NEG, TEST_VALUE_2_5);
        auto validator = createAllCloseValidator<OutputType>(tolerance, tolerance);
        EXPECT_TRUE(validator->allClose(expected, output));
    }

    void testUnarySingleElementTensor()
    {
        Tensor<Input1Type> input({1, 1, 1, 1});
        Tensor<OutputType> output({1, 1, 1, 1});

        input.setHostValue(safeTestTypeCast<Input1Type>(E), 0, 0, 0, 0); // e

        CpuReferencePointwiseImpl<OutputType, Input1Type>::pointwiseCompute(
            PointwiseMode::TANH_FWD, output, input);

        // Create expected tensor: tanh(e)
        Tensor<OutputType> expected({1, 1, 1, 1});
        expected.setHostValue(safeTestTypeCast<OutputType>(std::tanh(E)), 0, 0, 0, 0);

        auto tolerance = getDynamicTolerance(PointwiseMode::TANH_FWD, E);
        auto validator = createAllCloseValidator<OutputType>(tolerance, tolerance);
        EXPECT_TRUE(validator->allClose(expected, output));
    }

    void testIdentityOperation()
    {
        Tensor<Input1Type> input({1, 2, 2, 2});
        Tensor<OutputType> output({1, 2, 2, 2});

        // Fill with mix of positive, negative, and zero values
        input.setHostValue(safeTestTypeCast<Input1Type>(TEST_VALUE_1), 0, 0, 0, 0); // 1.0
        input.setHostValue(safeTestTypeCast<Input1Type>(-TEST_VALUE_2), 0, 0, 0, 1); // -2.0
        input.setHostValue(safeTestTypeCast<Input1Type>(0.0f), 0, 0, 1, 0); // 0.0
        input.setHostValue(safeTestTypeCast<Input1Type>(PI), 0, 0, 1, 1); // π
        input.setHostValue(safeTestTypeCast<Input1Type>(E), 0, 1, 0, 0); // e
        input.setHostValue(safeTestTypeCast<Input1Type>(-TEST_VALUE_1_5), 0, 1, 0, 1); // -1.5
        input.setHostValue(safeTestTypeCast<Input1Type>(TEST_VALUE_2_5), 0, 1, 1, 0); // 2.5
        input.setHostValue(safeTestTypeCast<Input1Type>(SQRT_2), 0, 1, 1, 1); // √2

        CpuReferencePointwiseImpl<OutputType, Input1Type>::pointwiseCompute(
            PointwiseMode::IDENTITY, output, input);

        // Create expected tensor: identity(x) = x
        Tensor<OutputType> expected({1, 2, 2, 2});
        expected.setHostValue(safeTestTypeCast<OutputType>(TEST_VALUE_1), 0, 0, 0, 0); // 1.0
        expected.setHostValue(safeTestTypeCast<OutputType>(-TEST_VALUE_2), 0, 0, 0, 1); // -2.0
        expected.setHostValue(safeTestTypeCast<OutputType>(0.0f), 0, 0, 1, 0); // 0.0
        expected.setHostValue(safeTestTypeCast<OutputType>(PI), 0, 0, 1, 1); // π
        expected.setHostValue(safeTestTypeCast<OutputType>(E), 0, 1, 0, 0); // e
        expected.setHostValue(safeTestTypeCast<OutputType>(-TEST_VALUE_1_5), 0, 1, 0, 1); // -1.5
        expected.setHostValue(safeTestTypeCast<OutputType>(TEST_VALUE_2_5), 0, 1, 1, 0); // 2.5
        expected.setHostValue(safeTestTypeCast<OutputType>(SQRT_2), 0, 1, 1, 1); // √2

        auto tolerance = getDynamicTolerance(PointwiseMode::IDENTITY, PI);
        auto validator = createAllCloseValidator<OutputType>(tolerance, tolerance);
        EXPECT_TRUE(validator->allClose(expected, output));
    }
};

using TestTypes = ::testing::Types<float, half, bfloat16, double, int8_t>;
// Empty third argument required for C++17 compatibility with TYPED_TEST_SUITE macro
TYPED_TEST_SUITE(CpuReferencePointwiseFixture, TestTypes, );

// Binary operation tests
TYPED_TEST(CpuReferencePointwiseFixture, BinaryAddOperation)
{
    this->testBinaryAddOperation();
}

TYPED_TEST(CpuReferencePointwiseFixture, BinarySubtractOperation)
{
    this->testBinarySubtractOperation();
}

TYPED_TEST(CpuReferencePointwiseFixture, BinaryAddOperationSanityValidation)
{
    this->testBinaryAddOperationSanityValidation();
}

TYPED_TEST(CpuReferencePointwiseFixture, BinarySubtractOperationSanityValidation)
{
    this->testBinarySubtractOperationSanityValidation();
}

TYPED_TEST(CpuReferencePointwiseFixture, BinarySingleElementTensors)
{
    this->testBinarySingleElementTensors();
}

TYPED_TEST(CpuReferencePointwiseFixture, BinaryNumericalPrecision)
{
    // This test uses fractional precision test values,
    // which don't make sense for integer types since they'd be truncated to 0.
    if constexpr(std::is_integral_v<TypeParam>)
    {
        GTEST_SKIP() << "Skipping numerical precision test for integer types";
    }
    this->testBinaryNumericalPrecision();
}

TYPED_TEST(CpuReferencePointwiseFixture, BinaryElementwise1D)
{
    this->testElementwise1D();
}

TYPED_TEST(CpuReferencePointwiseFixture, BinaryBroadcast2Dx1D)
{
    this->testBroadcast2Dx1D();
}

TYPED_TEST(CpuReferencePointwiseFixture, BinaryBroadcast3D)
{
    this->testBroadcast3D();
}

TYPED_TEST(CpuReferencePointwiseFixture, BinaryBroadcast3DImplicitLeading)
{
    this->testBroadcast3DImplicitLeading();
}

TYPED_TEST(CpuReferencePointwiseFixture, BinaryBroadcast4Dx4D)
{
    this->testBroadcast4Dx4D();
}

TYPED_TEST(CpuReferencePointwiseFixture, BinaryBroadcastComplexND)
{
    this->testBroadcastComplexND();
}

TYPED_TEST(CpuReferencePointwiseFixture, BinaryBroadcast5D)
{
    this->testBroadcast5D();
}

// Unary operation tests
TYPED_TEST(CpuReferencePointwiseFixture, UnaryReluForward)
{
    this->testReluForwardOperation();
}

TYPED_TEST(CpuReferencePointwiseFixture, UnaryReluBackward)
{
    this->testReluBackwardOperation();
}

TYPED_TEST(CpuReferencePointwiseFixture, UnaryParameterizedReluForward)
{
    this->testParameterizedReluForwardOperation();
}

TYPED_TEST(CpuReferencePointwiseFixture, UnaryParameterizedReluBackward)
{
    this->testParameterizedReluBackwardOperation();
}

TYPED_TEST(CpuReferencePointwiseFixture, UnarySigmoidForward)
{
    this->testSigmoidForwardOperation();
}

TYPED_TEST(CpuReferencePointwiseFixture, UnarySigmoidBackward)
{
    this->testSigmoidBackwardOperation();
}

TYPED_TEST(CpuReferencePointwiseFixture, UnaryTanhForward)
{
    this->testTanhForwardOperation();
}

TYPED_TEST(CpuReferencePointwiseFixture, UnaryTanhBackward)
{
    this->testTanhBackwardOperation();
}

TYPED_TEST(CpuReferencePointwiseFixture, UnaryAbsoluteValue)
{
    this->testAbsoluteValueOperation();
}

TYPED_TEST(CpuReferencePointwiseFixture, UnaryNegation)
{
    this->testNegationOperation();
}

TYPED_TEST(CpuReferencePointwiseFixture, UnaryGeluForward)
{
    this->testGeluForwardOperation();
}

TYPED_TEST(CpuReferencePointwiseFixture, UnaryGeluApproxTanhForward)
{
    this->testGeluApproxTanhForwardOperation();
}

TYPED_TEST(CpuReferencePointwiseFixture, UnarySwishForward)
{
    this->testSwishForwardOperation();
}

TYPED_TEST(CpuReferencePointwiseFixture, Unary1DOperation)
{
    this->testUnary1DOperation();
}

TYPED_TEST(CpuReferencePointwiseFixture, Unary2DOperation)
{
    this->testUnary2DOperation();
}

TYPED_TEST(CpuReferencePointwiseFixture, Unary3DOperation)
{
    this->testUnary3DOperation();
}

TYPED_TEST(CpuReferencePointwiseFixture, UnarySingleElementTensor)
{
    this->testUnarySingleElementTensor();
}

TYPED_TEST(CpuReferencePointwiseFixture, UnaryIdentity)
{
    this->testIdentityOperation();
}

// Mixed-type binary test instantiations
using TestCpuReferencePointwiseBinaryMixed1Bfp16
    = CpuReferencePointwiseFixture<float, half, bfloat16>;
using TestCpuReferencePointwiseBinaryMixed2Bfp16
    = CpuReferencePointwiseFixture<half, float, bfloat16>;
using TestCpuReferencePointwiseBinaryMixed1Fp16 = CpuReferencePointwiseFixture<float, float, half>;
using TestCpuReferencePointwiseBinaryMixed2Fp16
    = CpuReferencePointwiseFixture<bfloat16, float, half>;
using TestCpuReferencePointwiseBinaryMixed1Fp32
    = CpuReferencePointwiseFixture<bfloat16, half, float>;
using TestCpuReferencePointwiseBinaryMixed3Fp16
    = CpuReferencePointwiseFixture<float, bfloat16, half>;
using TestCpuReferencePointwiseBinaryMixed2Fp32
    = CpuReferencePointwiseFixture<half, bfloat16, float>;

// Test a sample of mixed-type binary operations
TEST_F(TestCpuReferencePointwiseBinaryMixed1Bfp16, BinaryMixedTypeAddOperation)
{
    this->testBinaryAddOperation();
}

TEST_F(TestCpuReferencePointwiseBinaryMixed1Bfp16, BinaryMixedTypeSubtractOperation)
{
    this->testBinarySubtractOperation();
}

TEST_F(TestCpuReferencePointwiseBinaryMixed2Bfp16, BinaryMixedTypeAddOperation)
{
    this->testBinaryAddOperation();
}

TEST_F(TestCpuReferencePointwiseBinaryMixed2Bfp16, BinaryMixedTypeSubtractOperation)
{
    this->testBinarySubtractOperation();
}

TEST_F(TestCpuReferencePointwiseBinaryMixed1Fp16, BinaryMixedTypeAddOperation)
{
    this->testBinaryAddOperation();
}

TEST_F(TestCpuReferencePointwiseBinaryMixed1Fp16, BinaryMixedTypeSubtractOperation)
{
    this->testBinarySubtractOperation();
}

TEST_F(TestCpuReferencePointwiseBinaryMixed2Fp16, BinaryMixedTypeAddOperation)
{
    this->testBinaryAddOperation();
}

TEST_F(TestCpuReferencePointwiseBinaryMixed2Fp16, BinaryMixedTypeSubtractOperation)
{
    this->testBinarySubtractOperation();
}

TEST_F(TestCpuReferencePointwiseBinaryMixed1Fp32, BinaryMixedTypeAddOperation)
{
    this->testBinaryAddOperation();
}

TEST_F(TestCpuReferencePointwiseBinaryMixed1Fp32, BinaryMixedTypeSubtractOperation)
{
    this->testBinarySubtractOperation();
}

TEST_F(TestCpuReferencePointwiseBinaryMixed3Fp16, BinaryMixedTypeAddOperation)
{
    this->testBinaryAddOperation();
}

TEST_F(TestCpuReferencePointwiseBinaryMixed3Fp16, BinaryMixedTypeSubtractOperation)
{
    this->testBinarySubtractOperation();
}

TEST_F(TestCpuReferencePointwiseBinaryMixed2Fp32, BinaryMixedTypeAddOperation)
{
    this->testBinaryAddOperation();
}

TEST_F(TestCpuReferencePointwiseBinaryMixed2Fp32, BinaryMixedTypeSubtractOperation)
{
    this->testBinarySubtractOperation();
}

// Mixed-type unary test instantiations
using TestCpuReferencePointwiseUnaryMixedFp16 = CpuReferencePointwiseFixture<float, float, half>;
using TestCpuReferencePointwiseUnaryMixed1Fp32 = CpuReferencePointwiseFixture<half, half, float>;
using TestCpuReferencePointwiseUnaryMixed1Bfp16
    = CpuReferencePointwiseFixture<float, float, bfloat16>;
using TestCpuReferencePointwiseUnaryMixed2Fp32
    = CpuReferencePointwiseFixture<bfloat16, bfloat16, float>;
using TestCpuReferencePointwiseUnaryMixed2Bfp16
    = CpuReferencePointwiseFixture<half, half, bfloat16>;
using TestCpuReferencePointwiseUnaryMixed2Fp16
    = CpuReferencePointwiseFixture<bfloat16, bfloat16, half>;

// Test a sample of mixed-type unary operations
TEST_F(TestCpuReferencePointwiseUnaryMixedFp16, UnaryMixedTypeReluForward)
{
    this->testReluForwardOperation();
}

TEST_F(TestCpuReferencePointwiseUnaryMixedFp16, UnaryMixedTypeSigmoidForward)
{
    this->testSigmoidForwardOperation();
}

TEST_F(TestCpuReferencePointwiseUnaryMixedFp16, UnaryMixedTypeGeluForward)
{
    this->testGeluForwardOperation();
}

TEST_F(TestCpuReferencePointwiseUnaryMixedFp16, UnaryMixedTypeGeluApproxTanhForward)
{
    this->testGeluApproxTanhForwardOperation();
}

TEST_F(TestCpuReferencePointwiseUnaryMixedFp16, UnaryMixedTypeSwishForward)
{
    this->testSwishForwardOperation();
}

TEST_F(TestCpuReferencePointwiseUnaryMixed1Fp32, UnaryMixedTypeAbsoluteValue)
{
    this->testAbsoluteValueOperation();
}

TEST_F(TestCpuReferencePointwiseUnaryMixed1Fp32, UnaryMixedTypeNegation)
{
    this->testNegationOperation();
}

TEST_F(TestCpuReferencePointwiseUnaryMixed1Bfp16, UnaryMixedTypeTanhForward)
{
    this->testTanhForwardOperation();
}

TEST_F(TestCpuReferencePointwiseUnaryMixed1Bfp16, UnaryMixedTypeGeluForward)
{
    this->testGeluForwardOperation();
}

TEST_F(TestCpuReferencePointwiseUnaryMixed1Bfp16, UnaryMixedTypeGeluApproxTanhForward)
{
    this->testGeluApproxTanhForwardOperation();
}

TEST_F(TestCpuReferencePointwiseUnaryMixed1Bfp16, UnaryMixedTypeSwishForward)
{
    this->testSwishForwardOperation();
}

TEST_F(TestCpuReferencePointwiseUnaryMixed1Bfp16, UnaryMixedTypeReluBackward)
{
    this->testReluBackwardOperation();
}

TEST_F(TestCpuReferencePointwiseUnaryMixed2Fp32, UnaryMixedTypeParameterizedRelu)
{
    this->testParameterizedReluForwardOperation();
}

TEST_F(TestCpuReferencePointwiseUnaryMixed2Fp32, UnaryMixedTypeSigmoidBackward)
{
    this->testSigmoidBackwardOperation();
}

TEST_F(TestCpuReferencePointwiseUnaryMixed2Bfp16, UnaryMixedTypeTanhBackward)
{
    this->testTanhBackwardOperation();
}

TEST_F(TestCpuReferencePointwiseUnaryMixed2Bfp16, UnaryMixedTypeAbsoluteValue)
{
    this->testAbsoluteValueOperation();
}

TEST_F(TestCpuReferencePointwiseUnaryMixed2Fp16, UnaryMixedTypeReluForward)
{
    this->testReluForwardOperation();
}

TEST_F(TestCpuReferencePointwiseUnaryMixed2Fp16, UnaryMixedTypeParameterizedReluBackward)
{
    this->testParameterizedReluBackwardOperation();
}

TEST_F(TestCpuReferencePointwiseUnaryMixed2Fp16, UnaryMixedTypeIdentity)
{
    this->testIdentityOperation();
}

TEST_F(TestCpuReferencePointwiseUnaryMixed2Fp32, UnaryMixedTypeGeluForward)
{
    this->testGeluForwardOperation();
}

TEST_F(TestCpuReferencePointwiseUnaryMixed2Fp32, UnaryMixedTypeGeluApproxTanhForward)
{
    this->testGeluApproxTanhForwardOperation();
}

TEST_F(TestCpuReferencePointwiseUnaryMixed2Fp32, UnaryMixedTypeSwishForward)
{
    this->testSwishForwardOperation();
}

// --- Regression tests ---

TEST(TestCpuReferencePointwiseRegressions, ParameterizedReluBwdBfloat16LowerClipEdgeCase)
{
    // In bfloat16, a value higher than lowerClip can be equal to lowerClip if lowerClip itself is cast to bfloat16. However, this value should still register as higher than lowerClip.

    Tensor<bfloat16> dy({1, 1, 1, 1});
    Tensor<bfloat16> y({1, 1, 1, 1});
    Tensor<bfloat16> output({1, 1, 1, 1});

    const float lowerClip = 0.1f;
    const float upperClip = 0.5f;
    const float lowerSlope = 0.2f;

    // Set y to a value slightly above lowerClip. In bfloat16, this is equal to lowerClip
    dy.setHostValue(safeTestTypeCast<bfloat16>(1.0f), 0, 0, 0, 0);
    dy.markHostModified();
    y.setHostValue(safeTestTypeCast<bfloat16>(0.10009765625f), 0, 0, 0, 0);
    y.markHostModified();

    EXPECT_NE(lowerClip, safeTestTypeCast<float>(y.getHostValue(0, 0, 0, 0)));

    CpuReferencePointwiseImpl<bfloat16, bfloat16, bfloat16>::pointwiseCompute(
        PointwiseMode::RELU_BWD, output, dy, y, lowerClip, upperClip, lowerSlope);

    output.markDeviceModified();

    EXPECT_EQ(output.getHostValue(0, 0, 0, 0), dy.getHostValue(0, 0, 0, 0));
}