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

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

#include <cmath>
#include <cstdint>
#include <gtest/gtest.h>
#include <limits>
#include <vector>

#include <hipdnn_data_sdk/types.hpp>
#include <hipdnn_data_sdk/utilities/ShapeUtilities.hpp>
#include <hipdnn_data_sdk/utilities/Tensor.hpp>
#include <hipdnn_test_sdk/utilities/DynamicTolerancesCommon.hpp>

using namespace hipdnn_test_sdk::utilities;
using hipdnn_data_sdk::types::bfloat16;
using hipdnn_data_sdk::types::half;

// =================================================================================================
// TestComputeGamma — verifies computeGamma against hand-derived formulas
//
// computeGamma selects between two regimes based on nU = 2*k*epsilon:
//   nU < 0.01  → linear:      gamma = nU / (1 - nU)
//   nU >= 0.01 → statistical: gamma = 6 * sqrt(2*k) * epsilon
//
// Tests below compute expected values from the raw formulas, NOT by calling computeGamma.
// =================================================================================================

// --- Linear regime (nU < 0.01) ---

TEST(TestComputeGamma, LinearRegimeFloatK1)
{
    // float eps = 2^-23, k=1: nU = 2 * 2^-23 = 2^-22 ≈ 2.38e-7 << 0.01
    auto u = static_cast<double>(std::numeric_limits<float>::epsilon());
    const double nU = 2.0 * 1.0 * u;
    const double expected = nU / (1.0 - nU);

    EXPECT_NEAR(computeGamma(1, u), expected, 1e-15);
}

TEST(TestComputeGamma, LinearRegimeFloatK100)
{
    // float eps = 2^-23, k=100: nU = 200 * 2^-23 ≈ 2.38e-5 << 0.01
    auto u = static_cast<double>(std::numeric_limits<float>::epsilon());
    const double nU = 2.0 * 100.0 * u;
    const double expected = nU / (1.0 - nU);

    EXPECT_NEAR(computeGamma(100, u), expected, 1e-15);
}

TEST(TestComputeGamma, LinearRegimeHalfK1)
{
    // half eps = 2^-10, k=1: nU = 2 * 2^-10 = 2^-9 ≈ 0.00195 < 0.01
    auto u = static_cast<double>(std::numeric_limits<half>::epsilon());
    const double nU = 2.0 * 1.0 * u;
    const double expected = nU / (1.0 - nU); // = 1/511 ≈ 0.001957

    EXPECT_NEAR(computeGamma(1, u), expected, 1e-10);
}

TEST(TestComputeGamma, LinearRegimeHalfK5)
{
    // half eps = 2^-10, k=5: nU = 10 * 2^-10 = 10/1024 ≈ 0.00977 < 0.01 (boundary)
    auto u = static_cast<double>(std::numeric_limits<half>::epsilon());
    const double nU = 2.0 * 5.0 * u;
    const double expected = nU / (1.0 - nU); // = 5/507 ≈ 0.009862

    EXPECT_NEAR(computeGamma(5, u), expected, 1e-10);
}

TEST(TestComputeGamma, LinearRegimeDoubleK1000)
{
    // double eps = 2^-52, k=1000: nU = 2000 * 2^-52 ≈ 4.44e-13 << 0.01
    const auto u = std::numeric_limits<double>::epsilon();
    const double nU = 2.0 * 1000.0 * u;
    const double expected = nU / (1.0 - nU);

    EXPECT_NEAR(computeGamma(1000, u), expected, 1e-20);
}

// --- Statistical regime (nU >= 0.01) ---

TEST(TestComputeGamma, StatisticalRegimeBf16K1)
{
    // bf16 eps = 2^-7, k=1: nU = 2 * 2^-7 = 2^-6 = 0.015625 >= 0.01
    auto u = static_cast<double>(std::numeric_limits<bfloat16>::epsilon());
    const double expected = 6.0 * std::sqrt(2.0 * 1.0) * u; // 6 * sqrt(2) * 2^-7

    EXPECT_NEAR(computeGamma(1, u), expected, 1e-10);
}

TEST(TestComputeGamma, StatisticalRegimeBf16K10)
{
    // bf16 eps = 2^-7, k=10: nU = 20 * 2^-7 ≈ 0.156 >= 0.01
    auto u = static_cast<double>(std::numeric_limits<bfloat16>::epsilon());
    const double expected = 6.0 * std::sqrt(2.0 * 10.0) * u; // 6 * sqrt(20) * 2^-7

    EXPECT_NEAR(computeGamma(10, u), expected, 1e-10);
}

TEST(TestComputeGamma, StatisticalRegimeBf16K100)
{
    // bf16 eps = 2^-7, k=100: nU = 200 * 2^-7 ≈ 1.56 >= 0.01
    // gamma = 6 * sqrt(200) * 2^-7 ≈ 0.663 (exceeds 0.5 — useful for overflow tests)
    auto u = static_cast<double>(std::numeric_limits<bfloat16>::epsilon());
    const double expected = 6.0 * std::sqrt(2.0 * 100.0) * u;

    EXPECT_NEAR(computeGamma(100, u), expected, 1e-10);
    EXPECT_GT(computeGamma(100, u), 0.5); // Confirms this triggers validateGamma overflow
}

TEST(TestComputeGamma, StatisticalRegimeHalfK6)
{
    // half eps = 2^-10, k=6: nU = 12 * 2^-10 ≈ 0.01172 >= 0.01 (just above threshold)
    auto u = static_cast<double>(std::numeric_limits<half>::epsilon());
    const double expected = 6.0 * std::sqrt(2.0 * 6.0) * u; // 6 * sqrt(12) * 2^-10

    EXPECT_NEAR(computeGamma(6, u), expected, 1e-10);
}

TEST(TestComputeGamma, StatisticalRegimeHalfK10)
{
    // half eps = 2^-10, k=10: nU = 20 * 2^-10 ≈ 0.0195 >= 0.01
    auto u = static_cast<double>(std::numeric_limits<half>::epsilon());
    const double expected = 6.0 * std::sqrt(2.0 * 10.0) * u;

    EXPECT_NEAR(computeGamma(10, u), expected, 1e-10);
}

// --- Threshold boundary ---

TEST(TestComputeGamma, ThresholdBoundaryExactlyAtThreshold)
{
    // Construct epsilon such that nU = 2*k*eps = exactly 0.01
    // k=1, eps=0.005 → nU = 0.01 → should use statistical (nU >= 0.01)
    const double eps = 0.005;
    const double expected = 6.0 * std::sqrt(2.0) * eps;

    EXPECT_NEAR(computeGamma(1, eps), expected, 1e-15);
}

TEST(TestComputeGamma, ThresholdBoundaryJustBelowThreshold)
{
    // k=1, eps=0.004999 → nU = 0.009998 < 0.01 → should use linear
    const double eps = 0.004999;
    const double nU = 2.0 * eps;
    const double expected = nU / (1.0 - nU);

    EXPECT_NEAR(computeGamma(1, eps), expected, 1e-15);
}

// --- Edge cases ---

TEST(TestComputeGamma, ZeroAccumulations)
{
    auto u = static_cast<double>(std::numeric_limits<float>::epsilon());
    // k=0: nU = 0 < 0.01 → linear: 0/(1-0) = 0
    EXPECT_DOUBLE_EQ(computeGamma(0, u), 0.0);
}

TEST(TestComputeGamma, ZeroEpsilon)
{
    // Hypothetical perfect precision: gamma should be 0
    EXPECT_DOUBLE_EQ(computeGamma(100, 0.0), 0.0);
}

// =================================================================================================
// TestValidateGamma
// =================================================================================================

TEST(TestValidateGamma, DoesNotThrowBelowThreshold)
{
    EXPECT_NO_THROW(validateGamma(0.0));
    EXPECT_NO_THROW(validateGamma(0.1));
    EXPECT_NO_THROW(validateGamma(0.499));
}

TEST(TestValidateGamma, ThrowsAtThreshold)
{
    EXPECT_THROW(validateGamma(0.5), std::overflow_error);
}

TEST(TestValidateGamma, ThrowsAboveThreshold)
{
    EXPECT_THROW(validateGamma(0.9), std::overflow_error);
    EXPECT_THROW(validateGamma(1.0), std::overflow_error);
}

TEST(TestValidateGamma, CustomThreshold)
{
    EXPECT_NO_THROW(validateGamma(0.09, 0.1));
    EXPECT_THROW(validateGamma(0.1, 0.1), std::overflow_error);
    EXPECT_THROW(validateGamma(0.5, 0.1), std::overflow_error);
}

// =================================================================================================
// TestComputeInputCastingError
// =================================================================================================

TEST(TestComputeInputCastingError, HigherPrecisionInputReturnsError)
{
    // double→float downcast: inputEps < computeEps → error = 2 * signal * computeEps
    auto computeEps = static_cast<double>(std::numeric_limits<float>::epsilon());
    const double signal = 100.0;
    const double expected = 2.0 * signal * computeEps;

    EXPECT_NEAR((computeInputCastingError<double, float>(signal)), expected, 1e-15);
}

TEST(TestComputeInputCastingError, LowerPrecisionInputReturnsZero)
{
    // half→float upcast: inputEps > computeEps → no error
    EXPECT_DOUBLE_EQ((computeInputCastingError<half, float>(100.0)), 0.0);
}

TEST(TestComputeInputCastingError, SamePrecisionReturnsZero)
{
    // float→float: no casting → no error
    EXPECT_DOUBLE_EQ((computeInputCastingError<float, float>(100.0)), 0.0);
}

TEST(TestComputeInputCastingError, SingleInputFactor)
{
    // Self-product (e.g., x^2): numDistinctInputs=1
    auto computeEps = static_cast<double>(std::numeric_limits<float>::epsilon());
    const double signal = 50.0;
    const double expected = 1.0 * signal * computeEps;

    EXPECT_NEAR((computeInputCastingError<double, float>(signal, 1)), expected, 1e-15);
}

// =================================================================================================
// TestComputeOutputCastingError
// =================================================================================================

TEST(TestComputeOutputCastingError, LowerPrecisionOutputReturnsError)
{
    // float→half downcast: outputEps > computeEps → error = magnitude * outputEps
    auto outputEps = static_cast<double>(std::numeric_limits<half>::epsilon());
    const double magnitude = 200.0;
    const double expected = magnitude * outputEps;

    EXPECT_NEAR((computeOutputCastingError<half, float>(magnitude)), expected, 1e-15);
}

TEST(TestComputeOutputCastingError, HigherPrecisionOutputReturnsZero)
{
    // float→double upcast: outputEps < computeEps → no error
    EXPECT_DOUBLE_EQ((computeOutputCastingError<double, float>(200.0)), 0.0);
}

TEST(TestComputeOutputCastingError, SamePrecisionReturnsZero)
{
    EXPECT_DOUBLE_EQ((computeOutputCastingError<float, float>(200.0)), 0.0);
}

TEST(TestComputeOutputCastingError, Bf16OutputReturnsError)
{
    // float→bf16 downcast
    auto outputEps = static_cast<double>(std::numeric_limits<bfloat16>::epsilon());
    const double magnitude = 100.0;
    const double expected = magnitude * outputEps;

    EXPECT_NEAR((computeOutputCastingError<bfloat16, float>(magnitude)), expected, 1e-15);
}

// =================================================================================================
// TestValidateToleranceRange
// =================================================================================================

TEST(TestValidateToleranceRange, WithinRangeNoThrow)
{
    EXPECT_NO_THROW((validateToleranceRange<float>(1.0)));
    EXPECT_NO_THROW((validateToleranceRange<half>(100.0)));
    EXPECT_NO_THROW((validateToleranceRange<bfloat16>(100.0)));
}

TEST(TestValidateToleranceRange, ExceedsHalfMaxThrows)
{
    // half max ≈ 65504
    EXPECT_THROW((validateToleranceRange<half>(70000.0)), std::overflow_error);
}

TEST(TestValidateToleranceRange, ExceedsBf16MaxThrows)
{
    // bf16 max ≈ 3.39e+38 (same exponent range as float, less mantissa)
    // Use a value larger than float max
    EXPECT_THROW((validateToleranceRange<bfloat16>(1e39)), std::overflow_error);
}

TEST(TestValidateToleranceRange, AtExactMaxNoThrow)
{
    auto halfMax = static_cast<double>(std::numeric_limits<half>::max());
    EXPECT_NO_THROW((validateToleranceRange<half>(halfMax)));
}

// =================================================================================================
// TestComputeMatrixInfNorm
// =================================================================================================

namespace
{
// Helper to create a tensor and fill it element-by-element from a flat vector of values.
// Values are stored in row-major order.
template <typename T>
hipdnn_data_sdk::utilities::Tensor<T> createTensorFromFlatValues(const std::vector<int64_t>& dims,
                                                                 const std::vector<double>& values)
{
    using hipdnn_data_sdk::utilities::iterateAlongDimensions;

    hipdnn_data_sdk::utilities::Tensor<T> tensor(dims);

    size_t idx = 0;
    iterateAlongDimensions(dims, [&](const std::vector<int64_t>& indices) {
        tensor.setHostValue(static_cast<T>(values[idx++]), indices);
    });

    return tensor;
}
} // namespace

TEST(TestComputeMatrixInfNorm, IdentityMatrix2x2)
{
    // [[1, 0], [0, 1]] -> row sums: 1, 1 -> inf-norm = 1.0
    auto tensor = createTensorFromFlatValues<float>({2, 2}, {1.0, 0.0, 0.0, 1.0});
    EXPECT_DOUBLE_EQ(computeMatrixInfNorm<float>(tensor), 1.0);
}

TEST(TestComputeMatrixInfNorm, AsymmetricRows)
{
    // [[1, 2, 3], [4, 5, 6]] -> row sums: 6, 15 -> inf-norm = 15.0
    auto tensor = createTensorFromFlatValues<float>({2, 3}, {1.0, 2.0, 3.0, 4.0, 5.0, 6.0});
    EXPECT_DOUBLE_EQ(computeMatrixInfNorm<float>(tensor), 15.0);
}

TEST(TestComputeMatrixInfNorm, NegativeValues)
{
    // [[-3, 2], [1, -1]] -> row sums: |−3|+|2|=5, |1|+|−1|=2 -> inf-norm = 5.0
    auto tensor = createTensorFromFlatValues<float>({2, 2}, {-3.0, 2.0, 1.0, -1.0});
    EXPECT_DOUBLE_EQ(computeMatrixInfNorm<float>(tensor), 5.0);
}

TEST(TestComputeMatrixInfNorm, SingleElement)
{
    // [[7]] -> inf-norm = 7.0
    auto tensor = createTensorFromFlatValues<float>({1, 1}, {7.0});
    EXPECT_DOUBLE_EQ(computeMatrixInfNorm<float>(tensor), 7.0);
}

TEST(TestComputeMatrixInfNorm, SingleRow)
{
    // [[2, 3, 5]] -> row sum: 10 -> inf-norm = 10.0
    auto tensor = createTensorFromFlatValues<float>({1, 3}, {2.0, 3.0, 5.0});
    EXPECT_DOUBLE_EQ(computeMatrixInfNorm<float>(tensor), 10.0);
}

TEST(TestComputeMatrixInfNorm, BatchedTensor)
{
    // Batch of 2 matrices, each 2x2:
    // Batch 0: [[1, 2], [3, 4]] -> row sums: 3, 7
    // Batch 1: [[5, 5], [1, 1]] -> row sums: 10, 2
    // inf-norm = max(3, 7, 10, 2) = 10.0
    auto tensor
        = createTensorFromFlatValues<float>({2, 2, 2}, {1.0, 2.0, 3.0, 4.0, 5.0, 5.0, 1.0, 1.0});
    EXPECT_DOUBLE_EQ(computeMatrixInfNorm<float>(tensor), 10.0);
}

TEST(TestComputeMatrixInfNorm, HalfPrecision)
{
    // [[1, 2], [3, 4]] -> row sums: 3, 7 -> inf-norm = 7.0
    auto tensor = createTensorFromFlatValues<half>({2, 2}, {1.0, 2.0, 3.0, 4.0});
    EXPECT_DOUBLE_EQ(computeMatrixInfNorm<half>(tensor), 7.0);
}