TheRock/rocm-libraries/projects/hipdnn/backend/tests/heuristics/TestEngineOverrideConfig.cpp

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

/**
 * @file TestEngineOverrideConfig.cpp
 * @brief Unit tests for the rule-matching internals of EngineOverrideConfig
 *        (op + dim + stride wildcards, exact/wildcard partition ordering,
 *        JSON parsing). The end-to-end policy behavior driven by
 *        HIPDNN_HEUR_CONFIG_PATH is covered in TestConfigBuiltIn.cpp.
 */

#include "heuristics/config/EngineOverrideConfig.hpp"

#include <gtest/gtest.h>
#include <hipdnn_data_sdk/detail/AutotuneConfigNames.hpp>
#include <hipdnn_data_sdk/utilities/EngineNames.hpp>
#include <nlohmann/json.hpp>

#include <cstdint>
#include <string>
#include <string_view>
#include <vector>

using namespace hipdnn_backend::heuristics::config;
using namespace hipdnn_data_sdk::utilities;

namespace
{
namespace config_criterion = hipdnn_data_sdk::detail::autotune_config::criterion;
// NOLINTNEXTLINE(misc-unused-alias-decls)
namespace config_json = hipdnn_data_sdk::detail::autotune_config::json;
namespace config_op = hipdnn_data_sdk::detail::autotune_config::op;
namespace config_tensor = hipdnn_data_sdk::detail::autotune_config::tensor;
// NOLINTNEXTLINE(misc-unused-alias-decls)
namespace config_version = hipdnn_data_sdk::detail::autotune_config::version;

struct TensorData
{
    std::vector<int64_t> dim;
    std::vector<int64_t> stride;
};

TensorView viewOf(const TensorData& t)
{
    return TensorView{"", &t.dim, &t.stride};
}

TensorView namedViewOf(std::string_view tensorId, const TensorData& t)
{
    return TensorView{tensorId, &t.dim, &t.stride};
}

std::vector<TensorView> viewsOf(const std::vector<TensorData>& ts)
{
    std::vector<TensorView> views;
    views.reserve(ts.size());
    for(const auto& t : ts)
    {
        views.push_back(viewOf(t));
    }
    return views;
}

TensorPattern makePattern(std::vector<int64_t> dim)
{
    TensorPattern p;
    p.dim = std::move(dim);
    return p;
}

TensorPattern makePatternWithStride(std::vector<int64_t> dim, std::vector<int64_t> stride)
{
    TensorPattern p;
    p.dim = std::move(dim);
    p.stride = std::move(stride);
    return p;
}

TensorPattern makeNamedPattern(std::string tensorId, std::vector<int64_t> dim)
{
    TensorPattern p;
    p.tensorId = std::move(tensorId);
    p.dim = std::move(dim);
    return p;
}

EngineOverrideConfig makeConfig(std::vector<OperationRule> rules)
{
    return EngineOverrideConfig(std::move(rules));
}

} // namespace

// ── Test 1: exact dim match, single rule ────────────────────────────────────

TEST(TestEngineOverrideConfig, ExactDimMatchSingleRule)
{
    OperationRule rule;
    rule.op = "conv_fprop";
    rule.engineName = MIOPEN_ENGINE_NAME;
    rule.tensors = {makePattern({1, 3, 224, 224}), makePattern({64, 3, 7, 7})};

    const auto config = makeConfig({std::move(rule)});

    const std::vector<TensorData> tensors = {{{1, 3, 224, 224}, {}}, {{64, 3, 7, 7}, {}}};

    auto result = config.matchOperation("conv_fprop", viewsOf(tensors));
    ASSERT_TRUE(result.has_value());
    EXPECT_EQ(*result, MIOPEN_ENGINE_ID);
}

// ── Test 2: first matching rule wins ────────────────────────────────────────

TEST(TestEngineOverrideConfig, FirstMatchingRuleWins)
{
    OperationRule rule1;
    rule1.op = "conv_fprop";
    rule1.engineName = MIOPEN_ENGINE_NAME;
    rule1.tensors = {makePattern({1, 3, 224, 224})};

    OperationRule rule2;
    rule2.op = "conv_fprop";
    rule2.engineName = HIPBLASLT_ENGINE_NAME;
    rule2.tensors = {makePattern({1, 3, 224, 224})};

    const auto config = makeConfig({std::move(rule1), std::move(rule2)});

    const std::vector<TensorData> tensors = {{{1, 3, 224, 224}, {}}};

    auto result = config.matchOperation("conv_fprop", viewsOf(tensors));
    ASSERT_TRUE(result.has_value());
    EXPECT_EQ(*result, MIOPEN_ENGINE_ID);
}

// ── Test 3: no rule matches (wrong dims) ────────────────────────────────────

TEST(TestEngineOverrideConfig, NoRuleMatchesWrongDims)
{
    OperationRule rule;
    rule.op = "conv_fprop";
    rule.engineName = MIOPEN_ENGINE_NAME;
    rule.tensors = {makePattern({1, 3, 224, 224})};

    const auto config = makeConfig({std::move(rule)});

    const std::vector<TensorData> tensors = {{{1, 3, 112, 112}, {}}};

    auto result = config.matchOperation("conv_fprop", viewsOf(tensors));
    EXPECT_FALSE(result.has_value());
}

// ── Test 4: wildcard (-1) in one dimension ──────────────────────────────────

TEST(TestEngineOverrideConfig, WildcardInOneDimension)
{
    OperationRule rule;
    rule.op = "conv_fprop";
    rule.engineName = HIPBLASLT_ENGINE_NAME;
    rule.tensors = {makePattern({-1, 64, 56, 56})};

    const auto config = makeConfig({std::move(rule)});

    for(const int64_t batch : {1, 4, 8, 32})
    {
        const std::vector<TensorData> tensors = {{{batch, 64, 56, 56}, {}}};
        auto result = config.matchOperation("conv_fprop", viewsOf(tensors));
        ASSERT_TRUE(result.has_value()) << "batch=" << batch << " should match";
        EXPECT_EQ(*result, HIPBLASLT_ENGINE_ID);
    }

    const std::vector<TensorData> tensors = {{{4, 128, 56, 56}, {}}};
    EXPECT_FALSE(config.matchOperation("conv_fprop", viewsOf(tensors)).has_value());
}

// ── Test 5: all-wildcard rule matches any shape ─────────────────────────────

TEST(TestEngineOverrideConfig, AllWildcardRuleMatchesAnyShape)
{
    OperationRule rule;
    rule.op = "conv_fprop";
    rule.engineName = HIP_MLOPS_ENGINE_NAME;
    rule.tensors = {makePattern({-1, -1, -1, -1})};

    const auto config = makeConfig({std::move(rule)});

    for(const auto& shape :
        std::vector<std::vector<int64_t>>{{1, 3, 224, 224}, {8, 64, 56, 56}, {32, 256, 14, 14}})
    {
        const std::vector<TensorData> tensors = {{shape, {}}};
        auto result = config.matchOperation("conv_fprop", viewsOf(tensors));
        ASSERT_TRUE(result.has_value());
        EXPECT_EQ(*result, HIP_MLOPS_ENGINE_ID);
    }
}

// ── Test 6: wrong op name → nullopt ─────────────────────────────────────────

TEST(TestEngineOverrideConfig, WrongOpNameReturnsNullopt)
{
    OperationRule rule;
    rule.op = "conv_fprop";
    rule.engineName = MIOPEN_ENGINE_NAME;
    rule.tensors = {makePattern({1, 3, 224, 224})};

    const auto config = makeConfig({std::move(rule)});

    const std::vector<TensorData> tensors = {{{1, 3, 224, 224}, {}}};

    EXPECT_FALSE(config.matchOperation("conv_dgrad", viewsOf(tensors)).has_value());
    EXPECT_FALSE(config.matchOperation("conv_wgrad", viewsOf(tensors)).has_value());
    EXPECT_FALSE(config.matchOperation("matmul", viewsOf(tensors)).has_value());
}

TEST(TestEngineOverrideConfig, CriteriaMustMatch)
{
    OperationRule rule;
    rule.op = "pointwise";
    rule.engineName = MIOPEN_ENGINE_NAME;
    rule.criteria = {Criterion{"pointwise_mode", 2}};
    rule.tensors = {makePattern({2, 4, 16, 16}), makePattern({2, 4, 16, 16})};

    const auto config = makeConfig({std::move(rule)});
    const std::vector<TensorData> tensors = {{{2, 4, 16, 16}, {}}, {{2, 4, 16, 16}, {}}};

    auto result
        = config.matchOperation("pointwise", {Criterion{"pointwise_mode", 2}}, viewsOf(tensors));
    ASSERT_TRUE(result.has_value());
    EXPECT_EQ(*result, MIOPEN_ENGINE_ID);

    EXPECT_FALSE(
        config.matchOperation("pointwise", {Criterion{"pointwise_mode", 34}}, viewsOf(tensors))
            .has_value());
}

TEST(TestEngineOverrideConfig, CriteriaOrderIsNormalizedBeforeMatch)
{
    OperationRule rule;
    rule.op = config_op::RESAMPLE_FWD;
    rule.engineName = MIOPEN_ENGINE_NAME;
    rule.criteria = {Criterion{config_criterion::RESAMPLE_MODE, 1},
                     Criterion{config_criterion::PADDING_MODE, 2}};
    rule.tensors = {makePattern({2, 4, 16, 16})};

    const auto config = makeConfig({std::move(rule)});
    const std::vector<TensorData> tensors = {{{2, 4, 16, 16}, {}}};

    auto result = config.matchOperation(config_op::RESAMPLE_FWD,
                                        {Criterion{config_criterion::PADDING_MODE, 2},
                                         Criterion{config_criterion::RESAMPLE_MODE, 1}},
                                        viewsOf(tensors));
    ASSERT_TRUE(result.has_value());
    EXPECT_EQ(*result, MIOPEN_ENGINE_ID);
}

TEST(TestEngineOverrideConfig, MissingCriteriaDoesNotMatchSpecificRule)
{
    OperationRule rule;
    rule.op = "pointwise";
    rule.engineName = MIOPEN_ENGINE_NAME;
    rule.criteria = {Criterion{"pointwise_mode", 2}};
    rule.tensors = {makePattern({2, 4, 16, 16})};

    const auto config = makeConfig({std::move(rule)});
    const std::vector<TensorData> tensors = {{{2, 4, 16, 16}, {}}};

    EXPECT_FALSE(config.matchOperation("pointwise", viewsOf(tensors)).has_value());
}

TEST(TestEngineOverrideConfig, SameShapeDifferentCriteriaSelectsDifferentEngines)
{
    OperationRule addRule;
    addRule.op = "pointwise";
    addRule.engineName = MIOPEN_ENGINE_NAME;
    addRule.criteria = {Criterion{"pointwise_mode", 2}};
    addRule.tensors = {makePattern({2, 4, 16, 16}), makePattern({2, 4, 16, 16})};

    OperationRule mulRule;
    mulRule.op = "pointwise";
    mulRule.engineName = HIPBLASLT_ENGINE_NAME;
    mulRule.criteria = {Criterion{"pointwise_mode", 30}};
    mulRule.tensors = {makePattern({2, 4, 16, 16}), makePattern({2, 4, 16, 16})};

    const auto config = makeConfig({std::move(addRule), std::move(mulRule)});
    const std::vector<TensorData> tensors = {{{2, 4, 16, 16}, {}}, {{2, 4, 16, 16}, {}}};

    auto addResult
        = config.matchOperation("pointwise", {Criterion{"pointwise_mode", 2}}, viewsOf(tensors));
    ASSERT_TRUE(addResult.has_value());
    EXPECT_EQ(*addResult, MIOPEN_ENGINE_ID);

    auto mulResult
        = config.matchOperation("pointwise", {Criterion{"pointwise_mode", 30}}, viewsOf(tensors));
    ASSERT_TRUE(mulResult.has_value());
    EXPECT_EQ(*mulResult, HIPBLASLT_ENGINE_ID);
}

TEST(TestEngineOverrideConfig, NamedTensorRulesIgnoreConfigTensorOrder)
{
    auto config = EngineOverrideConfig::loadFromContent(nlohmann::json{
        {hipdnn_data_sdk::detail::autotune_config::json::VERSION,
         hipdnn_data_sdk::detail::autotune_config::version::CURRENT},
        {hipdnn_data_sdk::detail::autotune_config::json::ENGINE_OVERRIDES,
         nlohmann::json::array(
             {{{hipdnn_data_sdk::detail::autotune_config::json::OP, config_op::POINTWISE},
               {hipdnn_data_sdk::detail::autotune_config::json::CRITERIA,
                {{config_criterion::POINTWISE_MODE, 2}}},
               {hipdnn_data_sdk::detail::autotune_config::json::ENGINE_NAME, MIOPEN_ENGINE_NAME},
               {hipdnn_data_sdk::detail::autotune_config::json::TENSORS,
                nlohmann::json::array(
                    {{{hipdnn_data_sdk::detail::autotune_config::json::TENSOR_ID,
                       config_tensor::IN_1},
                      {hipdnn_data_sdk::detail::autotune_config::json::DIM, {5, 6}}},
                     {{hipdnn_data_sdk::detail::autotune_config::json::TENSOR_ID,
                       config_tensor::IN_0},
                      {hipdnn_data_sdk::detail::autotune_config::json::DIM, {2, 4}}}})}}})}}
                                                            .dump());
    ASSERT_TRUE(config.has_value());

    const TensorData x{{2, 4}, {}};
    const TensorData y{{5, 6}, {}};
    const std::vector<TensorView> tensors{namedViewOf(config_tensor::IN_0, x),
                                          namedViewOf(config_tensor::IN_1, y)};

    auto result = config->matchOperation(
        config_op::POINTWISE, {Criterion{config_criterion::POINTWISE_MODE, 2}}, tensors);
    ASSERT_TRUE(result.has_value());
    EXPECT_EQ(*result, MIOPEN_ENGINE_ID);
}

TEST(TestEngineOverrideConfig, MissingVersionUsesLegacyPositionalTensorMatchingEvenWithTensorIds)
{
    auto config = EngineOverrideConfig::loadFromContent(nlohmann::json{
        {hipdnn_data_sdk::detail::autotune_config::json::ENGINE_OVERRIDES,
         nlohmann::json::array(
             {{{hipdnn_data_sdk::detail::autotune_config::json::OP, config_op::POINTWISE},
               {hipdnn_data_sdk::detail::autotune_config::json::CRITERIA,
                {{config_criterion::POINTWISE_MODE, 2}}},
               {hipdnn_data_sdk::detail::autotune_config::json::ENGINE_NAME, MIOPEN_ENGINE_NAME},
               {hipdnn_data_sdk::detail::autotune_config::json::TENSORS,
                nlohmann::json::array(
                    {{{hipdnn_data_sdk::detail::autotune_config::json::TENSOR_ID,
                       config_tensor::IN_1},
                      {hipdnn_data_sdk::detail::autotune_config::json::DIM, {2, 4}}},
                     {{hipdnn_data_sdk::detail::autotune_config::json::TENSOR_ID,
                       config_tensor::IN_0},
                      {hipdnn_data_sdk::detail::autotune_config::json::DIM, {5, 6}}}})}}})}}
                                                            .dump());
    ASSERT_TRUE(config.has_value());

    const TensorData x{{2, 4}, {}};
    const TensorData y{{5, 6}, {}};
    const std::vector<TensorView> tensors{namedViewOf(config_tensor::IN_0, x),
                                          namedViewOf(config_tensor::IN_1, y)};

    auto result = config->matchOperation(
        config_op::POINTWISE, {Criterion{config_criterion::POINTWISE_MODE, 2}}, tensors);
    ASSERT_TRUE(result.has_value());
    EXPECT_EQ(*result, MIOPEN_ENGINE_ID);
}

TEST(TestEngineOverrideConfig, NewerVersionUsesNamedTensorMatching)
{
    auto config = EngineOverrideConfig::loadFromContent(nlohmann::json{
        {config_json::VERSION, config_version::CURRENT + 1},
        {config_json::ENGINE_OVERRIDES,
         nlohmann::json::array(
             {{{config_json::OP, config_op::POINTWISE},
               {config_json::CRITERIA, {{config_criterion::POINTWISE_MODE, 2}}},
               {config_json::ENGINE_NAME, MIOPEN_ENGINE_NAME},
               {config_json::TENSORS,
                nlohmann::json::array(
                    {{{config_json::TENSOR_ID, config_tensor::IN_1}, {config_json::DIM, {5, 6}}},
                     {{config_json::TENSOR_ID, config_tensor::IN_0},
                      {config_json::DIM, {2, 4}}}})}}})}}.dump());
    ASSERT_TRUE(config.has_value());

    const TensorData x{{2, 4}, {}};
    const TensorData y{{5, 6}, {}};
    const std::vector<TensorView> tensors{namedViewOf(config_tensor::IN_0, x),
                                          namedViewOf(config_tensor::IN_1, y)};

    auto result = config->matchOperation(
        config_op::POINTWISE, {Criterion{config_criterion::POINTWISE_MODE, 2}}, tensors);
    ASSERT_TRUE(result.has_value());
    EXPECT_EQ(*result, MIOPEN_ENGINE_ID);
}

TEST(TestEngineOverrideConfig, LoadFromContentRejectsDuplicateTensorIdsInNamedConfig)
{
    auto config = EngineOverrideConfig::loadFromContent(nlohmann::json{
        {config_json::VERSION, config_version::CURRENT},
        {config_json::ENGINE_OVERRIDES,
         nlohmann::json::array(
             {{{config_json::OP, config_op::POINTWISE},
               {config_json::CRITERIA, {{config_criterion::POINTWISE_MODE, 2}}},
               {config_json::ENGINE_NAME, MIOPEN_ENGINE_NAME},
               {config_json::TENSORS,
                nlohmann::json::array(
                    {{{config_json::TENSOR_ID, config_tensor::IN_0}, {config_json::DIM, {2, 4}}},
                     {{config_json::TENSOR_ID, config_tensor::IN_0},
                      {config_json::DIM, {5, 6}}}})}}})}}.dump());
    EXPECT_FALSE(config.has_value());
}

TEST(TestEngineOverrideConfig, LoadFromContentRejectsMissingTensorIdInNamedConfig)
{
    auto config = EngineOverrideConfig::loadFromContent(nlohmann::json{
        {config_json::VERSION, config_version::CURRENT},
        {config_json::ENGINE_OVERRIDES,
         nlohmann::json::array(
             {{{config_json::OP, config_op::POINTWISE},
               {config_json::CRITERIA, {{config_criterion::POINTWISE_MODE, 2}}},
               {config_json::ENGINE_NAME, MIOPEN_ENGINE_NAME},
               {config_json::TENSORS,
                nlohmann::json::array(
                    {{{config_json::TENSOR_ID, config_tensor::IN_0}, {config_json::DIM, {2, 4}}},
                     {{config_json::DIM, {5, 6}}}})}}})}}.dump());
    EXPECT_FALSE(config.has_value());
}

TEST(TestEngineOverrideConfig, FirstRuleDimMismatchFallsThroughToSecondMatchingRule)
{
    OperationRule first;
    first.op = config_op::POINTWISE;
    first.engineName = MIOPEN_ENGINE_NAME;
    first.criteria = {Criterion{config_criterion::POINTWISE_MODE, 2}};
    first.tensors = {makePattern({9, 9}), makePattern({5, 6})};

    OperationRule second;
    second.op = config_op::POINTWISE;
    second.engineName = HIPBLASLT_ENGINE_NAME;
    second.criteria = {Criterion{config_criterion::POINTWISE_MODE, 2}};
    second.tensors = {makePattern({2, 4}), makePattern({5, 6})};

    const auto config = makeConfig({std::move(first), std::move(second)});
    const std::vector<TensorData> tensors = {{{2, 4}, {}}, {{5, 6}, {}}};
    auto result = config.matchOperation(
        config_op::POINTWISE, {Criterion{config_criterion::POINTWISE_MODE, 2}}, viewsOf(tensors));
    ASSERT_TRUE(result.has_value());
    EXPECT_EQ(*result, HIPBLASLT_ENGINE_ID);
}

TEST(TestEngineOverrideConfig, FirstRuleCriteriaMismatchFallsThroughToSecondMatchingRule)
{
    OperationRule first;
    first.op = config_op::POINTWISE;
    first.engineName = MIOPEN_ENGINE_NAME;
    first.criteria = {Criterion{config_criterion::POINTWISE_MODE, 30}};
    first.tensors = {makePattern({2, 4}), makePattern({5, 6})};

    OperationRule second;
    second.op = config_op::POINTWISE;
    second.engineName = HIPBLASLT_ENGINE_NAME;
    second.criteria = {Criterion{config_criterion::POINTWISE_MODE, 2}};
    second.tensors = {makePattern({2, 4}), makePattern({5, 6})};

    const auto config = makeConfig({std::move(first), std::move(second)});
    const std::vector<TensorData> tensors = {{{2, 4}, {}}, {{5, 6}, {}}};
    auto result = config.matchOperation(
        config_op::POINTWISE, {Criterion{config_criterion::POINTWISE_MODE, 2}}, viewsOf(tensors));
    ASSERT_TRUE(result.has_value());
    EXPECT_EQ(*result, HIPBLASLT_ENGINE_ID);
}

TEST(TestEngineOverrideConfig, FirstRuleNamedTensorRoleMismatchFallsThroughToSecondMatchingRule)
{
    OperationRule first;
    first.op = config_op::POINTWISE;
    first.engineName = MIOPEN_ENGINE_NAME;
    first.criteria = {Criterion{config_criterion::POINTWISE_MODE, 2}};
    first.useNamedTensorIds = true;
    first.tensors = {makeNamedPattern(config_tensor::IN_0, {5, 6}),
                     makeNamedPattern(config_tensor::IN_1, {2, 4})};

    OperationRule second;
    second.op = config_op::POINTWISE;
    second.engineName = HIPBLASLT_ENGINE_NAME;
    second.criteria = {Criterion{config_criterion::POINTWISE_MODE, 2}};
    second.useNamedTensorIds = true;
    second.tensors = {makeNamedPattern(config_tensor::IN_0, {2, 4}),
                      makeNamedPattern(config_tensor::IN_1, {5, 6})};

    const auto config = makeConfig({std::move(first), std::move(second)});
    const TensorData x{{2, 4}, {}};
    const TensorData y{{5, 6}, {}};
    const std::vector<TensorView> tensors{namedViewOf(config_tensor::IN_0, x),
                                          namedViewOf(config_tensor::IN_1, y)};
    auto result = config.matchOperation(
        config_op::POINTWISE, {Criterion{config_criterion::POINTWISE_MODE, 2}}, tensors);
    ASSERT_TRUE(result.has_value());
    EXPECT_EQ(*result, HIPBLASLT_ENGINE_ID);
}

TEST(TestEngineOverrideConfig, NamedTensorRulesRejectSameShapeWrongRole)
{
    OperationRule rule;
    rule.op = config_op::SDPA_FWD;
    rule.engineName = MIOPEN_ENGINE_NAME;
    rule.useNamedTensorIds = true;
    rule.tensors = {makeNamedPattern(config_tensor::Q, {2, 4}),
                    makeNamedPattern(config_tensor::K, {2, 4}),
                    makeNamedPattern(config_tensor::V, {2, 4}),
                    makeNamedPattern(config_tensor::SCALE, {1})};

    const auto config = makeConfig({std::move(rule)});
    const TensorData q{{2, 4}, {}};
    const TensorData k{{2, 4}, {}};
    const TensorData v{{2, 4}, {}};
    const TensorData bias{{1}, {}};
    const std::vector<TensorView> tensors{namedViewOf(config_tensor::Q, q),
                                          namedViewOf(config_tensor::K, k),
                                          namedViewOf(config_tensor::V, v),
                                          namedViewOf(config_tensor::ATTN_MASK, bias)};

    EXPECT_FALSE(config.matchOperation(config_op::SDPA_FWD, tensors).has_value());
}

TEST(TestEngineOverrideConfig, LegacyRuleWithoutTensorIdsUsesPositionalOrder)
{
    constexpr const char* CONTENTS = R"({
  "engine_overrides": [
    {
      "op": "conv_fprop",
      "engine_name": "MIOPEN_ENGINE",
      "tensors": [
        { "dim": [1, 3, 4, 4] },
        { "dim": [2, 3, 1, 1] }
      ]
    }
  ]
})";

    auto config = EngineOverrideConfig::loadFromContent(CONTENTS);
    ASSERT_TRUE(config.has_value());

    const TensorData x{{1, 3, 4, 4}, {}};
    const TensorData w{{2, 3, 1, 1}, {}};
    const std::vector<TensorView> matching{namedViewOf("x_tensor_uid", x),
                                           namedViewOf("w_tensor_uid", w)};
    const std::vector<TensorView> reversed{namedViewOf("w_tensor_uid", w),
                                           namedViewOf("x_tensor_uid", x)};

    EXPECT_TRUE(config->matchOperation("conv_fprop", matching).has_value());
    EXPECT_FALSE(config->matchOperation("conv_fprop", reversed).has_value());
}

// ── Test 7: wrong tensor count in rule → nullopt ────────────────────────────

TEST(TestEngineOverrideConfig, WrongTensorCountReturnsNullopt)
{
    OperationRule rule;
    rule.op = "conv_fprop";
    rule.engineName = MIOPEN_ENGINE_NAME;
    rule.tensors = {makePattern({1, 3, 224, 224}), makePattern({64, 3, 7, 7})};

    const auto config = makeConfig({std::move(rule)});

    const std::vector<TensorData> tensors1 = {{{1, 3, 224, 224}, {}}};
    EXPECT_FALSE(config.matchOperation("conv_fprop", viewsOf(tensors1)).has_value());

    const std::vector<TensorData> tensors3
        = {{{1, 3, 224, 224}, {}}, {{64, 3, 7, 7}, {}}, {{64, 1, 1, 1}, {}}};
    EXPECT_FALSE(config.matchOperation("conv_fprop", viewsOf(tensors3)).has_value());
}

// ── Tests 11–12: cross-partition ordering (exact vs wildcard) ───────────────

TEST(TestEngineOverrideConfig, WildcardBeforeExactBothMatch)
{
    OperationRule wildcard;
    wildcard.op = "conv_fprop";
    wildcard.engineName = HIP_MLOPS_ENGINE_NAME;
    wildcard.tensors = {makePattern({-1, 3, 224, 224})};

    OperationRule exact;
    exact.op = "conv_fprop";
    exact.engineName = HIPBLASLT_ENGINE_NAME;
    exact.tensors = {makePattern({1, 3, 224, 224})};

    const auto config = makeConfig({std::move(wildcard), std::move(exact)});

    const std::vector<TensorData> tensors = {{{1, 3, 224, 224}, {}}};
    auto result = config.matchOperation("conv_fprop", viewsOf(tensors));
    ASSERT_TRUE(result.has_value());
    EXPECT_EQ(*result, HIP_MLOPS_ENGINE_ID);
}

TEST(TestEngineOverrideConfig, ExactBeforeWildcardBothMatch)
{
    OperationRule exact;
    exact.op = "conv_fprop";
    exact.engineName = HIPBLASLT_ENGINE_NAME;
    exact.tensors = {makePattern({1, 3, 224, 224})};

    OperationRule wildcard;
    wildcard.op = "conv_fprop";
    wildcard.engineName = HIP_MLOPS_ENGINE_NAME;
    wildcard.tensors = {makePattern({-1, 3, 224, 224})};

    const auto config = makeConfig({std::move(exact), std::move(wildcard)});

    const std::vector<TensorData> tensors = {{{1, 3, 224, 224}, {}}};
    auto result = config.matchOperation("conv_fprop", viewsOf(tensors));
    ASSERT_TRUE(result.has_value());
    EXPECT_EQ(*result, HIPBLASLT_ENGINE_ID);
}

// ── Stride matching tests ────────────────────────────────────────────────────

TEST(TestEngineOverrideConfig, ExactStrideMatchSelectsEngine)
{
    OperationRule rule;
    rule.op = "conv_fprop";
    rule.engineName = MIOPEN_ENGINE_NAME;
    rule.tensors = {makePatternWithStride({1, 3, 224, 224}, {150528, 50176, 224, 1})};

    const auto config = makeConfig({std::move(rule)});

    const std::vector<TensorData> matching = {{{1, 3, 224, 224}, {150528, 50176, 224, 1}}};
    auto result = config.matchOperation("conv_fprop", viewsOf(matching));
    ASSERT_TRUE(result.has_value());
    EXPECT_EQ(*result, MIOPEN_ENGINE_ID);

    const std::vector<TensorData> wrongStride
        = {{{1, 3, 224, 224}, {1, 224, int64_t{224} * 3, int64_t{224} * 3 * 224}}};
    EXPECT_FALSE(config.matchOperation("conv_fprop", viewsOf(wrongStride)).has_value());
}

TEST(TestEngineOverrideConfig, WildcardStrideElement)
{
    OperationRule rule;
    rule.op = "conv_fprop";
    rule.engineName = HIPBLASLT_ENGINE_NAME;
    rule.tensors = {makePatternWithStride({1, 3, 224, 224}, {150528, 50176, -1, -1})};

    const auto config = makeConfig({std::move(rule)});

    for(const int64_t s2 : {224, 112, 56})
    {
        const std::vector<TensorData> tensors = {{{1, 3, 224, 224}, {150528, 50176, s2, 1}}};
        auto result = config.matchOperation("conv_fprop", viewsOf(tensors));
        ASSERT_TRUE(result.has_value()) << "stride[2]=" << s2;
        EXPECT_EQ(*result, HIPBLASLT_ENGINE_ID);
    }

    const std::vector<TensorData> wrongStride = {{{1, 3, 224, 224}, {999, 50176, 224, 1}}};
    EXPECT_FALSE(config.matchOperation("conv_fprop", viewsOf(wrongStride)).has_value());
}

TEST(TestEngineOverrideConfig, EmptyStridePatternMatchesAnyStride)
{
    OperationRule rule;
    rule.op = "conv_fprop";
    rule.engineName = HIP_MLOPS_ENGINE_NAME;
    rule.tensors = {makePattern({1, 3, 224, 224})};

    const auto config = makeConfig({std::move(rule)});

    for(const auto& strides : std::vector<std::vector<int64_t>>{
            {150528, 50176, 224, 1}, {1, 3, 672, 150528}, {999, 888, 777, 666}})
    {
        const std::vector<TensorData> tensors = {{{1, 3, 224, 224}, strides}};
        auto result = config.matchOperation("conv_fprop", viewsOf(tensors));
        ASSERT_TRUE(result.has_value());
        EXPECT_EQ(*result, HIP_MLOPS_ENGINE_ID);
    }
}

// ── JSON-dependent ──────────────────────────────────────────────────────────

TEST(TestEngineOverrideConfig, LoadFromValidJsonFile)
{
    constexpr const char* CONTENTS = R"({
  "engine_overrides": [
    {
      "comment": "test rule for ResNet first conv",
      "op": "conv_fprop",
      "engine_name": "MIOPEN_ENGINE",
      "tensors": [
        { "dim": [1, 3, 224, 224] },
        { "dim": [64, 3, 7, 7] }
      ]
    },
    {
      "comment": "wildcard catch-all",
      "op": "conv_fprop",
      "engine_name": "HIP_MLOPS_ENGINE",
      "tensors": [
        { "dim": [-1, -1, -1, -1] },
        { "dim": [-1, -1, -1, -1] }
      ]
    }
  ]
})";

    auto config = EngineOverrideConfig::loadFromContent(CONTENTS);
    ASSERT_TRUE(config.has_value());

    const std::vector<TensorData> exact = {{{1, 3, 224, 224}, {}}, {{64, 3, 7, 7}, {}}};
    auto r1 = config->matchOperation("conv_fprop", viewsOf(exact));
    ASSERT_TRUE(r1.has_value());
    EXPECT_EQ(*r1, MIOPEN_ENGINE_ID);

    const std::vector<TensorData> other = {{{8, 64, 56, 56}, {}}, {{64, 64, 3, 3}, {}}};
    auto r2 = config->matchOperation("conv_fprop", viewsOf(other));
    ASSERT_TRUE(r2.has_value());
    EXPECT_EQ(*r2, HIP_MLOPS_ENGINE_ID);
}

TEST(TestEngineOverrideConfig, LoadFromJsonParsesCriteria)
{
    constexpr const char* CONTENTS = R"({
  "engine_overrides": [
    {
      "op": "pointwise",
      "criteria": { "pointwise_mode": 2 },
      "engine_name": "MIOPEN_ENGINE",
      "tensors": [
        { "dim": [2, 4, 16, 16] },
        { "dim": [2, 4, 16, 16] }
      ]
    }
  ]
})";

    auto config = EngineOverrideConfig::loadFromContent(CONTENTS);
    ASSERT_TRUE(config.has_value());

    const std::vector<TensorData> tensors = {{{2, 4, 16, 16}, {}}, {{2, 4, 16, 16}, {}}};
    auto result
        = config->matchOperation("pointwise", {Criterion{"pointwise_mode", 2}}, viewsOf(tensors));
    ASSERT_TRUE(result.has_value());
    EXPECT_EQ(*result, MIOPEN_ENGINE_ID);
    EXPECT_FALSE(
        config->matchOperation("pointwise", {Criterion{"pointwise_mode", 34}}, viewsOf(tensors))
            .has_value());
}

TEST(TestEngineOverrideConfig, LoadFromMissingFileReturnsNullopt)
{
    auto config = EngineOverrideConfig::load("/nonexistent/path/hipdnn_no_such_file.json");
    EXPECT_FALSE(config.has_value());
}

TEST(TestEngineOverrideConfig, EnvVarUnsetReturnsNullopt)
{
    EXPECT_FALSE(EngineOverrideConfig::loadFromEnv().has_value());
}

// ── Malformed JSON: parser must return nullopt, not throw ──────────────────

TEST(TestEngineOverrideConfig, LoadFromContentRejectsInvalidJsonSyntax)
{
    // Trailing comma + unterminated array — nlohmann::json::parse_error.
    constexpr const char* CONTENTS = R"({ "engine_overrides": [)";
    EXPECT_FALSE(EngineOverrideConfig::loadFromContent(CONTENTS).has_value());
}

TEST(TestEngineOverrideConfig, LoadFromContentRejectsMissingTopLevelKey)
{
    // Valid JSON, but no "engine_overrides" key — at() throws out_of_range.
    constexpr const char* CONTENTS = R"({ "other_key": [] })";
    EXPECT_FALSE(EngineOverrideConfig::loadFromContent(CONTENTS).has_value());
}

TEST(TestEngineOverrideConfig, LoadFromContentRejectsMissingEntryFields)
{
    // engine_overrides entry missing "engine_name" — at() throws out_of_range.
    constexpr const char* CONTENTS = R"({
  "engine_overrides": [
    { "op": "conv_fprop", "tensors": [ { "dim": [1, 3, 224, 224] } ] }
  ]
})";
    EXPECT_FALSE(EngineOverrideConfig::loadFromContent(CONTENTS).has_value());
}

TEST(TestEngineOverrideConfig, LoadFromContentRejectsMissingTensorDim)
{
    // tensor entry missing "dim" — at() throws out_of_range.
    constexpr const char* CONTENTS = R"({
  "engine_overrides": [
    {
      "op": "conv_fprop",
      "engine_name": "MIOPEN_ENGINE",
      "tensors": [ { "stride": [1, 2, 3, 4] } ]
    }
  ]
})";
    EXPECT_FALSE(EngineOverrideConfig::loadFromContent(CONTENTS).has_value());
}

TEST(TestEngineOverrideConfig, LoadFromContentRejectsWrongFieldType)
{
    // "op" is a number, not a string — get<string> throws type_error.
    constexpr const char* CONTENTS = R"({
  "engine_overrides": [
    {
      "op": 42,
      "engine_name": "MIOPEN_ENGINE",
      "tensors": [ { "dim": [1, 3, 224, 224] } ]
    }
  ]
})";
    EXPECT_FALSE(EngineOverrideConfig::loadFromContent(CONTENTS).has_value());
}

TEST(TestEngineOverrideConfig, JsonWithStrideConstraint)
{
    constexpr const char* CONTENTS = R"({
  "engine_overrides": [
    {
      "op": "conv_fprop",
      "engine_name": "MIOPEN_ENGINE",
      "tensors": [
        { "dim": [1, 3, 224, 224], "stride": [150528, 50176, 224, 1] },
        { "dim": [64, 3, 7, 7] }
      ]
    }
  ]
})";

    auto config = EngineOverrideConfig::loadFromContent(CONTENTS);
    ASSERT_TRUE(config.has_value());

    const std::vector<TensorData> matching
        = {{{1, 3, 224, 224}, {150528, 50176, 224, 1}}, {{64, 3, 7, 7}, {}}};
    auto r1 = config->matchOperation("conv_fprop", viewsOf(matching));
    ASSERT_TRUE(r1.has_value());
    EXPECT_EQ(*r1, MIOPEN_ENGINE_ID);

    const std::vector<TensorData> wrong
        = {{{1, 3, 224, 224}, {1, 224, int64_t{224} * 3, int64_t{224} * 3 * 224}},
           {{64, 3, 7, 7}, {}}};
    EXPECT_FALSE(config->matchOperation("conv_fprop", viewsOf(wrong)).has_value());
}