TheRock/rocm-libraries/dnn-providers/miopen-provider/integration_tests/IntegrationGpuBatchnormUnhappyActivation.cpp

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

#include <gtest/gtest.h>
#include <hip/hip_runtime.h>
#include <hipdnn_frontend.hpp>
#include <hipdnn_frontend/Graph.hpp>

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

using namespace hipdnn_frontend;
using namespace hipdnn_frontend::graph;
using namespace hipdnn_data_sdk::utilities;

namespace
{

/// Describes a single invalid BatchNorm + Activation configuration.
struct UnhappyBnActivationCase
{
    PointwiseMode pointwiseMode;
    bool isBackward;
    const char* name;
};

/// Builds a minimal BN + Pointwise graph that should fail at build time
/// due to unsupported activation fusion.
Graph makeGraph(const UnhappyBnActivationCase& tc)
{
    Graph g;

    g.set_name("IntegrationGpuBatchnormUnhappyActivation");
    g.set_compute_data_type(DataType::FLOAT)
        .set_intermediate_data_type(DataType::FLOAT)
        .set_io_data_type(DataType::FLOAT);

    // Valid 4D tensor (NCHW)
    const std::vector<int64_t> dims = {2, 4, 8, 8};
    const std::vector<int64_t> cDims = getDerivedShape(dims);

    // Input
    auto x = std::make_shared<TensorAttributes>(
        makeTensorAttributes("x", DataType::FLOAT, dims, generateStrides(dims)));

    // BN parameters
    auto mean = std::make_shared<TensorAttributes>(
        makeTensorAttributes("mean", DataType::FLOAT, cDims, generateStrides(cDims)));
    auto invVar = std::make_shared<TensorAttributes>(
        makeTensorAttributes("inv_variance", DataType::FLOAT, cDims, generateStrides(cDims)));
    auto scale = std::make_shared<TensorAttributes>(
        makeTensorAttributes("scale", DataType::FLOAT, cDims, generateStrides(cDims)));
    auto bias = std::make_shared<TensorAttributes>(
        makeTensorAttributes("bias", DataType::FLOAT, cDims, generateStrides(cDims)));

    std::shared_ptr<TensorAttributes> bnOut;

    if(!tc.isBackward)
    {
        // Forward BatchNorm
        const BatchnormInferenceAttributes bn;
        bnOut = g.batchnorm_inference(x, mean, invVar, scale, bias, bn);
    }
    else
    {
        // Backward BatchNorm
        auto dY = std::make_shared<TensorAttributes>(
            makeTensorAttributes("dY", DataType::FLOAT, dims, generateStrides(dims)));

        const BatchnormBackwardAttributes bn;

        auto bnOutputs = g.batchnorm_backward(dY, x, scale, bn);
        bnOut = bnOutputs[0]; // dX
    }

    // Add activation via pointwise op
    PointwiseAttributes pw;
    pw.set_mode(tc.pointwiseMode);

    // LeakyReLU workaround
    if(std::string(tc.name).find("LeakyRelu") != std::string::npos)
    {
        pw.set_relu_lower_clip_slope(0.01f);
    }

    auto z = g.pointwise(bnOut, pw);
    z->set_output(true);

    return g;
}

// Test fixture
class IntegrationGpuBatchnormUnhappyActivation
    : public ::testing::TestWithParam<UnhappyBnActivationCase>
{
protected:
    hipdnnHandle_t _handle = nullptr;
    hipStream_t _stream = nullptr;
    int _deviceId = 0;

    void SetUp() override
    {
        SKIP_IF_NO_DEVICES();

        ASSERT_EQ(hipInit(0), hipSuccess);
        ASSERT_EQ(hipGetDevice(&_deviceId), hipSuccess);

        auto pluginPath = std::filesystem::weakly_canonical(
            hipdnn_data_sdk::utilities::getCurrentExecutableDirectory() / PLUGIN_PATH);

        const std::string pluginPathStr = pluginPath.string();
        const std::array<const char*, 1> paths = {pluginPathStr.c_str()};

        ASSERT_EQ(hipdnnSetEnginePluginPaths_ext(
                      paths.size(), paths.data(), HIPDNN_PLUGIN_LOADING_ABSOLUTE),
                  HIPDNN_STATUS_SUCCESS);

        ASSERT_EQ(hipdnnCreate(&_handle), HIPDNN_STATUS_SUCCESS);
        ASSERT_EQ(hipStreamCreate(&_stream), hipSuccess);
        ASSERT_EQ(hipdnnSetStream(_handle, _stream), HIPDNN_STATUS_SUCCESS);
    }

    void TearDown() override
    {
        if(_handle != nullptr)
        {
            ASSERT_EQ(hipdnnDestroy(_handle), HIPDNN_STATUS_SUCCESS);
        }

        if(_stream != nullptr)
        {
            ASSERT_EQ(hipStreamDestroy(_stream), hipSuccess);
        }
    }
};

} // anonymous namespace

// Test body
TEST_P(IntegrationGpuBatchnormUnhappyActivation, RejectsUnsupportedActivations)
{
    const auto& tc = GetParam();

    auto graph = makeGraph(tc);
    auto result = graph.build(_handle);

    // Must fail
    EXPECT_NE(result.code, ErrorCode::OK) << "Unexpected success. err_msg: " << result.err_msg;

    // Must provide an error message
    EXPECT_FALSE(result.err_msg.empty()) << "Expected non-empty error message";

    // Classify failure
    const bool isBackendFailure = result.code == ErrorCode::HIPDNN_BACKEND_ERROR;

    const bool isFrontendFailure
        = result.code != ErrorCode::OK && result.code != ErrorCode::HIPDNN_BACKEND_ERROR;

    EXPECT_TRUE(isBackendFailure || isFrontendFailure)
        << "Unexpected error code: " << static_cast<int>(result.code)
        << ", err_msg: " << result.err_msg;
}

// Since ReLU is the only supported fusion (not leakyReLU),
// also test that it is accepted in the forward case to validate that the test is correctly set up
// to allow supported fusions and reject unsupported ones.
TEST_F(IntegrationGpuBatchnormUnhappyActivation, AcceptsReluForwardFusion)
{
    Graph g;

    g.set_name("IntegrationGpuBatchnormHappyRelu");
    g.set_compute_data_type(DataType::FLOAT)
        .set_intermediate_data_type(DataType::FLOAT)
        .set_io_data_type(DataType::FLOAT);

    // Valid 4D tensor
    const std::vector<int64_t> dims = {2, 4, 8, 8};
    const std::vector<int64_t> cDims = getDerivedShape(dims);

    // Input
    auto x = std::make_shared<TensorAttributes>(
        makeTensorAttributes("x", DataType::FLOAT, dims, generateStrides(dims)));

    // BN parameters
    auto mean = std::make_shared<TensorAttributes>(
        makeTensorAttributes("mean", DataType::FLOAT, cDims, generateStrides(cDims)));
    auto invVar = std::make_shared<TensorAttributes>(
        makeTensorAttributes("inv_variance", DataType::FLOAT, cDims, generateStrides(cDims)));
    auto scale = std::make_shared<TensorAttributes>(
        makeTensorAttributes("scale", DataType::FLOAT, cDims, generateStrides(cDims)));
    auto bias = std::make_shared<TensorAttributes>(
        makeTensorAttributes("bias", DataType::FLOAT, cDims, generateStrides(cDims)));

    // Forward BatchNorm
    const BatchnormInferenceAttributes bn;
    auto bnOut = g.batchnorm_inference(x, mean, invVar, scale, bias, bn);

    // ReLU activation (supported fusion)
    PointwiseAttributes pw;
    pw.set_mode(PointwiseMode::RELU_FWD);

    auto z = g.pointwise(bnOut, pw);
    z->set_output(true);

    // Build graph
    auto result = g.build(_handle);

    // Must succeed
    EXPECT_EQ(result.code, ErrorCode::OK)
        << "Expected successful build but got error: " << result.err_msg;
}

// Test cases
// LeakyReLU has been emulated by using RELU_FWD with a non-zero negative slope,
// since frontend does not directly expose a LeakyReLU mode.
INSTANTIATE_TEST_SUITE_P(
    Smoke,
    IntegrationGpuBatchnormUnhappyActivation,
    ::testing::Values(UnhappyBnActivationCase{PointwiseMode::SIGMOID_FWD, false, "SigmoidFwd"},
                      UnhappyBnActivationCase{PointwiseMode::TANH_FWD, false, "TanhFwd"},
                      UnhappyBnActivationCase{PointwiseMode::RELU_FWD, false, "LeakyReluFwd"},
                      UnhappyBnActivationCase{PointwiseMode::SIGMOID_BWD, true, "SigmoidBwd"},
                      UnhappyBnActivationCase{PointwiseMode::RELU_FWD, true, "ReluFwdInBackward"}),
    [](const ::testing::TestParamInfo<UnhappyBnActivationCase>& info) {
        return std::string(info.param.name);
    });