TheRock/rocm-libraries/projects/hipdnn/frontend/tests/TestConvolutionWgradNode.cpp

// Copyright © Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier:  MIT
#include <gtest/gtest.h>
#include <hipdnn_frontend/Error.hpp>
#include <hipdnn_frontend/attributes/ConvolutionWgradAttributes.hpp>
#include <hipdnn_frontend/node/ConvolutionWgradNode.hpp>
#include <hipdnn_test_sdk/constants/ConvFpropConstants.hpp>
#include <hipdnn_test_sdk/utilities/ToVec.hpp>

#include "fake_backend/MockHipdnnBackend.hpp"

#include <gmock/gmock.h>

using namespace hipdnn_frontend;
using namespace hipdnn_frontend::graph;
using namespace hipdnn_tests::constants;
using hipdnn_tests::toVec;
using namespace ::testing;

TEST(TestConvolutionWgradNode, PreValidateNode)
{
    ConvWgradAttributes convAttributes;

    auto xTensor = std::make_shared<TensorAttributes>();
    xTensor->set_dim({1, 3, 32, 32});
    xTensor->set_stride({3072, 1024, 32, 1});
    convAttributes.set_x(xTensor);

    auto dyTensor = std::make_shared<TensorAttributes>();
    dyTensor->set_dim({1, 64, 32, 32});
    dyTensor->set_stride({65536, 1024, 32, 1});
    convAttributes.set_dy(dyTensor);

    auto dwTensor = std::make_shared<TensorAttributes>();
    dwTensor->set_dim({64, 3, 3, 3});
    convAttributes.set_dw(dwTensor);

    convAttributes.set_pre_padding({1, 1});
    convAttributes.set_post_padding({1, 1});
    convAttributes.set_stride({1, 1});
    convAttributes.set_dilation({1, 1});

    const GraphAttributes graphAttributes;
    const ConvolutionWgradNode node(std::move(convAttributes), graphAttributes);

    auto error = node.pre_validate_node();
    EXPECT_EQ(error.code, error_code_t::OK) << error.err_msg;
}

TEST(TestConvolutionWgradNode, PreValidateNodeMissingXTensor)
{
    ConvWgradAttributes convAttributes;

    auto dyTensor = std::make_shared<TensorAttributes>();
    dyTensor->set_dim({1, 64, 32, 32});
    dyTensor->set_stride({65536, 1024, 32, 1});
    convAttributes.set_dy(dyTensor);

    auto dwTensor = std::make_shared<TensorAttributes>();
    dwTensor->set_dim({64, 3, 3, 3});
    dwTensor->set_stride({27, 9, 3, 1});
    convAttributes.set_dw(dwTensor);

    convAttributes.set_pre_padding({1, 1});
    convAttributes.set_post_padding({1, 1});
    convAttributes.set_stride({1, 1});
    convAttributes.set_dilation({1, 1});

    // X tensor is missing
    const GraphAttributes graphAttributes;
    const ConvolutionWgradNode node(std::move(convAttributes), graphAttributes);

    auto error = node.pre_validate_node();
    EXPECT_EQ(error.code, error_code_t::ATTRIBUTE_NOT_SET);
}

TEST(TestConvolutionWgradNode, PreValidateNodeMissingDyTensor)
{
    ConvWgradAttributes convAttributes;

    auto xTensor = std::make_shared<TensorAttributes>();
    xTensor->set_dim({1, 3, 32, 32});
    xTensor->set_stride({3072, 1024, 32, 1});
    convAttributes.set_x(xTensor);

    auto dwTensor = std::make_shared<TensorAttributes>();
    dwTensor->set_dim({64, 3, 3, 3});
    dwTensor->set_stride({27, 9, 3, 1});
    convAttributes.set_dw(dwTensor);

    convAttributes.set_pre_padding({1, 1});
    convAttributes.set_post_padding({1, 1});
    convAttributes.set_stride({1, 1});
    convAttributes.set_dilation({1, 1});

    // Dy tensor is missing
    const GraphAttributes graphAttributes;
    const ConvolutionWgradNode node(std::move(convAttributes), graphAttributes);

    auto error = node.pre_validate_node();
    EXPECT_EQ(error.code, error_code_t::ATTRIBUTE_NOT_SET);
}

TEST(TestConvolutionWgradNode, PreValidateNodeMissingDwTensor)
{
    ConvWgradAttributes convAttributes;

    auto xTensor = std::make_shared<TensorAttributes>();
    xTensor->set_dim({1, 3, 32, 32});
    xTensor->set_stride({3072, 1024, 32, 1});
    convAttributes.set_x(xTensor);

    auto dyTensor = std::make_shared<TensorAttributes>();
    dyTensor->set_dim({1, 64, 32, 32});
    dyTensor->set_stride({65536, 1024, 32, 1});
    convAttributes.set_dy(dyTensor);

    convAttributes.set_pre_padding({1, 1});
    convAttributes.set_post_padding({1, 1});
    convAttributes.set_stride({1, 1});
    convAttributes.set_dilation({1, 1});

    // Dw tensor is missing
    const GraphAttributes graphAttributes;
    const ConvolutionWgradNode node(std::move(convAttributes), graphAttributes);

    auto error = node.pre_validate_node();
    EXPECT_EQ(error.code, error_code_t::ATTRIBUTE_NOT_SET);
}

TEST(TestConvolutionWgradNode, PreValidateNodeMissingDwDims)
{
    ConvWgradAttributes convAttributes;

    auto xTensor = std::make_shared<TensorAttributes>();
    xTensor->set_dim({1, 3, 32, 32});
    xTensor->set_stride({3072, 1024, 32, 1});
    convAttributes.set_x(xTensor);

    auto dyTensor = std::make_shared<TensorAttributes>();
    dyTensor->set_dim({1, 64, 32, 32});
    dyTensor->set_stride({65536, 1024, 32, 1});
    convAttributes.set_dy(dyTensor);

    // Dw tensor is present but its dimensions are omitted; inference is not
    // supported, so pre-validation must reject at the first gate (issue #8530).
    auto dwTensor = std::make_shared<TensorAttributes>();
    convAttributes.set_dw(dwTensor);

    convAttributes.set_pre_padding({1, 1});
    convAttributes.set_post_padding({1, 1});
    convAttributes.set_stride({1, 1});
    convAttributes.set_dilation({1, 1});

    const GraphAttributes graphAttributes;
    const ConvolutionWgradNode node(std::move(convAttributes), graphAttributes);

    auto error = node.pre_validate_node();
    EXPECT_EQ(error.code, error_code_t::ATTRIBUTE_NOT_SET);
    // Assert the dw-specific gate fired, not some other ATTRIBUTE_NOT_SET gate.
    EXPECT_NE(error.err_msg.find("set dw dimensions explicitly"), std::string::npos)
        << error.err_msg;
}

// Regression: a zero stride must be rejected by the positivity check before the
// stride division in the spatial-compatibility loop, never divide-by-zero.
TEST(TestConvolutionWgradNode, PreValidateNodeZeroStride)
{
    ConvWgradAttributes convAttributes;

    auto xTensor = std::make_shared<TensorAttributes>();
    xTensor->set_dim({1, 3, 32, 32});
    xTensor->set_stride({3072, 1024, 32, 1});
    convAttributes.set_x(xTensor);

    auto dyTensor = std::make_shared<TensorAttributes>();
    dyTensor->set_dim({1, 64, 32, 32});
    dyTensor->set_stride({65536, 1024, 32, 1});
    convAttributes.set_dy(dyTensor);

    auto dwTensor = std::make_shared<TensorAttributes>();
    dwTensor->set_dim({64, 3, 3, 3});
    dwTensor->set_stride({27, 9, 3, 1});
    convAttributes.set_dw(dwTensor);

    convAttributes.set_pre_padding({1, 1});
    convAttributes.set_post_padding({1, 1});
    convAttributes.set_stride({1, 0}); // Zero stride in the second spatial dim
    convAttributes.set_dilation({1, 1});

    const GraphAttributes graphAttributes;
    const ConvolutionWgradNode node(std::move(convAttributes), graphAttributes);

    auto error = node.pre_validate_node();
    EXPECT_EQ(error.code, error_code_t::INVALID_VALUE);
    EXPECT_NE(error.err_msg.find("Stride must be > 0"), std::string::npos) << error.err_msg;
}

TEST(TestConvolutionWgradNode, PreValidateNodeMissingConvolutionParameters)
{
    ConvWgradAttributes convAttributes;

    auto xTensor = std::make_shared<TensorAttributes>();
    xTensor->set_dim({1, 3, 32, 32});
    xTensor->set_stride({3072, 1024, 32, 1});
    convAttributes.set_x(xTensor);

    auto dyTensor = std::make_shared<TensorAttributes>();
    dyTensor->set_dim({1, 64, 32, 32});
    dyTensor->set_stride({65536, 1024, 32, 1});
    convAttributes.set_dy(dyTensor);

    auto dwTensor = std::make_shared<TensorAttributes>();
    dwTensor->set_dim({64, 3, 3, 3});
    dwTensor->set_stride({27, 9, 3, 1});
    convAttributes.set_dw(dwTensor);

    // Convolution parameters are missing
    const GraphAttributes graphAttributes;
    const ConvolutionWgradNode node(std::move(convAttributes), graphAttributes);

    auto error = node.pre_validate_node();
    EXPECT_EQ(error.code, error_code_t::ATTRIBUTE_NOT_SET);
}

TEST(TestConvolutionWgradNode, PreValidateNodeAllValuesSet)
{
    ConvWgradAttributes convAttributes;

    auto xTensor = std::make_shared<TensorAttributes>();
    xTensor->set_dim({1, 3, 32, 32});
    xTensor->set_stride({3072, 1024, 32, 1});
    convAttributes.set_x(xTensor);

    auto dyTensor = std::make_shared<TensorAttributes>();
    dyTensor->set_dim({1, 64, 32, 32});
    dyTensor->set_stride({65536, 1024, 32, 1});
    convAttributes.set_dy(dyTensor);

    auto dwTensor = std::make_shared<TensorAttributes>();
    dwTensor->set_dim({64, 3, 3, 3});
    dwTensor->set_stride({27, 9, 3, 1});
    convAttributes.set_dw(dwTensor);

    convAttributes.set_pre_padding({1, 1});
    convAttributes.set_post_padding({1, 1});
    convAttributes.set_stride({1, 1});
    convAttributes.set_dilation({1, 1});

    const GraphAttributes graphAttributes;
    const ConvolutionWgradNode node(std::move(convAttributes), graphAttributes);

    auto error = node.pre_validate_node();
    EXPECT_EQ(error.code, error_code_t::OK) << error.err_msg;
}

TEST(TestConvolutionWgradNode, InferPropertiesNodeMissingXTensor)
{
    ConvWgradAttributes convAttributes;
    const GraphAttributes graphAttributes;
    ConvolutionWgradNode node(std::move(convAttributes), graphAttributes);

    auto error = node.infer_properties_node();
    EXPECT_EQ(error.code, error_code_t::ATTRIBUTE_NOT_SET);
}

TEST(TestConvolutionWgradNode, InferPropertiesNodeMissingDyTensor)
{
    ConvWgradAttributes convAttributes;
    convAttributes.set_x(std::make_shared<TensorAttributes>());
    const GraphAttributes graphAttributes;
    ConvolutionWgradNode node(std::move(convAttributes), graphAttributes);

    auto error = node.infer_properties_node();
    EXPECT_EQ(error.code, error_code_t::ATTRIBUTE_NOT_SET);
}

TEST(TestConvolutionWgradNode, InferPropertiesNodeMissingDwTensor)
{
    ConvWgradAttributes convAttributes;
    convAttributes.set_x(std::make_shared<TensorAttributes>());
    convAttributes.set_dy(std::make_shared<TensorAttributes>());
    const GraphAttributes graphAttributes;
    ConvolutionWgradNode node(std::move(convAttributes), graphAttributes);

    auto error = node.infer_properties_node();
    EXPECT_EQ(error.code, error_code_t::ATTRIBUTE_NOT_SET);
}

TEST(TestConvolutionWgradNode, InferPropertiesNodeMissingDwDimensions)
{
    ConvWgradAttributes convAttributes;

    auto xTensor = std::make_shared<TensorAttributes>();
    xTensor->set_dim({1, 3, 32, 32});
    xTensor->set_stride({3072, 1024, 32, 1});
    convAttributes.set_x(xTensor);

    auto dyTensor = std::make_shared<TensorAttributes>();
    dyTensor->set_dim({1, 64, 32, 32});
    dyTensor->set_stride({65536, 1024, 32, 1});
    convAttributes.set_dy(dyTensor);

    auto dwTensor = std::make_shared<TensorAttributes>();
    convAttributes.set_dw(dwTensor);

    convAttributes.set_pre_padding({1, 1});
    convAttributes.set_post_padding({1, 1});
    convAttributes.set_stride({1, 1});
    convAttributes.set_dilation({1, 1});

    const GraphAttributes graphAttributes;
    ConvolutionWgradNode node(std::move(convAttributes), graphAttributes);

    auto error = node.infer_properties_node();
    EXPECT_EQ(error.code, error_code_t::ATTRIBUTE_NOT_SET);
    EXPECT_TRUE(dwTensor->get_dim().empty());
}

TEST(TestConvolutionWgradNode, InferPropertiesNodeMissingDwDimensionsIssue8530Ambiguous)
{
    ConvWgradAttributes convAttributes;

    auto xTensor = std::make_shared<TensorAttributes>();
    xTensor->set_dim({2, 128, 56, 56});
    xTensor->set_stride({401408, 3136, 56, 1});
    convAttributes.set_x(xTensor);

    auto dyTensor = std::make_shared<TensorAttributes>();
    dyTensor->set_dim({2, 128, 28, 28});
    dyTensor->set_stride({100352, 784, 28, 1});
    convAttributes.set_dy(dyTensor);

    auto dwTensor = std::make_shared<TensorAttributes>();
    convAttributes.set_dw(dwTensor);

    convAttributes.set_pre_padding({1, 1});
    convAttributes.set_post_padding({1, 1});
    convAttributes.set_stride({2, 2});
    convAttributes.set_dilation({1, 1});

    const GraphAttributes graphAttributes;
    ConvolutionWgradNode node(std::move(convAttributes), graphAttributes);

    auto error = node.infer_properties_node();
    EXPECT_EQ(error.code, error_code_t::ATTRIBUTE_NOT_SET);
    EXPECT_TRUE(dwTensor->get_dim().empty());
}

TEST(TestConvolutionWgradNode, InferPropertiesNode2DConvolutionSuccess)
{
    ConvWgradAttributes convAttributes;

    auto xTensor = std::make_shared<TensorAttributes>();
    xTensor->set_dim({1, 3, 32, 32});
    xTensor->set_stride({3072, 1024, 32, 1});
    convAttributes.set_x(xTensor);

    auto dyTensor = std::make_shared<TensorAttributes>();
    dyTensor->set_dim({1, 64, 32, 32});
    dyTensor->set_stride({65536, 1024, 32, 1});
    convAttributes.set_dy(dyTensor);

    auto dwTensor = std::make_shared<TensorAttributes>();
    dwTensor->set_dim({64, 3, 3, 3});
    convAttributes.set_dw(dwTensor);

    convAttributes.set_pre_padding({1, 1});
    convAttributes.set_post_padding({1, 1});
    convAttributes.set_stride({1, 1});
    convAttributes.set_dilation({1, 1});

    const GraphAttributes graphAttributes;
    ConvolutionWgradNode node(std::move(convAttributes), graphAttributes);

    auto error = node.infer_properties_node();
    EXPECT_EQ(error.code, error_code_t::OK) << error.err_msg;

    auto dwDims = dwTensor->get_dim();
    EXPECT_EQ(dwDims.size(), 4);
    EXPECT_EQ(dwDims[0], 64); // Output channels
    EXPECT_EQ(dwDims[1], 3); // Input channels
    // Kernel height: ((32 + 1 + 1 - 1 * (32 - 1) - 1) / 1) + 1 = (32 + 2 - 31 - 1) / 1 + 1 = 2 / 1 + 1 = 3
    EXPECT_EQ(dwDims[2], 3);
    EXPECT_EQ(dwDims[3], 3);
}

TEST(TestConvolutionWgradNode, InferPropertiesNode3DConvolutionSuccess)
{
    ConvWgradAttributes convAttributes;

    auto xTensor = std::make_shared<TensorAttributes>();
    xTensor->set_dim({2, 16, 8, 16, 16});
    xTensor->set_stride({32768, 2048, 256, 16, 1});
    convAttributes.set_x(xTensor);

    auto dyTensor = std::make_shared<TensorAttributes>();
    dyTensor->set_dim({2, 32, 6, 16, 16});
    dyTensor->set_stride({49152, 1536, 256, 16, 1});
    convAttributes.set_dy(dyTensor);

    auto dwTensor = std::make_shared<TensorAttributes>();
    dwTensor->set_dim({32, 16, 3, 3, 3});
    convAttributes.set_dw(dwTensor);

    convAttributes.set_pre_padding({0, 1, 1});
    convAttributes.set_post_padding({0, 1, 1});
    convAttributes.set_stride({1, 1, 1});
    convAttributes.set_dilation({1, 1, 1});

    const GraphAttributes graphAttributes;
    ConvolutionWgradNode node(std::move(convAttributes), graphAttributes);

    auto error = node.infer_properties_node();
    EXPECT_EQ(error.code, error_code_t::OK) << error.err_msg;

    auto dwDims = dwTensor->get_dim();
    EXPECT_EQ(dwDims.size(), 5);
    EXPECT_EQ(dwDims[0], 32); // Output channels
    EXPECT_EQ(dwDims[1], 16); // Input channels
    // Depth kernel: ((8 + 0 + 0 - 1 * (6 - 1) - 1) / 1) + 1 = (8 - 5 - 1) / 1 + 1 = 2 / 1 + 1 = 3
    EXPECT_EQ(dwDims[2], 3);
    // Height kernel: ((16 + 1 + 1 - 1 * (16 - 1) - 1) / 1) + 1 = (16 + 2 - 15 - 1) / 1 + 1 = 2 / 1 + 1 = 3
    EXPECT_EQ(dwDims[3], 3);
    EXPECT_EQ(dwDims[4], 3);
}

TEST(TestConvolutionWgradNode, InferPropertiesNodeWithStride2x2)
{
    ConvWgradAttributes convAttributes;

    auto xTensor = std::make_shared<TensorAttributes>();
    xTensor->set_dim({1, 3, 31, 31});
    xTensor->set_stride({2883, 961, 31, 1});
    convAttributes.set_x(xTensor);

    auto dyTensor = std::make_shared<TensorAttributes>();
    dyTensor->set_dim({1, 64, 16, 16});
    dyTensor->set_stride({16384, 256, 16, 1});
    convAttributes.set_dy(dyTensor);

    auto dwTensor = std::make_shared<TensorAttributes>();
    dwTensor->set_dim({64, 3, 3, 3});
    convAttributes.set_dw(dwTensor);

    convAttributes.set_pre_padding({1, 1});
    convAttributes.set_post_padding({1, 1});
    convAttributes.set_stride({2, 2}); // 2x2 stride
    convAttributes.set_dilation({1, 1});

    const GraphAttributes graphAttributes;
    ConvolutionWgradNode node(std::move(convAttributes), graphAttributes);

    auto error = node.infer_properties_node();
    EXPECT_EQ(error.code, error_code_t::OK) << error.err_msg;

    auto dwDims = dwTensor->get_dim();
    EXPECT_EQ(dwDims.size(), 4);
    EXPECT_EQ(dwDims[0], 64); // Output channels
    EXPECT_EQ(dwDims[1], 3); // Input channels
    // Kernel height: ((31 + 1 + 1 - 2 * (16 - 1) - 1) / 1) + 1 = (31 + 2 - 30 - 1) / 1 + 1 = 2 / 1 + 1 = 3
    EXPECT_EQ(dwDims[2], 3);
    EXPECT_EQ(dwDims[3], 3);
}

TEST(TestConvolutionWgradNode, InferPropertiesNodeWithDilation2x2)
{
    ConvWgradAttributes convAttributes;

    auto xTensor = std::make_shared<TensorAttributes>();
    xTensor->set_dim({1, 16, 20, 20});
    xTensor->set_stride({6400, 400, 20, 1});
    convAttributes.set_x(xTensor);

    auto dyTensor = std::make_shared<TensorAttributes>();
    dyTensor->set_dim({1, 32, 20, 20});
    dyTensor->set_stride({12800, 400, 20, 1});
    convAttributes.set_dy(dyTensor);

    auto dwTensor = std::make_shared<TensorAttributes>();
    dwTensor->set_dim({32, 16, 3, 3});
    convAttributes.set_dw(dwTensor);

    convAttributes.set_pre_padding({2, 2});
    convAttributes.set_post_padding({2, 2});
    convAttributes.set_stride({1, 1});
    convAttributes.set_dilation({2, 2}); // 2x2 dilation

    const GraphAttributes graphAttributes;
    ConvolutionWgradNode node(std::move(convAttributes), graphAttributes);

    auto error = node.infer_properties_node();
    EXPECT_EQ(error.code, error_code_t::OK) << error.err_msg;

    auto dwDims = dwTensor->get_dim();
    EXPECT_EQ(dwDims.size(), 4);
    EXPECT_EQ(dwDims[0], 32); // Output channels
    EXPECT_EQ(dwDims[1], 16); // Input channels
    // Kernel height: ((20 + 2 + 2 - 1 * (20 - 1) - 1) / 2) + 1 = (20 + 4 - 19 - 1) / 2 + 1 = 4 / 2 + 1 = 3
    EXPECT_EQ(dwDims[2], 3);
    EXPECT_EQ(dwDims[3], 3);
}

TEST(TestConvolutionWgradNode, GatherHipdnnTensor)
{
    ConvWgradAttributes convAttributes;
    auto xTensor = std::make_shared<TensorAttributes>();
    xTensor->set_uid(1).set_name("XTensor");
    convAttributes.set_x(xTensor);

    auto dyTensor = std::make_shared<TensorAttributes>();
    dyTensor->set_uid(2).set_name("DyTensor");
    convAttributes.set_dy(dyTensor);

    auto dwTensor = std::make_shared<TensorAttributes>();
    dwTensor->set_uid(3).set_name("DwTensor");
    convAttributes.set_dw(dwTensor);

    convAttributes.set_pre_padding({1, 1});
    convAttributes.set_post_padding({1, 1});
    convAttributes.set_stride({1, 1});
    convAttributes.set_dilation({1, 1});

    const GraphAttributes graphAttributes;
    const ConvolutionWgradNode node(std::move(convAttributes), graphAttributes);

    std::unordered_set<std::shared_ptr<TensorAttributes>> allTensors;
    node.gather_hipdnn_tensors(allTensors);

    EXPECT_TRUE(allTensors.find(dwTensor) != allTensors.end());
    EXPECT_TRUE(allTensors.find(dyTensor) != allTensors.end());
    EXPECT_TRUE(allTensors.find(xTensor) != allTensors.end());
    EXPECT_EQ(allTensors.size(), 3);
}

TEST(TestConvolutionWgradNode, StrideInferenceNchwLayoutSuccess)
{
    ConvWgradAttributes convAttributes;

    auto xTensor = std::make_shared<TensorAttributes>();
    xTensor->set_dim({1, 3, 32, 32});
    xTensor->set_stride({3072, 1024, 32, 1}); // NCHW layout
    convAttributes.set_x(xTensor);

    auto dyTensor = std::make_shared<TensorAttributes>();
    dyTensor->set_dim({1, 64, 32, 32});
    dyTensor->set_stride({65536, 1024, 32, 1});
    convAttributes.set_dy(dyTensor);

    auto dwTensor = std::make_shared<TensorAttributes>();
    dwTensor->set_dim({64, 3, 3, 3});
    // No stride set - should be inferred
    convAttributes.set_dw(dwTensor);

    convAttributes.set_pre_padding({1, 1});
    convAttributes.set_post_padding({1, 1});
    convAttributes.set_stride({1, 1});
    convAttributes.set_dilation({1, 1});

    const GraphAttributes graphAttributes;
    ConvolutionWgradNode node(std::move(convAttributes), graphAttributes);

    auto error = node.infer_properties_node();
    EXPECT_EQ(error.code, error_code_t::OK) << error.err_msg;

    auto inferredStrides = dwTensor->get_stride();
    EXPECT_EQ(inferredStrides.size(), 4);
    EXPECT_EQ(inferredStrides[0], 27); // K stride: 3 * 3 * 3
    EXPECT_EQ(inferredStrides[1], 9); // C stride: 3 * 3
    EXPECT_EQ(inferredStrides[2], 3); // R stride: 3
    EXPECT_EQ(inferredStrides[3], 1); // S stride: 1
}

TEST(TestConvolutionWgradNode, StrideInferenceNhwcLayoutSuccess)
{
    ConvWgradAttributes convAttributes;

    auto xTensor = std::make_shared<TensorAttributes>();
    xTensor->set_dim({1, 3, 32, 32}); // Dims in NCHW order
    xTensor->set_stride({3072, 1, 96, 3}); // NHWC strides (channels last)
    convAttributes.set_x(xTensor);

    auto dyTensor = std::make_shared<TensorAttributes>();
    dyTensor->set_dim({1, 64, 32, 32}); // Dims in NCHW order
    dyTensor->set_stride({65536, 1, 2048, 64}); // NHWC strides (channels last)
    convAttributes.set_dy(dyTensor);

    auto dwTensor = std::make_shared<TensorAttributes>();
    dwTensor->set_dim({64, 3, 3, 3}); // Dims in KCRS order
    convAttributes.set_dw(dwTensor);

    convAttributes.set_pre_padding({1, 1});
    convAttributes.set_post_padding({1, 1});
    convAttributes.set_stride({1, 1});
    convAttributes.set_dilation({1, 1});

    const GraphAttributes graphAttributes;
    ConvolutionWgradNode node(std::move(convAttributes), graphAttributes);

    auto error = node.infer_properties_node();
    EXPECT_EQ(error.code, error_code_t::OK) << error.err_msg;

    auto inferredStrides = dwTensor->get_stride();
    EXPECT_EQ(inferredStrides.size(), 4);
    EXPECT_EQ(inferredStrides[0], 27); // K stride: 3 * 3 * 3
    EXPECT_EQ(inferredStrides[1], 1); // C stride: 1 (channels last)
    EXPECT_EQ(inferredStrides[2], 9); // R stride: 3 * 3
    EXPECT_EQ(inferredStrides[3], 3); // S stride: 3
}

TEST(TestConvolutionWgradNode, InferPropertiesGroupedConv2Groups)
{
    ConvWgradAttributes convAttributes;

    auto xTensor = std::make_shared<TensorAttributes>();
    xTensor->set_dim({1, 64, 32, 32}); // 64 input channels
    xTensor->set_stride({65536, 1024, 32, 1});
    convAttributes.set_x(xTensor);

    auto dyTensor = std::make_shared<TensorAttributes>();
    dyTensor->set_dim({1, 128, 32, 32}); // 128 output channels
    dyTensor->set_stride({131072, 1024, 32, 1});
    convAttributes.set_dy(dyTensor);

    auto dwTensor = std::make_shared<TensorAttributes>();
    dwTensor->set_dim({128, 64, 3, 3});
    convAttributes.set_dw(dwTensor);

    convAttributes.set_pre_padding({1, 1});
    convAttributes.set_post_padding({1, 1});
    convAttributes.set_stride({1, 1});
    convAttributes.set_dilation({1, 1});

    const GraphAttributes graphAttributes;
    ConvolutionWgradNode node(std::move(convAttributes), graphAttributes);

    auto error = node.infer_properties_node();
    EXPECT_EQ(error.code, error_code_t::OK) << error.err_msg;

    auto dwDims = dwTensor->get_dim();
    EXPECT_EQ(dwDims.size(), 4);
    EXPECT_EQ(dwDims[0], 128); // Output channels
    EXPECT_EQ(dwDims[1], 64); // Input channels
    EXPECT_EQ(dwDims[2], 3); // Kernel height
    EXPECT_EQ(dwDims[3], 3); // Kernel width
}

TEST(TestConvolutionWgradNode, PreValidateTensorDimsTooFew)
{
    ConvWgradAttributes convAttributes;

    // Test with too few dimensions (less than 3)
    auto xTensor = std::make_shared<TensorAttributes>();
    xTensor->set_dim({1, 3}); // Only 2 dimensions
    xTensor->set_stride({3, 1});
    convAttributes.set_x(xTensor);

    auto dyTensor = std::make_shared<TensorAttributes>();
    dyTensor->set_dim({1, 64});
    dyTensor->set_stride({64, 1});
    convAttributes.set_dy(dyTensor);

    auto dwTensor = std::make_shared<TensorAttributes>();
    dwTensor->set_dim({64, 3});
    convAttributes.set_dw(dwTensor);

    convAttributes.set_pre_padding({1});
    convAttributes.set_post_padding({1});
    convAttributes.set_stride({1});
    convAttributes.set_dilation({1});

    const GraphAttributes graphAttributes;
    const ConvolutionWgradNode node(std::move(convAttributes), graphAttributes);

    auto error = node.pre_validate_node();
    EXPECT_EQ(error.code, error_code_t::INVALID_VALUE);
}

TEST(TestConvolutionWgradNode, PreValidateDyDimsMismatch)
{
    ConvWgradAttributes convAttributes;

    auto xTensor = std::make_shared<TensorAttributes>();
    xTensor->set_dim({1, 3, 32, 32});
    xTensor->set_stride({3072, 1024, 32, 1});
    convAttributes.set_x(xTensor);

    auto dyTensor = std::make_shared<TensorAttributes>();
    dyTensor->set_dim({1, 64, 32, 32, 32}); // 5D dy for 4D x
    dyTensor->set_stride({2097152, 32768, 1024, 32, 1});
    convAttributes.set_dy(dyTensor);

    auto dwTensor = std::make_shared<TensorAttributes>();
    dwTensor->set_dim({64, 3, 3, 3});
    convAttributes.set_dw(dwTensor);

    convAttributes.set_pre_padding({1, 1});
    convAttributes.set_post_padding({1, 1});
    convAttributes.set_stride({1, 1});
    convAttributes.set_dilation({1, 1});

    const GraphAttributes graphAttributes;
    const ConvolutionWgradNode node(std::move(convAttributes), graphAttributes);

    auto error = node.pre_validate_node();
    EXPECT_EQ(error.code, error_code_t::INVALID_VALUE);
}

TEST(TestConvolutionWgradNode, PreValidateBatchSizeMismatch)
{
    ConvWgradAttributes convAttributes;

    auto xTensor = std::make_shared<TensorAttributes>();
    xTensor->set_dim({1, 3, 32, 32}); // Batch size 1
    xTensor->set_stride({3072, 1024, 32, 1});
    convAttributes.set_x(xTensor);

    auto dyTensor = std::make_shared<TensorAttributes>();
    dyTensor->set_dim({2, 64, 32, 32}); // Batch size 2
    dyTensor->set_stride({131072, 1024, 32, 1});
    convAttributes.set_dy(dyTensor);

    auto dwTensor = std::make_shared<TensorAttributes>();
    dwTensor->set_dim({64, 3, 3, 3});
    convAttributes.set_dw(dwTensor);

    convAttributes.set_pre_padding({1, 1});
    convAttributes.set_post_padding({1, 1});
    convAttributes.set_stride({1, 1});
    convAttributes.set_dilation({1, 1});

    const GraphAttributes graphAttributes;
    const ConvolutionWgradNode node(std::move(convAttributes), graphAttributes);

    auto error = node.pre_validate_node();
    EXPECT_EQ(error.code, error_code_t::INVALID_VALUE);
}

TEST(TestConvolutionWgradNode, PreValidateSpatialParamMismatch)
{
    ConvWgradAttributes convAttributes;

    auto xTensor = std::make_shared<TensorAttributes>();
    xTensor->set_dim({1, 3, 32, 32, 32}); // 3D spatial
    xTensor->set_stride({98304, 32768, 1024, 32, 1});
    convAttributes.set_x(xTensor);

    auto dyTensor = std::make_shared<TensorAttributes>();
    dyTensor->set_dim({1, 64, 32, 32, 32});
    dyTensor->set_stride({2097152, 32768, 1024, 32, 1});
    convAttributes.set_dy(dyTensor);

    auto dwTensor = std::make_shared<TensorAttributes>();
    dwTensor->set_dim({64, 3, 3, 3, 3});
    convAttributes.set_dw(dwTensor);

    // Only 2 spatial parameters for 3D spatial dimensions
    convAttributes.set_pre_padding({1, 1});
    convAttributes.set_post_padding({1, 1, 1});
    convAttributes.set_stride({1, 1, 1});
    convAttributes.set_dilation({1, 1, 1});

    const GraphAttributes graphAttributes;
    const ConvolutionWgradNode node(std::move(convAttributes), graphAttributes);

    auto error = node.pre_validate_node();
    EXPECT_EQ(error.code, error_code_t::INVALID_VALUE);
}

TEST(TestConvolutionWgradNode, PreValidateNegativeStride)
{
    ConvWgradAttributes convAttributes;

    auto xTensor = std::make_shared<TensorAttributes>();
    xTensor->set_dim({1, 3, 32, 32});
    xTensor->set_stride({3072, 1024, 32, 1});
    convAttributes.set_x(xTensor);

    auto dyTensor = std::make_shared<TensorAttributes>();
    dyTensor->set_dim({1, 64, 32, 32});
    dyTensor->set_stride({65536, 1024, 32, 1});
    convAttributes.set_dy(dyTensor);

    auto dwTensor = std::make_shared<TensorAttributes>();
    dwTensor->set_dim({64, 3, 3, 3});
    convAttributes.set_dw(dwTensor);

    convAttributes.set_pre_padding({1, 1});
    convAttributes.set_post_padding({1, 1});
    convAttributes.set_stride({1, -1}); // Negative stride
    convAttributes.set_dilation({1, 1});

    const GraphAttributes graphAttributes;
    const ConvolutionWgradNode node(std::move(convAttributes), graphAttributes);

    auto error = node.pre_validate_node();
    EXPECT_EQ(error.code, error_code_t::INVALID_VALUE);
}

TEST(TestConvolutionWgradNode, PreValidateZeroDilation)
{
    ConvWgradAttributes convAttributes;

    auto xTensor = std::make_shared<TensorAttributes>();
    xTensor->set_dim({1, 3, 32, 32});
    xTensor->set_stride({3072, 1024, 32, 1});
    convAttributes.set_x(xTensor);

    auto dyTensor = std::make_shared<TensorAttributes>();
    dyTensor->set_dim({1, 64, 32, 32});
    dyTensor->set_stride({65536, 1024, 32, 1});
    convAttributes.set_dy(dyTensor);

    auto dwTensor = std::make_shared<TensorAttributes>();
    dwTensor->set_dim({64, 3, 3, 3});
    convAttributes.set_dw(dwTensor);

    convAttributes.set_pre_padding({1, 1});
    convAttributes.set_post_padding({1, 1});
    convAttributes.set_stride({1, 1});
    convAttributes.set_dilation({1, 0}); // Zero dilation

    const GraphAttributes graphAttributes;
    const ConvolutionWgradNode node(std::move(convAttributes), graphAttributes);

    auto error = node.pre_validate_node();
    EXPECT_EQ(error.code, error_code_t::INVALID_VALUE);
}

TEST(TestConvolutionWgradNode, PreValidateNegativePrePadding)
{
    ConvWgradAttributes convAttributes;

    auto xTensor = std::make_shared<TensorAttributes>();
    xTensor->set_dim({1, 3, 32, 32});
    xTensor->set_stride({3072, 1024, 32, 1});
    convAttributes.set_x(xTensor);

    auto dyTensor = std::make_shared<TensorAttributes>();
    dyTensor->set_dim({1, 64, 32, 32});
    dyTensor->set_stride({65536, 1024, 32, 1});
    convAttributes.set_dy(dyTensor);

    auto dwTensor = std::make_shared<TensorAttributes>();
    dwTensor->set_dim({64, 3, 3, 3});
    convAttributes.set_dw(dwTensor);

    convAttributes.set_pre_padding({-1, 1}); // Negative padding
    convAttributes.set_post_padding({1, 1});
    convAttributes.set_stride({1, 1});
    convAttributes.set_dilation({1, 1});

    const GraphAttributes graphAttributes;
    const ConvolutionWgradNode node(std::move(convAttributes), graphAttributes);

    auto error = node.pre_validate_node();
    EXPECT_EQ(error.code, error_code_t::INVALID_VALUE);
}

TEST(TestConvolutionWgradNode, InferPropertiesWithLargeStride)
{
    ConvWgradAttributes convAttributes;

    auto xTensor = std::make_shared<TensorAttributes>();
    xTensor->set_dim({1, 3, 22, 22});
    xTensor->set_stride({1452, 484, 22, 1});
    convAttributes.set_x(xTensor);

    auto dyTensor = std::make_shared<TensorAttributes>();
    dyTensor->set_dim({1, 64, 8, 8});
    dyTensor->set_stride({4096, 64, 8, 1});
    convAttributes.set_dy(dyTensor);

    auto dwTensor = std::make_shared<TensorAttributes>();
    dwTensor->set_dim({64, 3, 5, 5});
    convAttributes.set_dw(dwTensor);

    convAttributes.set_pre_padding({2, 2});
    convAttributes.set_post_padding({2, 2});
    convAttributes.set_stride({3, 3}); // Large stride
    convAttributes.set_dilation({1, 1});

    const GraphAttributes graphAttributes;
    ConvolutionWgradNode node(std::move(convAttributes), graphAttributes);

    auto error = node.infer_properties_node();
    EXPECT_EQ(error.code, error_code_t::OK) << error.err_msg;

    auto dwDims = dwTensor->get_dim();
    EXPECT_EQ(dwDims.size(), 4);
    EXPECT_EQ(dwDims[0], 64); // Output channels
    EXPECT_EQ(dwDims[1], 3); // Input channels
    // Kernel height: ((22 + 2 + 2 - 3 * (8 - 1) - 1) / 1) + 1 = (22 + 4 - 21 - 1) / 1 + 1 = 4 / 1 + 1 = 5
    EXPECT_EQ(dwDims[2], 5);
    EXPECT_EQ(dwDims[3], 5);
}

TEST(TestConvolutionWgradNode, InferPropertiesZeroPadding)
{
    ConvWgradAttributes convAttributes;

    auto xTensor = std::make_shared<TensorAttributes>();
    xTensor->set_dim({1, 64, 32, 32});
    xTensor->set_stride({65536, 1024, 32, 1});
    convAttributes.set_x(xTensor);

    auto dyTensor = std::make_shared<TensorAttributes>();
    dyTensor->set_dim({1, 128, 28, 28});
    dyTensor->set_stride({100352, 784, 28, 1});
    convAttributes.set_dy(dyTensor);

    auto dwTensor = std::make_shared<TensorAttributes>();
    dwTensor->set_dim({128, 64, 5, 5});
    convAttributes.set_dw(dwTensor);

    convAttributes.set_pre_padding({0, 0}); // No padding
    convAttributes.set_post_padding({0, 0});
    convAttributes.set_stride({1, 1});
    convAttributes.set_dilation({1, 1});

    const GraphAttributes graphAttributes;
    ConvolutionWgradNode node(std::move(convAttributes), graphAttributes);

    auto error = node.infer_properties_node();
    EXPECT_EQ(error.code, error_code_t::OK) << error.err_msg;

    auto dwDims = dwTensor->get_dim();
    EXPECT_EQ(dwDims.size(), 4);
    EXPECT_EQ(dwDims[0], 128); // Output channels
    EXPECT_EQ(dwDims[1], 64); // Input channels
    // Kernel height: ((32 + 0 + 0 - 1 * (28 - 1) - 1) / 1) + 1 = (32 - 27 - 1) / 1 + 1 = 4 / 1 + 1 = 5
    EXPECT_EQ(dwDims[2], 5);
    EXPECT_EQ(dwDims[3], 5);
}

TEST(TestConvolutionWgradNode, InferPropertiesAsymmetricPadding)
{
    ConvWgradAttributes convAttributes;

    auto xTensor = std::make_shared<TensorAttributes>();
    xTensor->set_dim({2, 32, 16, 16});
    xTensor->set_stride({8192, 256, 16, 1});
    convAttributes.set_x(xTensor);

    auto dyTensor = std::make_shared<TensorAttributes>();
    dyTensor->set_dim({2, 64, 15, 15});
    dyTensor->set_stride({14400, 225, 15, 1});
    convAttributes.set_dy(dyTensor);

    auto dwTensor = std::make_shared<TensorAttributes>();
    dwTensor->set_dim({64, 32, 3, 3});
    convAttributes.set_dw(dwTensor);

    convAttributes.set_pre_padding({0, 1}); // Asymmetric padding
    convAttributes.set_post_padding({1, 0});
    convAttributes.set_stride({1, 1});
    convAttributes.set_dilation({1, 1});

    const GraphAttributes graphAttributes;
    ConvolutionWgradNode node(std::move(convAttributes), graphAttributes);

    auto error = node.infer_properties_node();
    EXPECT_EQ(error.code, error_code_t::OK) << error.err_msg;

    auto dwDims = dwTensor->get_dim();
    EXPECT_EQ(dwDims.size(), 4);
    EXPECT_EQ(dwDims[0], 64); // Output channels
    EXPECT_EQ(dwDims[1], 32); // Input channels
    // Kernel height: ((16 + 0 + 1 - 1 * (15 - 1) - 1) / 1) + 1 = (16 + 1 - 14 - 1) / 1 + 1 = 2 / 1 + 1 = 3
    EXPECT_EQ(dwDims[2], 3);
    // Kernel width: ((16 + 1 + 0 - 1 * (15 - 1) - 1) / 1) + 1 = (16 + 1 - 14 - 1) / 1 + 1 = 2 / 1 + 1 = 3
    EXPECT_EQ(dwDims[3], 3);
}

TEST(TestConvolutionWgradNode, PreValidateGroupedConvInvalidOutputChannels)
{
    ConvWgradAttributes convAttributes;

    auto xTensor = std::make_shared<TensorAttributes>();
    xTensor->set_dim({1, 64, 32, 32}); // 64 input channels
    xTensor->set_stride({65536, 1024, 32, 1});
    convAttributes.set_x(xTensor);

    auto dyTensor = std::make_shared<TensorAttributes>();
    dyTensor->set_dim({1, 63, 32, 32}); // 63 output channels
    dyTensor->set_stride({64512, 1024, 32, 1});
    convAttributes.set_dy(dyTensor);

    auto dwTensor = std::make_shared<TensorAttributes>();
    dwTensor->set_dim({63, 32, 3, 3}); // Groups = 64/32 = 2, but 63 is not divisible by 2
    dwTensor->set_stride({288, 9, 3, 1});
    convAttributes.set_dw(dwTensor);

    convAttributes.set_pre_padding({1, 1});
    convAttributes.set_post_padding({1, 1});
    convAttributes.set_stride({1, 1});
    convAttributes.set_dilation({1, 1});

    const GraphAttributes graphAttributes;
    const ConvolutionWgradNode node(std::move(convAttributes), graphAttributes);

    auto error = node.pre_validate_node();
    EXPECT_EQ(error.code, error_code_t::INVALID_VALUE);
}

TEST(TestConvolutionWgradNode, PreValidateGroupedConv2GroupsWithDwSet)
{
    ConvWgradAttributes convAttributes;

    auto xTensor = std::make_shared<TensorAttributes>();
    xTensor->set_dim({1, 64, 32, 32}); // 64 input channels
    xTensor->set_stride({65536, 1024, 32, 1});
    convAttributes.set_x(xTensor);

    auto dyTensor = std::make_shared<TensorAttributes>();
    dyTensor->set_dim({1, 64, 32, 32}); // 64 output channels
    dyTensor->set_stride({65536, 1024, 32, 1});
    convAttributes.set_dy(dyTensor);

    auto dwTensor = std::make_shared<TensorAttributes>();
    // 64 output channels, 32 input channels per group (2 groups)
    dwTensor->set_dim({64, 32, 3, 3});
    dwTensor->set_stride({288, 9, 3, 1});
    convAttributes.set_dw(dwTensor);

    convAttributes.set_pre_padding({1, 1});
    convAttributes.set_post_padding({1, 1});
    convAttributes.set_stride({1, 1});
    convAttributes.set_dilation({1, 1});

    const GraphAttributes graphAttributes;
    const ConvolutionWgradNode node(std::move(convAttributes), graphAttributes);

    auto error = node.pre_validate_node();
    EXPECT_EQ(error.code, error_code_t::OK) << error.err_msg;
}

TEST(TestConvolutionWgradNode, PreValidateGroupedConv4GroupsWithDwSet)
{
    ConvWgradAttributes convAttributes;

    auto xTensor = std::make_shared<TensorAttributes>();
    xTensor->set_dim({2, 64, 16, 16}); // 64 input channels
    xTensor->set_stride({16384, 256, 16, 1});
    convAttributes.set_x(xTensor);

    auto dyTensor = std::make_shared<TensorAttributes>();
    dyTensor->set_dim({2, 128, 16, 16}); // 128 output channels
    dyTensor->set_stride({32768, 256, 16, 1});
    convAttributes.set_dy(dyTensor);

    auto dwTensor = std::make_shared<TensorAttributes>();
    // 128 output channels, 16 input channels per group (4 groups)
    dwTensor->set_dim({128, 16, 3, 3});
    dwTensor->set_stride({144, 9, 3, 1});
    convAttributes.set_dw(dwTensor);

    convAttributes.set_pre_padding({1, 1});
    convAttributes.set_post_padding({1, 1});
    convAttributes.set_stride({1, 1});
    convAttributes.set_dilation({1, 1});

    const GraphAttributes graphAttributes;
    const ConvolutionWgradNode node(std::move(convAttributes), graphAttributes);

    auto error = node.pre_validate_node();
    EXPECT_EQ(error.code, error_code_t::OK) << error.err_msg;
}

TEST(TestConvolutionWgradNode, PreValidateGroupedConvInvalidInputChannels)
{
    ConvWgradAttributes convAttributes;

    auto xTensor = std::make_shared<TensorAttributes>();
    xTensor->set_dim({1, 63, 32, 32}); // 63 input channels
    xTensor->set_stride({64512, 1024, 32, 1});
    convAttributes.set_x(xTensor);

    auto dyTensor = std::make_shared<TensorAttributes>();
    dyTensor->set_dim({1, 64, 32, 32});
    dyTensor->set_stride({65536, 1024, 32, 1});
    convAttributes.set_dy(dyTensor);

    auto dwTensor = std::make_shared<TensorAttributes>();
    // 63 is not divisible by 32
    dwTensor->set_dim({64, 32, 3, 3});
    dwTensor->set_stride({288, 9, 3, 1});
    convAttributes.set_dw(dwTensor);

    convAttributes.set_pre_padding({1, 1});
    convAttributes.set_post_padding({1, 1});
    convAttributes.set_stride({1, 1});
    convAttributes.set_dilation({1, 1});

    const GraphAttributes graphAttributes;
    const ConvolutionWgradNode node(std::move(convAttributes), graphAttributes);

    auto error = node.pre_validate_node();
    EXPECT_EQ(error.code, error_code_t::INVALID_VALUE);
}

TEST(TestConvolutionWgradNode, InferGroupedConvStrideInferenceNchwLayout)
{
    ConvWgradAttributes convAttributes;

    auto xTensor = std::make_shared<TensorAttributes>();
    xTensor->set_dim({1, 64, 32, 32});
    xTensor->set_stride({65536, 1024, 32, 1}); // NCHW layout
    convAttributes.set_x(xTensor);

    auto dyTensor = std::make_shared<TensorAttributes>();
    dyTensor->set_dim({1, 128, 32, 32});
    dyTensor->set_stride({131072, 1024, 32, 1});
    convAttributes.set_dy(dyTensor);

    auto dwTensor = std::make_shared<TensorAttributes>();
    // Set dimensions to establish groups = 2
    dwTensor->set_dim({128, 32, 3, 3});
    // No stride set - should be inferred
    convAttributes.set_dw(dwTensor);

    convAttributes.set_pre_padding({1, 1});
    convAttributes.set_post_padding({1, 1});
    convAttributes.set_stride({1, 1});
    convAttributes.set_dilation({1, 1});

    const GraphAttributes graphAttributes;
    ConvolutionWgradNode node(std::move(convAttributes), graphAttributes);

    auto error = node.infer_properties_node();
    EXPECT_EQ(error.code, error_code_t::OK) << error.err_msg;

    auto inferredStrides = dwTensor->get_stride();
    EXPECT_EQ(inferredStrides.size(), 4);
    EXPECT_EQ(inferredStrides[0], 288); // N stride: 32 * 3 * 3
    EXPECT_EQ(inferredStrides[1], 9); // C stride: 3 * 3
    EXPECT_EQ(inferredStrides[2], 3); // H stride: 3
    EXPECT_EQ(inferredStrides[3], 1); // W stride: 1
}

TEST(TestConvolutionWgradNode, InferGroupedConvStrideInferenceNhwcLayout)
{
    ConvWgradAttributes convAttributes;

    auto xTensor = std::make_shared<TensorAttributes>();
    xTensor->set_dim({1, 64, 32, 32}); // Dims in NCHW order
    xTensor->set_stride({65536, 1, 2048, 64}); // NHWC strides
    convAttributes.set_x(xTensor);

    auto dyTensor = std::make_shared<TensorAttributes>();
    dyTensor->set_dim({1, 128, 32, 32}); // Dims in NCHW order
    dyTensor->set_stride({131072, 1, 4096, 128}); // NHWC
    convAttributes.set_dy(dyTensor);

    auto dwTensor = std::make_shared<TensorAttributes>();
    // Set dimensions to establish groups = 2
    dwTensor->set_dim({128, 32, 3, 3}); // Dims in KCRS order
    // No stride set - should be inferred
    convAttributes.set_dw(dwTensor);

    convAttributes.set_pre_padding({1, 1});
    convAttributes.set_post_padding({1, 1});
    convAttributes.set_stride({1, 1});
    convAttributes.set_dilation({1, 1});

    const GraphAttributes graphAttributes;
    ConvolutionWgradNode node(std::move(convAttributes), graphAttributes);

    auto error = node.infer_properties_node();
    EXPECT_EQ(error.code, error_code_t::OK) << error.err_msg;

    auto inferredStrides = dwTensor->get_stride();
    EXPECT_EQ(inferredStrides.size(), 4);
    // KRSC strides
    EXPECT_EQ(inferredStrides[0], 288); // K stride: 32 * 3 * 3
    EXPECT_EQ(inferredStrides[1], 1); // C stride: 1 (channels last)
    EXPECT_EQ(inferredStrides[2], 96); // R stride: 3 * 32
    EXPECT_EQ(inferredStrides[3], 32); // S stride: 32
}

TEST(TestConvolutionWgradNode, InferGroupedConvWithDilation)
{
    ConvWgradAttributes convAttributes;

    auto xTensor = std::make_shared<TensorAttributes>();
    xTensor->set_dim({1, 32, 20, 20});
    xTensor->set_stride({12800, 400, 20, 1});
    convAttributes.set_x(xTensor);

    auto dyTensor = std::make_shared<TensorAttributes>();
    dyTensor->set_dim({1, 64, 20, 20});
    dyTensor->set_stride({25600, 400, 20, 1});
    convAttributes.set_dy(dyTensor);

    auto dwTensor = std::make_shared<TensorAttributes>();
    // 64 output channels, 16 input channels per group (groups = 32/16 = 2)
    dwTensor->set_dim({64, 16, 3, 3});
    convAttributes.set_dw(dwTensor);

    convAttributes.set_pre_padding({2, 2});
    convAttributes.set_post_padding({2, 2});
    convAttributes.set_stride({1, 1});
    convAttributes.set_dilation({2, 2}); // 2x2 dilation

    const GraphAttributes graphAttributes;
    ConvolutionWgradNode node(std::move(convAttributes), graphAttributes);

    auto error = node.infer_properties_node();
    EXPECT_EQ(error.code, error_code_t::OK) << error.err_msg;

    auto inferredStrides = dwTensor->get_stride();
    EXPECT_EQ(inferredStrides.size(), 4);
    EXPECT_EQ(inferredStrides[0], 144); // K stride: 16 * 3 * 3
    EXPECT_EQ(inferredStrides[1], 9); // C stride: 3 * 3
    EXPECT_EQ(inferredStrides[2], 3); // R stride: 3
    EXPECT_EQ(inferredStrides[3], 1); // S stride: 1
}

TEST(TestConvolutionWgradNode, InferGroupedConvWithLargeStride)
{
    ConvWgradAttributes convAttributes;

    auto xTensor = std::make_shared<TensorAttributes>();
    xTensor->set_dim({2, 64, 22, 22});
    xTensor->set_stride({30976, 484, 22, 1});
    convAttributes.set_x(xTensor);

    auto dyTensor = std::make_shared<TensorAttributes>();
    dyTensor->set_dim({2, 128, 8, 8});
    dyTensor->set_stride({8192, 64, 8, 1});
    convAttributes.set_dy(dyTensor);

    auto dwTensor = std::make_shared<TensorAttributes>();
    // 128 output channels, 8 input channels per group (groups = 64/8 = 8)
    dwTensor->set_dim({128, 8, 5, 5});
    convAttributes.set_dw(dwTensor);

    convAttributes.set_pre_padding({2, 2});
    convAttributes.set_post_padding({2, 2});
    convAttributes.set_stride({3, 3}); // Large stride
    convAttributes.set_dilation({1, 1});

    const GraphAttributes graphAttributes;
    ConvolutionWgradNode node(std::move(convAttributes), graphAttributes);

    auto error = node.infer_properties_node();
    EXPECT_EQ(error.code, error_code_t::OK) << error.err_msg;

    auto inferredStrides = dwTensor->get_stride();
    EXPECT_EQ(inferredStrides.size(), 4);
    EXPECT_EQ(inferredStrides[0], 200); // N stride: 8 * 5 * 5
    EXPECT_EQ(inferredStrides[1], 25); // C stride: 5 * 5
    EXPECT_EQ(inferredStrides[2], 5); // H stride: 5
    EXPECT_EQ(inferredStrides[3], 1); // W stride: 1
}

TEST(TestConvolutionWgradNode, InferGroupedConv3D)
{
    ConvWgradAttributes convAttributes;

    auto xTensor = std::make_shared<TensorAttributes>();
    xTensor->set_dim({2, 32, 8, 16, 16});
    xTensor->set_stride({65536, 2048, 256, 16, 1});
    convAttributes.set_x(xTensor);

    auto dyTensor = std::make_shared<TensorAttributes>();
    dyTensor->set_dim({2, 64, 8, 16, 16});
    dyTensor->set_stride({131072, 2048, 256, 16, 1});
    convAttributes.set_dy(dyTensor);

    auto dwTensor = std::make_shared<TensorAttributes>();
    // 64 output channels, 8 input channels per group (groups = 32/8 = 4)
    dwTensor->set_dim({64, 8, 3, 3, 3});
    convAttributes.set_dw(dwTensor);

    convAttributes.set_pre_padding({1, 1, 1});
    convAttributes.set_post_padding({1, 1, 1});
    convAttributes.set_stride({1, 1, 1});
    convAttributes.set_dilation({1, 1, 1});

    const GraphAttributes graphAttributes;
    ConvolutionWgradNode node(std::move(convAttributes), graphAttributes);

    auto error = node.infer_properties_node();
    EXPECT_EQ(error.code, error_code_t::OK) << error.err_msg;

    auto inferredStrides = dwTensor->get_stride();
    EXPECT_EQ(inferredStrides.size(), 5);
    EXPECT_EQ(inferredStrides[0], 216); // N stride: 8 * 3 * 3 * 3
    EXPECT_EQ(inferredStrides[1], 27); // C stride: 3 * 3 * 3
    EXPECT_EQ(inferredStrides[2], 9); // D stride: 3 * 3
    EXPECT_EQ(inferredStrides[3], 3); // H stride: 3
    EXPECT_EQ(inferredStrides[4], 1); // W stride: 1
}

TEST(TestConvolutionWgradNode, InferGroupedConvDepthwiseSeparable)
{
    ConvWgradAttributes convAttributes;

    auto xTensor = std::make_shared<TensorAttributes>();
    xTensor->set_dim({1, 32, 112, 112}); // 32 input channels
    xTensor->set_stride({401408, 12544, 112, 1});
    convAttributes.set_x(xTensor);

    auto dyTensor = std::make_shared<TensorAttributes>();
    dyTensor->set_dim({1, 32, 112, 112}); // 32 output channels (same as input for depthwise)
    dyTensor->set_stride({401408, 12544, 112, 1});
    convAttributes.set_dy(dyTensor);

    auto dwTensor = std::make_shared<TensorAttributes>();
    // Depthwise: 32 output channels, 1 input channel per group (32 groups)
    dwTensor->set_dim({32, 1, 3, 3});
    convAttributes.set_dw(dwTensor);

    convAttributes.set_pre_padding({1, 1});
    convAttributes.set_post_padding({1, 1});
    convAttributes.set_stride({1, 1});
    convAttributes.set_dilation({1, 1});

    const GraphAttributes graphAttributes;
    ConvolutionWgradNode node(std::move(convAttributes), graphAttributes);

    auto error = node.infer_properties_node();
    EXPECT_EQ(error.code, error_code_t::OK) << error.err_msg;

    auto inferredStrides = dwTensor->get_stride();
    EXPECT_EQ(inferredStrides.size(), 4);
    EXPECT_EQ(inferredStrides[0], 9); // N stride: 1 * 3 * 3
    EXPECT_EQ(inferredStrides[1], 9); // C stride: 3 * 3
    EXPECT_EQ(inferredStrides[2], 3); // H stride: 3
    EXPECT_EQ(inferredStrides[3], 1); // W stride: 1
}

TEST(TestConvolutionWgradNode, PreValidateInvalidSpatialDimensions)
{
    ConvWgradAttributes convAttributes;

    auto xTensor = std::make_shared<TensorAttributes>();
    xTensor->set_dim({1, 3, 3, 3}); // Very small input
    xTensor->set_stride({27, 9, 3, 1});
    convAttributes.set_x(xTensor);

    auto dyTensor = std::make_shared<TensorAttributes>();
    dyTensor->set_dim({1, 64, 10, 10}); // Large output
    dyTensor->set_stride({6400, 100, 10, 1});
    convAttributes.set_dy(dyTensor);

    auto dwTensor = std::make_shared<TensorAttributes>();
    dwTensor->set_dim({64, 3, 3, 3});
    convAttributes.set_dw(dwTensor);

    convAttributes.set_pre_padding({0, 0});
    convAttributes.set_post_padding({0, 0});
    convAttributes.set_stride({1, 1});
    convAttributes.set_dilation({1, 1});

    const GraphAttributes graphAttributes;
    const ConvolutionWgradNode node(std::move(convAttributes), graphAttributes);

    auto error = node.pre_validate_node();
    // Will fail because: 3 + 0 + 0 - 1 * (10 - 1) - 1 = 3 - 9 - 1 = -7 (negative)
    EXPECT_EQ(error.code, error_code_t::INVALID_VALUE);
}

TEST(TestConvolutionWgradNode, PreValidateIncompatibleDilation)
{
    ConvWgradAttributes convAttributes;

    auto xTensor = std::make_shared<TensorAttributes>();
    xTensor->set_dim({1, 16, 20, 20});
    xTensor->set_stride({6400, 400, 20, 1});
    convAttributes.set_x(xTensor);

    auto dyTensor = std::make_shared<TensorAttributes>();
    dyTensor->set_dim({1, 32, 19, 19});
    dyTensor->set_stride({11552, 361, 19, 1});
    convAttributes.set_dy(dyTensor);

    auto dwTensor = std::make_shared<TensorAttributes>();
    dwTensor->set_dim({32, 16, 3, 3});
    convAttributes.set_dw(dwTensor);

    convAttributes.set_pre_padding({2, 2});
    convAttributes.set_post_padding({2, 2});
    convAttributes.set_stride({1, 1});
    convAttributes.set_dilation({2, 2}); // 2x2 dilation

    const GraphAttributes graphAttributes;
    const ConvolutionWgradNode node(std::move(convAttributes), graphAttributes);

    auto error = node.pre_validate_node();
    // numerator = 20 + 2 + 2 - 1 * (19 - 1) - 1 = 20 + 4 - 18 - 1 = 5
    // 5 % 2 = 1 (not divisible by dilation)
    EXPECT_EQ(error.code, error_code_t::INVALID_VALUE);
}

TEST(TestConvolutionWgradNode, InferProperties1x1Convolution)
{
    ConvWgradAttributes convAttributes;

    auto xTensor = std::make_shared<TensorAttributes>();
    xTensor->set_dim({1, 256, 14, 14});
    xTensor->set_stride({50176, 196, 14, 1});
    convAttributes.set_x(xTensor);

    auto dyTensor = std::make_shared<TensorAttributes>();
    dyTensor->set_dim({1, 512, 14, 14});
    dyTensor->set_stride({100352, 196, 14, 1});
    convAttributes.set_dy(dyTensor);

    auto dwTensor = std::make_shared<TensorAttributes>();
    dwTensor->set_dim({512, 256, 1, 1});
    convAttributes.set_dw(dwTensor);

    convAttributes.set_pre_padding({0, 0});
    convAttributes.set_post_padding({0, 0});
    convAttributes.set_stride({1, 1});
    convAttributes.set_dilation({1, 1});

    const GraphAttributes graphAttributes;
    ConvolutionWgradNode node(std::move(convAttributes), graphAttributes);

    auto error = node.infer_properties_node();
    EXPECT_EQ(error.code, error_code_t::OK) << error.err_msg;

    auto dwDims = dwTensor->get_dim();
    EXPECT_EQ(dwDims.size(), 4);
    EXPECT_EQ(dwDims[0], 512); // Output channels
    EXPECT_EQ(dwDims[1], 256); // Input channels
    // Kernel height: ((14 + 0 + 0 - 1 * (14 - 1) - 1) / 1) + 1 = (14 - 13 - 1) / 1 + 1 = 0 / 1 + 1 = 1
    EXPECT_EQ(dwDims[2], 1);
    EXPECT_EQ(dwDims[3], 1);
}

TEST(TestConvolutionWgradNode, InferPropertiesLargeKernel7x7)
{
    ConvWgradAttributes convAttributes;

    auto xTensor = std::make_shared<TensorAttributes>();
    xTensor->set_dim({1, 3, 224, 224});
    xTensor->set_stride({150528, 50176, 224, 1});
    convAttributes.set_x(xTensor);

    auto dyTensor = std::make_shared<TensorAttributes>();
    dyTensor->set_dim({1, 64, 112, 112});
    dyTensor->set_stride({802816, 12544, 112, 1});
    convAttributes.set_dy(dyTensor);

    auto dwTensor = std::make_shared<TensorAttributes>();
    dwTensor->set_dim({64, 3, 8, 8});
    convAttributes.set_dw(dwTensor);

    convAttributes.set_pre_padding({3, 3});
    convAttributes.set_post_padding({3, 3});
    convAttributes.set_stride({2, 2});
    convAttributes.set_dilation({1, 1});

    const GraphAttributes graphAttributes;
    ConvolutionWgradNode node(std::move(convAttributes), graphAttributes);

    auto error = node.infer_properties_node();
    EXPECT_EQ(error.code, error_code_t::OK) << error.err_msg;

    auto dwDims = dwTensor->get_dim();
    EXPECT_EQ(dwDims.size(), 4);
    EXPECT_EQ(dwDims[0], 64); // Output channels
    EXPECT_EQ(dwDims[1], 3); // Input channels
    EXPECT_EQ(dwDims[2], 8);
    EXPECT_EQ(dwDims[3], 8);
}

TEST(TestConvolutionWgradNode, InferPropertiesWithComplexDilation)
{
    ConvWgradAttributes convAttributes;

    auto xTensor = std::make_shared<TensorAttributes>();
    xTensor->set_dim({1, 32, 16, 16});
    xTensor->set_stride({8192, 256, 16, 1});
    convAttributes.set_x(xTensor);

    auto dyTensor = std::make_shared<TensorAttributes>();
    dyTensor->set_dim({1, 64, 14, 14});
    dyTensor->set_stride({12544, 196, 14, 1});
    convAttributes.set_dy(dyTensor);

    auto dwTensor = std::make_shared<TensorAttributes>();
    dwTensor->set_dim({64, 32, 3, 5});
    convAttributes.set_dw(dwTensor);

    convAttributes.set_pre_padding({1, 1});
    convAttributes.set_post_padding({1, 1});
    convAttributes.set_stride({1, 1});
    convAttributes.set_dilation({2, 1}); // Dilation only in one dimension

    const GraphAttributes graphAttributes;
    ConvolutionWgradNode node(std::move(convAttributes), graphAttributes);

    auto error = node.infer_properties_node();
    EXPECT_EQ(error.code, error_code_t::OK) << error.err_msg;

    auto dwDims = dwTensor->get_dim();
    EXPECT_EQ(dwDims.size(), 4);
    EXPECT_EQ(dwDims[0], 64); // Output channels
    EXPECT_EQ(dwDims[1], 32); // Input channels
    // Kernel height with dilation=2: ((16 + 1 + 1 - 1 * (14 - 1) - 1) / 2) + 1 = (18 - 13 - 1) / 2 + 1 = 4 / 2 + 1 = 3
    EXPECT_EQ(dwDims[2], 3);
    // Kernel width with dilation=1: ((16 + 1 + 1 - 1 * (14 - 1) - 1) / 1) + 1 = (18 - 13 - 1) / 1 + 1 = 4 / 1 + 1 = 5
    EXPECT_EQ(dwDims[3], 5);
}

TEST(TestConvolutionWgradNode, PreValidateOutputChannelMismatch)
{
    ConvWgradAttributes convAttributes;

    auto xTensor = std::make_shared<TensorAttributes>();
    xTensor->set_dim({1, 3, 32, 32});
    xTensor->set_stride({3072, 1024, 32, 1});
    convAttributes.set_x(xTensor);

    auto dyTensor = std::make_shared<TensorAttributes>();
    dyTensor->set_dim({1, 128, 32, 32}); // 128 output channels
    dyTensor->set_stride({131072, 1024, 32, 1});
    convAttributes.set_dy(dyTensor);

    auto dwTensor = std::make_shared<TensorAttributes>();
    dwTensor->set_dim({64, 3, 3, 3}); // 64 output channels, doesn't match dy
    dwTensor->set_stride({27, 9, 3, 1});
    convAttributes.set_dw(dwTensor);

    convAttributes.set_pre_padding({1, 1});
    convAttributes.set_post_padding({1, 1});
    convAttributes.set_stride({1, 1});
    convAttributes.set_dilation({1, 1});

    const GraphAttributes graphAttributes;
    const ConvolutionWgradNode node(std::move(convAttributes), graphAttributes);

    auto error = node.pre_validate_node();
    EXPECT_EQ(error.code, error_code_t::INVALID_VALUE);
}

TEST(TestConvolutionWgradNode, InferPropertiesLargeKernel5x5WithZeroPadding)
{
    ConvWgradAttributes convAttributes;

    auto xTensor = std::make_shared<TensorAttributes>();
    xTensor->set_dim({1, 64, 28, 28});
    xTensor->set_stride({50176, 784, 28, 1});
    convAttributes.set_x(xTensor);

    auto dyTensor = std::make_shared<TensorAttributes>();
    dyTensor->set_dim({1, 128, 24, 24});
    dyTensor->set_stride({73728, 576, 24, 1});
    convAttributes.set_dy(dyTensor);

    auto dwTensor = std::make_shared<TensorAttributes>();
    dwTensor->set_dim({128, 64, 5, 5});
    convAttributes.set_dw(dwTensor);

    convAttributes.set_pre_padding({0, 0});
    convAttributes.set_post_padding({0, 0});
    convAttributes.set_stride({1, 1});
    convAttributes.set_dilation({1, 1});

    const GraphAttributes graphAttributes;
    ConvolutionWgradNode node(std::move(convAttributes), graphAttributes);

    auto error = node.infer_properties_node();
    EXPECT_EQ(error.code, error_code_t::OK) << error.err_msg;

    auto dwDims = dwTensor->get_dim();
    EXPECT_EQ(dwDims.size(), 4);
    EXPECT_EQ(dwDims[0], 128); // Output channels
    EXPECT_EQ(dwDims[1], 64); // Input channels
    // Kernel height: ((28 + 0 + 0 - 1 * (24 - 1) - 1) / 1) + 1 = (28 - 23 - 1) / 1 + 1 = 4 / 1 + 1 = 5
    EXPECT_EQ(dwDims[2], 5);
    EXPECT_EQ(dwDims[3], 5);
}

TEST(TestConvolutionWgradNode, PreValidateDwDimsMismatch)
{
    ConvWgradAttributes convAttributes;

    auto xTensor = std::make_shared<TensorAttributes>();
    xTensor->set_dim({1, 3, 32, 32});
    xTensor->set_stride({3072, 1024, 32, 1});
    convAttributes.set_x(xTensor);

    auto dyTensor = std::make_shared<TensorAttributes>();
    dyTensor->set_dim({1, 64, 32, 32});
    dyTensor->set_stride({65536, 1024, 32, 1});
    convAttributes.set_dy(dyTensor);

    auto dwTensor = std::make_shared<TensorAttributes>();
    dwTensor->set_dim({64, 3, 3, 3, 3}); // 5D weight for 4D input
    dwTensor->set_stride({81, 27, 9, 3, 1});
    convAttributes.set_dw(dwTensor);

    convAttributes.set_pre_padding({1, 1});
    convAttributes.set_post_padding({1, 1});
    convAttributes.set_stride({1, 1});
    convAttributes.set_dilation({1, 1});

    const GraphAttributes graphAttributes;
    const ConvolutionWgradNode node(std::move(convAttributes), graphAttributes);

    auto error = node.pre_validate_node();
    EXPECT_EQ(error.code, error_code_t::INVALID_VALUE);
}

TEST(TestConvolutionWgradNode, PreValidateNodeNegativeStrideValues)
{
    ConvWgradAttributes convAttributes;

    auto xTensor = std::make_shared<TensorAttributes>();
    xTensor->set_dim({1, 3, 32, 32});
    xTensor->set_stride({3072, 1024, 32, 1});
    convAttributes.set_x(xTensor);

    auto dyTensor = std::make_shared<TensorAttributes>();
    dyTensor->set_dim({1, 64, 32, 32});
    dyTensor->set_stride({65536, 1024, 32, 1});
    convAttributes.set_dy(dyTensor);

    auto dwTensor = std::make_shared<TensorAttributes>();
    dwTensor->set_dim({64, 3, 3, 3});
    convAttributes.set_dw(dwTensor);

    convAttributes.set_pre_padding({1, 1});
    convAttributes.set_post_padding({1, 1});
    convAttributes.set_stride({-1, 1}); // Negative stride
    convAttributes.set_dilation({1, 1});

    const GraphAttributes graphAttributes;
    const ConvolutionWgradNode node(std::move(convAttributes), graphAttributes);

    auto error = node.pre_validate_node();
    EXPECT_EQ(error.code, error_code_t::INVALID_VALUE);
}

TEST(TestConvolutionWgradNode, PreValidateNodeNegativeDilationValues)
{
    ConvWgradAttributes convAttributes;

    auto xTensor = std::make_shared<TensorAttributes>();
    xTensor->set_dim({1, 3, 32, 32});
    xTensor->set_stride({3072, 1024, 32, 1});
    convAttributes.set_x(xTensor);

    auto dyTensor = std::make_shared<TensorAttributes>();
    dyTensor->set_dim({1, 64, 32, 32});
    dyTensor->set_stride({65536, 1024, 32, 1});
    convAttributes.set_dy(dyTensor);

    auto dwTensor = std::make_shared<TensorAttributes>();
    dwTensor->set_dim({64, 3, 3, 3});
    convAttributes.set_dw(dwTensor);

    convAttributes.set_pre_padding({1, 1});
    convAttributes.set_post_padding({1, 1});
    convAttributes.set_stride({1, 1});
    convAttributes.set_dilation({1, -1}); // Negative dilation

    const GraphAttributes graphAttributes;
    const ConvolutionWgradNode node(std::move(convAttributes), graphAttributes);

    auto error = node.pre_validate_node();
    EXPECT_EQ(error.code, error_code_t::INVALID_VALUE);
}

TEST(TestConvolutionWgradNode, PreValidateNodeNegativePrePaddingValues)
{
    ConvWgradAttributes convAttributes;

    auto xTensor = std::make_shared<TensorAttributes>();
    xTensor->set_dim({1, 3, 32, 32});
    xTensor->set_stride({3072, 1024, 32, 1});
    convAttributes.set_x(xTensor);

    auto dyTensor = std::make_shared<TensorAttributes>();
    dyTensor->set_dim({1, 64, 32, 32});
    dyTensor->set_stride({65536, 1024, 32, 1});
    convAttributes.set_dy(dyTensor);

    auto dwTensor = std::make_shared<TensorAttributes>();
    dwTensor->set_dim({64, 3, 3, 3});
    convAttributes.set_dw(dwTensor);

    convAttributes.set_pre_padding({-1, 1}); // Negative pre-padding
    convAttributes.set_post_padding({1, 1});
    convAttributes.set_stride({1, 1});
    convAttributes.set_dilation({1, 1});

    const GraphAttributes graphAttributes;
    const ConvolutionWgradNode node(std::move(convAttributes), graphAttributes);

    auto error = node.pre_validate_node();
    EXPECT_EQ(error.code, error_code_t::INVALID_VALUE);
}

TEST(TestConvolutionWgradNode, PreValidateNodeNegativePostPaddingValues)
{
    ConvWgradAttributes convAttributes;

    auto xTensor = std::make_shared<TensorAttributes>();
    xTensor->set_dim({1, 3, 32, 32});
    xTensor->set_stride({3072, 1024, 32, 1});
    convAttributes.set_x(xTensor);

    auto dyTensor = std::make_shared<TensorAttributes>();
    dyTensor->set_dim({1, 64, 32, 32});
    dyTensor->set_stride({65536, 1024, 32, 1});
    convAttributes.set_dy(dyTensor);

    auto dwTensor = std::make_shared<TensorAttributes>();
    dwTensor->set_dim({64, 3, 3, 3});
    convAttributes.set_dw(dwTensor);

    convAttributes.set_pre_padding({1, 1});
    convAttributes.set_post_padding({1, -2}); // Negative post-padding
    convAttributes.set_stride({1, 1});
    convAttributes.set_dilation({1, 1});

    const GraphAttributes graphAttributes;
    const ConvolutionWgradNode node(std::move(convAttributes), graphAttributes);

    auto error = node.pre_validate_node();
    EXPECT_EQ(error.code, error_code_t::INVALID_VALUE);
}

TEST(TestConvolutionWgradNode, PreValidateNegativePostPadding)
{
    ConvWgradAttributes convAttributes;

    auto xTensor = std::make_shared<TensorAttributes>();
    xTensor->set_dim({1, 3, 32, 32});
    xTensor->set_stride({3072, 1024, 32, 1});
    convAttributes.set_x(xTensor);

    auto dyTensor = std::make_shared<TensorAttributes>();
    dyTensor->set_dim({1, 64, 32, 32});
    dyTensor->set_stride({65536, 1024, 32, 1});
    convAttributes.set_dy(dyTensor);

    auto dwTensor = std::make_shared<TensorAttributes>();
    dwTensor->set_dim({64, 3, 3, 3});
    convAttributes.set_dw(dwTensor);

    convAttributes.set_pre_padding({1, 1});
    convAttributes.set_post_padding({1, -1}); // Negative post-padding
    convAttributes.set_stride({1, 1});
    convAttributes.set_dilation({1, 1});

    const GraphAttributes graphAttributes;
    const ConvolutionWgradNode node(std::move(convAttributes), graphAttributes);

    auto error = node.pre_validate_node();
    EXPECT_EQ(error.code, error_code_t::INVALID_VALUE);
}

TEST(TestConvolutionWgradNode, PreValidateSpatialDimensionMismatch)
{
    ConvWgradAttributes convAttributes;

    auto xTensor = std::make_shared<TensorAttributes>();
    xTensor->set_dim({1, 3, 32, 32});
    xTensor->set_stride({3072, 1024, 32, 1});
    convAttributes.set_x(xTensor);

    auto dyTensor = std::make_shared<TensorAttributes>();
    dyTensor->set_dim({1, 64, 28, 28}); // Incorrect spatial dimensions
    dyTensor->set_stride({50176, 784, 28, 1});
    convAttributes.set_dy(dyTensor);

    auto dwTensor = std::make_shared<TensorAttributes>();
    dwTensor->set_dim({64, 3, 3, 3}); // 3x3 kernel
    dwTensor->set_stride({27, 9, 3, 1});
    convAttributes.set_dw(dwTensor);

    convAttributes.set_pre_padding({1, 1});
    convAttributes.set_post_padding({1, 1});
    convAttributes.set_stride({1, 1});
    convAttributes.set_dilation({1, 1});

    const GraphAttributes graphAttributes;
    const ConvolutionWgradNode node(std::move(convAttributes), graphAttributes);

    auto error = node.pre_validate_node();
    // Expected dy size: ((32 + 1 + 1 - 3) / 1) + 1 = 32, but we have 28
    EXPECT_EQ(error.code, error_code_t::INVALID_VALUE);
}

TEST(TestConvolutionWgradNode, InferPropertiesMissingXStrides)
{
    ConvWgradAttributes convAttributes;

    auto xTensor = std::make_shared<TensorAttributes>();
    xTensor->set_dim({1, 3, 32, 32});
    // No stride set on x tensor
    convAttributes.set_x(xTensor);

    auto dyTensor = std::make_shared<TensorAttributes>();
    dyTensor->set_dim({1, 64, 32, 32});
    dyTensor->set_stride({65536, 1024, 32, 1});
    convAttributes.set_dy(dyTensor);

    auto dwTensor = std::make_shared<TensorAttributes>();
    dwTensor->set_dim({64, 3, 3, 3});
    // No stride set on dw tensor - will try to infer stride from x
    convAttributes.set_dw(dwTensor);

    convAttributes.set_pre_padding({1, 1});
    convAttributes.set_post_padding({1, 1});
    convAttributes.set_stride({1, 1});
    convAttributes.set_dilation({1, 1});

    const GraphAttributes graphAttributes;
    ConvolutionWgradNode node(std::move(convAttributes), graphAttributes);

    auto error = node.infer_properties_node();
    EXPECT_EQ(error.code, error_code_t::ATTRIBUTE_NOT_SET);
}

TEST(TestConvolutionWgradNode, InferPropertiesStrideDimensionMismatch)
{
    ConvWgradAttributes convAttributes;

    auto xTensor = std::make_shared<TensorAttributes>();
    xTensor->set_dim({1, 3, 32, 32});
    xTensor->set_stride({3072, 1024, 32}); // Only 3 strides for 4D tensor
    convAttributes.set_x(xTensor);

    auto dyTensor = std::make_shared<TensorAttributes>();
    dyTensor->set_dim({1, 64, 32, 32});
    dyTensor->set_stride({65536, 1024, 32, 1});
    convAttributes.set_dy(dyTensor);

    auto dwTensor = std::make_shared<TensorAttributes>();
    dwTensor->set_dim({64, 3, 3, 3});
    // No stride set on dw tensor - will try to infer stride from x
    convAttributes.set_dw(dwTensor);

    convAttributes.set_pre_padding({1, 1});
    convAttributes.set_post_padding({1, 1});
    convAttributes.set_stride({1, 1});
    convAttributes.set_dilation({1, 1});

    const GraphAttributes graphAttributes;
    ConvolutionWgradNode node(std::move(convAttributes), graphAttributes);

    auto error = node.infer_properties_node();
    EXPECT_EQ(error.code, error_code_t::ATTRIBUTE_NOT_SET);
}

TEST(TestConvolutionWgradNode, InferPropertiesWithDroppedPixelsAndDilation)
{
    ConvWgradAttributes convAttributes;

    auto xTensor = std::make_shared<TensorAttributes>();
    xTensor->set_dim({1, 32, 16, 16});
    xTensor->set_stride({8192, 256, 16, 1});
    convAttributes.set_x(xTensor);

    auto dyTensor = std::make_shared<TensorAttributes>();
    dyTensor->set_dim({1, 4, 5, 5});
    dyTensor->set_stride({100, 25, 5, 1});
    convAttributes.set_dy(dyTensor);

    auto dwTensor = std::make_shared<TensorAttributes>();
    dwTensor->set_dim({4, 32, 3, 3});
    convAttributes.set_dw(dwTensor);

    convAttributes.set_pre_padding({0, 0});
    convAttributes.set_post_padding({0, 0});
    convAttributes.set_stride({2, 2});
    convAttributes.set_dilation({3, 3});

    const GraphAttributes graphAttributes;
    ConvolutionWgradNode node(std::move(convAttributes), graphAttributes);

    auto error = node.infer_properties_node();
    EXPECT_EQ(error.code, error_code_t::OK) << error.err_msg;

    auto dwDims = dwTensor->get_dim();
    EXPECT_EQ(dwDims.size(), 4);
    EXPECT_EQ(dwDims[0], 4); // Output channels
    EXPECT_EQ(dwDims[1], 32); // Input channels
    EXPECT_EQ(dwDims[2], 3);
    EXPECT_EQ(dwDims[3], 3);
}

TEST(TestConvolutionWgradNode, PreValidateIncompatibleStrideDilationCombination)
{
    ConvWgradAttributes convAttributes;

    auto xTensor = std::make_shared<TensorAttributes>();
    // xSize = 5
    xTensor->set_dim({1, 1, 5, 5});
    xTensor->set_stride({25, 25, 5, 1});
    convAttributes.set_x(xTensor);

    auto dyTensor = std::make_shared<TensorAttributes>();
    // dySize = 2
    dyTensor->set_dim({1, 1, 2, 2});
    dyTensor->set_stride({4, 4, 2, 1});
    convAttributes.set_dy(dyTensor);

    auto dwTensor = std::make_shared<TensorAttributes>();
    dwTensor->set_dim({1, 1, 1, 1});
    convAttributes.set_dw(dwTensor);

    convAttributes.set_pre_padding({0, 0});
    convAttributes.set_post_padding({0, 0});
    convAttributes.set_stride({2, 2});
    convAttributes.set_dilation({3, 3});

    // numerator = 5 + 0 + 0 - (2 * (2 - 1)) - 1 = 2
    // remainder = 2 % 3 = 2
    // remainder (2) >= stride (2) -> Error because dropped pixels cannot exceed stride

    const GraphAttributes graphAttributes;
    const ConvolutionWgradNode node(std::move(convAttributes), graphAttributes);

    auto error = node.pre_validate_node();
    EXPECT_EQ(error.code, error_code_t::INVALID_VALUE);
}

TEST(TestConvolutionWgradNode, PreValidateNodeInvalidSpatialDimensions)
{
    ConvWgradAttributes convAttributes;

    auto xTensor = std::make_shared<TensorAttributes>();
    xTensor->set_dim({1, 3, 32, 32});
    xTensor->set_stride({3072, 1024, 32, 1});
    convAttributes.set_x(xTensor);

    auto dyTensor = std::make_shared<TensorAttributes>();
    dyTensor->set_dim({1, 64, 30, 30}); // Incorrect dy size (should be 32x32)
    dyTensor->set_stride({57600, 900, 30, 1});
    convAttributes.set_dy(dyTensor);

    auto dwTensor = std::make_shared<TensorAttributes>();
    dwTensor->set_dim({64, 3, 3, 3});
    dwTensor->set_stride({27, 9, 3, 1});
    convAttributes.set_dw(dwTensor);

    convAttributes.set_pre_padding({1, 1});
    convAttributes.set_post_padding({1, 1});
    convAttributes.set_stride({1, 1});
    convAttributes.set_dilation({1, 1});

    const GraphAttributes graphAttributes;
    const ConvolutionWgradNode node(std::move(convAttributes), graphAttributes);

    auto error = node.pre_validate_node();
    EXPECT_EQ(error.code, error_code_t::INVALID_VALUE);
}

TEST(TestConvolutionWgradNode, PreValidateNodeWithDroppedPixels)
{
    ConvWgradAttributes convAttributes;

    // For each spatial dim Index (0, 1, 2) & (2, 3, 4) are the two 3x3 kernels
    // applied with stride 2 on 6x6 dx 5th is dropped / unused due to stride
    auto xTensor = std::make_shared<TensorAttributes>();
    xTensor->set_dim({1, 3, 6, 6}); // Input 6x6
    xTensor->set_stride({108, 36, 6, 1});
    convAttributes.set_x(xTensor);

    auto dyTensor = std::make_shared<TensorAttributes>();
    dyTensor->set_dim({1, 64, 2, 2}); // Output 2x2
    dyTensor->set_stride({256, 4, 2, 1});
    convAttributes.set_dy(dyTensor);

    auto dwTensor = std::make_shared<TensorAttributes>();
    dwTensor->set_dim({64, 3, 3, 3}); // Kernel 3x3
    dwTensor->set_stride({27, 9, 3, 1});
    convAttributes.set_dw(dwTensor);

    convAttributes.set_pre_padding({0, 0});
    convAttributes.set_post_padding({0, 0});
    convAttributes.set_stride({2, 2});
    convAttributes.set_dilation({1, 1});

    const GraphAttributes graphAttributes;
    const ConvolutionWgradNode node(std::move(convAttributes), graphAttributes);

    auto error = node.pre_validate_node();
    EXPECT_EQ(error.code, error_code_t::OK) << error.err_msg;
}

TEST(TestConvolutionWgradNode, PreValidateNodeInputTooSmall)
{
    ConvWgradAttributes convAttributes;

    auto xTensor = std::make_shared<TensorAttributes>();
    xTensor->set_dim({1, 3, 2, 2}); // Input smaller than kernel
    xTensor->set_stride({12, 4, 2, 1});
    convAttributes.set_x(xTensor);

    auto dyTensor = std::make_shared<TensorAttributes>();
    dyTensor->set_dim({1, 64, 1, 1});
    dyTensor->set_stride({64, 1, 1, 1});
    convAttributes.set_dy(dyTensor);

    auto dwTensor = std::make_shared<TensorAttributes>();
    dwTensor->set_dim({64, 3, 3, 3}); // Kernel 3x3
    dwTensor->set_stride({27, 9, 3, 1});
    convAttributes.set_dw(dwTensor);

    convAttributes.set_pre_padding({0, 0});
    convAttributes.set_post_padding({0, 0});
    convAttributes.set_stride({1, 1});
    convAttributes.set_dilation({1, 1});

    const GraphAttributes graphAttributes;
    const ConvolutionWgradNode node(std::move(convAttributes), graphAttributes);

    auto error = node.pre_validate_node();
    EXPECT_EQ(error.code, error_code_t::INVALID_VALUE);
}

// =============================================================================
// create_operation() Tests
// =============================================================================

class TestConvolutionWgradNodeCreateOperation : public ::testing::Test
{
protected:
    std::shared_ptr<Mock_hipdnn_backend> _mockBackend;

    void SetUp() override
    {
        _mockBackend = std::make_shared<Mock_hipdnn_backend>();
        detail::IHipdnnBackend::setInstance(_mockBackend);
    }
    void TearDown() override
    {
        detail::IHipdnnBackend::resetInstance();
        _mockBackend.reset();
    }

    static ConvolutionWgradNode makeConvWgradNode()
    {
        ConvWgradAttributes attrs;
        auto x = std::make_shared<TensorAttributes>();
        x->set_uid(K_FPROP_TENSOR_X_UID)
            .set_name("X")
            .set_data_type(DataType::FLOAT)
            .set_dim(toVec(K_FPROP_TENSOR_X_DIMS))
            .set_stride(toVec(K_FPROP_TENSOR_X_STRIDES));
        auto dy = std::make_shared<TensorAttributes>();
        dy->set_uid(K_FPROP_TENSOR_Y_UID)
            .set_name("DY")
            .set_data_type(DataType::FLOAT)
            .set_dim(toVec(K_FPROP_TENSOR_Y_DIMS))
            .set_stride(toVec(K_FPROP_TENSOR_Y_STRIDES));
        auto dw = std::make_shared<TensorAttributes>();
        dw->set_uid(K_FPROP_TENSOR_W_UID)
            .set_name("DW")
            .set_data_type(DataType::FLOAT)
            .set_dim(toVec(K_FPROP_TENSOR_W_DIMS))
            .set_stride(toVec(K_FPROP_TENSOR_W_STRIDES));

        attrs.set_x(x);
        attrs.set_dy(dy);
        attrs.set_dw(dw);
        attrs.set_pre_padding(toVec(K_FPROP_CONV_PADDING));
        attrs.set_post_padding(toVec(K_FPROP_CONV_PADDING));
        attrs.set_stride(toVec(K_FPROP_CONV_STRIDE));
        attrs.set_dilation(toVec(K_FPROP_CONV_DILATION));
        attrs.compute_data_type = DataType::FLOAT;

        const GraphAttributes graphAttrs;
        return {std::move(attrs), graphAttrs};
    }
};

TEST_F(TestConvolutionWgradNodeCreateOperation, PropagatesBackendError)
{
    EXPECT_CALL(*_mockBackend, backendCreateDescriptor(HIPDNN_BACKEND_TENSOR_DESCRIPTOR, _))
        .WillRepeatedly(Return(HIPDNN_STATUS_SUCCESS));
    EXPECT_CALL(
        *_mockBackend,
        backendCreateDescriptor(HIPDNN_BACKEND_OPERATION_CONVOLUTION_BACKWARD_FILTER_DESCRIPTOR, _))
        .WillOnce(Return(HIPDNN_STATUS_INTERNAL_ERROR));
    EXPECT_CALL(*_mockBackend, backendDestroyDescriptor(_))
        .WillRepeatedly(Return(HIPDNN_STATUS_SUCCESS));
    EXPECT_CALL(*_mockBackend, backendSetAttribute(_, _, _, _, _))
        .WillRepeatedly(Return(HIPDNN_STATUS_SUCCESS));
    EXPECT_CALL(*_mockBackend, backendFinalize(_)).WillRepeatedly(Return(HIPDNN_STATUS_SUCCESS));
    EXPECT_CALL(*_mockBackend, getLastErrorString(_, _)).Times(AnyNumber());

    auto node = makeConvWgradNode();
    std::unordered_map<int64_t, detail::ScopedHipdnnBackendDescriptor> tensorDescs;
    std::vector<detail::ScopedHipdnnBackendDescriptor> operations;

    auto err = node.create_operation(tensorDescs, operations);
    EXPECT_TRUE(err.is_bad());
    EXPECT_EQ(err.code, ErrorCode::HIPDNN_BACKEND_ERROR);
}

TEST_F(TestConvolutionWgradNodeCreateOperation, SuccessCreatesThreeTensorsAndOneOperation)
{
    EXPECT_CALL(*_mockBackend, backendCreateDescriptor(HIPDNN_BACKEND_TENSOR_DESCRIPTOR, _))
        .Times(3)
        .WillRepeatedly(Return(HIPDNN_STATUS_SUCCESS));
    EXPECT_CALL(
        *_mockBackend,
        backendCreateDescriptor(HIPDNN_BACKEND_OPERATION_CONVOLUTION_BACKWARD_FILTER_DESCRIPTOR, _))
        .WillOnce(Return(HIPDNN_STATUS_SUCCESS));
    EXPECT_CALL(*_mockBackend, backendDestroyDescriptor(_))
        .WillRepeatedly(Return(HIPDNN_STATUS_SUCCESS));
    EXPECT_CALL(*_mockBackend, backendSetAttribute(_, _, _, _, _))
        .WillRepeatedly(Return(HIPDNN_STATUS_SUCCESS));

    EXPECT_CALL(*_mockBackend, backendFinalize(_))
        .Times(4)
        .WillRepeatedly(Return(HIPDNN_STATUS_SUCCESS));

    auto node = makeConvWgradNode();
    std::unordered_map<int64_t, detail::ScopedHipdnnBackendDescriptor> tensorDescs;
    std::vector<detail::ScopedHipdnnBackendDescriptor> operations;

    auto err = node.create_operation(tensorDescs, operations);
    EXPECT_TRUE(err.is_good()) << err.err_msg;
    EXPECT_EQ(tensorDescs.size(), 3u);
    EXPECT_EQ(operations.size(), 1u);
}

TEST(TestConvolutionWgradNode, GetNodeTypeReturnsConvolutionWgrad)
{
    const GraphAttributes graphAttrs;
    const ConvolutionWgradNode node(ConvWgradAttributes{}, graphAttrs);
    EXPECT_EQ(node.getNodeType(), NodeType::CONVOLUTION_WGRAD);
}