TheRock/rocm-libraries/projects/miopen/test/gtest/find_2_conv.cpp

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

#include <gtest/gtest.h>
#include "get_handle.hpp"
#include "gtest_common.hpp"
#include "../lib_env_var.hpp"
#include "../workspace.hpp"

#include <miopen/miopen.h>
#include <miopen/convolution.hpp>
#include <miopen/solution.hpp>
#include <miopen/solver_id.hpp>
#include <nlohmann/json.hpp>

#include <vector>
#include <limits>
#include <optional>
#include <algorithm>

namespace {

struct Find2ConvTestCase
{
    miopenProblemDirection_t direction;
    int tune;
    bool preallocate;
    std::size_t workspace_limit;
    bool attach_binaries;

    friend std::ostream& operator<<(std::ostream& os, const Find2ConvTestCase& tc)
    {
        return os << "direction:" << tc.direction << " tune:" << tc.tune
                  << " preallocate:" << tc.preallocate << " ws_limit:" << tc.workspace_limit
                  << " attach_binaries:" << tc.attach_binaries;
    }
};

void RunFind2ConvTest(const Find2ConvTestCase& test_case)
{
    auto& handle_deref = get_handle();

    // Generate tensors
    tensor<float> x{16, 192, 28, 28};
    x.generate(tensor_elem_gen_integer{17});
    tensor<float> w{32, 192, 5, 5};
    w.generate(tensor_elem_gen_integer{17});

    miopen::ConvolutionDescriptor filter = {
        2, miopenConvolution, miopenPaddingDefault, {1, 1}, {1, 1}, {1, 1}};
    tensor<float> y{filter.GetForwardOutputTensor(x.desc, w.desc)};

    auto x_dev = handle_deref.Write(x.data);
    auto w_dev = handle_deref.Write(w.data);
    auto y_dev = handle_deref.Write(y.data);

    // Test convolution
    miopenHandle_t handle = &handle_deref;
    miopenProblem_t problem;

    EXPECT_EQ(miopenCreateConvProblem(&problem, &filter, test_case.direction), miopenStatusSuccess);

    // Add tensor descriptors
    EXPECT_EQ(miopenSetProblemTensorDescriptor(problem, miopenTensorConvolutionX, &x.desc),
              miopenStatusSuccess);
    EXPECT_EQ(miopenSetProblemTensorDescriptor(problem, miopenTensorConvolutionW, &w.desc),
              miopenStatusSuccess);
    EXPECT_EQ(miopenSetProblemTensorDescriptor(problem, miopenTensorConvolutionY, &y.desc),
              miopenStatusSuccess);

    // Adding x descriptor again to validate that error is produced
    EXPECT_EQ(miopenSetProblemTensorDescriptor(problem, miopenTensorConvolutionX, &x.desc),
              miopenStatusBadParm);

    // Test FindSolutions
    auto solutions    = std::vector<miopenSolution_t>{};
    std::size_t found = 0;
    solutions.resize(100);

    EXPECT_EQ(
        miopenFindSolutions(handle, problem, nullptr, solutions.data(), &found, solutions.size()),
        miopenStatusSuccess);
    EXPECT_GE(found, 0);
    solutions.resize(found);

    // Test FindSolutions with options
    solutions.resize(100);
    found = 0;

    {
        miopenFindOptions_t options;
        EXPECT_EQ(miopenCreateFindOptions(&options), miopenStatusSuccess);

        EXPECT_EQ(miopenSetFindOptionTuning(options, test_case.tune), miopenStatusSuccess);
        EXPECT_EQ(miopenSetFindOptionResultsOrder(options, miopenFindResultsOrderByTime),
                  miopenStatusSuccess);
        EXPECT_EQ(miopenSetFindOptionWorkspaceLimit(options, test_case.workspace_limit),
                  miopenStatusSuccess);
        EXPECT_EQ(miopenSetFindOptionAttachBinaries(options, test_case.attach_binaries),
                  miopenStatusSuccess);

        Workspace wspace{}; // This GPU buffer may be used by miopenFindSolutions
        if(test_case.preallocate)
        {
            std::size_t workspace_max = [&]() {
                std::size_t workspace_max_ = 0;
                switch(test_case.direction)
                {
                case miopenProblemDirectionForward:
                    EXPECT_EQ(miopenConvolutionForwardGetWorkSpaceSize(
                                  handle, &x.desc, &w.desc, &filter, &y.desc, &workspace_max_),
                              miopenStatusSuccess);
                    return workspace_max_;
                case miopenProblemDirectionBackward:
                    EXPECT_EQ(miopenConvolutionBackwardDataGetWorkSpaceSize(
                                  handle, &y.desc, &w.desc, &filter, &x.desc, &workspace_max_),
                              miopenStatusSuccess);
                    return workspace_max_;
                case miopenProblemDirectionBackwardWeights:
                    EXPECT_EQ(miopenConvolutionBackwardWeightsGetWorkSpaceSize(
                                  handle, &y.desc, &x.desc, &filter, &w.desc, &workspace_max_),
                              miopenStatusSuccess);
                    return workspace_max_;

#ifdef MIOPEN_BETA_API
                case miopenProblemDirectionInference: break;
#endif
                }

                MIOPEN_THROW(miopenStatusNotImplemented);
            }();

            const auto workspace_size = std::min(test_case.workspace_limit, workspace_max);
            wspace.resize(workspace_size);

            EXPECT_EQ(
                miopenSetFindOptionPreallocatedWorkspace(options, wspace.ptr(), wspace.size()),
                miopenStatusSuccess);

            EXPECT_EQ(miopenSetFindOptionPreallocatedTensor(
                          options, miopenTensorConvolutionX, x_dev.get()),
                      miopenStatusSuccess);

            EXPECT_EQ(miopenSetFindOptionPreallocatedTensor(
                          options, miopenTensorConvolutionW, w_dev.get()),
                      miopenStatusSuccess);

            EXPECT_EQ(miopenSetFindOptionPreallocatedTensor(
                          options, miopenTensorConvolutionY, y_dev.get()),
                      miopenStatusSuccess);
        }

        EXPECT_EQ(miopenFindSolutions(
                      handle, problem, options, solutions.data(), &found, solutions.size()),
                  miopenStatusSuccess);

        EXPECT_EQ(miopenDestroyFindOptions(options), miopenStatusSuccess);
    }

    EXPECT_GE(found, 0);
    solutions.resize(found);

    // Test solution attributes
    for(const auto& solution : solutions)
    {
        float time;
        std::size_t workspace_size;
        uint64_t solver_id;
        miopenConvAlgorithm_t algo;

        EXPECT_EQ(miopenGetSolutionTime(solution, &time), miopenStatusSuccess);
        EXPECT_EQ(miopenGetSolutionWorkspaceSize(solution, &workspace_size), miopenStatusSuccess);
        EXPECT_EQ(miopenGetSolutionSolverId(solution, &solver_id), miopenStatusSuccess);
        EXPECT_EQ(miopenGetSolverIdConvAlgorithm(solver_id, &algo), miopenStatusSuccess);
    }

    // Test running solutions
    miopenTensorDescriptor_t x_desc = &x.desc, w_desc = &w.desc, y_desc = &y.desc;

    for(const auto& solution : solutions)
    {
        uint64_t solver_id;
        EXPECT_EQ(miopenGetSolutionSolverId(solution, &solver_id), miopenStatusSuccess);

        miopenTensorArgumentId_t names[3] = {
            miopenTensorConvolutionX, miopenTensorConvolutionW, miopenTensorConvolutionY};
        void* buffers[3]                        = {x_dev.get(), w_dev.get(), y_dev.get()};
        miopenTensorDescriptor_t descriptors[3] = {x_desc, w_desc, y_desc};

        // Run solution
        std::size_t workspace_size;
        EXPECT_EQ(miopenGetSolutionWorkspaceSize(solution, &workspace_size), miopenStatusSuccess);

        Workspace wspace{workspace_size};

        auto arguments = std::make_unique<miopenTensorArgument_t[]>(3);
        for(auto i = 0; i < 3; ++i)
        {
            arguments[i].id         = names[i];
            arguments[i].descriptor = &descriptors[i];
            arguments[i].buffer     = buffers[i];
        }

        EXPECT_EQ(
            miopenRunSolution(handle, solution, 3, arguments.get(), wspace.ptr(), wspace.size()),
            miopenStatusSuccess);

        // Save-load cycle
        std::size_t solution_size;
        EXPECT_EQ(miopenGetSolutionSize(solution, &solution_size), miopenStatusSuccess);
        EXPECT_GT(solution_size, 0);

        auto solution_binary = std::vector<char>{};
        solution_binary.resize(solution_size);

        EXPECT_EQ(miopenSaveSolution(solution, solution_binary.data()), miopenStatusSuccess);
        EXPECT_EQ(miopenDestroySolution(solution), miopenStatusSuccess);

        miopenSolution_t read_solution;
        EXPECT_EQ(
            miopenLoadSolution(&read_solution, solution_binary.data(), solution_binary.size()),
            miopenStatusSuccess);

        // Run loaded solution
        auto read_arguments = std::make_unique<miopenTensorArgument_t[]>(3);
        for(auto i = 0; i < 3; ++i)
        {
            read_arguments[i].id         = names[i];
            read_arguments[i].descriptor = &descriptors[i];
            read_arguments[i].buffer     = buffers[i];
        }

        EXPECT_EQ(miopenRunSolution(
                      handle, read_solution, 3, read_arguments.get(), wspace.ptr(), wspace.size()),
                  miopenStatusSuccess);
        EXPECT_EQ(miopenDestroySolution(read_solution), miopenStatusSuccess);
    }

    EXPECT_EQ(miopenDestroyProblem(problem), miopenStatusSuccess);
}

} // namespace

class GPU_Find2Conv_FP32 : public testing::TestWithParam<
                               std::tuple<miopenProblemDirection_t, int, bool, std::size_t, bool>>
{
protected:
    void SetUp() override
    {
        prng::reset_seed();

        // Parity with CTest: Set log level to 6
        if(MIOPEN_LOG_LEVEL)
        {
            prev_log_level = lib_env::value<int>(MIOPEN_LOG_LEVEL);
        }
        lib_env::update(MIOPEN_LOG_LEVEL, 6);
    }

    void TearDown() override
    {
        // Reset internal environment values as per Wiki guidelines
        if(prev_log_level.has_value())
        {
            lib_env::update(MIOPEN_LOG_LEVEL, *prev_log_level);
        }
        else
        {
            lib_env::clear(MIOPEN_LOG_LEVEL);
        }
    }

private:
    std::optional<int> prev_log_level;
};

// TEST_INFO is "Find2ConvTest" for descriptiveness.
// Specificity is provided by Parameter Name Generator in the test output.
TEST_P(GPU_Find2Conv_FP32, Find2ConvTest)
{
    auto param = GetParam();
    Find2ConvTestCase tc{std::get<0>(param),
                         std::get<1>(param),
                         std::get<2>(param),
                         std::get<3>(param),
                         std::get<4>(param)};
    RunFind2ConvTest(tc);
}

inline std::string
GetFind2ConvTestCaseName(const testing::TestParamInfo<GPU_Find2Conv_FP32::ParamType>& info)
{
    Find2ConvTestCase tc{std::get<0>(info.param),
                         std::get<1>(info.param),
                         std::get<2>(info.param),
                         std::get<3>(info.param),
                         std::get<4>(info.param)};
    std::ostringstream os;
    os << tc;
    std::string name = os.str();
    std::replace_if(name.begin(), name.end(), [](char c) { return !std::isalnum(c); }, '_');
    return name;
}

INSTANTIATE_TEST_SUITE_P(Smoke,
                         GPU_Find2Conv_FP32,
                         testing::Combine(testing::Values(miopenProblemDirectionForward,
                                                          miopenProblemDirectionBackward,
                                                          miopenProblemDirectionBackwardWeights),
                                          testing::Values(0, 1),
                                          testing::Values(false, true),
                                          testing::Values(std::numeric_limits<std::size_t>::max(),
                                                          static_cast<std::size_t>(0)),
                                          testing::Values(false, true)),
                         GetFind2ConvTestCaseName);