TheRock/rocm-libraries/projects/hipblaslt/tensilelite/include/Tensile/GranularitySelectionLibrary.hpp

/*******************************************************************************
 *
 * MIT License
 *
 * Copyright (C) 2022-2023 Advanced Micro Devices, Inc. All rights reserved.
 *
 * Permission is hereby granted, free of charge, to any person obtaining a copy
 * of this software and associated documentation files (the "Software"), to deal
 * in the Software without restriction, including without limitation the rights
 * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
 * copies of the Software, and to permit persons to whom the Software is
 * furnished to do so, subject to the following conditions:
 *
 * The above copyright notice and this permission notice shall be included in
 * all copies or substantial portions of the Software.
 *
 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
 * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
 * SOFTWARE.
 *
 *******************************************************************************/

#pragma once

#include <set>
#include <vector>

#include <Tensile/Debug.hpp>
#include <Tensile/PredicateDebugger.hpp>
#include <Tensile/Properties.hpp>
#include <Tensile/Utils.hpp>

#include <Tensile/PropertyMatching.hpp>

#include <tensilelitehost/export.h>

namespace TensileLite
{
    /**
 * \ingroup SolutionLibrary
 */
    struct ExactSelectionTableEntry
    {
        std::vector<size_t> key;
        int                 value;
    };

    /**
 * \ingroup SolutionLibrary
 *
 * Compares the tile sizes of each kernel, the dimensions of the problem,
 * and the number of compute units on the target GPU to select a kernel
 * that fits the best on the GPU with the lowest amount of waste
 * ("granularity loss").
 */
    template <typename MyProblem, typename MySolution = typename MyProblem::Solution>
    struct GranularitySelectionLibrary : public SolutionLibrary<MyProblem, MySolution>
    {
        std::map<int, std::shared_ptr<MySolution>> solutions;
        std::map<std::vector<size_t>, int>         exactMap;

        static std::string Type()
        {
            return "GranularitySelection";
        }
        virtual std::string type() const override
        {
            return Type();
        }
        virtual std::string description() const override
        {
            std::string rv = this->type();

            return rv;
        }

        virtual std::shared_ptr<MySolution> getSolutionByIndex(MyProblem const& problem,
                                                               Hardware const&  hardware,
                                                               const int index) const override
        {
            if(index >= solutions.size())
            {
                return std::shared_ptr<MySolution>();
            }
            return solutions.at(index);
        }

        virtual std::shared_ptr<MySolution> findBestSolution(MyProblem const& problem,
                                                             Hardware const&  hardware,
                                                             double*          fitness
                                                             = nullptr) const override
        {
            const bool debug = Debug::Instance().printPropertyEvaluation();

            std::vector<size_t> key;
            size_t              M = problem.freeSizeA(0);
            key.push_back(M);
            size_t N = problem.freeSizeB(0);
            key.push_back(N);
            size_t NumBatches = problem.batchSize(0);
            key.push_back(NumBatches);
            size_t K = problem.boundSize(0);
            key.push_back(K);

            auto exactMatch = exactMap.find(key);
            if(exactMatch != this->exactMap.end())
            {
                int index = exactMatch->second;

                auto rv = solutions.at(index);

                Task task(hardware, problem, *rv);
                bool predicateMatch = (*rv->problemPredicate)(problem)
                                      && (*rv->taskPredicate)(task)
                                      && (*rv->hardwarePredicate)(hardware);
                if(debug)
                {
                    PredicateDebugger::printHeader(
                        std::cout, "Granularity (Exact): " + rv->name());
                    rv->hardwarePredicate->debugEval(hardware, std::cout);
                    rv->problemPredicate->debugEval(problem, std::cout);
                    rv->taskPredicate->debugEval(task, std::cout);
                    PredicateDebugger::printFooter(std::cout, predicateMatch);
                }

                if(predicateMatch)
                {
                    return rv;
                }
            }

            double                      bestPerformance = 0.0;
            std::shared_ptr<MySolution> bestSolution;

            for(auto const& row : solutions)
            {
                auto myPerformance
                    = row.second->projectedPerformance(problem, hardware).speedGFlops;

                if(debug)
                {
                    std::cout << row.second->description() << ": " << myPerformance;
                }

                if(myPerformance > bestPerformance)
                {
                    Task task(hardware, problem, *(row.second));
                    bool predicateMatch = (*row.second->problemPredicate)(problem)
                                          && (*row.second->taskPredicate)(task)
                                          && (*row.second->hardwarePredicate)(hardware);

                    if(debug)
                    {
                        PredicateDebugger::printHeader(
                            std::cout, "Granularity: " + row.second->name());
                        row.second->hardwarePredicate->debugEval(hardware, std::cout);
                        row.second->problemPredicate->debugEval(problem, std::cout);
                        row.second->taskPredicate->debugEval(task, std::cout);
                        PredicateDebugger::printFooter(std::cout, predicateMatch);
                    }

                    if(predicateMatch)
                    {
                        bestPerformance = myPerformance;
                        bestSolution    = row.second;
                    }
                }
            }

            return bestSolution;
        }

        virtual SolutionSet<MySolution>
            findAllSolutions(MyProblem const&          problem,
                             Hardware const&           hardware,
                             SolutionLibrarySearchType searchType
                             = SolutionLibrarySearchType::DEFAULT) const override
        {
            bool debug = Debug::Instance().printPropertyEvaluation();

            SolutionSet<MySolution> rv;

            for(auto const& row : solutions)
            {
                Task task(hardware, problem, *(row.second));
                bool predicateMatch = softwarePredicate(searchType, task, hardware, *(row.second), problem)
                                      && (*row.second->hardwarePredicate)(hardware);

                if(debug)
                {
                    PredicateDebugger::printHeader(std::cout, "Granularity: " + row.second->name());
                    row.second->hardwarePredicate->debugEval(hardware, std::cout);
                    if(searchType == SolutionLibrarySearchType::DEFAULT)
                    {
                        row.second->problemPredicate->debugEval(problem, std::cout);
                        row.second->taskPredicate->debugEval(task, std::cout);
                    }
                    PredicateDebugger::printFooter(std::cout, predicateMatch);
                }

                if(predicateMatch)
                {
                    rv.insert(row.second);
                }
            }

            return rv;
        }

        virtual SolutionSet<MySolution>
            findAllSolutionsGroupedGemm(std::vector<MyProblem> const& problems,
                                        Hardware const&               hardware,
                                        SolutionLibrarySearchType     searchType
                                        = SolutionLibrarySearchType::DEFAULT) const override
        {
            bool debug = Debug::Instance().printPropertyEvaluation();

            SolutionSet<MySolution> rv;

            for(auto const& row : solutions)
            {
                bool useSolution = false;

                if((*row.second->hardwarePredicate)(hardware))
                {
                    useSolution = true;
                    if(searchType == SolutionLibrarySearchType::DEFAULT)
                    {
                        size_t ws = (*row.second).requiredWorkspaceSizeGroupedGemm(problems, hardware);

                        for(size_t idx = 0; idx < problems.size(); idx++)
                        {
                            auto problem = problems[idx];
                            Task task(hardware, problem, *(row.second));
                            problem.setWorkspaceSizeGroupedGemm(ws);
                            problem.setGroupedGemmCount(problems.size());
                            if(!(*row.second->problemPredicate)(problem)
                               || !(*row.second->taskPredicate)(task))
                                useSolution = false;
                        }
                    }
                    else if(searchType == SolutionLibrarySearchType::GEMM_TYPE_ONLY)
                    {
                        if(!isGemmTypeSame((*row.second), problems[0]))
                            useSolution = false;
                    }

                    if(useSolution)
                    {
                        if((*row.second).requiredHostWorkspaceSizePerProblem
                           == static_cast<size_t>(-1))
                        {
                            (*row.second).requiredHostWorkspaceSizePerProblem
                                = (*row.second)
                                      .requiredHostSizeGroupedGemmSingle(problems[0], hardware);
                        }
                        rv.insert(row.second);
                    }
                }

                if(debug)
                {
                    PredicateDebugger::printHeader(
                        std::cout, "Granularity (Grouped): " + row.second->name());
                    row.second->hardwarePredicate->debugEval(hardware, std::cout);
                    if(searchType == SolutionLibrarySearchType::DEFAULT)
                    {
                        for(size_t idx = 0; idx < problems.size(); idx++)
                        {
                            auto problem = problems[idx];
                            Task task(hardware, problem, *(row.second));
                            row.second->problemPredicate->debugEval(problem, std::cout);
                            row.second->taskPredicate->debugEval(task, std::cout);
                        }
                    }
                    PredicateDebugger::printFooter(std::cout, useSolution);
                }
            }

            return rv;
        }
    };
} // namespace TensileLite