MNN/test/expr/GatherTest.cpp

//
//  GatherTest.cpp
//  MNNTests
//
//  Created by MNN on 2019/09/17.
//  Copyright © 2018, Alibaba Group Holding Limited
//

/*
 Test Case From https://www.tensorflow.org/api_docs/cc/class/tensorflow/ops/gather-nd
 */
#include <math.h>
#include <MNN/expr/ExprCreator.hpp>
#include <MNN/expr/Module.hpp>
#include "MNNTestSuite.h"
#include "MNN_generated.h"
#include "TestUtils.h"
using namespace MNN;
using namespace MNN::Express;

class GatherExprTest : public MNNTestCase {
public:
    virtual bool run(int precision) {
        auto executor = cloneCurrentExecutor();
        ExecutorScope scope(executor);

        std::unique_ptr<MNN::OpT> gatherOp(new MNN::OpT);
        gatherOp->type = MNN::OpType_GatherND;
        auto parameter = _Input({2, 2}, NHWC, halide_type_of<int32_t>());
        parameter->setName("param");
        auto indice    = _Input({2, 2}, NHWC, halide_type_of<int32_t>());
        indice->setName("indice");
        auto y         = Variable::create(Expr::create(gatherOp.get(), {parameter, indice}));
        y->setName("y");
        {
            parameter->resize({2, 2});
            auto ptr = parameter->writeMap<float>();
            ptr[0]   = 7.0;
            ptr[1]   = 2.0;
            ptr[2]   = 4.0;
            ptr[3]   = 6.0;
        }
        {
            auto indicePtr = indice->writeMap<int32_t>();
            indicePtr[0]   = 0;
            indicePtr[1]   = 0;
            indicePtr[2]   = 1;
            indicePtr[3]   = 1;
            auto size      = y->getInfo()->size;
            if (size != 2) {
                return false;
            }
            auto yPtr = y->readMap<float>();
            if (fabs(yPtr[0] - 7.0) > 0.001 || fabs(yPtr[1] - 6.0) > 0.001) {
                return false;
            }
        }
        {
            indice->resize({2, 1});
            auto indicePtr = indice->writeMap<int32_t>();
            indicePtr[0]   = 1;
            indicePtr[1]   = 0;
            auto size      = y->getInfo()->size;
            if (4 != size) {
                return false;
            }
            auto yPtr = y->readMap<float>();
            if (fabs(yPtr[0] - 4.0) > 0.001 || fabs(yPtr[1] - 6.0) > 0.001 || fabs(yPtr[2] - 7.0) > 0.001 ||
                fabs(yPtr[3] - 2.0) > 0.001) {
                return false;
            }
        }
        {
            indice->resize({1, 1});
            auto indicePtr = indice->writeMap<int32_t>();
            indicePtr[0]   = 1;
            parameter->resize({2, 2, 2});
            auto parameterPtr = parameter->writeMap<float>();
            for (int i = 0; i < parameter->getInfo()->size; ++i) {
                parameterPtr[i] = 1.0 * i;
            }
            auto size = y->getInfo()->size;
            if (4 != size) {
                return false;
            }
            auto yPtr = y->readMap<float>();
            for (int i = 0; i < size; ++i) {
                if (fabs(yPtr[i] - 4.0 - i) > 0.001) {
                    return false;
                }
            }
        }
        // Run as Module
        flatbuffers::FlatBufferBuilder builderOutput(1024);
        {
            std::unique_ptr<MNN::NetT> net(new NetT);
            Variable::save({y}, net.get());
            y = nullptr;
            auto len = MNN::Net::Pack(builderOutput, net.get());
            builderOutput.Finish(len);
        }
        int sizeOutput    = builderOutput.GetSize();
        auto bufferOutput = builderOutput.GetBufferPointer();
        std::shared_ptr<MNN::Express::Module> module(Module::load(std::vector<std::string>{"param", "indice"}, std::vector<std::string>{"y"}, bufferOutput, sizeOutput));
        {
            {
                parameter->resize({2, 2});
                auto ptr = parameter->writeMap<float>();
                ptr[0]   = 7.0;
                ptr[1]   = 2.0;
                ptr[2]   = 4.0;
                ptr[3]   = 6.0;
            }
            {
                indice->resize({2, 2});
                auto indicePtr = indice->writeMap<int32_t>();
                indicePtr[0]   = 0;
                indicePtr[1]   = 0;
                indicePtr[2]   = 1;
                indicePtr[3]   = 1;
                auto y2 = module->onForward({parameter, indice})[0];
                auto size      = y2->getInfo()->size;
                if (size != 2) {
                    return false;
                }
                auto yPtr = y2->readMap<float>();
                if (fabs(yPtr[0] - 7.0) > 0.001 || fabs(yPtr[1] - 6.0) > 0.001) {
                    return false;
                }
            }
            {
                indice->resize({2, 1});
                auto indicePtr = indice->writeMap<int32_t>();
                indicePtr[0]   = 1;
                indicePtr[1]   = 0;
                auto y2 = module->onForward({parameter, indice})[0];
                auto size      = y2->getInfo()->size;
                if (4 != size) {
                    return false;
                }
                auto yPtr = y2->readMap<float>();
                if (fabs(yPtr[0] - 4.0) > 0.001 || fabs(yPtr[1] - 6.0) > 0.001 || fabs(yPtr[2] - 7.0) > 0.001 ||
                    fabs(yPtr[3] - 2.0) > 0.001) {
                    return false;
                }
            }
            {
                indice->resize({1, 1});
                auto indicePtr = indice->writeMap<int32_t>();
                indicePtr[0]   = 1;
                parameter->resize({2, 2, 2});
                auto parameterPtr = parameter->writeMap<float>();
                for (int i = 0; i < parameter->getInfo()->size; ++i) {
                    parameterPtr[i] = 1.0 * i;
                }
                auto y2 = module->onForward({parameter, indice})[0];
                auto yPtr = y2->readMap<float>();
                auto size = y2->getInfo()->size;
                if (4 != size) {
                    return false;
                }
                for (int i = 0; i < size; ++i) {
                    if (fabs(yPtr[i] - 4.0 - i) > 0.001) {
                        return false;
                    }
                }
            }
            {
                const float inpudata[]                  = {-1.0, -2.0, 3.0, 4.0};
                const int indices_data[]                = {0, 0, 1, 1};
                auto params                             = _Const(inpudata, {2, 2}, NHWC, halide_type_of<float>());
                auto indices                            = _Const(indices_data, {2, 2}, NHWC, halide_type_of<int>());
                auto x1 = _GatherND(params, indices);
                x1->setName("input1");
                auto shape = x1->getInfo()->dim;
                auto x0 = _Input(shape, NHWC, halide_type_of<float>());
                float x0data[] = {1.0f, 2.0f};
                ::memcpy(x0->writeMap<float>(), x0data, 2 * sizeof(float));
                x0->setName("input0");
                auto res = _Add(x0, x1);
                res->setName("GatherNd_output_0");
                
                flatbuffers::FlatBufferBuilder builderOutput(1024);
                {
                    std::unique_ptr<MNN::NetT> net(new NetT);
                    Variable::save({res}, net.get());
                    y = nullptr;
                    auto len = MNN::Net::Pack(builderOutput, net.get());
                    builderOutput.Finish(len);
                }
                int sizeOutput    = builderOutput.GetSize();
                auto bufferOutput = builderOutput.GetBufferPointer();
                const char* cacheFileName = ".tempcache";
                MNN::ScheduleConfig config;
                config.numThread = 1;
                
                BackendConfig bnConfig;
                bnConfig.precision = (MNN::BackendConfig::PrecisionMode)precision;
                config.backendConfig = &bnConfig;
                std::shared_ptr<Executor::RuntimeManager> rtmgr(Executor::RuntimeManager::createRuntimeManager(config));
                rtmgr->setCache(cacheFileName);
                MNN::Express::Module::Config mConfig;
                /*
                 ScheduleConfig config;
                 BackendConfig bnConfig;
                 bnConfig.precision = (MNN::BackendConfig::PrecisionMode)precision;
                 config.numThread = 1;
                 config.type = ExecutorScope::Current()->getAttr()->firstType.first;
                 config.backendConfig = &bnConfig;
                 */
                
                std::shared_ptr<MNN::Express::Module> module_2(Module::load(std::vector<std::string>{"input0"}, std::vector<std::string>{"GatherNd_output_0"}, bufferOutput, sizeOutput, rtmgr, &mConfig));
                
                auto y2 = module_2->onForward({x0})[0];
                
                const float inpudata1[]                  = {-5.0, -6.0, 7.0, 8.0};
                x0->resize({2, 2});
                auto parameterPtr = params->writeMap<float>();
                ::memcpy(parameterPtr, inpudata1, 4*sizeof(float));
                
                y2 = module_2->onForward({x0})[0];
            }
        }
        return true;
    }
};

class GatherNdReComputeTest : public MNNTestCase {
public:
    virtual bool run(int precision) {
        auto executor = cloneCurrentExecutor();
        ExecutorScope scope(executor);
        const float inpudata[]                  = {-1.0, -2.0, 3.0, 4.0};
        const int indices_data[]                = {0, 0, 1, 1};
        auto params                             = _Const(inpudata, {2, 2}, NHWC, halide_type_of<float>());
        auto indices                            = _Const(indices_data, {2, 2}, NHWC, halide_type_of<int>());
        auto x1 = _GatherND(params, indices);
        x1->setName("input1");
        auto shape = x1->getInfo()->dim;
        auto x0 = _Input(shape, NHWC, halide_type_of<float>());
        x0->setName("input0");
        auto res = _Add(x0, x1);
        res->setName("GatherNd_output_0");
        
        flatbuffers::FlatBufferBuilder builderOutput(1024);
        {
            std::unique_ptr<MNN::NetT> net(new NetT);
            Variable::save({res}, net.get());
            auto len = MNN::Net::Pack(builderOutput, net.get());
            builderOutput.Finish(len);
        }
        int sizeOutput    = builderOutput.GetSize();
        auto bufferOutput = builderOutput.GetBufferPointer();
        
        std::shared_ptr<MNN::Express::Module> module(Module::load(std::vector<std::string>{"input0"}, std::vector<std::string>{"GatherNd_output_0"}, bufferOutput, sizeOutput));
        
        // first run, call GatherNd compute function when resize.
        float data0[] = {1.0f, 2.0f};
        ::memcpy(x0->writeMap<float>(), data0, 2 * sizeof(float));
        auto y = module->onForward({x0});

        // resize input and test GatherNd recompute function when risize.
        const float data1[]                  = {-5.0, -6.0, 7.0, 8.0};
        x0->resize({2, 2});
        ::memcpy(x0->writeMap<float>(), data1, 4 * sizeof(float));
        y = module->onForward({x0});
        return true;
    }
    
};
MNNTestSuiteRegister(GatherExprTest, "expr/Gather");
MNNTestSuiteRegister(GatherNdReComputeTest, "expr/GatherNdRecomputeTest");