MNN/source/backend/cpu/CPUBackend.hpp

//
//  CPUBackend.hpp
//  MNN
//
//  Created by MNN on 2018/07/06.
//  Copyright © 2018, Alibaba Group Holding Limited
//

#ifndef CPUBackend_hpp
#define CPUBackend_hpp

#include <map>
#include <memory>
#include "core/Backend.hpp"
#include "core/Execution.hpp"
#include "core/BufferAllocator.hpp"
#include "MNN_generated.h"

namespace MNN {
class CPURuntime : public Runtime {
public:
    friend class CPUBackend;
    CPURuntime(const Backend::Info& info);
    virtual ~ CPURuntime();
    int onGetRuntimeStatus(RuntimeStatus statusEnum) const override;
    virtual Backend* onCreate(const BackendConfig* config) const override;
    virtual void onGabageCollect(int level) override;
    virtual float onGetMemoryInMB() override;
    virtual CompilerType onGetCompilerType() const override {
        return Compiler_Loop;
    }
    void onConcurrencyBegin() const;
    void onConcurrencyEnd() const;
    virtual bool onCheckInfo(Backend::Info& info) const override;


private:
    std::shared_ptr<EagerBufferAllocator> mStaticAllocator;
    int mThreadNumber;
    mutable int mTaskIndex;
    BackendConfig::MemoryMode mMemory;
    BackendConfig::PowerMode mPower;
    BackendConfig::PrecisionMode mPrecision;

    // Backend features
    // CPU features
    float mFlops = 0.0f;
    static Backend*(*gExtraCreate)(const Runtime* runtime);
    size_t mFlags = 0;
    int mAllocator = 0;
};
struct CoreFunctions;
struct CoreInt8Functions;

class CPUResizeCache;
class CPUMemObj : public Backend::MemObj {
public:
    CPUMemObj(BufferAllocator* allocator, MemChunk chunk, int size) : mAllocator(allocator), mChunk(chunk), mSize(size) {}
    virtual ~ CPUMemObj() {
        if (mAllocator) {
            mAllocator->free(mChunk);
        }
    }
    virtual MemChunk chunk() {
        return mChunk;
    }
    inline int getSize() const {
        return mSize;
    }
private:
    BufferAllocator* mAllocator;
    MemChunk mChunk;
    int mSize;
};
class CPUBackend : public Backend {
public:
    CPUBackend(const CPURuntime* runtime, BackendConfig::PrecisionMode precision, BackendConfig::MemoryMode memory, MNNForwardType type = MNN_FORWARD_CPU, size_t flags = 0);
    virtual ~CPUBackend();

    // Return sizeDivide, scheduleNumber aligned memory
    std::pair<int, int> multiThreadDivide(int size) const;
public:
    virtual MemObj* onAcquire(const Tensor* nativeTensor, StorageType storageType) override;
    virtual bool onClearBuffer() override;
    virtual void onCopyBuffer(const Tensor* srcTensor, const Tensor* dstTensor) const override;

    virtual Execution* onCreate(const std::vector<Tensor*>& inputs, const std::vector<Tensor*>& outputs,
                                const MNN::Op* op) override;

    virtual void onExecuteBegin() const override;
    virtual void onExecuteEnd() const override;
    
    virtual void onResizeBegin() override;
    virtual void onResizeEnd() override;

    const CoreFunctions* functions() const {
        return mCoreFunctions;
    }
    // Return element size for Tensor, conside pack
    int getTensorSize(const Tensor* tensor, bool multiBytes = false) const;
    const CoreInt8Functions* int8Functions() const {
        return mInt8CoreFunctions;
    }
public:
    class Creator {
    public:
        virtual Execution* onCreate(const std::vector<Tensor*>& inputs, const std::vector<Tensor*>& outputs,
                                    const MNN::Op* op, Backend* backend) const = 0;
    };

    static bool addCreator(OpType t, Creator* c);

    int threadNumber() const {
        return mRuntime->mThreadNumber;
    }

    BufferAllocator* getBufferAllocator(bool defer_allocator = true) const {
        return mDynamicAllocator.get();
    }

    BackendConfig::MemoryMode memoryMode() const {
        return mMemory;
    }
    BackendConfig::PrecisionMode precisionMode() const {
        return mPrecisionMode;
    }
    CPUResizeCache* getCache() const {
        return mCache;
    }

    virtual const Runtime* getRuntime() override;

#ifdef MNN_USE_THREAD_POOL
    inline int taskIndex() const {return mRuntime->mTaskIndex;}
#endif
    static void initCreatorMap();
    static int getBytes(const Backend* backend, const Tensor* output);
    static DataType getDataType(const Tensor* tensor);


protected:
    MemObj* allocBuffer(int size, Tensor* dest,  StorageType storageType);
    const CoreFunctions* mCoreFunctions;
    const CoreInt8Functions* mInt8CoreFunctions;
private:
    std::shared_ptr<EagerBufferAllocator> mStaticAllocator;
    std::shared_ptr<BufferAllocator> mDynamicAllocator;
    CPURuntime* mRuntime;
    BackendConfig::PrecisionMode mPrecisionMode;
    BackendConfig::MemoryMode mMemory;
    static std::map<OpType, CPUBackend::Creator*>* gCreator;
    CPUResizeCache* mCache;
};
/** execution cast wrapper. insert tensor cast dynamic. */
class CastWrapExecution : public Execution {
public:
    CastWrapExecution(Backend* backend, DataType runT)
                    : Execution(backend), mRunType(runT) {}
    virtual ErrorCode onExecute(const std::vector<Tensor*>& inputs, const std::vector<Tensor*>& outputs) override;
private:
    DataType mRunType;
};


#define REGISTER_CPU_OP_CREATOR(name, opType)     \
    void ___##name##__##opType##__() {            \
        static name _temp;\
        CPUBackend::addCreator(opType, &_temp); \
    }

#ifdef MNN_SUPPORT_DEPRECATED_OP
#define REGISTER_CPU_OP_CREATOR_OLD(name, opType)     \
    void ___##name##__##opType##__() {            \
        static name _temp;\
        CPUBackend::addCreator(opType, &_temp); \
    }

#else
#define REGISTER_CPU_OP_CREATOR_OLD(name, opType)     \
    void ___##name##__##opType##__() {            \
    }
#endif

} // namespace MNN

#endif /* CPUBackend_hpp */