Method CPUBackend

source/backend/cpu/CPUBackend.cpp:450–517

    return new EagerBufferAllocator(BufferAllocator::Allocator::createRecurse(mStaticAllocator.get()));
}
CPUBackend::CPUBackend(const CPURuntime* runtime, BackendConfig::PrecisionMode precision, BackendConfig::MemoryMode memory, MNNForwardType type, size_t flags) : Backend(type) {
#ifdef LOG_VERBOSE
    MNN_PRINT("cpu backend create\n");
#endif
    mMemory = memory;
    mRuntime = const_cast<CPURuntime*>(runtime);
    auto core = MNNGetCoreFunctions();
    mThreadNumber = mRuntime->mThreadNumber;
    mRelatedFunctions = &core->int8MatmulRelatedFunctions;

    // Compute Group Rate
    do {
        if (mThreadNumber <= 1 || mRuntime->mPower == BackendConfig::Power_Low) {
            break;
        }
        auto rate = mRuntime->hint().cpuDecreaseRate;
        if (rate >= 100 || rate <= 0) {
            break;
        }
        auto cpuInfo = MNNGetCPUInfo();
        if (cpuInfo->groups.size() < 2) {
            break;
        }
        if (cpuInfo->i8mm) {
            mComputeI = 28.f;
        } else if (cpuInfo->dot) {
            mComputeI = 14.f;
        } else {
            mComputeI = 7.f;
        }
        mGroupWithComputeRate.clear();
        float decreaseRate = (float)(rate) / 100.0f;
        int validCpuSize = (int)(cpuInfo->groups[cpuInfo->groups.size()-1].ids.size());
        int groupIndex = (int)cpuInfo->groups.size()-2;
        validCpuSize = ALIMIN(validCpuSize, mThreadNumber);
        float totalComputeRate = 1.0f * validCpuSize;
        mGroupWithComputeRate.emplace_back(std::make_pair(totalComputeRate, validCpuSize));
        float currentRate = 1.0f;
        while (validCpuSize < mThreadNumber && groupIndex >= 0) {
            auto& group = cpuInfo->groups[groupIndex];
            int selectSize = ALIMIN(mThreadNumber - validCpuSize, (int)group.ids.size());
            validCpuSize += group.ids.size();
            currentRate *= decreaseRate;
            totalComputeRate += currentRate * selectSize;
            mGroupWithComputeRate.emplace_back(std::make_pair(currentRate * selectSize, selectSize));
            groupIndex--;
        }
        for (auto& g : mGroupWithComputeRate) {
            g.first = g.first / totalComputeRate;
        }
    } while (false);
    auto dynamicAlloc = mRuntime->mSharedDmaInfo;
    if (nullptr == dynamicAlloc.get()) {
        mDmaInfo.reset(new CPURuntime::DynamicAllocator);
        mDmaInfo->mDynamicAllocator.reset(mRuntime->createDynamicBufferAlloctor(0));
        mDmaInfo->mCurrentDynamicAllocator = mDmaInfo->mDynamicAllocator.get();
        mDmaInfo->mCacheGroup.resize(MNN_CPU_MAX_BUFFER_INDEX);
        for (int i=0; i<mDmaInfo->mCacheGroup.size(); ++i) {