Function codegen

codegen/OpFuse.cpp:307–415

}

bool codegen(std::vector<Schedule::OpCacheInfo>& infos, std::vector<std::vector<Node*>>& fuseSets, MNNForwardType type, BackendConfig::PrecisionMode precision) {
    // generate Kernel
    std::unique_ptr<Target> target;
    switch (type) {
#ifdef MNN_CODEGEN_OPENCL
        case MNN_FORWARD_OPENCL:
            target.reset(new OpenCLTarget(precision));
            break;
#endif
#ifdef MNN_CODEGEN_METAL
        case MNN_FORWARD_METAL:
            target.reset(new MetalTarget(precision));
            break;
#endif
#ifdef MNN_CODEGEN_CUDA
        case MNN_FORWARD_CUDA:
            target.reset(new CUDATarget(precision));
            break;
#endif
        default:
            return false;
    }
#if 0
    if (fuseSets.size() > 0) {
        MNN_PRINT(">>>>>>>>>>>>> fuseSets.size = %lu\n", fuseSets.size());
    }
#endif
    std::map<std::string, int> mapKernelSources;
    for (int i = 0; i < fuseSets.size(); i++) {
        auto& compSet = fuseSets[i];
        /*
        for (auto comp : compSet) {
            dumpCmd(comp->cmd);
        }
        */
        bool fuseKernelVectorize = true;
        for (auto& node : compSet) {
            auto cmd = node->cmd;
            if(!cmd->canVectorize) {
                fuseKernelVectorize = false;
                break;
            }
        }
        target->setFuseKernelVectorize(fuseKernelVectorize);
        SourceModule fuseModule(target.get());
        InOutTensors tensors = fuseModule.buildKernel(compSet, i);
        auto inputs = tensors.first;
        auto outputs = tensors.second;
        // build Plugin Op
        std::shared_ptr<Command> cmdPlugin;
        {
            auto sourceCode = fuseModule.codegen();
            if(mapKernelSources.find(sourceCode) == mapKernelSources.end()) {
                int kernelCount = mapKernelSources.size();
                mapKernelSources.insert(std::pair<std::string, int>(sourceCode, kernelCount));
            }
            std::string kernelName = "kernel_" + std::to_string(mapKernelSources[sourceCode]);
            sourceCode.insert(fuseModule.strIndexForKernelNum(), kernelName);