# Neural Network Runtime对接AI推理框架开发指导 ## 场景介绍 Neural Network Runtime作为AI推理引擎和加速芯片的桥梁，为AI推理引擎提供精简的Native接口，满足推理引擎通过加速芯片执行端到端推理的需求。 本文以图1展示的Add单算子模型为例，介绍Neural Network Runtime的开发流程。Add算子包含两个输入、一个参数和一个输出，其中的activation参数用于指定Add算子中激活函数的类型。 **图1** Add单算子网络示意图 ![](https://contentcenter-vali-drcn.dbankcdn.cn/pvt_2/DeveloperAlliance_scene_100_1/e2/v3/dYb2Ykc8SLCpRaCMOxt_nQ/zh-cn_image_0000002779093155.png?HW-CC-KV=V1&HW-CC-Date=20260929T121643Z&HW-CC-Expire=31536000000&HW-CC-Sign=268B1A6A169A126AF80A4F59FEF329E70B649C18D3CA3A3A7DA6536396A06B09) ## 环境准备 ### 环境要求 Neural Network Runtime部件的环境要求如下： * 开发环境：Ubuntu 18.04及以上。 * 接入设备：系统定义的标准设备，系统中内置AI硬件驱动并已接入Neural Network Runtime。 由于Neural Network Runtime通过Native API对外开放，需要下载对应的SDK并通过Native开发套件编译Neural Network Runtime应用。可以使用DevEco Studio来搭建环境和编译代码。 ### 环境搭建 1. 使用Ubuntu编译服务器的终端。 2. 指定native工具链路径来编译代码，可以使用DevEco Studio来下载对应的SDK来进行编译。 3. DevEco Studio安装目录下的SDK路径可以在DevEco Studio工程界面，点击File > Settings... > 在settings中搜索SDK，下载对应的SDK即可。 ## 接口说明 以下为Neural Network Runtime开发流程中的常用接口，具体可见[NeuralNetworkRuntime](https://developer.huawei.com/consumer/cn/doc/harmonyos-references/capi-neuralnetworkruntime)。 ### 结构体 |结构体名称|描述| |:-----------------------------------------------|:---------------------------------------------------------| |typedef struct OH_NNModel OH_NNModel|Neural Network Runtime的模型句柄，用于构造模型。| |typedef struct OH_NNCompilation OH_NNCompilation|Neural Network Runtime的编译器句柄，用于编译AI模型。| |typedef struct OH_NNExecutor OH_NNExecutor|Neural Network Runtime的执行器句柄，用于在指定设备上执行推理计算。| |typedef struct NN_QuantParam NN_QuantParam|Neural Network Runtime的量化参数句柄，用于在构造模型时指定张量的量化参数。| |typedef struct NN_TensorDesc NN_TensorDesc|Neural Network Runtime的张量描述句柄，用于描述张量的各类属性，例如数据布局、数据类型、形状等。| |typedef struct NN_Tensor NN_Tensor|Neural Network Runtime的张量句柄，用于设置执行器的推理输入和输出张量。| ### 模型构造接口 |接口名称|描述| |:--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|:-------------------| |OH_NNModel_Construct()|创建OH_NNModel类型的模型实例。| |OH_NN_ReturnCode OH_NNModel_AddTensorToModel(OH_NNModel *model, const NN_TensorDesc *tensorDesc)|向模型实例中添加张量。| |OH_NN_ReturnCode OH_NNModel_SetTensorData(OH_NNModel *model, uint32_t index, const void *dataBuffer, size_t length)|设置张量的数值。| |OH_NN_ReturnCode OH_NNModel_AddOperation(OH_NNModel *model, OH_NN_OperationType op, const OH_NN_UInt32Array *paramIndices, const OH_NN_UInt32Array *inputIndices, const OH_NN_UInt32Array *outputIndices)|向模型实例中添加算子。| |OH_NN_ReturnCode OH_NNModel_SpecifyInputsAndOutputs(OH_NNModel *model, const OH_NN_UInt32Array *inputIndices, const OH_NN_UInt32Array *outputIndices)|指定模型的输入和输出张量的索引值。| |OH_NN_ReturnCode OH_NNModel_Finish(OH_NNModel *model)|完成模型构图。| |void OH_NNModel_Destroy(OH_NNModel **model)|销毁模型实例。| ### 模型编译接口 |接口名称|描述| |:-----------------------------------------------------------------------------------------------------------------------------------------------------------------|:-------------------------------------------| |OH_NNCompilation *OH_NNCompilation_Construct(const OH_NNModel *model)|基于模型实例创建OH_NNCompilation类型的编译实例。| |OH_NNCompilation *OH_NNCompilation_ConstructWithOfflineModelFile(const char *modelPath)|基于离线模型文件路径创建OH_NNCompilation类型的编译实例。| |OH_NNCompilation *OH_NNCompilation_ConstructWithOfflineModelBuffer(const void *modelBuffer, size_t modelSize)|基于离线模型文件内存创建OH_NNCompilation类型的编译实例。| |OH_NNCompilation *OH_NNCompilation_ConstructForCache()|创建一个空的编译实例，以便稍后从模型缓存中恢复。| |OH_NN_ReturnCode OH_NNCompilation_ExportCacheToBuffer(OH_NNCompilation *compilation, const void *buffer, size_t length, size_t *modelSize)|将模型缓存写入到指定内存区域。| |OH_NN_ReturnCode OH_NNCompilation_ImportCacheFromBuffer(OH_NNCompilation *compilation, const void *buffer, size_t modelSize)|从指定内存区域读取模型缓存。| |OH_NN_ReturnCode OH_NNCompilation_AddExtensionConfig(OH_NNCompilation *compilation, const char *configName, const void *configValue, const size_t configValueSize)|为自定义硬件属性添加扩展配置，具体硬件的扩展属性名称和属性值需要从硬件厂商的文档中获取。| |OH_NN_ReturnCode OH_NNCompilation_SetDevice(OH_NNCompilation *compilation, size_t deviceID)|指定模型编译和计算的硬件，可通过设备管理接口获取。| |OH_NN_ReturnCode OH_NNCompilation_SetCache(OH_NNCompilation *compilation, const char *cachePath, uint32_t version)|设置编译模型的缓存目录和版本。| |OH_NN_ReturnCode OH_NNCompilation_SetPerformanceMode(OH_NNCompilation *compilation, OH_NN_PerformanceMode performanceMode)|设置模型计算的性能模式。| |OH_NN_ReturnCode OH_NNCompilation_SetPriority(OH_NNCompilation *compilation, OH_NN_Priority priority)|设置模型计算的优先级。| |OH_NN_ReturnCode OH_NNCompilation_EnableFloat16(OH_NNCompilation *compilation, bool enableFloat16)|是否以float16的浮点数精度计算。| |OH_NN_ReturnCode OH_NNCompilation_Build(OH_NNCompilation *compilation)|执行模型编译。| |void OH_NNCompilation_Destroy(OH_NNCompilation **compilation)|销毁编译实例。| ### 张量描述接口 |接口名称|描述| |:---------------------------------------------------------------------------------------------------------------|:--------------------------| |NN_TensorDesc *OH_NNTensorDesc_Create()|创建一个张量描述实例，用于后续创建张量。| |OH_NN_ReturnCode OH_NNTensorDesc_SetName(NN_TensorDesc *tensorDesc, const char *name)|设置张量描述的名称。| |OH_NN_ReturnCode OH_NNTensorDesc_GetName(const NN_TensorDesc *tensorDesc, const char **name)|获取张量描述的名称。| |OH_NN_ReturnCode OH_NNTensorDesc_SetDataType(NN_TensorDesc *tensorDesc, OH_NN_DataType dataType)|设置张量描述的数据类型。| |OH_NN_ReturnCode OH_NNTensorDesc_GetDataType(const NN_TensorDesc *tensorDesc, OH_NN_DataType *dataType)|获取张量描述的数据类型。| |OH_NN_ReturnCode OH_NNTensorDesc_SetShape(NN_TensorDesc *tensorDesc, const int32_t *shape, size_t shapeLength)|设置张量描述的形状。| |OH_NN_ReturnCode OH_NNTensorDesc_GetShape(const NN_TensorDesc *tensorDesc, int32_t **shape, size_t *shapeLength)|获取张量描述的形状。| |OH_NN_ReturnCode OH_NNTensorDesc_SetFormat(NN_TensorDesc *tensorDesc, OH_NN_Format format)|设置张量描述的数据布局。| |OH_NN_ReturnCode OH_NNTensorDesc_GetFormat(const NN_TensorDesc *tensorDesc, OH_NN_Format *format)|获取张量描述的数据布局。| |OH_NN_ReturnCode OH_NNTensorDesc_GetElementCount(const NN_TensorDesc *tensorDesc, size_t *elementCount)|获取张量描述的元素个数。| |OH_NN_ReturnCode OH_NNTensorDesc_GetByteSize(const NN_TensorDesc *tensorDesc, size_t *byteSize)|获取基于张量描述的形状和数据类型计算的数据占用字节数。| |OH_NN_ReturnCode OH_NNTensorDesc_Destroy(NN_TensorDesc **tensorDesc)|销毁张量描述实例。| ### 张量接口 |接口名称|描述| |:------------------------------------------------------------------------------------------------------------------|:-------------------------------------------| |NN_Tensor* OH_NNTensor_Create(size_t deviceID, NN_TensorDesc *tensorDesc)|从张量描述创建张量实例，会申请设备共享内存。| |NN_Tensor* OH_NNTensor_CreateWithSize(size_t deviceID, NN_TensorDesc *tensorDesc, size_t size)|按照指定内存大小和张量描述创建张量实例，会申请设备共享内存。| |NN_Tensor* OH_NNTensor_CreateWithFd(size_t deviceID, NN_TensorDesc *tensorDesc, int fd, size_t size, size_t offset)|按照指定共享内存的文件描述符和张量描述创建张量实例，从而可以复用其他张量的设备共享内存。| |NN_TensorDesc* OH_NNTensor_GetTensorDesc(const NN_Tensor *tensor)|获取张量内部的张量描述实例指针，从而可读取张量的属性，例如数据类型、形状等。| |void* OH_NNTensor_GetDataBuffer(const NN_Tensor *tensor)|获取张量数据的内存地址，可以读写张量数据。| |OH_NN_ReturnCode OH_NNTensor_GetFd(const NN_Tensor *tensor, int *fd)|获取张量数据所在共享内存的文件描述符，文件描述符fd对应了一块设备共享内存。| |OH_NN_ReturnCode OH_NNTensor_GetSize(const NN_Tensor *tensor, size_t *size)|获取张量数据所在共享内存的大小。| |OH_NN_ReturnCode OH_NNTensor_GetOffset(const NN_Tensor *tensor, size_t *offset)|获取张量数据所在共享内存上的偏移量，张量数据可使用的大小为所在共享内存的大小减去偏移量。| |OH_NN_ReturnCode OH_NNTensor_Destroy(NN_Tensor **tensor)|销毁张量实例。| ### 执行推理接口 |接口名称|描述| |:--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|:------------------------------------------| |OH_NNExecutor *OH_NNExecutor_Construct(OH_NNCompilation *compilation)|创建OH_NNExecutor类型的执行器实例。| |OH_NN_ReturnCode OH_NNExecutor_GetOutputShape(OH_NNExecutor *executor, uint32_t outputIndex, int32_t **shape, uint32_t *shapeLength)|获取输出张量的维度信息，用于输出张量具有动态形状的情况。| |OH_NN_ReturnCode OH_NNExecutor_GetInputCount(const OH_NNExecutor *executor, size_t *inputCount)|获取输入张量的数量。| |OH_NN_ReturnCode OH_NNExecutor_GetOutputCount(const OH_NNExecutor *executor, size_t *outputCount)|获取输出张量的数量。| |NN_TensorDesc* OH_NNExecutor_CreateInputTensorDesc(const OH_NNExecutor *executor, size_t index)|由指定索引值创建一个输入张量的描述，用于读取张量的属性或创建张量实例。| |NN_TensorDesc* OH_NNExecutor_CreateOutputTensorDesc(const OH_NNExecutor *executor, size_t index)|由指定索引值创建一个输出张量的描述，用于读取张量的属性或创建张量实例。| |OH_NN_ReturnCode OH_NNExecutor_GetInputDimRange(const OH_NNExecutor *executor, size_t index, size_t **minInputDims, size_t **maxInputDims, size_t *shapeLength)|获取所有输入张量的维度范围。当输入张量具有动态形状时，不同设备可能支持不同的维度范围。| |OH_NN_ReturnCode OH_NNExecutor_SetOnRunDone(OH_NNExecutor *executor, NN_OnRunDone onRunDone)|设置异步推理结束后的回调处理函数，回调函数定义详见接口文档。| |OH_NN_ReturnCode OH_NNExecutor_SetOnServiceDied(OH_NNExecutor *executor, NN_OnServiceDied onServiceDied)|设置异步推理执行期间设备驱动服务突然死亡时的回调处理函数，回调函数定义详见接口文档。| |OH_NN_ReturnCode OH_NNExecutor_RunSync(OH_NNExecutor *executor, NN_Tensor *inputTensor[], size_t inputCount, NN_Tensor *outputTensor[], size_t outputCount)|执行同步推理。| |OH_NN_ReturnCode OH_NNExecutor_RunAsync(OH_NNExecutor *executor, NN_Tensor *inputTensor[], size_t inputCount, NN_Tensor *outputTensor[], size_t outputCount, int32_t timeout, void *userData)|执行异步推理。| |void OH_NNExecutor_Destroy(OH_NNExecutor **executor)|销毁执行器实例。| ### 设备管理接口 |接口名称|描述| |:-----------------------------------------------------------------------------------------------|:----------------------------------| |OH_NN_ReturnCode OH_NNDevice_GetAllDevicesID(const size_t **allDevicesID, uint32_t *deviceCount)|获取对接到Neural Network Runtime的所有硬件ID。| |OH_NN_ReturnCode OH_NNDevice_GetName(size_t deviceID, const char **name)|获取指定硬件的名称。| |OH_NN_ReturnCode OH_NNDevice_GetType(size_t deviceID, OH_NN_DeviceType *deviceType)|获取指定硬件的类别信息。| ## 开发步骤 Neural Network Runtime的开发流程主要包含**模型构造** 、**模型编译** 和**推理执行**三个阶段。以下开发步骤以Add单算子模型为例，介绍调用Neural Network Runtime接口，开发应用的过程。 1. 创建应用样例文件。 首先，创建Neural Network Runtime应用样例的源文件。在项目目录下执行以下命令，创建nnrt_example/目录，并在目录下创建 nnrt_example.cpp 源文件。 ```shell mkdir ~/nnrt_example && cd ~/nnrt_example touch nnrt_example.cpp ``` 2. 导入Neural Network Runtime。 在 nnrt_example.cpp 文件的开头添加以下代码，引入Neural Network Runtime。 ```cpp #include #include #include "neural_network_runtime/neural_network_runtime.h" ``` 3. 定义日志打印、设置输入数据、数据打印等辅助函数。 ```cpp // 返回值检查宏 #define CHECKNEQ(realRet, expectRet, retValue, ...) \ do { \ if ((realRet) != (expectRet)) { \ printf(__VA_ARGS__); \ return (retValue); \ } \ } while (0) #define CHECKEQ(realRet, expectRet, retValue, ...) \ do { \ if ((realRet) == (expectRet)) { \ printf(__VA_ARGS__); \ return (retValue); \ } \ } while (0) // 设置输入数据用于推理 OH_NN_ReturnCode SetInputData(NN_Tensor* inputTensor[], size_t inputSize) { OH_NN_DataType dataType(OH_NN_FLOAT32); OH_NN_ReturnCode ret{OH_NN_FAILED}; size_t elementCount = 0; for (size_t i = 0; i < inputSize; ++i) { // 获取张量的数据内存 auto data = OH_NNTensor_GetDataBuffer(inputTensor[i]); CHECKEQ(data, nullptr, OH_NN_FAILED, "Failed to get data buffer."); // 获取张量的描述 auto desc = OH_NNTensor_GetTensorDesc(inputTensor[i]); CHECKEQ(desc, nullptr, OH_NN_FAILED, "Failed to get desc."); // 获取张量的数据类型 ret = OH_NNTensorDesc_GetDataType(desc, &dataType); CHECKNEQ(ret, OH_NN_SUCCESS, OH_NN_FAILED, "Failed to get data type."); // 获取张量的元素个数 ret = OH_NNTensorDesc_GetElementCount(desc, &elementCount); CHECKNEQ(ret, OH_NN_SUCCESS, OH_NN_FAILED, "Failed to get element count."); switch(dataType) { case OH_NN_FLOAT32: { float* floatValue = reinterpret_cast(data); for (size_t j = 0; j < elementCount; ++j) { floatValue[j] = static_cast(j); } break; } case OH_NN_INT32: { int* intValue = reinterpret_cast(data); for (size_t j = 0; j < elementCount; ++j) { intValue[j] = static_cast(j); } break; } default: return OH_NN_FAILED; } } return OH_NN_SUCCESS; } OH_NN_ReturnCode Print(NN_Tensor* outputTensor[], size_t outputSize) { OH_NN_DataType dataType(OH_NN_FLOAT32); OH_NN_ReturnCode ret{OH_NN_FAILED}; size_t elementCount = 0; for (size_t i = 0; i < outputSize; ++i) { auto data = OH_NNTensor_GetDataBuffer(outputTensor[i]); CHECKEQ(data, nullptr, OH_NN_FAILED, "Failed to get data buffer."); auto desc = OH_NNTensor_GetTensorDesc(outputTensor[i]); CHECKEQ(desc, nullptr, OH_NN_FAILED, "Failed to get desc."); ret = OH_NNTensorDesc_GetDataType(desc, &dataType); CHECKNEQ(ret, OH_NN_SUCCESS, OH_NN_FAILED, "Failed to get data type."); ret = OH_NNTensorDesc_GetElementCount(desc, &elementCount); CHECKNEQ(ret, OH_NN_SUCCESS, OH_NN_FAILED, "Failed to get element count."); switch(dataType) { case OH_NN_FLOAT32: { float* floatValue = reinterpret_cast(data); for (size_t j = 0; j < elementCount; ++j) { std::cout << "Output index: " << j << ", value is: " << floatValue[j] << "." << std::endl; } break; } case OH_NN_INT32: { int* intValue = reinterpret_cast(data); for (size_t j = 0; j < elementCount; ++j) { std::cout << "Output index: " << j << ", value is: " << intValue[j] << "." << std::endl; } break; } default: return OH_NN_FAILED; } } return OH_NN_SUCCESS; } ``` 4. 构造模型。 使用Neural Network Runtime的模型构造接口，构造Add单算子样例模型。 ```cpp OH_NN_ReturnCode BuildModel(OH_NNModel** pmodel) { // 创建模型实例model，进行模型构造 OH_NNModel* model = OH_NNModel_Construct(); CHECKEQ(model, nullptr, OH_NN_FAILED, "Create model failed."); // 添加Add算子的第一个输入张量，类型为float32，张量形状为[1, 2, 2, 3] NN_TensorDesc* tensorDesc = OH_NNTensorDesc_Create(); CHECKEQ(tensorDesc, nullptr, OH_NN_FAILED, "Create TensorDesc failed."); int32_t inputDims[4] = {1, 2, 2, 3}; auto returnCode = OH_NNTensorDesc_SetShape(tensorDesc, inputDims, 4); CHECKNEQ(returnCode, OH_NN_SUCCESS, OH_NN_FAILED, "Set TensorDesc shape failed."); returnCode = OH_NNTensorDesc_SetDataType(tensorDesc, OH_NN_FLOAT32); CHECKNEQ(returnCode, OH_NN_SUCCESS, OH_NN_FAILED, "Set TensorDesc data type failed."); returnCode = OH_NNTensorDesc_SetFormat(tensorDesc, OH_NN_FORMAT_NONE); CHECKNEQ(returnCode, OH_NN_SUCCESS, OH_NN_FAILED, "Set TensorDesc format failed."); returnCode = OH_NNModel_AddTensorToModel(model, tensorDesc); CHECKNEQ(returnCode, OH_NN_SUCCESS, OH_NN_FAILED, "Add first TensorDesc to model failed."); returnCode = OH_NNModel_SetTensorType(model, 0, OH_NN_TENSOR); CHECKNEQ(returnCode, OH_NN_SUCCESS, OH_NN_FAILED, "Set model tensor type failed."); // 添加Add算子的第二个输入张量，类型为float32，张量形状为[1, 2, 2, 3] tensorDesc = OH_NNTensorDesc_Create(); CHECKEQ(tensorDesc, nullptr, OH_NN_FAILED, "Create TensorDesc failed."); returnCode = OH_NNTensorDesc_SetShape(tensorDesc, inputDims, 4); CHECKNEQ(returnCode, OH_NN_SUCCESS, OH_NN_FAILED, "Set TensorDesc shape failed."); returnCode = OH_NNTensorDesc_SetDataType(tensorDesc, OH_NN_FLOAT32); CHECKNEQ(returnCode, OH_NN_SUCCESS, OH_NN_FAILED, "Set TensorDesc data type failed."); returnCode = OH_NNTensorDesc_SetFormat(tensorDesc, OH_NN_FORMAT_NONE); CHECKNEQ(returnCode, OH_NN_SUCCESS, OH_NN_FAILED, "Set TensorDesc format failed."); returnCode = OH_NNModel_AddTensorToModel(model, tensorDesc); CHECKNEQ(returnCode, OH_NN_SUCCESS, OH_NN_FAILED, "Add second TensorDesc to model failed."); returnCode = OH_NNModel_SetTensorType(model, 1, OH_NN_TENSOR); CHECKNEQ(returnCode, OH_NN_SUCCESS, OH_NN_FAILED, "Set model tensor type failed."); // 添加Add算子的参数张量，该参数张量用于指定激活函数的类型，张量的数据类型为int8。 tensorDesc = OH_NNTensorDesc_Create(); CHECKEQ(tensorDesc, nullptr, OH_NN_FAILED, "Create TensorDesc failed."); int32_t activationDims = 1; returnCode = OH_NNTensorDesc_SetShape(tensorDesc, &activationDims, 1); CHECKNEQ(returnCode, OH_NN_SUCCESS, OH_NN_FAILED, "Set TensorDesc shape failed."); returnCode = OH_NNTensorDesc_SetDataType(tensorDesc, OH_NN_INT8); CHECKNEQ(returnCode, OH_NN_SUCCESS, OH_NN_FAILED, "Set TensorDesc data type failed."); returnCode = OH_NNTensorDesc_SetFormat(tensorDesc, OH_NN_FORMAT_NONE); CHECKNEQ(returnCode, OH_NN_SUCCESS, OH_NN_FAILED, "Set TensorDesc format failed."); returnCode = OH_NNModel_AddTensorToModel(model, tensorDesc); CHECKNEQ(returnCode, OH_NN_SUCCESS, OH_NN_FAILED, "Add second TensorDesc to model failed."); returnCode = OH_NNModel_SetTensorType(model, 2, OH_NN_ADD_ACTIVATIONTYPE); CHECKNEQ(returnCode, OH_NN_SUCCESS, OH_NN_FAILED, "Set model tensor type failed."); // 将激活函数类型设置为OH_NN_FUSED_NONE，表示该算子不添加激活函数。 int8_t activationValue = OH_NN_FUSED_NONE; returnCode = OH_NNModel_SetTensorData(model, 2, &activationValue, sizeof(int8_t)); CHECKNEQ(returnCode, OH_NN_SUCCESS, OH_NN_FAILED, "Set model tensor data failed."); // 设置Add算子的输出张量，类型为float32，张量形状为[1, 2, 2, 3] tensorDesc = OH_NNTensorDesc_Create(); CHECKEQ(tensorDesc, nullptr, OH_NN_FAILED, "Create TensorDesc failed."); returnCode = OH_NNTensorDesc_SetShape(tensorDesc, inputDims, 4); CHECKNEQ(returnCode, OH_NN_SUCCESS, OH_NN_FAILED, "Set TensorDesc shape failed."); returnCode = OH_NNTensorDesc_SetDataType(tensorDesc, OH_NN_FLOAT32); CHECKNEQ(returnCode, OH_NN_SUCCESS, OH_NN_FAILED, "Set TensorDesc data type failed."); returnCode = OH_NNTensorDesc_SetFormat(tensorDesc, OH_NN_FORMAT_NONE); CHECKNEQ(returnCode, OH_NN_SUCCESS, OH_NN_FAILED, "Set TensorDesc format failed."); returnCode = OH_NNModel_AddTensorToModel(model, tensorDesc); CHECKNEQ(returnCode, OH_NN_SUCCESS, OH_NN_FAILED, "Add forth TensorDesc to model failed."); returnCode = OH_NNModel_SetTensorType(model, 3, OH_NN_TENSOR); CHECKNEQ(returnCode, OH_NN_SUCCESS, OH_NN_FAILED, "Set model tensor type failed."); // 指定Add算子的输入张量、参数张量和输出张量的索引 uint32_t inputIndicesValues[2] = {0, 1}; uint32_t paramIndicesValues = 2; uint32_t outputIndicesValues = 3; OH_NN_UInt32Array paramIndices = {¶mIndicesValues, 1}; OH_NN_UInt32Array inputIndices = {inputIndicesValues, 2}; OH_NN_UInt32Array outputIndices = {&outputIndicesValues, 1}; // 向模型实例添加Add算子 returnCode = OH_NNModel_AddOperation(model, OH_NN_OPS_ADD, ¶mIndices, &inputIndices, &outputIndices); CHECKNEQ(returnCode, OH_NN_SUCCESS, OH_NN_FAILED, "Add operation to model failed."); // 设置模型实例的输入张量、输出张量的索引 returnCode = OH_NNModel_SpecifyInputsAndOutputs(model, &inputIndices, &outputIndices); CHECKNEQ(returnCode, OH_NN_SUCCESS, OH_NN_FAILED, "Specify model inputs and outputs failed."); // 完成模型实例的构建 returnCode = OH_NNModel_Finish(model); CHECKNEQ(returnCode, OH_NN_SUCCESS, OH_NN_FAILED, "Build model failed."); // 返回模型实例 *pmodel = model; return OH_NN_SUCCESS; } ``` 5. 查询Neural Network Runtime已经对接的AI加速芯片。 Neural Network Runtime支持通过HDI接口，对接多种AI加速芯片。在执行模型编译前，需要查询当前设备下，Neural Network Runtime已经对接的AI加速芯片。每个AI加速芯片对应唯一的ID值，在编译阶段需要通过设备ID，指定模型编译的芯片。 ```cpp void GetAvailableDevices(std::vector& availableDevice) { availableDevice.clear(); // 获取可用的硬件ID const size_t* devices = nullptr; uint32_t deviceCount = 0; OH_NN_ReturnCode ret = OH_NNDevice_GetAllDevicesID(&devices, &deviceCount); if (ret != OH_NN_SUCCESS) { std::cout << "GetAllDevicesID failed, get no available device." << std::endl; return; } for (uint32_t i = 0; i < deviceCount; i++) { availableDevice.emplace_back(devices[i]); } } ``` 6. 在指定的设备上编译模型。 Neural Network Runtime使用抽象的模型表达描述AI模型的拓扑结构。在AI加速芯片上执行前，需要通过Neural Network Runtime提供的编译模块来创建编译实例，并由编译实例将抽象的模型表达下发至芯片驱动层，转换成可以直接推理计算的格式，即模型编译。 ```cpp OH_NN_ReturnCode CreateCompilation(OH_NNModel* model, const std::vector& availableDevice, OH_NNCompilation** pCompilation) { // 创建编译实例compilation，将构图的模型实例或MSLite传下来的模型实例传入 OH_NNCompilation* compilation = OH_NNCompilation_Construct(model); CHECKEQ(compilation, nullptr, OH_NN_FAILED, "OH_NNCore_ConstructCompilationWithNNModel failed."); // 设置编译的硬件、缓存路径、性能模式、计算优先级、是否开启float16低精度计算等选项 // 选择在第一个设备上编译模型 auto returnCode = OH_NNCompilation_SetDevice(compilation, availableDevice[0]); CHECKNEQ(returnCode, OH_NN_SUCCESS, OH_NN_FAILED, "OH_NNCompilation_SetDevice failed."); // 将模型编译结果缓存在/data/local/tmp目录下，版本号指定为1 returnCode = OH_NNCompilation_SetCache(compilation, "/data/local/tmp", 1); CHECKNEQ(returnCode, OH_NN_SUCCESS, OH_NN_FAILED, "OH_NNCompilation_SetCache failed."); // 设置硬件性能模式 returnCode = OH_NNCompilation_SetPerformanceMode(compilation, OH_NN_PERFORMANCE_EXTREME); CHECKNEQ(returnCode, OH_NN_SUCCESS, OH_NN_FAILED, "OH_NNCompilation_SetPerformanceMode failed."); // 设置推理执行优先级 returnCode = OH_NNCompilation_SetPriority(compilation, OH_NN_PRIORITY_HIGH); CHECKNEQ(returnCode, OH_NN_SUCCESS, OH_NN_FAILED, "OH_NNCompilation_SetPriority failed."); // 是否开启FP16计算模式 returnCode = OH_NNCompilation_EnableFloat16(compilation, false); CHECKNEQ(returnCode, OH_NN_SUCCESS, OH_NN_FAILED, "OH_NNCompilation_EnableFloat16 failed."); // 执行模型编译 returnCode = OH_NNCompilation_Build(compilation); CHECKNEQ(returnCode, OH_NN_SUCCESS, OH_NN_FAILED, "OH_NNCompilation_Build failed."); *pCompilation = compilation; return OH_NN_SUCCESS; } ``` 7. 创建执行器。 完成模型编译后，需要调用Neural Network Runtime的执行模块，通过编译实例创建执行器。模型推理阶段中的设置模型输入、触发推理计算以及获取模型输出等操作均需要围绕执行器完成。 ```cpp OH_NNExecutor* CreateExecutor(OH_NNCompilation* compilation) { // 通过编译实例compilation创建执行器executor OH_NNExecutor *executor = OH_NNExecutor_Construct(compilation); CHECKEQ(executor, nullptr, nullptr, "OH_NNExecutor_Construct failed."); return executor; } ``` 8. 执行推理计算，并打印推理结果。 通过执行模块提供的接口，将推理计算所需要的输入数据传递给执行器，触发执行器完成一次推理计算，获取模型的推理结果并打印。 ```cpp OH_NN_ReturnCode Run(OH_NNExecutor* executor, const std::vector& availableDevice) { // 从executor获取输入输出信息 // 获取输入张量的个数 size_t inputCount = 0; auto returnCode = OH_NNExecutor_GetInputCount(executor, &inputCount); CHECKNEQ(returnCode, OH_NN_SUCCESS, OH_NN_FAILED, "OH_NNExecutor_GetInputCount failed."); std::vector inputTensorDescs; NN_TensorDesc* tensorDescTmp = nullptr; for (size_t i = 0; i < inputCount; ++i) { // 创建输入张量的描述 tensorDescTmp = OH_NNExecutor_CreateInputTensorDesc(executor, i); CHECKEQ(tensorDescTmp, nullptr, OH_NN_FAILED, "OH_NNExecutor_CreateInputTensorDesc failed."); inputTensorDescs.emplace_back(tensorDescTmp); } // 获取输出张量的个数 size_t outputCount = 0; returnCode = OH_NNExecutor_GetOutputCount(executor, &outputCount); CHECKNEQ(returnCode, OH_NN_SUCCESS, OH_NN_FAILED, "OH_NNExecutor_GetOutputCount failed."); std::vector outputTensorDescs; for (size_t i = 0; i < outputCount; ++i) { // 创建输出张量的描述 tensorDescTmp = OH_NNExecutor_CreateOutputTensorDesc(executor, i); CHECKEQ(tensorDescTmp, nullptr, OH_NN_FAILED, "OH_NNExecutor_CreateOutputTensorDesc failed."); outputTensorDescs.emplace_back(tensorDescTmp); } // 创建输入和输出张量 NN_Tensor* inputTensors[inputCount]; NN_Tensor* tensor = nullptr; for (size_t i = 0; i < inputCount; ++i) { tensor = nullptr; tensor = OH_NNTensor_Create(availableDevice[0], inputTensorDescs[i]); CHECKEQ(tensor, nullptr, OH_NN_FAILED, "OH_NNTensor_Create failed."); inputTensors[i] = tensor; } NN_Tensor* outputTensors[outputCount]; for (size_t i = 0; i < outputCount; ++i) { tensor = nullptr; tensor = OH_NNTensor_Create(availableDevice[0], outputTensorDescs[i]); CHECKEQ(tensor, nullptr, OH_NN_FAILED, "OH_NNTensor_Create failed."); outputTensors[i] = tensor; } // 设置输入张量的数据 returnCode = SetInputData(inputTensors, inputCount); CHECKNEQ(returnCode, OH_NN_SUCCESS, OH_NN_FAILED, "SetInputData failed."); // 执行推理 returnCode = OH_NNExecutor_RunSync(executor, inputTensors, inputCount, outputTensors, outputCount); CHECKNEQ(returnCode, OH_NN_SUCCESS, OH_NN_FAILED, "OH_NNExecutor_RunSync failed."); // 打印输出张量的数据 Print(outputTensors, outputCount); // 清理输入和输出张量以及张量描述 for (size_t i = 0; i < inputCount; ++i) { returnCode = OH_NNTensor_Destroy(&inputTensors[i]); CHECKNEQ(returnCode, OH_NN_SUCCESS, OH_NN_FAILED, "OH_NNTensor_Destroy failed."); returnCode = OH_NNTensorDesc_Destroy(&inputTensorDescs[i]); CHECKNEQ(returnCode, OH_NN_SUCCESS, OH_NN_FAILED, "OH_NNTensorDesc_Destroy failed."); } for (size_t i = 0; i < outputCount; ++i) { returnCode = OH_NNTensor_Destroy(&outputTensors[i]); CHECKNEQ(returnCode, OH_NN_SUCCESS, OH_NN_FAILED, "OH_NNTensor_Destroy failed."); returnCode = OH_NNTensorDesc_Destroy(&outputTensorDescs[i]); CHECKNEQ(returnCode, OH_NN_SUCCESS, OH_NN_FAILED, "OH_NNTensorDesc_Destroy failed."); } return OH_NN_SUCCESS; } ``` 9. 构建端到端模型构造-编译-执行流程。 步骤4-步骤8实现了模型的模型构造、编译和执行流程，并封装成多个函数，便于模块化开发。以下示例代码将串联这些函数， 形成一个完整的Neural Network Runtime使用流程。 ```cpp int main(int argc, char** argv) { OH_NNModel* model = nullptr; OH_NNCompilation* compilation = nullptr; OH_NNExecutor* executor = nullptr; std::vector availableDevices; // 模型构造 OH_NN_ReturnCode ret = BuildModel(&model); if (ret != OH_NN_SUCCESS) { std::cout << "BuildModel failed." << std::endl; OH_NNModel_Destroy(&model); return -1; } // 获取可执行的设备 GetAvailableDevices(availableDevices); if (availableDevices.empty()) { std::cout << "No available device." << std::endl; OH_NNModel_Destroy(&model); return -1; } // 模型编译 ret = CreateCompilation(model, availableDevices, &compilation); if (ret != OH_NN_SUCCESS) { std::cout << "CreateCompilation failed." << std::endl; OH_NNModel_Destroy(&model); OH_NNCompilation_Destroy(&compilation); return -1; } // 销毁模型实例 OH_NNModel_Destroy(&model); // 创建模型的推理执行器 executor = CreateExecutor(compilation); if (executor == nullptr) { std::cout << "CreateExecutor failed, no executor is created." << std::endl; OH_NNCompilation_Destroy(&compilation); return -1; } // 销毁编译实例 OH_NNCompilation_Destroy(&compilation); // 使用上一步创建的执行器，执行推理计算 ret = Run(executor, availableDevices); if (ret != OH_NN_SUCCESS) { std::cout << "Run failed." << std::endl; OH_NNExecutor_Destroy(&executor); return -1; } // 销毁执行器实例 OH_NNExecutor_Destroy(&executor); return 0; } ``` ## 调测验证 1. 准备应用样例的编译配置文件。 新建一个 CMakeLists.txt 文件，为开发步骤中的应用样例文件 nnrt_example.cpp 添加编译配置。以下提供简单的 CMakeLists.txt 示例： ```text cmake_minimum_required(VERSION 3.16) project(nnrt_example C CXX) add_executable(nnrt_example ./nnrt_example.cpp ) target_link_libraries(nnrt_example neural_network_runtime neural_network_core ) ``` 2. 编译应用样例。 执行以下命令，在当前目录下新建build/目录，在build/目录下编译 nnrt_example.cpp，得到二进制文件 nnrt_example。 ```shell mkdir build && cd build cmake -DCMAKE_TOOLCHAIN_FILE={交叉编译工具链的路径}/build/cmake/ohos.toolchain.cmake -DOHOS_ARCH=arm64-v8a -DOHOS_PLATFORM=OHOS -DOHOS_STL=c++_static .. make ``` 3. 执行以下代码，将样例推送到设备上执行。 ```shell # 将编译得到的 `nnrt_example` 推送到设备上，执行样例。 hdc_std file send ./nnrt_example /data/local/tmp/. # 给测试用例可执行文件加上权限。 hdc_std shell "chmod +x /data/local/tmp/nnrt_example" # 执行测试用例 hdc_std shell "/data/local/tmp/nnrt_example" ``` 如果样例执行正常，应该得到以下输出。 ```text Output index: 0, value is: 0.000000. Output index: 1, value is: 2.000000. Output index: 2, value is: 4.000000. Output index: 3, value is: 6.000000. Output index: 4, value is: 8.000000. Output index: 5, value is: 10.000000. Output index: 6, value is: 12.000000. Output index: 7, value is: 14.000000. Output index: 8, value is: 16.000000. Output index: 9, value is: 18.000000. Output index: 10, value is: 20.000000. Output index: 11, value is: 22.000000. ``` 4. 检查模型缓存（可选）。 如果在调测环境下，Neural Network Runtime对接的HDI服务支持模型缓存功能，执行完 nnrt_example, 可以在 /data/local/tmp 目录下找到生成的缓存文件。 > 说明 > > 模型的IR需要传递到硬件驱动层，由HDI服务将统一的IR图，编译成硬件专用的计算图，编译的过程非常耗时。Neural Network Runtime支持计算图缓存的特性，可以将HDI服务编译生成的计算图，缓存到设备存储中。当下一次在同一个加速芯片上编译同一个模型时，通过指定缓存的路径，Neural Network Runtime可以直接加载缓存文件中的计算图，减少编译消耗的时间。 检查缓存目录下的缓存文件： ```shell ls /data/local/tmp ``` 以下为打印结果： ```text # 0.nncache 1.nncache 2.nncache cache_info.nncache ``` 如果缓存不再使用，需要手动删除缓存，可以参考以下命令，删除缓存文件。 ```shell rm /data/local/tmp/*nncache ```