harmony 鸿蒙NeuralNetworkRuntime

2023-06-24
浏览 (768)

NeuralNetworkRuntime

Overview

Provides APIs of Neural Network Runtime for accelerating the model inference.

Provides APIs for accelerating the Neural Network Runtime model inference.

\@Syscap SystemCapability.Ai.NeuralNetworkRuntime

Since: 9

Summary

Files

Name	Description
neural_network_runtime.h	Defines the Neural Network Runtime APIs. The AI inference framework uses the Native APIs provided by Neural Network Runtime to construct and compile models and perform inference and computing on acceleration hardware. Note: Currently, the APIs of Neural Network Runtime do not support multi-thread calling. File to Include: Library: libneural_network_runtime.so
neural_network_runtime_type.h	Defines the structure and enumeration for Neural Network Runtime. File to Include: Library: libneural_network_runtime.so

Structs

Name	Description
OH_NN_UInt32Array	Defines a 32-bit unsigned integer array.
OH_NN_QuantParam	Defines quantization information.
OH_NN_Tensor	Defines the tensor structure.
OH_NN_Memory	Defines the memory structure.

Types

Name	Description
OH_NNModel	Defines the handles of models for Neural Network Runtime.
OH_NNCompilation	Defines the compiler handle for Neural Network Runtime.
OH_NNExecutor	Defines the executor handle for Neural Network Runtime.
OH_NN_UInt32Array	Defines a 32-bit unsigned integer array.
OH_NN_QuantParam	Defines quantization information.
OH_NN_Tensor	Defines the tensor structure.
OH_NN_Memory	Defines the memory structure.

Enums

Name	Description
OH_NN_PerformanceMode { OH_NN_PERFORMANCE_NONE = 0, OH_NN_PERFORMANCE_LOW = 1, OH_NN_PERFORMANCE_MEDIUM = 2, OH_NN_PERFORMANCE_HIGH = 3, OH_NN_PERFORMANCE_EXTREME = 4 }	Defines the hardware performance mode.
OH_NN_Priority { OH_NN_PRIORITY_NONE = 0, OH_NN_PRIORITY_LOW = 1, OH_NN_PRIORITY_MEDIUM = 2, OH_NN_PRIORITY_HIGH = 3 }	Defines the model inference task priority.
OH_NN_ReturnCode { OH_NN_SUCCESS = 0, OH_NN_FAILED = 1, OH_NN_INVALID_PARAMETER = 2, OH_NN_MEMORY_ERROR = 3, OH_NN_OPERATION_FORBIDDEN = 4, OH_NN_NULL_PTR = 5, OH_NN_INVALID_FILE = 6, OH_NN_UNAVALIDABLE_DEVICE = 7, OH_NN_INVALID_PATH = 8 }	Defines error codes for Neural Network Runtime.
OH_NN_FuseType : int8_t { OH_NN_FUSED_NONE = 0, OH_NN_FUSED_RELU = 1, OH_NN_FUSED_RELU6 = 2 }	Defines activation function types in the fusion operator for Neural Network Runtime.
OH_NN_Format { OH_NN_FORMAT_NONE = 0, OH_NN_FORMAT_NCHW = 1, OH_NN_FORMAT_NHWC = 2 }	Defines the layout type of tensor data.
OH_NN_DeviceType { OH_NN_OTHERS = 0, OH_NN_CPU = 1, OH_NN_GPU = 2, OH_NN_ACCELERATOR = 3 }	Defines device types supported by Neural Network Runtime.
OH_NN_DataType { OH_NN_UNKNOWN = 0, OH_NN_BOOL = 1, OH_NN_INT8 = 2, OH_NN_INT16 = 3, OH_NN_INT32 = 4, OH_NN_INT64 = 5, OH_NN_UINT8 = 6, OH_NN_UINT16 = 7, OH_NN_UINT32 = 8, OH_NN_UINT64 = 9, OH_NN_FLOAT16 = 10, OH_NN_FLOAT32 = 11, OH_NN_FLOAT64 = 12 }	Defines tensor data types supported by Neural Network Runtime.
OH_NN_OperationType { OH_NN_OPS_ADD = 1, OH_NN_OPS_AVG_POOL = 2, OH_NN_OPS_BATCH_NORM = 3, OH_NN_OPS_BATCH_TO_SPACE_ND = 4, OH_NN_OPS_BIAS_ADD = 5, OH_NN_OPS_CAST = 6, OH_NN_OPS_CONCAT = 7, OH_NN_OPS_CONV2D = 8, OH_NN_OPS_CONV2D_TRANSPOSE = 9, OH_NN_OPS_DEPTHWISE_CONV2D_NATIVE = 10, OH_NN_OPS_DIV = 11, OH_NN_OPS_ELTWISE = 12, OH_NN_OPS_EXPAND_DIMS = 13, OH_NN_OPS_FILL = 14, OH_NN_OPS_FULL_CONNECTION = 15, OH_NN_OPS_GATHER = 16, OH_NN_OPS_HSWISH = 17, OH_NN_OPS_LESS_EQUAL = 18, OH_NN_OPS_MATMUL = 19, OH_NN_OPS_MAXIMUM = 20, OH_NN_OPS_MAX_POOL = 21, OH_NN_OPS_MUL = 22, OH_NN_OPS_ONE_HOT = 23, OH_NN_OPS_PAD = 24, OH_NN_OPS_POW = 25, OH_NN_OPS_SCALE = 26, OH_NN_OPS_SHAPE = 27, OH_NN_OPS_SIGMOID = 28, OH_NN_OPS_SLICE = 29, OH_NN_OPS_SOFTMAX = 30, OH_NN_OPS_SPACE_TO_BATCH_ND = 31, OH_NN_OPS_SPLIT = 32, OH_NN_OPS_SQRT = 33, OH_NN_OPS_SQUARED_DIFFERENCE = 34, OH_NN_OPS_SQUEEZE = 35, OH_NN_OPS_STACK = 36, OH_NN_OPS_STRIDED_SLICE = 37, OH_NN_OPS_SUB = 38, OH_NN_OPS_TANH = 39, OH_NN_OPS_TILE = 40, OH_NN_OPS_TRANSPOSE = 41, OH_NN_OPS_REDUCE_MEAN = 42, OH_NN_OPS_RESIZE_BILINEAR = 43, OH_NN_OPS_RSQRT = 44, OH_NN_OPS_RESHAPE = 45, OH_NN_OPS_PRELU = 46, OH_NN_OPS_RELU = 47, OH_NN_OPS_RELU6 = 48, OH_NN_OPS_LAYER_NORM = 49, OH_NN_OPS_REDUCE_PROD = 50, OH_NN_OPS_REDUCE_ALL = 51, OH_NN_OPS_QUANT_DTYPE_CAST = 52, OH_NN_OPS_TOP_K = 53, OH_NN_OPS_ARG_MAX = 54, OH_NN_OPS_UNSQUEEZE = 55, OH_NN_OPS_GELU = 56 }	Defines operator types supported by Neural Network Runtime.
OH_NN_TensorType { OH_NN_TENSOR = 0, OH_NN_ADD_ACTIVATIONTYPE = 1, OH_NN_AVG_POOL_KERNEL_SIZE = 2, OH_NN_AVG_POOL_STRIDE = 3, OH_NN_AVG_POOL_PAD_MODE = 4, OH_NN_AVG_POOL_PAD = 5, OH_NN_AVG_POOL_ACTIVATION_TYPE = 6, OH_NN_BATCH_NORM_EPSILON = 7, OH_NN_BATCH_TO_SPACE_ND_BLOCKSIZE = 8, OH_NN_BATCH_TO_SPACE_ND_CROPS = 9, OH_NN_CONCAT_AXIS = 10, OH_NN_CONV2D_STRIDES = 11, OH_NN_CONV2D_PAD = 12, OH_NN_CONV2D_DILATION = 13, OH_NN_CONV2D_PAD_MODE = 14, OH_NN_CONV2D_ACTIVATION_TYPE = 15, OH_NN_CONV2D_GROUP = 16, OH_NN_CONV2D_TRANSPOSE_STRIDES = 17, OH_NN_CONV2D_TRANSPOSE_PAD = 18, OH_NN_CONV2D_TRANSPOSE_DILATION = 19, OH_NN_CONV2D_TRANSPOSE_OUTPUT_PADDINGS = 20, OH_NN_CONV2D_TRANSPOSE_PAD_MODE = 21, OH_NN_CONV2D_TRANSPOSE_ACTIVATION_TYPE = 22, OH_NN_CONV2D_TRANSPOSE_GROUP = 23, OH_NN_DEPTHWISE_CONV2D_NATIVE_STRIDES = 24, OH_NN_DEPTHWISE_CONV2D_NATIVE_PAD = 25, OH_NN_DEPTHWISE_CONV2D_NATIVE_DILATION = 26, OH_NN_DEPTHWISE_CONV2D_NATIVE_PAD_MODE = 27, OH_NN_DEPTHWISE_CONV2D_NATIVE_ACTIVATION_TYPE = 28, OH_NN_DIV_ACTIVATIONTYPE = 29, OH_NN_ELTWISE_MODE = 30, OH_NN_FULL_CONNECTION_AXIS = 31, OH_NN_FULL_CONNECTION_ACTIVATIONTYPE = 32, OH_NN_MATMUL_TRANSPOSE_A = 33, OH_NN_MATMUL_TRANSPOSE_B = 34, OH_NN_MATMUL_ACTIVATION_TYPE = 35, OH_NN_MAX_POOL_KERNEL_SIZE = 36, OH_NN_MAX_POOL_STRIDE = 37, OH_NN_MAX_POOL_PAD_MODE = 38, OH_NN_MAX_POOL_PAD = 39, OH_NN_MAX_POOL_ACTIVATION_TYPE = 40, OH_NN_MUL_ACTIVATION_TYPE = 41, OH_NN_ONE_HOT_AXIS = 42, OH_NN_PAD_CONSTANT_VALUE = 43, OH_NN_SCALE_ACTIVATIONTYPE = 44, OH_NN_SCALE_AXIS = 45, OH_NN_SOFTMAX_AXIS = 46, OH_NN_SPACE_TO_BATCH_ND_BLOCK_SHAPE = 47, OH_NN_SPACE_TO_BATCH_ND_PADDINGS = 48, OH_NN_SPLIT_AXIS = 49, OH_NN_SPLIT_OUTPUT_NUM = 50, OH_NN_SPLIT_SIZE_SPLITS = 51, OH_NN_SQUEEZE_AXIS = 52, OH_NN_STACK_AXIS = 53, OH_NN_STRIDED_SLICE_BEGIN_MASK = 54, OH_NN_STRIDED_SLICE_END_MASK = 55, OH_NN_STRIDED_SLICE_ELLIPSIS_MASK = 56, OH_NN_STRIDED_SLICE_NEW_AXIS_MASK = 57, OH_NN_STRIDED_SLICE_SHRINK_AXIS_MASK = 58, OH_NN_SUB_ACTIVATIONTYPE = 59, OH_NN_REDUCE_MEAN_KEEP_DIMS = 60, OH_NN_RESIZE_BILINEAR_NEW_HEIGHT = 61, OH_NN_RESIZE_BILINEAR_NEW_WIDTH = 62, OH_NN_RESIZE_BILINEAR_PRESERVE_ASPECT_RATIO = 63, OH_NN_RESIZE_BILINEAR_COORDINATE_TRANSFORM_MODE = 64, OH_NN_RESIZE_BILINEAR_EXCLUDE_OUTSIDE = 65, OH_NN_LAYER_NORM_BEGIN_NORM_AXIS = 66, OH_NN_LAYER_NORM_EPSILON = 67, OH_NN_LAYER_NORM_BEGIN_PARAM_AXIS = 68, OH_NN_LAYER_NORM_ELEMENTWISE_AFFINE = 69, OH_NN_REDUCE_PROD_KEEP_DIMS = 70, OH_NN_REDUCE_ALL_KEEP_DIMS = 71, OH_NN_QUANT_DTYPE_CAST_SRC_T = 72, OH_NN_QUANT_DTYPE_CAST_DST_T = 73, OH_NN_TOP_K_SORTED = 74, OH_NN_ARG_MAX_AXIS = 75, OH_NN_ARG_MAX_KEEPDIMS = 76, OH_NN_UNSQUEEZE_AXIS = 77 }	Enumerates the tensor data types.

Functions

Name	Description
OH_NNModel_Construct (void)	Creates a model instance of the OH_NNModel type and uses other APIs provided by OH_NNModel to construct the model instance.
OH_NNModel_AddTensor (OH_NNModel model, const OH_NN_Tensor tensor)	Adds a tensor to a model instance.
OH_NNModel_SetTensorData (OH_NNModel model, uint32_t index, const void dataBuffer, size_t length)	Sets the tensor value.
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)	Adds an operator to a model instance.
OH_NNModel_SpecifyInputsAndOutputs (OH_NNModel model, const OH_NN_UInt32Array inputIndices, const OH_NN_UInt32Array *outputIndices)	Specifies the inputs and outputs of a model.
OH_NNModel_Finish (OH_NNModel *model)	Completes model composition.
OH_NNModel_Destroy (OH_NNModel **model)	Releases a model instance.
OH_NNModel_GetAvailableOperations (OH_NNModel model, size_t deviceID, const bool isSupported, uint32_t opCount)	Queries whether the device supports operators in the model. The support status is indicated by the Boolean value.
OH_NNCompilation_Construct (const OH_NNModel *model)	Creates a compilation instance of the OH_NNCompilation type.
OH_NNCompilation_SetDevice (OH_NNCompilation *compilation, size_t deviceID)	Specifies the device for model compilation and computing.
OH_NNCompilation_SetCache (OH_NNCompilation compilation, const char cachePath, uint32_t version)	Set the cache directory and version of the compiled model.
OH_NNCompilation_SetPerformanceMode (OH_NNCompilation *compilation, OH_NN_PerformanceMode performanceMode)	Sets the performance mode for model computing.
OH_NNCompilation_SetPriority (OH_NNCompilation *compilation, OH_NN_Priority priority)	Sets the model computing priority.
OH_NNCompilation_EnableFloat16 (OH_NNCompilation *compilation, bool enableFloat16)	Enables float16 for computing.
OH_NNCompilation_Build (OH_NNCompilation *compilation)	Compiles a model.
OH_NNCompilation_Destroy (OH_NNCompilation **compilation)	Releases the Compilation object.
OH_NNExecutor_Construct (OH_NNCompilation *compilation)	OH_NNExecutor * Creates an executor instance of the OH_NNExecutor type.
OH_NNExecutor_SetInput (OH_NNExecutor executor, uint32_t inputIndex, const OH_NN_Tensor tensor, const void *dataBuffer, size_t length)	Sets the single input data for a model.
OH_NNExecutor_SetOutput (OH_NNExecutor executor, uint32_t outputIndex, void dataBuffer, size_t length)	Sets the buffer for a single output of a model.
OH_NNExecutor_GetOutputShape (OH_NNExecutor executor, uint32_t outputIndex, int32_t shape, uint32_t shapeLength)	Obtains the dimension information about the output tensor.
OH_NNExecutor_Run (OH_NNExecutor *executor)	Performs inference.
OH_NNExecutor_AllocateInputMemory (OH_NNExecutor *executor, uint32_t inputIndex, size_t length)	Allocates shared memory to a single input on a device.
OH_NNExecutor_AllocateOutputMemory (OH_NNExecutor *executor, uint32_t outputIndex, size_t length)	Allocates shared memory to a single output on a device.
OH_NNExecutor_DestroyInputMemory (OH_NNExecutor executor, uint32_t inputIndex, OH_NN_Memory *memory)	Releases the input memory to which the OH_NN_Memory instance points.
OH_NNExecutor_DestroyOutputMemory (OH_NNExecutor executor, uint32_t outputIndex, OH_NN_Memory *memory)	Releases the output memory to which the OH_NN_Memory instance points.
OH_NNExecutor_SetInputWithMemory (OH_NNExecutor executor, uint32_t inputIndex, const OH_NN_Tensor tensor, const OH_NN_Memory *memory)	Specifies the hardware shared memory pointed to by the OH_NN_Memory instance as the shared memory used by a single input.
OH_NNExecutor_SetOutputWithMemory (OH_NNExecutor executor, uint32_t outputIndex, const OH_NN_Memory memory)	Specifies the hardware shared memory pointed to by the OH_NN_Memory instance as the shared memory used by a single output.
OH_NNExecutor_Destroy (OH_NNExecutor **executor)	Destroys an executor instance to release the memory occupied by the executor.
OH_NNDevice_GetAllDevicesID (const size_t *allDevicesID, uint32_t deviceCount)	Obtains the ID of the device connected to Neural Network Runtime.
OH_NNDevice_GetName (size_t deviceID, const char **name)	Obtains the name of the specified device.
OH_NNDevice_GetType (size_t deviceID, OH_NN_DeviceType *deviceType)	Obtains the type information of the specified device.

Type Description

OH_NN_Memory

typedef struct OH_NN_MemoryOH_NN_Memory

Description
Defines the memory structure.

OH_NN_QuantParam

typedef struct OH_NN_QuantParamOH_NN_QuantParam

Description
Defines the quantization information.

In quantization scenarios, the 32-bit floating-point data type is quantized into the fixed-point data type according to the following formula:

s and z are quantization parameters, which are stored by scale and zeroPoint in OH_NN_QuantParam. r is a floating point number, q is the quantization result, q_min is the lower bound of the quantization result, and q_max is an upper bound of a quantization result. The calculation method is as follows:

The clamp function is defined as follows:

OH_NN_Tensor

typedef struct OH_NN_TensorOH_NN_Tensor

Description
Defines the tensor structure.

It is usually used to construct data nodes and operator parameters in a model graph. When constructing a tensor, you need to specify the data type, number of dimensions, dimension information, and quantization information.

OH_NN_UInt32Array

typedef struct OH_NN_UInt32ArrayOH_NN_UInt32Array

Description
Defines a 32-bit unsigned integer array.

OH_NNCompilation

typedef struct OH_NNCompilationOH_NNCompilation

Description
Defines the compiler handle for Neural Network Runtime.

OH_NNExecutor

typedef struct OH_NNExecutorOH_NNExecutor

Description
Defines the executor handle for Neural Network Runtime.

OH_NNModel

typedef struct OH_NNModelOH_NNModel

Description
Defines the handles of models for Neural Network Runtime.

Enum Description

OH_NN_DataType

enum OH_NN_DataType

Description
Defines tensor data types supported by Neural Network Runtime.

Name	Description
OH_NN_UNKNOWN	Unknown type
OH_NN_BOOL	bool
OH_NN_INT8	int8
OH_NN_INT16	int16
OH_NN_INT32	int32
OH_NN_INT64	int64
OH_NN_UINT8	uint8
OH_NN_UINT16	uint16
OH_NN_UINT32	uint32
OH_NN_UINT64	uint64
OH_NN_FLOAT16	float16
OH_NN_FLOAT32	float32
OH_NN_FLOAT64	float64

OH_NN_DeviceType

enum OH_NN_DeviceType

Description
Defines device types supported by Neural Network Runtime.

Name	Description
OH_NN_OTHERS	Devices that are not CPU, GPU, or dedicated accelerator
OH_NN_CPU	CPU device
OH_NN_GPU	GPU device
OH_NN_ACCELERATOR	Dedicated hardware accelerator

OH_NN_Format

enum OH_NN_Format

Description
Defines the layout type of tensor data.

Name	Description
OH_NN_FORMAT_NONE	The tensor does not have a specific layout type (such as scalar or vector).
OH_NN_FORMAT_NCHW	The tensor arranges data in NCHW format.
OH_NN_FORMAT_NHWC	The tensor arranges data in NHWC format.

OH_NN_FuseType

enum OH_NN_FuseType : int8_t

Description
Defines activation function types in the fusion operator for Neural Network Runtime.

Name	Description
OH_NN_FUSED_NONE	The fusion activation function is not specified.
OH_NN_FUSED_RELU	Fusion relu activation function
OH_NN_FUSED_RELU6	Fusion relu6 activation function

OH_NN_OperationType

enum OH_NN_OperationType

Description
Defines operator types supported by Neural Network Runtime.

Name	Description
OH_NN_OPS_ADD	Returns the tensor of the sum of the elements corresponding to two input tensors. Inputs: - input1: first input tensor, of the Boolean or number type. - input2: second input tensor, whose data type must be the same as that of the first tensor. Parameters: - activationType is an integer constant which is contained in FuseType. The specified activation function is called before output. Outputs: - output: sum of input1 and input2. The data shape is the same as that of the input after broadcasting, and the data type is the same as that of the input with a higher precision.
OH_NN_OPS_AVG_POOL	Apply 2D average pooling to the input tensor, which now must be in NHWC format. The int8 quantization input is supported. If the input contains the padMode parameter: Inputs: - input: tensor. Parameters: - kernelSize indicates the kernel size used to obtain the average value. It is an int array [kernel_height, kernel_width]. The first number indicates the kernel height, and the second number indicates the kernel width. - strides indicates the distance of kernel moving. The value is an int array [stride_height, stride_width]. The first number indicates the moving step in height, and the second number indicates the moving step in width. - padMode: padding mode, which is optional. The value is of the int type and can be 0 (same) or 1 (valid). The nearest neighbor value is used for padding. 0 (same): The height and width of the output are the same as those of the input. The total padding quantity is calculated horizontally and vertically and evenly distributed to the top, bottom, left, and right if possible. Otherwise, the last additional padding will be completed from the bottom and right. 1 (valid): The possible maximum height and width of the output will be returned in case of no padding. Excessive pixels will be discarded. - activationType is an integer constant which is contained in FuseType. The specified activation function is called before output. If the input contains the padList parameter: Inputs: - input: tensor. Parameters: - kernelSize indicates the kernel size used to obtain the average value. It is an int array [kernel_height, kernel_width]. The first number indicates the kernel height, and the second number indicates the kernel width. - strides indicates the distance of kernel moving. The value is an int array [stride_height, stride_width]. The first number indicates the moving step in height, and the second number indicates the moving step in width. - padList: padding around input. It is an int array [top, bottom, left, right], and the nearest neighbor values are used for padding. - activationType is an integer constant which is contained in FuseType. The specified activation function is called before output. Outputs: - output: average pooling result of the input.
OH_NN_OPS_BATCH_NORM	Batch normalization is performed on a tensor to scale and shift tensor elements, relieving potential covariate shift in a batch of data. Inputs: - input: n-dimensional tensor of shape [N, …, C]. The nth dimension is the number of channels. - scale: 1D tensor of the scaling factor used to scale the first normalized tensor. - offset: 1D tensor used to move to the first normalized tensor. - mean: 1D tensor of the overall mean value. It is used only for inference. In case of training, this parameter must be left empty. - variance: 1D tensor used for the overall variance. It is used only for inference. In case of training, this parameter must be left empty. Parameters: - epsilon: fixed small additional value. Outputs: - output: n-dimensional output tensor whose shape and data type are the same as those of the input.
OH_NN_OPS_BATCH_TO_SPACE_ND	Divides the batch dimension of a 4D tensor into small blocks by block_shape, and interleaves these blocks back into the spatial dimension. Parameters: - input: input tensor. The dimension will be divided into small blocks, and these blocks will be interleaved into the spatial dimension. Outputs: - blockSize: size of each block to be interleaved into the spatial dimension. The value is an array [height_block, width_block]. - crops: elements truncated from the spatial dimension of the output. The value is a 2D array [[crop0_start, crop0_end], [crop1_start, crop1_end]] with the shape of (2, 2). Outputs: - output. Assume that the shape of input is (n,h,w,c) and the shape of output is (n’,h’,w’,c’): n’ = n / (block_shape[0] * block_shape[1]) h’ = h * block_shape[0] - crops[0][0] - crops[0][1] w’ = w * block_shape[1] - crops[1][0] - crops[1][1] c’= c
OH_NN_OPS_BIAS_ADD	Offsets the data in each dimension of the input tensor. Inputs: - input: input tensor, which can have two to five dimensions. - bias: offset of the number of input dimensions. Outputs: - output: sum of the input tensor and the bias in each dimension.
OH_NN_OPS_CAST	Converts the data type in the input tensor. Inputs: - input: input tensor. - type: converted data type. Outputs: - output: converted tensor.
OH_NN_OPS_CONCAT	Connects tensors in a specified dimension. Inputs: - input: N input tensors. Parameters: - axis: dimension for connecting tensors. Outputs: - output: result of connecting N tensors along the axis.
OH_NN_OPS_CONV2D	2D convolutional layer. If the input contains the padMode parameter: Inputs: - input: input tensor. - weight: convolution weight in [outChannel, kernelHeight, kernelWidth, inChannel/group] format. The value of inChannel must be exactly divided by the value of group. - bias: bias of the convolution. It is an array with a length of [outChannel]. In quantization scenarios, the bias parameter does not require quantization parameters. The quantization version requires data input of the OH_NN_INT32 type. The actual quantization parameters are determined by input and weight. Parameters: - stride: movement stride of the convolution kernel in height and width. It is an int array [strideHeight, strideWidth]. - dilation: dilation size of the convolution kernel in height and width. It is an int array [dilationHeight, dilationWidth]. The value must be greater than or equal to 1 and cannot exceed the height and width of input. - padMode: padding mode of input. The value is of the int type and can be 0 (same) or 1 (valid). 0 (same): The height and width of the output are the same as those of the input. The total padding quantity is calculated horizontally and vertically and evenly distributed to the top, bottom, left, and right if possible. Otherwise, the last additional padding will be completed from the bottom and right. 1 (valid): The possible maximum height and width of the output will be returned in case of no padding. The excessive pixels will be discarded. - group: number of groups in which the input is divided by in_channel. The value is of the int type. If group is 1, it is a conventional convolution. If group is greater than 1 and less than or equal to in_channel, it is a group convolution. - activationType is an integer constant which is contained in FuseType. The specified activation function is called before output. If the input contains the padList parameter: Inputs: - input: input tensor. - weight: convolution weight in [outChannel, kernelHeight, kernelWidth, inChannel/group] format. The value of inChannel must be exactly divided by the value of group. - bias: bias of the convolution. It is an array with a length of [outChannel]. In quantization scenarios, the bias parameter does not require quantization parameters. The quantization version requires data input of the OH_NN_INT32 type. The actual quantization parameters are determined by input and weight. Parameters: - stride: movement stride of the convolution kernel in height and width. It is an int array [strideHeight, strideWidth]. - dilation: dilation size of the convolution kernel in height and width. It is an int array [dilationHeight, dilationWidth]. The value must be greater than or equal to 1 and cannot exceed the height and width of input. - padList: padding around input. It is an int array [top, bottom, left, right]. - group: number of groups in which the input is divided by in_channel. The value is of the int type. If group is 1, it is a conventional convolution. If group is in_channel, it is depthwiseConv2d. In this case, group==in_channel==out_channel. If group is greater than 1 and less than in_channel, it is a group convolution. In this case, out_channel==group. - activationType is an integer constant which is contained in FuseType. The specified activation function is called before output. Outputs: - output: convolution computing result.
OH_NN_OPS_CONV2D_TRANSPOSE	2D convolution transposition. If the input contains the padMode parameter: Inputs: - input: input tensor. - weight: convolution weight in [outChannel, kernelHeight, kernelWidth, inChannel/group] format. The value of inChannel must be exactly divided by the value of group. - bias: bias of the convolution. It is an array with a length of [outChannel]. In quantization scenarios, the bias parameter does not require quantization parameters. The quantization version requires data input of the OH_NN_INT32 type. The actual quantization parameters are determined by input and weight. - stride: movement stride of the convolution kernel in height and width. It is an int array [strideHeight, strideWidth]. Parameters: - dilation: dilation size of the convolution kernel in height and width. It is an int array [dilationHeight, dilationWidth]. The value must be greater than or equal to 1 and cannot exceed the height and width of input. - padMode: padding mode of input. The value is of the int type and can be 0 (same) or 1 (valid). 0 (same): The height and width of the output are the same as those of the input. The total padding quantity is calculated horizontally and vertically and evenly distributed to the top, bottom, left, and right if possible. Otherwise, the last additional padding will be completed from the bottom and right. 1 (valid): The possible maximum height and width of the output will be returned in case of no padding. The excessive pixels will be discarded. - group: number of groups in which the input is divided by in_channel. The value is of the int type. If group is 1, it is a conventional convolution. If group is greater than 1 and less than or equal to in_channel, it is a group convolution. - outputPads: padding along the height and width of the output tensor. The value is an int or a tuple. It can be a single integer to specify the same value for all spatial dimensions. The amount of output padding along a dimension must be less than the stride along this dimension. - activationType is an integer constant which is contained in FuseType. The specified activation function is called before output. If the input contains the padList parameter: Inputs: - input: input tensor. - weight: convolution weight in [outChannel, kernelHeight, kernelWidth, inChannel/group] format. The value of inChannel must be exactly divided by the value of group. - bias: bias of the convolution. It is an array with a length of [outChannel]. In quantization scenarios, the bias parameter does not require quantization parameters. The quantization version requires data input of the OH_NN_INT32 type. The actual quantization parameters are determined by input and weight. Parameters: - stride: movement stride of the convolution kernel in height and width. It is an int array [strideHeight, strideWidth]. - dilation: dilation size of the convolution kernel in height and width. It is an int array [dilationHeight, dilationWidth]. The value must be greater than or equal to 1 and cannot exceed the height and width of input. - padList: padding around input. It is an int array [top, bottom, left, right]. - group: number of groups in which the input is divided by in_channel. The value is of the int type. If group is 1, it is a conventional convolution. If group is greater than 1 and less than or equal to in_channel, it is a group convolution. - outputPads: padding along the height and width of the output tensor. The value is an int or a tuple. It can be a single integer to specify the same value for all spatial dimensions. The amount of output padding along a dimension must be less than the stride along this dimension. - activationType is an integer constant which is contained in FuseType. The specified activation function is called before output. Outputs: output: computing result after convolution and transposition.
OH_NN_OPS_DEPTHWISE_CONV2D_NATIVE	2D depthwise separable convolution. If the input contains the padMode parameter: Inputs: - input: input tensor. - weight: convolution weight in [outChannel, kernelHeight, kernelWidth, 1] format. outChannel is equal to channelMultiplier multiplied by inChannel. - bias: bias of the convolution. It is an array with a length of [outChannel]. In quantization scenarios, the bias parameter does not require quantization parameters. The quantization version requires data input of the OH_NN_INT32 type. The actual quantization parameters are determined by input and weight. Parameters: - stride: movement stride of the convolution kernel in height and width. It is an int array [strideHeight, strideWidth]. - dilation: dilation size of the convolution kernel in height and width. It is an int array [dilationHeight, dilationWidth]. The value must be greater than or equal to 1 and cannot exceed the height and width of input. - padMode: padding mode of input. The value is of the int type and can be 0 (same) or 1 (valid). 0 (same): The height and width of the output are the same as those of the input. The total padding quantity is calculated horizontally and vertically and evenly distributed to the top, bottom, left, and right if possible. Otherwise, the last additional padding will be completed from the bottom and right. 1 (valid): The possible maximum height and width of the output will be returned in case of no padding. The excessive pixels will be discarded. - activationType is an integer constant which is contained in FuseType. The specified activation function is called before output. If the input contains the padList parameter: Inputs: - input: input tensor. - weight: convolution weight in [outChannel, kernelHeight, kernelWidth, 1] format. outChannel is equal to channelMultiplier multiplied by inChannel. - bias: bias of the convolution. It is an array with a length of [outChannel]. In quantization scenarios, the bias parameter does not require quantization parameters. The quantization version requires data input of the OH_NN_INT32 type. The actual quantization parameters are determined by input and weight. Parameters: - stride: movement stride of the convolution kernel in height and width. It is an int array [strideHeight, strideWidth]. - dilation: dilation size of the convolution kernel in height and width. It is an int array [dilationHeight, dilationWidth]. The value must be greater than or equal to 1 and cannot exceed the height and width of input. - padList: padding around input. It is an int array [top, bottom, left, right]. - activationType is an integer constant which is contained in FuseType. The specified activation function is called before output. Outputs: - output: convolution computing result.
OH_NN_OPS_DIV	Divides two input scalars or tensors. Inputs: - input1: first input, which is a number, a bool, or a tensor whose data type is number or Boolean. - input2: second input, which must meet the following requirements: If the first input is a tensor, the second input can be a real number, a Boolean value, or a tensor whose data type is real number or Boolean value. If the first input is a real number or Boolean value, the second input must be a tensor whose data type is real number or Boolean value. Parameters: - activationType is an integer constant which is contained in FuseType. The specified activation function is called before output. Outputs: - output: result of dividing input1 by input2.
OH_NN_OPS_ELTWISE	Sets parameters to perform product (dot product), sum (addition and subtraction), or max (larger value) on the input. Inputs: - input1: first input tensor. - input2: second input tensor. Parameters: - mode: operation mode. The value is an enumerated value. Outputs: - output: computing result, which has the same data type and shape of output and input1.
OH_NN_OPS_EXPAND_DIMS	Adds an additional dimension to a tensor in the given dimension. Inputs: - input: input tensor. - axis: index of the dimension to be added. The value is of the int32_t type and must be a constant in the range [-dim-1, dim]. Outputs: - output: tensor after dimension expansion.
OH_NN_OPS_FILL	Creates a tensor of the specified dimensions and fills it with a scalar. Inputs: - value: scalar used to fill the tensor. - shape: dimensions of the tensor to be created. Outputs: - output: generated tensor, which has the same data type as value. The tensor shape is specified by the shape parameter.
OH_NN_OPS_FULL_CONNECTION	Full connection. The entire input is used as the feature map for feature extraction. Inputs: - input: full-connection input tensor. - weight: weight tensor for a full connection. - bias: full-connection bias. In quantization scenarios, no quantized parameter is required for this parameter. If quantization is required, the data must be of the OH_NN_INT32 type. The actual quantization parameters are determined by input and weight. Parameters: - activationType is an integer constant which is contained in FuseType. The specified activation function is called before output. Outputs: - output: computed tensor. If the input contains the axis parameter: Inputs: - input: full-connection input tensor. - weight: weight tensor for a full connection. - bias: full-connection bias. In quantization scenarios, no quantized parameter is required for this parameter. If quantization is required, the data must be of the OH_NN_INT32 type. The actual quantization parameters are determined by input and weight. Parameters: - axis: axis in which the full connection is applied. The specified axis and its following axes are converted into a 1D tensor for applying the full connection. - activationType is an integer constant which is contained in FuseType. The specified activation function is called before output. Outputs: - output: computed tensor.
OH_NN_OPS_GATHER	Returns the slice of the input tensor based on the specified index and axis. Inputs: - input: tensor to be sliced. - inputIndices: indices of the specified input on the axis. The value is an array of the int type and must be in the range [0,input.shape[axis]). - axis: axis on which input is sliced. The value is an array with one element of the int32_t type. Outputs: - output: sliced tensor.
OH_NN_OPS_HSWISH	Calculate the Hswish activation value of the input. Inputs: - An n-dimensional input tensor. Outputs: - n-dimensional Hswish activation value. The data type is the same as that of shape and input.
OH_NN_OPS_LESS_EQUAL	For input1 and input2, calculate the result of input1[i]<=input2[i] for each pair of elements, where i is the index of each element in the input tensor. Inputs: - input1, which can be a real number, Boolean value, or tensor whose data type is real number or NN_BOOL. - input2, which can be a real number or a Boolean value if input1 is a tensor and must be a tensor with the data type of real number or NN_BOOL if input1 is not a tensor. Outputs: - A tensor of the data type NN_BOOL. When a quantization model is used, the quantization parameters of the output cannot be omitted. However, values of the quantization parameters do not affect the result.
OH_NN_OPS_MATMUL	Calculate the inner product of input1 and input2. Inputs: - input1: n-dimensional input tensor. - input2: n-dimensional input tensor. Parameters: - TransposeX: Boolean value indicating whether to transpose input1. - TransposeY: Boolean value indicating whether to transpose input2. Outputs: - output: inner product obtained after calculation. In case of type!=NN_UNKNOWN, the output data type is determined by type. In case of type==NN_UNKNOWN, the output data type depends on the data type converted during computing of inputX and inputY.
OH_NN_OPS_MAXIMUM	Calculates the maximum of input1 and input2 element-wise. The inputs of input1 and input2 comply with the implicit type conversion rules to make the data types consistent. * The inputs must be two tensors or one tensor and one scalar. When the inputs are two tensors, their data types cannot be both NN_BOOL. Their shapes can be broadcast to the same size. When the inputs are one tensor and one scalar, the scalar must be a constant. Inputs: - input1: n-dimensional input tensor of the real number or NN_BOOL type. - input2: n-dimensional input tensor of the real number or NN_BOOL type. Outputs: - output: n-dimensional output tensor. The shape and data type of output are the same as those of the two inputs with a higher precision.
OH_NN_OPS_MAX_POOL	Applies 2D maximum pooling to the input tensor. If the input contains the padMode parameter: Inputs: - input: tensor. Parameters: - kernelSize: kernel size used to obtain the maximum. It is an int array [kernel_height, kernel_width]. The first number indicates the kernel height, and the second number indicates the kernel width. - strides indicates the distance of kernel moving. The value is an int array [stride_height, stride_width]. The first number indicates the moving step in height, and the second number indicates the moving step in width. - padMode: padding mode, which is optional. The value is of the int type and can be 0 (same) or 1 (valid). The nearest neighbor value is used for padding. 0 (same): The height and width of the output are the same as those of the input. The total padding quantity is calculated horizontally and vertically and evenly distributed to the top, bottom, left, and right if possible. Otherwise, the last additional padding will be completed from the bottom and right. 1 (valid): The possible maximum height and width of the output will be returned in case of no padding. The excessive pixels will be discarded. - activationType is an integer constant which is contained in FuseType. The specified activation function is called before output. If the input contains the padList parameter: Inputs: - input: tensor. Parameters: - kernelSize: kernel size used to obtain the maximum. It is an int array [kernel_height, kernel_width]. The first number indicates the kernel height, and the second number indicates the kernel width. - strides indicates the distance of kernel moving. The value is an int array [stride_height, stride_width]. The first number indicates the moving step in height, and the second number indicates the moving step in width. - padList: padding around input. It is an int array [top, bottom, left, right], and the nearest neighbor values are used for padding. - activationType is an integer constant which is contained in FuseType. The specified activation function is called before output. Outputs: - output: tensor obtained after maximum pooling is applied to the input.
OH_NN_OPS_MUL	Multiplies elements in the same positions of inputX and inputY to obtain the output. If inputX and inputY have different shapes, expand them to the same shape through broadcast and then perform multiplication. Inputs: - input1: n-dimensional tensor. - input2: n-dimensional tensor. Parameters: - activationType is an integer constant which is contained in FuseType. The specified activation function is called before output. Outputs: - Product of each element of input1 and input2.
OH_NN_OPS_ONE_HOT	Generates a one-hot tensor based on the positions specified by indices. The positions specified by indices are determined by on_value, and other positions are determined by off_value. Inputs: - indices: n-dimensional tensor. Each element in indices determines the position of on_value in each one-hot vector. - depth: integer scalar that determines the depth of the one-hot vector. The value of depth must be greater than 0. - on_value: scalar that specifies a valid value in the one-hot vector. - off_value: scalar that specifies the values of other posistions in the one-hot vector except the valid value. Parameters: - axis: integer scalar that specifies the dimension for inserting the one-hot. Assume that the shape of indices is [N, C], and the value of depth is D. When axis is 0, the shape of the output is [D, N, C]. When axis is -1, the shape of the output is [N, C, D]. When axis is 1, the shape of the output is [N, D, C]. Outputs: - output: (n+1)-dimensional tensor if indices is an n-dimensional tensor. The output shape is determined by indices and axis.
OH_NN_OPS_PAD	Pads inputX in the specified dimensions. Inputs: - inputX: n-dimensional tensor in [BatchSize, …] format. - paddings: 2D tensor that specifies the length to pad in each dimension. The shape is [n, 2]. For example, paddings[i][0] indicates the number of paddings to be added preceding inputX in the ith dimension. paddings[i][1] indicates the number of paddings to be added following inputX in the ith dimension. Parameters: - padValues: value to be added to the pad operation. The value is a constant with the same data type as inputX. Outputs: - output: n-dimensional tensor after padding, with the same dimensions and data type as inputX. The shape is determined by inputX and paddings. output.shape[i] = input.shape[i] + paddings[i][0]+paddings[i][1]
OH_NN_OPS_POW	Calculates the y power of each element in input. The inputs must be two tensors or one tensor and one scalar. When the inputs are two tensors, their data types cannot be both NN_BOOL, and their shapes must be the same. When the inputs are one tensor and one scalar, the scalar must be a constant. Inputs: - input: real number, Boolean value, or tensor whose data type is real number or NN_BOOL. - y: real number, Boolean value, or tensor whose data type is real number or NN_BOOL. Outputs: - output: tensor, whose shape is determined by the shape of input and y after broadcasting.
OH_NN_OPS_SCALE	Scales a tensor. Inputs: - input: n-dimensional tensor. - scale: scaling tensor. - bias: bias tensor. Parameters: - axis: dimensions to be scaled. - activationType is an integer constant which is contained in FuseType. The specified activation function is called before output. Outputs: - output: scaled n-dimensional tensor, whose data type is the same as that of input and shape is determined by axis.
OH_NN_OPS_SHAPE	Calculates the shape of the input tensor. Inputs: - input: n-dimensional tensor. Outputs: - output: integer array representing the dimensions of the input tensor.
OH_NN_OPS_SIGMOID	Applies the sigmoid operation to the input tensor. Inputs: - input: n-dimensional tensor. Outputs: - output: result of the sigmoid operation. It is an n-dimensional tensor with the same data type and shape as input.
OH_NN_OPS_SLICE	Slices a tensor of the specified size from the input in each dimension. Inputs: - input: n-dimensional input tensor. - begin: start of the slice, which is an array of integers greater than or equal to 0. - size: slice length, which is an array of integers greater than or equal to 0. Assume that a dimension is i and 1<=size[i]<=input.shape[i]-begin[i]. Outputs: - output: n-dimensional tensor obtained by slicing. The TensorType, shape, and size of the output are the same as those of the input.
OH_NN_OPS_SOFTMAX	Applies the softmax operation to the input tensor. Inputs: - input: n-dimensional input tensor. Parameters: - axis: dimension in which the softmax operation is performed. The value is of the int64 type. It is an integer in the range [-n, n). Outputs: - output: result of the softmax operation. It is an n-dimensional tensor with the same data type and shape as input.
OH_NN_OPS_SPACE_TO_BATCH_ND	Divides a 4D tensor into small blocks and combines these blocks in the original batch. The number of blocks is blockShape[0] multiplied by blockShape[1]. Inputs: - input: 4D tensor. Parameters: - blockShape: a pair of integers. Each of them is greater than or equal to 1. - paddings: a pair of arrays. Each of them consists of two integers. The four integers that form paddings must be greater than or equal to 0. paddings[0][0] and paddings[0][1] specify the number of paddings in the third dimension, and paddings[1][0] and paddings[1][1] specify the number of paddings in the fourth dimension. Outputs: - output: 4D tensor with the same data type as input. The shape is determined by input, blockShape, and paddings. Assume that the input shape is [n,c,h,w], then: output.shape[0] = n * blockShape[0] * blockShape[1] output.shape[1] = c output.shape[2] = (h + paddings[0][0] + paddings[0][1]) / blockShape[0] output.shape[3] = (w + paddings[1][0] + paddings[1][1]) / blockShape1 and (w + paddings[1][0] + paddings[1][1]) is exactly divisible by (h + paddings[0][0] + paddings[0][1]) and (w + paddings[1][0] + paddings[1][1]).
OH_NN_OPS_SPLIT	Splits the input into multiple tensors along the axis dimension. The number of tensors is specified by outputNum. Inputs: - input: n-dimensional tensor. Parameters: - outputNum: number of output tensors. The data type is long. - size_splits: size of each tensor split from the input. The value is a 1D tensor of the int type. If size_splits is empty, the input will be evenly split into tensors of the same size. In this case, input.shape[axis] can be exactly divisible by outputNum. If size_splits is not empty, the sum of all its elements must be equal to input.shape[axis]. - axis: splitting dimension of the int type. Outputs: - outputs: array of n-dimensional tensors, with the same data type and dimensions. The data type of each tensor is the same as that of input.
OH_NN_OPS_SQRT	Calculates the square root of a tensor. Inputs: - input: n-dimensional tensor. Outputs: - output: square root of the input. It is an n-dimensional tensor with the same data type and shape as input.
OH_NN_OPS_SQUARED_DIFFERENCE	Calculates the square of the difference between two tensors. The SquaredDifference operator supports tensor and tensor subtraction. If two tensors have different TensorTypes, the Sub operator converts the low-precision tensor to a high-precision one. If two tensors have different shapes, the two tensors can be extended to tensors with the same shape through broadcast. Inputs: - input1: minuend, which is a tensor of the NN_FLOAT16, NN_FLOAT32, NN_INT32, or NN_BOOL type. - input2: subtrahend, which is a tensor of the NN_FLOAT16, NN_FLOAT32, NN_INT32, or NN_BOOL type. Outputs: - output: square of the difference between two inputs. The output shape is determined byinput1 and input2. If they have the same shape, the output tensor has the same shape as them. If they have different shapes, perform the broadcast operation on input1 and input2 and perform subtraction. TensorType of the output is the same as that of the input tensor with higher precision.
OH_NN_OPS_SQUEEZE	Removes the dimension with a length of 1 from the specified axis. The int8 quantization input is supported. Assume that the input shape is [2, 1, 1, 2, 2] and axis is [0,1], the output shape is [2, 1, 2, 2], which means the dimension whose length is 0 between dimensions 0 and dimension 1 is removed. Inputs: - input: n-dimensional tensor. Parameters: - axis: dimension to be removed. The value is of int64_t type and can be an integer in the range [-n, n) or an array. Outputs: - output: output tensor.
OH_NN_OPS_STACK	Stacks multiple tensors along the specified axis. If each tensor has n dimensions before stacking, the output tensor will have n+1 dimensions. Inputs: - input: input for stacking, which can contain multiple n-dimensional tensors. Each of them must have the same shape and type. Parameters: - axis: dimension for tensor stacking, which is an integer. The value range is [-(n+1),(n+1)), which means a negative number is allowed. Outputs: - output: stacking result of the input along the axis dimension. The value is an n+1-dimensional tensor and has the same TensorType as the input.
OH_NN_OPS_STRIDED_SLICE	Slices a tensor with the specified stride. Inputs: - input: n-dimensional input tensor. - begin: start of slicing, which is a 1D tensor. The length of begin is n. begin[i] specifies the start of slicing in the ith dimension. - end: end of slicing, which is a 1D tensor. The length of end is n. end[i] specifies the end of slicing in the ith dimension. - strides: slicing stride, which is a 1D tensor. The length of strides is n. strides[i] specifies the stride at which the tensor is sliced in the ith dimension. Parameters: - beginMask: an integer used to mask begin. beginMask is represented in binary code. In case of binary(beginMask)[i]==1, for the ith dimension, elements are sliced from the first element at strides[i] until the end[i]-1 element. - endMask: an integer used to mask end. endMask is represented in binary code. In case of binary(endMask)[i]==1, elements are sliced from the element at the begin[i] position in the ith dimension until the tensor boundary at strides[i]. - ellipsisMask: integer used to mask begin and end. ellipsisMask is represented in binary code. In case of binary(ellipsisMask)[i]==1, elements are sliced from the first element at strides[i] in the ith dimension until the tensor boundary. Only one bit of binary(ellipsisMask) can be a non-zero value. - newAxisMask: new dimension, which is an integer. newAxisMask is represented in binary code. In case of binary(newAxisMask)[i]==1, a new dimension whose length is 1 is inserted into the ith dimension. - shrinkAxisMask: shrinking dimension, which is an integer. * shrinkAxisMask is represented in binary code. In the case of binary(shrinkAxisMask)[i]==1, all elements in the ith dimension will be discarded, and the length of the ith dimension is shrunk to 1. Outputs: - A tensor, with the same data type as input. The number of dimensions of the output tensor is rank(input[0])+1.
OH_NN_OPS_SUB	Calculates the difference between two tensors. Inputs: - input1: minuend, which is a tensor. - input2: subtrahend, which is a tensor. Parameters: - activationType is an integer constant which is contained in FuseType. The specified activation function is called before output. Outputs: - output: difference between the two tensors. The output shape is determined byinput1 and input2. If they have the same shape, the output tensor has the same shape as them. If they have different shapes, perform the broadcast operation on input1 and input2 and perform subtraction. TensorType of the output is the same as that of the input tensor with higher precision.
OH_NN_OPS_TANH	Computes hyperbolic tangent of the input tensor. Inputs: - input: n-dimensional tensor. Outputs: - output: hyperbolic tangent of the input. The TensorType and tensor shape are the same as those of the input.
OH_NN_OPS_TILE	Copies a tensor the specified times. Inputs: - input: n-dimensional tensor. - multiples: number of times that the input tensor is copied in each dimension. The value is a 1D tensor. The length m is not less than the number of dimensions, that is, n. Outputs: - An m-dimensional tensor whose TensorType is the same as that of the input. If input and multiples have the same length, input and output have the same number of dimensions. If the length of multiples is greater than n, 1 is used to fill the input dimension, and then the input is copied in each dimension the specified times to obtain the m-dimensional tensor.
OH_NN_OPS_TRANSPOSE	Transposes data of input 0 based on permutation. Inputs: - input: n-dimensional tensor to be transposed. - permutation: The value is a 1D tensor whose length is the same as the number of dimensions of input 0. Outputs: - output: n-dimensional tensor. TensorType of output 0 is the same as that of input 0, and the output shape is determined by the shape and permutation of input 0.
OH_NN_OPS_REDUCE_MEAN	Calculates the average value in the specified dimension. If keepDims is set to false, the number of dimensions is reduced for the input; if keepDims is set to true, the number of dimensions is retained. Inputs: - input: n-dimensional input tensor, where n is less than 8. - axis: dimension used to calculate the average value. The value is a 1D tensor. The value range of each element in axis is [–n, n). Parameters: - keepDims: indicates whether to retain the dimension. The value is a Boolean value. Outputs: - output: m-dimensional output tensor whose data type is the same as that of the input. If keepDims is false, m==n. If keepDims is true, m<n.
OH_NN_OPS_RESIZE_BILINEAR	The Bilinear method is used to deform the input based on the given parameters. Inputs: - input: 4D input tensor. Each element in the input cannot be less than 0. The input layout must be [batchSize, height, width, channels]. Parameters: - newHeight: resized height of the 4D tensor. - newWidth: resized width of the 4D tensor. - preserveAspectRatio: indicates whether to maintain the height/width ratio of input after resizing. - coordinateTransformMode: coordinate transformation method used by the resize operation. The value is an int32 integer. Currently, the following methods are supported: - excludeOutside: an int64 floating point number. When its value is 1, the sampling weight of the part that exceeds the boundary of input is set to 0, and other weights are normalized. Outputs: - output: n-dimensional tensor, with the same shape and data type as input.
OH_NN_OPS_RSQRT	Calculates the reciprocal of the square root of a tensor. Inputs: - input: n-dimensional tensor, where n is less than 8. Each element of the tensor cannot be less than 0. Outputs: - output: n-dimensional tensor, with the same shape and data type as input.
OH_NN_OPS_RESHAPE	Reshapes a tensor. Inputs: - input: n-dimensional input tensor. - InputShape: shape of the output tensor. The value is a 1D constant tensor. Outputs: - output: tensor whose data type is the same as that of input and shape is determined by InputShape.
OH_NN_OPS_PRELU	Calculates the PReLU activation value of input and weight. Inputs: - input: n-dimensional tensor. If n is greater than or equal to 2, inputX must be [BatchSize, …, Channels]. The second dimension is the number of channels. - weight: 1D tensor. The length of weight must be 1 or equal to the number of channels. If the length of weight is 1, all channels share the same weight. If the length of weight is equal to the number of channels, each channel exclusively has a weight. If n is less than 2 for inputX, the weight length must be 1. Outputs: output: PReLU activation value of x, with the same shape and data type as inputX.
OH_NN_OPS_RELU	Calculates the Relu activation value of input. Inputs: - input: n-dimensional input tensor. Outputs: - output: n-dimensional tensor, with the same data type and shape as the input tensor.
OH_NN_OPS_RELU6	Calculates the Relu6 activation value of the input, that is, calculate min(max(x, 0), 6) for each element x in the input. Inputs: - input: n-dimensional input tensor. Outputs: - output: n-dimensional Relu6 tensor, with the same data type and shape as the input tensor.
OH_NN_OPS_LAYER_NORM	Applies layer normalization for a tensor from the specified axis. Inputs: - input: n-dimensional input tensor. - gamma: m-dimensional tensor. The dimensions of gamma must be the same as the shape of the part of the input tensor to normalize. - beta: m-dimensional tensor with the same shape as gamma. Parameters: - beginAxis is an NN_INT32 scalar that specifies the axis from which normalization starts. The value range is [1, rank(input)). - epsilon is a scalar of NN_FLOAT32. It is a tiny amount in the normalization formula. The common value is 1e-7. Outputs: - output: n-dimensional tensor, with the same data type and shape as the input tensor.
OH_NN_OPS_REDUCE_PROD	Calculates the accumulated value for a tensor along the specified dimension. Inputs: - input: n-dimensional input tensor, where n is less than 8. - axis: dimension used to calculate the product. The value is a 1D tensor. The value range of each element in axis is [–n, n). Parameters: - keepDims: indicates whether to retain the dimension. The value is a Boolean value. When its value is true, the number of output dimensions is the same as that of the input. When its value is false, the number of output dimensions is reduced. Outputs: - output: m-dimensional output tensor whose data type is the same as that of the input. If keepDims is false, m==n. If keepDims is true, m<n.
OH_NN_OPS_REDUCE_ALL	Operates the logical OR in the specified dimension. If keepDims is set to false, the number of dimensions is reduced for the input; if keepDims is set to true, the number of dimensions is retained. Inputs: - A n-dimensional input tensor, where n is less than 8. - A 1D tensor specifying the dimension used to operate the logical OR. The value range of each element in axis is [–n, n). Parameters: - keepDims: indicates whether to retain the dimension. The value is a Boolean value. Outputs: - output: m-dimensional output tensor whose data type is the same as that of the input. If keepDims is false, m==n. If keepDims is true, m<n.
OH_NN_OPS_QUANT_DTYPE_CAST	Converts the data type. Inputs: - input: n-dimensional tensor. Parameters: - src_t: data type of the input. - dst_t: data type of the output. Outputs: - output: n-dimensional tensor. The data type is determined by input2. The output shape is the same as the input shape.
OH_NN_OPS_TOP_K	Obtains the values and indices of the largest k entries in the last dimension. Inputs: - input: n-dimensional tensor. - input k: first k records of data and their indices. Parameters: - sorted: order of sorting. The value true means descending and false means ascending. Outputs: - output0: largest k elements in each slice of the last dimension. - output1: index of the value in the last dimension of the input.
OH_NN_OPS_ARG_MAX	Returns the index of the maximum tensor value across axes. Inputs: - input: n-dimensional tensor (N, ∗), where ∗ means any number of additional dimensions. Parameters: - axis: dimension for calculating the index of the maximum. - keep_dims: indicates whether to maintain the input tensor dimension. The value is a Boolean value. Outputs: - output: index of the maximum input tensor on the axis. The value is a tensor.
OH_NN_OPS_UNSQUEEZE	Adds a dimension based on the value of axis. Inputs: - input: n-dimensional tensor. Parameters: - axis: dimension to be added. The value of axis can be an integer or an array of integers. The value range of the integer is [-n, n). Outputs: - output: output tensor.
OH_NN_OPS_GELU	Gaussian error linear unit activation function. The int quantization input is not supported. output=0.5∗input∗(1+tanh(input/2)) Inputs: - An n-dimensional input tensor. Outputs: - output: n-dimensional tensor, with the same data type and shape as the input tensor.

OH_NN_PerformanceMode

enum OH_NN_PerformanceMode

Description
Defines the hardware performance mode.

Name	Description
OH_NN_PERFORMANCE_NONE	No performance mode preference
OH_NN_PERFORMANCE_LOW	Low power consumption mode
OH_NN_PERFORMANCE_MEDIUM	Medium performance mode
OH_NN_PERFORMANCE_HIGH	High performance mode
OH_NN_PERFORMANCE_EXTREME	Ultimate performance mode

OH_NN_Priority

enum OH_NN_Priority

Description
Defines the model inference task priority.

Name	Description
OH_NN_PRIORITY_NONE	No priority preference
OH_NN_PRIORITY_LOW	Low priority
OH_NN_PRIORITY_MEDIUM	Medium priority
OH_NN_PRIORITY_HIGH	High priority

OH_NN_ReturnCode

enum OH_NN_ReturnCode

Description
Defines error codes for Neural Network Runtime.

Name	Description
OH_NN_SUCCESS	The operation is successful.
OH_NN_FAILED	The operation failed.
OH_NN_INVALID_PARAMETER	Invalid parameter.
OH_NN_MEMORY_ERROR	Memory-related error, for example, insufficient memory, memory data copy failure, or memory application failure.
OH_NN_OPERATION_FORBIDDEN	Invalid operation.
OH_NN_NULL_PTR	Null pointer exception
OH_NN_INVALID_FILE	Invalid file.
OH_NN_UNAVALIDABLE_DEVICE	A hardware error occurs, for example, HDL service crash.
OH_NN_INVALID_PATH	Invalid path.

OH_NN_TensorType

enum OH_NN_TensorType

Description
Enumerates the tensor data types.

Tensors are usually used to set the input, output, and operator parameters of a model. When a tensor is used as the input or output of a model (or operator), set the tensor type to OH_NN_TENSOR. When the tensor is used as an operator parameter, select an enumerated value other than OH_NN_TENSOR as the tensor type. Assume that the pad parameter of the OH_NN_OPS_CONV2D operator is being set. You need to set the type attribute of the OH_NN_Tensor instance to OH_NN_CONV2D_PAD. The settings of other operator parameters are similar. The enumerated values are named in the format OHNN{Operator name}_{Attribute name}.

Name	Description
OH_NN_TENSOR	This enumerated value is used when the tensor is used as the input or output of a model (or operator).
OH_NN_ADD_ACTIVATIONTYPE	This enumerated value is used when the tensor is used as the activationType parameter of the Add operator.
OH_NN_AVG_POOL_KERNEL_SIZE	This enumerated value is used when the tensor is used as the kernel_size parameter of the AvgPool operator.
OH_NN_AVG_POOL_STRIDE	This enumerated value is used when the tensor is used as the stride parameter of the AvgPool operator.
OH_NN_AVG_POOL_PAD_MODE	This enumerated value is used when the tensor is used as the pad_mode parameter of the AvgPool operator.
OH_NN_AVG_POOL_PAD	This enumerated value is used when the tensor is used as the pad parameter of the AvgPool operator.
OH_NN_AVG_POOL_ACTIVATION_TYPE	This enumerated value is used when the tensor is used as the activation_type parameter of the AvgPool operator.
OH_NN_BATCH_NORM_EPSILON	This enumerated value is used when the tensor is used as the eosilon parameter of the BatchNorm operator.
OH_NN_BATCH_TO_SPACE_ND_BLOCKSIZE	This enumerated value is used when the tensor is used as the blockSize parameter of the BatchToSpaceND operator.
OH_NN_BATCH_TO_SPACE_ND_CROPS	This enumerated value is used when the tensor is used as the crops parameter of the BatchToSpaceND operator.
OH_NN_CONCAT_AXIS	This enumerated value is used when the tensor is used as the axis parameter of the Concat operator.
OH_NN_CONV2D_STRIDES	This enumerated value is used when the tensor is used as the strides parameter of the Conv2D operator.
OH_NN_CONV2D_PAD	This enumerated value is used when the tensor is used as the pad parameter of the Conv2D operator.
OH_NN_CONV2D_DILATION	This enumerated value is used when the tensor is used as the dilation parameter of the Conv2D operator.
OH_NN_CONV2D_PAD_MODE	This enumerated value is used when the tensor is used as the padMode parameter of the Conv2D operator.
OH_NN_CONV2D_ACTIVATION_TYPE	This enumerated value is used when the tensor is used as the activationType parameter of the Conv2D operator.
OH_NN_CONV2D_GROUP	This enumerated value is used when the tensor is used as the group parameter of the Conv2D operator.
OH_NN_CONV2D_TRANSPOSE_STRIDES	This enumerated value is used when the tensor is used as the strides parameter of the Conv2DTranspose operator.
OH_NN_CONV2D_TRANSPOSE_PAD	This enumerated value is used when the tensor is used as the pad parameter of the Conv2DTranspose operator.
OH_NN_CONV2D_TRANSPOSE_DILATION	This enumerated value is used when the tensor is used as the dilation parameter of the Conv2DTranspose operator.
OH_NN_CONV2D_TRANSPOSE_OUTPUT_PADDINGS	This enumerated value is used when the tensor is used as the outputPaddings parameter of the Conv2DTranspose operator.
OH_NN_CONV2D_TRANSPOSE_PAD_MODE	This enumerated value is used when the tensor is used as the padMode parameter of the Conv2DTranspose operator.
OH_NN_CONV2D_TRANSPOSE_ACTIVATION_TYPE	This enumerated value is used when the tensor is used as the activationType parameter of the Conv2DTranspose operator.
OH_NN_CONV2D_TRANSPOSE_GROUP	This enumerated value is used when the tensor is used as the group parameter of the Conv2DTranspose operator.
OH_NN_DEPTHWISE_CONV2D_NATIVE_STRIDES	This enumerated value is used when the tensor is used as the strides parameter of the DepthwiseConv2dNative operator.
OH_NN_DEPTHWISE_CONV2D_NATIVE_PAD	This enumerated value is used when the tensor is used as the pad parameter of the DepthwiseConv2dNative operator.
OH_NN_DEPTHWISE_CONV2D_NATIVE_DILATION	This enumerated value is used when the tensor is used as the dilation parameter of the DepthwiseConv2dNative operator.
OH_NN_DEPTHWISE_CONV2D_NATIVE_PAD_MODE	This enumerated value is used when the tensor is used as the padMode parameter of the DepthwiseConv2dNative operator.
OH_NN_DEPTHWISE_CONV2D_NATIVE_ACTIVATION_TYPE	This enumerated value is used when the tensor is used as the activationType parameter of the DepthwiseConv2dNative operator.
OH_NN_DIV_ACTIVATIONTYPE	This enumerated value is used when the tensor is used as the activationType parameter of the Div operator.
OH_NN_ELTWISE_MODE	This enumerated value is used when the tensor is used as the mode parameter of the Eltwise operator.
OH_NN_FULL_CONNECTION_AXIS	This enumerated value is used when the tensor is used as the axis parameter of the FullConnection operator.
OH_NN_FULL_CONNECTION_ACTIVATIONTYPE	This enumerated value is used when the tensor is used as the activationType parameter of the FullConnection operator.
OH_NN_MATMUL_TRANSPOSE_A	This enumerated value is used when the tensor is used as the transposeA parameter of the Matmul operator.
OH_NN_MATMUL_TRANSPOSE_B	This enumerated value is used when the tensor is used as the transposeB parameter of the Matmul operator.
OH_NN_MATMUL_ACTIVATION_TYPE	This enumerated value is used when the tensor is used as the activationType parameter of the Matmul operator.
OH_NN_MAX_POOL_KERNEL_SIZE	This enumerated value is used when the tensor is used as the kernel_size parameter of the MaxPool operator.
OH_NN_MAX_POOL_STRIDE	This enumerated value is used when the tensor is used as the stride parameter of the MaxPool operator.
OH_NN_MAX_POOL_PAD_MODE	This enumerated value is used when the tensor is used as the pad_mode parameter of the MaxPool operator.
OH_NN_MAX_POOL_PAD	This enumerated value is used when the tensor is used as the pad parameter of the MaxPool operator.
OH_NN_MAX_POOL_ACTIVATION_TYPE	This enumerated value is used when the tensor is used as the activation_type parameter of the MaxPool operator.
OH_NN_MUL_ACTIVATION_TYPE	This enumerated value is used when the tensor is used as the activationType parameter of the Mul operator.
OH_NN_ONE_HOT_AXIS	This enumerated value is used when the tensor is used as the axis parameter of the OneHot operator.
OH_NN_PAD_CONSTANT_VALUE	This enumerated value is used when the tensor is used as the constant_value parameter of the Pad operator.
OH_NN_SCALE_ACTIVATIONTYPE	This enumerated value is used when the tensor is used as the activationType parameter of the Scale operator.
OH_NN_SCALE_AXIS	This enumerated value is used when the tensor is used as the axis parameter of the Scale operator.
OH_NN_SOFTMAX_AXIS	This enumerated value is used when the tensor is used as the axis parameter of the Softmax operator.
OH_NN_SPACE_TO_BATCH_ND_BLOCK_SHAPE	This enumerated value is used when the tensor is used as the BlockShape parameter of the SpaceToBatchND operator.
OH_NN_SPACE_TO_BATCH_ND_PADDINGS	This enumerated value is used when the tensor is used as the Paddings parameter of the SpaceToBatchND operator.
OH_NN_SPLIT_AXIS	This enumerated value is used when the tensor is used as the Axis parameter of the Split operator.
OH_NN_SPLIT_OUTPUT_NUM	This enumerated value is used when the tensor is used as the OutputNum parameter of the Split operator.
OH_NN_SPLIT_SIZE_SPLITS	This enumerated value is used when the tensor is used as the SizeSplits parameter of the Split operator.
OH_NN_SQUEEZE_AXIS	This enumerated value is used when the tensor is used as the Axis parameter of the Squeeze operator.
OH_NN_STACK_AXIS	This enumerated value is used when the tensor is used as the Axis parameter of the Stack operator.
OH_NN_STRIDED_SLICE_BEGIN_MASK	This enumerated value is used when the tensor is used as the BeginMask parameter of the StridedSlice operator.
OH_NN_STRIDED_SLICE_END_MASK	This enumerated value is used when the tensor is used as the EndMask parameter of the StridedSlice operator.
OH_NN_STRIDED_SLICE_ELLIPSIS_MASK	This enumerated value is used when the tensor is used as the EllipsisMask parameter of the StridedSlice operator.
OH_NN_STRIDED_SLICE_NEW_AXIS_MASK	This enumerated value is used when the tensor is used as the NewAxisMask parameter of the StridedSlice operator.
OH_NN_STRIDED_SLICE_SHRINK_AXIS_MASK	This enumerated value is used when the tensor is used as the ShrinkAxisMask parameter of the StridedSlice operator.
OH_NN_SUB_ACTIVATIONTYPE	This enumerated value is used when the tensor is used as the ActivationType parameter of the Sub operator.
OH_NN_REDUCE_MEAN_KEEP_DIMS	This enumerated value is used when the tensor is used as the keep_dims parameter of the ReduceMean operator.
OH_NN_RESIZE_BILINEAR_NEW_HEIGHT	This enumerated value is used when the tensor is used as the new_height parameter of the ResizeBilinear operator.
OH_NN_RESIZE_BILINEAR_NEW_WIDTH	This enumerated value is used when the tensor is used as the new_width parameter of the ResizeBilinear operator.
OH_NN_RESIZE_BILINEAR_PRESERVE_ASPECT_RATIO	This enumerated value is used when the tensor is used as the preserve_aspect_ratio parameter of the ResizeBilinear operator.
OH_NN_RESIZE_BILINEAR_COORDINATE_TRANSFORM_MODE	This enumerated value is used when the tensor is used as the coordinate_transform_mode parameter of the ResizeBilinear operator.
OH_NN_RESIZE_BILINEAR_EXCLUDE_OUTSIDE	This enumerated value is used when the tensor is used as the exclude_outside parameter of the ResizeBilinear operator.
OH_NN_LAYER_NORM_BEGIN_NORM_AXIS	This enumerated value is used when the tensor is used as the beginNormAxis parameter of the LayerNorm operator.
OH_NN_LAYER_NORM_EPSILON	This enumerated value is used when the tensor is used as the epsilon parameter of the LayerNorm operator.
OH_NN_LAYER_NORM_BEGIN_PARAM_AXIS	This enumerated value is used when the tensor is used as the beginParamsAxis parameter of the LayerNorm operator.
OH_NN_LAYER_NORM_ELEMENTWISE_AFFINE	This enumerated value is used when the tensor is used as the elementwiseAffine parameter of the LayerNorm operator.
OH_NN_REDUCE_PROD_KEEP_DIMS	This enumerated value is used when the tensor is used as the keep_dims parameter of the ReduceProd operator.
OH_NN_REDUCE_ALL_KEEP_DIMS	This enumerated value is used when the tensor is used as the keep_dims parameter of the ReduceAll operator.
OH_NN_QUANT_DTYPE_CAST_SRC_T	This enumerated value is used when the tensor is used as the src_t parameter of the QuantDTypeCast operator.
OH_NN_QUANT_DTYPE_CAST_DST_T	This enumerated value is used when the tensor is used as the dst_t parameter of the QuantDTypeCast operator.
OH_NN_TOP_K_SORTED	This enumerated value is used when the tensor is used as the Sorted parameter of the Topk operator.
OH_NN_ARG_MAX_AXIS	This enumerated value is used when the tensor is used as the axis parameter of the ArgMax operator.
OH_NN_ARG_MAX_KEEPDIMS	This enumerated value is used when the tensor is used as the keepDims parameter of the ArgMax operator.
OH_NN_UNSQUEEZE_AXIS	This enumerated value is used when the tensor is used as the Axis parameter of the Unsqueeze operator.

Function Description

OH_NNCompilation_Build()

OH_NN_ReturnCode OH_NNCompilation_Build (OH_NNCompilation * compilation)

Description
Compiles a model.

After the compilation configuration is complete, call this method to return the compilation result. The compilation instance pushes the model and compilation options to the device for compilation. After this method is called, additional compilation operations cannot be performed. If the OH_NNCompilation_SetDevice, OH_NNCompilation_SetCache, OH_NNCompilation_SetPerformanceMode, OH_NNCompilation_SetPriority, and OH_NNCompilation_EnableFloat16 methods are called, OH_NN_OPERATION_FORBIDDEN is returned.

Parameters

Name	Description
compilation	Pointer to the OH_NNCompilation instance.

Returns

Execution result of the function. If the operation is successful, OH_NN_SUCCESS is returned. If the operation fails, an error code is returned. For details about the error codes, see OH_NN_ReturnCode.

OH_NNCompilation_Construct()

OH_NNCompilation* OH_NNCompilation_Construct (const OH_NNModel * model)

Description
Creates a compilation instance of the OH_NNCompilation type.

After the OH_NNModel module completes model construction, APIs provided by the OH_NNCompilation module pass the model to underlying device for compilation. This method creates a OH_NNCompilation instance based on the passed OH_NNModel instance. The OH_NNCompilation_SetDevice method is called

to set the device to compile on, and OH_NNCompilation_Build is then called to complete compilation.

In addition to computing device selection, the OH_NNCompilation module supports features such as model caching, performance preference, priority setting, and float16 computing, which can be implemented by the following methods:

OH_NNCompilation_SetCache
OH_NNCompilation_SetPerformanceMode
OH_NNCompilation_SetPriority
OH_NNCompilation_EnableFloat16

After OH_NNCompilation is created by calling this method, the OH_NNModel instance can be released.

Parameters

Name	Description
model	Pointer to the OH_NNModel instance.

Returns

Returns the pointer to a OH_NNCompilation instance.

OH_NNCompilation_Destroy()

void OH_NNCompilation_Destroy (OH_NNCompilation ** compilation)

Description
Releases the Compilation object.

This method needs to be called to release the compilation instance created by calling OH_NNCompilation_Construct. Otherwise, memory leak will occur.

If compilation or *compilation is a null pointer, this method only prints warning logs and does not execute the release logic.

Parameters

Name	Description
compilation	Level-2 pointer to the OH_NNCompilation instance. After a compilation instance is destroyed, this method sets *compilation to a null pointer.

OH_NNCompilation_EnableFloat16()

OH_NN_ReturnCode OH_NNCompilation_EnableFloat16 (OH_NNCompilation * compilation, bool enableFloat16 )

Description
Enables float16 for computing.

Currently, Neural Network Runtime supports only float32 and int8. If this method is called on a device that supports float16, float16 will be used for computing the float32 model to reduce memory usage and execution time.

If this method is called on the device that does not support float16, the OH_NN_UNAVALIDABLE_DEVICE error code is returned.

Parameters

Name	Description
compilation	Pointer to the OH_NNCompilation instance.
enableFloat16	Indicates whether to enable float16. If this parameter is set to true, float16 inference is performed. If this parameter is set to false, float32 inference is performed.

Returns

OH_NNCompilation_SetCache()

OH_NN_ReturnCode OH_NNCompilation_SetCache (OH_NNCompilation * compilation, const char * cachePath, uint32_t version )

Description
Set the cache directory and version of the compiled model.

On the device that supports caching, a model can be saved as a cache file after being compiled at the device driver layer. The model can be directly read from the cache file in the next compilation, saving recompilation time. This method performs different operations based on the passed cache directory and version:

No file exists in the cache directory: Caches the compiled model to the directory and sets the cache version to version.
A complete cache file exists in the cache directory, and its version is version: Reads the cache file in the path and passes the data to the underlying device for conversion into executable model instances.
A complete cache file exists in the cache directory, and its version is earlier than version: When model compilation is complete on the underlying device, overwrites the cache file and changes the version number to version.
A complete cache file exists in the cache directory, and its version is later than version: Returns the OH_NN_INVALID_PARAMETER error code without reading the cache file.
The cache file in the cache directory is incomplete or you do not have the permission to access the cache file. Returns the OH_NN_INVALID_FILE error code.
The cache directory does not exist or you do not have the access permission. Returns the OH_NN_INVALID_PATH error code.

Parameters

Name	Description
compilation	Pointer to the OH_NNCompilation instance.
cachePath	Directory for storing model cache files. This method creates directories for different devices in the cachePath directory. You are advised to use a separate cache directory for each model.
version	Cache version.

Returns

OH_NNCompilation_SetDevice()

OH_NN_ReturnCode OH_NNCompilation_SetDevice (OH_NNCompilation * compilation, size_t deviceID )

Description
Specifies the device for model compilation and computing.

In the compilation phase, you need to specify the device for model compilation and computing. Call OH_NNDevice_GetAllDevicesID to obtain available device IDs. Call OH_NNDevice_GetType and OH_NNDevice_GetName to obtain device information and pass target device IDs to this method for setting.

Parameters

Name	Description
compilation	Pointer to the OH_NNCompilation instance.
deviceID	Device ID.

Returns

OH_NNCompilation_SetPerformanceMode()

OH_NN_ReturnCode OH_NNCompilation_SetPerformanceMode (OH_NNCompilation * compilation, OH_NN_PerformanceMode performanceMode )

Description
Sets the performance mode for model computing.

Neural Network Runtime allows you to set the performance mode for model computing to meet the requirements of low power consumption and ultimate performance. If this method is not called to set the performance mode in the compilation phase, the compilation instance assigns the OH_NN_PERFORMANCE_NONE mode for the model by default. In this case, the device performs computing in the default performance mode.

If this method is called on the device that does not support the setting of the performance mode, the OH_NN_UNAVALIDABLE_DEVICE error code is returned.

Parameters

Name	Description
compilation	Pointer to the OH_NNCompilation instance.
performanceMode	Performance mode. For details about the available performance modes, see OH_NN_PerformanceMode.

Returns

OH_NNCompilation_SetPriority()

OH_NN_ReturnCode OH_NNCompilation_SetPriority (OH_NNCompilation * compilation, OH_NN_Priority priority )

Description
Sets the model computing priority.

Neural Network Runtime allows you to set computing priorities for models.

The priorities apply only to models created by the process with the same UID. The settings will not affect models created by processes with different UIDs on different devices.

If this method is called on the device that does not support the priority setting, the OH_NN_UNAVALIDABLE_DEVICE error code is returned.

Parameters

Name	Description
compilation	Pointer to the OH_NNCompilation instance.
priority	Priority. For details about the optional priorities, see OH_NN_Priority.

Returns

OH_NNDevice_GetAllDevicesID()

OH_NN_ReturnCode OH_NNDevice_GetAllDevicesID (const size_t ** allDevicesID, uint32_t * deviceCount )

Description
Obtains the ID of the device connected to Neural Network Runtime.

Each device has a unique and fixed ID in Neural Network Runtime. This method returns device IDs on the current device through the uint32_t array.

Device IDs are returned through the size_t array. Each element of the array is the ID of a single device. The array memory is managed by Neural Network Runtime. The data pointer is valid before this method is called next time.

Parameters

Name	Description
allDevicesID	Pointer to the size_t array. The input *allDevicesID must be a null pointer. Otherwise, OH_NN_INVALID_PARAMETER is returned.
deviceCount	Pointer of the uint32_t type, which is used to return the length of *(allDevicesID)**.

Returns

OH_NNDevice_GetName()

OH_NN_ReturnCode OH_NNDevice_GetName (size_t deviceID, const char ** name )

Description
Obtains the name of the specified device.

deviceID specifies the device whose name will be obtained. The device ID needs to be obtained by calling OH_NNDevice_GetAllDevicesID.

Parameters

Name	Description
deviceID	Device ID.
name	Pointer to the char array. The passed *(char) must be a null pointer. Otherwise, OH_NN_INVALID_PARAMETER is returned. The value of (name)* is a C-style string ended with ’\0’.

Returns

OH_NNDevice_GetType()

OH_NN_ReturnCode OH_NNDevice_GetType (size_t deviceID, OH_NN_DeviceType * deviceType )

Description
Obtains the type information of the specified device.

deviceID specifies the device whose type will be obtained. Currently, Neural Network Runtime supports the following device types:

OH_NN_CPU: CPU device.
OH_NN_GPU: GPU device.
OH_NN_ACCELERATOR: machine learning dedicated accelerator.
OH_NN_OTHERS: other hardware types.

Parameters

Name	Description
deviceID	Device ID.
deviceType	Pointer to the OH_NN_DeviceType instance. The device type information is returned.

Returns

OH_NNExecutor_AllocateInputMemory()

OH_NN_Memory* OH_NNExecutor_AllocateInputMemory (OH_NNExecutor * executor, uint32_t inputIndex, size_t length )

Description
Allocates shared memory to a single input on a device.

Neural Network Runtime provides a method for proactively allocating shared memory on a device. By specifying the executor and input index value, this method allocates shared memory whose size is specified by length on the device associated with a single input and returns the operation result through the OH_NN_Memory instance.

Parameters

Name	Description
executor	Pointer to the OH_NNExecutor instance.
inputIndex	Input index value, which is in the same sequence of the data input when OH_NNModel_SpecifyInputsAndOutputs is called. Assume that the value of inputIndices is {1, 5, 9} when OH_NNModel_SpecifyInputsAndOutputs is called. In the memory input application, the index value for the three inputs is {0, 1, 2}.
length	Memory size to be applied for, in bytes.

Returns

Pointer to a OH_NN_Memory instance.

OH_NNExecutor_AllocateOutputMemory()

OH_NN_Memory* OH_NNExecutor_AllocateOutputMemory (OH_NNExecutor * executor, uint32_t outputIndex, size_t length )

Description
Allocates shared memory to a single output on a device.

Neural Network Runtime provides a method for proactively allocating shared memory on a device. By specifying the executor and output index value, this method allocates shared memory whose size is specified by length on the device associated with a single output and returns the operation result through the OH_NN_Memory instance.

Parameters

Name	Description
executor	Pointer to the OH_NNExecutor instance.
outputIndex	Output Index value, which is in the same sequence of the data output when OH_NNModel_SpecifyInputsAndOutputs is called. Assume that the value of outputIndices is {4, 6, 8} when OH_NNModel_SpecifyInputsAndOutputs is called. In output memory application, the index value for the three outputs is {0, 1, 2}.
length	Memory size to be applied for, in bytes.

Returns

Pointer to a OH_NN_Memory instance.

OH_NNExecutor_Construct()

OH_NNExecutor* OH_NNExecutor_Construct (OH_NNCompilation * compilation)

Description
Creates an executor instance of the OH_NNExecutor type.

This method constructs a model inference executor associated with the device based on the passed compiler. Use OH_NNExecutor_SetInput to set the model input data. After the input data is set, call OH_NNExecutor_Run to perform inference and then call OH_NNExecutor_SetOutput to obtain the computing result.

After calling this method to create the OH_NNExecutor instance, you can release the OH_NNCompilation instance if you do not need to create any other executors.

Parameters

Name	Description
compilation	Pointer to the OH_NNCompilation instance.

Returns

Pointer to a OH_NNExecutor instance.

OH_NNExecutor_Destroy()

void OH_NNExecutor_Destroy (OH_NNExecutor ** executor)

Description
Destroys an executor instance to release the memory occupied by the executor.

This method needs to be called to release the executor instance created by calling OH_NNExecutor_Construct. Otherwise, memory leak will occur.

If executor or *executor is a null pointer, this method only prints warning logs and does not execute the release logic.

Parameters

Name	Description
executor	Level-2 pointer to the OH_NNExecutor instance.

OH_NNExecutor_DestroyInputMemory()

void OH_NNExecutor_DestroyInputMemory (OH_NNExecutor * executor, uint32_t inputIndex, OH_NN_Memory ** memory )

Description
Releases the input memory to which the OH_NN_Memory instance points.

This method needs to be called to release the memory instance created by calling OH_NNExecutor_AllocateInputMemory. Otherwise, memory leak will occur. The mapping between inputIndex and memory must be the same as that in memory instance creation.

If memory or *memory is a null pointer, this method only prints warning logs and does not execute the release logic.

Parameters

Name	Description
executor	Pointer to the OH_NNExecutor instance.
inputIndex	Input index value, which is in the same sequence of the data input when OH_NNModel_SpecifyInputsAndOutputs is called. Assume that the value of inputIndices is {1, 5, 9} when OH_NNModel_SpecifyInputsAndOutputs is called. In memory input release, the index value for the three inputs is {0, 1, 2}.
memory	Level-2 pointer to the OH_NN_Memory instance. After shared memory is destroyed, this method sets *memory to a null pointer.

OH_NNExecutor_DestroyOutputMemory()

void OH_NNExecutor_DestroyOutputMemory (OH_NNExecutor * executor, uint32_t outputIndex, OH_NN_Memory ** memory )

Description
Releases the output memory to which the OH_NN_Memory instance points.

This method needs to be called to release the memory instance created by calling OH_NNExecutor_AllocateOutputMemory. Otherwise, memory leak will occur. The mapping between outputIndex and memory must be the same as that in memory instance creation.

If memory or *memory is a null pointer, this method only prints warning logs and does not execute the release logic.

Parameters

Name	Description
executor	Pointer to the OH_NNExecutor instance.
outputIndex	Output Index value, which is in the same sequence of the data output when OH_NNModel_SpecifyInputsAndOutputs is called. Assume that the value of outputIndices is {4, 6, 8} when OH_NNModel_SpecifyInputsAndOutputs is called. In output memory release, the index value for the three outputs is {0, 1, 2}.
memory	Level-2 pointer to the OH_NN_Memory instance. After shared memory is destroyed, this method sets *memory to a null pointer.

OH_NNExecutor_GetOutputShape()

OH_NN_ReturnCode OH_NNExecutor_GetOutputShape (OH_NNExecutor * executor, uint32_t outputIndex, int32_t ** shape, uint32_t * shapeLength )

Description
Obtains the dimension information about the output tensor.

After OH_NNExecutor_Run is called to complete a single inference, call this method to obtain the specified output dimension information and number of dimensions. It is commonly used in dynamic shape input and output scenarios.

Parameters

Name	Description
executor	Pointer to the OH_NNExecutor instance.
outputIndex	Output Index value, which is in the same sequence of the data output when OH_NNModel_SpecifyInputsAndOutputs is called. Assume that outputIndices is {4, 6, 8} when OH_NNModel_SpecifyInputsAndOutputs is called. When OH_NNExecutor_GetOutputShape is called to obtain dimension information about the output tensor, outputIndices is {0, 1, 2}.
shape	Pointer to the int32_t array. The value of each element in the array is the length of the output tensor in each dimension.
shapeLength	Pointer to the uint32_t type. The number of output dimensions is returned.

Returns

OH_NNExecutor_Run()

OH_NN_ReturnCode OH_NNExecutor_Run (OH_NNExecutor * executor)

Description
Performs inference.

Performs end-to-end inference and computing of the model on the device associated with the executor.

Parameters

Name	Description
executor	Pointer to the OH_NNExecutor instance.

Returns

OH_NNExecutor_SetInput()

OH_NN_ReturnCode OH_NNExecutor_SetInput (OH_NNExecutor * executor, uint32_t inputIndex, const OH_NN_Tensor * tensor, const void * dataBuffer, size_t length )

Description
Sets the single input data for a model.

This method copies the data whose length is specified by length (in bytes) in dataBuffer to the shared memory of the underlying device. inputIndex specifies the input to be set and tensor sets information such as the input shape, type, and quantization parameters.

Neural Network Runtime supports models with dynamical shape input. For fixed shape input and dynamic shape input scenarios, this method uses different processing policies.

Fixed shape input: The attributes of tensor must be the same as those of the tensor added by calling OH_NNModel_AddTensor in the composition phase.
Dynamic shape input: In the composition phase, because the shape is not fixed, each value in tensor.dimensions must be greater than 0 in the method calls to determine the shape input in the calculation phase. When setting the shape, you can modify only the dimension whose value is -1. Assume that [-1, 224, 224, 3] is input as the the dimension of A in the composition phase. When this method is called, only the size of the first dimension can be modified, for example, to [3, 224, 224, 3]. If other dimensions are adjusted, OH_NN_INVALID_PARAMETER is returned.

Parameters

Name	Description
executor	Pointer to the OH_NNExecutor instance.
inputIndex	Input index value, which is in the same sequence of the data input when OH_NNModel_SpecifyInputsAndOutputs is called. Assume that the value of inputIndices is {1, 5, 9} when OH_NNModel_SpecifyInputsAndOutputs is called. In input settings, the index value for the three inputs is {0, 1, 2}.
tensor	Sets the tensor corresponding to the input data.
dataBuffer	Pointer to the input data.
length	Length of the data buffer, in bytes.

Returns

OH_NNExecutor_SetInputWithMemory()

OH_NN_ReturnCode OH_NNExecutor_SetInputWithMemory (OH_NNExecutor * executor, uint32_t inputIndex, const OH_NN_Tensor * tensor, const OH_NN_Memory * memory )

Description
Specifies the hardware shared memory pointed to by the OH_NN_Memory instance as the shared memory used by a single input.

In scenarios where memory needs to be managed by yourself, this method binds the execution input to the OH_NN_Memory memory instance. During computing, the underlying device reads the input data from the shared memory pointed to by the memory instance. By using this method, concurrent execution of input setting, computing, and read can be implemented to improve inference efficiency of a data flow.

Parameters

Name	Description
executor	Pointer to the OH_NNExecutor instance.
inputIndex	Input index value, which is in the same sequence of the data input when OH_NNModel_SpecifyInputsAndOutputs is called. Assume that the value of inputIndices is {1, 5, 9} when OH_NNModel_SpecifyInputsAndOutputs is called. When the input shared memory is specified, the index value for the three inputs is {0, 1, 2}.
tensor	Pointer to OH_NN_Tensor, used to set the tensor corresponding to a single input.
memory	Pointer to OH_NN_Memory.

Returns

OH_NNExecutor_SetOutput()

OH_NN_ReturnCode OH_NNExecutor_SetOutput (OH_NNExecutor * executor, uint32_t outputIndex, void * dataBuffer, size_t length )

Description
Sets the buffer for a single output of a model.

This method binds the buffer to which dataBuffer points to the output specified by outputIndex. The length of the buffer is specified by length.

After OH_NNExecutor_Run is called to complete a single model inference, Neural Network Runtime compares the length of the buffer to which dataBuffer points with the length of the output data and returns different results based on the actual situation.

If the buffer length is greater than or equal to the data length, the inference result is copied to the buffer and OH_NN_SUCCESS is returned. You can read the inference result from dataBuffer.
If the buffer length is smaller than the data length, OH_NNExecutor_Run returns OH_NN_INVALID_PARAMETER and generates a log indicating that the buffer is too small.

Parameters

Name	Description
executor	Pointer to the OH_NNExecutor instance.
outputIndex	Output Index value, which is in the same sequence of the data output when OH_NNModel_SpecifyInputsAndOutputs is called. Assume that the value of outputIndices is {4, 6, 8} when OH_NNModel_SpecifyInputsAndOutputs is called. In output buffer settings, the index value for the three outputs is {0, 1, 2}.
dataBuffer	Pointer to the output data.
length	Length of the data buffer, in bytes.

Returns

OH_NNExecutor_SetOutputWithMemory()

OH_NN_ReturnCode OH_NNExecutor_SetOutputWithMemory (OH_NNExecutor * executor, uint32_t outputIndex, const OH_NN_Memory * memory )

Description
Specifies the hardware shared memory pointed to by the OH_NN_Memory instance as the shared memory used by a single output.

In scenarios where memory needs to be managed by yourself, this method binds the execution output to the OH_NN_Memory memory instance. When computing is performed, the underlying hardware directly writes the computing result to the shared memory to which the memory instance points. By using this method, concurrent execution of input setting, computing, and read can be implemented to improve inference efficiency of a data flow.

Parameters

Name	Description
executor	Executor.
outputIndex	Output Index value, which is in the same sequence of the data output when OH_NNModel_SpecifyInputsAndOutputs is called. Assume that the value of outputIndices is {4, 6, 8} when OH_NNModel_SpecifyInputsAndOutputs is called. When output shared memory is specified, the index value for the three outputs is {0, 1, 2}.
memory	Pointer to OH_NN_Memory.

Returns

OH_NNModel_AddOperation()

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 )

Description
Adds an operator to a model instance.

This method is used to add an operator to a model instance. The operator type is specified by op, and the operator parameters, inputs, and outputs are specified by paramIndices, inputIndices, and outputIndices respectively. This method verifies the attributes of operator parameters and the number of input and output parameters. These attributes must be correctly set when OH_NNModel_AddTensor is called to add tensors. For details about the expected parameters, input attributes, and output attributes of each operator, see OH_NN_OperationType.

paramIndices, inputIndices, and outputIndices store index values of tensors. Index values are determined by the sequence in which tensors are added to the model. For details about how to add a tensor, see OH_NNModel_AddTensor.

If unnecessary parameters are added for adding an operator, this method returns OH_NN_INVALID_PARAMETER. If no operator parameter is set, the operator uses the default parameter value. For details about the default values, see OH_NN_OperationType.

Parameters

Name	Description
model	Pointer to the OH_NNModel instance.
op	Specifies the type of an operator to be added. For details, see the enumerated values of OH_NN_OperationType.
paramIndices	Pointer to the OH_NN_UInt32Array instance, which is used to set operator parameters.
inputIndices	Pointer to the OH_NN_UInt32Array instance, which is used to set the operator input.
outputIndices	Pointer to the OH_NN_UInt32Array instance, which is used to set the operator output.

Returns

OH_NNModel_AddTensor()

OH_NN_ReturnCode OH_NNModel_AddTensor (OH_NNModel * model, const OH_NN_Tensor * tensor )

Description
Adds a tensor to a model instance.

The data node and operator parameters in the Neural Network Runtime model are composed of tensors of the model. This method is used to add tensors to a model instance based on the tensor parameter. The sequence of adding tensors is specified by the index value recorded in the model. The OH_NNModel_SetTensorData, OH_NNModel_AddOperation, and OH_NNModel_SpecifyInputsAndOutputs methods specifies tensors based on the index value.

Neural Network Runtime supports inputs and outputs of the dynamic shape. When adding a data node with a dynamic shape, you need to set the dimensions that support dynamic changes in tensor.dimensions to -1. For example, if tensor.dimensions of a four-dimensional tensor is set to [1, -1, 2, 2], the second dimension supports dynamic changes.

Parameters

Name	Description
model	Pointer to the OH_NNModel instance.
tensor	Pointer to the OH_NN_Tensor tensor. The tensor specifies the attributes of the tensor added to the model instance.

Returns

OH_NNModel_Construct()

OH_NNModel* OH_NNModel_Construct (void )

Description
Creates a model instance of the OH_NNModel type and uses other APIs provided by OH_NNModel to construct the model instance.

Before composition, call OH_NNModel_Construct to create a model instance. Based on the model topology, call the OH_NNModel_AddTensor, OH_NNModel_AddOperation, and OH_NNModel_SetTensorData methods to fill in the data and operator nodes of the model, and then call OH_NNModel_SpecifyInputsAndOutputs to specify the inputs and outputs of the model. After the model topology is constructed, call OH_NNModel_Finish to build the model.

After a model instance is used, you need to destroy it by calling OH_NNModel_Destroy to avoid memory leak.

Returns

Returns the pointer to a OH_NNModel instance.

OH_NNModel_Destroy()

void OH_NNModel_Destroy (OH_NNModel ** model)

Description
Releases a model instance.

This method needs to be called to release the model instance created by calling OH_NNModel_Construct. Otherwise, memory leak will occur.

If model or *model is a null pointer, this method only prints warning logs and does not execute the release logic.

Parameters

Name	Description
model	Level-2 pointer to the OH_NNModel instance. After a model instance is destroyed, this method sets *model to a null pointer.

OH_NNModel_Finish()

OH_NN_ReturnCode OH_NNModel_Finish (OH_NNModel * model)

Description
Completes model composition.

After the model topology is set up, call this method to indicate that the composition is complete. After this method is called, additional composition operations cannot be performed. If OH_NNModel_AddTensor, OH_NNModel_AddOperation, OH_NNModel_SetTensorData, and OH_NNModel_SpecifyInputsAndOutputs are called, OH_NN_OPERATION_FORBIDDEN is returned.

Before calling OH_NNModel_GetAvailableOperations and OH_NNCompilation_Construct, you must call this method to complete composition.

Parameters

Name	Description
model	Pointer to the OH_NNModel instance.

Returns

OH_NNModel_GetAvailableOperations()

OH_NN_ReturnCode OH_NNModel_GetAvailableOperations (OH_NNModel * model, size_t deviceID, const bool ** isSupported, uint32_t * opCount )

Description
Queries whether the device supports operators in the model. The support status is indicated by the Boolean value.

Queries whether underlying device supports operators in a model instance. The device is specified by deviceID, and the result is represented by the array pointed by isSupported. If the ith operator is supported, the value of (*isSupported)[i] is true. Otherwise, the value is false.

After this method is successfully executed, (*isSupported) points to the bool array that records the operator support status. The operator quantity for the array length is the same as that for the model instance. The memory corresponding to this array is managed by Neural Network Runtime and is automatically destroyed after the model instance is destroyed or this method is called again.

Parameters

Name	Description
model	Pointer to the OH_NNModel instance.
deviceID	Device ID to be queried, which can be obtained by using OH_NNDevice_GetAllDevicesID.
isSupported	Pointer to the bool array. When this method is called, *(isSupported)** must be a null pointer. Otherwise, OH_NN_INVALID_PARAMETER is returned.
opCount	Number of operators in a model instance, corresponding to the length of the *(isSupported)** array.

Returns

OH_NNModel_SetTensorData()

OH_NN_ReturnCode OH_NNModel_SetTensorData (OH_NNModel * model, uint32_t index, const void * dataBuffer, size_t length )

Description
Sets the tensor value.

For tensors with constant values (such as model weights), you need to use this method in the composition phase. The index value of a tensor is determined by the sequence in which the tensor is added to the model. For details about how to add a tensor, see OH_NNModel_AddTensor.

Parameters

Name	Description
model	Pointer to the OH_NNModel instance.
index	Index value of a tensor.
dataBuffer	Pointer to real data.
length	Length of the data buffer.

Returns

OH_NNModel_SpecifyInputsAndOutputs()

OH_NN_ReturnCode OH_NNModel_SpecifyInputsAndOutputs (OH_NNModel * model, const OH_NN_UInt32Array * inputIndices, const OH_NN_UInt32Array * outputIndices )

Description
Specifies the inputs and outputs of a model.

A tensor must be specified as the end-to-end inputs and outputs of a model instance. This type of tensor cannot be set using OH_NNModel_SetTensorData. The OH_NNExecutor method needs to be called in the execution phase to set the input and output data.

The index value of a tensor is determined by the sequence in which the tensor is added to the model. For details about how to add a tensor, see OH_NNModel_AddTensor.

Currently, the model inputs and outputs cannot be set asynchronously.

Parameters

Name	Description
model	Pointer to the OH_NNModel instance.
inputIndices	Pointer to the OH_NN_UInt32Array instance, which is used to set the operator input.
outputIndices	Pointer to the OH_NN_UInt32Array instance, which is used to set the operator output.

Returns

你可能感兴趣的鸿蒙文章

harmony 鸿蒙Native APIs

harmony 鸿蒙AVCapability

harmony 鸿蒙AVDemuxer

harmony 鸿蒙AVMuxer

harmony 鸿蒙AVScreenCapture

harmony 鸿蒙AVSource

harmony 鸿蒙AudioDecoder

harmony 鸿蒙AudioEncoder

harmony 鸿蒙CodecBase

harmony 鸿蒙Core

0 赞

所属分类： 后端技术
本文标签：
版权声明： 原创文章如转载，请注明本文链接: https://www.seaxiang.com/blog/f9c1cfef058b4f31b09707d157b6974f

harmony 鸿蒙NeuralNetworkRuntime

NeuralNetworkRuntime

Overview

Summary

Files

Structs

Types

Enums

Functions

Type Description

OH_NN_Memory

OH_NN_QuantParam

OH_NN_Tensor

OH_NN_UInt32Array

OH_NNCompilation

OH_NNExecutor

OH_NNModel

Enum Description

OH_NN_DataType

OH_NN_DeviceType

OH_NN_Format

OH_NN_FuseType

OH_NN_OperationType

OH_NN_PerformanceMode

OH_NN_Priority

OH_NN_ReturnCode

OH_NN_TensorType

Function Description

OH_NNCompilation_Build()

OH_NNCompilation_Construct()

OH_NNCompilation_Destroy()

OH_NNCompilation_EnableFloat16()

OH_NNCompilation_SetCache()

OH_NNCompilation_SetDevice()

OH_NNCompilation_SetPerformanceMode()

OH_NNCompilation_SetPriority()

OH_NNDevice_GetAllDevicesID()

OH_NNDevice_GetName()

OH_NNDevice_GetType()

OH_NNExecutor_AllocateInputMemory()

OH_NNExecutor_AllocateOutputMemory()

OH_NNExecutor_Construct()

OH_NNExecutor_Destroy()

OH_NNExecutor_DestroyInputMemory()

OH_NNExecutor_DestroyOutputMemory()

OH_NNExecutor_GetOutputShape()

OH_NNExecutor_Run()

OH_NNExecutor_SetInput()

OH_NNExecutor_SetInputWithMemory()

OH_NNExecutor_SetOutput()

OH_NNExecutor_SetOutputWithMemory()

OH_NNModel_AddOperation()

OH_NNModel_AddTensor()

OH_NNModel_Construct()

OH_NNModel_Destroy()

OH_NNModel_Finish()

OH_NNModel_GetAvailableOperations()

OH_NNModel_SetTensorData()

OH_NNModel_SpecifyInputsAndOutputs()

你可能感兴趣的鸿蒙文章

热门推荐

标签云

最新文章

本文目录