xref: /drivers/interface/nnrt/v1_0/NnrtTypes.idl

/*
 * Copyright (c) 2022 Huawei Device Co., Ltd.
 * Licensed under the Apache License, Version 2.0 (the "License");
 * you may not use this file except in compliance with the License.
 * You may obtain a copy of the License at
 *
 *     http://www.apache.org/licenses/LICENSE-2.0
 *
 * Unless required by applicable law or agreed to in writing, software
 * distributed under the License is distributed on an "AS IS" BASIS,
 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 * See the License for the specific language governing permissions and
 * limitations under the License.
 */

/**
 * @addtogroup NNRt
 * @{
 *
 * @brief Provides a unified interface for AI chip drivers to access OpenHarmony. 
 * Neural Network Runtime (NNRt) is a cross-chip inference computing runtime environment oriented to the AI field.
 * It serves as a bridge between the upper-layer AI inference framework and the underlying acceleration chip to implement cross-chip inference computing of AI models.
 * @since 3.2
 * @version 1.0
 */

/**
 * @file NnrtTypes.idl
 *
 * @brief Defines the types used in the HDI methods.
 *
 * @since 3.2
 * @version 1.0
 */

/**
 * @brief Defines the package path of the NNRt module.
 *
 * @since 3.2
 * @version 1.0
 */
package ohos.hdi.nnrt.v1_0;

/**
 * @brief Defines the shared memory data structure.
 *
 * @since 3.2
 * @version 1.0
 */
struct SharedBuffer {
    /** File descriptor of the shared memory. */
    FileDescriptor fd;
    /** Size of the shared memory, in bytes. */
    unsigned int bufferSize;
    /** Offset of the start address of the valid data in the shared memory. */
    unsigned int offset;
    /** Space occupied by the valid data, in bytes. */
    unsigned int dataSize;
};

/**
 * @brief Enumerates the AI chip types.
 *
 * @since 3.2
 * @version 1.0
 */
enum DeviceType: int {
    /** Other type */
    OTHER,
    /** CPU chip */
    CPU,
    /** GPU chip */
    GPU,
    /** AI acceleration chip, such as NPU chip and DSP chip */
    ACCELERATOR
};

/**
 * @brief Enumerates the AI chip states.
 *
 * @since 3.2
 * @version 1.0
 */
enum DeviceStatus: int {
    /** Available. */
    AVAILABLE,
    /** Busy. The chip in this state may not respond to computing tasks in a timely manner. */
    BUSY,
    /** Offline. The chip in this state cannot respond to computing tasks. */
    OFFLINE,
    /** Unknown state. */
    UNKNOWN
};

/**
 * @brief Enumerates the performance modes for a chip to perform AI computing.
 *
 * @since 3.2
 * @version 1.0
 */
enum PerformanceMode: int {
    /** No performance mode is specified. The specific running mode is defined by the chip. */
    PERFORMANCE_NONE,
    /** Low-performance mode, which provides slow AI computing but low power consumption. */
    PERFORMANCE_LOW,
    /** Medium-performance mode, which provides moderate computing speed and power consumption. */
    PERFORMANCE_MEDIUM,
    /** High-performance mode, which provides fast AI computing but high power consumption. */
    PERFORMANCE_HIGH,
    /** Extreme-performance mode, which provides the fastest AI computing but highest power consumption. */
    PERFORMANCE_EXTREME
};

/**
 * @brief Enumerates the AI computing task priorities.
 *
 * @since 3.2
 * @version 1.0
 */
enum Priority: int {
    /** No task priority is specified. The specific execution policy is defined by the chip. */
    PRIORITY_NONE,
    /** Low priority. A task with a higher priority will be executed first. */
    PRIORITY_LOW,
    /** Medium priority. A task with a higher priority will be executed first. */
    PRIORITY_MEDIUM,
    /** High priority. High-priority tasks are executed first. */
    PRIORITY_HIGH
};

/**
 * @brief Defines the parameters required for model building.
 *
 * @since 3.2
 * @version 1.0
 */
struct ModelConfig {
    /** Whether to run a Float32 model in Float16 precision. */
    boolean enableFloat16;
    /** Performance mode of the computing task. For details, see {@link PerformanceMode}. */
    enum PerformanceMode mode;
    /** Priority of the computing task. For details, see {@link Priority}. */
    enum Priority priority;
};

/**
 * @brief Enumerates the operator data formats. This parameter must be used together with {@link Tensor}.
 *
 * @since 3.2
 * @version 1.0
 */
enum Format : byte {
    /** Format initial value. */
    FORMAT_NONE = -1,
    /** NCHW, which indicates the number of data samples, image channels, image height, and image width in sequence. */
    FORMAT_NCHW = 0,
    /** NHWC. */
    FORMAT_NHWC = 1
};

/**
 * @brief Defines the quantization parameter structure.
 *
 * In the following formula, <b>q</b> is a quantized parameter, <b>r</b> is a real parameter, <b>\f$ r_{max} \f$</b> is the maximum value of the data to be quantized, <b>\f$ r_{min} \f$</b> is the minimum value of the data to be quantized, and <b>round(x)</b> means to round off <b>x</b> to an integer.
 \f[
  \text{clamp}(x,min,max) =
  \begin{cases}
       \text{max} & \text{ if } x > \text{ max } \\
       \text{min} & \text{ if } x < \text{ min } \\
       x & \text{ otherwise } \\
   \end{cases}
 \f]
 * Formula for transforming a real number from a floating-point representation to a fixed-point representation:
 \f[
    \text{q}(x_i) = clamp(round(\frac{r}{scale}+zeroPoint), min , max)
 \f]
 * Formula for transforming a real number from a fixed-point representation to a floating-point representation:
 \f[
    \text{r}= (q-zeroPoint)*scale
 \f]
 * <b>scale</b> is calculated by using the following formula:
 \f[
    scale = \frac{r_{max}-r_{min}}{q_{max}-q_{min}}
 \f]
  * <b>zeroPoint</b> is calculated by using the following formula:
 \f[
    zeroPoint = round(q_{min}-\frac{r_{min}}{scale})
 \f]
  * <b>\f$ q_{min},q_{max} \f$</b> is calculated by using the following formula:
 \f[
    q_{min} = -(1<<(numBits-1))
 \f]
 \f[
    q_{max} = (1<<(numBits-1))-1
 \f]
 * When \f$ r_{min} \f$ and \f$ r_{max} \f$ are <b>0</b>, <b>scale</b> and <b>zeroPoint</b> must be <b>0</b>.
 *
 * @since 3.2
 * @version 1.0
 */
struct QuantParam {
    /** Number of quantized bits */
    int numBits;
    /** Zero value */
    int zeroPoint;
    /** Step of the quantizer. */
    double scale;
};

/**
 * @brief Enumerates the tensor data types. This parameter must be used together with {@link Tensor}.
 *
 * @since 3.2
 * @version 1.0
 */
enum DataType : byte {
    /** Unknown type. */
    DATA_TYPE_UNKNOWN = 0,
    /** Boolean. */
    DATA_TYPE_BOOL = 30,
    /** INT8. */
    DATA_TYPE_INT8 = 32,
    /** INT16. */
    DATA_TYPE_INT16 = 33,
    /** INT32. */
    DATA_TYPE_INT32 = 34,
    /** INT64. */
    DATA_TYPE_INT64 = 35,
    /** UINT8. */
    DATA_TYPE_UINT8 = 37,
    /** UINT16. */
    DATA_TYPE_UINT16 = 38,
    /** UINT32. */
    DATA_TYPE_UINT32 = 39,
    /** UINT64. */
    DATA_TYPE_UINT64 = 40,
    /** FLOAT16. */
    DATA_TYPE_FLOAT16 = 42,
    /** FLOAT32. */
    DATA_TYPE_FLOAT32 = 43,
    /** FLOAT64. */
    DATA_TYPE_FLOAT64 = 44,
};

/**
 * @brief Defines the input and output tensors of an AI model.
 *
 * @since 3.2
 * @version 1.0
 */
struct IOTensor {
    /** Tensor name. */
    String name;
    /** Data type of the tensor. For details, see {@link DataType}. */
    enum DataType dataType;
    /** Dimensions of the tensor. */
    int[] dimensions;
    /** Format of the tensor. For details, see {@link Format}. */
    enum Format format;
    /** Tensor data, which is stored in the shared memory. For details about the shared memory, see {@link SharedBuffer}. */
    struct SharedBuffer data;
};

/**
 * @brief Enumerates the quantization types. This parameter must be used together with {@link Node}.
 *
 * @since 3.2
 * @version 1.0
 */
enum QuantType: byte {
    /** Do not use quantification. */
    QUANT_TYPE_NONE,
    /** INT8 quantization. */
    QUANT_TYPE_ALL,
};

/**
 * @brief Enumerates the operator types.
 *
 * @since 3.2
 * @version 1.0
 */
enum NodeType : unsigned int {
    /** None. */
    NODE_TYPE_NONE = 0,
    /** Activation function */
    NODE_TYPE_ACTIVATION = 2,
    /** ADD operator */
    NODE_TYPE_ADD_FUSION = 5,
    /** ArgMax operator */
    NODE_TYPE_ARGMAX_FUSION = 11,
    /** AVGPOOL operator */
    NODE_TYPE_AVGPOOL_FUSION = 17,
    /** BatchToSpaceND operator */
    NODE_TYPE_BATCH_TO_SPACE_ND = 22,
    /** BiasAdd operator */
    NODE_TYPE_BIAS_ADD = 23,
    /** Cast operator */
    NODE_TYPE_CAST = 28,
    /** Concat operator */
    NODE_TYPE_CONCAT = 31,
    /** Conv2D operator, including common convolution, separable convolution, and group convolution */
    NODE_TYPE_CONV2D_FUSION = 35,
    /** Two-dimensional deconvolution operator */
    NODE_TYPE_CONV2D_TRANSPOSE_FUSION = 36,
    /** Div operator */
    NODE_TYPE_DIV_FUSION = 47,
    /** Element-level operator */
    NODE_TYPE_ELTWISE = 52,
    /** ExpandDims operator */
    NODE_TYPE_EXPAND_DIMS = 56,
    /** Fill operator */
    NODE_TYPE_FILL = 66,
    /** FullConnection operator */
    NODE_TYPE_FULL_CONNECTION = 67,
    /** BatchNorm operator */
    NODE_TYPE_FUSED_BATCH_NORM = 68,
    /** Gather operator */
    NODE_TYPE_GATHER = 69,
    /** LayerNorm operator */
    NODE_TYPE_LAYER_NORM_FUSION = 75,
    /** LessEqual operator */
    NODE_TYPE_LESS_EQUAL = 78,
    /** MatMul operator */
    NODE_TYPE_MATMUL_FUSION = 89,
    /** Maximum operator */
    NODE_TYPE_MAXIMUM = 90,
    /** MaxPool operator */
    NODE_TYPE_MAX_POOL_FUSION = 92,
    /** Mul operator */
    NODE_TYPE_MUL_FUSION = 99,
    /** OneHot operator */
    NODE_TYPE_ONE_HOT = 105,
    /** Pad operator */
    NODE_TYPE_PAD_FUSION = 107,
    /** Pow operator */
    NODE_TYPE_POW_FUSION = 110,
    /** PReLU operator */
    NODE_TYPE_PRELU_FUSION = 112,
    /** QuantDTypeCast operator */
    NODE_TYPE_QUANT_DTYPE_CAST = 113,
    /** Reduce operator */
    NODE_TYPE_REDUCE_FUSION = 118,
    /** Reshape operator */
    NODE_TYPE_RESHAPE = 119,
    /** Resize operator */
    NODE_TYPE_RESIZE = 120,
    /** Rsqrt operator */
    NODE_TYPE_RSQRT = 126,
    /** Scale operator */
    NODE_TYPE_SCALE_FUSION = 127,
    /** Shape operator */
    NODE_TYPE_SHAPE = 130,
    /** Slice operator */
    NODE_TYPE_SLICE_FUSION = 135,
    /** Softmax operator */
    NODE_TYPE_SOFTMAX = 138,
    /** SpaceToBatchND operator */
    NODE_TYPE_SPACE_TO_BATCH_ND = 141,
    /** Split operator */
    NODE_TYPE_SPLIT = 145,
    /** Sqrt operator */
    NODE_TYPE_SQRT = 146,
    /** Squeeze operator */
    NODE_TYPE_SQUEEZE = 147,
    /** SquaredDifference operator */
    NODE_TYPE_SQUARED_DIFFERENCE = 149,
    /** Stack operator */
    NODE_TYPE_STACK = 150,
    /** StridedSlice operator */
    NODE_TYPE_STRIDED_SLICE = 151,
    /** Sub operator */
    NODE_TYPE_SUB_FUSION = 152,
    /** Tile operator */
    NODE_TYPE_TILE_FUSION = 160,
    /** TopK operator */
    NODE_TYPE_TOPK_FUSION = 161,
    /** Transpose operator */
    NODE_TYPE_TRANSPOSE = 162,
    /** Unsqueeze operator */
    NODE_TYPE_UNSQUEEZE = 165,
};

/**
 * @brief Enumerates the resize methods. It must be used together with the {@link Resize} operator.
 *
 * @since 3.2
 * @version 1.0
 */
enum ResizeMethod : byte {
    /** Unknown. This is the default value. */
    RESIZE_METHOD_UNKNOWN = -1,
    /** Bilinear interpolation.
     * For example, calculate the value of an unknown function <b>f</b> at point \f$ (x,y) \f$, where \f$ x_1< x < x_2, y_1< y < y_2 \f$.
     * The values of the four coordinate points are \f$ Q_{11} = (x_1, y_1), Q_{12} = (x1, y2), Q_{21} = (x_2, y_1), and Q_{22} = (x_2, y_2) \f$.
     * \f$f(Q_{11}), f(Q_{12}), f(Q_{21}), and f(Q_{22}) \f$ represent the values of the four points. The value of \f$ f(x,y) \f$ can be calculated by using the following formula:
      \f[
         f(x,y_1) = \frac{x_2-x}{x_2-x_1}f(Q_{11})+\frac{x-x_1}{x_2-x_1}f(Q_{21})
      \f]

      \f[
         f(x,y_2) = \frac{x_2-x}{x_2-x_1}f(Q_{12})+\frac{x-x_1}{x_2-x_1}f(Q_{22})
      \f]

      \f[
         f(x,y) = \frac{y_2-y}{y_2-y_1}f(x,y_1)+\frac{y-y_1}{y_2-y_1}f(x,y_2)
      \f]
     */
    RESIZE_METHOD_LINEAR = 0,
    /** Nearest neighbor interpolation.
     * For example, calculate the value of an unknown function <b>f</b> at point \f$ (x,y) \f$, where \f$ x_1< x < x_2, y_1< y < y_2 \f$.
     * The values of the four coordinate points are \f$ Q_{11} = (x_1, y_1), Q_{12} = (x1, y2), Q_{21} = (x_2, y_1), and Q_{22} = (x_2, y_2) \f$.
     * Then, the value of the point closest to the point \f$(x,y) \f$ is the value of \f$ f(x,y) \f$.
     */
    RESIZE_METHOD_NEAREST = 1,
    /** Bicubic interpolation.
     * Bicubic interpolation obtains the value of a sampling point by calculating the weighted average of the values of 16 points around the sampling point. This parameter must be used together with <b>cubicCoeff</b> and <b>coordinateTransformMode</b> of {@link Resize}.
     * When coordinateTransformMode==COORDINATE_TRANSFORM_MODE_HALF_PIXEL, <b>cubicCoeff</b> is <b>-0.5</b>. In other cases, cubicCoeff is <b>-0.75</b>. The weight function of the interpolation is as follows:
      \f[
         W(x) =
         \begin{cases}
            (cubicCoeff+2)|x|^3 - (cubicCoeff+3)|x|^2 +1 , &\text{if } |x| \leq 1; \cr
            cubicCoeff|x|^3 - 5cubicCoeff|x|^2 + 8cubicCoeff|x| - 4a, &\text{if } 1 \lt |x| \leq 2; \cr
            0, &\text{otherwise.}
        \end{cases}
      \f]
     */
    RESIZE_METHOD_CUBIC = 2
};

/**
 * @brief Enumerates the coordinate transformation modes. Only the {@link Resize} operator uses this parameter.
 * For example, the width coordinates are transformed, where:
 * <b>new_i</b> is the ith coordinate of the resized tensor along the x axis.
 * <b>old_i</b> is the coordinate of the input tensor along the x axis.
 * <b>newWidth</b> is the length of the resized tensor along the x axis.
 * <b>oldWidth</b> is the length of the input tensor along the x axis.
 * <b>old_i</b> can be calculated by using the following formula:
 *
 * COORDINATE_TRANSFORM_MODE_ASYMMETRIC: \f$ old_i = newWidth != 0 ? new_i * oldWidth / newWidth : 0 \f$ <br>
 * COORDINATE_TRANSFORM_MODE_ALIGN_CORNERS: \f$ old_i = newWidth != 1 ? new_i * (oldWidth - 1) / (newWidth - 1) \f$<br>
 * COORDINATE_TRANSFORM_MODE_HALF_PIXEL: \f$ old_i = newWidth > 1 ? (new_x + 0.5) * oldWidth / newWidth - 0.5 : 0 \f$<br>
 *
 * @since 3.2
 * @version 1.0
 */
enum CoordinateTransformMode : byte {
    /** Scale based on the ratio without alignment. */
    COORDINATE_TRANSFORM_MODE_ASYMMETRIC = 0,
    /** Align the four corners of the image. */
    COORDINATE_TRANSFORM_MODE_ALIGN_CORNERS = 1,
    /** Align with the pixel center. */
    COORDINATE_TRANSFORM_MODE_HALF_PIXEL = 2
};

/**
 * @brief Enumerates the nearest neighbor interpolation types. It must be used together with the {@link Resize} operator.
 *
 * @since 3.2
 * @version 1.0
 */
enum NearestMode : byte {
    /** Round off. */
    NEAREST_MODE_NORMAL = 0,
    /** Round toward negative infinity. For example, 23.5 is rounded to 23, and −23.5 is rounded to −24. */
    NEAREST_MODE_ROUND_HALF_DOWN = 1,
    /** Round toward positive infinity. For example, 23.5 is rounded to 24, and −23.5 is rounded to −23. */
    NEAREST_MODE_ROUND_HALF_UP = 2,
    /** Round down to the nearest integer. For example, 23.5 is rounded down to 23, and −23.5 is rounded down to −24. */
    NEAREST_MODE_FLOOR = 3,
    /** Round up to the nearest integer. For example, 23.5 is rounded up to 24, and −23.5 is rounded up to −23. */
    NEAREST_MODE_CEIL = 4
};

/**
 * @brief Enumerates the activation function types. Activation functions introduce nonlinearity to neural networks. This allows the use of neural network models in nonlinear models.
 * If an operator in the {@link NodeAttrTypes.idl} file has <b>ActivationType</b> parameters, the corresponding activation function will be called after the operator calculation is complete.
 *
 * @since 3.2
 * @version 1.0
 */
enum ActivationType : byte {
    /** No activation function. */
    ACTIVATION_TYPE_NO_ACTIVATION = 0,
    /**
     * ReLU activation function.
     * ReLU calculates \f$ max(x_i, 0) \f$ element by element. It outputs the value directly if it is positive; otherwise, it outputs <b>0</b>.
     \f[
        \text{ReLU}(x_i) = (x_i)^+ = \max(x_i, 0),
     \f]
     * <b>\f$ x_i \f$</b> is the input element.
     */
    ACTIVATION_TYPE_RELU = 1,
    /**
     * Sigmoid activation function.
     * Execute the sigmoid activation function element-wise.
     * The sigmoid function is defined as follows:
     \f[
        \text{Sigmoid}(x_i) = \frac{1}{1 + \exp(-x_i)}
     \f]
     * <b>\f$ x_i \f$</b> is the input element.
     */
    ACTIVATION_TYPE_SIGMOID = 2,
    /**
     * ReLU6 activation function.
     * ReLU6 is similar to ReLU. The difference is ReLU6 has an upper limit of <b>6</b>. If the input is greater than 6, the output is limited to <b>6</b>.
     * The ReLU6 function is defined as follows:
     \f[
        \text{ReLU6}(x_i) = \min(\max(0, x_i), 6)
     \f]
     * <b>\f$ x_i \f$</b> is the input element.
     */
    ACTIVATION_TYPE_RELU6 = 3,
    /**
     * Exponential Linear Unit (ELU) activation function.
     * ELU calculates the ELU for each input element.
     * The ELU function is defined as follows:
     \f[
        ELU(x_{i}) =
        \begin{cases}
        x_i, &\text{if } x_i \geq 0; \cr
        \alpha * (\exp(x_i) - 1), &\text{otherwise.}
        \end{cases}
     \f]
     * <b>\f$ x_i \f$</b> indicates the input element, and <b>\f$ \alpha \f$</b> indicates the alpha parameter, which is set by {@link Activation}.
     */
    ACTIVATION_TYPE_ELU = 4,
    /**
     * LeakyReLU activation function.
     * The LeakyReLU function is defined as follows:
     \f[
        \text{LeakyReLU}(x_i) =
        \begin{cases}
            x_i, &\text{if } x_i \geq 0; \cr
            {\alpha} * x_i, &\text{otherwise.}
        \end{cases}
     \f]
     * <b>\f$ x_i \f$</b> indicates the input element, and <b>\f$ \alpha \f$</b> indicates the alpha parameter, which is set by {@link Activation}.
     */
    ACTIVATION_TYPE_LEAKY_RELU = 5,
    /**
     * Activation function for calculating the absolute value.
     * The function is defined as follows:
     \f[
        \text{abs}(x_i) = |x_i|
     \f]
     * <b>\f$ x_i \f$</b> is the input element.
     */
    ACTIVATION_TYPE_ABS = 6,
    /**
     * ReLU1 activation function.
     * The ReLU1 function is defined as follows:
     \f[
        \text{ReLU1}(x_i)= \min(\max(0, x_i), 1)
     \f]
     * <b>\f$ x_i \f$</b> is the input element.
     */
    ACTIVATION_TYPE_RELU1 = 7,
    /**
     * SoftSign activation function.
     * The SoftSign function is defined as follows:
     \f[
        \text{SoftSign}(x_i) = \frac{x_i}{1 + |x_i|}
     \f]
     * <b>\f$ x_i \f$</b> is the input.
     */
    ACTIVATION_TYPE_SOFTSIGN = 8,
    /**
     * Softplus activation function.
     * Softplus is a smooth approximation to ReLU. It can be used to constrain the output to always be positive.
     * The Softplus function is defined as follows:
     \f[
        \text{Softplus}(x_i) = \log(1 + \exp(x_i))
     \f]
     * <b>\f$ x_i \f$</b> is the input element.
     */
    ACTIVATION_TYPE_SOFTPLUS = 9,
    /**
     * Tanh activation function.
     * The Tanh function is defined as follows:
     \f[
        tanh(x) = \frac{\exp(x_i) - \exp(-x_i)}{\exp(x_i) + \exp(-x_i)} = \frac{\exp(2x_i) - 1}{\exp(2x_i) + 1}
     \f]
     * <b>\f$ x_i \f$</b> is the input element.
     */
    ACTIVATION_TYPE_TANH = 10,
    /**
     * Scaled exponential Linear Unit (SELU) activation function.
     * The SELU function is defined as follows:
     \f[
        SELU(x_{i}) =
        scale *
        \begin{cases}
        x_{i}, &\text{if } x_{i} \geq 0; \cr
        \text{alpha} * (\exp(x_i) - 1), &\text{otherwise.}
        \end{cases}
     \f]
     * <b>\f$ x_i \f$</b> is the input element, and <b>\f$ \alpha \f$</b> and <b>\f$ scale \f$</b> are predefined constants (\f$ \alpha = 1.67326324 \f$, \f$ scale = 1.05070098 \f$).
     */
    ACTIVATION_TYPE_SELU = 11,
    /**
     * Hard Swish activation function.
     *
     \f[
        \text{Hardswish}(x_{i}) = x_{i} * \frac{ReLU6(x_{i} + 3)}{6}
     \f]
     * <b>\f$ x_i \f$</b> is the input element.
     */
    ACTIVATION_TYPE_HSWISH = 12,
    /**
     * Hard sigmoid activation function.
     * The hard sigmoid function is defined as follows:
     \f[
        \text{Hardsigmoid}(x_{i}) = max(0, min(1, \frac{x_{i} + 3}{6}))
     \f]
     * <b>\f$ x_i \f$</b> is the input element.
     */
    ACTIVATION_TYPE_HSIGMOID = 13,
    /**
     * ThresholdedReLU activation function.
     * ThresholdedReLU is similar to ReLU. The <b>ThresholdedReLU</b> function is defined as follows:
     \f[
        \text{ThresholdedReLU}(x_i) = \min(\max(0, x_i), t)
     \f]
     * <b>\f$ x_i \f$</b> is the input element, and <b>\f$ t \f$</b> is the maximum value.
     */
    ACTIVATION_TYPE_THRESHOLDRELU = 14,
    /**
     * Linear activation function.
     * The Linear function is defined as follows:
     \f[
        \text{Linear}(x_i) = x_i
     \f]
     * <b>\f$ x_i \f$</b> is the input element.
     */
    ACTIVATION_TYPE_LINEAR = 15,
    /**
     * HardTanh activation function.
     * The HardTanh function is defined as follows:
     \f[
       \text{HardTanh}(x_i) =
       \begin{cases}
            \text{max_val} & \text{ if } x_i > \text{ max_val } \\
            \text{min_val} & \text{ if } x_i < \text{ min_val } \\
            x_i & \text{ otherwise } \\
        \end{cases}
     \f]
     * <b>\f$ x_i \f$</b> is the input, <b>\f$ max\_val \f$</b> is the maximum value, and <b>\f$ min\_val \f$</b> is the minimum value. The two parameters are set by {@link Activation}.
     */
    ACTIVATION_TYPE_HARD_TANH = 16,
    /**
     * Sign activation function.
     * The Sign function is defined as follows:
     \f[
        Sign(x_i) = \begin{cases} -1, &if\ x_i < 0 \cr
        0, &if\ x_i = 0 \cr
        1, &if\ x_i > 0\end{cases}
     \f]
     * <b>\f$ x_i \f$</b> is the input.
     */
    ACTIVATION_TYPE_SIGN = 17,
    /**
     * Swish activation function.
     * The Swish function is defined as follows:
     \f[
        \text{Swish}(x_i) = x_i * Sigmoid(x_i)
     \f]
     * <b>\f$ x_i \f$</b> is the input.
     */
    ACTIVATION_TYPE_SWISH = 18,
    /**
     * Gaussian error linear unit (GELU) activation function.
     * The GELU function is defined as follows:
     \f[
        GELU(x_i) = x_i*P(X < x_i)
     \f]
     * <b>\f$ x_i \f$</b> is the input element, and <b>\f$ P \f$</b> is a cumulative distribution function of the standard Gaussian distribution.
     * You need to use the <b>approximate</b> parameter of {@link Activation} to specify whether to use approximation.
     */
    ACTIVATION_TYPE_GELU = 19,
    /** Unknown */
    ACTIVATION_TYPE_UNKNOWN = 20
};

/**
 * @brief Enumerates the modes for removing dimensions. It must be used together with the {@link ReduceFusion} operator.
 *
 * @since 3.2
 * @version 1.0
 */
enum ReduceMode : byte {
    /** Use the average value of all elements of a dimension to replace other elements of the dimension. */
    REDUCE_MODE_MEAN = 0,
    /** Use the maximum value of all elements of a dimension to replace other elements of the dimension. */
    REDUCE_MODE_MAX = 1,
    /** Use the minimum value of all elements of a dimension to replace other elements of the dimension. */
    REDUCE_MODE_MIN = 2,
    /** Use the product of all elements of a dimension to replace other elements of the dimension. */
    REDUCE_MODE_PROD = 3,
    /** Use the sum of all elements of a dimension to replace other elements of the dimension. */
    REDUCE_MODE_SUM = 4,
    /** Use the sum of squares of all elements of a dimension to replace other elements of the dimension. */
    REDUCE_MODE_SUM_SQUARE = 5,
    /** Use the sum of absolute values of all elements of a dimension to replace other elements of the dimension. */
    REDUCE_MODE_ASUM = 6,
    /** Use the logical AND of all elements of a dimension to replace other elements of the dimension. */
    REDUCE_MODE_ALL = 7
};

/**
 * @brief Enumerates the calculation types supported by elements. It must be used together with the {@link Eltwise} operator.
 *
 * @since 3.2
 * @version 1.0
 */
enum EltwiseMode : byte {
    /** Product of the elements of two tensors */
    ELTWISE_MODE_PROD = 0,
    /** Difference between the elements of two tensors */
    ELTWISE_MODE_SUM = 1,
    /** Maximum value of the elements of two tensors */
    ELTWISE_MODE_MAXIMUM = 2,
    /** Unknown type */
    ELTWISE_MODE_UNKNOWN = 3
};

/**
 * @brief Enumerates the padding types. It must be used together with {@link AvgPoolFusion}, {@link AvgPoolFusion}, {@link Conv2DFusion}, and {@link MaxPoolFusion}.
 *
 * @since 3.2
 * @version 1.0
 */
enum PadMode : byte {
    /**
     * Adds 0s in the input height and width directions.
     * If this mode is used, the padding parameter of the operator must be greater than or equal to 0.
     */
    PAD_MODE_PAD = 0,
    /**
     * The output height and width are obtained by dividing the input height and width by a stride and rounding off the quotient to an integer.
     * If this mode is used, the padding parameter of the operator must be <b>0</b>.
     */
    PAD_MODE_SAME = 1,
    /**
     * Return the output of a valid calculation without padding. Pixels that do not meet the calculation requirements will be discarded.
     * If this mode is used, the padding parameter of the operator must be <b>0</b>.
     */
    PAD_MODE_VALID = 2,
};

/**
 * @brief Enumerates the algorithms for rounding off decimals. It must be used together with the {@link AvgPoolFusion} operator.
 *
 * @since 3.2
 * @version 1.0
 */
enum RoundMode : byte {
    /** Round down to the nearest integer. For example, 23.5 is rounded down to 23, and −23.5 is rounded down to −24. */
    ROUND_MODE_FLOOR = 0,
    /** Round up to the nearest integer. For example, 23.5 is rounded up to 24, and −23.5 is rounded up to −23. */
    ROUND_MODE_CEIL = 1
};

/**
 * @brief Enumerates the padding modes. It must be used together with the {@link PadFusion} operator.
 *
 * When <b>x</b> is \f$[[1,2,3],[4,5,6],[7,8,9]]\f$ and <b>paddings</b>is \f$[[2,2], [2,2]] \f$, the effect is as follows: <br>
 * If paddingMode==PADDING_MODE_CONSTANT and constantValue = 0, the output is as follows:
 *
      \f$[[0. 0. 0. 0. 0. 0. 0.],\\
          [0. 0. 0. 0. 0. 0. 0.],\\
          [0. 0. 1. 2. 3. 0. 0.],\\
          [0. 0. 4. 5. 6. 0. 0.],\\
          [0. 0. 7. 8. 9. 0. 0.],\\
          [0. 0. 0. 0. 0. 0. 0.],\\
          [0. 0. 0. 0. 0. 0. 0.]]\\ \f$
 *
 * If paddingMode==PADDING_MODE_REFLECT, the output is as follows:
 *
      \f$[[9. 8. 7. 8. 9. 8. 7.],\\
          [6. 5. 4. 5. 6. 5. 4.],\\
          [3. 2. 1. 2. 3. 2. 1.],\\
          [6. 5. 4. 5. 6. 5. 4.],\\
          [9. 8. 7. 8. 9. 8. 7.],\\
          [6. 5. 4. 5. 6. 5. 4.],\\
          [3. 2. 1. 2. 3. 2. 1.]]\\ \f$
 *
 * If paddingMode==PADDING_MODE_SYMMETRIC, the output is as follows:
 *
      \f$[[5. 4. 4. 5. 6. 6. 5.],\\
          [2. 1. 1. 2. 3. 3. 2.],\\
          [2. 1. 1. 2. 3. 3. 2.],\\
          [5. 4. 4. 5. 6. 6. 5.],\\
          [8. 7. 7. 8. 9. 9. 8.],\\
          [8. 7. 7. 8. 9. 9. 8.],\\
          [5. 4. 4. 5. 6. 6. 5.]]\\ \f$
 *
 * @since 3.2
 * @version 1.0
 */
enum PaddingMode : byte {
    /** Constant (0 by default) padding. */
    PADDING_MODE_CONSTANT = 0,
    /** Reflection padding, which uses the content next to the input data to pad the values directly next to it. */
    PADDING_MODE_REFLECT = 1,
    /** Symmetric padding, which is similar to {@link PADDING_MODE_REFLECT}. Symmetric padding makes a copy of the input. */
    PADDING_MODE_SYMMETRIC = 2,
    /** Reserved. */
    PADDING_MODE_RESERVED = 3
};

/** @} */