/*
 * Copyright (C) 2017 The Android Open Source Project
 *
 * 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.
 */

package android.hardware.neuralnetworks@1.0;

/**
 * Operand types.
 *
 * The type of an operand in a model.
 *
 * Types prefaced with TENSOR_* must be used for tensor data (i.e., tensors
 * with at least one dimension). Types not prefaced by TENSOR_* represent
 * scalar values and must have no dimensions.
 *
 * Although we define many types, most operators accept just a few
 * types. Most used are {@link OperandType::TENSOR_FLOAT32},
 * {@link OperandType::TENSOR_QUANT8_ASYMM},
 * and {@link OperandType::INT32}.
 */
enum OperandType : int32_t {
    /** A 32 bit floating point scalar value. */
    FLOAT32 = 0,
    /** A signed 32 bit integer scalar value. */
    INT32 = 1,
    /** An unsigned 32 bit integer scalar value. */
    UINT32 = 2,
    /** A tensor of 32 bit floating point values. */
    TENSOR_FLOAT32 = 3,
    /** A tensor of 32 bit integer values. */
    TENSOR_INT32 = 4,
    /**
     * A tensor of 8 bit unsigned integers that represent real numbers.
     *
     * Attached to this tensor are two numbers that can be used to convert the
     * 8 bit integer to the real value and vice versa. These two numbers are:
     * - scale: a 32 bit floating point value greater than zero.
     * - zeroPoint: a 32 bit integer, in range [0, 255].
     *
     * The formula is:
     *   real_value = (integer_value - zeroPoint) * scale.
     */
    TENSOR_QUANT8_ASYMM = 5,

    /**
     * DEPRECATED. Since HAL version 1.2, extensions are the preferred
     * alternative to OEM operation and data types.
     *
     * OEM specific scalar value.
     */
    OEM                 = 10000,

    /**
     * DEPRECATED. Since HAL version 1.2, extensions are the preferred
     * alternative to OEM operation and data types.
     *
     * A tensor of OEM specific values.
     */
    TENSOR_OEM_BYTE     = 10001,
};

/**
 * Operation types.
 *
 * The type of an operation in a model.
 */
enum OperationType : int32_t {
    /**
     * Adds two tensors, element-wise.
     *
     * Takes two input tensors of identical {@link OperandType} and compatible
     * dimensions. The output is the sum of both input tensors, optionally
     * modified by an activation function.
     *
     * Two dimensions are compatible when:
     *     1. they are equal, or
     *     2. one of them is 1
     *
     * The size of the output is the maximum size along each dimension of the
     * input operands. It starts with the trailing dimensions, and works its
     * way forward.
     *
     * Example:
     *
     *     input1.dimension = {4, 1, 2}
     *     input2.dimension = {5, 4, 3, 1}
     *     output.dimension = {5, 4, 3, 2}
     *
     * Supported tensor {@link OperandType}:
     * * {@link OperandType::TENSOR_FLOAT32}
     * * {@link OperandType::TENSOR_QUANT8_ASYMM}
     *
     * Supported tensor rank: up to 4
     *
     * Inputs:
     * * 0: A tensor.
     * * 1: A tensor of the same {@link OperandType}, and compatible dimensions
     *      as input0.
     *      For a {@link OperandType::TENSOR_QUANT8_ASYMM} tensor,
     *      the scales and zeroPoint can be different from input0 scale and zeroPoint.
     * * 2: An {@link OperandType::INT32} scalar, and has to be one of the
     *      {@link FusedActivationFunc} values. Specifies the activation to
     *      invoke on the result.
     *
     * Outputs:
     * * 0: The sum, a tensor of the same {@link OperandType} as input0.
     *      For a {@link OperandType::TENSOR_QUANT8_ASYMM} tensor,
     *      the scale and zeroPoint can be different from inputs' scale and zeroPoint.
     */
    ADD = 0,

    /**
     * Performs a 2-D average pooling operation.
     *
     * The output dimensions are functions of the filter dimensions, stride, and
     * padding.
     *
     * The values in the output tensor are computed as:
     *
     *     output[b, i, j, channel] =
     *         sum_{di, dj}(
     *             input[b, strides[1] * i + di, strides[2] * j + dj, channel]
     *         ) / sum(1)
     *
     * Supported tensor {@link OperandType}:
     * * {@link OperandType::TENSOR_FLOAT32}
     * * {@link OperandType::TENSOR_QUANT8_ASYMM}
     *
     * Supported tensor rank: 4, with "NHWC" (i.e., Num_samples, Height, Width,
     * and Channels) data layout.
     *
     * Both explicit padding and implicit padding are supported.
     *
     * Inputs (explicit padding):
     * * 0: A 4-D tensor, of shape [batches, height, width, depth], specifying
     *      the input.
     * * 1: An {@link OperandType::INT32} scalar, specifying the padding on
     *      the left, in the ‘width’ dimension.
     * * 2: An {@link OperandType::INT32} scalar, specifying the padding on
     *      the right, in the ‘width’ dimension.
     * * 3: An {@link OperandType::INT32} scalar, specifying the padding on
     *      the top, in the ‘height’ dimension.
     * * 4: An {@link OperandType::INT32} scalar, specifying the padding on
     *      the bottom, in the ‘height’ dimension.
     * * 5: An {@link OperandType::INT32} scalar, specifying the stride when
     *      walking through input in the ‘width’ dimension.
     * * 6: An {@link OperandType::INT32} scalar, specifying the stride when
     *      walking through input in the ‘height’ dimension.
     * * 7: An {@link OperandType::INT32} scalar, specifying the filter
     *      width.
     * * 8: An {@link OperandType::INT32} scalar, specifying the filter
     *      height.
     * * 9: An {@link OperandType::INT32} scalar, and has to be one of the
     *      {@link FusedActivationFunc} values. Specifies the activation to
     *      invoke on the result.
     *
     * Inputs (implicit padding):
     * * 0: A 4-D tensor, of shape [batches, height, width, depth], specifying
     *      the input.
     * * 1: An {@link OperandType::INT32} scalar, specifying the implicit
     *      padding scheme, has to be one of the
     *      following values: {0 (NONE), 1 (SAME), 2 (VALID)}.
     * * 2: An {@link OperandType::INT32} scalar, specifying the stride when
     *      walking through input in the ‘width’ dimension.
     * * 3: An {@link OperandType::INT32} scalar, specifying the stride when
     *      walking through input in the ‘height’ dimension.
     * * 4: An {@link OperandType::INT32} scalar, specifying the filter
     *      width.
     * * 5: An {@link OperandType::INT32} scalar, specifying the filter
     *      height.
     * * 6: An {@link OperandType::INT32} scalar, and has to be one of the
     *      {@link FusedActivationFunc} values. Specifies the activation to
     *      invoke on the result.
     *
     * Outputs:
     * * 0: The output 4-D tensor, of shape
     *      [batches, out_height, out_width, depth].
     *      For a {@link OperandType::TENSOR_QUANT8_ASYMM} tensor,
     *      the scale and zeroPoint must be the same as input0.
     */
    AVERAGE_POOL_2D = 1,

    /**
     * Concatenates the input tensors along the given dimension.
     *
     * The input tensors must have identical {@link OperandType} and the same
     * dimensions except the dimension along the concatenation axis.
     *
     * Supported tensor {@link OperandType}:
     * * {@link OperandType::TENSOR_FLOAT32}
     * * {@link OperandType::TENSOR_QUANT8_ASYMM}
     *
     * Supported tensor rank: up to 4
     *
     * Inputs:
     * * 0 ~ n-1: The list of n input tensors, of shape
     *            [D0, D1, ..., Daxis(i), ..., Dm].
     *            All input tensors of
     *            {@link OperandType::TENSOR_QUANT8_ASYMM}
     *            must have the same scale and zeroPoint as the output tensor.
     * * n: An {@link OperandType::INT32} scalar, specifying the
     *      concatenation axis.
     *
     * Outputs:
     * * 0: The output, a tensor of the same {@link OperandType} as the input
     *      tensors. The output shape is [D0, D1, ..., sum(Daxis(i)), ..., Dm].
     *      For a {@link OperandType::TENSOR_QUANT8_ASYMM} tensor, the scale and zeroPoint
     *      values must be the same as the input tensors'.
     */
    CONCATENATION = 2,

    /**
     * Performs a 2-D convolution operation.
     *
     * The CONV_2D op sweeps a 2-D filter that can mix channels together over a
     * batch of images, applying the filter to each window of each image of the
     * appropriate size.
     *
     * The output dimensions are functions of the filter dimensions, stride, and
     * padding.
     *
     * The values in the output tensor are computed as:
     *
     *     output[b, i, j, channel] =
     *         sum_{di, dj, k} (
     *             input[b, strides[1] * i + di, strides[2] * j + dj, k] *
     *             filter[channel, di, dj, k]
     *         ) + bias[channel]
     *
     * Supported tensor {@link OperandType} configurations:
     * * 32 bit floating point:
     * * * {@link OperandType::TENSOR_FLOAT32} for input, filter, output, and bias.
     *
     * * Quantized:
     * * * {@link OperandType::TENSOR_QUANT8_ASYMM} for input, filter, and output.
     * * * {@link OperandType::TENSOR_INT32} for bias (with scale set to
     * * * input.scale * filter.scale).
     *
     * Supported tensor rank: 4, with "NHWC" (i.e., Num_samples, Height, Width,
     * and Channels) data layout.
     *
     * Both explicit padding and implicit padding are supported.
     *
     * Inputs (explicit padding):
     * * 0: A 4-D tensor, of shape [batches, height, width, depth_in],
     *      specifying the input.
     * * 1: A 4-D tensor, of shape
     *      [depth_out, filter_height, filter_width, depth_in], specifying the
     *      filter.
     * * 2: A 1-D tensor, of shape [depth_out], specifying the bias. For input
     *      tensor of type {@link OperandType::TENSOR_FLOAT32}
     *       the bias must be of the same type.
     *      For filter tensor of {@link OperandType::TENSOR_QUANT8_ASYMM},
     *      the bias should be of {@link OperandType::TENSOR_INT32}, with zeroPoint
     *      of 0 and bias_scale == input_scale * filter_scale.
     * * 3: An {@link OperandType::INT32} scalar, specifying the padding on
     *      the left, in the ‘width’ dimension.
     * * 4: An {@link OperandType::INT32} scalar, specifying the padding on
     *      the right, in the ‘width’ dimension.
     * * 5: An {@link OperandType::INT32} scalar, specifying the padding on
     *      the top, in the ‘height’ dimension.
     * * 6: An {@link OperandType::INT32} scalar, specifying the padding on
     *      the bottom, in the ‘height’ dimension.
     * * 7: An {@link OperandType::INT32} scalar, specifying the stride when
     *      walking through input in the ‘width’ dimension.
     * * 8: An {@link OperandType::INT32} scalar, specifying the stride when
     *      walking through input in the ‘height’ dimension.
     * * 9: An {@link OperandType::INT32} scalar, and has to be one of the
     *      {@link FusedActivationFunc} values. Specifies the activation to
     *      invoke on the result.
     *
     * Inputs (implicit padding):
     * * 0: A 4-D tensor, of shape [batches, height, width, depth_in],
     *      specifying the input.
     * * 1: A 4-D tensor, of shape
     *      [depth_out, filter_height, filter_width, depth_in], specifying the
     *      filter.
     * * 2: A 1-D tensor, of shape [depth_out], specifying the bias. For input
     *      tensor of type {@link OperandType::TENSOR_FLOAT32}
     *       the bias must be of the same
     *      type.
     *      For filter tensor of {@link OperandType::TENSOR_QUANT8_ASYMM},
     *      the bias should be of {@link OperandType::TENSOR_INT32}, with zeroPoint
     *      of 0 and bias_scale == input_scale * filter_scale.
     * * 3: An {@link OperandType::INT32} scalar, specifying the implicit
     *      padding scheme, has to be one of the
     *      following values: {0 (NONE), 1 (SAME), 2 (VALID)}.
     * * 4: An {@link OperandType::INT32} scalar, specifying the stride when
     *      walking through input in the ‘width’ dimension.
     * * 5: An {@link OperandType::INT32} scalar, specifying the stride when
     *      walking through input in the ‘height’ dimension.
     * * 6: An {@link OperandType::INT32} scalar, and has to be one of the
     *      {@link FusedActivationFunc} values. Specifies the activation to
     *      invoke on the result.
     *
     * Outputs:
     * * 0: The output 4-D tensor, of shape
     *      [batches, out_height, out_width, depth_out].
     *      For output tensor of
     *      {@link OperandType::TENSOR_QUANT8_ASYMM}, the following condition must
     *      be satisfied: output_scale > input_scale * filter_scale
     */
    CONV_2D = 3,

    /**
     * Performs a depthwise 2-D convolution operation.
     *
     * Given an input tensor of shape [batches, height, width, depth_in] and a
     * filter tensor of shape [1, filter_height, filter_width, depth_out]
     * containing depth_out convolutional filters of depth 1, DEPTHWISE_CONV
     * applies a different filter to each input channel (expanding from 1
     * channel to channel_multiplier channels for each), then concatenates the
     * results together.
     *
     * The output has depth_out = depth_in * depth_multiplier channels.
     * The output dimensions are functions of the filter dimensions, stride, and
     * padding.
     *
     * The values in the output tensor are computed as:
     *
     *     output[b, i, j, k * channel_multiplier + q] =
     *         sum_{di, dj} (
     *             input[b, strides[1] * i + di, strides[2] * j + dj, k] *
     *             filter[1, di, dj, k * channel_multiplier + q]
     *         ) + bias[k * channel_multiplier + q]
     *
     * Supported tensor {@link OperandType} configurations:
     * * 32 bit floating point:
     * * * {@link OperandType::TENSOR_FLOAT32} for input, filter, output, and bias.
     *
     * * Quantized:
     * * * {@link OperandType::TENSOR_QUANT8_ASYMM} for input, filter, and output.
     * * * {@link OperandType::TENSOR_INT32} for bias (with scale set to
     * * * input.scale * filter.scale).
     *
     * Supported tensor rank: 4, with "NHWC" (i.e., Num_samples, Height, Width,
     * and Channels) data layout.
     *
     * Both explicit padding and implicit padding are supported.
     *
     * Inputs (explicit padding):
     * * 0: A 4-D tensor, of shape [batches, height, width, depth_in],
     *      specifying the input.
     * * 1: A 4-D tensor, of shape [1, filter_height, filter_width, depth_out],
     *      specifying the filter.
     * * 2: A 1-D tensor, of shape [depth_out], specifying the bias. For input
     *      tensor of type {@link OperandType::TENSOR_FLOAT32}
     *       the bias must be of the same type.
     *      For filter tensor of {@link OperandType::TENSOR_QUANT8_ASYMM},
     *      the bias should be of {@link OperandType::TENSOR_INT32}, with zeroPoint
     *      of 0 and bias_scale == input_scale * filter_scale.
     * * 3: An {@link OperandType::INT32} scalar, specifying the padding on
     *      the left, in the ‘width’ dimension.
     * * 4: An {@link OperandType::INT32} scalar, specifying the padding on
     *      the right, in the ‘width’ dimension.
     * * 5: An {@link OperandType::INT32} scalar, specifying the padding on
     *      the top, in the ‘height’ dimension.
     * * 6: An {@link OperandType::INT32} scalar, specifying the padding on
     *      the bottom, in the ‘height’ dimension.
     * * 7: An {@link OperandType::INT32} scalar, specifying the stride when
     *      walking through input in the ‘width’ dimension.
     * * 8: An {@link OperandType::INT32} scalar, specifying the stride when
     *      walking through input in the ‘height’ dimension.
     * * 9: An {@link OperandType::INT32} scalar, specifying the depthwise
     *      multiplier.
     * * 10: An {@link OperandType::INT32} scalar, and has to be one of the
     *       {@link FusedActivationFunc} values. Specifies the activation to
     *       invoke on the result.
     *
     * Inputs (implicit padding):
     * * 0: A 4-D tensor, of shape [batches, height, width, depth_in],
     *      specifying the input.
     * * 1: A 4-D tensor, of shape [1, filter_height, filter_width, depth_out],
     *      specifying the filter.
     * * 2: A 1-D tensor, of shape [depth_out], specifying the bias. For input
     *      tensor of type {@link OperandType::TENSOR_FLOAT32}
     *       the bias must be of the same type.
     *      For filter tensor of {@link OperandType::TENSOR_QUANT8_ASYMM},
     *      the bias should be of {@link OperandType::TENSOR_INT32}, with zeroPoint
     *      of 0 and bias_scale == input_scale * filter_scale.
     * * 3: An {@link OperandType::INT32} scalar, specifying the implicit
     *      padding scheme, has to be one of the
     *      following values: {0 (NONE), 1 (SAME), 2 (VALID)}.
     * * 4: An {@link OperandType::INT32} scalar, specifying the stride when
     *      walking through input in the ‘width’ dimension.
     * * 5: An {@link OperandType::INT32} scalar, specifying the stride when
     *      walking through input in the ‘height’ dimension.
     * * 6: An {@link OperandType::INT32} scalar, specifying the depthwise
     *      multiplier.
     * * 7: An {@link OperandType::INT32} scalar, and has to be one of the
     *      {@link FusedActivationFunc} values. Specifies the activation to
     *      invoke on the result.
     *
     * Outputs:
     * * 0: The output 4-D tensor, of shape
     *      [batches, out_height, out_width, depth_out]. For
     *      output tensor of {@link OperandType::TENSOR_QUANT8_ASYMM},
     *      the following condition must be satisfied:
     *      output_scale > input_scale * filter_scale
     */
    DEPTHWISE_CONV_2D = 4,

    /**
     * Rearranges data from depth into blocks of spatial data.
     *
     * More specifically, this op outputs a copy of the input tensor where
     * values from the depth dimension are moved in spatial blocks to the height
     * and width dimensions. The value block_size indicates the input block size
     * and how the data is moved.
     *
     * Chunks of data of size block_size * block_size from depth are rearranged
     * into non-overlapping blocks of size block_size x block_size.
     *
     * The width of the output tensor is input_depth * block_size, whereas the
     * height is input_height * block_size. The depth of the input tensor must
     * be divisible by block_size * block_size
     *
     * Supported tensor {@link OperandType}:
     * * {@link OperandType::TENSOR_FLOAT32}
     * * {@link OperandType::TENSOR_QUANT8_ASYMM}
     *
     * Supported tensor rank: 4, with "NHWC" (i.e., Num_samples, Height, Width,
     * and Channels) data layout.
     *
     * Inputs:
     * * 0: A 4-D tensor, of shape [batches, height, width, depth_in],
     *      specifying the input.
     * * 1: An {@link OperandType::INT32} scalar, specifying the block_size.
     *      block_size must be >=1 and block_size * block_size must be a divisor
     *      of the input depth.
     *
     * Outputs:
     * * 0: The output 4-D tensor, of shape [batch, height*block_size,
     *      width*block_size, depth/(block_size*block_size)].
     *      For a {@link OperandType::TENSOR_QUANT8_ASYMM} tensor,
     *      the scale and zeroPoint must be the same as input0.
     */
    DEPTH_TO_SPACE = 5,

    /**
     * Dequantizes the input tensor.
     *
     * The formula is:
     *
     *     output = (input - zeroPoint) * scale.
     *
     * Supported input tensor {@link OperandType}:
     * * {@link OperandType::TENSOR_QUANT8_ASYMM}
     *
     * Supported output tensor {@link OperandType}:
     * * {@link OperandType::TENSOR_FLOAT32}.
     *
     * Supported tensor rank: up to 4
     *
     * Inputs:
     * * 0: A tensor.
     *
     * Outputs:
     * * 0: A tensor with the same shape as input0.
     */
    DEQUANTIZE = 6,

    /**
     * Looks up sub-tensors in the input tensor.
     *
     * This operator takes for input a tensor of values (Values) and
     * a one-dimensional tensor of selection indices (Lookups).
     * The output tensor is the concatenation of sub-tensors of Values as
     * selected by Lookups.
     *
     * Think of Values as being sliced along its first dimension:
     * The entries in Lookups select which slices are concatenated together
     * to create the output tensor.
     *
     * For example, if Values has shape of [40, 200, 300] and
     * Lookups has shape of [3], all three values found in Lookups are
     * expected to be between 0 and 39. The resulting tensor must
     * have shape of [3, 200, 300].
     *
     * If a value in Lookups is out of bounds, the operation must fail
     * and an error must be reported.
     *
     * Supported value tensor {@link OperandType}:
     * * {@link OperandType::TENSOR_FLOAT32}
     *
     * Supported value tensor rank: from 2
     *
     * Inputs:
     * * 0: Lookups. A 1-D tensor of {@link OperandType::TENSOR_INT32}.
     *      The values are indices into the first dimension of Values.
     * * 1: Values. An n-D tensor, where n >= 2, from which sub-tensors are
     *      extracted.
     *
     * Output:
     * * 0: A n-D tensor with the same rank and shape as the Values
     *      tensor, except for the first dimension which has the same size
     *      as Lookups' only dimension.
     *      For a {@link OperandType::TENSOR_QUANT8_ASYMM} tensor,
     *      the scale and zeroPoint must be the same as input1.
     */
    EMBEDDING_LOOKUP = 7,

    /**
     * Computes element-wise floor() on the input tensor.
     *
     * Supported tensor {@link OperandType}:
     * * {@link OperandType::TENSOR_FLOAT32}
     *
     * Supported tensor rank: up to 4
     *
     * Inputs:
     * * 0: A tensor.
     *
     * Outputs:
     * * 0: The output tensor, of the same {@link OperandType} and dimensions as
     *      the input tensor.
     */
    FLOOR = 8,

    /**
     * Denotes a fully (densely) connected layer, which connects all elements
     * in the input tensor with each element in the output tensor.
     *
     * This layer implements the operation:
     *
     *     outputs = activation(inputs * weights’ + bias)
     *
     * Supported tensor {@link OperandType}:
     * * {@link OperandType::TENSOR_FLOAT32}
     * * {@link OperandType::TENSOR_QUANT8_ASYMM}
     *
     * Supported tensor rank: up to 4.
     *
     * Inputs:
     * * 0: A tensor of at least rank 2, specifying the input. If rank is
     *      greater than 2, then it gets flattened to a 2-D Tensor. The
     *      (flattened) 2-D Tensor is reshaped (if necessary) to
     *      [batch_size, input_size], where "input_size" corresponds to the
     *      number of inputs to the layer, matching the second dimension of
     *      weights, and "batch_size" is calculated by dividing the number of
     *      elements by "input_size".
     * * 1: A 2-D tensor, specifying the weights, of shape
     *      [num_units, input_size], where "num_units" corresponds to the number
     *      of output nodes.
     * * 2: A 1-D tensor, of shape [num_units], specifying the bias. For input
     *      tensor of {@link OperandType::TENSOR_FLOAT32}, the bias should
     *      also be of {@link OperandType::TENSOR_FLOAT32}.
     *      For input tensor of {@link OperandType::TENSOR_QUANT8_ASYMM},
     *      the bias should be of {@link OperandType::TENSOR_INT32},
     *      with zeroPoint of 0 and bias_scale == input_scale * filter_scale.
     * * 3: An {@link OperandType::INT32} scalar, and has to be one of the
     *      {@link FusedActivationFunc} values. Specifies the activation to
     *      invoke on the result.
     *
     * Outputs:
     * * 0: The output tensor, of shape [batch_size, num_units]. For
     *      output tensor of {@link OperandType::TENSOR_QUANT8_ASYMM}, the following
     *      condition must be satisfied: output_scale > input_scale * filter_scale.
     */
    FULLY_CONNECTED = 9,

    /**
     * Looks up sub-tensors in the input tensor using a key-value map.
     *
     * This operator takes for input a tensor of values (Values),
     * a one-dimensional tensor of selection values (Lookups) and
     * a one-dimensional tensor that maps these values to Values
     * indexes. The output tensor is the concatenation of sub-tensors of
     * Values as selected by Lookups via Keys.
     *
     * Think of Values as being sliced along its outer-most dimension.
     * The output is a concatenation of selected slices, with one slice
     * for each entry of Lookups. The slice selected is the one at the
     * same index as the Maps entry that matches the value in Lookups.
     *
     * For a hit, the corresponding sub-tensor of Values is included
     * in the Output tensor. For a miss, the corresponding sub-tensor in
     * Output must have zero values.
     *
     * For example, if Values has shape of [40, 200, 300],
     * Keys should have a shape of [40]. If Lookups tensor has shape
     * of [3], three slices are being concatenated, so the resulting tensor
     * must have the shape of [3, 200, 300]. If the first entry in Lookups
     * has the value 123456, that value must be located in Keys tensor.
     * If the sixth entry of Keys contains 123456, the sixth slice of Values
     * must be selected. If no entry in Keys has 123456, a slice of zeroes
     * must be concatenated.
     *
     * Supported value tensor {@link OperandType}:
     * * {@link OperandType::TENSOR_FLOAT32}
     * * {@link OperandType::TENSOR_INT32}
     * * {@link OperandType::TENSOR_QUANT8_ASYMM}
     *
     * Supported value tensor rank: from 2
     *
     * Inputs:
     * * 0: Lookups. A 1-D {@link OperandType::TENSOR_INT32} tensor with
     *      shape [ k ].
     * * 1: Keys. A 1-D {@link OperandType::TENSOR_INT32} tensor with shape
     *      [ n ]; Keys and Values pair represent a map, i.e., the ith element
     *      in Keys (Keys[i]) is the key to select the ith sub-tensor in Values
     *      (Values[i]), where 0 <= i <= n-1. Keys tensor *MUST* be sorted in
     *      ascending order.
     * * 2: Values. A tensor with shape of [ n, … ]; i.e., the first dimension
     *      must be n.
     *
     * Outputs:
     * * 0: Output. A tensor with shape [ k …].
     *      For a {@link OperandType::TENSOR_QUANT8_ASYMM} tensor,
     *      the scale and zeroPoint must be the same as input2.
     * * 1: Hits. A boolean tensor with shape [ k ] indicates whether the lookup
     *      hits (True) or not (False).
     *      Stored as {@link OperandType::TENSOR_QUANT8_ASYMM} with offset 0
     *      and scale 1.0f.
     *      A non-zero byte represents True, a hit. A zero indicates otherwise.
     */
    HASHTABLE_LOOKUP = 10,

    /**
     * Applies L2 normalization along the axis dimension.
     *
     * The values in the output tensor are computed as:
     *
     *     output[batch, row, col, channel] =
     *         input[batch, row, col, channel] /
     *         sqrt(sum_{c} pow(input[batch, row, col, c], 2))
     *
     * Supported tensor {@link OperandType}:
     * * {@link OperandType::TENSOR_FLOAT32}
     *
     * Supported tensor rank: 4, with "NHWC" data layout (i.e., Num_samples,
     * Height, Width, and Channels).
     *
     * Inputs:
     * * 0: A 4-D tensor, specifying the tensor to be normalized.
     *
     * Outputs:
     * * 0: A tensor of the same {@link OperandType} and same shape as input0.
     */
    L2_NORMALIZATION = 11,

    /**
     * Performs an 2-D L2 pooling operation.
     *
     * The output dimensions are functions of the filter dimensions, stride, and
     * padding.
     *
     * The values in the output tensor are computed as:
     *
     *     output[b, i, j, c] =
     *         sqrt(sum_{di, dj} pow(input[b, strides[1] * i + di, strides[2] * j + dj, c], 2) /
     *              sum(1))
     *
     * Supported tensor {@link OperandType}:
     * * {@link OperandType::TENSOR_FLOAT32}
     *
     * Supported tensor rank: 4, with "NHWC" (i.e., Num_samples, Height, Width,
     * and Channels) data layout.
     *
     * Both explicit padding and implicit padding are supported.
     *
     * Inputs (explicit padding):
     * * 0: A 4-D tensor, of shape [batches, height, width, depth], specifying
     *      the input.
     * * 1: An {@link OperandType::INT32} scalar, specifying the padding on
     *      the left, in the ‘width’ dimension.
     * * 2: An {@link OperandType::INT32} scalar, specifying the padding on
     *      the right, in the ‘width’ dimension.
     * * 3: An {@link OperandType::INT32} scalar, specifying the padding on
     *      the top, in the ‘height’ dimension.
     * * 4: An {@link OperandType::INT32} scalar, specifying the padding on
     *      the bottom, in the ‘height’ dimension.
     * * 5: An {@link OperandType::INT32} scalar, specifying the stride when
     *      walking through input in the ‘width’ dimension.
     * * 6: An {@link OperandType::INT32} scalar, specifying the stride when
     *      walking through input in the ‘height’ dimension.
     * * 7: An {@link OperandType::INT32} scalar, specifying the filter
     *      width.
     * * 8: An {@link OperandType::INT32} scalar, specifying the filter
     *      height.
     * * 9: An {@link OperandType::INT32} scalar, and has to be one of the
     *      {@link FusedActivationFunc} values. Specifies the activation to
     *      invoke on the result.
     *
     * Inputs (implicit padding):
     * * 0: A 4-D tensor, of shape [batches, height, width, depth], specifying
     *      the input.
     * * 1: An {@link OperandType::INT32} scalar, specifying the implicit
     *      padding scheme, has to be one of the
     *      following values: {0 (NONE), 1 (SAME), 2 (VALID)}.
     * * 2: An {@link OperandType::INT32} scalar, specifying the stride when
     *      walking through input in the ‘width’ dimension.
     * * 3: An {@link OperandType::INT32} scalar, specifying the stride when
     *      walking through input in the ‘height’ dimension.
     * * 4: An {@link OperandType::INT32} scalar, specifying the filter
     *      width.
     * * 5: An {@link OperandType::INT32} scalar, specifying the filter
     *      height.
     * * 6: An {@link OperandType::INT32} scalar, and has to be one of the
     *      {@link FusedActivationFunc} values. Specifies the activation to
     *      invoke on the result.
     *
     * Outputs:
     * * 0: The output 4-D tensor, of shape
     *      [batches, out_height, out_width, depth].
     */
    L2_POOL_2D = 12,

    /**
     * Applies Local Response Normalization along the depth dimension.
     *
     * The 4-D input tensor is treated as a 3-D array of 1-D vectors (along the
     * last dimension), and each vector is normalized independently. Within a
     * given vector, each component is divided by the weighted, squared sum of
     * inputs within depth_radius.
     *
     * The output is calculated using this formula:
     *
     *     sqr_sum[a, b, c, d] = sum(
     *         pow(input[a, b, c, d - depth_radius : d + depth_radius + 1], 2))
     *     output = input / pow((bias + alpha * sqr_sum), beta)
     *
     * Supported tensor {@link OperandType}:
     * * {@link OperandType::TENSOR_FLOAT32}
     *
     * Supported tensor rank: 4, with "NHWC" data layout.
     *
     * Inputs:
     * * 0: A 4-D tensor, of shape [batches, height, width, depth], specifying
     *      the input.
     * * 1: An {@link OperandType::INT32} scalar, specifying the radius of
     *      the normalization window.
     * * 2: A scalar, specifying the bias, must not be zero.
     *      For input tensor of {@link OperandType::TENSOR_FLOAT32}, the bias
     *      value must be of {@link OperandType::FLOAT32}.
     * * 3: A scalar, specifying the scale factor, alpha.
     *      For input tensor of {@link OperandType::TENSOR_FLOAT32}, the
     *      alpha value must be of {@link OperandType::FLOAT32}.
     * * 4: A scalar, specifying the exponent, beta.
     *      For input tensor of {@link OperandType::TENSOR_FLOAT32}, the beta
     *      value must be of {@link OperandType::FLOAT32}.
     *
     * Outputs:
     * * 0: The output tensor of same shape as input0.
     */
    LOCAL_RESPONSE_NORMALIZATION = 13,

    /**
     * Computes sigmoid activation on the input tensor element-wise.
     *
     * The output is calculated using this formula:
     *
     *     output = 1 / (1 + exp(-input))
     *
     * Supported tensor {@link OperandType}:
     * * {@link OperandType::TENSOR_FLOAT32}
     * * {@link OperandType::TENSOR_QUANT8_ASYMM}
     *
     * Supported tensor rank: up to 4.
     *
     * Inputs:
     * * 0: A tensor, specifying the input.
     *
     * Outputs:
     * * 0: The output tensor of same shape as input0.
     *      For {@link OperandType::TENSOR_QUANT8_ASYMM},
     *      the scale must be 1.f / 256 and the zeroPoint must be 0.
     */
    LOGISTIC = 14,

    /**
     * Projects an input to a bit vector via locality senstive hashing.
     *
     * Supported input tensor {@link OperandType}:
     * * {@link OperandType::TENSOR_FLOAT32}
     * * {@link OperandType::TENSOR_INT32}
     * * {@link OperandType::TENSOR_QUANT8_ASYMM}
     *
     * Supported input tensor rank: from 1
     *
     * Inputs:
     * * 0: Hash functions. Dim.size == 2, DataType: Float.
     *      Tensor[0].Dim[0]: Number of hash functions.
     *      Tensor[0].Dim[1]: Number of projected output bits generated by each
     *      hash function.
     *      If the projection type is Sparse:
     *      Tensor[0].Dim[1] + ceil(log2(Tensor[0].Dim[0])) <= 32
     *
     * * 1: Input. Dim.size >= 1, no restriction on DataType.
     * * 2: Weight. Optional. Dim.size == 1, DataType: Float.
     *      If not set, each input element is considered to have the same weight
     *      of 1.0.
     *      Tensor[1].Dim[0] == Tensor[2].Dim[0]
     * * 3: Type:
     *        Sparse:
     *          Value LSHProjectionType_SPARSE(=1).
     *          Computed bit vector is considered to be sparse.
     *          Each output element is an int32 made up of multiple bits
     *          computed from hash functions.
     *
     *        Dense:
     *          Value LSHProjectionType_DENSE(=2).
     *          Computed bit vector is considered to be dense. Each output
     *          element represents a bit and can take the value of either
     *          0 or 1.
     *
     * Outputs:
     * * 0: If the projection type is Sparse:
     *      Output.Dim == { Tensor[0].Dim[0] }
     *      A tensor of int32 that represents hash signatures.
     *
     *      If the projection type is Dense:
     *      Output.Dim == { Tensor[0].Dim[0] * Tensor[0].Dim[1] }
     *      A flattened tensor that represents projected bit vectors.
     */
    LSH_PROJECTION = 15,

    /**
     * Performs a single time step in a Long Short-Term Memory (LSTM) layer
     *
     * The LSTM operation is described by the following equations.
     *
     * \f{eqnarray*}{
     * i_t =& \sigma(W_{xi}x_t+W_{hi}h_{t-1}+W_{ci}C_{t-1}+b_i) & \\
     * f_t =& \sigma(W_{xf}x_t+W_{hf}h_{t-1}+W_{cf}C_{t-1}+b_f) & \\
     * C_t =& clip(f_t \odot C_{t-1} + i_t \odot
     *        g(W_{xc}x_t+W_{hc}h_{t-1}+b_c),\ t_{cell}) & \\
     * o_t =& \sigma(W_{xo}x_t+W_{ho}h_{t-1}+W_{co}C_t+b_o) & \\
     *      & & \\
     *      & clip(W_{proj}(o_t \odot g(C_t))+b_{proj},\ t_{proj})
     *      & if\ there\ is\ a\ projection; \\
     * h_t =& & \\
     *      & o_t \odot g(C_t) & otherwise. \\
     * \f}
     * Where:
     * * \f$x_t\f$ is the input,
     * * \f$i_t\f$ is the input gate,
     * * \f$f_t\f$ is the forget gate,
     * * \f$C_t\f$ is the cell state,
     * * \f$o_t\f$ is the output,
     * * \f$h_t\f$ is the output state,
     * * \f$\sigma\f$ is the logistic sigmoid function,
     * * \f$g\f$ is the cell input and cell output activation function, usually
     *   \f$tahn\f$,
     * * \f$W_{xi}\f$ is the input-to-input weight matrix,
     * * \f$W_{hi}\f$ is the recurrent to input weight matrix,
     * * \f$W_{ci}\f$ is the cell-to-input weight matrix,
     * * \f$b_i\f$ is the input gate bias,
     * * \f$W_{xf}\f$ is the input-to-forget weight matrix,
     * * \f$W_{hf}\f$ is the recurrent-to-forget weight matrix,
     * * \f$W_{cf}\f$ is the cell-to-forget weight matrix,
     * * \f$b_f\f$ is the forget gate bias,
     * * \f$W_{xc}\f$ is the input-to-cell weight matrix,
     * * \f$W_{hc}\f$ is the recurrent-to-cell weight matrix,
     * * \f$b_c\f$ is the cell bias,
     * * \f$W_{xo}\f$ is the input-to-output weight matrix,
     * * \f$W_{ho}\f$ is the recurrent-to-output weight matrix,
     * * \f$W_{co}\f$ is the cell-to-output weight matrix,
     * * \f$b_o\f$ is the output gate bias,
     * * \f$W_{proj}\f$ is the projection weight matrix,
     * * \f$b_{proj}\f$ is the projection bias,
     * * \f$t_{cell}\f$ is the threshold for clipping the cell state, and
     * * \f$t_{proj}\f$ is the threshold for clipping the projected output.
     * * \f$\odot\f$ is the
     *   <a href="https://en.wikipedia.org/wiki/Hadamard_product_(matrices)">
     *   Hadamard product</a> that takes two matrices and produces another
     *   matrix, each element of which is the product of the corresponding
     *   elements of the input matrices.
     *
     * The operation has the following independently optional inputs:
     * * The cell-to-input weights (\f$W_{ci}\f$), cell-to-forget weights
     *   (\f$W_{cf}\f$) and cell-to-output weights (\f$W_{co}\f$) either all
     *   have values or neither of them have values (i.e., all set to null). If
     *   they have values, the peephole optimization is used.
     * * The input-to-input weights (\f$W_{xi}\f$), recurrent-to-input weights
     *   (\f$W_{hi}\f$) and input gate bias (\f$b_i\f$) either all have values,
     *   or none of them have values. If they have no values, coupling of input
     *   and forget gates (CIFG) is used, in which case the input gate
     *   (\f$i_t\f$) is calculated using the following equation instead.
     *   \f{eqnarray*}{
     *   i_t = 1 - f_t
     *   \f}
     *   In case peephole optimization is used and CIFG is not used
     *   cell-to-input (\f$W_{ci}\f$) weights must be present. Otherwise, the
     *   cell-to-input weights must have no value.
     * * The projection weights (\f$W_{proj}\f$) is required only for the
     *   recurrent projection layer, and should otherwise have no value.
     * * The projection bias (\f$b_{proj}\f$) may (but not required to) have a
     *   value if the recurrent projection layer exists, and should otherwise
     *   have no value.
     *
     * References:
     *
     * The default non-peephole non-CIFG implementation is based on:
     * http://www.bioinf.jku.at/publications/older/2604.pdf
     * S. Hochreiter and J. Schmidhuber. "Long Short-Term Memory". Neural
     * Computation, 9(8):1735-1780, 1997.
     *
     * The peephole implementation and projection layer is based on:
     * https://research.google.com/pubs/archive/43905.pdf
     * Hasim Sak, Andrew Senior, and Francoise Beaufays. "Long short-term memory
     * recurrent neural network architectures for large scale acoustic
     * modeling." INTERSPEECH, 2014.
     * (However, the concept of peephole optimization was introduced in work
     * prior to this paper.)
     *
     * The coupling of input and forget gate (CIFG) is based on:
     * http://arxiv.org/pdf/1503.04069.pdf
     * Greff et al. "LSTM: A Search Space Odyssey"
     *
     * Supported tensor {@link OperandType}:
     * * {@link OperandType::TENSOR_FLOAT32}
     *
     * All input and output tensors must be of the same type.
     *
     * Inputs:
     * * 0: The input (\f$x_t\f$).
     *      A 2-D tensor of shape [batch_size, input_size], where “batch_size”
     *      corresponds to the batching dimension, and “input_size” is the size
     *      of the input.
     * * 1: The input-to-input weights (\f$W_{xi}\f$). Optional.
     *      A 2-D tensor of shape [num_units, input_size], where “num_units”
     *      corresponds to the number of cell units.
     * * 2: The input-to-forget weights (\f$W_{xf}\f$).
     *      A 2-D tensor of shape [num_units, input_size].
     * * 3: The input-to-cell weights (\f$W_{xc}\f$).
     *      A 2-D tensor of shape [num_units, input_size].
     * * 4: The input-to-output weights (\f$W_{xo}\f$).
     *      A 2-D tensor of shape [num_units, input_size].
     * * 5: The recurrent-to-input weights (\f$W_{hi}\f$). Optional.
     *      A 2-D tensor of shape [num_units, output_size], where “output_size”
     *      corresponds to either the number of cell units (i.e., “num_units”),
     *      or the second dimension of the “projection_weights”, if defined.
     * * 6: The recurrent-to-forget weights (\f$W_{hf}\f$).
     *      A 2-D tensor of shape [num_units, output_size].
     * * 7: The recurrent-to-cell weights (\f$W_{hc}\f$).
     *      A 2-D tensor of shape [num_units, output_size].
     * * 8: The recurrent-to-output weights (\f$W_{ho}\f$).
     *      A 2-D tensor of shape [num_units, output_size].
     * * 9: The cell-to-input weights (\f$W_{ci}\f$). Optional.
     *      A 1-D tensor of shape [num_units].
     * * 10:The cell-to-forget weights (\f$W_{cf}\f$). Optional.
     *      A 1-D tensor of shape [num_units].
     * * 11:The cell-to-output weights (\f$W_{co}\f$). Optional.
     *      A 1-D tensor of shape [num_units].
     * * 12:The input gate bias (\f$b_i\f$). Optional.
     *      A 1-D tensor of shape [num_units].
     * * 13:The forget gate bias (\f$b_f\f$).
     *      A 1-D tensor of shape [num_units].
     * * 14:The cell bias (\f$b_c\f$).
     *      A 1-D tensor of shape [num_units].
     * * 15:The output gate bias (\f$b_o\f$).
     *      A 1-D tensor of shape [num_units].
     * * 16:The projection weights (\f$W_{proj}\f$). Optional.
     *      A 2-D tensor of shape [output_size, num_units].
     * * 17:The projection bias (\f$b_{proj}\f$). Optional.
     *      A 1-D tensor of shape [output_size].
     * * 18:The output state (in) (\f$h_{t-1}\f$).
     *      A 2-D tensor of shape [batch_size, output_size].
     * * 19:The cell state (in) (\f$C_{t-1}\f$).
     *      A 2-D tensor of shape [batch_size, num_units].
     * * 20:The activation function (\f$g\f$).
     *      A value indicating the activation function:
     *      <ul>
     *      <li>0: None;
     *      <li>1: Relu;
     *      <li>3: Relu6;
     *      <li>4: Tanh;
     *      <li>6: Sigmoid.
     *      </ul>
     * * 21:The clipping threshold (\f$t_{cell}\f$) for the cell state, such
     *      that values are bound within [-cell_clip, cell_clip]. If set to 0.0
     *      then clipping is disabled.
     * * 22:The clipping threshold (\f$t_{proj}\f$) for the output from the
     *      projection layer, such that values are bound within
     *      [-proj_clip, proj_clip]. If set to 0.0 then clipping is disabled.
     *
     * Outputs:
     * * 0: The scratch buffer.
     *      A 2-D tensor of shape [batch_size, num_units * 3] with CIFG, or
     *      [batch_size, num_units * 4] without CIFG.
     * * 1: The output state (out) (\f$h_t\f$).
     *      A 2-D tensor of shape [batch_size, output_size].
     * * 2: The cell state (out) (\f$C_t\f$).
     *      A 2-D tensor of shape [batch_size, num_units].
     * * 3: The output (\f$o_t\f$).
     *      A 2-D tensor of shape [batch_size, output_size]. This is effectively
     *      the same as the current “output state (out)” value.
     */
    LSTM = 16,

    /**
     * Performs an 2-D max pooling operation.
     *
     * The output dimensions are functions of the filter dimensions, stride, and
     * padding.
     *
     * The values in the output tensor are computed as:
     *
     *     output[b, i, j, channel] =
     *         max_{di, dj} (
     *             input[b, strides[1] * i + di, strides[2] * j + dj, channel]
     *         )
     *
     * Supported tensor {@link OperandType}:
     * * {@link OperandType::TENSOR_FLOAT32}
     * * {@link OperandType::TENSOR_QUANT8_ASYMM}
     *
     * Supported tensor rank: 4, with "NHWC" (i.e., Num_samples, Height, Width,
     * and Channels) data layout.
     *
     * Both explicit padding and implicit padding are supported.
     *
     * Inputs (explicit padding):
     * * 0: A 4-D tensor, of shape [batches, height, width, depth], specifying
     *      the input.
     * * 1: An {@link OperandType::INT32} scalar, specifying the padding on
     *      the left, in the ‘width’ dimension.
     * * 2: An {@link OperandType::INT32} scalar, specifying the padding on
     *      the right, in the ‘width’ dimension.
     * * 3: An {@link OperandType::INT32} scalar, specifying the padding on
     *      the top, in the ‘height’ dimension.
     * * 4: An {@link OperandType::INT32} scalar, specifying the padding on
     *      the bottom, in the ‘height’ dimension.
     * * 5: An {@link OperandType::INT32} scalar, specifying the stride when
     *      walking through input in the ‘width’ dimension.
     * * 6: An {@link OperandType::INT32} scalar, specifying the stride when
     *      walking through input in the ‘height’ dimension.
     * * 7: An {@link OperandType::INT32} scalar, specifying the filter
     *      width.
     * * 8: An {@link OperandType::INT32} scalar, specifying the filter
     *      height.
     * * 9: An {@link OperandType::INT32} scalar, and has to be one of the
     *      {@link FusedActivationFunc} values. Specifies the activation to
     *      invoke on the result.
     *
     * Inputs (implicit padding):
     * * 0: A 4-D tensor, of shape [batches, height, width, depth], specifying
     *      the input.
     * * 1: An {@link OperandType::INT32} scalar, specifying the implicit
     *      padding scheme, has to be one of the
     *      following values: {0 (NONE), 1 (SAME), 2 (VALID)}.
     * * 2: An {@link OperandType::INT32} scalar, specifying the stride when
     *      walking through input in the ‘width’ dimension.
     * * 3: An {@link OperandType::INT32} scalar, specifying the stride when
     *      walking through input in the ‘height’ dimension.
     * * 4: An {@link OperandType::INT32} scalar, specifying the filter
     *      width.
     * * 5: An {@link OperandType::INT32} scalar, specifying the filter
     *      height.
     * * 6: An {@link OperandType::INT32} scalar, and has to be one of the
     *      {@link FusedActivationFunc} values. Specifies the activation to
     *      invoke on the result.
     *
     * Outputs:
     * * 0: The output 4-D tensor, of shape
     *      [batches, out_height, out_width, depth].
     *      For a {@link OperandType::TENSOR_QUANT8_ASYMM} tensor,
     *      the scale and zeroPoint must be the same as input0.
     */
    MAX_POOL_2D = 17,

    /**
     * Multiplies two tensors, element-wise.
     *
     * Takes two input tensors of identical {@link OperandType} and compatible
     * dimensions. The output is the product of both input tensors, optionally
     * modified by an activation function.
     *
     * Two dimensions are compatible when:
     *     1. they are equal, or
     *     2. one of them is 1
     *
     * The size of the resulting output is the maximum size along each dimension
     * of the input operands. It starts with the trailing dimensions, and works
     * its way forward.
     *
     * Supported tensor {@link OperandType}:
     * * {@link OperandType::TENSOR_FLOAT32}
     * * {@link OperandType::TENSOR_QUANT8_ASYMM}
     *
     * Supported tensor rank: up to 4
     *
     * Inputs:
     * * 0: A tensor.
     * * 1: A tensor of the same {@link OperandType}, and compatible dimensions
     *      as input0.
     * * 2: An {@link OperandType::INT32} scalar, and has to be one of the
     *      {@link FusedActivationFunc} values. Specifies the activation to
     *      invoke on the result.
     *
     * Outputs:
     * * 0: The product, a tensor of the same {@link OperandType} as input0.
     *      For output tensor of {@link OperandType::TENSOR_QUANT8_ASYMM},
     *      the following condition must be satisfied:
     *      output_scale > input1_scale * input2_scale.
     */
    MUL = 18,

    /**
     * Computes rectified linear activation on the input tensor element-wise.
     *
     * The output is calculated using this formula:
     *
     *     output = max(0, input)
     *
     * Supported tensor {@link OperandType}:
     * * {@link OperandType::TENSOR_FLOAT32}
     * * {@link OperandType::TENSOR_QUANT8_ASYMM}
     *
     * Supported tensor rank: up to 4.
     *
     * Inputs:
     * * 0: A tensor, specifying the input.
     *
     * Outputs:
     * * 0: The output tensor of same shape as input0.
     *      For a {@link OperandType::TENSOR_QUANT8_ASYMM} tensor,
     *      the scale and zeroPoint must be the same as input0.
     */
    RELU = 19,

    /**
     * Computes rectified linear 1 activation on the input tensor element-wise.
     *
     * The output is calculated using this formula:
     *
     *     output = min(1.f, max(-1.f, input))
     *
     * Supported tensor {@link OperandType}:
     * * {@link OperandType::TENSOR_FLOAT32}
     * * {@link OperandType::TENSOR_QUANT8_ASYMM}
     *
     * Supported tensor rank: up to 4.
     *
     * Inputs:
     * * 0: A tensor, specifying the input.
     *
     * Outputs:
     * * 0: The output tensor of the same shape as input0.
     *      For a {@link OperandType::TENSOR_QUANT8_ASYMM} tensor,
     *      the scale and zeroPoint must be the same as input0.
     */
    RELU1 = 20,

    /**
     * Computes rectified linear 6 activation on the input tensor element-wise.
     *
     * The output is calculated using this formula:
     *
     *     output = min(6, max(0, input))
     *
     * Supported tensor {@link OperandType}:
     * * {@link OperandType::TENSOR_FLOAT32}
     * * {@link OperandType::TENSOR_QUANT8_ASYMM}
     *
     * Supported tensor rank: up to 4.
     *
     * Inputs:
     * * 0: A tensor, specifying the input.
     *
     * Outputs:
     * * 0: The output tensor of same shape as input0.
     *      For a {@link OperandType::TENSOR_QUANT8_ASYMM} tensor,
     *      the scale and zeroPoint must be the same as input0.
     */
    RELU6 = 21,

    /**
     * Reshapes a tensor.
     *
     * Given tensor, this operation returns a tensor that has the same values as
     * tensor, but with a newly specified shape.
     *
     * Supported tensor {@link OperandType}:
     * * {@link OperandType::TENSOR_FLOAT32}
     * * {@link OperandType::TENSOR_QUANT8_ASYMM}
     *
     * Supported tensor rank: up to 4.
     *
     * Inputs:
     * * 0: A tensor, specifying the tensor to be reshaped.
     * * 1: A 1-D tensor of {@link OperandType::TENSOR_INT32}, defining the
     *      shape of the output tensor. The number of elements implied by shape
     *      must be the same as the number of elements in the input tensor.
     *
     *      If one component of shape is the special value -1, the size of that
     *      dimension is computed so that the total size remains constant. In
     *      particular, a shape of [-1] flattens into 1-D. At most one component
     *      of shape can be -1.
     *
     * Outputs:
     * * 0: The output tensor, of shape specified by the input shape.
     *      For a {@link OperandType::TENSOR_QUANT8_ASYMM} tensor,
     *      the scale and zeroPoint must be the same as input0.
     */
    RESHAPE = 22,

    /**
     * Resizes images to given size using the bilinear interpretation.
     *
     * Resized images must be distorted if their output aspect ratio is not the
     * same as input aspect ratio. The corner pixels of output may not be the
     * same as corner pixels of input.
     *
     * Supported tensor {@link OperandType}:
     * * {@link OperandType::TENSOR_FLOAT32}
     *
     * Supported tensor rank: 4, with "NHWC" (i.e., Num_samples, Height, Width,
     * and Channels) data layout.
     *
     * Inputs (resizing by shape):
     * * 0: A 4-D tensor, of shape [batches, height, width, depth], specifying
     *      the input.
     * * 1: An {@link OperandType::INT32} scalar, specifying the output
     *      width of the output tensor.
     * * 2: An {@link OperandType::INT32} scalar, specifying the output
     *      height of the output tensor.
     *
     * Outputs:
     * * 0: The output 4-D tensor, of shape
     *      [batches, new_height, new_width, depth].
     */
    RESIZE_BILINEAR = 23,

    /**
     * A basic recurrent neural network layer.
     *
     * This layer implements the operation:
     * outputs = state = activation(inputs * input_weights +
     *                              state * recurrent_weights + bias)
     *
     * Where:
     * * “input_weights” is a weight matrix that multiplies the inputs;
     * * “recurrent_weights” is a weight matrix that multiplies the current
     *    “state” which itself is the output from the previous time step
     *    computation;
     * * “bias” is a bias vector (added to each output vector in the batch);
     * * “activation” is the function passed as the “fused_activation_function”
     *   argument (if not “NONE”).
     *
     * Supported tensor {@link OperandType}:
     * * {@link OperandType::TENSOR_FLOAT32}
     *
     * The input tensors must all be the same type.
     *
     * Inputs:
     * * 0: input.
     *      A 2-D tensor of shape [batch_size, input_size], where “batch_size”
     *      corresponds to the batching dimension, and “input_size” is the size
     *      of the input.
     * * 1: weights.
     *      A 2-D tensor of shape [num_units, input_size], where “num_units”
     *      corresponds to the number of units.
     * * 2: recurrent_weights.
     *      A 2-D tensor of shape [num_units, num_units], with columns
     *      corresponding to the weights from each unit.
     * * 3: bias.
     *      A 1-D tensor of shape [num_units].
     * * 4: hidden state (in).
     *      A 2-D tensor of shape [batch_size, num_units].
     * * 5: fused_activation_function.
     *      An optional {@link FusedActivationFunc} value indicating the
     *      activation function. If “NONE” is specified then it results in a
     *      linear activation.
     *
     * Outputs:
     * * 0: hidden state (out).
     *      A 2-D tensor of shape [batch_size, num_units].
     *
     * * 1: output.
     *      A 2-D tensor of shape [batch_size, num_units]. This is effectively
     *      the same as the current state value.
     */
    RNN = 24,

    /**
     * Computes the softmax activation on the input tensor element-wise, per
     * batch, by normalizing the input vector so the maximum coefficient is
     * zero.
     *
     * The output is calculated using this formula:
     *
     *     output[batch, i] =
     *         exp((input[batch, i] - max(input[batch, :])) * beta) /
     *         sum_{k}{exp((input[batch, k] - max(input[batch, :])) * beta)}
     *
     * For input tensor with rank other than 2, the activation will be applied
     * independently on each 1-D slice along specified dimension.
     *
     * Supported tensor {@link OperandType}:
     * * {@link OperandType::TENSOR_FLOAT32}
     * * {@link OperandType::TENSOR_QUANT8_ASYMM}
     *
     * Supported tensor rank: 2 or 4.
     *
     * Inputs:
     * * 0: A 2-D or 4-D tensor, specifying the tensor to be reshaped.
     * * 1: A scalar, specifying the positive scaling factor for the exponent,
     *      beta. If input0 is of {@link OperandType::TENSOR_FLOAT32} or
     *      {@link OperandType::TENSOR_QUANT8_ASYMM}, the scalar must be of
     *      {@link OperandType::FLOAT32}.
     *
     * Outputs:
     * * 0: The output tensor of same shape as input0.
     *      For {@link OperandType::TENSOR_QUANT8_ASYMM},
     *      the scale must be 1.f / 256 and the zeroPoint must be 0.
     */
    SOFTMAX = 25,

    /**
     * Rearranges blocks of spatial data, into depth.
     *
     * More specifically, this op outputs a copy of the input tensor where
     * values from the height and width dimensions are moved to the depth
     * dimension. The value block_size indicates the input block size and how
     * the data is moved.
     *
     * Chunks of data of size block_size * block_size from depth are rearranged
     * into non-overlapping blocks of size block_size x block_size.
     *
     * The depth of the output tensor is input_depth * block_size * block_size.
     * The input tensor's height and width must be divisible by block_size.
     *
     * Supported tensor {@link OperandType}:
     * * {@link OperandType::TENSOR_FLOAT32}
     * * {@link OperandType::TENSOR_QUANT8_ASYMM}
     *
     * Supported tensor rank: 4, with "NHWC" (i.e., Num_samples, Height, Width,
     * and Channels) data layout.
     *
     * Inputs:
     * * 0: A 4-D tensor, of shape [batches, height, width, depth_in],
     *      specifying the input.
     * * 1: An {@link OperandType::INT32} scalar, specifying the block_size.
     *      block_size must be >=1 and block_size must be a divisor of both the
     *      input height and width.
     *
     * Outputs:
     * * 0: The output 4-D tensor, of shape [batches, height/block_size,
     *      width/block_size, depth_in*block_size*block_size].
     *      For a {@link OperandType::TENSOR_QUANT8_ASYMM} tensor,
     *      the scale and zeroPoint must be the same as input0.
     */
    SPACE_TO_DEPTH = 26,

    /**
     * SVDF op is a kind of stateful layer derived from the notion that a
     * densely connected layer that's processing a sequence of input frames can
     * be approximated by using a singular value decomposition of each of its
     * nodes. The implementation is based on:
     *
     * https://research.google.com/pubs/archive/43813.pdf
     *
     * P. Nakkiran, R. Alvarez, R. Prabhavalkar, C. Parada.
     * “Compressing Deep Neural Networks using a Rank-Constrained Topology”.
     * INTERSPEECH, 2015.
     *
     * It processes the incoming input using a 2-stage filtering mechanism:
     * * stage 1 performs filtering on the "features" dimension, whose outputs
     *   get pushed into a memory of fixed-size memory_size.
     * * stage 2 performs filtering on the "time" dimension of the memory_size
     *   memoized outputs of stage 1.
     *
     * Specifically, for rank 1, this layer implements the operation:
     *
     *     memory = push(conv1d(inputs, weights_feature, feature_dim,
     *                          "PADDING_VALID"));
     *     outputs = activation(memory * weights_time + bias);
     *
     * Where:
     * * “weights_feature” is a weights matrix that processes the inputs (by
     *   convolving the input with every “feature filter”), and whose outputs
     *   get pushed, stacked in order, into the fixed-size “memory” (the oldest
     *   entry gets dropped);
     * * “weights_time” is a weights matrix that processes the “memory” (by a
     *   batched matrix multiplication on the num_units);
     * * “bias” is an optional bias vector (added to each output vector in the
     *   batch); and
     * * “activation” is the function passed as the “fused_activation_function”
     *   argument (if not “NONE”).
     *
     * Each rank adds a dimension to the weights matrices by means of stacking
     * the filters.
     *
     * Supported tensor {@link OperandType}:
     * * {@link OperandType::TENSOR_FLOAT32}
     *
     * All input tensors must be the same type.
     *
     * Inputs:
     * * 0: input.
     *      A 2-D tensor of shape [batch_size, input_size], where “batch_size”
     *      corresponds to the batching dimension, and “input_size” is the size
     *      of the input.
     * * 1: weights_feature.
     *      A 2-D tensor of shape [num_units, input_size], where “num_units”
     *      corresponds to the number of units.
     * * 2: weights_time.
     *      A 2-D tensor of shape [num_units, memory_size], where “memory_size”
     *      corresponds to the fixed-size of the memory.
     * * 3: bias.
     *      An optional 1-D tensor of shape [num_units].
     * * 4: state (in).
     *      A 2-D tensor of shape [batch_size, (memory_size - 1) * num_units * rank].
     * * 5: rank.
     *      The rank of the SVD approximation.
     * * 6: fused_activation_function.
     *      An optional {@link FusedActivationFunc} value indicating the
     *      activation function. If “NONE” is specified then it results in a
     *      linear activation.
     *
     * Outputs:
     * * 0: state (out).
     *      A 2-D tensor of the same {@link OperandType} as the inputs, with shape
     *      [batch_size, (memory_size - 1) * num_units * rank].
     * * 1: output.
     *      A 2-D tensor of the same {@link OperandType} as the inputs, with shape
     *      [batch_size, num_units].
     */
    SVDF = 27,

    /**
     * Computes hyperbolic tangent of input tensor element-wise.
     *
     * The output is calculated using this formula:
     *
     *     output = tanh(input)
     *
     * Supported tensor {@link OperandType}:
     * * {@link OperandType::TENSOR_FLOAT32}
     *
     * Supported tensor rank: up to 4.
     *
     * Inputs:
     * * 0: A tensor, specifying the input.
     *
     * Outputs:
     * * 0: The output tensor of same shape as input0.
     */
    TANH = 28,

    /**
     * DEPRECATED. Since NNAPI 1.2, extensions are the preferred alternative to
     * OEM operation and data types.
     *
     * This operation is OEM specific. It should only be used for OEM
     * applications.
     */
    OEM_OPERATION = 10000,
};

/**
 * Fused activation function types.
 */
enum FusedActivationFunc : int32_t {
    NONE  = 0,
    RELU  = 1,
    RELU1 = 2,
    RELU6 = 3,
};

/**
 * How an operand is used.
 */
enum OperandLifeTime : int32_t {
    /**
     * The operand is internal to the model. It's created by an operation and
     * consumed by other operations. It must be an output operand of
     * exactly one operation.
     */
    TEMPORARY_VARIABLE,

    /**
     * The operand is an input of the model. It must not be an output
     * operand of any operation.
     *
     * An operand can't be both input and output of a model.
     */
    MODEL_INPUT,

    /**
     * The operand is an output of the model. It must be an output
     * operand of exactly one operation.
     *
     * An operand can't be both input and output of a model.
     */
    MODEL_OUTPUT,

    /**
     * The operand is a constant found in Model.operandValues. It must
     * not be an output operand of any operation.
     */
    CONSTANT_COPY,

    /**
     * The operand is a constant that was specified via a Memory
     * object. It must not be an output operand of any operation.
     */
    CONSTANT_REFERENCE,

    /**
     * The operand does not have a value. This is valid only for optional
     * arguments of operations.
     */
    NO_VALUE,
};

/**
 * Status of a device.
 */
enum DeviceStatus : int32_t {
    AVAILABLE,
    BUSY,
    OFFLINE,
    UNKNOWN,
};

/**
 * Performance information for the reference workload.
 *
 * Used by a driver to report its performance characteristics.
 */
struct PerformanceInfo {
    /**
     * Ratio of the time taken by the driver to execute the
     * workload compared to the time the CPU would take for the
     * same workload. A lower number is better.
     */
    float execTime;

    /**
     * Ratio of the energy used by the driver compared to what
     * the CPU would use for doing the same workload. A lower number
     * is better.
     */
    float powerUsage;
};

/**
 * The capabilities of a driver.
 */
struct Capabilities {
    /**
     * Driver performance when operating on float32 data.
     */
    PerformanceInfo float32Performance;

    /**
     * Driver performance when operating on asymmetric 8-bit quantized data.
     */
    PerformanceInfo quantized8Performance;
};

/**
 * Describes the location of a data object.
 */
struct DataLocation {
    /**
     * The index of the memory pool where this location is found.
     */
    uint32_t poolIndex;

    /**
     * Offset in bytes from the start of the pool.
     */
    uint32_t offset;

    /**
     * The length of the data in bytes.
     */
    uint32_t length;
};

/**
 * Describes one operand of the model's graph.
 */
struct Operand {
    /**
     * Data type of the operand.
     */
    OperandType type;

    /**
     * Dimensions of the operand.
     *
     * For a scalar operand, dimensions.size() must be 0.
     *
     * For a tensor operand, dimensions.size() must be at least 1;
     * however, any of the dimensions may be unspecified.
     *
     * A tensor operand with all dimensions specified has "fully
     * specified" dimensions. Whenever possible (i.e., whenever the
     * dimensions are known at model construction time), a tensor
     * operand should have (but is not required to have) fully
     * specified dimensions, in order to enable the best possible
     * performance.
     *
     * If a tensor operand's dimensions are not fully specified, the
     * dimensions of the operand are deduced from the operand
     * dimensions and values of the operation for which that operand
     * is an output.
     *
     * In the following situations, a tensor operand's dimensions must
     * be fully specified:
     *
     *     . The operand has lifetime CONSTANT_COPY or
     *       CONSTANT_REFERENCE.
     *
     *     . The operand has lifetime MODEL_INPUT or MODEL_OUTPUT. Fully
     *       specified dimensions must either be present in the
     *       Operand or they must be provided in the corresponding
     *       RequestArgument.
     *       EXCEPTION: If the input or output is optional and omitted
     *       (by setting the hasNoValue field of the corresponding
     *       RequestArgument to true) then it need not have fully
     *       specified dimensions.
     *
     * A tensor operand with some number of unspecified dimensions is
     * represented by setting each unspecified dimension to 0.
     */
    vec<uint32_t> dimensions;

    /**
     * The number of times this operand appears as an operation input.
     *
     * (For example, if this operand appears once in one operation's
     * input list, and three times in another operation's input list,
     * then numberOfConsumers = 4.)
     */
    uint32_t numberOfConsumers;

    /**
     * Quantized scale of the operand.
     *
     * Only applicable if the operand is of type TENSOR_QUANT8_ASYMM or
     * TENSOR_INT32.
     */
    float scale;

    /**
     * Quantized zero-point offset of the operand.
     *
     * Only applicable if the operand is of type TENSOR_QUANT8_ASYMM.
     */
    int32_t zeroPoint;

    /**
     * How the operand is used.
     */
    OperandLifeTime lifetime;

    /**
     * Where to find the data for this operand.
     * If the lifetime is TEMPORARY_VARIABLE, MODEL_INPUT, MODEL_OUTPUT, or
     * NO_VALUE:
     * - All the fields must be 0.
     * If the lifetime is CONSTANT_COPY:
     * - location.poolIndex is 0.
     * - location.offset is the offset in bytes into Model.operandValues.
     * - location.length is set.
     * If the lifetime is CONSTANT_REFERENCE:
     * - location.poolIndex is set.
     * - location.offset is the offset in bytes into the specified pool.
     * - location.length is set.
     */
    DataLocation location;
};

/**
 * Describes one operation of the model's graph.
 */
struct Operation {
    /**
     * The operation type.
     */
    OperationType type;

    /**
     * Describes the table that contains the indexes of the inputs of the
     * operation. The offset is the index in the operandIndexes table.
     */
    vec<uint32_t> inputs;

    /**
     * Describes the table that contains the indexes of the outputs of the
     * operation. The offset is the index in the operandIndexes table.
     */
    vec<uint32_t> outputs;
};

/**
 * A Neural Network Model.
 *
 * This includes not only the execution graph, but also constant data such as
 * weights or scalars added at construction time. The only information that
 * might not be known is the shape of the input tensors.
 */
struct Model {
    /**
     * All operands included in the model.
     */
    vec<Operand> operands;

    /**
     * All operations included in the model.
     *
     * The operations are sorted into execution order. Every operand
     * with lifetime MODEL_OUTPUT or TEMPORARY_VARIABLE must be
     * written before it is read.
     */
    vec<Operation> operations;

    /**
     * Input indexes of the model. There must be at least one.
     *
     * Each value corresponds to the index of the operand in "operands".
     */
    vec<uint32_t> inputIndexes;

    /**
     * Output indexes of the model. There must be at least one.
     *
     * Each value corresponds to the index of the operand in "operands".
     */
    vec<uint32_t> outputIndexes;

    /**
     * A byte buffer containing operand data that were copied into the model.
     *
     * An operand's value must be located here if and only if Operand::lifetime
     * equals OperandLifeTime::CONSTANT_COPY.
     */
    vec<uint8_t> operandValues;

    /**
     * A collection of shared memory pools containing operand values.
     *
     * An operand's value must be located here if and only if Operand::lifetime
     * equals OperandLifeTime::CONSTANT_REFERENCE.
     */
    vec<memory> pools;
};

/**
 * Metadata information specifying the location of the input or output data and
 * any updates to the input or output operand.
 */
struct RequestArgument {
    /**
     * If true, the argument does not have a value. This can be used for
     * operations that take optional arguments. If true, the fields of location
     * are set to 0 and the dimensions vector is left empty.
     */
    bool hasNoValue;

    /**
     * The location within one of the memory pools passed in the Request.
     */
    DataLocation location;

    /**
     * Updated dimension information.
     *
     * If dimensions.size() > 0, dimension information was provided
     * along with the argument. This can be the case for models that
     * accept inputs of varying size. This can't change the rank, just
     * the value of the dimensions that were unspecified in the
     * model. If dimensions.size() > 0, then all dimensions must be
     * specified here; and any dimension that was specified in the
     * model must have the same value here.
     *
     * If the dimensions in the model are not fully specified, then
     * they must be fully specified here, unless hasNoValue is set to
     * true. If the dimensions in the model are fully specified, then
     * either dimensions.size() may be 0, or the dimensions in the
     * model must be identical to the dimensions here.
     */
    vec<uint32_t> dimensions;
};

/**
 * Inputs to be sent to and outputs to be retrieved from a prepared model.
 *
 * A Request serves two primary tasks:
 * 1) Provides the input and output data to be used when executing the model.
 * 2) Specifies any updates to the input operand metadata that were left
 *    unspecified at model preparation time.
 *
 * An output must not overlap with any other output, with an input, or
 * with an operand of lifetime CONSTANT_REFERENCE.
 */
struct Request {
    /**
     * Input data and information to be used in the execution of a prepared
     * model.
     *
     * The index of the input corresponds to the index in Model.inputIndexes.
     *   E.g., input[i] corresponds to Model.inputIndexes[i].
     */
    vec<RequestArgument> inputs;

    /**
     * Output data and information to be used in the execution of a prepared
     * model.
     *
     * The index of the output corresponds to the index in Model.outputIndexes.
     *   E.g., output[i] corresponds to Model.outputIndexes[i].
     */
    vec<RequestArgument> outputs;

    /**
     * A collection of shared memory pools containing operand data for both the
     * inputs and the outputs to a model.
     */
    vec<memory> pools;
};

/**
 * Return status of a function.
 */
enum ErrorStatus : int32_t {
    NONE,
    DEVICE_UNAVAILABLE,
    GENERAL_FAILURE,
    OUTPUT_INSUFFICIENT_SIZE,
    INVALID_ARGUMENT,
};