graphengine/third_party/fwkacllib/inc/ops/rnn.h

/**
 * Copyright 2019 Huawei Technologies 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.
 */

/*!
 * \file rnn.h
 * \brief
 */
#ifndef OPS_BUILT_IN_OP_PROTO_INC_RNN_H_
#define OPS_BUILT_IN_OP_PROTO_INC_RNN_H_

#include "graph/operator_reg.h"

namespace ge {
/**
*@brief: Basic LSTM Cell forward calculation.
*@par Inputs:
*five inputs:
*@li x:A 4D Tensor. Must be one of the following types: float16.
*@li h:A 4D Tensor. Must be one of the following types: float16.
*@li c:A 4D Tensor. Must be one of the following types: float16, float32.
*@li w:A 4D Tensor. Must be one of the following types: float16.
*@li b:A 1D Tensor. Must be one of the following types: float16. The format must be ND . \n
*@li mask:A 1D Tensor. Must be one of the following types: uint8.

*@par Attributes:
*@li keep_prob:An integer identifying the keep prob in the op. Default to 1.
*@li forget_bias:An integer identifying the forget bias in the op. Default to 1.
*@li state_is_tuple:An bool identifying if the hidden state and cell state is tuple. Default to true.
*@li activation:An string identifying the type of activation function in the op. Default to "tanh". Only tanh is currently supported . \n

*@par Outputs:
*seven outputs:
*@li ct:A 4D Tensor. Must be one of the following types: float16, float32.
*@li ht:A 4D Tensor. Must be one of the following types: float16.
*@li it:A 4D Tensor. Must be one of the following types: float16, float32.
*@li jt:A 4D Tensor. Must be one of the following types: float16, float32.
*@li ft:A 4D Tensor. Must be one of the following types: float16, float32.
*@li ot:A 4D Tensor. Must be one of the following types: float16, float32.
*@li tanhct:A 4D Tensor. Must be one of the following types: float16, float32.
*/
REG_OP(BasicLSTMCell)
    .INPUT(x, TensorType({DT_FLOAT16}))
    .INPUT(h, TensorType({DT_FLOAT16}))
    .INPUT(c, TensorType({DT_FLOAT16, DT_FLOAT}))
    .INPUT(w, TensorType({DT_FLOAT16}))
    .INPUT(b, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OPTIONAL_INPUT(mask, TensorType({DT_UINT8}))
    .OUTPUT(ct, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OUTPUT(ht, TensorType({DT_FLOAT16}))
    .OUTPUT(it, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OUTPUT(jt, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OUTPUT(ft, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OUTPUT(ot, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OUTPUT(tanhct, TensorType({DT_FLOAT16, DT_FLOAT}))
    .ATTR(keep_prob, Float, 1.0)
    .ATTR(forget_bias, Float, 1.0)
    .ATTR(state_is_tuple, Bool, true)
    .ATTR(activation, String, "tanh")
    .OP_END_FACTORY_REG(BasicLSTMCell)

/**
*@brief: Dynamic LSTM forward calculation . \n

*@par Inputs:
*@li x:A 4D Tensor. Must be the type float32.
*@li w:A 4D Tensor. Must be the type float32.
*@li b:A 1D Tensor. Must be the type float32. The format must be ND . \n

*@par Outputs:
*output_h:A Tensor of output. Must be the type float32.

*@par Restrictions:
*Warning: THIS FUNCTION IS EXPERIMENTAL. Please do not use.
*/
REG_OP(DynamicLSTM)
    .INPUT(x, TensorType({DT_FLOAT32}))
    .INPUT(w, TensorType({DT_FLOAT32}))
    .INPUT(b, TensorType({DT_FLOAT32}))
    .OUTPUT(output_h, TensorType({DT_FLOAT32}))
    .OP_END_FACTORY_REG(DynamicLSTM)

/**
*@brief: DynamicRNNGrad calculation.
*@par Inputs:
*ten inputs: \n
*@li x:A 4D Tensor. Must be one of the following types: float16, float32.
*@li w:A 4D Tensor. Must be one of the following types: float16, float32.
*@li b:A 1D Tensor. Must be one of the following types: float16, float32.
*@li y:A 1D Tensor. Must be one of the following types: int32.
*@li init_h:A 4D Tensor. Must be one of the following types: float16, float32.
*@li init_c:A 4D Tensor. Must be one of the following types: float16, float32.
*@li h:A 4D Tensor. Must be one of the following types: float16, float32.
*@li c:A 4D Tensor. Must be one of the following types: float16, float32.
*@li dy:A 4D Tensor. Must be one of the following types: float16, float32.
*@li dh:A 4D Tensor. Must be one of the following types: float16, float32.
*@li dc:A 4D Tensor. Must be one of the following types: float16, float32.
*@li i:A 4D Tensor. Must be one of the following types: float16, float32.
*@li j:A 4D Tensor. Must be one of the following types: float16, float32.
*@li f:A 4D Tensor. Must be one of the following types: float16, float32.
*@li o:A 4D Tensor. Must be one of the following types: float16, float32.
*@li tanhct:A 4D Tensor. Must be one of the following types: float16, float32.
*@li seq_length:A 1D Tensor. Must be one of the following types: int32.
*@li mask:A 1D Tensor. Must be one of the following types: int8.
*@li wci:A 4D Tensor. Must be one of the following types: float16, float32.
*@li wcf:A 4D Tensor. Must be one of the following types: float16, float32.
*@li wco:A 4D Tensor. Must be one of the following types: float16, float32.

*@par Attributes:
*@li cell_type:An string identifying the cell type in the op. Default to "LSTM". Only LSTM is currently supported.
*@li direction:An string identifying the direction in the op. Default to "UNIDIRECTIONAL". Only UNIDIRECTIONAL is currently supported.
*@li cell_depth:An integer identifying the cell depth in the op. Default to 1.
*@li use_peephole:An bool identifying if use peephole in the op. Default to false.
*@li keep_prob:An float identifying the keep prob in the op. Default to 1.
*@li cell_clip:An float identifying the cell clip in the op. Default to -1.
*@li num_proj:An integer identifying the num projection in the op. Default to 0.
*@li time_major:An bool identifying the time major in the op. Default to false.
*@li forget_bias:An float identifying the forget bias in the op. Default to 0.

*@par Outputs:
*eight outputs: \n
*@li dw:A 4D Tensor. Must be one of the following types: float16, float32.
*@li db:A 4D Tensor. Must be one of the following types: float16, float32.
*@li dx:A 4D Tensor. Must be one of the following types: float16, float32.
*@li dh_prev:A 4D Tensor. Must be one of the following types: float16, float32.
*@li dc_prev:A 4D Tensor. Must be one of the following types: float16, float32.
*@li dwci:A 4D Tensor. Must be one of the following types: float16, float32.
*@li dwcf:A 4D Tensor. Must be one of the following types: float16, float32.
*@li dwco:A 4D Tensor. Must be one of the following types: float16, float32.
*/
REG_OP(DynamicRNNGrad)
    .INPUT(x, TensorType({DT_FLOAT16, DT_FLOAT}))
    .INPUT(w, TensorType({DT_FLOAT16, DT_FLOAT}))
    .INPUT(b, TensorType({DT_FLOAT16, DT_FLOAT}))
    .INPUT(y, TensorType({DT_FLOAT16, DT_FLOAT}))
    .INPUT(init_h, TensorType({DT_FLOAT16, DT_FLOAT}))
    .INPUT(init_c, TensorType({DT_FLOAT16, DT_FLOAT}))
    .INPUT(h, TensorType({DT_FLOAT16, DT_FLOAT}))
    .INPUT(c, TensorType({DT_FLOAT16, DT_FLOAT}))
    .INPUT(dy, TensorType({DT_FLOAT16, DT_FLOAT}))
    .INPUT(dh, TensorType({DT_FLOAT16, DT_FLOAT}))
    .INPUT(dc, TensorType({DT_FLOAT16, DT_FLOAT}))
    .INPUT(i, TensorType({DT_FLOAT16, DT_FLOAT}))
    .INPUT(j, TensorType({DT_FLOAT16, DT_FLOAT}))
    .INPUT(f, TensorType({DT_FLOAT16, DT_FLOAT}))
    .INPUT(o, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OPTIONAL_INPUT(tanhct, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OPTIONAL_INPUT(seq_length, TensorType({DT_INT32}))
    .OPTIONAL_INPUT(mask, TensorType({DT_UINT8}))
    .OPTIONAL_INPUT(wci, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OPTIONAL_INPUT(wcf, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OPTIONAL_INPUT(wco, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OUTPUT(dw, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OUTPUT(db, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OUTPUT(dx, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OUTPUT(dh_prev, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OUTPUT(dc_prev, TensorType({DT_FLOAT16, DT_FLOAT}))
    .DYNAMIC_OUTPUT(dwci, TensorType({DT_FLOAT16, DT_FLOAT}))
    .DYNAMIC_OUTPUT(dwcf, TensorType({DT_FLOAT16, DT_FLOAT}))
    .DYNAMIC_OUTPUT(dwco, TensorType({DT_FLOAT16, DT_FLOAT}))
    .ATTR(cell_type, String, "LSTM")
    .ATTR(direction, String, "UNIDIRECTIONAL")
    .ATTR(cell_depth, Int, 0)
    .ATTR(use_peephole, Bool, false)
    .ATTR(keep_prob, Float, -1.0)
    .ATTR(cell_clip, Float, -1.0)
    .ATTR(num_proj, Int, 0)
    .ATTR(time_major, Bool, true)
    .ATTR(forget_bias, Float, 0.0)
    .OP_END_FACTORY_REG(DynamicRNNGrad)

/**
*@brief: DynamicRNN calculation.
*@par Inputs:
*ten inputs:
*@li x:A required 4D Tensor. Must be one of the following types: float16, float32.
*@li w:A required 4D Tensor. Must be one of the following types: float16, float32.
*@li b:A required 1D Tensor. Must be one of the following types: float16, float32. The format must be ND.
*@li seq_length:A optional Tensor. Only Support int32 in ND.
*@li init_h:A optional 4D Tensor. Must be one of the following types: float16, float32.
*@li init_c:A optional 4D Tensor. Must be one of the following types: float16, float32.
*@li wci:A 4D optional Tensor. Must be one of the following types: float16, float32.
*@li wcf:A 4D optional Tensor. Must be one of the following types: float16, float32.
*@li wco:A 4D optional Tensor. Must be one of the following types: float16, float32.
*@li mask:A 1D optional Tensor. Must be one of the following types: uint8. The format must be ND . \n

*@par Attributes:
*@li cell_type:An string identifying the cell type in the op. Default to "LSTM". Only LSTM is currently supported.
*@li direction:An string identifying the direction in the op. Default to "UNIDIRECTIONAL". Only UNIDIRECTIONAL is currently supported.
*@li cell_depth:An integer identifying the cell depth in the op. Default to 1.
*@li use_peephole:An bool identifying if use peephole in the op. Default to false.
*@li keep_prob:An float identifying the keep prob in the op. Default to 1.
*@li cell_clip:An float identifying the cell clip in the op. Default to -1.
*@li num_proj:An integer identifying the num projection in the op. Default to 0.
*@li time_major:An bool identifying the time major in the op. Default to true.
*@li activation:An string identifying the type of activation function in the op. Default to "tanh". Only tanh is currently supported.
*@li forget_bias:An float identifying the forget bias in the op. Default to 0.
*@li gate_order:An string identifying the type of gate order in the op. Support "ijfo" and "ifjo". Default to "ijfo".
*@li is_training:An bool identifying is training in the op. Default to true . \n

*@par Outputs:
*eight outputs:
*@li y:A 4D Tensor. Must be one of the following types: float16, float32.
*@li output_h:A 4D Tensor. Must be one of the following types: float16, float32.
*@li output_c:A 4D Tensor. Must be one of the following types: float16, float32.
*@li i:A 4D Tensor. Must be one of the following types: float16, float32.
*@li j:A 4D Tensor. Must be one of the following types: float16, float32.
*@li f:A 4D Tensor. Must be one of the following types: float16, float32.
*@li o:A 4D Tensor. Must be one of the following types: float16, float32.
*@li tanhct:A 4D Tensor. Must be one of the following types: float16, float32.
*@par Third-party framework compatibility:
* Compatible with the TF operator LSTM.
*/
REG_OP(DynamicRNN)
    .INPUT(x, TensorType({DT_FLOAT16, DT_FLOAT}))
    .INPUT(w, TensorType({DT_FLOAT16, DT_FLOAT}))
    .INPUT(b, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OPTIONAL_INPUT(seq_length, TensorType({DT_INT32, DT_FLOAT16}))
    .OPTIONAL_INPUT(init_h, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OPTIONAL_INPUT(init_c, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OPTIONAL_INPUT(wci, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OPTIONAL_INPUT(wcf, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OPTIONAL_INPUT(wco, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OPTIONAL_INPUT(mask, TensorType({DT_UINT8}))
    .OUTPUT(y, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OUTPUT(output_h, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OUTPUT(output_c, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OUTPUT(i, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OUTPUT(j, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OUTPUT(f, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OUTPUT(o, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OUTPUT(tanhc, TensorType({DT_FLOAT16, DT_FLOAT}))
    .ATTR(cell_type, String, "LSTM")
    .ATTR(direction, String, "UNIDIRECTIONAL")
    .ATTR(cell_depth, Int, 1)
    .ATTR(use_peephole, Bool, false)
    .ATTR(keep_prob, Float, 1.0)
    .ATTR(cell_clip, Float, -1.0)
    .ATTR(num_proj, Int, 0)
    .ATTR(time_major, Bool, true)
    .ATTR(activation, String, "tanh")
    .ATTR(forget_bias, Float, 0.0)
    .ATTR(gate_order, String, "ijfo")
    .ATTR(is_training, Bool, true)
    .OP_END_FACTORY_REG(DynamicRNN)

/**
*@brief: DynamicRNNV2 calculation.
*@par Inputs:
*ten inputs:
*@li x:A required 4D Tensor. Must be one of the following types: float16, float32.
*@li weight_input:A required 4D Tensor. Must be one of the following types: float16, float32.
*@li weight_hidden:A required 4D Tensor. Must be one of the following types: float16, float32.
*@li b:A required 1D Tensor. Must be one of the following types: float16, float32. The format must be ND.
*@li seq_length:A optional 1D Tensor. Must be one of the following types: int32. The format must be ND.
*@li init_h:A optional 4D Tensor. Must be one of the following types: float16, float32.
*@li init_c:A optional 4D Tensor. Must be one of the following types: float16, float32.
*@li wci:A 4D optional Tensor. Must be one of the following types: float16, float32.
*@li wcf:A 4D optional Tensor. Must be one of the following types: float16, float32.
*@li wco:A 4D optional Tensor. Must be one of the following types: float16, float32.
*@li mask:A 1D optional Tensor. Must be one of the following types: uint8. The format must be ND . \n

*@par Attributes:
*@li cell_type:An string identifying the cell type in the op. Default to "LSTM". Only LSTM is currently supported.
*@li direction:An string identifying the direction in the op. Default to "UNIDIRECTIONAL".
*Only UNIDIRECTIONAL is currently supported.
*@li cell_depth:An integer identifying the cell depth in the op. Default to 1.
*@li use_peephole:An bool identifying if use peephole in the op. Default to false.
*@li keep_prob:An float identifying the keep prob in the op. Default to 1.
*@li cell_clip:An float identifying the cell clip in the op. Default to -1.
*@li num_proj:An integer identifying the num projection in the op. Default to 0.
*@li time_major:An bool identifying the time major in the op. Default to true.
*@li activation:An string identifying the type of activation function in the op. Default to "tanh".
*Support "tanh" and "clip".
*@li recurrent_activation:An string identifying the type of activation function in the op. Default to "sigmoid".
*Support "sigmoid" and "hard_sigmoid". In general, set "hard_sigmoid" for TF Keras LSTM.
*@li forget_bias:An float identifying the forget bias in the op. Default to 0.
*@li gate_order:An string identifying the type of gate order in the op. Support "ijfo" and "ifco". Default to "ijfo".
*Set "ijfo" for TF operator LSTM, Set "ifco" for TF Keras LSTM.
*@li stateful: An bool identifying the type of stateful in the op. Default to fasle.Only false is currently supported.
*@li merge_mode: An string identifying the type of merge_modein the op. Default to "concat".
*Only "concat" is currently supported
*@li is_training:An bool identifying is training in the op. Default to true . \n

*@par Outputs:
*eight outputs:
*@li y:A 4D Tensor. Must be one of the following types: float16, float32.
*@li output_h:A 4D Tensor. Must be one of the following types: float16, float32.
*Return the last output_h.
*@li output_c:A 4D Tensor. Must be one of the following types: float16, float32.
*Return the last output_c.
*@li i:A 4D Tensor. Must be one of the following types: float16, float32.
*@li j:A 4D Tensor. Must be one of the following types: float16, float32.
*@li f:A 4D Tensor. Must be one of the following types: float16, float32.
*@li o:A 4D Tensor. Must be one of the following types: float16, float32.
*@li tanhct:A 4D Tensor. Must be one of the following types: float16, float32.
*@par Third-party framework compatibility:
* Compatible with the TF operator LSTM or TF keras operator LSTM.
*/

REG_OP(DynamicRNNV2)
    .INPUT(x, TensorType({DT_FLOAT16, DT_FLOAT}))
    .INPUT(weight_input, TensorType({DT_FLOAT16, DT_FLOAT}))
    .INPUT(weight_hidden, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OPTIONAL_INPUT(b, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OPTIONAL_INPUT(seq_length, TensorType({DT_INT32}))
    .OPTIONAL_INPUT(init_h, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OPTIONAL_INPUT(init_c, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OPTIONAL_INPUT(wci, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OPTIONAL_INPUT(wcf, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OPTIONAL_INPUT(wco, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OPTIONAL_INPUT(mask, TensorType({DT_UINT8}))
    .OUTPUT(y, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OUTPUT(output_h, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OUTPUT(output_c, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OUTPUT(i, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OUTPUT(j, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OUTPUT(f, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OUTPUT(o, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OUTPUT(tanhc, TensorType({DT_FLOAT16, DT_FLOAT}))
    .ATTR(cell_type, String, "LSTM")
    .ATTR(direction, String, "UNIDIRECTIONAL")
    .ATTR(cell_depth, Int, 1)
    .ATTR(use_peephole, Bool, false)
    .ATTR(keep_prob, Float, 1.0)
    .ATTR(cell_clip, Float, -1.0)
    .ATTR(num_proj, Int, 0)
    .ATTR(time_major, Bool, true)
    .ATTR(activation, String, "tanh")
    .ATTR(recurrent_activation, String, "sigmoid")
    .ATTR(forget_bias, Float, 0.0)
    .ATTR(gate_order, String, "ijfo")
    .ATTR(stateful, Bool, false)
    .ATTR(merge_mode, String, "concat")
    .ATTR(is_training, Bool, true)
    .OP_END_FACTORY_REG(DynamicRNNV2)

/**
* @brief: DynamicRNNV2Grad calculation.
* @par Inputs:
* twenty-one inputs:
* @li x:A required 4D Tensor. Must be one of the following types: float16, float32.
* @li w_x:A required 4D Tensor. Must be one of the following types: float16, float32.
* @li w_h:A required 4D Tensor. Must be one of the following types: float16, float32.
* @li y:A 4D Tensor. Must be one of the following types: float16, float32.
* @li init_h:A 4D Tensor. Must be one of the following types: float16, float32.
* @li init_c:A 4D Tensor. Must be one of the following types: float16, float32.
* @li h:A 4D Tensor. Must be one of the following types: float16, float32.
* @li c:A 4D Tensor. Must be one of the following types: float16, float32.
* @li dy:A 4D Tensor. Must be one of the following types: float16, float32.
* @li dh:A 4D Tensor. Must be one of the following types: float16, float32.
* @li dc:A 4D Tensor. Must be one of the following types: float16, float32.
* @li i:A 4D Tensor. Must be one of the following types: float16, float32.
* @li j:A 4D Tensor. Must be one of the following types: float16, float32.
* @li f:A 4D Tensor. Must be one of the following types: float16, float32.
* @li o:A 4D Tensor. Must be one of the following types: float16, float32.
* @li tanhct:A 4D Tensor. Must be one of the following types: float16, float32.
* @li seq_length:A 1D Tensor. Must be one of the following types: int32.
* @li wci:A 4D Tensor. Must be one of the following types: float16, float32.
* @li wcf:A 4D Tensor. Must be one of the following types: float16, float32.
* @li wco:A 4D Tensor. Must be one of the following types: float16, float32.
* @li mask:A 1D Tensor. Must be one of the following types: int8. \n

* @par Attributes:
* @li cell_type:An string identifying the cell type in the op. Default to "LSTM". Only LSTM is currently supported.
* @li direction:An string identifying the direction in the op. Default to "UNIDIRECTIONAL".
* Only UNIDIRECTIONAL is currently supported.
* @li cell_depth:An integer identifying the cell depth in the op. Default to 1. Only 1 is currently supported.
* @li use_peephole:An bool identifying if use peephole in the op. Default to false.
* Only false is currently supported.
* @li keep_prob:An float identifying the keep prob in the op. Default to 1. Only 1 is currently supported.
* @li cell_clip:An float identifying the cell clip in the op. Default to -1. Only -1 is currently supported.
* @li num_proj:An integer identifying the num projection in the op. Default to 0. Only 0 is currently supported.
* @li time_major:An bool identifying the time major in the op. Default to true. Only true is currently supported.
* @li activation:An string identifying the type of activation function in the op. Default to "tanh".
* Only "tanh" is currently supported.
* @li recurrent_activation:An string identifying the type of activation function in the op. Default to "sigmoid".
* Only "sigmoid" is currently supported.
* @li gate_order:An string identifying the type of gate order in the op. Support "ijfo" and "ifco". Default to "ijfo".
* Set "ijfo" for TF operator LSTM, Set "ifco" for TF Keras/Pytorch LSTM .
* @li stateful: An bool identifying the type of stateful in the op. Default to fasle.Only false is currently supported.
* @li merge_mode: An string identifying the type of merge_modein the op. Default to "concat".
* Only "concat" is currently supported. \n

* @par Outputs:
* nine outputs:
* @li dw_x:A 4D Tensor. Must be one of the following types: float16, float32.
* @li dw_h:A 4D Tensor. Must be one of the following types: float16, float32.
* @li db:A 4D Tensor. Must be one of the following types: float16, float32.
* @li dx:A 4D Tensor. Must be one of the following types: float16, float32.
* @li dh_prev:A 4D Tensor. Must be one of the following types: float16, float32.
* @li dc_prev:A 4D Tensor. Must be one of the following types: float16, float32.
* @li dwci:A 4D Tensor. Must be one of the following types: float16, float32.
* @li dwcf:A 4D Tensor. Must be one of the following types: float16, float32.
* @li dwco:A 4D Tensor. Must be one of the following types: float16, float32.

* @par Third-party framework compatibility:
* Compatible with the TF operator LSTM or TF keras operator LSTM.

* @par Restrictions:
* Warning: THIS FUNCTION IS EXPERIMENTAL. Please do not use.
*/
REG_OP(DynamicRNNV2Grad)
    .INPUT(x, TensorType({DT_FLOAT16, DT_FLOAT}))
    .INPUT(w_x, TensorType({DT_FLOAT16, DT_FLOAT}))
    .INPUT(w_h, TensorType({DT_FLOAT16, DT_FLOAT}))
    .INPUT(y, TensorType({DT_FLOAT16, DT_FLOAT}))
    .INPUT(init_h, TensorType({DT_FLOAT16, DT_FLOAT}))
    .INPUT(init_c, TensorType({DT_FLOAT16, DT_FLOAT}))
    .INPUT(h, TensorType({DT_FLOAT16, DT_FLOAT}))
    .INPUT(c, TensorType({DT_FLOAT16, DT_FLOAT}))
    .INPUT(dy, TensorType({DT_FLOAT16, DT_FLOAT}))
    .INPUT(dh, TensorType({DT_FLOAT16, DT_FLOAT}))
    .INPUT(dc, TensorType({DT_FLOAT16, DT_FLOAT}))
    .INPUT(i, TensorType({DT_FLOAT16, DT_FLOAT}))
    .INPUT(j, TensorType({DT_FLOAT16, DT_FLOAT}))
    .INPUT(f, TensorType({DT_FLOAT16, DT_FLOAT}))
    .INPUT(o, TensorType({DT_FLOAT16, DT_FLOAT}))
    .INPUT(tanhct, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OPTIONAL_INPUT(seq_length, TensorType({DT_INT32}))
    .OPTIONAL_INPUT(wci, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OPTIONAL_INPUT(wcf, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OPTIONAL_INPUT(wco, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OPTIONAL_INPUT(mask, TensorType({DT_UINT8}))
    .OUTPUT(dw_x, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OUTPUT(dw_h, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OUTPUT(db, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OUTPUT(dx, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OUTPUT(dh_prev, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OUTPUT(dc_prev, TensorType({DT_FLOAT16, DT_FLOAT}))
    .DYNAMIC_OUTPUT(dwci, TensorType({DT_FLOAT16, DT_FLOAT}))
    .DYNAMIC_OUTPUT(dwcf, TensorType({DT_FLOAT16, DT_FLOAT}))
    .DYNAMIC_OUTPUT(dwco, TensorType({DT_FLOAT16, DT_FLOAT}))
    .ATTR(cell_type, String, "LSTM")
    .ATTR(direction, String, "UNIDIRECTIONAL")
    .ATTR(cell_depth, Int, 1)
    .ATTR(use_peephole, Bool, false)
    .ATTR(keep_prob, Float, 1.0)
    .ATTR(cell_clip, Float, -1.0)
    .ATTR(num_proj, Int, 0)
    .ATTR(time_major, Bool, true)
    .ATTR(activation, String, "tanh")
    .ATTR(recurrent_activation, String, "sigmoid")
    .ATTR(gate_order, String, "ijfo")
    .ATTR(stateful, Bool, false)
    .ATTR(merge_mode, String, "concat")
    .OP_END_FACTORY_REG(DynamicRNNV2Grad)

/**
*@brief: DynamicRNNV3 calculation.
*@par Inputs:
*ten inputs:
*@li x:A required 4D Tensor. Must be one of the following types: float16, float32.
*@li w:A required 4D Tensor. Must be one of the following types: float16, float32.
*@li b:A required 1D Tensor. Must be one of the following types: float16, float32. The format must be ND.
*@li seq_length:A optional 1D Tensor. Must be one of the following types: int32. The format must be ND.
*@li init_h:A optional 4D Tensor. Must be one of the following types: float16, float32.
*@li init_c:A optional 4D Tensor. Must be one of the following types: float16, float32.
*@li wci:A 4D optional Tensor. Must be one of the following types: float16, float32.
*@li wcf:A 4D optional Tensor. Must be one of the following types: float16, float32.
*@li wco:A 4D optional Tensor. Must be one of the following types: float16, float32.
*@li mask:A 1D optional Tensor. Must be one of the following types: uint8. The format must be ND . \n
*@li real_mask:A 4D optional Tensor. Must be one of the following types: float16, float32.
*@li project:A 4D optional Tensor. Must be one of the following types: float16, float32.

*@par Attributes:
*@li cell_type:An string identifying the cell type in the op. Default to "LSTM". Only LSTM is currently supported.
*@li direction:An string identifying the direction in the op. Default to "UNIDIRECTIONAL". Only UNIDIRECTIONAL is currently supported.
*@li cell_depth:An integer identifying the cell depth in the op. Default to 1.
*@li use_peephole:An bool identifying if use peephole in the op. Default to false.
*@li keep_prob:An float identifying the keep prob in the op. Default to 1.
*@li cell_clip:An float identifying the cell clip in the op. Default to -1.
*@li num_proj:An integer identifying the num projection in the op. Default to 0.
*@li time_major:An bool identifying the time major in the op. Default to true.
*@li activation:An string identifying the type of activation function in the op. Default to "tanh". Only tanh is currently supported.
*@li forget_bias:An float identifying the forget bias in the op. Default to 0.
*@li is_training:An bool identifying is training in the op. Default to true . \n

*@par Outputs:
*eight outputs:
*@li y:A 4D Tensor. Must be one of the following types: float16, float32.
*@li output_h:A 4D Tensor. Must be one of the following types: float16, float32.
*@li output_c:A 4D Tensor. Must be one of the following types: float16, float32.
*@li i:A 4D Tensor. Must be one of the following types: float16, float32.
*@li j:A 4D Tensor. Must be one of the following types: float16, float32.
*@li f:A 4D Tensor. Must be one of the following types: float16, float32.
*@li o:A 4D Tensor. Must be one of the following types: float16, float32.
*@li tanhct:A 4D Tensor. Must be one of the following types: float16, float32.
*@par Third-party framework compatibility:
* Compatible with the TF operator LSTM.
*/
REG_OP(DynamicRNNV3)
    .INPUT(x, TensorType({DT_FLOAT16, DT_FLOAT}))
    .INPUT(w, TensorType({DT_FLOAT16, DT_FLOAT}))
    .INPUT(b, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OPTIONAL_INPUT(seq_length, TensorType({DT_INT32}))
    .OPTIONAL_INPUT(init_h, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OPTIONAL_INPUT(init_c, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OPTIONAL_INPUT(wci, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OPTIONAL_INPUT(wcf, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OPTIONAL_INPUT(wco, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OPTIONAL_INPUT(mask, TensorType({DT_UINT8}))
    .OPTIONAL_INPUT(real_mask, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OPTIONAL_INPUT(project, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OUTPUT(y, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OUTPUT(output_h, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OUTPUT(output_c, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OUTPUT(i, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OUTPUT(j, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OUTPUT(f, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OUTPUT(o, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OUTPUT(tanhc, TensorType({DT_FLOAT16, DT_FLOAT}))
    .ATTR(cell_type, String, "LSTM")
    .ATTR(direction, String, "UNIDIRECTIONAL")
    .ATTR(cell_depth, Int, 1)
    .ATTR(use_peephole, Bool, false)
    .ATTR(keep_prob, Float, 1.0)
    .ATTR(cell_clip, Float, -1.0)
    .ATTR(num_proj, Int, 0)
    .ATTR(time_major, Bool, true)
    .ATTR(activation, String, "tanh")
    .ATTR(forget_bias, Float, 0.0)
    .ATTR(is_training, Bool, true)
    .OP_END_FACTORY_REG(DynamicRNNV3)

/**
*@brief: DynamicLSTMV2 calculation.
*@par Inputs:
*ten inputs:
*@li x:A required 4D Tensor. Must be one of the following types: float16, float32.
*@li w:A required 4D Tensor. Must be one of the following types: float16, float32.
*@li b:A required 1D Tensor. Must be one of the following types: float16, float32. The format must be ND.
*@li cont:A required 2D Tensor. Must be one of the following types: float16, float32. The format must be ND.
*@li w_xc_x_static:A optional 2D Tensor. Must be one of the following types: float16, float32. The format must be ND.
*@li h0:A optional 4D Tensor. Must be one of the following types: float16, float32.
*@li c0:A optional 4D Tensor. Must be one of the following types: float16, float32.
*@li wci:A optional 4D Tensor. Must be one of the following types: float16, float32.
*@li wcf:A optional 4D Tensor. Must be one of the following types: float16, float32.
*@li wco:A optional 4D Tensor. Must be one of the following types: float16, float32.
*@li mask:A optional 1D Tensor. Must be one of the following types: uint8. The format must be ND .

*@par Attributes:
*@li num_output:An integer identifying the num projection in the op. Default to 0.
*@li expose_hidden:An bool identifying the expose_hidden in the op. Default to flase.
*@li need_output_last:An bool identifying the time major in the op. Default to true.
*@li forget_bias:An float identifying the forget bias in the op. Default to 0.

*@par Outputs:
*eight outputs:
*@li y:A 4D Tensor. Must be one of the following types: float16, float32.
*@li output_h:A 4D Tensor. Must be one of the following types: float16, float32.
*@li output_c:A 4D Tensor. Must be one of the following types: float16, float32.
*@li last_output_h:A 4D Tensor. Must be one of the following types: float16, float32.
*@li last_output_c:A 4D Tensor. Must be one of the following types: float16, float32.
*@par Third-party framework compatibility:
* Compatible with the Caffe operator LSTM.
*@par Restrictions:
* Warning: THIS FUNCTION IS EXPERIMENTAL. Please do not use.
*/
REG_OP(DynamicLSTMV2)
    .INPUT(x, TensorType({DT_FLOAT16, DT_FLOAT}))
    .INPUT(w, TensorType({DT_FLOAT16, DT_FLOAT}))
    .INPUT(b, TensorType({DT_FLOAT16, DT_FLOAT}))
    .INPUT(cont, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OPTIONAL_INPUT(w_xc_x_static, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OPTIONAL_INPUT(h0, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OPTIONAL_INPUT(c0, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OPTIONAL_INPUT(wci, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OPTIONAL_INPUT(wcf, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OPTIONAL_INPUT(wco, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OPTIONAL_INPUT(mask, TensorType({DT_UINT8}))
    .OUTPUT(y, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OUTPUT(output_h, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OUTPUT(output_c, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OUTPUT(last_output_h, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OUTPUT(last_output_c, TensorType({DT_FLOAT16, DT_FLOAT}))
    .ATTR(num_output, Int, 0)
    .ATTR(expose_hidden, Bool, false)
    .ATTR(need_output_last, Bool, false)
    .ATTR(forget_bias, Float, 0.0)
    .OP_END_FACTORY_REG(DynamicLSTMV2)

/**
*@brief: LSTMInputGrad calculation.
*@par Inputs:
*ten inputs: \n
*@li w:A 4D Tensor. Must be one of the following types: float16, float32.
*@li init_c:A 4D Tensor. Must be one of the following types: float16, float32.
*@li c:A 4D Tensor. Must be one of the following types: float16, float32.
*@li dy:A 4D Tensor. Must be one of the following types: float16, float32.
*@li dh:A 4D Tensor. Must be one of the following types: float16, float32.
*@li dc:A 4D Tensor. Must be one of the following types: float16, float32.
*@li i:A 4D Tensor. Must be one of the following types: float16, float32.
*@li j:A 4D Tensor. Must be one of the following types: float16, float32.
*@li f:A 4D Tensor. Must be one of the following types: float16, float32.
*@li o:A 4D Tensor. Must be one of the following types: float16, float32.
*@li tanhct:A 4D Tensor. Must be one of the following types: float16, float32.


*@par Outputs:
*four outputs: \n
*@li dx:A 4D Tensor. Must be one of the following types: float16, float32.
*@li dh_prev:A 4D Tensor. Must be one of the following types: float16, float32.
*@li dc_prev:A 4D Tensor. Must be one of the following types: float16, float32.
*@li dgate:A 4D Tensor. Must be one of the following types: float16.
*/
REG_OP(LSTMInputGrad)
    .INPUT(w, TensorType({DT_FLOAT16, DT_FLOAT}))
    .INPUT(init_c, TensorType({DT_FLOAT16, DT_FLOAT}))
    .INPUT(c, TensorType({DT_FLOAT16, DT_FLOAT}))
    .INPUT(dy, TensorType({DT_FLOAT16, DT_FLOAT}))
    .INPUT(dh, TensorType({DT_FLOAT16, DT_FLOAT}))
    .INPUT(dc, TensorType({DT_FLOAT16, DT_FLOAT}))
    .INPUT(i, TensorType({DT_FLOAT16, DT_FLOAT}))
    .INPUT(j, TensorType({DT_FLOAT16, DT_FLOAT}))
    .INPUT(f, TensorType({DT_FLOAT16, DT_FLOAT}))
    .INPUT(o, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OPTIONAL_INPUT(tanhct, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OUTPUT(dx, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OUTPUT(dh_prev, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OUTPUT(dc_prev, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OUTPUT(dgate, TensorType({DT_FLOAT16}))
    .OP_END_FACTORY_REG(LSTMInputGrad)



/**
*@brief: Dynamic LSTM Cell grad calculation.Calculate the gradient of gates and cell state.
*@par Inputs:
*twelve inputs:
*@li init_c:A 4D Tensor. Must be one of the following types: float16, float32.
*@li c:A 4D Tensor. Must be one of the following types: float16, float32.
*@li dy:A 4D Tensor. Must be one of the following types: float16, float32.
*@li dh:A 4D Tensor. Must be one of the following types: float16, float32.
*@li dc:A 4D Tensor. Must be one of the following types: float16, float32.
*@li i:A 4D Tensor. Must be one of the following types: float16, float32.
*@li j:A 4D Tensor. Must be one of the following types: float16, float32.
*@li f:A 4D Tensor. Must be one of the following types: float16, float32.
*@li o:A 4D Tensor. Must be one of the following types: float16, float32.
*@li tanhct:A 4D Tensor. Must be one of the following types: float16, float32.
*@li mask:A 4D Tensor. Must be one of the following types: float16, float32.
*@li t_state:A 4D Tensor. Must be one of the following types: float16, float32. . \n

*@par Attributes:
*@li forget_bias:An integer identifying the forget bias in the op. Default to 1.
*@li activation:An string identifying the type of activation function in the op. Default to "tanh". Only tanh is currently supported . \n
*@li direction:An string that marks the calculation sequence of the operator. Default to "Forward".
*@li gate_order:An string mark the order of output 4 gate. Default to "ijfo".

*@par Outputs:
*two outputs:
*@li dgate:A 4D Tensor. Must be one of the following types: float16.
*@li dct_1:A 4D Tensor. Must be one of the following types: float16, float32.

*@par Restrictions:
*Warning: THIS FUNCTION IS EXPERIMENTAL.  Please do not use.
*/
REG_OP(DynamicLSTMGradCell)
  .INPUT(init_c, TensorType({DT_FLOAT16, DT_FLOAT}))
  .INPUT(c, TensorType({DT_FLOAT16, DT_FLOAT}))
  .INPUT(dy, TensorType({DT_FLOAT16, DT_FLOAT}))
  .INPUT(dh, TensorType({DT_FLOAT16, DT_FLOAT}))
  .INPUT(dc, TensorType({DT_FLOAT16, DT_FLOAT}))
  .INPUT(i, TensorType({DT_FLOAT16, DT_FLOAT}))
  .INPUT(j, TensorType({DT_FLOAT16, DT_FLOAT}))
  .INPUT(f, TensorType({DT_FLOAT16, DT_FLOAT}))
  .INPUT(o, TensorType({DT_FLOAT16, DT_FLOAT}))
  .INPUT(tanhct, TensorType({DT_FLOAT16, DT_FLOAT}))
  .INPUT(t_state, TensorType({DT_INT32, DT_INT32}))
  .INPUT(mask, TensorType({DT_FLOAT16, DT_FLOAT}))
  .OUTPUT(dgate, TensorType({DT_FLOAT16, DT_FLOAT}))
  .OUTPUT(dct_1, TensorType({DT_FLOAT16, DT_FLOAT}))
  .ATTR(forget_bias, Float, 1.0)
  .ATTR(activation, String, "tanh")
  .ATTR(direction, String, "UNIDIRECTIONAL")
  .ATTR(gate_order, String, "ijfo")
  .OP_END_FACTORY_REG(DynamicLSTMGradCell)


/**
*@brief: Basic LSTM Cell backward calculation.Calculate the gradient of input and hidden state.
*@par Inputs:
*three inputs:
*@li dgate:A 4D Tensor. Must be one of the following types: float16.
*@li w:A 4D Tensor. Must be one of the following types: float16.
*@li dropout_mask:A 1D Tensor. Must be one of the following types: uint8. The format must be ND . \n

*@par Attributes:
*keep_prob:An integer identifying the keep prob in the op. Default to 1 . \n

*@par Outputs:
*two outputs:
*@li dxt:A 4D Tensor. Must be one of the following types: float16, float32.
*@li dht:A 4D Tensor. Must be one of the following types: float16, float32.

*@par Restrictions:
*Warning: THIS FUNCTION IS EXPERIMENTAL.  Please do not use.
*/
REG_OP(BasicLSTMCellInputGrad)
    .INPUT(dgate, TensorType({DT_FLOAT16}))
    .INPUT(w, TensorType({DT_FLOAT16}))
    .OPTIONAL_INPUT(dropout_mask, TensorType({DT_UINT8}))
    .OUTPUT(dxt, TensorType({DT_FLOAT16, DT_FLOAT32}))
    .OUTPUT(dht, TensorType({DT_FLOAT16, DT_FLOAT32}))
    .ATTR(keep_prob, Float, 1.0)
    .OP_END_FACTORY_REG(BasicLSTMCellInputGrad)

/**
*@brief: Basic LSTM Cell backward calculation.Calculate the gradient of weight and bias.
*@par Inputs:
*three inputs:
*@li x:A 4D Tensor. Must be one of the following types: float16.
*@li h:A 4D Tensor. Must be one of the following types: float16.
*@li dgate:A 4D Tensor. Must be one of the following types: uint8. \n

*@par Outputs:
*two outputs:
*@li dw:A 4D Tensor. Must be one of the following types: float16.
*@li db:A 4D Tensor. Must be one of the following types: float16, float32.

*@par Restrictions:
*Warning: THIS FUNCTION IS EXPERIMENTAL.  Please do not use.
*/
REG_OP(BasicLSTMCellWeightGrad)
    .INPUT(x, TensorType({DT_FLOAT16}))
    .INPUT(h, TensorType({DT_FLOAT16}))
    .INPUT(dgate, TensorType({DT_FLOAT16}))
    .OUTPUT(dw, TensorType({DT_FLOAT16}))
    .OUTPUT(db, TensorType({DT_FLOAT16, DT_FLOAT32}))
    .OP_END_FACTORY_REG(BasicLSTMCellWeightGrad)

/**
*@brief: Basic LSTM Cell backward calculation.Calculate the gradient of gates and cell state.
*@par Inputs:
*eight inputs:
*@li c:A 4D Tensor. Must be one of the following types: float16, float32.
*@li dht:A 4D Tensor. Must be one of the following types: float16, float32.
*@li dct:A 4D Tensor. Must be one of the following types: float16, float32.
*@li it:A 4D Tensor. Must be one of the following types: float16, float32.
*@li jt:A 4D Tensor. Must be one of the following types: float16, float32.
*@li ft:A 4D Tensor. Must be one of the following types: float16, float32.
*@li ot:A 4D Tensor. Must be one of the following types: float16, float32.
*@li tanhct:A 4D Tensor. Must be one of the following types: float16, float32. \n

*@par Attributes:
*@li forget_bias:An integer identifying the forget bias in the op. Default to 1.
*@li activation:An string identifying the type of activation function in the op. Default to "tanh". Only tanh is currently supported . \n

*@par Outputs:
*two outputs:
*@li dgate:A 4D Tensor. Must be one of the following types: float16.
*@li dct_1:A 4D Tensor. Must be one of the following types: float16, float32.

*@par Restrictions:
*Warning: THIS FUNCTION IS EXPERIMENTAL.  Please do not use.
*/
REG_OP(BasicLSTMCellCStateGrad)
    .INPUT(c, TensorType({DT_FLOAT16, DT_FLOAT}))
    .INPUT(dht, TensorType({DT_FLOAT16, DT_FLOAT}))
    .INPUT(dct, TensorType({DT_FLOAT16, DT_FLOAT}))
    .INPUT(it, TensorType({DT_FLOAT16, DT_FLOAT}))
    .INPUT(jt, TensorType({DT_FLOAT16, DT_FLOAT}))
    .INPUT(ft, TensorType({DT_FLOAT16, DT_FLOAT}))
    .INPUT(ot, TensorType({DT_FLOAT16, DT_FLOAT}))
    .INPUT(tanhct, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OUTPUT(dgate, TensorType({DT_FLOAT16}))
    .OUTPUT(dct_1, TensorType({DT_FLOAT16, DT_FLOAT}))
    .ATTR(forget_bias, Float, 1.0)
    .ATTR(activation, String, "tanh")
    .OP_END_FACTORY_REG(BasicLSTMCellCStateGrad)

/**
*@brief: RNN operator.
*@par Inputs:
*eight inputs:
*@li x:A 4D Tensor. Must be one of the following types: float16.
*@li cont:A 1D Tensor. Must be one of the following types: float16. The format must be ND.
*@li x_static:A 4D Tensor. Must be one of the following types: float16.
*@li h_0:A 4D Tensor. Must be one of the following types: float16, float32.
*@li w_xh:A 4D Tensor. Must be one of the following types: float16.
*@li w_sh:A 4D Tensor. Must be one of the following types: float16.
*@li w_hh:A 4D Tensor. Must be one of the following types: float16.
*@li w_ho:A 4D Tensor. Must be one of the following types: float16.
*@li bias_h:A 1D Tensor. Must be one of the following types: float16, float32. The format must be ND.
*@li bias_o:A 1D Tensor. Must be one of the following types: float16, float32. The format must be ND . \n

*@par Attributes:
*@li expose_hidden:An bool identifying if expose the hidden state of last time step. Default to false.
*@li num_output:An integer identifying the number of output features. Default to 0 . \n

*@par Outputs:
*two outputs:
*@li o:A 4D Tensor. Must be one of the following types: float16, float32.
*@li h_t:A 4D Tensor. Must be one of the following types: float16, float32.
*@par Restrictions:
*Warning: THIS FUNCTION IS EXPERIMENTAL. Please do not use.
*/
REG_OP(RNN)
    .INPUT(x, TensorType({DT_FLOAT16}))
    .INPUT(cont, TensorType({DT_FLOAT16}))
    .OPTIONAL_INPUT(x_static, TensorType({DT_FLOAT16}))
    .OPTIONAL_INPUT(h_0, TensorType({DT_FLOAT16, DT_FLOAT}))
    .INPUT(w_xh, TensorType({DT_FLOAT16}))
    .INPUT(bias_h, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OPTIONAL_INPUT(w_sh, TensorType({DT_FLOAT16}))
    .INPUT(w_hh, TensorType({DT_FLOAT16}))
    .INPUT(w_ho, TensorType({DT_FLOAT16}))
    .INPUT(bias_o, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OUTPUT(o, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OUTPUT(h_t, TensorType({DT_FLOAT16, DT_FLOAT}))
    .ATTR(num_output, Int, 0)
    .ATTR(expose_hidden, Bool, false)
    .OP_END_FACTORY_REG(RNN)

/**
*@brief: BasicRNNCell operator.
*@par Inputs:
*eight inputs:
*@li x:A 4D Tensor. Must be one of the following types: float16.
*@li cont:A 1D Tensor. Must be one of the following types: float16. The format must be ND.
*@li w_xh_x_static:A 4D Tensor. Must be one of the following types: float16.
*@li h_0:A 4D Tensor. Must be one of the following types: float16, float32.
*@li w_xh:A 4D Tensor. Must be one of the following types: float16.
*@li w_hh:A 4D Tensor. Must be one of the following types: float16.
*@li w_ho:A 4D Tensor. Must be one of the following types: float16.
*@li bias_h:A 1D Tensor. Must be one of the following types: float16, float32. The format must be ND.
*@li bias_o:A 1D Tensor. Must be one of the following types: float16, float32. The format must be ND . \n

*@par Attributes:
*@li expose_hidden:An bool identifying if expose the hidden state of last time step. Default to false.
*@li num_output:An integer identifying the number of output features. Default to 0 . \n

*@par Outputs:
*two outputs:
*@li o_t:A 4D Tensor. Must be one of the following types: float16, float32.
*@li h_t:A 4D Tensor. Must be one of the following types: float16, float32.

*@par Restrictions:
*Warning: THIS FUNCTION IS EXPERIMENTAL.  Please do not use.
*/
REG_OP(BasicRNNCell)
    .INPUT(x, TensorType({DT_FLOAT16}))
    .OPTIONAL_INPUT(cont, TensorType({DT_FLOAT16}))
    .OPTIONAL_INPUT(w_xh_x_static, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OPTIONAL_INPUT(h_0, TensorType({DT_FLOAT16, DT_FLOAT}))
    .INPUT(w_xh, TensorType({DT_FLOAT16}))
    .INPUT(bias_h, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OPTIONAL_INPUT(w_hh, TensorType({DT_FLOAT16}))
    .INPUT(w_ho, TensorType({DT_FLOAT16}))
    .INPUT(bias_o, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OUTPUT(o_t, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OUTPUT(h_t, TensorType({DT_FLOAT16, DT_FLOAT}))
    .ATTR(expose_hidden, Bool, false)
    .ATTR(num_output, Int, 0)
    .OP_END_FACTORY_REG(BasicRNNCell)

/**
*@brief DynamicGRU calculation.
*@par Inputs:
*seven inputs: 
*@li x:Must be one of the following types: float16.
*@li w:Must be one of the following types: float16.
*@li b:Must be one of the following types: float16, float32. The format must be ND.
*@li cw:Must be one of the following types: float16.
*@li cb:Must be one of the following types: float16, float32. The format must be ND.
*@li seq_length:Must be one of the following types: int32. The format must be ND.
*@li init_h:Must be one of the following types: float16, float32.

*@par Attributes:
*@li direction:An string identifying the direction in the op. Default to "UNIDIRECTIONAL". Only UNIDIRECTIONAL is currently supported.
*@li cell_depth:An integer identifying the cell depth in the op. Default to 1.
*@li keep_prob:An float identifying the keep prob in the op. Default to 1.
*@li cell_clip:An float identifying the cell clip in the op. Default to -1.
*@li num_proj:An integer identifying the num projection in the op. Default to 0.
*@li time_major:An bool identifying the time major in the op. Default to true.
*@li activation:An string identifying the type of activation function in the op. Default to "tanh". Only tanh is currently supported.
*@li is_training:An bool identifying is training in the op. Default to true.

*@par Outputs:
*five outputs: 
*@li y:Must be one of the following types: float16, float32.
*@li output_h:Must be one of the following types: float16, float32.
*@li r:Must be one of the following types: float16, float32.
*@li i:Must be one of the following types: float16, float32.
*@li n:Must be one of the following types: float16, float32.

*@par Restrictions:
*Warning: THIS FUNCTION IS EXPERIMENTAL. Please do not use.
*/
REG_OP(DynamicGRU)
    .INPUT(x, TensorType({DT_FLOAT16}))
    .INPUT(w, TensorType({DT_FLOAT16}))
    .INPUT(b, TensorType({DT_FLOAT16, DT_FLOAT}))
    .INPUT(cw, TensorType({DT_FLOAT16}))
    .INPUT(cb, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OPTIONAL_INPUT(seq_length, TensorType({DT_INT32}))
    .OPTIONAL_INPUT(init_h, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OUTPUT(y, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OUTPUT(output_h, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OUTPUT(r, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OUTPUT(i, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OUTPUT(n, TensorType({DT_FLOAT16, DT_FLOAT}))
    .ATTR(direction, String, "UNIDIRECTIONAL")
    .ATTR(cell_depth, Int, 1)
    .ATTR(keep_prob, Float, 1.0)
    .ATTR(cell_clip, Float, -1.0)
    .ATTR(num_proj, Int, 0)
    .ATTR(time_major, Bool, true)
    .ATTR(activation, String, "tanh")
    .ATTR(is_training, Bool, true)
    .OP_END_FACTORY_REG(DynamicGRU)

/**
*@brief DynamicGRUV2 calculation.
*@par Inputs:
*seven inputs: 
*@li x:Must be one of the following types: float16.
*@li weight_input:Must be one of the following types: float16.
*@li weight_hidden:Must be one of the following types: float16.
*@li bias_input:Must be one of the following types: float16, float32. The format must be ND.
*@li bias_hidden:Must be one of the following types: float16, float32. The format must be ND.
*@li seq_length:Must be one of the following types: int32 in ND.
*@li init_h:Must be one of the following types: float16, float32.

*@par Attributes:
*@li direction:An string identifying the direction in the op. Default to "UNIDIRECTIONAL". Only UNIDIRECTIONAL is currently supported.
*@li cell_depth:An integer identifying the cell depth in the op. Default to 1.
*@li keep_prob:An float identifying the keep prob in the op. Default to 1.
*@li cell_clip:An float identifying the cell clip in the op. Default to -1.
*@li num_proj:An integer identifying the num projection in the op. Default to 0.
*@li time_major:An bool identifying the time major in the op. Default to true.
*@li activation:An string identifying the type of activation function in the op. Default to "tanh". Only tanh is currently supported.
*@li gate_order:An string identifying the gate order in weight and bias. Default to "zrh". "rzh" is another option.
*@li reset_after:An bool identifying whether to apply reset gate after matrix multiplication. Default to true.
*@li is_training:An bool identifying is training in the op. Default to true.

*@par Outputs:
*six outputs: 
*@li y:Must be one of the following types: float16, float32.
*@li output_h:Must be one of the following types: float16, float32.
*@li update:Must be one of the following types: float16, float32.
*@li reset:Must be one of the following types: float16, float32.
*@li new:Must be one of the following types: float16, float32.
*@li hidden_new:Must be one of the following types: float16, float32.
*/
REG_OP(DynamicGRUV2)
    .INPUT(x, TensorType({DT_FLOAT16}))
    .INPUT(weight_input, TensorType({DT_FLOAT16}))
    .INPUT(weight_hidden, TensorType({DT_FLOAT16}))
    .OPTIONAL_INPUT(bias_input, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OPTIONAL_INPUT(bias_hidden, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OPTIONAL_INPUT(seq_length, TensorType({DT_INT32, DT_FLOAT16}))
    .OPTIONAL_INPUT(init_h, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OUTPUT(y, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OUTPUT(output_h, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OUTPUT(update, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OUTPUT(reset, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OUTPUT(new, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OUTPUT(hidden_new, TensorType({DT_FLOAT16, DT_FLOAT}))
    .ATTR(direction, String, "UNIDIRECTIONAL")
    .ATTR(cell_depth, Int, 1)
    .ATTR(keep_prob, Float, 1.0)
    .ATTR(cell_clip, Float, -1.0)
    .ATTR(num_proj, Int, 0)
    .ATTR(time_major, Bool, true)
    .ATTR(activation, String, "tanh")
    .ATTR(gate_order, String, "zrh")
    .ATTR(reset_after, Bool, true)
    .ATTR(is_training, Bool, true)
    .OP_END_FACTORY_REG(DynamicGRUV2)


/**
*@brief DynamicGRUV2Hidden calculation.
*@par Inputs:
*five inputs: 
*@li x_weight_input:Must be one of the following types: float32.
*@li weight_hidden:Must be one of the following types: float16.
*@li bias_hidden:Must be one of the following types: float16, float32. The format must be ND.
*@li seq_length:Must be one of the following types: int32 in ND.
*@li init_h:Must be one of the following types: float16, float32.

*@par Attributes:
*@li direction:An string identifying the direction in the op. Default to "UNIDIRECTIONAL". 
Only UNIDIRECTIONAL is currently supported.
*@li cell_depth:An integer identifying the cell depth in the op. Default to 1.
*@li keep_prob:An float identifying the keep prob in the op. Default to 1.
*@li cell_clip:An float identifying the cell clip in the op. Default to -1.
*@li num_proj:An integer identifying the num projection in the op. Default to 0.
*@li time_major:An bool identifying the time major in the op. Default to true.
*@li activation:An string identifying the type of activation function in the op. Default to "tanh". 
Only tanh is currently supported.
*@li gate_order:An string identifying the gate order in weight and bias. Default to "zrh". "rzh" is another option.
*@li reset_after:An bool identifying whether to apply reset gate after matrix multiplication. Default to true.
*@li is_training:An bool identifying is training in the op. Default to true.

*@par Outputs:
*six outputs: 
*@li y:Must be one of the following types: float16, float32.
*@li output_h:Must be one of the following types: float16, float32.
*@li update:Must be one of the following types: float16, float32.
*@li reset:Must be one of the following types: float16, float32.
*@li new:Must be one of the following types: float16, float32.
*@li hidden_new:Must be one of the following types: float16, float32.

*@par Restrictions:
*Warning: THIS FUNCTION IS EXPERIMENTAL. Please do not use.
*/
REG_OP(DynamicGRUV2Hidden)
    .INPUT(x_weight_input, TensorType({DT_FLOAT32}))
    .INPUT(weight_hidden, TensorType({DT_FLOAT16}))
    .OPTIONAL_INPUT(bias_hidden, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OPTIONAL_INPUT(seq_length, TensorType({DT_INT32, DT_FLOAT16}))
    .OPTIONAL_INPUT(init_h, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OUTPUT(y, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OUTPUT(output_h, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OUTPUT(update, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OUTPUT(reset, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OUTPUT(new, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OUTPUT(hidden_new, TensorType({DT_FLOAT16, DT_FLOAT}))
    .ATTR(direction, String, "UNIDIRECTIONAL")
    .ATTR(cell_depth, Int, 1)
    .ATTR(keep_prob, Float, 1.0)
    .ATTR(cell_clip, Float, -1.0)
    .ATTR(num_proj, Int, 0)
    .ATTR(time_major, Bool, true)
    .ATTR(activation, String, "tanh")
    .ATTR(gate_order, String, "zrh")
    .ATTR(reset_after, Bool, true)
    .ATTR(is_training, Bool, true)
    .OP_END_FACTORY_REG(DynamicGRUV2Hidden)


/**
*@brief: DynamicGRUV2Grad calculation.
*@par Inputs:
*fourteen inputs: \n
*@li x:A 4D Tensor. Must be one of the following types: float16, float32.
*@li weight_input:A 4D Tensor. Must be one of the following types: float16, float32.
*@li weight_hidden:A 4D Tensor. Must be one of the following types: float16, float32.
*@li y:A 4D Tensor. Must be one of the following types: float16, float32.
*@li init_h:A 4D Tensor. Must be one of the following types: float16, float32.
*@li h:A 4D Tensor. Must be one of the following types: float16, float32.
*@li dy:A 4D Tensor. Must be one of the following types: float16, float32.
*@li dh:A 4D Tensor. Must be one of the following types: float16, float32.
*@li update:A 4D Tensor. Must be one of the following types: float16, float32.
*@li reset:A 4D Tensor. Must be one of the following types: float16, float32.
*@li new:A 4D Tensor. Must be one of the following types: float16, float32.
*@li hidden_new:A 4D Tensor. Must be one of the following types: float16, float32.
*@li seq_length:A 4D Tensor. Must be one of the following types: float16, float32.
*@li mask:A 4D Tensor. Must be one of the following types: float16, float32.

*@par Attributes:
*@li direction:An string identifying the direction in the op. Default to "UNIDIRECTIONAL". Only UNIDIRECTIONAL is currently supported.
*@li cell_depth:An integer identifying the cell depth in the op. Default to 1.
*@li keep_prob:An float identifying the keep prob in the op. Default to 1.
*@li cell_clip:An float identifying the cell clip in the op. Default to -1.
*@li num_proj:An integer identifying the num projection in the op. Default to 0.
*@li time_major:An bool identifying the time major in the op. Default to true.
*@li gate_order:An string identifying the gate order in weight and bias. Default to "zrh". "rzh" is another option.
*@li reset_after:An bool identifying whether to apply reset gate after matrix multiplication. Default to true.

*@par Outputs:
*six outputs: \n
*@li dw_input:A 4D Tensor. Must be one of the following types: float16, float32.
*@li dw_hidden:A 4D Tensor. Must be one of the following types: float16, float32.
*@li db_input:A 4D Tensor. Must be one of the following types: float16, float32.
*@li db_hidden:A 4D Tensor. Must be one of the following types: float16, float32.
*@li dx:A 4D Tensor. Must be one of the following types: float16, float32.
*@li dh_prev:A 4D Tensor. Must be one of the following types: float16, float32.

*@par Restrictions:
*Warning: THIS FUNCTION IS EXPERIMENTAL. Please do not use.
*/
REG_OP(DynamicGRUV2Grad)
    .INPUT(x, TensorType({DT_FLOAT16, DT_FLOAT}))
    .INPUT(weight_input, TensorType({DT_FLOAT16, DT_FLOAT}))
    .INPUT(weight_hidden, TensorType({DT_FLOAT16, DT_FLOAT}))
    .INPUT(y, TensorType({DT_FLOAT16, DT_FLOAT}))
    .INPUT(init_h, TensorType({DT_FLOAT16, DT_FLOAT}))
    .INPUT(h, TensorType({DT_FLOAT16, DT_FLOAT}))
    .INPUT(dy, TensorType({DT_FLOAT16, DT_FLOAT}))
    .INPUT(dh, TensorType({DT_FLOAT16, DT_FLOAT}))
    .INPUT(update, TensorType({DT_FLOAT16, DT_FLOAT}))
    .INPUT(reset, TensorType({DT_FLOAT16, DT_FLOAT}))
    .INPUT(new, TensorType({DT_FLOAT16, DT_FLOAT}))
    .INPUT(hidden_new, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OPTIONAL_INPUT(seq_length, TensorType({DT_INT32}))
    .OPTIONAL_INPUT(mask, TensorType({DT_UINT8}))
    .OUTPUT(dw_input, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OUTPUT(dw_hidden, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OUTPUT(db_input, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OUTPUT(db_hidden, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OUTPUT(dx, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OUTPUT(dh_prev, TensorType({DT_FLOAT16, DT_FLOAT}))
    .ATTR(direction, String, "UNIDIRECTIONAL")
    .ATTR(cell_depth, Int, 0)
    .ATTR(keep_prob, Float, -1.0)
    .ATTR(cell_clip, Float, -1.0)
    .ATTR(num_proj, Int, 0)
    .ATTR(time_major, Bool, true)
    .ATTR(gate_order, String, "zrh")
    .ATTR(reset_after, Bool, true)
    .OP_END_FACTORY_REG(DynamicGRUV2Grad)

/**
*@brief: GRUV2HiddenGrad calculation.
*@par Inputs:
*nine inputs: \n
*@li dh_pre_t:A 4D Tensor. Must be one of the following types: float16, float32.
*@li init_h:A 4D Tensor. Must be one of the following types: float16, float32.
*@li h:A 4D Tensor. Must be one of the following types: float16, float32.
*@li dy:A 4D Tensor. Must be one of the following types: float16, float32.
*@li dh:A 4D Tensor. Must be one of the following types: float16, float32.
*@li update:A 4D Tensor. Must be one of the following types: float16, float32.
*@li reset:A 4D Tensor. Must be one of the following types: float16, float32.
*@li new:A 4D Tensor. Must be one of the following types: float16, float32.
*@li hidden_new:A 4D Tensor. Must be one of the following types: float16, float32.

*@par Attributes:
*@li t_state:An Int identifying the current t state. Default to [0, 4].
*@li gate_order:An string identifying the gate order in weight and bias. Default to "zrh". "rzh" is another option.

*@par Outputs:
*three outputs: \n
*@li dh_prev:A 4D Tensor. Must be one of the following types: float16, float32.
*@li dgate_h:A 4D Tensor. Must be one of the following types: float16, float32.
*@li dnt_x:A 4D Tensor. Must be one of the following types: float16, float32.

*@par Restrictions:
*Warning: THIS FUNCTION IS EXPERIMENTAL. Please do not use.
*/
REG_OP(GRUV2HiddenGradCell)
    .INPUT(dh_pre_t, TensorType({DT_FLOAT16, DT_FLOAT}))
    .INPUT(h, TensorType({DT_FLOAT16, DT_FLOAT}))
    .INPUT(dy, TensorType({DT_FLOAT16, DT_FLOAT}))
    .INPUT(dh, TensorType({DT_FLOAT16, DT_FLOAT}))
    .INPUT(update, TensorType({DT_FLOAT16, DT_FLOAT}))
    .INPUT(reset, TensorType({DT_FLOAT16, DT_FLOAT}))
    .INPUT(new, TensorType({DT_FLOAT16, DT_FLOAT}))
    .INPUT(hidden_new, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OUTPUT(dh_prev, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OUTPUT(dgate_h, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OUTPUT(dnt_x, TensorType({DT_FLOAT16, DT_FLOAT}))
    .ATTR(t_state, Int, 0)
    .ATTR(gate_order, String, "zrh")
    .OP_END_FACTORY_REG(GRUV2HiddenGradCell)

/**
*@brief: DynamicGRUCellGrad calculation.
*@par Inputs:
*ten inputs: \n
*@li dh_pre_t:A 4D Tensor. Must be one of the following types: float16, float32.
*@li h:A 4D Tensor. Must be one of the following types: float16, float32.
*@li dy:A 4D Tensor. Must be one of the following types: float16, float32.
*@li dh:A 4D Tensor. Must be one of the following types: float16, float32.
*@li update:A 4D Tensor. Must be one of the following types: float16, float32.
*@li reset:A 4D Tensor. Must be one of the following types: float16, float32.
*@li new:A 4D Tensor. Must be one of the following types: float16, float32.
*@li hidden_new:A 4D Tensor. Must be one of the following types: float16, float32.+
*@li init_h:A 4D Tensor. Must be one of the following types: float16, float32.
*@li t_state:A 1D Tensor. Must be one of the following types: int32. The format must be ND.

*@par Attributes:
*gate_order:An string identifying the gate order in weight and bias. Default to "zrh". "rzh" is another option.

*@par Outputs:
*three outputs: \n
*@li dh_prev:A 4D Tensor. Must be one of the following types: float16, float32.
*@li dgate_h:A 4D Tensor. Must be one of the following types: float16, float32.
*@li dnt_x:A 4D Tensor. Must be one of the following types: float16, float32.

*@par Restrictions:
*Warning: THIS FUNCTION IS EXPERIMENTAL. Please do not use.
*/
REG_OP(DynamicGRUCellGrad)
    .INPUT(dh_pre_t, TensorType({DT_FLOAT16, DT_FLOAT}))
    .INPUT(h, TensorType({DT_FLOAT16, DT_FLOAT}))
    .INPUT(dy, TensorType({DT_FLOAT16, DT_FLOAT}))
    .INPUT(dh, TensorType({DT_FLOAT16, DT_FLOAT}))
    .INPUT(update, TensorType({DT_FLOAT16, DT_FLOAT}))
    .INPUT(reset, TensorType({DT_FLOAT16, DT_FLOAT}))
    .INPUT(new, TensorType({DT_FLOAT16, DT_FLOAT}))
    .INPUT(hidden_new, TensorType({DT_FLOAT16, DT_FLOAT}))
    .INPUT(init_h, TensorType({DT_FLOAT16, DT_FLOAT}))
    .INPUT(t_state, TensorType({DT_INT32, DT_INT32}))
    .OUTPUT(dh_prev, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OUTPUT(dgate_h, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OUTPUT(dnt_x, TensorType({DT_FLOAT16, DT_FLOAT}))
    .ATTR(gate_order, String, "zrh")
    .OP_END_FACTORY_REG(DynamicGRUCellGrad)

/**
* @brief Calculates the reversed outputs of the function "embedding". \n

* @par Inputs:
* Two inputs, including:
* @li grad: A mutable Tensor of word grad. Must be one of the following types:
*     float32.
* @li indices: A mutable word index Tensor of the int32 type.\n

* @par Attributes:
* @li num_weights: An int attr which use to judge how many words in dict. \n

* @li padding_idx: An int attr judge which word to fill zeros. Defaults to "-1". \n

* @li scale_grad_by_freq: An optional bool. Defaults to "False".
*     If "True", "grad_weight" will be scale by word_frequency.
*     If "False", "grad_weight" will not be scale by word_frequency. \n

* @par Outputs:
* y: A mutable output Tensor of new word grad has the same type as "grads". \n

* @par Third-party framework compatibility
* Compatible with the Pytorch operator EmbeddingDenseGrad.
*/
REG_OP(EmbeddingDenseGrad)
    .INPUT(grad, TensorType({ DT_FLOAT32 }))  /* "First operand." */
    .INPUT(indices, TensorType({ DT_INT32 })) /* "Second operand." */
    .OUTPUT(y, TensorType({ DT_FLOAT32 }))    /* "Result, has same element type as two inputs" */
    .REQUIRED_ATTR(num_weights, Int)
    .ATTR(padding_idx, Int, -1)
    .ATTR(scale_grad_by_freq, Bool, false)
    .OP_END_FACTORY_REG(EmbeddingDenseGrad)

/**
*@brief CommonLSTM calculation.
*@par Inputs:
*eight inputs: \n
*@li x:Each time step is a 4D Tensor. Must be one of the following types: float16, float32.
*@li w:Each direction is a 4D Tensor. Must be one of the following types: float16, float32.
*@li r:Each direction is a 4D Tensor. Must be one of the following types: float16, float32.
*@li b:An optional input. Each direction is a 1D Tensor. Must be one of the following types: float16, float32. The format must be ND.
*@li sequence_lens:An optional input. A 1D Tensor.Must be one of the following types: int32. The format must be ND.
*@li initial_h:An optional input. Each direction is a 4D Tensor. Must be one of the following types: float16, float32.
*@li initial_c:An optional input. Each direction is a 4D Tensor. Must be one of the following types: float16, float32.
*@li p:An optional input. Each direction is a 1D Tensor.Must be one of the following types: float16, float32. The format must be ND.

*@par Attributes:
*@li activation_alpha:Optional scaling values used by some activation functions. Empty is currently supported.
*@li activation_beta:Optional scaling values used by some activation functions. Empty is currently supported.
*@li activations:The list of activation functions. Empty is currently supported.
*@li clip:An float identifying the cell clip in the op. Default to -1.
*@li direction:Specify if the RNN is forward, reverse, or bidirectional. Must be one of forward(default), reverse, or bidirectional.
*@li hidden_size:Number of neurons in the hidden layer. Reserved.
*@li input_forget:Couple the input and forget gates if 1. Reserved.

*@par Outputs:
*three outputs: \n
*@li y:First dimension is time step, second dimension is direction, others is a 4D Tensor. Must be one of the following types: float16, float32.
*@li y_h:Each direction is a 4D Tensor. Must be one of the following types: float16, float32.
*@li y_c:Each direction is a 4D Tensor. Must be one of the following types: float16, float32.
*/

REG_OP(CommonLSTM)
    .INPUT(x, TensorType({DT_FLOAT16, DT_FLOAT}))
    .INPUT(w, TensorType({DT_FLOAT16, DT_FLOAT}))
    .INPUT(r, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OPTIONAL_INPUT(b, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OPTIONAL_INPUT(sequence_lens, TensorType({DT_INT32}))
    .OPTIONAL_INPUT(initial_h, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OPTIONAL_INPUT(initial_c, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OPTIONAL_INPUT(p, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OUTPUT(y, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OUTPUT(y_h, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OUTPUT(y_c, TensorType({DT_FLOAT16, DT_FLOAT}))
    .ATTR(activation_alpha, ListFloat, {})
    .ATTR(activation_beta, ListFloat, {})
    .ATTR(activations, ListString, {})
    .ATTR(clip, Float, -1.0)
    .ATTR(direction, String, "forward")
    .REQUIRED_ATTR(hidden_size, Int)
    .ATTR(input_forget, Int, 0)
    .OP_END_FACTORY_REG(CommonLSTM)

/**
 * @brief Calculate the mask. According to hidden_size and num_step, convert seq_length to mask.
 *
 * @par Inputs:
 * @li seq_length: A 1D Tensor. Must be one of the following types: int32. Record the current length of each batch. [batch_size].
 * @li x: A 3D Tensor. Must be one of the following types: fp16/fp32. Record the num_step/batch_size/input_size. [num_step, batch_size, input_size].
 * @li hidden_size: An optional attribute of type int32. pass the hidden_size. \n
 *
 * @par Outputs:
 * seq_mask: A 3D Tensor. Must be one of the following types: fp16/fp32. with the shape of [num_step, batch_size, hidden_size]. And has the same type as "b" \n
 *
 * @par Restrictions:
 * Warning: THIS FUNCTION IS EXPERIMENTAL. Please do not use.
 */
REG_OP(RnnGenMaskV2)
    .INPUT(seq_length, TensorType({DT_INT32}))
    .INPUT(x, TensorType({DT_FLOAT16, DT_FLOAT}))
    .REQUIRED_ATTR(hidden_size, Int)
    .OUTPUT(seq_mask, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OP_END_FACTORY_REG(RnnGenMaskV2)

/**
* @brief Common GRU calculation.

* @par Inputs:
* Eight inputs, including:
* @li x: The input sequences packed (and pontentially padded) into on 3D Tesnor(float16).
* @li w: The weight tensor for the gates is 3D Tensor(float16).
* @li r: The recurrence weight tesnor is 3D Tensor(float16).
* @li b: The bias tensor for the gates. The format must be ND
* @li sequence_lens: Optional tensor specifying lengths of sequences(int32). The format must be ND
* @li init_h: Optional initial value of the hidden(float16,float32).

* @par Attributes:
* @li activation_alpha: Optional scaling values used by some activation functions.  \n
* @li activation_beta: Optional scaling values used by some activation functions.  \n
* @li activations: A list of 2 (or 4 if bidirectional) activation functions for update, reset, and hidden gates.  \n
* @li clip: Cell clip threshold. \n
* @li direction: Specify if the RNN is forward, reverse, or bidirectional. \n
* @li hidden_size: Number of neurons in the hidden layer. \n
* @li linear_before_reset: When computing the output of the hidden gate, apply the linear transformation before multiplying by the output of the reset gate. \n

* @par Outputs:
* @li y: A Tensor that concats all the intermediate output values of the hidden(float16,float32).
* @li y_h: The last output value of the hidden(float16,float32).
*/
REG_OP(CommonGRU)
    .INPUT(x, TensorType({DT_FLOAT16, DT_FLOAT}))
    .INPUT(w, TensorType({DT_FLOAT16, DT_FLOAT}))
    .INPUT(r, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OPTIONAL_INPUT(b, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OPTIONAL_INPUT(sequence_lens, TensorType({DT_INT32}))
    .OPTIONAL_INPUT(initial_h, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OUTPUT(y, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OUTPUT(y_h, TensorType({DT_FLOAT16, DT_FLOAT}))
    .ATTR(activation_alpha, ListFloat, {})
    .ATTR(activation_beta , ListFloat, {})
    .ATTR(activations , ListString, {})
    .ATTR(clip, Float, -1.0)
    .ATTR(direction, String, "forward")
    .REQUIRED_ATTR(hidden_size, Int)
    .ATTR(linear_before_reset , Int, 0)
    .OP_END_FACTORY_REG(CommonGRU)
/**
* @brief Calculates the reversed outputs of the function "embedding". \n

* @par Inputs:
* Four inputs, including:
* @li weight: A mutable Tensor of word grad. Must be one of the following types:
*     float32.
* @li indices: A mutable word index Tensor of the int32 type.\n
* @li offsets: A mutable word index Tensor of the int32 type.\n
* @li per_sample_weights: to indicate all weights should be taken to be 1.
*     If specified, per_sample_weights must have exactly the same shape as input
*     and is treated as having the same offsets, if those are not None.
*     Only supported for mode='sum'.\n

* @par Attributes:
* @li mode: An string attr which use "sum"``, ``"mean"`` or ``"max"``. Specifies the way to reduce the bag. \n

* @li scale_grad_by_freq: An optional bool. Defaults to "False".
*     If "True", "grad_weight" will be scale by word_frequency.
*     If "False", "grad_weight" will not be scale by word_frequency. \n
* @li sparse: if True, gradient w.r.t.attr weight matrix will be a sparse tensor. \n
* @li include_last_offset: if True, attr offsets  has one additional element, where the last element
*     is equivalent to the size of indices. This matches the CSR format. \n

* @par Outputs:
* y: A mutable output Tensor of new word grad has the same type as "grads". \n

* @par Third-party framework compatibility
* Compatible with the Pytorch operator EmbeddingBag.
*/
REG_OP(EmbeddingBag)
    .INPUT(weight, TensorType({ DT_FLOAT32 }))
    .INPUT(indices, TensorType({ DT_INT32 }))
    .OPTIONAL_INPUT(offsets, TensorType({DT_INT32}))
    .OPTIONAL_INPUT(per_sample_weights, TensorType({DT_FLOAT32}))
    .OUTPUT(y, TensorType({ DT_FLOAT32 }))
    .ATTR(mode, String, "mean")
    .ATTR(scale_grad_by_freq, Bool, false)
    .ATTR(sparse, Bool, false)
    .ATTR(include_last_offset, Bool, false)
    .OP_END_FACTORY_REG(EmbeddingBag)
/**
 * @brief:LSTMP calculation
 * @par Inputs:
 * eight inputs:
 * @li x:A required Tensor(seq, batch, dim). Must be one of the following types: float16, float32.
 * @li real_mask:A optional Tensor(seq, batch). Must be one of the following types: float16, float32.
 * @li init_h:A optional Tensor(batch, state). Must be one of the following types: float16, float32.
 * @li init_c:A optional Tensor(batch, hidden). Must be one of the following types: float16, float32.
 * @li wx:A required Tensor(4*hidden, dim). Must be one of the following types: float16, float32.
 * @li wr:A required Tensor(4*hidden, state). Must be one of the following types: float16, float32.
 * @li bias:A optional Tensor(hidden). Must be one of the following types: float16, float32. The format must be ND.
 * @li project: A optional Tensor. Must be one of the following types: float16, float32.
 *
 * @par Outputs:
 *three outputs:
 *@li y:A Tensor. Must be one of the following types: float16, float32.
 *@li output_h:A Tensor. Must be one of the following types: float16, float32.
 *@li output_c:A Tensor. Must be one of the following types: float16, float32.
 *
 *@par Attributes:
 *time_major:An bool identifying the time major in the op. Default to false.
 * @par Restrictions:
 * Warning: THIS FUNCTION IS EXPERIMENTAL. Please do not use.
*/
REG_OP(LSTMP)
    .INPUT(x, TensorType({DT_FLOAT16, DT_FLOAT}))
    .INPUT(wx, TensorType({DT_FLOAT16, DT_FLOAT}))
    .INPUT(bias, TensorType({DT_FLOAT16, DT_FLOAT}))
    .INPUT(wr, TensorType({DT_FLOAT16, DT_FLOAT}))
    .INPUT(project, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OPTIONAL_INPUT(real_mask, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OPTIONAL_INPUT(init_h, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OPTIONAL_INPUT(init_c, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OUTPUT(y, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OUTPUT(output_h, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OUTPUT(output_c, TensorType({DT_FLOAT16, DT_FLOAT}))
    .ATTR(time_major, Bool, false)
    .OP_END_FACTORY_REG(LSTMP)
}  // namespace ge

#endif  // OPS_BUILT_IN_OP_PROTO_INC_RNN_H_