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

/**
 * Copyright 2019-2020 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.
 */

#ifndef GE_OP_BATCH_OPS_H_
#define GE_OP_BATCH_OPS_H_

#include "graph/operator_reg.h"

namespace ge {

/**
*@brief Creates batches of tensors in tensors.

*@par Inputs:
*The input x_tensors can be a list or a dictionary of tensors. Inputs include: \n
*x_tensors:The list or dictionary of tensors to enqueue.

*@par Attributes:
*@li num_batch_threads:The number of threads enqueuing tensors. The batching will be nondeterministic if num_batch_threads > 1.
*@li max_batch_size:Max batch size pulled from the queue.
*@li max_enqueued_batches:Maximum number of batches pulled from the queue.
*@li batch_timeout_micros:Batch processing timeout in microseconds unit.
*@li allowed_batch_sizes:Allowed batch size pulled from the queue.
*@li grad_timeout_micros:Calculate the gradient batch processing timeout in microseconds unit.
*@li container:If non-empty, this queue is placed in the given container. Otherwise, a default container is used.
*@li shared_name:If set, this queue will be shared under the given name across multiple sessions.
*@li batching_queue:queue resource container.

*@par Outputs:
*@li y_index:Tensor, index of a BatchTensor. Must be in row-major order.
*@li y_id:Tensor, id of a BatchTensor. Must be in row-major order.
*@li y_tensors:A list or dictionary of tensors with the same types as tensors.

*@attention Constraints: \n
*-The implementation for Batch on Ascend uses AI CPU, with bad performance. \n

*@par Quantization supported or not
*Not supported
*@par Quantized inference supported or not
*Supported
*@par L2 convergence supported or not
*@par Multiple batches supported or not
*/

REG_OP(Batch)
  .DYNAMIC_INPUT(x_tensors, TensorType({DT_FLOAT16, DT_FLOAT, DT_INT8, \
      DT_INT16, DT_UINT16, DT_UINT8, DT_INT32, DT_INT64, DT_BOOL, DT_DOUBLE}))
  .OUTPUT(y_index, TensorType({ DT_INT64 }))
  .OUTPUT(y_id, TensorType({ DT_INT64 }))
  .DYNAMIC_OUTPUT(y_tensors, TensorType({DT_INT8, DT_UINT8, DT_INT16, \
      DT_UINT16, DT_INT32, DT_INT64, DT_FLOAT, DT_FLOAT16, DT_DOUBLE, DT_BOOL}))
  .REQUIRED_ATTR(num_batch_threads, Int)
  .REQUIRED_ATTR(max_batch_size, Int)
  .ATTR(max_enqueued_batches, Int, 10)
  .REQUIRED_ATTR(batch_timeout_micros, Int)
  .ATTR(allowed_batch_sizes, ListInt, {})
  .REQUIRED_ATTR(grad_timeout_micros, Int)
  .ATTR(container, String, "")
  .ATTR(shared_name, String, "")
  .ATTR(batching_queue, String, "")
  .OP_END_FACTORY_REG(Batch)

/**
*@brief Reverses the operation of Batch for a single output Tensor.

*@par Inputs:
*The input x_tensors can be a list or a dictionary of tensors. Inputs include: \n
* @li x_tensors:The list or dictionary of tensors to enqueue.
* @li index:The matching batch_index obtained from Batch.
* @li id:The id scalar emitted by Batch.

*@par Attributes:
*@li timeout_micros:Calculate the unbatch processing timeout in microseconds unit.
*@li container:If non-empty, this queue is placed in the given container. Otherwise, a default container is used.
*@li shared_name:If set, this queue will be shared under the given name across multiple sessions.

*@par Outputs:
*y_tensor:A list or dictionary of tensors with the same types as tensors.

*@attention Constraints: \n
*-The implementation for Unbatch on Ascend uses AI CPU, with bad performance. \n

*@par Quantization supported or not
*Not supported
*@par Quantized inference supported or not
*Supported
*@par L2 convergence supported or not
*@par Multiple batches supported or not
*/

REG_OP(Unbatch)
  .INPUT(x_tensor, TensorType({DT_INT8, DT_UINT8, DT_INT16, DT_UINT16, \
      DT_INT32, DT_INT64, DT_BOOL, DT_FLOAT, DT_DOUBLE}))
  .INPUT(index, TensorType({DT_INT64}))
  .INPUT(id, TensorType({DT_INT64}))
  .OUTPUT(y_tensor, TensorType({DT_INT8, DT_UINT8, DT_INT16, DT_UINT16, \
      DT_INT32, DT_INT64, DT_BOOL, DT_FLOAT, DT_DOUBLE}))
  .REQUIRED_ATTR(timeout_micros, Int)
  .ATTR(container, String, "")
  .ATTR(shared_name, String, "")
  .OP_END_FACTORY_REG(Unbatch)

/**
*@brief Acts like Batch but using the given batch_index index of batching things as they become available.

*@par Inputs:
*The input x_input can be a list or a dictionary of tensors. Inputs include: \n
* @li x_input:The input to the Unbatch operation.
* @li index:The batch_index given to the Unbatch operation.
* @li id:The id scalar emitted by Batch.
* @li grad:The downstream gradient.

*@par Attributes:
*@li container:If non-empty, this queue is placed in the given container. Otherwise, a default container is used.
*@li shared_name:If set, this queue will be shared under the given name across multiple sessions.

*@par Outputs:
*y_grad:The return value, either an empty tensor or the batched gradient.

*@attention Constraints: \n
*-The implementation for UnbatchGrad on Ascend uses AI CPU, with bad performance. \n

*@par Quantization supported or not
*Not supported
*@par Quantized inference supported or not
*Supported
*@par L2 convergence supported or not
*@par Multiple batches supported or not
*/

REG_OP(UnbatchGrad)
  .INPUT(x_input, TensorType({DT_INT8, DT_UINT8, DT_INT16, DT_UINT16, \
      DT_INT32, DT_INT64, DT_BOOL, DT_FLOAT, DT_DOUBLE}))
  .INPUT(index, TensorType({DT_INT64}))
  .INPUT(grad, TensorType({DT_INT8, DT_UINT8, DT_INT16, DT_UINT16, \
      DT_INT32, DT_INT64, DT_BOOL, DT_FLOAT, DT_DOUBLE}))
  .INPUT(id, TensorType({DT_INT64}))
  .OUTPUT(y_grad, TensorType({DT_INT8, DT_UINT8, DT_INT16, DT_UINT16, \
      DT_INT32, DT_INT64, DT_BOOL, DT_FLOAT, DT_DOUBLE}))
  .ATTR(container, String, "")
  .ATTR(shared_name, String, "")
  .OP_END_FACTORY_REG(UnbatchGrad)
}  // namespace ge

#endif  // GE_OP_BATCH_OPS_H_