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

/**
 * Copyright (c) Huawei Technologies Co., Ltd. 2020-2021. All rights reserved.
 *
 * 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 experiment_ops.h
 * \brief
 */
#ifndef OPS_BUILT_IN_OP_PROTO_INC_EXPERIMENT_OPS_H_
#define OPS_BUILT_IN_OP_PROTO_INC_EXPERIMENT_OPS_H_

#include "graph/operator_reg.h"
namespace ge {
/**
* @brief Updates "var" according to the AdamW algorithm.
*
* @attention Constraints:
*  The input tensors must have the same shape.*
*
* @par Inputs:
* @li var: A mutable Tensor of the type TensorType::NumberType().
*     Should be from a Variable().
* @li m: A mutable Tensor of the same type as "var".
*     Should be from a Variable().
* @li v: A mutable Tensor of the same type as "var".
*     Should be from a Variable().
* @li beta1_power: A scalar of the same type as "var".
* @li beta2_power: A scalar of the same type as "var".
* @li lr: learning_rate. A scalar of the same type as "var".
* @li weight_decay: learning_rate. A scalar of the same type as "var".
* @li beta1: A scalar of the same type as "var".
* @li beta2: A scalar of the same type as "var".
* @li epsilon: A scalar of the same type as "var".
* @li grad: A Tensor of the same type as "var", for the gradient.
* @li max_grad_norm: A mutable Tensor of the same type as "var", an optional input.
*     Should be from a Variable().
*
* @par Attributes:
* @li amsgrad: An optional bool. Defaults to "False".
*     If "True", max_grad_norm input and output must be entered.
* @li maximize: An optional bool. Defaults to "False".
*
* @par Outputs:
* @li var: A mutable tensor. Has the same type as input "var".
* @li m: A mutable tensor. Has the same type as input "m".
* @li v: A mutable tensor. Has the same type as input "v". \n
*/
REG_OP(ApplyAdamW)
    .INPUT(var, TensorType::NumberType())
    .INPUT(m, TensorType::NumberType())
    .INPUT(v, TensorType::NumberType())
    .INPUT(beta1_power, TensorType::NumberType())
    .INPUT(beta2_power, TensorType::NumberType())
    .INPUT(lr, TensorType::NumberType())
    .INPUT(weight_decay, TensorType::NumberType())
    .INPUT(beta1, TensorType::NumberType())
    .INPUT(beta2, TensorType::NumberType())
    .INPUT(epsilon, TensorType::NumberType())
    .INPUT(grad, TensorType::NumberType())
    .OPTIONAL_INPUT(max_grad_norm, TensorType::NumberType())
    .OUTPUT(var, TensorType::NumberType())
    .OUTPUT(m, TensorType::NumberType())
    .OUTPUT(v, TensorType::NumberType())
    .ATTR(amsgrad, Bool, false)
    .ATTR(maximize, Bool, false)
    .OP_END_FACTORY_REG(ApplyAdamW)

/**
* @brief Multiplies matrix "a" by matrix "b", producing "a * b". \n
* @par Inputs:
* Four inputs, including:
* @li x1: A matrix Tensor. Must be one of the following types: float32,
* float16, int32, int8, int4, bf16. 3D. Has format ND.
* @li x2: A matrix Tensor. Must be one of the following types: float32,
* float16, int32, int8, int4, bf16. 3D. Has format ND.
* @li bias: A optional Tensor. Must be one of the following types:
* float32, float16, int32, bf16. 1D. Has format ND.
* @li offset_w: A optional Tensor. Must be one of the following types:
* int8, int4. Has format ND. \n

* @par Attributes:
* Three attributes, including:
* @li perm_x1: A list int. "x1" is permuted to shape [B, M, K] before multiplication.
* @li perm_x2: A list int. "x2" is permuted to shape [B, K, N] before multiplication.
* @li perm_y: A list int. "y" is permuted after multiplication.
* @li offset_x: An optional integer for quantized TransposeBatchMatMul.
* The negative offset added to the input "x1" for int8, int4 type. Ensure offset_x
* within the effective range of input data type. Defaults to "0". \n

* @par Outputs:
* y: The result matrix Tensor. 3D. Must be one of the following
* types: float32, float16, int32, bf16. 3D. Has format ND. \n
*/
REG_OP(TransposeBatchMatMul)
    .INPUT(x1, TensorType({DT_FLOAT, DT_FLOAT16, DT_INT32, DT_INT8, DT_INT4, DT_BF16}))
    .INPUT(x2, TensorType({DT_FLOAT, DT_FLOAT16, DT_INT32, DT_INT8, DT_INT4, DT_BF16}))
    .OPTIONAL_INPUT(bias, TensorType({DT_FLOAT, DT_FLOAT16, DT_INT32, DT_BF16}))
    .OPTIONAL_INPUT(offset_w, TensorType({DT_INT8, DT_INT4}))
    .OUTPUT(y, TensorType({DT_FLOAT, DT_FLOAT16, DT_INT32, DT_BF16}))
    .ATTR(perm_x1, ListInt, {})
    .ATTR(perm_x2, ListInt, {})
    .ATTR(perm_y, ListInt, {})
    .ATTR(offset_x, Int, 0)
    .OP_END_FACTORY_REG(TransposeBatchMatMul)

/**
* @brief Performs non-maximum suppression (NMS) on the rotated boxes according
* to their intersection-over-union (IoU). Rotated NMS interatively removes lower
* scoring rotated boxes which have an IoU greater than iou_threshold with
* another (higher scoring) rotated box.

* @par Inputs:
* Three inputs, including:
* @li boxes: A 2D Tensor of float16 or float32 with shape (N, 5). Rotated boxes to
* perform NMS on. They are expected to be in (x1, y1, x2, y2, angle_degress) format.
* @li scores: A 1D Tensor of float16 or float32 with shape (N). Scores for each one of
* the rotated boxes.
* @li labels: A 1D Tensor of int32 or int64 with shape (N). Labels for each one of
* the rotated boxes.

* @par Attributes:
* iou_threshold: A required float attribute. Discards all overlapping rotated
* boxes with IoU < iou_threshold.

* @par Outputs:
* Two outputs, including:
* @li selected_detections: A 2D Tensor of float16 or float32 with shape (N, 5).
* The selected boxes that kept by Rotated NMS, sorted in decreasing order of scores.
* @li keep_indices: A 1D Tensor of int32 or int64 with shape (N). The indices of
* selected_detections.

* @attention Constraints:
* Currently, the tensor type of input (boxes, scores) only support float.
* The tensor type of keep_indices only support int32.
*/
REG_OP(RotatedNMS)
    .INPUT(boxes, TensorType({DT_FLOAT16, DT_FLOAT}))
    .INPUT(scores, TensorType({DT_FLOAT16, DT_FLOAT}))
    .INPUT(labels, TensorType({DT_INT32, DT_INT64}))
    .OUTPUT(selected_detections, TensorType({DT_FLOAT16, DT_FLOAT}))
    .OUTPUT(keep_indices, TensorType({DT_INT32, DT_INT64}))
    .REQUIRED_ATTR(iou_threshold, Float)
    .OP_END_FACTORY_REG(RotatedNMS)
}  // namespace ge

#endif  // OPS_BUILT_IN_OP_PROTO_INC_EXPERIMENT_OPS_H_