• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * Copyright (c) 2021 Arm Limited.
3  *
4  * SPDX-License-Identifier: MIT
5  *
6  * Permission is hereby granted, free of charge, to any person obtaining a copy
7  * of this software and associated documentation files (the "Software"), to
8  * deal in the Software without restriction, including without limitation the
9  * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
10  * sell copies of the Software, and to permit persons to whom the Software is
11  * furnished to do so, subject to the following conditions:
12  *
13  * The above copyright notice and this permission notice shall be included in all
14  * copies or substantial portions of the Software.
15  *
16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
19  * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
21  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
22  * SOFTWARE.
23  */
24 #ifndef ARM_COMPUTE_EXPERIMENTAL_IPOSTOP
25 #define ARM_COMPUTE_EXPERIMENTAL_IPOSTOP
26 
27 #include <memory>
28 #include <numeric>
29 #include <vector>
30 
31 namespace arm_compute
32 {
33 namespace experimental
34 {
35 /** Type of Post Op */
36 enum class PostOpType
37 {
38     Activation,
39     Eltwise_Add,
40     Eltwise_PRelu
41 };
42 /** An ordered sequence of type of Post Ops */
43 using PostOpTypeSequence = std::vector<PostOpType>;
44 /** An elementwise n-ary operation that can be appended to and fused with (at kernel-level) other operators
45  *  It contains:
46  *      1. The attributes of the original operator.
47  *      2. Any additional tensor argument.
48  *      3. The position of the previous op's dst tensor in its argument list ( @ref prev_dst_pos )
49  *
50  *  For example, a series of chained ops:
51  *
52  *          div(src1, relu(conv(src0, weights, bias, conv_info), act_info), div_info)
53  *
54  *      translates to
55  *
56  *          dst = conv(src0, weights, bias, conv_info)  // main op
57  *          dst = relu(dst, act_info)                   // previous dst is placed in the first (and only) argument
58  *          dst = div(src1, dst, div_info)              // previous dst is placed in the second argument
59  *
60  *      which in turn translates to:
61  *
62  *          main op: conv(src0, weights, bias, conv_info)
63  *          post op1: relu(act_info, prev_dst_pos = 0)
64  *          post op2: div(div_info, src1, prev_dst_pos = 1)
65  *
66  *  @note: On Broadcasting
67  *      For n-ary post ops, the tensor arguments must not "widen" the dst tensor of the main op
68  *      For example, for a dst of shape [14, 1, 34]:
69  *          * post_op_arg1 = [1, 1, 34] is allowed: broadcast in dim 0
70  *          * post_op_arg1 = [14, 1, 34] is allowed: no broadcast
71  *          * post_op_arg1 = [1, 1, 34] is allowed: broadcast in dims 0 and 1
72  *          * post_op_arg1 = [14, 15, 34] is NOT allowed: broadcast widens the dst tensor
73  *
74  * @note: On Data layout
75  *      All post ops are data layout agnostic. This means post ops do not have an inherent idea of "width", "height" and so on.
76  *      Should we want to perform a post op with 2 tensors of different data layouts (where data layouts are significant to both),
77  *      then we need to perform necessary permutation op beforehand to unify their data layout before they can be fused with a post op
78  *
79  *      Note although post ops themselves should be able to support any data layout, the main op they fuse to may impose
80  *      additional restrictions in the presence of post ops. For example, the implementation of a gemm op may only allow
81  *      NHWC data layout if post ops are provided. Such restrictions are main op implementation specific.
82  *
83  *  @note: PostOps do not own any resources pointed to by TensorRelatedT if it's a pointer type
84  *  @note: If TensorRelatedT points to a resource, IPostOp assumes that resource is valid throughout its lifetime
85  *        and the lifetime of its copies. This is almost guaranteed as IPostOp is only meant to be used at configure time
86  *        after the ITensor or ITensorInfo objects are already constructed
87  */
88 template <typename TensorRelatedT>
89 struct IPostOp
90 {
91     /** Get the arity of the post op
92      * @note: that this is one fewer than the arity of the original op, because we implicitly pass the previous op's dst
93      *       tensor as one of the arguments
94      */
arityIPostOp95     size_t arity() const
96     {
97         return arguments().size();
98     }
99     /** The position of previous op's dst in current op's argument list */
100     virtual int prev_dst_pos() const = 0;
101     /** The IPostOp type */
102     virtual PostOpType type() const = 0;
103     /** The argument tensors
104      * The order of the argument tensor is strictly preserved
105      */
106     virtual std::vector<TensorRelatedT *>       arguments()       = 0;
107     virtual std::vector<const TensorRelatedT *> arguments() const = 0;
108     /** Clone method used in cases where PostOps are owned by unique_ptr
109      * @note: This performs a shallow copy of the TensorRelatedT if TensorRelatedT points to a resource
110      */
111     virtual std::unique_ptr<IPostOp<TensorRelatedT>> clone() const = 0;
~IPostOpIPostOp112     virtual ~IPostOp()
113     {
114     }
115 };
116 
117 /** A sequence of PostOps that can be appended to the end of other operators */
118 template <typename TensorRelatedT>
119 class PostOpList
120 {
121 public:
122     /** Constructor */
123     PostOpList() = default;
124     /** Destructor */
125     ~PostOpList() = default;
PostOpList(const PostOpList & other)126     PostOpList(const PostOpList &other)
127     {
128         for(const auto &op : other._post_ops)
129         {
130             this->_post_ops.push_back(op->clone());
131         }
132     }
133     PostOpList &operator=(const PostOpList &other)
134     {
135         PostOpList tmp{ other };
136         std::swap(tmp, *this);
137         return *this;
138     }
139     PostOpList(PostOpList &&other) = default;
140     PostOpList &operator=(PostOpList &&other) = default;
141 
142     /** Add a new post op at the end of the list */
143     template <typename OpT, typename... Args>
push_back_op(Args &&...args)144     void push_back_op(Args &&... args)
145     {
146         _post_ops.push_back(std::make_unique<OpT>(std::forward<Args>(args)...));
147     }
148 
149     /** Number of post ops */
size()150     size_t size() const
151     {
152         return _post_ops.size();
153     }
154 
155     /** Total number of post ops */
total_num_arguments()156     size_t total_num_arguments() const
157     {
158         return std::accumulate(_post_ops.begin(), _post_ops.end(), 0, [](size_t op1_arity, const auto & op2)
159         {
160             return op1_arity + op2->arity();
161         });
162     }
163 
164     /** Get the underlying post op list */
get_list()165     std::vector<std::unique_ptr<IPostOp<TensorRelatedT>>> &get_list()
166     {
167         return _post_ops;
168     }
get_list()169     const std::vector<std::unique_ptr<IPostOp<TensorRelatedT>>> &get_list() const
170     {
171         return _post_ops;
172     }
173 
174 private:
175     std::vector<std::unique_ptr<IPostOp<TensorRelatedT>>> _post_ops{};
176 };
177 
178 } // namespace experimental
179 } // namespace arm_compute
180 #endif //ARM_COMPUTE_EXPERIMENTAL_IPOSTOP
181