xref: /aosp_15_r20/external/ComputeLibrary/src/core/CL/CLUtils.h (revision c217d954acce2dbc11938adb493fc0abd69584f3)
1 /*
2  * Copyright (c) 2020-2021, 2023 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 
25 #ifndef ARM_COMPUTE_CL_CLUTILS_H
26 #define ARM_COMPUTE_CL_CLUTILS_H
27 
28 #include "arm_compute/core/CL/OpenCL.h"
29 #include "arm_compute/core/experimental/IPostOp.h"
30 
31 namespace arm_compute
32 {
33 class TensorShape;
34 class CLBuildOptions;
35 class ITensorInfo;
36 
37 /** OpenCL Image2D types */
38 enum class CLImage2DType
39 {
40     ReadOnly,
41     WriteOnly
42 };
43 
44 /** Create a cl::Image2D object from an OpenCL buffer
45  *
46  * @note The following conditions are required to create a OpenCL image object from OpenCL buffer,
47  *       -# The platform should support the OpenCL cl_khr_image2d_from_buffer extension
48  *       -# The stride Y for the input1 should satisfy the OpenCL pitch alignment requirement
49  *       -# input width should be less or equal to (CL_DEVICE_IMAGE2D_MAX_WIDTH * 4)
50  *       -# input height should be less or equal to CL_DEVICE_IMAGE2D_MAX_HEIGHT
51  *
52  * It is user responsibility to ensure the above conditions are satisfied since no checks are performed within this function
53  *
54  * @param[in] ctx             cl::Context object
55  * @param[in] buffer          cl::Buffer object from which the OpenCL image2d object is created
56  * @param[in] shape2d         2D tensor shape
57  * @param[in] data_type       DataType to use. Only supported: F32,F16
58  * @param[in] image_row_pitch Image row pitch (a.k.a. stride Y) to be used in the image2d object
59  * @param[in] image_type      Image 2D type (@ref CLImage2DType)
60  *
61  * @return cl::Image2D object
62  */
63 cl::Image2D create_image2d_from_buffer(const cl::Context &ctx, const cl::Buffer &buffer, const TensorShape &shape2d, DataType data_type, size_t image_row_pitch, CLImage2DType image_type);
64 
65 namespace experimental
66 {
67 /** @name (EXPERIMENTAL_POST_OPS)
68  * @{
69  */
70 
71 /** Manage validation, building and configurations of PostOp CL kernels */
72 class PostOpCLKernelUtils final
73 {
74 public:
75     /** CL kernel name postfix for post ops */
76     using NamePostfix = std::string;
77     /** CL kernels that supports post ops assign each post op to a 'slot', in accordance with the postfix
78      * For example, for a kernel with postfix '_act_prelu_eltwiseadd', there are 3 slots
79      * slot 1: (unary) activation, slot 2: pRelu, slot 3: elementwise addition
80      *
81      * Some kernels may allow some slots to be optional, to support multiple combinations of post op sequences.
82      * In such cases, we need to explicitly set up a mapping between each post op and the slots for that kernel.
83      * For example, suppose we have 2 kernels with postfixes: _eltwiseadd_prelu, _act_eltwiseadd_act_prelu, where the activations in the
84      * second kernel are optional. Say we want to support an eltwise addition, followed by a prelu (sequence { eltwiseadd, prelu }).
85      * Now we can choose which one of the 2 kernels to use, since they both support this post op sequence.
86      * We can either:
87      *  1. assign the elementwise to slot 1 and prelu to slot 2 of kernel 1
88      *  { { Eltwise_Add, PRelu } -> {"_eltwise_act", {1, 2} } } or
89      *  2. assign the elementwise to slot 2 and prelu to slot 4 of kernel 1
90      *  { { Eltwise_Add, PRelu } -> {"_act_eltwiseadd_act_prelu", {2, 4} } }
91      */
92     using Slots  = std::vector<unsigned int>;
93     using Config = std::map<PostOpTypeSequence, std::tuple<NamePostfix, Slots>>;
94 
95 public:
96     explicit PostOpCLKernelUtils(const Config &config);
97 
98     /** Check if post op argument tensor shapes are compliant
99      * All post ops must not alter the shape of the original dst tensor (even after broadcasting)
100      *
101      * @param[in] dst      Dst tensor to apply the post ops to
102      * @param[in] post_ops Post ops
103      *
104      * @return true if shapes are compliant and false otherwise
105      */
106     static bool are_post_op_shapes_compliant(const ITensorInfo *dst, const experimental::PostOpList<ITensorInfo *> &post_ops);
107     /** Check if the post op sequence is supported in the current configuration
108      *
109      * @param[in] post_ops Post ops
110      *
111      * @return true if the post op sequence is supported and false otherwise
112      */
113     bool is_post_op_sequence_supported(const PostOpList<ITensorInfo *> &post_ops) const;
114     /** Helper function to set PostOp related build options
115      * @note Convention
116      *      1. Each post op "slot" is prefixed with "P<slot number>", followed by the usual parameters for that post op.
117      *      E.g. If the first slot is an activation, we need to pass 3 definitions in this way:
118      *          -P1_ACTIVATION_TYPE=...  -P1_ACTIVATION_A_VAL=...   -P1_ACTIVATION_B_VAL=...
119      *
120      *      2. For multi-ary post ops, to pass the position of the previous op's dest tensor,
121      *         we append "_X_POS_<pos>" to the post op type.
122      *      E.g. for a single post op add(dst, x), where dst is the result of the main op.
123      *         In this case, the position of the previous op's dest is 0, so we pass
124      *         -P1_ELTWISE_OP=ADD_X_POS_0
125      *
126      * @param[out] built_opts OpenCL kernel build options
127      * @param[in]  post_ops   Post ops
128      *
129      */
130     void set_post_ops_cl_build_options(CLBuildOptions &built_opts, const PostOpList<ITensorInfo *> &post_ops) const;
131     /** Helper function to set PostOp kernel name
132      *
133      * @param[out] kernel_name OpenCL kernel name
134      * @param[in]  post_ops    Post ops
135      *
136      */
137     void set_post_ops_cl_kernel_name(std::string &kernel_name, const PostOpList<ITensorInfo *> &post_ops) const;
138 
139 private:
140     Config _supported_config{};
141 };
142 /** @} */ // end of group (EXPERIMENTAL_POST_OPS)
143 
144 } // namespace experimental
145 
146 } // arm_compute
147 
148 #endif /* ARM_COMPUTE_CL_CLUTILS_H */
149