• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * Copyright (c) 2018-2020 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_NEGEMMASSEMBLYDISPATCH_H
25 #define ARM_COMPUTE_NEGEMMASSEMBLYDISPATCH_H
26 
27 #include "arm_compute/runtime/IFunction.h"
28 #include "arm_compute/runtime/IMemoryManager.h"
29 #include "arm_compute/runtime/IWeightsManager.h"
30 #include "arm_compute/runtime/MemoryGroup.h"
31 #include "arm_compute/runtime/Tensor.h"
32 
33 namespace arm_compute
34 {
35 /* Convolution method supported by the assembly gemm interface */
36 enum class AsmConvMethod
37 {
38     Im2Col,
39     Indirect,
40     Conv
41 };
42 
43 struct AsmGemmInfo
44 {
45     AsmConvMethod           method{ AsmConvMethod::Im2Col };
46     PadStrideInfo           ps_info{};
47     ActivationLayerInfo     activation_info{};
48     GEMMLowpOutputStageInfo output_stage{};
49     bool                    negated_offsets{ true };
50     bool                    reinterpret_input_as_3d{ false };
51     bool                    depth_output_gemm3d{ false };
52     int64_t                 padding_top{ 0 };
53     int64_t                 padding_left{ 0 };
54     float                   padding_value{ 0.f };
55 };
56 
57 /** Assembly kernel glue */
58 class NEGEMMAssemblyDispatch : public IFunction
59 {
60 public:
61     /** Constructor */
62     NEGEMMAssemblyDispatch(std::shared_ptr<IMemoryManager> memory_manager = nullptr, IWeightsManager *weights_manager = nullptr);
63     /** Prevent instances of this class from being copy constructed */
64     NEGEMMAssemblyDispatch(const NEGEMMAssemblyDispatch &) = delete;
65     /** Prevent instances of this class from being copied */
66     NEGEMMAssemblyDispatch &operator=(const NEGEMMAssemblyDispatch &) = delete;
67     NEGEMMAssemblyDispatch(NEGEMMAssemblyDispatch &&)                 = default;
68     NEGEMMAssemblyDispatch &operator=(NEGEMMAssemblyDispatch &&) = default;
69     ~NEGEMMAssemblyDispatch()                                    = default;
70 
71     class IFallback
72     {
73     public:
74         virtual void run()                 = 0;
75         virtual void prepare()             = 0;
76         virtual bool is_configured() const = 0;
77         virtual ~IFallback()               = default;
78     };
79 
80 public:
81     /** If supported create a Compute Library function else fallback to the arm_gemm function.
82      *
83      * @param[in]  a    Input tensor (Matrix A)
84      * @param[in]  b    Input tensor (Matrix B)
85      * @param[in]  c    Input tensor (Matrix C) used to pass the bias for quantized calculations
86      * @param[out] d    Output tensor to store the result of matrix multiplication. Data type supported: same as @p input0.
87      * @param[in]  info GEMM meta-data
88      */
89     void configure(const ITensor *a, const ITensor *b, const ITensor *c, ITensor *d, const AsmGemmInfo &info);
90 
91     /** Indicates whether or not this function can be used to process the given parameters.
92      *
93      * @param[in] a    Input tensor info (Matrix A)
94      * @param[in] b    Input tensor info (Matrix B)
95      * @param[in] c    Input tensor info (Matrix C) used to pass the bias for quantized calculations
96      * @param[in] d    Output tensor to store the result of matrix multiplication. Data type supported: same as @p input0.
97      * @param[in] info GEMM meta-data
98      *
99      * @return a status.
100      */
101     static Status validate(const ITensorInfo *a, const ITensorInfo *b, const ITensorInfo *c, const ITensorInfo *d, const AsmGemmInfo &info);
102     /** Checks if activation is supported by the gemm assembly dispatcher
103      *
104      * @param[in] activation Activation to check
105      *
106      * @return True if activation is supported else false
107      */
108     static bool is_activation_supported(const ActivationLayerInfo &activation);
109     /** Was the function successfully configured ?
110      *
111      * @return True if the function is configured and ready to run
112      */
113     bool is_configured() const;
114 
115     // Inherited methods overridden:
116     void prepare() override;
117     void run() override;
118 
119 private:
120     std::unique_ptr<IFallback> _arm_gemm;        /** Interface for the arm_gemm fallback */
121     MemoryGroup                _memory_group;    /**< Function memory group */
122     IWeightsManager           *_weights_manager; /**< Pointer to the weights manager */
123 };
124 } // namespace arm_compute
125 #endif /* ARM_COMPUTE_NEGEMMASSEMBLYDISPATCH_H */
126