xref: /aosp_15_r20/external/ComputeLibrary/src/cpu/operators/CpuGemmDirectConv2d.h (revision c217d954acce2dbc11938adb493fc0abd69584f3)
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_CPU_GEMM_DIRECT_CONV_2D_H
25 #define ARM_COMPUTE_CPU_GEMM_DIRECT_CONV_2D_H
26 
27 #include "arm_compute/core/TensorInfo.h"
28 #include "src/core/common/Macros.h"
29 #include "src/cpu/ICpuOperator.h"
30 #include "src/cpu/operators/CpuActivation.h"
31 #include "src/cpu/operators/CpuPermute.h"
32 #include "src/cpu/operators/internal/CpuGemmAssemblyDispatch.h"
33 
34 namespace arm_compute
35 {
36 // Forward declarations
37 class ITensor;
38 struct Conv2dInfo;
39 namespace cpu
40 {
41 class CpuGemmDirectConv2d : public ICpuOperator
42 {
43 public:
44     CpuGemmDirectConv2d();
45     ARM_COMPUTE_DISALLOW_COPY_ALLOW_MOVE(CpuGemmDirectConv2d);
46     ~CpuGemmDirectConv2d();
47     /** Set the input and output tensors.
48      *
49      * Valid data layouts:
50      * - All
51      *
52      * Valid data type configurations:
53      * |src0           |src1           |src2           |dst            |
54      * |:--------------|:--------------|:--------------|:--------------|
55      * |QASYMM8        |QASYMM8        |S32            |QASYMM8        |
56      * |QASYMM8_SIGNED |QASYMM8_SIGNED |S32            |QASYMM8_SIGNED |
57      * |F16            |F16            |F16            |F16            |
58      * |F32            |F32            |F32            |F32            |
59      * |BFLOAT16       |BFLOAT16       |BFLOAT16       |BFLOAT16       |
60      *
61      * @param[in] src     Source tensor info. 3 lower dimensions represent a single input [width, height, IFM],
62      *                    while every optional dimension from 4 and above represent a batch of inputs.
63      *                    Data types supported: QASYMM8/QASYMM8_SIGNED/BFLOAT16/F16/F32.
64      * @param[in] weights Weights tensor info. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM].
65      *                    Data type supported: QASYMM8/QASYMM8_SIGNED/QSYMM8_PER_CHANNEL/BFLOAT16/F16/F32.
66      * @param[in] biases  Biases tensor info. Shared biases supported. Biases are 1D tensor with dimensions [OFM].
67      *                    Data type supported: Should match @p input data type, except for input of QASYMM8/QASYMM8_SIGNED type where biases should be of S32 type.
68      * @param[in] dst     Destination tensor info. 3 lower dimensions represent a single output [width, height, OFM], while the rest represent batch of outputs.
69      *                    Data types supported: Same as @p input.
70      * @param[in] info    Contains padding and stride information described in @ref PadStrideInfo.
71      */
72     void configure(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *biases, ITensorInfo *dst, const Conv2dInfo &info);
73     /** Static function to check if given info will lead to a valid configuration of @ref CpuGemmDirectConv2d
74      *
75      * Similar to CpuGemmDirectConv2d::configure()
76      *
77      * @return a status
78      */
79     static Status validate(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *dst, const Conv2dInfo &info);
80 
81     // Inherited methods overridden:
82     void run(ITensorPack &tensors) override;
83     void prepare(ITensorPack &constants) override;
84     experimental::MemoryRequirements workspace() const override;
85 
86 private:
87     enum AuxTensorIdx
88     {
89         AsmGemmWorkspace = 0,
90         Pretranspose,
91         PermutedWeights,
92         Count
93     };
94 
95     std::unique_ptr<CpuGemmAssemblyDispatch> _gemm_asm_func;
96     std::unique_ptr<CpuActivation>           _activation_func;
97     std::unique_ptr<CpuPermute>              _weights_permute_func;
98     experimental::MemoryRequirements         _aux_mem;
99     TensorInfo                               _perm_weights;
100     bool                                     _run_activation;
101     bool                                     _is_prepared;
102 };
103 } // namespace cpu
104 } // namespace arm_compute
105 
106 #endif /* ARM_COMPUTE_CPU_GEMM_DIRECT_CONV_2D_H */
107