xref: /aosp_15_r20/external/ComputeLibrary/src/cpu/kernels/CpuGemmMatrixMultiplyKernel.h (revision c217d954acce2dbc11938adb493fc0abd69584f3)
1 /*
2  * Copyright (c) 2017-2022 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_MATRIX_MULTIPLY_KERNEL_H
25 #define ARM_COMPUTE_CPU_GEMM_MATRIX_MULTIPLY_KERNEL_H
26 
27 #include "src/core/common/Macros.h"
28 #include "src/cpu/ICpuKernel.h"
29 
30 namespace arm_compute
31 {
32 namespace cpu
33 {
34 namespace kernels
35 {
36 /** Kernel to multiply two input matrices "A" and "B". All elements of the output matrix/vector will be multiplied by alpha after the matrix multiplication
37  *
38  * @note If the output tensor is a matrix, the implementation assumes that the input tensors @p lhs and @p rhs are both matrices and reshaped respectively with @ref CpuGemmInterleave4x4Kernel" and @ref CpuGemmTranspose1xWKernel
39  * @note If the output tensor is a vector and the data type is F32, the implementation assumes that the first input tensor @p lhs is a vector and the second input tensor @p rhs a matrix. The implementation also assumes that both tensors have not been reshaped
40  *
41  */
42 class CpuGemmMatrixMultiplyKernel : public ICpuKernel<CpuGemmMatrixMultiplyKernel>
43 {
44 private:
45     using GemmMatrixMulKernelPtr = std::add_pointer<void(const ITensor *, const ITensor *, ITensor *, const Window &, const ThreadInfo &, float, const bool)>::type;
46 
47 public:
48     struct GemmMatrixMulKernel
49     {
50         const char                  *name;
51         const DataTypeISASelectorPtr is_selected;
52         GemmMatrixMulKernelPtr       ukernel;
53     };
54 
55     CpuGemmMatrixMultiplyKernel() = default;
56     ARM_COMPUTE_DISALLOW_COPY_ALLOW_MOVE(CpuGemmMatrixMultiplyKernel);
57     /** Initialise the kernel's input and output.
58      *
59      * @note If the output tensor is a matrix, the input matrices @p lhs and @p rhs should be the output of the kernels: @ref CpuGemmInterleave4x4Kernel and @ref CpuGemmTranspose1xWKernel
60      *       These two kernels change the layout of the original matrices to be more cache-friendly.
61      *
62      * @param[in]  lhs            Left-handside tensor info containing the interleaved Matrix A or the vector A. Data types supported: F16/F32
63      * @param[in]  rhs            Right-handside tensor info containing the transposed Matrix B if the first input tensor A is not a vector.
64      *                            If the output tensor is a vector, rhs must contain the matrix B not reshaped. Data type supported: same as @p lhs
65      * @param[out] dst            Output tensor to store the result of matrix multiplication. Data type supported: same as @p lhs.
66      * @param[in]  alpha          Weight of the matrix product
67      * @param[in]  is_interleaved (Optional) True if lhs and rhs have been reshaped respectively using @ref CpuGemmInterleave4x4Kernel and @ref CpuGemmTranspose1xWKernel
68      * @param[in]  reshape_info   (Optional) GEMM reshape info. If is_interleaved_transposed = true, this object must contain the information to understand how @p lhs and @p rhs have been reshaped
69      */
70     void configure(const ITensorInfo *lhs, const ITensorInfo *rhs, ITensorInfo *dst, float alpha, bool is_interleaved, const GEMMReshapeInfo &reshape_info = GEMMReshapeInfo());
71     /** Static function to check if given info will lead to a valid configuration of @ref CpuGemmMatrixMultiplyKernel
72      *
73      * Similar to @ref CpuGemmMatrixMultiplyKernel::configure()
74      *
75      * @return a status
76      */
77     static Status validate(const ITensorInfo *lhs, const ITensorInfo *rhs, const ITensorInfo *dst, float alpha, bool is_interleaved, const GEMMReshapeInfo &reshape_info);
78 
79     // Inherited methods overridden:
80     void run_op(ITensorPack &tensors, const Window &window, const ThreadInfo &info) override;
81     const char *name() const override;
82 
83     static const std::vector<GemmMatrixMulKernel> &get_available_kernels();
84 
85 private:
86     /** Common signature for all the matrix multiply functions
87      *
88      * @param[in]  lhs    Left-handside input tensor. Data types supported: F16/F32
89      * @param[in]  rhs    Right-handside input tensor. Data types supported: same as @p lhs
90      * @param[out] dst    The output tensor. Data type supported: same as @p rhs
91      * @param[in]  window Region on which to execute the kernel.
92      * @param[in]  info   Thread info metadata.
93      * @param[in]  alpha  Weight of the matrix product.
94      */
95 
96     /** Matrix multiply function to use for the particular tensor types passed to configure() */
97     GemmMatrixMulKernelPtr _func{ nullptr };
98     float                  _alpha{ 1.f };
99 };
100 } // namespace kernels
101 } // namespace cpu
102 } // namespace arm_compute
103 #endif /* ARM_COMPUTE_CPU_GEMM_MATRIX_MULTIPLY_KERNEL_H */
104