Michele Di Giorgio | 53832b2 | 2021-06-21 14:45:44 +0100 | [diff] [blame] | 1 | /* |
Giorgio Arena | 5ae8d80 | 2021-11-18 18:02:13 +0000 | [diff] [blame] | 2 | * Copyright (c) 2017-2022 Arm Limited. |
Michele Di Giorgio | 53832b2 | 2021-06-21 14:45:44 +0100 | [diff] [blame] | 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" |
Georgios Pinitas | 7891a73 | 2021-08-20 21:39:25 +0100 | [diff] [blame] | 28 | #include "src/cpu/ICpuKernel.h" |
Michele Di Giorgio | 53832b2 | 2021-06-21 14:45:44 +0100 | [diff] [blame] | 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 | */ |
Yair Schwarzbaum | 46d44d2 | 2022-01-12 16:38:58 +0200 | [diff] [blame] | 42 | class CpuGemmMatrixMultiplyKernel : public ICpuKernel<CpuGemmMatrixMultiplyKernel> |
Michele Di Giorgio | 53832b2 | 2021-06-21 14:45:44 +0100 | [diff] [blame] | 43 | { |
Dana Zlotnik | 256ac62 | 2022-02-02 15:06:11 +0200 | [diff] [blame] | 44 | private: |
Felix Thomasmathibalan | afd38f0 | 2023-09-27 17:46:17 +0100 | [diff] [blame] | 45 | using GemmMatrixMulKernelPtr = std::add_pointer<void( |
| 46 | const ITensor *, const ITensor *, ITensor *, const Window &, const ThreadInfo &, float, const bool)>::type; |
Dana Zlotnik | 256ac62 | 2022-02-02 15:06:11 +0200 | [diff] [blame] | 47 | |
Michele Di Giorgio | 53832b2 | 2021-06-21 14:45:44 +0100 | [diff] [blame] | 48 | public: |
Dana Zlotnik | 256ac62 | 2022-02-02 15:06:11 +0200 | [diff] [blame] | 49 | struct GemmMatrixMulKernel |
| 50 | { |
| 51 | const char *name; |
| 52 | const DataTypeISASelectorPtr is_selected; |
| 53 | GemmMatrixMulKernelPtr ukernel; |
| 54 | }; |
| 55 | |
Michele Di Giorgio | 53832b2 | 2021-06-21 14:45:44 +0100 | [diff] [blame] | 56 | CpuGemmMatrixMultiplyKernel() = default; |
| 57 | ARM_COMPUTE_DISALLOW_COPY_ALLOW_MOVE(CpuGemmMatrixMultiplyKernel); |
| 58 | /** Initialise the kernel's input and output. |
| 59 | * |
| 60 | * @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 |
| 61 | * These two kernels change the layout of the original matrices to be more cache-friendly. |
| 62 | * |
| 63 | * @param[in] lhs Left-handside tensor info containing the interleaved Matrix A or the vector A. Data types supported: F16/F32 |
| 64 | * @param[in] rhs Right-handside tensor info containing the transposed Matrix B if the first input tensor A is not a vector. |
| 65 | * If the output tensor is a vector, rhs must contain the matrix B not reshaped. Data type supported: same as @p lhs |
| 66 | * @param[out] dst Output tensor to store the result of matrix multiplication. Data type supported: same as @p lhs. |
| 67 | * @param[in] alpha Weight of the matrix product |
| 68 | * @param[in] is_interleaved (Optional) True if lhs and rhs have been reshaped respectively using @ref CpuGemmInterleave4x4Kernel and @ref CpuGemmTranspose1xWKernel |
| 69 | * @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 |
| 70 | */ |
Felix Thomasmathibalan | afd38f0 | 2023-09-27 17:46:17 +0100 | [diff] [blame] | 71 | void configure(const ITensorInfo *lhs, |
| 72 | const ITensorInfo *rhs, |
| 73 | ITensorInfo *dst, |
| 74 | float alpha, |
| 75 | bool is_interleaved, |
| 76 | const GEMMReshapeInfo &reshape_info = GEMMReshapeInfo()); |
Michele Di Giorgio | 53832b2 | 2021-06-21 14:45:44 +0100 | [diff] [blame] | 77 | /** Static function to check if given info will lead to a valid configuration of @ref CpuGemmMatrixMultiplyKernel |
| 78 | * |
| 79 | * Similar to @ref CpuGemmMatrixMultiplyKernel::configure() |
| 80 | * |
| 81 | * @return a status |
| 82 | */ |
Felix Thomasmathibalan | afd38f0 | 2023-09-27 17:46:17 +0100 | [diff] [blame] | 83 | static Status validate(const ITensorInfo *lhs, |
| 84 | const ITensorInfo *rhs, |
| 85 | const ITensorInfo *dst, |
| 86 | float alpha, |
| 87 | bool is_interleaved, |
| 88 | const GEMMReshapeInfo &reshape_info); |
Michele Di Giorgio | 53832b2 | 2021-06-21 14:45:44 +0100 | [diff] [blame] | 89 | |
| 90 | // Inherited methods overridden: |
Felix Thomasmathibalan | afd38f0 | 2023-09-27 17:46:17 +0100 | [diff] [blame] | 91 | void run_op(ITensorPack &tensors, const Window &window, const ThreadInfo &info) override; |
Michele Di Giorgio | 53832b2 | 2021-06-21 14:45:44 +0100 | [diff] [blame] | 92 | const char *name() const override; |
| 93 | |
Dana Zlotnik | 256ac62 | 2022-02-02 15:06:11 +0200 | [diff] [blame] | 94 | static const std::vector<GemmMatrixMulKernel> &get_available_kernels(); |
| 95 | |
Michele Di Giorgio | 53832b2 | 2021-06-21 14:45:44 +0100 | [diff] [blame] | 96 | private: |
| 97 | /** Common signature for all the matrix multiply functions |
| 98 | * |
| 99 | * @param[in] lhs Left-handside input tensor. Data types supported: F16/F32 |
| 100 | * @param[in] rhs Right-handside input tensor. Data types supported: same as @p lhs |
| 101 | * @param[out] dst The output tensor. Data type supported: same as @p rhs |
| 102 | * @param[in] window Region on which to execute the kernel. |
| 103 | * @param[in] info Thread info metadata. |
| 104 | * @param[in] alpha Weight of the matrix product. |
| 105 | */ |
Dana Zlotnik | 256ac62 | 2022-02-02 15:06:11 +0200 | [diff] [blame] | 106 | |
Michele Di Giorgio | 53832b2 | 2021-06-21 14:45:44 +0100 | [diff] [blame] | 107 | /** Matrix multiply function to use for the particular tensor types passed to configure() */ |
Felix Thomasmathibalan | afd38f0 | 2023-09-27 17:46:17 +0100 | [diff] [blame] | 108 | GemmMatrixMulKernelPtr _func{nullptr}; |
| 109 | float _alpha{1.f}; |
Michele Di Giorgio | 53832b2 | 2021-06-21 14:45:44 +0100 | [diff] [blame] | 110 | }; |
| 111 | } // namespace kernels |
| 112 | } // namespace cpu |
| 113 | } // namespace arm_compute |
Georgios Pinitas | 2eb5d16 | 2021-07-02 09:01:49 +0100 | [diff] [blame] | 114 | #endif /* ARM_COMPUTE_CPU_GEMM_MATRIX_MULTIPLY_KERNEL_H */ |