Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1 | /* |
Michele Di Giorgio | d9eaf61 | 2020-07-08 11:12:57 +0100 | [diff] [blame] | 2 | * Copyright (c) 2017-2020 Arm Limited. |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +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 | */ |
Michalis Spyrou | f464337 | 2019-11-29 16:17:13 +0000 | [diff] [blame] | 24 | #ifndef ARM_COMPUTE_NEGEMM_H |
| 25 | #define ARM_COMPUTE_NEGEMM_H |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 26 | |
| 27 | #include "arm_compute/core/NEON/kernels/NEFillBorderKernel.h" |
| 28 | #include "arm_compute/core/NEON/kernels/NEGEMMInterleave4x4Kernel.h" |
| 29 | #include "arm_compute/core/NEON/kernels/NEGEMMMatrixAdditionKernel.h" |
| 30 | #include "arm_compute/core/NEON/kernels/NEGEMMMatrixMultiplyKernel.h" |
| 31 | #include "arm_compute/core/NEON/kernels/NEGEMMTranspose1xWKernel.h" |
| 32 | #include "arm_compute/runtime/IFunction.h" |
Georgios Pinitas | 658039b | 2017-09-15 16:30:50 +0100 | [diff] [blame] | 33 | #include "arm_compute/runtime/IMemoryManager.h" |
Michalis Spyrou | 1a569a3 | 2019-09-10 17:20:34 +0100 | [diff] [blame] | 34 | #include "arm_compute/runtime/IWeightsManager.h" |
Georgios Pinitas | 658039b | 2017-09-15 16:30:50 +0100 | [diff] [blame] | 35 | #include "arm_compute/runtime/MemoryGroup.h" |
Georgios Pinitas | 48b3ef8 | 2019-10-14 19:03:09 +0100 | [diff] [blame] | 36 | #include "arm_compute/runtime/NEON/functions/NEActivationLayer.h" |
Michalis Spyrou | 173ba9b | 2020-06-23 17:25:43 +0100 | [diff] [blame] | 37 | #include "arm_compute/runtime/NEON/functions/NEArithmeticAddition.h" |
Anthony Barbier | 71d9b57 | 2018-07-06 17:05:59 +0100 | [diff] [blame] | 38 | #include "arm_compute/runtime/NEON/functions/NEGEMMAssemblyDispatch.h" |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 39 | #include "arm_compute/runtime/Tensor.h" |
| 40 | |
| 41 | namespace arm_compute |
| 42 | { |
| 43 | /** Basic function to execute GEMM on NEON. This function calls the following NEON kernels: |
| 44 | * |
Georgios Pinitas | 48b3ef8 | 2019-10-14 19:03:09 +0100 | [diff] [blame] | 45 | * If optimized assembly is available: |
| 46 | * -# @ref NEGEMMAssemblyDispatch |
| 47 | * -# @ref NEActivationLayer (if alpha != 1.0) |
| 48 | * Else: |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 49 | * -# @ref NEGEMMInterleave4x4Kernel (if the output tensor is a matrix) |
| 50 | * -# @ref NEGEMMTranspose1xWKernel (if the output tensor is a matrix) |
| 51 | * -# @ref NEGEMMMatrixMultiplyKernel |
Georgios Pinitas | 48b3ef8 | 2019-10-14 19:03:09 +0100 | [diff] [blame] | 52 | * In both cases: |
| 53 | * -# @ref NEGEMMMatrixAdditionKernel (if c != nullptr and beta != 0.0 and is not reshaped once) |
| 54 | * Else: |
| 55 | * -# @ref NEArithmeticAdditionKernel (if c != nullptr and is reshaped once and not optimized assembly in place) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 56 | * |
Georgios Pinitas | 48b3ef8 | 2019-10-14 19:03:09 +0100 | [diff] [blame] | 57 | * -# @ref NEActivationLayer (if activation is specified in GEMMInfo) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 58 | */ |
| 59 | class NEGEMM : public IFunction |
| 60 | { |
| 61 | public: |
| 62 | /** Constructor */ |
Michalis Spyrou | 1a569a3 | 2019-09-10 17:20:34 +0100 | [diff] [blame] | 63 | NEGEMM(std::shared_ptr<IMemoryManager> memory_manager = nullptr, IWeightsManager *weights_manager = nullptr); |
Georgios Pinitas | 7221933 | 2018-06-05 14:56:06 +0100 | [diff] [blame] | 64 | /** Prevent instances of this class from being copied (As this class contains pointers) */ |
| 65 | NEGEMM(const NEGEMM &) = delete; |
| 66 | /** Default move constructor */ |
| 67 | NEGEMM(NEGEMM &&) = default; |
| 68 | /** Prevent instances of this class from being copied (As this class contains pointers) */ |
| 69 | NEGEMM &operator=(const NEGEMM &) = delete; |
| 70 | /** Default move assignment operator */ |
| 71 | NEGEMM &operator=(NEGEMM &&) = default; |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 72 | /** Initialise the kernel's inputs, output |
| 73 | * |
| 74 | * @note GEMM: General Matrix Multiply - [alpha * A * B + beta * C]. |
| 75 | * @note GEMM: The tensors a, b, c, d must have the same data type. You should not mix data types when calling this function. |
| 76 | * |
Michele Di Giorgio | a602f03 | 2020-03-12 19:34:33 +0000 | [diff] [blame] | 77 | * @param[in] a First input tensor (Matrix A or Vector A). Data type supported: BFLOAT16/F16/F32 |
Gian Marco | 1d25ed5 | 2017-12-16 19:33:50 +0000 | [diff] [blame] | 78 | * @param[in] b Second input tensor (Matrix B). Data type supported: same as @p a |
| 79 | * @param[in] c Third input tensor (Matrix C). It can be a nullptr if just the multiplication between @p a and @p b is needed. Data type supported: same as @p a |
| 80 | * @param[out] d Output tensor. Data type supported: same as @p a |
| 81 | * @param[in] alpha Weight of the matrix product |
| 82 | * @param[in] beta Weight of matrix C |
| 83 | * @param[in] gemm_info (Optional) Specifies if the matrix A and/or matrix B have been reshaped and |
| 84 | * if the reshape of matrix B should happen only for the first run |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 85 | */ |
Gian Marco | 1d25ed5 | 2017-12-16 19:33:50 +0000 | [diff] [blame] | 86 | void configure(const ITensor *a, const ITensor *b, const ITensor *c, ITensor *d, float alpha, float beta, const GEMMInfo &gemm_info = GEMMInfo()); |
Giorgio Arena | a855af1 | 2018-07-16 17:20:38 +0100 | [diff] [blame] | 87 | /** Static function to check if given info will lead to a valid configuration of @ref NEGEMM. |
| 88 | * |
Michele Di Giorgio | a602f03 | 2020-03-12 19:34:33 +0000 | [diff] [blame] | 89 | * @param[in] a First input tensor info (Matrix or Vector A). Data types supported: BFLOAT16/F16/F32 |
Giorgio Arena | a855af1 | 2018-07-16 17:20:38 +0100 | [diff] [blame] | 90 | * @param[in] b Second input tensor info (Matrix B). Data type supported: same as @p a. |
| 91 | * @param[in] c Third input tensor info (Matrix C). It can be a nullptr if just the multiplication between @p a and @p b is needed. Data type supported: same as @p a. |
| 92 | * @param[out] output Output tensor info. Data type supported: same as @p a |
| 93 | * @param[in] alpha Weight of the matrix product |
| 94 | * @param[in] beta Weight of matrix C |
| 95 | * @param[in] gemm_info (Optional) Specifies if the matrix A and/or matrix B have been reshaped and |
| 96 | * if the reshape of matrix B should happen only for the first run |
| 97 | * |
| 98 | * @return a status |
| 99 | */ |
| 100 | static Status validate(const ITensorInfo *a, const ITensorInfo *b, const ITensorInfo *c, const ITensorInfo *output, float alpha, float beta, const GEMMInfo &gemm_info = GEMMInfo()); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 101 | |
| 102 | // Inherited methods overridden: |
| 103 | void run() override; |
Georgios Pinitas | 7221933 | 2018-06-05 14:56:06 +0100 | [diff] [blame] | 104 | void prepare() override; |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 105 | |
| 106 | private: |
Pablo Tello | eb82fd2 | 2018-02-23 13:43:50 +0000 | [diff] [blame] | 107 | MemoryGroup _memory_group; |
Michalis Spyrou | 1a569a3 | 2019-09-10 17:20:34 +0100 | [diff] [blame] | 108 | IWeightsManager *_weights_manager; |
Pablo Tello | eb82fd2 | 2018-02-23 13:43:50 +0000 | [diff] [blame] | 109 | NEGEMMInterleave4x4Kernel _interleave_kernel; |
| 110 | NEGEMMTranspose1xWKernel _transpose_kernel; |
| 111 | NEGEMMMatrixMultiplyKernel _mm_kernel; |
Anthony Barbier | eaefd00 | 2018-07-20 17:49:35 +0100 | [diff] [blame] | 112 | NEGEMMAssemblyDispatch _asm_glue; |
Pablo Tello | eb82fd2 | 2018-02-23 13:43:50 +0000 | [diff] [blame] | 113 | NEGEMMMatrixAdditionKernel _ma_kernel; |
Georgios Pinitas | 48b3ef8 | 2019-10-14 19:03:09 +0100 | [diff] [blame] | 114 | NEActivationLayer _alpha_scale_func; |
Michalis Spyrou | 173ba9b | 2020-06-23 17:25:43 +0100 | [diff] [blame] | 115 | NEArithmeticAddition _add_bias; |
Georgios Pinitas | 48b3ef8 | 2019-10-14 19:03:09 +0100 | [diff] [blame] | 116 | NEActivationLayer _activation_func; |
| 117 | |
| 118 | Tensor _tmp_a; |
| 119 | Tensor _tmp_b; |
| 120 | Tensor _tmp_d; |
| 121 | const ITensor *_original_b; |
| 122 | bool _run_vector_matrix_multiplication; |
| 123 | bool _run_alpha_scale; |
| 124 | bool _run_addition; |
| 125 | bool _run_bias_addition; |
| 126 | bool _run_activation; |
| 127 | bool _reshape_b_only_on_first_run; |
| 128 | bool _is_prepared; |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 129 | }; |
Georgios Pinitas | 7221933 | 2018-06-05 14:56:06 +0100 | [diff] [blame] | 130 | } // namespace arm_compute |
Michalis Spyrou | f464337 | 2019-11-29 16:17:13 +0000 | [diff] [blame] | 131 | #endif /*ARM_COMPUTE_NEGEMM_H */ |