Moritz Pflanzer | 4dfc235 | 2017-08-02 14:51:36 +0100 | [diff] [blame] | 1 | /* |
Michalis Spyrou | d1d7722 | 2020-04-08 14:10:15 +0100 | [diff] [blame] | 2 | * Copyright (c) 2017-2020 ARM Limited. |
Moritz Pflanzer | 4dfc235 | 2017-08-02 14:51:36 +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 | #include "GEMM.h" |
| 25 | |
Michalis Spyrou | d1d7722 | 2020-04-08 14:10:15 +0100 | [diff] [blame] | 26 | #include "arm_compute/core/Helpers.h" |
Georgios Pinitas | 583137c | 2017-08-31 18:12:42 +0100 | [diff] [blame] | 27 | #include "arm_compute/core/Types.h" |
Moritz Pflanzer | 4dfc235 | 2017-08-02 14:51:36 +0100 | [diff] [blame] | 28 | |
| 29 | namespace arm_compute |
| 30 | { |
| 31 | namespace test |
| 32 | { |
| 33 | namespace validation |
| 34 | { |
| 35 | namespace reference |
| 36 | { |
| 37 | template <typename T, typename std::enable_if<is_floating_point<T>::value, int>::type> |
| 38 | SimpleTensor<T> gemm(const SimpleTensor<T> &a, const SimpleTensor<T> &b, const SimpleTensor<T> &c, float alpha, float beta) |
| 39 | { |
| 40 | // Create reference |
Vidhya Sudhan Loganathan | 014333d | 2018-07-02 09:13:49 +0100 | [diff] [blame] | 41 | SimpleTensor<T> dst{ c.shape(), c.data_type(), 1 }; |
Moritz Pflanzer | 4dfc235 | 2017-08-02 14:51:36 +0100 | [diff] [blame] | 42 | |
| 43 | // Compute reference |
Gian Marco Iodice | 2213d4b | 2018-04-27 10:39:06 +0100 | [diff] [blame] | 44 | const int M = a.shape().y(); |
| 45 | const int N = b.shape().x(); |
Moritz Pflanzer | 4dfc235 | 2017-08-02 14:51:36 +0100 | [diff] [blame] | 46 | const int K = a.shape().x(); |
Gian Marco Iodice | 2213d4b | 2018-04-27 10:39:06 +0100 | [diff] [blame] | 47 | const int D = a.shape().z(); // Number of matrices in a batch |
| 48 | const int W = a.shape()[3]; // Number of batched-gemm (Winograd case) |
Moritz Pflanzer | 4dfc235 | 2017-08-02 14:51:36 +0100 | [diff] [blame] | 49 | |
Gian Marco Iodice | 2213d4b | 2018-04-27 10:39:06 +0100 | [diff] [blame] | 50 | const int a_stride_z = K * M; |
| 51 | const int a_stride_w = K * M * D; |
| 52 | |
| 53 | const int b_stride_z = b.shape().num_dimensions() > 2 ? N * K : 0; // Do not slide the matrix B along the 3th dimension in case matrix B has less than 3 dimensions |
| 54 | const int b_stride_w = b.shape().num_dimensions() > 3 ? K * N * D : 0; // Do not slide the matrix B along the 4th dimension in case matrix B has less than 4 dimensions |
| 55 | |
| 56 | const int c_stride_z = N * M; |
| 57 | const int c_stride_w = N * M * D; |
| 58 | |
Manuel Bottini | 9329175 | 2020-05-18 15:00:10 +0100 | [diff] [blame] | 59 | #if defined(_OPENMP) && !( defined(__arm__) && defined(__ANDROID__)) |
Michalis Spyrou | d1d7722 | 2020-04-08 14:10:15 +0100 | [diff] [blame] | 60 | #pragma omp parallel for collapse(2) |
| 61 | #endif /* _OPENMP */ |
Gian Marco Iodice | 2213d4b | 2018-04-27 10:39:06 +0100 | [diff] [blame] | 62 | for(int w = 0; w < W; ++w) |
Moritz Pflanzer | 4dfc235 | 2017-08-02 14:51:36 +0100 | [diff] [blame] | 63 | { |
Gian Marco Iodice | 2213d4b | 2018-04-27 10:39:06 +0100 | [diff] [blame] | 64 | for(int depth = 0; depth < D; ++depth) |
Moritz Pflanzer | 4dfc235 | 2017-08-02 14:51:36 +0100 | [diff] [blame] | 65 | { |
Gian Marco Iodice | 2213d4b | 2018-04-27 10:39:06 +0100 | [diff] [blame] | 66 | const int base_addr_a = depth * a_stride_z + w * a_stride_w; |
| 67 | const int base_addr_b = depth * b_stride_z + w * b_stride_w; |
| 68 | const int base_addr_c = depth * c_stride_z + w * c_stride_w; |
Moritz Pflanzer | 4dfc235 | 2017-08-02 14:51:36 +0100 | [diff] [blame] | 69 | |
Gian Marco Iodice | 2213d4b | 2018-04-27 10:39:06 +0100 | [diff] [blame] | 70 | for(int row = 0; row < M; ++row) |
Moritz Pflanzer | 4dfc235 | 2017-08-02 14:51:36 +0100 | [diff] [blame] | 71 | { |
Gian Marco Iodice | 2213d4b | 2018-04-27 10:39:06 +0100 | [diff] [blame] | 72 | for(int col = 0; col < N; ++col) |
| 73 | { |
| 74 | T acc(0); |
Moritz Pflanzer | 4dfc235 | 2017-08-02 14:51:36 +0100 | [diff] [blame] | 75 | |
Gian Marco Iodice | 2213d4b | 2018-04-27 10:39:06 +0100 | [diff] [blame] | 76 | for(int k = 0; k < K; ++k) |
| 77 | { |
| 78 | acc += a[base_addr_a + k + row * K] * b[base_addr_b + col + k * N]; |
| 79 | } |
| 80 | |
| 81 | // Finalize the result: alpha * A * B + beta * C |
| 82 | dst[base_addr_c + col + row * N] = alpha * acc + beta * c[base_addr_c + col + row * N]; |
| 83 | } |
| 84 | } |
Moritz Pflanzer | 4dfc235 | 2017-08-02 14:51:36 +0100 | [diff] [blame] | 85 | } |
| 86 | } |
| 87 | |
| 88 | return dst; |
| 89 | } |
| 90 | |
Gian Marco Iodice | 0c17aa2 | 2019-09-27 09:23:15 +0100 | [diff] [blame] | 91 | template <typename T, typename std::enable_if<is_floating_point<T>::value, int>::type> |
| 92 | SimpleTensor<T> gemm_mixed_precision(const SimpleTensor<T> &a, const SimpleTensor<T> &b, const SimpleTensor<T> &c, float alpha, float beta) |
| 93 | { |
| 94 | // GEMM mixed-precision combines F32 accumulators with F16 multiplications |
| 95 | // Create reference |
| 96 | SimpleTensor<T> dst{ c.shape(), c.data_type(), 1 }; |
| 97 | |
| 98 | // Compute reference |
| 99 | const int M = a.shape().y(); |
| 100 | const int N = b.shape().x(); |
| 101 | const int K = a.shape().x(); |
| 102 | const int D = a.shape().z(); // Number of matrices in a batch |
| 103 | const int W = a.shape()[3]; // Number of batched-gemm (Winograd case) |
| 104 | |
| 105 | const int a_stride_z = K * M; |
| 106 | const int a_stride_w = K * M * D; |
| 107 | |
| 108 | const int b_stride_z = b.shape().num_dimensions() > 2 ? N * K : 0; // Do not slide the matrix B along the 3th dimension in case matrix B has less than 3 dimensions |
| 109 | const int b_stride_w = b.shape().num_dimensions() > 3 ? K * N * D : 0; // Do not slide the matrix B along the 4th dimension in case matrix B has less than 4 dimensions |
| 110 | |
| 111 | const int c_stride_z = N * M; |
| 112 | const int c_stride_w = N * M * D; |
| 113 | |
Manuel Bottini | d5af1a0 | 2020-05-20 12:28:45 +0100 | [diff] [blame] | 114 | #if defined(_OPENMP) && !( defined(__arm__) && defined(__ANDROID__)) |
Michalis Spyrou | d1d7722 | 2020-04-08 14:10:15 +0100 | [diff] [blame] | 115 | #pragma omp parallel for collapse(2) |
| 116 | #endif /* _OPENMP */ |
Gian Marco Iodice | 0c17aa2 | 2019-09-27 09:23:15 +0100 | [diff] [blame] | 117 | for(int w = 0; w < W; ++w) |
| 118 | { |
| 119 | for(int depth = 0; depth < D; ++depth) |
| 120 | { |
| 121 | const int base_addr_a = depth * a_stride_z + w * a_stride_w; |
| 122 | const int base_addr_b = depth * b_stride_z + w * b_stride_w; |
| 123 | const int base_addr_c = depth * c_stride_z + w * c_stride_w; |
| 124 | |
| 125 | for(int row = 0; row < M; ++row) |
| 126 | { |
| 127 | for(int col = 0; col < N; ++col) |
| 128 | { |
| 129 | float acc(0); |
| 130 | |
| 131 | for(int k = 0; k < K; ++k) |
| 132 | { |
| 133 | acc += static_cast<float>(a[base_addr_a + k + row * K] * b[base_addr_b + col + k * N]); |
| 134 | } |
| 135 | |
| 136 | // Finalize the result: alpha * A * B + beta * C |
| 137 | dst[base_addr_c + col + row * N] = static_cast<T>(alpha * acc + beta * c[base_addr_c + col + row * N]); |
| 138 | } |
| 139 | } |
| 140 | } |
| 141 | } |
| 142 | |
| 143 | return dst; |
| 144 | } |
| 145 | |
Moritz Pflanzer | 4dfc235 | 2017-08-02 14:51:36 +0100 | [diff] [blame] | 146 | template SimpleTensor<float> gemm(const SimpleTensor<float> &a, const SimpleTensor<float> &b, const SimpleTensor<float> &c, float alpha, float beta); |
Georgios Pinitas | 583137c | 2017-08-31 18:12:42 +0100 | [diff] [blame] | 147 | template SimpleTensor<half> gemm(const SimpleTensor<half> &a, const SimpleTensor<half> &b, const SimpleTensor<half> &c, float alpha, float beta); |
Gian Marco Iodice | 0c17aa2 | 2019-09-27 09:23:15 +0100 | [diff] [blame] | 148 | template SimpleTensor<half> gemm_mixed_precision(const SimpleTensor<half> &a, const SimpleTensor<half> &b, const SimpleTensor<half> &c, float alpha, float beta); |
Moritz Pflanzer | 4dfc235 | 2017-08-02 14:51:36 +0100 | [diff] [blame] | 149 | } // namespace reference |
| 150 | } // namespace validation |
| 151 | } // namespace test |
| 152 | } // namespace arm_compute |