Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1 | /* |
Gian Marco | 20d7848 | 2018-01-11 15:10:58 +0000 | [diff] [blame] | 2 | * Copyright (c) 2017-2018 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 | */ |
| 24 | #include "arm_compute/runtime/CL/functions/CLGEMM.h" |
| 25 | |
| 26 | #include "arm_compute/core/CL/ICLTensor.h" |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 27 | #include "arm_compute/core/Error.h" |
Gian Marco Iodice | 750641d | 2018-05-08 12:01:57 +0100 | [diff] [blame] | 28 | #include "arm_compute/core/GPUTarget.h" |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 29 | #include "arm_compute/core/Helpers.h" |
| 30 | #include "arm_compute/core/TensorInfo.h" |
| 31 | #include "arm_compute/core/Types.h" |
Gian Marco Iodice | bb36a8e | 2018-04-19 12:05:08 +0100 | [diff] [blame] | 32 | #include "arm_compute/core/Utils.h" |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 33 | #include "arm_compute/core/Validate.h" |
Gian Marco Iodice | 750641d | 2018-05-08 12:01:57 +0100 | [diff] [blame] | 34 | #include "arm_compute/core/utils/misc/ShapeCalculator.h" |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 35 | #include "arm_compute/runtime/CL/CLScheduler.h" |
| 36 | #include "arm_compute/runtime/ITensorAllocator.h" |
| 37 | |
| 38 | using namespace arm_compute; |
Gian Marco Iodice | 750641d | 2018-05-08 12:01:57 +0100 | [diff] [blame] | 39 | using namespace arm_compute::misc::shape_calculator; |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 40 | |
Gian Marco | 36a0a46 | 2018-01-12 10:21:40 +0000 | [diff] [blame] | 41 | namespace |
| 42 | { |
| 43 | inline bool is_interleaved_transposed(int m, int n, int k, DataType data_type, bool reshape_b_only_on_first_run, GPUTarget gpu_target) |
| 44 | { |
| 45 | bool flag = true; |
| 46 | |
Gian Marco Iodice | 513fe2e | 2018-06-04 18:08:48 +0100 | [diff] [blame] | 47 | if(gpu_target_is_in(gpu_target, GPUTarget::G71, GPUTarget::G72)) |
Gian Marco | 36a0a46 | 2018-01-12 10:21:40 +0000 | [diff] [blame] | 48 | { |
| 49 | // COMPMID-852 |
Gian Marco Iodice | bb36a8e | 2018-04-19 12:05:08 +0100 | [diff] [blame] | 50 | if(k > 256 && m > 4 && is_data_type_float(data_type) && reshape_b_only_on_first_run) |
Gian Marco | 36a0a46 | 2018-01-12 10:21:40 +0000 | [diff] [blame] | 51 | { |
Gian Marco Iodice | 513fe2e | 2018-06-04 18:08:48 +0100 | [diff] [blame] | 52 | constexpr float alpha = 3.2f; |
| 53 | constexpr float fact0 = 1.51f; |
| 54 | constexpr float fact1 = 1.66f; |
| 55 | constexpr float ops = 12.0f; |
| 56 | const float scale = k > 1024 ? 1.07f : 1.0f; |
| 57 | flag = alpha + ((n * fact0) / ops) < ((fact1 * n * scale) / ops); |
Gian Marco | 36a0a46 | 2018-01-12 10:21:40 +0000 | [diff] [blame] | 58 | } |
| 59 | else |
| 60 | { |
| 61 | flag = false; |
| 62 | } |
| 63 | } |
Gian Marco Iodice | cda0c38 | 2018-04-23 16:16:22 +0100 | [diff] [blame] | 64 | else |
| 65 | { |
| 66 | // We reshape the matrices only if we do not have the vector-by-matrix case and we reshape the matrix B only once |
| 67 | flag = m != 1 && reshape_b_only_on_first_run; |
| 68 | } |
Gian Marco | 36a0a46 | 2018-01-12 10:21:40 +0000 | [diff] [blame] | 69 | |
| 70 | return flag; |
| 71 | } |
| 72 | } // namespace |
| 73 | |
Georgios Pinitas | 8a94e7c | 2017-09-15 19:06:47 +0100 | [diff] [blame] | 74 | CLGEMM::CLGEMM(std::shared_ptr<IMemoryManager> memory_manager) |
Georgios Pinitas | 82b5148 | 2018-04-24 15:14:12 +0100 | [diff] [blame] | 75 | : _memory_group(std::move(memory_manager)), _interleave_kernel(), _transpose_kernel(), _mm_kernel(), _ma_kernel(), _tmp_a(), _tmp_b(), _original_b(nullptr), _is_interleaved_transposed(false), |
Georgios Pinitas | e043767 | 2018-05-02 14:07:55 +0100 | [diff] [blame] | 76 | _run_addition(false), _reshape_b_only_on_first_run(false), _is_prepared(false) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 77 | { |
| 78 | } |
| 79 | |
Gian Marco | 1d25ed5 | 2017-12-16 19:33:50 +0000 | [diff] [blame] | 80 | void CLGEMM::configure(const ICLTensor *a, const ICLTensor *b, const ICLTensor *c, ICLTensor *output, float alpha, float beta, const GEMMInfo &gemm_info) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 81 | { |
Georgios Pinitas | 78c0090 | 2018-01-09 17:33:11 +0000 | [diff] [blame] | 82 | ARM_COMPUTE_ERROR_ON_NULLPTR(a, b, output); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 83 | |
Georgios Pinitas | 78c0090 | 2018-01-09 17:33:11 +0000 | [diff] [blame] | 84 | // Perform validation step |
Gian Marco Iodice | 750641d | 2018-05-08 12:01:57 +0100 | [diff] [blame] | 85 | ARM_COMPUTE_ERROR_THROW_ON(validate(a->info(), b->info(), c != nullptr ? c->info() : nullptr, output->info(), alpha, beta, gemm_info)); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 86 | |
Gian Marco | 1d25ed5 | 2017-12-16 19:33:50 +0000 | [diff] [blame] | 87 | // Check if we need to reshape the matrix B only on the first run |
| 88 | _reshape_b_only_on_first_run = gemm_info.reshape_b_only_on_first_run(); |
Georgios Pinitas | e043767 | 2018-05-02 14:07:55 +0100 | [diff] [blame] | 89 | _is_prepared = false; |
Georgios Pinitas | 7221933 | 2018-06-05 14:56:06 +0100 | [diff] [blame] | 90 | _original_b = b; |
Gian Marco Iodice | 1246b63 | 2017-08-16 18:38:32 +0100 | [diff] [blame] | 91 | |
Gian Marco Iodice | edfa9f4 | 2017-08-15 11:45:22 +0100 | [diff] [blame] | 92 | const ICLTensor *matrix_a = a; |
| 93 | const ICLTensor *matrix_b = b; |
| 94 | |
Gian Marco | 36a0a46 | 2018-01-12 10:21:40 +0000 | [diff] [blame] | 95 | // Get the GPU target |
| 96 | const GPUTarget gpu_target = CLScheduler::get().target(); |
| 97 | |
| 98 | // Set the target for the kernels |
| 99 | _interleave_kernel.set_target(gpu_target); |
| 100 | _mm_kernel.set_target(gpu_target); |
| 101 | |
| 102 | // Arguments used by GEMMReshapeInfo |
| 103 | // If we pass the matrix A and matrix B reshaped to CLGEMMMatrixMultiplyKernel, we need to pass m, n, k, mult_transpose1xW_width and mult_interleave4x4_height to CLGEMMReshapeInfo |
| 104 | // in order to know how the matrices have been reshaped |
| 105 | const int m = a->info()->dimension(1); |
| 106 | const int n = b->info()->dimension(0); |
| 107 | const int k = a->info()->dimension(0); |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 108 | const int depth_output_gemm3d = gemm_info.depth_output_gemm3d(); |
Gian Marco | 36a0a46 | 2018-01-12 10:21:40 +0000 | [diff] [blame] | 109 | int mult_transpose1xW_width = 1; |
| 110 | int mult_interleave4x4_height = 1; |
| 111 | |
Gian Marco Iodice | 750641d | 2018-05-08 12:01:57 +0100 | [diff] [blame] | 112 | if(get_arch_from_target(gpu_target) == GPUTarget::BIFROST) |
Gian Marco | 36a0a46 | 2018-01-12 10:21:40 +0000 | [diff] [blame] | 113 | { |
| 114 | mult_transpose1xW_width = 4; |
| 115 | mult_interleave4x4_height = 2; |
| 116 | } |
| 117 | |
| 118 | // Check if we need to reshape the matrix A and matrix B |
| 119 | _is_interleaved_transposed = is_interleaved_transposed(m, n, k, a->info()->data_type(), _reshape_b_only_on_first_run, gpu_target); |
Gian Marco | b5311a6 | 2017-12-13 12:48:03 +0000 | [diff] [blame] | 120 | |
Gian Marco Iodice | edfa9f4 | 2017-08-15 11:45:22 +0100 | [diff] [blame] | 121 | if(_is_interleaved_transposed) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 122 | { |
Gian Marco Iodice | edfa9f4 | 2017-08-15 11:45:22 +0100 | [diff] [blame] | 123 | matrix_a = &_tmp_a; |
| 124 | matrix_b = &_tmp_b; |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 125 | |
Gian Marco | 19835e5 | 2018-01-30 13:35:54 +0000 | [diff] [blame] | 126 | // Manage intermediate buffers |
| 127 | _memory_group.manage(&_tmp_a); |
Georgios Pinitas | ae4ce7b | 2018-03-19 17:50:45 +0000 | [diff] [blame] | 128 | if(!_reshape_b_only_on_first_run) |
Georgios Pinitas | d8734b5 | 2017-12-22 15:27:52 +0000 | [diff] [blame] | 129 | { |
| 130 | _memory_group.manage(&_tmp_b); |
| 131 | } |
Gian Marco | 20d7848 | 2018-01-11 15:10:58 +0000 | [diff] [blame] | 132 | // _tmp_a and _tmp_b will be auto configured in _interleave_kernel and in _transpose_kernel |
Georgios Pinitas | 8a94e7c | 2017-09-15 19:06:47 +0100 | [diff] [blame] | 133 | |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 134 | // Configure interleave kernel |
Gian Marco | 36a0a46 | 2018-01-12 10:21:40 +0000 | [diff] [blame] | 135 | _interleave_kernel.configure(a, &_tmp_a, mult_interleave4x4_height); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 136 | |
| 137 | // Configure transpose kernel |
Gian Marco | 36a0a46 | 2018-01-12 10:21:40 +0000 | [diff] [blame] | 138 | _transpose_kernel.configure(b, &_tmp_b, mult_transpose1xW_width); |
Gian Marco Iodice | edfa9f4 | 2017-08-15 11:45:22 +0100 | [diff] [blame] | 139 | } |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 140 | |
Georgios Pinitas | 17812ba | 2018-06-04 19:27:13 +0100 | [diff] [blame] | 141 | // Configure and tune matrix multiply kernel |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 142 | _mm_kernel.configure(matrix_a, matrix_b, output, alpha, _is_interleaved_transposed, GEMMReshapeInfo(m, n, k, mult_transpose1xW_width, mult_interleave4x4_height, depth_output_gemm3d)); |
Georgios Pinitas | 17812ba | 2018-06-04 19:27:13 +0100 | [diff] [blame] | 143 | CLScheduler::get().tune_kernel_static(_mm_kernel); |
Gian Marco Iodice | edfa9f4 | 2017-08-15 11:45:22 +0100 | [diff] [blame] | 144 | |
| 145 | if(_is_interleaved_transposed) |
| 146 | { |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 147 | // Allocate intermediate tensors |
| 148 | _tmp_a.allocator()->allocate(); |
Georgios Pinitas | e043767 | 2018-05-02 14:07:55 +0100 | [diff] [blame] | 149 | if(!_reshape_b_only_on_first_run) |
| 150 | { |
| 151 | _tmp_b.allocator()->allocate(); |
| 152 | } |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 153 | } |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 154 | |
| 155 | // Configure matrix addition kernel |
| 156 | if(beta != 0 && c != nullptr) |
| 157 | { |
| 158 | _ma_kernel.configure(c, output, beta); |
| 159 | _run_addition = true; |
| 160 | } |
| 161 | } |
| 162 | |
Gian Marco Iodice | 750641d | 2018-05-08 12:01:57 +0100 | [diff] [blame] | 163 | Status CLGEMM::validate(const ITensorInfo *a, const ITensorInfo *b, const ITensorInfo *c, const ITensorInfo *output, float alpha, float beta, const GEMMInfo &gemm_info) |
Georgios Pinitas | 78c0090 | 2018-01-09 17:33:11 +0000 | [diff] [blame] | 164 | { |
Gian Marco Iodice | 750641d | 2018-05-08 12:01:57 +0100 | [diff] [blame] | 165 | ARM_COMPUTE_UNUSED(alpha); |
| 166 | |
| 167 | // Check if we need to reshape the matrix B only on the first run |
| 168 | const bool reshape_b_only_on_first_run = gemm_info.reshape_b_only_on_first_run(); |
| 169 | |
| 170 | const ITensorInfo *matrix_a_info = a; |
| 171 | const ITensorInfo *matrix_b_info = b; |
| 172 | |
| 173 | TensorInfo tmp_a_info{}; |
| 174 | TensorInfo tmp_b_info{}; |
| 175 | TensorInfo tmp_output_info = *output->clone(); |
| 176 | |
| 177 | // Get the GPU target |
| 178 | const GPUTarget gpu_target = CLScheduler::get().target(); |
| 179 | |
| 180 | // Arguments used by GEMMReshapeInfo |
| 181 | // If we pass the matrix A and matrix B reshaped to CLGEMMMatrixMultiplyKernel, we need to pass m, n, k, mult_transpose1xW_width and mult_interleave4x4_height to CLGEMMReshapeInfo |
| 182 | // in order to know how the matrices have been reshaped |
| 183 | const int m = a->dimension(1); |
| 184 | const int n = b->dimension(0); |
| 185 | const int k = a->dimension(0); |
| 186 | int mult_transpose1xW_width = 1; |
| 187 | int mult_interleave4x4_height = 1; |
| 188 | |
| 189 | if(get_arch_from_target(gpu_target) == GPUTarget::BIFROST) |
| 190 | { |
| 191 | mult_transpose1xW_width = 4; |
| 192 | mult_interleave4x4_height = 2; |
| 193 | } |
| 194 | |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 195 | const GEMMReshapeInfo reshape_info = GEMMReshapeInfo(m, n, k, mult_transpose1xW_width, mult_interleave4x4_height, gemm_info.depth_output_gemm3d()); |
Gian Marco Iodice | 750641d | 2018-05-08 12:01:57 +0100 | [diff] [blame] | 196 | |
| 197 | // Check if we need to reshape the matrix A and matrix B |
| 198 | const bool run_interleave_transpose = is_interleaved_transposed(m, n, k, a->data_type(), reshape_b_only_on_first_run, gpu_target); |
| 199 | |
| 200 | if(run_interleave_transpose) |
| 201 | { |
| 202 | matrix_a_info = &tmp_a_info; |
| 203 | matrix_b_info = &tmp_b_info; |
| 204 | |
| 205 | // Validate interleave kernel |
| 206 | auto_init_if_empty(tmp_a_info, a->clone()->set_tensor_shape(compute_interleaved_shape(*a, mult_interleave4x4_height))); |
| 207 | ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMInterleave4x4Kernel::validate(a, &tmp_a_info, mult_interleave4x4_height)); |
| 208 | |
| 209 | // Validate transpose kernel |
| 210 | auto_init_if_empty(tmp_b_info, b->clone()->set_tensor_shape(compute_transpose1xW_with_element_size_shape(*b, mult_transpose1xW_width))); |
| 211 | ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMTranspose1xWKernel::validate(b, &tmp_b_info, mult_transpose1xW_width)); |
| 212 | } |
| 213 | |
| 214 | // Validate matrix multiply |
| 215 | auto_init_if_empty(tmp_output_info, matrix_a_info->clone()->set_tensor_shape(compute_mm_shape(*matrix_a_info, *matrix_b_info, run_interleave_transpose, reshape_info))); |
| 216 | ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMMatrixMultiplyKernel::validate(matrix_a_info, matrix_b_info, &tmp_output_info, alpha, run_interleave_transpose, reshape_info, gpu_target)); |
| 217 | |
| 218 | if(beta != 0 && c != nullptr) |
| 219 | { |
| 220 | // Validate matrix addition kernel |
| 221 | ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMMatrixAdditionKernel::validate(c, &tmp_output_info, beta)); |
| 222 | } |
| 223 | |
Georgios Pinitas | 78c0090 | 2018-01-09 17:33:11 +0000 | [diff] [blame] | 224 | return Status{}; |
| 225 | } |
| 226 | |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 227 | void CLGEMM::run() |
| 228 | { |
Georgios Pinitas | e043767 | 2018-05-02 14:07:55 +0100 | [diff] [blame] | 229 | prepare(); |
| 230 | |
Georgios Pinitas | 8a94e7c | 2017-09-15 19:06:47 +0100 | [diff] [blame] | 231 | _memory_group.acquire(); |
| 232 | |
Gian Marco Iodice | edfa9f4 | 2017-08-15 11:45:22 +0100 | [diff] [blame] | 233 | if(_is_interleaved_transposed) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 234 | { |
| 235 | // Run interleave kernel |
| 236 | CLScheduler::get().enqueue(_interleave_kernel, false); |
| 237 | |
Georgios Pinitas | e043767 | 2018-05-02 14:07:55 +0100 | [diff] [blame] | 238 | if(!_reshape_b_only_on_first_run) |
Gian Marco | 1d25ed5 | 2017-12-16 19:33:50 +0000 | [diff] [blame] | 239 | { |
| 240 | // Run transpose kernel |
| 241 | CLScheduler::get().enqueue(_transpose_kernel, false); |
| 242 | } |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 243 | } |
| 244 | |
| 245 | // Run matrix multiply kernel |
| 246 | CLScheduler::get().enqueue(_mm_kernel, !_run_addition); |
| 247 | |
| 248 | // Run matrix addition kernel |
| 249 | if(_run_addition) |
| 250 | { |
| 251 | CLScheduler::get().enqueue(_ma_kernel); |
| 252 | } |
Georgios Pinitas | 8a94e7c | 2017-09-15 19:06:47 +0100 | [diff] [blame] | 253 | |
| 254 | _memory_group.release(); |
Georgios Pinitas | e043767 | 2018-05-02 14:07:55 +0100 | [diff] [blame] | 255 | } |
Georgios Pinitas | 82b5148 | 2018-04-24 15:14:12 +0100 | [diff] [blame] | 256 | |
Georgios Pinitas | e043767 | 2018-05-02 14:07:55 +0100 | [diff] [blame] | 257 | void CLGEMM::prepare() |
| 258 | { |
| 259 | if(!_is_prepared) |
| 260 | { |
| 261 | if(_is_interleaved_transposed && _reshape_b_only_on_first_run) |
| 262 | { |
Georgios Pinitas | 7221933 | 2018-06-05 14:56:06 +0100 | [diff] [blame] | 263 | // Run transpose kernel and mark original weights tensor as unused |
Georgios Pinitas | e043767 | 2018-05-02 14:07:55 +0100 | [diff] [blame] | 264 | _tmp_b.allocator()->allocate(); |
| 265 | CLScheduler::get().enqueue(_transpose_kernel, false); |
| 266 | _original_b->mark_as_unused(); |
| 267 | } |
| 268 | CLScheduler::get().queue().finish(); |
| 269 | _is_prepared = true; |
| 270 | } |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 271 | } |