giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 1 | /* |
Giorgio Arena | 9f7d55a | 2021-02-08 13:20:24 +0000 | [diff] [blame] | 2 | * Copyright (c) 2019-2021 Arm Limited. |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +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 | */ |
Sang-Hoon Park | bef7fa2 | 2020-10-21 15:58:54 +0100 | [diff] [blame] | 24 | #include "src/core/CL/kernels/CLGEMMMatrixMultiplyNativeKernel.h" |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 25 | |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 26 | #include "arm_compute/core/CL/CLHelpers.h" |
| 27 | #include "arm_compute/core/CL/CLKernelLibrary.h" |
| 28 | #include "arm_compute/core/CL/ICLTensor.h" |
| 29 | #include "arm_compute/core/CL/OpenCL.h" |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 30 | #include "arm_compute/core/Helpers.h" |
| 31 | #include "arm_compute/core/TensorInfo.h" |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 32 | #include "arm_compute/core/Utils.h" |
| 33 | #include "arm_compute/core/Validate.h" |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 34 | #include "arm_compute/core/utils/misc/ShapeCalculator.h" |
Sang-Hoon Park | 68dd25f | 2020-10-19 16:00:11 +0100 | [diff] [blame] | 35 | #include "src/core/AccessWindowStatic.h" |
| 36 | #include "src/core/helpers/AutoConfiguration.h" |
| 37 | #include "src/core/helpers/WindowHelpers.h" |
| 38 | #include "src/core/utils/helpers/float_ops.h" |
Matthew Bentham | 758b5ba | 2020-03-05 23:37:48 +0000 | [diff] [blame] | 39 | #include "support/StringSupport.h" |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 40 | |
| 41 | #include <cstddef> |
| 42 | #include <cstdint> |
| 43 | #include <tuple> |
| 44 | |
| 45 | using namespace arm_compute::misc::shape_calculator; |
| 46 | |
| 47 | namespace arm_compute |
| 48 | { |
| 49 | namespace |
| 50 | { |
| 51 | using ElementsProcessed = Steps; |
| 52 | |
Gian Marco Iodice | 944170e | 2019-06-24 14:40:30 +0100 | [diff] [blame] | 53 | Status validate_arguments(const ITensorInfo *input0, const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output, float alpha, float beta, const GEMMLHSMatrixInfo &lhs_info, |
| 54 | const GEMMRHSMatrixInfo &rhs_info, |
Gian Marco Iodice | 7026b30 | 2019-06-26 17:18:11 +0100 | [diff] [blame] | 55 | const GEMMKernelInfo &gemm_info) |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 56 | { |
| 57 | ARM_COMPUTE_UNUSED(alpha); |
| 58 | ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input0, input1, output); |
Gian Marco Iodice | d820db6 | 2019-08-05 14:23:23 +0100 | [diff] [blame] | 59 | ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input0, 1, DataType::F32); |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 60 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input0, input1); |
| 61 | ARM_COMPUTE_RETURN_ERROR_ON_MSG(input0->num_dimensions() > 4, "The number of dimensions for the LHS matrix must be <= 4"); |
| 62 | ARM_COMPUTE_RETURN_ERROR_ON_MSG(input1->num_dimensions() > 3, "The number of dimensions for the RHS matrix must be <= 3"); |
| 63 | ARM_COMPUTE_RETURN_ERROR_ON_MSG(((rhs_info.k0 & (rhs_info.k0 - 1)) && rhs_info.k0 != 3), "Only 2,3,4,8,16 are supported for k0"); |
| 64 | ARM_COMPUTE_RETURN_ERROR_ON(rhs_info.k0 > 16); |
| 65 | ARM_COMPUTE_RETURN_ERROR_ON(lhs_info.m0 < 1 || lhs_info.m0 > 8); |
| 66 | ARM_COMPUTE_RETURN_ERROR_ON_MSG(((rhs_info.n0 & (rhs_info.n0 - 1)) && rhs_info.n0 != 3), "Only 2,3,4,8,16 are supported for n0"); |
Gian Marco Iodice | b238f5f | 2019-08-02 09:09:53 +0100 | [diff] [blame] | 67 | ARM_COMPUTE_RETURN_ERROR_ON_MSG((gemm_info.reinterpret_input_as_3d || gemm_info.depth_output_gemm3d != 0) && (input2 != nullptr) |
Gian Marco Iodice | d820db6 | 2019-08-05 14:23:23 +0100 | [diff] [blame] | 68 | && (!gemm_info.broadcast_bias), |
| 69 | "Bias addition only supported with broadcast mode in case the input or output has to be reinterpreted as 3D"); |
Gian Marco Iodice | 0c17aa2 | 2019-09-27 09:23:15 +0100 | [diff] [blame] | 70 | ARM_COMPUTE_RETURN_ERROR_ON_MSG(gemm_info.fp_mixed_precision, "Mixed precision not supported"); |
Gian Marco Iodice | dd717c3 | 2020-05-28 10:22:03 +0100 | [diff] [blame] | 71 | ARM_COMPUTE_RETURN_ERROR_ON_MSG(rhs_info.export_to_cl_image, "Export to CLImage not supported for GEMM native"); |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 72 | |
Gian Marco Iodice | 7026b30 | 2019-06-26 17:18:11 +0100 | [diff] [blame] | 73 | const unsigned int m = gemm_info.m; |
| 74 | const unsigned int n = gemm_info.n; |
| 75 | const unsigned int k = gemm_info.k; |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 76 | |
| 77 | ARM_COMPUTE_UNUSED(m); |
| 78 | ARM_COMPUTE_UNUSED(n); |
| 79 | ARM_COMPUTE_UNUSED(k); |
| 80 | |
Gian Marco Iodice | 7026b30 | 2019-06-26 17:18:11 +0100 | [diff] [blame] | 81 | ARM_COMPUTE_RETURN_ERROR_ON(input0->dimension(0) != k); |
| 82 | ARM_COMPUTE_RETURN_ERROR_ON(input1->dimension(0) != n); |
| 83 | ARM_COMPUTE_RETURN_ERROR_ON(input1->dimension(1) != k); |
| 84 | if(gemm_info.reinterpret_input_as_3d) |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 85 | { |
Gian Marco Iodice | 7026b30 | 2019-06-26 17:18:11 +0100 | [diff] [blame] | 86 | ARM_COMPUTE_RETURN_ERROR_ON(input0->dimension(1) * input0->dimension(2) != m); |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 87 | } |
| 88 | else |
| 89 | { |
Gian Marco Iodice | 7026b30 | 2019-06-26 17:18:11 +0100 | [diff] [blame] | 90 | ARM_COMPUTE_RETURN_ERROR_ON(input0->dimension(1) != m); |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 91 | } |
| 92 | |
Gian Marco Iodice | 944170e | 2019-06-24 14:40:30 +0100 | [diff] [blame] | 93 | if(input2 != nullptr && !(helpers::float_ops::is_zero(beta))) |
| 94 | { |
Gian Marco Iodice | 7026b30 | 2019-06-26 17:18:11 +0100 | [diff] [blame] | 95 | const unsigned int input2_dim0 = input2->dimension(0); |
| 96 | const unsigned int input2_dim1 = input2->dimension(1); |
Gian Marco Iodice | 944170e | 2019-06-24 14:40:30 +0100 | [diff] [blame] | 97 | |
| 98 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input2, input1); |
Gian Marco Iodice | 7026b30 | 2019-06-26 17:18:11 +0100 | [diff] [blame] | 99 | if(gemm_info.broadcast_bias) |
Gian Marco Iodice | 944170e | 2019-06-24 14:40:30 +0100 | [diff] [blame] | 100 | { |
| 101 | ARM_COMPUTE_RETURN_ERROR_ON_MSG((input2_dim1 != 1 || input2_dim0 != n), "Incorrect dimension of bias matrix which is to be broadcasted"); |
| 102 | } |
| 103 | else |
| 104 | { |
| 105 | ARM_COMPUTE_RETURN_ERROR_ON_MSG((input2_dim0 != n || input2_dim1 != m), "Incorrect dimension of bias matrix"); |
| 106 | } |
| 107 | } |
| 108 | |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 109 | if(output->total_size() != 0) |
| 110 | { |
| 111 | const TensorInfo tensor_info_output = output->clone()->set_tensor_shape(compute_mm_shape(*input0, *input1, gemm_info)); |
| 112 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(output, &tensor_info_output); |
| 113 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input0, output); |
| 114 | } |
| 115 | |
| 116 | return Status{}; |
| 117 | } |
| 118 | |
Gian Marco Iodice | 944170e | 2019-06-24 14:40:30 +0100 | [diff] [blame] | 119 | std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input0, ITensorInfo *input1, ITensorInfo *input2, ITensorInfo *output, const GEMMLHSMatrixInfo &lhs_info, |
| 120 | const GEMMRHSMatrixInfo &rhs_info, |
Gian Marco Iodice | 7026b30 | 2019-06-26 17:18:11 +0100 | [diff] [blame] | 121 | const GEMMKernelInfo &gemm_info, ElementsProcessed &num_elements_processed) |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 122 | { |
| 123 | unsigned int &num_elems_processed_per_iteration_x = num_elements_processed[0]; |
| 124 | unsigned int &num_elems_processed_per_iteration_y = num_elements_processed[1]; |
Gian Marco Iodice | 7026b30 | 2019-06-26 17:18:11 +0100 | [diff] [blame] | 125 | bool reinterpret_input_as_3d = gemm_info.reinterpret_input_as_3d; |
| 126 | bool reinterpret_output_as_3d = gemm_info.depth_output_gemm3d != 0; |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 127 | |
| 128 | Window win{}; |
| 129 | Window win_out{}; |
| 130 | bool window_changed = false; |
| 131 | |
| 132 | // In case both input and output have to be reinterpreted as 3D tensors, |
| 133 | // force reinterpret_input_as_3d and reinterpret_output_as_3d to be false. |
| 134 | if(reinterpret_input_as_3d == reinterpret_output_as_3d) |
| 135 | { |
| 136 | reinterpret_output_as_3d = false; |
| 137 | } |
| 138 | |
| 139 | // Output tensor auto initialization if not yet initialized |
| 140 | auto_init_if_empty(*output, input0->clone()->set_tensor_shape(compute_mm_shape(*input0, *input1, gemm_info))); |
| 141 | |
| 142 | TensorInfo tmp_info(*output); |
| 143 | |
| 144 | if(reinterpret_output_as_3d) |
| 145 | { |
| 146 | // Since the output tensor has to be reinterpreted as 3D and the execute window is based on a 2D GEMM, |
| 147 | // the window needs to be constructed on the 2D collapsed version of the tensor |
| 148 | TensorShape tmp_shape(output->tensor_shape()); |
| 149 | tmp_shape.collapse(2U, 1U); |
| 150 | tmp_info.set_tensor_shape(tmp_shape); |
| 151 | } |
| 152 | |
| 153 | // Configure kernel window |
| 154 | num_elems_processed_per_iteration_x = rhs_info.n0; |
| 155 | num_elems_processed_per_iteration_y = lhs_info.m0; |
| 156 | |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 157 | win = calculate_max_window(tmp_info, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y)); |
| 158 | win_out = calculate_max_window(*output, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y)); |
| 159 | |
| 160 | AccessWindowStatic input0_access(input0, 0, 0, |
| 161 | input0->dimension(0), |
SiCong Li | 406a13f | 2020-07-15 12:09:58 +0100 | [diff] [blame] | 162 | input0->dimension(1)); |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 163 | AccessWindowStatic input1_access(input1, 0, 0, |
| 164 | ceil_to_multiple(input1->dimension(0), num_elems_processed_per_iteration_x), |
| 165 | input1->dimension(1)); |
| 166 | AccessWindowStatic output_access(output, 0, 0, |
SiCong Li | 3a50166 | 2020-06-26 10:02:06 +0100 | [diff] [blame] | 167 | output->dimension(0), |
| 168 | output->dimension(1)); |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 169 | |
Gian Marco Iodice | 944170e | 2019-06-24 14:40:30 +0100 | [diff] [blame] | 170 | if(input2 != nullptr) |
| 171 | { |
| 172 | const int bias_processed_per_iteration_x = num_elems_processed_per_iteration_x; |
| 173 | |
Gian Marco Iodice | 944170e | 2019-06-24 14:40:30 +0100 | [diff] [blame] | 174 | AccessWindowStatic input2_access(input2, 0, 0, |
| 175 | ceil_to_multiple(input2->dimension(0), bias_processed_per_iteration_x), |
SiCong Li | 406a13f | 2020-07-15 12:09:58 +0100 | [diff] [blame] | 176 | input2->dimension(1)); |
Gian Marco Iodice | 944170e | 2019-06-24 14:40:30 +0100 | [diff] [blame] | 177 | |
| 178 | window_changed = update_window_and_padding(win, input0_access, input1_access, input2_access) || // window used by the execute_window_loop |
| 179 | update_window_and_padding(win_out, output_access); // window used to update the padding requirements of output tensor |
| 180 | } |
| 181 | else |
| 182 | { |
| 183 | window_changed = update_window_and_padding(win, input0_access, input1_access) || // window used by the execute_window_loop |
| 184 | update_window_and_padding(win_out, output_access); // window used to update the padding requirements of output tensor |
| 185 | } |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 186 | |
| 187 | output_access.set_valid_region(win_out, ValidRegion(Coordinates(), output->tensor_shape())); |
| 188 | |
| 189 | // Collapse along the Z direction |
| 190 | // This collapse needs to be here in order to tune the Z dimension of LWS |
| 191 | Window collapsed = win; |
| 192 | const unsigned int dimension_to_collapse = std::min(static_cast<unsigned int>(output->num_dimensions()), 2u); |
| 193 | collapsed = win.collapse(win, dimension_to_collapse); |
| 194 | |
| 195 | Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{}; |
| 196 | return std::make_pair(err, collapsed); |
| 197 | } |
| 198 | } // namespace |
| 199 | |
| 200 | CLGEMMMatrixMultiplyNativeKernel::CLGEMMMatrixMultiplyNativeKernel() |
Gian Marco Iodice | 944170e | 2019-06-24 14:40:30 +0100 | [diff] [blame] | 201 | : _input0(nullptr), _input1(nullptr), _input2(nullptr), _output(nullptr), _slide_matrix_b(true), _reinterpret_input_as_3d(false), _reinterpret_output_as_3d(false), _use_dummy_work_items(false), |
| 202 | _add_bias(false), _broadcast_bias(false) |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 203 | { |
| 204 | } |
| 205 | |
Gian Marco Iodice | 944170e | 2019-06-24 14:40:30 +0100 | [diff] [blame] | 206 | void CLGEMMMatrixMultiplyNativeKernel::configure(const ICLTensor *input0, const ICLTensor *input1, const ICLTensor *input2, ICLTensor *output, float alpha, float beta, |
| 207 | const GEMMLHSMatrixInfo &lhs_info, |
Gian Marco Iodice | 7026b30 | 2019-06-26 17:18:11 +0100 | [diff] [blame] | 208 | const GEMMRHSMatrixInfo &rhs_info, const GEMMKernelInfo &gemm_info) |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 209 | { |
Manuel Bottini | 4c6bd51 | 2020-04-08 10:15:51 +0100 | [diff] [blame] | 210 | configure(CLKernelLibrary::get().get_compile_context(), input0, input1, input2, output, alpha, beta, lhs_info, rhs_info, gemm_info); |
| 211 | } |
| 212 | |
Manuel Bottini | 679fc96 | 2020-04-21 16:08:53 +0100 | [diff] [blame] | 213 | void CLGEMMMatrixMultiplyNativeKernel::configure(const CLCompileContext &compile_context, const ICLTensor *input0, const ICLTensor *input1, const ICLTensor *input2, ICLTensor *output, float alpha, |
Manuel Bottini | 4c6bd51 | 2020-04-08 10:15:51 +0100 | [diff] [blame] | 214 | float beta, |
| 215 | const GEMMLHSMatrixInfo &lhs_info, |
| 216 | const GEMMRHSMatrixInfo &rhs_info, const GEMMKernelInfo &gemm_info) |
| 217 | { |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 218 | ARM_COMPUTE_ERROR_ON_NULLPTR(input0, input1, output); |
| 219 | |
Gian Marco Iodice | 944170e | 2019-06-24 14:40:30 +0100 | [diff] [blame] | 220 | ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input0->info(), input1->info(), (input2 != nullptr ? input2->info() : nullptr), output->info(), alpha, beta, lhs_info, rhs_info, gemm_info)); |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 221 | |
SiCong Li | 04a0706 | 2020-11-17 14:09:01 +0000 | [diff] [blame] | 222 | auto padding_info = get_padding_info({ input0, output }); |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 223 | _input0 = input0; |
| 224 | _input1 = input1; |
Gian Marco Iodice | 944170e | 2019-06-24 14:40:30 +0100 | [diff] [blame] | 225 | _input2 = helpers::float_ops::is_zero(beta) ? nullptr : input2; |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 226 | _output = output; |
Gian Marco Iodice | 7026b30 | 2019-06-26 17:18:11 +0100 | [diff] [blame] | 227 | _reinterpret_input_as_3d = gemm_info.reinterpret_input_as_3d; |
| 228 | _reinterpret_output_as_3d = gemm_info.depth_output_gemm3d != 0; |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 229 | _use_dummy_work_items = preferred_dummy_work_items_support(CLKernelLibrary::get().get_device()); |
Gian Marco Iodice | 944170e | 2019-06-24 14:40:30 +0100 | [diff] [blame] | 230 | _add_bias = _input2 != nullptr; |
Gian Marco Iodice | 7026b30 | 2019-06-26 17:18:11 +0100 | [diff] [blame] | 231 | _broadcast_bias = gemm_info.broadcast_bias; |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 232 | |
| 233 | // In case both input and output have to be reinterpreted as 3D tensors, |
| 234 | // force reinterpret_input_as_3d and reinterpret_output_as_3d to be false. |
| 235 | if(_reinterpret_input_as_3d == _reinterpret_output_as_3d) |
| 236 | { |
| 237 | _reinterpret_input_as_3d = false; |
| 238 | _reinterpret_output_as_3d = false; |
| 239 | } |
| 240 | |
| 241 | // Check if we need to slide the matrix B |
| 242 | const unsigned int num_dimensions_input0 = _input0->info()->num_dimensions(); |
| 243 | _slide_matrix_b = (_input1->info()->num_dimensions() >= num_dimensions_input0); |
| 244 | |
| 245 | ElementsProcessed num_elements_processed{}; |
| 246 | |
| 247 | // Configure kernel window |
Gian Marco Iodice | 944170e | 2019-06-24 14:40:30 +0100 | [diff] [blame] | 248 | auto win_config = validate_and_configure_window(input0->info(), input1->info(), input2 != nullptr ? input2->info() : nullptr, output->info(), lhs_info, rhs_info, gemm_info, num_elements_processed); |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 249 | ARM_COMPUTE_ERROR_THROW_ON(win_config.first); |
| 250 | ICLKernel::configure_internal(win_config.second); |
| 251 | |
Gian Marco Iodice | 7b9d7ca | 2019-09-19 16:37:39 +0100 | [diff] [blame] | 252 | // If _reinterpret_input_as_3d = _reinterpret_output_as_3d = true, |
| 253 | // we will dispatch a batched-GEMM to reduce the complexity of the address calculation within the OpenCL kernel. |
| 254 | // This means that the actual m used by the kernel is given by output->info()->dimension(1) and not by gemm_info.m |
| 255 | const unsigned int internal_m = _reinterpret_output_as_3d ? gemm_info.m : output->info()->dimension(1); |
| 256 | |
| 257 | const unsigned int h_gemm_3d = _reinterpret_output_as_3d ? output->info()->dimension(1) : input0->info()->dimension(1); |
| 258 | const unsigned int d_gemm_3d = _reinterpret_output_as_3d ? output->info()->dimension(2) : input0->info()->dimension(2); |
| 259 | |
SiCong Li | 3a50166 | 2020-06-26 10:02:06 +0100 | [diff] [blame] | 260 | // Calculate partial (store instead of load) M0 and partial N0 for the partial blocks at the end of a row/column if any. This is to avoid padding. |
| 261 | const unsigned int partial_store_m0 = internal_m % lhs_info.m0; |
| 262 | const unsigned int partial_store_n0 = gemm_info.n % rhs_info.n0; |
| 263 | |
SiCong Li | b972ae6 | 2020-08-03 15:39:45 +0100 | [diff] [blame] | 264 | // Shrink M0 to be always <= M (internal_m) to prevent out-of-bounds reads. |
| 265 | // NOTE: This might have implications on heuristics and performance |
| 266 | const unsigned int internal_m0 = std::min(internal_m, lhs_info.m0); |
| 267 | |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 268 | // Create build options |
| 269 | CLBuildOptions build_opts; |
| 270 | build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(input0->info()->data_type())); |
Gian Marco Iodice | 82d9dd1 | 2019-06-10 16:45:40 +0100 | [diff] [blame] | 271 | build_opts.add_option_if(!(helpers::float_ops::is_one(alpha)), "-DALPHA=" + float_to_string_with_full_precision(alpha)); |
Gian Marco Iodice | 944170e | 2019-06-24 14:40:30 +0100 | [diff] [blame] | 272 | build_opts.add_option_if(_input2 != nullptr, "-DBETA=" + float_to_string_with_full_precision(beta)); |
| 273 | build_opts.add_option_if(helpers::float_ops::is_one(beta), "-DUNIT_BETA"); |
Gian Marco Iodice | 7026b30 | 2019-06-26 17:18:11 +0100 | [diff] [blame] | 274 | build_opts.add_option_if(gemm_info.broadcast_bias, "-DBROADCAST_BIAS"); |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 275 | build_opts.add_option_if(_reinterpret_input_as_3d, "-DREINTERPRET_INPUT_AS_3D"); |
| 276 | build_opts.add_option_if(_reinterpret_output_as_3d, "-DREINTERPRET_OUTPUT_AS_3D"); |
Gian Marco Iodice | 7b9d7ca | 2019-09-19 16:37:39 +0100 | [diff] [blame] | 277 | build_opts.add_option_if(_reinterpret_input_as_3d || _reinterpret_output_as_3d, "-DHEIGHT_GEMM3D=" + support::cpp11::to_string(h_gemm_3d)); |
| 278 | build_opts.add_option_if(_reinterpret_input_as_3d || _reinterpret_output_as_3d, "-DDEPTH_GEMM3D=" + support::cpp11::to_string(d_gemm_3d)); |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 279 | build_opts.add_option_if(!_slide_matrix_b, "-DMATRIX_B_DEPTH=" + support::cpp11::to_string(input1->info()->dimension(2))); |
| 280 | build_opts.add_option_if(_use_dummy_work_items, "-DDUMMY_WORK_ITEMS"); |
Gian Marco Iodice | 7b9d7ca | 2019-09-19 16:37:39 +0100 | [diff] [blame] | 281 | build_opts.add_option("-DM=" + support::cpp11::to_string(internal_m)); |
Gian Marco Iodice | 7026b30 | 2019-06-26 17:18:11 +0100 | [diff] [blame] | 282 | build_opts.add_option("-DN=" + support::cpp11::to_string(gemm_info.n)); |
| 283 | build_opts.add_option("-DK=" + support::cpp11::to_string(gemm_info.k)); |
SiCong Li | b972ae6 | 2020-08-03 15:39:45 +0100 | [diff] [blame] | 284 | build_opts.add_option("-DM0=" + support::cpp11::to_string(internal_m0)); |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 285 | build_opts.add_option("-DN0=" + support::cpp11::to_string(rhs_info.n0)); |
| 286 | build_opts.add_option("-DK0=" + support::cpp11::to_string(rhs_info.k0)); |
SiCong Li | 3a50166 | 2020-06-26 10:02:06 +0100 | [diff] [blame] | 287 | build_opts.add_option("-DPARTIAL_STORE_M0=" + support::cpp11::to_string(partial_store_m0)); |
| 288 | build_opts.add_option("-DPARTIAL_STORE_N0=" + support::cpp11::to_string(partial_store_n0)); |
Gian Marco Iodice | ca1f460 | 2019-07-16 15:46:48 +0100 | [diff] [blame] | 289 | build_opts.add_option_if(gemm_info.activation_info.enabled(), "-DACTIVATION_TYPE=" + lower_string(string_from_activation_func(gemm_info.activation_info.activation()))); |
| 290 | build_opts.add_option_if(gemm_info.activation_info.enabled(), "-DA_VAL=" + float_to_string_with_full_precision(gemm_info.activation_info.a())); |
| 291 | build_opts.add_option_if(gemm_info.activation_info.enabled(), "-DB_VAL=" + float_to_string_with_full_precision(gemm_info.activation_info.b())); |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 292 | |
| 293 | std::string kernel_name("gemm_mm_native"); |
| 294 | |
| 295 | // Create kernel |
Manuel Bottini | 4c6bd51 | 2020-04-08 10:15:51 +0100 | [diff] [blame] | 296 | _kernel = create_kernel(compile_context, kernel_name, build_opts.options()); |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 297 | |
| 298 | // Set config_id for enabling LWS tuning |
| 299 | _config_id = kernel_name; |
| 300 | _config_id += "_"; |
Gian Marco Iodice | 944170e | 2019-06-24 14:40:30 +0100 | [diff] [blame] | 301 | _config_id += (_add_bias ? "add_bias_" : ""); |
| 302 | _config_id += (_broadcast_bias ? "broadcast_bias_" : ""); |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 303 | _config_id += (_reinterpret_input_as_3d ? "3di_" : ""); |
| 304 | _config_id += (_reinterpret_output_as_3d ? "3do_" : ""); |
Gian Marco Iodice | ca1f460 | 2019-07-16 15:46:48 +0100 | [diff] [blame] | 305 | _config_id += (gemm_info.activation_info.enabled() ? "fused_activation_" : ""); |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 306 | _config_id += lower_string(string_from_data_type(input0->info()->data_type())); |
| 307 | _config_id += "_"; |
| 308 | _config_id += support::cpp11::to_string(output->info()->dimension(1)); |
| 309 | _config_id += "_"; |
| 310 | _config_id += support::cpp11::to_string(output->info()->dimension(0)); |
| 311 | _config_id += "_"; |
Gian Marco Iodice | 7026b30 | 2019-06-26 17:18:11 +0100 | [diff] [blame] | 312 | _config_id += support::cpp11::to_string(gemm_info.k); |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 313 | _config_id += "_"; |
| 314 | _config_id += support::cpp11::to_string(output->info()->dimension(2)); |
| 315 | _config_id += "_"; |
| 316 | _config_id += support::cpp11::to_string(lhs_info.m0); |
| 317 | _config_id += "_"; |
| 318 | _config_id += support::cpp11::to_string(rhs_info.n0); |
| 319 | _config_id += "_"; |
| 320 | _config_id += support::cpp11::to_string(rhs_info.k0); |
SiCong Li | 04a0706 | 2020-11-17 14:09:01 +0000 | [diff] [blame] | 321 | |
| 322 | ARM_COMPUTE_ERROR_ON(has_padding_changed(padding_info)); |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 323 | } |
| 324 | |
Gian Marco Iodice | 944170e | 2019-06-24 14:40:30 +0100 | [diff] [blame] | 325 | Status CLGEMMMatrixMultiplyNativeKernel::validate(const ITensorInfo *input0, const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output, float alpha, float beta, |
| 326 | const GEMMLHSMatrixInfo &lhs_info, |
Gian Marco Iodice | 7026b30 | 2019-06-26 17:18:11 +0100 | [diff] [blame] | 327 | const GEMMRHSMatrixInfo &rhs_info, const GEMMKernelInfo &gemm_info) |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 328 | { |
| 329 | ElementsProcessed num_elements_processed{}; |
Gian Marco Iodice | 944170e | 2019-06-24 14:40:30 +0100 | [diff] [blame] | 330 | ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input0, input1, input2, output, alpha, beta, lhs_info, rhs_info, gemm_info)); |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 331 | ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input0->clone().get(), |
| 332 | input1->clone().get(), |
Gian Marco Iodice | 944170e | 2019-06-24 14:40:30 +0100 | [diff] [blame] | 333 | input2 != nullptr ? input2->clone().get() : nullptr, |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 334 | output->clone().get(), |
| 335 | lhs_info, |
| 336 | rhs_info, |
| 337 | gemm_info, |
| 338 | num_elements_processed) |
| 339 | .first); |
| 340 | |
| 341 | return Status{}; |
| 342 | } |
| 343 | |
| 344 | void CLGEMMMatrixMultiplyNativeKernel::run(const Window &window, cl::CommandQueue &queue) |
| 345 | { |
| 346 | ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); |
| 347 | ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window); |
| 348 | |
| 349 | if(_input1->info()->num_dimensions() < 3) |
| 350 | { |
Giorgio Arena | edc524e | 2021-02-10 11:54:47 +0000 | [diff] [blame] | 351 | // The stride_z for matrix B must be zero if we do not slice |
| 352 | ARM_COMPUTE_ERROR_ON(_input1->info()->strides_in_bytes()[3] != 0); |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 353 | } |
| 354 | |
| 355 | Window slice = window.first_slice_window_3D(); |
| 356 | Window slice_matrix_b = slice; |
| 357 | |
| 358 | slice_matrix_b.set(Window::DimX, Window::Dimension(0, 1, 1)); |
| 359 | slice_matrix_b.set(Window::DimY, Window::Dimension(0, 1, 1)); |
| 360 | |
| 361 | if(_reinterpret_input_as_3d) |
| 362 | { |
| 363 | // Pass bottom paddings to the kernel if the input has to be reinterpreted as 3D tensor |
Gian Marco Iodice | 944170e | 2019-06-24 14:40:30 +0100 | [diff] [blame] | 364 | unsigned int idx0; |
| 365 | if(_add_bias) |
| 366 | { |
| 367 | idx0 = 4 * num_arguments_per_2D_tensor() + 4; |
| 368 | } |
| 369 | else |
| 370 | { |
| 371 | idx0 = 3 * num_arguments_per_2D_tensor() + 3; |
| 372 | } |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 373 | const unsigned int total_cross_plane_pad = _input0->info()->padding().top + _input0->info()->padding().bottom; |
| 374 | _kernel.setArg<cl_uint>(idx0, static_cast<unsigned int>(total_cross_plane_pad)); |
| 375 | } |
| 376 | |
| 377 | if(_reinterpret_output_as_3d) |
| 378 | { |
| 379 | // Pass bottom paddings to the kernel if the output has to be reinterpreted as 3D tensor |
Gian Marco Iodice | 944170e | 2019-06-24 14:40:30 +0100 | [diff] [blame] | 380 | unsigned int idx0; |
| 381 | if(_add_bias) |
| 382 | { |
| 383 | idx0 = 4 * num_arguments_per_2D_tensor() + 4 + (_reinterpret_input_as_3d ? 1 : 0); |
| 384 | } |
| 385 | else |
| 386 | { |
| 387 | idx0 = 3 * num_arguments_per_2D_tensor() + 3 + (_reinterpret_input_as_3d ? 1 : 0); |
| 388 | } |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 389 | const unsigned int total_cross_plane_pad = _output->info()->padding().top + _output->info()->padding().bottom; |
| 390 | _kernel.setArg<cl_uint>(idx0, static_cast<unsigned int>(total_cross_plane_pad)); |
| 391 | } |
| 392 | |
| 393 | do |
| 394 | { |
| 395 | Window slice_b = slice; |
| 396 | // Don't slice matrix B along the z dimension if matrix B has just 2 dimensions and matrix A more than 2 |
| 397 | // This scenario can happen when the matrix multiplication is used to perform a convolution operation |
| 398 | if(!_slide_matrix_b) |
| 399 | { |
| 400 | slice_b = slice_matrix_b; |
| 401 | } |
| 402 | |
| 403 | unsigned int idx = 0; |
| 404 | add_2D_tensor_argument(idx, _input0, slice); |
| 405 | add_2D_tensor_argument(idx, _input1, slice_b); |
Gian Marco Iodice | 944170e | 2019-06-24 14:40:30 +0100 | [diff] [blame] | 406 | if(_add_bias) |
| 407 | { |
| 408 | add_2D_tensor_argument(idx, _input2, slice); |
| 409 | } |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 410 | add_2D_tensor_argument(idx, _output, slice); |
| 411 | _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_input0->info()->strides_in_bytes()[2])); |
| 412 | _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_input1->info()->strides_in_bytes()[2])); |
Gian Marco Iodice | 944170e | 2019-06-24 14:40:30 +0100 | [diff] [blame] | 413 | if(_add_bias) |
| 414 | { |
| 415 | _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_input2->info()->strides_in_bytes()[2])); |
| 416 | } |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 417 | _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_output->info()->strides_in_bytes()[2])); |
| 418 | enqueue(queue, *this, slice, lws_hint(), _use_dummy_work_items); |
| 419 | } |
| 420 | while(window.slide_window_slice_3D(slice)); |
| 421 | } |
| 422 | } // namespace arm_compute |