Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1 | /* |
giuros01 | 1c9efeb | 2019-01-11 14:04:43 +0000 | [diff] [blame] | 2 | * Copyright (c) 2017-2019 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/core/CL/kernels/CLGEMMMatrixMultiplyKernel.h" |
| 25 | |
| 26 | #include "arm_compute/core/AccessWindowStatic.h" |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 27 | #include "arm_compute/core/CL/CLHelpers.h" |
| 28 | #include "arm_compute/core/CL/CLKernelLibrary.h" |
Vidhya Sudhan Loganathan | f1f4906 | 2018-05-25 13:21:26 +0100 | [diff] [blame] | 29 | #include "arm_compute/core/CL/CLValidate.h" |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 30 | #include "arm_compute/core/CL/ICLTensor.h" |
| 31 | #include "arm_compute/core/CL/OpenCL.h" |
| 32 | #include "arm_compute/core/Error.h" |
| 33 | #include "arm_compute/core/Helpers.h" |
Isabella Gottardi | d56e770 | 2018-02-28 14:29:36 +0000 | [diff] [blame] | 34 | #include "arm_compute/core/TensorInfo.h" |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 35 | #include "arm_compute/core/Types.h" |
| 36 | #include "arm_compute/core/Utils.h" |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 37 | #include "arm_compute/core/Window.h" |
Gian Marco | 36a0a46 | 2018-01-12 10:21:40 +0000 | [diff] [blame] | 38 | #include "arm_compute/core/utils/misc/ShapeCalculator.h" |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 39 | |
| 40 | #include <set> |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 41 | #include <string> |
| 42 | |
giuros01 | 1c9efeb | 2019-01-11 14:04:43 +0000 | [diff] [blame] | 43 | namespace arm_compute |
| 44 | { |
Gian Marco | 36a0a46 | 2018-01-12 10:21:40 +0000 | [diff] [blame] | 45 | using namespace arm_compute::misc::shape_calculator; |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 46 | |
Georgios Pinitas | 358ca20 | 2017-12-07 16:47:52 +0000 | [diff] [blame] | 47 | namespace |
| 48 | { |
| 49 | using ElementsProcessed = Steps; |
| 50 | |
Michele Di Giorgio | ebc3a90 | 2018-11-16 16:04:25 +0000 | [diff] [blame] | 51 | inline Status validate_arguments(const ITensorInfo *input0, const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output, float beta, |
| 52 | bool is_interleaved_transposed, const GEMMReshapeInfo &reshape_info, bool fp_mixed_precision) |
Georgios Pinitas | 358ca20 | 2017-12-07 16:47:52 +0000 | [diff] [blame] | 53 | { |
Georgios Pinitas | 78c0090 | 2018-01-09 17:33:11 +0000 | [diff] [blame] | 54 | ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input0, input1, output); |
Vidhya Sudhan Loganathan | f1f4906 | 2018-05-25 13:21:26 +0100 | [diff] [blame] | 55 | ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input0); |
Vidhya Sudhan Loganathan | 7485d5a | 2018-07-04 09:34:00 +0100 | [diff] [blame] | 56 | ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input0, 1, DataType::F16, DataType::F32); |
Gian Marco | 36a0a46 | 2018-01-12 10:21:40 +0000 | [diff] [blame] | 57 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input0, input1); |
Vidhya Sudhan Loganathan | a25d16c | 2018-11-16 11:33:12 +0000 | [diff] [blame] | 58 | ARM_COMPUTE_RETURN_ERROR_ON_MSG((fp_mixed_precision && (input0->data_type() != DataType::F16)), "Mixed precision floating point is supported only for F16 data"); |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 59 | ARM_COMPUTE_RETURN_ERROR_ON_MSG(input0->num_dimensions() > 4, "The number of dimensions for the matrix A must be <= 4"); |
Gian Marco Iodice | d2fab73 | 2018-03-02 11:18:12 +0000 | [diff] [blame] | 60 | ARM_COMPUTE_RETURN_ERROR_ON_MSG(input1->num_dimensions() > 3, "The number of dimensions for the matrix B must be <= 3"); |
Gian Marco Iodice | 68a3f56 | 2018-07-26 11:44:03 +0100 | [diff] [blame] | 61 | ARM_COMPUTE_RETURN_ERROR_ON_MSG(is_interleaved_transposed && reshape_info.reinterpret_input_as_3d(), "The input tensor cannot be reinterpreted as 3D if is_interleaved_transposed is true"); |
Gian Marco Iodice | d39e2b1 | 2018-08-06 14:31:15 +0100 | [diff] [blame] | 62 | ARM_COMPUTE_RETURN_ERROR_ON_MSG(input1->num_dimensions() > 2 && reshape_info.reinterpret_input_as_3d(), "The input1 tensor cannot have more than 2 dimensions if input0 has to be reinterpreted as 3D"); |
Gian Marco | 36a0a46 | 2018-01-12 10:21:40 +0000 | [diff] [blame] | 63 | |
Michele Di Giorgio | ebc3a90 | 2018-11-16 16:04:25 +0000 | [diff] [blame] | 64 | const bool is_beta_one = std::abs(1.0f - beta) < 0.00001f; |
| 65 | const bool has_vec_c = input2 != nullptr && beta != 0.f; |
| 66 | ARM_COMPUTE_RETURN_ERROR_ON_MSG(has_vec_c && !is_beta_one, "Adding input2 is only supported for beta equal to 1"); |
| 67 | |
Georgios Pinitas | 358ca20 | 2017-12-07 16:47:52 +0000 | [diff] [blame] | 68 | if(!is_interleaved_transposed) |
| 69 | { |
| 70 | ARM_COMPUTE_RETURN_ERROR_ON(input0->dimension(0) != input1->dimension(1)); |
Michele Di Giorgio | ebc3a90 | 2018-11-16 16:04:25 +0000 | [diff] [blame] | 71 | |
| 72 | if(has_vec_c) |
| 73 | { |
| 74 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input0, input2); |
| 75 | ARM_COMPUTE_RETURN_ERROR_ON_MSG(input2->num_dimensions() > 1, "input2 must be a 1D tensor"); |
| 76 | ARM_COMPUTE_RETURN_ERROR_ON_MSG(input2->dimension(0) != input1->dimension(0), "Length of Vector C must match the number of columns of matrix B"); |
| 77 | } |
Gian Marco | 36a0a46 | 2018-01-12 10:21:40 +0000 | [diff] [blame] | 78 | } |
| 79 | else |
| 80 | { |
giuros01 | 8b6b4a9 | 2018-12-18 19:01:33 +0000 | [diff] [blame] | 81 | GEMMRHSMatrixInfo rhs_info; |
giuros01 | 1c9efeb | 2019-01-11 14:04:43 +0000 | [diff] [blame] | 82 | GEMMLHSMatrixInfo lhs_info; |
giuros01 | 8b6b4a9 | 2018-12-18 19:01:33 +0000 | [diff] [blame] | 83 | const int m = reshape_info.m(); |
| 84 | const int n = reshape_info.n(); |
| 85 | const int k = reshape_info.k(); |
| 86 | const int mult_transpose1xW_width = reshape_info.mult_transpose1xW_width(); |
| 87 | const int mult_interleave4x4_height = reshape_info.mult_interleave4x4_height(); |
| 88 | rhs_info.n0 = 16 / input1->element_size(); |
| 89 | rhs_info.k0 = 1; |
| 90 | rhs_info.h0 = mult_transpose1xW_width; |
| 91 | rhs_info.interleave = false; |
| 92 | rhs_info.transpose = false; |
giuros01 | 1c9efeb | 2019-01-11 14:04:43 +0000 | [diff] [blame] | 93 | lhs_info.m0 = 4; |
| 94 | lhs_info.k0 = 4; |
| 95 | lhs_info.v0 = mult_interleave4x4_height; |
| 96 | lhs_info.interleave = true; |
| 97 | lhs_info.transpose = true; |
Gian Marco | 36a0a46 | 2018-01-12 10:21:40 +0000 | [diff] [blame] | 98 | |
| 99 | TensorShape tensor_shape0{ input0->tensor_shape() }; |
| 100 | tensor_shape0.set(0, k); |
| 101 | tensor_shape0.set(1, m); |
| 102 | |
| 103 | TensorShape tensor_shape1{ input1->tensor_shape() }; |
| 104 | tensor_shape1.set(0, n); |
| 105 | tensor_shape1.set(1, k); |
| 106 | |
| 107 | const TensorInfo tensor_info0 = input0->clone()->set_tensor_shape(tensor_shape0); |
| 108 | const TensorInfo tensor_info1 = input1->clone()->set_tensor_shape(tensor_shape1); |
| 109 | |
giuros01 | 1c9efeb | 2019-01-11 14:04:43 +0000 | [diff] [blame] | 110 | const TensorInfo tensor_info_reshaped0 = input0->clone()->set_tensor_shape(compute_lhs_reshaped_shape(tensor_info0, lhs_info)); |
giuros01 | 8b6b4a9 | 2018-12-18 19:01:33 +0000 | [diff] [blame] | 111 | const TensorInfo tensor_info_reshaped1 = input1->clone()->set_tensor_shape(compute_rhs_reshaped_shape(tensor_info1, rhs_info)); |
Gian Marco | 36a0a46 | 2018-01-12 10:21:40 +0000 | [diff] [blame] | 112 | |
| 113 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input0, &tensor_info_reshaped0); |
| 114 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input1, &tensor_info_reshaped1); |
Michele Di Giorgio | ebc3a90 | 2018-11-16 16:04:25 +0000 | [diff] [blame] | 115 | |
| 116 | if(has_vec_c) |
| 117 | { |
| 118 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input0, input2); |
| 119 | ARM_COMPUTE_RETURN_ERROR_ON_MSG(input2->num_dimensions() > 1, "input2 must be a 1D tensor"); |
| 120 | } |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 121 | } |
Gian Marco | 36a0a46 | 2018-01-12 10:21:40 +0000 | [diff] [blame] | 122 | |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 123 | if(output->total_size() != 0) |
| 124 | { |
| 125 | const TensorInfo tensor_info_output = output->clone()->set_tensor_shape(compute_mm_shape(*input0, *input1, is_interleaved_transposed, reshape_info)); |
| 126 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(output, &tensor_info_output); |
| 127 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input0, output); |
Georgios Pinitas | 358ca20 | 2017-12-07 16:47:52 +0000 | [diff] [blame] | 128 | } |
| 129 | |
| 130 | return Status{}; |
| 131 | } |
| 132 | |
Michele Di Giorgio | ebc3a90 | 2018-11-16 16:04:25 +0000 | [diff] [blame] | 133 | inline std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input0, ITensorInfo *input1, ITensorInfo *input2, ITensorInfo *output, |
| 134 | float beta, bool is_interleaved_transposed, const GEMMReshapeInfo &reshape_info, GPUTarget gpu_target, |
Georgios Pinitas | 358ca20 | 2017-12-07 16:47:52 +0000 | [diff] [blame] | 135 | ElementsProcessed &num_elements_processed) |
| 136 | { |
Michele Di Giorgio | ebc3a90 | 2018-11-16 16:04:25 +0000 | [diff] [blame] | 137 | ARM_COMPUTE_UNUSED(beta); |
Georgios Pinitas | 358ca20 | 2017-12-07 16:47:52 +0000 | [diff] [blame] | 138 | bool window_changed = false; |
| 139 | Window win{}; |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 140 | Window win_out{}; |
Georgios Pinitas | 358ca20 | 2017-12-07 16:47:52 +0000 | [diff] [blame] | 141 | |
| 142 | const DataType data_type = input0->data_type(); |
| 143 | unsigned int &num_elems_processed_per_iteration_x = num_elements_processed[0]; |
| 144 | unsigned int &num_elems_processed_per_iteration_y = num_elements_processed[1]; |
Gian Marco Iodice | d39e2b1 | 2018-08-06 14:31:15 +0100 | [diff] [blame] | 145 | bool reinterpret_input_as_3d = reshape_info.reinterpret_input_as_3d(); |
Gian Marco Iodice | 3139f03 | 2018-11-05 14:26:32 +0000 | [diff] [blame] | 146 | bool reinterpret_output_as_3d = (reshape_info.depth_output_gemm3d() != 0); |
Michele Di Giorgio | ebc3a90 | 2018-11-16 16:04:25 +0000 | [diff] [blame] | 147 | const bool has_vec_c = input2 != nullptr && beta != 0.f; |
Gian Marco Iodice | d39e2b1 | 2018-08-06 14:31:15 +0100 | [diff] [blame] | 148 | |
| 149 | // In case both input and output have to be reinterpreted as 3D tensors, |
| 150 | // force reinterpret_input_as_3d and reinterpret_output_as_3d to be false. |
| 151 | if(reinterpret_input_as_3d == reinterpret_output_as_3d) |
| 152 | { |
| 153 | reinterpret_input_as_3d = false; |
| 154 | reinterpret_output_as_3d = false; |
| 155 | } |
Georgios Pinitas | 358ca20 | 2017-12-07 16:47:52 +0000 | [diff] [blame] | 156 | |
Gian Marco Iodice | 750641d | 2018-05-08 12:01:57 +0100 | [diff] [blame] | 157 | // Output tensor auto inizialitation if not yet initialized |
Isabella Gottardi | c4f582e | 2018-10-11 19:14:55 +0100 | [diff] [blame] | 158 | auto_init_if_empty(*output, input0->clone()->set_tensor_shape(compute_mm_shape(*input0, *input1, is_interleaved_transposed, reshape_info))); |
Gian Marco Iodice | 750641d | 2018-05-08 12:01:57 +0100 | [diff] [blame] | 159 | |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 160 | TensorInfo tmp_info(*output); |
| 161 | |
Gian Marco Iodice | d39e2b1 | 2018-08-06 14:31:15 +0100 | [diff] [blame] | 162 | if(reinterpret_output_as_3d) |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 163 | { |
| 164 | // Since the output tensor has to be reinterpreted as 3D and the execute window is based on a 2D GEMM, |
| 165 | // the window needs to be constructed on the 2D collapsed version of the tensor |
| 166 | TensorShape tmp_shape(output->tensor_shape()); |
| 167 | tmp_shape.collapse(2U, 1U); |
| 168 | tmp_info.set_tensor_shape(tmp_shape); |
| 169 | } |
| 170 | |
Georgios Pinitas | 358ca20 | 2017-12-07 16:47:52 +0000 | [diff] [blame] | 171 | if(is_interleaved_transposed) |
| 172 | { |
Gian Marco Iodice | 68a3f56 | 2018-07-26 11:44:03 +0100 | [diff] [blame] | 173 | // reinterpret_input_as_3d is not supported if is_interleaved_transposed is set |
Isabella Gottardi | c4f582e | 2018-10-11 19:14:55 +0100 | [diff] [blame] | 174 | ARM_COMPUTE_ERROR_ON(reshape_info.reinterpret_input_as_3d()); |
Gian Marco Iodice | 68a3f56 | 2018-07-26 11:44:03 +0100 | [diff] [blame] | 175 | |
Georgios Pinitas | 358ca20 | 2017-12-07 16:47:52 +0000 | [diff] [blame] | 176 | // Configure kernel window |
| 177 | num_elems_processed_per_iteration_x = max_cl_vector_width / data_size_from_type(data_type); |
| 178 | num_elems_processed_per_iteration_y = 4; |
| 179 | |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 180 | // Note: bottom paddings are calculated manually as the output can be reinterpreted as 3D tensor |
| 181 | // The only way to set properly the paddings, it is to set those explicitly through the AccessWindowStatic |
| 182 | const int m = reshape_info.m(); |
| 183 | const int bottom_pad = (num_elems_processed_per_iteration_y - (m % num_elems_processed_per_iteration_y)) % num_elems_processed_per_iteration_y; |
| 184 | |
| 185 | win = calculate_max_window(tmp_info, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y)); |
| 186 | win_out = calculate_max_window(*output, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y)); |
Georgios Pinitas | 358ca20 | 2017-12-07 16:47:52 +0000 | [diff] [blame] | 187 | |
Michele Di Giorgio | 17a01a3 | 2019-01-03 15:12:27 +0000 | [diff] [blame] | 188 | AccessWindowStatic input0_access(input0, 0, 0, input0->dimension(0), input0->dimension(1)); |
| 189 | AccessWindowStatic input1_access(input1, 0, 0, |
| 190 | ceil_to_multiple(input1->dimension(0), num_elems_processed_per_iteration_x), |
| 191 | ceil_to_multiple(input1->dimension(1), num_elems_processed_per_iteration_y)); |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 192 | AccessWindowStatic output_access(output, 0, 0, |
| 193 | ceil_to_multiple(output->dimension(0), num_elems_processed_per_iteration_x), |
| 194 | output->dimension(1) + bottom_pad); |
Georgios Pinitas | 358ca20 | 2017-12-07 16:47:52 +0000 | [diff] [blame] | 195 | |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 196 | window_changed = update_window_and_padding(win, input0_access, input1_access) || // window used by the execute_window_loop |
| 197 | update_window_and_padding(win_out, output_access); // window used to update the padding requirements of output tensor |
Michele Di Giorgio | ebc3a90 | 2018-11-16 16:04:25 +0000 | [diff] [blame] | 198 | if(has_vec_c) |
| 199 | { |
| 200 | AccessWindowHorizontal input2_access(input2, 0, num_elems_processed_per_iteration_x); |
| 201 | window_changed = window_changed || update_window_and_padding(win, input2_access); |
| 202 | } |
Georgios Pinitas | 358ca20 | 2017-12-07 16:47:52 +0000 | [diff] [blame] | 203 | |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 204 | output_access.set_valid_region(win_out, ValidRegion(Coordinates(0, 0), output->tensor_shape())); |
Georgios Pinitas | 358ca20 | 2017-12-07 16:47:52 +0000 | [diff] [blame] | 205 | } |
| 206 | else // The input tensors have not been reshaped |
| 207 | { |
| 208 | // Special case for 1xN, 2xN, 3xN and 4xN input0 tensor. num_elems_processed_per_iteration_x is set up for the default case. |
| 209 | num_elems_processed_per_iteration_x = max_cl_vector_width / data_size_from_type(data_type); |
| 210 | num_elems_processed_per_iteration_y = std::min(static_cast<int>(output->dimension(1)), 4); |
| 211 | |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 212 | // Note: bottom paddings are calculated manually as the output can be reinterpreted as 3D tensor |
| 213 | // The only way to set properly the paddings, it is to set those explicitly through the AccessWindowStatic |
Gian Marco Iodice | d39e2b1 | 2018-08-06 14:31:15 +0100 | [diff] [blame] | 214 | const int m = reinterpret_input_as_3d ? input0->tensor_shape()[1] * input0->tensor_shape()[2] : input0->tensor_shape()[1]; |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 215 | const int bottom_pad = (num_elems_processed_per_iteration_y - (m % num_elems_processed_per_iteration_y)) % num_elems_processed_per_iteration_y; |
| 216 | |
Georgios Pinitas | 358ca20 | 2017-12-07 16:47:52 +0000 | [diff] [blame] | 217 | // Create kernels according to the architecture, data type and input size. |
Michalis Spyrou | a967611 | 2018-02-22 18:07:43 +0000 | [diff] [blame] | 218 | GPUTarget arch_target = get_arch_from_target(gpu_target); |
| 219 | if(arch_target == GPUTarget::BIFROST && data_type == DataType::F32) |
Georgios Pinitas | 358ca20 | 2017-12-07 16:47:52 +0000 | [diff] [blame] | 220 | { |
Gian Marco | 1d25ed5 | 2017-12-16 19:33:50 +0000 | [diff] [blame] | 221 | num_elems_processed_per_iteration_x = (input1->dimension(0) <= 1000 && input0->num_dimensions() == 1) ? 2 : 4; |
Georgios Pinitas | 358ca20 | 2017-12-07 16:47:52 +0000 | [diff] [blame] | 222 | } |
| 223 | |
| 224 | // Configure window |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 225 | win = calculate_max_window(tmp_info, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y)); |
| 226 | win_out = calculate_max_window(*output, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y)); |
Georgios Pinitas | 358ca20 | 2017-12-07 16:47:52 +0000 | [diff] [blame] | 227 | |
Gian Marco Iodice | 68a3f56 | 2018-07-26 11:44:03 +0100 | [diff] [blame] | 228 | AccessWindowStatic input0_access(input0, 0, 0, input0->dimension(0), input0->dimension(1) + bottom_pad); |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 229 | AccessWindowStatic input1_access(input1, 0, 0, ceil_to_multiple(input1->dimension(0), num_elems_processed_per_iteration_x), input1->dimension(1)); |
| 230 | AccessWindowStatic output_access(output, 0, 0, |
| 231 | ceil_to_multiple(output->dimension(0), num_elems_processed_per_iteration_x), |
| 232 | output->dimension(1) + bottom_pad); |
Georgios Pinitas | 358ca20 | 2017-12-07 16:47:52 +0000 | [diff] [blame] | 233 | |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 234 | window_changed = update_window_and_padding(win, input0_access, input1_access) || // window used by the execute_window_loop |
| 235 | update_window_and_padding(win_out, output_access); // window used to update the padding requirements of output tensor |
Michele Di Giorgio | ebc3a90 | 2018-11-16 16:04:25 +0000 | [diff] [blame] | 236 | if(has_vec_c) |
| 237 | { |
| 238 | AccessWindowHorizontal input2_access(input2, 0, num_elems_processed_per_iteration_x); |
| 239 | window_changed = window_changed || update_window_and_padding(win, input2_access); |
| 240 | } |
Georgios Pinitas | 358ca20 | 2017-12-07 16:47:52 +0000 | [diff] [blame] | 241 | |
| 242 | Coordinates coord; |
| 243 | coord.set_num_dimensions(output->num_dimensions()); |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 244 | output_access.set_valid_region(win_out, ValidRegion(coord, output->tensor_shape())); |
Georgios Pinitas | 358ca20 | 2017-12-07 16:47:52 +0000 | [diff] [blame] | 245 | } |
| 246 | |
Gian Marco | ae2af74 | 2018-02-15 12:35:44 +0000 | [diff] [blame] | 247 | // Collapse along the Z direction |
| 248 | // This collapse needs to be here in order to tune the Z dimension of LWS |
Gian Marco Iodice | 81b28c4 | 2018-03-29 10:29:36 +0100 | [diff] [blame] | 249 | Window collapsed = win; |
| 250 | const unsigned int dimension_to_collapse = std::min(static_cast<unsigned int>(output->num_dimensions()), 2u); |
| 251 | collapsed = win.collapse(win, dimension_to_collapse); |
Gian Marco | ae2af74 | 2018-02-15 12:35:44 +0000 | [diff] [blame] | 252 | |
Georgios Pinitas | 358ca20 | 2017-12-07 16:47:52 +0000 | [diff] [blame] | 253 | Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{}; |
Gian Marco | ae2af74 | 2018-02-15 12:35:44 +0000 | [diff] [blame] | 254 | return std::make_pair(err, collapsed); |
Georgios Pinitas | 358ca20 | 2017-12-07 16:47:52 +0000 | [diff] [blame] | 255 | } |
| 256 | } // namespace |
| 257 | |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 258 | CLGEMMMatrixMultiplyKernel::CLGEMMMatrixMultiplyKernel() |
Michele Di Giorgio | ebc3a90 | 2018-11-16 16:04:25 +0000 | [diff] [blame] | 259 | : _input0(nullptr), _input1(nullptr), _input2(nullptr), _output(nullptr), _slide_matrix_b(true), _reinterpret_input_as_3d(false), _reinterpret_output_as_3d(false), _has_vec_c(false) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 260 | { |
| 261 | } |
| 262 | |
Michele Di Giorgio | ebc3a90 | 2018-11-16 16:04:25 +0000 | [diff] [blame] | 263 | void CLGEMMMatrixMultiplyKernel::configure(const ICLTensor *input0, const ICLTensor *input1, const ICLTensor *input2, ICLTensor *output, float alpha, float beta, |
| 264 | bool is_interleaved_transposed, const GEMMReshapeInfo &reshape_info, bool fp_mixed_precision) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 265 | { |
Georgios Pinitas | 358ca20 | 2017-12-07 16:47:52 +0000 | [diff] [blame] | 266 | ARM_COMPUTE_ERROR_ON_NULLPTR(input0, input1, output); |
| 267 | |
| 268 | // Perform validate step |
Michele Di Giorgio | ebc3a90 | 2018-11-16 16:04:25 +0000 | [diff] [blame] | 269 | ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input0->info(), input1->info(), (input2 != nullptr) ? input2->info() : nullptr, output->info(), beta, |
| 270 | is_interleaved_transposed, reshape_info, fp_mixed_precision)); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 271 | |
Gian Marco Iodice | 68a3f56 | 2018-07-26 11:44:03 +0100 | [diff] [blame] | 272 | _input0 = input0; |
| 273 | _input1 = input1; |
Michele Di Giorgio | ebc3a90 | 2018-11-16 16:04:25 +0000 | [diff] [blame] | 274 | _input2 = input2; |
Gian Marco Iodice | 68a3f56 | 2018-07-26 11:44:03 +0100 | [diff] [blame] | 275 | _output = output; |
| 276 | _reinterpret_input_as_3d = reshape_info.reinterpret_input_as_3d(); |
Gian Marco Iodice | 3139f03 | 2018-11-05 14:26:32 +0000 | [diff] [blame] | 277 | _reinterpret_output_as_3d = (reshape_info.depth_output_gemm3d() != 0); |
Gian Marco Iodice | 68a3f56 | 2018-07-26 11:44:03 +0100 | [diff] [blame] | 278 | |
Gian Marco Iodice | d39e2b1 | 2018-08-06 14:31:15 +0100 | [diff] [blame] | 279 | // In case both input and output have to be reinterpreted as 3D tensors, |
| 280 | // force reinterpret_input_as_3d and reinterpret_output_as_3d to be false. |
| 281 | if(_reinterpret_input_as_3d == _reinterpret_output_as_3d) |
| 282 | { |
| 283 | _reinterpret_input_as_3d = false; |
| 284 | _reinterpret_output_as_3d = false; |
| 285 | } |
| 286 | |
Gian Marco Iodice | 68a3f56 | 2018-07-26 11:44:03 +0100 | [diff] [blame] | 287 | // Check if we need to slide the matrix B |
| 288 | const unsigned int num_dimensions_input0 = _reinterpret_input_as_3d ? _input0->info()->num_dimensions() - 1 : _input0->info()->num_dimensions(); |
| 289 | |
| 290 | _slide_matrix_b = (_input1->info()->num_dimensions() >= num_dimensions_input0); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 291 | |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 292 | const DataType data_type = input0->info()->data_type(); |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 293 | |
| 294 | // Get target architecture |
| 295 | GPUTarget gpu_target = get_target(); |
| 296 | |
Georgios Pinitas | 358ca20 | 2017-12-07 16:47:52 +0000 | [diff] [blame] | 297 | ElementsProcessed num_elements_processed{}; |
| 298 | |
| 299 | // Configure kernel window |
Michele Di Giorgio | ebc3a90 | 2018-11-16 16:04:25 +0000 | [diff] [blame] | 300 | auto win_config = validate_and_configure_window(input0->info(), input1->info(), (input2 != nullptr) ? input2->info() : nullptr, output->info(), beta, is_interleaved_transposed, reshape_info, |
| 301 | gpu_target, num_elements_processed); |
Georgios Pinitas | 358ca20 | 2017-12-07 16:47:52 +0000 | [diff] [blame] | 302 | ARM_COMPUTE_ERROR_THROW_ON(win_config.first); |
Anthony Barbier | b6eb353 | 2018-08-08 13:20:04 +0100 | [diff] [blame] | 303 | ICLKernel::configure_internal(win_config.second); |
Georgios Pinitas | 358ca20 | 2017-12-07 16:47:52 +0000 | [diff] [blame] | 304 | |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 305 | // Create build options |
| 306 | CLBuildOptions build_opts; |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 307 | |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 308 | // Only define ALPHA when alpha is not 1.0f. This avoids performing unnecessary multiplications. |
Georgios Pinitas | 358ca20 | 2017-12-07 16:47:52 +0000 | [diff] [blame] | 309 | if(std::abs(1.0f - alpha) > 0.00001f) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 310 | { |
Vidhya Sudhan Loganathan | 7485d5a | 2018-07-04 09:34:00 +0100 | [diff] [blame] | 311 | build_opts.add_option("-DALPHA=" + float_to_string_with_full_precision(alpha)); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 312 | } |
Gian Marco Iodice | 68a3f56 | 2018-07-26 11:44:03 +0100 | [diff] [blame] | 313 | build_opts.add_option_if(_reinterpret_input_as_3d, "-DREINTERPRET_INPUT_AS_3D"); |
| 314 | build_opts.add_option_if(_reinterpret_output_as_3d, "-DREINTERPRET_OUTPUT_AS_3D"); |
| 315 | build_opts.add_option_if(_reinterpret_input_as_3d || _reinterpret_output_as_3d, "-DHEIGHT_GEMM3D=" + support::cpp11::to_string(output->info()->dimension(1))); |
| 316 | build_opts.add_option_if(_reinterpret_input_as_3d || _reinterpret_output_as_3d, "-DDEPTH_GEMM3D=" + support::cpp11::to_string(output->info()->dimension(2))); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 317 | |
Gian Marco Iodice | d2fab73 | 2018-03-02 11:18:12 +0000 | [diff] [blame] | 318 | // Do not slide matrix B if _slide_matrix_b = false |
| 319 | build_opts.add_option_if(!_slide_matrix_b, "-DMATRIX_B_DEPTH=" + support::cpp11::to_string(input1->info()->dimension(2))); |
| 320 | |
Gian Marco Iodice | bb36a8e | 2018-04-19 12:05:08 +0100 | [diff] [blame] | 321 | const bool is_bifrost = get_arch_from_target(gpu_target) == GPUTarget::BIFROST; |
| 322 | |
Michele Di Giorgio | ebc3a90 | 2018-11-16 16:04:25 +0000 | [diff] [blame] | 323 | _has_vec_c = input2 != nullptr && beta != 0.f; |
| 324 | |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 325 | std::string kernel_name; |
Gian Marco Iodice | edfa9f4 | 2017-08-15 11:45:22 +0100 | [diff] [blame] | 326 | if(is_interleaved_transposed) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 327 | { |
Gian Marco | 36a0a46 | 2018-01-12 10:21:40 +0000 | [diff] [blame] | 328 | const int mult_transpose1xW_width = reshape_info.mult_transpose1xW_width(); |
| 329 | const int mult_interleave4x4_height = reshape_info.mult_interleave4x4_height(); |
| 330 | |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 331 | build_opts.add_option("-DCOLS_B=" + support::cpp11::to_string(input1->info()->dimension(0))); |
Gian Marco | 36a0a46 | 2018-01-12 10:21:40 +0000 | [diff] [blame] | 332 | build_opts.add_option("-DMULT_TRANSPOSE1XW_WIDTH=" + support::cpp11::to_string(mult_transpose1xW_width)); |
| 333 | build_opts.add_option("-DMULT_INTERLEAVE4X4_HEIGHT=" + support::cpp11::to_string(mult_interleave4x4_height)); |
| 334 | |
Gian Marco Iodice | bb36a8e | 2018-04-19 12:05:08 +0100 | [diff] [blame] | 335 | if(is_data_type_float(data_type) && is_bifrost) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 336 | { |
Gian Marco Iodice | bb36a8e | 2018-04-19 12:05:08 +0100 | [diff] [blame] | 337 | kernel_name = "gemm_mm_interleaved_transposed_" + lower_string(string_from_data_type(data_type)) + "_bifrost"; |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 338 | } |
| 339 | else |
| 340 | { |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 341 | kernel_name = "gemm_mm_interleaved_transposed_" + lower_string(string_from_data_type(data_type)); |
Vidhya Sudhan Loganathan | 38d93bd | 2018-11-20 15:38:13 +0000 | [diff] [blame] | 342 | if(fp_mixed_precision && data_type == DataType::F16) |
| 343 | { |
| 344 | // currently wider accumulator is only supported for fp16 kernels. |
| 345 | kernel_name += "_acc32"; |
| 346 | } |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 347 | } |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 348 | } |
Gian Marco Iodice | edfa9f4 | 2017-08-15 11:45:22 +0100 | [diff] [blame] | 349 | else // The input tensors have not been reshaped |
| 350 | { |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 351 | build_opts.add_option("-DCOLS_A=" + support::cpp11::to_string(input0->info()->dimension(0))); |
Gian Marco Iodice | e52a300 | 2018-04-11 15:59:10 +0100 | [diff] [blame] | 352 | build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(data_type)); |
Gian Marco Iodice | edfa9f4 | 2017-08-15 11:45:22 +0100 | [diff] [blame] | 353 | |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 354 | // Create kernels according to the architecture, data type and input size. |
Gian Marco Iodice | bb36a8e | 2018-04-19 12:05:08 +0100 | [diff] [blame] | 355 | if(is_data_type_float(data_type) && is_bifrost) |
Gian Marco Iodice | edfa9f4 | 2017-08-15 11:45:22 +0100 | [diff] [blame] | 356 | { |
Gian Marco Iodice | e52a300 | 2018-04-11 15:59:10 +0100 | [diff] [blame] | 357 | kernel_name = "gemm_mm_floating_point"; |
| 358 | |
| 359 | if(input0->info()->num_dimensions() != 1) |
Gian Marco Iodice | fd68311 | 2018-04-17 09:52:44 +0100 | [diff] [blame] | 360 | { |
Gian Marco Iodice | e52a300 | 2018-04-11 15:59:10 +0100 | [diff] [blame] | 361 | kernel_name += "_" + lower_string(string_from_data_type(data_type)) + "_bifrost"; |
Vidhya Sudhan Loganathan | 38d93bd | 2018-11-20 15:38:13 +0000 | [diff] [blame] | 362 | if(fp_mixed_precision && data_type == DataType::F16) |
| 363 | { |
| 364 | // currently wider accumulator is only supported for fp16 kernels. |
| 365 | kernel_name += "_acc32"; |
| 366 | } |
Gian Marco Iodice | e52a300 | 2018-04-11 15:59:10 +0100 | [diff] [blame] | 367 | } |
| 368 | else if(input1->info()->dimension(0) <= 1000 && data_type == DataType::F32) |
| 369 | { |
| 370 | // The first kernel is optimized for the case of 1000 or less output elements (e.g. FC8 of AlexNet and VGG-16, and |
| 371 | // FC1 of Inception v3). The second kernel is optimized for the case of greater than 1000 output elements (e.g. |
| 372 | // FC6 and FC7 of AlexNet and VGG-16). |
| 373 | kernel_name += "_" + lower_string(string_from_data_type(data_type)) + "_bifrost_1000"; |
Gian Marco Iodice | fd68311 | 2018-04-17 09:52:44 +0100 | [diff] [blame] | 374 | } |
Georgios Pinitas | 358ca20 | 2017-12-07 16:47:52 +0000 | [diff] [blame] | 375 | |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 376 | // The work-group size equal to the Bifrost quad size has been proved to be optimal for these kernels |
| 377 | // via exhaustive autotuning over a range of representative layer configurations. |
Anthony Barbier | b6eb353 | 2018-08-08 13:20:04 +0100 | [diff] [blame] | 378 | set_lws_hint(cl::NDRange(4)); |
Gian Marco Iodice | edfa9f4 | 2017-08-15 11:45:22 +0100 | [diff] [blame] | 379 | } |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 380 | else // (MIDGARD and F32) or (F16) |
| 381 | { |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 382 | kernel_name = "gemm_mm_floating_point"; |
| 383 | } |
Georgios Pinitas | 358ca20 | 2017-12-07 16:47:52 +0000 | [diff] [blame] | 384 | build_opts.add_option("-DNUM_ELEMS_PROCESSED_PER_THREAD_Y=" + support::cpp11::to_string(num_elements_processed.y())); |
| 385 | build_opts.add_option("-DNUM_ELEMS_PROCESSED_PER_THREAD_X=" + support::cpp11::to_string(num_elements_processed.x())); |
Gian Marco Iodice | edfa9f4 | 2017-08-15 11:45:22 +0100 | [diff] [blame] | 386 | } |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 387 | |
Michele Di Giorgio | ebc3a90 | 2018-11-16 16:04:25 +0000 | [diff] [blame] | 388 | // Configure matrix C addition if necessary |
| 389 | build_opts.add_option_if(_has_vec_c, "-DADD_VEC_C"); |
| 390 | |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 391 | // Create kernel |
| 392 | _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel(kernel_name, build_opts.options())); |
| 393 | |
| 394 | // Set config_id for enabling LWS tuning |
| 395 | _config_id = "gemm_"; |
| 396 | _config_id += (is_interleaved_transposed ? "reshaped_" : ""); |
Vidhya Sudhan Loganathan | a25d16c | 2018-11-16 11:33:12 +0000 | [diff] [blame] | 397 | _config_id += (fp_mixed_precision ? "fp_mixed_" : ""); |
Gian Marco Iodice | 68a3f56 | 2018-07-26 11:44:03 +0100 | [diff] [blame] | 398 | _config_id += (_reinterpret_input_as_3d ? "3di_" : ""); |
| 399 | _config_id += (_reinterpret_output_as_3d ? "3do_" : ""); |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 400 | _config_id += lower_string(string_from_data_type(input0->info()->data_type())); |
| 401 | _config_id += "_"; |
| 402 | _config_id += support::cpp11::to_string(output->info()->dimension(1)); |
| 403 | _config_id += "_"; |
| 404 | _config_id += support::cpp11::to_string(output->info()->dimension(0)); |
| 405 | _config_id += "_"; |
Gian Marco | ae2af74 | 2018-02-15 12:35:44 +0000 | [diff] [blame] | 406 | _config_id += support::cpp11::to_string(output->info()->dimension(2)); |
| 407 | _config_id += "_"; |
| 408 | _config_id += support::cpp11::to_string(output->info()->dimension(3)); |
| 409 | _config_id += "_"; |
Anton Lokhmotov | 3e80c7f | 2017-11-20 11:02:10 +0000 | [diff] [blame] | 410 | _config_id += (is_interleaved_transposed ? support::cpp11::to_string(input1->info()->dimension(0)) : support::cpp11::to_string(input1->info()->dimension(1))); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 411 | } |
| 412 | |
Michele Di Giorgio | ebc3a90 | 2018-11-16 16:04:25 +0000 | [diff] [blame] | 413 | Status CLGEMMMatrixMultiplyKernel::validate(const ITensorInfo *input0, const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output, float alpha, float beta, |
| 414 | bool is_interleaved_transposed, const GEMMReshapeInfo &reshape_info, GPUTarget gpu_target, bool fp_mixed_precision) |
Georgios Pinitas | 358ca20 | 2017-12-07 16:47:52 +0000 | [diff] [blame] | 415 | { |
Gian Marco | 36a0a46 | 2018-01-12 10:21:40 +0000 | [diff] [blame] | 416 | // Note: num_elements_processed will be set in validate_and_configure_window() |
Georgios Pinitas | 358ca20 | 2017-12-07 16:47:52 +0000 | [diff] [blame] | 417 | ElementsProcessed num_elements_processed{}; |
| 418 | ARM_COMPUTE_UNUSED(alpha); |
Michele Di Giorgio | ebc3a90 | 2018-11-16 16:04:25 +0000 | [diff] [blame] | 419 | ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input0, input1, input2, output, beta, is_interleaved_transposed, reshape_info, fp_mixed_precision)); |
Georgios Pinitas | 358ca20 | 2017-12-07 16:47:52 +0000 | [diff] [blame] | 420 | ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input0->clone().get(), |
| 421 | input1->clone().get(), |
Michele Di Giorgio | ebc3a90 | 2018-11-16 16:04:25 +0000 | [diff] [blame] | 422 | (input2 != nullptr) ? input2->clone().get() : nullptr, |
Georgios Pinitas | 358ca20 | 2017-12-07 16:47:52 +0000 | [diff] [blame] | 423 | output->clone().get(), |
Michele Di Giorgio | ebc3a90 | 2018-11-16 16:04:25 +0000 | [diff] [blame] | 424 | beta, |
Georgios Pinitas | 358ca20 | 2017-12-07 16:47:52 +0000 | [diff] [blame] | 425 | is_interleaved_transposed, |
Gian Marco Iodice | 750641d | 2018-05-08 12:01:57 +0100 | [diff] [blame] | 426 | reshape_info, |
Georgios Pinitas | 358ca20 | 2017-12-07 16:47:52 +0000 | [diff] [blame] | 427 | gpu_target, |
| 428 | num_elements_processed) |
| 429 | .first); |
| 430 | |
| 431 | return Status{}; |
| 432 | } |
| 433 | |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 434 | void CLGEMMMatrixMultiplyKernel::run(const Window &window, cl::CommandQueue &queue) |
| 435 | { |
| 436 | ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); |
| 437 | ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window); |
| 438 | |
Gian Marco | ae2af74 | 2018-02-15 12:35:44 +0000 | [diff] [blame] | 439 | if(_input1->info()->num_dimensions() < 3) |
| 440 | { |
| 441 | // The stride_z for matrix B must be zero if we do not slice |
| 442 | ARM_COMPUTE_ERROR_ON(_input1->info()->strides_in_bytes()[3] != 0); |
| 443 | } |
| 444 | |
| 445 | Window slice = window.first_slice_window_3D(); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 446 | Window slice_matrix_b = slice; |
Gian Marco Iodice | edfa9f4 | 2017-08-15 11:45:22 +0100 | [diff] [blame] | 447 | |
| 448 | slice_matrix_b.set(Window::DimX, Window::Dimension(0, 1, 1)); |
| 449 | slice_matrix_b.set(Window::DimY, Window::Dimension(0, 1, 1)); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 450 | |
Michele Di Giorgio | ebc3a90 | 2018-11-16 16:04:25 +0000 | [diff] [blame] | 451 | const unsigned int num_arguments_vec_c = (_has_vec_c) ? num_arguments_per_1D_tensor() : 0; |
| 452 | |
Gian Marco Iodice | 68a3f56 | 2018-07-26 11:44:03 +0100 | [diff] [blame] | 453 | if(_reinterpret_input_as_3d) |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 454 | { |
Isabella Gottardi | b92805b | 2018-09-28 18:24:27 +0100 | [diff] [blame] | 455 | // Pass bottom paddings to the kernel if the input has to be reinterpreted as 3D tensor |
Michele Di Giorgio | ebc3a90 | 2018-11-16 16:04:25 +0000 | [diff] [blame] | 456 | const unsigned int idx0 = 3 * num_arguments_per_2D_tensor() + 3 + num_arguments_vec_c; |
Gian Marco Iodice | 68a3f56 | 2018-07-26 11:44:03 +0100 | [diff] [blame] | 457 | const unsigned int total_cross_plane_pad = _input0->info()->padding().top + _input0->info()->padding().bottom; |
| 458 | _kernel.setArg<cl_uint>(idx0, static_cast<unsigned int>(total_cross_plane_pad)); |
| 459 | } |
| 460 | |
| 461 | if(_reinterpret_output_as_3d) |
| 462 | { |
| 463 | // Pass bottom paddings to the kernel if the output has to be reinterpreted as 3D tensor |
Michele Di Giorgio | ebc3a90 | 2018-11-16 16:04:25 +0000 | [diff] [blame] | 464 | const unsigned int idx0 = 3 * num_arguments_per_2D_tensor() + 3 + (_reinterpret_input_as_3d ? 1 : 0) + num_arguments_vec_c; |
Georgios Pinitas | e8bd2c7 | 2018-07-11 15:54:56 +0100 | [diff] [blame] | 465 | const unsigned int total_cross_plane_pad = _output->info()->padding().top + _output->info()->padding().bottom; |
| 466 | _kernel.setArg<cl_uint>(idx0, static_cast<unsigned int>(total_cross_plane_pad)); |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 467 | } |
| 468 | |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 469 | do |
| 470 | { |
| 471 | Window slice_b = slice; |
| 472 | // Don't slice matrix B along the z dimension if matrix B has just 2 dimensions and matrix A more than 2 |
Gian Marco | ae2af74 | 2018-02-15 12:35:44 +0000 | [diff] [blame] | 473 | // This scenario can happen when the matrix multiplication is used to perform a convolution operation |
Gian Marco Iodice | d2fab73 | 2018-03-02 11:18:12 +0000 | [diff] [blame] | 474 | if(!_slide_matrix_b) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 475 | { |
| 476 | slice_b = slice_matrix_b; |
| 477 | } |
| 478 | |
| 479 | unsigned int idx = 0; |
| 480 | add_2D_tensor_argument(idx, _input0, slice); |
| 481 | add_2D_tensor_argument(idx, _input1, slice_b); |
Michele Di Giorgio | ebc3a90 | 2018-11-16 16:04:25 +0000 | [diff] [blame] | 482 | if(_has_vec_c) |
| 483 | { |
| 484 | add_1D_tensor_argument(idx, _input2, slice); |
| 485 | } |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 486 | add_2D_tensor_argument(idx, _output, slice); |
Gian Marco Iodice | d2fab73 | 2018-03-02 11:18:12 +0000 | [diff] [blame] | 487 | _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_input0->info()->strides_in_bytes()[2])); |
| 488 | _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_input1->info()->strides_in_bytes()[2])); |
| 489 | _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_output->info()->strides_in_bytes()[2])); |
Anthony Barbier | b6eb353 | 2018-08-08 13:20:04 +0100 | [diff] [blame] | 490 | enqueue(queue, *this, slice, lws_hint()); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 491 | } |
Gian Marco | ae2af74 | 2018-02-15 12:35:44 +0000 | [diff] [blame] | 492 | while(window.slide_window_slice_3D(slice)); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 493 | } |
giuros01 | 1c9efeb | 2019-01-11 14:04:43 +0000 | [diff] [blame] | 494 | } // namespace arm_compute |