Gian Marco Iodice | 62251f7 | 2019-03-11 16:07:12 +0000 | [diff] [blame] | 1 | /* |
Michele Di Giorgio | b54ba28 | 2020-01-14 15:31:55 +0000 | [diff] [blame] | 2 | * Copyright (c) 2019-2020 ARM Limited. |
Gian Marco Iodice | 62251f7 | 2019-03-11 16:07:12 +0000 | [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/CLGEMMLowpMatrixMultiplyReshapedOnlyRHSKernel.h" |
| 25 | |
| 26 | #include "arm_compute/core/AccessWindowStatic.h" |
| 27 | #include "arm_compute/core/CL/CLHelpers.h" |
| 28 | #include "arm_compute/core/CL/CLKernelLibrary.h" |
| 29 | #include "arm_compute/core/CL/ICLTensor.h" |
| 30 | #include "arm_compute/core/CL/OpenCL.h" |
| 31 | #include "arm_compute/core/Error.h" |
| 32 | #include "arm_compute/core/Helpers.h" |
| 33 | #include "arm_compute/core/TensorInfo.h" |
| 34 | #include "arm_compute/core/Types.h" |
| 35 | #include "arm_compute/core/Utils.h" |
| 36 | #include "arm_compute/core/Validate.h" |
| 37 | #include "arm_compute/core/Window.h" |
| 38 | #include "arm_compute/core/utils/misc/ShapeCalculator.h" |
| 39 | #include "support/ToolchainSupport.h" |
| 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 | |
Michele Di Giorgio | b54ba28 | 2020-01-14 15:31:55 +0000 | [diff] [blame] | 53 | Status validate_arguments(const ITensorInfo *input0, const ITensorInfo *input1, const ITensorInfo *output, const GEMMKernelInfo &gemm_info, |
| 54 | const ITensorInfo *vector_sum_col, const ITensorInfo *vector_sum_row, const ITensorInfo *bias, |
| 55 | const ITensorInfo *output_multipliers, const ITensorInfo *output_shifts) |
Gian Marco Iodice | 62251f7 | 2019-03-11 16:07:12 +0000 | [diff] [blame] | 56 | { |
| 57 | ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input0, input1, output); |
Michele Di Giorgio | f9179d3 | 2019-11-27 16:17:30 +0000 | [diff] [blame] | 58 | ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input0, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED); |
Gian Marco Iodice | 62251f7 | 2019-03-11 16:07:12 +0000 | [diff] [blame] | 59 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input0, input1); |
| 60 | ARM_COMPUTE_RETURN_ERROR_ON_MSG(input0->num_dimensions() > 4, "The number of dimensions for the LHS matrix must be <= 4"); |
| 61 | ARM_COMPUTE_RETURN_ERROR_ON_MSG(input1->num_dimensions() > 3, "The number of dimensions for the RHS matrix must be <= 3"); |
Michele Di Giorgio | b54ba28 | 2020-01-14 15:31:55 +0000 | [diff] [blame] | 62 | |
| 63 | const GEMMRHSMatrixInfo rhs_info = gemm_info.rhs_info; |
| 64 | const GEMMLHSMatrixInfo lhs_info = gemm_info.lhs_info; |
| 65 | const GEMMLowpOutputStageInfo output_stage = gemm_info.output_stage; |
| 66 | |
Gian Marco Iodice | 62251f7 | 2019-03-11 16:07:12 +0000 | [diff] [blame] | 67 | ARM_COMPUTE_RETURN_ERROR_ON_MSG((((rhs_info.k0 & (rhs_info.k0 - 1)) && rhs_info.k0 != 3) || (rhs_info.k0 > 16)), "Only 2,3,4,8,16 are supported for k0"); |
| 68 | ARM_COMPUTE_RETURN_ERROR_ON(lhs_info.m0 < 1 || lhs_info.m0 > 8); |
| 69 | ARM_COMPUTE_RETURN_ERROR_ON_MSG((((rhs_info.n0 & (rhs_info.n0 - 1)) && rhs_info.n0 != 3) || rhs_info.n0 > 16), "Only 2,3,4,8,16 are supported for n0"); |
| 70 | |
Michele Di Giorgio | b54ba28 | 2020-01-14 15:31:55 +0000 | [diff] [blame] | 71 | const int m = gemm_info.m; |
| 72 | const int n = gemm_info.n; |
| 73 | const int k = gemm_info.k; |
Gian Marco Iodice | 62251f7 | 2019-03-11 16:07:12 +0000 | [diff] [blame] | 74 | |
Gian Marco Iodice | 62251f7 | 2019-03-11 16:07:12 +0000 | [diff] [blame] | 75 | TensorShape tensor_shape1{ input1->tensor_shape() }; |
| 76 | tensor_shape1.set(0, n); |
| 77 | tensor_shape1.set(1, k); |
| 78 | |
Gian Marco Iodice | 86cfffe | 2019-04-02 11:02:20 +0100 | [diff] [blame] | 79 | const TensorInfo tensor_info1 = input1->clone()->set_tensor_shape(tensor_shape1); |
Gian Marco Iodice | 62251f7 | 2019-03-11 16:07:12 +0000 | [diff] [blame] | 80 | const TensorInfo tensor_info_reshaped1 = input1->clone()->set_tensor_shape(compute_rhs_reshaped_shape(tensor_info1, rhs_info)); |
| 81 | |
Gian Marco Iodice | 86cfffe | 2019-04-02 11:02:20 +0100 | [diff] [blame] | 82 | ARM_COMPUTE_RETURN_ERROR_ON(input0->dimension(0) != static_cast<unsigned int>(k)); |
Michele Di Giorgio | b54ba28 | 2020-01-14 15:31:55 +0000 | [diff] [blame] | 83 | if(gemm_info.reinterpret_input_as_3d) |
Gian Marco Iodice | 86cfffe | 2019-04-02 11:02:20 +0100 | [diff] [blame] | 84 | { |
| 85 | ARM_COMPUTE_RETURN_ERROR_ON(input0->dimension(1) * input0->dimension(2) != static_cast<unsigned int>(m)); |
| 86 | } |
| 87 | else |
| 88 | { |
| 89 | ARM_COMPUTE_RETURN_ERROR_ON(input0->dimension(1) != static_cast<unsigned int>(m)); |
| 90 | } |
Gian Marco Iodice | 62251f7 | 2019-03-11 16:07:12 +0000 | [diff] [blame] | 91 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input1, &tensor_info_reshaped1); |
| 92 | |
Michele Di Giorgio | b54ba28 | 2020-01-14 15:31:55 +0000 | [diff] [blame] | 93 | const TensorShape expected_output_shape = compute_mm_shape(*input0, *input1, gemm_info); |
Gian Marco Iodice | 62251f7 | 2019-03-11 16:07:12 +0000 | [diff] [blame] | 94 | if(output->total_size() != 0) |
| 95 | { |
Michele Di Giorgio | b54ba28 | 2020-01-14 15:31:55 +0000 | [diff] [blame] | 96 | const TensorInfo tensor_info_output = output->clone()->set_tensor_shape(expected_output_shape); |
Gian Marco Iodice | 62251f7 | 2019-03-11 16:07:12 +0000 | [diff] [blame] | 97 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(output, &tensor_info_output); |
Michele Di Giorgio | b54ba28 | 2020-01-14 15:31:55 +0000 | [diff] [blame] | 98 | if(output_stage.type == GEMMLowpOutputStageType::NONE) |
| 99 | { |
| 100 | ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::S32); |
| 101 | } |
| 102 | else |
| 103 | { |
| 104 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input0, output); |
| 105 | } |
Gian Marco Iodice | 62251f7 | 2019-03-11 16:07:12 +0000 | [diff] [blame] | 106 | } |
| 107 | |
Michele Di Giorgio | b54ba28 | 2020-01-14 15:31:55 +0000 | [diff] [blame] | 108 | if(bias != nullptr) |
| 109 | { |
| 110 | ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(bias, 1, DataType::S32); |
| 111 | ARM_COMPUTE_RETURN_ERROR_ON(bias->num_dimensions() > 1); |
| 112 | ARM_COMPUTE_RETURN_ERROR_ON(expected_output_shape[0] != bias->dimension(0)); |
| 113 | } |
| 114 | |
| 115 | ARM_COMPUTE_RETURN_ERROR_ON_MSG((output_stage.type == GEMMLowpOutputStageType::QUANTIZE_DOWN) || (output_stage.type == GEMMLowpOutputStageType::QUANTIZE_DOWN_FLOAT), |
| 116 | "Only GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT is supported"); |
| 117 | |
| 118 | // Checks performed if the output stage needs to be fused |
| 119 | if(output_stage.type == GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT) |
| 120 | { |
| 121 | // If a_offset == 0, vector_sum_col can be a nullptr |
| 122 | if(gemm_info.a_offset != 0) |
| 123 | { |
| 124 | ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(vector_sum_col, 1, DataType::S32); |
| 125 | ARM_COMPUTE_RETURN_ERROR_ON(vector_sum_col->dimension(0) != expected_output_shape[0]); |
| 126 | } |
| 127 | |
| 128 | // If b_offset == 0, vector_sum_row can be a nullptr |
| 129 | if(gemm_info.b_offset != 0) |
| 130 | { |
| 131 | ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(vector_sum_row, 1, DataType::S32); |
| 132 | |
| 133 | // Check if mm result is a 3D reinterpretation |
| 134 | const bool reinterpret_as_3d = expected_output_shape.num_dimensions() > 1 && expected_output_shape.y() != vector_sum_row->tensor_shape().x(); |
| 135 | |
| 136 | // Validate input |
| 137 | ARM_COMPUTE_RETURN_ERROR_ON(reinterpret_as_3d && vector_sum_row->dimension(0) != (expected_output_shape[1] * expected_output_shape[2])); |
| 138 | ARM_COMPUTE_RETURN_ERROR_ON(!reinterpret_as_3d && vector_sum_row->dimension(0) != expected_output_shape[1]); |
| 139 | |
| 140 | if(expected_output_shape.num_dimensions() > 1) |
| 141 | { |
| 142 | const unsigned int output_batch_idx = reinterpret_as_3d ? 3 : 2; |
| 143 | |
| 144 | TensorShape vector_sum_row_shape = vector_sum_row->tensor_shape(); |
| 145 | vector_sum_row_shape.collapse_from(1); |
| 146 | TensorShape collapsed_output_shape(expected_output_shape); |
| 147 | collapsed_output_shape.collapse_from(output_batch_idx); |
| 148 | |
| 149 | ARM_COMPUTE_RETURN_ERROR_ON_MSG(vector_sum_row_shape[1] != collapsed_output_shape[output_batch_idx], |
| 150 | "vector_sum_row must have the same number of batches of output tensor"); |
| 151 | |
| 152 | if(gemm_info.a_offset != 0) |
| 153 | { |
| 154 | TensorShape vector_sum_col_shape = vector_sum_col->tensor_shape(); |
| 155 | vector_sum_col_shape.collapse_from(1); |
| 156 | |
| 157 | ARM_COMPUTE_RETURN_ERROR_ON_MSG(vector_sum_col_shape[1] != 1 && vector_sum_col_shape[1] != vector_sum_row_shape[1], |
| 158 | "vector_sum_col tensor must have the same number of batches of vector_sum_row_shape or the number of batches must be set to 1"); |
| 159 | } |
| 160 | } |
| 161 | } |
| 162 | |
Michele Di Giorgio | b54ba28 | 2020-01-14 15:31:55 +0000 | [diff] [blame] | 163 | if(output->total_size() != 0) |
| 164 | { |
| 165 | ARM_COMPUTE_RETURN_ERROR_ON(output_stage.output_data_type != output->data_type()); |
Michele Di Giorgio | b54ba28 | 2020-01-14 15:31:55 +0000 | [diff] [blame] | 166 | } |
Michele Di Giorgio | 398b8e4 | 2020-03-06 13:56:54 +0000 | [diff] [blame^] | 167 | ARM_COMPUTE_RETURN_ERROR_ON(output_stage.gemmlowp_min_bound > output_stage.gemmlowp_max_bound); |
Michele Di Giorgio | b54ba28 | 2020-01-14 15:31:55 +0000 | [diff] [blame] | 168 | |
| 169 | if(output_multipliers != nullptr && output_shifts != nullptr) |
| 170 | { |
| 171 | ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output_multipliers, 1, DataType::S32); |
| 172 | ARM_COMPUTE_RETURN_ERROR_ON(output_multipliers->num_dimensions() > 1); |
| 173 | ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output_shifts, 1, DataType::S32); |
| 174 | ARM_COMPUTE_RETURN_ERROR_ON(output_shifts->num_dimensions() > 1); |
| 175 | if(output_stage.is_quantized_per_channel) |
| 176 | { |
| 177 | ARM_COMPUTE_RETURN_ERROR_ON(expected_output_shape[0] != output_shifts->dimension(0)); |
| 178 | ARM_COMPUTE_RETURN_ERROR_ON(expected_output_shape[0] != output_multipliers->dimension(0)); |
| 179 | } |
| 180 | } |
| 181 | } |
Gian Marco Iodice | 62251f7 | 2019-03-11 16:07:12 +0000 | [diff] [blame] | 182 | return Status{}; |
| 183 | } |
| 184 | |
Michele Di Giorgio | b54ba28 | 2020-01-14 15:31:55 +0000 | [diff] [blame] | 185 | std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input0, ITensorInfo *input1, ITensorInfo *output, const GEMMKernelInfo &gemm_info, |
| 186 | ITensorInfo *vector_sum_col, ITensorInfo *vector_sum_row, ITensorInfo *bias, |
| 187 | ITensorInfo *output_multipliers, ITensorInfo *output_shifts, ElementsProcessed &num_elements_processed) |
Gian Marco Iodice | 62251f7 | 2019-03-11 16:07:12 +0000 | [diff] [blame] | 188 | { |
Michele Di Giorgio | b54ba28 | 2020-01-14 15:31:55 +0000 | [diff] [blame] | 189 | const GEMMLowpOutputStageInfo output_stage = gemm_info.output_stage; |
| 190 | |
Gian Marco Iodice | 62251f7 | 2019-03-11 16:07:12 +0000 | [diff] [blame] | 191 | unsigned int &num_elems_processed_per_iteration_x = num_elements_processed[0]; |
| 192 | unsigned int &num_elems_processed_per_iteration_y = num_elements_processed[1]; |
Michele Di Giorgio | b54ba28 | 2020-01-14 15:31:55 +0000 | [diff] [blame] | 193 | bool reinterpret_input_as_3d = gemm_info.reinterpret_input_as_3d; |
| 194 | bool reinterpret_output_as_3d = (gemm_info.depth_output_gemm3d != 0); |
Gian Marco Iodice | 62251f7 | 2019-03-11 16:07:12 +0000 | [diff] [blame] | 195 | |
| 196 | Window win{}; |
| 197 | Window win_out{}; |
| 198 | bool window_changed = false; |
| 199 | |
| 200 | // In case both input and output have to be reinterpreted as 3D tensors, |
| 201 | // force reinterpret_input_as_3d and reinterpret_output_as_3d to be false. |
Gian Marco Iodice | 86cfffe | 2019-04-02 11:02:20 +0100 | [diff] [blame] | 202 | if(reinterpret_input_as_3d == reinterpret_output_as_3d) |
| 203 | { |
| 204 | reinterpret_output_as_3d = false; |
| 205 | } |
Gian Marco Iodice | 62251f7 | 2019-03-11 16:07:12 +0000 | [diff] [blame] | 206 | |
| 207 | // Output tensor auto initialization if not yet initialized |
Michele Di Giorgio | b54ba28 | 2020-01-14 15:31:55 +0000 | [diff] [blame] | 208 | const TensorShape expected_output_shape = compute_mm_shape(*input0, *input1, gemm_info); |
| 209 | if(output_stage.type != GEMMLowpOutputStageType::NONE) |
| 210 | { |
| 211 | auto_init_if_empty(*output, input0->clone()->set_tensor_shape(expected_output_shape).set_data_type(output_stage.output_data_type)); |
| 212 | } |
| 213 | else |
| 214 | { |
| 215 | auto_init_if_empty(*output, input0->clone()->set_tensor_shape(expected_output_shape).set_data_type(DataType::S32)); |
| 216 | } |
Gian Marco Iodice | 62251f7 | 2019-03-11 16:07:12 +0000 | [diff] [blame] | 217 | |
| 218 | TensorInfo tmp_info(*output); |
| 219 | |
| 220 | if(reinterpret_output_as_3d) |
| 221 | { |
| 222 | // Since the output tensor has to be reinterpreted as 3D and the execute window is based on a 2D GEMM, |
| 223 | // the window needs to be constructed on the 2D collapsed version of the tensor |
| 224 | TensorShape tmp_shape(output->tensor_shape()); |
| 225 | tmp_shape.collapse(2U, 1U); |
| 226 | tmp_info.set_tensor_shape(tmp_shape); |
| 227 | } |
| 228 | |
| 229 | // Configure kernel window |
Michele Di Giorgio | b54ba28 | 2020-01-14 15:31:55 +0000 | [diff] [blame] | 230 | num_elems_processed_per_iteration_x = gemm_info.rhs_info.n0; |
| 231 | num_elems_processed_per_iteration_y = gemm_info.lhs_info.m0; |
Gian Marco Iodice | 62251f7 | 2019-03-11 16:07:12 +0000 | [diff] [blame] | 232 | |
| 233 | // Note: bottom paddings are calculated manually as the output can be reinterpreted as 3D tensor |
| 234 | // The only way to set properly the paddings, it is to set those explicitly through the AccessWindowStatic |
Michele Di Giorgio | b54ba28 | 2020-01-14 15:31:55 +0000 | [diff] [blame] | 235 | const int m = reinterpret_output_as_3d ? gemm_info.m : input0->dimension(1); |
Gian Marco Iodice | 62251f7 | 2019-03-11 16:07:12 +0000 | [diff] [blame] | 236 | const int bottom_pad = (num_elems_processed_per_iteration_y - (m % num_elems_processed_per_iteration_y)) % num_elems_processed_per_iteration_y; |
| 237 | |
| 238 | win = calculate_max_window(tmp_info, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y)); |
| 239 | win_out = calculate_max_window(*output, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y)); |
| 240 | |
| 241 | AccessWindowStatic input0_access(input0, 0, 0, |
Michele Di Giorgio | b54ba28 | 2020-01-14 15:31:55 +0000 | [diff] [blame] | 242 | ceil_to_multiple(input0->dimension(0), gemm_info.lhs_info.k0), |
Gian Marco Iodice | 62251f7 | 2019-03-11 16:07:12 +0000 | [diff] [blame] | 243 | input0->dimension(1) + bottom_pad); |
| 244 | AccessWindowStatic input1_access(input1, 0, 0, |
| 245 | input1->dimension(0), |
| 246 | input1->dimension(1)); |
| 247 | AccessWindowStatic output_access(output, 0, 0, |
| 248 | ceil_to_multiple(output->dimension(0), num_elems_processed_per_iteration_x), |
| 249 | output->dimension(1) + bottom_pad); |
| 250 | |
| 251 | window_changed = update_window_and_padding(win, input0_access, input1_access) || // window used by the execute_window_loop |
| 252 | update_window_and_padding(win_out, output_access); // window used to update the padding requirements of output tensor |
| 253 | |
Michele Di Giorgio | b54ba28 | 2020-01-14 15:31:55 +0000 | [diff] [blame] | 254 | if(output_stage.type == GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT) |
| 255 | { |
| 256 | if(gemm_info.a_offset != 0) |
| 257 | { |
| 258 | AccessWindowHorizontal vector_sum_col_access(vector_sum_col, 0, num_elems_processed_per_iteration_x); |
| 259 | window_changed = window_changed || update_window_and_padding(win_out, vector_sum_col_access); |
| 260 | } |
| 261 | // No access window needed for vector_sum_row |
| 262 | ARM_COMPUTE_UNUSED(vector_sum_row); |
| 263 | |
| 264 | if(bias != nullptr) |
| 265 | { |
| 266 | AccessWindowHorizontal bias_access(bias, 0, num_elems_processed_per_iteration_x); |
| 267 | window_changed = window_changed || update_window_and_padding(win_out, bias_access); |
| 268 | } |
| 269 | |
| 270 | if(output_multipliers != nullptr && output_multipliers->dimension(0) > 1) |
| 271 | { |
| 272 | AccessWindowHorizontal output_multipliers_access(output_multipliers, 0, num_elems_processed_per_iteration_x); |
| 273 | AccessWindowHorizontal output_shifts_access(output_shifts, 0, num_elems_processed_per_iteration_x); |
| 274 | window_changed = window_changed || update_window_and_padding(win_out, output_multipliers_access, output_shifts_access); |
| 275 | } |
| 276 | } |
| 277 | |
Gian Marco Iodice | 2ec6c1e | 2019-04-09 12:03:05 +0100 | [diff] [blame] | 278 | output_access.set_valid_region(win_out, ValidRegion(Coordinates(), output->tensor_shape())); |
Gian Marco Iodice | 62251f7 | 2019-03-11 16:07:12 +0000 | [diff] [blame] | 279 | |
| 280 | // Collapse along the Z direction |
| 281 | // This collapse needs to be here in order to tune the Z dimension of LWS |
| 282 | Window collapsed = win; |
| 283 | const unsigned int dimension_to_collapse = std::min(static_cast<unsigned int>(output->num_dimensions()), 2u); |
| 284 | collapsed = win.collapse(win, dimension_to_collapse); |
| 285 | |
| 286 | Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{}; |
| 287 | return std::make_pair(err, collapsed); |
| 288 | } |
| 289 | } // namespace |
| 290 | |
| 291 | CLGEMMLowpMatrixMultiplyReshapedOnlyRHSKernel::CLGEMMLowpMatrixMultiplyReshapedOnlyRHSKernel() |
Michele Di Giorgio | b54ba28 | 2020-01-14 15:31:55 +0000 | [diff] [blame] | 292 | : _input0(nullptr), |
| 293 | _input1(nullptr), |
| 294 | _output(nullptr), |
| 295 | _vector_sum_col(nullptr), |
| 296 | _vector_sum_row(nullptr), |
| 297 | _bias(nullptr), |
| 298 | _output_multipliers(nullptr), |
| 299 | _output_shifts(nullptr), |
| 300 | _slide_matrix_b(true), |
| 301 | _reinterpret_input_as_3d(false), |
| 302 | _reinterpret_output_as_3d(false), |
| 303 | _use_dummy_work_items(false), |
| 304 | _is_quantized_per_channel(false), |
| 305 | _fuse_output_stage(false) |
Gian Marco Iodice | 62251f7 | 2019-03-11 16:07:12 +0000 | [diff] [blame] | 306 | { |
| 307 | } |
| 308 | |
Michele Di Giorgio | b54ba28 | 2020-01-14 15:31:55 +0000 | [diff] [blame] | 309 | void CLGEMMLowpMatrixMultiplyReshapedOnlyRHSKernel::configure(const ICLTensor *input0, const ICLTensor *input1, ICLTensor *output, const GEMMKernelInfo &gemm_info, |
| 310 | const ICLTensor *vector_sum_col, const ICLTensor *vector_sum_row, const ICLTensor *bias, |
| 311 | const ICLTensor *output_multipliers, const ICLTensor *output_shifts) |
Gian Marco Iodice | 62251f7 | 2019-03-11 16:07:12 +0000 | [diff] [blame] | 312 | { |
| 313 | ARM_COMPUTE_ERROR_ON_NULLPTR(input0, input1, output); |
Michele Di Giorgio | b54ba28 | 2020-01-14 15:31:55 +0000 | [diff] [blame] | 314 | ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input0->info(), |
| 315 | input1->info(), |
| 316 | output->info(), |
| 317 | gemm_info, |
| 318 | vector_sum_col != nullptr ? vector_sum_col->info() : nullptr, |
| 319 | vector_sum_row != nullptr ? vector_sum_row->info() : nullptr, |
| 320 | bias != nullptr ? bias->info() : nullptr, |
| 321 | output_multipliers != nullptr ? output_multipliers->info() : nullptr, |
| 322 | output_shifts != nullptr ? output_shifts->info() : nullptr)); |
Gian Marco Iodice | 62251f7 | 2019-03-11 16:07:12 +0000 | [diff] [blame] | 323 | |
Michele Di Giorgio | b54ba28 | 2020-01-14 15:31:55 +0000 | [diff] [blame] | 324 | const GEMMRHSMatrixInfo rhs_info = gemm_info.rhs_info; |
| 325 | const GEMMLHSMatrixInfo lhs_info = gemm_info.lhs_info; |
| 326 | const GEMMLowpOutputStageInfo output_stage = gemm_info.output_stage; |
| 327 | const int32_t a_offset = gemm_info.a_offset; |
| 328 | const int32_t b_offset = gemm_info.b_offset; |
Gian Marco Iodice | 62251f7 | 2019-03-11 16:07:12 +0000 | [diff] [blame] | 329 | |
| 330 | _input0 = input0; |
| 331 | _input1 = input1; |
| 332 | _output = output; |
Michele Di Giorgio | b54ba28 | 2020-01-14 15:31:55 +0000 | [diff] [blame] | 333 | _vector_sum_col = vector_sum_col; |
| 334 | _vector_sum_row = vector_sum_row; |
| 335 | _bias = bias; |
| 336 | _output_multipliers = output_multipliers; |
| 337 | _output_shifts = output_shifts; |
| 338 | _reinterpret_input_as_3d = gemm_info.reinterpret_input_as_3d; |
| 339 | _reinterpret_output_as_3d = (gemm_info.depth_output_gemm3d != 0); |
Gian Marco Iodice | 86cfffe | 2019-04-02 11:02:20 +0100 | [diff] [blame] | 340 | _use_dummy_work_items = preferred_dummy_work_items_support(CLKernelLibrary::get().get_device()); |
Michele Di Giorgio | b54ba28 | 2020-01-14 15:31:55 +0000 | [diff] [blame] | 341 | _is_quantized_per_channel = output_stage.is_quantized_per_channel; |
Gian Marco Iodice | 62251f7 | 2019-03-11 16:07:12 +0000 | [diff] [blame] | 342 | |
| 343 | // In case both input and output have to be reinterpreted as 3D tensors, |
| 344 | // force reinterpret_input_as_3d and reinterpret_output_as_3d to be false. |
Gian Marco Iodice | 86cfffe | 2019-04-02 11:02:20 +0100 | [diff] [blame] | 345 | if(_reinterpret_input_as_3d == _reinterpret_output_as_3d) |
| 346 | { |
| 347 | _reinterpret_input_as_3d = false; |
| 348 | _reinterpret_output_as_3d = false; |
| 349 | } |
Gian Marco Iodice | 62251f7 | 2019-03-11 16:07:12 +0000 | [diff] [blame] | 350 | |
| 351 | // Check if we need to slide the matrix B |
| 352 | const unsigned int num_dimensions_input0 = _input0->info()->num_dimensions(); |
| 353 | _slide_matrix_b = (_input1->info()->num_dimensions() >= num_dimensions_input0); |
| 354 | |
| 355 | ElementsProcessed num_elements_processed{}; |
| 356 | |
| 357 | // Configure kernel window |
Michele Di Giorgio | b54ba28 | 2020-01-14 15:31:55 +0000 | [diff] [blame] | 358 | auto win_config = validate_and_configure_window(input0->info(), |
| 359 | input1->info(), |
| 360 | output->info(), |
| 361 | gemm_info, |
| 362 | vector_sum_col != nullptr ? vector_sum_col->info() : nullptr, |
| 363 | vector_sum_row != nullptr ? vector_sum_row->info() : nullptr, |
| 364 | bias != nullptr ? bias->info() : nullptr, |
| 365 | output_multipliers != nullptr ? output_multipliers->info() : nullptr, |
| 366 | output_shifts != nullptr ? output_shifts->info() : nullptr, |
| 367 | num_elements_processed); |
Gian Marco Iodice | 62251f7 | 2019-03-11 16:07:12 +0000 | [diff] [blame] | 368 | ARM_COMPUTE_ERROR_THROW_ON(win_config.first); |
| 369 | ICLKernel::configure_internal(win_config.second); |
| 370 | |
| 371 | // Create build options |
| 372 | CLBuildOptions build_opts; |
| 373 | build_opts.add_option_if(_reinterpret_input_as_3d, "-DREINTERPRET_INPUT_AS_3D"); |
| 374 | build_opts.add_option_if(_reinterpret_output_as_3d, "-DREINTERPRET_OUTPUT_AS_3D"); |
| 375 | build_opts.add_option_if(_reinterpret_input_as_3d || _reinterpret_output_as_3d, "-DHEIGHT_GEMM3D=" + support::cpp11::to_string(output->info()->dimension(1))); |
| 376 | build_opts.add_option_if(_reinterpret_input_as_3d || _reinterpret_output_as_3d, "-DDEPTH_GEMM3D=" + support::cpp11::to_string(output->info()->dimension(2))); |
| 377 | build_opts.add_option_if(!_slide_matrix_b, "-DMATRIX_B_DEPTH=" + support::cpp11::to_string(input1->info()->dimension(2))); |
| 378 | build_opts.add_option_if(rhs_info.interleave, "-DRHS_INTERLEAVE"); |
Gian Marco Iodice | 86cfffe | 2019-04-02 11:02:20 +0100 | [diff] [blame] | 379 | build_opts.add_option_if(_use_dummy_work_items, "-DDUMMY_WORK_ITEMS"); |
| 380 | build_opts.add_option("-DM=" + support::cpp11::to_string(input0->info()->dimension(1))); |
Michele Di Giorgio | b54ba28 | 2020-01-14 15:31:55 +0000 | [diff] [blame] | 381 | build_opts.add_option("-DN=" + support::cpp11::to_string(gemm_info.n)); |
| 382 | build_opts.add_option("-DK=" + support::cpp11::to_string(gemm_info.k)); |
Gian Marco Iodice | 62251f7 | 2019-03-11 16:07:12 +0000 | [diff] [blame] | 383 | build_opts.add_option("-DM0=" + support::cpp11::to_string(lhs_info.m0)); |
| 384 | build_opts.add_option("-DN0=" + support::cpp11::to_string(rhs_info.n0)); |
| 385 | build_opts.add_option("-DK0=" + support::cpp11::to_string(rhs_info.k0)); |
| 386 | build_opts.add_option("-DH0=" + support::cpp11::to_string(rhs_info.h0)); |
Michele Di Giorgio | f9179d3 | 2019-11-27 16:17:30 +0000 | [diff] [blame] | 387 | build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(input0->info()->data_type())); |
| 388 | build_opts.add_option("-DACC_DATA_TYPE=" + get_cl_dot8_acc_type_from_data_type(input0->info()->data_type())); |
Gian Marco Iodice | 62251f7 | 2019-03-11 16:07:12 +0000 | [diff] [blame] | 389 | |
| 390 | std::string kernel_name("gemmlowp_mm_reshaped_only_rhs_"); |
| 391 | kernel_name += rhs_info.transpose ? "t" : "nt"; |
| 392 | |
Michele Di Giorgio | b54ba28 | 2020-01-14 15:31:55 +0000 | [diff] [blame] | 393 | if(output_stage.type == GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT) |
| 394 | { |
| 395 | kernel_name += "_fused_output_stage_fixedpoint"; |
| 396 | _fuse_output_stage = true; |
| 397 | // If a_offset == 0, vector_sum_col can be a nullptr |
| 398 | if(a_offset != 0) |
| 399 | { |
| 400 | build_opts.add_option("-DA_OFFSET=" + support::cpp11::to_string(a_offset)); |
| 401 | build_opts.add_option_if(vector_sum_col->info()->tensor_shape().num_dimensions() > 1, "-DSUM_COL_HAS_BATCHES"); |
| 402 | } |
| 403 | // If b_offset == 0, vector_sum_row can be a nullptr |
| 404 | build_opts.add_option_if(b_offset != 0, "-DB_OFFSET=" + support::cpp11::to_string(b_offset)); |
| 405 | build_opts.add_option("-DK_OFFSET=" + support::cpp11::to_string(a_offset * b_offset * input0->info()->dimension(0))); |
| 406 | build_opts.add_option_if(bias != nullptr, "-DADD_BIAS"); |
| 407 | build_opts.add_option("-DRESULT_OFFSET=" + support::cpp11::to_string(output_stage.gemmlowp_offset)); |
| 408 | build_opts.add_option("-DRESULT_MULTIPLIER=" + support::cpp11::to_string(output_stage.gemmlowp_multipliers[0])); |
| 409 | build_opts.add_option("-DRESULT_SHIFT=" + support::cpp11::to_string(output_stage.gemmlowp_shifts[0])); |
| 410 | build_opts.add_option_if(_is_quantized_per_channel, "-DPER_CHANNEL_QUANTIZATION"); |
| 411 | |
| 412 | const int min = output_stage.gemmlowp_min_bound; |
| 413 | const int max = output_stage.gemmlowp_max_bound; |
| 414 | |
| 415 | PixelValue min_val{}; |
| 416 | PixelValue max_val{}; |
| 417 | std::tie(min_val, max_val) = get_min_max(output->info()->data_type()); |
Michele Di Giorgio | 398b8e4 | 2020-03-06 13:56:54 +0000 | [diff] [blame^] | 418 | build_opts.add_option_if(min != min_val.get<int32_t>(), "-DMIN_BOUND=" + support::cpp11::to_string(min)); |
| 419 | build_opts.add_option_if(max != max_val.get<int32_t>(), "-DMAX_BOUND=" + support::cpp11::to_string(max)); |
Michele Di Giorgio | b54ba28 | 2020-01-14 15:31:55 +0000 | [diff] [blame] | 420 | } |
| 421 | |
Gian Marco Iodice | 62251f7 | 2019-03-11 16:07:12 +0000 | [diff] [blame] | 422 | // Create kernel |
| 423 | _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel(kernel_name, build_opts.options())); |
| 424 | |
| 425 | // Set config_id for enabling LWS tuning |
| 426 | _config_id = kernel_name; |
| 427 | _config_id += "_"; |
Gian Marco Iodice | 43a129e | 2019-05-14 10:14:08 +0100 | [diff] [blame] | 428 | _config_id += dot8_supported(CLKernelLibrary::get().get_device()) ? "_dot8" : ""; |
| 429 | _config_id += "_"; |
Gian Marco Iodice | 62251f7 | 2019-03-11 16:07:12 +0000 | [diff] [blame] | 430 | _config_id += (_reinterpret_input_as_3d ? "3di_" : ""); |
| 431 | _config_id += (_reinterpret_output_as_3d ? "3do_" : ""); |
| 432 | _config_id += support::cpp11::to_string(output->info()->dimension(1)); |
| 433 | _config_id += "_"; |
| 434 | _config_id += support::cpp11::to_string(output->info()->dimension(0)); |
| 435 | _config_id += "_"; |
Michele Di Giorgio | b54ba28 | 2020-01-14 15:31:55 +0000 | [diff] [blame] | 436 | _config_id += support::cpp11::to_string(gemm_info.k); |
Gian Marco Iodice | 62251f7 | 2019-03-11 16:07:12 +0000 | [diff] [blame] | 437 | _config_id += "_"; |
| 438 | _config_id += support::cpp11::to_string(output->info()->dimension(2)); |
| 439 | _config_id += "_"; |
| 440 | _config_id += support::cpp11::to_string(lhs_info.m0); |
| 441 | _config_id += "_"; |
| 442 | _config_id += support::cpp11::to_string(rhs_info.n0); |
| 443 | _config_id += "_"; |
| 444 | _config_id += support::cpp11::to_string(rhs_info.k0); |
| 445 | _config_id += "_"; |
| 446 | _config_id += support::cpp11::to_string(rhs_info.h0); |
| 447 | _config_id += "_"; |
| 448 | _config_id += support::cpp11::to_string(rhs_info.interleave); |
| 449 | } |
| 450 | |
Michele Di Giorgio | b54ba28 | 2020-01-14 15:31:55 +0000 | [diff] [blame] | 451 | Status CLGEMMLowpMatrixMultiplyReshapedOnlyRHSKernel::validate(const ITensorInfo *input0, const ITensorInfo *input1, const ITensorInfo *output, const GEMMKernelInfo &gemm_info, |
| 452 | const ITensorInfo *vector_sum_col, const ITensorInfo *vector_sum_row, const ITensorInfo *bias, |
| 453 | const ITensorInfo *output_multipliers, const ITensorInfo *output_shifts) |
Gian Marco Iodice | 62251f7 | 2019-03-11 16:07:12 +0000 | [diff] [blame] | 454 | { |
| 455 | ElementsProcessed num_elements_processed{}; |
Michele Di Giorgio | b54ba28 | 2020-01-14 15:31:55 +0000 | [diff] [blame] | 456 | ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input0, input1, output, gemm_info, vector_sum_col, vector_sum_row, bias, output_multipliers, output_shifts)); |
Gian Marco Iodice | 62251f7 | 2019-03-11 16:07:12 +0000 | [diff] [blame] | 457 | ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input0->clone().get(), |
| 458 | input1->clone().get(), |
| 459 | output->clone().get(), |
Gian Marco Iodice | 62251f7 | 2019-03-11 16:07:12 +0000 | [diff] [blame] | 460 | gemm_info, |
Michele Di Giorgio | b54ba28 | 2020-01-14 15:31:55 +0000 | [diff] [blame] | 461 | vector_sum_col != nullptr ? vector_sum_col->clone().get() : nullptr, |
| 462 | vector_sum_row != nullptr ? vector_sum_row->clone().get() : nullptr, |
| 463 | bias != nullptr ? bias->clone().get() : nullptr, |
| 464 | output_multipliers != nullptr ? output_multipliers->clone().get() : nullptr, |
| 465 | output_shifts != nullptr ? output_shifts->clone().get() : nullptr, |
Gian Marco Iodice | 62251f7 | 2019-03-11 16:07:12 +0000 | [diff] [blame] | 466 | num_elements_processed) |
| 467 | .first); |
| 468 | |
| 469 | return Status{}; |
| 470 | } |
| 471 | |
| 472 | void CLGEMMLowpMatrixMultiplyReshapedOnlyRHSKernel::run(const Window &window, cl::CommandQueue &queue) |
| 473 | { |
| 474 | ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); |
| 475 | ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window); |
| 476 | |
| 477 | if(_input1->info()->num_dimensions() < 3) |
| 478 | { |
| 479 | // The stride_z for matrix B must be zero if we do not slice |
| 480 | ARM_COMPUTE_ERROR_ON(_input1->info()->strides_in_bytes()[3] != 0); |
| 481 | } |
| 482 | |
| 483 | Window slice = window.first_slice_window_3D(); |
| 484 | Window slice_matrix_b = slice; |
| 485 | |
| 486 | slice_matrix_b.set(Window::DimX, Window::Dimension(0, 1, 1)); |
| 487 | slice_matrix_b.set(Window::DimY, Window::Dimension(0, 1, 1)); |
| 488 | |
| 489 | if(_reinterpret_input_as_3d) |
| 490 | { |
| 491 | // Pass bottom paddings to the kernel if the input has to be reinterpreted as 3D tensor |
| 492 | const unsigned int idx0 = 3 * num_arguments_per_2D_tensor() + 3; |
| 493 | const unsigned int total_cross_plane_pad = _input0->info()->padding().top + _input0->info()->padding().bottom; |
| 494 | _kernel.setArg<cl_uint>(idx0, static_cast<unsigned int>(total_cross_plane_pad)); |
| 495 | } |
| 496 | |
| 497 | if(_reinterpret_output_as_3d) |
| 498 | { |
| 499 | // Pass bottom paddings to the kernel if the output has to be reinterpreted as 3D tensor |
| 500 | const unsigned int idx0 = 3 * num_arguments_per_2D_tensor() + 3 + (_reinterpret_input_as_3d ? 1 : 0); |
| 501 | const unsigned int total_cross_plane_pad = _output->info()->padding().top + _output->info()->padding().bottom; |
| 502 | _kernel.setArg<cl_uint>(idx0, static_cast<unsigned int>(total_cross_plane_pad)); |
| 503 | } |
| 504 | |
Michele Di Giorgio | b54ba28 | 2020-01-14 15:31:55 +0000 | [diff] [blame] | 505 | // Set window for vector_sum_col |
| 506 | Window win_vector_sum_col = slice; |
| 507 | win_vector_sum_col.set(Window::DimY, Window::Dimension(0, 0, 0)); |
| 508 | win_vector_sum_col.set(Window::DimZ, Window::Dimension(0, 0, 0)); |
| 509 | |
| 510 | // Set window for vector_sum_row |
| 511 | Window win_vector_sum_row = slice; |
| 512 | win_vector_sum_row.set(Window::DimX, Window::Dimension(0, 0, 0)); |
| 513 | win_vector_sum_row.set(Window::DimY, Window::Dimension(0, 0, 0)); |
| 514 | win_vector_sum_col.set(Window::DimZ, Window::Dimension(0, 0, 0)); |
| 515 | |
| 516 | Window biases_slice = slice; |
| 517 | biases_slice.set(Window::DimY, Window::Dimension(0, 1, 1)); |
| 518 | biases_slice.set(Window::DimZ, Window::Dimension(0, 1, 1)); |
| 519 | |
Gian Marco Iodice | 62251f7 | 2019-03-11 16:07:12 +0000 | [diff] [blame] | 520 | do |
| 521 | { |
| 522 | Window slice_b = slice; |
| 523 | // Don't slice matrix B along the z dimension if matrix B has just 2 dimensions and matrix A more than 2 |
| 524 | // This scenario can happen when the matrix multiplication is used to perform a convolution operation |
| 525 | if(!_slide_matrix_b) |
| 526 | { |
| 527 | slice_b = slice_matrix_b; |
| 528 | } |
| 529 | |
| 530 | unsigned int idx = 0; |
| 531 | add_2D_tensor_argument(idx, _input0, slice); |
| 532 | add_2D_tensor_argument(idx, _input1, slice_b); |
| 533 | add_2D_tensor_argument(idx, _output, slice); |
| 534 | _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_input0->info()->strides_in_bytes()[2])); |
| 535 | _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_input1->info()->strides_in_bytes()[2])); |
| 536 | _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_output->info()->strides_in_bytes()[2])); |
Michele Di Giorgio | b54ba28 | 2020-01-14 15:31:55 +0000 | [diff] [blame] | 537 | if(_reinterpret_input_as_3d) |
| 538 | { |
| 539 | // Pass bottom paddings to the kernel if the input has to be reinterpreted as 3D tensor |
| 540 | idx++; |
| 541 | } |
| 542 | |
| 543 | if(_reinterpret_output_as_3d) |
| 544 | { |
| 545 | // Pass bottom paddings to the kernel if the output has to be reinterpreted as 3D tensor |
| 546 | idx++; |
| 547 | } |
| 548 | |
| 549 | if(_fuse_output_stage) |
| 550 | { |
| 551 | add_2D_tensor_argument_if((_vector_sum_col != nullptr), idx, _vector_sum_col, win_vector_sum_col); |
| 552 | add_2D_tensor_argument_if((_vector_sum_row != nullptr), idx, _vector_sum_row, win_vector_sum_row); |
| 553 | add_1D_tensor_argument_if((_bias != nullptr), idx, _bias, biases_slice); |
| 554 | add_1D_tensor_argument_if(_is_quantized_per_channel, idx, _output_multipliers, biases_slice); |
| 555 | add_1D_tensor_argument_if(_is_quantized_per_channel, idx, _output_shifts, biases_slice); |
| 556 | } |
Gian Marco Iodice | 86cfffe | 2019-04-02 11:02:20 +0100 | [diff] [blame] | 557 | enqueue(queue, *this, slice, lws_hint(), _use_dummy_work_items); |
Gian Marco Iodice | 62251f7 | 2019-03-11 16:07:12 +0000 | [diff] [blame] | 558 | } |
| 559 | while(window.slide_window_slice_3D(slice)); |
| 560 | } |
giuros01 | 46a49a0 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 561 | } // namespace arm_compute |