giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 1 | /* |
| 2 | * Copyright (c) 2019 ARM Limited. |
| 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/CLGEMMMatrixMultiplyNativeKernel.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 | |
| 53 | Status validate_arguments(const ITensorInfo *input0, const ITensorInfo *input1, const ITensorInfo *output, float alpha, const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info, |
| 54 | const GEMMReshapeInfo &gemm_info) |
| 55 | { |
| 56 | ARM_COMPUTE_UNUSED(alpha); |
| 57 | ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input0, input1, output); |
| 58 | ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input0, 1, DataType::F32, DataType::F16); |
| 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"); |
| 62 | 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"); |
| 63 | ARM_COMPUTE_RETURN_ERROR_ON(rhs_info.k0 > 16); |
| 64 | ARM_COMPUTE_RETURN_ERROR_ON(lhs_info.m0 < 1 || lhs_info.m0 > 8); |
| 65 | 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"); |
| 66 | |
| 67 | const int m = gemm_info.m(); |
| 68 | const int n = gemm_info.n(); |
| 69 | const int k = gemm_info.k(); |
| 70 | |
| 71 | ARM_COMPUTE_UNUSED(m); |
| 72 | ARM_COMPUTE_UNUSED(n); |
| 73 | ARM_COMPUTE_UNUSED(k); |
| 74 | |
| 75 | ARM_COMPUTE_RETURN_ERROR_ON(input0->dimension(0) != static_cast<unsigned int>(k)); |
| 76 | ARM_COMPUTE_RETURN_ERROR_ON(input1->dimension(0) != static_cast<unsigned int>(n)); |
| 77 | ARM_COMPUTE_RETURN_ERROR_ON(input1->dimension(1) != static_cast<unsigned int>(k)); |
| 78 | if(gemm_info.reinterpret_input_as_3d()) |
| 79 | { |
| 80 | ARM_COMPUTE_RETURN_ERROR_ON(input0->dimension(1) * input0->dimension(2) != static_cast<unsigned int>(m)); |
| 81 | } |
| 82 | else |
| 83 | { |
| 84 | ARM_COMPUTE_RETURN_ERROR_ON(input0->dimension(1) != static_cast<unsigned int>(m)); |
| 85 | } |
| 86 | |
| 87 | if(output->total_size() != 0) |
| 88 | { |
| 89 | const TensorInfo tensor_info_output = output->clone()->set_tensor_shape(compute_mm_shape(*input0, *input1, gemm_info)); |
| 90 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(output, &tensor_info_output); |
| 91 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input0, output); |
| 92 | } |
| 93 | |
| 94 | return Status{}; |
| 95 | } |
| 96 | |
| 97 | std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input0, ITensorInfo *input1, ITensorInfo *output, const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info, |
| 98 | const GEMMReshapeInfo &gemm_info, ElementsProcessed &num_elements_processed) |
| 99 | { |
| 100 | unsigned int &num_elems_processed_per_iteration_x = num_elements_processed[0]; |
| 101 | unsigned int &num_elems_processed_per_iteration_y = num_elements_processed[1]; |
| 102 | bool reinterpret_input_as_3d = gemm_info.reinterpret_input_as_3d(); |
| 103 | bool reinterpret_output_as_3d = (gemm_info.depth_output_gemm3d() != 0); |
| 104 | |
| 105 | Window win{}; |
| 106 | Window win_out{}; |
| 107 | bool window_changed = false; |
| 108 | |
| 109 | // In case both input and output have to be reinterpreted as 3D tensors, |
| 110 | // force reinterpret_input_as_3d and reinterpret_output_as_3d to be false. |
| 111 | if(reinterpret_input_as_3d == reinterpret_output_as_3d) |
| 112 | { |
| 113 | reinterpret_output_as_3d = false; |
| 114 | } |
| 115 | |
| 116 | // Output tensor auto initialization if not yet initialized |
| 117 | auto_init_if_empty(*output, input0->clone()->set_tensor_shape(compute_mm_shape(*input0, *input1, gemm_info))); |
| 118 | |
| 119 | TensorInfo tmp_info(*output); |
| 120 | |
| 121 | if(reinterpret_output_as_3d) |
| 122 | { |
| 123 | // Since the output tensor has to be reinterpreted as 3D and the execute window is based on a 2D GEMM, |
| 124 | // the window needs to be constructed on the 2D collapsed version of the tensor |
| 125 | TensorShape tmp_shape(output->tensor_shape()); |
| 126 | tmp_shape.collapse(2U, 1U); |
| 127 | tmp_info.set_tensor_shape(tmp_shape); |
| 128 | } |
| 129 | |
| 130 | // Configure kernel window |
| 131 | num_elems_processed_per_iteration_x = rhs_info.n0; |
| 132 | num_elems_processed_per_iteration_y = lhs_info.m0; |
| 133 | |
| 134 | // Note: bottom paddings are calculated manually as the output can be reinterpreted as 3D tensor |
| 135 | // The only way to set properly the paddings, it is to set those explicitly through the AccessWindowStatic |
Gian Marco Iodice | c625acd | 2019-06-04 12:39:23 +0100 | [diff] [blame] | 136 | const int m = reinterpret_output_as_3d ? gemm_info.m() : input0->dimension(1); |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 137 | const int bottom_pad = (num_elems_processed_per_iteration_y - (m % num_elems_processed_per_iteration_y)) % num_elems_processed_per_iteration_y; |
| 138 | |
| 139 | win = calculate_max_window(tmp_info, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y)); |
| 140 | win_out = calculate_max_window(*output, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y)); |
| 141 | |
| 142 | AccessWindowStatic input0_access(input0, 0, 0, |
| 143 | input0->dimension(0), |
| 144 | input0->dimension(1) + bottom_pad); |
| 145 | AccessWindowStatic input1_access(input1, 0, 0, |
| 146 | ceil_to_multiple(input1->dimension(0), num_elems_processed_per_iteration_x), |
| 147 | input1->dimension(1)); |
| 148 | AccessWindowStatic output_access(output, 0, 0, |
| 149 | ceil_to_multiple(output->dimension(0), num_elems_processed_per_iteration_x), |
| 150 | output->dimension(1) + bottom_pad); |
| 151 | |
| 152 | window_changed = update_window_and_padding(win, input0_access, input1_access) || // window used by the execute_window_loop |
| 153 | update_window_and_padding(win_out, output_access); // window used to update the padding requirements of output tensor |
| 154 | |
| 155 | output_access.set_valid_region(win_out, ValidRegion(Coordinates(), output->tensor_shape())); |
| 156 | |
| 157 | // Collapse along the Z direction |
| 158 | // This collapse needs to be here in order to tune the Z dimension of LWS |
| 159 | Window collapsed = win; |
| 160 | const unsigned int dimension_to_collapse = std::min(static_cast<unsigned int>(output->num_dimensions()), 2u); |
| 161 | collapsed = win.collapse(win, dimension_to_collapse); |
| 162 | |
| 163 | Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{}; |
| 164 | return std::make_pair(err, collapsed); |
| 165 | } |
| 166 | } // namespace |
| 167 | |
| 168 | CLGEMMMatrixMultiplyNativeKernel::CLGEMMMatrixMultiplyNativeKernel() |
| 169 | : _input0(nullptr), _input1(nullptr), _output(nullptr), _slide_matrix_b(true), _reinterpret_input_as_3d(false), _reinterpret_output_as_3d(false), _use_dummy_work_items(false) |
| 170 | { |
| 171 | } |
| 172 | |
| 173 | void CLGEMMMatrixMultiplyNativeKernel::configure(const ICLTensor *input0, const ICLTensor *input1, ICLTensor *output, float alpha, const GEMMLHSMatrixInfo &lhs_info, |
| 174 | const GEMMRHSMatrixInfo &rhs_info, const GEMMReshapeInfo &gemm_info) |
| 175 | { |
| 176 | ARM_COMPUTE_ERROR_ON_NULLPTR(input0, input1, output); |
| 177 | |
| 178 | ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input0->info(), input1->info(), output->info(), alpha, lhs_info, rhs_info, gemm_info)); |
| 179 | |
| 180 | _input0 = input0; |
| 181 | _input1 = input1; |
| 182 | _output = output; |
| 183 | _reinterpret_input_as_3d = gemm_info.reinterpret_input_as_3d(); |
| 184 | _reinterpret_output_as_3d = (gemm_info.depth_output_gemm3d() != 0); |
| 185 | _use_dummy_work_items = preferred_dummy_work_items_support(CLKernelLibrary::get().get_device()); |
| 186 | |
| 187 | // In case both input and output have to be reinterpreted as 3D tensors, |
| 188 | // force reinterpret_input_as_3d and reinterpret_output_as_3d to be false. |
| 189 | if(_reinterpret_input_as_3d == _reinterpret_output_as_3d) |
| 190 | { |
| 191 | _reinterpret_input_as_3d = false; |
| 192 | _reinterpret_output_as_3d = false; |
| 193 | } |
| 194 | |
| 195 | // Check if we need to slide the matrix B |
| 196 | const unsigned int num_dimensions_input0 = _input0->info()->num_dimensions(); |
| 197 | _slide_matrix_b = (_input1->info()->num_dimensions() >= num_dimensions_input0); |
| 198 | |
| 199 | ElementsProcessed num_elements_processed{}; |
| 200 | |
| 201 | // Configure kernel window |
| 202 | auto win_config = validate_and_configure_window(input0->info(), input1->info(), output->info(), lhs_info, rhs_info, gemm_info, num_elements_processed); |
| 203 | ARM_COMPUTE_ERROR_THROW_ON(win_config.first); |
| 204 | ICLKernel::configure_internal(win_config.second); |
| 205 | |
| 206 | // Create build options |
| 207 | CLBuildOptions build_opts; |
| 208 | build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(input0->info()->data_type())); |
| 209 | build_opts.add_option_if(std::abs(1.0f - alpha) > 0.00001f, "-DALPHA=" + float_to_string_with_full_precision(alpha)); |
| 210 | build_opts.add_option_if(_reinterpret_input_as_3d, "-DREINTERPRET_INPUT_AS_3D"); |
| 211 | build_opts.add_option_if(_reinterpret_output_as_3d, "-DREINTERPRET_OUTPUT_AS_3D"); |
| 212 | build_opts.add_option_if(_reinterpret_input_as_3d || _reinterpret_output_as_3d, "-DHEIGHT_GEMM3D=" + support::cpp11::to_string(output->info()->dimension(1))); |
| 213 | build_opts.add_option_if(_reinterpret_input_as_3d || _reinterpret_output_as_3d, "-DDEPTH_GEMM3D=" + support::cpp11::to_string(output->info()->dimension(2))); |
| 214 | build_opts.add_option_if(!_slide_matrix_b, "-DMATRIX_B_DEPTH=" + support::cpp11::to_string(input1->info()->dimension(2))); |
| 215 | build_opts.add_option_if(_use_dummy_work_items, "-DDUMMY_WORK_ITEMS"); |
| 216 | build_opts.add_option("-DM=" + support::cpp11::to_string(input0->info()->dimension(1))); |
| 217 | build_opts.add_option("-DN=" + support::cpp11::to_string(gemm_info.n())); |
| 218 | build_opts.add_option("-DK=" + support::cpp11::to_string(gemm_info.k())); |
| 219 | build_opts.add_option("-DM0=" + support::cpp11::to_string(lhs_info.m0)); |
| 220 | build_opts.add_option("-DN0=" + support::cpp11::to_string(rhs_info.n0)); |
| 221 | build_opts.add_option("-DK0=" + support::cpp11::to_string(rhs_info.k0)); |
| 222 | |
| 223 | std::string kernel_name("gemm_mm_native"); |
| 224 | |
| 225 | // Create kernel |
| 226 | _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel(kernel_name, build_opts.options())); |
| 227 | |
| 228 | // Set config_id for enabling LWS tuning |
| 229 | _config_id = kernel_name; |
| 230 | _config_id += "_"; |
| 231 | _config_id += (_reinterpret_input_as_3d ? "3di_" : ""); |
| 232 | _config_id += (_reinterpret_output_as_3d ? "3do_" : ""); |
| 233 | _config_id += lower_string(string_from_data_type(input0->info()->data_type())); |
| 234 | _config_id += "_"; |
| 235 | _config_id += support::cpp11::to_string(output->info()->dimension(1)); |
| 236 | _config_id += "_"; |
| 237 | _config_id += support::cpp11::to_string(output->info()->dimension(0)); |
| 238 | _config_id += "_"; |
| 239 | _config_id += support::cpp11::to_string(gemm_info.k()); |
| 240 | _config_id += "_"; |
| 241 | _config_id += support::cpp11::to_string(output->info()->dimension(2)); |
| 242 | _config_id += "_"; |
| 243 | _config_id += support::cpp11::to_string(lhs_info.m0); |
| 244 | _config_id += "_"; |
| 245 | _config_id += support::cpp11::to_string(rhs_info.n0); |
| 246 | _config_id += "_"; |
| 247 | _config_id += support::cpp11::to_string(rhs_info.k0); |
| 248 | } |
| 249 | |
| 250 | Status CLGEMMMatrixMultiplyNativeKernel::validate(const ITensorInfo *input0, const ITensorInfo *input1, const ITensorInfo *output, float alpha, const GEMMLHSMatrixInfo &lhs_info, |
| 251 | const GEMMRHSMatrixInfo &rhs_info, const GEMMReshapeInfo &gemm_info) |
| 252 | { |
| 253 | ElementsProcessed num_elements_processed{}; |
| 254 | ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input0, input1, output, alpha, lhs_info, rhs_info, gemm_info)); |
| 255 | ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input0->clone().get(), |
| 256 | input1->clone().get(), |
| 257 | output->clone().get(), |
| 258 | lhs_info, |
| 259 | rhs_info, |
| 260 | gemm_info, |
| 261 | num_elements_processed) |
| 262 | .first); |
| 263 | |
| 264 | return Status{}; |
| 265 | } |
| 266 | |
| 267 | void CLGEMMMatrixMultiplyNativeKernel::run(const Window &window, cl::CommandQueue &queue) |
| 268 | { |
| 269 | ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); |
| 270 | ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window); |
| 271 | |
| 272 | if(_input1->info()->num_dimensions() < 3) |
| 273 | { |
| 274 | // The stride_z for matrix B must be zero if we do not slice |
| 275 | ARM_COMPUTE_ERROR_ON(_input1->info()->strides_in_bytes()[3] != 0); |
| 276 | } |
| 277 | |
| 278 | Window slice = window.first_slice_window_3D(); |
| 279 | Window slice_matrix_b = slice; |
| 280 | |
| 281 | slice_matrix_b.set(Window::DimX, Window::Dimension(0, 1, 1)); |
| 282 | slice_matrix_b.set(Window::DimY, Window::Dimension(0, 1, 1)); |
| 283 | |
| 284 | if(_reinterpret_input_as_3d) |
| 285 | { |
| 286 | // Pass bottom paddings to the kernel if the input has to be reinterpreted as 3D tensor |
| 287 | const unsigned int idx0 = 3 * num_arguments_per_2D_tensor() + 3; |
| 288 | const unsigned int total_cross_plane_pad = _input0->info()->padding().top + _input0->info()->padding().bottom; |
| 289 | _kernel.setArg<cl_uint>(idx0, static_cast<unsigned int>(total_cross_plane_pad)); |
| 290 | } |
| 291 | |
| 292 | if(_reinterpret_output_as_3d) |
| 293 | { |
| 294 | // Pass bottom paddings to the kernel if the output has to be reinterpreted as 3D tensor |
| 295 | const unsigned int idx0 = 3 * num_arguments_per_2D_tensor() + 3 + (_reinterpret_input_as_3d ? 1 : 0); |
| 296 | const unsigned int total_cross_plane_pad = _output->info()->padding().top + _output->info()->padding().bottom; |
| 297 | _kernel.setArg<cl_uint>(idx0, static_cast<unsigned int>(total_cross_plane_pad)); |
| 298 | } |
| 299 | |
| 300 | do |
| 301 | { |
| 302 | Window slice_b = slice; |
| 303 | // Don't slice matrix B along the z dimension if matrix B has just 2 dimensions and matrix A more than 2 |
| 304 | // This scenario can happen when the matrix multiplication is used to perform a convolution operation |
| 305 | if(!_slide_matrix_b) |
| 306 | { |
| 307 | slice_b = slice_matrix_b; |
| 308 | } |
| 309 | |
| 310 | unsigned int idx = 0; |
| 311 | add_2D_tensor_argument(idx, _input0, slice); |
| 312 | add_2D_tensor_argument(idx, _input1, slice_b); |
| 313 | add_2D_tensor_argument(idx, _output, slice); |
| 314 | _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_input0->info()->strides_in_bytes()[2])); |
| 315 | _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_input1->info()->strides_in_bytes()[2])); |
| 316 | _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_output->info()->strides_in_bytes()[2])); |
| 317 | enqueue(queue, *this, slice, lws_hint(), _use_dummy_work_items); |
| 318 | } |
| 319 | while(window.slide_window_slice_3D(slice)); |
| 320 | } |
| 321 | } // namespace arm_compute |