Gian Marco Iodice | bf9731e | 2018-12-12 10:18:04 +0000 | [diff] [blame^] | 1 | /* |
| 2 | * Copyright (c) 2018 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/CLGEMMMatrixMultiplyReshapedKernel.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; |
| 46 | using namespace arm_compute::misc::shape_calculator; |
| 47 | |
| 48 | namespace arm_compute |
| 49 | { |
| 50 | class Coordinates; |
| 51 | } // namespace arm_compute |
| 52 | |
| 53 | namespace |
| 54 | { |
| 55 | using ElementsProcessed = Steps; |
| 56 | |
| 57 | Status validate_arguments(const ITensorInfo *input0, const ITensorInfo *input1, const ITensorInfo *output, float alpha, const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info, |
| 58 | const GEMMReshapeInfo &gemm_info) |
| 59 | { |
| 60 | ARM_COMPUTE_UNUSED(alpha); |
| 61 | ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input0, input1, output); |
| 62 | ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input0, 1, DataType::F32, DataType::F16); |
| 63 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input0, input1); |
| 64 | ARM_COMPUTE_RETURN_ERROR_ON_MSG(input0->num_dimensions() > 4, "The number of dimensions for the LHS matrix must be <= 4"); |
| 65 | ARM_COMPUTE_RETURN_ERROR_ON_MSG(input1->num_dimensions() > 3, "The number of dimensions for the RHS matrix must be <= 3"); |
| 66 | ARM_COMPUTE_RETURN_ERROR_ON(lhs_info.transpose); |
| 67 | ARM_COMPUTE_RETURN_ERROR_ON(!rhs_info.transpose); |
| 68 | ARM_COMPUTE_RETURN_ERROR_ON(lhs_info.k0 != rhs_info.k0); |
| 69 | ARM_COMPUTE_RETURN_ERROR_ON(lhs_info.m0 < 2 || lhs_info.m0 > 8); |
| 70 | |
| 71 | const int m = gemm_info.m(); |
| 72 | const int n = gemm_info.n(); |
| 73 | const int k = gemm_info.k(); |
| 74 | |
| 75 | TensorShape tensor_shape0{ input0->tensor_shape() }; |
| 76 | tensor_shape0.set(0, k); |
| 77 | tensor_shape0.set(1, m); |
| 78 | |
| 79 | TensorShape tensor_shape1{ input1->tensor_shape() }; |
| 80 | tensor_shape1.set(0, n); |
| 81 | tensor_shape1.set(1, k); |
| 82 | |
| 83 | const TensorInfo tensor_info0 = input0->clone()->set_tensor_shape(tensor_shape0); |
| 84 | const TensorInfo tensor_info1 = input1->clone()->set_tensor_shape(tensor_shape1); |
| 85 | |
| 86 | const TensorInfo tensor_info_reshaped0 = input0->clone()->set_tensor_shape(compute_lhs_reshaped_shape(tensor_info0, lhs_info)); |
| 87 | const TensorInfo tensor_info_reshaped1 = input1->clone()->set_tensor_shape(compute_rhs_reshaped_shape(tensor_info1, rhs_info)); |
| 88 | |
| 89 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input0, &tensor_info_reshaped0); |
| 90 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input1, &tensor_info_reshaped1); |
| 91 | |
| 92 | if(output->total_size() != 0) |
| 93 | { |
| 94 | const TensorInfo tensor_info_output = output->clone()->set_tensor_shape(compute_mm_shape(*input0, *input1, gemm_info)); |
| 95 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(output, &tensor_info_output); |
| 96 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input0, output); |
| 97 | } |
| 98 | |
| 99 | return Status{}; |
| 100 | } |
| 101 | |
| 102 | std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input0, ITensorInfo *input1, ITensorInfo *output, const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info, |
| 103 | const GEMMReshapeInfo &gemm_info, ElementsProcessed &num_elements_processed) |
| 104 | { |
| 105 | unsigned int &num_elems_processed_per_iteration_x = num_elements_processed[0]; |
| 106 | unsigned int &num_elems_processed_per_iteration_y = num_elements_processed[1]; |
| 107 | bool reinterpret_output_as_3d = (gemm_info.depth_output_gemm3d() != 0); |
| 108 | |
| 109 | Window win{}; |
| 110 | Window win_out{}; |
| 111 | bool window_changed = false; |
| 112 | |
| 113 | // Output tensor auto initialization if not yet initialized |
| 114 | auto_init_if_empty(*output, input0->clone()->set_tensor_shape(compute_mm_shape(*input0, *input1, gemm_info))); |
| 115 | |
| 116 | TensorInfo tmp_info(*output); |
| 117 | |
| 118 | if(reinterpret_output_as_3d) |
| 119 | { |
| 120 | // Since the output tensor has to be reinterpreted as 3D and the execute window is based on a 2D GEMM, |
| 121 | // the window needs to be constructed on the 2D collapsed version of the tensor |
| 122 | TensorShape tmp_shape(output->tensor_shape()); |
| 123 | tmp_shape.collapse(2U, 1U); |
| 124 | tmp_info.set_tensor_shape(tmp_shape); |
| 125 | } |
| 126 | |
| 127 | // Configure kernel window |
| 128 | num_elems_processed_per_iteration_x = rhs_info.n0; |
| 129 | num_elems_processed_per_iteration_y = lhs_info.m0; |
| 130 | |
| 131 | // Note: bottom paddings are calculated manually as the output can be reinterpreted as 3D tensor |
| 132 | // The only way to set properly the paddings, it is to set those explicitly through the AccessWindowStatic |
| 133 | const int m = gemm_info.m(); |
| 134 | const int bottom_pad = (num_elems_processed_per_iteration_y - (m % num_elems_processed_per_iteration_y)) % num_elems_processed_per_iteration_y; |
| 135 | |
| 136 | win = calculate_max_window(tmp_info, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y)); |
| 137 | win_out = calculate_max_window(*output, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y)); |
| 138 | |
| 139 | AccessWindowStatic input0_access(input0, 0, 0, |
| 140 | ceil_to_multiple(input0->dimension(0), num_elems_processed_per_iteration_y), |
| 141 | input0->dimension(1)); |
| 142 | AccessWindowStatic input1_access(input1, 0, 0, |
| 143 | ceil_to_multiple(input1->dimension(0), num_elems_processed_per_iteration_x), |
| 144 | input1->dimension(1)); |
| 145 | AccessWindowStatic output_access(output, 0, 0, |
| 146 | ceil_to_multiple(output->dimension(0), num_elems_processed_per_iteration_x), |
| 147 | output->dimension(1) + bottom_pad); |
| 148 | |
| 149 | window_changed = update_window_and_padding(win, input0_access, input1_access) || // window used by the execute_window_loop |
| 150 | update_window_and_padding(win_out, output_access); // window used to update the padding requirements of output tensor |
| 151 | |
| 152 | output_access.set_valid_region(win_out, ValidRegion(Coordinates(0, 0), output->tensor_shape())); |
| 153 | |
| 154 | // Collapse along the Z direction |
| 155 | // This collapse needs to be here in order to tune the Z dimension of LWS |
| 156 | Window collapsed = win; |
| 157 | const unsigned int dimension_to_collapse = std::min(static_cast<unsigned int>(output->num_dimensions()), 2u); |
| 158 | collapsed = win.collapse(win, dimension_to_collapse); |
| 159 | |
| 160 | Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{}; |
| 161 | return std::make_pair(err, collapsed); |
| 162 | } |
| 163 | } // namespace |
| 164 | |
| 165 | CLGEMMMatrixMultiplyReshapedKernel::CLGEMMMatrixMultiplyReshapedKernel() |
| 166 | : _input0(nullptr), _input1(nullptr), _output(nullptr), _slide_matrix_b(true), _reinterpret_output_as_3d(false) |
| 167 | { |
| 168 | } |
| 169 | |
| 170 | void CLGEMMMatrixMultiplyReshapedKernel::configure(const ICLTensor *input0, const ICLTensor *input1, ICLTensor *output, float alpha, const GEMMLHSMatrixInfo &lhs_info, |
| 171 | const GEMMRHSMatrixInfo &rhs_info, const GEMMReshapeInfo &gemm_info) |
| 172 | { |
| 173 | ARM_COMPUTE_ERROR_ON_NULLPTR(input0, input1, output); |
| 174 | |
| 175 | ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input0->info(), input1->info(), output->info(), alpha, lhs_info, rhs_info, gemm_info)); |
| 176 | |
| 177 | _input0 = input0; |
| 178 | _input1 = input1; |
| 179 | _output = output; |
| 180 | _reinterpret_output_as_3d = (gemm_info.depth_output_gemm3d() != 0); |
| 181 | |
| 182 | // Check if we need to slide the matrix B |
| 183 | const unsigned int num_dimensions_input0 = _input0->info()->num_dimensions(); |
| 184 | _slide_matrix_b = (_input1->info()->num_dimensions() >= num_dimensions_input0); |
| 185 | |
| 186 | ElementsProcessed num_elements_processed{}; |
| 187 | |
| 188 | // Configure kernel window |
| 189 | auto win_config = validate_and_configure_window(input0->info(), input1->info(), output->info(), lhs_info, rhs_info, gemm_info, num_elements_processed); |
| 190 | ARM_COMPUTE_ERROR_THROW_ON(win_config.first); |
| 191 | ICLKernel::configure_internal(win_config.second); |
| 192 | |
| 193 | // Create build options |
| 194 | CLBuildOptions build_opts; |
| 195 | build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(input0->info()->data_type())); |
| 196 | build_opts.add_option_if(std::abs(1.0f - alpha) > 0.00001f, "-DALPHA=" + float_to_string_with_full_precision(alpha)); |
| 197 | build_opts.add_option_if(_reinterpret_output_as_3d, "-DREINTERPRET_OUTPUT_AS_3D"); |
| 198 | build_opts.add_option_if(_reinterpret_output_as_3d, "-DHEIGHT_GEMM3D=" + support::cpp11::to_string(output->info()->dimension(1))); |
| 199 | build_opts.add_option_if(_reinterpret_output_as_3d, "-DDEPTH_GEMM3D=" + support::cpp11::to_string(output->info()->dimension(2))); |
| 200 | build_opts.add_option_if(!_slide_matrix_b, "-DMATRIX_B_DEPTH=" + support::cpp11::to_string(input1->info()->dimension(2))); |
| 201 | build_opts.add_option_if(lhs_info.interleave, "-DLHS_INTERLEAVE"); |
| 202 | build_opts.add_option_if(rhs_info.interleave, "-DRHS_INTERLEAVE"); |
| 203 | build_opts.add_option("-DK=" + support::cpp11::to_string(gemm_info.k())); |
| 204 | build_opts.add_option("-DM0=" + support::cpp11::to_string(lhs_info.m0)); |
| 205 | build_opts.add_option("-DN0=" + support::cpp11::to_string(rhs_info.n0)); |
| 206 | build_opts.add_option("-DK0=" + support::cpp11::to_string(lhs_info.k0)); |
| 207 | build_opts.add_option("-DV0=" + support::cpp11::to_string(lhs_info.v0)); |
| 208 | build_opts.add_option("-DH0=" + support::cpp11::to_string(rhs_info.h0)); |
| 209 | |
| 210 | std::string kernel_name("gemm_mm_reshaped_"); |
| 211 | kernel_name += lhs_info.transpose ? "lhs_t_" : "lhs_nt_"; |
| 212 | kernel_name += rhs_info.transpose ? "rhs_t" : "rhs_nt"; |
| 213 | |
| 214 | // Create kernel |
| 215 | _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel(kernel_name, build_opts.options())); |
| 216 | |
| 217 | // Set config_id for enabling LWS tuning |
| 218 | _config_id = kernel_name; |
| 219 | _config_id += "_"; |
| 220 | _config_id += (_reinterpret_output_as_3d ? "3do_" : ""); |
| 221 | _config_id += lower_string(string_from_data_type(input0->info()->data_type())); |
| 222 | _config_id += "_"; |
| 223 | _config_id += support::cpp11::to_string(output->info()->dimension(1)); |
| 224 | _config_id += "_"; |
| 225 | _config_id += support::cpp11::to_string(output->info()->dimension(0)); |
| 226 | _config_id += "_"; |
| 227 | _config_id += support::cpp11::to_string(gemm_info.k()); |
| 228 | _config_id += "_"; |
| 229 | _config_id += support::cpp11::to_string(output->info()->dimension(2)); |
| 230 | _config_id += "_"; |
| 231 | _config_id += support::cpp11::to_string(lhs_info.m0); |
| 232 | _config_id += "_"; |
| 233 | _config_id += support::cpp11::to_string(rhs_info.n0); |
| 234 | _config_id += "_"; |
| 235 | _config_id += support::cpp11::to_string(lhs_info.k0); |
| 236 | _config_id += "_"; |
| 237 | _config_id += support::cpp11::to_string(lhs_info.v0); |
| 238 | _config_id += "_"; |
| 239 | _config_id += support::cpp11::to_string(rhs_info.h0); |
| 240 | _config_id += "_"; |
| 241 | _config_id += support::cpp11::to_string(lhs_info.interleave); |
| 242 | _config_id += "_"; |
| 243 | _config_id += support::cpp11::to_string(rhs_info.interleave); |
| 244 | } |
| 245 | |
| 246 | Status CLGEMMMatrixMultiplyReshapedKernel::validate(const ITensorInfo *input0, const ITensorInfo *input1, const ITensorInfo *output, float alpha, const GEMMLHSMatrixInfo &lhs_info, |
| 247 | const GEMMRHSMatrixInfo &rhs_info, const GEMMReshapeInfo &gemm_info) |
| 248 | { |
| 249 | ElementsProcessed num_elements_processed{}; |
| 250 | ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input0, input1, output, alpha, lhs_info, rhs_info, gemm_info)); |
| 251 | ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input0->clone().get(), |
| 252 | input1->clone().get(), |
| 253 | output->clone().get(), |
| 254 | lhs_info, |
| 255 | rhs_info, |
| 256 | gemm_info, |
| 257 | num_elements_processed) |
| 258 | .first); |
| 259 | |
| 260 | return Status{}; |
| 261 | } |
| 262 | |
| 263 | void CLGEMMMatrixMultiplyReshapedKernel::run(const Window &window, cl::CommandQueue &queue) |
| 264 | { |
| 265 | ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); |
| 266 | ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window); |
| 267 | |
| 268 | if(_input1->info()->num_dimensions() < 3) |
| 269 | { |
| 270 | // The stride_z for matrix B must be zero if we do not slice |
| 271 | ARM_COMPUTE_ERROR_ON(_input1->info()->strides_in_bytes()[3] != 0); |
| 272 | } |
| 273 | |
| 274 | Window slice = window.first_slice_window_3D(); |
| 275 | Window slice_matrix_b = slice; |
| 276 | |
| 277 | slice_matrix_b.set(Window::DimX, Window::Dimension(0, 1, 1)); |
| 278 | slice_matrix_b.set(Window::DimY, Window::Dimension(0, 1, 1)); |
| 279 | |
| 280 | if(_reinterpret_output_as_3d) |
| 281 | { |
| 282 | // Pass bottom paddings to the kernel if the output has to be reinterpreted as 3D tensor |
| 283 | const unsigned int idx0 = 3 * num_arguments_per_2D_tensor() + 3; |
| 284 | const unsigned int total_cross_plane_pad = _output->info()->padding().top + _output->info()->padding().bottom; |
| 285 | _kernel.setArg<cl_uint>(idx0, static_cast<unsigned int>(total_cross_plane_pad)); |
| 286 | } |
| 287 | |
| 288 | do |
| 289 | { |
| 290 | Window slice_b = slice; |
| 291 | // Don't slice matrix B along the z dimension if matrix B has just 2 dimensions and matrix A more than 2 |
| 292 | // This scenario can happen when the matrix multiplication is used to perform a convolution operation |
| 293 | if(!_slide_matrix_b) |
| 294 | { |
| 295 | slice_b = slice_matrix_b; |
| 296 | } |
| 297 | |
| 298 | unsigned int idx = 0; |
| 299 | add_2D_tensor_argument(idx, _input0, slice); |
| 300 | add_2D_tensor_argument(idx, _input1, slice_b); |
| 301 | add_2D_tensor_argument(idx, _output, slice); |
| 302 | _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_input0->info()->strides_in_bytes()[2])); |
| 303 | _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_input1->info()->strides_in_bytes()[2])); |
| 304 | _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_output->info()->strides_in_bytes()[2])); |
| 305 | enqueue(queue, *this, slice, lws_hint()); |
| 306 | } |
| 307 | while(window.slide_window_slice_3D(slice)); |
| 308 | } |