Georgios Pinitas | 856f66e | 2021-04-22 21:13:21 +0100 | [diff] [blame] | 1 | /* |
| 2 | * Copyright (c) 2019-2021 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 "src/core/gpu/cl/kernels/ClGemmMatrixMultiplyReshapedOnlyRhsKernel.h" |
| 25 | |
| 26 | #include "arm_compute/core/CL/ICLTensor.h" |
| 27 | #include "arm_compute/core/Helpers.h" |
| 28 | #include "arm_compute/core/TensorInfo.h" |
| 29 | #include "arm_compute/core/Utils.h" |
| 30 | #include "arm_compute/core/utils/misc/ShapeCalculator.h" |
| 31 | #include "src/core/AccessWindowStatic.h" |
| 32 | #include "src/core/CL/CLUtils.h" |
| 33 | #include "src/core/CL/CLValidate.h" |
| 34 | #include "src/core/gpu/cl/kernels/gemm/ClGemmHelpers.h" |
| 35 | #include "src/core/helpers/AutoConfiguration.h" |
| 36 | #include "src/core/helpers/WindowHelpers.h" |
| 37 | #include "src/core/utils/helpers/float_ops.h" |
| 38 | #include "support/Cast.h" |
| 39 | #include "support/StringSupport.h" |
| 40 | |
| 41 | namespace arm_compute |
| 42 | { |
| 43 | namespace opencl |
| 44 | { |
| 45 | namespace kernels |
| 46 | { |
| 47 | namespace |
| 48 | { |
| 49 | using ElementsProcessed = Steps; |
| 50 | |
| 51 | Status validate_arguments(const ITensorInfo *src0, const ITensorInfo *src1, const ITensorInfo *src2, const ITensorInfo *dst, float alpha, float beta, |
| 52 | const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info, const GEMMKernelInfo &gemm_info) |
| 53 | { |
| 54 | ARM_COMPUTE_UNUSED(alpha); |
| 55 | ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src0, src1, dst); |
| 56 | ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(src0); |
| 57 | ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src0, 1, DataType::F16, DataType::F32); |
| 58 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src0, src1); |
| 59 | ARM_COMPUTE_RETURN_ERROR_ON_MSG(src0->num_dimensions() > 4, "The number of dimensions for the LHS matrix must be <= 4"); |
| 60 | ARM_COMPUTE_RETURN_ERROR_ON_MSG(src1->num_dimensions() > 3, "The number of dimensions for the RHS matrix must be <= 3"); |
| 61 | ARM_COMPUTE_RETURN_ERROR_ON_MSG(lhs_info.m0 < 1 || lhs_info.m0 > 8, "Only 1,2,3,4,5,6,7,8 are supported for m0"); |
| 62 | ARM_COMPUTE_RETURN_ERROR_ON(rhs_info.k0 > 16 || rhs_info.k0 < 2); |
| 63 | ARM_COMPUTE_RETURN_ERROR_ON_MSG(((rhs_info.k0 & (rhs_info.k0 - 1)) && rhs_info.k0 != 3), "Only 2,3,4,8,16 are supported for k0"); |
| 64 | ARM_COMPUTE_RETURN_ERROR_ON(rhs_info.n0 > 16 || rhs_info.n0 < 2); |
| 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 | ARM_COMPUTE_RETURN_ERROR_ON_MSG((gemm_info.reinterpret_input_as_3d || gemm_info.depth_output_gemm3d != 0) && (src2 != nullptr) |
| 67 | && (!gemm_info.broadcast_bias), |
| 68 | "Bias addition only supported with broadcast mode in case the input or dst has to be reinterpreted as 3D"); |
| 69 | ARM_COMPUTE_RETURN_ERROR_ON_MSG(gemm_info.fp_mixed_precision, "Mixed precision not supported"); |
| 70 | ARM_COMPUTE_RETURN_ON_ERROR(gemm::validate_image2d_support_on_rhs(*src1, rhs_info)); |
| 71 | |
| 72 | const unsigned int m = gemm_info.m; |
| 73 | const unsigned int n = gemm_info.n; |
| 74 | const unsigned int k = gemm_info.k; |
| 75 | |
| 76 | TensorShape tensor_shape1{ src1->tensor_shape() }; |
| 77 | tensor_shape1.set(0, n); |
| 78 | tensor_shape1.set(1, k); |
| 79 | |
| 80 | if(src2 != nullptr && !(helpers::float_ops::is_zero(beta))) |
| 81 | { |
| 82 | const unsigned int src2_dim0 = src2->dimension(0); |
| 83 | const unsigned int src2_dim1 = src2->dimension(1); |
| 84 | |
| 85 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src2, src0); |
| 86 | if(gemm_info.broadcast_bias) |
| 87 | { |
| 88 | ARM_COMPUTE_RETURN_ERROR_ON_MSG((src2_dim1 != 1 || src2_dim0 != n), "Incorrect dimension of bias matrix which is to be broadcasted"); |
| 89 | } |
| 90 | else |
| 91 | { |
| 92 | ARM_COMPUTE_RETURN_ERROR_ON_MSG((src2_dim0 != n || src2_dim1 != m), "Incorrect dimension of bias matrix"); |
| 93 | } |
| 94 | } |
| 95 | |
| 96 | const TensorInfo tensor_info1 = src1->clone()->set_tensor_shape(tensor_shape1); |
| 97 | |
| 98 | const TensorInfo tensor_info_reshaped1 = src1->clone()->set_tensor_shape(misc::shape_calculator::compute_rhs_reshaped_shape(tensor_info1, rhs_info)); |
| 99 | |
| 100 | ARM_COMPUTE_RETURN_ERROR_ON(src0->dimension(0) != k); |
| 101 | if(gemm_info.reinterpret_input_as_3d) |
| 102 | { |
| 103 | ARM_COMPUTE_RETURN_ERROR_ON(src0->dimension(1) * src0->dimension(2) != m); |
| 104 | } |
| 105 | else |
| 106 | { |
| 107 | ARM_COMPUTE_RETURN_ERROR_ON(src0->dimension(1) != m); |
| 108 | } |
| 109 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(src1, &tensor_info_reshaped1); |
| 110 | |
| 111 | if(dst->total_size() != 0) |
| 112 | { |
| 113 | const TensorInfo tensor_info_dst = dst->clone()->set_tensor_shape(misc::shape_calculator::compute_mm_shape(*src0, *src1, gemm_info)); |
| 114 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(dst, &tensor_info_dst); |
| 115 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src0, dst); |
| 116 | } |
| 117 | |
| 118 | return Status{}; |
| 119 | } |
| 120 | |
| 121 | std::pair<Status, Window> validate_and_configure_window(ITensorInfo *src0, ITensorInfo *src1, ITensorInfo *src2, ITensorInfo *dst, const GEMMLHSMatrixInfo &lhs_info, |
| 122 | const GEMMRHSMatrixInfo &rhs_info, const GEMMKernelInfo &gemm_info, ElementsProcessed &num_elements_processed) |
| 123 | { |
| 124 | unsigned int &num_elems_processed_per_iteration_x = num_elements_processed[0]; |
| 125 | unsigned int &num_elems_processed_per_iteration_y = num_elements_processed[1]; |
| 126 | bool reinterpret_input_as_3d = gemm_info.reinterpret_input_as_3d; |
| 127 | bool reinterpret_output_as_3d = gemm_info.depth_output_gemm3d != 0; |
| 128 | |
| 129 | Window win{}; |
| 130 | Window win_out{}; |
| 131 | bool window_changed = false; |
| 132 | |
| 133 | // In case both input and dst have to be reinterpreted as 3D tensors, |
| 134 | // force reinterpret_input_as_3d and reinterpret_output_as_3d to be false. |
| 135 | // This approach should only be used when the input/dst tensors have pad on the y direction |
| 136 | if((reinterpret_input_as_3d == reinterpret_output_as_3d) && gemm_info.has_pad_y) |
| 137 | { |
| 138 | reinterpret_output_as_3d = false; |
| 139 | } |
| 140 | |
| 141 | // dst tensor auto initialization if not yet initialized |
| 142 | auto_init_if_empty(*dst, src0->clone()->set_tensor_shape(misc::shape_calculator::compute_mm_shape(*src0, *src1, gemm_info))); |
| 143 | |
| 144 | TensorInfo tmp_info(*dst); |
| 145 | |
| 146 | if(reinterpret_output_as_3d) |
| 147 | { |
| 148 | // Since the dst tensor has to be reinterpreted as 3D and the execute window is based on a 2D GEMM, |
| 149 | // the window needs to be constructed on the 2D collapsed version of the tensor |
| 150 | TensorShape tmp_shape(dst->tensor_shape()); |
| 151 | tmp_shape.collapse(2U, 1U); |
| 152 | tmp_info.set_tensor_shape(tmp_shape); |
| 153 | } |
| 154 | |
| 155 | // Configure kernel window |
| 156 | num_elems_processed_per_iteration_x = rhs_info.n0; |
| 157 | num_elems_processed_per_iteration_y = lhs_info.m0; |
| 158 | |
| 159 | win = calculate_max_window(tmp_info, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y)); |
| 160 | win_out = calculate_max_window(*dst, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y)); |
| 161 | |
| 162 | if(src2 != nullptr) |
| 163 | { |
| 164 | const int bias_processed_per_iteration_x = num_elems_processed_per_iteration_x; |
| 165 | |
| 166 | AccessWindowStatic src2_access(src2, 0, 0, |
| 167 | ceil_to_multiple(src2->dimension(0), bias_processed_per_iteration_x), |
| 168 | src2->dimension(1)); |
| 169 | |
| 170 | window_changed = update_window_and_padding(win, src2_access); |
| 171 | } |
| 172 | |
| 173 | // Collapse along the Z direction |
| 174 | // This collapse needs to be here in order to tune the Z dimension of LWS |
| 175 | Window collapsed = win; |
| 176 | const unsigned int dimension_to_collapse = std::min(static_cast<unsigned int>(dst->num_dimensions()), 2u); |
| 177 | collapsed = win.collapse(win, dimension_to_collapse); |
| 178 | |
| 179 | Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{}; |
| 180 | return std::make_pair(err, collapsed); |
| 181 | } |
| 182 | } // namespace |
| 183 | |
Giorgio Arena | 4a95bba | 2021-06-28 11:00:27 +0100 | [diff] [blame] | 184 | ClGemmMatrixMultiplyReshapedOnlyRhsKernel::ClGemmMatrixMultiplyReshapedOnlyRhsKernel() |
| 185 | { |
| 186 | _type = CLKernelType::GEMM; |
| 187 | } |
| 188 | |
Georgios Pinitas | 856f66e | 2021-04-22 21:13:21 +0100 | [diff] [blame] | 189 | void ClGemmMatrixMultiplyReshapedOnlyRhsKernel::configure(const CLCompileContext &compile_context, |
| 190 | ITensorInfo *src0, ITensorInfo *src1, ITensorInfo *src2, ITensorInfo *dst, float alpha, float beta, |
| 191 | const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info, const GEMMKernelInfo &gemm_info) |
| 192 | { |
| 193 | ARM_COMPUTE_ERROR_ON_NULLPTR(src0, src1, dst); |
| 194 | |
| 195 | ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(src0, src1, src2, dst, alpha, beta, lhs_info, rhs_info, gemm_info)); |
| 196 | |
| 197 | _reinterpret_input_as_3d = gemm_info.reinterpret_input_as_3d; |
| 198 | _reinterpret_output_as_3d = gemm_info.depth_output_gemm3d != 0; |
| 199 | _use_dummy_work_items = preferred_dummy_work_items_support(CLKernelLibrary::get().get_device()); |
| 200 | _add_bias = src2 != nullptr; |
| 201 | _export_to_cl_image = rhs_info.export_to_cl_image; |
| 202 | _has_pad_y = gemm_info.has_pad_y; |
| 203 | |
| 204 | auto padding_info = get_padding_info({ src0, src1, dst }); |
| 205 | |
| 206 | // In case both input and dst have to be reinterpreted as 3D tensors, |
| 207 | // force reinterpret_input_as_3d and reinterpret_output_as_3d to be false. |
| 208 | if((_reinterpret_input_as_3d == _reinterpret_output_as_3d) && _has_pad_y) |
| 209 | { |
| 210 | _reinterpret_input_as_3d = false; |
| 211 | _reinterpret_output_as_3d = false; |
| 212 | } |
| 213 | |
| 214 | // Check if we need to slide the matrix B |
| 215 | const unsigned int num_dimensions_src0 = src0->num_dimensions(); |
| 216 | _slide_matrix_b = (src1->num_dimensions() >= num_dimensions_src0); |
| 217 | |
| 218 | ElementsProcessed num_elements_processed{}; |
| 219 | |
| 220 | // Configure kernel window |
| 221 | auto win_config = validate_and_configure_window(src0, src1, src2, dst, lhs_info, rhs_info, gemm_info, num_elements_processed); |
| 222 | ARM_COMPUTE_ERROR_THROW_ON(win_config.first); |
| 223 | ICLKernel::configure_internal(win_config.second); |
| 224 | |
| 225 | // If _reinterpret_input_as_3d = reinterpret_output_as_3d = true, |
| 226 | // we will dispatch a batched-GEMM to reduce the complexity of the address calculation within the OpenCL kernel. |
| 227 | // This means that the actual m used by the kernel is given by dst->dimension(1) and not by gemm_info.m |
| 228 | const unsigned int internal_m = _reinterpret_output_as_3d ? gemm_info.m : dst->dimension(1); |
| 229 | |
| 230 | // These variables are used only if gemm_info.has_pad_y == true |
| 231 | const unsigned int h_gemm_3d = _reinterpret_output_as_3d ? dst->dimension(1) : src0->dimension(1); |
| 232 | const unsigned int d_gemm_3d = _reinterpret_output_as_3d ? dst->dimension(2) : src0->dimension(2); |
| 233 | |
| 234 | // Shrink M0 to be always <= M (internal_m) to prevent out-of-bounds reads. |
| 235 | // NOTE: This might have implications on heuristics and performance |
| 236 | const unsigned int internal_m0 = std::min(internal_m, lhs_info.m0); |
| 237 | |
| 238 | // Calculate partial (store instead of load) M0 and partial N0 for the partial blocks at the end of a row/column if any. This is to avoid padding. |
| 239 | const unsigned int partial_store_m0 = internal_m % internal_m0; |
| 240 | const unsigned int partial_store_n0 = gemm_info.n % rhs_info.n0; |
| 241 | |
| 242 | // Create build options |
| 243 | CLBuildOptions build_opts; |
| 244 | build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(src0->data_type())); |
| 245 | build_opts.add_option_if(!(helpers::float_ops::is_one(alpha)), "-DALPHA=" + float_to_string_with_full_precision(alpha)); |
| 246 | build_opts.add_option_if(src2 != nullptr, "-DBETA=" + float_to_string_with_full_precision(beta)); |
| 247 | build_opts.add_option_if(helpers::float_ops::is_one(beta), "-DUNIT_BETA"); |
| 248 | build_opts.add_option_if(gemm_info.broadcast_bias, "-DBROADCAST_BIAS"); |
| 249 | build_opts.add_option_if(!_slide_matrix_b, "-DMATRIX_B_DEPTH=" + support::cpp11::to_string(src1->dimension(2))); |
| 250 | build_opts.add_option_if(rhs_info.interleave, "-DRHS_INTERLEAVE"); |
| 251 | build_opts.add_option_if(_use_dummy_work_items, "-DDUMMY_WORK_ITEMS"); |
| 252 | build_opts.add_option_if(rhs_info.export_to_cl_image, "-DOPENCL_IMAGE_SUPPORT"); |
| 253 | build_opts.add_option("-DRHS_HEIGHT=" + support::cpp11::to_string(src1->dimension(1))); |
| 254 | build_opts.add_option("-DM=" + support::cpp11::to_string(internal_m)); |
| 255 | build_opts.add_option("-DN=" + support::cpp11::to_string(gemm_info.n)); |
| 256 | build_opts.add_option("-DK=" + support::cpp11::to_string(gemm_info.k)); |
| 257 | build_opts.add_option("-DM0=" + support::cpp11::to_string(internal_m0)); |
| 258 | build_opts.add_option("-DN0=" + support::cpp11::to_string(rhs_info.n0)); |
| 259 | build_opts.add_option("-DK0=" + support::cpp11::to_string(rhs_info.k0)); |
| 260 | build_opts.add_option("-DH0=" + support::cpp11::to_string(rhs_info.h0)); |
| 261 | build_opts.add_option("-DPARTIAL_STORE_M0=" + support::cpp11::to_string(partial_store_m0)); |
| 262 | build_opts.add_option("-DPARTIAL_STORE_N0=" + support::cpp11::to_string(partial_store_n0)); |
| 263 | build_opts.add_option_if(gemm_info.activation_info.enabled(), "-DACTIVATION_TYPE=" + lower_string(string_from_activation_func(gemm_info.activation_info.activation()))); |
| 264 | build_opts.add_option_if(gemm_info.activation_info.enabled(), "-DA_VAL=" + float_to_string_with_full_precision(gemm_info.activation_info.a())); |
| 265 | build_opts.add_option_if(gemm_info.activation_info.enabled(), "-DB_VAL=" + float_to_string_with_full_precision(gemm_info.activation_info.b())); |
| 266 | if(_has_pad_y) |
| 267 | { |
| 268 | build_opts.add_option_if(_reinterpret_input_as_3d, "-DREINTERPRET_INPUT_AS_3D"); |
| 269 | build_opts.add_option_if(_reinterpret_output_as_3d, "-DREINTERPRET_OUTPUT_AS_3D"); |
| 270 | build_opts.add_option_if(_reinterpret_input_as_3d || _reinterpret_output_as_3d, "-DHEIGHT_GEMM3D=" + support::cpp11::to_string(h_gemm_3d)); |
| 271 | build_opts.add_option_if(_reinterpret_input_as_3d || _reinterpret_output_as_3d, "-DDEPTH_GEMM3D=" + support::cpp11::to_string(d_gemm_3d)); |
| 272 | } |
| 273 | |
| 274 | std::string kernel_name("gemm_mm_reshaped_only_rhs_"); |
| 275 | kernel_name += rhs_info.transpose ? "t" : "nt"; |
| 276 | kernel_name += rhs_info.export_to_cl_image ? "_texture" : ""; |
| 277 | |
| 278 | // Create kernel |
| 279 | _kernel = create_kernel(compile_context, kernel_name, build_opts.options()); |
| 280 | |
| 281 | // Set config_id for enabling LWS tuning |
| 282 | _config_id = kernel_name; |
| 283 | _config_id += "_"; |
| 284 | _config_id += (_has_pad_y ? "" : "no_pad_y_"); |
| 285 | _config_id += (_add_bias ? "add_bias_" : ""); |
| 286 | _config_id += (gemm_info.broadcast_bias ? "broadcast_bias_" : ""); |
| 287 | _config_id += (_reinterpret_input_as_3d ? "3di_" : ""); |
| 288 | _config_id += (_reinterpret_output_as_3d ? "3do_" : ""); |
| 289 | _config_id += (gemm_info.activation_info.enabled() ? "fused_activation_" : ""); |
| 290 | _config_id += lower_string(string_from_data_type(src0->data_type())); |
| 291 | _config_id += "_"; |
| 292 | _config_id += support::cpp11::to_string(dst->dimension(1)); |
| 293 | _config_id += "_"; |
| 294 | _config_id += support::cpp11::to_string(dst->dimension(0)); |
| 295 | _config_id += "_"; |
| 296 | _config_id += support::cpp11::to_string(gemm_info.k); |
| 297 | _config_id += "_"; |
| 298 | _config_id += support::cpp11::to_string(dst->dimension(2)); |
| 299 | _config_id += "_"; |
| 300 | _config_id += support::cpp11::to_string(lhs_info.m0); |
| 301 | _config_id += "_"; |
| 302 | _config_id += support::cpp11::to_string(rhs_info.n0); |
| 303 | _config_id += "_"; |
| 304 | _config_id += support::cpp11::to_string(rhs_info.k0); |
| 305 | _config_id += "_"; |
| 306 | _config_id += support::cpp11::to_string(rhs_info.h0); |
| 307 | _config_id += "_"; |
| 308 | _config_id += support::cpp11::to_string(rhs_info.interleave); |
| 309 | |
| 310 | ARM_COMPUTE_ERROR_ON(has_padding_changed(padding_info)); |
| 311 | } |
| 312 | |
| 313 | Status ClGemmMatrixMultiplyReshapedOnlyRhsKernel::validate(const ITensorInfo *src0, const ITensorInfo *src1, const ITensorInfo *src2, const ITensorInfo *dst, float alpha, float beta, |
| 314 | const GEMMLHSMatrixInfo &lhs_info, |
| 315 | const GEMMRHSMatrixInfo &rhs_info, const GEMMKernelInfo &gemm_info) |
| 316 | { |
| 317 | ElementsProcessed num_elements_processed{}; |
| 318 | ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(src0, src1, src2, dst, alpha, beta, lhs_info, rhs_info, gemm_info)); |
| 319 | ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(src0->clone().get(), |
| 320 | src1->clone().get(), |
| 321 | src2 != nullptr ? src2->clone().get() : nullptr, |
| 322 | dst->clone().get(), |
| 323 | lhs_info, |
| 324 | rhs_info, |
| 325 | gemm_info, |
| 326 | num_elements_processed) |
| 327 | .first); |
| 328 | |
| 329 | return Status{}; |
| 330 | } |
| 331 | |
| 332 | void ClGemmMatrixMultiplyReshapedOnlyRhsKernel::run_op(ITensorPack &tensors, const Window &window, cl::CommandQueue &queue) |
| 333 | { |
| 334 | ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); |
| 335 | ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window); |
| 336 | |
| 337 | const auto src0 = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_0)); |
| 338 | const auto src1 = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_1)); |
| 339 | const auto src2 = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_2)); |
| 340 | auto dst = utils::cast::polymorphic_downcast<ICLTensor *>(tensors.get_tensor(TensorType::ACL_DST)); |
| 341 | |
| 342 | ARM_COMPUTE_ERROR_ON_NULLPTR(src0, src1, dst); |
| 343 | ARM_COMPUTE_ERROR_ON(_add_bias && src2 == nullptr); |
| 344 | |
| 345 | if(src1->info()->num_dimensions() < 3) |
| 346 | { |
| 347 | // The stride_z for matrix B must be zero if we do not slice |
| 348 | ARM_COMPUTE_ERROR_ON(src1->info()->strides_in_bytes()[3] != 0); |
| 349 | } |
| 350 | |
| 351 | const size_t lhs_idx_batch_size = _reinterpret_input_as_3d && !_has_pad_y ? 3u : 2u; |
| 352 | const size_t rhs_idx_batch_size = 2u; |
| 353 | const size_t bia_idx_batch_size = 2u; |
| 354 | const size_t out_idx_batch_size = _reinterpret_output_as_3d && !_has_pad_y ? 3u : 2u; |
| 355 | |
| 356 | Window slice = window.first_slice_window_3D(); |
| 357 | Window slice_matrix_b = slice; |
| 358 | |
| 359 | slice_matrix_b.set(Window::DimX, Window::Dimension(0, 1, 1)); |
| 360 | slice_matrix_b.set(Window::DimY, Window::Dimension(0, 1, 1)); |
| 361 | |
| 362 | // Get cross plane pads |
| 363 | const unsigned int total_cross_plane_pad_lhs = src0->info()->padding().top + src0->info()->padding().bottom; |
| 364 | const unsigned int total_cross_plane_pad_out = dst->info()->padding().top + dst->info()->padding().bottom; |
| 365 | |
| 366 | // The execution should fail if we try to run with has_pad_y = false but we have padding in either the LHS or DST tensor |
| 367 | ARM_COMPUTE_ERROR_ON(!_has_pad_y && ((total_cross_plane_pad_lhs != 0) || (total_cross_plane_pad_out != 0))); |
| 368 | |
| 369 | cl::Image2D src1_image2d; |
| 370 | |
| 371 | if(_export_to_cl_image) |
| 372 | { |
| 373 | const TensorShape shape2d(src1->info()->dimension(0) / 4, src1->info()->dimension(1) * src1->info()->dimension(2)); |
| 374 | const size_t image_row_pitch = src1->info()->strides_in_bytes()[1]; |
| 375 | |
| 376 | src1_image2d = create_image2d_from_buffer(CLKernelLibrary::get().context(), src1->cl_buffer(), shape2d, src1->info()->data_type(), image_row_pitch); |
| 377 | } |
| 378 | |
| 379 | do |
| 380 | { |
| 381 | Window slice_b = slice; |
| 382 | // Don't slice matrix B along the z dimension if matrix B has just 2 dimensions and matrix A more than 2 |
| 383 | // This scenario can happen when the matrix multiplication is used to perform a convolution operation |
| 384 | if(!_slide_matrix_b) |
| 385 | { |
| 386 | slice_b = slice_matrix_b; |
| 387 | } |
| 388 | |
| 389 | unsigned int idx = 0; |
| 390 | |
| 391 | // LHS buffer |
| 392 | add_2D_tensor_argument(idx, src0, slice); |
| 393 | |
| 394 | // RHS buffer or RHS OpenCL image (_export_to_cl_image == true) |
| 395 | if(_export_to_cl_image) |
| 396 | { |
| 397 | _kernel.setArg(idx++, src1_image2d); |
| 398 | } |
| 399 | else |
| 400 | { |
| 401 | add_2D_tensor_argument(idx, src1, slice_b); |
| 402 | } |
| 403 | |
| 404 | // Bias buffer (_add_bias == true) |
| 405 | add_2D_tensor_argument_if(_add_bias, idx, src2, slice); |
| 406 | |
| 407 | // dst buffer |
| 408 | add_2D_tensor_argument(idx, dst, slice); |
| 409 | |
| 410 | // LHS stride_z |
| 411 | _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(src0->info()->strides_in_bytes()[lhs_idx_batch_size])); |
| 412 | |
| 413 | // RHS stride_z (not used if _export_to_cl_image == true) |
| 414 | _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(src1->info()->strides_in_bytes()[rhs_idx_batch_size])); |
| 415 | |
| 416 | // Bias stride_z (if _add_bias == true) |
| 417 | if(_add_bias) |
| 418 | { |
| 419 | _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(src2->info()->strides_in_bytes()[bia_idx_batch_size])); |
| 420 | } |
| 421 | |
| 422 | // dst stride_z |
| 423 | _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(dst->info()->strides_in_bytes()[out_idx_batch_size])); |
| 424 | |
| 425 | // Cross-plan padding (if _reinterpret_input_as_3d = true) |
| 426 | if(_reinterpret_input_as_3d && _has_pad_y) |
| 427 | { |
| 428 | _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(total_cross_plane_pad_lhs)); |
| 429 | } |
| 430 | |
| 431 | // Cross-plan padding (if reinterpret_output_as_3d = true) |
| 432 | if(_reinterpret_output_as_3d && _has_pad_y) |
| 433 | { |
| 434 | _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(total_cross_plane_pad_out)); |
| 435 | } |
| 436 | |
| 437 | enqueue(queue, *this, slice, lws_hint(), _use_dummy_work_items); |
| 438 | } |
| 439 | while(window.slide_window_slice_3D(slice)); |
| 440 | } |
| 441 | } // namespace kernels |
| 442 | } // namespace opencl |
| 443 | } // namespace arm_compute |