Giorgio Arena | 232c452 | 2022-03-03 10:09:01 +0000 | [diff] [blame] | 1 | /* |
| 2 | * Copyright (c) 2022 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 | |
| 25 | #if defined(ENABLE_EXPERIMENTAL_DYNAMIC_FUSION) |
| 26 | |
| 27 | #include "src/gpu/cl/kernels/experimental/dynamic_fusion/ClCompositeKernel.h" |
| 28 | |
| 29 | #include "src/core/utils/helpers/float_ops.h" |
| 30 | #include "src/gpu/cl/kernels/ClElementwiseKernel.h" |
| 31 | #include "src/gpu/cl/kernels/ClGemmMatrixMultiplyNativeKernel.h" |
| 32 | #include "tests/CL/CLAccessor.h" |
| 33 | #include "tests/framework/Macros.h" |
| 34 | #include "tests/framework/datasets/Datasets.h" |
| 35 | #include "tests/validation/Validation.h" |
| 36 | #include "tests/validation/reference/ElementwiseOperations.h" |
| 37 | #include "tests/validation/reference/GEMM.h" |
| 38 | |
| 39 | #include "arm_compute/core/utils/misc/ShapeCalculator.h" |
| 40 | #include "src/core/AccessWindowStatic.h" |
| 41 | #include "src/core/helpers/AutoConfiguration.h" |
| 42 | #include "src/core/helpers/WindowHelpers.h" |
| 43 | |
| 44 | #include <chrono> |
| 45 | |
| 46 | using namespace arm_compute::experimental::dynamic_fusion; |
| 47 | |
| 48 | namespace arm_compute |
| 49 | { |
| 50 | namespace test |
| 51 | { |
| 52 | namespace validation |
| 53 | { |
| 54 | namespace |
| 55 | { |
| 56 | /** Macros which measures the wall clock time, and records it into a map measurement_map with name clock_name */ |
| 57 | #define TICK(clock_name) \ |
| 58 | auto clock_name##_tick = std::chrono::high_resolution_clock::now(); |
| 59 | #define TOCK(clock_name, measurement_map) \ |
| 60 | auto clock_name##_tock = std::chrono::high_resolution_clock::now(); \ |
| 61 | measurement_map["\"" #clock_name "\""] = duration_cast<microseconds>(clock_name##_tock - clock_name##_tick); |
| 62 | #define TOCK_AVG(clock_name, measurement_map, num_iterations) \ |
| 63 | auto clock_name##_tock = std::chrono::high_resolution_clock::now(); \ |
| 64 | measurement_map["\"" #clock_name "\""] = duration_cast<microseconds>((clock_name##_tock - clock_name##_tick) / (num_iterations)); |
| 65 | |
| 66 | template <typename T, typename U> |
| 67 | void fill(U &&tensor, int seed) |
| 68 | { |
| 69 | static_assert(std::is_floating_point<T>::value || std::is_same<T, half>::value, "Only floating point data types supported."); |
| 70 | using DistributionType = typename std::conditional<std::is_same<T, half>::value, arm_compute::utils::uniform_real_distribution_16bit<T>, std::uniform_real_distribution<T>>::type; |
| 71 | |
| 72 | DistributionType distribution{ T(-1.0f), T(1.0f) }; |
| 73 | library->fill(tensor, distribution, seed); |
| 74 | |
| 75 | // Fill border with infinity in order to check the presence of NaN values (i.e. inf * 0) |
| 76 | DistributionType distribution_inf{ T(std::numeric_limits<float>::infinity()), T(std::numeric_limits<float>::infinity()) }; |
| 77 | library->fill_borders_with_garbage(tensor, distribution_inf, seed); |
| 78 | } |
| 79 | |
| 80 | using ElementsProcessed = Steps; |
| 81 | std::pair<Status, Window> mock_gemm_native_validate_and_configure_window(ITensorInfo *src0, ITensorInfo *src1, ITensorInfo *src2, ITensorInfo *dst, const GEMMLHSMatrixInfo &lhs_info, |
| 82 | const GEMMRHSMatrixInfo &rhs_info, |
| 83 | const GEMMKernelInfo &gemm_info, ElementsProcessed &num_elements_processed) |
| 84 | { |
| 85 | unsigned int &num_elems_processed_per_iteration_x = num_elements_processed[0]; |
| 86 | unsigned int &num_elems_processed_per_iteration_y = num_elements_processed[1]; |
| 87 | bool reinterpret_input_as_3d = gemm_info.reinterpret_input_as_3d; |
| 88 | bool reinterpret_output_as_3d = gemm_info.depth_output_gemm3d != 0; |
| 89 | |
| 90 | Window win{}; |
| 91 | Window win_out{}; |
| 92 | bool window_changed = false; |
| 93 | |
| 94 | // In case both input and dst have to be reinterpreted as 3D tensors, |
| 95 | // force reinterpret_input_as_3d and reinterpret_output_as_3d to be false. |
| 96 | if(reinterpret_input_as_3d == reinterpret_output_as_3d) |
| 97 | { |
| 98 | reinterpret_output_as_3d = false; |
| 99 | } |
| 100 | |
| 101 | // dst tensor auto initialization if not yet initialized |
| 102 | auto_init_if_empty(*dst, src0->clone()->set_tensor_shape(misc::shape_calculator::compute_mm_shape(*src0, *src1, gemm_info))); |
| 103 | |
| 104 | TensorInfo tmp_info(*dst); |
| 105 | |
| 106 | if(reinterpret_output_as_3d) |
| 107 | { |
| 108 | // Since the dst tensor has to be reinterpreted as 3D and the execute window is based on a 2D GEMM, |
| 109 | // the window needs to be constructed on the 2D collapsed version of the tensor |
| 110 | TensorShape tmp_shape(dst->tensor_shape()); |
| 111 | tmp_shape.collapse(2U, 1U); |
| 112 | tmp_info.set_tensor_shape(tmp_shape); |
| 113 | } |
| 114 | |
| 115 | // Configure kernel window |
| 116 | num_elems_processed_per_iteration_x = rhs_info.n0; |
| 117 | num_elems_processed_per_iteration_y = lhs_info.m0; |
| 118 | |
| 119 | win = calculate_max_window(tmp_info, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y)); |
| 120 | win_out = calculate_max_window(*dst, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y)); |
| 121 | |
| 122 | AccessWindowStatic src0_access(src0, 0, 0, |
| 123 | src0->dimension(0), |
| 124 | src0->dimension(1)); |
| 125 | AccessWindowStatic src1_access(src1, 0, 0, |
| 126 | ceil_to_multiple(src1->dimension(0), num_elems_processed_per_iteration_x), |
| 127 | src1->dimension(1)); |
| 128 | AccessWindowStatic dst_access(dst, 0, 0, |
| 129 | dst->dimension(0), |
| 130 | dst->dimension(1)); |
| 131 | |
| 132 | if(src2 != nullptr) |
| 133 | { |
| 134 | const int bias_processed_per_iteration_x = num_elems_processed_per_iteration_x; |
| 135 | |
| 136 | AccessWindowStatic src2_access(src2, 0, 0, |
| 137 | ceil_to_multiple(src2->dimension(0), bias_processed_per_iteration_x), |
| 138 | src2->dimension(1)); |
| 139 | |
| 140 | window_changed = update_window_and_padding(win, src0_access, src1_access, src2_access) || // window used by the execute_window_loop |
| 141 | update_window_and_padding(win_out, dst_access); // window used to update the padding requirements of dst tensor |
| 142 | } |
| 143 | else |
| 144 | { |
| 145 | window_changed = update_window_and_padding(win, src0_access, src1_access) || // window used by the execute_window_loop |
| 146 | update_window_and_padding(win_out, dst_access); // window used to update the padding requirements of dst tensor |
| 147 | } |
| 148 | |
| 149 | // Collapse along the Z direction |
| 150 | // This collapse needs to be here in order to tune the Z dimension of LWS |
| 151 | Window collapsed = win; |
| 152 | const unsigned int dimension_to_collapse = std::min(static_cast<unsigned int>(dst->num_dimensions()), 2u); |
| 153 | collapsed = win.collapse(win, dimension_to_collapse); |
| 154 | |
| 155 | Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{}; |
| 156 | return std::make_pair(err, collapsed); |
| 157 | } |
| 158 | |
| 159 | void set_build_options(ClKernelCode &cl_code, GemmNativeDescriptor gemm_native_desc, |
| 160 | const TensorInfo &t_lhs_info, |
| 161 | const TensorInfo &t_rhs_info, |
| 162 | const TensorInfo *t_bias_info, |
| 163 | const TensorInfo &t_dst_info) |
| 164 | { |
| 165 | CLBuildOptions ref_cl_build_options; |
| 166 | { |
| 167 | // If reinterpret_input_as_3d = reinterpret_output_as_3d = true, |
| 168 | // we will dispatch a batched-GEMM to reduce the complexity of the address calculation within the OpenCL kernel. |
| 169 | // This means that the actual m used by the kernel is given by dst->dimension(1) and not by gemm_info.m |
| 170 | auto reinterpret_input_as_3d = gemm_native_desc.reinterpret_input_as_3d; |
| 171 | auto reinterpret_output_as_3d = gemm_native_desc.depth_output_gemm3d != 0; |
| 172 | auto _slide_matrix_b = (t_rhs_info.num_dimensions() >= t_lhs_info.num_dimensions()); |
| 173 | auto _use_dummy_work_items = false; |
| 174 | // In case both input and dst have to be reinterpreted as 3D tensors, |
| 175 | // force reinterpret_input_as_3d and reinterpret_output_as_3d to be false. |
| 176 | if(reinterpret_input_as_3d == reinterpret_output_as_3d) |
| 177 | { |
| 178 | reinterpret_input_as_3d = false; |
| 179 | reinterpret_output_as_3d = false; |
| 180 | } |
| 181 | |
| 182 | const unsigned int internal_m = reinterpret_output_as_3d ? gemm_native_desc.m : t_dst_info.dimension(1); |
| 183 | |
| 184 | const unsigned int h_gemm_3d = reinterpret_output_as_3d ? t_dst_info.dimension(1) : t_lhs_info.dimension(1); |
| 185 | const unsigned int d_gemm_3d = reinterpret_output_as_3d ? t_dst_info.dimension(2) : t_lhs_info.dimension(2); |
| 186 | |
| 187 | // 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. |
| 188 | const unsigned int partial_store_m0 = internal_m % gemm_native_desc.lhs_info.m0; |
| 189 | const unsigned int partial_store_n0 = gemm_native_desc.n % gemm_native_desc.rhs_info.n0; |
| 190 | |
| 191 | // Shrink M0 to be always <= M (internal_m) to prevent out-of-bounds reads. |
| 192 | const unsigned int internal_m0 = std::min(internal_m, gemm_native_desc.lhs_info.m0); |
| 193 | |
| 194 | ref_cl_build_options.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(t_dst_info.data_type())); |
| 195 | ref_cl_build_options.add_option_if(!(helpers::float_ops::is_one(gemm_native_desc.alpha)), "-DALPHA=" + float_to_string_with_full_precision(gemm_native_desc.alpha)); |
| 196 | ref_cl_build_options.add_option_if(t_bias_info != nullptr, "-DBETA=" + float_to_string_with_full_precision(gemm_native_desc.beta)); |
| 197 | ref_cl_build_options.add_option_if(helpers::float_ops::is_one(gemm_native_desc.beta), "-DUNIT_BETA"); |
| 198 | ref_cl_build_options.add_option_if(gemm_native_desc.broadcast_bias, "-DBROADCAST_BIAS"); |
| 199 | ref_cl_build_options.add_option_if(reinterpret_input_as_3d, "-DREINTERPRET_INPUT_AS_3D"); |
| 200 | ref_cl_build_options.add_option_if(reinterpret_output_as_3d, "-DREINTERPRET_OUTPUT_AS_3D"); |
| 201 | ref_cl_build_options.add_option_if(reinterpret_input_as_3d || reinterpret_output_as_3d, "-DHEIGHT_GEMM3D=" + support::cpp11::to_string(h_gemm_3d)); |
| 202 | ref_cl_build_options.add_option_if(reinterpret_input_as_3d || reinterpret_output_as_3d, "-DDEPTH_GEMM3D=" + support::cpp11::to_string(d_gemm_3d)); |
| 203 | ref_cl_build_options.add_option_if(!_slide_matrix_b, "-DMATRIX_B_DEPTH=" + support::cpp11::to_string(t_rhs_info.dimension(2))); |
| 204 | ref_cl_build_options.add_option_if(_use_dummy_work_items, "-DDUMMY_WORK_ITEMS"); |
| 205 | ref_cl_build_options.add_option("-DM=" + support::cpp11::to_string(internal_m)); |
| 206 | ref_cl_build_options.add_option("-DN=" + support::cpp11::to_string(gemm_native_desc.n)); |
| 207 | ref_cl_build_options.add_option("-DK=" + support::cpp11::to_string(gemm_native_desc.k)); |
| 208 | ref_cl_build_options.add_option("-DM0=" + support::cpp11::to_string(internal_m0)); |
| 209 | ref_cl_build_options.add_option("-DN0=" + support::cpp11::to_string(gemm_native_desc.rhs_info.n0)); |
| 210 | ref_cl_build_options.add_option("-DK0=" + support::cpp11::to_string(gemm_native_desc.rhs_info.k0)); |
| 211 | ref_cl_build_options.add_option("-DPARTIAL_STORE_M0=" + support::cpp11::to_string(partial_store_m0)); |
| 212 | ref_cl_build_options.add_option("-DPARTIAL_STORE_N0=" + support::cpp11::to_string(partial_store_n0)); |
| 213 | // Manually add PostOps |
| 214 | { |
| 215 | ref_cl_build_options.add_option("-DOP=ADD_X_POS_1"); |
| 216 | ref_cl_build_options.add_option("-DP2_ELTWISE_ARG1_HEIGHT=" + support::cpp11::to_string(t_dst_info.dimension(1))); |
| 217 | ref_cl_build_options.add_option("-DP2_ELTWISE_ARG1_WIDTH=" + support::cpp11::to_string(t_dst_info.dimension(0))); |
| 218 | } |
| 219 | } |
| 220 | cl_code.build_options = ref_cl_build_options; |
| 221 | } |
| 222 | } // namespace |
| 223 | |
| 224 | TEST_SUITE(CL) |
| 225 | TEST_SUITE(UNIT) |
| 226 | TEST_SUITE(DYNAMIC_FUSION) |
| 227 | TEST_SUITE(ClCompositeKernel) |
| 228 | TEST_SUITE(Validate) |
| 229 | |
| 230 | TEST_CASE(MoveNet_SubGraph_1, framework::DatasetMode::ALL) |
| 231 | { |
| 232 | /* Computation: |
| 233 | * out = add(addend, gemm_native(lhs, rhs, bias)) (non-broadcast) |
| 234 | */ |
| 235 | const auto data_type = DataType::F32; |
| 236 | const auto m = 5U; |
| 237 | const auto n = 4U; |
| 238 | const auto k = 3U; |
| 239 | const auto t_lhs_shape = TensorShape(k, m); |
| 240 | const auto t_rhs_shape = TensorShape(n, k); |
| 241 | const auto t_dst_shape = TensorShape(n, m); |
| 242 | auto t_lhs_info = TensorInfo(t_lhs_shape, 1, data_type); |
| 243 | auto t_rhs_info = TensorInfo(t_rhs_shape, 1, data_type); |
| 244 | const auto t_bias_info = TensorInfo(TensorShape(), 1, DataType::F32); |
| 245 | auto t_dst_info = TensorInfo(t_dst_shape, 1, data_type); |
| 246 | |
| 247 | const ClTensorDescriptor t_lhs_desc{ &t_lhs_info, 2 }; |
| 248 | const ClTensorDescriptor t_rhs_desc{ &t_rhs_info, 2 }; |
| 249 | const ClTensorDescriptor t_bias_desc{ &t_bias_info, 2 }; |
| 250 | const ClTensorDescriptor t_addend_desc{ &t_dst_info, 2 }; |
| 251 | const ClTensorDescriptor t_dst_desc{ &t_dst_info, 2 }; |
| 252 | |
| 253 | ClKernelBlueprint bp; |
| 254 | ArgumentID tid_lhs; |
| 255 | ArgumentID tid_rhs; |
| 256 | ArgumentID tid_l0_bias = g_arg_placeholder; |
| 257 | ArgumentID tid_l1_addend; |
| 258 | ArgumentID tid_dst; |
| 259 | auto st = add_tensor_argument(bp, t_lhs_desc, tid_lhs); |
| 260 | st = add_tensor_argument(bp, t_rhs_desc, tid_rhs); |
| 261 | st = add_tensor_argument(bp, t_addend_desc, tid_l1_addend); |
| 262 | st = add_tensor_argument(bp, t_dst_desc, tid_dst); |
| 263 | |
| 264 | const auto common_kernel_desc = ClKernelComponentDescriptor{}; |
| 265 | const GemmNativeDescriptor gemm_native_desc{ 1.0, 1.0, m, n, k }; |
| 266 | const GEMMKernelInfo gemm_info{ m, n, k, 0, false, false, false, false, ActivationLayerInfo{}, 1, 1, gemm_native_desc.lhs_info, gemm_native_desc.rhs_info, 0, 0 }; |
| 267 | const EltwiseAddDescriptor eltwise_add_desc{ ConvertPolicy::WRAP }; |
| 268 | const TileDescriptor store_tile_info{}; |
| 269 | |
| 270 | ArgumentID tid_acc; |
| 271 | st = add_tensor_intermed(bp, tid_acc); |
| 272 | st = add_kcomp_gemm_native(bp, common_kernel_desc, gemm_native_desc, tid_lhs, tid_rhs, tid_l0_bias, tid_acc); |
| 273 | |
| 274 | st = add_kcomp_eltwise_add(bp, common_kernel_desc, EltwiseAddDescriptor{}, tid_l1_addend, tid_acc, tid_acc); |
| 275 | st = add_kcomp_store(bp, common_kernel_desc, tid_acc, tid_dst, StoreType::StoreBlockBoundaryAware); |
| 276 | |
| 277 | ClKernelCode cl_code; |
| 278 | |
| 279 | st = set_tile_info(bp, store_tile_info); |
| 280 | st = build(cl_code, ClCodeBuilderContext{ GpuInfo{ GPUTarget::G71 } }, bp); |
| 281 | |
| 282 | set_build_options(cl_code, gemm_native_desc, t_lhs_info, t_rhs_info, nullptr, t_dst_info); |
| 283 | ElementsProcessed num_elements_processed{}; |
| 284 | auto win_config = mock_gemm_native_validate_and_configure_window(&t_lhs_info, &t_rhs_info, nullptr, &t_dst_info, gemm_native_desc.lhs_info, gemm_native_desc.rhs_info, gemm_info, |
| 285 | num_elements_processed); |
| 286 | ARM_COMPUTE_ERROR_THROW_ON(win_config.first); |
| 287 | cl_code.window = win_config.second; |
| 288 | |
| 289 | ClExecutionDescriptor exec_desc; |
| 290 | st = tune_static(exec_desc, cl_code); |
| 291 | |
| 292 | CLScheduler::get().default_init(); |
| 293 | ClCompositeKernel kernel; |
| 294 | kernel.configure(CLKernelLibrary::get().get_compile_context(), cl_code); |
| 295 | |
| 296 | // Construct tensors |
| 297 | CLTensor t_lhs{}; |
| 298 | CLTensor t_rhs{}; |
| 299 | CLTensor t_l1_addend{}; |
| 300 | CLTensor t_dst{}; |
| 301 | // Init tensors |
| 302 | { |
| 303 | t_lhs.allocator()->init(t_lhs_info); |
| 304 | t_rhs.allocator()->init(t_rhs_info); |
| 305 | t_l1_addend.allocator()->init(t_dst_info); |
| 306 | t_dst.allocator()->init(t_dst_info); |
| 307 | } |
| 308 | // "Pack" tensors |
| 309 | TensorBinding tensors({ { tid_lhs, &t_lhs }, |
| 310 | { tid_rhs, &t_rhs }, |
| 311 | { tid_l1_addend, &t_l1_addend }, |
| 312 | { tid_dst, &t_dst } |
| 313 | }); |
| 314 | // Allocate and fill tensors |
| 315 | { |
| 316 | t_lhs.allocator()->allocate(); |
| 317 | t_rhs.allocator()->allocate(); |
| 318 | t_l1_addend.allocator()->allocate(); |
| 319 | t_dst.allocator()->allocate(); |
| 320 | fill<float>(CLAccessor(t_lhs), 0); |
| 321 | fill<float>(CLAccessor(t_rhs), 1); |
| 322 | fill<float>(CLAccessor(t_l1_addend), 2); |
| 323 | } |
| 324 | |
| 325 | CLScheduler::get().enqueue_op(kernel, tensors, exec_desc, true); |
| 326 | |
| 327 | // Create reference |
| 328 | SimpleTensor<float> ref_t_lhs{ t_lhs_shape, data_type, 1 }; |
| 329 | SimpleTensor<float> ref_t_rhs{ t_rhs_shape, data_type, 1 }; |
| 330 | SimpleTensor<float> ref_t_bias_placeholder{ t_dst_shape, data_type, 1 }; |
| 331 | SimpleTensor<float> ref_t_l1_addend{ t_dst_shape, data_type, 1 }; |
| 332 | |
| 333 | // Fill reference |
| 334 | fill<float>(ref_t_lhs, 0); |
| 335 | fill<float>(ref_t_rhs, 1); |
| 336 | fill<float>(ref_t_l1_addend, 2); |
| 337 | const auto ref_t_dst = reference::arithmetic_operation( |
| 338 | ArithmeticOperation::ADD, |
| 339 | ref_t_l1_addend, |
| 340 | reference::gemm(ref_t_lhs, ref_t_rhs, ref_t_bias_placeholder, gemm_native_desc.alpha, 0.f /* To disable bias */), |
| 341 | data_type, |
| 342 | eltwise_add_desc.convert_policy); |
| 343 | |
| 344 | RelativeTolerance<float> tolerance_f32(0.001f); /**< Tolerance value for comparing reference's output against implementation's output for floating point data types */ |
| 345 | validate(CLAccessor(t_dst), ref_t_dst, tolerance_f32); |
| 346 | } |
| 347 | |
| 348 | TEST_SUITE_END() // Validate |
| 349 | |
| 350 | TEST_SUITE(Benchmark) |
| 351 | TEST_CASE(MoveNet_SubGraph_1, framework::DatasetMode::ALL) |
| 352 | { |
| 353 | using std::chrono::duration_cast; |
| 354 | using std::chrono::microseconds; |
| 355 | const int num_iterations = 200; |
| 356 | std::map<std::string, std::chrono::microseconds> measurements; |
| 357 | /* Computation: |
| 358 | * out = add(addend, gemm_native(lhs, rhs, bias)) |
| 359 | */ |
| 360 | const auto data_type = DataType::F32; |
| 361 | const unsigned int m = 12 * 12; |
| 362 | const unsigned int n = 64; |
| 363 | const unsigned int k = 384; |
| 364 | const auto t_lhs_shape = TensorShape(k, m); |
| 365 | const auto t_rhs_shape = TensorShape(n, k); |
| 366 | const auto t_dst_shape = TensorShape(n, m); |
| 367 | auto t_lhs_info = TensorInfo(t_lhs_shape, 1, data_type); |
| 368 | auto t_rhs_info = TensorInfo(t_rhs_shape, 1, data_type); |
| 369 | auto t_bias_info = TensorInfo(TensorShape(), 1, data_type); |
| 370 | auto t_l0_dst_info = TensorInfo(t_dst_shape, 1, data_type); // Intermediate tensor for cond3 |
| 371 | auto t_l1_rhs_info = TensorInfo(t_dst_shape, 1, data_type); |
| 372 | auto t_dst_info = TensorInfo(t_dst_shape, 1, data_type); |
| 373 | |
| 374 | const auto common_kernel_desc = ClKernelComponentDescriptor{}; |
| 375 | const GemmNativeDescriptor gemm_native_desc{ 1.0, 0.0, m, n, k }; |
| 376 | const GEMMKernelInfo gemm_info{ m, n, k, 0, false, false, false, false, ActivationLayerInfo{}, 1, 1, gemm_native_desc.lhs_info, gemm_native_desc.rhs_info, 0, 0 }; |
| 377 | const EltwiseAddDescriptor eltwise_add_desc{ ConvertPolicy::WRAP }; |
| 378 | const TileDescriptor store_tile_info{}; |
| 379 | |
| 380 | // Create reference |
| 381 | SimpleTensor<float> ref_t_lhs{ t_lhs_shape, data_type, 1 }; |
| 382 | SimpleTensor<float> ref_t_rhs{ t_rhs_shape, data_type, 1 }; |
| 383 | SimpleTensor<float> ref_t_bias_placeholder{ t_dst_shape, data_type, 1 }; |
| 384 | SimpleTensor<float> ref_t_l1_addend{ t_dst_shape, data_type, 1 }; |
| 385 | |
| 386 | // Fill reference |
| 387 | fill<float>(ref_t_lhs, 0); |
| 388 | fill<float>(ref_t_rhs, 1); |
| 389 | fill<float>(ref_t_l1_addend, 2); |
| 390 | const auto ref_t_dst = reference::arithmetic_operation( |
| 391 | ArithmeticOperation::ADD, |
| 392 | ref_t_l1_addend, |
| 393 | reference::gemm(ref_t_lhs, ref_t_rhs, ref_t_bias_placeholder, gemm_native_desc.alpha, 0.f /* To disable bias */), |
| 394 | data_type, |
| 395 | eltwise_add_desc.convert_policy); |
| 396 | |
| 397 | CLScheduler::get().default_init(); |
| 398 | |
| 399 | /* Condition 0: Dynamic Fused Kernel */ |
| 400 | CLTensor cond0_t_dst{}; |
| 401 | { |
| 402 | TICK(cond0_0_startup_time); |
| 403 | |
| 404 | ClKernelBlueprint bp; |
| 405 | ArgumentID tid_lhs; |
| 406 | ArgumentID tid_rhs; |
| 407 | ArgumentID tid_l0_bias = g_arg_placeholder; |
| 408 | ArgumentID tid_l1_addend; |
| 409 | ArgumentID tid_dst; |
| 410 | |
| 411 | const ClTensorDescriptor t_lhs_desc{ &t_lhs_info, 2 }; |
| 412 | const ClTensorDescriptor t_rhs_desc{ &t_rhs_info, 2 }; |
| 413 | const ClTensorDescriptor t_bias_desc{ &t_bias_info, 2 }; |
| 414 | const ClTensorDescriptor t_addend_desc{ &t_dst_info, 2 }; |
| 415 | const ClTensorDescriptor t_dst_desc{ &t_dst_info, 2 }; |
| 416 | |
| 417 | ClKernelCode cl_code; |
| 418 | TICK(cond0_build_time) |
| 419 | auto st = add_tensor_argument(bp, t_lhs_desc, tid_lhs); |
| 420 | st = add_tensor_argument(bp, t_rhs_desc, tid_rhs); |
| 421 | st = add_tensor_argument(bp, t_addend_desc, tid_l1_addend); |
| 422 | st = add_tensor_argument(bp, t_dst_desc, tid_dst); |
| 423 | |
| 424 | ArgumentID tid_acc; |
| 425 | st = add_tensor_intermed(bp, tid_acc); |
| 426 | st = add_kcomp_gemm_native(bp, common_kernel_desc, gemm_native_desc, tid_lhs, tid_rhs, tid_l0_bias, tid_acc); |
| 427 | |
| 428 | st = add_kcomp_eltwise_add(bp, common_kernel_desc, EltwiseAddDescriptor{}, tid_l1_addend, tid_acc, tid_acc); |
| 429 | |
| 430 | st = add_kcomp_store(bp, common_kernel_desc, tid_acc, tid_dst, StoreType::StoreBlockBoundaryAware); |
| 431 | |
| 432 | st = set_tile_info(bp, store_tile_info); |
| 433 | st = build(cl_code, ClCodeBuilderContext{ GpuInfo{ GPUTarget::G71 } }, bp); |
| 434 | set_build_options(cl_code, gemm_native_desc, t_lhs_info, t_rhs_info, nullptr, t_dst_info); |
| 435 | ElementsProcessed num_elements_processed{}; |
| 436 | auto win_config = mock_gemm_native_validate_and_configure_window(&t_lhs_info, &t_rhs_info, nullptr, &t_dst_info, gemm_native_desc.lhs_info, gemm_native_desc.rhs_info, gemm_info, |
| 437 | num_elements_processed); |
| 438 | ARM_COMPUTE_ERROR_THROW_ON(win_config.first); |
| 439 | cl_code.window = win_config.second; |
| 440 | TOCK(cond0_build_time, measurements) |
| 441 | |
| 442 | TICK(cond0_tune_time) |
| 443 | ClExecutionDescriptor exec_desc; |
| 444 | st = tune_static(exec_desc, cl_code); |
| 445 | TOCK(cond0_tune_time, measurements) |
| 446 | |
| 447 | TICK(cond0_configure_time) |
| 448 | ClCompositeKernel kernel; |
| 449 | kernel.configure(CLKernelLibrary::get().get_compile_context(), cl_code); |
| 450 | TOCK(cond0_configure_time, measurements) |
| 451 | |
| 452 | // Construct tensors |
| 453 | CLTensor t_lhs{}; |
| 454 | CLTensor t_rhs{}; |
| 455 | CLTensor t_l1_addend{}; |
| 456 | |
| 457 | // Init tensors |
| 458 | { |
| 459 | t_lhs.allocator()->init(t_lhs_info); |
| 460 | t_rhs.allocator()->init(t_rhs_info); |
| 461 | t_l1_addend.allocator()->init(t_dst_info); |
| 462 | cond0_t_dst.allocator()->init(t_dst_info); |
| 463 | } |
| 464 | // Allocate tensors |
| 465 | { |
| 466 | t_lhs.allocator()->allocate(); |
| 467 | t_rhs.allocator()->allocate(); |
| 468 | t_l1_addend.allocator()->allocate(); |
| 469 | cond0_t_dst.allocator()->allocate(); |
| 470 | fill<float>(CLAccessor(t_lhs), 0); |
| 471 | fill<float>(CLAccessor(t_rhs), 1); |
| 472 | fill<float>(CLAccessor(t_l1_addend), 2); |
| 473 | } |
| 474 | |
| 475 | // "Pack" tensors |
| 476 | TensorBinding tensors({ { tid_lhs, &t_lhs }, { tid_rhs, &t_rhs }, { tid_l1_addend, &t_l1_addend }, { tid_dst, &cond0_t_dst } }); |
| 477 | |
| 478 | CLScheduler::get().enqueue_op(kernel, tensors, exec_desc, true); |
| 479 | CLScheduler::get().sync(); |
| 480 | TOCK(cond0_0_startup_time, measurements) |
| 481 | |
| 482 | TICK(cond0_1_latency) |
| 483 | for(int i = 0; i < num_iterations; ++i) |
| 484 | { |
| 485 | CLScheduler::get().enqueue_op(kernel, tensors, exec_desc, true); |
| 486 | } |
| 487 | CLScheduler::get().sync(); |
| 488 | TOCK_AVG(cond0_1_latency, measurements, num_iterations) |
| 489 | } |
| 490 | /* Condition 1: Dynamic Unfused Kernel */ |
| 491 | /* Condition 2: Static Fused Kernel (current) */ |
| 492 | CLTensor cond2_t_dst{}; |
| 493 | { |
| 494 | TICK(cond2_0_startup_time); |
| 495 | arm_compute::opencl::kernels::ClGemmMatrixMultiplyNativeKernel l0_gemm_mm; |
| 496 | |
| 497 | TICK(cond2_configure_time); |
| 498 | experimental::PostOpList<ITensorInfo *> post_ops; |
| 499 | post_ops.push_back_op<experimental::PostOpEltwiseAdd<ITensorInfo *>>(&t_dst_info, 1, eltwise_add_desc.convert_policy); |
| 500 | GEMMKernelInfo gemm_info{ m, n, k, 0, false, false, false, false, ActivationLayerInfo{}, 1, 1, gemm_native_desc.lhs_info, gemm_native_desc.rhs_info, 0, 0, post_ops }; |
| 501 | l0_gemm_mm.configure(CLKernelLibrary::get().get_compile_context(), &t_lhs_info, &t_rhs_info, nullptr, &t_dst_info, gemm_native_desc.alpha, gemm_native_desc.beta, gemm_native_desc.lhs_info, |
| 502 | gemm_native_desc.rhs_info, gemm_info); |
| 503 | TOCK(cond2_configure_time, measurements); |
| 504 | |
| 505 | // Construct tensors |
| 506 | CLTensor t_lhs{}; |
| 507 | CLTensor t_rhs{}; |
| 508 | CLTensor t_l1_addend{}; |
| 509 | |
| 510 | // Init tensors |
| 511 | { |
| 512 | t_lhs.allocator()->init(t_lhs_info); |
| 513 | t_rhs.allocator()->init(t_rhs_info); |
| 514 | t_l1_addend.allocator()->init(t_dst_info); |
| 515 | cond2_t_dst.allocator()->init(t_dst_info); |
| 516 | } |
| 517 | // Allocate tensors |
| 518 | { |
| 519 | t_lhs.allocator()->allocate(); |
| 520 | t_rhs.allocator()->allocate(); |
| 521 | t_l1_addend.allocator()->allocate(); |
| 522 | cond2_t_dst.allocator()->allocate(); |
| 523 | fill<float>(CLAccessor(t_lhs), 0); |
| 524 | fill<float>(CLAccessor(t_rhs), 1); |
| 525 | fill<float>(CLAccessor(t_l1_addend), 2); |
| 526 | } |
| 527 | |
| 528 | // "Pack" tensors |
| 529 | ITensorPack tensors |
| 530 | { |
| 531 | { ACL_SRC_0, &t_lhs }, |
| 532 | { ACL_SRC_1, &t_rhs }, |
| 533 | { EXPERIMENTAL_ACL_POST_OP_ARG_FIRST, &t_l1_addend }, |
| 534 | { ACL_DST, &cond2_t_dst }, |
| 535 | }; |
| 536 | CLScheduler::get().enqueue_op(l0_gemm_mm, tensors, true); |
| 537 | CLScheduler::get().sync(); |
| 538 | TOCK(cond2_0_startup_time, measurements); |
| 539 | |
| 540 | TICK(cond2_1_latency); |
| 541 | for(int i = 0; i < num_iterations; ++i) |
| 542 | { |
| 543 | CLScheduler::get().enqueue_op(l0_gemm_mm, tensors, true); |
| 544 | } |
| 545 | CLScheduler::get().sync(); |
| 546 | TOCK_AVG(cond2_1_latency, measurements, num_iterations); |
| 547 | } |
| 548 | /* Condition 3: Static Unfused Kernel (current) */ |
| 549 | CLTensor cond3_t_dst{}; |
| 550 | { |
| 551 | TICK(cond3_0_startup_time); |
| 552 | arm_compute::opencl::kernels::ClGemmMatrixMultiplyNativeKernel l0_gemm_mm; |
| 553 | arm_compute::opencl::kernels::ClSaturatedArithmeticKernel l1_add; |
| 554 | |
| 555 | TICK(cond3_configure_time); |
| 556 | GEMMKernelInfo gemm_info{ m, n, k, 0, false, false, false, false, ActivationLayerInfo{}, 1, 1, gemm_native_desc.lhs_info, gemm_native_desc.rhs_info, 0, 0 }; |
| 557 | l0_gemm_mm.configure(CLKernelLibrary::get().get_compile_context(), &t_lhs_info, &t_rhs_info, nullptr, &t_l0_dst_info, gemm_native_desc.alpha, gemm_native_desc.beta, gemm_native_desc.lhs_info, |
| 558 | gemm_native_desc.rhs_info, gemm_info); |
| 559 | l1_add.configure(CLKernelLibrary::get().get_compile_context(), ArithmeticOperation::ADD, &t_l0_dst_info, &t_l1_rhs_info, &t_dst_info, eltwise_add_desc.convert_policy); |
| 560 | TOCK(cond3_configure_time, measurements); |
| 561 | |
| 562 | // Construct tensors |
| 563 | CLTensor t_lhs{}; |
| 564 | CLTensor t_rhs{}; |
| 565 | CLTensor t_l0_dst{}; |
| 566 | CLTensor t_l1_addend{}; |
| 567 | |
| 568 | // Init tensors |
| 569 | { |
| 570 | t_lhs.allocator()->init(t_lhs_info); |
| 571 | t_rhs.allocator()->init(t_rhs_info); |
| 572 | t_l0_dst.allocator()->init(t_l0_dst_info); |
| 573 | t_l1_addend.allocator()->init(t_dst_info); |
| 574 | cond3_t_dst.allocator()->init(t_dst_info); |
| 575 | } |
| 576 | // Allocate tensors |
| 577 | { |
| 578 | t_lhs.allocator()->allocate(); |
| 579 | t_rhs.allocator()->allocate(); |
| 580 | t_l0_dst.allocator()->allocate(); |
| 581 | t_l1_addend.allocator()->allocate(); |
| 582 | cond3_t_dst.allocator()->allocate(); |
| 583 | fill<float>(CLAccessor(t_lhs), 0); |
| 584 | fill<float>(CLAccessor(t_rhs), 1); |
| 585 | fill<float>(CLAccessor(t_l1_addend), 2); |
| 586 | } |
| 587 | |
| 588 | // "Pack" tensors |
| 589 | ITensorPack tensors_l0 |
| 590 | { |
| 591 | { ACL_SRC_0, &t_lhs }, |
| 592 | { ACL_SRC_1, &t_rhs }, |
| 593 | { ACL_DST, &t_l0_dst }, |
| 594 | }; |
| 595 | ITensorPack tensors_l1 |
| 596 | { |
| 597 | { ACL_SRC_0, &t_l0_dst }, |
| 598 | { ACL_SRC_1, &t_l1_addend }, |
| 599 | { ACL_DST, &cond3_t_dst }, |
| 600 | }; |
| 601 | CLScheduler::get().enqueue_op(l0_gemm_mm, tensors_l0, true); |
| 602 | CLScheduler::get().enqueue_op(l1_add, tensors_l1, true); |
| 603 | CLScheduler::get().sync(); |
| 604 | TOCK(cond3_0_startup_time, measurements); |
| 605 | |
| 606 | TICK(cond3_1_latency); |
| 607 | for(int i = 0; i < num_iterations; ++i) |
| 608 | { |
| 609 | CLScheduler::get().enqueue_op(l0_gemm_mm, tensors_l0, true); |
| 610 | CLScheduler::get().enqueue_op(l1_add, tensors_l1, true); |
| 611 | } |
| 612 | CLScheduler::get().sync(); |
| 613 | TOCK_AVG(cond3_1_latency, measurements, num_iterations); |
| 614 | } |
| 615 | |
| 616 | RelativeTolerance<float> tolerance_f32(0.001f); /**< Tolerance value for comparing reference's output against implementation's output for floating point data types */ |
| 617 | std::cout << "cond0 validation: " << std::endl; |
| 618 | validate(CLAccessor(cond0_t_dst), ref_t_dst, tolerance_f32); |
| 619 | std::cout << "cond2 validation: " << std::endl; |
| 620 | validate(CLAccessor(cond2_t_dst), ref_t_dst, tolerance_f32); |
| 621 | std::cout << "cond3 validation: " << std::endl; |
| 622 | validate(CLAccessor(cond3_t_dst), ref_t_dst, tolerance_f32); |
| 623 | |
| 624 | /* Report */ |
| 625 | std::cout << "Performance comparison (gemm native + add)" << std::endl; |
| 626 | std::cout << "cond0: dynamic fusion module" << std::endl; |
| 627 | std::cout << "cond2: static fused with post ops" << std::endl; |
| 628 | std::cout << "cond3: static unfused" << std::endl; |
| 629 | for(auto m : measurements) |
| 630 | { |
| 631 | std::cout << m.first << ": " << m.second.count() << "us" << std::endl; |
| 632 | } |
| 633 | } |
| 634 | TEST_SUITE_END() // Benchmark |
| 635 | TEST_SUITE_END() // ClCompositeKernel |
| 636 | TEST_SUITE_END() // DYNAMIC_FUSION |
| 637 | TEST_SUITE_END() // UNIT |
| 638 | TEST_SUITE_END() // CL |
| 639 | } // namespace validation |
| 640 | } // namespace test |
| 641 | } // namespace arm_compute |
| 642 | |
| 643 | #endif // defined(ENABLE_EXPERIMENTAL_DYNAMIC_FUSION) |