Gunes Bayir | 16c5697 | 2022-03-28 21:32:33 +0100 | [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 | */ |
SiCong Li | 4e9f568 | 2022-05-10 10:15:59 +0100 | [diff] [blame] | 24 | #ifdef ENABLE_EXPERIMENTAL_DYNAMIC_FUSION |
Gunes Bayir | 16c5697 | 2022-03-28 21:32:33 +0100 | [diff] [blame] | 25 | |
| 26 | #include "src/core/experimental/dynamic_fusion/ClKernelBuildingImpl/components/ClDirectConvolutionKernelComponent.h" |
| 27 | |
| 28 | #include "arm_compute/core/utils/misc/ShapeCalculator.h" |
| 29 | #include "src/core/CL/ICLKernel.h" |
| 30 | #include "src/core/helpers/AutoConfiguration.h" |
| 31 | #include "src/core/helpers/WindowHelpers.h" |
| 32 | #include "src/gpu/cl/kernels/gemm/ClGemmHelpers.h" |
| 33 | |
SiCong Li | b63b119 | 2022-01-28 18:24:39 +0000 | [diff] [blame] | 34 | #include "arm_compute/runtime/CL/CLScheduler.h" |
Gunes Bayir | 16c5697 | 2022-03-28 21:32:33 +0100 | [diff] [blame] | 35 | namespace arm_compute |
| 36 | { |
| 37 | namespace experimental |
| 38 | { |
| 39 | namespace dynamic_fusion |
| 40 | { |
| 41 | ComponentType ClDirectConvolutionKernelComponent::get_component_type() const |
| 42 | { |
| 43 | return ComponentType::Complex; |
| 44 | } |
| 45 | |
| 46 | std::set<std::string> ClDirectConvolutionKernelComponent::get_headers_list() const |
| 47 | { |
SiCong Li | b63b119 | 2022-01-28 18:24:39 +0000 | [diff] [blame] | 48 | return std::set<std::string> { "helpers.h", "tile_helpers.h" }; |
Gunes Bayir | 16c5697 | 2022-03-28 21:32:33 +0100 | [diff] [blame] | 49 | } |
| 50 | |
| 51 | Window ClDirectConvolutionKernelComponent::get_window() const |
| 52 | { |
| 53 | const auto src_info = _blueprint->impl().get_kernel_argument_info(_src.arg_id); |
| 54 | const auto weight_info = _blueprint->impl().get_kernel_argument_info(_weight.arg_id); |
| 55 | auto dst_info = _blueprint->impl().get_kernel_argument_info(_blueprint->impl().get_dst_id()); |
| 56 | |
| 57 | // Get dst shape |
SiCong Li | b63b119 | 2022-01-28 18:24:39 +0000 | [diff] [blame] | 58 | PadStrideInfo pad_stride_info |
| 59 | { |
| 60 | static_cast<unsigned int>(_desc.conv2d.stride.x()), |
| 61 | static_cast<unsigned int>(_desc.conv2d.stride.y()), |
| 62 | static_cast<unsigned int>(_desc.conv2d.pad.left), |
| 63 | static_cast<unsigned int>(_desc.conv2d.pad.right), |
| 64 | static_cast<unsigned int>(_desc.conv2d.pad.top), |
| 65 | static_cast<unsigned int>(_desc.conv2d.pad.bottom), |
| 66 | DimensionRoundingType::FLOOR /*default rounding type*/ |
| 67 | }; |
| 68 | TensorShape output_shape = misc::shape_calculator::compute_deep_convolution_shape(*src_info, *weight_info, pad_stride_info); |
Gunes Bayir | 16c5697 | 2022-03-28 21:32:33 +0100 | [diff] [blame] | 69 | |
| 70 | // Output auto initialization if not yet initialized |
| 71 | auto_init_if_empty(*dst_info, output_shape, |
| 72 | 1, |
| 73 | src_info->data_type(), |
| 74 | src_info->quantization_info()); |
| 75 | |
| 76 | const unsigned int vec_size = std::min(static_cast<unsigned int>(dst_info->tensor_shape()[0]), 4u); |
| 77 | const unsigned int num_rows = (dst_info->tensor_shape()[0] > 16) ? ((src_info->data_type() == DataType::F32) ? 2U : 4U) : 1U; |
SiCong Li | b63b119 | 2022-01-28 18:24:39 +0000 | [diff] [blame] | 78 | // const unsigned int num_rows = 1; |
| 79 | // const unsigned int vec_size = tile_info.tile_dims.x(); |
| 80 | // const unsigned int num_rows = tile_info.tile_dims.y(); |
Gunes Bayir | 16c5697 | 2022-03-28 21:32:33 +0100 | [diff] [blame] | 81 | |
| 82 | // Create and configure kernel window |
| 83 | Window win = calculate_max_window(output_shape, Steps(vec_size, num_rows)); |
| 84 | |
| 85 | const size_t dim_y_collapsed = ceil_to_multiple(output_shape[1] * output_shape[2], num_rows); |
| 86 | win.set(Window::DimY, Window::Dimension(0, dim_y_collapsed, num_rows)); |
| 87 | win.set(Window::DimZ, Window::Dimension(0, output_shape.total_size_upper(3), 1)); |
| 88 | |
| 89 | return win; |
| 90 | } |
| 91 | |
| 92 | std::string ClDirectConvolutionKernelComponent::get_additional_macros() const |
| 93 | { |
| 94 | return R"_()_"; // no macros |
| 95 | } |
| 96 | |
| 97 | std::string ClDirectConvolutionKernelComponent::get_component_code() const |
| 98 | { |
| 99 | const auto src_info = _blueprint->impl().get_kernel_argument_info(_src.arg_id); |
| 100 | const auto bias_info = _blueprint->impl().get_kernel_argument_info(_bias.arg_id); |
| 101 | |
| 102 | ARM_COMPUTE_ERROR_ON_MSG(src_info->data_layout() != DataLayout::NHWC, "Only NHWC data layout is supported by this component."); |
| 103 | |
| 104 | const auto channel_idx = get_data_layout_dimension_index(src_info->data_layout(), DataLayoutDimension::CHANNEL); |
| 105 | const auto k0 = adjust_vec_size(is_data_type_quantized(src_info->data_type()) ? 16u : 8u, src_info->dimension(channel_idx)); |
| 106 | const bool leftover_loop = (src_info->dimension(channel_idx) % k0) != 0; |
| 107 | |
| 108 | std::string code = R"_( |
| 109 | //------------------ START KERNEL {{meta_kernel_id}} --------------------- |
| 110 | // IN_0(src) {{src}} |
| 111 | // IN_1(wei) {{weight}} |
SiCong Li | b63b119 | 2022-01-28 18:24:39 +0000 | [diff] [blame] | 112 | )_"; |
| 113 | if(bias_info != nullptr) |
| 114 | { |
| 115 | code += R"_( |
Gunes Bayir | 16c5697 | 2022-03-28 21:32:33 +0100 | [diff] [blame] | 116 | // IN_1(bia) {{bias}} |
SiCong Li | b63b119 | 2022-01-28 18:24:39 +0000 | [diff] [blame] | 117 | )_"; |
| 118 | } |
| 119 | code += R"_( |
Gunes Bayir | 16c5697 | 2022-03-28 21:32:33 +0100 | [diff] [blame] | 120 | // OUT(dst, accum) {{dst}} |
| 121 | |
Gunes Bayir | 16c5697 | 2022-03-28 21:32:33 +0100 | [diff] [blame] | 122 | // Initialize the accumulators |
| 123 | TILE({{ACC_DATA_TYPE}}, M0, N0, {{dst}}); |
| 124 | { |
| 125 | // All the tensor dimensions are passed at compile time. |
| 126 | // In case of dynamic tensor support, the following dimensions should be passed as function argument. |
SiCong Li | b63b119 | 2022-01-28 18:24:39 +0000 | [diff] [blame] | 127 | #define _IWEI_WIDTH {{WEI_WIDTH}} |
| 128 | #define _IWEI_HEIGHT {{WEI_HEIGHT}} |
Gunes Bayir | 16c5697 | 2022-03-28 21:32:33 +0100 | [diff] [blame] | 129 | #define _ISRC_WIDTH {{src}}_w |
| 130 | #define _ISRC_HEIGHT {{src}}_h |
| 131 | #define _ISRC_CHANNELS {{src}}_c |
SiCong Li | b63b119 | 2022-01-28 18:24:39 +0000 | [diff] [blame] | 132 | #define _IDST_WIDTH {{arg_dst}}_w |
| 133 | #define _IDST_HEIGHT {{arg_dst}}_h |
| 134 | #define _IDST_CHANNELS {{arg_dst}}_c |
| 135 | #define _IY_MULTIPLIER (_IWEI_WIDTH * _IWEI_HEIGHT) |
Gunes Bayir | 16c5697 | 2022-03-28 21:32:33 +0100 | [diff] [blame] | 136 | |
| 137 | // .v = access the whole vector (OpenCL vector) |
| 138 | // .s[x] = access the vector element at position x (scalar access) |
| 139 | TILE(int, M0, 1, xi); |
| 140 | TILE(int, M0, 1, yi); |
| 141 | |
| 142 | // Convert the linear index to coordinate |
| 143 | LOOP_UNROLLING(int, i, 0, 1, M0, |
| 144 | { |
| 145 | xi[i].v = ((mout + i) % _IDST_WIDTH) * {{STRIDE_X}}; |
| 146 | yi[i].v = ((mout + i) / _IDST_WIDTH) * {{STRIDE_Y}}; |
| 147 | xi[i].v -= {{PAD_LEFT}}; |
| 148 | yi[i].v -= {{PAD_TOP}}; |
| 149 | }) |
| 150 | |
| 151 | LOOP_UNROLLING(int, i, 0, 1, M0, |
| 152 | { |
| 153 | {{dst}}[i].v = 0; |
| 154 | }) |
| 155 | |
SiCong Li | b63b119 | 2022-01-28 18:24:39 +0000 | [diff] [blame] | 156 | for(int i = 0; i < (_IWEI_WIDTH * _IWEI_HEIGHT); ++i) |
Gunes Bayir | 16c5697 | 2022-03-28 21:32:33 +0100 | [diff] [blame] | 157 | { |
| 158 | int ck = 0; |
SiCong Li | b63b119 | 2022-01-28 18:24:39 +0000 | [diff] [blame] | 159 | int xk = i % _IWEI_WIDTH; |
| 160 | int yk = i / _IWEI_HEIGHT; |
Gunes Bayir | 16c5697 | 2022-03-28 21:32:33 +0100 | [diff] [blame] | 161 | |
| 162 | int k = 0; |
| 163 | for(; k <= (_ISRC_CHANNELS - K0); k += K0) |
| 164 | { |
| 165 | TILE({{SRC_DATA_TYPE}}, M0, K0, a); |
| 166 | TILE({{WEI_DATA_TYPE}}, N0, K0, b); |
| 167 | |
| 168 | LOOP_UNROLLING(int, i, 0, 1, M0, |
| 169 | { |
| 170 | a[i].v = {{ZERO_VALUE}}; |
| 171 | }) |
| 172 | |
| 173 | // Load tile from the src tensor |
| 174 | T_LOAD_NHWC_INDIRECT({{SRC_DATA_TYPE}}, M0, K0, {{SRC_TENSOR_TYPE}}, {{src}}, bout, yk, xk, ck, _ISRC_WIDTH, _ISRC_HEIGHT, {{src}}_stride_y, xi, yi, a); |
| 175 | |
| 176 | // Load tile from the weights tensor |
| 177 | T_LOAD({{WEI_DATA_TYPE}}, N0, K0, {{WEI_TENSOR_TYPE}}, {{weight}}, ck, cout * _IY_MULTIPLIER + i, _IY_MULTIPLIER, {{weight}}_stride_y, b); |
| 178 | |
| 179 | // Compute the matrix multiplication between two tiles |
| 180 | T_MMUL({{SRC_DATA_TYPE}}, {{WEI_DATA_TYPE}}, {{ACC_DATA_TYPE}}, M0, N0, K0, NT, T, a, b, {{dst}}); |
| 181 | |
| 182 | ck += K0; |
| 183 | } |
| 184 | |
| 185 | // We voluntarily use SRC_CHANNELS rather than _DSRC_CHANNELS |
| 186 | // This #if directive should be removed in case of dynamic tensor support |
| 187 | )_"; |
| 188 | |
| 189 | if(leftover_loop) |
| 190 | { |
| 191 | code += R"_( |
| 192 | // Left-over accumulations |
| 193 | for(; k < _ISRC_CHANNELS; ++k) |
| 194 | { |
| 195 | TILE({{SRC_DATA_TYPE}}, M0, 1, a); |
| 196 | TILE({{WEI_DATA_TYPE}}, N0, 1, b); |
| 197 | |
| 198 | LOOP_UNROLLING(int, i, 0, 1, M0, |
| 199 | { |
| 200 | a[i].v = {{ZERO_VALUE}}; |
| 201 | }) |
| 202 | |
| 203 | // Load tile from the src tensor |
| 204 | T_LOAD_NHWC_INDIRECT({{SRC_DATA_TYPE}}, M0, 1, {{SRC_TENSOR_TYPE}}, {{src}}, bout, yk, xk, ck, _ISRC_WIDTH, _ISRC_HEIGHT, {{src}}_stride_y, xi, yi, a); |
| 205 | |
| 206 | // Load tile from the weights tensor |
| 207 | // The T_LOAD for the left-over elements can only use BUFFER because we load one element per iteration |
| 208 | T_LOAD({{WEI_DATA_TYPE}}, N0, 1, BUFFER, {{weight}}, ck, cout * _IY_MULTIPLIER + i, _IY_MULTIPLIER, {{weight}}_stride_y, b); |
| 209 | |
| 210 | // Compute the matrix multiplication between two tiles |
| 211 | T_MMUL({{SRC_DATA_TYPE}}, {{WEI_DATA_TYPE}}, {{ACC_DATA_TYPE}}, M0, N0, 1, NT, T, a, b, {{dst}}); |
| 212 | |
| 213 | ++ck; |
| 214 | } |
| 215 | )_"; |
| 216 | } |
| 217 | |
| 218 | code += R"_( |
SiCong Li | b63b119 | 2022-01-28 18:24:39 +0000 | [diff] [blame] | 219 | #undef _I_WEI_WIDTH |
| 220 | #undef _I_WEI_HEIGHT |
| 221 | #undef _ISRC_WIDTH |
| 222 | #undef _ISRC_HEIGHT |
| 223 | #undef _ISRC_CHANNELS |
| 224 | #undef _IDST_WIDTH |
| 225 | #undef _IDST_HEIGHT |
| 226 | #undef _IDST_CHANNELS |
| 227 | #undef _IY_MULTIPLIER |
| 228 | |
Gunes Bayir | 16c5697 | 2022-03-28 21:32:33 +0100 | [diff] [blame] | 229 | } |
| 230 | )_"; |
| 231 | |
| 232 | if(bias_info != nullptr) |
| 233 | { |
| 234 | code += R"_( |
| 235 | TILE({{BIA_DATA_TYPE}}, 1, N0, bias0); |
| 236 | |
| 237 | T_LOAD({{BIA_DATA_TYPE}}, 1, N0, BUFFER, {{bias}}, cout, 0, 1, 0, bias0); |
| 238 | |
| 239 | // c = c + bias[broadcasted] |
Michalis Spyrou | b1fcefd | 2022-06-15 19:02:28 +0100 | [diff] [blame] | 240 | T_ELTWISE_BROADCAST_ADD_X({{ACC_DATA_TYPE}}, M0, N0, {{dst}}, bias0, {{dst}}); |
Gunes Bayir | 16c5697 | 2022-03-28 21:32:33 +0100 | [diff] [blame] | 241 | )_"; |
| 242 | } |
| 243 | |
| 244 | code += R"_( |
Gunes Bayir | 16c5697 | 2022-03-28 21:32:33 +0100 | [diff] [blame] | 245 | } |
Gunes Bayir | 16c5697 | 2022-03-28 21:32:33 +0100 | [diff] [blame] | 246 | //------------------ END KERNEL {{meta_kernel_id}} --------------------- |
| 247 | )_"; |
| 248 | return code.c_str(); |
| 249 | } |
| 250 | |
| 251 | bool export_to_cl_image_support(const ITensorInfo *tensor, GPUTarget gpu_target, DataLayout data_layout) |
| 252 | { |
| 253 | if(tensor->tensor_shape()[0] % 4 || (data_layout != DataLayout::NHWC)) |
| 254 | { |
| 255 | return false; |
| 256 | } |
| 257 | |
| 258 | // If not floating point |
| 259 | if(!is_data_type_float(tensor->data_type())) |
| 260 | { |
| 261 | return false; |
| 262 | } |
| 263 | |
| 264 | if(gpu_target == GPUTarget::G71 || get_arch_from_target(gpu_target) == GPUTarget::MIDGARD) |
| 265 | { |
| 266 | return false; |
| 267 | } |
| 268 | |
| 269 | // Check if the cl_khr_image2d_from_buffer extension is supported on the target platform |
| 270 | if(!image2d_from_buffer_supported(CLKernelLibrary::get().get_device())) |
| 271 | { |
| 272 | return false; |
| 273 | } |
| 274 | |
| 275 | // Check cl image pitch alignment |
| 276 | if(get_cl_image_pitch_alignment(CLKernelLibrary::get().get_device()) == 0) |
| 277 | { |
| 278 | return false; |
| 279 | } |
| 280 | |
| 281 | const size_t image_w = tensor->tensor_shape()[0] / 4; |
| 282 | const size_t image_h = tensor->tensor_shape()[1] * tensor->tensor_shape()[2] * tensor->tensor_shape()[3]; |
| 283 | const size_t max_image_w = CLKernelLibrary::get().get_device().getInfo<CL_DEVICE_IMAGE2D_MAX_WIDTH>(); |
| 284 | const size_t max_image_h = CLKernelLibrary::get().get_device().getInfo<CL_DEVICE_IMAGE2D_MAX_HEIGHT>(); |
| 285 | |
| 286 | if(image_w > max_image_w || image_h > max_image_h) |
| 287 | { |
| 288 | return false; |
| 289 | } |
| 290 | |
| 291 | return true; |
| 292 | } |
| 293 | |
| 294 | CLBuildOptions ClDirectConvolutionKernelComponent::generate_build_options() const |
| 295 | { |
| 296 | const auto src_info = _blueprint->impl().get_kernel_argument_info(_src.arg_id); |
SiCong Li | b63b119 | 2022-01-28 18:24:39 +0000 | [diff] [blame] | 297 | auto weight_info = _blueprint->impl().get_kernel_argument_info(_weight.arg_id); |
Gunes Bayir | 16c5697 | 2022-03-28 21:32:33 +0100 | [diff] [blame] | 298 | const auto dst_info = _blueprint->impl().get_kernel_argument_info(_blueprint->impl().get_dst_id()); |
SiCong Li | b63b119 | 2022-01-28 18:24:39 +0000 | [diff] [blame] | 299 | // const auto tile_info = _blueprint->impl().get_tile_info(); |
Gunes Bayir | 16c5697 | 2022-03-28 21:32:33 +0100 | [diff] [blame] | 300 | |
| 301 | const unsigned int channel_idx = get_data_layout_dimension_index(src_info->data_layout(), DataLayoutDimension::CHANNEL); |
| 302 | const DataType data_type = src_info->data_type(); |
SiCong Li | b63b119 | 2022-01-28 18:24:39 +0000 | [diff] [blame] | 303 | const GPUTarget gpu_target = CLScheduler::get().target(); |
Gunes Bayir | 16c5697 | 2022-03-28 21:32:33 +0100 | [diff] [blame] | 304 | |
SiCong Li | b63b119 | 2022-01-28 18:24:39 +0000 | [diff] [blame] | 305 | const unsigned int n0 = _blueprint->impl().get_execution_window().x().step(); |
| 306 | const unsigned int m0 = _blueprint->impl().get_execution_window().y().step(); |
Gunes Bayir | 16c5697 | 2022-03-28 21:32:33 +0100 | [diff] [blame] | 307 | const unsigned int k0 = adjust_vec_size(is_data_type_quantized(data_type) ? 16u : 8u, src_info->dimension(channel_idx)); |
SiCong Li | b63b119 | 2022-01-28 18:24:39 +0000 | [diff] [blame] | 308 | const unsigned int partial_store_n0 = dst_info->dimension(0) % n0; |
Gunes Bayir | 16c5697 | 2022-03-28 21:32:33 +0100 | [diff] [blame] | 309 | const bool export_to_cl_image = export_to_cl_image_support(weight_info, gpu_target, src_info->data_layout()); |
| 310 | |
| 311 | // Update the padding for the weights tensor if we can export to cl_image |
| 312 | if(export_to_cl_image) |
| 313 | { |
| 314 | arm_compute::opencl::kernels::gemm::update_padding_for_cl_image(weight_info); |
| 315 | } |
| 316 | |
| 317 | CLBuildOptions build_opts{}; |
| 318 | build_opts.add_option("-cl-fast-relaxed-math"); |
| 319 | build_opts.add_option("-DIS_TILED"); |
| 320 | build_opts.add_option("-DN0=" + support::cpp11::to_string(n0)); |
| 321 | build_opts.add_option("-DM0=" + support::cpp11::to_string(m0)); |
| 322 | build_opts.add_option("-DK0=" + support::cpp11::to_string(k0)); |
| 323 | build_opts.add_option("-DPARTIAL_N0=" + support::cpp11::to_string(partial_store_n0)); |
| 324 | |
| 325 | return build_opts; |
| 326 | } |
| 327 | |
SiCong Li | b63b119 | 2022-01-28 18:24:39 +0000 | [diff] [blame] | 328 | void ClDirectConvolutionKernelComponent::allocate_shared_vars(SharedVarTable &vtable) const |
| 329 | { |
| 330 | const auto src_info = _blueprint->impl().get_kernel_argument_info(_src.arg_id); |
| 331 | const auto weight_info = _blueprint->impl().get_kernel_argument_info(_weight.arg_id); |
| 332 | |
| 333 | vtable.add(_src, _blueprint->impl().group(_src.arg_id), ClKernelArgDescriptor(_src.arg_id, ClKernelTensorArgType::Tensor_4D_t_Buffer), "src"); |
| 334 | |
| 335 | const GPUTarget gpu_target = CLScheduler::get().target(); |
| 336 | const bool export_to_cl_image = export_to_cl_image_support(weight_info, gpu_target, src_info->data_layout()); |
| 337 | const ClKernelTensorArgType weight_type = export_to_cl_image ? ClKernelTensorArgType::Tensor_4D_t_Image : ClKernelTensorArgType::Tensor_4D_t_Buffer; |
| 338 | vtable.add(_weight, _blueprint->impl().group(_weight.arg_id), ClKernelArgDescriptor(_weight.arg_id, weight_type), "weight"); |
| 339 | |
| 340 | if(!_bias.is_empty()) // optional bias |
| 341 | { |
| 342 | vtable.add(_bias, _blueprint->impl().group(_bias.arg_id), ClKernelArgDescriptor(_bias.arg_id, ClKernelTensorArgType::Vector), "bias"); |
| 343 | } |
| 344 | vtable.add(_dst, _blueprint->impl().group(_dst.arg_id), ClKernelArgDescriptor(_dst.arg_id, ClKernelTensorArgType::Tensor_4D_t_Buffer), "dst"); |
| 345 | } |
| 346 | |
| 347 | ClDirectConvolutionKernelComponent::TagLUT ClDirectConvolutionKernelComponent::get_tag_lut(const SharedVarTable &vtable) const |
Gunes Bayir | 16c5697 | 2022-03-28 21:32:33 +0100 | [diff] [blame] | 348 | { |
| 349 | TagLUT lut{}; |
| 350 | |
| 351 | const auto src_info = _blueprint->impl().get_kernel_argument_info(_src.arg_id); |
| 352 | const auto weight_info = _blueprint->impl().get_kernel_argument_info(_weight.arg_id); |
| 353 | const auto bias_info = _blueprint->impl().get_kernel_argument_info(_bias.arg_id); |
Gunes Bayir | 16c5697 | 2022-03-28 21:32:33 +0100 | [diff] [blame] | 354 | |
SiCong Li | b63b119 | 2022-01-28 18:24:39 +0000 | [diff] [blame] | 355 | // Arguments and global shared variables |
| 356 | lut["src"] = vtable.get(_src); |
| 357 | lut["weight"] = vtable.get(_weight); |
Gunes Bayir | 16c5697 | 2022-03-28 21:32:33 +0100 | [diff] [blame] | 358 | |
| 359 | if(!_bias.is_empty()) // optional bias |
| 360 | { |
SiCong Li | b63b119 | 2022-01-28 18:24:39 +0000 | [diff] [blame] | 361 | lut["bias"] = vtable.get(_bias); |
Gunes Bayir | 16c5697 | 2022-03-28 21:32:33 +0100 | [diff] [blame] | 362 | lut["BIA_DATA_TYPE"] = get_cl_type_from_data_type(bias_info->data_type()); |
| 363 | } |
SiCong Li | b63b119 | 2022-01-28 18:24:39 +0000 | [diff] [blame] | 364 | lut["dst"] = vtable.get(_dst); |
| 365 | |
| 366 | const auto dst_argument = _blueprint->impl().get_argument_shared_vars().get_dst_var(); |
| 367 | lut["arg_dst"] = dst_argument.uniq_name; |
Gunes Bayir | 16c5697 | 2022-03-28 21:32:33 +0100 | [diff] [blame] | 368 | |
| 369 | // Local build options |
SiCong Li | b63b119 | 2022-01-28 18:24:39 +0000 | [diff] [blame] | 370 | lut["meta_kernel_id"] = id(); |
| 371 | lut["ACC_DATA_TYPE"] = src_info->data_type(); |
| 372 | lut["SRC_DATA_TYPE"] = src_info->data_type(); |
| 373 | lut["WEI_DATA_TYPE"] = weight_info->data_type(); |
Gunes Bayir | 16c5697 | 2022-03-28 21:32:33 +0100 | [diff] [blame] | 374 | |
| 375 | lut["SRC_TENSOR_TYPE"] = "BUFFER"; |
SiCong Li | b63b119 | 2022-01-28 18:24:39 +0000 | [diff] [blame] | 376 | switch(vtable.get(_weight).desc.tensor_arg_type) |
| 377 | { |
| 378 | case ClKernelTensorArgType::Image_Export_To_ClImage2D: |
| 379 | case ClKernelTensorArgType::Image_3D_Export_To_ClImage2D: |
| 380 | case ClKernelTensorArgType::Tensor_4D_t_Image: |
| 381 | { |
| 382 | lut["WEI_TENSOR_TYPE"] = "IMAGE"; |
| 383 | break; |
| 384 | } |
| 385 | default: |
| 386 | { |
| 387 | lut["WEI_TENSOR_TYPE"] = "BUFFER"; |
| 388 | break; |
| 389 | } |
| 390 | } |
| 391 | const auto width_idx = get_data_layout_dimension_index(src_info->data_layout(), DataLayoutDimension::WIDTH); |
| 392 | const auto height_idx = get_data_layout_dimension_index(src_info->data_layout(), DataLayoutDimension::HEIGHT); |
| 393 | lut["WEI_WIDTH"] = weight_info->dimension(width_idx); |
| 394 | lut["WEI_HEIGHT"] = weight_info->dimension(height_idx); |
Gunes Bayir | 16c5697 | 2022-03-28 21:32:33 +0100 | [diff] [blame] | 395 | |
SiCong Li | b63b119 | 2022-01-28 18:24:39 +0000 | [diff] [blame] | 396 | lut["STRIDE_X"] = _desc.conv2d.stride.x(); |
| 397 | lut["STRIDE_Y"] = _desc.conv2d.stride.y(); |
Gunes Bayir | 16c5697 | 2022-03-28 21:32:33 +0100 | [diff] [blame] | 398 | |
SiCong Li | b63b119 | 2022-01-28 18:24:39 +0000 | [diff] [blame] | 399 | lut["PAD_LEFT"] = _desc.conv2d.pad.left; |
| 400 | lut["PAD_TOP"] = _desc.conv2d.pad.top; |
Gunes Bayir | 16c5697 | 2022-03-28 21:32:33 +0100 | [diff] [blame] | 401 | |
| 402 | lut["ZERO_VALUE"] = 0; |
| 403 | |
| 404 | return lut; |
| 405 | } |
| 406 | } // namespace dynamic_fusion |
| 407 | } // namespace experimental |
SiCong Li | 4e9f568 | 2022-05-10 10:15:59 +0100 | [diff] [blame] | 408 | } // namespace arm_compute |
| 409 | #endif /* ENABLE_EXPERIMENTAL_DYNAMIC_FUSION */ |