Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1 | |
Kevin Cheng | 3a47857 | 2021-01-22 17:21:02 -0800 | [diff] [blame] | 2 | // Copyright (c) 2020-2021, ARM Limited. |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 3 | // |
| 4 | // Licensed under the Apache License, Version 2.0 (the "License"); |
| 5 | // you may not use this file except in compliance with the License. |
| 6 | // You may obtain a copy of the License at |
| 7 | // |
| 8 | // http://www.apache.org/licenses/LICENSE-2.0 |
| 9 | // |
| 10 | // Unless required by applicable law or agreed to in writing, software |
| 11 | // distributed under the License is distributed on an "AS IS" BASIS, |
| 12 | // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 13 | // See the License for the specific language governing permissions and |
| 14 | // limitations under the License. |
| 15 | |
| 16 | #include "tensor_ops.h" |
| 17 | #include "quant_util.h" |
| 18 | #include "template_types.h" |
| 19 | |
| 20 | using namespace TosaReference; |
| 21 | using namespace Eigen; |
| 22 | using namespace tosa; |
| 23 | |
| 24 | template <int Rank, DType Dtype> |
Kevin Cheng | acb550f | 2021-06-29 15:32:19 -0700 | [diff] [blame] | 25 | OpArgMax<Rank, Dtype>::OpArgMax(SubgraphTraverser* sgt_, |
| 26 | TosaAttributeBase* attribute_, |
| 27 | TosaQuantInfoBase* qinfo_, |
| 28 | uint64_t id_) |
| 29 | : GraphNode(sgt_, Op_ARGMAX, id_) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 30 | { |
| 31 | setRequiredOperands(1, 1); |
| 32 | setRequiredRank(0, 6); |
| 33 | |
| 34 | INIT_ATTRIBUTE(Axis); |
| 35 | } |
| 36 | |
| 37 | template <int Rank, DType Dtype> |
| 38 | OpArgMax<Rank, Dtype>::~OpArgMax() |
| 39 | { |
| 40 | if (attribute) |
| 41 | delete attribute; |
| 42 | } |
| 43 | |
| 44 | template <int Rank, DType Dtype> |
| 45 | int OpArgMax<Rank, Dtype>::checkTensorAttributes() |
| 46 | { |
| 47 | if (validateRequiredOperands()) |
| 48 | return 1; |
| 49 | |
| 50 | if (validateRequiredRank(inputs[0]) || validateRequiredRank(outputs[0])) |
| 51 | { |
| 52 | return 1; |
| 53 | } |
| 54 | |
| 55 | input = dynamic_cast<TosaReference::TensorTemplate<TIn>*>(inputs[0]); |
| 56 | output = dynamic_cast<TosaReference::TensorTemplate<TOut>*>(outputs[0]); |
| 57 | |
| 58 | return 0; |
| 59 | } |
| 60 | |
| 61 | template <int Rank, DType Dtype> |
| 62 | int OpArgMax<Rank, Dtype>::eval() |
| 63 | { |
| 64 | Eigen::Tensor<DenseIndex, Rank - 1> index = this->input->getTensor().argmax(attribute->axis()); |
| 65 | |
| 66 | this->output->getTensor() = index.unaryExpr([](DenseIndex in) -> OutEigenType { return (OutEigenType)in; }); |
| 67 | |
| 68 | return GraphNode::eval(); |
| 69 | } |
| 70 | |
| 71 | template <DType Dtype> |
Kevin Cheng | acb550f | 2021-06-29 15:32:19 -0700 | [diff] [blame] | 72 | OpAvgPool2d<Dtype>::OpAvgPool2d(SubgraphTraverser* sgt_, |
| 73 | TosaAttributeBase* attribute_, |
| 74 | TosaQuantInfoBase* qinfo_, |
| 75 | uint64_t id_) |
| 76 | : GraphNode(sgt_, Op_AVG_POOL2D, id_) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 77 | { |
| 78 | setRequiredOperands(1, 1); |
| 79 | setRequiredRank(4); |
| 80 | |
Kevin Cheng | 93a1628 | 2021-08-31 16:14:03 -0700 | [diff] [blame] | 81 | INIT_ATTRIBUTE(Pool); |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 82 | INIT_QINFO(Unary); |
| 83 | } |
| 84 | |
| 85 | template <DType Dtype> |
| 86 | OpAvgPool2d<Dtype>::~OpAvgPool2d() |
| 87 | { |
| 88 | if (attribute) |
| 89 | delete attribute; |
| 90 | } |
| 91 | |
| 92 | template <DType Dtype> |
| 93 | int OpAvgPool2d<Dtype>::checkTensorAttributes() |
| 94 | { |
| 95 | if (validateRequiredOperands()) |
| 96 | return 1; |
| 97 | |
| 98 | if (validateRequiredRank(inputs[0]) || validateRequiredRank(outputs[0])) |
| 99 | { |
| 100 | return 1; |
| 101 | } |
| 102 | |
| 103 | if (inputs[0]->matchType(*outputs[0])) |
| 104 | { |
| 105 | printNodeValidationError("OpAvgPool2d: input and output tensor type mismatch"); |
| 106 | return 1; |
| 107 | } |
| 108 | |
| 109 | in = dynamic_cast<TosaReference::TensorTemplate<TIn>*>(inputs[0]); |
| 110 | out = dynamic_cast<TosaReference::TensorTemplate<TOut>*>(outputs[0]); |
| 111 | |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 112 | if (attribute->padding().size() != 4) |
| 113 | { |
| 114 | printNodeValidationError("OpAvgPool2d: illegal size for attribute padding"); |
| 115 | return 1; |
| 116 | } |
| 117 | |
| 118 | if (attribute->kernel().size() != 2) |
| 119 | { |
| 120 | printNodeValidationError("OpAvgPool2d: illegal size for attribute kernel"); |
| 121 | return 1; |
| 122 | } |
| 123 | |
| 124 | if (attribute->stride().size() != 2) |
| 125 | { |
| 126 | printNodeValidationError("OpAvgPool2d: illegal size for attribute stride"); |
| 127 | return 1; |
| 128 | } |
| 129 | |
| 130 | return 0; |
| 131 | } |
| 132 | |
| 133 | template <DType Dtype> |
| 134 | ETensor1<int32_t> OpAvgPool2d<Dtype>::calculate_div_map_1d(int in_size, int out_size, int kernel_size, int stride) |
| 135 | { |
| 136 | ETensor1<int32_t> result(out_size); |
| 137 | |
| 138 | int32_t total_pad = (out_size - 1) * stride + kernel_size - in_size; |
| 139 | total_pad = total_pad < 0 ? 0 : total_pad; |
| 140 | |
| 141 | int32_t pad_left = total_pad >> 1; |
| 142 | int32_t pad_right = total_pad - pad_left; |
| 143 | |
| 144 | result.setConstant(kernel_size); |
| 145 | |
| 146 | // the index left to 'left_index' and index right to 'right_index' indicates |
| 147 | // the input window of this output covers a pad bit |
| 148 | int32_t left_index = pad_left / stride; |
| 149 | int32_t right_index = pad_right / stride; |
| 150 | |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 151 | // minus the number of pad bit this index cover |
| 152 | while (left_index >= 0) |
| 153 | { |
| 154 | result(left_index) -= (pad_left - left_index * stride); |
| 155 | left_index--; |
| 156 | } |
| 157 | |
| 158 | while (right_index >= 0) |
| 159 | { |
| 160 | result(out_size - 1 - right_index) -= (pad_right - right_index * stride); |
| 161 | right_index--; |
| 162 | } |
| 163 | |
| 164 | return result; |
| 165 | } |
| 166 | |
| 167 | // assuming input and output tensor have same scales like tflite reference |
| 168 | // so no need to scale input and output |
| 169 | template <DType Dtype> |
| 170 | int OpAvgPool2d<Dtype>::eval() |
| 171 | { |
| 172 | int in_batch = this->in->getShape()[0]; |
| 173 | int in_height = this->in->getShape()[1]; |
| 174 | int in_width = this->in->getShape()[2]; |
| 175 | int in_channels = this->in->getShape()[3]; |
| 176 | |
| 177 | int out_batch = this->out->getShape()[0]; |
| 178 | int out_height = this->out->getShape()[1]; |
| 179 | int out_width = this->out->getShape()[2]; |
| 180 | int out_channels = this->out->getShape()[3]; |
| 181 | |
Kevin Cheng | acb550f | 2021-06-29 15:32:19 -0700 | [diff] [blame] | 182 | ERROR_IF(in_batch != out_batch, "OpAvgPool2d: tensor batch mismatch %d != %d", in_batch, out_batch); |
| 183 | ERROR_IF(in_channels != out_channels, "OpAvgPool2d: tensor channel mismatch %d != %d", in_channels, out_channels); |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 184 | |
| 185 | int padding_top = this->attribute->padding()[0]; |
| 186 | int padding_bottom = this->attribute->padding()[1]; |
| 187 | int padding_left = this->attribute->padding()[2]; |
| 188 | int padding_right = this->attribute->padding()[3]; |
| 189 | int kernel_h = this->attribute->kernel()[0]; |
| 190 | int kernel_w = this->attribute->kernel()[1]; |
| 191 | int stride_h = this->attribute->stride()[0]; |
| 192 | int stride_w = this->attribute->stride()[1]; |
| 193 | |
| 194 | DEBUG_INFO(OP, |
| 195 | "perform AvgPool2d, input.shape=[%d,%d,%d,%d], output.shape=[%d,%d,%d,%d], kernel=[%d,%d], " |
| 196 | "stride=[%d,%d], padding=[%d,%d,%d,%d]", |
| 197 | in_batch, in_height, in_width, in_channels, out_batch, out_height, out_width, out_channels, kernel_h, |
| 198 | kernel_w, stride_h, stride_w, padding_top, padding_bottom, padding_left, padding_right); |
| 199 | |
| 200 | Eigen::array<Eigen::Index, 2> im2col_input_dims; |
| 201 | im2col_input_dims[0] = kernel_h * kernel_w; |
| 202 | im2col_input_dims[1] = out_batch * out_height * out_width * out_channels; |
| 203 | |
| 204 | Eigen::array<Eigen::Index, 4> col2im_output_dims; |
| 205 | col2im_output_dims[0] = out_batch; |
| 206 | col2im_output_dims[1] = out_height; |
| 207 | col2im_output_dims[2] = out_width; |
| 208 | col2im_output_dims[3] = out_channels; |
| 209 | |
| 210 | Eigen::array<std::pair<int32_t, int32_t>, 4> padding; |
| 211 | padding[0] = std::make_pair(0, 0); |
| 212 | padding[1] = std::make_pair(padding_top, padding_bottom); |
| 213 | padding[2] = std::make_pair(padding_left, padding_right); |
| 214 | padding[3] = std::make_pair(0, 0); |
| 215 | |
| 216 | ETensor4<InEigenType> input_val = this->in->getTensor(); |
| 217 | if (this->qinfo) |
| 218 | { |
| 219 | input_val = input_val - (InEigenType)this->qinfo->input_zp(); |
| 220 | } |
| 221 | |
| 222 | ETensor4<InEigenType> input_padded = input_val.pad(padding); |
| 223 | |
| 224 | // assuming input and output have same scales |
| 225 | // so input and output scaling is not required |
| 226 | // TODO: check if this assumption TOSA made |
| 227 | |
| 228 | // extract_image_patches() output [N, KH, KW, H * W, C] |
| 229 | // transpose to [KH, KW, N, H * W, C] |
| 230 | // reshape to [KH * KW, N * H * W * C] |
| 231 | ETensor2<InEigenType> input_extract_patches = |
| 232 | input_padded.extract_image_patches(kernel_h, kernel_w, stride_h, stride_w, 1, 1, Eigen::PADDING_VALID) |
| 233 | .shuffle(Eigen::array<Eigen::Index, 5>{ 1, 2, 0, 3, 4 }) |
| 234 | .reshape(im2col_input_dims); |
| 235 | |
| 236 | // 1D result with [N * H * W * C] |
| 237 | ETensor1<AccEigenType> out_1d(this->out->getElementCount()); |
| 238 | out_1d.setZero(); |
| 239 | |
| 240 | // sum pool |
| 241 | for (size_t i = 0; i < this->out->getElementCount(); i++) |
| 242 | { |
| 243 | for (int32_t j = 0; j < kernel_h * kernel_w; j++) |
| 244 | { |
| 245 | out_1d(i) += (AccEigenType)input_extract_patches(j, i); |
| 246 | } |
| 247 | } |
| 248 | |
| 249 | // reshape result to [N, H, W, C] and divide with div_map |
| 250 | ETensor4<AccEigenType> sum = out_1d.reshape(col2im_output_dims); |
| 251 | |
| 252 | // calculate 1d height/width div_map (number of elements this pooling window covers) |
| 253 | // and outer product to get 2d div_map, then reshape/broadcast to [N, H, W, C] |
| 254 | ETensor1<int32_t> div_map_h = calculate_div_map_1d(in_height, out_height, kernel_h, stride_h); |
| 255 | ETensor1<int32_t> div_map_w = calculate_div_map_1d(in_width, out_width, kernel_w, stride_w); |
| 256 | Eigen::array<Eigen::IndexPair<Eigen::Index>, 1> contract_dims = { Eigen::IndexPair<Eigen::Index>(1, 0) }; |
| 257 | Eigen::array<Eigen::Index, 4> bcast{ out_batch, 1, 1, out_channels }; |
| 258 | |
| 259 | ETensor4<int32_t> div_map = |
| 260 | div_map_h.reshape(Eigen::array<Eigen::Index, 2>{ out_height, 1 }) |
| 261 | .contract(div_map_w.reshape(Eigen::array<Eigen::Index, 2>{ 1, out_width }), contract_dims) |
| 262 | .reshape(Eigen::array<Eigen::Index, 4>{ 1, out_height, out_width, 1 }) |
| 263 | .broadcast(bcast); |
| 264 | |
| 265 | if (Dtype != DType_FLOAT) |
| 266 | { |
Kevin Cheng | acb550f | 2021-06-29 15:32:19 -0700 | [diff] [blame] | 267 | try |
| 268 | { |
| 269 | this->out->getTensor() = sum.binaryExpr(div_map, [](AccEigenType value, int32_t div) -> OutEigenType { |
| 270 | int32_t multiplier, shift; |
| 271 | TosaReference::QuantUtil::reciprocal_scale(div, multiplier, shift); |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 272 | |
Kevin Cheng | acb550f | 2021-06-29 15:32:19 -0700 | [diff] [blame] | 273 | return (OutEigenType)TosaReference::QuantUtil::apply_scale_32(value, multiplier, shift, false); |
| 274 | }); |
| 275 | } |
| 276 | catch (std::string desc) |
| 277 | { |
| 278 | REQUIRE(false, "OpAvgPool2d apply_scale_32() fails: %s.", desc.c_str()); |
| 279 | } |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 280 | this->out->getTensor() = this->out->getTensor() + (OutEigenType)(this->qinfo->output_zp()); |
| 281 | this->out->getTensor() = this->out->getTensor().cwiseMax((OutEigenType)QMin); |
| 282 | this->out->getTensor() = this->out->getTensor().cwiseMin((OutEigenType)QMax); |
| 283 | } |
| 284 | else |
| 285 | { |
| 286 | this->out->getTensor() = (sum / div_map.template cast<AccEigenType>()).template cast<OutEigenType>(); |
| 287 | } |
| 288 | |
| 289 | return GraphNode::eval(); |
| 290 | } |
| 291 | |
| 292 | template <DType InDtype, DType WeightDtype> |
Kevin Cheng | acb550f | 2021-06-29 15:32:19 -0700 | [diff] [blame] | 293 | OpConv2d<InDtype, WeightDtype>::OpConv2d(SubgraphTraverser* sgt_, |
| 294 | TosaAttributeBase* attribute_, |
| 295 | TosaQuantInfoBase* qinfo_, |
| 296 | uint64_t id_) |
| 297 | : GraphNode(sgt_, Op_CONV2D, id_) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 298 | { |
| 299 | setRequiredOperands(3, 1); |
| 300 | setRequiredRank(4); |
| 301 | |
Kevin Cheng | 93a1628 | 2021-08-31 16:14:03 -0700 | [diff] [blame] | 302 | INIT_ATTRIBUTE(Conv); |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 303 | INIT_QINFO(Conv); |
| 304 | } |
| 305 | |
| 306 | template <DType InDtype, DType WeightDtype> |
| 307 | OpConv2d<InDtype, WeightDtype>::~OpConv2d() |
| 308 | { |
| 309 | if (attribute) |
| 310 | delete attribute; |
| 311 | if (qinfo) |
| 312 | delete qinfo; |
| 313 | } |
| 314 | |
| 315 | template <DType InDtype, DType WeightDtype> |
| 316 | int OpConv2d<InDtype, WeightDtype>::checkTensorAttributes() |
| 317 | { |
| 318 | if (validateRequiredOperands()) |
| 319 | return 1; |
| 320 | |
| 321 | if (validateRequiredRank(inputs[0]) || validateRequiredRank(inputs[1]) || validateRequiredRank(outputs[0])) |
| 322 | { |
| 323 | return 1; |
| 324 | } |
| 325 | |
| 326 | // 'bias' checked separatedly since it doens't make sense to make required rank ranging from 1 to 4 |
| 327 | if (inputs[2]->getRank() != 1) |
| 328 | { |
| 329 | printNodeValidationError("OpConv2d: bias tensor must be rank 1"); |
| 330 | } |
| 331 | |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 332 | input = dynamic_cast<TosaReference::TensorTemplate<TIn>*>(inputs[0]); |
| 333 | weight = dynamic_cast<TosaReference::TensorTemplate<TWeight>*>(inputs[1]); |
| 334 | bias = dynamic_cast<TosaReference::TensorTemplate<TBias>*>(inputs[2]); |
| 335 | output = dynamic_cast<TosaReference::TensorTemplate<TAcc>*>(outputs[0]); |
| 336 | |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 337 | if (attribute->padding().size() != 4) |
| 338 | { |
| 339 | printNodeValidationError("OpConv2d: illegal size for attribute padding"); |
| 340 | return 1; |
| 341 | } |
| 342 | |
| 343 | if (attribute->stride().size() != 2) |
| 344 | { |
| 345 | printNodeValidationError("OpConv2d: illegal size for attribute stride"); |
| 346 | return 1; |
| 347 | } |
| 348 | |
| 349 | if (attribute->dilation().size() != 2) |
| 350 | { |
| 351 | printNodeValidationError("OpConv2d: illegal size for attribute dilation"); |
| 352 | return 1; |
| 353 | } |
| 354 | |
| 355 | return 0; |
| 356 | } |
| 357 | |
| 358 | template <DType InDtype, DType WeightDtype> |
| 359 | int OpConv2d<InDtype, WeightDtype>::eval() |
| 360 | { |
| 361 | int in_batch = this->input->getShape()[0]; |
| 362 | int in_height = this->input->getShape()[1]; |
| 363 | int in_width = this->input->getShape()[2]; |
| 364 | int in_channels = this->input->getShape()[3]; |
| 365 | |
| 366 | int f_out_channels = this->weight->getShape()[0]; |
| 367 | int f_height = this->weight->getShape()[1]; |
| 368 | int f_width = this->weight->getShape()[2]; |
| 369 | int f_in_channels = this->weight->getShape()[3]; |
| 370 | |
| 371 | int b_out_channels = this->bias->getShape()[0]; |
| 372 | |
| 373 | int out_batch = this->output->getShape()[0]; |
| 374 | int out_height = this->output->getShape()[1]; |
| 375 | int out_width = this->output->getShape()[2]; |
| 376 | int out_channels = this->output->getShape()[3]; |
| 377 | |
Kevin Cheng | acb550f | 2021-06-29 15:32:19 -0700 | [diff] [blame] | 378 | ERROR_IF(in_batch != out_batch, "OpConv2d: tensor batch mismatch %d != %d", in_batch, out_batch); |
| 379 | ERROR_IF(f_in_channels != in_channels, "OpConv2d: tensor input channel mismatch %d != %d", f_in_channels, |
| 380 | in_channels); |
| 381 | ERROR_IF(f_out_channels != out_channels, "OpConv2d: tensor output channel mismatch %d != %d", f_out_channels, |
| 382 | out_channels); |
| 383 | ERROR_IF(b_out_channels != out_channels, "OpConv2d: bias channel mismatch %d != %d", b_out_channels, out_channels); |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 384 | |
| 385 | int padding_top = this->attribute->padding()[0]; |
| 386 | int padding_bottom = this->attribute->padding()[1]; |
| 387 | int padding_left = this->attribute->padding()[2]; |
| 388 | int padding_right = this->attribute->padding()[3]; |
| 389 | int stride_h = this->attribute->stride()[0]; |
| 390 | int stride_w = this->attribute->stride()[1]; |
| 391 | int dilation_h = this->attribute->dilation()[0]; |
| 392 | int dilation_w = this->attribute->dilation()[1]; |
| 393 | |
| 394 | DEBUG_INFO(OP, |
| 395 | "perform OpConv2d, input.shape=[%d,%d,%d,%d], weight.shape=[%d,%d,%d,%d], output.shape=[%d,%d,%d,%d], " |
| 396 | "stride=[%d,%d], dilation=[%d,%d], padding=[%d,%d,%d,%d]", |
| 397 | in_batch, in_height, in_width, in_channels, f_height, f_width, f_in_channels, f_out_channels, out_batch, |
| 398 | out_height, out_width, out_channels, stride_h, stride_w, dilation_h, dilation_w, padding_top, |
| 399 | padding_bottom, padding_left, padding_right); |
| 400 | |
| 401 | // GEMM-conv2d, left matrix is input, right matrix is weight |
| 402 | Eigen::array<Eigen::Index, 2> im2col_input_dims; |
| 403 | im2col_input_dims[0] = out_batch * out_height * out_width; |
| 404 | im2col_input_dims[1] = f_height * f_width * f_in_channels; |
| 405 | |
| 406 | Eigen::array<Eigen::Index, 2> im2col_weight_dims; |
| 407 | im2col_weight_dims[0] = f_height * f_width * f_in_channels; |
| 408 | im2col_weight_dims[1] = f_out_channels; |
| 409 | |
| 410 | Eigen::array<Eigen::Index, 2> bias_reshaped_dims; |
| 411 | bias_reshaped_dims[0] = 1; |
| 412 | bias_reshaped_dims[1] = b_out_channels; |
| 413 | |
| 414 | Eigen::array<Eigen::Index, 4> weight_zp_bcast_dims; |
| 415 | weight_zp_bcast_dims[0] = f_height; |
| 416 | weight_zp_bcast_dims[1] = f_width; |
| 417 | weight_zp_bcast_dims[2] = f_in_channels; |
| 418 | |
| 419 | Eigen::array<Eigen::Index, 2> bias_bcast_dims; |
| 420 | bias_bcast_dims[0] = out_batch * out_height * out_width; |
| 421 | bias_bcast_dims[1] = 1; |
| 422 | |
| 423 | Eigen::array<Eigen::Index, 4> col2im_output_dims; |
| 424 | col2im_output_dims[0] = out_batch; |
| 425 | col2im_output_dims[1] = out_height; |
| 426 | col2im_output_dims[2] = out_width; |
| 427 | col2im_output_dims[3] = out_channels; |
| 428 | |
| 429 | Eigen::array<Eigen::IndexPair<Eigen::Index>, 1> contract_dims = { Eigen::IndexPair<Eigen::Index>(1, 0) }; |
| 430 | |
| 431 | Eigen::array<std::pair<int32_t, int32_t>, 4> padding; |
| 432 | padding[0] = std::make_pair(0, 0); |
| 433 | padding[1] = std::make_pair(padding_top, padding_bottom); |
| 434 | padding[2] = std::make_pair(padding_left, padding_right); |
| 435 | padding[3] = std::make_pair(0, 0); |
| 436 | |
| 437 | TIn input_val = this->input->getTensor(); |
| 438 | TWeight weight_val = this->weight->getTensor(); |
| 439 | if (this->qinfo) |
| 440 | { |
| 441 | input_val = input_val - (InEigenType)this->qinfo->input_zp(); |
| 442 | weight_val = weight_val - (WeightEigenType)this->qinfo->weight_zp(); |
| 443 | } |
| 444 | |
| 445 | ETensor4<InEigenType> input_padded = input_val.pad(padding); |
| 446 | |
| 447 | // extract_image_patches() output [N, KH, KW, H * W, C] |
| 448 | // need to transpose to [N, H * W, KH, KW, C] |
| 449 | ETensor5<InEigenType> input_extract_patches = |
| 450 | input_padded |
| 451 | .extract_image_patches(f_height, f_width, stride_h, stride_w, dilation_h, dilation_w, Eigen::PADDING_VALID) |
| 452 | .shuffle(Eigen::array<Eigen::Index, 5>{ 0, 3, 1, 2, 4 }); |
| 453 | |
| 454 | // reshape input to [N * H * W, KH * KW * C] |
| 455 | ETensor2<InEigenType> im2col_input = input_extract_patches.reshape(im2col_input_dims); |
| 456 | |
| 457 | // transpose and reshape weight from [OC, H, W, IC] to [H * W * IC, OC] |
| 458 | ETensor2<WeightEigenType> im2col_weight = |
| 459 | weight_val.shuffle(Eigen::array<Eigen::Index, 4>({ 1, 2, 3, 0 })).reshape(im2col_weight_dims); |
| 460 | |
| 461 | // don't need to apply bias_multiplier ( * bias_scale and >> bias_shift) since tflite already scale it |
| 462 | // and reshaped from [C] to [1, C], and broadcast to [N * H * W, C] |
| 463 | ETensor2<AccEigenType> bias_2d = this->bias->getTensor().reshape(bias_reshaped_dims).broadcast(bias_bcast_dims); |
| 464 | |
| 465 | // output matrix is [N * H * W, C] |
| 466 | ETensor2<AccEigenType> contracted_result = |
| 467 | im2col_input.template cast<AccEigenType>().contract(im2col_weight.template cast<AccEigenType>(), contract_dims); |
| 468 | |
| 469 | // adding bias |
| 470 | ETensor2<AccEigenType> biased_output = contracted_result + bias_2d.template cast<AccEigenType>(); |
| 471 | |
| 472 | // reshape back to [N, H, W, C] |
| 473 | this->output->getTensor() = biased_output.reshape(col2im_output_dims); |
| 474 | |
| 475 | if (AccDtype == DType_INT48) |
| 476 | { |
| 477 | this->output->getTensor() = this->output->getTensor().cwiseMax((AccEigenType)AccQMin); |
| 478 | this->output->getTensor() = this->output->getTensor().cwiseMin((AccEigenType)AccQMax); |
| 479 | } |
| 480 | |
| 481 | return GraphNode::eval(); |
| 482 | } |
| 483 | |
| 484 | template <DType InDtype, DType WeightDtype> |
Kevin Cheng | 1533b85 | 2021-09-01 12:51:58 -0700 | [diff] [blame^] | 485 | OpConv3d<InDtype, WeightDtype>::OpConv3d(SubgraphTraverser* sgt_, |
| 486 | TosaAttributeBase* attribute_, |
| 487 | TosaQuantInfoBase* qinfo_, |
| 488 | uint64_t id_) |
| 489 | : GraphNode(sgt_, Op_CONV3D, id_) |
| 490 | { |
| 491 | setRequiredOperands(3, 1); |
| 492 | setRequiredRank(5); |
| 493 | |
| 494 | INIT_ATTRIBUTE(Conv); |
| 495 | INIT_QINFO(Conv); |
| 496 | } |
| 497 | |
| 498 | template <DType InDtype, DType WeightDtype> |
| 499 | OpConv3d<InDtype, WeightDtype>::~OpConv3d() |
| 500 | { |
| 501 | if (attribute) |
| 502 | delete attribute; |
| 503 | if (qinfo) |
| 504 | delete qinfo; |
| 505 | } |
| 506 | |
| 507 | template <DType InDtype, DType WeightDtype> |
| 508 | int OpConv3d<InDtype, WeightDtype>::checkTensorAttributes() |
| 509 | { |
| 510 | if (validateRequiredOperands()) |
| 511 | return 1; |
| 512 | |
| 513 | if (validateRequiredRank(inputs[0]) || validateRequiredRank(inputs[1]) || validateRequiredRank(outputs[0])) |
| 514 | { |
| 515 | return 1; |
| 516 | } |
| 517 | |
| 518 | // 'bias' checked separatedly since it doens't make sense to make required rank ranging from 1 to 4 |
| 519 | if (inputs[2]->getRank() != 1) |
| 520 | { |
| 521 | printNodeValidationError("OpConv3d: bias tensor must be rank 1"); |
| 522 | } |
| 523 | |
| 524 | input = dynamic_cast<TosaReference::TensorTemplate<TIn>*>(inputs[0]); |
| 525 | weight = dynamic_cast<TosaReference::TensorTemplate<TWeight>*>(inputs[1]); |
| 526 | bias = dynamic_cast<TosaReference::TensorTemplate<TBias>*>(inputs[2]); |
| 527 | output = dynamic_cast<TosaReference::TensorTemplate<TAcc>*>(outputs[0]); |
| 528 | |
| 529 | if (attribute->padding().size() != 6) |
| 530 | { |
| 531 | printNodeValidationError("OpConv3d: illegal size for attribute padding"); |
| 532 | return 1; |
| 533 | } |
| 534 | |
| 535 | if (attribute->stride().size() != 3) |
| 536 | { |
| 537 | printNodeValidationError("OpConv3d: illegal size for attribute stride"); |
| 538 | return 1; |
| 539 | } |
| 540 | |
| 541 | if (attribute->dilation().size() != 3) |
| 542 | { |
| 543 | printNodeValidationError("OpConv3d: illegal size for attribute dilation"); |
| 544 | return 1; |
| 545 | } |
| 546 | |
| 547 | return 0; |
| 548 | } |
| 549 | |
| 550 | template <DType InDtype, DType WeightDtype> |
| 551 | int OpConv3d<InDtype, WeightDtype>::eval() |
| 552 | { |
| 553 | int in_batch = this->input->getShape()[0]; |
| 554 | int in_depth = this->input->getShape()[1]; |
| 555 | int in_height = this->input->getShape()[2]; |
| 556 | int in_width = this->input->getShape()[3]; |
| 557 | int in_channels = this->input->getShape()[4]; |
| 558 | |
| 559 | int f_out_channels = this->weight->getShape()[0]; |
| 560 | int f_depth = this->weight->getShape()[1]; |
| 561 | int f_height = this->weight->getShape()[2]; |
| 562 | int f_width = this->weight->getShape()[3]; |
| 563 | int f_in_channels = this->weight->getShape()[4]; |
| 564 | |
| 565 | int b_out_channels = this->bias->getShape()[0]; |
| 566 | |
| 567 | int out_batch = this->output->getShape()[0]; |
| 568 | int out_depth = this->output->getShape()[1]; |
| 569 | int out_height = this->output->getShape()[2]; |
| 570 | int out_width = this->output->getShape()[3]; |
| 571 | int out_channels = this->output->getShape()[4]; |
| 572 | |
| 573 | ERROR_IF(in_batch != out_batch, "OpConv3d: tensor batch mismatch %d != %d", in_batch, out_batch); |
| 574 | ERROR_IF(f_in_channels != in_channels, "OpConv3d: tensor input channel mismatch %d != %d", f_in_channels, |
| 575 | in_channels); |
| 576 | ERROR_IF(f_out_channels != out_channels, "OpConv3d: tensor output channel mismatch %d != %d", f_out_channels, |
| 577 | out_channels); |
| 578 | ERROR_IF(b_out_channels != out_channels, "OpConv3d: bias channel mismatch %d != %d", b_out_channels, out_channels); |
| 579 | |
| 580 | int padding_d0 = this->attribute->padding()[0]; |
| 581 | int padding_d1 = this->attribute->padding()[1]; |
| 582 | int padding_top = this->attribute->padding()[2]; |
| 583 | int padding_bottom = this->attribute->padding()[3]; |
| 584 | int padding_left = this->attribute->padding()[4]; |
| 585 | int padding_right = this->attribute->padding()[5]; |
| 586 | int stride_d = this->attribute->stride()[0]; |
| 587 | int stride_h = this->attribute->stride()[1]; |
| 588 | int stride_w = this->attribute->stride()[2]; |
| 589 | int dilation_d = this->attribute->dilation()[0]; |
| 590 | int dilation_h = this->attribute->dilation()[1]; |
| 591 | int dilation_w = this->attribute->dilation()[2]; |
| 592 | |
| 593 | DEBUG_INFO( |
| 594 | OP, |
| 595 | "perform OpConv3d, input.shape=[%d,%d,%d,%d,%d], weight.shape=[%d,%d,%d,%d,%d], output.shape=[%d,%d,%d,%d,%d], " |
| 596 | "stride=[%d,%d,%d], dilation=[%d,%d,%d], padding=[%d,%d,%d,%d,%d,%d]", |
| 597 | in_batch, in_depth, in_height, in_width, in_channels, f_out_channels, f_depth, f_height, f_width, f_in_channels, |
| 598 | out_batch, out_depth, out_height, out_width, out_channels, stride_d, stride_h, stride_w, dilation_d, dilation_h, |
| 599 | dilation_w, padding_d0, padding_d1, padding_top, padding_bottom, padding_left, padding_right); |
| 600 | |
| 601 | Eigen::array<std::pair<int32_t, int32_t>, 5> padding; |
| 602 | padding[0] = std::make_pair(0, 0); |
| 603 | padding[1] = std::make_pair(padding_d0, padding_d1); |
| 604 | padding[2] = std::make_pair(padding_top, padding_bottom); |
| 605 | padding[3] = std::make_pair(padding_left, padding_right); |
| 606 | padding[4] = std::make_pair(0, 0); |
| 607 | |
| 608 | TIn input_val = this->input->getTensor(); |
| 609 | TWeight weight_val = this->weight->getTensor(); |
| 610 | if (this->qinfo) |
| 611 | { |
| 612 | input_val = input_val - (InEigenType)this->qinfo->input_zp(); |
| 613 | weight_val = weight_val - (WeightEigenType)this->qinfo->weight_zp(); |
| 614 | } |
| 615 | |
| 616 | ETensor5<InEigenType> input_padded = input_val.pad(padding); |
| 617 | |
| 618 | // 1. initialize with bias |
| 619 | Eigen::array<Eigen::Index, 5> reshape_dim; |
| 620 | reshape_dim.fill(1); |
| 621 | reshape_dim[4] = b_out_channels; |
| 622 | |
| 623 | Eigen::array<Eigen::Index, 5> bcast; |
| 624 | bcast[0] = out_batch; |
| 625 | bcast[1] = out_depth; |
| 626 | bcast[2] = out_height; |
| 627 | bcast[3] = out_width; |
| 628 | bcast[4] = 1; |
| 629 | this->output->getTensor() = this->bias->getTensor().reshape(reshape_dim).broadcast(bcast); |
| 630 | |
| 631 | // 2. direct convolution |
| 632 | AccEigenType acc = 0; |
| 633 | int d_idx, h_idx, w_idx; |
| 634 | |
| 635 | for (int ob = 0; ob < out_batch; ob++) |
| 636 | { |
| 637 | for (int od = 0; od < out_depth; od++) |
| 638 | { |
| 639 | for (int oh = 0; oh < out_height; oh++) |
| 640 | { |
| 641 | for (int ow = 0; ow < out_width; ow++) |
| 642 | { |
| 643 | for (int oc = 0; oc < out_channels; oc++) |
| 644 | { |
| 645 | acc = 0; |
| 646 | for (int fd = 0; fd < f_depth; fd++) |
| 647 | { |
| 648 | d_idx = od * stride_d + fd * dilation_d; |
| 649 | for (int fh = 0; fh < f_height; fh++) |
| 650 | { |
| 651 | h_idx = oh * stride_h + fh * dilation_h; |
| 652 | for (int fw = 0; fw < f_width; fw++) |
| 653 | { |
| 654 | w_idx = ow * stride_w + fw * dilation_w; |
| 655 | for (int ic = 0; ic < in_channels; ic++) |
| 656 | { |
| 657 | acc += ((AccEigenType)input_padded(ob, d_idx, h_idx, w_idx, ic) * |
| 658 | (AccEigenType)weight_val(oc, fd, fh, fw, ic)); |
| 659 | } |
| 660 | } |
| 661 | } |
| 662 | } |
| 663 | this->output->getTensor()(ob, od, oh, ow, oc) = acc; |
| 664 | } |
| 665 | } |
| 666 | } |
| 667 | } |
| 668 | } |
| 669 | |
| 670 | if (AccDtype == DType_INT48) |
| 671 | { |
| 672 | this->output->getTensor() = this->output->getTensor().cwiseMax((AccEigenType)AccQMin); |
| 673 | this->output->getTensor() = this->output->getTensor().cwiseMin((AccEigenType)AccQMax); |
| 674 | } |
| 675 | |
| 676 | return GraphNode::eval(); |
| 677 | } |
| 678 | |
| 679 | template <DType InDtype, DType WeightDtype> |
Kevin Cheng | acb550f | 2021-06-29 15:32:19 -0700 | [diff] [blame] | 680 | OpDepthwiseConv2d<InDtype, WeightDtype>::OpDepthwiseConv2d(SubgraphTraverser* sgt_, |
| 681 | TosaAttributeBase* attribute_, |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 682 | TosaQuantInfoBase* qinfo_, |
| 683 | uint64_t id_) |
Kevin Cheng | acb550f | 2021-06-29 15:32:19 -0700 | [diff] [blame] | 684 | : GraphNode(sgt_, Op_DEPTHWISE_CONV2D, id_) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 685 | { |
| 686 | setRequiredOperands(3, 1); |
| 687 | setRequiredRank(4); |
| 688 | |
Kevin Cheng | 93a1628 | 2021-08-31 16:14:03 -0700 | [diff] [blame] | 689 | INIT_ATTRIBUTE(Conv); |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 690 | INIT_QINFO(Conv); |
| 691 | } |
| 692 | |
| 693 | template <DType InDtype, DType WeightDtype> |
| 694 | OpDepthwiseConv2d<InDtype, WeightDtype>::~OpDepthwiseConv2d() |
| 695 | { |
| 696 | if (attribute) |
| 697 | delete attribute; |
| 698 | if (qinfo) |
| 699 | delete qinfo; |
| 700 | } |
| 701 | |
| 702 | template <DType InDtype, DType WeightDtype> |
| 703 | int OpDepthwiseConv2d<InDtype, WeightDtype>::checkTensorAttributes() |
| 704 | { |
| 705 | if (validateRequiredOperands()) |
| 706 | return 1; |
| 707 | |
| 708 | if (validateRequiredRank(inputs[0]) || validateRequiredRank(inputs[1]) || validateRequiredRank(outputs[0])) |
| 709 | { |
| 710 | return 1; |
| 711 | } |
| 712 | |
| 713 | // 'bias' checked separatedly since it doens't make sense to make required rank ranging from 1 to 4 |
| 714 | if (inputs[2]->getRank() != 1) |
| 715 | { |
| 716 | printNodeValidationError("OpDepthwiseConv2d: bias tensor must be rank 1"); |
| 717 | } |
| 718 | |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 719 | input = dynamic_cast<TosaReference::TensorTemplate<TIn>*>(inputs[0]); |
| 720 | weight = dynamic_cast<TosaReference::TensorTemplate<TWeight>*>(inputs[1]); |
| 721 | bias = dynamic_cast<TosaReference::TensorTemplate<TBias>*>(inputs[2]); |
| 722 | output = dynamic_cast<TosaReference::TensorTemplate<TAcc>*>(outputs[0]); |
| 723 | |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 724 | if (attribute->padding().size() != 4) |
| 725 | { |
| 726 | printNodeValidationError("OpDepthwiseConv2d: illegal size for attribute padding"); |
| 727 | return 1; |
| 728 | } |
| 729 | |
| 730 | if (attribute->stride().size() != 2) |
| 731 | { |
| 732 | printNodeValidationError("OpDepthwiseConv2d: illegal size for attribute stride"); |
| 733 | return 1; |
| 734 | } |
| 735 | |
| 736 | if (attribute->dilation().size() != 2) |
| 737 | { |
| 738 | printNodeValidationError("OpDepthwiseConv2d: illegal size for attribute dilation"); |
| 739 | return 1; |
| 740 | } |
| 741 | |
| 742 | return 0; |
| 743 | } |
| 744 | |
| 745 | template <DType InDtype, DType WeightDtype> |
| 746 | int OpDepthwiseConv2d<InDtype, WeightDtype>::eval() |
| 747 | { |
| 748 | int in_batch = this->input->getShape()[0]; |
| 749 | int in_height = this->input->getShape()[1]; |
| 750 | int in_width = this->input->getShape()[2]; |
| 751 | int in_channels = this->input->getShape()[3]; |
| 752 | |
| 753 | int f_height = this->weight->getShape()[0]; |
| 754 | int f_width = this->weight->getShape()[1]; |
| 755 | int f_in_channels = this->weight->getShape()[2]; |
| 756 | int f_multiplier = this->weight->getShape()[3]; |
| 757 | |
| 758 | int b_out_channels = this->bias->getShape()[0]; |
| 759 | |
| 760 | int out_batch = this->output->getShape()[0]; |
| 761 | int out_height = this->output->getShape()[1]; |
| 762 | int out_width = this->output->getShape()[2]; |
| 763 | int out_channels = this->output->getShape()[3]; |
| 764 | |
Kevin Cheng | acb550f | 2021-06-29 15:32:19 -0700 | [diff] [blame] | 765 | ERROR_IF(in_batch != out_batch, "OpDepthwiseConv2d: tensor batch mismatch %d != %d", in_batch, out_batch); |
| 766 | ERROR_IF(f_in_channels != in_channels, "OpDepthwiseConv2d: tensor input channel mismatch %d != %d", f_in_channels, |
| 767 | in_channels); |
| 768 | ERROR_IF(in_channels * f_multiplier != out_channels, "OpDepthwiseConv2d: tensor output channel mismatch %d != %d", |
| 769 | in_channels * f_multiplier, out_channels); |
| 770 | ERROR_IF(b_out_channels != out_channels, "OpDepthwiseConv2d: bias channels mismatch %d != %d", b_out_channels, |
| 771 | out_channels); |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 772 | |
| 773 | int padding_top = this->attribute->padding()[0]; |
| 774 | int padding_bottom = this->attribute->padding()[1]; |
| 775 | int padding_left = this->attribute->padding()[2]; |
| 776 | int padding_right = this->attribute->padding()[3]; |
| 777 | int stride_h = this->attribute->stride()[0]; |
| 778 | int stride_w = this->attribute->stride()[1]; |
| 779 | int dilation_h = this->attribute->dilation()[0]; |
| 780 | int dilation_w = this->attribute->dilation()[1]; |
| 781 | |
| 782 | DEBUG_INFO(OP, |
| 783 | "perform OpDepthwiseConv2d, input.shape=[%d,%d,%d,%d], weight.shape=[%d,%d,%d,%d], " |
| 784 | "output.shape=[%d,%d,%d,%d], stride=[%d,%d], dilation=[%d,%d], padding=[%d,%d,%d,%d]", |
| 785 | in_batch, in_height, in_width, in_channels, f_height, f_width, f_in_channels, f_multiplier, out_batch, |
| 786 | out_height, out_width, out_channels, stride_h, stride_w, dilation_h, dilation_w, padding_top, |
| 787 | padding_bottom, padding_left, padding_right); |
| 788 | |
| 789 | Eigen::array<std::pair<int32_t, int32_t>, 4> padding; |
| 790 | padding[0] = std::make_pair(0, 0); |
| 791 | padding[1] = std::make_pair(padding_top, padding_bottom); |
| 792 | padding[2] = std::make_pair(padding_left, padding_right); |
| 793 | padding[3] = std::make_pair(0, 0); |
| 794 | |
| 795 | TIn input_val = this->input->getTensor(); |
| 796 | TWeight weight_val = this->weight->getTensor(); |
| 797 | if (this->qinfo) |
| 798 | { |
| 799 | input_val = input_val - (InEigenType)this->qinfo->input_zp(); |
| 800 | weight_val = weight_val - (WeightEigenType)this->qinfo->weight_zp(); |
| 801 | } |
| 802 | |
| 803 | ETensor4<InEigenType> input_padded = input_val.pad(padding); |
| 804 | |
| 805 | // GEMM doesn't fit well with DepthwiseConv2d |
| 806 | // 1. use extract_image_patches() to handle stride/dilation/padding |
| 807 | // 2. perform direct convolution |
| 808 | |
| 809 | // 1. extract_image_patches() output [N, KH, KW, OH * OW, IC] |
| 810 | ETensor5<InEigenType> input_extract_patches = input_padded.extract_image_patches( |
| 811 | f_height, f_width, stride_h, stride_w, dilation_h, dilation_w, Eigen::PADDING_VALID); |
| 812 | |
| 813 | Eigen::array<Eigen::Index, 4> reshape_dim; |
| 814 | reshape_dim.fill(1); |
| 815 | reshape_dim[3] = b_out_channels; |
| 816 | |
| 817 | Eigen::array<Eigen::Index, 4> bcast; |
| 818 | bcast[0] = out_batch; |
| 819 | bcast[1] = out_height; |
| 820 | bcast[2] = out_width; |
| 821 | bcast[3] = 1; |
| 822 | |
| 823 | // initialize with bias |
| 824 | this->output->getTensor() = this->bias->getTensor().reshape(reshape_dim).broadcast(bcast); |
| 825 | |
| 826 | // 2. direct depthwise convolution |
| 827 | for (int ob = 0; ob < out_batch; ob++) |
| 828 | { |
| 829 | for (int oh = 0; oh < out_height; oh++) |
| 830 | { |
| 831 | for (int ow = 0; ow < out_width; ow++) |
| 832 | { |
| 833 | for (int ic = 0; ic < in_channels; ic++) |
| 834 | { |
| 835 | for (int cm = 0; cm < f_multiplier; cm++) |
| 836 | { |
| 837 | for (int fh = 0; fh < f_height; fh++) |
| 838 | { |
| 839 | for (int fw = 0; fw < f_width; fw++) |
| 840 | { |
| 841 | this->output->getTensor()(ob, oh, ow, ic * f_multiplier + cm) += |
| 842 | ((AccEigenType)input_extract_patches(ob, fh, fw, ow * out_height + oh, ic) * |
| 843 | (AccEigenType)weight_val(fh, fw, ic, cm)); |
| 844 | } |
| 845 | } |
| 846 | } |
| 847 | } |
| 848 | } |
| 849 | } |
| 850 | } |
| 851 | |
| 852 | if (AccDtype == DType_INT48) |
| 853 | { |
| 854 | this->output->getTensor() = this->output->getTensor().cwiseMax((AccEigenType)AccQMin); |
| 855 | this->output->getTensor() = this->output->getTensor().cwiseMin((AccEigenType)AccQMax); |
| 856 | } |
| 857 | |
| 858 | return GraphNode::eval(); |
| 859 | } |
| 860 | |
| 861 | template <DType InDtype, DType WeightDtype> |
Kevin Cheng | acb550f | 2021-06-29 15:32:19 -0700 | [diff] [blame] | 862 | OpFullyConnected<InDtype, WeightDtype>::OpFullyConnected(SubgraphTraverser* sgt_, |
| 863 | TosaAttributeBase* attribute_, |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 864 | TosaQuantInfoBase* qinfo_, |
| 865 | uint64_t id_) |
Kevin Cheng | acb550f | 2021-06-29 15:32:19 -0700 | [diff] [blame] | 866 | : GraphNode(sgt_, Op_FULLY_CONNECTED, id_) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 867 | { |
| 868 | setRequiredOperands(3, 1); |
| 869 | setRequiredRank(2); |
| 870 | |
| 871 | INIT_QINFO(Conv); |
| 872 | } |
| 873 | |
| 874 | template <DType InDtype, DType WeightDtype> |
| 875 | OpFullyConnected<InDtype, WeightDtype>::~OpFullyConnected() |
| 876 | { |
| 877 | if (qinfo) |
| 878 | delete qinfo; |
| 879 | } |
| 880 | |
| 881 | template <DType InDtype, DType WeightDtype> |
| 882 | int OpFullyConnected<InDtype, WeightDtype>::checkTensorAttributes() |
| 883 | { |
| 884 | if (validateRequiredOperands()) |
| 885 | return 1; |
| 886 | |
| 887 | if (validateRequiredRank(inputs[0]) || validateRequiredRank(inputs[1]) || validateRequiredRank(outputs[0])) |
| 888 | { |
| 889 | return 1; |
| 890 | } |
| 891 | |
| 892 | input = dynamic_cast<TosaReference::TensorTemplate<TIn>*>(inputs[0]); |
| 893 | weight = dynamic_cast<TosaReference::TensorTemplate<TWeight>*>(inputs[1]); |
| 894 | bias = dynamic_cast<TosaReference::TensorTemplate<TBias>*>(inputs[2]); |
| 895 | |
| 896 | if (input->getShape()[1] != weight->getShape()[1]) |
| 897 | { |
| 898 | printNodeValidationError("OpFullyConnected operator input.shape[1] should match weight.shape[1]"); |
| 899 | return 1; |
| 900 | } |
| 901 | |
| 902 | if (weight->getShape()[0] != bias->getShape()[0]) |
| 903 | { |
| 904 | printNodeValidationError("OpFullyConnected operator bias.shape[0] should match weight.shape[0]"); |
| 905 | return 1; |
| 906 | } |
| 907 | |
| 908 | output = dynamic_cast<TosaReference::TensorTemplate<TAcc>*>(outputs[0]); |
| 909 | |
| 910 | return 0; |
| 911 | } |
| 912 | |
| 913 | template <DType InDtype, DType WeightDtype> |
| 914 | int OpFullyConnected<InDtype, WeightDtype>::eval() |
| 915 | { |
| 916 | typedef Eigen::Tensor<int, 1>::DimensionPair DimPair; |
| 917 | Eigen::array<DimPair, 1> dims{ { DimPair(1, 0) } }; |
| 918 | |
| 919 | Eigen::array<Eigen::Index, 2> weight_shuffle{ 1, 0 }; |
| 920 | |
| 921 | Eigen::array<Eigen::Index, 2> bias_reshape; |
| 922 | bias_reshape[0] = 1; |
| 923 | bias_reshape[1] = this->bias->getShape()[0]; |
| 924 | |
| 925 | Eigen::array<Eigen::Index, 2> bias_bcast; |
| 926 | bias_bcast[0] = this->input->getShape()[0]; |
| 927 | bias_bcast[1] = 1; |
| 928 | |
| 929 | TIn input_val = this->input->getTensor(); |
| 930 | TWeight weight_val = this->weight->getTensor().shuffle(weight_shuffle); |
| 931 | if (this->qinfo) |
| 932 | { |
| 933 | input_val = input_val - (InEigenType)this->qinfo->input_zp(); |
| 934 | weight_val = weight_val - (WeightEigenType)this->qinfo->weight_zp(); |
| 935 | } |
| 936 | |
| 937 | this->output->getTensor() = |
| 938 | input_val.template cast<AccEigenType>().contract(weight_val.template cast<AccEigenType>(), dims) + |
| 939 | this->bias->getTensor().reshape(bias_reshape).broadcast(bias_bcast); |
| 940 | |
| 941 | if (AccDtype == DType_INT48) |
| 942 | { |
| 943 | this->output->getTensor() = this->output->getTensor().cwiseMax((AccEigenType)AccQMin); |
| 944 | this->output->getTensor() = this->output->getTensor().cwiseMin((AccEigenType)AccQMax); |
| 945 | } |
| 946 | return GraphNode::eval(); |
| 947 | } |
| 948 | |
| 949 | template <DType Dtype> |
Kevin Cheng | acb550f | 2021-06-29 15:32:19 -0700 | [diff] [blame] | 950 | OpMatMul<Dtype>::OpMatMul(SubgraphTraverser* sgt_, |
| 951 | TosaAttributeBase* attribute_, |
| 952 | TosaQuantInfoBase* qinfo_, |
| 953 | uint64_t id_) |
| 954 | : GraphNode(sgt_, Op_MATMUL, id_) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 955 | { |
| 956 | setRequiredOperands(2, 1); |
Kevin Cheng | 2d60f00 | 2021-06-09 14:18:32 -0700 | [diff] [blame] | 957 | setRequiredRank(3); |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 958 | |
| 959 | INIT_QINFO(MatMul); |
| 960 | } |
| 961 | |
| 962 | template <DType Dtype> |
| 963 | OpMatMul<Dtype>::~OpMatMul() |
| 964 | { |
| 965 | if (qinfo) |
| 966 | delete qinfo; |
| 967 | } |
| 968 | |
| 969 | template <DType Dtype> |
| 970 | int OpMatMul<Dtype>::checkTensorAttributes() |
| 971 | { |
| 972 | if (validateRequiredOperands()) |
| 973 | return 1; |
| 974 | |
| 975 | if (validateRequiredRank(inputs[0]) || validateRequiredRank(inputs[1]) || validateRequiredRank(outputs[0])) |
| 976 | { |
| 977 | return 1; |
| 978 | } |
| 979 | |
Kevin Cheng | 2d60f00 | 2021-06-09 14:18:32 -0700 | [diff] [blame] | 980 | a = dynamic_cast<TosaReference::TensorTemplate<TIn>*>(inputs[0]); |
| 981 | b = dynamic_cast<TosaReference::TensorTemplate<TIn>*>(inputs[1]); |
| 982 | output = dynamic_cast<TosaReference::TensorTemplate<TAcc>*>(outputs[0]); |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 983 | |
Kevin Cheng | 2d60f00 | 2021-06-09 14:18:32 -0700 | [diff] [blame] | 984 | ASSERT_MEM(a && b && output); |
| 985 | |
| 986 | // a: [N, H, C] |
| 987 | // b: [N, C, W] |
| 988 | // c: [N, H, W] |
| 989 | |
| 990 | // Check N |
| 991 | if (a->getShape()[0] != b->getShape()[0] || a->getShape()[0] != output->getShape()[0]) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 992 | { |
Kevin Cheng | 2d60f00 | 2021-06-09 14:18:32 -0700 | [diff] [blame] | 993 | printNodeValidationError("OpMatMul operator a.shape[0], b.shape[0] and output.shape[0] should match"); |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 994 | return 1; |
| 995 | } |
Kevin Cheng | 2d60f00 | 2021-06-09 14:18:32 -0700 | [diff] [blame] | 996 | N = a->getShape()[0]; |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 997 | |
Kevin Cheng | 2d60f00 | 2021-06-09 14:18:32 -0700 | [diff] [blame] | 998 | // Check C |
| 999 | if (a->getShape()[2] != b->getShape()[1]) |
| 1000 | { |
| 1001 | printNodeValidationError("OpMatMul operator a.shape[2] should match b.shape[1]"); |
| 1002 | return 1; |
| 1003 | } |
| 1004 | C = a->getShape()[2]; |
| 1005 | |
| 1006 | // Check H |
| 1007 | if (a->getShape()[1] != output->getShape()[1]) |
| 1008 | { |
| 1009 | printNodeValidationError("OpMatMul operator a.shape[1] should match output.shape[1]"); |
| 1010 | return 1; |
| 1011 | } |
| 1012 | H = a->getShape()[1]; |
| 1013 | |
| 1014 | // Check W |
| 1015 | if (b->getShape()[2] != output->getShape()[2]) |
| 1016 | { |
| 1017 | printNodeValidationError("OpMatMul operator output.shape[2] should match output.shape[2]"); |
| 1018 | return 1; |
| 1019 | } |
| 1020 | W = b->getShape()[2]; |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1021 | |
| 1022 | return 0; |
| 1023 | } |
| 1024 | |
| 1025 | template <DType Dtype> |
| 1026 | int OpMatMul<Dtype>::eval() |
| 1027 | { |
| 1028 | typedef Eigen::Tensor<int, 1>::DimensionPair DimPair; |
| 1029 | Eigen::array<DimPair, 1> dims{ { DimPair(1, 0) } }; |
| 1030 | |
| 1031 | TIn a_val = this->a->getTensor(); |
| 1032 | TIn b_val = this->b->getTensor(); |
| 1033 | if (this->qinfo) |
| 1034 | { |
| 1035 | a_val = a_val - (InEigenType)this->qinfo->a_zp(); |
| 1036 | b_val = b_val - (InEigenType)this->qinfo->b_zp(); |
| 1037 | } |
| 1038 | |
Kevin Cheng | 2d60f00 | 2021-06-09 14:18:32 -0700 | [diff] [blame] | 1039 | Eigen::array<Eigen::Index, 2> a_rank2_shape({ H, C }); |
| 1040 | Eigen::array<Eigen::Index, 2> b_rank2_shape({ C, W }); |
| 1041 | Eigen::array<Eigen::Index, 3> output_rank3_shape({ 1, H, W }); |
| 1042 | |
| 1043 | Eigen::array<Eigen::Index, 3> a_size_array({ 1, H, C }); |
| 1044 | Eigen::array<Eigen::Index, 3> b_size_array({ 1, C, W }); |
| 1045 | |
| 1046 | Eigen::array<Eigen::Index, 3> a_begin_array({ 0, 0, 0 }); |
| 1047 | Eigen::array<Eigen::Index, 3> b_begin_array({ 0, 0, 0 }); |
| 1048 | |
| 1049 | // Iterate N dimension. |
| 1050 | for (int i = 0; i < N; i++) |
| 1051 | { |
| 1052 | a_begin_array[0] = i; |
| 1053 | b_begin_array[0] = i; |
| 1054 | |
| 1055 | TInRank2 a_rank2_val = a_val.slice(a_begin_array, a_size_array).reshape(a_rank2_shape); |
| 1056 | TInRank2 b_rank2_val = b_val.slice(b_begin_array, b_size_array).reshape(b_rank2_shape); |
| 1057 | TAccRank2 output_rank2_val = |
| 1058 | a_rank2_val.template cast<AccEigenType>().contract(b_rank2_val.template cast<AccEigenType>(), dims); |
| 1059 | TAcc output_rank3_val = output_rank2_val.reshape(output_rank3_shape); |
| 1060 | if (i == 0) |
| 1061 | { |
| 1062 | this->output->getTensor() = output_rank3_val; |
| 1063 | } |
| 1064 | else |
| 1065 | { |
| 1066 | TAcc temp = this->output->getTensor().concatenate(output_rank3_val, 0); |
| 1067 | this->output->getTensor() = temp; |
| 1068 | } |
| 1069 | } |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1070 | |
| 1071 | if (AccDtype == DType_INT48) |
| 1072 | { |
Kevin Cheng | 2d60f00 | 2021-06-09 14:18:32 -0700 | [diff] [blame] | 1073 | this->output->getTensor() = this->output->getTensor().cwiseMax((AccEigenType)AccQMin); |
| 1074 | this->output->getTensor() = this->output->getTensor().cwiseMin((AccEigenType)AccQMax); |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1075 | } |
| 1076 | |
| 1077 | return GraphNode::eval(); |
| 1078 | } |
| 1079 | |
| 1080 | template <DType Dtype> |
Kevin Cheng | acb550f | 2021-06-29 15:32:19 -0700 | [diff] [blame] | 1081 | OpMaxPool2d<Dtype>::OpMaxPool2d(SubgraphTraverser* sgt_, |
| 1082 | TosaAttributeBase* attribute_, |
| 1083 | TosaQuantInfoBase* qinfo_, |
| 1084 | uint64_t id_) |
| 1085 | : GraphNode(sgt_, Op_MAX_POOL2D, id_) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1086 | { |
| 1087 | setRequiredOperands(1, 1); |
| 1088 | setRequiredRank(4); |
| 1089 | |
Kevin Cheng | 93a1628 | 2021-08-31 16:14:03 -0700 | [diff] [blame] | 1090 | INIT_ATTRIBUTE(Pool); |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1091 | } |
| 1092 | |
| 1093 | template <DType Dtype> |
| 1094 | OpMaxPool2d<Dtype>::~OpMaxPool2d() |
| 1095 | { |
| 1096 | if (attribute) |
| 1097 | delete attribute; |
| 1098 | } |
| 1099 | |
| 1100 | template <DType Dtype> |
| 1101 | int OpMaxPool2d<Dtype>::checkTensorAttributes() |
| 1102 | { |
| 1103 | if (validateRequiredOperands()) |
| 1104 | return 1; |
| 1105 | |
| 1106 | if (validateRequiredRank(inputs[0]) || validateRequiredRank(outputs[0])) |
| 1107 | { |
| 1108 | return 1; |
| 1109 | } |
| 1110 | |
| 1111 | if (inputs[0]->matchType(*outputs[0])) |
| 1112 | { |
| 1113 | printNodeValidationError("OpMaxPool2d: input and output tensor type mismatch"); |
| 1114 | return 1; |
| 1115 | } |
| 1116 | |
| 1117 | in = dynamic_cast<TosaReference::TensorTemplate<TIn>*>(inputs[0]); |
| 1118 | out = dynamic_cast<TosaReference::TensorTemplate<TOut>*>(outputs[0]); |
| 1119 | |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1120 | if (attribute->padding().size() != 4) |
| 1121 | { |
| 1122 | printNodeValidationError("OpMaxPool2d: illegal size for attribute padding"); |
| 1123 | return 1; |
| 1124 | } |
| 1125 | |
| 1126 | if (attribute->kernel().size() != 2) |
| 1127 | { |
| 1128 | printNodeValidationError("OpMaxPool2d: illegal size for attribute kernel"); |
| 1129 | return 1; |
| 1130 | } |
| 1131 | |
| 1132 | if (attribute->stride().size() != 2) |
| 1133 | { |
| 1134 | printNodeValidationError("OpMaxPool2d: illegal size for attribute stride"); |
| 1135 | return 1; |
| 1136 | } |
| 1137 | |
| 1138 | return 0; |
| 1139 | } |
| 1140 | |
| 1141 | template <DType Dtype> |
| 1142 | int OpMaxPool2d<Dtype>::eval() |
| 1143 | { |
| 1144 | int in_batch = this->in->getShape()[0]; |
| 1145 | int in_height = this->in->getShape()[1]; |
| 1146 | int in_width = this->in->getShape()[2]; |
| 1147 | int in_channels = this->in->getShape()[3]; |
| 1148 | |
| 1149 | int out_batch = this->out->getShape()[0]; |
| 1150 | int out_height = this->out->getShape()[1]; |
| 1151 | int out_width = this->out->getShape()[2]; |
| 1152 | int out_channels = this->out->getShape()[3]; |
| 1153 | |
Kevin Cheng | acb550f | 2021-06-29 15:32:19 -0700 | [diff] [blame] | 1154 | ERROR_IF(in_batch != out_batch, "OpMaxPool2d: tensor batch mismatch %d != %d", in_batch, out_batch); |
| 1155 | ERROR_IF(in_channels != out_channels, "OpMaxPool2d: tensor channel mismatch %d != %d", in_channels, out_channels); |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1156 | |
| 1157 | int padding_top = this->attribute->padding()[0]; |
| 1158 | int padding_bottom = this->attribute->padding()[1]; |
| 1159 | int padding_left = this->attribute->padding()[2]; |
| 1160 | int padding_right = this->attribute->padding()[3]; |
| 1161 | int kernel_h = this->attribute->kernel()[0]; |
| 1162 | int kernel_w = this->attribute->kernel()[1]; |
| 1163 | int stride_h = this->attribute->stride()[0]; |
| 1164 | int stride_w = this->attribute->stride()[1]; |
| 1165 | |
| 1166 | DEBUG_INFO(OP, |
| 1167 | "perform MaxPool2d, input.shape=[%d,%d,%d,%d], output.shape=[%d,%d,%d,%d], kernel=[%d,%d], " |
| 1168 | "stride=[%d,%d], padding=[%d,%d,%d,%d]", |
| 1169 | in_batch, in_height, in_width, in_channels, out_batch, out_height, out_width, out_channels, kernel_h, |
| 1170 | kernel_w, stride_h, stride_w, padding_top, padding_bottom, padding_left, padding_right); |
| 1171 | |
| 1172 | Eigen::array<Eigen::Index, 2> im2col_input_dims; |
| 1173 | im2col_input_dims[0] = kernel_h * kernel_w; |
| 1174 | im2col_input_dims[1] = out_batch * out_height * out_width * out_channels; |
| 1175 | |
| 1176 | Eigen::array<Eigen::Index, 4> col2im_output_dims; |
| 1177 | col2im_output_dims[0] = out_batch; |
| 1178 | col2im_output_dims[1] = out_height; |
| 1179 | col2im_output_dims[2] = out_width; |
| 1180 | col2im_output_dims[3] = out_channels; |
| 1181 | |
| 1182 | Eigen::array<std::pair<int32_t, int32_t>, 4> padding; |
| 1183 | padding[0] = std::make_pair(0, 0); |
| 1184 | padding[1] = std::make_pair(padding_top, padding_bottom); |
| 1185 | padding[2] = std::make_pair(padding_left, padding_right); |
| 1186 | padding[3] = std::make_pair(0, 0); |
| 1187 | |
| 1188 | ETensor4<InEigenType> input_padded = this->in->getTensor().pad(padding, std::numeric_limits<InEigenType>::lowest()); |
| 1189 | |
| 1190 | // extract_image_patches() output [N, KH, KW, H * W, C] |
| 1191 | // transpose to [KH, KW, N, H * W, C] |
| 1192 | // reshape to [KH * KW, N * H * W * C] |
| 1193 | // |
| 1194 | // Set the padding value to be the most negative value that can be |
| 1195 | // represented by the datatype to ensure that any padding values will be equal |
| 1196 | // to or smaller than the actual maximum in the KH x KW patch. |
| 1197 | ETensor2<InEigenType> input_extract_patches = |
| 1198 | input_padded |
| 1199 | .extract_image_patches(kernel_h, kernel_w, stride_h, stride_w, 1, 1, Eigen::PADDING_VALID, |
| 1200 | std::numeric_limits<InEigenType>::lowest()) |
| 1201 | .shuffle(Eigen::array<Eigen::Index, 5>{ 1, 2, 0, 3, 4 }) |
| 1202 | .reshape(im2col_input_dims); |
| 1203 | |
| 1204 | // Get the maximum of the KHxHW patches along axis 0 |
| 1205 | Eigen::Tensor<DenseIndex, 1> tensor_argmax = input_extract_patches.argmax(0); |
| 1206 | |
| 1207 | // 1D result with [N * H * W * C] |
| 1208 | ETensor1<OutEigenType> out_1d(this->out->getElementCount()); |
| 1209 | |
| 1210 | // index input_patches with argmax array should give the result |
| 1211 | for (size_t i = 0; i < this->out->getElementCount(); i++) |
| 1212 | { |
| 1213 | out_1d(i) = (OutEigenType)input_extract_patches(tensor_argmax(i), i); |
| 1214 | } |
| 1215 | |
| 1216 | // reshape result to [N, H, W, C] |
| 1217 | this->out->getTensor() = out_1d.reshape(col2im_output_dims); |
| 1218 | |
| 1219 | return GraphNode::eval(); |
| 1220 | } |
| 1221 | |
| 1222 | template <DType InDtype, DType OutDtype> |
Kevin Cheng | acb550f | 2021-06-29 15:32:19 -0700 | [diff] [blame] | 1223 | OpTransposeConv2d<InDtype, OutDtype>::OpTransposeConv2d(SubgraphTraverser* sgt_, |
| 1224 | TosaAttributeBase* attribute_, |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1225 | TosaQuantInfoBase* qinfo_, |
| 1226 | uint64_t id_) |
Kevin Cheng | acb550f | 2021-06-29 15:32:19 -0700 | [diff] [blame] | 1227 | : GraphNode(sgt_, Op_TRANSPOSE_CONV2D, id_) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1228 | { |
| 1229 | setRequiredOperands(3, 1); |
| 1230 | setRequiredRank(4); |
| 1231 | |
Kevin Cheng | 93a1628 | 2021-08-31 16:14:03 -0700 | [diff] [blame] | 1232 | INIT_ATTRIBUTE(TransposeConv); |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1233 | INIT_QINFO(Conv); |
| 1234 | } |
| 1235 | |
| 1236 | template <DType InDtype, DType OutDtype> |
| 1237 | OpTransposeConv2d<InDtype, OutDtype>::~OpTransposeConv2d() |
| 1238 | { |
| 1239 | if (attribute) |
| 1240 | delete attribute; |
| 1241 | if (qinfo) |
| 1242 | delete qinfo; |
| 1243 | } |
| 1244 | |
| 1245 | template <DType InDtype, DType OutDtype> |
| 1246 | int OpTransposeConv2d<InDtype, OutDtype>::checkTensorAttributes() |
| 1247 | { |
| 1248 | if (validateRequiredOperands()) |
| 1249 | return 1; |
| 1250 | |
| 1251 | if (validateRequiredRank(inputs[0]) || validateRequiredRank(inputs[1]) || validateRequiredRank(outputs[0])) |
| 1252 | { |
| 1253 | return 1; |
| 1254 | } |
| 1255 | |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1256 | input = dynamic_cast<TosaReference::TensorTemplate<TIn>*>(inputs[0]); |
| 1257 | weight = dynamic_cast<TosaReference::TensorTemplate<TWeight>*>(inputs[1]); |
| 1258 | bias = dynamic_cast<TosaReference::TensorTemplate<TBias>*>(inputs[2]); |
| 1259 | output = dynamic_cast<TosaReference::TensorTemplate<TAcc>*>(outputs[0]); |
| 1260 | |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1261 | if (attribute->outpad().size() != 2) |
| 1262 | { |
| 1263 | printNodeValidationError("OpTransposeConv2d: illegal size for attribute outpad"); |
| 1264 | return 1; |
| 1265 | } |
| 1266 | |
| 1267 | if (attribute->stride().size() != 2) |
| 1268 | { |
| 1269 | printNodeValidationError("OpTransposeConv2d: illegal size for attribute stride"); |
| 1270 | return 1; |
| 1271 | } |
| 1272 | |
| 1273 | if (attribute->dilation().size() != 2) |
| 1274 | { |
| 1275 | printNodeValidationError("OpTransposeConv2d: illegal size for attribute dilation"); |
| 1276 | return 1; |
| 1277 | } |
| 1278 | |
| 1279 | if (attribute->output_shape().size() != 4) |
| 1280 | { |
| 1281 | printNodeValidationError("OpTransposeConv2d: illegal size for attribute output_shape"); |
| 1282 | return 1; |
| 1283 | } |
| 1284 | |
| 1285 | for (int d = 0; d < 4; d++) |
| 1286 | { |
| 1287 | if (attribute->output_shape()[d] != this->output->getShape()[d]) |
| 1288 | { |
| 1289 | printNodeValidationError("OpTransposeConv2d: illegal size for attribute output_shape"); |
| 1290 | return 1; |
| 1291 | } |
| 1292 | } |
| 1293 | |
| 1294 | return 0; |
| 1295 | } |
| 1296 | |
| 1297 | template <DType InDtype, DType OutDtype> |
| 1298 | int OpTransposeConv2d<InDtype, OutDtype>::eval() |
| 1299 | { |
| 1300 | int in_batch = this->input->getShape()[0]; |
| 1301 | int in_height = this->input->getShape()[1]; |
| 1302 | int in_width = this->input->getShape()[2]; |
| 1303 | int in_channels = this->input->getShape()[3]; |
| 1304 | |
| 1305 | int f_out_channels = this->weight->getShape()[0]; |
| 1306 | int f_height = this->weight->getShape()[1]; |
| 1307 | int f_width = this->weight->getShape()[2]; |
| 1308 | int f_in_channels = this->weight->getShape()[3]; |
| 1309 | |
| 1310 | int b_out_channels = this->bias->getShape()[0]; |
| 1311 | |
| 1312 | int out_batch = this->output->getShape()[0]; |
| 1313 | int out_height = this->output->getShape()[1]; |
| 1314 | int out_width = this->output->getShape()[2]; |
| 1315 | int out_channels = this->output->getShape()[3]; |
| 1316 | |
| 1317 | int padding_top = this->attribute->outpad()[0]; |
| 1318 | int padding_left = this->attribute->outpad()[1]; |
| 1319 | int stride_h = this->attribute->stride()[0]; |
| 1320 | int stride_w = this->attribute->stride()[1]; |
| 1321 | int dilation_h = this->attribute->dilation()[0]; |
| 1322 | int dilation_w = this->attribute->dilation()[1]; |
| 1323 | |
Kevin Cheng | acb550f | 2021-06-29 15:32:19 -0700 | [diff] [blame] | 1324 | ERROR_IF(in_batch != out_batch, "OpTransposeConv2d: tensor batch mismatch %d != %d", in_batch, out_batch); |
| 1325 | ERROR_IF(f_in_channels != in_channels, "OpTransposeConv2d: tensor input channel mismatch %d != %d", f_in_channels, |
| 1326 | in_channels); |
| 1327 | ERROR_IF(f_out_channels != out_channels, "OpTransposeConv2d: tensor output channel mismatch %d != %d", |
| 1328 | f_out_channels, out_channels); |
| 1329 | ERROR_IF(b_out_channels != out_channels, "OpDepthwiseConv2d: bias channels mismatch %d != %d", b_out_channels, |
| 1330 | out_channels); |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1331 | |
| 1332 | DEBUG_INFO(OP, |
| 1333 | "perform OpTransposeConv2d, input.shape=[%d,%d,%d,%d], weight.shape=[%d,%d,%d,%d], " |
| 1334 | "output.shape=[%d,%d,%d,%d], stride=[%d,%d], dilation=[%d,%d], padding=[%d,%d]", |
| 1335 | in_batch, in_height, in_width, in_channels, f_height, f_width, f_out_channels, f_in_channels, out_batch, |
| 1336 | out_height, out_width, out_channels, stride_h, stride_w, dilation_h, dilation_w, padding_top, |
| 1337 | padding_left); |
| 1338 | |
| 1339 | TIn input_val = this->input->getTensor(); |
| 1340 | TWeight weight_val = this->weight->getTensor(); |
| 1341 | if (this->qinfo) |
| 1342 | { |
| 1343 | input_val = input_val - (InEigenType)this->qinfo->input_zp(); |
| 1344 | weight_val = weight_val - (WeightEigenType)this->qinfo->weight_zp(); |
| 1345 | } |
| 1346 | |
| 1347 | Eigen::array<Eigen::Index, 4> reshape_dim; |
| 1348 | reshape_dim.fill(1); |
| 1349 | reshape_dim[3] = b_out_channels; |
| 1350 | |
| 1351 | Eigen::array<Eigen::Index, 4> bcast; |
| 1352 | bcast[0] = out_batch; |
| 1353 | bcast[1] = out_height; |
| 1354 | bcast[2] = out_width; |
| 1355 | bcast[3] = 1; |
| 1356 | |
| 1357 | // initialize with bias |
| 1358 | this->output->getTensor() = this->bias->getTensor().reshape(reshape_dim).broadcast(bcast); |
| 1359 | |
| 1360 | int out_x_origin, out_y_origin; |
| 1361 | int out_x, out_y; |
| 1362 | |
| 1363 | // reference implementation from: tensorflow/tensorflow/lite/kernels/internal/reference/reference_ops.h |
| 1364 | for (int ob = 0; ob < out_batch; ob++) |
| 1365 | { |
| 1366 | for (int ih = 0; ih < in_height; ih++) |
| 1367 | { |
| 1368 | for (int iw = 0; iw < in_width; iw++) |
| 1369 | { |
| 1370 | out_x_origin = iw * stride_w - padding_left; |
| 1371 | out_y_origin = ih * stride_h - padding_top; |
| 1372 | for (int ic = 0; ic < in_channels; ic++) |
| 1373 | { |
| 1374 | for (int fh = 0; fh < f_height; fh++) |
| 1375 | { |
| 1376 | for (int fw = 0; fw < f_width; fw++) |
| 1377 | { |
| 1378 | out_x = out_x_origin + fw * dilation_w; |
| 1379 | out_y = out_y_origin + fh * dilation_h; |
| 1380 | for (int oc = 0; oc < out_channels; oc++) |
| 1381 | { |
| 1382 | if ((out_x >= 0 && out_x < out_width) && (out_y >= 0 && out_y < out_height)) |
| 1383 | { |
| 1384 | this->output->getTensor()(ob, out_y, out_x, oc) += |
| 1385 | ((AccEigenType)input_val(ob, ih, iw, ic) * |
| 1386 | (AccEigenType)weight_val(oc, fh, fw, ic)); |
| 1387 | } |
| 1388 | } |
| 1389 | } |
| 1390 | } |
| 1391 | } |
| 1392 | } |
| 1393 | } |
| 1394 | } |
| 1395 | |
| 1396 | if (AccDtype == DType_INT48) |
| 1397 | { |
| 1398 | this->output->getTensor() = this->output->getTensor().cwiseMax((AccEigenType)AccQMin); |
| 1399 | this->output->getTensor() = this->output->getTensor().cwiseMin((AccEigenType)AccQMax); |
| 1400 | } |
| 1401 | |
| 1402 | return GraphNode::eval(); |
| 1403 | } |
| 1404 | |
| 1405 | // template explicit instantiation |
| 1406 | DEF_INSTANTIATE_RANK1_6_ONE_RANK_ONE_TYPE(OpArgMax, FLOAT); |
Kevin Cheng | 3a47857 | 2021-01-22 17:21:02 -0800 | [diff] [blame] | 1407 | DEF_INSTANTIATE_RANK1_6_ONE_RANK_ONE_TYPE(OpArgMax, INT8); |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1408 | DEF_INSTANTIATE_RANK1_6_ONE_RANK_ONE_TYPE(OpArgMax, INT16); |
| 1409 | |
| 1410 | DEF_INSTANTIATE_ONE_TYPE(OpAvgPool2d, FLOAT) |
Kevin Cheng | 3a47857 | 2021-01-22 17:21:02 -0800 | [diff] [blame] | 1411 | DEF_INSTANTIATE_ONE_TYPE(OpAvgPool2d, INT8) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1412 | DEF_INSTANTIATE_ONE_TYPE(OpAvgPool2d, INT16) |
| 1413 | |
| 1414 | DEF_INSTANTIATE_TWO_TYPE(OpConv2d, FLOAT, FLOAT); |
Kevin Cheng | 3a47857 | 2021-01-22 17:21:02 -0800 | [diff] [blame] | 1415 | DEF_INSTANTIATE_TWO_TYPE(OpConv2d, INT8, INT4); |
| 1416 | DEF_INSTANTIATE_TWO_TYPE(OpConv2d, INT8, INT8); |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1417 | DEF_INSTANTIATE_TWO_TYPE(OpConv2d, INT16, INT8); |
| 1418 | |
Kevin Cheng | 1533b85 | 2021-09-01 12:51:58 -0700 | [diff] [blame^] | 1419 | DEF_INSTANTIATE_TWO_TYPE(OpConv3d, FLOAT, FLOAT); |
| 1420 | DEF_INSTANTIATE_TWO_TYPE(OpConv3d, INT8, INT4); |
| 1421 | DEF_INSTANTIATE_TWO_TYPE(OpConv3d, INT8, INT8); |
| 1422 | DEF_INSTANTIATE_TWO_TYPE(OpConv3d, INT16, INT8); |
| 1423 | |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1424 | DEF_INSTANTIATE_TWO_TYPE(OpDepthwiseConv2d, FLOAT, FLOAT); |
Kevin Cheng | 3a47857 | 2021-01-22 17:21:02 -0800 | [diff] [blame] | 1425 | DEF_INSTANTIATE_TWO_TYPE(OpDepthwiseConv2d, INT8, INT4); |
| 1426 | DEF_INSTANTIATE_TWO_TYPE(OpDepthwiseConv2d, INT8, INT8); |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1427 | DEF_INSTANTIATE_TWO_TYPE(OpDepthwiseConv2d, INT16, INT8); |
| 1428 | |
| 1429 | DEF_INSTANTIATE_TWO_TYPE(OpFullyConnected, FLOAT, FLOAT); |
Kevin Cheng | 3a47857 | 2021-01-22 17:21:02 -0800 | [diff] [blame] | 1430 | DEF_INSTANTIATE_TWO_TYPE(OpFullyConnected, INT8, INT4); |
| 1431 | DEF_INSTANTIATE_TWO_TYPE(OpFullyConnected, INT8, INT8); |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1432 | DEF_INSTANTIATE_TWO_TYPE(OpFullyConnected, INT16, INT8); |
| 1433 | |
Kevin Cheng | 3a47857 | 2021-01-22 17:21:02 -0800 | [diff] [blame] | 1434 | DEF_INSTANTIATE_ONE_TYPE(OpMatMul, INT8); |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1435 | DEF_INSTANTIATE_ONE_TYPE(OpMatMul, INT16); |
| 1436 | DEF_INSTANTIATE_ONE_TYPE(OpMatMul, FLOAT); |
| 1437 | |
| 1438 | DEF_INSTANTIATE_ONE_TYPE(OpMaxPool2d, FLOAT); |
Kevin Cheng | 3a47857 | 2021-01-22 17:21:02 -0800 | [diff] [blame] | 1439 | DEF_INSTANTIATE_ONE_TYPE(OpMaxPool2d, INT8); |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1440 | DEF_INSTANTIATE_ONE_TYPE(OpMaxPool2d, INT16); |
| 1441 | |
| 1442 | DEF_INSTANTIATE_TWO_TYPE(OpTransposeConv2d, FLOAT, FLOAT); |
Kevin Cheng | 3a47857 | 2021-01-22 17:21:02 -0800 | [diff] [blame] | 1443 | DEF_INSTANTIATE_TWO_TYPE(OpTransposeConv2d, INT8, INT4); |
| 1444 | DEF_INSTANTIATE_TWO_TYPE(OpTransposeConv2d, INT8, INT8); |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1445 | DEF_INSTANTIATE_TWO_TYPE(OpTransposeConv2d, INT16, INT8); |