op_teller.cc 59.5 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
// Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
//     http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.

#include "paddle/fluid/inference/tensorrt/op_teller.h"
16
#include <bitset>
17
#include "paddle/fluid/framework/block_desc.h"
18
#include "paddle/fluid/framework/data_layout.h"
19

W
wanghuancoder 已提交
20 21 22 23 24 25
namespace paddle {
namespace framework {
class OpDesc;
}  // namespace framework
}  // namespace paddle

26 27 28 29 30 31
namespace paddle {
namespace inference {
namespace tensorrt {

// Just tell by the op_types.
struct SimpleOpTypeSetTeller : public Teller {
32 33 34
  SimpleOpTypeSetTeller() {
#if IS_TRT_VERSION_GE(5130)
    teller_set.insert("relu6");
35
    teller_set.insert("hard_sigmoid");
P
Pei Yang 已提交
36
    teller_set.insert("clip");
37 38
    int8_teller_set.insert("relu6");
    int8_teller_set.insert("hard_sigmoid");
P
Pei Yang 已提交
39
    int8_teller_set.insert("clip");
40 41 42 43 44
#endif
#if IS_TRT_VERSION_GE(6000)
    teller_set.insert("fused_embedding_eltwise_layernorm");
    teller_set.insert("multihead_matmul");
    teller_set.insert("skip_layernorm");
45
    teller_set.insert("slice");
C
ceci3 已提交
46
    int8_teller_set.insert("fused_embedding_eltwise_layernorm");
47 48 49
    int8_teller_set.insert("multihead_matmul");
    int8_teller_set.insert("skip_layernorm");
    int8_teller_set.insert("slice");
C
ceci3 已提交
50
#endif
51 52 53 54 55
// TODO(baoachun) The group_norm trt plugin will check input's dim
// not -1 failed when dynamic shape mode.
// #if IS_TRT_VERSION_GE(7130)
//     teller_set.insert("group_norm");
// #endif
W
wenbin 已提交
56 57 58
#if IS_TRT_VERSION_GE(7000)
    teller_set.insert("tile");
#endif
W
wenbin 已提交
59
#if CUDA_VERSION >= 10020
W
Wangzheee 已提交
60 61
    teller_set.insert("reshape");
    teller_set.insert("reshape2");
62 63
    int8_teller_set.insert("reshape");
    int8_teller_set.insert("reshape2");
64 65
#endif
  }
66

67 68 69 70 71 72 73
  bool operator()(const std::string& op_type, const framework::OpDesc& desc,
                  bool use_no_calib_int8) override {
    if (use_no_calib_int8) {
      return int8_teller_set.count(op_type);
    } else {
      return teller_set.count(op_type);
    }
74 75 76
  }

 private:
77
  // use this set for no calib int8.
78
  std::unordered_set<std::string> int8_teller_set{"mul",
C
ceci3 已提交
79
                                                  "matmul",
80
                                                  "conv2d",
81
                                                  "conv2d_fusion",
82 83 84
                                                  "pool2d",
                                                  "relu",
                                                  "softmax",
85
                                                  "sigmoid",
86 87
                                                  "hard_swish",
                                                  "depthwise_conv2d",
88
                                                  "batch_norm",
89 90 91
                                                  "concat",
                                                  "tanh",
                                                  "pad",
92
                                                  "elementwise_add",
93 94 95 96 97
                                                  "elementwise_mul",
                                                  "dropout",
                                                  "prelu",
                                                  "conv2d_transpose",
                                                  "depthwise_conv2d_transpose",
98 99
                                                  "leaky_relu",
                                                  "fc",
100 101 102 103 104 105
                                                  "shuffle_channel",
                                                  "swish",
                                                  "split",
                                                  "instance_norm",
                                                  "gelu",
                                                  "layer_norm",
106
                                                  "scale",
107 108
                                                  "stack",
                                                  "transpose2",
109
                                                  "transpose",
110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126
                                                  "flatten2",
                                                  "flatten",
                                                  "gather",
                                                  "gather_nd",
                                                  "yolo_box",
                                                  "roi_align",
                                                  "affine_channel",
                                                  "nearest_interp",
                                                  "anchor_generator",
                                                  "reduce_sum",
                                                  "reduce_mean",
                                                  "conv3d",
                                                  "conv3d_transpose",
                                                  "mish",
                                                  "nearest_interp_v2",
                                                  "pool3d",
                                                  "deformable_conv"};
W
wenbin 已提交
127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170
  std::unordered_set<std::string> teller_set{"mul",
                                             "matmul",
                                             "conv2d",
                                             "conv2d_fusion",
                                             "pool2d",
                                             "relu",
                                             "softmax",
                                             "sigmoid",
                                             "hard_swish",
                                             "depthwise_conv2d",
                                             "batch_norm",
                                             "concat",
                                             "tanh",
                                             "pad",
                                             "elementwise_add",
                                             "elementwise_mul",
                                             "dropout",
                                             "prelu",
                                             "conv2d_transpose",
                                             "depthwise_conv2d_transpose",
                                             "leaky_relu",
                                             "fc",
                                             "shuffle_channel",
                                             "swish",
                                             "split",
                                             "instance_norm",
                                             "gelu",
                                             "layer_norm",
                                             "scale",
                                             "stack",
                                             "transpose2",
                                             "transpose",
                                             "flatten2",
                                             "flatten",
                                             "gather",
                                             "gather_nd",
                                             "yolo_box",
                                             "roi_align",
                                             "affine_channel",
                                             "nearest_interp",
                                             "anchor_generator",
                                             "reduce_sum",
                                             "reduce_mean",
                                             "conv3d",
W
wangxinxin08 已提交
171
                                             "conv3d_transpose",
172
                                             "mish",
F
feng_shuai 已提交
173
                                             "nearest_interp_v2",
W
wangxinxin08 已提交
174 175
                                             "pool3d",
                                             "deformable_conv"};
176 177
};

178 179 180 181
bool OpTeller::Tell(const framework::ir::Node* node, bool use_no_calib_int8,
                    bool with_dynamic_shape) {
  const std::string op_type = node->Op()->Type();
  const framework::OpDesc desc = *node->Op();
182
  // do not support the op which is labeled the `skip_quant`
183
  if ((desc.HasAttr("namescope") &&
184
       BOOST_GET_CONST(std::string, desc.GetAttr("op_namescope")) ==
185 186
           "/skip_quant_2/") ||
      desc.HasAttr("skip_quant"))
187
    return false;
188

189
  for (auto& teller : tellers_) {
J
JingZhuangzhuang 已提交
190 191 192
    if (op_type == "relu" || op_type == "relu6" || op_type == "tanh" ||
        op_type == "sigmoid") {
      auto* block = desc.Block();
193 194 195 196 197 198
      if (block == nullptr) {
        VLOG(3) << "The block desc is nullptr, we can't continue to analyze. "
                   "Developers need to check whether block_desc is passed in "
                   "the pass.";
        return false;
      }
J
JingZhuangzhuang 已提交
199 200 201 202 203 204 205 206 207 208
      auto x_var_name = desc.Input("X")[0];
      auto* x_var_desc = block->FindVar(x_var_name);
      const auto x_shape = x_var_desc->GetShape();
      if (x_shape.size() == 1) {
        VLOG(3) << op_type
                << " op does not support input's dim is 1 in tensorrt.";
        return false;
      }
    }

209 210 211
    if (op_type == "pool2d") {
      std::vector<int> paddings =
          BOOST_GET_CONST(std::vector<int>, desc.GetAttr("paddings"));
212 213
      if (paddings.size() > 2) {
        return false;
214
      }
215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234
      if (desc.Input("X").size() != 1) {
        VLOG(3) << "TRT Pool2d expect 1 input, but got "
                << desc.Input("X").size();
        return false;
      }
      if (desc.Output("Out").size() != 1) {
        VLOG(3) << "TRT Pool2d has only 1 output, but got "
                << desc.Output("Out").size();
        return false;
      }
      if (!desc.HasAttr("pooling_type")) {
        return false;
      } else {
        std::string pool_type =
            BOOST_GET_CONST(std::string, desc.GetAttr("pooling_type"));
        if (pool_type != "max" && pool_type != "avg") {
          VLOG(3) << "Wrong pool op type, the trt do not support the "
                  << pool_type << " pool type.";
          return false;
        }
235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254
        if (pool_type == "avg") {
          if (desc.HasAttr("global_pooling")) {
            if (!BOOST_GET_CONST(bool, desc.GetAttr("global_pooling"))) {
              if (desc.HasAttr("exclusive")) {
                if (BOOST_GET_CONST(bool, desc.GetAttr("exclusive"))) {
                  std::vector<int> ksize =
                      BOOST_GET_CONST(std::vector<int>, desc.GetAttr("ksize"));
                  for (size_t i = 0; i < ksize.size(); i++) {
                    if (ksize[i] <= paddings[i]) {
                      VLOG(3) << "the padding size should be less than the "
                                 "filter size "
                                 "for exclusive-counting pooling.";
                      return false;
                    }
                  }
                }
              }
            }
          }
        }
255 256 257 258
      }
    }

    if (op_type == "conv2d" || op_type == "conv2d_transpose" ||
259 260
        op_type == "conv2d_fusion" || op_type == "depthwise_conv2d" ||
        op_type == "depthwise_conv2d_transpose") {
261 262 263 264 265 266 267 268 269 270 271 272
      if (desc.Input("Input").size() != 1) {
        VLOG(3) << "TRT Conv2d expect 1 input, but got "
                << desc.Input("Input").size() << " input.";
        return false;
      }

      if (desc.Input("Filter").size() != 1) {
        VLOG(3) << "TRT Conv2d expect 1 filter, but got "
                << desc.Input("Filter").size() << " filter.";
        return false;
      }

273 274 275
      if (desc.HasAttr("padding_algorithm")) {
        auto padding_algorithm =
            BOOST_GET_CONST(std::string, desc.GetAttr("padding_algorithm"));
276
        if (padding_algorithm == "VALID") {
277 278
          return false;
        }
279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300
        if (padding_algorithm == "SAME") {
          if (desc.HasAttr("dilations")) {
            const std::vector<int> dilations =
                BOOST_GET_CONST(std::vector<int>, desc.GetAttr("dilations"));
            if (dilations[0] != 1 || dilations[1] != 1) {
              VLOG(3) << "In Same mode, Dilations must be (1, 1) for "
                         "tensorRT, but given ("
                      << dilations[0] << ", " << dilations[1] << ")";
              return false;
            }
          }
        }
      }

      if (use_no_calib_int8) {
        if (desc.HasAttr("padding_algorithm")) {
          auto padding_algorithm =
              BOOST_GET_CONST(std::string, desc.GetAttr("padding_algorithm"));
          if (padding_algorithm == "SAME") {
            return false;
          }
        }
301 302
      }

303 304 305 306 307 308 309 310 311 312 313
      if (desc.HasAttr("enable_int8")) {
        if (op_type == "conv2d" || op_type == "conv2d_fusion") {
          if (!desc.HasAttr("Input_scale")) {
            VLOG(3) << "Input scale not found. TRT int8"
                       " requires conv/deconv to have "
                       "input quantization scales.";
            return false;
          }
        }
      }

314 315
      if (op_type == "conv2d_transpose" ||
          op_type == "depthwise_conv2d_transpose") {
316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334
        if (!desc.HasAttr("dilations")) {
          return false;
        } else {
          const std::vector<int> dilations =
              BOOST_GET_CONST(std::vector<int>, desc.GetAttr("dilations"));
          if (dilations[0] != 1 || dilations[1] != 1) {
            VLOG(3) << "In conv2d_transpose, Dilations must be (1, 1) for "
                       "tensorRT, but given ("
                    << dilations[0] << ", " << dilations[1] << ")";
            return false;
          }
        }
      }

      if (desc.Output("Output").size() != 1) {
        VLOG(3) << "TRT Conv2d expect 1 output, but got "
                << desc.Output("Output").size() << " output.";
        return false;
      }
335

W
wenbin 已提交
336
// strides > 1 and 'SAME' is only supported by trt7.0 above
337
#if !IS_TRT_VERSION_GE(7000)
W
wenbin 已提交
338 339 340 341 342 343 344 345 346 347 348 349 350 351
      if (op_type == "conv2d" || op_type == "conv2d_fusion" ||
          op_type == "depthwise_conv2d") {
        if (desc.HasAttr("padding_algorithm") && with_dynamic_shape) {
          auto padding_algorithm =
              BOOST_GET_CONST(std::string, desc.GetAttr("padding_algorithm"));
          if (padding_algorithm == "SAME" && desc.HasAttr("strides")) {
            const std::vector<int> strides =
                BOOST_GET_CONST(std::vector<int>, desc.GetAttr("strides"));
            // there is no issue if strides.size() less than 2
            if (strides.size() > 1) {
              for (size_t i = 0; i < strides.size(); i++) {
                if (strides[i] > 1) return false;
              }
            }
352 353 354 355
          }
        }
      }
#endif
356 357
    }

W
wangxinxin08 已提交
358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402
    if (op_type == "deformable_conv") {
      if (with_dynamic_shape) {
        VLOG(3) << "Deformable conv trt plugin does not support dynamic shape";
        return false;
      }
      auto* block = desc.Block();
      auto input_name = desc.Input("Input")[0];
      auto* input_desc = block->FindVar(input_name);
      const auto input_shape = input_desc->GetShape();

      if (input_shape.size() != 4) {
        VLOG(3) << "Input of deformable conv should be 4-D Tensor, but got "
                << input_shape.size();
        return false;
      }

      auto filter_name = desc.Input("Filter")[0];
      auto* filter_desc = block->FindVar(filter_name);
      const auto filter_shape = filter_desc->GetShape();

      int groups = BOOST_GET_CONST(int, desc.GetAttr("groups"));
      if (input_shape[1] != filter_shape[1] * groups) {
        VLOG(3) << "The number of input channels should be equal to filter "
                << "channels * groups. But got input channels "
                << input_shape[1] << "filter channels " << filter_shape[1];
        return false;
      }

      const std::vector<int> strides =
          BOOST_GET_CONST(std::vector<int>, desc.GetAttr("strides"));
      if (strides.size() != 2) {
        VLOG(3) << "The size of strides should be 2, but got "
                << strides.size();
        return false;
      }

      const std::vector<int> paddings =
          BOOST_GET_CONST(std::vector<int>, desc.GetAttr("paddings"));
      if (paddings.size() != 2) {
        VLOG(3) << "The size of paddings shoule be 2, but got "
                << paddings.size();
        return false;
      }
    }

403 404
    if (op_type == "matmul") {
      auto* block = desc.Block();
405 406 407 408 409 410
      if (block == nullptr) {
        VLOG(3) << "The block desc is nullptr, we can't continue to analyze. "
                   "Developers need to check whether block_desc is passed in "
                   "the pass.";
        return false;
      }
411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430

      // not support broadcast
      auto* x_var_desc = block->FindVar(desc.Input("X")[0]);
      auto* y_var_desc = block->FindVar(desc.Input("Y")[0]);
      const auto x_shape = x_var_desc->GetShape();
      const auto y_shape = y_var_desc->GetShape();
      if (x_shape.size() != y_shape.size()) {
        VLOG(3)
            << "matmul op not support broadcast, please check inputs'shape. ";
        return false;
      }
      uint64_t dims = 2;
      for (size_t i = 0; i < x_shape.size() - dims; ++i) {
        if (x_shape[i] != y_shape[i] && (x_shape[i] == 1 || y_shape[i] == 1)) {
          VLOG(3) << "matmul op not support broadcast, please check "
                     "inputs'shape[i]. ";
          return false;
        }
      }

431 432 433 434 435
      for (auto& param_name : desc.Inputs()) {
        for (auto& var_name : param_name.second) {
          auto* var_desc = block->FindVar(var_name);
          const auto shape = var_desc->GetShape();
          if (shape.size() < 3) {
436
            VLOG(3)
P
Pei Yang 已提交
437 438
                << "matmul op dims < 3 not supported in tensorrt, but got dims "
                << shape.size() << ", so jump it.";
439 440 441 442 443
            return false;
          }
        }
      }
    }
W
Wilber 已提交
444 445 446 447 448 449 450 451 452 453 454 455
    if (op_type == "softmax") {
      auto* block = desc.Block();
      if (block == nullptr) {
        VLOG(3) << "The block desc is nullptr, we can't continue to analyze. "
                   "Developers need to check whether block_desc is passed in "
                   "the pass.";
        return false;
      }
      auto x_var_name = desc.Input("X")[0];
      auto* x_var_desc = block->FindVar(x_var_name);
      const auto x_shape = x_var_desc->GetShape();
    }
456
    if (op_type == "group_norm") {
457
      if (!with_dynamic_shape) return false;
458 459 460 461 462 463 464 465 466
      bool has_attrs = (desc.HasAttr("epsilon") && desc.HasAttr("groups"));
      if (has_attrs == false) return false;

      auto registry = GetPluginRegistry();
      if (registry == nullptr) return false;
    }
    if (op_type == "concat") {
      if (!desc.HasAttr("axis")) {
        return false;
W
Wilber 已提交
467 468 469 470
      }
      int axis = BOOST_GET_CONST(int, desc.GetAttr("axis"));
      if (with_dynamic_shape) {
        if (axis < 0) return false;
471
      } else {
W
Wilber 已提交
472 473 474 475 476 477
        if (axis <= 0) return false;
      }
      auto concat_inputs = desc.Inputs();
      if (concat_inputs.find("AxisTensor") != concat_inputs.end()) {
        if (desc.Input("AxisTensor").size() >= 1) {
          return false;
478
        }
479 480
      }
    }
481 482 483
    if (op_type == "transpose2" || op_type == "transpose") {
      if (!desc.HasAttr("axis")) {
        return false;
484 485 486 487 488 489 490 491
      }
      std::vector<int> axis =
          BOOST_GET_CONST(std::vector<int>, desc.GetAttr("axis"));
      if (!with_dynamic_shape && axis[0] != 0) return false;
      if (axis.size() >= nvinfer1::Dims::MAX_DIMS) return false;
      if (axis[0] == 0 && axis.size() == 2) return false;

      auto* block = desc.Block();
492 493 494 495 496 497
      if (block == nullptr) {
        VLOG(3) << "The block desc is nullptr, we can't continue to analyze. "
                   "Developers need to check whether block_desc is passed in "
                   "the pass.";
        return false;
      }
498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519
      auto x_var_name = desc.Input("X")[0];
      auto* x_var_desc = block->FindVar(x_var_name);
      const auto x_shape = x_var_desc->GetShape();
      int dims = x_shape.size();
      std::vector<int> perm(nvinfer1::Dims::MAX_DIMS);
      for (int i = 0; i < dims; i++) {
        perm[i] = axis[i];
      }
      auto is_valid_permutation = [&](int dims,
                                      const std::vector<int>& permutation) {
        std::bitset<nvinfer1::Dims::MAX_DIMS> found;
        for (int i = 0; i < dims; ++i) {
          const int x = permutation[i];
          if ((x < 0) || (x >= dims) || found[x])
            return false;  // Out of bounds or duplicate
          found.set(x);
        }
        return true;
      };
      if (!is_valid_permutation(dims, perm)) {
        VLOG(3) << "Invalid permutation dimensions for trt transpose op "
                   "converter: duplicate or out of bound.";
520 521
      }
    }
522
    if (op_type == "flatten2" || op_type == "flatten") {
523 524 525
      if (!desc.HasAttr("axis")) {
        return false;
      } else {
526 527
#if IS_TRT_VERSION_GE(7130)
#else
528
        if (with_dynamic_shape) return false;
529
#endif
530 531 532 533
        int axis = BOOST_GET_CONST(int, desc.GetAttr("axis"));
        if (axis != 1) return false;
      }
    }
534

535
    if (op_type == "gather") {
536 537 538 539 540 541 542 543 544
      auto gather_inputs = desc.Inputs();
      if (gather_inputs.find("Axis") != gather_inputs.end()) {
        if (desc.Input("Axis").size() >= 1) {
          return false;
        }
      }
      if (!with_dynamic_shape) {
        return false;
      } else {
545
        auto* block = desc.Block();
546 547 548 549 550 551
        if (block == nullptr) {
          VLOG(3) << "The block desc is nullptr, we can't continue to analyze. "
                     "Developers need to check whether block_desc is passed in "
                     "the pass.";
          return false;
        }
552 553 554 555 556 557 558
        auto* x_var_desc = block->FindVar(desc.Input("X")[0]);
        const auto x_shape = x_var_desc->GetShape();
        if (x_shape.size() == 1) {
          VLOG(3) << "Gather does not support 1-dimensional input in tensorrt";
          return false;
        }
      }
559
    }
Z
zlsh80826 已提交
560

561
    if (op_type == "gather_nd") {
562 563
      if (!with_dynamic_shape) return false;

564
      auto* block = desc.Block();
565 566 567 568 569 570
      if (block == nullptr) {
        VLOG(3) << "The block desc is nullptr, we can't continue to analyze. "
                   "Developers need to check whether block_desc is passed in "
                   "the pass.";
        return false;
      }
571 572 573 574 575 576 577 578 579 580 581 582 583 584
      auto x_var_name = desc.Input("X")[0];
      auto index_var_name = desc.Input("Index")[0];
      auto* x_var_desc = block->FindVar(x_var_name);
      auto* index_var_desc = block->FindVar(index_var_name);

      // The index input must be int32 datatype.
      if (index_var_desc->GetDataType() !=
          paddle::framework::proto::VarType_Type::VarType_Type_INT32) {
        VLOG(3) << "gather_nd op Index input data type must be int32";
        return false;
      }

      const auto index_shape = index_var_desc->GetShape();
      const auto x_shape = x_var_desc->GetShape();
585 586 587 588 589 590
      if (x_shape.size() <= 2) {
        VLOG(3) << "gather_nd op requires the input's dimension to be greater "
                   "than 2";
        return false;
      }

591 592 593 594 595 596 597
      if (x_shape.size() != index_shape.size()) {
        VLOG(3) << "gather_nd op Index input dims size [" << index_shape.size()
                << " ] not equal to x dims size [" << x_shape.size() << "]";
        return false;
      }
    }

598 599 600 601
    if (op_type == "anchor_generator") {
      if (!with_dynamic_shape) return false;
    }

Z
zlsh80826 已提交
602 603 604 605 606 607
    if (op_type == "yolo_box") {
      if (with_dynamic_shape) return false;
      bool has_attrs =
          (desc.HasAttr("class_num") && desc.HasAttr("anchors") &&
           desc.HasAttr("downsample_ratio") && desc.HasAttr("conf_thresh") &&
           desc.HasAttr("clip_bbox") && desc.HasAttr("scale_x_y"));
Z
zlsh80826 已提交
608
      if (!has_attrs) return false;
Z
zlsh80826 已提交
609 610
    }

611 612 613 614 615
    if (op_type == "affine_channel") {
      if (!desc.HasAttr("data_layout")) return false;
      auto data_layout = framework::StringToDataLayout(
          BOOST_GET_CONST(std::string, desc.GetAttr("data_layout")));
      if (data_layout != framework::DataLayout::kNCHW) return false;
616 617

      auto* block = desc.Block();
618 619 620 621 622 623
      if (block == nullptr) {
        VLOG(3) << "The block desc is nullptr, we can't continue to analyze. "
                   "Developers need to check whether block_desc is passed in "
                   "the pass.";
        return false;
      }
624 625 626 627 628 629
      auto x_var_name = desc.Input("X")[0];
      auto* x_var_desc = block->FindVar(x_var_name);
      const auto x_shape = x_var_desc->GetShape();
      if (x_shape.size() == 2) {
        return false;
      }
630 631
    }

Z
zlsh80826 已提交
632 633 634
    if (op_type == "multiclass_nms") {
      if (with_dynamic_shape) return false;
      auto* block = desc.Block();
635 636 637 638 639 640
      if (block == nullptr) {
        VLOG(3) << "The block desc is nullptr, we can't continue to analyze. "
                   "Developers need to check whether block_desc is passed in "
                   "the pass.";
        return false;
      }
Z
zlsh80826 已提交
641 642 643 644 645
      for (auto& param_name : desc.Inputs()) {
        for (auto& var_name : param_name.second) {
          auto* var_desc = block->FindVar(var_name);
          const auto shape = var_desc->GetShape();
          if (shape.size() != 3) {
646
            VLOG(3) << "multiclass_nms op dims != 3 not supported in tensorrt, "
Z
zlsh80826 已提交
647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668
                       "but got dims "
                    << shape.size() << ", so jump it.";
            return false;
          }
        }
      }
      bool has_attrs =
          (desc.HasAttr("background_label") &&
           desc.HasAttr("score_threshold") && desc.HasAttr("nms_top_k") &&
           desc.HasAttr("keep_top_k") && desc.HasAttr("normalized"));
      if (has_attrs == false) return false;

      auto nms_top_k = BOOST_GET_CONST(int, desc.GetAttr("nms_top_k"));
      if (nms_top_k < 0) return false;

      auto keep_top_k = BOOST_GET_CONST(int, desc.GetAttr("keep_top_k"));
      if (keep_top_k < 0) return false;

      auto registry = GetPluginRegistry();
      if (registry == nullptr) return false;
    }

669
    if (op_type == "nearest_interp") {
670 671
      std::vector<std::string> attrs{"interp_method", "align_corners", "scale",
                                     "out_h", "out_w"};
672 673 674
      for (auto const attr : attrs) {
        if (!desc.HasAttr(attr)) return false;
      }
675 676 677 678 679 680 681
      if (desc.HasAttr("data_layout")) {
        auto data_layout = framework::StringToDataLayout(
            BOOST_GET_CONST(std::string, desc.GetAttr("data_layout")));
        if (data_layout != framework::DataLayout::kNCHW &&
            data_layout != framework::DataLayout::kNHWC)
          return false;
      }
682 683 684
      auto interp_method =
          BOOST_GET_CONST(std::string, desc.GetAttr("interp_method"));
      if (interp_method != "nearest") return false;
685 686 687 688 689 690 691 692
      auto scale = BOOST_GET_CONST(float, desc.GetAttr("scale"));
      auto out_h = BOOST_GET_CONST(int, desc.GetAttr("out_h"));
      auto out_w = BOOST_GET_CONST(int, desc.GetAttr("out_w"));
      auto align_corners = BOOST_GET_CONST(bool, desc.GetAttr("align_corners"));
      if (!(scale > 0.f && (out_h <= 0 && out_w <= 0))) {
        if (out_h <= 0) {
          VLOG(3) << "out_h must be greater than 0 if scale is not set.";
          return false;
693
        }
694 695
        if (out_w <= 0) {
          VLOG(3) << "out_w must be greater than 0 if scale is not set.";
已提交
696 697
          return false;
        }
698
      }
699 700 701 702 703 704 705 706 707
      if ((scale <= 0.f) && with_dynamic_shape) {
        VLOG(3) << "dynamic shape not support scale not set.";
        return false;
      }
      // When align_corners = true, the paddle's and trt_layer's results has
      // diff
      if (align_corners && scale != 1) {
        return false;
      }
708
    }
709

710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728
    if (op_type == "nearest_interp_v2") {
      std::vector<std::string> attrs{"data_layout",   "interp_method",
                                     "align_corners", "scale",
                                     "out_h",         "out_w"};
      for (auto const attr : attrs) {
        if (!desc.HasAttr(attr)) return false;
      }
      auto data_layout = framework::StringToDataLayout(
          BOOST_GET_CONST(std::string, desc.GetAttr("data_layout")));
      if (data_layout != framework::DataLayout::kNCHW &&
          data_layout != framework::DataLayout::kNHWC)
        return false;
      auto interp_method =
          BOOST_GET_CONST(std::string, desc.GetAttr("interp_method"));
      if (interp_method != "nearest") return false;
      auto scale = BOOST_GET_CONST(std::vector<float>, desc.GetAttr("scale"));
      auto out_h = BOOST_GET_CONST(int, desc.GetAttr("out_h"));
      auto out_w = BOOST_GET_CONST(int, desc.GetAttr("out_w"));
      if (!(out_h > 0 && out_w > 0)) {
W
wenbin 已提交
729
        if (scale.size() < 2) return false;
730 731 732 733 734 735 736 737
        if (scale[0] <= 0.f || scale[1] <= 0.f) {
          VLOG(3) << "scale factor must be greater than 0 if out_h or out_w is "
                     "not set.";
          return false;
        }
      }
    }

738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759
    if (op_type == "roi_align") {
      if (!with_dynamic_shape) return false;

      std::vector<std::string> attrs{"pooled_height", "pooled_width",
                                     "spatial_scale", "sampling_ratio"};
      for (auto const attr : attrs) {
        if (!desc.HasAttr(attr)) return false;
      }

      const auto pooled_height =
          BOOST_GET_CONST(int, desc.GetAttr("pooled_height"));
      if (pooled_height <= 0) return false;

      const auto pooled_width =
          BOOST_GET_CONST(int, desc.GetAttr("pooled_width"));
      if (pooled_width <= 0) return false;

      const auto spatial_scale =
          BOOST_GET_CONST(float, desc.GetAttr("spatial_scale"));
      if (spatial_scale <= 0.f) return false;
    }

760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785
    if (op_type == "hard_swish") {
      if (desc.Input("X").size() != 1) {
        VLOG(3) << "HardSwish op has only 1 input, but got "
                << desc.Input("X").size();
        return false;
      }

      if (desc.Output("Out").size() != 1) {
        VLOG(3) << "HardSwish op has only 1 output, but got "
                << desc.Output("Out").size();
        return false;
      }
    }

    if (op_type == "batch_norm") {
      const std::vector<std::string> bn_inputs = {"X", "Bias", "Mean", "Scale",
                                                  "Variance"};
      for (unsigned int i = 0; i < bn_inputs.size(); i++) {
        if (desc.Input(bn_inputs[i]).size() != 1) {
          VLOG(3) << "Invalid " << bn_inputs[i]
                  << "'s size of batch_norm TRT "
                     "converter. Expected 1, received "
                  << desc.Input(bn_inputs[i]).size() << ".";
          return false;
        }
      }
786 787 788 789 790 791
      auto batch_norm_inputs = desc.Inputs();
      if (batch_norm_inputs.find("MomentumTensor") != batch_norm_inputs.end()) {
        if (desc.Input("MomentumTensor").size() >= 1) {
          return false;
        }
      }
792 793 794 795 796 797
      if (desc.Output("Y").size() != 1) {
        VLOG(3) << "Invalid output Y's size of batch_norm TRT "
                   "converter. Expected 1, received "
                << desc.Output("Y").size() << ".";
        return false;
      }
W
Wilber 已提交
798 799 800 801 802 803 804 805 806 807
      auto* block = desc.Block();
      if (block == nullptr) {
        VLOG(3) << "The block desc is nullptr, we can't continue to analyze. "
                   "Developers need to check whether block_desc is passed in "
                   "the pass.";
        return false;
      }
      auto x_var_name = desc.Input("X")[0];
      auto* x_var_desc = block->FindVar(x_var_name);
      const auto x_shape = x_var_desc->GetShape();
808 809 810 811 812 813 814 815 816
    }

    if (op_type == "split") {
      if (desc.Input("X").size() != 1) {
        VLOG(3) << "Invalid input X's size of split TRT converter. "
                   "Expected 1, received "
                << desc.Input("X").size() << ".";
        return false;
      }
817 818 819 820 821 822 823 824 825 826 827
      auto split_inputs = desc.Inputs();
      if (split_inputs.find("AxisTensor") != split_inputs.end()) {
        if (desc.Input("AxisTensor").size() >= 1) {
          return false;
        }
      }
      if (split_inputs.find("SectionsTensorList") != split_inputs.end()) {
        if (desc.Input("SectionsTensorList").size() >= 1) {
          return false;
        }
      }
828 829
      if (!desc.HasAttr("axis")) {
        return false;
830 831 832 833 834 835 836 837 838
      }
      int axis = BOOST_GET_CONST(int, desc.GetAttr("axis"));

      if (axis == 0) {
        VLOG(3) << "Invalid split axis. Split on batch is not supported in "
                   "TensorRT";
        return false;
      }
      auto* block = desc.Block();
839 840 841 842 843 844
      if (block == nullptr) {
        VLOG(3) << "The block desc is nullptr, we can't continue to analyze. "
                   "Developers need to check whether block_desc is passed in "
                   "the pass.";
        return false;
      }
845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888
      auto x_var_name = desc.Input("X")[0];
      auto* x_var_desc = block->FindVar(x_var_name);
      const auto x_shape = x_var_desc->GetShape();
      size_t output_num = desc.Output("Out").size();
      std::vector<int> output_lengths;
      int num = 0;
      if (desc.HasAttr("num")) {
        num = BOOST_GET_CONST(int, desc.GetAttr("num"));
      }
      if (desc.HasAttr("sections")) {
        output_lengths =
            BOOST_GET_CONST(std::vector<int>, desc.GetAttr("sections"));
      }
      if (output_lengths.size() == 0 && num == 0) {
        VLOG(3) << "sections and num cannot be equal to 0 at the same time";
        return false;
      }
      if (with_dynamic_shape) {
#if IS_TRT_VERSION_GE(6000)
#else
        VLOG(3) << "You are running the TRT Dynamic Shape mode, need to "
                   "confirm that "
                   "your TRT version is no less than 6.0";
        return false;
#endif
      }
      axis += (axis < 0) ? x_shape.size() : 0;
      if (x_shape[axis] == -1) {
        VLOG(3) << "The (" << axis << ") dim of input should not be -1";
        return false;
      }
      if (output_lengths.size() == 0) {
        if (num > 0) {
          int64_t in_axis_dim = x_shape[axis];
          if (in_axis_dim % num != 0) {
            VLOG(3) << "Invalid number to split. Tensor split does not result"
                       " in an equal division of dimensions. Axis dim = "
                    << in_axis_dim << " num = " << num << "!= 0";
            return false;
          }
          size_t out_axis_dim = in_axis_dim / num;
          for (int i = 0; i < num; ++i) {
            output_lengths.push_back(out_axis_dim);
          }
889 890
        }
      }
891 892 893 894
      if (output_lengths.size() != output_num) {
        VLOG(3) << "The output_length should be equal to the output size.";
        return false;
      }
895
    }
896

897 898 899 900 901 902 903 904
    if (op_type == "scale") {
      auto scale_inputs = desc.Inputs();
      if (scale_inputs.find("ScaleTensor") != scale_inputs.end()) {
        if (desc.Input("ScaleTensor").size() >= 1) {
          return false;
        }
      }
      auto* block = desc.Block();
905 906 907 908 909 910
      if (block == nullptr) {
        VLOG(3) << "The block desc is nullptr, we can't continue to analyze. "
                   "Developers need to check whether block_desc is passed in "
                   "the pass.";
        return false;
      }
911 912 913
      auto x_var_name = desc.Input("X")[0];
      auto* x_var_desc = block->FindVar(x_var_name);
      const auto x_shape = x_var_desc->GetShape();
914 915 916 917
      if (!with_dynamic_shape && x_shape.size() == 1) {
        VLOG(3) << "Scale op does not support 1-dimensional input in tensorrt";
        return false;
      }
918
    }
919

920
    if (op_type == "slice") {
921 922 923 924 925 926 927 928 929 930
      if (desc.HasAttr("decrease_axis")) {
        std::vector<int> decrease_axis =
            BOOST_GET_CONST(std::vector<int>, desc.GetAttr("decrease_axis"));
        if (decrease_axis.size() > 0) {
          VLOG(3) << "Invalid slice decrease_axis. decrease_axis.size() > 0"
                     "is not supported in TensorRT";
          return false;
        }
      }

931
      if (!desc.HasAttr("axes") || !desc.HasAttr("starts") ||
932 933 934
          !desc.HasAttr("ends")) {
        VLOG(3) << "The necessary attributes of the slice operator axes "
                   "or starts or ends are missing.";
935 936 937 938 939 940 941 942
        return false;
      } else {
        std::vector<int> axes =
            BOOST_GET_CONST(std::vector<int>, desc.GetAttr("axes"));
        std::vector<int> starts =
            BOOST_GET_CONST(std::vector<int>, desc.GetAttr("starts"));
        std::vector<int> ends =
            BOOST_GET_CONST(std::vector<int>, desc.GetAttr("ends"));
943

944
        if (axes.size() != starts.size() || axes.size() != ends.size()) {
945 946
          VLOG(3) << "The shape of attributes of the slice operator axes "
                     "or starts or ends are not equal.";
已提交
947 948
          return false;
        }
949 950 951 952 953 954 955 956
        if (!with_dynamic_shape) {
          for (size_t i = 0; i < axes.size(); i++) {
            if (axes[i] == 0) {
              VLOG(3) << "Invalid slice axis. Slice on batch axis is not "
                         "supported in TensorRT";
              return false;
            }
          }
S
Shang Zhizhou 已提交
957 958 959 960 961 962 963 964 965
        } else {
          for (size_t i = 0; i < axes.size(); i++) {
            if (starts[i] < 0 || ends[i] < 0) {
              VLOG(3) << "Invalid slice attribute 'starts' or 'ends'. "
                         "Negative starts or ends not supported in TensorRT "
                         "when running in dynamic shape mode.";
              return false;
            }
          }
966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988
        }
      }
    }

    if (op_type == "elementwise_add" || op_type == "elementwise_mul") {
      if (desc.Input("X").size() != 1) {
        VLOG(3) << "The input op's Input(\"X\").size() "
                   "should equal to 1, but received Input(\"X\").size() = "
                << desc.Input("X").size() << ".";
        return false;
      }
      if (desc.Input("Y").size() != 1) {
        VLOG(3) << "The input op's Input(\"Y\").size() "
                   "should equal to 1, but received Input(\"Y\").size() = "
                << desc.Input("Y").size() << ".";
        return false;
      }
      if (desc.Output("Out").size() != 1) {
        VLOG(3) << "The input op's Output(\"Out\").size() "
                   "should equal to 1, but reveceid Output(\"Out\").size() = "
                << desc.Output("Out").size() << ".";
        return false;
      }
989
      auto* block = desc.Block();
990 991 992 993 994 995
      if (block == nullptr) {
        VLOG(3) << "The block desc is nullptr, we can't continue to analyze. "
                   "Developers need to check whether block_desc is passed in "
                   "the pass.";
        return false;
      }
996 997 998 999 1000 1001 1002 1003
      auto* x_var_desc = block->FindVar(desc.Input("X")[0]);
      auto* y_var_desc = block->FindVar(desc.Input("Y")[0]);
      const auto x_shape = x_var_desc->GetShape();
      const auto y_shape = y_var_desc->GetShape();
      if (x_shape.size() == 1 && y_shape.size() == 1) {
        VLOG(3) << "Now trt may not support two 1d tensor elementwise op.";
        return false;
      }
1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040
    }

    if (op_type == "stack") {
      if (!with_dynamic_shape) {
        VLOG(3)
            << "static shape mode is not supported for TRT stack.\n"
               "You can use the config.SetTRTDynamicShapeInfo(...) interface"
               " to set the shape information to run the dynamic shape "
               "mode.";
        return false;
      }
    }

    if (op_type == "fused_embedding_eltwise_layernorm") {
      if (!with_dynamic_shape) {
        VLOG(3) << "fused_embedding_eltwise_layernorm should run on dynamic "
                   "shape mode.";
        return false;
      }
      if (desc.Input("Ids").size() != desc.Input("Embs").size()) {
        VLOG(3) << "The id and emb size of fused EmbEltwiseLayerNormOp "
                   "should be same ";
        return false;
      }
    }

    if (op_type == "gelu") {
      if (desc.Input("X").size() != 1) {
        VLOG(3) << "gelu op has only 1 input, but got "
                << desc.Input("X").size();
        return false;
      }
      if (desc.Output("Out").size() != 1) {
        VLOG(3) << "gelu op has only 1 output, but got "
                << desc.Output("Out").size();
        return false;
      }
1041 1042 1043 1044 1045 1046

      if (desc.HasAttr("approximate")) {
        if (BOOST_GET_CONST(bool, desc.GetAttr("approximate"))) return false;
      }

      auto* block = desc.Block();
1047 1048 1049 1050 1051 1052
      if (block == nullptr) {
        VLOG(3) << "The block desc is nullptr, we can't continue to analyze. "
                   "Developers need to check whether block_desc is passed in "
                   "the pass.";
        return false;
      }
1053 1054 1055 1056 1057 1058 1059
      auto x_var_name = desc.Input("X")[0];
      auto* x_var_desc = block->FindVar(x_var_name);
      const auto x_shape = x_var_desc->GetShape();
      if (x_shape.size() == 1) {
        VLOG(3) << "gelu op does not support input's dim is 1 in tensorrt.";
        return false;
      }
1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084
    }

    if (op_type == "layer_norm") {
      if (desc.Input("X").size() != 1) {
        VLOG(3) << "input of layer_norm op converter should be 1, got "
                << desc.Input("X").size();
        return false;
      }
      if (desc.Input("Bias").size() != 1) {
        VLOG(3) << "Bias of layer_norm op converter should be 1, got "
                << desc.Input("Bias").size();
        return false;
      }
      if (desc.Input("Scale").size() != 1) {
        VLOG(3) << "Scale of layer_norm op converter should be 1, got "
                << desc.Input("Scale").size();
        return false;
      }
      if (desc.Output("Y").size() != 1) {
        VLOG(3) << "output of layer_norm op converter should be 1, got "
                << desc.Output("Y").size();
        return false;
      }
    }

已提交
1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109
    if (op_type == "instance_norm") {
      if (with_dynamic_shape) {
        VLOG(3) << "trt instance_norm op does not support dynamic shape ";
        return false;
      }
      if (desc.Input("X").size() != 1) {
        VLOG(3) << "input of instance_norm op converter should be 1, got "
                << desc.Input("X").size();
        return false;
      }
      if (desc.Input("Bias").size() != 1) {
        VLOG(3) << "Bias of instance_norm op converter should be 1, got "
                << desc.Input("Bias").size();
        return false;
      }
      if (desc.Input("Scale").size() != 1) {
        VLOG(3) << "Scale of instance_norm op converter should be 1, got "
                << desc.Input("Scale").size();
        return false;
      }
      if (desc.Output("Y").size() != 1) {
        VLOG(3) << "output of layer_norm op converter should be 1, got "
                << desc.Output("Y").size();
        return false;
      }
1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125

      auto* block = desc.Block();
      if (block == nullptr) {
        VLOG(3) << "The block desc is nullptr, we can't continue to analyze. "
                   "Developers need to check whether block_desc is passed in "
                   "the pass.";
        return false;
      }
      auto x_var_name = desc.Input("X")[0];
      auto* x_var_desc = block->FindVar(x_var_name);
      const auto x_shape = x_var_desc->GetShape();
      if (x_shape.size() != 4) {
        VLOG(3) << "The instance_norm op only support 4-dimensional input in "
                   "tensorrt.";
        return false;
      }
已提交
1126 1127
    }

1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147
    if (op_type == "leaky_relu") {
      if (desc.Input("X").size() != 1) {
        VLOG(3) << "Invalid number of TRT leaky_relu op converter "
                   "inputs. Expected 1, but received "
                << desc.Input("X").size();
        return false;
      }
      if (desc.Output("Out").size() != 1) {
        VLOG(3) << "output of leaky_relu op converter should be 1, got "
                << desc.Output("Out").size();
        return false;
      }
    }

    if (op_type == "pad") {
      const float pad_value = BOOST_GET_CONST(float, desc.GetAttr("pad_value"));
      if (pad_value != 0.0f) {
        VLOG(3) << "The pad layer of TRT only support zero.";
        return false;
      }
已提交
1148 1149
      std::vector<int64_t> shape;
      auto* block = desc.Block();
1150 1151 1152 1153 1154 1155
      if (block == nullptr) {
        VLOG(3) << "The block desc is nullptr, we can't continue to analyze. "
                   "Developers need to check whether block_desc is passed in "
                   "the pass.";
        return false;
      }
已提交
1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176
      for (auto& param_name : desc.Inputs()) {
        for (auto& var_name : param_name.second) {
          auto* var_desc = block->FindVar(var_name);
          shape = var_desc->GetShape();
        }
      }
      int nbDims = shape.size();
      std::vector<int> paddings =
          BOOST_GET_CONST(std::vector<int>, desc.GetAttr("paddings"));
      int pad_size = paddings.size();
      if (nbDims < 2) {
        return false;
      }
      if (nbDims * 2 != pad_size) {
        return false;
      }
      for (int i = 0; i < pad_size - 4; i++) {
        if (paddings[i] != 0) {
          return false;
        }
      }
1177 1178
    }

1179 1180
    if (op_type == "swish") {
      auto* block = desc.Block();
1181 1182 1183 1184 1185 1186
      if (block == nullptr) {
        VLOG(3) << "The block desc is nullptr, we can't continue to analyze. "
                   "Developers need to check whether block_desc is passed in "
                   "the pass.";
        return false;
      }
1187 1188 1189 1190 1191 1192 1193 1194 1195
      auto x_var_name = desc.Input("X")[0];
      auto* x_var_desc = block->FindVar(x_var_name);
      const auto x_shape = x_var_desc->GetShape();
      if (x_shape.size() == 1) {
        VLOG(3) << "swish op does not support input's dim is 1 in tensorrt.";
        return false;
      }
    }

1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208
    if (op_type == "prelu") {
      if (desc.Input("X").size() != 1) {
        VLOG(3) << "Invalid input X's size of prelu TRT converter. "
                   "Expected 1, received "
                << desc.Input("X").size() << ".";
        return false;
      }
      if (desc.Output("Out").size() != 1) {
        VLOG(3) << "Invalid output Out's size of prelu TRT converter. "
                   "Expected 1, received "
                << desc.Output("Out").size() << ".";
        return false;
      }
1209 1210

      auto* block = desc.Block();
1211 1212 1213 1214 1215 1216
      if (block == nullptr) {
        VLOG(3) << "The block desc is nullptr, we can't continue to analyze. "
                   "Developers need to check whether block_desc is passed in "
                   "the pass.";
        return false;
      }
1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230
      auto* var_desc = block->FindVar(desc.Input("Alpha")[0]);
      if (!var_desc) {
        VLOG(3) << "Variable Alpha of prelu TRT converter not found.";
        return false;
      }

      auto x_var_name = desc.Input("X")[0];
      auto* x_var_desc = block->FindVar(x_var_name);
      const auto x_shape = x_var_desc->GetShape();
      if (x_shape.size() == 1) {
        VLOG(3) << "prelu op does not support input's dim is 1 in tensorrt.";
        return false;
      }

W
Wilber 已提交
1231 1232 1233 1234 1235 1236 1237
#if IS_TRT_VERSION_LT(7000)
      if (!with_dynamic_shape) {
        // TODO(inference): fix trt6 static plugin error.
        VLOG(3) << "prelu static plugin in trt6 has bug.";
        return false;
      }
#endif
1238 1239
    }

W
wangxinxin08 已提交
1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277
    if (op_type == "mish") {
      if (desc.Input("X").size() != 1) {
        VLOG(3) << "Invalid input X's size of mish TRT converter. "
                   "Expected 1, received "
                << desc.Input("X").size() << ".";
        return false;
      }
      if (desc.Output("Out").size() != 1) {
        VLOG(3) << "Invalid output Out's size of mish TRT converter. "
                   "Expected 1, received "
                << desc.Output("Out").size() << ".";
        return false;
      }

      auto* block = desc.Block();
      if (block == nullptr) {
        VLOG(3) << "The block desc is nullptr, we can't continue to analyze. "
                   "Developers need to check whether block_desc is passed in "
                   "the pass.";
        return false;
      }

      auto x_var_name = desc.Input("X")[0];
      auto* x_var_desc = block->FindVar(x_var_name);
      const auto x_shape = x_var_desc->GetShape();
      if (x_shape.size() == 1) {
        VLOG(3) << "mish op does not support input's dim is 1 in tensorrt.";
        return false;
      }

      if (!with_dynamic_shape) {
        if (x_shape.size() == 2) {
          VLOG(3) << "mish op does not support input's dim is 2 in tensorrt.";
          return false;
        }
      }
    }

1278 1279 1280 1281 1282 1283 1284
    if (op_type == "roi_align") {
      if (!with_dynamic_shape) {
        VLOG(3) << "TRT roi align plugin only accept the dynamic shape, "
                   "because that "
                   "the roi_align will change the batch size.";
        return false;
      }
1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312 1313 1314
      std::vector<std::string> attrs{"pooled_height", "pooled_width",
                                     "spatial_scale", "sampling_ratio"};
      for (auto const attr : attrs) {
        if (!desc.HasAttr(attr)) return false;
      }

      const auto pooled_height =
          BOOST_GET_CONST(int, desc.GetAttr("pooled_height"));
      if (pooled_height <= 0) return false;

      const auto pooled_width =
          BOOST_GET_CONST(int, desc.GetAttr("pooled_width"));
      if (pooled_width <= 0) return false;

      const auto spatial_scale =
          BOOST_GET_CONST(float, desc.GetAttr("spatial_scale"));
      if (spatial_scale <= 0.f) return false;

      const auto sampling_ratio =
          BOOST_GET_CONST(int, desc.GetAttr("sampling_ratio"));
      const auto aligned = BOOST_GET_CONST(bool, desc.GetAttr("aligned"));

      if (sampling_ratio == -1 && aligned == true) return false;

      auto roi_align_inputs = desc.Inputs();
      if (roi_align_inputs.find("RoisNum") != roi_align_inputs.end()) {
        if (desc.Input("RoisNum").size() >= 1) {
          return false;
        }
      }
1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336
    }

    if (op_type == "shuffle_channel") {
      if (with_dynamic_shape) {
        VLOG(3) << "You are running the TRT Dynamic Shape mode, "
                   "the shuffle_channel op does not support dynamic shape yet";
        return false;
      }
    }

    if (op_type == "skip_layernorm") {
      if (!with_dynamic_shape) {
        VLOG(3) << "the skip_layernorm does not support static shape yet";
        return false;
      }
    }

    if (op_type == "multihead_matmul") {
      if (!with_dynamic_shape) {
        VLOG(3) << "the multihead_matmul does not support static shape yet";
        return false;
      }
1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351 1352 1353 1354 1355 1356 1357 1358 1359 1360 1361 1362 1363 1364 1365 1366 1367 1368 1369 1370 1371 1372

      if (desc.HasAttr("enable_int8") && !desc.HasAttr("Input_scale")) {
        VLOG(3) << "Multihead layers must have input scale in int8 mode.";
        return false;
      }

      auto* block = desc.Block();
      if (block == nullptr) {
        VLOG(3) << "The block desc is nullptr, we can't continue to analyze. "
                   "Developers need to check whether block_desc is passed in "
                   "the pass.";
        return false;
      }
      auto* input_desc = block->FindVar(desc.Input("Input").front());
      const auto input_shape = input_desc->GetShape();
      const auto head_number =
          BOOST_GET_CONST(int, desc.GetAttr("head_number"));

      auto* biasqk_desc = block->FindVar(desc.Input("BiasQK").front());
      const auto biasqk_shape = biasqk_desc->GetShape();
      // The BiasQK's shape requires to be
      // [batch, 1, 1, length] or [batch, head, length, length].
      bool has_same_shape = head_number == biasqk_shape[1] &&
                            input_shape[1] == biasqk_shape[2] &&
                            input_shape[1] == biasqk_shape[3];
      bool is_broadcastable = biasqk_shape[1] == 1 && biasqk_shape[2] == 1 &&
                              input_shape[1] == biasqk_shape[3];
      if (!(has_same_shape || is_broadcastable)) {
        VLOG(3) << "The BiasQK's shape is invalid, expect [" << input_shape[0]
                << ", 1, 1, " << input_shape[1] << "] or [" << input_shape[0]
                << ", " << head_number << ", " << input_shape[1] << ", "
                << input_shape[1] << "] but [" << biasqk_shape[0] << ", "
                << biasqk_shape[1] << ", " << biasqk_shape[2] << ", "
                << biasqk_shape[3] << "].";
        return false;
      }
1373 1374
    }

1375
    if (op_type == "fc") {
1376 1377 1378 1379 1380 1381 1382 1383 1384 1385 1386 1387 1388 1389 1390 1391 1392 1393 1394 1395 1396 1397 1398 1399 1400 1401 1402 1403 1404 1405 1406 1407 1408 1409 1410 1411 1412 1413 1414 1415 1416
      auto* block = desc.Block();
      if (block == nullptr) {
        VLOG(3) << "The block desc is nullptr, we can't continue to analyze. "
                   "Developers need to check whether block_desc is passed in "
                   "the pass.";
        return false;
      }

      // y'shapes == 2
      auto fc_inputs = desc.Inputs();
      std::string fc_y = "";
      if (fc_inputs.find("Y") != fc_inputs.end()) {
        fc_y = "Y";
      } else if (fc_inputs.find("W") != fc_inputs.end()) {
        fc_y = "W";
      } else {
        VLOG(3) << " input_y(fc_op) must be Y or W ";
        return false;
      }

      //  There is currently no input: Y(weight) more than two dimensions
      /*
      auto* y_var_desc = block->FindVar(desc.Input(fc_y)[0]);
      const auto y_shape = y_var_desc->GetShape();
      if (y_shape.size() != 2) {
        VLOG(3)
            << " input_y(fc_op)'shapes must be 2, but input_y(fc_op)'shapes = "
            << y_shape.size();
        return false;
      }
      // y_num_col_dims ==1
      if (desc.HasAttr("y_num_col_dims")) {
        int y_num_col_dims =
            BOOST_GET_CONST(int, desc.GetAttr("y_num_col_dims"));
        if (y_num_col_dims != 1) {
          VLOG(3) << " fc_op'y_num_col_dims must be 1, but y_num_col_dims = "
                  << y_num_col_dims;
          return false;
        }
      }
      */
1417 1418 1419 1420 1421 1422 1423
      int x_num_col_dims =
          desc.HasAttr("x_num_col_dims")
              ? BOOST_GET_CONST(int, desc.GetAttr("x_num_col_dims"))
              : (desc.HasAttr("in_num_col_dims")
                     ? BOOST_GET_CONST(int, desc.GetAttr("in_num_col_dims"))
                     : 1);
      if (x_num_col_dims < 1) {
1424 1425 1426
        VLOG(3) << "fc_op expects x_num_col_dims >= 1, "
                   "but x_num_col_dims = "
                << x_num_col_dims;
1427 1428 1429
        return false;
      }
    }
1430

W
Wangzheee 已提交
1431 1432 1433
    if (op_type == "reshape" || op_type == "reshape2") {
      if (!desc.HasAttr("shape")) {
        return false;
W
Wilber 已提交
1434 1435
      }
      // Paddle-TRT does not support the input tensors: Shape and ShapeTensor
1436
      auto reshape_inputs = desc.Inputs();
1437 1438 1439 1440 1441 1442 1443 1444 1445
      if (reshape_inputs.find("Shape") != reshape_inputs.end()) {
        if (desc.Input("Shape").size() >= 1) {
          return false;
        }
      }
      if (reshape_inputs.find("ShapeTensor") != reshape_inputs.end()) {
        if (desc.Input("ShapeTensor").size() >= 1) {
          return false;
        }
W
Wangzheee 已提交
1446
      }
W
Wilber 已提交
1447 1448 1449
      std::vector<int> shape =
          BOOST_GET_CONST(std::vector<int>, desc.GetAttr("shape"));
      if (shape.size() >= nvinfer1::Dims::MAX_DIMS) return false;
1450 1451
      if (!with_dynamic_shape && (shape[0] == -1 || shape.size() == 1))
        return false;
W
Wangzheee 已提交
1452
    }
1453

1454 1455 1456 1457 1458 1459 1460 1461 1462 1463 1464 1465 1466 1467 1468
    if (op_type == "clip") {
      // Paddle-TRT does not support the input tensors: Min and Max
      auto clip_inputs = desc.Inputs();
      if (clip_inputs.find("Min") != clip_inputs.end()) {
        if (desc.Input("Min").size() >= 1) {
          return false;
        }
      }
      if (clip_inputs.find("Max") != clip_inputs.end()) {
        if (desc.Input("Max").size() >= 1) {
          return false;
        }
      }

      auto* block = desc.Block();
1469 1470 1471 1472 1473 1474
      if (block == nullptr) {
        VLOG(3) << "The block desc is nullptr, we can't continue to analyze. "
                   "Developers need to check whether block_desc is passed in "
                   "the pass.";
        return false;
      }
1475 1476 1477 1478 1479 1480 1481 1482 1483
      auto x_var_name = desc.Input("X")[0];
      auto* x_var_desc = block->FindVar(x_var_name);
      const auto x_shape = x_var_desc->GetShape();
      if (x_shape.size() == 1) {
        VLOG(3) << "clip op does not support input's dim is 1 in tensorrt.";
        return false;
      }
    }

W
wenbin 已提交
1484
    if (op_type == "reduce_sum" || op_type == "reduce_mean") {
1485 1486
      if (!(desc.HasAttr("keep_dim") && desc.HasAttr("dim") &&
            desc.HasAttr("reduce_all"))) {
W
wenbin 已提交
1487 1488
        VLOG(3) << "the " << op_type
                << " does not have attr (keep_dim or dim or "
1489
                   "reduce_all)";
W
wenbin 已提交
1490 1491
        std::cout << "attr " << desc.HasAttr("keep_dim") << " "
                  << desc.HasAttr("dim") << " " << desc.HasAttr("reduce_all");
1492 1493
        return false;
      }
W
wenbin 已提交
1494 1495 1496

      // The batch size dimension cannot be reduced if it's not dynamic shape.
      if (!with_dynamic_shape) {
W
wenbin 已提交
1497
        if (BOOST_GET_CONST(bool, desc.GetAttr("reduce_all"))) return false;
W
wenbin 已提交
1498 1499 1500 1501 1502
        std::vector<int32_t> dim =
            BOOST_GET_CONST(std::vector<int32_t>, desc.GetAttr("dim"));
        for (auto x : dim) {
          if (!x) return false;
        }
1503 1504 1505 1506 1507
      } else {
        if (BOOST_GET_CONST(bool, desc.GetAttr("reduce_all")) &&
            !BOOST_GET_CONST(bool, desc.GetAttr("keep_dim")))
          return false;
      }
1508
      if (desc.HasAttr("out_dtype")) {
1509 1510 1511 1512
        int out_dtype = BOOST_GET_CONST(int32_t, desc.GetAttr("out_dtype"));
        if (out_dtype != -1) {
          return false;
        }
W
wenbin 已提交
1513
      }
1514
    }
W
wenbin 已提交
1515 1516 1517
#if IS_TRT_VERSION_GE(7000)
    if (op_type == "tile") {
      // Paddle-TRT does not support the input tensors.
1518 1519 1520
      auto tile_inputs = desc.Inputs();
      if (tile_inputs.find("repeat_times_tensor") != tile_inputs.end()) {
        if (desc.Input("repeat_times_tensor").size() >= 1) {
W
wenbin 已提交
1521
          return false;
1522 1523 1524 1525
        }
      }
      if (tile_inputs.find("RepeatTimes") != tile_inputs.end()) {
        if (desc.Input("RepeatTimes").size() >= 1) {
W
wenbin 已提交
1526
          return false;
1527
        }
W
wenbin 已提交
1528 1529 1530 1531 1532
      }
      if (with_dynamic_shape) return false;
      if (!with_dynamic_shape && !desc.HasAttr("repeat_times")) return false;
    }
#endif
1533

W
wenbin 已提交
1534 1535 1536 1537 1538 1539 1540 1541 1542 1543 1544 1545 1546 1547 1548 1549 1550 1551 1552 1553 1554 1555 1556 1557 1558 1559 1560 1561 1562 1563 1564 1565 1566 1567 1568 1569 1570 1571 1572 1573 1574 1575 1576 1577 1578 1579 1580 1581 1582 1583 1584 1585 1586 1587 1588 1589 1590 1591 1592
    if (op_type == "conv3d" || op_type == "conv3d_transpose") {
      if (desc.HasAttr("padding_algorithm")) {
        std::string padding_algorithm =
            BOOST_GET_CONST(std::string, desc.GetAttr("padding_algorithm"));

        // trt error is arised if conv3d_transpose and SAME
        if (op_type == "conv3d_transpose" && padding_algorithm == "SAME" &&
            !with_dynamic_shape) {
          return false;
        }
      }

#if !IS_TRT_VERSION_GE(7000)
      // looks like some issues with trt6.0
      if (with_dynamic_shape) {
        return false;
      }
#endif
      std::vector<int> paddings =
          BOOST_GET_CONST(std::vector<int>, desc.GetAttr("paddings"));

      // conv3d and conv3d_transpose need padding check
      if (paddings.size() > 3) return false;

      if (desc.Input("Input").size() != 1) {
        VLOG(3) << "TRT Conv3d expect 1 input, but got "
                << desc.Input("Input").size() << " input.";
        return false;
      }

      if (desc.Input("Filter").size() != 1) {
        VLOG(3) << "TRT Conv3d expect 1 filter, but got "
                << desc.Input("Filter").size() << " filter.";
        return false;
      }

      if (op_type == "conv3d_transpose") {
        if (!desc.HasAttr("dilations")) {
          return false;
        } else {
          const std::vector<int> dilations =
              BOOST_GET_CONST(std::vector<int>, desc.GetAttr("dilations"));
          if (dilations[0] != 1 || dilations[1] != 1 || dilations[2] != 1) {
            VLOG(3) << "In conv3d_transpose, Dilations must be (1, 1, 1) for "
                       "tensorRT, but given ("
                    << dilations[0] << ", " << dilations[1] << ", "
                    << dilations[2] << ")";
            return false;
          }
        }
      }

      if (desc.Output("Output").size() != 1) {
        VLOG(3) << "TRT Conv3d expect 1 output, but got "
                << desc.Output("Output").size() << " output.";
        return false;
      }
    }

1593 1594 1595 1596
    if (op_type == "hard_sigmoid") {
      if (!with_dynamic_shape) {
        auto* block = desc.Block();
        if (block == nullptr) {
1597 1598 1599
          VLOG(3) << "The block desc is nullptr, we can't continue to analyze. "
                     "Developers need to check whether block_desc is passed in "
                     "the pass.";
1600 1601 1602 1603 1604
          return false;
        }
        auto x_var_name = desc.Input("X")[0];
        auto* x_var_desc = block->FindVar(x_var_name);
        const auto x_shape = x_var_desc->GetShape();
1605 1606 1607
        if (x_shape.size() == 1) {
          VLOG(3) << "Hard sigmoid does not support 1-dimensional input in "
                     "tensorrt";
1608 1609 1610 1611 1612
          return false;
        }
      }
    }

1613
    if ((*teller)(op_type, desc, use_no_calib_int8)) return true;
1614
  }
W
wenbin 已提交
1615 1616

  VLOG(3) << "trt unsupported op " << op_type;
1617 1618 1619 1620 1621 1622 1623 1624
  return false;
}

OpTeller::OpTeller() { tellers_.emplace_back(new SimpleOpTypeSetTeller); }

}  // namespace tensorrt
}  // namespace inference
}  // namespace paddle