op_teller.cc 48.7 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 51 52
#endif
#if IS_TRT_VERSION_GE(7130)
    teller_set.insert("group_norm");
W
Wangzheee 已提交
53
#endif
W
wenbin 已提交
54 55 56
#if IS_TRT_VERSION_GE(7000)
    teller_set.insert("tile");
#endif
W
wenbin 已提交
57
#if CUDA_VERSION >= 10020
W
Wangzheee 已提交
58 59
    teller_set.insert("reshape");
    teller_set.insert("reshape2");
60 61
#endif
  }
62

63 64 65 66 67 68 69
  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);
    }
70 71 72
  }

 private:
73
  // use this set for no calib int8.
74 75
  std::unordered_set<std::string> int8_teller_set{"mul",
                                                  "conv2d",
C
ceci3 已提交
76 77
                                                  "matmul",
                                                  "stack",
78
                                                  "conv2d_fusion",
79 80 81 82
                                                  "pool2d",
                                                  "relu",
                                                  "depthwise_conv2d",
                                                  "softmax",
83
                                                  "sigmoid",
84 85 86 87
                                                  "batch_norm",
                                                  "elementwise_add",
                                                  "leaky_relu",
                                                  "fc",
88 89 90
                                                  "concat",
                                                  "scale",
                                                  "elementwise_mul",
91 92
                                                  "conv2d_transpose",
                                                  "hard_swish"};
W
wenbin 已提交
93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137
  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",
                                             "conv3d_transpose"};
138 139
};

140 141 142 143
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();
144
  // do not support the op which is labeled the `skip_quant`
145
  if ((desc.HasAttr("namescope") &&
146
       BOOST_GET_CONST(std::string, desc.GetAttr("op_namescope")) ==
147 148
           "/skip_quant_2/") ||
      desc.HasAttr("skip_quant"))
149
    return false;
150

151
  for (auto& teller : tellers_) {
J
JingZhuangzhuang 已提交
152 153 154
    if (op_type == "relu" || op_type == "relu6" || op_type == "tanh" ||
        op_type == "sigmoid") {
      auto* block = desc.Block();
155 156 157 158 159 160
      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
161 162 163 164 165 166 167 168 169 170
      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;
      }
    }

171 172 173 174
    if (op_type == "pool2d") {
      std::vector<int> paddings =
          BOOST_GET_CONST(std::vector<int>, desc.GetAttr("paddings"));
      if (paddings.size() > 2) return false;
175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190
      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;
            }
          }
        }
      }
      if (desc.HasAttr("ceil_mode")) {
        if (BOOST_GET_CONST(bool, desc.GetAttr("ceil_mode"))) return false;
      }
191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214
      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;
        }
      }
    }

    if (op_type == "conv2d" || op_type == "conv2d_transpose" ||
215 216
        op_type == "conv2d_fusion" || op_type == "depthwise_conv2d" ||
        op_type == "depthwise_conv2d_transpose") {
217 218 219 220 221 222 223 224 225 226 227 228 229 230 231
      std::vector<int> paddings =
          BOOST_GET_CONST(std::vector<int>, desc.GetAttr("paddings"));

      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;
      }

232 233 234 235 236 237 238 239
      if (desc.HasAttr("padding_algorithm")) {
        auto padding_algorithm =
            BOOST_GET_CONST(std::string, desc.GetAttr("padding_algorithm"));
        if (padding_algorithm == "SAME" || padding_algorithm == "VALID") {
          return false;
        }
      }

240 241 242 243 244 245 246 247 248 249 250
      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;
          }
        }
      }

251 252
      if (op_type == "conv2d_transpose" ||
          op_type == "depthwise_conv2d_transpose") {
253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271
        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;
      }
272

W
wenbin 已提交
273
// strides > 1 and 'SAME' is only supported by trt7.0 above
274
#if !IS_TRT_VERSION_GE(7000)
W
wenbin 已提交
275 276 277 278 279 280 281 282 283 284 285 286 287 288
      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;
              }
            }
289 290 291 292
          }
        }
      }
#endif
293 294
    }

295 296
    if (op_type == "matmul") {
      auto* block = desc.Block();
297 298 299 300 301 302
      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;
      }
303 304 305 306 307
      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) {
308
            VLOG(3)
P
Pei Yang 已提交
309 310
                << "matmul op dims < 3 not supported in tensorrt, but got dims "
                << shape.size() << ", so jump it.";
311 312 313 314 315
            return false;
          }
        }
      }
    }
316
    if (op_type == "group_norm") {
317
      if (!with_dynamic_shape) return false;
318 319 320 321 322 323 324 325 326 327 328
      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;
      } else {
        int axis = BOOST_GET_CONST(int, desc.GetAttr("axis"));
329 330 331 332 333
        if (with_dynamic_shape) {
          if (axis < 0) return false;
        } else {
          if (axis <= 0) return false;
        }
334 335 336 337 338 339
        auto concat_inputs = desc.Inputs();
        if (concat_inputs.find("AxisTensor") != concat_inputs.end()) {
          if (desc.Input("AxisTensor").size() >= 1) {
            return false;
          }
        }
340 341
      }
    }
342 343 344
    if (op_type == "transpose2" || op_type == "transpose") {
      if (!desc.HasAttr("axis")) {
        return false;
345 346 347 348 349 350 351 352
      }
      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();
353 354 355 356 357 358
      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;
      }
359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380
      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.";
381 382
      }
    }
383
    if (op_type == "flatten2" || op_type == "flatten") {
384 385 386
      if (!desc.HasAttr("axis")) {
        return false;
      } else {
387 388
#if IS_TRT_VERSION_GE(7130)
#else
389
        if (with_dynamic_shape) return false;
390
#endif
391 392 393 394
        int axis = BOOST_GET_CONST(int, desc.GetAttr("axis"));
        if (axis != 1) return false;
      }
    }
395

396
    if (op_type == "gather") {
397 398 399 400 401 402 403 404 405
      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 {
406
        auto* block = desc.Block();
407 408 409 410 411 412
        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;
        }
413 414 415 416 417 418 419
        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;
        }
      }
420
    }
Z
zlsh80826 已提交
421

422
    if (op_type == "gather_nd") {
423 424
      if (!with_dynamic_shape) return false;

425
      auto* block = desc.Block();
426 427 428 429 430 431
      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;
      }
432 433 434 435 436 437 438 439 440 441 442 443 444 445
      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();
446 447 448 449 450 451
      if (x_shape.size() <= 2) {
        VLOG(3) << "gather_nd op requires the input's dimension to be greater "
                   "than 2";
        return false;
      }

452 453 454 455 456 457 458
      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;
      }
    }

459 460 461 462
    if (op_type == "anchor_generator") {
      if (!with_dynamic_shape) return false;
    }

Z
zlsh80826 已提交
463 464 465 466 467 468
    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
469
      if (!has_attrs) return false;
Z
zlsh80826 已提交
470 471
    }

472 473 474 475 476
    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;
477 478

      auto* block = desc.Block();
479 480 481 482 483 484
      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;
      }
485 486 487 488 489 490
      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;
      }
491 492
    }

Z
zlsh80826 已提交
493 494 495
    if (op_type == "multiclass_nms") {
      if (with_dynamic_shape) return false;
      auto* block = desc.Block();
496 497 498 499 500 501
      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
502 503 504 505 506
      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) {
507
            VLOG(3) << "multiclass_nms op dims != 3 not supported in tensorrt, "
Z
zlsh80826 已提交
508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529
                       "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;
    }

530 531 532 533 534 535 536 537 538 539 540 541 542 543 544
    if (op_type == "nearest_interp") {
      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;
545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562

      if (!desc.HasAttr("scale") || !desc.HasAttr("out_h") ||
          !desc.HasAttr("out_w")) {
        return false;
      } else {
        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"));
        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;
          }
          if (out_w <= 0) {
            VLOG(3) << "out_w must be greater than 0 if scale is not set.";
            return false;
          }
        }
已提交
563 564 565 566
        if ((scale <= 0.f) && with_dynamic_shape) {
          VLOG(3) << "dynamic shape not support scale not set.";
          return false;
        }
567
      }
568
    }
569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591

    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;
    }

592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617
    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;
        }
      }
618 619 620 621 622 623
      auto batch_norm_inputs = desc.Inputs();
      if (batch_norm_inputs.find("MomentumTensor") != batch_norm_inputs.end()) {
        if (desc.Input("MomentumTensor").size() >= 1) {
          return false;
        }
      }
624 625 626 627 628 629 630 631 632 633 634 635 636 637 638
      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;
      }
    }

    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;
      }
639 640 641 642 643 644 645 646 647 648 649
      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;
        }
      }
650 651
      if (!desc.HasAttr("axis")) {
        return false;
652 653 654 655 656 657 658 659 660
      }
      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();
661 662 663 664 665 666
      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;
      }
667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710
      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);
          }
711 712
        }
      }
713 714 715 716
      if (output_lengths.size() != output_num) {
        VLOG(3) << "The output_length should be equal to the output size.";
        return false;
      }
717
    }
718 719 720 721 722 723 724 725
    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();
726 727 728 729 730 731
      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;
      }
732 733 734 735 736
      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 (!with_dynamic_shape && x_shape.size() == 1) return false;
    }
737 738
    if (op_type == "slice") {
      if (!desc.HasAttr("axes") || !desc.HasAttr("starts") ||
已提交
739
          !desc.HasAttr("ends") || !desc.HasAttr("decrease_axis")) {
740 741 742 743 744 745 746 747
        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"));
已提交
748 749
        std::vector<int> decrease_axis =
            BOOST_GET_CONST(std::vector<int>, desc.GetAttr("decrease_axis"));
750 751 752
        if (axes.size() != starts.size() || axes.size() != ends.size()) {
          return false;
        }
已提交
753 754 755 756 757
        if (decrease_axis.size() > 0) {
          VLOG(3) << "Invalid slice decrease_axis. decrease_axis.size() > 0"
                     "is not supported in TensorRT";
          return false;
        }
758 759 760 761 762 763 764 765
        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
766 767 768 769 770 771 772 773 774
        } 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;
            }
          }
775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797
        }
      }
    }

    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;
      }
798
      auto* block = desc.Block();
799 800 801 802 803 804
      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;
      }
805 806 807 808 809 810 811 812
      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;
      }
813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849
    }

    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;
      }
850 851 852 853 854 855

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

      auto* block = desc.Block();
856 857 858 859 860 861
      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;
      }
862 863 864 865 866 867 868
      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;
      }
869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893
    }

    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
894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920
    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;
      }
    }

921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940
    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
941 942
      std::vector<int64_t> shape;
      auto* block = desc.Block();
943 944 945 946 947 948
      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
949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969
      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;
        }
      }
970 971
    }

972 973
    if (op_type == "scale") {
      auto* block = desc.Block();
974 975 976 977 978 979
      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;
      }
980 981 982 983 984 985 986 987 988
      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) << "dropout op does not support input's dim is 1 in tensorrt.";
        return false;
      }
    }

989 990
    if (op_type == "swish") {
      auto* block = desc.Block();
991 992 993 994 995 996
      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;
      }
997 998 999 1000 1001 1002 1003 1004 1005
      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;
      }
    }

1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018
    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;
      }
1019 1020

      auto* block = desc.Block();
1021 1022 1023 1024 1025 1026
      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;
      }
1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046
      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;
      }

      if (!with_dynamic_shape) {
        if (x_shape.size() == 2) {
          VLOG(3) << "prelu op does not support input's dim is 2 in tensorrt.";
          return false;
        }
      }
1047 1048 1049 1050 1051 1052 1053 1054 1055
    }

    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;
      }
1056 1057 1058 1059 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 1085
      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;
        }
      }
1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107
    }

    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;
      }
1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143

      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;
      }
1144 1145
    }

1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158
    if (op_type == "fc") {
      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) {
        VLOG(3) << "converter expects x_num_col_dims >= 1, "
                   "but x_num_col_dims = %d.";
        return false;
      }
    }
1159

W
Wangzheee 已提交
1160 1161 1162
    if (op_type == "reshape" || op_type == "reshape2") {
      if (!desc.HasAttr("shape")) {
        return false;
W
Wilber 已提交
1163 1164
      }
      // Paddle-TRT does not support the input tensors: Shape and ShapeTensor
1165
      auto reshape_inputs = desc.Inputs();
1166 1167 1168 1169 1170 1171 1172 1173 1174
      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
1175
      }
W
Wilber 已提交
1176 1177 1178
      std::vector<int> shape =
          BOOST_GET_CONST(std::vector<int>, desc.GetAttr("shape"));
      if (shape.size() >= nvinfer1::Dims::MAX_DIMS) return false;
1179 1180
      if (!with_dynamic_shape && (shape[0] == -1 || shape.size() == 1))
        return false;
W
Wangzheee 已提交
1181
    }
1182

1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197
    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();
1198 1199 1200 1201 1202 1203
      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;
      }
1204 1205 1206 1207 1208 1209 1210 1211 1212
      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
1213
    if (op_type == "reduce_sum" || op_type == "reduce_mean") {
1214 1215
      if (!(desc.HasAttr("keep_dim") && desc.HasAttr("dim") &&
            desc.HasAttr("reduce_all"))) {
W
wenbin 已提交
1216 1217
        VLOG(3) << "the " << op_type
                << " does not have attr (keep_dim or dim or "
1218
                   "reduce_all)";
W
wenbin 已提交
1219 1220
        std::cout << "attr " << desc.HasAttr("keep_dim") << " "
                  << desc.HasAttr("dim") << " " << desc.HasAttr("reduce_all");
1221 1222
        return false;
      }
W
wenbin 已提交
1223 1224 1225

      // The batch size dimension cannot be reduced if it's not dynamic shape.
      if (!with_dynamic_shape) {
W
wenbin 已提交
1226
        if (BOOST_GET_CONST(bool, desc.GetAttr("reduce_all"))) return false;
W
wenbin 已提交
1227 1228 1229 1230 1231
        std::vector<int32_t> dim =
            BOOST_GET_CONST(std::vector<int32_t>, desc.GetAttr("dim"));
        for (auto x : dim) {
          if (!x) return false;
        }
1232 1233 1234 1235 1236 1237 1238 1239 1240 1241
      } else {
        if (BOOST_GET_CONST(bool, desc.GetAttr("reduce_all")) &&
            !BOOST_GET_CONST(bool, desc.GetAttr("keep_dim")))
          return false;
      }
      if (desc.HasAttr("reduce_all")) {
        int out_dtype = BOOST_GET_CONST(int32_t, desc.GetAttr("out_dtype"));
        if (out_dtype != -1) {
          return false;
        }
W
wenbin 已提交
1242
      }
1243
    }
W
wenbin 已提交
1244 1245 1246
#if IS_TRT_VERSION_GE(7000)
    if (op_type == "tile") {
      // Paddle-TRT does not support the input tensors.
1247 1248 1249
      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
1250
          return false;
1251 1252 1253 1254
        }
      }
      if (tile_inputs.find("RepeatTimes") != tile_inputs.end()) {
        if (desc.Input("RepeatTimes").size() >= 1) {
W
wenbin 已提交
1255
          return false;
1256
        }
W
wenbin 已提交
1257 1258 1259 1260 1261
      }
      if (with_dynamic_shape) return false;
      if (!with_dynamic_shape && !desc.HasAttr("repeat_times")) return false;
    }
#endif
1262

W
wenbin 已提交
1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 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 1315 1316 1317 1318 1319 1320 1321
    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;
      }
    }

1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339
    if (op_type == "hard_sigmoid") {
      if (!with_dynamic_shape) {
        auto* block = desc.Block();
        if (block == nullptr) {
          VLOG(3) << "The block is null.";
          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() <= 2) {
          VLOG(3) << "hard_sigmoid op does not support input's dim less than 3 "
                     "in tensorrt.";
          return false;
        }
      }
    }

1340
    if ((*teller)(op_type, desc, use_no_calib_int8)) return true;
1341
  }
W
wenbin 已提交
1342 1343

  VLOG(3) << "trt unsupported op " << op_type;
1344 1345 1346 1347 1348 1349 1350 1351
  return false;
}

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

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