op_teller.cc 94.8 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

17
#include <bitset>
18

19
#include "paddle/fluid/framework/block_desc.h"
20
#include "paddle/fluid/framework/data_layout.h"
W
weishengying 已提交
21 22 23 24 25
#include "paddle/fluid/framework/op_meta_info_helper.h"
#include "paddle/fluid/framework/phi_utils.h"
#include "paddle/fluid/inference/tensorrt/dynamic_shape_infermeta_factory.h"
#include "paddle/phi/core/compat/op_utils.h"
#include "paddle/phi/core/kernel_factory.h"
26

W
wanghuancoder 已提交
27 28 29 30 31 32
namespace paddle {
namespace framework {
class OpDesc;
}  // namespace framework
}  // namespace paddle

33 34 35 36 37 38
namespace paddle {
namespace inference {
namespace tensorrt {

// Just tell by the op_types.
struct SimpleOpTypeSetTeller : public Teller {
39
  SimpleOpTypeSetTeller() {
40
#if IS_TRT_VERSION_GE(7130)
Z
Zhang Jun 已提交
41
    // use TensorRT plugin
42
    teller_set.insert("group_norm");
Z
Zhang Jun 已提交
43 44
    teller_set.insert("multiclass_nms3");
    teller_set.insert("multiclass_nms");
45 46
    int8_teller_set.insert("multiclass_nms3");
    int8_teller_set.insert("multiclass_nms");
47
#endif
W
wenbin 已提交
48 49
#if IS_TRT_VERSION_GE(7000)
    teller_set.insert("tile");
50
    teller_set.insert("flatten_contiguous_range");
51
    int8_teller_set.insert("flatten_contiguous_range");
Z
zhoutianzi666 已提交
52 53 54 55
    teller_set.insert("rnn");
    int8_teller_set.insert("rnn");
    teller_set.insert("fill_constant_batch_size_like");
    int8_teller_set.insert("fill_constant_batch_size_like");
W
wenbin 已提交
56
#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
    int8_teller_set.insert("reshape");
    int8_teller_set.insert("reshape2");
62 63 64 65 66 67
#endif
#if IS_TRT_VERSION_GE(8000)
    teller_set.insert("sparse_fc");
    int8_teller_set.insert("sparse_fc");
    teller_set.insert("sparse_multihead_matmul");
    int8_teller_set.insert("sparse_multihead_matmul");
68 69 70 71
#endif
#if IS_TRT_VERSION_GE(8200)
    teller_set.insert("round");
    int8_teller_set.insert("round");
72 73
#endif
  }
74

W
weishengying 已提交
75 76 77 78 79 80 81 82 83 84
  bool operator()(const framework::OpDesc& desc,
                  bool use_no_calib_int8 = false,
                  bool with_dynamic_shape = false) override {
    const std::string op_type = desc.Type();
    // do not support the op which is labeled the `skip_quant`
    if ((desc.HasAttr("namescope") &&
         PADDLE_GET_CONST(std::string, desc.GetAttr("op_namescope")) ==
             "/skip_quant_2/") ||
        desc.HasAttr("skip_quant"))
      return false;
85
    std::unordered_set<std::string> act_op_list = {
86 87 88 89 90 91 92 93 94 95 96 97
        "relu",       "relu6",       "sigmoid",
        "elu",        "selu",        "softsign",
        "softplus",   "stanh",       "thresholded_relu",
        "exp",        "log",         "sqrt",
        "abs",        "sin",         "cos",
        "tan",        "tanh",        "sinh",
        "cosh",       "asin",        "acos",
        "atan",       "asinh",       "acosh",
        "atanh",      "ceil",        "celu",
        "erf",        "floor",       "round",
        "sign",       "silu",        "logical_not",
        "reciprocal", "tanh_shrink", "logsigmoid"};
98
    if (act_op_list.find(op_type) != act_op_list.end()) {
J
JingZhuangzhuang 已提交
99
      auto* block = desc.Block();
100 101 102 103 104 105
      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
106 107 108 109 110 111 112 113
      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;
      }
114 115 116 117 118 119
#if !IS_TRT_VERSION_GE(7000)
      if (op_type == "erf") {
        VLOG(3) << op_type << " op does not support tensorrt.";
        return false;
      }
#endif
J
JingZhuangzhuang 已提交
120 121
    }

122 123 124 125 126 127
    // In static shape mode in TRT, we can't allow that op's input is a
    // 1D-tensor So we filter it here. Some op like elementwise having "Y" too,
    // but that is dealt with in the specified op, here just the common case
    if (!with_dynamic_shape) {
      std::string X_name;
      auto inputs = desc.Inputs();
128
      if (inputs.count("X") && !desc.Input("X").empty()) {
129
        X_name = desc.Input("X")[0];
130
      } else if (inputs.count("Input") && !desc.Input("Input").empty()) {
131 132 133 134 135 136 137 138 139 140 141 142 143
        X_name = desc.Input("Input")[0];
      }
      auto* block = desc.Block();
      if (block) {
        auto* x_var_desc = block->FindVar(X_name);
        // Can't get feed op's TensorDesc
        if (op_type != "feed" && x_var_desc && !x_var_desc->Persistable()) {
          const auto x_shape = x_var_desc->GetShape();
          if (x_shape.size() == 1) return false;
        }
      }
    }

144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159
    if (op_type == "dropout") {
      /*
       * Some OpDescs Attribute support both constant value and dynamic
       * runtime value (which is a Variable(s) type). But TensorRT maybe
       * only support constant value Attribute, so we shall distinguish
       * this case in time and return False in OpTeller.Tell().
       * If Attribute is Variable(s), HasAttr() will return False
       */
      if (!desc.HasAttr("dropout_prob", /*with_attr_var=*/false)) {
        VLOG(3)
            << "Skip to convert into TRT while found Attribute('dropout_prob') "
               "is Variable type in dropout.";
        return false;
      }
    }

160
    if (op_type == "pool2d") {
161 162 163 164 165 166 167
      // If Attribute is Variable(s), HasAttr() will return False
      if (!desc.HasAttr("ksize", /*with_attr_var=*/false)) {
        VLOG(3) << "Skip to convert into TRT while found Attribute('ksize') is "
                   "Variable type in pool2d.";
        return false;
      }

168
      std::vector<int> paddings =
R
Ruibiao Chen 已提交
169
          PADDLE_GET_CONST(std::vector<int>, desc.GetAttr("paddings"));
170 171
      if (paddings.size() > 2) {
        return false;
172
      }
173 174 175 176 177 178 179 180 181 182
      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;
      }
W
wenbin 已提交
183 184
      if (desc.HasAttr("data_format")) {
        std::string data_format =
R
Ruibiao Chen 已提交
185
            PADDLE_GET_CONST(std::string, desc.GetAttr("data_format"));
W
wenbin 已提交
186 187 188 189
        if (data_format == "NHWC" || data_format == "NDHWC") {
          return false;
        }
      }
190 191 192 193
      if (!desc.HasAttr("pooling_type")) {
        return false;
      } else {
        std::string pool_type =
R
Ruibiao Chen 已提交
194
            PADDLE_GET_CONST(std::string, desc.GetAttr("pooling_type"));
195 196 197 198 199
        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;
        }
200 201
        if (pool_type == "avg") {
          if (desc.HasAttr("global_pooling")) {
R
Ruibiao Chen 已提交
202
            if (!PADDLE_GET_CONST(bool, desc.GetAttr("global_pooling"))) {
203
              if (desc.HasAttr("exclusive")) {
R
Ruibiao Chen 已提交
204
                if (PADDLE_GET_CONST(bool, desc.GetAttr("exclusive"))) {
205
                  std::vector<int> ksize =
R
Ruibiao Chen 已提交
206
                      PADDLE_GET_CONST(std::vector<int>, desc.GetAttr("ksize"));
207 208 209 210 211 212 213 214 215 216 217 218 219
                  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;
                    }
                  }
                }
              }
            }
          }
        }
220 221 222 223
      }
    }

    if (op_type == "conv2d" || op_type == "conv2d_transpose" ||
224 225
        op_type == "conv2d_fusion" || op_type == "depthwise_conv2d" ||
        op_type == "depthwise_conv2d_transpose") {
226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248
      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;
      }

      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;
          }
        }
      }

249 250
      if (op_type == "conv2d_transpose" ||
          op_type == "depthwise_conv2d_transpose") {
251 252 253 254
        if (!desc.HasAttr("dilations")) {
          return false;
        } else {
          const std::vector<int> dilations =
R
Ruibiao Chen 已提交
255
              PADDLE_GET_CONST(std::vector<int>, desc.GetAttr("dilations"));
256 257 258 259 260 261 262 263 264 265 266 267 268 269
          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;
      }
270

W
wenbin 已提交
271
// strides > 1 and 'SAME' is only supported by trt7.0 above
272
#if !IS_TRT_VERSION_GE(7000)
W
wenbin 已提交
273 274 275 276
      if (op_type == "conv2d" || op_type == "conv2d_fusion" ||
          op_type == "depthwise_conv2d") {
        if (desc.HasAttr("padding_algorithm") && with_dynamic_shape) {
          auto padding_algorithm =
R
Ruibiao Chen 已提交
277
              PADDLE_GET_CONST(std::string, desc.GetAttr("padding_algorithm"));
W
wenbin 已提交
278 279
          if (padding_algorithm == "SAME" && desc.HasAttr("strides")) {
            const std::vector<int> strides =
R
Ruibiao Chen 已提交
280
                PADDLE_GET_CONST(std::vector<int>, desc.GetAttr("strides"));
W
wenbin 已提交
281 282 283 284 285 286
            // 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;
              }
            }
287 288 289 290
          }
        }
      }
#endif
291 292 293 294 295 296 297 298 299
      auto* block = desc.Block();
      if (block) {
        auto* filter_var_desc = block->FindVar(desc.Input("Filter")[0]);
        if (!filter_var_desc->Persistable()) {
          VLOG(3) << "Trt not support filter is  a intermediate tensor in "
                     "conv2d op.";
          return false;
        }
      }
300 301
    }

W
wangxinxin08 已提交
302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321
    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();

R
Ruibiao Chen 已提交
322
      int groups = PADDLE_GET_CONST(int, desc.GetAttr("groups"));
W
wangxinxin08 已提交
323 324 325 326 327 328 329 330
      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 =
R
Ruibiao Chen 已提交
331
          PADDLE_GET_CONST(std::vector<int>, desc.GetAttr("strides"));
W
wangxinxin08 已提交
332 333 334 335 336 337 338
      if (strides.size() != 2) {
        VLOG(3) << "The size of strides should be 2, but got "
                << strides.size();
        return false;
      }

      const std::vector<int> paddings =
R
Ruibiao Chen 已提交
339
          PADDLE_GET_CONST(std::vector<int>, desc.GetAttr("paddings"));
W
wangxinxin08 已提交
340 341 342 343 344 345 346
      if (paddings.size() != 2) {
        VLOG(3) << "The size of paddings shoule be 2, but got "
                << paddings.size();
        return false;
      }
    }

347 348 349 350 351 352
    if (op_type == "bmm") {
      if (!with_dynamic_shape) {
        return false;
      }
    }

353 354 355 356 357 358
    if (op_type == "range") {
      if (!with_dynamic_shape) {
        return false;
      }
    }

359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381
    if (op_type == "sign") {
#if IS_TRT_VERSION_GE(8200)
      if (!with_dynamic_shape) {
        return false;
      }
#else
      VLOG(3) << "sign op is only supported by trt8.2 above ";
      return false;
#endif
    }

    if (op_type == "logical_not") {
#if IS_TRT_VERSION_GE(8400)
      if (!with_dynamic_shape) {
        return false;
      }
#else
      VLOG(3) << "logical_not op is only supported by trt8.4 above because of "
                 "cast op";
      return false;
#endif
    }

382 383 384 385 386 387 388 389 390 391 392 393 394 395
    if (op_type == "matmul_v2") {
      if (!with_dynamic_shape) {
        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;
      }
      return true;
    }

396 397
    if (op_type == "matmul") {
      auto* block = desc.Block();
398 399 400 401 402 403
      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;
      }
404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423

      // 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;
        }
      }

424 425 426 427 428
      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) {
429
            VLOG(3)
P
Pei Yang 已提交
430 431
                << "matmul op dims < 3 not supported in tensorrt, but got dims "
                << shape.size() << ", so jump it.";
432 433 434 435 436
            return false;
          }
        }
      }
    }
W
Wilber 已提交
437 438 439 440 441 442 443 444 445 446 447 448
    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();
    }
449 450 451 452 453
    if (op_type == "group_norm") {
      bool has_attrs = (desc.HasAttr("epsilon") && desc.HasAttr("groups"));
      if (has_attrs == false) return false;
      auto registry = GetPluginRegistry();
      if (registry == nullptr) return false;
454 455 456 457 458 459 460
      std::string layout_str =
          PADDLE_GET_CONST(std::string, desc.GetAttr("data_layout"));
      if (layout_str != "NCHW") {
        VLOG(3) << "Group norm trt plugin only support NCHW layout, but got "
                << layout_str;
        return false;
      }
461 462 463 464
    }
    if (op_type == "concat") {
      if (!desc.HasAttr("axis")) {
        return false;
W
Wilber 已提交
465
      }
R
Ruibiao Chen 已提交
466
      int axis = PADDLE_GET_CONST(int, desc.GetAttr("axis"));
467 468
      if (!with_dynamic_shape) {
        if (axis == 0) return false;
W
Wilber 已提交
469 470 471 472 473
      }
      auto concat_inputs = desc.Inputs();
      if (concat_inputs.find("AxisTensor") != concat_inputs.end()) {
        if (desc.Input("AxisTensor").size() >= 1) {
          return false;
474
        }
475 476
      }
    }
477 478 479
    if (op_type == "transpose2" || op_type == "transpose") {
      if (!desc.HasAttr("axis")) {
        return false;
480 481
      }
      std::vector<int> axis =
R
Ruibiao Chen 已提交
482
          PADDLE_GET_CONST(std::vector<int>, desc.GetAttr("axis"));
483 484 485 486
      if (!with_dynamic_shape && axis[0] != 0) return false;
      if (axis.size() >= nvinfer1::Dims::MAX_DIMS) return false;

      auto* block = desc.Block();
487 488 489 490 491 492
      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;
      }
493 494 495
      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();
W
wenbin 已提交
496
      if (axis.size() != x_shape.size()) return false;
497
      int dims = x_shape.size();
W
wenbin 已提交
498

499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516
      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.";
W
wenbin 已提交
517
        return false;
518 519
      }
    }
520
    if (op_type == "flatten2" || op_type == "flatten") {
521 522 523
      if (!desc.HasAttr("axis")) {
        return false;
      } else {
524 525
#if IS_TRT_VERSION_GE(7130)
#else
526
        if (with_dynamic_shape) return false;
527
#endif
R
Ruibiao Chen 已提交
528
        int axis = PADDLE_GET_CONST(int, desc.GetAttr("axis"));
529 530 531
        if (axis != 1) return false;
      }
    }
532 533
    if (op_type == "flatten_contiguous_range") {
      if (!with_dynamic_shape) {
R
Ruibiao Chen 已提交
534 535
        int start_axis = PADDLE_GET_CONST(int, desc.GetAttr("start_axis"));
        int stop_axis = PADDLE_GET_CONST(int, desc.GetAttr("stop_axis"));
536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562
        auto x_var_name = desc.Input("X")[0];
        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_desc = block->FindVar(x_var_name);
        const auto x_shape = x_var_desc->GetShape();
        int dims = x_shape.size();
        if (start_axis < 0) start_axis += dims;
        if (start_axis == 0) {
          VLOG(3) << "TRT flatten_contiguous_range not support the "
                     "batch-dimension being changed";
          return false;
        }
        if (stop_axis < 0) stop_axis += dims;
        for (int i = start_axis; i <= stop_axis; ++i) {
          if (x_shape[i] < 0) {
            VLOG(3) << "On TRT static shape,flatten_contiguous_range input dim "
                       "should be > 0";
            return false;
          }
        }
      }
    }
563

564
    if (op_type == "gather") {
565 566 567 568 569 570 571 572 573
      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 {
574
        auto* block = desc.Block();
575 576 577 578 579 580
        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;
        }
F
feng_shuai 已提交
581 582 583 584 585 586 587 588 589 590

        auto index_var_name = desc.Input("Index")[0];
        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 op Index input data type must be int32";
          return false;
        }
F
feng_shuai 已提交
591
#if !IS_TRT_VERSION_GE(7000)
592 593 594 595 596 597
        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;
        }
F
feng_shuai 已提交
598
#endif
599
      }
600
    }
Z
zlsh80826 已提交
601

602
    if (op_type == "gather_nd") {
603 604
      if (!with_dynamic_shape) return false;

605
      auto* block = desc.Block();
606 607 608 609 610 611
      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;
      }
612
#if IS_TRT_VERSION_LT(8200)
613 614
      auto index_var_name = desc.Input("Index")[0];
      auto* index_var_desc = block->FindVar(index_var_name);
615 616
      auto x_var_name = desc.Input("X")[0];
      auto* x_var_desc = block->FindVar(x_var_name);
617 618
      const auto index_shape = index_var_desc->GetShape();
      const auto x_shape = x_var_desc->GetShape();
619 620 621 622 623 624
      if (x_shape.size() <= 2) {
        VLOG(3) << "gather_nd op requires the input's dimension to be greater "
                   "than 2";
        return false;
      }

625 626 627 628 629
      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;
      }
630
#endif
631 632
    }

633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662
    if (op_type == "take_along_axis") {
#if IS_TRT_VERSION_GE(8200)
      if (!with_dynamic_shape) return false;
      auto* block = desc.Block();
      auto input_var_name = desc.Input("Input")[0];
      auto index_var_name = desc.Input("Index")[0];
      auto* input_var_desc = block->FindVar(input_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) << "take_along_axis op Index input data type must be int32";
        return false;
      }

      const auto input_shape = input_var_desc->GetShape();
      const auto index_shape = index_var_desc->GetShape();
      if (input_shape.size() != index_shape.size()) {
        VLOG(3) << "take_along_axis op Index input dims size ["
                << index_shape.size() << " ] not equal to input dims size ["
                << input_shape.size() << "]";
        return false;
      }
#else
      VLOG(3) << "take_along_axis op is only supported by trt8.2 above ";
      return false;
#endif
    }

663 664 665 666
    if (op_type == "anchor_generator") {
      if (!with_dynamic_shape) return false;
    }

Z
zlsh80826 已提交
667 668 669 670 671 672
    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
673
      if (!has_attrs) return false;
Z
zlsh80826 已提交
674 675
    }

676 677 678 679 680 681
    if (op_type == "yolo_box_head") {
      if (with_dynamic_shape) return false;
      bool has_attrs = desc.HasAttr("class_num") && desc.HasAttr("anchors");
      if (!has_attrs) return false;
    }

682
    if (op_type == "arg_max") {
683 684 685 686 687 688
      if (!desc.HasAttr("axis", /*with_attr_var=*/false)) {
        VLOG(3) << "Skip to convert into TRT while found Attribute('axis') is "
                   "Variable type in arg_max.";
        return false;
      }

689
      int axis = desc.HasAttr("axis")
R
Ruibiao Chen 已提交
690
                     ? PADDLE_GET_CONST(int64_t, desc.GetAttr("axis"))
691
                     : -1;
R
Ruibiao Chen 已提交
692 693
      bool flatten = PADDLE_GET_CONST(bool, desc.GetAttr("flatten"));
      int dtype = PADDLE_GET_CONST(int, desc.GetAttr("dtype"));
694
      if (axis == 0 || flatten || (dtype != 2 && dtype != 3)) return false;
695 696
    }

697 698 699 700 701 702 703 704 705 706 707 708 709 710 711
    if (op_type == "arg_min") {
      if (!desc.HasAttr("axis", /*with_attr_var=*/false)) {
        VLOG(3) << "Skip to convert into TRT while found Attribute('axis') is "
                   "Variable type in arg_min.";
        return false;
      }

      int axis = desc.HasAttr("axis")
                     ? PADDLE_GET_CONST(int64_t, desc.GetAttr("axis"))
                     : -1;
      bool flatten = PADDLE_GET_CONST(bool, desc.GetAttr("flatten"));
      int dtype = PADDLE_GET_CONST(int, desc.GetAttr("dtype"));
      if (axis == 0 || flatten || dtype != 2) return false;
    }

712 713
    if (op_type == "affine_channel") {
      if (!desc.HasAttr("data_layout")) return false;
714
      auto data_layout = phi::StringToDataLayout(
R
Ruibiao Chen 已提交
715
          PADDLE_GET_CONST(std::string, desc.GetAttr("data_layout")));
716
      if (data_layout != phi::DataLayout::kNCHW) return false;
717 718

      auto* block = desc.Block();
719 720 721 722 723 724
      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;
      }
725 726 727 728 729 730
      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;
      }
731 732
    }

733
    if (op_type == "multiclass_nms" || op_type == "multiclass_nms3") {
Z
zlsh80826 已提交
734
      auto* block = desc.Block();
735 736 737 738 739 740
      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;
      }
741 742 743 744 745 746 747 748
      auto multiclass_nms_inputs = desc.Inputs();
      if (multiclass_nms_inputs.find("RoisNum") !=
          multiclass_nms_inputs.end()) {
        if (desc.Input("RoisNum").size() >= 1) {
          return false;
        }
      }
      for (auto& param_name : multiclass_nms_inputs) {
Z
zlsh80826 已提交
749 750 751 752
        for (auto& var_name : param_name.second) {
          auto* var_desc = block->FindVar(var_name);
          const auto shape = var_desc->GetShape();
          if (shape.size() != 3) {
753
            VLOG(3) << "multiclass_nms op dims != 3 not supported in tensorrt, "
Z
zlsh80826 已提交
754 755 756 757 758 759 760 761 762 763 764 765
                       "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;

766 767 768
      // TODO(wangxinxin08): tricky solution because the outputs of batchedNMS
      // plugin are not constient with those of multiclass_nms3
      if (desc.HasAttr("nms_eta") == false) return false;
R
Ruibiao Chen 已提交
769
      auto nms_eta = PADDLE_GET_CONST(float, desc.GetAttr("nms_eta"));
770 771
      if (nms_eta <= 1.0) return false;

R
Ruibiao Chen 已提交
772
      auto nms_top_k = PADDLE_GET_CONST(int, desc.GetAttr("nms_top_k"));
Z
zlsh80826 已提交
773 774
      if (nms_top_k < 0) return false;

R
Ruibiao Chen 已提交
775
      auto keep_top_k = PADDLE_GET_CONST(int, desc.GetAttr("keep_top_k"));
Z
zlsh80826 已提交
776 777 778 779 780 781
      if (keep_top_k < 0) return false;

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

782
    if (op_type == "nearest_interp") {
C
ccrrong 已提交
783 784
      std::vector<std::string> attrs{
          "interp_method", "align_corners", "scale", "out_h", "out_w"};
785 786 787
      for (auto const attr : attrs) {
        if (!desc.HasAttr(attr)) return false;
      }
788
      if (desc.HasAttr("data_layout")) {
789
        auto data_layout = phi::StringToDataLayout(
R
Ruibiao Chen 已提交
790
            PADDLE_GET_CONST(std::string, desc.GetAttr("data_layout")));
791 792
        if (data_layout != phi::DataLayout::kNCHW &&
            data_layout != phi::DataLayout::kNHWC)
793 794
          return false;
      }
795
      auto interp_method =
R
Ruibiao Chen 已提交
796
          PADDLE_GET_CONST(std::string, desc.GetAttr("interp_method"));
797
      if (interp_method != "nearest") return false;
R
Ruibiao Chen 已提交
798 799 800 801 802
      auto scale = PADDLE_GET_CONST(float, desc.GetAttr("scale"));
      auto out_h = PADDLE_GET_CONST(int, desc.GetAttr("out_h"));
      auto out_w = PADDLE_GET_CONST(int, desc.GetAttr("out_w"));
      auto align_corners =
          PADDLE_GET_CONST(bool, desc.GetAttr("align_corners"));
803 804 805 806
      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;
807
        }
808 809
        if (out_w <= 0) {
          VLOG(3) << "out_w must be greater than 0 if scale is not set.";
已提交
810 811
          return false;
        }
812
      }
813 814 815 816 817 818 819 820 821
      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;
      }
822
    }
823

824
    if (op_type == "nearest_interp_v2") {
C
ccrrong 已提交
825 826 827 828 829 830
      std::vector<std::string> attrs{"data_layout",
                                     "interp_method",
                                     "align_corners",
                                     "scale",
                                     "out_h",
                                     "out_w"};
831 832 833
      for (auto const attr : attrs) {
        if (!desc.HasAttr(attr)) return false;
      }
834
      auto data_layout = phi::StringToDataLayout(
R
Ruibiao Chen 已提交
835
          PADDLE_GET_CONST(std::string, desc.GetAttr("data_layout")));
836 837
      if (data_layout != phi::DataLayout::kNCHW &&
          data_layout != phi::DataLayout::kNHWC)
838 839
        return false;
      auto interp_method =
R
Ruibiao Chen 已提交
840
          PADDLE_GET_CONST(std::string, desc.GetAttr("interp_method"));
841
      if (interp_method != "nearest") return false;
R
Ruibiao Chen 已提交
842 843 844
      auto scale = PADDLE_GET_CONST(std::vector<float>, desc.GetAttr("scale"));
      auto out_h = PADDLE_GET_CONST(int, desc.GetAttr("out_h"));
      auto out_w = PADDLE_GET_CONST(int, desc.GetAttr("out_w"));
845
      if (!(out_h > 0 && out_w > 0)) {
W
wenbin 已提交
846
        if (scale.size() < 2) return false;
847 848 849 850 851 852 853 854
        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;
        }
      }
    }

855
    if (op_type == "bilinear_interp_v2") {
C
ccrrong 已提交
856 857 858 859 860 861
      std::vector<std::string> attrs{"data_layout",
                                     "interp_method",
                                     "align_corners",
                                     "scale",
                                     "out_h",
                                     "out_w"};
862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880
      for (auto const attr : attrs) {
        if (!desc.HasAttr(attr)) {
          VLOG(3) << "The op_type " << op_type << " doesn't have the attr "
                  << attr << " and return false";
          return false;
        }
      }

      auto resize_inputs = desc.Inputs();
      if (resize_inputs.find("SizeTensor") != resize_inputs.end()) {
        if (desc.Input("SizeTensor").size() >= 1) {
          VLOG(3)
              << "The Paddle-TRT doesn't support the SizeTensor for op_type "
              << op_type;
          return false;
        }
      }

      if (resize_inputs.find("OutSize") != resize_inputs.end()) {
881 882
        if (!with_dynamic_shape) {
          VLOG(3) << "Static shape don't support the OutSize for op_type "
883 884 885 886 887
                  << op_type;
          return false;
        }
      }

888
      auto data_layout = phi::StringToDataLayout(
R
Ruibiao Chen 已提交
889
          PADDLE_GET_CONST(std::string, desc.GetAttr("data_layout")));
890 891
      if (data_layout != phi::DataLayout::kNCHW &&
          data_layout != phi::DataLayout::kNHWC) {
892 893 894 895 896
        VLOG(3) << "The op_type " << op_type
                << " is not NCHW or NHWC return false";
        return false;
      }
      auto interp_method =
R
Ruibiao Chen 已提交
897
          PADDLE_GET_CONST(std::string, desc.GetAttr("interp_method"));
898 899 900 901 902 903
      if (interp_method != "bilinear") {
        VLOG(3) << "The interp_method of op_type " << op_type
                << " is not bilinear";
        return false;
      }

R
Ruibiao Chen 已提交
904 905
      auto align_corners =
          PADDLE_GET_CONST(bool, desc.GetAttr("align_corners"));
906 907 908 909 910 911 912 913 914 915 916
      if (align_corners != false) {
        VLOG(3)
            << "The bilinear_interp_v2 only supports align_corners with false.";
        return false;
      }

      bool has_scale_input_size =
          (resize_inputs.find("Scale") != resize_inputs.end());

      if (has_scale_input_size && desc.Input("Scale").size() != 1) {
        const std::vector<float> scale =
R
Ruibiao Chen 已提交
917
            PADDLE_GET_CONST(std::vector<float>, desc.GetAttr("scale"));
918 919 920 921 922 923 924
        if (scale.size() <= 1) {
          if (!desc.HasAttr("out_h") || !desc.HasAttr("out_w")) {
            VLOG(3) << "The op_type " << op_type
                    << " doesn't have Scale and the scale size <=1 and without "
                       "out_h / out_w, it will return false";
            return false;
          }
R
Ruibiao Chen 已提交
925 926
          auto out_h = PADDLE_GET_CONST(int, desc.GetAttr("out_h"));
          auto out_w = PADDLE_GET_CONST(int, desc.GetAttr("out_w"));
927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951
          if (!(out_h <= 0 && out_w <= 0)) {
            if (out_h <= 0) {
              VLOG(3) << "The op_type " << op_type
                      << "'s out_h must be greater than 0 if scale is not set.";
              return false;
            }
            if (out_w <= 0) {
              VLOG(3) << "The op_type " << op_type
                      << "'s out_w must be greater than 0 if scale is not set.";
              return false;
            }
          }
        } else {
          for (size_t i = 0; i < scale.size(); i++) {
            if (scale[i] <= 0 && with_dynamic_shape) {
              VLOG(3) << "dynamic shape not support Attr(scale[" << i << "]) "
                      << scale[i]
                      << " less than 1 and Input(Scale) vector not set.";
              return false;
            }
          }
        }
      }
    }

952 953 954 955 956 957 958 959 960 961 962 963 964 965
    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;
      }
    }

966
    if (op_type == "squeeze2") {
967 968 969 970 971 972 973
      // If Attribute is Variable(s), HasAttr() will return False
      if (!desc.HasAttr("axes", /*with_attr_var=*/false)) {
        VLOG(3) << "Skip to convert into TRT while found Attribute('axes') is "
                   "Variable type in squeeze2.";
        return false;
      }

974 975
      std::vector<int> axes;
      if (desc.HasAttr("axes")) {
R
Ruibiao Chen 已提交
976
        axes = PADDLE_GET_CONST(std::vector<int>, desc.GetAttr("axes"));
977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994
      }
      if (axes.size() == 0) {
        VLOG(3) << "The necessary attributes of the squeeze2 operator axes is "
                   "missing.";
        return false;
      }
      if (!with_dynamic_shape) {
        if (std::find(axes.begin(), axes.end(), 0) != axes.end()) {
          VLOG(3) << "Invalid squeeze axes. Axes having batch axis is not "
                     "supported in static shape";
          return false;
        }
      }
    }

    if (op_type == "unsqueeze2") {
      std::vector<int> axes;
      if (desc.HasAttr("axes")) {
R
Ruibiao Chen 已提交
995
        axes = PADDLE_GET_CONST(std::vector<int>, desc.GetAttr("axes"));
996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010
      }
      if (axes.size() == 0) {
        VLOG(3) << "The necessary attributes of the squeeze2 operator axes is "
                   "missing.";
        return false;
      }
      if (!with_dynamic_shape) {
        if (std::find(axes.begin(), axes.end(), 0) != axes.end()) {
          VLOG(3) << "Invalid squeeze axes. Axes having batch axis is not "
                     "supported in static shape";
          return false;
        }
      }
    }

1011
    if (op_type == "batch_norm") {
C
ccrrong 已提交
1012 1013
      const std::vector<std::string> bn_inputs = {
          "X", "Bias", "Mean", "Scale", "Variance"};
1014 1015 1016 1017 1018 1019 1020 1021 1022
      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;
        }
      }
1023 1024 1025 1026 1027 1028
      auto batch_norm_inputs = desc.Inputs();
      if (batch_norm_inputs.find("MomentumTensor") != batch_norm_inputs.end()) {
        if (desc.Input("MomentumTensor").size() >= 1) {
          return false;
        }
      }
1029 1030 1031 1032 1033 1034
      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
1035 1036 1037 1038 1039 1040 1041 1042 1043 1044
      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();
1045 1046 1047 1048 1049 1050 1051 1052 1053
    }

    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;
      }
1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064
      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;
        }
      }
1065 1066
      if (!desc.HasAttr("axis")) {
        return false;
1067
      }
R
Ruibiao Chen 已提交
1068
      int axis = PADDLE_GET_CONST(int, desc.GetAttr("axis"));
1069 1070 1071 1072 1073 1074 1075

      if (axis == 0) {
        VLOG(3) << "Invalid split axis. Split on batch is not supported in "
                   "TensorRT";
        return false;
      }
      auto* block = desc.Block();
1076 1077 1078 1079 1080 1081
      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;
      }
1082 1083 1084 1085 1086 1087 1088
      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")) {
R
Ruibiao Chen 已提交
1089
        num = PADDLE_GET_CONST(int, desc.GetAttr("num"));
1090 1091 1092
      }
      if (desc.HasAttr("sections")) {
        output_lengths =
R
Ruibiao Chen 已提交
1093
            PADDLE_GET_CONST(std::vector<int>, desc.GetAttr("sections"));
1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125
      }
      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);
          }
1126 1127
        }
      }
1128 1129 1130 1131
      if (output_lengths.size() != output_num) {
        VLOG(3) << "The output_length should be equal to the output size.";
        return false;
      }
1132
    }
1133

1134 1135 1136 1137 1138 1139 1140 1141
    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();
1142 1143 1144 1145 1146 1147
      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;
      }
1148 1149 1150
      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();
1151
      auto dtype = x_var_desc->GetDataType();
W
wenbin 已提交
1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169
      if (!with_dynamic_shape) {
        // At present, only support float32 or float16 into trt.
        if (!(dtype == framework::proto::VarType::FP32 ||
              dtype == framework::proto::VarType::FP16)) {
          return false;
        }
        if (x_shape.size() == 1) {
          VLOG(3)
              << "Scale op does not support 1-dimensional input in tensorrt";
          return false;
        }
      } else {
        // At present, only support float32 or float16 or int32 into trt.
        if (!(dtype == framework::proto::VarType::FP32 ||
              dtype == framework::proto::VarType::FP16 ||
              dtype == framework::proto::VarType::INT32)) {
          return false;
        }
1170
      }
1171
    }
1172

F
feng_shuai 已提交
1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183
    if (op_type == "roll") {
#if !IS_TRT_VERSION_GE(7000)
      VLOG(3) << "roll converter does not support trt versions below 7.0";
      return false;
#endif
      if (!with_dynamic_shape) {
        return false;
      }
    }

    if (op_type == "strided_slice") {
1184 1185 1186 1187 1188
#if !IS_TRT_VERSION_GE(7000)
      VLOG(3)
          << "strided_slice converter does not support trt versions below 7.0";
      return false;
#endif
F
feng_shuai 已提交
1189 1190 1191 1192 1193 1194 1195 1196
      if (!desc.HasAttr("axes") || !desc.HasAttr("starts") ||
          !desc.HasAttr("ends") || !desc.HasAttr("strides")) {
        VLOG(3)
            << "The necessary attributes of the strided_slice operator miss ";
        return false;
      }
    }

Z
zhoutianzi666 已提交
1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247
    if (op_type == "rnn") {
      if (!with_dynamic_shape) {
        return false;
      }
      if (desc.HasAttr("mode")) {
        std::string mode = PADDLE_GET_CONST(std::string, desc.GetAttr("mode"));
        if (mode != "LSTM") return false;
      }
      if (desc.HasAttr("dropout_prob")) {
        float dropout_prob =
            PADDLE_GET_CONST(float, desc.GetAttr("dropout_prob"));
        if (dropout_prob > 1e-5) return false;
      }
      // not support following four inputs for rnn in paddle-trt
      auto rnn_inputs = desc.Inputs();
      if (rnn_inputs.find("SequenceLength") != rnn_inputs.end()) {
        if (desc.Input("SequenceLength").size()) {
          return false;
        }
      }
    }

    if (op_type == "fill_constant_batch_size_like") {
      if (!with_dynamic_shape) {
        return false;
      }
      if (!desc.HasAttr("input_dim_idx")) {
        return false;
      }
      if (!desc.HasAttr("output_dim_idx")) {
        return false;
      }
      if (!desc.HasAttr("shape")) {
        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("Input")[0];
      auto* x_var_desc = block->FindVar(x_var_name);
      auto dtype = x_var_desc->GetDataType();
      // At present, only support float32 into trt.
      if (dtype != 5) {
        return false;
      }
    }

1248 1249 1250 1251 1252 1253
    if (op_type == "fill_any_like") {
      if (!with_dynamic_shape) {
        VLOG(3) << "the fill_any_like does not support static shape yet";
        return false;
      }
      int dtype = PADDLE_GET_CONST(int, desc.GetAttr("dtype"));
1254 1255 1256 1257 1258 1259 1260 1261 1262 1263
      auto* block = desc.Block();
      auto* x_var_desc = block->FindVar(desc.Input("X")[0]);
      auto input_type = x_var_desc->GetDataType();
#if IS_TRT_VERSION_GE(8400)
      if (dtype == 0 ||
          (dtype == -1 && input_type == framework::proto::VarType::BOOL)) {
        VLOG(3) << "the fill_any_like supports input of BOOL by trt8.4 above";
        return true;
      }
#endif
1264
      if (dtype != -1 && dtype != 2 && dtype != 5) {
1265 1266
        VLOG(3) << "the fill_any_like only supports int32 and float32 by "
                   "trt8.4 below";
1267 1268 1269 1270 1271
        return false;
      }
      if (dtype == -1) {
        if (input_type != framework::proto::VarType::INT32 &&
            input_type != framework::proto::VarType::FP32) {
1272 1273
          VLOG(3) << "the fill_any_like only supports int32 and float32 by "
                     "trt8.4 below";
1274 1275 1276 1277 1278
          return false;
        }
      }
    }

1279
    if (op_type == "slice") {
1280 1281
      if (desc.HasAttr("decrease_axis")) {
        std::vector<int> decrease_axis =
R
Ruibiao Chen 已提交
1282
            PADDLE_GET_CONST(std::vector<int>, desc.GetAttr("decrease_axis"));
1283 1284 1285
        if (!with_dynamic_shape) {
          if (decrease_axis.end() !=
              std::find(decrease_axis.begin(), decrease_axis.end(), 0)) {
1286 1287
            return false;
          }
1288 1289
        }
      }
1290 1291
      std::vector<int> axes;
      if (!desc.HasAttr("axes")) {
1292
        VLOG(3) << "The necessary attributes of the slice operator axes "
1293
                   " are missing.";
1294 1295
        return false;
      } else {
1296
        axes = PADDLE_GET_CONST(std::vector<int>, desc.GetAttr("axes"));
1297 1298 1299 1300 1301 1302 1303 1304 1305 1306
        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;
            }
          }
        }
      }
1307 1308
      // not support following four inputs for slice in paddle-trt
      auto slice_inputs = desc.Inputs();  // its size == 5
1309 1310 1311 1312 1313 1314 1315 1316
      if (slice_inputs.find("StartsTensor") != slice_inputs.end() &&
          desc.Input("StartsTensor").size()) {
        VLOG(3) << "The Slice has StartsTensor input.";
      } else {
        if (!desc.HasAttr("starts")) {
          VLOG(3) << "The necessary attributes of the slice operator starts or "
                     "StartsTensor"
                     " are missing.";
1317
          return false;
1318 1319 1320 1321 1322 1323 1324 1325
        } else {
          std::vector<int> starts =
              PADDLE_GET_CONST(std::vector<int>, desc.GetAttr("starts"));
          if (axes.size() != starts.size()) {
            VLOG(3) << "The shape of attributes of the slice operator axes "
                       "and starts are not equal.";
            return false;
          }
1326 1327
        }
      }
1328 1329 1330 1331 1332 1333 1334 1335
      if (slice_inputs.find("EndsTensor") != slice_inputs.end() &&
          desc.Input("EndsTensor").size()) {
        VLOG(3) << "The Slice has EndsTensor input.";
      } else {
        if (!desc.HasAttr("ends")) {
          VLOG(3) << "The necessary attributes of the slice operator ends or "
                     "EndsTensor"
                     " are missing.";
1336
          return false;
1337 1338 1339 1340 1341 1342 1343 1344
        } else {
          std::vector<int> ends =
              PADDLE_GET_CONST(std::vector<int>, desc.GetAttr("ends"));
          if (axes.size() != ends.size()) {
            VLOG(3) << "The shape of attributes of the slice operator axes "
                       "and ends are not equal.";
            return false;
          }
1345 1346 1347 1348 1349 1350 1351 1352 1353 1354 1355 1356
        }
      }
      if (slice_inputs.find("StartsTensorList") != slice_inputs.end()) {
        if (desc.Input("StartsTensorList").size()) {
          return false;
        }
      }
      if (slice_inputs.find("EndsTensorList") != slice_inputs.end()) {
        if (desc.Input("EndsTensorList").size()) {
          return false;
        }
      }
1357 1358
    }

1359 1360 1361 1362 1363 1364 1365 1366 1367 1368 1369 1370 1371 1372 1373 1374 1375 1376 1377 1378 1379 1380 1381 1382 1383 1384
    if (op_type == "less_than" || op_type == "greater_than" ||
        op_type == "logical_or" || op_type == "logical_xor" ||
        op_type == "logical_and" || op_type == "less_equal") {
#if IS_TRT_VERSION_GE(8400)
      if (!with_dynamic_shape) {
        VLOG(3) << "these ops do not support static shape yet";
        return false;
      }
      if (op_type == "logical_or" || op_type == "logical_xor" ||
          op_type == "logical_and") {
        auto* block = desc.Block();
        auto* x_var_desc = block->FindVar(desc.Input("X")[0]);
        auto* y_var_desc = block->FindVar(desc.Input("Y")[0]);
        auto x_dtype = x_var_desc->GetDataType();
        auto y_dtype = y_var_desc->GetDataType();
        if (x_dtype != framework::proto::VarType::BOOL ||
            y_dtype != framework::proto::VarType::BOOL) {
          VLOG(3) << "the op only support input of BOOL.";
          return false;
        }
      }
#else
      VLOG(3) << "these are not supported when TensorRT < 8.4";
      return false;
#endif
    }
1385
    if (op_type == "elementwise_add" || op_type == "elementwise_mul" ||
S
shentanyue 已提交
1386
        op_type == "elementwise_sub" || op_type == "elementwise_div" ||
1387
        op_type == "elementwise_pow" || op_type == "elementwise_min" ||
W
wenbin 已提交
1388
        op_type == "elementwise_max" || op_type == "elementwise_floordiv") {
1389 1390 1391 1392 1393 1394 1395 1396 1397 1398 1399 1400 1401 1402 1403 1404 1405 1406
      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;
      }
1407
      auto* block = desc.Block();
1408 1409 1410 1411 1412 1413
      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;
      }
1414 1415 1416 1417
      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();
1418 1419 1420 1421 1422 1423 1424

      // The case when x_shape.size() == 1 is dealt with in common case
      if (!with_dynamic_shape && (!y_var_desc->Persistable()) &&
          y_shape.size() == 1) {
        VLOG(3) << "Static shape in trt not support y is  a 1D intermediate "
                   "tensor in "
                   "elementwise op.";
1425 1426
        return false;
      }
1427 1428 1429 1430
      if (x_var_desc->Persistable() && !with_dynamic_shape) {
        VLOG(3)
            << "Input X is a parameter which is not supported for "
               "elementwise in tensorrt's static shape, swap x and y will work";
S
shentanyue 已提交
1431
        return false;
1432
      }
1433 1434 1435 1436 1437 1438 1439 1440 1441 1442 1443 1444
    }

    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;
      }
    }
1445 1446 1447 1448 1449 1450 1451 1452
    // remember that 1D input in static shape mode is filtered at the beginning
    if (op_type == "sum") {
      return true;
    }

    if (op_type == "shape" && !with_dynamic_shape) {
      return false;
    }
1453 1454 1455 1456 1457 1458 1459 1460 1461 1462 1463 1464

    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()) {
        return false;
      }
    }

1465 1466
    if (op_type == "fused_preln_embedding_eltwise_layernorm") {
      if (!with_dynamic_shape) {
1467 1468 1469
        VLOG(3) << "fused_preln_embedding_eltwise_layernorm should run on "
                   "dynamic "
                   "shape mode.";
1470 1471 1472 1473 1474 1475 1476 1477 1478 1479 1480 1481 1482
        return false;
      }
      if (desc.Input("Ids").size() != desc.Input("Embs").size()) {
        VLOG(3) << "The id and emb size of fused PrelnEmbEltwiseLayerNormOp "
                   "should be same ";
        return false;
      }
      if (!desc.HasAttr("enable_int8")) {
        VLOG(3) << "PrelnEmbEltwiseLayerNormOp must use int8 mode.";
        return false;
      }
    }

1483 1484 1485 1486 1487 1488 1489 1490 1491 1492 1493
    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;
      }
1494

1495
#if IS_TRT_VERSION_LT(7000)
1496
      if (desc.HasAttr("approximate")) {
1497
        VLOG(3) << "approximate gelu op needs TensorRT 7.0 and after";
R
Ruibiao Chen 已提交
1498
        if (PADDLE_GET_CONST(bool, desc.GetAttr("approximate"))) return false;
1499
      }
1500
#endif
1501 1502

      auto* block = desc.Block();
1503 1504 1505 1506 1507 1508
      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;
      }
1509

1510 1511 1512 1513 1514 1515 1516
      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;
      }
1517 1518 1519 1520 1521 1522 1523 1524 1525 1526 1527 1528 1529 1530 1531 1532 1533 1534 1535 1536 1537 1538 1539 1540 1541
    }

    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;
      }
    }

1542 1543 1544 1545 1546 1547 1548 1549 1550 1551 1552 1553 1554 1555
    if (op_type == "fill_constant") {
      auto fill_constant_inputs = desc.Inputs();
      if (fill_constant_inputs.find("ValueTensor") !=
          fill_constant_inputs.end()) {
        if (desc.Input("ValueTensor").size()) return false;
      }
      if (fill_constant_inputs.find("ShapeTensor") !=
          fill_constant_inputs.end()) {
        if (desc.Input("ShapeTensor").size()) return false;
      }
      if (fill_constant_inputs.find("ShapeTensorList") !=
          fill_constant_inputs.end()) {
        if (desc.Input("ShapeTensorList").size()) return false;
      }
R
Ruibiao Chen 已提交
1556
      int dtype = PADDLE_GET_CONST(int, desc.GetAttr("dtype"));
1557 1558 1559 1560 1561 1562
      // only support int32, int64, float32
      if (!(dtype == 2 || dtype == 3 || dtype == 5)) {
        return false;
      }
    }

已提交
1563 1564 1565 1566 1567 1568 1569 1570 1571 1572 1573 1574 1575 1576 1577 1578 1579 1580 1581 1582 1583
    if (op_type == "instance_norm") {
      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;
      }
1584 1585 1586 1587 1588 1589 1590 1591 1592 1593 1594 1595 1596 1597 1598 1599

      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
1600 1601
    }

1602 1603 1604 1605 1606 1607 1608 1609 1610 1611 1612 1613 1614 1615 1616
    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") {
R
Ruibiao Chen 已提交
1617 1618
      const float pad_value =
          PADDLE_GET_CONST(float, desc.GetAttr("pad_value"));
1619 1620 1621 1622
      if (pad_value != 0.0f) {
        VLOG(3) << "The pad layer of TRT only support zero.";
        return false;
      }
已提交
1623 1624
      std::vector<int64_t> shape;
      auto* block = desc.Block();
1625 1626 1627 1628 1629 1630
      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
1631 1632 1633 1634 1635 1636 1637 1638
      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 =
R
Ruibiao Chen 已提交
1639
          PADDLE_GET_CONST(std::vector<int>, desc.GetAttr("paddings"));
已提交
1640 1641 1642 1643 1644 1645 1646 1647 1648 1649 1650 1651
      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;
        }
      }
1652 1653
    }

1654 1655
    if (op_type == "swish") {
      auto* block = desc.Block();
1656 1657 1658 1659 1660 1661
      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;
      }
1662 1663 1664 1665 1666 1667 1668 1669 1670
      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;
      }
    }

1671 1672 1673 1674 1675 1676 1677 1678 1679 1680 1681 1682 1683
    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;
      }
1684 1685

      auto* block = desc.Block();
1686 1687 1688 1689 1690 1691
      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;
      }
1692 1693 1694 1695 1696 1697 1698 1699 1700
      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();
1701 1702 1703
      if (!with_dynamic_shape && x_shape.size() == 1) {
        VLOG(3) << "prelu op does not support input's dim is 1 in tensorrt "
                   "with static shape.";
1704 1705 1706
        return false;
      }

W
Wilber 已提交
1707 1708 1709 1710 1711 1712 1713
#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
1714 1715
    }

W
wangxinxin08 已提交
1716 1717 1718 1719 1720 1721 1722 1723 1724 1725 1726 1727 1728 1729 1730 1731 1732 1733 1734 1735 1736 1737 1738 1739 1740 1741 1742 1743 1744 1745 1746
    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;
      }
    }

1747 1748 1749 1750 1751 1752 1753
    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;
      }
C
ccrrong 已提交
1754 1755 1756 1757
      std::vector<std::string> attrs{"pooled_height",
                                     "pooled_width",
                                     "spatial_scale",
                                     "sampling_ratio",
F
fengkuangxiaxia 已提交
1758
                                     "aligned"};
1759 1760 1761 1762 1763
      for (auto const attr : attrs) {
        if (!desc.HasAttr(attr)) return false;
      }

      const auto pooled_height =
R
Ruibiao Chen 已提交
1764
          PADDLE_GET_CONST(int, desc.GetAttr("pooled_height"));
1765 1766 1767
      if (pooled_height <= 0) return false;

      const auto pooled_width =
R
Ruibiao Chen 已提交
1768
          PADDLE_GET_CONST(int, desc.GetAttr("pooled_width"));
1769 1770 1771
      if (pooled_width <= 0) return false;

      const auto spatial_scale =
R
Ruibiao Chen 已提交
1772
          PADDLE_GET_CONST(float, desc.GetAttr("spatial_scale"));
1773 1774 1775 1776 1777 1778 1779 1780
      if (spatial_scale <= 0.f) 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;
        }
      }
1781 1782 1783
    }

    if (op_type == "shuffle_channel") {
1784
#if !IS_TRT_VERSION_GE(8000)
1785 1786
      if (with_dynamic_shape) {
        VLOG(3) << "You are running the TRT Dynamic Shape mode, "
1787 1788
                   "the shuffle_channel op does not support dynamic shape "
                   "trt versions below 8.0 yet";
1789 1790
        return false;
      }
1791
#endif
1792 1793
    }

1794 1795 1796 1797 1798 1799 1800 1801 1802 1803 1804
    if (op_type == "where") {
#if !IS_TRT_VERSION_GE(8400)
      VLOG(3) << "where is not supported when TensorRT < 8.4";
      return false;
#endif
      if (!with_dynamic_shape) {
        VLOG(3) << "the where op does not support static shape yet";
        return false;
      }
    }

1805 1806 1807 1808 1809 1810 1811 1812 1813 1814 1815 1816 1817 1818 1819 1820 1821 1822 1823 1824 1825 1826 1827 1828 1829 1830 1831 1832 1833 1834 1835 1836 1837 1838 1839 1840 1841 1842 1843
    if (op_type == "one_hot" || op_type == "one_hot_v2") {
#if IS_TRT_VERSION_LT(8510)
      VLOG(3) << "one_hot/one_hot_v2 is not supported when TensorRT < 8.5.1";
      return false;
#endif
      if (!with_dynamic_shape) {
        VLOG(3)
            << "the one_hot/one_hot_v2 op does not support static shape yet";
        return false;
      }
      if (desc.HasAttr("allow_out_of_range")) {
        VLOG(3)
            << "allow_out_of_range one_hot/one_hot_v2 op is not supported now.";
        if (PADDLE_GET_CONST(bool, desc.GetAttr("allow_out_of_range")))
          return false;
      }
      if (desc.HasAttr("dtype")) {
        const int dtype = PADDLE_GET_CONST(int, desc.GetAttr("dtype"));
        if (dtype != 2 && dtype != 3 && dtype != 5) {
          VLOG(3) << "one_hot/one_hot_v2 op only support int32, int64, float.";
          return false;
        }
      }
      auto one_hot_inputs = desc.Inputs();
      if (one_hot_inputs.find("depth_tensor") != one_hot_inputs.end()) {
        if (desc.Input("depth_tensor").size() != 0) {
          return true;
        }
      }

      if (desc.HasAttr("depth")) {
        const int depth = PADDLE_GET_CONST(int, desc.GetAttr("depth"));
        if (depth <= 0) {
          VLOG(3) << "depth only support positive in one_hot/one_hot_v2 op.";
          return false;
        }
      }
    }

1844 1845 1846 1847 1848 1849 1850
    if (op_type == "skip_layernorm") {
      if (!with_dynamic_shape) {
        VLOG(3) << "the skip_layernorm does not support static shape yet";
        return false;
      }
    }

1851 1852 1853 1854 1855 1856 1857 1858 1859 1860 1861
    if (op_type == "preln_skip_layernorm") {
      if (!with_dynamic_shape) {
        VLOG(3) << "the preln_skip_layernorm does not support static shape yet";
        return false;
      }
      if (!desc.HasAttr("enable_int8")) {
        VLOG(3) << "PrelnEmbEltwiseLayerNormOp must use int8 mode.";
        return false;
      }
    }

1862 1863 1864 1865 1866
    if (op_type == "multihead_matmul") {
      if (!with_dynamic_shape) {
        VLOG(3) << "the multihead_matmul does not support static shape yet";
        return false;
      }
1867 1868 1869 1870 1871 1872 1873 1874 1875 1876 1877 1878 1879 1880 1881 1882

      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 =
R
Ruibiao Chen 已提交
1883
          PADDLE_GET_CONST(int, desc.GetAttr("head_number"));
F
feng_shuai 已提交
1884 1885 1886 1887 1888 1889 1890 1891 1892
      auto inputs = desc.Inputs();
      bool has_bias_qk = (inputs.find("BiasQK") == inputs.end()) ? false : true;
      if (has_bias_qk) {
        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] &&
1893
                              input_shape[1] == biasqk_shape[3];
F
feng_shuai 已提交
1894 1895
        bool is_broadcastable = biasqk_shape[1] == 1 && biasqk_shape[2] == 1 &&
                                input_shape[1] == biasqk_shape[3];
1896 1897 1898 1899
        is_broadcastable =
            is_broadcastable || (biasqk_shape[0] == 1 && biasqk_shape[1] == 1 &&
                                 input_shape[1] == biasqk_shape[2] &&
                                 input_shape[1] == biasqk_shape[3]);
F
feng_shuai 已提交
1900 1901
        if (!(has_same_shape || is_broadcastable)) {
          VLOG(3) << "The BiasQK's shape is invalid, expect [" << input_shape[0]
1902 1903 1904 1905 1906 1907 1908
                  << ", 1, 1, " << input_shape[1] << "] "
                  << "or [" << input_shape[0] << ", " << head_number << ", "
                  << input_shape[1] << ", " << input_shape[1] << "] "
                  << "or [" << input_shape[0] << "/1, " << 1 << ", "
                  << input_shape[1] << ", " << input_shape[1] << "] "
                  << "but got [" << biasqk_shape[0] << ", " << biasqk_shape[1]
                  << ", " << biasqk_shape[2] << ", " << biasqk_shape[3] << "].";
F
feng_shuai 已提交
1909 1910 1911
          return false;
        }
      } else {
1912 1913 1914
#if (IS_TRT_VERSION_GE(8000) && IS_TRT_VERSION_LT(8100)) || \
    (IS_TRT_VERSION_LT(7200))
        VLOG(3) << "There are some bugs with trt 8.0";
1915
        return false;
F
feng_shuai 已提交
1916
#endif
1917
      }
1918 1919
    }

1920 1921 1922 1923 1924 1925 1926 1927 1928 1929 1930 1931 1932 1933 1934 1935 1936 1937 1938 1939 1940 1941 1942 1943 1944 1945 1946 1947 1948 1949 1950 1951 1952 1953 1954 1955 1956 1957 1958 1959 1960 1961 1962 1963 1964 1965 1966 1967 1968 1969 1970 1971
    if (op_type == "multihead_matmul_roformer") {
      if (!with_dynamic_shape) {
        VLOG(3) << "the multihead_matmul_roformer does not support static "
                   "shape yet";
        return false;
      }

      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 =
          PADDLE_GET_CONST(int, desc.GetAttr("head_number"));
      auto inputs = desc.Inputs();
      bool has_bias_qk = (inputs.find("BiasQK") == inputs.end()) ? false : true;
      if (has_bias_qk) {
        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;
        }
      } else {
#if !IS_TRT_VERSION_GE(8000)
        VLOG(3) << "The version of TRT must be greater than 8000";
        return false;
#endif
      }
    }

1972
    if (op_type == "fc") {
1973 1974 1975 1976 1977 1978 1979 1980 1981 1982 1983 1984 1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998
      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)
1999 2000
            << " input_y(fc_op)'shapes must be 2, but input_y(fc_op)'shapes =
      "
2001 2002 2003 2004 2005 2006
            << y_shape.size();
        return false;
      }
      // y_num_col_dims ==1
      if (desc.HasAttr("y_num_col_dims")) {
        int y_num_col_dims =
R
Ruibiao Chen 已提交
2007
            PADDLE_GET_CONST(int, desc.GetAttr("y_num_col_dims"));
2008 2009 2010 2011 2012 2013 2014
        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;
        }
      }
      */
2015 2016
      int x_num_col_dims =
          desc.HasAttr("x_num_col_dims")
R
Ruibiao Chen 已提交
2017
              ? PADDLE_GET_CONST(int, desc.GetAttr("x_num_col_dims"))
2018
              : (desc.HasAttr("in_num_col_dims")
R
Ruibiao Chen 已提交
2019
                     ? PADDLE_GET_CONST(int, desc.GetAttr("in_num_col_dims"))
2020 2021
                     : 1);
      if (x_num_col_dims < 1) {
2022 2023 2024
        VLOG(3) << "fc_op expects x_num_col_dims >= 1, "
                   "but x_num_col_dims = "
                << x_num_col_dims;
2025 2026 2027
        return false;
      }
    }
2028

W
Wangzheee 已提交
2029 2030 2031
    if (op_type == "reshape" || op_type == "reshape2") {
      if (!desc.HasAttr("shape")) {
        return false;
W
Wilber 已提交
2032
      }
2033 2034 2035 2036
      if (with_dynamic_shape) {
        return true;
      }
      // Static shape does not support the input tensors: Shape and ShapeTensor
2037
      auto reshape_inputs = desc.Inputs();
2038 2039 2040 2041 2042 2043 2044 2045 2046
      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
2047
      }
W
Wilber 已提交
2048
      std::vector<int> shape =
R
Ruibiao Chen 已提交
2049
          PADDLE_GET_CONST(std::vector<int>, desc.GetAttr("shape"));
W
Wilber 已提交
2050
      if (shape.size() >= nvinfer1::Dims::MAX_DIMS) return false;
X
xiaoxiaohehe001 已提交
2051 2052 2053 2054 2055 2056 2057 2058 2059 2060 2061
      if (!with_dynamic_shape) {
        if (shape.size() == 1) {
          return false;
        }
        if (shape[0] == 0) {
          return true;
        } else {
          auto* block = desc.Block();
          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();
C
ccrrong 已提交
2062 2063 2064 2065
          int input_num = std::accumulate(
              x_shape.begin() + 1, x_shape.end(), 1, std::multiplies<int>());
          int shape_num = std::accumulate(
              shape.begin() + 1, shape.end(), 1, std::multiplies<int>());
X
xiaoxiaohehe001 已提交
2066 2067 2068 2069
          if (input_num == shape_num) {
            return true;
          }
        }
2070
        return false;
X
xiaoxiaohehe001 已提交
2071
      }
W
Wangzheee 已提交
2072
    }
2073

2074 2075 2076 2077 2078 2079 2080 2081 2082 2083 2084 2085 2086 2087 2088
    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();
2089 2090 2091 2092 2093 2094
      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;
      }
2095 2096 2097 2098 2099
      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();
    }

2100 2101
    if (op_type == "reduce_sum" || op_type == "reduce_mean" ||
        op_type == "reduce_max") {
2102 2103 2104 2105 2106 2107 2108
      if (!desc.HasAttr("dim", /*with_attr_var=*/false)) {
        VLOG(3) << "Skip to convert into TRT while found Attribute('dim') is "
                   "Variable type in "
                << desc.Type();
        return false;
      }

2109 2110
      if (!(desc.HasAttr("keep_dim") && desc.HasAttr("dim") &&
            desc.HasAttr("reduce_all"))) {
W
wenbin 已提交
2111 2112
        VLOG(3) << "the " << op_type
                << " does not have attr (keep_dim or dim or "
2113
                   "reduce_all)";
2114 2115 2116 2117 2118 2119 2120 2121
        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.";
2122 2123
        return false;
      }
W
wenbin 已提交
2124 2125

      // The batch size dimension cannot be reduced if it's not dynamic shape.
2126
      auto* x_var_desc = block->FindVar(desc.Input("X")[0]);
W
wenbin 已提交
2127
      if (!with_dynamic_shape) {
R
Ruibiao Chen 已提交
2128
        if (PADDLE_GET_CONST(bool, desc.GetAttr("reduce_all"))) return false;
W
wenbin 已提交
2129
        std::vector<int32_t> dim =
R
Ruibiao Chen 已提交
2130
            PADDLE_GET_CONST(std::vector<int32_t>, desc.GetAttr("dim"));
2131
        const auto input_shape = x_var_desc->GetShape();
W
wenbin 已提交
2132
        for (auto x : dim) {
2133
          if (x == 0 || (x + input_shape.size() == 0)) return false;
W
wenbin 已提交
2134
        }
2135

2136
      } else {
R
Ruibiao Chen 已提交
2137 2138
        if (PADDLE_GET_CONST(bool, desc.GetAttr("reduce_all")) &&
            !PADDLE_GET_CONST(bool, desc.GetAttr("keep_dim")))
2139 2140
          return false;
      }
2141 2142 2143 2144 2145 2146 2147

      auto dtype = x_var_desc->GetDataType();
#if IS_TRT_VERSION_GE(7000)
      if (dtype != framework::proto::VarType::INT32 &&
          dtype != framework::proto::VarType::FP32) {
        VLOG(3) << "reduce op input data type must be int32 or float32";
        return false;
W
wenbin 已提交
2148
      }
2149 2150
#else
      if (dtype != framework::proto::VarType::FP32) {
2151 2152
        VLOG(3) << "reduce op input data type must be float32 using TensorRT "
                   "< 7.0";
2153 2154 2155
        return false;
      }
#endif
2156
    }
W
wenbin 已提交
2157 2158 2159
#if IS_TRT_VERSION_GE(7000)
    if (op_type == "tile") {
      // Paddle-TRT does not support the input tensors.
2160 2161 2162
      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
2163
          return false;
2164 2165 2166 2167
        }
      }
      if (tile_inputs.find("RepeatTimes") != tile_inputs.end()) {
        if (desc.Input("RepeatTimes").size() >= 1) {
W
wenbin 已提交
2168
          return false;
2169
        }
W
wenbin 已提交
2170 2171 2172 2173 2174
      }
      if (with_dynamic_shape) return false;
      if (!with_dynamic_shape && !desc.HasAttr("repeat_times")) return false;
    }
#endif
2175

2176 2177 2178 2179 2180
    // conv3d_transpose
    if (op_type == "conv3d_transpose") {
      // trt doen't support output_padding when < 8406
      // output_padding is usually set when stride > 1
#if !IS_TRT_VERSION_GE(8400)
2181 2182
      if (desc.HasAttr("output_padding")) {
        const std::vector<int> output_padding =
R
Ruibiao Chen 已提交
2183
            PADDLE_GET_CONST(std::vector<int>, desc.GetAttr("output_padding"));
2184 2185 2186 2187 2188 2189
        if (output_padding.size() > 0) {
          int max_padding =
              *std::max_element(output_padding.begin(), output_padding.end());
          if (max_padding > 0) return false;
        }
      }
2190
#endif
2191 2192
    }

W
wenbin 已提交
2193 2194 2195
    if (op_type == "conv3d" || op_type == "conv3d_transpose") {
      if (desc.HasAttr("padding_algorithm")) {
        std::string padding_algorithm =
R
Ruibiao Chen 已提交
2196
            PADDLE_GET_CONST(std::string, desc.GetAttr("padding_algorithm"));
W
wenbin 已提交
2197 2198 2199 2200 2201 2202 2203 2204 2205 2206 2207 2208 2209 2210 2211

        // 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 =
R
Ruibiao Chen 已提交
2212
          PADDLE_GET_CONST(std::vector<int>, desc.GetAttr("paddings"));
W
wenbin 已提交
2213 2214 2215 2216 2217 2218 2219 2220 2221 2222 2223 2224 2225 2226 2227 2228 2229 2230 2231 2232 2233

      // 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 =
R
Ruibiao Chen 已提交
2234
              PADDLE_GET_CONST(std::vector<int>, desc.GetAttr("dilations"));
W
wenbin 已提交
2235 2236 2237 2238 2239 2240 2241 2242 2243 2244 2245 2246 2247 2248 2249 2250 2251
          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;
      }
    }

2252 2253 2254 2255
    if (op_type == "hard_sigmoid") {
      if (!with_dynamic_shape) {
        auto* block = desc.Block();
        if (block == nullptr) {
2256 2257 2258
          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.";
2259 2260 2261 2262 2263
          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();
2264 2265 2266
        if (x_shape.size() == 1) {
          VLOG(3) << "Hard sigmoid does not support 1-dimensional input in "
                     "tensorrt";
2267 2268 2269 2270 2271
          return false;
        }
      }
    }

C
ccrrong 已提交
2272
    if (op_type == "cast") {
Z
zhoutianzi666 已提交
2273 2274 2275 2276
// trt 6015 result in Windows ppyolo_mbv3 TRT fp32 diff
#if !IS_TRT_VERSION_GE(7000)
      return false;
#endif
C
ccrrong 已提交
2277 2278 2279 2280 2281 2282
      if (!(desc.HasAttr("in_dtype") && desc.HasAttr("out_dtype"))) {
        VLOG(3) << "the " << op_type
                << " does not have attr (in_dtype or "
                   "out_dtype)";
        return false;
      }
R
Ruibiao Chen 已提交
2283 2284
      int in_dtype = PADDLE_GET_CONST(int, desc.GetAttr("in_dtype"));
      int out_dtype = PADDLE_GET_CONST(int, desc.GetAttr("out_dtype"));
C
ccrrong 已提交
2285 2286 2287 2288
      if ((in_dtype == 4 || in_dtype == 5) && out_dtype == 4) {
        VLOG(3) << "unsupport data type conversion";
        return false;
      }
2289 2290 2291 2292 2293 2294 2295
#if IS_TRT_VERSION_GE(8400)
      if (in_dtype == 0 || out_dtype == 0) {
        if (with_dynamic_shape) {
          VLOG(3) << "the cast op supports inputs and outputs of BOOL by "
                     "trt8.4 above ";
          return true;
        }
2296
      }
2297
#endif
2298
      if (!((in_dtype == 5 || in_dtype == 4 || in_dtype == 2) &&
C
ccrrong 已提交
2299
            (out_dtype == 5 || out_dtype == 4 || out_dtype == 2))) {
2300 2301
        VLOG(3) << "only valid conversions are: "
                   "(kFLOAT | kHALF | kINT32) -> (kFLOAT | kHALF | kINT32)";
C
ccrrong 已提交
2302 2303 2304 2305
        return false;
      }
    }

2306 2307 2308 2309 2310 2311 2312 2313 2314 2315 2316
    if (op_type == "top_k_v2" || op_type == "top_k") {
      auto* block = desc.Block();
      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) << "top_k/top_k_v2 does not support 1-dimensional input in "
                   "tensorrt";
        return false;
      }
      if (desc.HasAttr("axis")) {
R
Ruibiao Chen 已提交
2317
        int axis = PADDLE_GET_CONST(int, desc.GetAttr("axis"));
2318 2319 2320 2321 2322 2323 2324
        if (axis == 0) {
          VLOG(3) << "top_k_v2 does not support axis == 0 in "
                     "tensorrt";
          return false;
        }
      }
      if (desc.HasAttr("sorted")) {
R
Ruibiao Chen 已提交
2325
        bool sorted = PADDLE_GET_CONST(bool, desc.GetAttr("sorted"));
2326 2327 2328 2329 2330 2331 2332 2333
        if (!sorted) {
          VLOG(3) << "top_k_v2 does not support results not sorted in "
                     "tensorrt";
          return false;
        }
      }
    }

2334 2335 2336 2337 2338 2339 2340 2341 2342 2343
#if IS_TRT_VERSION_GE(8000)
    if (op_type == "sparse_fc" || op_type == "sparse_multihead_matmul") {
      if (!with_dynamic_shape) {
        VLOG(3) << "the sparse_fc and sparse_multihead_matmul does not support "
                   "static shape yet";
        return false;
      }
    }
#endif

S
Sanbu 已提交
2344
    if (op_type == "equal" || op_type == "not_equal") {
C
ccrrong 已提交
2345 2346 2347 2348
#if !IS_TRT_VERSION_GE(8000)
      VLOG(3) << "compare is not supported when TensorRT < 8.0";
      return false;
#else
R
Ruibiao Chen 已提交
2349
      int axis = PADDLE_GET_CONST(int, desc.GetAttr("axis"));
C
ccrrong 已提交
2350 2351 2352 2353 2354 2355 2356 2357 2358 2359 2360 2361 2362
      if (axis == 0) {
        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;
      }
#endif
    }

W
wenbin 已提交
2363 2364 2365 2366 2367 2368 2369
    if (op_type == "layernorm_shift_partition") {
      if (!with_dynamic_shape) {
        VLOG(3) << "the layernorm_shift_partition does not support "
                   "static shape yet";
        return false;
      }
    }
W
wenbin 已提交
2370 2371 2372 2373 2374 2375 2376 2377 2378

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

W
Wang Bojun 已提交
2379 2380 2381 2382 2383 2384 2385
    if (op_type == "merge_layernorm") {
      if (!with_dynamic_shape) {
        VLOG(3) << "The merge_layernorm op does not support "
                   "static shape yet";
        return false;
      }
    }
W
wenbin 已提交
2386

W
Wang Bojun 已提交
2387 2388 2389 2390 2391 2392 2393
    if (op_type == "reverse_roll") {
      if (!with_dynamic_shape) {
        VLOG(3) << "The reverse roll fused op does not support static shape "
                   "mode yet.";
        return false;
      }
    }
W
wenbin 已提交
2394 2395 2396 2397 2398 2399 2400 2401
    if (op_type == "skip_merge_layernorm") {
      if (!with_dynamic_shape) {
        VLOG(3) << "The merge_layernorm op does not support "
                   "static shape yet";
        return false;
      }
    }

2402 2403 2404 2405 2406 2407 2408 2409
    if (op_type == "lookup_table") {
      if (!with_dynamic_shape) {
        VLOG(3) << "the lookup_table does not support "
                   "static shape yet";
        return false;
      }
    }

2410 2411 2412 2413 2414 2415 2416 2417 2418 2419 2420 2421 2422 2423 2424 2425 2426 2427 2428 2429 2430
    if (op_type == "expand_v2") {
      if (!with_dynamic_shape) {
        return false;
      }
      if (!desc.HasAttr("shape")) {
        return false;
      }
      auto expand_v2_inputs = desc.Inputs();
      if (expand_v2_inputs.find("Shape") != expand_v2_inputs.end()) {
        if (desc.Input("Shape").size() >= 1) {
          return false;
        }
      }
      if (expand_v2_inputs.find("expand_shapes_tensor") !=
          expand_v2_inputs.end()) {
        if (desc.Input("expand_shapes_tensor").size() >= 1) {
          return false;
        }
      }
    }

W
weishengying 已提交
2431 2432 2433 2434 2435
    if (use_no_calib_int8) {
      return int8_teller_set.count(op_type);
    } else {
      return teller_set.count(op_type);
    }
2436
  }
W
wenbin 已提交
2437

W
weishengying 已提交
2438 2439 2440 2441 2442
 private:
  // use this set for no calib int8.
  std::unordered_set<std::string> int8_teller_set{
      "mul",
      "matmul",
2443
      "matmul_v2",
2444
      "bmm",
2445
      "range",
W
weishengying 已提交
2446 2447 2448 2449 2450 2451 2452 2453 2454 2455 2456 2457 2458 2459 2460 2461 2462 2463 2464 2465 2466 2467 2468
      "conv2d",
      "conv2d_fusion",
      "pool2d",
      "relu",
      "elu",
      "selu",
      "softsign",
      "softplus",
      "stanh",
      "thresholded_relu",
      "exp",
      "log",
      "sqrt",
      "abs",
      "sin",
      "cos",
      "tan",
      "sinh",
      "cosh",
      "asin",
      "acos",
      "atan",
      "asinh",
2469
      "acosh",
W
weishengying 已提交
2470 2471 2472
      "atanh",
      "ceil",
      "floor",
G
gem5 已提交
2473
      "rsqrt",
2474
      "sign",
G
gem5 已提交
2475
      "reciprocal",
2476
      "logical_not",
W
weishengying 已提交
2477
      "erf",
2478
      "square",
W
weishengying 已提交
2479 2480 2481 2482 2483 2484 2485 2486 2487 2488 2489 2490 2491
      "softmax",
      "sigmoid",
      "hard_swish",
      "depthwise_conv2d",
      "batch_norm",
      "concat",
      "tanh",
      "pad",
      "elementwise_add",
      "elementwise_sub",
      "elementwise_mul",
      "elementwise_div",
      "elementwise_pow",
2492 2493
      "elementwise_min",
      "elementwise_max",
W
wenbin 已提交
2494
      "elementwise_floordiv",
W
weishengying 已提交
2495
      "equal",
S
Sanbu 已提交
2496
      "not_equal",
2497 2498 2499 2500 2501 2502
      "less_than",
      "greater_than",
      "logical_or",
      "logical_xor",
      "logical_and",
      "less_equal",
W
weishengying 已提交
2503
      "dropout",
2504
      "fill_any_like",
W
weishengying 已提交
2505 2506 2507 2508 2509 2510
      "prelu",
      "conv2d_transpose",
      "depthwise_conv2d_transpose",
      "leaky_relu",
      "fc",
      "shuffle_channel",
2511
      "where",
2512 2513
      "one_hot",
      "one_hot_v2",
W
weishengying 已提交
2514 2515
      "swish",
      "silu",
2516
      "celu",
W
weishengying 已提交
2517 2518 2519 2520 2521 2522 2523 2524 2525 2526 2527 2528 2529 2530 2531 2532 2533
      "split",
      "instance_norm",
      "gelu",
      "layer_norm",
      "scale",
      "stack",
      "transpose2",
      "transpose",
      "top_k",
      "top_k_v2",
      "flatten2",
      "flatten",
      "gather",
      "gather_nd",
      "yolo_box",
      "yolo_box_head",
      "arg_max",
2534
      "arg_min",
W
weishengying 已提交
2535 2536 2537 2538
      "roi_align",
      "affine_channel",
      "nearest_interp",
      "anchor_generator",
2539
      "reduce_max",
W
weishengying 已提交
2540
      "reduce_mean",
2541
      "reduce_sum",
W
weishengying 已提交
2542 2543 2544 2545 2546 2547 2548 2549 2550 2551 2552 2553
      "conv3d",
      "conv3d_transpose",
      "mish",
      "nearest_interp_v2",
      "bilinear_interp_v2",
      "pool3d",
      "deformable_conv",
      "relu6",
      "hard_sigmoid",
      "clip",
      "fused_embedding_eltwise_layernorm",
      "multihead_matmul",
2554
      "multihead_matmul_roformer",
W
weishengying 已提交
2555 2556 2557 2558 2559 2560 2561 2562 2563 2564 2565 2566 2567 2568 2569 2570 2571 2572 2573 2574
      "skip_layernorm",
      "slice",
      "strided_slice",
      "fused_preln_embedding_eltwise_layernorm",
      "preln_residual_bias",
      "c_allreduce_sum",
      "c_allreduce_min",
      "c_allreduce_max",
      "c_allreduce_prod",
      "roll",
      "cast",
      "preln_skip_layernorm",
      "transformer_input_convert",
      "recover_padding",
      "remove_padding",
      "fill_constant",
      "sum",
      "shape",
      "squeeze2",
      "unsqueeze2",
2575
      "layernorm_shift_partition",
W
Wang Bojun 已提交
2576
      "reverse_roll",
2577
      "take_along_axis",
2578 2579
      "tanh_shrink",
      "logsigmoid",
W
wenbin 已提交
2580
      "preln_layernorm_shift_partition",
2581
      "lookup_table",
W
wenbin 已提交
2582 2583
      "merge_layernorm",
      "skip_merge_layernorm",
2584
      "lookup_table_v2",
2585
      "expand_v2"};
W
wenbin 已提交
2586

W
weishengying 已提交
2587 2588 2589
  std::unordered_set<std::string> teller_set{
      "mul",
      "matmul",
2590
      "matmul_v2",
2591
      "bmm",
2592
      "range",
W
weishengying 已提交
2593 2594 2595 2596 2597 2598 2599 2600 2601 2602 2603 2604 2605 2606 2607 2608 2609 2610 2611 2612 2613 2614 2615
      "conv2d",
      "conv2d_fusion",
      "pool2d",
      "relu",
      "elu",
      "selu",
      "softsign",
      "softplus",
      "stanh",
      "thresholded_relu",
      "exp",
      "log",
      "sqrt",
      "abs",
      "sin",
      "cos",
      "tan",
      "sinh",
      "cosh",
      "asin",
      "acos",
      "atan",
      "asinh",
2616
      "acosh",
W
weishengying 已提交
2617 2618 2619
      "atanh",
      "ceil",
      "floor",
G
gem5 已提交
2620
      "rsqrt",
2621
      "sign",
G
gem5 已提交
2622
      "reciprocal",
2623
      "logical_not",
W
weishengying 已提交
2624
      "erf",
2625
      "square",
W
weishengying 已提交
2626 2627 2628 2629 2630 2631 2632 2633 2634 2635 2636 2637 2638
      "softmax",
      "sigmoid",
      "hard_swish",
      "depthwise_conv2d",
      "batch_norm",
      "concat",
      "tanh",
      "pad",
      "elementwise_add",
      "elementwise_sub",
      "elementwise_mul",
      "elementwise_div",
      "elementwise_pow",
2639 2640
      "elementwise_min",
      "elementwise_max",
W
wenbin 已提交
2641
      "elementwise_floordiv",
W
weishengying 已提交
2642
      "equal",
S
Sanbu 已提交
2643
      "not_equal",
2644 2645 2646 2647 2648 2649
      "less_than",
      "greater_than",
      "logical_or",
      "logical_xor",
      "logical_and",
      "less_equal",
W
weishengying 已提交
2650
      "dropout",
2651
      "fill_any_like",
W
weishengying 已提交
2652 2653 2654 2655 2656 2657
      "prelu",
      "conv2d_transpose",
      "depthwise_conv2d_transpose",
      "leaky_relu",
      "fc",
      "shuffle_channel",
2658
      "where",
2659 2660
      "one_hot",
      "one_hot_v2",
W
weishengying 已提交
2661 2662
      "swish",
      "silu",
2663
      "celu",
W
weishengying 已提交
2664 2665 2666 2667 2668 2669 2670 2671 2672 2673 2674 2675 2676 2677 2678 2679 2680
      "split",
      "instance_norm",
      "gelu",
      "layer_norm",
      "scale",
      "stack",
      "transpose2",
      "transpose",
      "top_k",
      "top_k_v2",
      "flatten2",
      "flatten",
      "gather",
      "gather_nd",
      "yolo_box",
      "yolo_box_head",
      "arg_max",
2681
      "arg_min",
W
weishengying 已提交
2682 2683 2684 2685
      "roi_align",
      "affine_channel",
      "nearest_interp",
      "anchor_generator",
2686
      "reduce_max",
W
weishengying 已提交
2687
      "reduce_mean",
2688
      "reduce_sum",
W
weishengying 已提交
2689 2690 2691 2692 2693 2694 2695 2696 2697 2698 2699 2700
      "conv3d",
      "conv3d_transpose",
      "mish",
      "bilinear_interp_v2",
      "nearest_interp_v2",
      "pool3d",
      "deformable_conv",
      "relu6",
      "hard_sigmoid",
      "clip",
      "fused_embedding_eltwise_layernorm",
      "multihead_matmul",
2701
      "multihead_matmul_roformer",
W
weishengying 已提交
2702 2703 2704 2705 2706 2707 2708 2709 2710 2711 2712 2713 2714 2715 2716 2717 2718 2719 2720 2721 2722
      "skip_layernorm",
      "slice",
      "strided_slice",
      "fused_preln_embedding_eltwise_layernorm",
      "preln_skip_layernorm",
      "preln_residual_bias",
      "c_allreduce_sum",
      "c_allreduce_min",
      "c_allreduce_max",
      "c_allreduce_prod",
      "roll",
      "cast",
      "transformer_input_convert",
      "recover_padding",
      "remove_padding",
      "fill_constant",
      "sum",
      "shape",
      "squeeze2",
      "unsqueeze2",
      "fused_token_prune",
2723
      "layernorm_shift_partition",
W
Wang Bojun 已提交
2724
      "reverse_roll",
2725
      "tanh_shrink",
2726
      "take_along_axis",
2727
      "logsigmoid",
W
wenbin 已提交
2728
      "preln_layernorm_shift_partition",
W
Wang Bojun 已提交
2729
      "merge_layernorm",
W
wenbin 已提交
2730
      "skip_merge_layernorm",
2731
      "lookup_table",
2732
      "lookup_table_v2",
2733
      "expand_v2"};
W
weishengying 已提交
2734 2735 2736 2737 2738 2739 2740 2741 2742 2743 2744 2745 2746
};

struct GenericPluginTeller : public Teller {
 public:
  GenericPluginTeller() {}
  bool operator()(const framework::OpDesc& desc,
                  bool use_no_calib_int8 = false,
                  bool with_dynamic_shape = false) override {
    const std::string op_type = desc.Type();
    // only consider dynamic_shape mode
    if (!with_dynamic_shape) {
      return false;
    }
2747 2748 2749 2750
    if (op_type == "yolo_box") {
      if (!desc.HasAttr("iou_aware") && !desc.HasAttr("iou_aware_factor"))
        return false;
    }
2751 2752 2753 2754 2755 2756 2757 2758
    if (op_type == "pad3d") {
      auto pad3d_inputs = desc.Inputs();
      if (pad3d_inputs.find("Paddings") != pad3d_inputs.end()) {
        if (desc.Input("Paddings").size() >= 1) {
          return false;
        }
      }
    }
W
weishengying 已提交
2759 2760 2761 2762 2763 2764 2765 2766 2767 2768 2769 2770 2771 2772 2773 2774 2775 2776 2777 2778 2779 2780 2781 2782 2783 2784 2785 2786 2787 2788 2789 2790 2791 2792 2793 2794 2795 2796 2797 2798 2799 2800 2801 2802 2803 2804 2805 2806 2807 2808 2809 2810 2811 2812 2813 2814 2815 2816
    if (use_no_calib_int8) {
      return false;
    } else {
      framework::InitDefaultKernelSignatureMap();
      bool res = phi::OpUtilsMap::Instance().HasArgumentMappingFn(op_type) ||
                 phi::DefaultKernelSignatureMap::Instance().Has(op_type);
      if (!res) {
        VLOG(3) << op_type << " has no KernelSignature";
        return false;
      }
      res = phi::KernelFactory::Instance().HasCompatiblePhiKernel(op_type);
      if (!res) {
        VLOG(3) << op_type << " has no CompatiblePhiKernel in phi.";
        return false;
      }
      auto& dynamic_infermeta_factory =
          tensorrt::DynamicMetaFnFactory::Instance();
      res = dynamic_infermeta_factory.Contains(op_type);
      if (!res) {
        VLOG(3) << op_type << " has no DynamicMetaFn.";
        return false;
      }
      return true;
    }
  }
};

struct CustomPluginTeller : public Teller {
 public:
  CustomPluginTeller() {}
  bool operator()(const framework::OpDesc& desc,
                  bool use_no_calib_int8 = false,
                  bool with_dynamic_shape = false) override {
    const std::string op_type = desc.Type();
    std::string expect_plugin_name;

    if (with_dynamic_shape) {
      expect_plugin_name = op_type + "_paddle_trt_dynamic_plugin";
    } else {
      expect_plugin_name = op_type + "_paddle_trt_plugin";
    }

    int num = 0;
    auto creators = GetPluginRegistry()->getPluginCreatorList(&num);

    for (int i = 0; i < num; i++) {
      if (std::string(creators[i]->getPluginName()) == expect_plugin_name)
        return true;
    }
    return false;
  }
};

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();
W
Wangzheee 已提交
2817 2818 2819 2820 2821 2822
  // do not support the op which is labeled the `skip_quant`
  if ((desc.HasAttr("namescope") &&
       PADDLE_GET_CONST(std::string, desc.GetAttr("op_namescope")) ==
           "/skip_quant_2/") ||
      desc.HasAttr("skip_quant"))
    return false;
W
weishengying 已提交
2823 2824 2825 2826 2827 2828 2829 2830 2831 2832 2833 2834 2835 2836 2837
  auto& default_teller = GetDefaultTeller();
  if ((*default_teller)(desc, use_no_calib_int8, with_dynamic_shape)) {
    SetOpConverterType(op_type, OpConverterType::Default);
    return true;
  }
  auto& generic_plugin_teller = GetGenericPluginTeller();
  if ((*generic_plugin_teller)(desc, use_no_calib_int8, with_dynamic_shape)) {
    SetOpConverterType(op_type, OpConverterType::GenericPluginCreater);
    return true;
  }
  auto& custom_plugin_teller = GetCustomPluginTeller();
  if ((*custom_plugin_teller)(desc, use_no_calib_int8, with_dynamic_shape)) {
    SetOpConverterType(op_type, OpConverterType::CustomPluginCreater);
    return true;
  }
2838 2839
  return false;
}
2840

W
weishengying 已提交
2841 2842 2843 2844 2845
OpTeller::OpTeller() {
  tellers_.emplace_back(new tensorrt::SimpleOpTypeSetTeller);
  tellers_.emplace_back(new tensorrt::GenericPluginTeller);
  tellers_.emplace_back(new tensorrt::CustomPluginTeller);
}
2846 2847 2848
}  // namespace tensorrt
}  // namespace inference
}  // namespace paddle