op_teller.cc 10.3 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
// Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
//     http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.

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

W
wanghuancoder 已提交
19 20 21 22 23 24
namespace paddle {
namespace framework {
class OpDesc;
}  // namespace framework
}  // namespace paddle

25 26 27 28 29 30
namespace paddle {
namespace inference {
namespace tensorrt {

// Just tell by the op_types.
struct SimpleOpTypeSetTeller : public Teller {
31 32 33
  SimpleOpTypeSetTeller() {
#if IS_TRT_VERSION_GE(5130)
    teller_set.insert("relu6");
34
    teller_set.insert("hard_sigmoid");
P
Pei Yang 已提交
35
    teller_set.insert("clip");
36 37
    int8_teller_set.insert("relu6");
    int8_teller_set.insert("hard_sigmoid");
P
Pei Yang 已提交
38
    int8_teller_set.insert("clip");
39 40 41 42 43
#endif
#if IS_TRT_VERSION_GE(6000)
    teller_set.insert("fused_embedding_eltwise_layernorm");
    teller_set.insert("multihead_matmul");
    teller_set.insert("skip_layernorm");
44
    teller_set.insert("slice");
45 46 47
#endif
#if IS_TRT_VERSION_GE(7130)
    teller_set.insert("group_norm");
48 49
#endif
  }
50

51 52 53 54 55 56 57
  bool operator()(const std::string& op_type, const framework::OpDesc& desc,
                  bool use_no_calib_int8) override {
    if (use_no_calib_int8) {
      return int8_teller_set.count(op_type);
    } else {
      return teller_set.count(op_type);
    }
58 59 60
  }

 private:
61
  // use this set for no calib int8.
62 63
  std::unordered_set<std::string> int8_teller_set{"mul",
                                                  "conv2d",
64
                                                  "conv2d_fusion",
65 66 67 68
                                                  "pool2d",
                                                  "relu",
                                                  "depthwise_conv2d",
                                                  "softmax",
69
                                                  "sigmoid",
70 71 72 73
                                                  "batch_norm",
                                                  "elementwise_add",
                                                  "leaky_relu",
                                                  "fc",
74 75 76
                                                  "concat",
                                                  "scale",
                                                  "elementwise_mul",
77 78
                                                  "conv2d_transpose",
                                                  "hard_swish"};
79
  std::unordered_set<std::string> teller_set{
80
      "mul",
81
      "matmul",
82
      "conv2d",
83
      "conv2d_fusion",
84 85 86 87
      "pool2d",
      "relu",
      "softmax",
      "sigmoid",
88
      "hard_swish",
89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106
      "depthwise_conv2d",
      "batch_norm",
      "concat",
      "tanh",
      "pad",
      "elementwise_add",
      "elementwise_mul",
      "dropout",
      "prelu",
      "conv2d_transpose",
      "leaky_relu",
      "fc",
      "shuffle_channel",
      "swish",
      "split",
      "instance_norm",
      "gelu",
      "layer_norm",
107
      "scale",
108
      "stack",
109 110 111 112
      "transpose2",
      "transpose",
      "flatten2",
      "flatten",
113
      "gather",
Z
zlsh80826 已提交
114
      "yolo_box",
115
      "roi_align",
116
      "affine_channel",
Z
zlsh80826 已提交
117
      "multiclass_nms",
118
      "nearest_interp",
119
      "anchor_generator",
120
  };
121 122
};

123 124 125 126
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();
127
  // do not support the op which is labeled the `skip_quant`
128
  if ((desc.HasAttr("namescope") &&
129
       BOOST_GET_CONST(std::string, desc.GetAttr("op_namescope")) ==
130 131
           "/skip_quant_2/") ||
      desc.HasAttr("skip_quant"))
132
    return false;
133

134
  for (auto& teller : tellers_) {
135 136 137
    if (op_type == "pool2d" || op_type == "conv2d" ||
        op_type == "depthwise_conv2d" || op_type == "conv2d_transpose") {
      std::vector<int> paddings =
138
          BOOST_GET_CONST(std::vector<int>, desc.GetAttr("paddings"));
139

140
      if (paddings.size() > 2) return false;
141
    }
142 143 144 145 146 147 148
    if (op_type == "matmul") {
      auto* block = desc.Block();
      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) {
P
Pei Yang 已提交
149 150 151
            VLOG(1)
                << "matmul op dims < 3 not supported in tensorrt, but got dims "
                << shape.size() << ", so jump it.";
152 153 154 155 156
            return false;
          }
        }
      }
    }
157
    if (op_type == "group_norm") {
158
      if (!with_dynamic_shape) return false;
159 160 161 162 163 164 165 166 167 168 169
      bool has_attrs = (desc.HasAttr("epsilon") && desc.HasAttr("groups"));
      if (has_attrs == false) return false;

      auto registry = GetPluginRegistry();
      if (registry == nullptr) return false;
    }
    if (op_type == "concat") {
      if (!desc.HasAttr("axis")) {
        return false;
      } else {
        int axis = BOOST_GET_CONST(int, desc.GetAttr("axis"));
170 171 172 173 174
        if (with_dynamic_shape) {
          if (axis < 0) return false;
        } else {
          if (axis <= 0) return false;
        }
175 176
      }
    }
177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196
    if (op_type == "transpose2" || op_type == "transpose") {
      if (!desc.HasAttr("axis")) {
        return false;
      } else {
        std::vector<int> axis =
            BOOST_GET_CONST(std::vector<int>, desc.GetAttr("axis"));
        if (!with_dynamic_shape && axis[0] != 0) return false;
        if (axis.size() >= nvinfer1::Dims::MAX_DIMS) return false;
      }
    }
    if (op_type == "flatten2" || op_type == "flatten") {
      // flatten doesn't support dynamic shape currently
      if (!desc.HasAttr("axis")) {
        return false;
      } else {
        if (with_dynamic_shape) return false;
        int axis = BOOST_GET_CONST(int, desc.GetAttr("axis"));
        if (axis != 1) return false;
      }
    }
197

198 199 200 201
    if (op_type == "gather") {
      // current not support axis from input, use default 0
      if (!with_dynamic_shape || desc.Input("Axis").size() > 0) return false;
    }
Z
zlsh80826 已提交
202

Z
zlsh80826 已提交
203 204 205 206 207 208 209 210 211
    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"));
      return has_attrs;
    }

212 213 214 215 216 217 218
    if (op_type == "affine_channel") {
      if (!desc.HasAttr("data_layout")) return false;
      auto data_layout = framework::StringToDataLayout(
          BOOST_GET_CONST(std::string, desc.GetAttr("data_layout")));
      if (data_layout != framework::DataLayout::kNCHW) return false;
    }

Z
zlsh80826 已提交
219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249
    if (op_type == "multiclass_nms") {
      if (with_dynamic_shape) return false;
      auto* block = desc.Block();
      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) {
            VLOG(1) << "multiclass_nms op dims != 3 not supported in tensorrt, "
                       "but got dims "
                    << shape.size() << ", so jump it.";
            return false;
          }
        }
      }
      bool has_attrs =
          (desc.HasAttr("background_label") &&
           desc.HasAttr("score_threshold") && desc.HasAttr("nms_top_k") &&
           desc.HasAttr("keep_top_k") && desc.HasAttr("normalized"));
      if (has_attrs == false) return false;

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

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

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

250 251 252 253 254 255 256 257 258 259 260
    if (op_type == "fc" || op_type == "mul") {
      const int x_num_col_dims =
          desc.HasAttr("x_num_col_dims")
              ? BOOST_GET_CONST(int, desc.GetAttr("x_num_col_dims"))
              : (desc.HasAttr("in_num_col_dims")
                     ? BOOST_GET_CONST(int, desc.GetAttr("in_num_col_dims"))
                     : 1);
      if (x_num_col_dims != 1 && x_num_col_dims != 2) {
        return false;
      }
    }
261

262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277
    if (op_type == "nearest_interp") {
      std::vector<std::string> attrs{"data_layout",   "interp_method",
                                     "align_corners", "scale",
                                     "out_h",         "out_w"};
      for (auto const attr : attrs) {
        if (!desc.HasAttr(attr)) return false;
      }
      auto data_layout = framework::StringToDataLayout(
          BOOST_GET_CONST(std::string, desc.GetAttr("data_layout")));
      if (data_layout != framework::DataLayout::kNCHW &&
          data_layout != framework::DataLayout::kNHWC)
        return false;
      auto interp_method =
          BOOST_GET_CONST(std::string, desc.GetAttr("interp_method"));
      if (interp_method != "nearest") return false;
    }
278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300

    if (op_type == "roi_align") {
      if (!with_dynamic_shape) return false;

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

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

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

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

301
    if ((*teller)(op_type, desc, use_no_calib_int8)) return true;
302 303 304 305 306 307 308 309 310
  }
  return false;
}

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

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