generic_plugin.cu 19.7 KB
Newer Older
W
weishengying 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218
// Copyright (c) 2022 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/plugin/generic_plugin.h"
#include "paddle/fluid/framework/framework.pb.h"
#include "paddle/fluid/framework/op_kernel_type.h"
#include "paddle/fluid/framework/phi_utils.h"
#include "paddle/fluid/inference/tensorrt/dynamic_shape_infermeta_registry.h"
#include "paddle/phi/backends/gpu/gpu_context.h"
#include "paddle/phi/core/compat/op_utils.h"
#include "paddle/phi/core/kernel_context.h"
#include "paddle/phi/core/kernel_factory.h"

namespace paddle {
namespace inference {
namespace tensorrt {
namespace plugin {

void BuildPhiKernelContextAttr(const framework::OpDesc& op_desc,
                               phi::KernelContext* kernel_context,
                               const phi::KernelSignature& signature,
                               const phi::Kernel& phi_kernel) {
  const phi::KernelArgsDef& args_def = phi_kernel.args_def();
  const auto& attr_names = signature.attr_names;
  const auto& attr_defs = args_def.attribute_defs();

  PADDLE_ENFORCE_EQ(
      attr_names.size(),
      attr_defs.size(),
      platform::errors::InvalidArgument(
          "The attr_names.size() should be equal to attr_defs.size()."));

  framework::AttrReader attr_reader(op_desc.GetAttrMap());

  for (size_t k = 0; k < attr_names.size(); ++k) {
    auto attr_name = attr_names[k];
    auto* attr_ptr = attr_reader.GetAttr(attr_name);
    if (attr_ptr) {
      switch (attr_defs[k].type_index) {
        case phi::AttributeType::SCALAR: {
          auto& attr = *attr_ptr;
          switch (AttrTypeID(attr)) {
            case framework::proto::AttrType::FLOAT:
              return kernel_context->EmplaceBackAttr(
                  phi::Scalar(PADDLE_GET_CONST(float, attr)));
              break;
            case framework::proto::AttrType::INT:
              return kernel_context->EmplaceBackAttr(
                  phi::Scalar(PADDLE_GET_CONST(int, attr)));
              break;
            case framework::proto::AttrType::STRING:
              return kernel_context->EmplaceBackAttr(
                  phi::Scalar(PADDLE_GET_CONST(std::string, attr)));
              break;
            default:
              PADDLE_THROW(platform::errors::Unimplemented(
                  "Unsupported cast op attribute `%s` to Scalar when "
                  "ProtoAttr2PhiAttr.",
                  attr_name));
          }
        } break;

        case phi::AttributeType::INT_ARRAY: {
          auto& attr = *attr_ptr;
          switch (AttrTypeID(attr)) {
            case framework::proto::AttrType::INTS:
              kernel_context->EmplaceBackAttr(std::move(
                  phi::IntArray(PADDLE_GET_CONST(std::vector<int32_t>, attr))));
              break;
            case framework::proto::AttrType::LONGS:
              kernel_context->EmplaceBackAttr(std::move(
                  phi::IntArray(PADDLE_GET_CONST(std::vector<int64_t>, attr))));
              break;
            case framework::proto::AttrType::INT:
              kernel_context->EmplaceBackAttr(
                  phi::IntArray({PADDLE_GET_CONST(int, attr)}));
              break;
            default:
              PADDLE_THROW(platform::errors::Unimplemented(
                  "Unsupported cast op attribute `%s` to IntArray when "
                  "ProtoAttr2PhiAttr.",
                  attr_name));
          }
        } break;

        case phi::AttributeType::SCALARS: {
          auto& attr = *attr_ptr;
          switch (AttrTypeID(attr)) {
            case framework::proto::AttrType::INTS: {
              const auto& vec = PADDLE_GET_CONST(std::vector<int32_t>, attr);
              std::vector<phi::Scalar> scalar_list;
              scalar_list.reserve(vec.size());
              for (const auto& val : vec) {
                scalar_list.emplace_back(val);
              }
              kernel_context->EmplaceBackAttr(std::move(scalar_list));
            } break;
            case framework::proto::AttrType::LONGS: {
              const auto& vec = PADDLE_GET_CONST(std::vector<int64_t>, attr);
              std::vector<phi::Scalar> scalar_list;
              scalar_list.reserve(vec.size());
              for (const auto& val : vec) {
                scalar_list.emplace_back(val);
              }
              kernel_context->EmplaceBackAttr(std::move(scalar_list));
            } break;
            case framework::proto::AttrType::FLOATS: {
              const auto& vec = PADDLE_GET_CONST(std::vector<float>, attr);
              std::vector<phi::Scalar> scalar_list;
              scalar_list.reserve(vec.size());
              for (const auto& val : vec) {
                scalar_list.emplace_back(val);
              }
              kernel_context->EmplaceBackAttr(std::move(scalar_list));
            } break;
            case framework::proto::AttrType::FLOAT64S: {
              const auto& vec = PADDLE_GET_CONST(std::vector<double>, attr);
              std::vector<phi::Scalar> scalar_list;
              scalar_list.reserve(vec.size());
              for (const auto& val : vec) {
                scalar_list.emplace_back(val);
              }
              kernel_context->EmplaceBackAttr(std::move(scalar_list));
            } break;
            default:
              PADDLE_THROW(platform::errors::Unimplemented(
                  "Unsupported cast op attribute `%s` to vector<Scalar> when "
                  "ProtoAttr2PhiAttr.",
                  attr_name));
          }
        } break;

        default: {
          auto& attr = *attr_ptr;
          switch (attr_defs[k].type_index) {
            case phi::AttributeType::FLOAT32:
              kernel_context->EmplaceBackAttr(PADDLE_GET_CONST(float, attr));
              break;
            case phi::AttributeType::INT32:
              kernel_context->EmplaceBackAttr(PADDLE_GET_CONST(int, attr));
              break;
            case phi::AttributeType::BOOL:
              kernel_context->EmplaceBackAttr(PADDLE_GET_CONST(bool, attr));
              break;
            case phi::AttributeType::INT64:
              kernel_context->EmplaceBackAttr(PADDLE_GET_CONST(int64_t, attr));
              break;
            case phi::AttributeType::INT32S:
              kernel_context->EmplaceBackAttr(
                  PADDLE_GET_CONST(std::vector<int>, attr));
              break;
            case phi::AttributeType::DATA_TYPE: {
              auto data_type = paddle::framework::TransToPhiDataType(
                  static_cast<framework::proto::VarType::Type>(
                      PADDLE_GET_CONST(int, attr)));
              kernel_context->EmplaceBackAttr(data_type);
            } break;
            case phi::AttributeType::STRING:
              kernel_context->EmplaceBackAttr(
                  PADDLE_GET_CONST(std::string, attr));
              break;
            case phi::AttributeType::INT64S:
              switch (AttrTypeID(attr)) {
                case framework::proto::AttrType::LONGS:
                  kernel_context->EmplaceBackAttr(
                      PADDLE_GET_CONST(std::vector<int64_t>, attr));
                  break;
                case framework::proto::AttrType::INTS: {
                  const auto& vector_int_attr =
                      PADDLE_GET_CONST(std::vector<int>, attr);
                  const std::vector<int64_t> vector_int64_attr(
                      vector_int_attr.begin(), vector_int_attr.end());
                  kernel_context->EmplaceBackAttr(vector_int64_attr);
                } break;
                default:
                  PADDLE_THROW(platform::errors::Unimplemented(
                      "Unsupported cast op attribute `%s` to vector<int64_t> "
                      "when ProtoAttr2PhiAttr.",
                      attr_name));
              }
              break;
            case phi::AttributeType::FLOAT32S:
              kernel_context->EmplaceBackAttr(
                  PADDLE_GET_CONST(std::vector<float>, attr));
              break;
            case phi::AttributeType::STRINGS:
              kernel_context->EmplaceBackAttr(
                  PADDLE_GET_CONST(std::vector<std::string>, attr));
              break;
            case phi::AttributeType::BOOLS:
              kernel_context->EmplaceBackAttr(
                  PADDLE_GET_CONST(std::vector<bool>, attr));
              break;
            case phi::AttributeType::FLOAT64S:
              kernel_context->EmplaceBackAttr(
                  PADDLE_GET_CONST(std::vector<double>, attr));
              break;
            default:
              PADDLE_THROW(platform::errors::Unimplemented(
                  "Unsupported cast op attribute `%s` when construct "
                  "ProtoAttr2PhiAttr.",
                  attr_name));
          }
        }
      }
    }
  }
219
  CHECK_EQ(attr_names.size(), kernel_context->AttrsSize());
W
weishengying 已提交
220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289
}

GenericPlugin::GenericPlugin(
    const paddle::framework::proto::OpDesc& proto_op_desc,
    const InputOutPutVarInfo& in_out_info) {
  proto_op_desc_ = proto_op_desc;
  op_desc_ = std::move(framework::OpDesc(proto_op_desc_, nullptr));
  proto_op_desc_.SerializeToString(&op_meta_data_);
  inputs_data_type_ = in_out_info.inputs_data_type;
  outputs_data_type_ = in_out_info.outputs_data_type;
}

GenericPlugin::GenericPlugin(
    const paddle::framework::proto::OpDesc& proto_op_desc,
    const std::vector<int>& inputs_data_type,
    const std::vector<int>& outputs_data_type) {
  proto_op_desc_ = proto_op_desc;
  op_desc_ = std::move(framework::OpDesc(proto_op_desc_, nullptr));
  proto_op_desc_.SerializeToString(&op_meta_data_);
  inputs_data_type_ = inputs_data_type;
  outputs_data_type_ = outputs_data_type;
}

GenericPlugin::GenericPlugin(void const* serial_data, size_t serial_length) {
  DeserializeValue(&serial_data, &serial_length, &inputs_data_type_);
  DeserializeValue(&serial_data, &serial_length, &outputs_data_type_);
  std::string op_meta_data((char*)(serial_data), serial_length);  // NOLINT
  op_meta_data_ = std::move(op_meta_data);
  proto_op_desc_.ParseFromString(op_meta_data_);
  op_desc_ = std::move(framework::OpDesc(proto_op_desc_, nullptr));
}

int GenericPlugin::getNbOutputs() const TRT_NOEXCEPT {
  int res = 0;
  for (auto& i : op_desc_.Outputs()) {
    if (!i.second.empty()) res += i.second.size();
  }
  return res;
}

int GenericPlugin::getNbInputs() const TRT_NOEXCEPT {
  int res = 0;
  for (auto& i : op_desc_.Inputs()) {
    if (!i.second.empty()) res += i.second.size();
  }
  return res;
}

nvinfer1::IPluginV2DynamicExt* GenericPlugin::clone() const TRT_NOEXCEPT {
  nvinfer1::IPluginV2DynamicExt* plugin =
      new GenericPlugin(proto_op_desc_, inputs_data_type_, outputs_data_type_);
  plugin->initialize();
  return plugin;
}

void GenericPlugin::serialize(void* buffer) const TRT_NOEXCEPT {
  // inputs_data_type_
  SerializeValue(&buffer, inputs_data_type_);
  // outputs_data_type_
  SerializeValue(&buffer, outputs_data_type_);
  // serialize op_meta_data_
  std::memcpy(buffer, op_meta_data_.c_str(), op_meta_data_.size());
  reinterpret_cast<char*&>(buffer) += op_meta_data_.size();
}

bool GenericPlugin::supportsFormatCombination(
    int pos,
    const nvinfer1::PluginTensorDesc* in_out,
    int nb_inputs,
    int nb_outputs) TRT_NOEXCEPT {
290
  if (op_desc_.Type() == "gather_nd" || op_desc_.Type() == "yolo_box") {
291 292 293 294 295 296 297 298 299
    if (pos == 0)
      return (in_out[pos].type == nvinfer1::DataType::kFLOAT) &&
             (in_out[pos].format == nvinfer1::TensorFormat::kLINEAR);
    if (pos == 1)
      return (in_out[pos].type == nvinfer1::DataType::kINT32) &&
             (in_out[pos].format == nvinfer1::TensorFormat::kLINEAR);
    if (pos == 2)
      return (in_out[pos].type == nvinfer1::DataType::kFLOAT) &&
             (in_out[pos].format == nvinfer1::TensorFormat::kLINEAR);
300
  } else if (op_desc_.Type() == "scatter_nd_add") {
301 302 303 304 305 306 307 308 309 310 311 312
    if (pos == 0)
      return (in_out[pos].type == nvinfer1::DataType::kFLOAT) &&
             (in_out[pos].format == nvinfer1::TensorFormat::kLINEAR);
    if (pos == 1)
      return (in_out[pos].type == nvinfer1::DataType::kINT32) &&
             (in_out[pos].format == nvinfer1::TensorFormat::kLINEAR);
    if (pos == 2)
      return (in_out[pos].type == nvinfer1::DataType::kFLOAT) &&
             (in_out[pos].format == nvinfer1::TensorFormat::kLINEAR);
    if (pos == 3)
      return (in_out[pos].type == nvinfer1::DataType::kFLOAT) &&
             (in_out[pos].format == nvinfer1::TensorFormat::kLINEAR);
313
  } else {
314 315
    return (in_out[pos].type == nvinfer1::DataType::kFLOAT) &&
           (in_out[pos].format == nvinfer1::TensorFormat::kLINEAR);
316
  }
W
weishengying 已提交
317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362
}

nvinfer1::DataType GenericPlugin::getOutputDataType(
    int index,
    const nvinfer1::DataType* input_types,
    int nb_inputs) const TRT_NOEXCEPT {
  return input_types[0];
}

int GenericPlugin::initialize() TRT_NOEXCEPT {
  std::string op_type = op_desc_.Type();

  phi::KernelSignature phi_kernel_signature;
  if (phi::OpUtilsMap::Instance().HasArgumentMappingFn(op_type)) {
    const phi::ArgumentMappingFn* argument_mapping_func =
        phi::OpUtilsMap::Instance().GetArgumentMappingFn(op_type);
    PluginArgumentMappingContext argument_mapping_context(&op_desc_);
    phi_kernel_signature = (*argument_mapping_func)(argument_mapping_context);
  } else {
    phi_kernel_signature =
        phi::DefaultKernelSignatureMap::Instance().Get(op_type);
  }

  phi::KernelKey phi_kernel_key(
      phi::Backend::GPU, phi::DataLayout::ANY, phi::DataType::FLOAT32);

  PADDLE_ENFORCE_EQ(
      phi::KernelFactory::Instance().HasCompatiblePhiKernel(op_type),
      true,
      platform::errors::Fatal("%s has no compatible phi kernel!",
                              op_type.c_str()));

  const phi::Kernel& phi_kernel = phi::KernelFactory::Instance().SelectKernel(
      phi_kernel_signature.name, phi_kernel_key);
  phi_kernel_ = &phi_kernel;

  PADDLE_ENFORCE_EQ(phi_kernel_->IsValid(),
                    true,
                    platform::errors::Fatal("%s phi kernel is invalid!.",
                                            phi_kernel_signature.name));

  paddle::platform::DeviceContextPool& pool =
      paddle::platform::DeviceContextPool::Instance();
  platform::CUDAPlace place(platform::GetCurrentDeviceId());
  auto* dev_ctx = static_cast<phi::GPUContext*>(pool.Get(place));

363 364 365 366 367 368 369 370 371
  if (!phi_kernel_context_) {
    phi_kernel_context_ = new phi::KernelContext(dev_ctx);
    BuildPhiKernelContextAttr(
        op_desc_, phi_kernel_context_, phi_kernel_signature, phi_kernel);
  }
  if (!dense_tensor_inputs_)
    dense_tensor_inputs_ = new std::vector<phi::DenseTensor>(getNbInputs());
  if (!dense_tensor_outputs_)
    dense_tensor_outputs_ = new std::vector<phi::DenseTensor>(getNbOutputs());
W
weishengying 已提交
372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420

  return 0;
}

nvinfer1::DimsExprs GenericPlugin::getOutputDimensions(
    int output_index,
    const nvinfer1::DimsExprs* inputs,
    int nb_inputs,
    nvinfer1::IExprBuilder& expr_builder) TRT_NOEXCEPT {
  CHECK(output_index < getNbOutputs());
  auto& dynamic_infermeta_factory = tensorrt::DynamicMetaFnFactory::Instance();
  PADDLE_ENFORCE_EQ(dynamic_infermeta_factory.Contains(op_desc_.Type()),
                    true,
                    platform::errors::InvalidArgument(
                        "The %s op has no dynamic plugin infershape function!",
                        op_desc_.Type().c_str()));

  auto* infershape_func = dynamic_infermeta_factory.Get(op_desc_.Type());
  return infershape_func(
      output_index, inputs, nb_inputs, expr_builder, op_desc_);
}

void GenericPlugin::configurePlugin(
    const nvinfer1::DynamicPluginTensorDesc* in,
    int nb_inputs,
    const nvinfer1::DynamicPluginTensorDesc* out,
    int nb_outputs) TRT_NOEXCEPT {
  CHECK(phi_kernel_context_);
  CHECK(phi_kernel_);
  CHECK(nb_inputs == getNbInputs());
  CHECK(nb_outputs == getNbOutputs());
}

// Shutdown the layer. This is called when the engine is destroyed
void GenericPlugin::terminate() TRT_NOEXCEPT {
  delete phi_kernel_context_;
  delete dense_tensor_inputs_;
  delete dense_tensor_outputs_;
}

int GenericPlugin::enqueue(const nvinfer1::PluginTensorDesc* input_desc,
                           const nvinfer1::PluginTensorDesc* output_desc,
                           const void* const* inputs,
                           void* const* outputs,
                           void* workspace,
                           cudaStream_t stream) TRT_NOEXCEPT {
  platform::CUDAPlace place(platform::GetCurrentDeviceId());

  // [TODO]now generic plugin do not support FP16 and INT8 precision
421
  auto protoType2PhiType = [](int proto_type) -> std::pair<phi::DataType, int> {
W
weishengying 已提交
422 423
    if (proto_type ==
        static_cast<int>(framework::proto::VarType_Type::VarType_Type_FP32))
424
      return {phi::DataType::FLOAT32, sizeof(float)};
W
weishengying 已提交
425 426 427 428 429 430
    else if (proto_type ==
                 static_cast<int>(
                     framework::proto::VarType_Type::VarType_Type_INT64) ||
             proto_type ==
                 static_cast<int>(
                     framework::proto::VarType_Type::VarType_Type_INT32))
431
      return {phi::DataType::INT32, sizeof(int32_t)};
W
weishengying 已提交
432 433 434
    else if (proto_type ==
             static_cast<int>(
                 framework::proto::VarType_Type::VarType_Type_BOOL))
435
      return {phi::DataType::BOOL, sizeof(bool)};
W
weishengying 已提交
436 437 438 439 440
    else
      CHECK(false) << "precision is not supported";
  };

  // input
441 442
  phi_kernel_context_->ClearInputOutput();

W
weishengying 已提交
443 444 445 446 447 448 449 450 451 452
  for (int i = 0; i < getNbInputs(); i++) {
    auto const& input_dims = input_desc[i].dims;

    std::vector<int> input_shape;
    for (int j = 0; j < input_dims.nbDims; j++)
      input_shape.push_back(input_dims.d[j]);

    int input_numel = 1;
    for (int k = 0; k < input_shape.size(); k++) input_numel *= input_shape[k];

453 454
    auto data_type_and_size = protoType2PhiType(inputs_data_type_[i]);
    phi::DenseTensorMeta input_meta(data_type_and_size.first,
W
weishengying 已提交
455 456 457
                                    phi::make_ddim(input_shape));
    std::shared_ptr<phi::Allocation> input_alloc(
        new phi::Allocation((void*)(inputs[i]),  // NOLINT
458
                            input_numel * data_type_and_size.second,
W
weishengying 已提交
459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476
                            place));
    (*dense_tensor_inputs_)[i] =
        std::move(phi::DenseTensor(input_alloc, input_meta));
    phi_kernel_context_->EmplaceBackInput(&((*dense_tensor_inputs_)[i]));
  }

  // output
  for (int i = 0; i < getNbOutputs(); i++) {
    auto const& output_dims = output_desc[i].dims;

    std::vector<int> output_shape;
    for (int j = 0; j < output_dims.nbDims; j++)
      output_shape.push_back(output_dims.d[j]);

    int output_numel = 1;
    for (int k = 0; k < output_shape.size(); k++)
      output_numel *= output_shape[k];

477 478
    auto data_type_and_size = protoType2PhiType(inputs_data_type_[i]);
    phi::DenseTensorMeta output_meta(data_type_and_size.first,
W
weishengying 已提交
479 480 481
                                     phi::make_ddim(output_shape));
    std::shared_ptr<phi::Allocation> output_alloc(
        new phi::Allocation(reinterpret_cast<void*>(outputs[i]),
482
                            output_numel * data_type_and_size.second,
W
weishengying 已提交
483 484 485 486 487 488 489
                            place));
    phi::DenseTensor output_densetonsor(output_alloc, output_meta);
    (*dense_tensor_outputs_)[i] =
        std::move(phi::DenseTensor(output_alloc, output_meta));
    phi_kernel_context_->EmplaceBackOutput(&((*dense_tensor_outputs_)[i]));
  }

490 491 492
  CHECK_EQ(phi_kernel_context_->InputsSize(), getNbInputs());
  CHECK_EQ(phi_kernel_context_->OutputsSize(), getNbOutputs());

W
weishengying 已提交
493 494 495 496 497 498 499 500 501
  (*phi_kernel_)(phi_kernel_context_);

  return cudaGetLastError() != cudaSuccess;
}

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