pten_utils.cc 7.0 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17
/* Copyright (c) 2021 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 <sstream>

#include "paddle/fluid/framework/pten_utils.h"
18
#include "paddle/pten/core/compat/op_utils.h"
19
#include "paddle/pten/core/convert_utils.h"
Z
Zeng Jinle 已提交
20
#include "paddle/pten/core/kernel_factory.h"
21 22

#include "paddle/fluid/framework/lod_tensor.h"
Z
Zeng Jinle 已提交
23
#include "paddle/fluid/framework/op_info.h"
24
#include "paddle/fluid/framework/selected_rows_utils.h"
25 26 27 28 29 30
#include "paddle/fluid/framework/variable.h"
#include "paddle/fluid/string/string_helper.h"

namespace paddle {
namespace framework {

Z
Zeng Jinle 已提交
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
class KernelArgsNameMakerByOpProto : public KernelArgsNameMaker {
 public:
  explicit KernelArgsNameMakerByOpProto(
      const framework::proto::OpProto* op_proto)
      : op_proto_(op_proto) {
    PADDLE_ENFORCE_NOT_NULL(op_proto_, platform::errors::InvalidArgument(
                                           "Op proto cannot be nullptr."));
  }

  ~KernelArgsNameMakerByOpProto() {}

  const paddle::SmallVector<std::string>& GetInputArgsNames() override;
  const paddle::SmallVector<std::string>& GetOutputArgsNames() override;
  const paddle::SmallVector<std::string>& GetAttrsArgsNames() override;

  KernelSignature GetKernelSignature();

 private:
  DISABLE_COPY_AND_ASSIGN(KernelArgsNameMakerByOpProto);

 private:
  const framework::proto::OpProto* op_proto_;

  paddle::SmallVector<std::string> input_names_;
  paddle::SmallVector<std::string> output_names_;
  paddle::SmallVector<std::string> attr_names_;
};

59 60 61 62 63
OpKernelType TransPtenKernelKeyToOpKernelType(
    const pten::KernelKey& kernel_key) {
  proto::VarType::Type data_type =
      pten::TransToProtoVarType(kernel_key.dtype());
  platform::Place place = pten::TransToFluidPlace(kernel_key.backend());
64
  DataLayout data_layout = kernel_key.layout();
65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86
  LibraryType library_type = LibraryType::kPlain;
  if (kernel_key.backend() == pten::Backend::MKLDNN) {
    library_type = LibraryType::kMKLDNN;
  } else if (kernel_key.backend() == pten::Backend::CUDNN) {
    library_type = LibraryType::kCUDNN;
  } else {
    // do nothing
  }
  // TODO(chenweihang): the customized_type_value is lost
  return OpKernelType(data_type, place, data_layout, library_type);
}

pten::KernelKey TransOpKernelTypeToPtenKernelKey(
    const OpKernelType& kernel_type) {
  pten::Backend backend = pten::TransToPtenBackend(kernel_type.place_);
  if (kernel_type.library_type_ == LibraryType::kMKLDNN) {
    backend = pten::Backend::MKLDNN;
  } else if (kernel_type.library_type_ == LibraryType::kCUDNN) {
    backend = pten::Backend::CUDNN;
  } else {
    // do
  }
87
  paddle::experimental::DataLayout layout = kernel_type.data_layout_;
88 89 90 91 92 93 94 95 96 97 98
  paddle::experimental::DataType dtype =
      pten::TransToPtenDataType(kernel_type.data_type_);
  return pten::KernelKey(backend, layout, dtype);
}

const paddle::SmallVector<std::string>&
KernelArgsNameMakerByOpProto::GetInputArgsNames() {
  for (int i = 0; i < op_proto_->inputs_size(); ++i) {
    auto& in = op_proto_->inputs()[i];
    auto& in_name = in.name();
    if ((in.has_extra() && in.extra()) || (in.has_quant() && in.quant())) {
99
      VLOG(6) << "Parse PtenKernel input: skip extra & quant input - "
100 101 102 103 104 105
              << in_name;
      continue;
    }
    // If contains dispensable input, we should override the
    // GetExpectedPtenKernelArgs method self
    if (in.has_dispensable() && in.dispensable()) {
106
      VLOG(6) << "Parse PtenKernel input: skip dispensable input - " << in_name;
107 108
      continue;
    }
109
    VLOG(6) << "Parse PtenKernel input: " << in_name;
110 111 112 113 114 115 116 117 118 119 120
    input_names_.emplace_back(in_name);
  }
  return input_names_;
}

const paddle::SmallVector<std::string>&
KernelArgsNameMakerByOpProto::GetOutputArgsNames() {
  for (int i = 0; i < op_proto_->outputs_size(); ++i) {
    auto& out = op_proto_->outputs()[i];
    auto& out_name = out.name();
    // TODO(chenweihang): outputs also need skip some cases
121
    VLOG(6) << "Parse PtenKernel output: " << out_name;
122 123 124 125 126 127 128 129 130 131 132 133 134
    output_names_.emplace_back(out_name);
  }
  return output_names_;
}

const paddle::SmallVector<std::string>&
KernelArgsNameMakerByOpProto::GetAttrsArgsNames() {
  for (int i = 0; i < op_proto_->attrs_size(); ++i) {
    auto& attr = op_proto_->attrs()[i];
    auto& attr_name = attr.name();
    if (attr_name == "use_mkldnn" || attr_name == "op_role" ||
        attr_name == "op_role_var" || attr_name == "op_namescope" ||
        attr_name == "op_callstack" || attr_name == "op_device") {
135
      VLOG(6) << "Parse PtenKernel attribute: skip needless attr - "
136 137 138 139 140
              << attr_name;
      continue;
    }
    if ((attr.has_extra() && attr.extra()) ||
        (attr.has_quant() && attr.quant())) {
141
      VLOG(6) << "Parse PtenKernel attribute: skip extra & quant attr - "
142 143 144
              << attr_name;
      continue;
    }
145
    VLOG(6) << "Parse PtenKernel attribute: " << attr_name;
146 147 148 149 150 151 152
    attr_names_.emplace_back(attr_name);
  }

  return attr_names_;
}

KernelSignature KernelArgsNameMakerByOpProto::GetKernelSignature() {
153 154 155
  return KernelSignature(pten::TransToPtenKernelName(op_proto_->type()),
                         GetInputArgsNames(), GetAttrsArgsNames(),
                         GetOutputArgsNames());
156 157
}

158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175
std::once_flag kernel_sig_map_init_flag;

void InitDefaultKernelSignatureMap() {
  std::call_once(kernel_sig_map_init_flag, [] {
    for (const auto& pair : paddle::framework::OpInfoMap::Instance().map()) {
      const auto& op_type = pair.first;
      const auto* op_proto = pair.second.proto_;
      if (pten::KernelFactory::Instance().HasCompatiblePtenKernel(op_type) &&
          op_proto) {
        paddle::framework::KernelArgsNameMakerByOpProto maker(op_proto);
        VLOG(10) << "Register kernel signature for " << op_type;
        pten::DefaultKernelSignatureMap::Instance().Insert(
            op_type, std::move(maker.GetKernelSignature()));
      }
    }
  });
}

176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193
void SetAllocationForOutputTenosr(pten::DenseTensor* tensor,
                                  const platform::Place& place) {
  if (!tensor->IsInitialized() || !(tensor->place() == place)) {
    int dtype_size = tensor->dtype() == DataType::UNDEFINED
                         ? 0
                         : experimental::SizeOf(tensor->dtype());
    int64_t numels = product(tensor->dims());
    numels = numels < 0 ? 0 : numels;
    auto tmp_allocation_ptr = memory::Alloc(place, numels * dtype_size);
    auto& deleter = tmp_allocation_ptr.get_deleter();
    auto* allocation_ptr = tmp_allocation_ptr.release();
    auto shared_allocation =
        std::shared_ptr<pten::Allocation>(allocation_ptr, deleter);

    tensor->ResetHolder(shared_allocation);
  }
}

194 195
}  // namespace framework
}  // namespace paddle