pten_utils.cc 9.2 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/convert_utils.h"
19
#include "paddle/pten/core/compat/op_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
OpKernelType TransPtenKernelKeyToOpKernelType(
    const pten::KernelKey& kernel_key) {
  proto::VarType::Type data_type =
      pten::TransToProtoVarType(kernel_key.dtype());
63 64
  // no need to set current device id here
  platform::Place place = pten::TransToFluidPlace(kernel_key.backend(), false);
65
  DataLayout data_layout = kernel_key.layout();
66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87
  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
  }
88
  paddle::experimental::DataLayout layout = kernel_type.data_layout_;
89 90 91 92 93
  paddle::experimental::DataType dtype =
      pten::TransToPtenDataType(kernel_type.data_type_);
  return pten::KernelKey(backend, layout, dtype);
}

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
pten::KernelKey FallBackToCpu(const OpKernelType& expected_kernel_key,
                              const pten::KernelKey& kernel_key,
                              const framework::OperatorBase& op) {
#ifdef PADDLE_WITH_XPU
  if (platform::is_xpu_place(expected_kernel_key.place_) ||
      paddle::platform::is_in_xpu_black_list(op.Type())) {
    VLOG(3) << "pten missing XPU kernel: " << op.Type()
            << ", expected_kernel_key:" << expected_kernel_key
            << ", fallbacking to CPU one!";
    return pten::KernelKey(pten::Backend::CPU, kernel_key.layout(),
                           kernel_key.dtype());
  }
#endif
#ifdef PADDLE_WITH_ASCEND_CL
  if (platform::is_npu_place(expected_kernel_key.place_)) {
    VLOG(3) << "pten missing NPU kernel: " << op.Type()
            << ", expected_kernel_key:" << expected_kernel_key
            << ", fallbacking to CPU one!";
    return pten::KernelKey(pten::Backend::CPU, kernel_key.layout(),
                           kernel_key.dtype());
  }
#endif
#ifdef PADDLE_WITH_MLU
  if (platform::is_mlu_place(expected_kernel_key.place_)) {
    VLOG(3) << "pten missing MLU kernel: " << op.Type()
            << ", expected_kernel_key:" << expected_kernel_key
            << ", fallbacking to CPU one!";
    return pten::KernelKey(pten::Backend::CPU, kernel_key.layout(),
                           kernel_key.dtype());
  }
#endif
  return pten::KernelKey();
}

128 129 130 131 132 133
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())) {
134
      VLOG(6) << "Parse PtenKernel input: skip extra & quant input - "
135 136 137 138 139 140
              << in_name;
      continue;
    }
    // If contains dispensable input, we should override the
    // GetExpectedPtenKernelArgs method self
    if (in.has_dispensable() && in.dispensable()) {
141
      VLOG(6) << "Parse PtenKernel input: skip dispensable input - " << in_name;
142 143
      continue;
    }
144
    VLOG(6) << "Parse PtenKernel input: " << in_name;
145 146 147 148 149 150 151 152 153 154 155
    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
156
    VLOG(6) << "Parse PtenKernel output: " << out_name;
157 158 159 160 161 162 163 164 165 166 167 168 169
    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") {
170
      VLOG(6) << "Parse PtenKernel attribute: skip needless attr - "
171 172 173 174 175
              << attr_name;
      continue;
    }
    if ((attr.has_extra() && attr.extra()) ||
        (attr.has_quant() && attr.quant())) {
176
      VLOG(6) << "Parse PtenKernel attribute: skip extra & quant attr - "
177 178 179
              << attr_name;
      continue;
    }
180
    VLOG(6) << "Parse PtenKernel attribute: " << attr_name;
181 182 183 184 185 186 187
    attr_names_.emplace_back(attr_name);
  }

  return attr_names_;
}

KernelSignature KernelArgsNameMakerByOpProto::GetKernelSignature() {
188 189
  return KernelSignature(op_proto_->type(), GetInputArgsNames(),
                         GetAttrsArgsNames(), GetOutputArgsNames());
190 191
}

192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209
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()));
      }
    }
  });
}

210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226
static void SetAllocationForUninitializedDenseTensor(
    pten::DenseTensor* dense_tensor, const platform::Place& place) {
  int dtype_size = dense_tensor->dtype() == DataType::UNDEFINED
                       ? 0
                       : experimental::SizeOf(dense_tensor->dtype());
  int64_t numels = product(dense_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);

  dense_tensor->ResetHolder(shared_allocation);
}

void SetAllocationForOutputTenosr(pten::TensorBase* tensor,
227
                                  const platform::Place& place) {
228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243
  if (pten::DenseTensor::classof(tensor)) {
    auto* dense_tensor = static_cast<pten::DenseTensor*>(tensor);
    if (!dense_tensor->IsInitialized() || !(dense_tensor->place() == place)) {
      SetAllocationForUninitializedDenseTensor(dense_tensor, place);
    }
  } else if (pten::SelectedRows::classof(tensor)) {
    auto* selected_rows = static_cast<pten::SelectedRows*>(tensor);
    if (!selected_rows->value().IsInitialized() ||
        !(selected_rows->place() == place)) {
      SetAllocationForUninitializedDenseTensor(selected_rows->mutable_value(),
                                               place);
    }
  } else {
    PADDLE_THROW(platform::errors::Unimplemented(
        "Unsupported tensor type is received when setting allocation for "
        "output tensor."));
244 245 246
  }
}

247 248
}  // namespace framework
}  // namespace paddle