dense_tensor_kernels.cc 10.9 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14
// 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.

15
#include "paddle/infrt/kernel/phi/dense_tensor_kernels.h"
W
Wilber 已提交
16
#include "llvm/Support/ErrorHandling.h"
17
#include "paddle/infrt/common/string.h"
18 19
#include "paddle/infrt/dialect/phi/data_type.h"
#include "paddle/infrt/kernel/phi/context_kernels.h"
20 21
#include "paddle/infrt/paddle/model_parser.h"
#include "paddle/infrt/paddle/scope.h"
W
Wilber 已提交
22 23 24 25 26 27
#include "paddle/phi/backends/all_context.h"
#include "paddle/phi/common/place.h"

#ifdef INFRT_WITH_GPU
#include <cuda_runtime.h>
#endif
28

29 30 31 32 33 34 35 36 37 38 39 40
namespace paddle {
namespace platform {
using DeviceContext = ::phi::DeviceContext;
}  // namespace platform
namespace framework {
using LoDTensor = ::phi::DenseTensor;
void DeserializeFromStream(std::istream& is,
                           LoDTensor* tensor,
                           const platform::DeviceContext& dev_ctx);
}
}  // namespace paddle

41 42
namespace infrt {
namespace kernel {
43
namespace phi {
44

45 46
::phi::DenseTensor CreateDenseTensor(
    const ::phi::CPUContext& context,
47
    host_context::Attribute<std::vector<int64_t>> dims,
48
    host_context::Attribute<std::vector<int64_t>> lod,
49
    host_context::Attribute<::infrt::LayoutType> layout,
50 51 52
    host_context::Attribute<::infrt::PrecisionType> precision) {
  return ::phi::DenseTensor(
      const_cast<::phi::Allocator*>(&context.GetAllocator()),
53
      ::phi::DenseTensorMeta(ConvertPrecisionToPhi(precision.get()),
54
                             ::phi::make_ddim(dims.get()),
55
                             ConvertLayoutToPhi(layout.get()),
56
                             {}));
57 58
}

W
Wilber 已提交
59 60 61 62 63 64 65 66 67 68 69 70 71 72
::phi::DenseTensor CreateGPUDenseTensor(
    const ::phi::GPUContext& context,
    host_context::Attribute<std::vector<int64_t>> dims,
    host_context::Attribute<std::vector<int64_t>> lod,
    host_context::Attribute<::infrt::LayoutType> layout,
    host_context::Attribute<::infrt::PrecisionType> precision) {
  return ::phi::DenseTensor(
      const_cast<::phi::Allocator*>(&context.GetAllocator()),
      ::phi::DenseTensorMeta(ConvertPrecisionToPhi(precision.get()),
                             ::phi::make_ddim(dims.get()),
                             ConvertLayoutToPhi(layout.get()),
                             {}));
}

73
void FillDenseTensorF32(::phi::DenseTensor* dense_tensor,
74
                        host_context::Attribute<std::vector<float>> value) {
W
Wilber 已提交
75
  auto place = dense_tensor->place();
76
  float* a_data = dense_tensor->mutable_data<float>(place);
W
Wilber 已提交
77 78 79 80 81 82 83 84 85 86 87 88 89 90
  if (place.GetType() == ::phi::AllocationType::CPU) {
    for (int64_t i = 0; i < dense_tensor->numel(); ++i) {
      a_data[i] = (value.get())[i];
    }
  } else if (place.GetType() == ::phi::AllocationType::GPU) {
#ifdef INFRT_WITH_GPU
    // TODO(wilber): how to set the stream parameter to copy with stream.
    cudaMemcpy(a_data,
               value.get().data(),
               sizeof(float) * value.get().size(),
               cudaMemcpyHostToDevice);
#endif
  } else {
    llvm_unreachable("temporarily not support other target.");
91 92
  }
}
93

94
void PrintDenseTensor(::phi::DenseTensor* dense_tensor) {
W
Wilber 已提交
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
#ifndef INFRT_WITH_GPU
#define PRINT_META_DATA(PHI_DATATYPE, DTYPE)                \
  case ::phi::DataType::PHI_DATATYPE: {                     \
    auto place = dense_tensor->place();                     \
    if (place.GetType() == ::phi::AllocationType::CPU) {    \
      DTYPE* data = dense_tensor->data<DTYPE>();            \
      if (dense_tensor->numel() == 0) break;                \
      std::cout << data[0];                                 \
      for (int64_t i = 1; i < dense_tensor->numel(); i++) { \
        std::cout << "," << data[i];                        \
      }                                                     \
    }                                                       \
    break;                                                  \
  }
#else
#define PRINT_META_DATA(PHI_DATATYPE, DTYPE)                     \
  case ::phi::DataType::PHI_DATATYPE: {                          \
    auto place = dense_tensor->place();                          \
    DTYPE* data = dense_tensor->data<DTYPE>();                   \
    if (dense_tensor->numel() == 0) break;                       \
    if (place.GetType() == ::phi::AllocationType::CPU) {         \
      std::cout << data[0];                                      \
      for (int64_t i = 1; i < dense_tensor->numel(); i++) {      \
        std::cout << "," << data[i];                             \
      }                                                          \
    } else if (place.GetType() == ::phi::AllocationType::GPU) {  \
      std::vector<DTYPE> host_data(dense_tensor->numel(), 0);    \
      cudaMemcpy(host_data.data(),                               \
                 data,                                           \
                 sizeof(DTYPE) * dense_tensor->numel(),          \
                 cudaMemcpyDeviceToHost);                        \
      std::cout << host_data[0];                                 \
      for (int64_t i = 1; i < dense_tensor->numel(); i++) {      \
        std::cout << "," << host_data[i];                        \
      }                                                          \
    } else {                                                     \
      llvm_unreachable("temporarily not support other target."); \
    }                                                            \
    break;                                                       \
134
  }
W
Wilber 已提交
135
#endif
136 137 138

  ::phi::DDim dims = dense_tensor->dims();
  std::cout << "dense_tensor: shape=shape" << dims.to_str() << ","
139
            << " value=[";
140 141 142 143 144 145 146 147 148
  switch (dense_tensor->dtype()) {
    PRINT_META_DATA(FLOAT32, float);
    PRINT_META_DATA(INT32, int32_t);
    default:
      std::cout << "Error! Unsupported data type!\n";
  }
  std::cout << "]\n";
#undef PRINT_META_DATA
}
149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231

::infrt::phi::DenseTensorMap LoadParams(
    host_context::Attribute<std::string> path) {
  const auto& file_path = path.get();
  std::cout << "loading params from: " << file_path << std::endl;
  ::infrt::phi::DenseTensorMap map;

  const std::string model_path = file_path + "/__model__";
  auto pb_proto_prog = paddle::LoadProgram(model_path);
  auto main_block = pb_proto_prog->blocks(0);

  for (auto& var : main_block.vars()) {
    if (var.name() == "feed" || var.name() == "fetch" || !var.persistable())
      continue;
    std::string param_path = file_path + "/" + var.name();
    std::ifstream param_file(param_path, std::ios::binary);
    switch (var.type().type()) {
      case ::paddle::framework::proto::VarType_Type_LOD_TENSOR: {
        std::unique_ptr<::phi::DenseTensor> tensor{
            std::make_unique<::phi::DenseTensor>()};
        ::phi::CPUContext ctx;
        ::paddle::framework::DeserializeFromStream(
            param_file, tensor.get(), ctx);
        map.SetDenseTensor(var.name(), std::move(tensor));
      } break;
      default: {
        LOG(WARNING) << "Var `" << var.name() << "` type `"
                     << static_cast<int>(var.type().type())
                     << "` has not been supported now.";
      }
    }
  }
  return map;
}

::infrt::phi::DenseTensorMap LoadCombinedParams(
    host_context::Attribute<std::string> model_path,
    host_context::Attribute<std::string> params_path) {
  const auto& model = model_path.get();
  std::cout << "loading params from: " << model << std::endl;
  ::infrt::phi::DenseTensorMap map;

  auto pb_proto_prog = paddle::LoadProgram(model);
  auto main_block = pb_proto_prog->blocks(0);

  std::ifstream param_file(params_path.get(), std::ios::binary);

  std::set<std::string> tmp;
  for (auto& var : main_block.vars()) {
    if (var.name() == "feed" || var.name() == "fetch" || !var.persistable()) {
      continue;
    }
    if (var.type().type() ==
        ::paddle::framework::proto::VarType_Type_LOD_TENSOR) {
      tmp.emplace(var.name());
    } else {
      llvm_unreachable("the tensor type is illegal.");
    }
  }

  for (auto& var : tmp) {
    std::unique_ptr<::phi::DenseTensor> tensor{
        std::make_unique<::phi::DenseTensor>()};
    ::phi::CPUContext ctx;
    ::paddle::framework::DeserializeFromStream(param_file, tensor.get(), ctx);
    map.SetDenseTensor(var, std::move(tensor));
  }

  return map;
}

::phi::DenseTensor TensorMapGetTensor(
    const ::infrt::phi::DenseTensorMap& map,
    host_context::Attribute<std::string> name) {
  auto* tensor = map.GetDenseTensor(name.get());
  CHECK(tensor);
  return *tensor;
}

int32_t TensorMapGetSize(const ::infrt::phi::DenseTensorMap& map) {
  return map.size();
}

W
Wilber 已提交
232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294
#ifdef INFRT_WITH_GPU
inline size_t SizeOfDataType(::phi::DataType data_type) {
  switch (data_type) {
    case ::phi::DataType::BOOL:
    case ::phi::DataType::UINT8:
    case ::phi::DataType::INT8:
      return 1;
    case ::phi::DataType::BFLOAT16:
    case ::phi::DataType::FLOAT16:
    case ::phi::DataType::INT16:
    case ::phi::DataType::UINT16:
      return 2;
    case ::phi::DataType::FLOAT32:
    case ::phi::DataType::INT32:
    case ::phi::DataType::UINT32:
      return 4;
    case ::phi::DataType::FLOAT64:
    case ::phi::DataType::INT64:
    case ::phi::DataType::UINT64:
    case ::phi::DataType::COMPLEX64:
      return 8;
    case ::phi::DataType::COMPLEX128:
      return 16;
    case ::phi::DataType::UNDEFINED:
      return 0;
    default:
      llvm_unreachable("should not reach here");
      return 0;
  }
  return 0;
}
::phi::DenseTensor GpuMemCpy(const ::phi::DenseTensor& input,
                             const ::phi::GPUContext& context,
                             bool d2h) {
  if (d2h) {
    ::phi::DenseTensor ret(
        const_cast<::phi::Allocator*>(&context.GetHostAllocator()),
        input.meta());
    CHECK(input.place().GetType() == ::phi::AllocationType::GPU);
    // TODO(wilber): Add sync op and stream.
    cudaMemcpyAsync(ret.data(),
                    input.data(),
                    SizeOfDataType(input.dtype()) * input.numel(),
                    cudaMemcpyDeviceToHost,
                    nullptr);
    return ret;
  } else {
    // h2d
    ::phi::DenseTensor ret(
        const_cast<::phi::Allocator*>(&context.GetAllocator()), input.meta());
    CHECK(input.place().GetType() == ::phi::AllocationType::CPU ||
          input.place().GetType() == ::phi::AllocationType::GPUPINNED);
    // TODO(wilber): Add sync op and stream.
    cudaMemcpyAsync(ret.data(),
                    input.data(),
                    SizeOfDataType(input.dtype()) * input.numel(),
                    cudaMemcpyHostToDevice,
                    nullptr);
    return ret;
  }
}
#endif

295
}  // namespace phi
296 297
}  // namespace kernel
}  // namespace infrt