analysis_config.cc 9.5 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
// Copyright (c) 2018 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/framework/lod_tensor.h"
#include "paddle/fluid/framework/scope.h"
17
#include "paddle/fluid/inference/api/paddle_analysis_config.h"
18
#include "paddle/fluid/inference/api/paddle_inference_api.h"
19
#include "paddle/fluid/inference/api/paddle_pass_builder.h"
20
#include "paddle/fluid/platform/enforce.h"
21
#include "paddle/fluid/platform/gpu_info.h"
22 23 24

namespace paddle {

25
PassStrategy *AnalysisConfig::pass_builder() const {
26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41
  if (!pass_builder_.get()) {
    if (use_gpu_) {
      LOG(INFO) << "Create GPU IR passes";
      pass_builder_.reset(new GpuPassStrategy);
    } else {
      LOG(INFO) << "Create CPU IR passes";
      pass_builder_.reset(new CpuPassStrategy);
    }
  } else if (pass_builder_->use_gpu() ^ use_gpu()) {
    LOG(WARNING) << "The use_gpu flag is not compatible between Config and "
                    "PassBuilder, the flags are "
                 << use_gpu() << " " << pass_builder_->use_gpu();
    LOG(WARNING) << "Please make them compatible, still use the existing "
                    "PassBuilder.";
  }

42 43 44
  return pass_builder_.get();
}

45
AnalysisConfig::AnalysisConfig(const std::string &model_dir) {
46
  model_dir_ = model_dir;
Y
Yan Chunwei 已提交
47 48

  Update();
49
}
50 51
AnalysisConfig::AnalysisConfig(const std::string &prog_file,
                               const std::string &params_file) {
52 53
  prog_file_ = prog_file;
  params_file_ = params_file;
Y
Yan Chunwei 已提交
54 55

  Update();
56
}
57 58
void AnalysisConfig::SetModel(const std::string &prog_file_path,
                              const std::string &params_file_path) {
59 60
  prog_file_ = prog_file_path;
  params_file_ = params_file_path;
Y
Yan Chunwei 已提交
61 62

  Update();
63
}
64 65
void AnalysisConfig::EnableUseGpu(uint64_t memory_pool_init_size_mb,
                                  int device_id) {
66 67 68 69 70
#ifdef PADDLE_WITH_CUDA
  use_gpu_ = true;
  memory_pool_init_size_mb_ = memory_pool_init_size_mb;
  device_id_ = device_id;
#else
Y
Yan Chunwei 已提交
71
  LOG(ERROR) << "Please compile with gpu to EnableGpu()";
72 73
  use_gpu_ = false;
#endif
Y
Yan Chunwei 已提交
74 75 76

  Update();
}
77
void AnalysisConfig::DisableGpu() {
Y
Yan Chunwei 已提交
78 79 80
  use_gpu_ = false;

  Update();
81 82
}

83
AnalysisConfig::AnalysisConfig(const AnalysisConfig &other) {
84 85 86 87 88 89 90 91
#define CP_MEMBER(member__) member__ = other.member__;

  // Model related.
  CP_MEMBER(model_dir_);
  CP_MEMBER(prog_file_);
  CP_MEMBER(params_file_);
  CP_MEMBER(model_from_memory_);  // the memory model reuses prog_file_ and
                                  // params_file_ fields.
S
Sylwester Fraczek 已提交
92
  // Gpu related.
93 94 95
  CP_MEMBER(use_gpu_);
  CP_MEMBER(device_id_);
  CP_MEMBER(memory_pool_init_size_mb_);
Y
Yan Chunwei 已提交
96 97

  CP_MEMBER(enable_memory_optim_);
Y
Yan Chunwei 已提交
98 99
  CP_MEMBER(static_memory_optim_);
  CP_MEMBER(static_memory_optim_force_update_);
S
Sylwester Fraczek 已提交
100
  // TensorRT related.
101 102 103 104
  CP_MEMBER(use_tensorrt_);
  CP_MEMBER(tensorrt_workspace_size_);
  CP_MEMBER(tensorrt_max_batchsize_);
  CP_MEMBER(tensorrt_min_subgraph_size_);
N
nhzlx 已提交
105
  CP_MEMBER(tensorrt_precision_mode_);
N
nhzlx 已提交
106
  CP_MEMBER(trt_use_static_engine_);
S
Sylwester Fraczek 已提交
107
  // MKLDNN related.
108 109 110 111 112 113 114 115 116 117 118 119 120 121
  CP_MEMBER(use_mkldnn_);
  CP_MEMBER(mkldnn_enabled_op_types_);

  // Ir related.
  CP_MEMBER(enable_ir_optim_);
  CP_MEMBER(use_feed_fetch_ops_);
  CP_MEMBER(ir_debug_);
  CP_MEMBER(specify_input_name_);

  CP_MEMBER(cpu_math_library_num_threads_);

  CP_MEMBER(serialized_info_cache_);

  if (use_gpu_) {
122 123 124 125 126 127 128
    pass_builder_.reset(new GpuPassStrategy(
        *static_cast<GpuPassStrategy *>(other.pass_builder())));
  } else {
    pass_builder_.reset(new CpuPassStrategy(
        *static_cast<CpuPassStrategy *>(other.pass_builder())));
  }

129
#undef CP_MEMBER
Y
Yan Chunwei 已提交
130 131

  Update();
132 133
}

134
void AnalysisConfig::EnableMKLDNN() {
135 136 137 138 139 140 141
#ifdef PADDLE_WITH_MKLDNN
  pass_builder()->EnableMKLDNN();
  use_mkldnn_ = true;
#else
  LOG(ERROR) << "Please compile with MKLDNN first to use MKLDNN";
  use_mkldnn_ = false;
#endif
Y
Yan Chunwei 已提交
142 143

  Update();
144 145
}

146
void AnalysisConfig::EnableTensorRtEngine(
N
nhzlx 已提交
147
    int workspace_size, int max_batch_size, int min_subgraph_size,
N
nhzlx 已提交
148
    AnalysisConfig::Precision precision_mode, bool use_static) {
Y
Yan Chunwei 已提交
149 150 151 152 153 154
#ifdef PADDLE_WITH_CUDA
  if (!use_gpu()) {
    LOG(ERROR) << "To use TensorRT engine, please call EnableGpu() first";
    return;
  }

155 156 157
  use_tensorrt_ = true;
  tensorrt_workspace_size_ = workspace_size;
  tensorrt_max_batchsize_ = max_batch_size;
N
nhzlx 已提交
158
  tensorrt_min_subgraph_size_ = min_subgraph_size;
N
nhzlx 已提交
159
  tensorrt_precision_mode_ = precision_mode;
N
nhzlx 已提交
160
  trt_use_static_engine_ = use_static;
Y
Yan Chunwei 已提交
161

162
  Update();
Y
Yan Chunwei 已提交
163 164 165 166
#else
  LOG(ERROR)
      << "To use TensorRT engine, please compile inference lib with GPU first.";
#endif
167 168
}

Y
Yan Chunwei 已提交
169
// TODO(Superjomn) refactor this, buggy.
170
void AnalysisConfig::Update() {
171 172 173
  auto info = SerializeInfoCache();
  if (info == serialized_info_cache_) return;

Y
Yan Chunwei 已提交
174 175 176 177 178 179 180 181 182 183 184 185 186
  // Transfer pass_builder and copy the existing compatible passes.
  if (!pass_builder_ || ((use_gpu() ^ pass_builder_->use_gpu()))) {
    if (use_gpu()) {
      pass_builder_.reset(new GpuPassStrategy);

      if (use_tensorrt_) {
        // Append after the Affine_channel_conv_fuse pass.
        pass_builder()->InsertPass(3, "tensorrt_subgraph_pass");
      }
    } else {
      pass_builder_.reset(new CpuPassStrategy);
    }

187
  } else {
Y
Yan Chunwei 已提交
188 189 190 191 192 193 194 195
    if (use_gpu()) {
      pass_builder_.reset(new GpuPassStrategy(
          *static_cast<GpuPassStrategy *>(pass_builder_.get())));

    } else {
      pass_builder_.reset(new CpuPassStrategy(
          *static_cast<CpuPassStrategy *>(pass_builder_.get())));
    }
196 197 198
  }

  if (use_tensorrt_) {
Y
Yan Chunwei 已提交
199 200 201
    const auto &passes = pass_builder_->AllPasses();
    if (std::find(passes.begin(), passes.end(), "tensorrt_subgraph_pass") ==
        std::end(passes)) {
N
nhzlx 已提交
202 203
      // Append after the Affine_channel_conv_fuse pass.
      pass_builder()->InsertPass(3, "tensorrt_subgraph_pass");
204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220
    }
  }

  if (use_mkldnn_) {
    if (!enable_ir_optim_) {
      LOG(ERROR)
          << "EnableMKLDNN() only works when IR optimization is enabled.";
    }
#ifdef PADDLE_WITH_MKLDNN
    pass_builder()->EnableMKLDNN();
    use_mkldnn_ = true;
#else
    LOG(ERROR) << "Please compile with MKLDNN first to use MKLDNN";
    use_mkldnn_ = false;
#endif
  }

Y
Yan Chunwei 已提交
221
  if (enable_memory_optim_) {
222 223 224 225 226 227 228 229
    auto analysis_passes = pass_builder()->AnalysisPasses();
    auto memory_opti_pass_name = "memory_optimize_pass";
    bool already_exists =
        std::find(analysis_passes.begin(), analysis_passes.end(),
                  memory_opti_pass_name) != analysis_passes.end();
    if (!already_exists) {
      pass_builder()->AppendAnalysisPass(memory_opti_pass_name);
    }
Y
Yan Chunwei 已提交
230 231
  }

232 233 234 235 236
  if (ir_debug_) {
    pass_builder()->TurnOnDebug();
  }
}

237
std::string AnalysisConfig::SerializeInfoCache() {
238
  std::stringstream ss;
Y
Yan Chunwei 已提交
239 240 241 242
  ss << model_dir_;
  ss << prog_file_;
  ss << params_file_;

243
  ss << use_gpu_;
Y
Yan Chunwei 已提交
244
  ss << device_id_;
245 246 247 248 249
  ss << memory_pool_init_size_mb_;

  ss << use_tensorrt_;
  ss << tensorrt_workspace_size_;
  ss << tensorrt_max_batchsize_;
Y
Yan Chunwei 已提交
250 251 252
  ss << tensorrt_min_subgraph_size_;

  ss << enable_memory_optim_;
Y
Yan Chunwei 已提交
253 254
  ss << static_memory_optim_;
  ss << static_memory_optim_force_update_;
255 256

  ss << use_mkldnn_;
Y
Yan Chunwei 已提交
257 258 259 260 261
  for (auto &item : mkldnn_enabled_op_types_) ss << item;
  ss << ";";

  ss << model_from_memory_;

262 263 264 265
  ss << enable_ir_optim_;
  ss << use_feed_fetch_ops_;
  ss << ir_debug_;

Y
Yan Chunwei 已提交
266 267 268
  ss << specify_input_name_;
  ss << cpu_math_library_num_threads_;

269 270 271
  return ss.str();
}

272
void AnalysisConfig::SetCpuMathLibraryNumThreads(
273 274
    int cpu_math_library_num_threads) {
  cpu_math_library_num_threads_ = cpu_math_library_num_threads;
Y
Yan Chunwei 已提交
275 276

  Update();
277 278
}

279
float AnalysisConfig::fraction_of_gpu_memory_for_pool() const {
280 281 282 283 284 285 286 287 288 289 290 291
#ifdef PADDLE_WITH_CUDA
  // Get the GPU memory details and calculate the fraction of memory for the
  // GPU memory pool.
  size_t gpu_used, gpu_available;
  platform::GpuMemoryUsage(&gpu_used, &gpu_available);
  double total_gpu_memory = (gpu_used + gpu_available) / 1024. / 1024.;
  float fraction_of_gpu_memory =
      static_cast<double>(memory_pool_init_size_mb()) / total_gpu_memory;
  return fraction_of_gpu_memory;
#else
  return 0.;
#endif
292 293
}

294 295
void AnalysisConfig::EnableMemoryOptim(bool static_optim,
                                       bool force_update_static_cache) {
Y
Yan Chunwei 已提交
296
  enable_memory_optim_ = true;
Y
Yan Chunwei 已提交
297 298
  static_memory_optim_ = static_optim;
  static_memory_optim_force_update_ = force_update_static_cache;
Y
Yan Chunwei 已提交
299 300 301 302

  Update();
}

303
bool AnalysisConfig::enable_memory_optim() const {
Y
Yan Chunwei 已提交
304 305 306
  return enable_memory_optim_;
}

307 308 309 310
void AnalysisConfig::SetModelBuffer(const char *prog_buffer,
                                    size_t prog_buffer_size,
                                    const char *param_buffer,
                                    size_t param_buffer_size) {
311 312
  prog_file_ = std::string(prog_buffer, prog_buffer + prog_buffer_size);
  params_file_ = std::string(param_buffer, param_buffer + param_buffer_size);
T
Tao Luo 已提交
313
  model_from_memory_ = true;
Y
Yan Chunwei 已提交
314 315

  Update();
T
Tao Luo 已提交
316 317
}

318
NativeConfig AnalysisConfig::ToNativeConfig() const {
Y
Yan Chunwei 已提交
319 320 321 322 323 324 325 326 327 328 329
  NativeConfig config;
  config.model_dir = model_dir_;
  config.prog_file = prog_file_;
  config.param_file = params_file_;
  config.use_gpu = use_gpu_;
  config.device = device_id_;
  config.fraction_of_gpu_memory = fraction_of_gpu_memory_for_pool();
  config.specify_input_name = specify_input_name_;
  return config;
}

Y
Yan Chunwei 已提交
330 331 332 333 334
void AnalysisConfig::SwitchIrDebug(int x) {
  ir_debug_ = x;
  Update();
}

335
}  // namespace paddle