analysis_config.cc 11.6 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

namespace paddle {
24
extern const std::vector<std::string> kAnakinSubgraphPasses;
25

26
PassStrategy *AnalysisConfig::pass_builder() const {
27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42
  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.";
  }

43 44 45
  return pass_builder_.get();
}

46
AnalysisConfig::AnalysisConfig(const std::string &model_dir) {
47
  model_dir_ = model_dir;
Y
Yan Chunwei 已提交
48 49

  Update();
50
}
51 52
AnalysisConfig::AnalysisConfig(const std::string &prog_file,
                               const std::string &params_file) {
53 54
  prog_file_ = prog_file;
  params_file_ = params_file;
Y
Yan Chunwei 已提交
55 56

  Update();
57
}
58 59
void AnalysisConfig::SetModel(const std::string &prog_file_path,
                              const std::string &params_file_path) {
60 61
  prog_file_ = prog_file_path;
  params_file_ = params_file_path;
Y
Yan Chunwei 已提交
62 63

  Update();
64
}
65 66
void AnalysisConfig::EnableUseGpu(uint64_t memory_pool_init_size_mb,
                                  int device_id) {
67 68 69 70 71
#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
72
  LOG(ERROR) << "Please compile with gpu to EnableGpu()";
73 74
  use_gpu_ = false;
#endif
Y
Yan Chunwei 已提交
75 76 77

  Update();
}
78
void AnalysisConfig::DisableGpu() {
Y
Yan Chunwei 已提交
79 80 81
  use_gpu_ = false;

  Update();
82 83
}

84
AnalysisConfig::AnalysisConfig(const AnalysisConfig &other) {
85 86 87 88 89 90 91 92
#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
93
  // Gpu related.
94 95 96
  CP_MEMBER(use_gpu_);
  CP_MEMBER(device_id_);
  CP_MEMBER(memory_pool_init_size_mb_);
Y
Yan Chunwei 已提交
97 98

  CP_MEMBER(enable_memory_optim_);
Y
Yan Chunwei 已提交
99 100
  CP_MEMBER(static_memory_optim_);
  CP_MEMBER(static_memory_optim_force_update_);
S
Sylwester Fraczek 已提交
101
  // TensorRT related.
102 103 104 105
  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
106
  CP_MEMBER(tensorrt_precision_mode_);
N
nhzlx 已提交
107
  CP_MEMBER(trt_use_static_engine_);
S
Sylwester Fraczek 已提交
108
  // MKLDNN related.
109 110
  CP_MEMBER(use_mkldnn_);
  CP_MEMBER(mkldnn_enabled_op_types_);
111 112 113
  // Quantization related.
  CP_MEMBER(use_mkldnn_quantizer_);
  CP_MEMBER(mkldnn_quantizer_config_);
114

115 116
  CP_MEMBER(use_anakin_);
  CP_MEMBER(anakin_max_batchsize_);
117
  CP_MEMBER(anakin_max_input_shape_);
118
  CP_MEMBER(anakin_min_subgraph_size_);
119

120 121 122 123 124 125 126 127 128 129 130
  // 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_) {
131 132 133 134 135 136 137
    pass_builder_.reset(new GpuPassStrategy(
        *static_cast<GpuPassStrategy *>(other.pass_builder())));
  } else {
    pass_builder_.reset(new CpuPassStrategy(
        *static_cast<CpuPassStrategy *>(other.pass_builder())));
  }

138
#undef CP_MEMBER
Y
Yan Chunwei 已提交
139 140

  Update();
141 142
}

143
void AnalysisConfig::EnableMKLDNN() {
144 145 146 147 148 149 150
#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
151 152

  Update();
153 154
}

155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174
void AnalysisConfig::EnableMkldnnQuantizer() {
#ifdef PADDLE_WITH_MKLDNN
  if (!mkldnn_quantizer_config_)
    mkldnn_quantizer_config_.reset(new MkldnnQuantizerConfig());
  use_mkldnn_quantizer_ = true;
#else
  LOG(ERROR) << "Please compile with MKLDNN first to use MkldnnQuantizer";
  use_mkldnn_quantizer_ = false;
#endif

  Update();
}

std::shared_ptr<MkldnnQuantizerConfig> AnalysisConfig::mkldnn_quantizer_config()
    const {
  PADDLE_ENFORCE_NOT_NULL(mkldnn_quantizer_config_,
                          "MkldnnQuantizer was not enabled yet.");
  return mkldnn_quantizer_config_;
}

175
void AnalysisConfig::EnableTensorRtEngine(
N
nhzlx 已提交
176
    int workspace_size, int max_batch_size, int min_subgraph_size,
N
nhzlx 已提交
177
    AnalysisConfig::Precision precision_mode, bool use_static) {
Y
Yan Chunwei 已提交
178 179 180 181 182 183
#ifdef PADDLE_WITH_CUDA
  if (!use_gpu()) {
    LOG(ERROR) << "To use TensorRT engine, please call EnableGpu() first";
    return;
  }

184 185 186
  use_tensorrt_ = true;
  tensorrt_workspace_size_ = workspace_size;
  tensorrt_max_batchsize_ = max_batch_size;
N
nhzlx 已提交
187
  tensorrt_min_subgraph_size_ = min_subgraph_size;
N
nhzlx 已提交
188
  tensorrt_precision_mode_ = precision_mode;
N
nhzlx 已提交
189
  trt_use_static_engine_ = use_static;
Y
Yan Chunwei 已提交
190

191
  Update();
Y
Yan Chunwei 已提交
192 193 194 195
#else
  LOG(ERROR)
      << "To use TensorRT engine, please compile inference lib with GPU first.";
#endif
196 197
}

Y
Yan Chunwei 已提交
198
// TODO(Superjomn) refactor this, buggy.
199
void AnalysisConfig::Update() {
200 201 202
  auto info = SerializeInfoCache();
  if (info == serialized_info_cache_) return;

Y
Yan Chunwei 已提交
203 204 205 206 207 208 209 210 211 212 213 214 215
  // 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);
    }

216
  } else {
Y
Yan Chunwei 已提交
217 218 219 220 221 222 223 224
    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())));
    }
225 226 227
  }

  if (use_tensorrt_) {
Y
Yan Chunwei 已提交
228 229 230
    const auto &passes = pass_builder_->AllPasses();
    if (std::find(passes.begin(), passes.end(), "tensorrt_subgraph_pass") ==
        std::end(passes)) {
N
nhzlx 已提交
231 232
      // Append after the Affine_channel_conv_fuse pass.
      pass_builder()->InsertPass(3, "tensorrt_subgraph_pass");
233
    }
L
luotao1 已提交
234
    pass_builder()->DeletePass("runtime_context_cache_pass");
235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250
  }

  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
  }

251 252 253 254 255
  // Quantization passes must come after all other optimization passes
  if (use_mkldnn_quantizer_) {
    if (!enable_ir_optim_) {
      LOG(ERROR) << "EnableMkldnnQuantizer() only works when IR optimization "
                    "is enabled.";
256
    }
257 258 259 260 261 262 263 264 265 266 267 268 269 270 271
#ifdef PADDLE_WITH_MKLDNN
    pass_builder()->EnableMkldnnQuantizer();
#else
    LOG(ERROR) << "Please compile with MKLDNN first to use MkldnnQuantizer";
    use_mkldnn_quantizer_ = false;
#endif
  }

#ifdef PADDLE_WITH_MKLDNN
  // Do not optimize before quantization
  if (enable_memory_optim_ && !use_mkldnn_quantizer_) {
#else
  if (enable_memory_optim_) {
#endif
    pass_builder()->AppendAnalysisPass("memory_optimize_pass");
Y
Yan Chunwei 已提交
272 273
  }

274 275 276 277 278 279 280 281 282 283 284 285 286 287
  if (use_anakin_) {
    PADDLE_ENFORCE(!use_tensorrt_,
                   "Anakin sub-graph and TensorRT sub-graph are not allowed to "
                   "run at the same time!");
    PADDLE_ENFORCE(
        use_gpu_,
        "Anakin sub-graph engine need gpu, please use the EnableGpu API.");

    pass_builder()->ClearPasses();
    for (const auto &pass : kAnakinSubgraphPasses) {
      pass_builder()->AppendPass(pass);
    }
  }

288 289 290 291 292
  if (ir_debug_) {
    pass_builder()->TurnOnDebug();
  }
}

293
std::string AnalysisConfig::SerializeInfoCache() {
294
  std::stringstream ss;
Y
Yan Chunwei 已提交
295 296 297 298
  ss << model_dir_;
  ss << prog_file_;
  ss << params_file_;

299
  ss << use_gpu_;
Y
Yan Chunwei 已提交
300
  ss << device_id_;
301 302 303 304 305
  ss << memory_pool_init_size_mb_;

  ss << use_tensorrt_;
  ss << tensorrt_workspace_size_;
  ss << tensorrt_max_batchsize_;
Y
Yan Chunwei 已提交
306 307 308
  ss << tensorrt_min_subgraph_size_;

  ss << enable_memory_optim_;
Y
Yan Chunwei 已提交
309 310
  ss << static_memory_optim_;
  ss << static_memory_optim_force_update_;
311 312

  ss << use_mkldnn_;
Y
Yan Chunwei 已提交
313 314 315
  for (auto &item : mkldnn_enabled_op_types_) ss << item;
  ss << ";";

316
  ss << use_mkldnn_quantizer_;
Y
Yan Chunwei 已提交
317 318
  ss << model_from_memory_;

319 320 321 322
  ss << enable_ir_optim_;
  ss << use_feed_fetch_ops_;
  ss << ir_debug_;

Y
Yan Chunwei 已提交
323 324
  ss << specify_input_name_;
  ss << cpu_math_library_num_threads_;
325
  ss << use_anakin_;
326
  ss << anakin_min_subgraph_size_;
327 328 329
  return ss.str();
}

330
void AnalysisConfig::SetCpuMathLibraryNumThreads(
331 332
    int cpu_math_library_num_threads) {
  cpu_math_library_num_threads_ = cpu_math_library_num_threads;
Y
Yan Chunwei 已提交
333 334

  Update();
335 336
}

337
float AnalysisConfig::fraction_of_gpu_memory_for_pool() const {
338 339 340 341 342 343 344 345 346 347 348 349
#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
350 351
}

352 353
void AnalysisConfig::EnableMemoryOptim(bool static_optim,
                                       bool force_update_static_cache) {
Y
Yan Chunwei 已提交
354
  enable_memory_optim_ = true;
Y
Yan Chunwei 已提交
355 356
  static_memory_optim_ = static_optim;
  static_memory_optim_force_update_ = force_update_static_cache;
Y
Yan Chunwei 已提交
357 358 359 360

  Update();
}

361
bool AnalysisConfig::enable_memory_optim() const {
Y
Yan Chunwei 已提交
362 363 364
  return enable_memory_optim_;
}

365 366 367 368
void AnalysisConfig::SetModelBuffer(const char *prog_buffer,
                                    size_t prog_buffer_size,
                                    const char *param_buffer,
                                    size_t param_buffer_size) {
369 370
  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
371
  model_from_memory_ = true;
Y
Yan Chunwei 已提交
372 373

  Update();
T
Tao Luo 已提交
374 375
}

376 377 378 379 380
void AnalysisConfig::SetEngineOptInfo(
    std::map<std::string, std::string> engine_opt_info) {
  engine_opt_info_ = engine_opt_info;
}

381
NativeConfig AnalysisConfig::ToNativeConfig() const {
Y
Yan Chunwei 已提交
382 383 384 385 386 387 388 389 390 391 392
  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
393 394 395 396
void AnalysisConfig::SwitchIrDebug(int x) {
  ir_debug_ = x;
  Update();
}
397
void AnalysisConfig::EnableAnakinEngine(
398 399
    int max_batch_size, std::map<std::string, std::vector<int>> max_input_shape,
    int min_subgraph_size) {
400
  anakin_max_batchsize_ = max_batch_size;
401
  anakin_max_input_shape_ = max_input_shape;
402
  anakin_min_subgraph_size_ = min_subgraph_size;
403 404 405
  use_anakin_ = true;
  Update();
}
406
}  // namespace paddle