analysis_config.cc 13.1 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> kTRTSubgraphPasses;
石晓伟 已提交
25
extern const std::vector<std::string> kLiteSubgraphPasses;
26

27
PassStrategy *AnalysisConfig::pass_builder() const {
28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43
  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.";
  }

44 45 46
  return pass_builder_.get();
}

47
AnalysisConfig::AnalysisConfig(const std::string &model_dir) {
48
  model_dir_ = model_dir;
Y
Yan Chunwei 已提交
49 50

  Update();
51
}
52 53
AnalysisConfig::AnalysisConfig(const std::string &prog_file,
                               const std::string &params_file) {
54 55
  prog_file_ = prog_file;
  params_file_ = params_file;
Y
Yan Chunwei 已提交
56 57

  Update();
58
}
59 60
void AnalysisConfig::SetModel(const std::string &prog_file_path,
                              const std::string &params_file_path) {
61 62
  prog_file_ = prog_file_path;
  params_file_ = params_file_path;
Y
Yan Chunwei 已提交
63 64

  Update();
65
}
66 67
void AnalysisConfig::EnableUseGpu(uint64_t memory_pool_init_size_mb,
                                  int device_id) {
68 69 70 71 72
#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
73
  LOG(ERROR) << "Please compile with gpu to EnableGpu()";
74 75
  use_gpu_ = false;
#endif
Y
Yan Chunwei 已提交
76 77 78

  Update();
}
79
void AnalysisConfig::DisableGpu() {
Y
Yan Chunwei 已提交
80 81 82
  use_gpu_ = false;

  Update();
83 84
}

85 86 87 88 89 90
void AnalysisConfig::DisableFCPadding() {
  use_fc_padding_ = false;

  Update();
}

91
AnalysisConfig::AnalysisConfig(const AnalysisConfig &other) {
92 93 94 95 96 97
#define CP_MEMBER(member__) member__ = other.member__;

  // Model related.
  CP_MEMBER(model_dir_);
  CP_MEMBER(model_from_memory_);  // the memory model reuses prog_file_ and
                                  // params_file_ fields.
98

99
  CP_MEMBER(opt_cache_dir_);
100 101 102
  prog_file_ = std::move(other.prog_file_);
  params_file_ = std::move(other.params_file_);

103
  CP_MEMBER(use_fc_padding_);
104
  // GPU related.
105
  CP_MEMBER(use_gpu_);
106
  CP_MEMBER(use_cudnn_);
107 108
  CP_MEMBER(device_id_);
  CP_MEMBER(memory_pool_init_size_mb_);
Y
Yan Chunwei 已提交
109 110

  CP_MEMBER(enable_memory_optim_);
S
Sylwester Fraczek 已提交
111
  // TensorRT related.
112 113 114 115
  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
116
  CP_MEMBER(tensorrt_precision_mode_);
N
nhzlx 已提交
117
  CP_MEMBER(trt_use_static_engine_);
118
  CP_MEMBER(trt_use_calib_mode_);
S
Sylwester Fraczek 已提交
119
  // MKLDNN related.
120 121
  CP_MEMBER(use_mkldnn_);
  CP_MEMBER(mkldnn_enabled_op_types_);
122
  CP_MEMBER(mkldnn_cache_capacity_);
123 124 125
  // Quantization related.
  CP_MEMBER(use_mkldnn_quantizer_);
  CP_MEMBER(mkldnn_quantizer_config_);
126 127 128
  CP_MEMBER(min_input_shape_);
  CP_MEMBER(max_input_shape_);
  CP_MEMBER(optim_input_shape_);
129
  CP_MEMBER(disable_trt_plugin_fp16_);
130

石晓伟 已提交
131 132 133 134 135
  CP_MEMBER(use_lite_);
  CP_MEMBER(lite_precision_mode_);
  CP_MEMBER(lite_passes_filter_);
  CP_MEMBER(lite_ops_filter_);

136 137 138
  // profile related.
  CP_MEMBER(with_profile_);

139 140 141
  // glog related.
  CP_MEMBER(with_glog_info_);

142 143 144 145 146 147 148 149 150 151 152
  // 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_) {
153 154 155 156 157 158 159
    pass_builder_.reset(new GpuPassStrategy(
        *static_cast<GpuPassStrategy *>(other.pass_builder())));
  } else {
    pass_builder_.reset(new CpuPassStrategy(
        *static_cast<CpuPassStrategy *>(other.pass_builder())));
  }

160
#undef CP_MEMBER
Y
Yan Chunwei 已提交
161 162

  Update();
163 164
}

165 166 167 168 169 170 171 172 173 174 175
void AnalysisConfig::EnableCUDNN() {
#ifdef PADDLE_WITH_CUDA
  use_cudnn_ = use_gpu_;
#else
  LOG(ERROR) << "Please compile with CUDA first to use cuDNN";
  use_cudnn_ = false;
#endif

  Update();
}

176
void AnalysisConfig::EnableMKLDNN() {
177 178 179 180 181 182
#ifdef PADDLE_WITH_MKLDNN
  use_mkldnn_ = true;
#else
  LOG(ERROR) << "Please compile with MKLDNN first to use MKLDNN";
  use_mkldnn_ = false;
#endif
Y
Yan Chunwei 已提交
183 184

  Update();
185 186
}

187 188 189 190 191 192 193 194 195
void AnalysisConfig::SetMkldnnCacheCapacity(int capacity) {
#ifdef PADDLE_WITH_MKLDNN
  mkldnn_cache_capacity_ = capacity;
#else
  LOG(ERROR) << "Please compile with MKLDNN first to set MKLDNN Thread Id";
  mkldnn_cache_capacity_ = 0;
#endif
}

196 197 198 199 200 201 202 203 204 205 206 207 208
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();
}

209
MkldnnQuantizerConfig *AnalysisConfig::mkldnn_quantizer_config() const {
210 211
  PADDLE_ENFORCE_NOT_NULL(mkldnn_quantizer_config_,
                          "MkldnnQuantizer was not enabled yet.");
212
  return mkldnn_quantizer_config_.get();
213 214
}

215
void AnalysisConfig::EnableTensorRtEngine(
N
nhzlx 已提交
216
    int workspace_size, int max_batch_size, int min_subgraph_size,
217
    AnalysisConfig::Precision precision_mode, bool use_static,
218
    bool use_calib_mode) {
Y
Yan Chunwei 已提交
219 220 221 222 223 224
#ifdef PADDLE_WITH_CUDA
  if (!use_gpu()) {
    LOG(ERROR) << "To use TensorRT engine, please call EnableGpu() first";
    return;
  }

225 226 227
  use_tensorrt_ = true;
  tensorrt_workspace_size_ = workspace_size;
  tensorrt_max_batchsize_ = max_batch_size;
N
nhzlx 已提交
228
  tensorrt_min_subgraph_size_ = min_subgraph_size;
N
nhzlx 已提交
229
  tensorrt_precision_mode_ = precision_mode;
N
nhzlx 已提交
230
  trt_use_static_engine_ = use_static;
231
  trt_use_calib_mode_ = use_calib_mode;
Y
Yan Chunwei 已提交
232

233
  Update();
Y
Yan Chunwei 已提交
234 235 236 237
#else
  LOG(ERROR)
      << "To use TensorRT engine, please compile inference lib with GPU first.";
#endif
238 239
}

240 241 242 243 244 245 246 247 248 249 250
void AnalysisConfig::SetTRTDynamicShapeInfo(
    std::map<std::string, std::vector<int>> min_input_shape,
    std::map<std::string, std::vector<int>> max_input_shape,
    std::map<std::string, std::vector<int>> optim_input_shape,
    bool disable_trt_plugin_fp16) {
  min_input_shape_ = min_input_shape;
  max_input_shape_ = max_input_shape;
  optim_input_shape_ = optim_input_shape;
  disable_trt_plugin_fp16_ = disable_trt_plugin_fp16;
}

Y
Yan Chunwei 已提交
251
// TODO(Superjomn) refactor this, buggy.
252
void AnalysisConfig::Update() {
253 254 255
  auto info = SerializeInfoCache();
  if (info == serialized_info_cache_) return;

Y
Yan Chunwei 已提交
256 257 258 259 260 261 262 263 264 265 266 267
  // 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);
    }
268

269
  } else {
Y
Yan Chunwei 已提交
270 271 272 273 274 275 276 277
    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())));
    }
278 279 280
  }

  if (use_tensorrt_) {
281
    pass_builder()->ClearPasses();
282 283 284
    bool use_calib_int8 =
        (tensorrt_precision_mode_ == AnalysisConfig::Precision::kInt8) &&
        trt_use_calib_mode_;
285
    for (const auto &pass : kTRTSubgraphPasses) {
286 287 288 289
      if (use_calib_int8 &&
          (pass == "conv_bn_fuse_pass" || pass == "fc_fuse_pass")) {
        continue;
      }
290
      pass_builder()->AppendPass(pass);
291 292
    }
  }
293 294 295 296 297 298 299 300 301 302
  if (use_gpu() && use_cudnn_) {
#ifdef PADDLE_WITH_CUDA
    if (!enable_ir_optim_) {
      LOG(ERROR) << "EnableCUDNN() only works when IR optimization is enabled.";
    } else {
      pass_builder()->EnableCUDNN();
    }
#endif
  }

303
  if (use_mkldnn_) {
W
Wojciech Uss 已提交
304
#ifdef PADDLE_WITH_MKLDNN
305 306 307
    if (!enable_ir_optim_) {
      LOG(ERROR)
          << "EnableMKLDNN() only works when IR optimization is enabled.";
W
Wojciech Uss 已提交
308 309
    } else {
      pass_builder()->EnableMKLDNN();
310 311 312 313
    }
#endif
  }

314 315 316 317 318
  // 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.";
319 320
    }
#ifdef PADDLE_WITH_MKLDNN
321
    pass_builder()->EnableMkldnnQuantizer();
322 323 324
#endif
  }

325
#ifdef PADDLE_WITH_MKLDNN
326 327
  // Do not optimize when mkldnn is on
  if (enable_memory_optim_ && !use_mkldnn_) {
328
#else
Y
Yan Chunwei 已提交
329
  if (enable_memory_optim_) {
330 331
#endif
    pass_builder()->AppendAnalysisPass("memory_optimize_pass");
Y
Yan Chunwei 已提交
332 333
  }

石晓伟 已提交
334 335 336 337 338 339 340 341 342 343 344 345 346 347
  if (use_lite_) {
#ifndef PADDLE_WITH_LITE
    LOG(WARNING) << "You tried to enable the lite subgraph "
                    "but did not have the option -DWITH_LITE compiled.";
#endif
    pass_builder()->ClearPasses();
    for (const auto &pass : kLiteSubgraphPasses) {
      if (std::find(lite_passes_filter_.begin(), lite_passes_filter_.end(),
                    pass) == lite_passes_filter_.end()) {
        pass_builder()->AppendPass(pass);
      }
    }
  }

348 349 350 351 352
  if (ir_debug_) {
    pass_builder()->TurnOnDebug();
  }
}

353
std::string AnalysisConfig::SerializeInfoCache() {
354
  std::stringstream ss;
Y
Yan Chunwei 已提交
355 356 357 358
  ss << model_dir_;
  ss << prog_file_;
  ss << params_file_;

359
  ss << use_gpu_;
360
  ss << use_fc_padding_;
Y
Yan Chunwei 已提交
361
  ss << device_id_;
362 363 364 365 366
  ss << memory_pool_init_size_mb_;

  ss << use_tensorrt_;
  ss << tensorrt_workspace_size_;
  ss << tensorrt_max_batchsize_;
Y
Yan Chunwei 已提交
367 368 369
  ss << tensorrt_min_subgraph_size_;

  ss << enable_memory_optim_;
370 371

  ss << use_mkldnn_;
372
  ss << mkldnn_cache_capacity_;
Y
Yan Chunwei 已提交
373 374 375
  for (auto &item : mkldnn_enabled_op_types_) ss << item;
  ss << ";";

376
  ss << use_mkldnn_quantizer_;
Y
Yan Chunwei 已提交
377 378
  ss << model_from_memory_;

379 380
  ss << with_profile_;

381 382
  ss << with_glog_info_;

383 384 385 386
  ss << enable_ir_optim_;
  ss << use_feed_fetch_ops_;
  ss << ir_debug_;

Y
Yan Chunwei 已提交
387 388
  ss << specify_input_name_;
  ss << cpu_math_library_num_threads_;
石晓伟 已提交
389 390

  ss << use_lite_;
391

392 393 394
  return ss.str();
}

395
void AnalysisConfig::SetCpuMathLibraryNumThreads(
396 397
    int cpu_math_library_num_threads) {
  cpu_math_library_num_threads_ = cpu_math_library_num_threads;
Y
Yan Chunwei 已提交
398 399

  Update();
400 401
}

402
float AnalysisConfig::fraction_of_gpu_memory_for_pool() const {
403 404 405 406
#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;
407
  platform::SetDeviceId(device_id_);
408 409 410 411 412 413 414 415
  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
416 417
}

418
void AnalysisConfig::EnableMemoryOptim() {
Y
Yan Chunwei 已提交
419 420 421 422
  enable_memory_optim_ = true;
  Update();
}

423
bool AnalysisConfig::enable_memory_optim() const {
Y
Yan Chunwei 已提交
424 425 426
  return enable_memory_optim_;
}

427 428 429 430
void AnalysisConfig::SetModelBuffer(const char *prog_buffer,
                                    size_t prog_buffer_size,
                                    const char *param_buffer,
                                    size_t param_buffer_size) {
431 432
  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
433
  model_from_memory_ = true;
Y
Yan Chunwei 已提交
434 435

  Update();
T
Tao Luo 已提交
436 437
}

438
NativeConfig AnalysisConfig::ToNativeConfig() const {
Y
Yan Chunwei 已提交
439 440 441 442 443 444 445 446 447 448 449
  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
450 451 452 453
void AnalysisConfig::SwitchIrDebug(int x) {
  ir_debug_ = x;
  Update();
}
454 455 456 457 458 459

void AnalysisConfig::EnableProfile() {
  with_profile_ = true;
  Update();
}

460 461 462 463 464
void AnalysisConfig::DisableGlogInfo() {
  with_glog_info_ = false;
  Update();
}

石晓伟 已提交
465 466 467 468 469 470 471 472 473 474 475
void AnalysisConfig::EnableLiteEngine(
    AnalysisConfig::Precision precision_mode,
    const std::vector<std::string> &passes_filter,
    const std::vector<std::string> &ops_filter) {
  use_lite_ = true;
  lite_precision_mode_ = precision_mode;
  lite_passes_filter_ = passes_filter;
  lite_ops_filter_ = ops_filter;
  Update();
}

476 477 478 479 480 481 482
void AnalysisConfig::PartiallyRelease() {
  prog_file_.clear();
  prog_file_.shrink_to_fit();
  params_file_.clear();
  params_file_.shrink_to_fit();
}

483
}  // namespace paddle