executor.cpp

/* 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 "io/executor.h"
#include <operators/math/gemm.h>
#include <algorithm>
#include <vector>
#include "common/enforce.h"
#include "common/log.h"
#include "framework/framework.pb-c.h"
#include "framework/lod_tensor.h"
#include "framework/operator.h"
#include "framework/program/program-optimize/program_optimize.h"
#include "framework/program/program_desc.h"
#include "framework/program/var_desc.h"
#include "framework/scope.h"
#include "framework/tensor.h"
#ifdef PADDLE_EXECUTOR_MULTITHREAD
#include <queue>
#include <utility>
#include "common/threadpool.h"
#endif

#ifdef PADDLE_MOBILE_FPGA
#include "fpga/fpga_quantilization.h"
#endif

namespace paddle_mobile {
using framework::Variable;

char *Get_binary_data(std::string filename) {
  FILE *file = fopen(filename.c_str(), "rb");
  PADDLE_MOBILE_ENFORCE(file != nullptr, "can't open file: %s ",
                        filename.c_str());
  fseek(file, 0, SEEK_END);
  int64_t size = ftell(file);
  PADDLE_MOBILE_ENFORCE(size > 0, "size is too small");
  rewind(file);
  char *data = new char[size];
  size_t bytes_read = fread(data, 1, size, file);
  PADDLE_MOBILE_ENFORCE(bytes_read == size,
                        "read binary file bytes do not match with fseek");
  fclose(file);
  return data;
}

#pragma mark - executor
template <typename Dtype, Precision P>
Executor<Dtype, P>::Executor(const framework::Program<Dtype> p, int batch_size,
                             bool use_optimize)
    : program_(p), batch_size_(batch_size), use_optimize_(use_optimize) {
  if (use_optimize_) {
    to_predict_program_ = program_.optimizeProgram;
  } else {
    to_predict_program_ = program_.originProgram;
  }
  Variable *variable_ptr = program_.scope->Var("batch_size");
  variable_ptr[0].SetValue<int>(batch_size);
  const std::vector<std::shared_ptr<framework::BlockDesc>> blocks =
      to_predict_program_->Blocks();
#ifdef PADDLE_EXECUTOR_MULTITHREAD
  depManager.resize(blocks.size());
#endif
  for (int i = 0; i < blocks.size(); ++i) {
    std::shared_ptr<framework::BlockDesc> block_desc = blocks[i];
    std::vector<std::shared_ptr<framework::OpDesc>> ops = block_desc->Ops();
    for (int j = 0; j < ops.size(); ++j) {
      std::shared_ptr<framework::OpDesc> op = ops[j];
      DLOG << "create op: " << op->Type();
      auto op_base = framework::OpRegistry<Dtype>::CreateOp(
          op->Type(), op->GetInputs(), op->GetOutputs(), op->GetAttrMap(),
          program_.scope);
      op_base->InferShape();
      ops_of_block_[*block_desc.get()].push_back(op_base);
#ifdef PADDLE_EXECUTOR_MULTITHREAD
      depManager[i].analysisDep(ops_of_block_[*block_desc.get()]);
#endif
    }
  }
  if (program_.combined) {
    InitCombineMemory();
  } else {
    InitMemory();
  }

  std::shared_ptr<framework::BlockDesc> to_predict_block =
      to_predict_program_->Block(0);
  auto &ops = ops_of_block_[*to_predict_block.get()];
  for (const auto &op : ops) {
    op->Init();
  }
#ifdef PADDLE_MOBILE_FPGA
  for (const auto &op : ops) {
    quantilize_op(op, program_.scope);
  }
#endif
}

template <typename Dtype, Precision P>
void Executor<Dtype, P>::LoadMemory(const framework::VarDesc var_desc,
                                    framework::LoDTensor *tensor, char **data) {
  // 1. version
  uint32_t version = *reinterpret_cast<uint32_t *>(*data);

  (*data) += sizeof(uint32_t);

  // 2 Lod information
  uint64_t *lod_level_ptr = new uint64_t();
  memcpy(lod_level_ptr, (*data), sizeof(uint64_t));
  uint64_t lod_level = *lod_level_ptr;
  delete lod_level_ptr;
  (*data) += sizeof(uint64_t);

  auto &lod = *tensor->mutable_lod();
  lod.resize(lod_level);
  for (uint64_t i = 0; i < lod_level; ++i) {
    uint64_t size = *reinterpret_cast<uint64_t *>(*data);
    (*data) += sizeof(uint64_t);
    DLOG << "lod size: " << i << size;
    std::vector<size_t> tmp(size / sizeof(size_t));

    for (int k = 0; k < tmp.size(); ++k) {
      tmp[k] = *reinterpret_cast<size_t *>(*data);
      (*data) += sizeof(size_t);
    }

    for (auto j : tmp) {
      LOG(kLOG_DEBUG1) << "    lod - " << j;
    }
    lod[i] = tmp;
  }

  // 3. tensor version
  uint32_t tensor_version = *reinterpret_cast<uint32_t *>(*data);
  (*data) += sizeof(uint32_t);

  // 4. tensor desc
  int32_t size = *reinterpret_cast<int32_t *>(*data);
  (*data) += sizeof(int32_t);

  std::unique_ptr<char[]> buf(new char[size]);
  for (int m = 0; m < size; ++m) {
    buf.get()[m] = (*data)[m];
  }
  (*data) += (sizeof(char) * size);

  const framework::TensorDesc &desc = var_desc.Tensor_desc();
  int memory_size = 1;
  for (auto l : desc.Dims()) {
    memory_size *= l;
  }

  tensor->Resize(framework::make_ddim(desc.Dims()));

  void *memory = nullptr;
  int type_size = 0;
  switch (desc.DataType()) {
    case framework::VARTYPE_TYPE_FP16:
      type_size = 2;
      break;
    case framework::VARTYPE_TYPE_FP32:
      type_size = 4;
      memory = tensor->mutable_data<float>();
      break;
    case framework::VARTYPE_TYPE_FP64:
      type_size = 8;
      break;
    case framework::VARTYPE_TYPE_INT32:
      type_size = 4;
      break;
    case framework::VARTYPE_TYPE_INT64:
      type_size = 8;
      break;
    case framework::VARTYPE_TYPE_BOOL:
      type_size = 1;
      break;
    default:
      break;
  }
  if (program_.quantification) {
    float min_value;
    float max_value;

    memcpy(&min_value, *data, sizeof(float));
    memcpy(&max_value, *data + sizeof(float), sizeof(float));
    *data += 2 * sizeof(float);
    const float factor = (max_value - min_value) / 255.0;
    uint8_t *uint8_data = reinterpret_cast<uint8_t *>(*data);
    for (int k = 0; k < memory_size; ++k) {
      static_cast<float *>(memory)[k] = uint8_data[k] * factor + min_value;
    }
    *data += (memory_size * sizeof(uint8_t));
  } else {
    for (int n = 0; n < memory_size * type_size; ++n) {
      static_cast<char *>(memory)[n] = (*data)[n];
    }
    (*data) += (sizeof(char) * memory_size * type_size);
  }
}

template <typename Dtype, Precision P>
void Executor<Dtype, P>::InitMemory() {
  for (const auto &block : to_predict_program_->Blocks()) {
    for (const auto &var_desc : block->Vars()) {
      auto var = program_.scope->Var(var_desc->Name());
      if (var_desc->Persistable()) {
        auto tensor = var->template GetMutable<framework::LoDTensor>();
        if (var_desc->Name() == "feed" || var_desc->Name() == "fetch") {
          continue;
        }

        char *origin_data =
            Get_binary_data(program_.model_path + "/" + var_desc->Name());
        char *data = origin_data;
        LoadMemory(*var_desc, tensor, &data);
        delete origin_data;
      } else {
        if (var_desc->Type() == framework::VARTYPE_TYPE_LOD_TENSOR) {
          auto tensor = var->template GetMutable<framework::LoDTensor>();

          tensor->template mutable_data<Ptype>();
        }
      }
    }
  }
}

template <typename Dtype, Precision P>
void Executor<Dtype, P>::InitCombineMemory() {
  LOG(kLOG_INFO) << " begin init combine memory";
  char *origin_data = Get_binary_data(program_.para_path);
  char *data = origin_data;
  for (const auto &block : to_predict_program_->Blocks()) {
    for (const auto &var_desc : block->Vars()) {
      auto var = program_.scope->Var(var_desc->Name());
      if (var_desc->Persistable()) {
        auto tensor = var->template GetMutable<framework::LoDTensor>();
        if (var_desc->Name() == "feed" || var_desc->Name() == "fetch") {
          continue;
        }
        LoadMemory(*var_desc, tensor, &data);
      } else {
        if (var_desc->Type() == framework::VARTYPE_TYPE_LOD_TENSOR) {
          auto tensor = var->template GetMutable<framework::LoDTensor>();
          tensor->template mutable_data<Ptype>();
        }
      }
    }
  }
  delete origin_data;
  LOG(kLOG_INFO) << " end init combine memory ";
}

template <typename Dtype, Precision P>
std::shared_ptr<framework::Tensor> Executor<Dtype, P>::Predict(
    const framework::Tensor &t) {
  framework::Variable *g_feed_value = program_.scope->Var("feed");
  framework::Tensor *feed_tensor =
      g_feed_value->GetMutable<framework::LoDTensor>();
  feed_tensor->Resize(t.dims());
  feed_tensor->ShareDataWith(t);
  std::shared_ptr<framework::BlockDesc> to_predict_block =
      to_predict_program_->Block(0);
  auto &ops = ops_of_block_[*to_predict_block.get()];
#ifdef PADDLE_MOBILE_PROFILE
  std::vector<ProfInfo> profile(ops.size());
#endif
#ifdef PADDLE_EXECUTOR_MULTITHREAD
  std::mutex m;
  std::condition_variable cv;
  std::queue<int> next;
  next.push(0);
  int rsize = ops.size();
  std::vector<int> status(rsize, 0);
  auto &threadPool = ThreadPool::getThreadPool();
  auto &dep = depManager[0];
  auto finishF = [&ops, &m, &cv, &next, &status, &rsize, &dep](int opi) {
    std::lock_guard<std::mutex> lk(m);
    rsize--;
    status[opi] = 2;
    for (int i : dep.getNext(opi)) {
      bool ok = true;
      for (int j : dep.getDeps(i)) {
        if (status[j] != 2) {
          ok = false;
          break;
        }
      }
      if (ok && (status[i] == 0)) {
        next.push(i);
      }
    }
    cv.notify_one();
  };
  for (;;) {
    std::unique_lock<std::mutex> lk(m);
    cv.wait(lk, [&next, &rsize] { return rsize == 0 || !next.empty(); });
    if (rsize == 0) {
      break;
    }
    while (next.size() > 0) {
      int opi = next.front();
      next.pop();
      status[opi] = 1;
      threadPool.enqueue([opi, &ops, &finishF, &profile] {
        auto &op = ops[opi];
#ifdef PADDLE_MOBILE_PROFILE
        struct timespec ts;
        clock_gettime(CLOCK_MONOTONIC, &ts);
        profile[opi].runBegin = (uint64_t)ts.tv_sec * 1e9 + ts.tv_nsec;
        profile[opi].tid = ThreadPool::getThreadPoolThreadId();
#endif
        ops[opi]->Run();
#ifdef PADDLE_MOBILE_PROFILE
        clock_gettime(CLOCK_MONOTONIC, &ts);
        profile[opi].runEnd = (uint64_t)ts.tv_sec * 1e9 + ts.tv_nsec;
#endif
        finishF(opi);
      });
    }
  }
#else
  for (int i = 0; i < ops.size(); i++) {
#ifdef PADDLE_MOBILE_PROFILE
    struct timespec ts;
    clock_gettime(CLOCK_MONOTONIC, &ts);
    profile[i].runBegin = (uint64_t)ts.tv_sec * 1e9 + ts.tv_nsec;
#endif

    // to Run
    ops[i]->Run();
#ifdef PADDLE_MOBILE_PROFILE
    clock_gettime(CLOCK_MONOTONIC, &ts);
    profile[i].runEnd = (uint64_t)ts.tv_sec * 1e9 + ts.tv_nsec;
#endif
  }
#endif
  auto last_op = ops.rbegin();

  auto output_map = (*last_op)->Outputs();
  std::vector<std::string> out_keys = (*last_op)->GetOutKeys();
  PADDLE_MOBILE_ENFORCE(out_keys.size() > 0, "the last op contains no output");
  framework::LoDTensor *output_tensor =
      framework::GetVarValue<framework::LoDTensor>(out_keys[0], output_map,
                                                   *(program_.scope));
#ifdef PADDLE_MOBILE_PROFILE
#ifdef PADDLE_EXECUTOR_MULTITHREAD
  // TODO(haipeng): expose profile info as an interface, user can get them to
  // analysis
  //      the performance of their deepnet.
  FILE *df = fopen("net.dot", "w");
  fprintf(df, "digraph {\n");
  for (int i = 0; i < ops.size(); i++) {
    for (int j : dep.getNext(i)) {
      fprintf(df, "op_%d -> op_%d\n", i, j);
    }
  }
  for (int i = 0; i < ops.size(); i++) {
    fprintf(df, "op_%d[label=\"%s (%d)\"]\n", i, ops[i]->Type().c_str(), i);
  }
  fprintf(df, "}\n");
  fclose(df);
#endif

  //  FILE *pf = fopen("profile.out", "w");
  std::unordered_map<std::string, uint64_t> _tp;
  for (int i = 0; i < profile.size(); i++) {
    const auto &pInfo = profile[i];
    uint64_t timeCost = pInfo.runEnd - pInfo.runBegin;
    _tp[ops[i]->Type()] += timeCost;
    //    fprintf(pf, "%d\t%s\t%d\t%llu\t%llu\t%llu\n", i,
    //    ops[i]->Type().c_str(),
    //            pInfo.tid, pInfo.runBegin, pInfo.runEnd, timeCost);
  }
  //  fclose(pf);

  printf("====================[ profile ]======================\n");
  using prof_t = std::pair<std::string, uint64_t>;
  std::vector<prof_t> _tv(_tp.begin(), _tp.end());
  uint64_t _ptotal = 0;
  for (auto const &p : _tv) {
    _ptotal += p.second;
  }
  auto compf = [](const prof_t &a, const prof_t &b) {
    return a.second > b.second;
  };
  std::sort(_tv.begin(), _tv.end(), compf);
  _tv.push_back(std::make_pair("total", _ptotal));
  for (auto const &p : _tv) {
    printf("%-16s\t%-10.0f\t%-2.4f\n", p.first.c_str(),
           static_cast<float>(p.second),
           static_cast<float>(p.second) / _ptotal * 100.0);
  }
  printf("====================[---------]======================\n");
#endif

  return std::make_shared<framework::Tensor>(framework::Tensor(*output_tensor));
}
template <typename Dtype, Precision P>
std::shared_ptr<framework::Tensor> Executor<Dtype, P>::Predict(
    const framework::Tensor &t, int block_id) {
  return Predict(t);
}

template <typename Dtype, Precision P>
std::vector<typename Executor<Dtype, P>::Ptype> Executor<Dtype, P>::Predict(
    const std::vector<Ptype> &input, const std::vector<int64_t> &dims) {
  framework::Tensor tensor(input, framework::make_ddim(dims));
  std::shared_ptr<framework::Tensor> output_tensor = Predict(tensor, 0);
  Executor<Dtype, P>::Ptype *output_ptr =
      output_tensor->data<typename Executor<Dtype, P>::Ptype>();
  std::vector<typename Executor<Dtype, P>::Ptype> result_vector;
  for (int j = 0; j < output_tensor->numel(); ++j) {
    result_vector.push_back(output_ptr[j]);
  }
  return result_vector;
}

template class Executor<CPU, Precision::FP32>;
template class Executor<GPU_MALI, Precision::FP32>;
template class Executor<FPGA, Precision::FP32>;

}  // namespace paddle_mobile