mindspore/mindspore/ccsrc/debug/debug_services.cc

/**
 * Copyright 2019-2020 Huawei Technologies Co., Ltd
 *
 * 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 <algorithm>
#include <map>
#include "backend/session/anf_runtime_algorithm.h"
#include "debug/debug_services.h"
namespace mindspore {

DebugServices::DebugServices() {
  tensor_loader_ = new TensorLoader();
  uint32_t iter_num = -1;
  tensor_loader_->set_iter_num(iter_num);
}

DebugServices::DebugServices(const DebugServices &other) {
  tensor_loader_ = other.tensor_loader_;
  watchpoint_table = other.watchpoint_table;
}

DebugServices &DebugServices::operator=(const DebugServices &other) {
  if (this != &other) {
    tensor_loader_ = other.tensor_loader_;
    watchpoint_table = other.watchpoint_table;
  }
  return *this;
}

DebugServices::~DebugServices() { delete tensor_loader_; }

void DebugServices::AddWatchpoint(unsigned int id, unsigned int watch_condition, float parameter,
                                  const std::vector<std::tuple<std::string, bool>> &check_node_list,
                                  const std::vector<parameter_t> &parameter_list) {
  std::lock_guard<std::mutex> lg(lock_);

  watchpoint_t watchpoint_item;
  watchpoint_item.id = id;
  watchpoint_item.condition.type = static_cast<CONDITION_TYPE>(watch_condition);
  watchpoint_item.condition.parameter = parameter;
  if (watch_condition > 2 && watch_condition < 13)
    // odd indices are greater than conditions and even indices are less than
    watchpoint_item.condition.comparison = (watch_condition & 1) == 0 ? "LT" : "GT";
  watchpoint_item.check_node_list = check_node_list;
  watchpoint_item.parameter_list = parameter_list;
  watchpoint_table[id] = watchpoint_item;
}

void DebugServices::RemoveWatchpoint(unsigned int id) {
  std::lock_guard<std::mutex> lg(lock_);
  watchpoint_table.erase(id);
}

template <typename T>
DebugServices::tensor_stats DebugServices::SummarizeTensor(const T *start, const T *start_prev, unsigned int n,
                                                           bool need_min_max, bool need_mean_sd,
                                                           bool need_zero_percentage,
                                                           bool need_tensor_update_ratio_mean, bool need_allclose) {
  tensor_stats stats;
  double zero_count = 0.0;
  double rtol = 1.0e-5;
  double atol = 1.0e-8;
  double update_ratio_sum = 0.0;
  double epsilon = 1.0e-9;
  for (unsigned int i = 0; i < n; ++i) {
    auto val = static_cast<double>(start[i]);
    double val_prev = 0.0;
    if (start_prev) {
      val_prev = static_cast<double>(start_prev[i]);
    }
    stats.has_nan = stats.has_nan || std::isnan(val);
    stats.has_inf = stats.has_inf || std::isinf(val);
    if (stats.has_inf && stats.has_nan) {
      // other statistics don't make sense in this case
      break;
    }

    if (need_min_max) {
      stats.min = std::min(stats.min, val);
      stats.max = std::max(stats.max, val);
    }

    if (need_mean_sd) {
      double delta = val - stats.mean;
      stats.mean += delta / (i + 1);
      stats.m2 += delta * (val - stats.mean);
    }

    if (need_zero_percentage) {
      if (val == 0) zero_count++;
    }

    if (need_tensor_update_ratio_mean && start_prev) {
      update_ratio_sum += (std::abs(val) / (epsilon + std::abs(val_prev)));
    }

    if (need_allclose && start_prev) {
      stats.allclose &= (std::abs(val - val_prev) <= (atol + rtol * std::abs(val_prev)));
    }
  }
  if (need_tensor_update_ratio_mean && start_prev) {
    stats.tensor_update_ratio_mean = (update_ratio_sum / n);
  }
  stats.zero_percentage = (zero_count / n) * 100;
  stats.n = n;
  return stats;
}

void DebugServices::CheckWatchpoints(std::vector<std::string> *name, std::vector<std::string> *slot,
                                     std::vector<int> *condition, std::vector<unsigned int> *watchpoint_id,
                                     std::vector<std::vector<parameter_t>> *parameters,
                                     const std::vector<std::string> &op_overflows,
                                     const std::vector<std::shared_ptr<TensorData>> &tensor_list,
                                     const bool init_dbg_suspend) {
  std::lock_guard<std::mutex> lg(lock_);
  if (watchpoint_table.empty()) {
    return;
  }

  for (const auto &tensor : tensor_list) {
    const auto tensor_name = tensor->GetName();
    const auto tensor_name_no_slot = tensor_name.substr(0, tensor_name.find_first_of(':'));
    const auto tensor_slot = std::to_string(tensor->GetSlot());
    mindspore::tensor::TensorPtr tensor_ptr = tensor->GetTensor();
    int tensor_dtype = tensor_ptr->data_type_c();
    std::vector<unsigned int> hit_encountered;
    std::vector<std::vector<bool>> hit_parms;
    std::unordered_map<unsigned int, watchpoint_t> watchpoints_to_check_table;
    bool min_max_enabled = false;
    bool mean_sd_enabled = false;
    bool inf_nan_enabled = false;
    bool zero_percentage_enabled = false;
    bool tensor_update_ratio_mean_enabled = false;
    bool allclose_enabled = false;
    for (auto w_table_item : watchpoint_table) {
      auto wp = std::get<1>(w_table_item);
      if (wp.condition.type == INIT && !init_dbg_suspend) continue;
      if (wp.condition.type != IS_OVERFLOW && tensor_dtype == kNumberTypeBool) continue;
      if (wp.IsNodeIncluded(tensor_name_no_slot)) {
        min_max_enabled |= wp.min_max_enabled();
        mean_sd_enabled |= wp.mean_sd_enabled();
        inf_nan_enabled |= wp.inf_nan_enabled();
        zero_percentage_enabled |= wp.zero_percentage_enabled();
        tensor_update_ratio_mean_enabled |= wp.tensor_update_ratio_mean_enabled();
        allclose_enabled |= wp.allclose_enabled();
        watchpoints_to_check_table[w_table_item.second.id] = w_table_item.second;
      }
    }
    tensor_stats stats;
    uint num_elements = tensor_ptr->DataSize();
    if (min_max_enabled || mean_sd_enabled || inf_nan_enabled || zero_percentage_enabled ||
        tensor_update_ratio_mean_enabled || allclose_enabled) {
      bool need_prev = (tensor_update_ratio_mean_enabled || allclose_enabled);
      bool have_prev = tensor_loader_->GetPrevTensor(tensor_name) != NULL;
      switch (tensor_dtype) {
        case kNumberTypeUInt8: {
          auto start_addr = reinterpret_cast<uint8_t *>(tensor_ptr->data_c());
          auto start_addr_prev =
            (need_prev && have_prev
               ? reinterpret_cast<uint8_t *>(tensor_loader_->GetPrevTensor(tensor_name)->GetTensor()->data_c())
               : NULL);
          stats = SummarizeTensor(start_addr, start_addr_prev, num_elements, min_max_enabled, mean_sd_enabled,
                                  zero_percentage_enabled, tensor_update_ratio_mean_enabled, allclose_enabled);
          break;
        }
        case kNumberTypeInt8: {
          auto start_addr = reinterpret_cast<int8_t *>(tensor_ptr->data_c());
          auto start_addr_prev =
            (need_prev && have_prev
               ? reinterpret_cast<int8_t *>(tensor_loader_->GetPrevTensor(tensor_name)->GetTensor()->data_c())
               : NULL);
          stats = SummarizeTensor(start_addr, start_addr_prev, num_elements, min_max_enabled, mean_sd_enabled,
                                  zero_percentage_enabled, tensor_update_ratio_mean_enabled, allclose_enabled);
          break;
        }
        case kNumberTypeUInt16: {
          auto start_addr = reinterpret_cast<uint16_t *>(tensor_ptr->data_c());
          auto start_addr_prev =
            (need_prev && have_prev
               ? reinterpret_cast<uint16_t *>(tensor_loader_->GetPrevTensor(tensor_name)->GetTensor()->data_c())
               : NULL);
          stats = SummarizeTensor(start_addr, start_addr_prev, num_elements, min_max_enabled, mean_sd_enabled,
                                  zero_percentage_enabled, tensor_update_ratio_mean_enabled, allclose_enabled);
          break;
        }
        case kNumberTypeInt16: {
          auto start_addr = reinterpret_cast<int16_t *>(tensor_ptr->data_c());
          auto start_addr_prev =
            (need_prev && have_prev
               ? reinterpret_cast<int16_t *>(tensor_loader_->GetPrevTensor(tensor_name)->GetTensor()->data_c())
               : NULL);
          stats = SummarizeTensor(start_addr, start_addr_prev, num_elements, min_max_enabled, mean_sd_enabled,
                                  zero_percentage_enabled, tensor_update_ratio_mean_enabled, allclose_enabled);
          break;
        }
        case kNumberTypeUInt32: {
          auto start_addr = reinterpret_cast<uint32_t *>(tensor_ptr->data_c());
          auto start_addr_prev =
            (need_prev && have_prev
               ? reinterpret_cast<uint32_t *>(tensor_loader_->GetPrevTensor(tensor_name)->GetTensor()->data_c())
               : NULL);
          stats = SummarizeTensor(start_addr, start_addr_prev, num_elements, min_max_enabled, mean_sd_enabled,
                                  zero_percentage_enabled, tensor_update_ratio_mean_enabled, allclose_enabled);
          break;
        }
        case kNumberTypeInt32:
        case kNumberTypeInt: {
          auto start_addr = reinterpret_cast<int32_t *>(tensor_ptr->data_c());
          auto start_addr_prev =
            (need_prev && have_prev
               ? reinterpret_cast<int32_t *>(tensor_loader_->GetPrevTensor(tensor_name)->GetTensor()->data_c())
               : NULL);
          stats = SummarizeTensor(start_addr, start_addr_prev, num_elements, min_max_enabled, mean_sd_enabled,
                                  zero_percentage_enabled, tensor_update_ratio_mean_enabled, allclose_enabled);
          break;
        }
        case kNumberTypeUInt64: {
          auto start_addr = reinterpret_cast<uint64_t *>(tensor_ptr->data_c());
          auto start_addr_prev =
            (need_prev && have_prev
               ? reinterpret_cast<uint64_t *>(tensor_loader_->GetPrevTensor(tensor_name)->GetTensor()->data_c())
               : NULL);
          stats = SummarizeTensor(start_addr, start_addr_prev, num_elements, min_max_enabled, mean_sd_enabled,
                                  zero_percentage_enabled, tensor_update_ratio_mean_enabled, allclose_enabled);
          break;
        }
        case kNumberTypeInt64: {
          auto start_addr = reinterpret_cast<int64_t *>(tensor_ptr->data_c());
          auto start_addr_prev =
            (need_prev && have_prev
               ? reinterpret_cast<int64_t *>(tensor_loader_->GetPrevTensor(tensor_name)->GetTensor()->data_c())
               : NULL);
          stats = SummarizeTensor(start_addr, start_addr_prev, num_elements, min_max_enabled, mean_sd_enabled,
                                  zero_percentage_enabled, tensor_update_ratio_mean_enabled, allclose_enabled);
          break;
        }
        case kNumberTypeFloat16: {
          auto start_addr = reinterpret_cast<float16 *>(tensor_ptr->data_c());
          auto start_addr_prev =
            (need_prev && have_prev
               ? reinterpret_cast<float16 *>(tensor_loader_->GetPrevTensor(tensor_name)->GetTensor()->data_c())
               : NULL);
          stats = SummarizeTensor(start_addr, start_addr_prev, num_elements, min_max_enabled, mean_sd_enabled,
                                  zero_percentage_enabled, tensor_update_ratio_mean_enabled, allclose_enabled);
          break;
        }
        case kNumberTypeFloat32:
        case kNumberTypeFloat: {
          auto start_addr = reinterpret_cast<float *>(tensor_ptr->data_c());
          auto start_addr_prev =
            (need_prev && have_prev
               ? reinterpret_cast<float *>(tensor_loader_->GetPrevTensor(tensor_name)->GetTensor()->data_c())
               : NULL);
          stats = SummarizeTensor(start_addr, start_addr_prev, num_elements, min_max_enabled, mean_sd_enabled,
                                  zero_percentage_enabled, tensor_update_ratio_mean_enabled, allclose_enabled);
          break;
        }
        case kNumberTypeFloat64: {
          auto start_addr = reinterpret_cast<double *>(tensor_ptr->data_c());
          auto start_addr_prev =
            (need_prev && have_prev
               ? reinterpret_cast<double *>(tensor_loader_->GetPrevTensor(tensor_name)->GetTensor()->data_c())
               : NULL);
          stats = SummarizeTensor(start_addr, start_addr_prev, num_elements, min_max_enabled, mean_sd_enabled,
                                  zero_percentage_enabled, tensor_update_ratio_mean_enabled, allclose_enabled);
          break;
        }
        default:
          MS_LOG(INFO) << "Unsupported tensor type";
          break;
      }
    }

    for (auto &it : watchpoints_to_check_table) {
      auto wp_id = it.second.id;
      std::vector<bool> hit_p;
      CONDITION_TYPE enabled_condition = it.second.condition.type;
      bool hit = (enabled_condition == HAS_NAN && stats.has_nan) || (enabled_condition == HAS_INF && stats.has_inf) ||
                 (enabled_condition == GENERAL_OVERFLOW && (stats.has_nan || stats.has_inf)) ||
                 (enabled_condition == IS_OVERFLOW &&
                  std::find(op_overflows.begin(), op_overflows.end(), tensor_name_no_slot) != op_overflows.end());

      if (enabled_condition > 2 && enabled_condition != GENERAL_OVERFLOW) {
        if (stats.has_inf || stats.has_nan) {
          MS_LOG(WARNING) << "NaN or/and INF present in tensor: " << tensor_name << ". Cannot check "
                          << condition_label[enabled_condition] << " watchpoint.";
        } else if (enabled_condition < 13) {
          bool gt = stats.statLookup(enabled_condition) > it.second.condition.parameter;
          bool lt = stats.statLookup(enabled_condition) < it.second.condition.parameter;
          hit |= it.second.condition.comparison == "GT" ? gt : lt;
        } else {
          std::vector<parameter_t> parameter_list_item = it.second.parameter_list;
          for (auto &p : parameter_list_item) {
            if (p.disabled == false) {
              bool p_hit = false;
              if (p.name == "zero_percentage_ge") {
                p_hit = stats.parmLookup(STAT_ZERO_PERCENTAGE) >= p.value;
              } else if (p.name == "max_gt") {
                p_hit = stats.parmLookup(STAT_MAX) > p.value;
              } else if (p.name == "max_lt") {
                p_hit = stats.parmLookup(STAT_MAX) < p.value;
              } else if (p.name == "min_gt") {
                p_hit = stats.parmLookup(STAT_MIN) > p.value;
              } else if (p.name == "min_lt") {
                p_hit = stats.parmLookup(STAT_MIN) < p.value;
              } else if (p.name == "mean_gt") {
                p_hit = stats.parmLookup(STAT_MEAN) > p.value;
              } else if (p.name == "mean_lt") {
                p_hit = stats.parmLookup(STAT_MEAN) < p.value;
              } else if (p.name == "abs_mean_gt") {
                p_hit = std::abs(stats.parmLookup(STAT_MEAN)) > p.value;
              } else if (p.name == "abs_mean_lt") {
                p_hit = std::abs(stats.parmLookup(STAT_MEAN)) < p.value;
              } else if (p.name == "abs_update_ratio_mean_gt") {
                p_hit = stats.parmLookup(STAT_TENSOR_UPDATE_RATIO_MEAN) > p.value;
              } else if (p.name == "abs_update_ratio_mean_lt") {
                p_hit = stats.parmLookup(STAT_TENSOR_UPDATE_RATIO_MEAN) < p.value;
              }
              hit |= p_hit;
              hit_p.push_back(p_hit);
            } else {
              hit_p.push_back(false);
            }
          }

          hit |= (enabled_condition == NOT_CHANGED && stats.parmLookup(STAT_ALLCLOSE));

          if (hit) hit_parms.push_back(hit_p);
        }
      }
      if (hit) hit_encountered.push_back(wp_id);
    }

    unsigned int index_parm_list = 0;
    for (auto it_hit_id = hit_encountered.begin(); it_hit_id != hit_encountered.end(); ++it_hit_id) {
      if (watchpoint_table.find(*it_hit_id) != watchpoint_table.end()) {
        // return fully qualified name for weights and bias to MI
        auto found_dot = tensor_name_no_slot.find_last_of('.');
        if (found_dot != std::string::npos && (tensor_name_no_slot.substr(found_dot + 1) == "weight" ||
                                               tensor_name_no_slot.substr(found_dot + 1) == "bias")) {
          auto check_node_list = watchpoint_table.find(*it_hit_id)->second.check_node_list;
          for (auto check_node : check_node_list) {
            std::string w_name = std::get<0>(check_node);
            auto found_slash = w_name.find_last_of('/');
            if (found_slash != std::string::npos && w_name.substr(found_slash + 1) == tensor_name_no_slot) {
              name->push_back(w_name);
            }
          }
        } else {
          name->push_back(tensor_name_no_slot);
        }

        slot->push_back(tensor_slot);
        int condition_item = watchpoint_table.find(*it_hit_id)->second.condition.type;
        condition->push_back(condition_item);
        watchpoint_id->push_back(*it_hit_id);
        std::vector<parameter_t> parameter_list_item = watchpoint_table.find(*it_hit_id)->second.parameter_list;
        if (condition_item >= 13) {
          unsigned int index_hit_parm = 0;
          for (auto &p : parameter_list_item) {
            p.hit = hit_parms[index_parm_list][index_hit_parm];
            index_hit_parm++;
          }
          index_parm_list++;
        }
        parameters->push_back(parameter_list_item);
      }
      watchpoints_to_check_table.erase(*it_hit_id);
    }
  }
}

void DebugServices::ReadNodesTensors(std::vector<std::string> name, std::vector<std::string> *ret_name,
                                     std::vector<char *> *data_ptr, std::vector<unsigned int> *data_size,
                                     std::vector<TypePtr> *dtype, std::vector<std::vector<int>> *shape) {
  std::vector<std::tuple<std::string, std::shared_ptr<TensorData>>> result_list;
  tensor_loader_->SearchTensors(name, &result_list);

  for (auto result : result_list) {
    if (!std::get<1>(result)) {
      continue;
    }
    ret_name->push_back(std::get<0>(result));
    data_ptr->push_back(reinterpret_cast<char *>(std::get<1>(result)->GetTensor()->data_c()));
    data_size->push_back(std::get<1>(result)->GetTensor()->data().nbytes());
    dtype->push_back(std::get<1>(result)->GetTensor()->Dtype());
    shape->push_back(std::get<1>(result)->GetTensor()->shape());
  }
}

bool DebugServices::IsWatchPoint(std::string kernel_name, const CNodePtr &kernel) {
  bool ret = false;
  for (auto w_table_item : watchpoint_table) {
    auto check_node_list = std::get<1>(w_table_item).check_node_list;
    for (auto check_node : check_node_list) {
      std::string w_name = std::get<0>(check_node);
      bool w_type = std::get<1>(check_node);
      if ((w_type == true &&
           ((kernel_name.find(w_name) != string::npos && kernel_name.rfind(w_name, 0) == 0) || w_name == "*")) ||
          (w_type == false && (kernel_name == w_name || IsWatchPointNodeInput(w_name, kernel)))) {
        ret = true;
        return ret;
      }
    }
  }
  return ret;
}

bool DebugServices::IsWatchPointNodeInput(std::string w_name, const CNodePtr &kernel) {
  if (kernel) {
    auto input_size = AnfAlgo::GetInputTensorNum(kernel);
    for (size_t j = 0; j < input_size; ++j) {
      auto input_kernel = kernel->input(j + 1);
      std::string input_kernel_name = input_kernel->fullname_with_scope();
      auto found = w_name.find_last_of('/');
      if (found != std::string::npos && w_name.substr(found + 1) == input_kernel_name) return true;
    }
    return false;
  } else {
    return false;
  }
}

void DebugServices::AddWeightsBiasInputs(std::vector<std::shared_ptr<TensorData>> *tensor_list,
                                         const CNodePtr &kernel) {
  if (kernel) {
    auto input_size = AnfAlgo::GetInputTensorNum(kernel);
    for (size_t j = 0; j < input_size; ++j) {
      auto input_kernel = kernel->input(j + 1);
      std::string input_kernel_name = input_kernel->fullname_with_scope();
      std::string locate_tensor = input_kernel_name + ":0";
      std::map<std::string, std::shared_ptr<TensorData>> tensor_map = tensor_loader_->GetTensorMap();
      std::map<std::string, std::shared_ptr<TensorData>>::iterator iter;
      iter = tensor_map.find(locate_tensor);
      if (iter != tensor_map.end()) {
        tensor_list->push_back(iter->second);
      }
    }
  }
}

TensorLoader *DebugServices::tensor_loader() const { return tensor_loader_; }
std::unordered_map<unsigned int, DebugServices::watchpoint_t> DebugServices::GetWatchpointTable() {
  return watchpoint_table;
}

}  // namespace mindspore