plan_memory.cc 15.2 KB
Newer Older
1 2 3 4 5 6 7 8
/*
 * Licensed to the Apache Software Foundation (ASF) under one
 * or more contributor license agreements.  See the NOTICE file
 * distributed with this work for additional information
 * regarding copyright ownership.  The ASF licenses this file
 * to you 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
9
 *
10
 *   http://www.apache.org/licenses/LICENSE-2.0
11
 *
12 13 14 15 16 17 18 19
 * 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.
 */

20 21 22 23 24 25 26 27 28
/*!
 * \file plan_memory.cc
 * \brief Assign memory tag to each of the data entries.
 */
#include <nnvm/graph.h>
#include <nnvm/pass.h>
#include <nnvm/graph_attr_types.h>
#include <nnvm/op_attr_types.h>
#include <memory>
29
#include "graph_algorithm.h"
30 31 32

namespace nnvm {
namespace pass {
33
namespace {
34

35 36 37 38 39 40 41
// Return bytes of data flag.
static int GetDTypeSize(int type_flag) {
  switch (type_flag) {
    case kUint8:
    case kInt8:
      return 1;
    case kFloat16:
42
    case kBfloat16:
43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58
    case kInt16:
    case kUint16:
      return 2;
    case kFloat32:
    case kInt32:
    case kUint32:
      return 4;
    case kFloat64:
    case kInt64:
    case kUint64:
      return 8;
    default:
      LOG(FATAL) << "unknown type_flag=" << type_flag;
      return -1;
  }
}
59 60 61 62 63 64

// simple graph based allocator.
class GraphAllocator {
 public:
  // storage id equals integer.
  using StorageID = int;
65

66 67
  // bad storage id
  static const StorageID kBadStorageID = -1;
68 69
  // external storage id
  static const StorageID kExternalStorageID = -2;
70 71
  // dynamic storage id
  static const StorageID kDynamicStorageID = -3;
72

73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103
  // request a free storage
  StorageID Request(int dev_id, int dtype, TShape shape, uint32_t node_id) {
    if (shape.ndim() == 0) return kBadStorageID;
    // search memory block in [size / match_range_, size * match_range_)
    // TODO(tqchen) add size of the dtype, assume 4 bytes for now
    size_t size = shape.Size() * 4;
    if (match_range_ == 0) return this->Alloc(dev_id, size);
    auto begin = free_.lower_bound(size / match_range_);
    auto mid = free_.lower_bound(size);
    auto end = free_.upper_bound(size * match_range_);
    // search for memory blocks larger than requested
    for (auto it = mid; it != end; ++it) {
      StorageEntry *e = it->second;
      if (e->device_id != dev_id) continue;
      if (node_color_.size() != 0 &&
          node_color_[e->released_by_node] != node_color_[node_id]) continue;
      // Use exect matching strategy
      e->max_bytes = std::max(size, e->max_bytes);
      // find a exact match, erase from map and return
      free_.erase(it);
      return e->id;
    }
    // then search for memory blocks smaller than requested space
    for (auto it = mid; it != begin;) {
      --it;
      StorageEntry *e = it->second;
      if (e->device_id != dev_id) continue;
      if (node_color_.size() != 0 &&
          node_color_[e->released_by_node] != node_color_[node_id]) continue;
      // Use exect matching strategy
      e->max_bytes = std::max(size, e->max_bytes);
104
      // erase from map and return
105 106 107 108 109 110 111 112 113
      free_.erase(it);
      return e->id;
    }
    // cannot find anything return a new one.
    return this->Alloc(dev_id, size);
  }
  // release a memory space.
  void Release(StorageID id, uint32_t node_id) {
    CHECK_NE(id, kBadStorageID);
114
    if (id == kExternalStorageID || id == kDynamicStorageID) return;
115 116 117 118
    StorageEntry *e = data_[id].get();
    e->released_by_node = node_id;
    free_.insert({e->max_bytes, e});
  }
119

120 121 122 123 124 125 126 127 128 129
  // totoal number of bytes allocated
  size_t TotalAllocBytes() const {
    size_t total = 0;
    for (auto &p : data_) {
      total += p->max_bytes;
    }
    return total;
  }

  // constructor
130 131
  explicit GraphAllocator(const IndexedGraph* idx, const size_t match_range) : idx_(idx) {
    this->Init(match_range, dmlc::GetEnv("NNVM_EXEC_NUM_TEMP", 1));
132 133 134 135
  }

 private:
  // initialize the graph allocator
136
  void Init(const size_t match_range, const uint32_t num_match_color) {
137 138 139 140 141 142 143 144
    match_range_ = match_range;
    num_match_color_ = num_match_color;
    if (num_match_color_ > 1) {
      std::vector<uint32_t> importance(idx_->num_nodes(), 0);
      for (uint32_t nid = 0; nid < idx_->num_nodes(); ++nid) {
        if ((*idx_)[nid].source->is_variable()) continue;
        importance[nid] = 1;
      }
145
      num_match_color_ = pass::ColorNodeGroup(
146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185
          *idx_, importance, num_match_color_, &node_color_);
    }
  }

  StorageID Alloc(int dev_id, size_t size) {
    StorageID id = static_cast<StorageID>(data_.size());
    std::unique_ptr<StorageEntry> ptr(new StorageEntry());
    ptr->id = id;
    ptr->device_id = dev_id;
    ptr->max_bytes = size;
    data_.emplace_back(std::move(ptr));
    return id;
  }
  // internal storage entry
  struct StorageEntry {
    // the id of the entry.
    StorageID id;
    // the device id of the storage.
    int device_id;
    // maximum size of storage requested.
    size_t max_bytes{0};
    // node index that released it last time
    uint32_t released_by_node{0};
  };
  // scale used for rough match
  size_t match_range_;
  // whether use color based match algorithm
  uint32_t num_match_color_{1};
  // the size of each dtype
  std::vector<size_t> dtype_size_dict_;
  // free list of storage entry
  std::multimap<size_t, StorageEntry*> free_;
  // all the storage resources available
  std::vector<std::unique_ptr<StorageEntry> > data_;
  // color of nodes in the graph, used for auxiliary policy making.
  std::vector<uint32_t> node_color_;
  // internal indexed graph
  const IndexedGraph* idx_;
};

186 187 188
/*
 * Internal method to perform the memory allocation for a graph
 * */
189 190 191
size_t AllocMemory(const Graph& ret, const IndexedGraph& idx,
                   const std::pair<uint32_t, uint32_t>& node_range,
                   StorageVector* storage_ptr,
192 193
                   std::vector<int>* storage_inplace_index_ptr,
                   const std::vector<uint32_t>& entry_ref_count,
194
                   GraphAllocator* allocator) {
195 196
  static auto& finplace_option = Op::GetAttr<FInplaceOption>("FInplaceOption");
  static auto& finplace_identity = Op::GetAttr<FInplaceIdentity>("FInplaceIdentity");
197
  static auto& fignore_inputs = Op::GetAttr<FIgnoreInputs>("FIgnoreInputs");
198

199 200 201
  // Get reference
  auto &storage = *storage_ptr;
  auto &storage_inplace_index = *storage_inplace_index_ptr;
202

203
  // Get attributes from the graph
204 205 206 207 208 209 210 211
  const ShapeVector& shape_vec = ret.GetAttr<ShapeVector>("shape");
  const DTypeVector& dtype_vec = ret.GetAttr<DTypeVector>("dtype");
  const DeviceVector* device_vec = nullptr;

  if (ret.attrs.count("device") != 0) {
    device_vec = &(ret.GetAttr<DeviceVector>("device"));
  }
  size_t num_not_allocated = 0;
212
  std::vector<GraphAllocator::StorageID> storage_ref_count(idx.num_node_entries(), 0);
213

214
  for (uint32_t nid = node_range.first; nid < node_range.second; ++nid) {
215 216 217
    const auto& inode = idx[nid];
    if (inode.source->is_variable()) continue;
    // check inplace option
218 219
    if (finplace_option.count(inode.source->op()) != 0) {
      auto inplace_pairs = finplace_option[inode.source->op()](inode.source->attrs);
220 221 222 223 224 225 226 227 228 229 230 231
      std::vector<bool> identity;
      if (finplace_identity.count(inode.source->op()) != 0) {
        identity = finplace_identity[inode.source->op()](inode.source->attrs);
        CHECK_EQ(identity.size(), inplace_pairs.size())
            << "FInplaceOption and FInplaceIdentity returned vectors of different "
            << "size for operator " << inode.source->op()->name;
      } else {
        identity = std::vector<bool>(inplace_pairs.size(), false);
      }
      std::vector<bool> taken(inode.inputs.size(), false);
      for (size_t ipair = 0; ipair < inplace_pairs.size(); ++ipair) {
        const auto& kv = inplace_pairs[ipair];
232 233
        uint32_t eid_out = idx.entry_id(nid, kv.second);
        uint32_t eid_in = idx.entry_id(inode.inputs[kv.first]);
234 235
        auto sid_out = storage[eid_out];
        auto sid_in = storage[eid_in];
236 237 238
        bool ignore_all_inputs = (fignore_inputs.count(inode.source->op()) != 0 &&
                                  fignore_inputs[inode.source->op()](
                                      inode.source->attrs).size() == inode.source->num_inputs());
239 240 241 242
        // Identity should only be true if shape.Size() and types match
        bool real_identity = identity[ipair] &&
                             shape_vec[eid_out].Size() == shape_vec[eid_in].Size() &&
                             dtype_vec[eid_out] == dtype_vec[eid_in];
243 244 245
        if (taken[kv.first] == false &&
            sid_out == GraphAllocator::kBadStorageID &&
            sid_in >= 0 &&
246
            ((storage_ref_count[sid_in] == 1 && !ignore_all_inputs) || real_identity) &&
247
            entry_ref_count[eid_out] > 0 &&
248
            shape_vec[eid_out].Size() == shape_vec[eid_in].Size() &&
249 250
             (dtype_vec[eid_out] == dtype_vec[eid_in] ||
             GetDTypeSize(dtype_vec[eid_out]) == GetDTypeSize(dtype_vec[eid_in]))) {
Tianqi Chen committed
251
          // inplace optimization
252 253 254 255 256 257
          taken[kv.first] = true;
          storage[eid_out] = sid_in;
          // Reuse storage for output and add ref count of output
          // to storage. This will get substracted later in free
          // input section.
          storage_ref_count[sid_in] += entry_ref_count[eid_out];
Tianqi Chen committed
258
          storage_inplace_index[eid_out] = kv.first;
259 260 261 262 263
        }
      }
    }
    // normal allocation
    const int dev_id = (device_vec != nullptr) ? device_vec->at(nid) : 0;
264 265
    // sort output nodes based on size before allocating output
    std::multimap<size_t, uint32_t> eids;
266 267
    for (uint32_t index = 0; index < inode.source->num_outputs(); ++index) {
      uint32_t eid = idx.entry_id(nid, index);
268
      // only request memory for kBadStorageID
269
      if (storage[eid] == GraphAllocator::kBadStorageID) {
270 271 272 273
        auto &eshape = shape_vec[eid];
        size_t esize = 0;
        if (eshape.ndim() != 0) esize = eshape.Size();
        eids.insert(std::make_pair(esize, eid));
274 275
      }
    }
276 277
    for (auto rit = eids.rbegin(); rit != eids.rend(); ++rit) {
        uint32_t eid = rit->second;
278
        auto sid = allocator->Request(dev_id, dtype_vec[eid], shape_vec[eid], nid);
279 280 281
        if (sid >= 0) {
          storage_ref_count[sid] = entry_ref_count[eid];
        }
282
        storage[eid] = sid;
283
    }
284 285 286 287 288 289
    // check if certain inputs is ignored.
    std::vector<uint32_t> ignore_inputs;
    if (fignore_inputs.count(inode.source->op()) != 0) {
      ignore_inputs = fignore_inputs[inode.source->op()](inode.source->attrs);
      std::sort(ignore_inputs.begin(), ignore_inputs.end());
    }
290
    // then free inputs
291 292 293 294
    for (size_t i = 0; i < inode.inputs.size(); ++i) {
      // ref counter of ignored input is already decreased.
      if (std::binary_search(ignore_inputs.begin(), ignore_inputs.end(), i)) continue;
      const auto& e = inode.inputs[i];
295
      uint32_t eid = idx.entry_id(e);
296 297 298
      auto sid = storage[eid];
      // storage_ref_count == 0 means it is taken by inplace op
      if (sid < 0) continue;
299
      // if we decrease it to zero, we are ready to relase
300 301 302
      --storage_ref_count[sid];
      if (storage_ref_count[sid] == 0) {
        allocator->Release(sid, nid);
303 304 305
      }
    }
    // check if there are outputs that can be freeded immediately
306
    // these output are not referenced by any operator.
307 308
    for (uint32_t index = 0; index < inode.source->num_outputs(); ++index) {
      uint32_t eid = idx.entry_id(nid, index);
309 310 311
      auto sid = storage[eid];
      if (sid >= 0 && storage_ref_count[sid] == 0) {
        allocator->Release(sid, nid);
312 313
        // use -2 to indicate that the node was never touched.
        storage_inplace_index[eid] = -2;
314 315 316 317 318 319
      }
      if (storage[eid] == GraphAllocator::kBadStorageID) {
        ++num_not_allocated;
      }
    }
  }
320 321 322 323 324 325 326 327 328
  return num_not_allocated;
}


// function to plan memory
Graph PlanMemory(Graph ret) {
  // setup ref counter
  const IndexedGraph& idx = ret.indexed_graph();
  static auto& fignore_inputs = Op::GetAttr<FIgnoreInputs>("FIgnoreInputs");
329 330 331 332
  std::pair<uint32_t, uint32_t> node_range = {0, idx.num_nodes()};
  if (ret.attrs.count("node_range")) {
    node_range = ret.MoveCopyAttr<std::pair<uint32_t, uint32_t> >("node_range");
  }
333
  // reference counter of each node
334
  std::vector<uint32_t> ref_count;
335
  // step 1: initialize reference count
336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352
  if (ret.attrs.count("ref_count") != 0) {
    ref_count = ret.MoveCopyAttr<std::vector<uint32_t> >("ref_count");
  } else {
    ref_count.resize(idx.num_node_entries(), 0);
    for (uint32_t nid = 0; nid < idx.num_nodes(); ++nid) {
      const auto& inode = idx[nid];
      if (inode.source->is_variable()) continue;
      for (const auto& e : inode.inputs) {
        ++ref_count[idx.entry_id(e)];
      }
      // no dataflow dependency is needed for those are ignored.
      // revoke the dependency counter.
      if (fignore_inputs.count(inode.source->op()) != 0) {
        auto ignore_inputs = fignore_inputs[inode.source->op()](inode.source->attrs);
        for (uint32_t i : ignore_inputs) {
          --ref_count[idx.entry_id(inode.inputs[i])];
        }
353 354
      }
    }
355 356 357
    for (const auto& e : idx.outputs()) {
      ++ref_count[idx.entry_id(e)];
    }
358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381
  }
  // step 2: allocate memory.
  StorageVector storage;
  if (ret.attrs.count("storage") != 0) {
    storage = ret.MoveCopyAttr<StorageVector>("storage");
  } else {
    storage.resize(idx.num_node_entries(), -1);
  }

  // Search the best NNVM_EXEC_MATCH_RANGE parameter. This is turned off by default
  size_t min_allocated_bytes = -1;
  size_t max_match_range = dmlc::GetEnv("NNVM_EXEC_MATCH_RANGE", 16);
  size_t min_match_range =
         dmlc::GetEnv("NNVM_AUTO_SEARCH_MATCH_RANGE", false) ? 1 : max_match_range;
  for (size_t match_range = min_match_range; match_range <= max_match_range; match_range *= 2) {
    // Make a copy of related fields
    StorageVector storage_vec(storage);
    std::vector<int> storage_inplace_index(idx.num_node_entries(), -1);

    // the allocator
    GraphAllocator allocator(&idx, match_range);

    // number of entries that are not statically allocated.
    size_t storage_num_not_allocated =
382 383
      AllocMemory(ret, idx, node_range, &storage_vec, &storage_inplace_index,
                  ref_count, &allocator);
384
    size_t storage_allocated_bytes = allocator.TotalAllocBytes();
385

386 387 388 389 390 391 392 393
    // Choose the plan which leads to minimal memory usage
    if (min_allocated_bytes > storage_allocated_bytes) {
      ret.attrs["storage_id"] = std::make_shared<any>(std::move(storage_vec));
      ret.attrs["storage_inplace_index"] = std::make_shared<any>(std::move(storage_inplace_index));
      ret.attrs["storage_allocated_bytes"] = std::make_shared<any>(storage_allocated_bytes);
      ret.attrs["storage_num_not_allocated"] = std::make_shared<any>(storage_num_not_allocated);
      min_allocated_bytes = storage_allocated_bytes;
    }
394 395 396 397

    if (max_match_range == 0) {
      break;
    }
398
  }
399 400 401 402 403 404 405 406 407
  return ret;
}

NNVM_REGISTER_PASS(PlanMemory)
.describe("Plan the memory allocation of each node entries.")
.set_body(PlanMemory)
.set_change_graph(false)
.depend_graph_attr("dtype")
.depend_graph_attr("shape")
Tianqi Chen committed
408 409
.provide_graph_attr("storage_id")
.provide_graph_attr("storage_inplace_index");
410

411
}  // namespace
412 413
}  // namespace pass
}  // namespace nnvm