graph_plan_memory.cc 12.2 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19
/*
 * 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
 * 
 *   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.
 */

20 21 22 23 24 25 26 27
/*!
 *  Copyright (c) 2018 by Contributors
 * \file relay/backend/graph_mem_alloca.cc
 * \brief Memory index assignment pass for executing
 *   the program in the graph runtime.
 */
#include <tvm/relay/expr.h>
#include <tvm/relay/expr_functor.h>
Zhi committed
28
#include <tvm/relay/analysis.h>
29 30 31 32 33
#include "../../common/arena.h"

namespace tvm {
namespace relay {

34 35
using IntegerArray = Array<Integer>;

36 37 38 39 40 41 42
struct StorageToken {
  /*! \brief Reference counter */
  int ref_counter{0};
  /*! \brief number of bytes */
  size_t max_bytes{0};
  /*! \brief The corresponding tensor type node. */
  const TensorTypeNode* ttype{nullptr};
43 44 45
  /*! \brief virtual device index that corresponds to the device_type in
   * DLContext. */
  int device_type{0};
46 47 48 49 50 51 52 53 54 55 56
  /*! \brief The storage id */
  int64_t storage_id{-1};
};

class StorageAllocaBaseVisitor : public ExprVisitor {
 public:
  // run the visitor on a function.
  void Run(const Function& func) {
    for (Var param : func->params) {
      CreateToken(param.operator->(), false);
    }
57 58 59 60
    // must always keep output alive.
    for (StorageToken* tok : GetToken(func->body)) {
      tok->ref_counter += 1;
    }
61 62 63 64 65 66 67 68 69 70 71 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139
  }

  void VisitExpr_(const ConstantNode* op) final {
    this->CreateToken(op, false);
  }

  void VisitExpr_(const VarNode* op) final {
    // Do nothing.
  }

  void VisitExpr_(const FunctionNode* op) final {
    // do not recursive into sub function.
  }

  void VisitExpr_(const GlobalVarNode* op) final {
    // Do nothing.
  }

  void VisitExpr_(const OpNode* op) final {
    // Do nothing.
  }

  void VisitExpr_(const TupleNode* op) final {
    std::vector<StorageToken*> fields;
    for (Expr field : op->fields) {
      auto tok = GetToken(field);
      CHECK_EQ(tok.size(), 1U);
      fields.push_back(tok[0]);
    }
    token_map_[op] = fields;
  }

  void VisitExpr_(const TupleGetItemNode* op) final {
    const auto& tok = GetToken(op->tuple);
    CHECK_LT(static_cast<size_t>(op->index), tok.size());
    token_map_[op] = {tok[op->index]};
  }

  void VisitExpr_(const IfNode* op) final {
    LOG(FATAL) << "if is not supported.";
  }

  void VisitExpr_(const LetNode* op) final {
    auto token = GetToken(op->value);
    token_map_[op->var.operator->()] = token;
    token_map_[op] = GetToken(op->body);
  }

 protected:
  /*! \brief internal token map */
  std::unordered_map<const ExprNode*, std::vector<StorageToken*> > token_map_;

  /*!
   * \brief Get the necessary token.
   * \param expr The expression.
   * \return The corresponding token.
   */
  const std::vector<StorageToken*>& GetToken(const Expr& expr) {
    this->VisitExpr(expr);
    auto it = token_map_.find(expr.operator->());
    CHECK(it != token_map_.end());
    return it->second;
  }
  /*!
   * \brief Populate the token map to set op's tokens
   * \param op The node to be processed.
   * \param can_realloc Whether we can re-allocate the memory.
   */
  virtual void CreateToken(const ExprNode* op, bool can_realloc) = 0;
};

class StorageAllocaInit : protected StorageAllocaBaseVisitor {
 public:
  explicit StorageAllocaInit(common::Arena* arena)
      : arena_(arena) {}

  /*! \return The internal token map */
  std::unordered_map<const ExprNode*, std::vector<StorageToken*> >
  GetInitTokenMap(const Function& func) {
140
    node_device_map_ = CollectDeviceInfo(func);
141 142 143 144 145 146 147 148 149 150
    this->Run(func);
    return std::move(token_map_);
  }

 protected:
  using StorageAllocaBaseVisitor::VisitExpr_;

  void CreateToken(const ExprNode* op, bool can_realloc)  final {
    CHECK(!token_map_.count(op));
    std::vector<StorageToken*> tokens;
151 152 153
    int device_type = node_device_map_.count(GetRef<Expr>(op))
                          ? node_device_map_[GetRef<Expr>(op)]->value
                          : 0;
154 155 156 157 158 159
    if (const auto* tuple_type = op->checked_type().as<TupleTypeNode>()) {
      for (Type t : tuple_type->fields) {
        const auto* ttype = t.as<TensorTypeNode>();
        CHECK(ttype);
        StorageToken* token = arena_->make<StorageToken>();
        token->ttype = ttype;
160
        token->device_type = device_type;
161 162 163 164 165 166 167
        tokens.push_back(token);
      }
    } else {
      const auto* ttype = op->checked_type().as<TensorTypeNode>();
      CHECK(ttype);
      StorageToken* token = arena_->make<StorageToken>();
      token->ttype = ttype;
168
      token->device_type = device_type;
169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187
      tokens.push_back(token);
    }
    token_map_[op] = tokens;
  }

  void VisitExpr_(const CallNode* op) final {
    // create token for the call node.
    CreateToken(op, true);
    // for each input, visit argument token.
    for (Expr arg : op->args) {
      for (StorageToken* tok : GetToken(arg)) {
        tok->ref_counter += 1;
      }
    }
  }

 private:
  // allocator
  common::Arena* arena_;
188
  Map<Expr, Integer> node_device_map_;
189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204
};

class StorageAllocator : public StorageAllocaBaseVisitor {
 public:
  /*!
   * \return totoal number of bytes allocated
   */
  size_t TotalAllocBytes() const {
    size_t total = 0;
    for (const auto* p : data_) {
      total += p->max_bytes;
    }
    return total;
  }

  // Run storage allocation for a function.
205
  Map<Expr, Array<IntegerArray> > Plan(const Function& func) {
206 207 208
    prototype_ = StorageAllocaInit(&arena_).GetInitTokenMap(func);
    this->Run(func);

209 210 211 212 213
    // The value of smap contains two integer arrays where the first array
    // contains the planned storage ids and the second holds the device types.
    Map<Expr, Array<IntegerArray> > smap;
    int num_annotated_nodes = 0;
    int num_nodes = 0;
214 215

    for (const auto& kv : token_map_) {
216 217
      std::vector<Integer> storage_ids;
      std::vector<Integer> device_types;
218
      for (StorageToken* tok : kv.second) {
219 220 221 222 223 224
        if (tok->device_type) {
          num_annotated_nodes++;
        }
        num_nodes++;
        storage_ids.push_back(tok->storage_id);
        device_types.push_back(tok->device_type);
225
      }
226 227 228 229 230 231 232 233
      smap.Set(GetRef<Expr>(kv.first), Array<IntegerArray>({storage_ids, device_types}));
    }
    // Either all or none of the nodes should be annotated.
    if (num_annotated_nodes != 0 && num_annotated_nodes != num_nodes) {
      LOG(FATAL)
          << num_annotated_nodes << " out of " << num_nodes
          << "expressions are assigned with virtual device types. Either all "
             "or none of the expressions are expected to be annotated.";
234 235 236 237 238 239 240
    }
    return smap;
  }

 protected:
  using StorageAllocaBaseVisitor::VisitExpr_;
  // override create token by getting token as prototype requirements.
241
  void CreateToken(const ExprNode* op, bool can_realloc) final {
242 243 244 245 246 247 248 249 250 251
    CHECK(!token_map_.count(op));
    auto it = prototype_.find(op);
    CHECK(it != prototype_.end());
    std::vector<StorageToken*> tokens;
    for (StorageToken* tok : it->second) {
      if (can_realloc) {
        tokens.push_back(Request(tok));
      } else {
        // Allocate a new token,
        StorageToken* allocated_tok = Alloc(tok, GetMemorySize(tok));
252
        allocated_tok->device_type = tok->device_type;
253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301
        // ensure it never get de-allocated.
        allocated_tok->ref_counter += 1;
        tokens.push_back(allocated_tok);
      }
    }
    token_map_[op] = tokens;
  }
  // The call map
  void VisitExpr_(const CallNode* op) final {
    std::vector<StorageToken*> args;
    // for each input, visit argument token.
    for (Expr arg : op->args) {
      for (StorageToken* tok : GetToken(arg)) {
        args.push_back(tok);
      }
    }
    // create token for the call node.
    CreateToken(op, true);
    // check if there is orphaned output that can be released immediately.
    for (StorageToken* tok : token_map_.at(op)) {
      CheckForRelease(tok);
    }
    for (StorageToken* tok : args) {
      tok->ref_counter -= 1;
      CheckForRelease(tok);
    }
  }
  /*!
   * \brief ceil(size/word_size) to get number of words.
   * \param size The original size.
   * \param word_size The element size.
   */
  static size_t DivRoundUp(size_t size, size_t word_size) {
    return (size + word_size - 1) / word_size;
  }
  /*!
   * \brief Get the memory requirement.
   * \param prototype The prototype token.
   * \return The required memory size.
   */
  size_t GetMemorySize(StorageToken* prototype) {
    const TensorTypeNode* ttype = prototype->ttype;
    CHECK(ttype != nullptr);
    size_t size = 1;
    for (IndexExpr dim : ttype->shape) {
      const int64_t* pval = as_const_int(dim);
      CHECK(pval != nullptr)
          << "Cannot allocate memory symbolic tensor shape "
          << ttype->shape;
302 303 304
      CHECK_GE(*pval, 0)
          << "Cannot allocate memory for tensor with negative shape"
          << *pval;
305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327
      size *= static_cast<size_t>(pval[0]);
    }
    size *= DivRoundUp(ttype->dtype.bits() * ttype->dtype.lanes(), 8);
    return size;
  }
  /*!
   * \brief Request a storage token for a given prototype.
   * \param prototype. The prototype storage token.
   * \return The result token.
   */
  StorageToken* Request(StorageToken* prototype) {
    // calculate the size;
    size_t size = GetMemorySize(prototype);
    // search memory block in [size / match_range_, size * match_range_)
    if (match_range_ == 0) {
      return this->Alloc(prototype, 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) {
      StorageToken *tok = it->second;
328
      if (tok->device_type != prototype->device_type) continue;
329 330 331 332 333 334 335 336 337 338 339 340
      CHECK_EQ(tok->ref_counter, 0);
      // Use exect matching strategy
      tok->max_bytes = std::max(size, tok->max_bytes);
      tok->ref_counter = prototype->ref_counter;
      // find a exact match, erase from map and return
      free_.erase(it);
      return tok;
    }
    // then search for memory blocks smaller than requested space
    for (auto it = mid; it != begin;) {
      --it;
      StorageToken *tok = it->second;
341
      if (tok->device_type != prototype->device_type) continue;
342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388
      CHECK_EQ(tok->ref_counter, 0);
      // Use exect matching strategy
      tok->max_bytes = std::max(size, tok->max_bytes);
      tok->ref_counter = prototype->ref_counter;
      // erase from map and return
      free_.erase(it);
      return tok;
    }
    // cannot find anything return a new one.
    return this->Alloc(prototype, size);
  }
  /*!
   * \brief Allocate a storage token by consuming prototype
   * \param prototype The prototype token.
   * \param size The size of memory being requested.
   */
  StorageToken* Alloc(StorageToken* prototype, size_t size) {
    prototype->max_bytes = size;
    prototype->storage_id = static_cast<int64_t>(data_.size());
    data_.push_back(prototype);
    return prototype;
  }
  /*!
   * \brief Check if we can release token.
   * \tok The token to be released.
   */
  void CheckForRelease(StorageToken* tok) {
    CHECK_GE(tok->storage_id, 0);
    CHECK_GE(tok->ref_counter, 0);
    if (tok->ref_counter == 0) {
      free_.insert({tok->max_bytes, tok});
    }
  }

 private:
  // allocator
  common::Arena arena_;
  // scale used for rough match
  size_t match_range_{16};
  // free list of storage entry
  std::multimap<size_t, StorageToken*> free_;
  // all the storage resources available
  std::vector<StorageToken*> data_;
  /*! \brief internal prototype token map */
  std::unordered_map<const ExprNode*, std::vector<StorageToken*> > prototype_;
};

389
Map<Expr, Array<IntegerArray> > GraphPlanMemory(const Function& func) {
390 391 392 393
  return StorageAllocator().Plan(func);
}

TVM_REGISTER_GLOBAL("relay.backend.GraphPlanMemory")
394
.set_body_typed<Map<Expr, Array<IntegerArray> >(const Function&)>(GraphPlanMemory);
395 396 397

}  // namespace relay
}  // namespace tvm