storage_rewrite.cc 33.4 KB
Newer Older
1 2 3 4 5 6 7 8 9 10
/*!
 * Copyright (c) 2017 by Contributors
 * \file storage_rewrite.cc
 * \brief Memory access pattern analysis and optimization.
 *  Re-write data access to enable memory sharing when possible.
 */
#include <tvm/ir.h>
#include <tvm/ir_pass.h>
#include <tvm/ir_mutator.h>
#include <tvm/ir_visitor.h>
11
#include <tvm/target_info.h>
12 13 14 15
#include <map>
#include <unordered_set>
#include <unordered_map>
#include "./ir_util.h"
16
#include "../arithmetic/compute_expr.h"
17
#include "../runtime/thread_storage_scope.h"
18 19 20 21

namespace tvm {
namespace ir {

22
using runtime::StorageRank;
23 24
using runtime::StorageScope;

25
// Find a linear pattern of storage acess
26
// Used for liveness analysis.
27 28 29 30 31 32 33 34 35 36 37 38
// Composite scopes(loop/thread_launch/IfThen) is represented by two points:
// before_scope -> scope_body -> after_scope
//
// The linear_seq_ stores before_scope and after_scope.
// The access to the arrays are stored at the after_scope point.
//
// Define "scope" as the body of For/thread_launch/IfThenElse
// This pass tries to detect last point that we need to keep memory
// alive under the same scope as allocate.
// The storage need to be kept alive between allocate and last access.
// The free point is only inserted at the same scope of allocate.
//
39
class LinearAccessPatternFinder final : public IRVisitor {
40
 public:
41 42 43 44
  /*! \brief record the touch hist of statment. */
  struct StmtEntry {
    // The statment
    const Node* stmt;
45 46 47 48 49
    // The index in the linear_seq_ to point to end of the nested scope.
    // This is only set to non-zero if stmt is a nested scope.
    // if offset > 0, means this is the begin, the end entry is current_index + offset
    // if offset < 0, means this is the end, the begin entry is current_index + offset
    int64_t scope_pair_offset{0};
50 51 52
    // The buffer variables this statment touched.
    std::vector<const Variable*> touched;
  };
53 54 55 56 57 58 59 60 61
  // The scope of each allocation
  struct AllocEntry {
    // Scope used for allocation.
    StorageScope storage_scope;
    // scope level
    size_t level{0};
    // allocation stmt
    const Allocate* alloc{nullptr};
  };
62

63 64 65
  void Visit_(const Allocate* op) final {
    size_t level = scope_.size();
    const Variable* buf = op->buffer_var.get();
66 67 68 69 70
    auto it = alloc_info_.find(buf);
    CHECK(it != alloc_info_.end());
    CHECK(it->second.alloc == nullptr);
    it->second.alloc = op;
    it->second.level = level;
71 72 73 74 75 76 77 78
    IRVisitor::Visit_(op);
  }
  void Visit_(const Store* op) final {
    scope_.push_back(StmtEntry());
    // visit subexpr
    IRVisitor::Visit_(op);
    // Add write access.
    const Variable* buf = op->buffer_var.get();
79 80 81 82
    auto it = alloc_info_.find(buf);
    if (it != alloc_info_.end() && it->second.alloc) {
      CHECK_LT(it->second.level, scope_.size());
      scope_[it->second.level].touched.push_back(buf);
83 84 85
    }
    StmtEntry e = scope_.back();
    scope_.pop_back();
86
    if (e.touched.size() != 0) {
87 88 89 90
      e.stmt = op;
      linear_seq_.push_back(e);
    }
  }
91 92 93 94 95 96
  void Visit_(const Evaluate* op) final {
    scope_.push_back(StmtEntry());
    // visit subexpr
    IRVisitor::Visit_(op);
    StmtEntry e = scope_.back();
    scope_.pop_back();
97
    if (e.touched.size() != 0) {
98 99 100 101
      e.stmt = op;
      linear_seq_.push_back(e);
    }
  }
102 103 104 105
  void Visit_(const Load* op) final {
    // Add write access.
    IRVisitor::Visit_(op);
    const Variable* buf = op->buffer_var.get();
106 107 108
    auto it = alloc_info_.find(buf);
    if (it != alloc_info_.end() && it->second.alloc) {
      CHECK_LT(it->second.level, scope_.size())
109
          << "Load memory in places other than store.";
110
      scope_[it->second.level].touched.push_back(buf);
111 112
    }
  }
113 114 115 116 117 118 119 120
  void Visit_(const Call* op) final {
    if (op->is_intrinsic(intrinsic::tvm_address_of)) {
      const Load* l = op->args[0].as<Load>();
      this->Visit(l->index);
    } else {
      IRVisitor::Visit_(op);
    }
  }
121 122
  void Visit_(const Variable* buf) final {
    // Directly reference to the variable count as a read.
123 124 125 126 127
    auto it = alloc_info_.find(buf);
    if (it != alloc_info_.end() && it->second.alloc) {
      CHECK_LT(it->second.level, scope_.size())
          << " buf=" << buf->name_hint;
      scope_[it->second.level].touched.push_back(buf);
128 129 130 131 132 133 134
    }
  }
  template<typename T>
  void VisitNewScope(const T* op) {
    scope_.push_back(StmtEntry());
    StmtEntry e;
    e.stmt = op;
135
    int64_t begin_index =  static_cast<int64_t>(linear_seq_.size());
136 137 138 139
    // before scope.
    linear_seq_.push_back(e);
    IRVisitor::Visit_(op);
    // after scope.
140
    e.touched = std::move(scope_.back().touched);
141
    scope_.pop_back();
142 143 144
    int64_t end_index =  static_cast<int64_t>(linear_seq_.size());
    CHECK_GT(end_index, begin_index);
    e.scope_pair_offset = begin_index - end_index;
145
    linear_seq_.push_back(e);
146 147 148
    // record the pointer to end index.
    CHECK_NE(end_index, 0U);
    linear_seq_[begin_index].scope_pair_offset = end_index - begin_index;
149 150 151 152 153 154 155
  }
  void Visit_(const AttrStmt* op) final {
    // Only record the outer most thread extent.
    if (op->attr_key == attr::thread_extent && !in_thread_env_) {
      in_thread_env_ = true;
      VisitNewScope(op);
      in_thread_env_ = false;
156 157
    } else if (op->attr_key == attr::extern_scope) {
      VisitNewScope(op);
158 159
    } else if (op->attr_key == attr::virtual_thread) {
      VisitNewScope(op);
160 161
    } else if (op->attr_key == attr::storage_scope) {
      const Variable* buf = op->node.as<Variable>();
162
      alloc_info_[buf].storage_scope =
163 164 165 166 167 168 169 170 171 172
          StorageScope::make(op->value.as<StringImm>()->value);
      IRVisitor::Visit_(op);
    } else {
      IRVisitor::Visit_(op);
    }
  }
  void Visit_(const IfThenElse* op) final {
    VisitNewScope(op);
  }

173 174 175 176
  void Visit_(const For* op) final {
    VisitNewScope(op);
  }

177 178 179 180 181
  // linearized access sequence.
  std::vector<StmtEntry> linear_seq_;
  // The storage scope of each buffer
  std::unordered_map<const Variable*, AllocEntry> alloc_info_;

182 183 184 185 186
 private:
  // Whether already in thread env.
  bool in_thread_env_{false};
  // The scope stack.
  std::vector<StmtEntry> scope_;
187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309
};

// Verify if the statement can be run safely via inplace fashion
//
// Detect pattern: dst[index] = f(src[index])
//
// WARNING: the current detection algorithm cannot handle the case
// when a location in an array is written multiple times
//
// For example, the following program will pass the check,
// but we cannot make A and B to be the same array.
//
//  A[0] = B[0] + 1
//  A[0] = B[0] + 1
//
// The high level code generator needs to ensure that the generated
// code only write each location of the target array once.
//
// This is the case with IR generated by the current compute schedule.
// We explicitly return false if we find there is an extern block
// which can be arbitrary IR.
//
// Neve-the-less, inplace detector should be used with care in mind.
// We may also consider introduce a condition checker that checks
// if every index only visited once for an absolute sufficient condition.
//
// The code after inplace transformation is no longer idempotent.
//
class InplaceOpVerifier : public IRVisitor {
 public:
  bool Check(const Node* stmt,
             const Variable* dst,
             const Variable* src) {
    dst_ = dst;
    src_ = src;
    result_ = true;
    if (stmt->is_type<AttrStmt>()) {
      Visit_(static_cast<const AttrStmt*>(stmt));
    } else if (stmt->is_type<For>()) {
      Visit_(static_cast<const For*>(stmt));
    } else if (stmt->is_type<IfThenElse>()) {
      Visit_(static_cast<const IfThenElse*>(stmt));
    } else if (stmt->is_type<Store>()) {
      Visit_(static_cast<const Store*>(stmt));
    } else {
      return false;
    }
    return result_;
  }

  using IRVisitor::Visit_;

  void Visit(const NodeRef& e) final {
    if (!result_) return;
    IRVisitor::Visit(e);
  }

  void Visit_(const Variable* op) final {
    // assume all opaque access is unsafe
    if (op == dst_ || op == src_) {
      result_ = false; return;
    }
  }

  void Visit_(const Store* op) final {
    ++mem_nest_;
    this->Visit(op->index);
    --mem_nest_;
    if (op->buffer_var.get() == dst_) {
      store_ = op;
      this->Visit(op->value);
      this->Visit(op->predicate);
      store_ = nullptr;
    } else {
      this->Visit(op->value);
      this->Visit(op->predicate);
    }
  }

  void Visit_(const AttrStmt* op) final {
    // always reject extern code
    if (op->attr_key == attr::extern_scope ||
        op->attr_key == attr::volatile_scope) {
      result_ = false; return;
    }
    IRVisitor::Visit_(op);
  }

  void Visit_(const Load* op) final {
    const Variable* buf = op->buffer_var.get();
    // cannot read from dst_ (no reduction)
    if (buf == dst_) {
      result_ = false; return;
    }
    // do not allow indirect memory load
    if (mem_nest_ != 0) {
      result_ = false; return;
    }
    if (src_ == buf) {
      if (store_ == nullptr ||
          store_->value.type() != op->type ||
          !ir::Equal(store_->index, op->index)) {
        result_ = false; return;
      }
    }
    ++mem_nest_;
    IRVisitor::Visit_(op);
    --mem_nest_;
  }


 private:
  // result of the check
  bool result_{true};
  // destination memory
  const Variable* dst_;
  // source variable
  const Variable* src_;
  // counter of load,
  // it is not safe to inplace when there is nested load like A[B[i]]
  int mem_nest_{0};
  // The current store to be inspected
  const Store* store_{nullptr};
310 311 312 313 314
};

// Planner to plan and rewrite memory allocation.
class StoragePlanRewriter : public IRMutator {
 public:
315
  using StmtEntry = LinearAccessPatternFinder::StmtEntry;
316
  using AllocEntry = LinearAccessPatternFinder::AllocEntry;
317

318 319 320 321 322 323 324
  Stmt Rewrite(Stmt stmt, bool detect_inplace) {
    detect_inplace_ = detect_inplace;
    // plan the rewrite
    LinearAccessPatternFinder finder;
    finder.Visit(stmt);
    this->LivenessAnalysis(finder.linear_seq_);
    this->PlanMemory(finder.linear_seq_, finder.alloc_info_);
325
    this->PrepareNewAlloc();
326
    // start rewrite
327 328 329 330
    stmt = this->Mutate(stmt);
    if (attach_map_.count(nullptr)) {
      std::vector<Stmt> nest;
      for (StorageEntry* e : attach_map_.at(nullptr)) {
331
        // CHECK_EQ(e->scope.rank, 0);
332 333 334 335 336 337 338
        if (e->new_alloc.defined()) {
          nest.emplace_back(AttrStmt::make(
              e->alloc_var, attr::storage_scope,
              StringImm::make(e->scope.to_string()),
              Evaluate::make(0)));
          nest.push_back(e->new_alloc);
        }
339 340 341 342 343 344 345 346 347 348
      }
      stmt = MergeNest(nest, stmt);
    }
    return stmt;
  }
  Stmt Mutate_(const Store* op, const Stmt& s) final {
    Stmt stmt = IRMutator::Mutate_(op, s);
    op = stmt.as<Store>();
    auto it = alloc_map_.find(op->buffer_var.get());
    if (it == alloc_map_.end()) return stmt;
349 350 351 352
    return Store::make(it->second->alloc_var,
                       op->value,
                       RemapIndex(op->value.type(), op->index, it->second),
                       op->predicate);
353 354 355 356 357 358
  }
  Expr Mutate_(const Load* op, const Expr& e) final {
    Expr expr = IRMutator::Mutate_(op, e);
    op = expr.as<Load>();
    auto it = alloc_map_.find(op->buffer_var.get());
    if (it == alloc_map_.end()) return expr;
359 360 361 362
    return Load::make(op->type,
                      it->second->alloc_var,
                      RemapIndex(op->type, op->index, it->second),
                      op->predicate);
363 364 365 366
  }
  Expr Mutate_(const Variable* op, const Expr& e) final {
    auto it = alloc_map_.find(op);
    if (it != alloc_map_.end()) {
367
      if (it->second->bits_offset != 0) {
368 369
        LOG(WARNING) << "Use a merged buffer variable address, could cause error";
      }
370 371 372 373 374
      return it->second->alloc_var;
    } else {
      return e;
    }
  }
375 376 377 378 379 380 381
  Expr Mutate_(const Call* op, const Expr& e) final {
    if (op->is_intrinsic(intrinsic::tvm_access_ptr)) {
      CHECK_EQ(op->args.size(), 5U);
      Type dtype = op->args[0].type();
      const Variable* buffer = op->args[1].as<Variable>();
      auto it = alloc_map_.find(buffer);
       if (it == alloc_map_.end()) return IRMutator::Mutate_(op, e);
382
       const StorageEntry* se = it->second;
383 384
       Expr offset = Mutate(op->args[2]);
       Expr extent = Mutate(op->args[3]);
385 386 387 388
       uint64_t elem_bits = dtype.bits() * dtype.lanes();
       CHECK_EQ(se->bits_offset % elem_bits, 0U);
       if (se->bits_offset != 0) {
         offset = make_const(offset.type(), se->bits_offset / elem_bits) + offset;
389 390 391
       }
       return Call::make(
           op->type, op->name,
392
           {op->args[0], se->alloc_var, offset, extent, op->args[4]},
393 394 395 396 397
           op->call_type);
    } else {
      return IRMutator::Mutate_(op, e);
    }
  }
398

399 400 401
  Stmt Mutate_(const AttrStmt* op, const Stmt& s) final {
    if (op->attr_key == attr::storage_scope) {
      return this->Mutate(op->body);
402
    } else if (op->attr_key == attr::thread_extent ||
403
               op->attr_key == attr::virtual_thread ||
404
               attr::IsPragmaKey(op->attr_key)) {
405
      // remake all the allocation at the attach scope.
406
      if (attach_map_.count(op)) {
407
        auto& svec = attach_map_[op];
408 409 410
        Stmt stmt = IRMutator::Mutate_(op, s);
        op = stmt.as<AttrStmt>();
        return AttrStmt::make(
411 412
            op->node, op->attr_key, op->value,
            MakeAttach(svec, op->body));
413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429
      } else {
        return IRMutator::Mutate_(op, s);
      }
    } else if (op->attr_key == attr::volatile_scope) {
      Stmt stmt = IRMutator::Mutate_(op, s);
      op = stmt.as<AttrStmt>();
      auto it = alloc_map_.find(op->node.as<Variable>());
      if (it == alloc_map_.end()) return stmt;
      return AttrStmt::make(
          it->second->alloc_var, op->attr_key, op->value, op->body);
    } else {
      return IRMutator::Mutate_(op, s);
    }
  }
  Stmt Mutate_(const For* op, const Stmt& s) final {
    CHECK(op->for_type != ForType::Vectorized)
        << "VectorizeLoop before LiftStorageAlloc";
430 431 432 433 434 435 436 437 438 439 440
    // remake all the allocation at the attach scope.
    if (attach_map_.count(op)) {
      auto& svec = attach_map_[op];
      Stmt stmt = IRMutator::Mutate_(op, s);
      op = stmt.as<For>();
      return For::make(
          op->loop_var, op->min, op->extent, op->for_type, op->device_api,
          MakeAttach(svec, op->body));
    } else {
      return IRMutator::Mutate_(op, s);
    }
441
  }
442

443 444 445 446 447 448 449 450 451 452
  Stmt Mutate_(const Allocate* op, const Stmt& s) final {
    return this->Mutate(op->body);
  }

 private:
  struct StorageEntry {
    // The scope that this alloc attaches after
    // For shared/local memory it is beginning of the thread extent.
    // for global memory it is nullptr, means beginning of everything.
    const Node* attach_scope_{nullptr};
453
    // The constant size of the buffer in bits, only used if it is constant
454
    uint64_t const_nbits{0};
455 456 457 458
    // The storage scope.
    StorageScope scope;
    // Allocs that shares this entry.
    std::vector<const Allocate*> allocs;
459 460 461 462
    // The children of this entry, not including itself.
    std::vector<StorageEntry*> merged_children;
    // The replacement allocation, if any.
    Stmt new_alloc;
463 464
    // The var expr of new allocation.
    VarExpr alloc_var;
465 466 467
    // The allocation element type.
    Type elem_type;
    // This is non-zero if this allocate is folded into another one
468 469 470 471 472 473 474 475 476 477 478
    // the address(in bits) becomes alloc_var + bits_offset;
    // can be effectively converted to the element type.
    // We need to convert bit_offset to offset of specific element type later.
    //
    // We use bits(instead of bytes) to support non-conventional indexing in hardware.
    // When we are merging buffer together, the bits_offset are set to be aligned
    // to certain value given by the max_simd_bits property of the special memory.
    //
    // This allows effective sharing among different types as long as their alignment
    // requirement fits into the max_simd_bits.
    uint64_t bits_offset{0};
479
  };
480 481 482 483 484 485 486 487 488 489

  // Alllocate entry of node.
  // Event entry in liveness analysis
  struct EventEntry {
    // variables we generate
    std::vector<const Variable*> gen;
    // variables we kill
    std::vector<const Variable*> kill;
  };

490 491 492 493
  Stmt MakeAttach(const std::vector<StorageEntry*>& svec,
                  Stmt body) {
    std::vector<Stmt> nest;
    for (StorageEntry* e : svec) {
494 495 496 497 498 499 500
      if (e->new_alloc.defined()) {
        nest.emplace_back(AttrStmt::make(
            e->alloc_var, attr::storage_scope,
            StringImm::make(e->scope.to_string()),
            Evaluate::make(0)));
        nest.push_back(e->new_alloc);
      }
501 502 503
    }
    return MergeNest(nest, body);
  }
504 505
  // Remap the index
  Expr RemapIndex(Type dtype, Expr index, StorageEntry* e) {
506 507 508 509
    if (e->bits_offset == 0) return index;
    uint64_t elem_bits = dtype.bits() * dtype.lanes();
    CHECK_EQ(e->bits_offset % elem_bits, 0U);
    return make_const(index.type(), e->bits_offset / elem_bits) + index;
510
  }
511 512 513 514
  // Prepare the new allocations
  void PrepareNewAlloc() {
    for (size_t i = 0; i < alloc_vec_.size(); ++i) {
      StorageEntry* e = alloc_vec_[i].get();
515 516 517 518
      attach_map_[e->attach_scope_].push_back(e);
    }
    // find allocation via attach map.
    for (auto &kv : attach_map_) {
519
      // find the element with the most amount of bytes.
520 521 522 523 524 525 526 527 528 529 530 531 532
      std::vector<StorageEntry*>& vec = kv.second;
      // try to find merge, for tagged memory
      for (size_t i = 0; i < vec.size(); ++i) {
        StorageEntry* e = vec[i];
        if (e->scope.tag.length() != 0) {
          CHECK_NE(e->const_nbits, 0U)
              << "Special tagged memory must be const size";
          for (size_t j = 0; j < i; ++j) {
            if (e->scope == vec[j]->scope) {
              vec[j]->merged_children.push_back(e);
              break;
            }
          }
533 534
        }
      }
535 536 537 538
      // Start allocation
      for (size_t i = 0; i < vec.size(); ++i) {
        StorageEntry* e = vec[i];
        // already merged
539
        if (e->bits_offset != 0) continue;
540 541 542 543 544 545
        if (e->merged_children.size() != 0) {
          NewAllocTagMerged(e); continue;
        }
        // Get the allocation size;
        e->alloc_var = e->allocs[0]->buffer_var;
        Type alloc_type = e->allocs[0]->type;
546
        for (const Allocate* op : e->allocs) {
547 548
          if (op->type.lanes() > alloc_type.lanes()) {
            alloc_type = op->type;
549
          }
550 551 552 553 554 555
        }
        if (e->allocs.size() == 1) {
          // simply use the original allocation.
          e->new_alloc = Allocate::make(
              e->alloc_var, alloc_type, e->allocs[0]->extents,
              e->allocs[0]->condition, Evaluate::make(0));
556 557 558 559 560 561
          if (e->scope.tag.length() != 0) {
            MemoryInfo info = GetMemoryInfo(e->scope.to_string());
            uint64_t total_elem = e->const_nbits / e->elem_type.bits();
            CHECK_LE(total_elem * e->elem_type.bits(), info->max_num_bits)
                << "Allocation exceed bound of memory tag " << e->scope.to_string();
          }
562 563 564 565
        } else {
          // Build a merged allocation
          Expr combo_size;
          for (const Allocate* op : e->allocs) {
566
            Expr sz = arith::ComputeReduce<Mul>(op->extents, make_const(Int(32), 1));
567 568
            // transform to bits
            auto sz_nbits = sz * (op->type.bits() * op->type.lanes());
569
            if (combo_size.defined()) {
570
              combo_size = max(combo_size, sz_nbits);
571
            } else {
572
              combo_size = sz_nbits;
573
            }
574
          }
575 576 577 578 579 580 581 582
          // transform to alloc bytes
          auto type_bits = alloc_type.bits() * alloc_type.lanes();
          bool divided = can_prove(combo_size % type_bits == 0);
          combo_size = combo_size / type_bits;
          // round up for can not divided
          if (!divided) {
             combo_size += make_const(Int(32), 1);
          }
583 584 585 586
          combo_size = ir::Simplify(combo_size);
          e->new_alloc = Allocate::make(
              e->alloc_var, alloc_type, {combo_size}, const_true(),
              Evaluate::make(0));
587 588
        }
      }
589 590 591 592 593 594 595 596
    }
  }
  // New allocation for merged data
  void NewAllocTagMerged(StorageEntry* e) {
    CHECK_NE(e->scope.tag.length(), 0U);
    // allocate with element type.
    CHECK_NE(e->const_nbits, 0U);
    MemoryInfo info = GetMemoryInfo(e->scope.to_string());
597
    uint64_t total_bits = e->const_nbits;
598 599
    // By default, align to 32 bits.
    size_t align = 32;
600
    if (info.defined()) {
601
      align = info->max_simd_bits;
602
    }
603 604
    // Always align to max_simd_bits
    // so we can remap types by keeping this property
605 606
    if (total_bits % align != 0) {
      total_bits += align  - (total_bits % align);
607 608 609
    }
    e->alloc_var = e->allocs[0]->buffer_var;
    for (StorageEntry* child : e->merged_children) {
610 611
      CHECK_NE(child->const_nbits, 0U);
      CHECK_NE(total_bits, 0U);
612
      child->bits_offset = total_bits;
613
      child->alloc_var = e->alloc_var;
614 615 616
      total_bits += child->const_nbits;
      if (total_bits % align != 0) {
        total_bits += align  - (total_bits % align);
617 618
      }
    }
619
    uint64_t type_bits = e->elem_type.bits() * e->elem_type.lanes();
620
    Expr alloc_size = make_const(e->allocs[0]->extents[0].type(),
621
                                 (total_bits + type_bits - 1) / type_bits);
622 623 624 625
    e->new_alloc = Allocate::make(
        e->alloc_var, e->elem_type, {alloc_size}, const_true(),
        Evaluate::make(0));
    if (info.defined()) {
626
      CHECK_LE(total_bits, info->max_num_bits)
627
          << "Allocation exceed bound of memory tag " << e->scope.to_string();
628 629
    }
  }
630 631 632
  // Liveness analysis to find gen and kill point of each variable.
  void LivenessAnalysis(const std::vector<StmtEntry>& seq) {
    // find kill point, do a reverse linear scan.
633 634 635
    std::unordered_set<const Variable*> touched;
    for (size_t i = seq.size(); i != 0; --i) {
      const StmtEntry& s = seq[i - 1];
636 637 638
      for (const Variable* buffer : s.touched) {
        if (!touched.count(buffer)) {
          touched.insert(buffer);
639 640 641 642 643 644 645 646 647 648 649 650 651 652
          event_map_[s.stmt].kill.push_back(buffer);
        }
      }
    }
    // find gen point, do forward scan
    touched.clear();
    for (size_t i = 0; i < seq.size(); ++i) {
      int64_t offset = seq[i].scope_pair_offset;
      if (offset < 0) continue;
      const StmtEntry& s = seq[i + offset];
      for (const Variable* buffer : s.touched) {
        if (!touched.count(buffer)) {
          touched.insert(buffer);
          event_map_[s.stmt].gen.push_back(buffer);
653 654 655 656
        }
      }
    }
  }
657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680
  void PlanNewScope(const Node* op) {
    if (thread_scope_ != nullptr) {
      CHECK(thread_scope_ == op);
      // erase all memory atatched to this scope.
      for (auto it = const_free_map_.begin(); it != const_free_map_.end();) {
        if (it->second->attach_scope_ == op) {
          it = const_free_map_.erase(it);
        } else {
          ++it;
        }
      }
      for (auto it = sym_free_list_.begin(); it != sym_free_list_.end();) {
        if ((*it)->attach_scope_ == op) {
          it = sym_free_list_.erase(it);
        } else {
          ++it;
        }
      }
      thread_scope_ = nullptr;
    } else {
      thread_scope_ = op;
    }
  }

681
  // Memory plan algorithm
682 683 684 685
  void PlanMemory(const std::vector<StmtEntry>& seq,
                  const std::unordered_map<const Variable*, AllocEntry>& alloc_info) {
    std::unordered_set<const Variable*> inplace_flag;

686 687
    for (size_t i = 0; i < seq.size(); ++i) {
      const StmtEntry& s = seq[i];
688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704
      auto it = event_map_.find(seq[i].stmt);

      // scope_pair_offset >= 0 means it is either
      // - leaf stmt(offset = 0)
      // - beginning of scope(offset < 0)
      // In both cases, we need to handle the gen event correctly
      if (it != event_map_.end() && seq[i].scope_pair_offset >= 0) {
        // Inplace operation detection
        // specially handle this
        bool detect_inplace = detect_inplace_ && (it->second.gen.size() <= 2);

        for (const Variable* var : it->second.gen) {
          CHECK(alloc_info.count(var));
          const AllocEntry& ae = alloc_info.at(var);
          StorageEntry* dst_entry = nullptr;
          // inplace detection
          if (detect_inplace) {
705 706
            // only one inplace var for s.stmt
            bool inplace_found = false;
707 708 709 710 711 712 713 714 715 716 717 718
            for (const Variable* src : it->second.kill) {
              if (!inplace_flag.count(src) && alloc_map_.count(src)) {
                InplaceOpVerifier visitor;
                StorageEntry* src_entry = alloc_map_.at(src);
                if (src_entry->scope == ae.storage_scope &&
                    src_entry->attach_scope_ == thread_scope_ &&
                    src_entry->elem_type == ae.alloc->type.element_of() &&
                    visitor.Check(s.stmt, var, src)) {
                  uint64_t const_nbits = static_cast<uint64_t>(
                      ae.alloc->constant_allocation_size() *
                      ae.alloc->type.bits() *
                      ae.alloc->type.lanes());
719
                  if (src_entry->const_nbits == const_nbits && !inplace_found) {
720 721 722
                    // successfully inplace
                    dst_entry = src_entry;
                    inplace_flag.insert(src);
723
                    inplace_found = true;
724 725 726 727 728 729 730 731 732 733 734 735 736
                  }
                }
              }
            }
          }
          if (dst_entry == nullptr) {
            dst_entry = FindAlloc(ae.alloc, thread_scope_, ae.storage_scope);
          }
          dst_entry->allocs.emplace_back(ae.alloc);
          alloc_map_[var] = dst_entry;
        }
      }
      // enter/exit new scope
737 738
      if (s.stmt->is_type<AttrStmt>()) {
        const auto* op = static_cast<const AttrStmt*>(s.stmt);
739
        if (op->attr_key == attr::thread_extent ||
740 741
            op->attr_key == attr::virtual_thread ||
            attr::IsPragmaKey(op->attr_key)) {
742 743 744 745
          PlanNewScope(op);
        } else {
          CHECK(op->attr_key == attr::extern_scope);
        }
746 747 748 749 750
      } else if (s.stmt->is_type<For>()) {
        const auto* op = static_cast<const For*>(s.stmt);
        if (op->for_type == ForType::Parallel) {
          if (thread_scope_ == nullptr || thread_scope_ == op) {
            PlanNewScope(op);
751 752 753
          }
        }
      }
754 755 756 757 758 759 760 761 762 763
      // scope_pair_offset <= 0 means it is either
      // - leaf stmt(offset = 0)
      // - end of scope(offset < 0)
      // In both cases, we need to handle the kill event correctly
      if (it != event_map_.end() && seq[i].scope_pair_offset <= 0) {
        for (const Variable* var : it->second.kill) {
          // skip space which are already replaced by inplace
          if (!inplace_flag.count(var)) {
            this->Free(var);
          }
764 765 766 767 768 769
        }
      }
    }
  }
  // Allocate new storage entry.
  StorageEntry* NewAlloc(const Allocate* op,
770
                         const Node* attach_scope,
771
                         const StorageScope& scope,
772
                         size_t const_nbits) {
773
    CHECK(op != nullptr);
774 775
    // Re-use not successful, allocate a new buffer.
    std::unique_ptr<StorageEntry> entry(new StorageEntry());
776
    entry->attach_scope_ = attach_scope;
777
    entry->scope = scope;
778 779
    entry->elem_type = op->type.element_of();
    entry->const_nbits = const_nbits;
780 781 782 783
    StorageEntry* e = entry.get();
    alloc_vec_.emplace_back(std::move(entry));
    return e;
  }
784

785
  StorageEntry* FindAlloc(const Allocate* op,
786
                          const Node* attach_scope,
787
                          const StorageScope& scope) {
788
    CHECK(op != nullptr);
789 790
    // skip plan for local variable,
    // compiler can do a better job with register allocation.
791
    const uint64_t match_range = 16;
792
    uint64_t op_elem_bits = op->type.bits() * op->type.lanes();
793
    uint64_t const_nbits = static_cast<uint64_t>(
794
        op->constant_allocation_size() * op_elem_bits);
795
    // disable reuse of small arrays, they will be lowered to registers in LLVM
796 797
    // This rules only apply if we are using non special memory
    if (scope.tag.length() == 0) {
798
      if (scope.rank >= StorageRank::kWarp || op->type.is_handle()) {
799 800 801 802 803
        return NewAlloc(op, attach_scope, scope, const_nbits);
      }
      if (const_nbits > 0  &&  const_nbits <= 32) {
        return NewAlloc(op, attach_scope, scope, const_nbits);
      }
804
    }
805
    if (const_nbits != 0) {
806
      // constant allocation.
807 808 809
      auto begin = const_free_map_.lower_bound(const_nbits / match_range);
      auto mid = const_free_map_.lower_bound(const_nbits);
      auto end = const_free_map_.upper_bound(const_nbits * match_range);
810
      // start looking at the buffer that is bigger than the required size first
811 812
      for (auto it = mid; it != end; ++it) {
        StorageEntry *e = it->second;
813
        if (e->attach_scope_ != attach_scope) continue;
814
        if (e->scope != scope) continue;
815 816
        // when not divided, no reuse, eg, float4 vs float3
        if (e->bits_offset % op_elem_bits != 0) continue;
817
        e->const_nbits = std::max(const_nbits, e->const_nbits);
818 819 820
        const_free_map_.erase(it);
        return e;
      }
821
      // then start looking at smaller buffers.
822 823 824
      for (auto it = mid; it != begin;) {
        --it;
        StorageEntry *e = it->second;
825
        if (e->attach_scope_ != attach_scope) continue;
826 827
        if (e->scope != scope) continue;
        if (e->elem_type != op->type.element_of()) continue;
828
        e->const_nbits = std::max(const_nbits, e->const_nbits);
829 830 831 832 833 834 835 836
        const_free_map_.erase(it);
        return e;
      }
    } else {
      // Simple strategy: round roubin.
      for (auto it = sym_free_list_.begin();
           it != sym_free_list_.end(); ++it) {
        StorageEntry* e = *it;
837
        if (e->attach_scope_ != attach_scope) continue;
838
        if (e->scope != scope) continue;
839
        if (e->elem_type != op->type.element_of()) continue;
840 841 842 843
        sym_free_list_.erase(it);
        return e;
      }
    }
844
    return NewAlloc(op, attach_scope, scope, const_nbits);
845 846 847 848 849 850
  }
  // simulated free.
  void Free(const Variable* var) {
    auto it = alloc_map_.find(var);
    CHECK(it != alloc_map_.end());
    StorageEntry* e = it->second;
851
    CHECK_NE(e->allocs.size(), 0U);
852 853 854 855 856

    // disable reuse of small arrays, they will be lowered to registers in LLVM
    // This rules only apply if we are using non special memory
    if (e->scope.tag.length() == 0) {
      // Disable sharing of local memory.
857 858
      if (e->scope.rank >= StorageRank::kWarp ||
          e->allocs[0]->type.is_handle()) return;
859 860 861
      // disable reuse of small arrays
      if (e->const_nbits > 0 && e->const_nbits <= 32) return;
    }
862
    // normal free.
863 864
    if (e->const_nbits != 0) {
      const_free_map_.insert({e->const_nbits, e});
865 866 867 868 869 870
    } else {
      sym_free_list_.push_back(e);
    }
  }
  // thread scope.
  const Node* thread_scope_{nullptr};
871 872
  // whether enable inplace detection.
  bool detect_inplace_{false};
873
  // Locations of free ops.
874
  std::unordered_map<const Node*, EventEntry> event_map_;
875
  // constant size free map.
876
  std::multimap<uint64_t, StorageEntry*> const_free_map_;
877 878
  // symbolic free list, for non constant items.
  std::list<StorageEntry*> sym_free_list_;
879 880 881 882
  // The allocation attach map
  std::unordered_map<const Node*, std::vector<StorageEntry*> > attach_map_;
  // The allocation assign map
  std::unordered_map<const Variable*, StorageEntry*> alloc_map_;
883 884 885 886
  // The allocations
  std::vector<std::unique_ptr<StorageEntry> > alloc_vec_;
};

887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939
// Turn alloc into vector alloc
// if all its access is the same vector type.
class VectorAllocRewriter : public IRMutator {
 public:
  Expr Mutate_(const Load* op, const Expr& e) final {
    UpdateTypeMap(op->buffer_var.get(), op->type);
    return IRMutator::Mutate_(op, e);
  }

  Stmt Mutate_(const Store* op, const Stmt& s) final {
    UpdateTypeMap(op->buffer_var.get(), op->value.type());
    return IRMutator::Mutate_(op, s);
  }
  Expr Mutate_(const Call* op, const Expr& e) final {
    if (op->is_intrinsic(intrinsic::tvm_access_ptr)) {
      Type dtype = op->args[0].type();
      const Variable* buffer = op->args[1].as<Variable>();
      UpdateTypeMap(buffer, dtype);
    }
    return IRMutator::Mutate_(op, e);
  }

  Stmt Mutate_(const Allocate* op, const Stmt& s) final {
    Stmt stmt = IRMutator::Mutate_(op, s);
    op = stmt.as<Allocate>();
    const auto& tvec = acc_map_[op->buffer_var.get()];

    if (tvec.size() == 1 &&
        tvec[0].element_of() == op->type.element_of() &&
        tvec[0].lanes() % op->type.lanes() == 0 &&
        tvec[0].lanes() != op->type.lanes()) {
      int factor = tvec[0].lanes() / op->type.lanes();
      Array<Expr> extents = op->extents;
      arith::ModularEntry me = EvalModular(
          extents[extents.size() - 1],
          std::unordered_map<const Variable*, arith::ModularEntry>());
      if (me.base % factor == 0 && me.coeff % factor == 0) {
        extents.Set(extents.size() - 1,
                    extents[extents.size() - 1] / make_const(extents[0].type(), factor));
        return Allocate::make(
            op->buffer_var, tvec[0], extents,
            op->condition, op->body);
      }
    }
    return stmt;
  }

  void UpdateTypeMap(const Variable* buffer, Type t) {
    auto& tvec = acc_map_[buffer];
    if (std::find(tvec.begin(), tvec.end(), t) == tvec.end()) {
      tvec.push_back(t);
    }
  }
940

941
  // Internal access map
942
  std::unordered_map<const Variable*, std::vector<Type> > acc_map_;
943 944 945
};


946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969
LoweredFunc PointerValueTypeRewrite(LoweredFunc f) {
  std::shared_ptr<LoweredFuncNode> n =
      std::make_shared<LoweredFuncNode>(*f.operator->());
  VectorAllocRewriter rewriter;
  n->body = rewriter.Mutate(n->body);
  for (Var arg : f->args) {
    if (arg.type().is_handle()) {
      const auto& tvec = rewriter.acc_map_[arg.get()];
      if (tvec.size() == 1) {
        Expr dtype = make_const(tvec[0], 0);
        n->handle_data_type.Set(arg, dtype);
      } else {
        // always set data type to be non vectorized so
        // load/store can still work via scalarization
        if (tvec.size() != 0 && !n->handle_data_type.count(arg)) {
          Expr dtype = make_const(tvec[0].with_lanes(1), 0);
          n->handle_data_type.Set(arg, dtype);
        }
      }
    }
  }
  return LoweredFunc(n);
}

970
Stmt StorageRewrite(Stmt stmt) {
971
  stmt = StoragePlanRewriter().Rewrite(stmt, true);
972
  return VectorAllocRewriter().Mutate(stmt);
973 974 975
}
}  // namespace ir
}  // namespace tvm