storage_access.cc 9.99 KB
Newer Older
1 2 3 4
/*!
 *  Copyright (c) 2017 by Contributors
 * \file storage_access.cc
 */
5 6 7
#include <tvm/ir_pass.h>
#include <tvm/ir_mutator.h>
#include <tvm/target_info.h>
8 9 10
#include <string>
#include "ir_util.h"
#include "storage_access.h"
11
#include "../arithmetic/compute_expr.h"
12 13 14 15 16 17 18 19 20 21 22

namespace tvm {
namespace ir {

void StorageAccessVisitor::Visit_(const Load* op) {
  const Variable* buf = op->buffer_var.as<Variable>();
  StorageScope scope = GetScope(buf);
  if (Enabled(buf, scope)) {
    CHECK(allow_append_);
    AccessEntry e;
    e.threads = env_threads();
23
    e.buffer = op->buffer_var;
24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42
    e.dtype = op->type.element_of();
    e.touched = arith::IntSet::vector(op->index);
    e.type = kRead;
    e.scope = scope;
    curr_stmt_.access.emplace_back(std::move(e));
  }
  // traverse child
  IRVisitor::Visit_(op);
}

void StorageAccessVisitor::Visit_(const Store* op) {
  allow_append_ = true;
  CHECK_EQ(curr_stmt_.access.size(), 0U);
  curr_stmt_.stmt = op;
  const Variable* buf = op->buffer_var.as<Variable>();
  StorageScope scope = GetScope(buf);
  if (Enabled(buf, scope)) {
    AccessEntry e;
    e.threads = env_threads();
43
    e.buffer = op->buffer_var;
44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77
    e.dtype = op->value.type().element_of();
    e.touched = arith::IntSet::vector(op->index);
    e.type = kWrite;
    e.scope = scope;
    curr_stmt_.access.emplace_back(std::move(e));
  }
  // traverse child
  IRVisitor::Visit_(op);
  // push to the scope
  scope_.back().push_back(curr_stmt_);
  // clear access entry.
  curr_stmt_.access.clear();
  allow_append_ = false;
}

void StorageAccessVisitor::Visit_(const Evaluate* op) {
  allow_append_ = true;
  CHECK_EQ(curr_stmt_.access.size(), 0U);
  curr_stmt_.stmt = op;
  IRVisitor::Visit_(op);
  // push to the scope
  if (curr_stmt_.access.size() != 0) {
    scope_.back().push_back(curr_stmt_);
    curr_stmt_.access.clear();
  }
  allow_append_ = false;
}

void StorageAccessVisitor::Visit_(const AttrStmt* op) {
  if (op->attr_key == attr::storage_scope) {
    const Variable* buf = op->node.as<Variable>();
    storage_scope_[buf] =
        StorageScope::make(op->value.as<StringImm>()->value);
    IRVisitor::Visit_(op);
78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95
  } else if (op->attr_key == attr::double_buffer_write) {
    CHECK(double_buffer_write_ == nullptr);
    double_buffer_write_ = op->node.as<Variable>();
    scope_.push_back(std::vector<StmtEntry>());
    IRVisitor::Visit_(op);
    StmtEntry s;
    s.stmt = op;
    s.access = Summarize(std::move(scope_.back()), nullptr);
    scope_.pop_back();
    if (!s.access.empty()) {
      for (AccessEntry& e : s.access) {
        if (e.type == kWrite && e.buffer.get() == double_buffer_write_) {
          e.double_buffer_write = true;
        }
      }
      scope_.back().emplace_back(std::move(s));
    }
    double_buffer_write_ = nullptr;
96 97 98 99 100
  } else if (op->attr_key == attr::coproc_scope) {
    IterVar iv(op->node.node_);
    env_threads_.push_back(iv);
    IRVisitor::Visit_(op);
    env_threads_.CopyOnWrite()->data.pop_back();
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
  } else if (op->attr_key == attr::thread_extent) {
    IterVar iv(op->node.node_);
    env_threads_.push_back(iv);
    if (!in_device_env_) {
      in_device_env_ = true;
      scope_.push_back(std::vector<StmtEntry>());
      IRVisitor::Visit_(op);
      // no need to take the result as the thread barrier automatically syncs.
      Summarize(std::move(scope_.back()), nullptr);
      in_device_env_ = false;
      scope_.pop_back();
    } else {
      IRVisitor::Visit_(op);
    }
    env_threads_.CopyOnWrite()->data.pop_back();
  } else {
    IRVisitor::Visit_(op);
  }
}

void StorageAccessVisitor::Visit_(const For* op) {
  scope_.push_back(std::vector<StmtEntry>());
  IRVisitor::Visit_(op);
  StmtEntry s;
  s.stmt = op;
  s.access = Summarize(std::move(scope_.back()), op);
  scope_.pop_back();
  if (s.access.size() != 0) {
    // relax the touched set to contain all ranges in the loop.
    std::unordered_map<const Variable*, arith::IntSet> relax_map;
    relax_map[op->loop_var.get()] = arith::IntSet::range(
        Range::make_by_min_extent(op->min, op->extent));
    for (AccessEntry& e : s.access) {
134
      if (e.buffer.defined()) {
135 136 137 138
        CHECK(e.touched.defined());
        e.touched = arith::EvalSet(e.touched, relax_map);
      }
    }
139 140
  }
  if (!s.access.empty()) {
141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181
    scope_.back().emplace_back(std::move(s));
  }
}

void StorageAccessVisitor::Visit_(const IfThenElse* op) {
  ++condition_counter_;
  this->Visit(op->condition);
  scope_.push_back(std::vector<StmtEntry>());
  this->Visit(op->then_case);
  StmtEntry s;
  s.stmt = op;
  s.access = Summarize(std::move(scope_.back()), nullptr);
  scope_.pop_back();
  if (op->else_case.defined()) {
    scope_.push_back(std::vector<StmtEntry>());
    auto v = Summarize(std::move(scope_.back()), nullptr);
    scope_.pop_back();
    s.access.insert(s.access.end(), v.begin(), v.end());
  }
  scope_.back().emplace_back(std::move(s));
  --condition_counter_;
}

void StorageAccessVisitor::Visit_(const Call* op) {
  if (op->is_intrinsic(intrinsic::tvm_address_of)) {
    const Load *l = op->args[0].as<Load>();
    IRVisitor::Visit_(l);
  } else 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>();
    Expr offset = op->args[2];
    Expr extent = op->args[3];
    const IntImm* flag = op->args[4].as<IntImm>();
    StorageScope scope = GetScope(buffer);
    // The buffer scope.
    if (Enabled(buffer, scope)) {
      CHECK(allow_append_);
      AccessEntry e;
      e.threads = env_threads();
      e.dtype = dtype;
182
      e.buffer = VarExpr(op->args[1].node_);
183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213
      e.touched = arith::IntSet::range(
          Range::make_by_min_extent(offset, extent));
      e.scope = scope;
      if (flag->value & 1) {
        e.type = kRead;
        curr_stmt_.access.emplace_back(e);
      }
      if (flag->value & 2) {
        e.type = kWrite;
        curr_stmt_.access.emplace_back(e);
      }
    }
    IRVisitor::Visit_(op);
  } else if (op->is_intrinsic(intrinsic::tvm_storage_sync)) {
    CHECK(allow_append_);
    const std::string& s = op->args[0].as<StringImm>()->value;
    if (s != "warp") {
      StorageScope scope = StorageScope::make(s);
      AccessEntry e;
      e.threads = env_threads();
      e.type = kSync;
      e.scope = StorageScope::make(s);
      curr_stmt_.access.emplace_back(std::move(e));
    }
  } else {
    IRVisitor::Visit_(op);
  }
}

StorageScope StorageAccessVisitor::GetScope(const Variable* buf) const {
  auto it = storage_scope_.find(buf);
214 215
  StorageScope s;
  s.rank = StorageRank::kGlobal;
216 217 218
  if (it == storage_scope_.end()) return s;
  return it->second;
}
219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323

class StorageAccessInfoLower : public IRMutator {
 public:
  Stmt Mutate_(const Allocate* op, const Stmt& s) final {
    // Lower allocate to device allocate when needed.
    Stmt stmt = IRMutator::Mutate_(op, s);
    op = stmt.as<Allocate>();
    // For special memory, remove allocate, or use head expr
    auto it = storage_info_.find(op->buffer_var.get());
    if (it != storage_info_.end() && it->second.info.defined()) {
      const MemoryInfo& info = it->second.info;
      ++it->second.alloc_count;
      CHECK_LE(it->second.alloc_count, 1)
          << "Double allocation of " << it->second.scope.to_string();
      if (info->head_address.defined()) {
        return Allocate::make(
            op->buffer_var, op->type, op->extents, op->condition,
            op->body, info->head_address, "nop");
      }
      return op->body;
    } else {
      return stmt;
    }
  }
  Stmt Mutate_(const AttrStmt* op, const Stmt& s) final {
    if (op->attr_key == attr::storage_scope) {
      const Variable* buf = op->node.as<Variable>();
      StorageScope scope = StorageScope::make(op->value.as<StringImm>()->value);
      StorageEntry e;
      e.scope = scope;
      if (scope.tag.length() != 0) {
        e.info = GetMemoryInfo(op->value.as<StringImm>()->value);
        CHECK(e.info.defined()) << "Cannot find memory info of " << scope.to_string();
      }
      storage_info_[buf] = e;
      return IRMutator::Mutate_(op, s);

    } else {
      return IRMutator::Mutate_(op, s);
    }
  }

  Expr Mutate_(const Call* op, const Expr &e) final {
    if (op->is_intrinsic(intrinsic::tvm_access_ptr)) {
      return MakeAccessPtr(op, e);
    } else {
      return IRMutator::Mutate_(op, e);
    }
  }

 private:
  // tvm_access_ptr
  Expr MakeAccessPtr(const Call* op, const Expr& e) {
    // Specially handle the buffer packed intrinsic
    Expr expr = IRMutator::Mutate_(op, e);
    op = expr.as<Call>();
    CHECK_EQ(op->args.size(), 5U);
    Type dtype = op->args[0].type();
    const Variable* buffer = op->args[1].as<Variable>();
    Var buffer_var(op->args[1].node_);
    Expr offset = op->args[2];
    auto it = storage_info_.find(buffer);
    if (it != storage_info_.end() && it->second.info.defined()) {
      return MakeTaggedAccessPtr(
          op->type, buffer_var, dtype, offset,
          it->second.info);
    }
    CHECK(op->type.is_handle());
    // Change to address_of
    return AddressOffset(buffer_var, dtype, offset);
  }

  Expr MakeTaggedAccessPtr(Type ptr_type,
                           Var buffer_var,
                           Type dtype,
                           Expr offset,
                           const MemoryInfo& info) {
    if (ptr_type.is_handle()) {
      CHECK(info->head_address.defined())
          << buffer_var << " is not adddressable.";
      return AddressOffset(buffer_var, dtype, offset);
    }
    int dtype_bits = dtype.bits() * dtype.lanes();
    CHECK_EQ(info->unit_bits % dtype_bits, 0);
    return cast(ptr_type,
                   ir::Simplify(offset / make_const(
                       offset.type(), info->unit_bits / dtype_bits)));
  }
  // The storage entry.
  struct StorageEntry {
    // Whether it is tagged memory.
    StorageScope scope;
    // The memory info if any.
    MemoryInfo info;
    // Allocation counter
    int alloc_count{0};
  };
  // The storage scope of each buffer
  std::unordered_map<const Variable*, StorageEntry> storage_info_;
};

Stmt LowerStorageAccessInfo(Stmt stmt) {
  return StorageAccessInfoLower().Mutate(stmt);
}

324 325
}  // namespace ir
}  // namespace tvm