touch_extractor.cc 17.7 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 28 29 30
/*!
 *  Copyright (c) 2018 by Contributors
 * \file touch_extractor.cc
 * \brief Extract feature of touch pattern of axes in lowered IR
 */

#include "touch_extractor.h"

#include <set>
#include <algorithm>
#include <cmath>
31
#include <unordered_map>
32 33 34 35 36 37 38 39 40 41 42 43 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 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

namespace tvm {
namespace autotvm {

int ParallelLevel(AnnotationType ann) {
  switch (ann) {
    case kBlockX: case kBlockY: case kBlockZ:
      return 2;
    case kThreadX: case kThreadY: case kThreadZ: case kParallel:
      return 1;
    default:
      return 0;
  }
}

// get touch pattern from index expression
class IndexParser: public IRVisitor {
 public:
  void Parse(Expr expr) {
    pattern_map.clear();
    this->Visit(expr);
  }

  void Visit_(const Variable *op) {
    // TODO(lmzheng): handle more index types (multiple occurrence)
    if (pattern_map.count(op) == 0) {
      pattern_map[op] = TouchPattern();
      pattern_map[op].stride = next_stride_;
      next_stride_ = 1;
    }
  }

  void Visit_(const Mul *op) {
    if (op->a.as<Variable>()) {
      if (const auto stride = op->b.as<IntImm>()) {
        next_stride_ = stride->value;
      }
    }
    IRVisitor::Visit_(op);
  }

  std::unordered_map<const Variable*, TouchPattern> pattern_map;

 private:
  int64_t next_stride_ = 1;
};

// extract iter vars and their touch pattern from ir
bool TouchExtractor::EnterItervar_(VarExpr var, int64_t length, AnnotationType ann_type) {
  // do not insert duplicated occurrences of virtual thread
  if (ann_type == kVirtualThread && itervar_map.count(var) != 0) {
    skip_stack_size_.push_back(itervar_stack_.size());
    return true;
  } else {
    itervar_stack_.push_back(var);
    topdown_product_ *= length;

    if (itervar_map.count(var) != 0) {
      // find two duplicated axes
      // these happens when we create tvm.thread_axis("threadIdx.x") once and
      // bind it twice. Here we treat them as two axes
      // so we create a snapshot for the old one and freeze it
      VarExpr old = VarExpr(var.get()->name_hint);
      itervar_map.insert({old, itervar_map[var]});
      itervar_map.erase(var);
    }

    itervar_map.insert({var, ItervarFeature(var, length,
                                            static_cast<int>(itervar_stack_.size()),
                                            ann_type,
                                            topdown_product_,
                                            static_cast<int>(itervar_counter_++))});
  }

  return true;
}

void TouchExtractor::ExitItervar_() {
  if (!skip_stack_size_.empty() && skip_stack_size_.back() == itervar_stack_.size()) {
    skip_stack_size_.pop_back();
    return;
  }
  VarExpr var = itervar_stack_.back();

  // update count and reuse ratio for upper iter vars (includes self)
  for (auto kv : itervar_map[var].touch_feature) {
    if (kv.second.stride != 0) {  // multiply count
      for (auto stack_var : itervar_stack_) {
        auto touch_pattern = itervar_map[stack_var].touch_feature.find(kv.first);
        CHECK(touch_pattern != itervar_map[stack_var].touch_feature.end());
        touch_pattern->second.count *= itervar_map[var].length;
      }
    } else {                      // multiply reuse ratio
      for (auto stack_var : itervar_stack_) {
        auto touch_pattern = itervar_map[stack_var].touch_feature.find(kv.first);
        CHECK(touch_pattern != itervar_map[stack_var].touch_feature.end());
        touch_pattern->second.reuse *= itervar_map[var].length;
      }
    }
  }
  itervar_stack_.pop_back();

134 135 136
  int64_t length = itervar_map[var].length;
  if (length != 0)
      topdown_product_ /= length;
137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518
  int64_t bottomup_product = -1;
  for (auto kv : itervar_map[var].touch_feature) {
    bottomup_product = std::max(bottomup_product, kv.second.count * kv.second.reuse);
  }

  itervar_map[var].bottomup_product = bottomup_product;

  // push base to upper parallel axis
  int para_level = ParallelLevel(itervar_map[var].ann);
  // if is the separate line of parallel level, push the base to upper parallel level
  if (!itervar_stack_.empty() &&
      ParallelLevel(itervar_map[itervar_stack_.back()].ann) == para_level + 1) {
    for (auto kv : itervar_map[var].touch_feature) {
      for (auto stack_var : itervar_stack_) {
        if (ParallelLevel(itervar_map[stack_var].ann) == para_level + 1) {
          auto touch_pattern = itervar_map[stack_var].touch_feature.find(kv.first);
          CHECK(touch_pattern != itervar_map[stack_var].touch_feature.end());
          touch_pattern->second.thread_reuse = -kv.second.reuse;
          touch_pattern->second.thread_count = -kv.second.count;
          // NOTE: use minus as a flag to denote it is a base,
          // indicating it is not the final value
        }
      }
    }
  }

  for (auto kv : itervar_map[var].touch_feature) {
    if (kv.second.thread_count < 0) {
      itervar_map[var].touch_feature[kv.first].thread_count =
          kv.second.count / (-kv.second.thread_count);
      itervar_map[var].touch_feature[kv.first].thread_reuse =
          kv.second.reuse / (-kv.second.thread_reuse);
    }
  }
}

void TouchExtractor::EnterMem_(VarExpr buffer_var, Expr index) {
  std::string name = buffer_var.get()->name_hint;
  TouchedBuffer buf = name + "_" + std::to_string(buffer_counter_[name]++);

  // extract touch pattern from index
  IndexParser parser;
  parser.Parse(index);

  // push up mem access info
  for (auto var : itervar_stack_) {
    auto x = parser.pattern_map.find(var.get());
    if (x != parser.pattern_map.end()) {
      itervar_map[var].touch_feature[buf] = x->second;
    } else {
      itervar_map[var].touch_feature[buf] = TouchPattern();
    }
  }
}

void TouchExtractor::ExitMem_() {
}

/*!
 * \brief Get axis-based feature for all axes
 * \param stmt The statement to be extracted
 * \param bool Whether take log for numerical feature
 * \param ret_feature The buffer where the return value is stored
 *
 * \note The format of return value is
 * ((
 *   ('_itervar_',  var),
 *   ('_attr_',     length, nest_level, topdown, bottomup, one_hot_annotation),
 *   ('_arith_',    add_ct, mul_ct, div_ct),
 *   ('data_vec_0', stride, mod, count, reuse, thread_count, thread_reuse),
 *   ('conv_0',     stride, mod, count, reuse, thread_count, thread_reuse),
 * ),
 * (
 *   ('_itervar_',    var2),
 *   ('_attr_',       length, nest_level, one_hot_annotation),
 *   ('_arith_',      add_ct, mul_ct, div_ct),
 *   ('kernel_vec_0', stride, mod, count, reuse, thread_count, thread_reuse),
 *   ('conv_1',       stride, mod, count, reuse, thread_count, thread_reuse),
 * ))
 *
 * Itervars are sorted according to their first occurrence position in IR.
 * Buffers touched by an itervar are sorted by their unique names.
 *
 * \note If you want to flatten these features as the input of your model,
 * You can use the faster one GetItervarFeatureFlatten below.
 */
void GetItervarFeature(Stmt stmt, bool take_log, Array<Array<Array<Expr> > > *ret_feature) {
  // extract
  TouchExtractor touch_analyzer;
  touch_analyzer.Analyze(stmt);

  // sort according to order
  std::vector<VarExpr> vars;
  for (auto kv : touch_analyzer.itervar_map) {
    vars.push_back(kv.first);
  }
  std::sort(vars.begin(), vars.end(), [&](const VarExpr &lhs, const VarExpr &rhs) -> bool {
    return touch_analyzer.itervar_map[lhs].order < touch_analyzer.itervar_map[rhs].order;
  });

  // whether take log for numerical feature
  std::function<double(int64_t)> trans;
  if (take_log) {
    trans = [](int64_t x) {
      if (x < 0)
        return -std::log(-x+1) / std::log(2);
      x = x + 1;
      return std::log(x) / std::log(2);
    };
  } else {
    trans = [](int64_t x) {
      return x;
    };
  }

  // serialize for front end
  for (auto var : vars) {
    Array<Array<Expr> > feature_row;
    ItervarFeature &fea = touch_analyzer.itervar_map[var];
    feature_row.push_back(Array<Expr>{std::string("_itervar_"), var});

    Array<Expr> attr{std::string("_attr_"),
                     FloatImm::make(Float(32), trans(fea.length)),
                     IntImm::make(Int(32), fea.nest_level),
                     FloatImm::make(Float(32), trans(fea.topdown_product)),
                     FloatImm::make(Float(32), trans(fea.bottomup_product)),
    };
    // one hot annotation
    for (int i = 0; i < kNum; i++) {
      attr.push_back(i == fea.ann);
    }
    feature_row.push_back(attr);

    // arithmetic
    feature_row.push_back(Array<Expr>{std::string("_arith_"),
                                      FloatImm::make(Float(32), trans(fea.add_ct)),
                                      FloatImm::make(Float(32), trans(fea.mul_ct)),
                                      FloatImm::make(Float(32), trans(fea.div_ct)),
    });

    // touch map
    std::vector<TouchedBuffer> bufs;
    for (auto kv : fea.touch_feature) {
      bufs.push_back(kv.first);
    }
    std::sort(bufs.begin(), bufs.end());
    for (auto k : bufs) {
      TouchPattern &v = fea.touch_feature[k];
      feature_row.push_back(Array<Expr>{k,
                                        FloatImm::make(Float(32), trans(v.stride)),
                                        FloatImm::make(Float(32), trans(v.mod)),
                                        FloatImm::make(Float(32), trans(v.count)),
                                        FloatImm::make(Float(32), trans(v.reuse)),
                                        FloatImm::make(Float(32), trans(v.thread_count)),
                                        FloatImm::make(Float(32), trans(v.thread_reuse)),
      });
    }

    ret_feature->push_back(feature_row);
  }
}

/*!
 * \brief Get axis-based feature for all axes and flatten them into a one-dimensional vector.
 * \param stmt The statement to be extracted
 * \param bool Whether take log for numerical feature
 * \param ret_feature The buffer where the return value is stored
 *
 * \note See GetItervarFeature for more details about the return value.
 *       This is an optimized version of GetItervarFeature + Flatten. This runs much faster.
 */
void GetItervarFeatureFlatten(Stmt stmt, bool take_log, std::vector<float> *ret_feature) {
  // extract touch feature
  TouchExtractor touch_analyzer;
  touch_analyzer.Analyze(stmt);

  // sort according to order
  std::vector<VarExpr> vars;
  for (auto kv : touch_analyzer.itervar_map) {
    vars.push_back(kv.first);
  }
  std::sort(vars.begin(), vars.end(), [&](const VarExpr &lhs, const VarExpr &rhs) -> bool {
    return touch_analyzer.itervar_map[lhs].order < touch_analyzer.itervar_map[rhs].order;
  });

  // whether take log for numerical feature
  std::function<float(int64_t)> trans;
  if (take_log) {
    trans = [](int64_t x) {
      if (x < 0)
        return -std::log(-x+1) / std::log(2);
      x = x + 1;
      return std::log(x) / std::log(2);
    };
  } else {
    trans = [](int64_t x) {
      return x;
    };
  }

  // serialize for front end
  for (auto var : vars) {
    ItervarFeature &fea = touch_analyzer.itervar_map[var];

    ret_feature->push_back(trans(fea.length));
    ret_feature->push_back(fea.nest_level);
    ret_feature->push_back(trans(fea.topdown_product));
    ret_feature->push_back(trans(fea.bottomup_product));

    // one hot annotation
    for (int i = 0; i < kNum; i++) {
      ret_feature->push_back(i == fea.ann);
    }

    // arithmetic
    ret_feature->push_back(trans(fea.add_ct));
    ret_feature->push_back(trans(fea.mul_ct));
    ret_feature->push_back(trans(fea.div_ct));

    // touch map
    std::vector<TouchedBuffer> bufs;
    for (auto kv : fea.touch_feature) {
      bufs.push_back(kv.first);
    }
    std::sort(bufs.begin(), bufs.end());
    for (auto k : bufs) {
      TouchPattern &v = fea.touch_feature[k];
      ret_feature->push_back(trans(v.stride));
      ret_feature->push_back(trans(v.mod));
      ret_feature->push_back(trans(v.count));
      ret_feature->push_back(trans(v.reuse));
      ret_feature->push_back(trans(v.thread_count));
      ret_feature->push_back(trans(v.thread_reuse));
    }
  }
}

/*!
 * \brief Get curve sample feature (relation feature) and flatten them into a one-dimensional vector.
 * \param stmt The statement to be extracted
 * \param sample_n The number of points used for sampling a curve (along one dimension)
 * \param ret_feature The buffer where the return value is stored
 */
void GetCurveSampleFeatureFlatten(Stmt stmt, int sample_n, std::vector<float> *ret_feature) {
  // extract touch feature
  TouchExtractor touch_ext;
  touch_ext.Analyze(stmt);

  // sort according to order
  std::vector<VarExpr> vars;
  for (auto kv : touch_ext.itervar_map) {
    vars.push_back(kv.first);
  }
  std::sort(vars.begin(), vars.end(), [&](const VarExpr &lhs, const VarExpr &rhs) -> bool {
    return touch_ext.itervar_map[lhs].order < touch_ext.itervar_map[rhs].order;
  });

  int max_depth = 0;
  std::map<TouchedBuffer, std::vector<double> > reuse_curve;
  std::map<TouchedBuffer, std::vector<double> > count_curve;
  std::map<TouchedBuffer, std::vector<double> > topdown_curve;
  std::map<TouchedBuffer, std::vector<double> > bottomup_curve;
  std::set<TouchedBuffer> innermost_buffers;
  std::set<std::string> added;

  // find maximum depth of loop nest
  for (auto var : vars) {
    ItervarFeature &fea = touch_ext.itervar_map[var];
    max_depth = std::max(max_depth, fea.nest_level);
  }

  // mark inner most buffer
  for (auto iter = vars.rbegin(); iter != vars.rend(); iter++) {
    auto var = *iter;
    ItervarFeature &fea = touch_ext.itervar_map[var];
    if (fea.nest_level == max_depth) {
      for (auto kv : fea.touch_feature) {
        // delete buffer no (e.g. 'A_0' -> 'A', 'A_1' -> 'A')
        std::string raw_name = kv.first.substr(0, kv.first.rfind("_"));

        // delete memory scope (e.g. 'A.local' -> 'A', 'A.shared' -> 'A')
        size_t pos = raw_name.find(".");
        if (pos < kv.first.size())
          raw_name = raw_name.substr(0, pos);

        // If there are multiple innermost buffers that are derived from a same raw buffer
        // We only record the last occurrence (note the `iter` is in reverse order)
        // e.g. `A.local`, `A.shared` are derived from `A`, if they all occurred at the inner most
        // level, we will only record the last occurrence,
        if (added.find(raw_name) == added.end()) {
          innermost_buffers.insert(kv.first);
          added.insert(raw_name);
        }
      }
    }
  }

  // pad the first point (zero) for all curves
  for (auto buf : innermost_buffers) {
    reuse_curve[buf].push_back(0);
    count_curve[buf].push_back(0);
    topdown_curve[buf].push_back(0);
    bottomup_curve[buf].push_back(0);
  }

  // extract curves
  for (auto var : vars) {
    ItervarFeature &fea = touch_ext.itervar_map[var];
    for (auto kv : fea.touch_feature) {
      if (innermost_buffers.find(kv.first) != innermost_buffers.end()) {
        reuse_curve[kv.first].emplace_back(std::log(kv.second.reuse) / std::log(2));
        count_curve[kv.first].emplace_back(std::log(kv.second.count) / std::log(2));
        topdown_curve[kv.first].emplace_back(std::log(fea.topdown_product) / std::log(2));
        bottomup_curve[kv.first].emplace_back(std::log(fea.bottomup_product) / std::log(2));
      }
    }
  }

  // sample relation in the curve
  auto sample_curve = [&](const std::vector<double> &x, const std::vector<double> &y,
                          double weight) {
    for (int i = 0; i < sample_n; i++) {
      double xx = i * weight;
      for (int j = static_cast<int>(x.size()) - 1; j >= 0; j--) {
        if (xx > x[j] - 1e-6) {
          ret_feature->emplace_back(y[j]);
          ret_feature->emplace_back(xx - x[j]);
          break;
        }
      }
    }
  };

  // serialize to frontend
  for (auto k : innermost_buffers) {
    std::vector<double> &count = count_curve[k];
    std::vector<double> &reuse = reuse_curve[k];
    std::vector<double> &top_down = topdown_curve[k];

    std::sort(count.begin(), count.end());
    std::sort(reuse.begin(), reuse.end());
    std::sort(top_down.begin(), top_down.end());

    sample_curve(count, reuse, 1);
    sample_curve(reuse, count, 1);
    sample_curve(count, top_down, 1);
    sample_curve(top_down, count, 1);
  }
}


// register API for front end
TVM_REGISTER_API("autotvm.feature.GetItervarFeature")
.set_body([](TVMArgs args, TVMRetValue *ret) {
  Stmt stmt = args[0];
  bool take_log = args[1];
  Array<Array<Array<Expr > > > ret_feature;

  GetItervarFeature(stmt, take_log, &ret_feature);

  *ret = ret_feature;
});


TVM_REGISTER_API("autotvm.feature.GetItervarFeatureFlatten")
.set_body([](TVMArgs args, TVMRetValue *ret) {
  Stmt stmt = args[0];
  bool take_log = args[1];
  std::vector<float> ret_feature;

  GetItervarFeatureFlatten(stmt, take_log, &ret_feature);

  TVMByteArray arr;
  arr.size = sizeof(float) * ret_feature.size();
  arr.data = reinterpret_cast<char *>(ret_feature.data());
  *ret = arr;
});


TVM_REGISTER_API("autotvm.feature.GetCurveSampleFeatureFlatten")
.set_body([](TVMArgs args, TVMRetValue *ret) {
  Stmt stmt = args[0];
519
  int sample_n = args[1];
520 521
  std::vector<float> ret_feature;

522
  GetCurveSampleFeatureFlatten(stmt, sample_n, &ret_feature);
523 524 525 526 527 528 529 530 531 532

  TVMByteArray arr;
  arr.size = sizeof(float) * ret_feature.size();
  arr.data = reinterpret_cast<char *>(ret_feature.data());
  *ret = arr;
});


}  // namespace autotvm
}  // namespace tvm