llvm/tools/llvm-exegesis/lib/Clustering.cpp

   1 //===-- Clustering.cpp ------------------------------------------*- C++ -*-===//
   2 //
   3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
   4 // See https://llvm.org/LICENSE.txt for license information.
   5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
   6 //
   7 //===----------------------------------------------------------------------===//
   8
   9 #include "Clustering.h"
  10 #include "Error.h"
  11 #include "SchedClassResolution.h"
  12 #include "llvm/ADT/MapVector.h"
  13 #include "llvm/ADT/SetVector.h"
  14 #include "llvm/ADT/SmallSet.h"
  15 #include "llvm/ADT/SmallVector.h"
  16 #include <algorithm>
  17 #include <deque>
  18 #include <string>
  19 #include <vector>
  20
  21 namespace llvm {
  22 namespace exegesis {
  23
  24 // The clustering problem has the following characteristics:
  25 //  (A) - Low dimension (dimensions are typically proc resource units,
  26 //    typically < 10).
  27 //  (B) - Number of points : ~thousands (points are measurements of an MCInst)
  28 //  (C) - Number of clusters: ~tens.
  29 //  (D) - The number of clusters is not known /a priory/.
  30 //  (E) - The amount of noise is relatively small.
  31 // The problem is rather small. In terms of algorithms, (D) disqualifies
  32 // k-means and makes algorithms such as DBSCAN[1] or OPTICS[2] more applicable.
  33 //
  34 // We've used DBSCAN here because it's simple to implement. This is a pretty
  35 // straightforward and inefficient implementation of the pseudocode in [2].
  36 //
  37 // [1] https://en.wikipedia.org/wiki/DBSCAN
  38 // [2] https://en.wikipedia.org/wiki/OPTICS_algorithm
  39
  40 // Finds the points at distance less than sqrt(EpsilonSquared) of Q (not
  41 // including Q).
  42 void BenchmarkClustering::rangeQuery(
  43     const size_t Q, std::vector<size_t> &Neighbors) const {
  44   Neighbors.clear();
  45   Neighbors.reserve(Points_.size() - 1); // The Q itself isn't a neighbor.
  46   const auto &QMeasurements = Points_[Q].Measurements;
  47   for (size_t P = 0, NumPoints = Points_.size(); P < NumPoints; ++P) {
  48     if (P == Q)
  49       continue;
  50     const auto &PMeasurements = Points_[P].Measurements;
  51     if (PMeasurements.empty()) // Error point.
  52       continue;
  53     if (isNeighbour(PMeasurements, QMeasurements,
  54                     AnalysisClusteringEpsilonSquared_)) {
  55       Neighbors.push_back(P);
  56     }
  57   }
  58 }
  59
  60 // Given a set of points, checks that all the points are neighbours
  61 // up to AnalysisClusteringEpsilon. This is O(2*N).
  62 bool BenchmarkClustering::areAllNeighbours(
  63     ArrayRef<size_t> Pts) const {
  64   // First, get the centroid of this group of points. This is O(N).
  65   SchedClassClusterCentroid G;
  66   for (size_t P : Pts) {
  67     assert(P < Points_.size());
  68     ArrayRef<BenchmarkMeasure> Measurements = Points_[P].Measurements;
  69     if (Measurements.empty()) // Error point.
  70       continue;
  71     G.addPoint(Measurements);
  72   }
  73   const std::vector<BenchmarkMeasure> Centroid = G.getAsPoint();
  74
  75   // Since we will be comparing with the centroid, we need to halve the epsilon.
  76   double AnalysisClusteringEpsilonHalvedSquared =
  77       AnalysisClusteringEpsilonSquared_ / 4.0;
  78
  79   // And now check that every point is a neighbour of the centroid. Also O(N).
  80   return all_of(
  81       Pts, [this, &Centroid, AnalysisClusteringEpsilonHalvedSquared](size_t P) {
  82         assert(P < Points_.size());
  83         const auto &PMeasurements = Points_[P].Measurements;
  84         if (PMeasurements.empty()) // Error point.
  85           return true;             // Pretend that error point is a neighbour.
  86         return isNeighbour(PMeasurements, Centroid,
  87                            AnalysisClusteringEpsilonHalvedSquared);
  88       });
  89 }
  90
  91 BenchmarkClustering::BenchmarkClustering(
  92     const std::vector<Benchmark> &Points,
  93     const double AnalysisClusteringEpsilonSquared)
  94     : Points_(Points),
  95       AnalysisClusteringEpsilonSquared_(AnalysisClusteringEpsilonSquared),
  96       NoiseCluster_(ClusterId::noise()), ErrorCluster_(ClusterId::error()) {}
  97
  98 Error BenchmarkClustering::validateAndSetup() {
  99   ClusterIdForPoint_.resize(Points_.size());
 100   // Mark erroneous measurements out.
 101   // All points must have the same number of dimensions, in the same order.
 102   const std::vector<BenchmarkMeasure> *LastMeasurement = nullptr;
 103   for (size_t P = 0, NumPoints = Points_.size(); P < NumPoints; ++P) {
 104     const auto &Point = Points_[P];
 105     if (!Point.Error.empty()) {
 106       ClusterIdForPoint_[P] = ClusterId::error();
 107       ErrorCluster_.PointIndices.push_back(P);
 108       continue;
 109     }
 110     const auto *CurMeasurement = &Point.Measurements;
 111     if (LastMeasurement) {
 112       if (LastMeasurement->size() != CurMeasurement->size()) {
 113         return make_error<ClusteringError>(
 114             "inconsistent measurement dimensions");
 115       }
 116       for (size_t I = 0, E = LastMeasurement->size(); I < E; ++I) {
 117         if (LastMeasurement->at(I).Key != CurMeasurement->at(I).Key) {
 118           return make_error<ClusteringError>(
 119               "inconsistent measurement dimensions keys");
 120         }
 121       }
 122     }
 123     LastMeasurement = CurMeasurement;
 124   }
 125   if (LastMeasurement) {
 126     NumDimensions_ = LastMeasurement->size();
 127   }
 128   return Error::success();
 129 }
 130
 131 void BenchmarkClustering::clusterizeDbScan(const size_t MinPts) {
 132   std::vector<size_t> Neighbors; // Persistent buffer to avoid allocs.
 133   for (size_t P = 0, NumPoints = Points_.size(); P < NumPoints; ++P) {
 134     if (!ClusterIdForPoint_[P].isUndef())
 135       continue; // Previously processed in inner loop.
 136     rangeQuery(P, Neighbors);
 137     if (Neighbors.size() + 1 < MinPts) { // Density check.
 138       // The region around P is not dense enough to create a new cluster, mark
 139       // as noise for now.
 140       ClusterIdForPoint_[P] = ClusterId::noise();
 141       continue;
 142     }
 143
 144     // Create a new cluster, add P.
 145     Clusters_.emplace_back(ClusterId::makeValid(Clusters_.size()));
 146     Cluster &CurrentCluster = Clusters_.back();
 147     ClusterIdForPoint_[P] = CurrentCluster.Id; /* Label initial point */
 148     CurrentCluster.PointIndices.push_back(P);
 149
 150     // Process P's neighbors.
 151     SetVector<size_t, std::deque<size_t>> ToProcess;
 152     ToProcess.insert(Neighbors.begin(), Neighbors.end());
 153     while (!ToProcess.empty()) {
 154       // Retrieve a point from the set.
 155       const size_t Q = *ToProcess.begin();
 156       ToProcess.erase(ToProcess.begin());
 157
 158       if (ClusterIdForPoint_[Q].isNoise()) {
 159         // Change noise point to border point.
 160         ClusterIdForPoint_[Q] = CurrentCluster.Id;
 161         CurrentCluster.PointIndices.push_back(Q);
 162         continue;
 163       }
 164       if (!ClusterIdForPoint_[Q].isUndef()) {
 165         continue; // Previously processed.
 166       }
 167       // Add Q to the current custer.
 168       ClusterIdForPoint_[Q] = CurrentCluster.Id;
 169       CurrentCluster.PointIndices.push_back(Q);
 170       // And extend to the neighbors of Q if the region is dense enough.
 171       rangeQuery(Q, Neighbors);
 172       if (Neighbors.size() + 1 >= MinPts) {
 173         ToProcess.insert(Neighbors.begin(), Neighbors.end());
 174       }
 175     }
 176   }
 177   // assert(Neighbors.capacity() == (Points_.size() - 1));
 178   // ^ True, but it is not quaranteed to be true in all the cases.
 179
 180   // Add noisy points to noise cluster.
 181   for (size_t P = 0, NumPoints = Points_.size(); P < NumPoints; ++P) {
 182     if (ClusterIdForPoint_[P].isNoise()) {
 183       NoiseCluster_.PointIndices.push_back(P);
 184     }
 185   }
 186 }
 187
 188 void BenchmarkClustering::clusterizeNaive(
 189     const MCSubtargetInfo &SubtargetInfo, const MCInstrInfo &InstrInfo) {
 190   // Given an instruction Opcode, which sched class id's are represented,
 191   // and which are the benchmarks for each sched class?
 192   std::vector<SmallMapVector<unsigned, SmallVector<size_t, 1>, 1>>
 193       OpcodeToSchedClassesToPoints;
 194   const unsigned NumOpcodes = InstrInfo.getNumOpcodes();
 195   OpcodeToSchedClassesToPoints.resize(NumOpcodes);
 196   size_t NumClusters = 0;
 197   for (size_t P = 0, NumPoints = Points_.size(); P < NumPoints; ++P) {
 198     const Benchmark &Point = Points_[P];
 199     const MCInst &MCI = Point.keyInstruction();
 200     unsigned SchedClassId;
 201     std::tie(SchedClassId, std::ignore) =
 202         ResolvedSchedClass::resolveSchedClassId(SubtargetInfo, InstrInfo, MCI);
 203     const unsigned Opcode = MCI.getOpcode();
 204     assert(Opcode < NumOpcodes && "NumOpcodes is incorrect (too small)");
 205     auto &Points = OpcodeToSchedClassesToPoints[Opcode][SchedClassId];
 206     if (Points.empty()) // If we previously have not seen any points of
 207       ++NumClusters;    // this opcode's sched class, then new cluster begins.
 208     Points.emplace_back(P);
 209   }
 210   assert(NumClusters <= NumOpcodes &&
 211          "can't see more opcodes than there are total opcodes");
 212   assert(NumClusters <= Points_.size() &&
 213          "can't see more opcodes than there are total points");
 214
 215   Clusters_.reserve(NumClusters); // We already know how many clusters there is.
 216   for (const auto &SchedClassesOfOpcode : OpcodeToSchedClassesToPoints) {
 217     if (SchedClassesOfOpcode.empty())
 218       continue;
 219     for (ArrayRef<size_t> PointsOfSchedClass :
 220          make_second_range(SchedClassesOfOpcode)) {
 221       if (PointsOfSchedClass.empty())
 222         continue;
 223       // Create a new cluster.
 224       Clusters_.emplace_back(ClusterId::makeValid(
 225           Clusters_.size(),
 226           /*IsUnstable=*/!areAllNeighbours(PointsOfSchedClass)));
 227       Cluster &CurrentCluster = Clusters_.back();
 228       // Mark points as belonging to the new cluster.
 229       for (size_t P : PointsOfSchedClass)
 230         ClusterIdForPoint_[P] = CurrentCluster.Id;
 231       // And add all the points of this opcode's sched class to the new cluster.
 232       CurrentCluster.PointIndices.reserve(PointsOfSchedClass.size());
 233       CurrentCluster.PointIndices.assign(PointsOfSchedClass.begin(),
 234                                          PointsOfSchedClass.end());
 235       assert(CurrentCluster.PointIndices.size() == PointsOfSchedClass.size());
 236     }
 237   }
 238   assert(Clusters_.size() == NumClusters);
 239 }
 240
 241 // Given an instruction Opcode, we can make benchmarks (measurements) of the
 242 // instruction characteristics/performance. Then, to facilitate further analysis
 243 // we group the benchmarks with *similar* characteristics into clusters.
 244 // Now, this is all not entirely deterministic. Some instructions have variable
 245 // characteristics, depending on their arguments. And thus, if we do several
 246 // benchmarks of the same instruction Opcode, we may end up with *different*
 247 // performance characteristics measurements. And when we then do clustering,
 248 // these several benchmarks of the same instruction Opcode may end up being
 249 // clustered into *different* clusters. This is not great for further analysis.
 250 // We shall find every opcode with benchmarks not in just one cluster, and move
 251 // *all* the benchmarks of said Opcode into one new unstable cluster per Opcode.
 252 void BenchmarkClustering::stabilize(unsigned NumOpcodes) {
 253   // Given an instruction Opcode and Config, in which clusters do benchmarks of
 254   // this instruction lie? Normally, they all should be in the same cluster.
 255   struct OpcodeAndConfig {
 256     explicit OpcodeAndConfig(const Benchmark &IB)
 257         : Opcode(IB.keyInstruction().getOpcode()), Config(&IB.Key.Config) {}
 258     unsigned Opcode;
 259     const std::string *Config;
 260
 261     auto Tie() const -> auto { return std::tie(Opcode, *Config); }
 262
 263     bool operator<(const OpcodeAndConfig &O) const { return Tie() < O.Tie(); }
 264     bool operator!=(const OpcodeAndConfig &O) const { return Tie() != O.Tie(); }
 265   };
 266   std::map<OpcodeAndConfig, SmallSet<ClusterId, 1>> OpcodeConfigToClusterIDs;
 267   // Populate OpcodeConfigToClusterIDs and UnstableOpcodes data structures.
 268   assert(ClusterIdForPoint_.size() == Points_.size() && "size mismatch");
 269   for (auto Point : zip(Points_, ClusterIdForPoint_)) {
 270     const ClusterId &ClusterIdOfPoint = std::get<1>(Point);
 271     if (!ClusterIdOfPoint.isValid())
 272       continue; // Only process fully valid clusters.
 273     const OpcodeAndConfig Key(std::get<0>(Point));
 274     SmallSet<ClusterId, 1> &ClusterIDsOfOpcode = OpcodeConfigToClusterIDs[Key];
 275     ClusterIDsOfOpcode.insert(ClusterIdOfPoint);
 276   }
 277
 278   for (const auto &OpcodeConfigToClusterID : OpcodeConfigToClusterIDs) {
 279     const SmallSet<ClusterId, 1> &ClusterIDs = OpcodeConfigToClusterID.second;
 280     const OpcodeAndConfig &Key = OpcodeConfigToClusterID.first;
 281     // We only care about unstable instructions.
 282     if (ClusterIDs.size() < 2)
 283       continue;
 284
 285     // Create a new unstable cluster, one per Opcode.
 286     Clusters_.emplace_back(ClusterId::makeValidUnstable(Clusters_.size()));
 287     Cluster &UnstableCluster = Clusters_.back();
 288     // We will find *at least* one point in each of these clusters.
 289     UnstableCluster.PointIndices.reserve(ClusterIDs.size());
 290
 291     // Go through every cluster which we recorded as containing benchmarks
 292     // of this UnstableOpcode. NOTE: we only recorded valid clusters.
 293     for (const ClusterId &CID : ClusterIDs) {
 294       assert(CID.isValid() &&
 295              "We only recorded valid clusters, not noise/error clusters.");
 296       Cluster &OldCluster = Clusters_[CID.getId()]; // Valid clusters storage.
 297       // Within each cluster, go through each point, and either move it to the
 298       // new unstable cluster, or 'keep' it.
 299       // In this case, we'll reshuffle OldCluster.PointIndices vector
 300       // so that all the points that are *not* for UnstableOpcode are first,
 301       // and the rest of the points is for the UnstableOpcode.
 302       const auto it = std::stable_partition(
 303           OldCluster.PointIndices.begin(), OldCluster.PointIndices.end(),
 304           [this, &Key](size_t P) {
 305             return OpcodeAndConfig(Points_[P]) != Key;
 306           });
 307       assert(std::distance(it, OldCluster.PointIndices.end()) > 0 &&
 308              "Should have found at least one bad point");
 309       // Mark to-be-moved points as belonging to the new cluster.
 310       for (size_t P : llvm::make_range(it, OldCluster.PointIndices.end()))
 311         ClusterIdForPoint_[P] = UnstableCluster.Id;
 312       // Actually append to-be-moved points to the new cluster.
 313       UnstableCluster.PointIndices.insert(UnstableCluster.PointIndices.end(),
 314                                           it, OldCluster.PointIndices.end());
 315       // And finally, remove "to-be-moved" points from the old cluster.
 316       OldCluster.PointIndices.erase(it, OldCluster.PointIndices.end());
 317       // Now, the old cluster may end up being empty, but let's just keep it
 318       // in whatever state it ended up. Purging empty clusters isn't worth it.
 319     };
 320     assert(UnstableCluster.PointIndices.size() > 1 &&
 321            "New unstable cluster should end up with more than one point.");
 322     assert(UnstableCluster.PointIndices.size() >= ClusterIDs.size() &&
 323            "New unstable cluster should end up with no less points than there "
 324            "was clusters");
 325   }
 326 }
 327
 328 Expected<BenchmarkClustering> BenchmarkClustering::create(
 329     const std::vector<Benchmark> &Points, const ModeE Mode,
 330     const size_t DbscanMinPts, const double AnalysisClusteringEpsilon,
 331     const MCSubtargetInfo *SubtargetInfo, const MCInstrInfo *InstrInfo) {
 332   BenchmarkClustering Clustering(
 333       Points, AnalysisClusteringEpsilon * AnalysisClusteringEpsilon);
 334   if (auto Error = Clustering.validateAndSetup()) {
 335     return std::move(Error);
 336   }
 337   if (Clustering.ErrorCluster_.PointIndices.size() == Points.size()) {
 338     return Clustering; // Nothing to cluster.
 339   }
 340
 341   if (Mode == ModeE::Dbscan) {
 342     Clustering.clusterizeDbScan(DbscanMinPts);
 343
 344     if (InstrInfo)
 345       Clustering.stabilize(InstrInfo->getNumOpcodes());
 346   } else /*if(Mode == ModeE::Naive)*/ {
 347     if (!SubtargetInfo || !InstrInfo)
 348       return make_error<Failure>("'naive' clustering mode requires "
 349                                  "SubtargetInfo and InstrInfo to be present");
 350     Clustering.clusterizeNaive(*SubtargetInfo, *InstrInfo);
 351   }
 352
 353   return Clustering;
 354 }
 355
 356 void SchedClassClusterCentroid::addPoint(ArrayRef<BenchmarkMeasure> Point) {
 357   if (Representative.empty())
 358     Representative.resize(Point.size());
 359   assert(Representative.size() == Point.size() &&
 360          "All points should have identical dimensions.");
 361
 362   for (auto I : zip(Representative, Point))
 363     std::get<0>(I).push(std::get<1>(I));
 364 }
 365
 366 std::vector<BenchmarkMeasure> SchedClassClusterCentroid::getAsPoint() const {
 367   std::vector<BenchmarkMeasure> ClusterCenterPoint(Representative.size());
 368   for (auto I : zip(ClusterCenterPoint, Representative))
 369     std::get<0>(I).PerInstructionValue = std::get<1>(I).avg();
 370   return ClusterCenterPoint;
 371 }
 372
 373 bool SchedClassClusterCentroid::validate(
 374     Benchmark::ModeE Mode) const {
 375   size_t NumMeasurements = Representative.size();
 376   switch (Mode) {
 377   case Benchmark::Latency:
 378     if (NumMeasurements != 1) {
 379       errs()
 380           << "invalid number of measurements in latency mode: expected 1, got "
 381           << NumMeasurements << "\n";
 382       return false;
 383     }
 384     break;
 385   case Benchmark::Uops:
 386     // Can have many measurements.
 387     break;
 388   case Benchmark::InverseThroughput:
 389     if (NumMeasurements != 1) {
 390       errs() << "invalid number of measurements in inverse throughput "
 391                 "mode: expected 1, got "
 392              << NumMeasurements << "\n";
 393       return false;
 394     }
 395     break;
 396   default:
 397     llvm_unreachable("unimplemented measurement matching mode");
 398     return false;
 399   }
 400
 401   return true; // All good.
 402 }
 403
 404 } // namespace exegesis
 405 } // namespace llvm