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