1 //===-- Analysis.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 "Analysis.h" 10 #include "BenchmarkResult.h" 11 #include "llvm/ADT/STLExtras.h" 12 #include "llvm/MC/MCAsmInfo.h" 13 #include "llvm/Support/FormatVariadic.h" 14 #include <limits> 15 #include <unordered_set> 16 #include <vector> 17 18 namespace llvm { 19 namespace exegesis { 20 21 static const char kCsvSep = ','; 22 23 static unsigned resolveSchedClassId(const llvm::MCSubtargetInfo &STI, 24 unsigned SchedClassId, 25 const llvm::MCInst &MCI) { 26 const auto &SM = STI.getSchedModel(); 27 while (SchedClassId && SM.getSchedClassDesc(SchedClassId)->isVariant()) 28 SchedClassId = 29 STI.resolveVariantSchedClass(SchedClassId, &MCI, SM.getProcessorID()); 30 return SchedClassId; 31 } 32 33 namespace { 34 35 enum EscapeTag { kEscapeCsv, kEscapeHtml, kEscapeHtmlString }; 36 37 template <EscapeTag Tag> 38 void writeEscaped(llvm::raw_ostream &OS, const llvm::StringRef S); 39 40 template <> 41 void writeEscaped<kEscapeCsv>(llvm::raw_ostream &OS, const llvm::StringRef S) { 42 if (std::find(S.begin(), S.end(), kCsvSep) == S.end()) { 43 OS << S; 44 } else { 45 // Needs escaping. 46 OS << '"'; 47 for (const char C : S) { 48 if (C == '"') 49 OS << "\"\""; 50 else 51 OS << C; 52 } 53 OS << '"'; 54 } 55 } 56 57 template <> 58 void writeEscaped<kEscapeHtml>(llvm::raw_ostream &OS, const llvm::StringRef S) { 59 for (const char C : S) { 60 if (C == '<') 61 OS << "<"; 62 else if (C == '>') 63 OS << ">"; 64 else if (C == '&') 65 OS << "&"; 66 else 67 OS << C; 68 } 69 } 70 71 template <> 72 void writeEscaped<kEscapeHtmlString>(llvm::raw_ostream &OS, 73 const llvm::StringRef S) { 74 for (const char C : S) { 75 if (C == '"') 76 OS << "\\\""; 77 else 78 OS << C; 79 } 80 } 81 82 } // namespace 83 84 template <EscapeTag Tag> 85 static void 86 writeClusterId(llvm::raw_ostream &OS, 87 const InstructionBenchmarkClustering::ClusterId &CID) { 88 if (CID.isNoise()) 89 writeEscaped<Tag>(OS, "[noise]"); 90 else if (CID.isError()) 91 writeEscaped<Tag>(OS, "[error]"); 92 else 93 OS << CID.getId(); 94 } 95 96 template <EscapeTag Tag> 97 static void writeMeasurementValue(llvm::raw_ostream &OS, const double Value) { 98 // Given Value, if we wanted to serialize it to a string, 99 // how many base-10 digits will we need to store, max? 100 static constexpr auto MaxDigitCount = 101 std::numeric_limits<decltype(Value)>::max_digits10; 102 // Also, we will need a decimal separator. 103 static constexpr auto DecimalSeparatorLen = 1; // '.' e.g. 104 // So how long of a string will the serialization produce, max? 105 static constexpr auto SerializationLen = MaxDigitCount + DecimalSeparatorLen; 106 107 // WARNING: when changing the format, also adjust the small-size estimate ^. 108 static constexpr StringLiteral SimpleFloatFormat = StringLiteral("{0:F}"); 109 110 writeEscaped<Tag>( 111 OS, 112 llvm::formatv(SimpleFloatFormat.data(), Value).sstr<SerializationLen>()); 113 } 114 115 template <typename EscapeTag, EscapeTag Tag> 116 void Analysis::writeSnippet(llvm::raw_ostream &OS, 117 llvm::ArrayRef<uint8_t> Bytes, 118 const char *Separator) const { 119 llvm::SmallVector<std::string, 3> Lines; 120 // Parse the asm snippet and print it. 121 while (!Bytes.empty()) { 122 llvm::MCInst MI; 123 uint64_t MISize = 0; 124 if (!Disasm_->getInstruction(MI, MISize, Bytes, 0, llvm::nulls(), 125 llvm::nulls())) { 126 writeEscaped<Tag>(OS, llvm::join(Lines, Separator)); 127 writeEscaped<Tag>(OS, Separator); 128 writeEscaped<Tag>(OS, "[error decoding asm snippet]"); 129 return; 130 } 131 llvm::SmallString<128> InstPrinterStr; // FIXME: magic number. 132 llvm::raw_svector_ostream OSS(InstPrinterStr); 133 InstPrinter_->printInst(&MI, OSS, "", *SubtargetInfo_); 134 Bytes = Bytes.drop_front(MISize); 135 Lines.emplace_back(llvm::StringRef(InstPrinterStr).trim()); 136 } 137 writeEscaped<Tag>(OS, llvm::join(Lines, Separator)); 138 } 139 140 // Prints a row representing an instruction, along with scheduling info and 141 // point coordinates (measurements). 142 void Analysis::printInstructionRowCsv(const size_t PointId, 143 llvm::raw_ostream &OS) const { 144 const InstructionBenchmark &Point = Clustering_.getPoints()[PointId]; 145 writeClusterId<kEscapeCsv>(OS, Clustering_.getClusterIdForPoint(PointId)); 146 OS << kCsvSep; 147 writeSnippet<EscapeTag, kEscapeCsv>(OS, Point.AssembledSnippet, "; "); 148 OS << kCsvSep; 149 writeEscaped<kEscapeCsv>(OS, Point.Key.Config); 150 OS << kCsvSep; 151 assert(!Point.Key.Instructions.empty()); 152 const llvm::MCInst &MCI = Point.Key.Instructions[0]; 153 const unsigned SchedClassId = resolveSchedClassId( 154 *SubtargetInfo_, InstrInfo_->get(MCI.getOpcode()).getSchedClass(), MCI); 155 156 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) 157 const llvm::MCSchedClassDesc *const SCDesc = 158 SubtargetInfo_->getSchedModel().getSchedClassDesc(SchedClassId); 159 writeEscaped<kEscapeCsv>(OS, SCDesc->Name); 160 #else 161 OS << SchedClassId; 162 #endif 163 for (const auto &Measurement : Point.Measurements) { 164 OS << kCsvSep; 165 writeMeasurementValue<kEscapeCsv>(OS, Measurement.PerInstructionValue); 166 } 167 OS << "\n"; 168 } 169 170 Analysis::Analysis(const llvm::Target &Target, 171 const InstructionBenchmarkClustering &Clustering) 172 : Clustering_(Clustering) { 173 if (Clustering.getPoints().empty()) 174 return; 175 176 const InstructionBenchmark &FirstPoint = Clustering.getPoints().front(); 177 InstrInfo_.reset(Target.createMCInstrInfo()); 178 RegInfo_.reset(Target.createMCRegInfo(FirstPoint.LLVMTriple)); 179 AsmInfo_.reset(Target.createMCAsmInfo(*RegInfo_, FirstPoint.LLVMTriple)); 180 SubtargetInfo_.reset(Target.createMCSubtargetInfo(FirstPoint.LLVMTriple, 181 FirstPoint.CpuName, "")); 182 InstPrinter_.reset(Target.createMCInstPrinter( 183 llvm::Triple(FirstPoint.LLVMTriple), 0 /*default variant*/, *AsmInfo_, 184 *InstrInfo_, *RegInfo_)); 185 186 Context_ = llvm::make_unique<llvm::MCContext>(AsmInfo_.get(), RegInfo_.get(), 187 &ObjectFileInfo_); 188 Disasm_.reset(Target.createMCDisassembler(*SubtargetInfo_, *Context_)); 189 assert(Disasm_ && "cannot create MCDisassembler. missing call to " 190 "InitializeXXXTargetDisassembler ?"); 191 } 192 193 template <> 194 llvm::Error 195 Analysis::run<Analysis::PrintClusters>(llvm::raw_ostream &OS) const { 196 if (Clustering_.getPoints().empty()) 197 return llvm::Error::success(); 198 199 // Write the header. 200 OS << "cluster_id" << kCsvSep << "opcode_name" << kCsvSep << "config" 201 << kCsvSep << "sched_class"; 202 for (const auto &Measurement : Clustering_.getPoints().front().Measurements) { 203 OS << kCsvSep; 204 writeEscaped<kEscapeCsv>(OS, Measurement.Key); 205 } 206 OS << "\n"; 207 208 // Write the points. 209 const auto &Clusters = Clustering_.getValidClusters(); 210 for (size_t I = 0, E = Clusters.size(); I < E; ++I) { 211 for (const size_t PointId : Clusters[I].PointIndices) { 212 printInstructionRowCsv(PointId, OS); 213 } 214 OS << "\n\n"; 215 } 216 return llvm::Error::success(); 217 } 218 219 Analysis::ResolvedSchedClassAndPoints::ResolvedSchedClassAndPoints( 220 ResolvedSchedClass &&RSC) 221 : RSC(std::move(RSC)) {} 222 223 std::vector<Analysis::ResolvedSchedClassAndPoints> 224 Analysis::makePointsPerSchedClass() const { 225 std::vector<ResolvedSchedClassAndPoints> Entries; 226 // Maps SchedClassIds to index in result. 227 std::unordered_map<unsigned, size_t> SchedClassIdToIndex; 228 const auto &Points = Clustering_.getPoints(); 229 for (size_t PointId = 0, E = Points.size(); PointId < E; ++PointId) { 230 const InstructionBenchmark &Point = Points[PointId]; 231 if (!Point.Error.empty()) 232 continue; 233 assert(!Point.Key.Instructions.empty()); 234 // FIXME: we should be using the tuple of classes for instructions in the 235 // snippet as key. 236 const llvm::MCInst &MCI = Point.Key.Instructions[0]; 237 unsigned SchedClassId = InstrInfo_->get(MCI.getOpcode()).getSchedClass(); 238 const bool WasVariant = SchedClassId && SubtargetInfo_->getSchedModel() 239 .getSchedClassDesc(SchedClassId) 240 ->isVariant(); 241 SchedClassId = resolveSchedClassId(*SubtargetInfo_, SchedClassId, MCI); 242 const auto IndexIt = SchedClassIdToIndex.find(SchedClassId); 243 if (IndexIt == SchedClassIdToIndex.end()) { 244 // Create a new entry. 245 SchedClassIdToIndex.emplace(SchedClassId, Entries.size()); 246 ResolvedSchedClassAndPoints Entry( 247 ResolvedSchedClass(*SubtargetInfo_, SchedClassId, WasVariant)); 248 Entry.PointIds.push_back(PointId); 249 Entries.push_back(std::move(Entry)); 250 } else { 251 // Append to the existing entry. 252 Entries[IndexIt->second].PointIds.push_back(PointId); 253 } 254 } 255 return Entries; 256 } 257 258 // Uops repeat the same opcode over again. Just show this opcode and show the 259 // whole snippet only on hover. 260 static void writeUopsSnippetHtml(llvm::raw_ostream &OS, 261 const std::vector<llvm::MCInst> &Instructions, 262 const llvm::MCInstrInfo &InstrInfo) { 263 if (Instructions.empty()) 264 return; 265 writeEscaped<kEscapeHtml>(OS, InstrInfo.getName(Instructions[0].getOpcode())); 266 if (Instructions.size() > 1) 267 OS << " (x" << Instructions.size() << ")"; 268 } 269 270 // Latency tries to find a serial path. Just show the opcode path and show the 271 // whole snippet only on hover. 272 static void 273 writeLatencySnippetHtml(llvm::raw_ostream &OS, 274 const std::vector<llvm::MCInst> &Instructions, 275 const llvm::MCInstrInfo &InstrInfo) { 276 bool First = true; 277 for (const llvm::MCInst &Instr : Instructions) { 278 if (First) 279 First = false; 280 else 281 OS << " → "; 282 writeEscaped<kEscapeHtml>(OS, InstrInfo.getName(Instr.getOpcode())); 283 } 284 } 285 286 void Analysis::printSchedClassClustersHtml( 287 const std::vector<SchedClassCluster> &Clusters, 288 const ResolvedSchedClass &RSC, llvm::raw_ostream &OS) const { 289 const auto &Points = Clustering_.getPoints(); 290 OS << "<table class=\"sched-class-clusters\">"; 291 OS << "<tr><th>ClusterId</th><th>Opcode/Config</th>"; 292 assert(!Clusters.empty()); 293 for (const auto &Measurement : 294 Points[Clusters[0].getPointIds()[0]].Measurements) { 295 OS << "<th>"; 296 writeEscaped<kEscapeHtml>(OS, Measurement.Key); 297 OS << "</th>"; 298 } 299 OS << "</tr>"; 300 for (const SchedClassCluster &Cluster : Clusters) { 301 OS << "<tr class=\"" 302 << (Cluster.measurementsMatch(*SubtargetInfo_, RSC, Clustering_) 303 ? "good-cluster" 304 : "bad-cluster") 305 << "\"><td>"; 306 writeClusterId<kEscapeHtml>(OS, Cluster.id()); 307 OS << "</td><td><ul>"; 308 for (const size_t PointId : Cluster.getPointIds()) { 309 const auto &Point = Points[PointId]; 310 OS << "<li><span class=\"mono\" title=\""; 311 writeSnippet<EscapeTag, kEscapeHtmlString>(OS, Point.AssembledSnippet, 312 "\n"); 313 OS << "\">"; 314 switch (Point.Mode) { 315 case InstructionBenchmark::Latency: 316 writeLatencySnippetHtml(OS, Point.Key.Instructions, *InstrInfo_); 317 break; 318 case InstructionBenchmark::Uops: 319 case InstructionBenchmark::InverseThroughput: 320 writeUopsSnippetHtml(OS, Point.Key.Instructions, *InstrInfo_); 321 break; 322 default: 323 llvm_unreachable("invalid mode"); 324 } 325 OS << "</span> <span class=\"mono\">"; 326 writeEscaped<kEscapeHtml>(OS, Point.Key.Config); 327 OS << "</span></li>"; 328 } 329 OS << "</ul></td>"; 330 for (const auto &Stats : Cluster.getRepresentative()) { 331 OS << "<td class=\"measurement\">"; 332 writeMeasurementValue<kEscapeHtml>(OS, Stats.avg()); 333 OS << "<br><span class=\"minmax\">["; 334 writeMeasurementValue<kEscapeHtml>(OS, Stats.min()); 335 OS << ";"; 336 writeMeasurementValue<kEscapeHtml>(OS, Stats.max()); 337 OS << "]</span></td>"; 338 } 339 OS << "</tr>"; 340 } 341 OS << "</table>"; 342 } 343 344 // Return the non-redundant list of WriteProcRes used by the given sched class. 345 // The scheduling model for LLVM is such that each instruction has a certain 346 // number of uops which consume resources which are described by WriteProcRes 347 // entries. Each entry describe how many cycles are spent on a specific ProcRes 348 // kind. 349 // For example, an instruction might have 3 uOps, one dispatching on P0 350 // (ProcResIdx=1) and two on P06 (ProcResIdx = 7). 351 // Note that LLVM additionally denormalizes resource consumption to include 352 // usage of super resources by subresources. So in practice if there exists a 353 // P016 (ProcResIdx=10), then the cycles consumed by P0 are also consumed by 354 // P06 (ProcResIdx = 7) and P016 (ProcResIdx = 10), and the resources consumed 355 // by P06 are also consumed by P016. In the figure below, parenthesized cycles 356 // denote implied usage of superresources by subresources: 357 // P0 P06 P016 358 // uOp1 1 (1) (1) 359 // uOp2 1 (1) 360 // uOp3 1 (1) 361 // ============================= 362 // 1 3 3 363 // Eventually we end up with three entries for the WriteProcRes of the 364 // instruction: 365 // {ProcResIdx=1, Cycles=1} // P0 366 // {ProcResIdx=7, Cycles=3} // P06 367 // {ProcResIdx=10, Cycles=3} // P016 368 // 369 // Note that in this case, P016 does not contribute any cycles, so it would 370 // be removed by this function. 371 // FIXME: Move this to MCSubtargetInfo and use it in llvm-mca. 372 static llvm::SmallVector<llvm::MCWriteProcResEntry, 8> 373 getNonRedundantWriteProcRes(const llvm::MCSchedClassDesc &SCDesc, 374 const llvm::MCSubtargetInfo &STI) { 375 llvm::SmallVector<llvm::MCWriteProcResEntry, 8> Result; 376 const auto &SM = STI.getSchedModel(); 377 const unsigned NumProcRes = SM.getNumProcResourceKinds(); 378 379 // This assumes that the ProcResDescs are sorted in topological order, which 380 // is guaranteed by the tablegen backend. 381 llvm::SmallVector<float, 32> ProcResUnitUsage(NumProcRes); 382 for (const auto *WPR = STI.getWriteProcResBegin(&SCDesc), 383 *const WPREnd = STI.getWriteProcResEnd(&SCDesc); 384 WPR != WPREnd; ++WPR) { 385 const llvm::MCProcResourceDesc *const ProcResDesc = 386 SM.getProcResource(WPR->ProcResourceIdx); 387 if (ProcResDesc->SubUnitsIdxBegin == nullptr) { 388 // This is a ProcResUnit. 389 Result.push_back({WPR->ProcResourceIdx, WPR->Cycles}); 390 ProcResUnitUsage[WPR->ProcResourceIdx] += WPR->Cycles; 391 } else { 392 // This is a ProcResGroup. First see if it contributes any cycles or if 393 // it has cycles just from subunits. 394 float RemainingCycles = WPR->Cycles; 395 for (const auto *SubResIdx = ProcResDesc->SubUnitsIdxBegin; 396 SubResIdx != ProcResDesc->SubUnitsIdxBegin + ProcResDesc->NumUnits; 397 ++SubResIdx) { 398 RemainingCycles -= ProcResUnitUsage[*SubResIdx]; 399 } 400 if (RemainingCycles < 0.01f) { 401 // The ProcResGroup contributes no cycles of its own. 402 continue; 403 } 404 // The ProcResGroup contributes `RemainingCycles` cycles of its own. 405 Result.push_back({WPR->ProcResourceIdx, 406 static_cast<uint16_t>(std::round(RemainingCycles))}); 407 // Spread the remaining cycles over all subunits. 408 for (const auto *SubResIdx = ProcResDesc->SubUnitsIdxBegin; 409 SubResIdx != ProcResDesc->SubUnitsIdxBegin + ProcResDesc->NumUnits; 410 ++SubResIdx) { 411 ProcResUnitUsage[*SubResIdx] += RemainingCycles / ProcResDesc->NumUnits; 412 } 413 } 414 } 415 return Result; 416 } 417 418 Analysis::ResolvedSchedClass::ResolvedSchedClass( 419 const llvm::MCSubtargetInfo &STI, unsigned ResolvedSchedClassId, 420 bool WasVariant) 421 : SchedClassId(ResolvedSchedClassId), SCDesc(STI.getSchedModel().getSchedClassDesc(ResolvedSchedClassId)), 422 WasVariant(WasVariant), 423 NonRedundantWriteProcRes(getNonRedundantWriteProcRes(*SCDesc, STI)), 424 IdealizedProcResPressure(computeIdealizedProcResPressure( 425 STI.getSchedModel(), NonRedundantWriteProcRes)) { 426 assert((SCDesc == nullptr || !SCDesc->isVariant()) && 427 "ResolvedSchedClass should never be variant"); 428 } 429 430 void Analysis::SchedClassCluster::addPoint( 431 size_t PointId, const InstructionBenchmarkClustering &Clustering) { 432 PointIds.push_back(PointId); 433 const auto &Point = Clustering.getPoints()[PointId]; 434 if (ClusterId.isUndef()) { 435 ClusterId = Clustering.getClusterIdForPoint(PointId); 436 Representative.resize(Point.Measurements.size()); 437 } 438 for (size_t I = 0, E = Point.Measurements.size(); I < E; ++I) { 439 Representative[I].push(Point.Measurements[I]); 440 } 441 assert(ClusterId == Clustering.getClusterIdForPoint(PointId)); 442 } 443 444 // Returns a ProxResIdx by id or name. 445 static unsigned findProcResIdx(const llvm::MCSubtargetInfo &STI, 446 const llvm::StringRef NameOrId) { 447 // Interpret the key as an ProcResIdx. 448 unsigned ProcResIdx = 0; 449 if (llvm::to_integer(NameOrId, ProcResIdx, 10)) 450 return ProcResIdx; 451 // Interpret the key as a ProcRes name. 452 const auto &SchedModel = STI.getSchedModel(); 453 for (int I = 0, E = SchedModel.getNumProcResourceKinds(); I < E; ++I) { 454 if (NameOrId == SchedModel.getProcResource(I)->Name) 455 return I; 456 } 457 return 0; 458 } 459 460 bool Analysis::SchedClassCluster::measurementsMatch( 461 const llvm::MCSubtargetInfo &STI, const ResolvedSchedClass &RSC, 462 const InstructionBenchmarkClustering &Clustering) const { 463 const size_t NumMeasurements = Representative.size(); 464 std::vector<BenchmarkMeasure> ClusterCenterPoint(NumMeasurements); 465 std::vector<BenchmarkMeasure> SchedClassPoint(NumMeasurements); 466 // Latency case. 467 assert(!Clustering.getPoints().empty()); 468 const InstructionBenchmark::ModeE Mode = Clustering.getPoints()[0].Mode; 469 if (Mode == InstructionBenchmark::Latency) { 470 if (NumMeasurements != 1) { 471 llvm::errs() 472 << "invalid number of measurements in latency mode: expected 1, got " 473 << NumMeasurements << "\n"; 474 return false; 475 } 476 // Find the latency. 477 SchedClassPoint[0].PerInstructionValue = 0.0; 478 for (unsigned I = 0; I < RSC.SCDesc->NumWriteLatencyEntries; ++I) { 479 const llvm::MCWriteLatencyEntry *const WLE = 480 STI.getWriteLatencyEntry(RSC.SCDesc, I); 481 SchedClassPoint[0].PerInstructionValue = 482 std::max<double>(SchedClassPoint[0].PerInstructionValue, WLE->Cycles); 483 } 484 ClusterCenterPoint[0].PerInstructionValue = Representative[0].avg(); 485 } else if (Mode == InstructionBenchmark::Uops) { 486 for (int I = 0, E = Representative.size(); I < E; ++I) { 487 const auto Key = Representative[I].key(); 488 uint16_t ProcResIdx = findProcResIdx(STI, Key); 489 if (ProcResIdx > 0) { 490 // Find the pressure on ProcResIdx `Key`. 491 const auto ProcResPressureIt = 492 std::find_if(RSC.IdealizedProcResPressure.begin(), 493 RSC.IdealizedProcResPressure.end(), 494 [ProcResIdx](const std::pair<uint16_t, float> &WPR) { 495 return WPR.first == ProcResIdx; 496 }); 497 SchedClassPoint[I].PerInstructionValue = 498 ProcResPressureIt == RSC.IdealizedProcResPressure.end() 499 ? 0.0 500 : ProcResPressureIt->second; 501 } else if (Key == "NumMicroOps") { 502 SchedClassPoint[I].PerInstructionValue = RSC.SCDesc->NumMicroOps; 503 } else { 504 llvm::errs() << "expected `key` to be either a ProcResIdx or a ProcRes " 505 "name, got " 506 << Key << "\n"; 507 return false; 508 } 509 ClusterCenterPoint[I].PerInstructionValue = Representative[I].avg(); 510 } 511 } else if (Mode == InstructionBenchmark::InverseThroughput) { 512 for (int I = 0, E = Representative.size(); I < E; ++I) { 513 SchedClassPoint[I].PerInstructionValue = 514 MCSchedModel::getReciprocalThroughput(STI, *RSC.SCDesc); 515 ClusterCenterPoint[I].PerInstructionValue = Representative[I].min(); 516 } 517 } else { 518 llvm_unreachable("unimplemented measurement matching mode"); 519 return false; 520 } 521 return Clustering.isNeighbour(ClusterCenterPoint, SchedClassPoint); 522 } 523 524 void Analysis::printSchedClassDescHtml(const ResolvedSchedClass &RSC, 525 llvm::raw_ostream &OS) const { 526 OS << "<table class=\"sched-class-desc\">"; 527 OS << "<tr><th>Valid</th><th>Variant</th><th>NumMicroOps</th><th>Latency</" 528 "th><th>RThroughput</th><th>WriteProcRes</th><th title=\"This is the " 529 "idealized unit resource (port) pressure assuming ideal " 530 "distribution\">Idealized Resource Pressure</th></tr>"; 531 if (RSC.SCDesc->isValid()) { 532 const auto &SM = SubtargetInfo_->getSchedModel(); 533 OS << "<tr><td>✔</td>"; 534 OS << "<td>" << (RSC.WasVariant ? "✔" : "✕") << "</td>"; 535 OS << "<td>" << RSC.SCDesc->NumMicroOps << "</td>"; 536 // Latencies. 537 OS << "<td><ul>"; 538 for (int I = 0, E = RSC.SCDesc->NumWriteLatencyEntries; I < E; ++I) { 539 const auto *const Entry = 540 SubtargetInfo_->getWriteLatencyEntry(RSC.SCDesc, I); 541 OS << "<li>" << Entry->Cycles; 542 if (RSC.SCDesc->NumWriteLatencyEntries > 1) { 543 // Dismabiguate if more than 1 latency. 544 OS << " (WriteResourceID " << Entry->WriteResourceID << ")"; 545 } 546 OS << "</li>"; 547 } 548 OS << "</ul></td>"; 549 // inverse throughput. 550 OS << "<td>"; 551 writeMeasurementValue<kEscapeHtml>( 552 OS, 553 MCSchedModel::getReciprocalThroughput(*SubtargetInfo_, *RSC.SCDesc)); 554 OS << "</td>"; 555 // WriteProcRes. 556 OS << "<td><ul>"; 557 for (const auto &WPR : RSC.NonRedundantWriteProcRes) { 558 OS << "<li><span class=\"mono\">"; 559 writeEscaped<kEscapeHtml>(OS, 560 SM.getProcResource(WPR.ProcResourceIdx)->Name); 561 OS << "</span>: " << WPR.Cycles << "</li>"; 562 } 563 OS << "</ul></td>"; 564 // Idealized port pressure. 565 OS << "<td><ul>"; 566 for (const auto &Pressure : RSC.IdealizedProcResPressure) { 567 OS << "<li><span class=\"mono\">"; 568 writeEscaped<kEscapeHtml>(OS, SubtargetInfo_->getSchedModel() 569 .getProcResource(Pressure.first) 570 ->Name); 571 OS << "</span>: "; 572 writeMeasurementValue<kEscapeHtml>(OS, Pressure.second); 573 OS << "</li>"; 574 } 575 OS << "</ul></td>"; 576 OS << "</tr>"; 577 } else { 578 OS << "<tr><td>✕</td><td></td><td></td></tr>"; 579 } 580 OS << "</table>"; 581 } 582 583 static constexpr const char kHtmlHead[] = R"( 584 <head> 585 <title>llvm-exegesis Analysis Results</title> 586 <style> 587 body { 588 font-family: sans-serif 589 } 590 span.sched-class-name { 591 font-weight: bold; 592 font-family: monospace; 593 } 594 span.opcode { 595 font-family: monospace; 596 } 597 span.config { 598 font-family: monospace; 599 } 600 div.inconsistency { 601 margin-top: 50px; 602 } 603 table { 604 margin-left: 50px; 605 border-collapse: collapse; 606 } 607 table, table tr,td,th { 608 border: 1px solid #444; 609 } 610 table ul { 611 padding-left: 0px; 612 margin: 0px; 613 list-style-type: none; 614 } 615 table.sched-class-clusters td { 616 padding-left: 10px; 617 padding-right: 10px; 618 padding-top: 10px; 619 padding-bottom: 10px; 620 } 621 table.sched-class-desc td { 622 padding-left: 10px; 623 padding-right: 10px; 624 padding-top: 2px; 625 padding-bottom: 2px; 626 } 627 span.mono { 628 font-family: monospace; 629 } 630 td.measurement { 631 text-align: center; 632 } 633 tr.good-cluster td.measurement { 634 color: #292 635 } 636 tr.bad-cluster td.measurement { 637 color: #922 638 } 639 tr.good-cluster td.measurement span.minmax { 640 color: #888; 641 } 642 tr.bad-cluster td.measurement span.minmax { 643 color: #888; 644 } 645 </style> 646 </head> 647 )"; 648 649 template <> 650 llvm::Error Analysis::run<Analysis::PrintSchedClassInconsistencies>( 651 llvm::raw_ostream &OS) const { 652 const auto &FirstPoint = Clustering_.getPoints()[0]; 653 // Print the header. 654 OS << "<!DOCTYPE html><html>" << kHtmlHead << "<body>"; 655 OS << "<h1><span class=\"mono\">llvm-exegesis</span> Analysis Results</h1>"; 656 OS << "<h3>Triple: <span class=\"mono\">"; 657 writeEscaped<kEscapeHtml>(OS, FirstPoint.LLVMTriple); 658 OS << "</span></h3><h3>Cpu: <span class=\"mono\">"; 659 writeEscaped<kEscapeHtml>(OS, FirstPoint.CpuName); 660 OS << "</span></h3>"; 661 662 for (const auto &RSCAndPoints : makePointsPerSchedClass()) { 663 if (!RSCAndPoints.RSC.SCDesc) 664 continue; 665 // Bucket sched class points into sched class clusters. 666 std::vector<SchedClassCluster> SchedClassClusters; 667 for (const size_t PointId : RSCAndPoints.PointIds) { 668 const auto &ClusterId = Clustering_.getClusterIdForPoint(PointId); 669 if (!ClusterId.isValid()) 670 continue; // Ignore noise and errors. FIXME: take noise into account ? 671 auto SchedClassClusterIt = 672 std::find_if(SchedClassClusters.begin(), SchedClassClusters.end(), 673 [ClusterId](const SchedClassCluster &C) { 674 return C.id() == ClusterId; 675 }); 676 if (SchedClassClusterIt == SchedClassClusters.end()) { 677 SchedClassClusters.emplace_back(); 678 SchedClassClusterIt = std::prev(SchedClassClusters.end()); 679 } 680 SchedClassClusterIt->addPoint(PointId, Clustering_); 681 } 682 683 // Print any scheduling class that has at least one cluster that does not 684 // match the checked-in data. 685 if (llvm::all_of(SchedClassClusters, 686 [this, &RSCAndPoints](const SchedClassCluster &C) { 687 return C.measurementsMatch( 688 *SubtargetInfo_, RSCAndPoints.RSC, Clustering_); 689 })) 690 continue; // Nothing weird. 691 692 OS << "<div class=\"inconsistency\"><p>Sched Class <span " 693 "class=\"sched-class-name\">"; 694 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) 695 writeEscaped<kEscapeHtml>(OS, RSCAndPoints.RSC.SCDesc->Name); 696 #else 697 OS << RSCAndPoints.RSC.SchedClassId; 698 #endif 699 OS << "</span> contains instructions whose performance characteristics do" 700 " not match that of LLVM:</p>"; 701 printSchedClassClustersHtml(SchedClassClusters, RSCAndPoints.RSC, OS); 702 OS << "<p>llvm SchedModel data:</p>"; 703 printSchedClassDescHtml(RSCAndPoints.RSC, OS); 704 OS << "</div>"; 705 } 706 707 OS << "</body></html>"; 708 return llvm::Error::success(); 709 } 710 711 // Distributes a pressure budget as evenly as possible on the provided subunits 712 // given the already existing port pressure distribution. 713 // 714 // The algorithm is as follows: while there is remaining pressure to 715 // distribute, find the subunits with minimal pressure, and distribute 716 // remaining pressure equally up to the pressure of the unit with 717 // second-to-minimal pressure. 718 // For example, let's assume we want to distribute 2*P1256 719 // (Subunits = [P1,P2,P5,P6]), and the starting DensePressure is: 720 // DensePressure = P0 P1 P2 P3 P4 P5 P6 P7 721 // 0.1 0.3 0.2 0.0 0.0 0.5 0.5 0.5 722 // RemainingPressure = 2.0 723 // We sort the subunits by pressure: 724 // Subunits = [(P2,p=0.2), (P1,p=0.3), (P5,p=0.5), (P6, p=0.5)] 725 // We'll first start by the subunits with minimal pressure, which are at 726 // the beginning of the sorted array. In this example there is one (P2). 727 // The subunit with second-to-minimal pressure is the next one in the 728 // array (P1). So we distribute 0.1 pressure to P2, and remove 0.1 cycles 729 // from the budget. 730 // Subunits = [(P2,p=0.3), (P1,p=0.3), (P5,p=0.5), (P5,p=0.5)] 731 // RemainingPressure = 1.9 732 // We repeat this process: distribute 0.2 pressure on each of the minimal 733 // P2 and P1, decrease budget by 2*0.2: 734 // Subunits = [(P2,p=0.5), (P1,p=0.5), (P5,p=0.5), (P5,p=0.5)] 735 // RemainingPressure = 1.5 736 // There are no second-to-minimal subunits so we just share the remaining 737 // budget (1.5 cycles) equally: 738 // Subunits = [(P2,p=0.875), (P1,p=0.875), (P5,p=0.875), (P5,p=0.875)] 739 // RemainingPressure = 0.0 740 // We stop as there is no remaining budget to distribute. 741 void distributePressure(float RemainingPressure, 742 llvm::SmallVector<uint16_t, 32> Subunits, 743 llvm::SmallVector<float, 32> &DensePressure) { 744 // Find the number of subunits with minimal pressure (they are at the 745 // front). 746 llvm::sort(Subunits, [&DensePressure](const uint16_t A, const uint16_t B) { 747 return DensePressure[A] < DensePressure[B]; 748 }); 749 const auto getPressureForSubunit = [&DensePressure, 750 &Subunits](size_t I) -> float & { 751 return DensePressure[Subunits[I]]; 752 }; 753 size_t NumMinimalSU = 1; 754 while (NumMinimalSU < Subunits.size() && 755 getPressureForSubunit(NumMinimalSU) == getPressureForSubunit(0)) { 756 ++NumMinimalSU; 757 } 758 while (RemainingPressure > 0.0f) { 759 if (NumMinimalSU == Subunits.size()) { 760 // All units are minimal, just distribute evenly and be done. 761 for (size_t I = 0; I < NumMinimalSU; ++I) { 762 getPressureForSubunit(I) += RemainingPressure / NumMinimalSU; 763 } 764 return; 765 } 766 // Distribute the remaining pressure equally. 767 const float MinimalPressure = getPressureForSubunit(NumMinimalSU - 1); 768 const float SecondToMinimalPressure = getPressureForSubunit(NumMinimalSU); 769 assert(MinimalPressure < SecondToMinimalPressure); 770 const float Increment = SecondToMinimalPressure - MinimalPressure; 771 if (RemainingPressure <= NumMinimalSU * Increment) { 772 // There is not enough remaining pressure. 773 for (size_t I = 0; I < NumMinimalSU; ++I) { 774 getPressureForSubunit(I) += RemainingPressure / NumMinimalSU; 775 } 776 return; 777 } 778 // Bump all minimal pressure subunits to `SecondToMinimalPressure`. 779 for (size_t I = 0; I < NumMinimalSU; ++I) { 780 getPressureForSubunit(I) = SecondToMinimalPressure; 781 RemainingPressure -= SecondToMinimalPressure; 782 } 783 while (NumMinimalSU < Subunits.size() && 784 getPressureForSubunit(NumMinimalSU) == SecondToMinimalPressure) { 785 ++NumMinimalSU; 786 } 787 } 788 } 789 790 std::vector<std::pair<uint16_t, float>> computeIdealizedProcResPressure( 791 const llvm::MCSchedModel &SM, 792 llvm::SmallVector<llvm::MCWriteProcResEntry, 8> WPRS) { 793 // DensePressure[I] is the port pressure for Proc Resource I. 794 llvm::SmallVector<float, 32> DensePressure(SM.getNumProcResourceKinds()); 795 llvm::sort(WPRS, [](const llvm::MCWriteProcResEntry &A, 796 const llvm::MCWriteProcResEntry &B) { 797 return A.ProcResourceIdx < B.ProcResourceIdx; 798 }); 799 for (const llvm::MCWriteProcResEntry &WPR : WPRS) { 800 // Get units for the entry. 801 const llvm::MCProcResourceDesc *const ProcResDesc = 802 SM.getProcResource(WPR.ProcResourceIdx); 803 if (ProcResDesc->SubUnitsIdxBegin == nullptr) { 804 // This is a ProcResUnit. 805 DensePressure[WPR.ProcResourceIdx] += WPR.Cycles; 806 } else { 807 // This is a ProcResGroup. 808 llvm::SmallVector<uint16_t, 32> Subunits(ProcResDesc->SubUnitsIdxBegin, 809 ProcResDesc->SubUnitsIdxBegin + 810 ProcResDesc->NumUnits); 811 distributePressure(WPR.Cycles, Subunits, DensePressure); 812 } 813 } 814 // Turn dense pressure into sparse pressure by removing zero entries. 815 std::vector<std::pair<uint16_t, float>> Pressure; 816 for (unsigned I = 0, E = SM.getNumProcResourceKinds(); I < E; ++I) { 817 if (DensePressure[I] > 0.0f) 818 Pressure.emplace_back(I, DensePressure[I]); 819 } 820 return Pressure; 821 } 822 823 } // namespace exegesis 824 } // namespace llvm 825