| 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/MC/MCTargetOptions.h" |
| 14 | #include "llvm/Support/FormatVariadic.h" |
| 15 | #include <limits> |
| 16 | #include <vector> |
| 17 | |
| 18 | namespace llvm { |
| 19 | namespace exegesis { |
| 20 | |
| 21 | static const char kCsvSep = ','; |
| 22 | |
| 23 | namespace { |
| 24 | |
| 25 | enum EscapeTag { kEscapeCsv, kEscapeHtml, kEscapeHtmlString }; |
| 26 | |
| 27 | template <EscapeTag Tag> void writeEscaped(raw_ostream &OS, const StringRef S); |
| 28 | |
| 29 | template <> void writeEscaped<kEscapeCsv>(raw_ostream &OS, const StringRef S) { |
| 30 | if (!S.contains(C: kCsvSep)) { |
| 31 | OS << S; |
| 32 | } else { |
| 33 | // Needs escaping. |
| 34 | OS << '"'; |
| 35 | for (const char C : S) { |
| 36 | if (C == '"') |
| 37 | OS << "\"\"" ; |
| 38 | else |
| 39 | OS << C; |
| 40 | } |
| 41 | OS << '"'; |
| 42 | } |
| 43 | } |
| 44 | |
| 45 | template <> void writeEscaped<kEscapeHtml>(raw_ostream &OS, const StringRef S) { |
| 46 | for (const char C : S) { |
| 47 | if (C == '<') |
| 48 | OS << "<" ; |
| 49 | else if (C == '>') |
| 50 | OS << ">" ; |
| 51 | else if (C == '&') |
| 52 | OS << "&" ; |
| 53 | else |
| 54 | OS << C; |
| 55 | } |
| 56 | } |
| 57 | |
| 58 | template <> |
| 59 | void writeEscaped<kEscapeHtmlString>(raw_ostream &OS, const StringRef S) { |
| 60 | for (const char C : S) { |
| 61 | if (C == '"') |
| 62 | OS << "\\\"" ; |
| 63 | else |
| 64 | OS << C; |
| 65 | } |
| 66 | } |
| 67 | |
| 68 | } // namespace |
| 69 | |
| 70 | template <EscapeTag Tag> |
| 71 | static void |
| 72 | writeClusterId(raw_ostream &OS, |
| 73 | const BenchmarkClustering::ClusterId &CID) { |
| 74 | if (CID.isNoise()) |
| 75 | writeEscaped<Tag>(OS, "[noise]" ); |
| 76 | else if (CID.isError()) |
| 77 | writeEscaped<Tag>(OS, "[error]" ); |
| 78 | else |
| 79 | OS << CID.getId(); |
| 80 | } |
| 81 | |
| 82 | template <EscapeTag Tag> |
| 83 | static void writeMeasurementValue(raw_ostream &OS, const double Value) { |
| 84 | // Given Value, if we wanted to serialize it to a string, |
| 85 | // how many base-10 digits will we need to store, max? |
| 86 | static constexpr auto MaxDigitCount = |
| 87 | std::numeric_limits<decltype(Value)>::max_digits10; |
| 88 | // Also, we will need a decimal separator. |
| 89 | static constexpr auto DecimalSeparatorLen = 1; // '.' e.g. |
| 90 | // So how long of a string will the serialization produce, max? |
| 91 | static constexpr auto SerializationLen = MaxDigitCount + DecimalSeparatorLen; |
| 92 | |
| 93 | // WARNING: when changing the format, also adjust the small-size estimate ^. |
| 94 | static constexpr StringLiteral SimpleFloatFormat = StringLiteral("{0:F}" ); |
| 95 | |
| 96 | writeEscaped<Tag>( |
| 97 | OS, formatv(Fmt: SimpleFloatFormat.data(), Vals: Value).sstr<SerializationLen>()); |
| 98 | } |
| 99 | |
| 100 | template <typename EscapeTag, EscapeTag Tag> |
| 101 | void Analysis::writeSnippet(raw_ostream &OS, ArrayRef<uint8_t> Bytes, |
| 102 | const char *Separator) const { |
| 103 | SmallVector<std::string, 3> Lines; |
| 104 | // Parse the asm snippet and print it. |
| 105 | while (!Bytes.empty()) { |
| 106 | MCInst MI; |
| 107 | uint64_t MISize = 0; |
| 108 | if (!DisasmHelper_->decodeInst(MI, MISize, Bytes)) { |
| 109 | writeEscaped<Tag>(OS, join(R&: Lines, Separator)); |
| 110 | writeEscaped<Tag>(OS, Separator); |
| 111 | writeEscaped<Tag>(OS, "[error decoding asm snippet]" ); |
| 112 | return; |
| 113 | } |
| 114 | SmallString<128> InstPrinterStr; // FIXME: magic number. |
| 115 | raw_svector_ostream OSS(InstPrinterStr); |
| 116 | DisasmHelper_->printInst(MI: &MI, OS&: OSS); |
| 117 | Bytes = Bytes.drop_front(N: MISize); |
| 118 | Lines.emplace_back(Args: InstPrinterStr.str().trim()); |
| 119 | } |
| 120 | writeEscaped<Tag>(OS, join(R&: Lines, Separator)); |
| 121 | } |
| 122 | |
| 123 | // Prints a row representing an instruction, along with scheduling info and |
| 124 | // point coordinates (measurements). |
| 125 | void Analysis::printInstructionRowCsv(const size_t PointId, |
| 126 | raw_ostream &OS) const { |
| 127 | const Benchmark &Point = Clustering_.getPoints()[PointId]; |
| 128 | writeClusterId<kEscapeCsv>(OS, CID: Clustering_.getClusterIdForPoint(P: PointId)); |
| 129 | OS << kCsvSep; |
| 130 | writeSnippet<EscapeTag, kEscapeCsv>(OS, Bytes: Point.AssembledSnippet, Separator: "; " ); |
| 131 | OS << kCsvSep; |
| 132 | writeEscaped<kEscapeCsv>(OS, S: Point.Key.Config); |
| 133 | OS << kCsvSep; |
| 134 | assert(!Point.Key.Instructions.empty()); |
| 135 | const MCInst &MCI = Point.keyInstruction(); |
| 136 | unsigned SchedClassId; |
| 137 | std::tie(args&: SchedClassId, args: std::ignore) = ResolvedSchedClass::resolveSchedClassId( |
| 138 | SubtargetInfo: State_.getSubtargetInfo(), InstrInfo: State_.getInstrInfo(), MCI); |
| 139 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) |
| 140 | StringRef SCDescName = |
| 141 | State_.getSubtargetInfo().getSchedModel().getSchedClassName(SchedClassId); |
| 142 | writeEscaped<kEscapeCsv>(OS, SCDescName); |
| 143 | #else |
| 144 | OS << SchedClassId; |
| 145 | #endif |
| 146 | for (const auto &Measurement : Point.Measurements) { |
| 147 | OS << kCsvSep; |
| 148 | writeMeasurementValue<kEscapeCsv>(OS, Value: Measurement.PerInstructionValue); |
| 149 | } |
| 150 | OS << "\n" ; |
| 151 | } |
| 152 | |
| 153 | Analysis::Analysis(const LLVMState &State, |
| 154 | const BenchmarkClustering &Clustering, |
| 155 | double AnalysisInconsistencyEpsilon, |
| 156 | bool AnalysisDisplayUnstableOpcodes) |
| 157 | : Clustering_(Clustering), State_(State), |
| 158 | AnalysisInconsistencyEpsilonSquared_(AnalysisInconsistencyEpsilon * |
| 159 | AnalysisInconsistencyEpsilon), |
| 160 | AnalysisDisplayUnstableOpcodes_(AnalysisDisplayUnstableOpcodes) { |
| 161 | if (Clustering.getPoints().empty()) |
| 162 | return; |
| 163 | |
| 164 | DisasmHelper_ = std::make_unique<DisassemblerHelper>(args: State); |
| 165 | } |
| 166 | |
| 167 | template <> |
| 168 | Error Analysis::run<Analysis::PrintClusters>(raw_ostream &OS) const { |
| 169 | if (Clustering_.getPoints().empty()) |
| 170 | return Error::success(); |
| 171 | |
| 172 | // Write the header. |
| 173 | OS << "cluster_id" << kCsvSep << "opcode_name" << kCsvSep << "config" |
| 174 | << kCsvSep << "sched_class" ; |
| 175 | for (const auto &Measurement : Clustering_.getPoints().front().Measurements) { |
| 176 | OS << kCsvSep; |
| 177 | writeEscaped<kEscapeCsv>(OS, S: Measurement.Key); |
| 178 | } |
| 179 | OS << "\n" ; |
| 180 | |
| 181 | // Write the points. |
| 182 | for (const auto &ClusterIt : Clustering_.getValidClusters()) { |
| 183 | for (const size_t PointId : ClusterIt.PointIndices) { |
| 184 | printInstructionRowCsv(PointId, OS); |
| 185 | } |
| 186 | OS << "\n\n" ; |
| 187 | } |
| 188 | return Error::success(); |
| 189 | } |
| 190 | |
| 191 | Analysis::ResolvedSchedClassAndPoints::ResolvedSchedClassAndPoints( |
| 192 | ResolvedSchedClass &&RSC) |
| 193 | : RSC(std::move(RSC)) {} |
| 194 | |
| 195 | std::vector<Analysis::ResolvedSchedClassAndPoints> |
| 196 | Analysis::makePointsPerSchedClass() const { |
| 197 | std::vector<ResolvedSchedClassAndPoints> Entries; |
| 198 | // Maps SchedClassIds to index in result. |
| 199 | std::unordered_map<unsigned, size_t> SchedClassIdToIndex; |
| 200 | const auto &Points = Clustering_.getPoints(); |
| 201 | for (size_t PointId = 0, E = Points.size(); PointId < E; ++PointId) { |
| 202 | const Benchmark &Point = Points[PointId]; |
| 203 | if (!Point.Error.empty()) |
| 204 | continue; |
| 205 | assert(!Point.Key.Instructions.empty()); |
| 206 | // FIXME: we should be using the tuple of classes for instructions in the |
| 207 | // snippet as key. |
| 208 | const MCInst &MCI = Point.keyInstruction(); |
| 209 | unsigned SchedClassId; |
| 210 | bool WasVariant; |
| 211 | std::tie(args&: SchedClassId, args&: WasVariant) = |
| 212 | ResolvedSchedClass::resolveSchedClassId(SubtargetInfo: State_.getSubtargetInfo(), |
| 213 | InstrInfo: State_.getInstrInfo(), MCI); |
| 214 | const auto IndexIt = SchedClassIdToIndex.find(x: SchedClassId); |
| 215 | if (IndexIt == SchedClassIdToIndex.end()) { |
| 216 | // Create a new entry. |
| 217 | SchedClassIdToIndex.emplace(args&: SchedClassId, args: Entries.size()); |
| 218 | ResolvedSchedClassAndPoints Entry(ResolvedSchedClass( |
| 219 | State_.getSubtargetInfo(), SchedClassId, WasVariant)); |
| 220 | Entry.PointIds.push_back(x: PointId); |
| 221 | Entries.push_back(x: std::move(Entry)); |
| 222 | } else { |
| 223 | // Append to the existing entry. |
| 224 | Entries[IndexIt->second].PointIds.push_back(x: PointId); |
| 225 | } |
| 226 | } |
| 227 | return Entries; |
| 228 | } |
| 229 | |
| 230 | // Parallel benchmarks repeat the same opcode multiple times. Just show this |
| 231 | // opcode and show the whole snippet only on hover. |
| 232 | static void writeParallelSnippetHtml(raw_ostream &OS, |
| 233 | const std::vector<MCInst> &Instructions, |
| 234 | const MCInstrInfo &InstrInfo) { |
| 235 | if (Instructions.empty()) |
| 236 | return; |
| 237 | writeEscaped<kEscapeHtml>(OS, S: InstrInfo.getName(Opcode: Instructions[0].getOpcode())); |
| 238 | if (Instructions.size() > 1) |
| 239 | OS << " (x" << Instructions.size() << ")" ; |
| 240 | } |
| 241 | |
| 242 | // Latency tries to find a serial path. Just show the opcode path and show the |
| 243 | // whole snippet only on hover. |
| 244 | static void writeLatencySnippetHtml(raw_ostream &OS, |
| 245 | const std::vector<MCInst> &Instructions, |
| 246 | const MCInstrInfo &InstrInfo) { |
| 247 | bool First = true; |
| 248 | for (const MCInst &Instr : Instructions) { |
| 249 | if (First) |
| 250 | First = false; |
| 251 | else |
| 252 | OS << " → " ; |
| 253 | writeEscaped<kEscapeHtml>(OS, S: InstrInfo.getName(Opcode: Instr.getOpcode())); |
| 254 | } |
| 255 | } |
| 256 | |
| 257 | void Analysis::printPointHtml(const Benchmark &Point, raw_ostream &OS) const { |
| 258 | OS << "<li><span class=\"mono\" title=\"" ; |
| 259 | writeSnippet<EscapeTag, kEscapeHtmlString>(OS, Bytes: Point.AssembledSnippet, Separator: "\n" ); |
| 260 | OS << "\">" ; |
| 261 | switch (Point.Mode) { |
| 262 | case Benchmark::Latency: |
| 263 | writeLatencySnippetHtml(OS, Instructions: Point.Key.Instructions, InstrInfo: State_.getInstrInfo()); |
| 264 | break; |
| 265 | case Benchmark::Uops: |
| 266 | case Benchmark::InverseThroughput: |
| 267 | writeParallelSnippetHtml(OS, Instructions: Point.Key.Instructions, InstrInfo: State_.getInstrInfo()); |
| 268 | break; |
| 269 | default: |
| 270 | llvm_unreachable("invalid mode" ); |
| 271 | } |
| 272 | OS << "</span> <span class=\"mono\">" ; |
| 273 | writeEscaped<kEscapeHtml>(OS, S: Point.Key.Config); |
| 274 | OS << "</span></li>" ; |
| 275 | } |
| 276 | |
| 277 | void Analysis::printSchedClassClustersHtml( |
| 278 | const std::vector<SchedClassCluster> &Clusters, |
| 279 | const ResolvedSchedClass &RSC, raw_ostream &OS) const { |
| 280 | const auto &Points = Clustering_.getPoints(); |
| 281 | OS << "<table class=\"sched-class-clusters\">" ; |
| 282 | OS << "<tr><th>ClusterId</th><th>Opcode/Config</th>" ; |
| 283 | assert(!Clusters.empty()); |
| 284 | for (const auto &Measurement : |
| 285 | Points[Clusters[0].getPointIds()[0]].Measurements) { |
| 286 | OS << "<th>" ; |
| 287 | writeEscaped<kEscapeHtml>(OS, S: Measurement.Key); |
| 288 | OS << "</th>" ; |
| 289 | } |
| 290 | OS << "</tr>" ; |
| 291 | for (const SchedClassCluster &Cluster : Clusters) { |
| 292 | OS << "<tr class=\"" |
| 293 | << (Cluster.measurementsMatch(STI: State_.getSubtargetInfo(), SC: RSC, |
| 294 | Clustering: Clustering_, |
| 295 | AnalysisInconsistencyEpsilonSquared_) |
| 296 | ? "good-cluster" |
| 297 | : "bad-cluster" ) |
| 298 | << "\"><td>" ; |
| 299 | writeClusterId<kEscapeHtml>(OS, CID: Cluster.id()); |
| 300 | OS << "</td><td><ul>" ; |
| 301 | for (const size_t PointId : Cluster.getPointIds()) { |
| 302 | printPointHtml(Point: Points[PointId], OS); |
| 303 | } |
| 304 | OS << "</ul></td>" ; |
| 305 | for (const auto &Stats : Cluster.getCentroid().getStats()) { |
| 306 | OS << "<td class=\"measurement\">" ; |
| 307 | writeMeasurementValue<kEscapeHtml>(OS, Value: Stats.avg()); |
| 308 | OS << "<br><span class=\"minmax\">[" ; |
| 309 | writeMeasurementValue<kEscapeHtml>(OS, Value: Stats.min()); |
| 310 | OS << ";" ; |
| 311 | writeMeasurementValue<kEscapeHtml>(OS, Value: Stats.max()); |
| 312 | OS << "]</span></td>" ; |
| 313 | } |
| 314 | OS << "</tr>" ; |
| 315 | } |
| 316 | OS << "</table>" ; |
| 317 | } |
| 318 | |
| 319 | void Analysis::SchedClassCluster::addPoint( |
| 320 | size_t PointId, const BenchmarkClustering &Clustering) { |
| 321 | PointIds.push_back(x: PointId); |
| 322 | const auto &Point = Clustering.getPoints()[PointId]; |
| 323 | if (ClusterId.isUndef()) |
| 324 | ClusterId = Clustering.getClusterIdForPoint(P: PointId); |
| 325 | assert(ClusterId == Clustering.getClusterIdForPoint(PointId)); |
| 326 | |
| 327 | Centroid.addPoint(Point: Point.Measurements); |
| 328 | } |
| 329 | |
| 330 | bool Analysis::SchedClassCluster::measurementsMatch( |
| 331 | const MCSubtargetInfo &STI, const ResolvedSchedClass &RSC, |
| 332 | const BenchmarkClustering &Clustering, |
| 333 | const double AnalysisInconsistencyEpsilonSquared_) const { |
| 334 | assert(!Clustering.getPoints().empty()); |
| 335 | const Benchmark::ModeE Mode = Clustering.getPoints()[0].Mode; |
| 336 | |
| 337 | if (!Centroid.validate(Mode)) |
| 338 | return false; |
| 339 | |
| 340 | const std::vector<BenchmarkMeasure> ClusterCenterPoint = |
| 341 | Centroid.getAsPoint(); |
| 342 | |
| 343 | const std::vector<BenchmarkMeasure> SchedClassPoint = |
| 344 | RSC.getAsPoint(Mode, STI, Representative: Centroid.getStats()); |
| 345 | if (SchedClassPoint.empty()) |
| 346 | return false; // In Uops mode validate() may not be enough. |
| 347 | |
| 348 | assert(ClusterCenterPoint.size() == SchedClassPoint.size() && |
| 349 | "Expected measured/sched data dimensions to match." ); |
| 350 | |
| 351 | return Clustering.isNeighbour(P: ClusterCenterPoint, Q: SchedClassPoint, |
| 352 | EpsilonSquared_: AnalysisInconsistencyEpsilonSquared_); |
| 353 | } |
| 354 | |
| 355 | void Analysis::printSchedClassDescHtml(const ResolvedSchedClass &RSC, |
| 356 | raw_ostream &OS) const { |
| 357 | OS << "<table class=\"sched-class-desc\">" ; |
| 358 | OS << "<tr><th>Valid</th><th>Variant</th><th>NumMicroOps</th><th>Latency</" |
| 359 | "th><th>RThroughput</th><th>WriteProcRes</th><th title=\"This is the " |
| 360 | "idealized unit resource (port) pressure assuming ideal " |
| 361 | "distribution\">Idealized Resource Pressure</th></tr>" ; |
| 362 | if (RSC.SCDesc->isValid()) { |
| 363 | const auto &SI = State_.getSubtargetInfo(); |
| 364 | const auto &SM = SI.getSchedModel(); |
| 365 | OS << "<tr><td>✔</td>" ; |
| 366 | OS << "<td>" << (RSC.WasVariant ? "✔" : "✕" ) << "</td>" ; |
| 367 | OS << "<td>" << RSC.SCDesc->NumMicroOps << "</td>" ; |
| 368 | // Latencies. |
| 369 | OS << "<td><ul>" ; |
| 370 | for (int I = 0, E = RSC.SCDesc->NumWriteLatencyEntries; I < E; ++I) { |
| 371 | const auto *const Entry = SI.getWriteLatencyEntry(SC: RSC.SCDesc, DefIdx: I); |
| 372 | OS << "<li>" << Entry->Cycles; |
| 373 | if (RSC.SCDesc->NumWriteLatencyEntries > 1) { |
| 374 | // Dismabiguate if more than 1 latency. |
| 375 | OS << " (WriteResourceID " << Entry->WriteResourceID << ")" ; |
| 376 | } |
| 377 | OS << "</li>" ; |
| 378 | } |
| 379 | OS << "</ul></td>" ; |
| 380 | // inverse throughput. |
| 381 | OS << "<td>" ; |
| 382 | writeMeasurementValue<kEscapeHtml>( |
| 383 | OS, Value: MCSchedModel::getReciprocalThroughput(STI: SI, SCDesc: *RSC.SCDesc)); |
| 384 | OS << "</td>" ; |
| 385 | // WriteProcRes. |
| 386 | OS << "<td><ul>" ; |
| 387 | for (const auto &WPR : RSC.NonRedundantWriteProcRes) { |
| 388 | OS << "<li><span class=\"mono\">" ; |
| 389 | writeEscaped<kEscapeHtml>(OS, |
| 390 | S: SM.getProcResource(ProcResourceIdx: WPR.ProcResourceIdx)->Name); |
| 391 | OS << "</span>: " << WPR.ReleaseAtCycle << "</li>" ; |
| 392 | } |
| 393 | OS << "</ul></td>" ; |
| 394 | // Idealized port pressure. |
| 395 | OS << "<td><ul>" ; |
| 396 | for (const auto &Pressure : RSC.IdealizedProcResPressure) { |
| 397 | OS << "<li><span class=\"mono\">" ; |
| 398 | writeEscaped<kEscapeHtml>( |
| 399 | OS, S: SI.getSchedModel().getProcResource(ProcResourceIdx: Pressure.first)->Name); |
| 400 | OS << "</span>: " ; |
| 401 | writeMeasurementValue<kEscapeHtml>(OS, Value: Pressure.second); |
| 402 | OS << "</li>" ; |
| 403 | } |
| 404 | OS << "</ul></td>" ; |
| 405 | OS << "</tr>" ; |
| 406 | } else { |
| 407 | OS << "<tr><td>✕</td><td></td><td></td></tr>" ; |
| 408 | } |
| 409 | OS << "</table>" ; |
| 410 | } |
| 411 | |
| 412 | void Analysis::printClusterRawHtml(const BenchmarkClustering::ClusterId &Id, |
| 413 | StringRef display_name, |
| 414 | raw_ostream &OS) const { |
| 415 | const auto &Points = Clustering_.getPoints(); |
| 416 | const auto &Cluster = Clustering_.getCluster(Id); |
| 417 | if (Cluster.PointIndices.empty()) |
| 418 | return; |
| 419 | |
| 420 | OS << "<div class=\"inconsistency\"><p>" << display_name << " Cluster (" |
| 421 | << Cluster.PointIndices.size() << " points)</p>" ; |
| 422 | OS << "<table class=\"sched-class-clusters\">" ; |
| 423 | // Table Header. |
| 424 | OS << "<tr><th>ClusterId</th><th>Opcode/Config</th>" ; |
| 425 | for (const auto &Measurement : Points[Cluster.PointIndices[0]].Measurements) { |
| 426 | OS << "<th>" ; |
| 427 | writeEscaped<kEscapeHtml>(OS, S: Measurement.Key); |
| 428 | OS << "</th>" ; |
| 429 | } |
| 430 | OS << "</tr>" ; |
| 431 | |
| 432 | // Point data. |
| 433 | for (const auto &PointId : Cluster.PointIndices) { |
| 434 | OS << "<tr class=\"bad-cluster\"><td>" << display_name << "</td><td><ul>" ; |
| 435 | printPointHtml(Point: Points[PointId], OS); |
| 436 | OS << "</ul></td>" ; |
| 437 | for (const auto &Measurement : Points[PointId].Measurements) { |
| 438 | OS << "<td class=\"measurement\">" ; |
| 439 | writeMeasurementValue<kEscapeHtml>(OS, Value: Measurement.PerInstructionValue); |
| 440 | } |
| 441 | OS << "</tr>" ; |
| 442 | } |
| 443 | OS << "</table>" ; |
| 444 | |
| 445 | OS << "</div>" ; |
| 446 | |
| 447 | } // namespace exegesis |
| 448 | |
| 449 | static constexpr const char kHtmlHead[] = R"( |
| 450 | <head> |
| 451 | <title>llvm-exegesis Analysis Results</title> |
| 452 | <style> |
| 453 | body { |
| 454 | font-family: sans-serif |
| 455 | } |
| 456 | span.sched-class-name { |
| 457 | font-weight: bold; |
| 458 | font-family: monospace; |
| 459 | } |
| 460 | span.opcode { |
| 461 | font-family: monospace; |
| 462 | } |
| 463 | span.config { |
| 464 | font-family: monospace; |
| 465 | } |
| 466 | div.inconsistency { |
| 467 | margin-top: 50px; |
| 468 | } |
| 469 | table { |
| 470 | margin-left: 50px; |
| 471 | border-collapse: collapse; |
| 472 | } |
| 473 | table, table tr,td,th { |
| 474 | border: 1px solid #444; |
| 475 | } |
| 476 | table ul { |
| 477 | padding-left: 0px; |
| 478 | margin: 0px; |
| 479 | list-style-type: none; |
| 480 | } |
| 481 | table.sched-class-clusters td { |
| 482 | padding-left: 10px; |
| 483 | padding-right: 10px; |
| 484 | padding-top: 10px; |
| 485 | padding-bottom: 10px; |
| 486 | } |
| 487 | table.sched-class-desc td { |
| 488 | padding-left: 10px; |
| 489 | padding-right: 10px; |
| 490 | padding-top: 2px; |
| 491 | padding-bottom: 2px; |
| 492 | } |
| 493 | span.mono { |
| 494 | font-family: monospace; |
| 495 | } |
| 496 | td.measurement { |
| 497 | text-align: center; |
| 498 | } |
| 499 | tr.good-cluster td.measurement { |
| 500 | color: #292 |
| 501 | } |
| 502 | tr.bad-cluster td.measurement { |
| 503 | color: #922 |
| 504 | } |
| 505 | tr.good-cluster td.measurement span.minmax { |
| 506 | color: #888; |
| 507 | } |
| 508 | tr.bad-cluster td.measurement span.minmax { |
| 509 | color: #888; |
| 510 | } |
| 511 | </style> |
| 512 | </head> |
| 513 | )" ; |
| 514 | |
| 515 | template <> |
| 516 | Error Analysis::run<Analysis::PrintSchedClassInconsistencies>( |
| 517 | raw_ostream &OS) const { |
| 518 | const auto &FirstPoint = Clustering_.getPoints()[0]; |
| 519 | // Print the header. |
| 520 | OS << "<!DOCTYPE html><html>" << kHtmlHead << "<body>" ; |
| 521 | OS << "<h1><span class=\"mono\">llvm-exegesis</span> Analysis Results</h1>" ; |
| 522 | OS << "<h3>Triple: <span class=\"mono\">" ; |
| 523 | writeEscaped<kEscapeHtml>(OS, S: FirstPoint.LLVMTriple); |
| 524 | OS << "</span></h3><h3>Cpu: <span class=\"mono\">" ; |
| 525 | writeEscaped<kEscapeHtml>(OS, S: FirstPoint.CpuName); |
| 526 | OS << "</span></h3>" ; |
| 527 | OS << "<h3>Epsilon: <span class=\"mono\">" |
| 528 | << format(Fmt: "%0.2f" , Vals: std::sqrt(x: AnalysisInconsistencyEpsilonSquared_)) |
| 529 | << "</span></h3>" ; |
| 530 | |
| 531 | const auto &SI = State_.getSubtargetInfo(); |
| 532 | for (const auto &RSCAndPoints : makePointsPerSchedClass()) { |
| 533 | if (!RSCAndPoints.RSC.SCDesc) |
| 534 | continue; |
| 535 | // Bucket sched class points into sched class clusters. |
| 536 | std::vector<SchedClassCluster> SchedClassClusters; |
| 537 | for (const size_t PointId : RSCAndPoints.PointIds) { |
| 538 | const auto &ClusterId = Clustering_.getClusterIdForPoint(P: PointId); |
| 539 | if (!ClusterId.isValid()) |
| 540 | continue; // Ignore noise and errors. FIXME: take noise into account ? |
| 541 | if (ClusterId.isUnstable() ^ AnalysisDisplayUnstableOpcodes_) |
| 542 | continue; // Either display stable or unstable clusters only. |
| 543 | auto SchedClassClusterIt = |
| 544 | find_if(Range&: SchedClassClusters, P: [ClusterId](const SchedClassCluster &C) { |
| 545 | return C.id() == ClusterId; |
| 546 | }); |
| 547 | if (SchedClassClusterIt == SchedClassClusters.end()) { |
| 548 | SchedClassClusters.emplace_back(); |
| 549 | SchedClassClusterIt = std::prev(x: SchedClassClusters.end()); |
| 550 | } |
| 551 | SchedClassClusterIt->addPoint(PointId, Clustering: Clustering_); |
| 552 | } |
| 553 | |
| 554 | // Print any scheduling class that has at least one cluster that does not |
| 555 | // match the checked-in data. |
| 556 | if (all_of(Range&: SchedClassClusters, P: [this, &RSCAndPoints, |
| 557 | &SI](const SchedClassCluster &C) { |
| 558 | return C.measurementsMatch(STI: SI, RSC: RSCAndPoints.RSC, Clustering: Clustering_, |
| 559 | AnalysisInconsistencyEpsilonSquared_); |
| 560 | })) |
| 561 | continue; // Nothing weird. |
| 562 | |
| 563 | OS << "<div class=\"inconsistency\"><p>Sched Class <span " |
| 564 | "class=\"sched-class-name\">" ; |
| 565 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) |
| 566 | writeEscaped<kEscapeHtml>(OS, SI.getSchedModel().getSchedClassName( |
| 567 | RSCAndPoints.RSC.SchedClassId)); |
| 568 | #else |
| 569 | OS << RSCAndPoints.RSC.SchedClassId; |
| 570 | #endif |
| 571 | OS << "</span> contains instructions whose performance characteristics do" |
| 572 | " not match that of LLVM:</p>" ; |
| 573 | printSchedClassClustersHtml(Clusters: SchedClassClusters, RSC: RSCAndPoints.RSC, OS); |
| 574 | OS << "<p>llvm SchedModel data:</p>" ; |
| 575 | printSchedClassDescHtml(RSC: RSCAndPoints.RSC, OS); |
| 576 | OS << "</div>" ; |
| 577 | } |
| 578 | |
| 579 | printClusterRawHtml(Id: BenchmarkClustering::ClusterId::noise(), |
| 580 | display_name: "[noise]" , OS); |
| 581 | |
| 582 | OS << "</body></html>" ; |
| 583 | return Error::success(); |
| 584 | } |
| 585 | |
| 586 | } // namespace exegesis |
| 587 | } // namespace llvm |
| 588 | |