tools/llvm-mca/Views/SchedulerStatistics.cpp

   1 //===--------------------- SchedulerStatistics.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 /// \file
   9 ///
  10 /// This file implements the SchedulerStatistics interface.
  11 ///
  12 //===----------------------------------------------------------------------===//
  13
  14 #include "Views/SchedulerStatistics.h"
  15 #include "llvm/Support/Format.h"
  16 #include "llvm/Support/FormattedStream.h"
  17
  18 namespace llvm {
  19 namespace mca {
  20
  21 SchedulerStatistics::SchedulerStatistics(const llvm::MCSubtargetInfo &STI)
  22     : SM(STI.getSchedModel()), LQResourceID(0), SQResourceID(0), NumIssued(0),
  23       NumCycles(0), MostRecentLoadDispatched(~0U),
  24       MostRecentStoreDispatched(~0U),
  25       Usage(STI.getSchedModel().NumProcResourceKinds, {0, 0, 0}) {
  26   if (SM.hasExtraProcessorInfo()) {
  27     const MCExtraProcessorInfo &EPI = SM.getExtraProcessorInfo();
  28     LQResourceID = EPI.LoadQueueID;
  29     SQResourceID = EPI.StoreQueueID;
  30   }
  31 }
  32
  33 // FIXME: This implementation works under the assumption that load/store queue
  34 // entries are reserved at 'instruction dispatched' stage, and released at
  35 // 'instruction executed' stage. This currently matches the behavior of LSUnit.
  36 //
  37 // The current design minimizes the number of events generated by the
  38 // Dispatch/Execute stages, at the cost of doing extra bookkeeping in method
  39 // `onEvent`. However, it introduces a subtle dependency between this view and
  40 // how the LSUnit works.
  41 //
  42 // In future we should add a new "memory queue" event type, so that we stop
  43 // making assumptions on how LSUnit internally works (See PR39828).
  44 void SchedulerStatistics::onEvent(const HWInstructionEvent &Event) {
  45   if (Event.Type == HWInstructionEvent::Issued) {
  46     const Instruction &Inst = *Event.IR.getInstruction();
  47     NumIssued += Inst.getDesc().NumMicroOps;
  48   } else if (Event.Type == HWInstructionEvent::Dispatched) {
  49     const Instruction &Inst = *Event.IR.getInstruction();
  50     const unsigned Index = Event.IR.getSourceIndex();
  51     if (LQResourceID && Inst.getDesc().MayLoad &&
  52         MostRecentLoadDispatched != Index) {
  53       Usage[LQResourceID].SlotsInUse++;
  54       MostRecentLoadDispatched = Index;
  55     }
  56     if (SQResourceID && Inst.getDesc().MayStore &&
  57         MostRecentStoreDispatched != Index) {
  58       Usage[SQResourceID].SlotsInUse++;
  59       MostRecentStoreDispatched = Index;
  60     }
  61   } else if (Event.Type == HWInstructionEvent::Executed) {
  62     const Instruction &Inst = *Event.IR.getInstruction();
  63     if (LQResourceID && Inst.getDesc().MayLoad) {
  64       assert(Usage[LQResourceID].SlotsInUse);
  65       Usage[LQResourceID].SlotsInUse--;
  66     }
  67     if (SQResourceID && Inst.getDesc().MayStore) {
  68       assert(Usage[SQResourceID].SlotsInUse);
  69       Usage[SQResourceID].SlotsInUse--;
  70     }
  71   }
  72 }
  73
  74 void SchedulerStatistics::onReservedBuffers(const InstRef & /* unused */,
  75                                             ArrayRef<unsigned> Buffers) {
  76   for (const unsigned Buffer : Buffers) {
  77     if (Buffer == LQResourceID || Buffer == SQResourceID)
  78       continue;
  79     Usage[Buffer].SlotsInUse++;
  80   }
  81 }
  82
  83 void SchedulerStatistics::onReleasedBuffers(const InstRef & /* unused */,
  84                                             ArrayRef<unsigned> Buffers) {
  85   for (const unsigned Buffer : Buffers) {
  86     if (Buffer == LQResourceID || Buffer == SQResourceID)
  87       continue;
  88     Usage[Buffer].SlotsInUse--;
  89   }
  90 }
  91
  92 void SchedulerStatistics::updateHistograms() {
  93   for (BufferUsage &BU : Usage) {
  94     BU.CumulativeNumUsedSlots += BU.SlotsInUse;
  95     BU.MaxUsedSlots = std::max(BU.MaxUsedSlots, BU.SlotsInUse);
  96   }
  97
  98   IssueWidthPerCycle[NumIssued]++;
  99   NumIssued = 0;
 100 }
 101
 102 void SchedulerStatistics::printSchedulerStats(raw_ostream &OS) const {
 103   OS << "\n\nSchedulers - "
 104      << "number of cycles where we saw N micro opcodes issued:\n";
 105   OS << "[# issued], [# cycles]\n";
 106
 107   bool HasColors = OS.has_colors();
 108   const auto It =
 109       std::max_element(IssueWidthPerCycle.begin(), IssueWidthPerCycle.end());
 110   for (const std::pair<unsigned, unsigned> &Entry : IssueWidthPerCycle) {
 111     unsigned NumIssued = Entry.first;
 112     if (NumIssued == It->first && HasColors)
 113       OS.changeColor(raw_ostream::SAVEDCOLOR, true, false);
 114
 115     unsigned IPC = Entry.second;
 116     OS << " " << NumIssued << ",          " << IPC << "  ("
 117        << format("%.1f", ((double)IPC / NumCycles) * 100) << "%)\n";
 118     if (HasColors)
 119       OS.resetColor();
 120   }
 121 }
 122
 123 void SchedulerStatistics::printSchedulerUsage(raw_ostream &OS) const {
 124   assert(NumCycles && "Unexpected number of cycles!");
 125
 126   OS << "\nScheduler's queue usage:\n";
 127   if (all_of(Usage, [](const BufferUsage &BU) { return !BU.MaxUsedSlots; })) {
 128     OS << "No scheduler resources used.\n";
 129     return;
 130   }
 131
 132   OS << "[1] Resource name.\n"
 133      << "[2] Average number of used buffer entries.\n"
 134      << "[3] Maximum number of used buffer entries.\n"
 135      << "[4] Total number of buffer entries.\n\n"
 136      << " [1]            [2]        [3]        [4]\n";
 137
 138   formatted_raw_ostream FOS(OS);
 139   bool HasColors = FOS.has_colors();
 140   for (unsigned I = 0, E = SM.getNumProcResourceKinds(); I < E; ++I) {
 141     const MCProcResourceDesc &ProcResource = *SM.getProcResource(I);
 142     if (ProcResource.BufferSize <= 0)
 143       continue;
 144
 145     const BufferUsage &BU = Usage[I];
 146     double AvgUsage = (double)BU.CumulativeNumUsedSlots / NumCycles;
 147     double AlmostFullThreshold = (double)(ProcResource.BufferSize * 4) / 5;
 148     unsigned NormalizedAvg = floor((AvgUsage * 10) + 0.5) / 10;
 149     unsigned NormalizedThreshold = floor((AlmostFullThreshold * 10) + 0.5) / 10;
 150
 151     FOS << ProcResource.Name;
 152     FOS.PadToColumn(17);
 153     if (HasColors && NormalizedAvg >= NormalizedThreshold)
 154       FOS.changeColor(raw_ostream::YELLOW, true, false);
 155     FOS << NormalizedAvg;
 156     if (HasColors)
 157       FOS.resetColor();
 158     FOS.PadToColumn(28);
 159     if (HasColors &&
 160         BU.MaxUsedSlots == static_cast<unsigned>(ProcResource.BufferSize))
 161       FOS.changeColor(raw_ostream::RED, true, false);
 162     FOS << BU.MaxUsedSlots;
 163     if (HasColors)
 164       FOS.resetColor();
 165     FOS.PadToColumn(39);
 166     FOS << ProcResource.BufferSize << '\n';
 167   }
 168
 169   FOS.flush();
 170 }
 171
 172 void SchedulerStatistics::printView(raw_ostream &OS) const {
 173   printSchedulerStats(OS);
 174   printSchedulerUsage(OS);
 175 }
 176
 177 } // namespace mca
 178 } // namespace llvm