1 //===--------------------- SummaryView.cpp -------------------*- C++ -*-===//
2 //
3 // The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 /// \file
10 ///
11 /// This file implements the functionalities used by the SummaryView to print
12 /// the report information.
13 ///
14 //===----------------------------------------------------------------------===//
15
16 #include "SummaryView.h"
17 #include "Support.h"
18 #include "llvm/ADT/SmallVector.h"
19 #include "llvm/Support/Format.h"
20
21 namespace mca {
22
23 #define DEBUG_TYPE "llvm-mca"
24
25 using namespace llvm;
26
SummaryView(const llvm::MCSchedModel & Model,const SourceMgr & S,unsigned Width)27 SummaryView::SummaryView(const llvm::MCSchedModel &Model, const SourceMgr &S,
28 unsigned Width)
29 : SM(Model), Source(S), DispatchWidth(Width), TotalCycles(0),
30 NumMicroOps(0), ProcResourceUsage(Model.getNumProcResourceKinds(), 0),
31 ProcResourceMasks(Model.getNumProcResourceKinds(), 0) {
32 computeProcResourceMasks(SM, ProcResourceMasks);
33 }
34
onEvent(const HWInstructionEvent & Event)35 void SummaryView::onEvent(const HWInstructionEvent &Event) {
36 // We are only interested in the "instruction dispatched" events generated by
37 // the dispatch stage for instructions that are part of iteration #0.
38 if (Event.Type != HWInstructionEvent::Dispatched)
39 return;
40
41 if (Event.IR.getSourceIndex() >= Source.size())
42 return;
43
44 // Update the cumulative number of resource cycles based on the processor
45 // resource usage information available from the instruction descriptor. We
46 // need to compute the cumulative number of resource cycles for every
47 // processor resource which is consumed by an instruction of the block.
48 const Instruction &Inst = *Event.IR.getInstruction();
49 const InstrDesc &Desc = Inst.getDesc();
50 NumMicroOps += Desc.NumMicroOps;
51 for (const std::pair<uint64_t, const ResourceUsage> &RU : Desc.Resources) {
52 if (RU.second.size()) {
53 const auto It = find(ProcResourceMasks, RU.first);
54 assert(It != ProcResourceMasks.end() &&
55 "Invalid processor resource mask!");
56 ProcResourceUsage[std::distance(ProcResourceMasks.begin(), It)] +=
57 RU.second.size();
58 }
59 }
60 }
61
printView(raw_ostream & OS) const62 void SummaryView::printView(raw_ostream &OS) const {
63 unsigned Iterations = Source.getNumIterations();
64 unsigned Instructions = Source.size();
65 unsigned TotalInstructions = Instructions * Iterations;
66 double IPC = (double)TotalInstructions / TotalCycles;
67 double BlockRThroughput = computeBlockRThroughput(
68 SM, DispatchWidth, NumMicroOps, ProcResourceUsage);
69
70 std::string Buffer;
71 raw_string_ostream TempStream(Buffer);
72 TempStream << "Iterations: " << Iterations;
73 TempStream << "\nInstructions: " << TotalInstructions;
74 TempStream << "\nTotal Cycles: " << TotalCycles;
75 TempStream << "\nDispatch Width: " << DispatchWidth;
76 TempStream << "\nIPC: " << format("%.2f", IPC);
77
78 // Round to the block reciprocal throughput to the nearest tenth.
79 TempStream << "\nBlock RThroughput: "
80 << format("%.1f", floor((BlockRThroughput * 10) + 0.5) / 10)
81 << '\n';
82 TempStream.flush();
83 OS << Buffer;
84 }
85 } // namespace mca.
86