tools/llvm-mca/Views/SummaryView.h

   1 //===--------------------- SummaryView.h ---------------------*- 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 summary view.
  11 ///
  12 /// The goal of the summary view is to give a very quick overview of the
  13 /// performance throughput. Below is an example of summary view:
  14 ///
  15 ///
  16 /// Iterations:        300
  17 /// Instructions:      900
  18 /// Total Cycles:      610
  19 /// Dispatch Width:    2
  20 /// IPC:               1.48
  21 /// Block RThroughput: 2.0
  22 ///
  23 /// The summary view collects a few performance numbers. The two main
  24 /// performance indicators are 'Total Cycles' and IPC (Instructions Per Cycle).
  25 ///
  26 //===----------------------------------------------------------------------===//
  27
  28 #ifndef LLVM_TOOLS_LLVM_MCA_SUMMARYVIEW_H
  29 #define LLVM_TOOLS_LLVM_MCA_SUMMARYVIEW_H
  30
  31 #include "Views/View.h"
  32 #include "llvm/ADT/DenseMap.h"
  33 #include "llvm/MC/MCSchedule.h"
  34 #include "llvm/Support/raw_ostream.h"
  35
  36 namespace llvm {
  37 namespace mca {
  38
  39 /// A view that collects and prints a few performance numbers.
  40 class SummaryView : public View {
  41   const llvm::MCSchedModel &SM;
  42   llvm::ArrayRef<llvm::MCInst> Source;
  43   const unsigned DispatchWidth;
  44   unsigned LastInstructionIdx;
  45   unsigned TotalCycles;
  46   // The total number of micro opcodes contributed by a block of instructions.
  47   unsigned NumMicroOps;
  48
  49   // For each processor resource, this vector stores the cumulative number of
  50   // resource cycles consumed by the analyzed code block.
  51   llvm::SmallVector<unsigned, 8> ProcResourceUsage;
  52
  53   // Each processor resource is associated with a so-called processor resource
  54   // mask. This vector allows to correlate processor resource IDs with processor
  55   // resource masks. There is exactly one element per each processor resource
  56   // declared by the scheduling model.
  57   llvm::SmallVector<uint64_t, 8> ProcResourceMasks;
  58
  59   // Used to map resource indices to actual processor resource IDs.
  60   llvm::SmallVector<unsigned, 8> ResIdx2ProcResID;
  61
  62   // Compute the reciprocal throughput for the analyzed code block.
  63   // The reciprocal block throughput is computed as the MAX between:
  64   //   - NumMicroOps / DispatchWidth
  65   //   - Total Resource Cycles / #Units   (for every resource consumed).
  66   double getBlockRThroughput() const;
  67
  68 public:
  69   SummaryView(const llvm::MCSchedModel &Model, llvm::ArrayRef<llvm::MCInst> S,
  70               unsigned Width);
  71
  72   void onCycleEnd() override { ++TotalCycles; }
  73   void onEvent(const HWInstructionEvent &Event) override;
  74   void printView(llvm::raw_ostream &OS) const override;
  75 };
  76
  77 } // namespace mca
  78 } // namespace llvm
  79
  80 #endif