brintos

brintos / llvm-project-archived public Read only

0
0
Text · 7.5 KiB · 81be824 Raw
189 lines · c
1//===--------------------- TimelineView.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-exception6//7//===----------------------------------------------------------------------===//8/// \brief9///10/// This file implements a timeline view for the llvm-mca tool.11///12/// Class TimelineView observes events generated by the pipeline. For every13/// instruction executed by the pipeline, it stores information related to14/// state transition. It then plots that information in the form of a table15/// as reported by the example below:16///17/// Timeline view:18///     	          012345619/// Index	012345678920///21/// [0,0]	DeER .    .    ..	vmovshdup  %xmm0, %xmm122/// [0,1]	DeER .    .    ..	vpermilpd  $1, %xmm0, %xmm223/// [0,2]	.DeER.    .    ..	vpermilps  $231, %xmm0, %xmm524/// [0,3]	.DeeeER   .    ..	vaddss  %xmm1, %xmm0, %xmm325/// [0,4]	. D==eeeER.    ..	vaddss  %xmm3, %xmm2, %xmm426/// [0,5]	. D=====eeeER  ..	vaddss  %xmm4, %xmm5, %xmm627///28/// [1,0]	.  DeE------R  ..	vmovshdup  %xmm0, %xmm129/// [1,1]	.  DeE------R  ..	vpermilpd  $1, %xmm0, %xmm230/// [1,2]	.   DeE-----R  ..	vpermilps  $231, %xmm0, %xmm531/// [1,3]	.   D=eeeE--R  ..	vaddss  %xmm1, %xmm0, %xmm332/// [1,4]	.    D===eeeER ..	vaddss  %xmm3, %xmm2, %xmm433/// [1,5]	.    D======eeeER	vaddss  %xmm4, %xmm5, %xmm634///35/// There is an entry for every instruction in the input assembly sequence.36/// The first field is a pair of numbers obtained from the instruction index.37/// The first element of the pair is the iteration index, while the second38/// element of the pair is a sequence number (i.e. a position in the assembly39/// sequence).40/// The second field of the table is the actual timeline information; each41/// column is the information related to a specific cycle of execution.42/// The timeline of an instruction is described by a sequence of character43/// where each character represents the instruction state at a specific cycle.44///45/// Possible instruction states are:46///  D: Instruction Dispatched47///  e: Instruction Executing48///  E: Instruction Executed (write-back stage)49///  R: Instruction retired50///  =: Instruction waiting in the Scheduler's queue51///  -: Instruction executed, waiting to retire in order.52///53/// dots ('.') and empty spaces are cycles where the instruction is not54/// in-flight.55///56/// The last column is the assembly instruction associated to the entry.57///58/// Based on the timeline view information from the example, instruction 059/// at iteration 0 was dispatched at cycle 0, and was retired at cycle 3.60/// Instruction [0,1] was also dispatched at cycle 0, and it retired at61/// the same cycle than instruction [0,0].62/// Instruction [0,4] has been dispatched at cycle 2. However, it had to63/// wait for two cycles before being issued. That is because operands64/// became ready only at cycle 5.65///66/// This view helps further understanding bottlenecks and the impact of67/// resource pressure on the code.68///69/// To better understand why instructions had to wait for multiple cycles in70/// the scheduler's queue, class TimelineView also reports extra timing info71/// in another table named "Average Wait times" (see example below).72///73///74/// Average Wait times (based on the timeline view):75/// [0]: Executions76/// [1]: Average time spent waiting in a scheduler's queue77/// [2]: Average time spent waiting in a scheduler's queue while ready78/// [3]: Average time elapsed from WB until retire stage79///80///	[0]	[1]	[2]	[3]81/// 0.	 2	1.0	1.0	3.0	vmovshdup  %xmm0, %xmm182/// 1.	 2	1.0	1.0	3.0	vpermilpd  $1, %xmm0, %xmm283/// 2.	 2	1.0	1.0	2.5	vpermilps  $231, %xmm0, %xmm584/// 3.	 2	1.5	0.5	1.0	vaddss  %xmm1, %xmm0, %xmm385/// 4.	 2	3.5	0.0	0.0	vaddss  %xmm3, %xmm2, %xmm486/// 5.	 2	6.5	0.0	0.0	vaddss  %xmm4, %xmm5, %xmm687///      2	2.4	0.6	1.6     <total>88///89/// By comparing column [2] with column [1], we get an idea about how many90/// cycles were spent in the scheduler's queue due to data dependencies.91///92/// In this example, instruction 5 spent an average of ~6 cycles in the93/// scheduler's queue. As soon as operands became ready, the instruction94/// was immediately issued to the pipeline(s).95/// That is expected because instruction 5 cannot transition to the "ready"96/// state until %xmm4 is written by instruction 4.97///98//===----------------------------------------------------------------------===//99 100#ifndef LLVM_TOOLS_LLVM_MCA_TIMELINEVIEW_H101#define LLVM_TOOLS_LLVM_MCA_TIMELINEVIEW_H102 103#include "Views/InstructionView.h"104#include "llvm/ADT/ArrayRef.h"105#include "llvm/MC/MCInst.h"106#include "llvm/MC/MCInstPrinter.h"107#include "llvm/MC/MCSubtargetInfo.h"108#include "llvm/Support/FormattedStream.h"109#include "llvm/Support/JSON.h"110#include "llvm/Support/raw_ostream.h"111 112namespace llvm {113namespace mca {114 115/// This class listens to instruction state transition events116/// in order to construct a timeline information.117///118/// For every instruction executed by the Pipeline, this class constructs119/// a TimelineViewEntry object. TimelineViewEntry objects are then used120/// to print the timeline information, as well as the "average wait times"121/// for every instruction in the input assembly sequence.122class TimelineView : public InstructionView {123  unsigned CurrentCycle;124  unsigned MaxCycle;125  unsigned LastCycle;126 127  struct TimelineViewEntry {128    int CycleDispatched; // A negative value is an "invalid cycle".129    unsigned CycleReady;130    unsigned CycleIssued;131    unsigned CycleExecuted;132    unsigned CycleRetired;133  };134  std::vector<TimelineViewEntry> Timeline;135 136  struct WaitTimeEntry {137    unsigned CyclesSpentInSchedulerQueue;138    unsigned CyclesSpentInSQWhileReady;139    unsigned CyclesSpentAfterWBAndBeforeRetire;140  };141  std::vector<WaitTimeEntry> WaitTime;142 143  // This field is used to map instructions to buffered resources.144  // Elements of this vector are <resourceID, BufferSizer> pairs.145  std::vector<std::pair<unsigned, int>> UsedBuffer;146 147  void printTimelineViewEntry(llvm::formatted_raw_ostream &OS,148                              const TimelineViewEntry &E, unsigned Iteration,149                              unsigned SourceIndex) const;150  void printWaitTimeEntry(llvm::formatted_raw_ostream &OS,151                          const WaitTimeEntry &E, unsigned Index,152                          unsigned Executions) const;153 154  // Display characters for the TimelineView report output.155  struct DisplayChar {156    static const char Dispatched = 'D';157    static const char Executed = 'E';158    static const char Retired = 'R';159    static const char Waiting = '='; // Instruction is waiting in the scheduler.160    static const char Executing = 'e';161    static const char RetireLag = '-'; // The instruction is waiting to retire.162  };163 164public:165  TimelineView(const llvm::MCSubtargetInfo &sti, llvm::MCInstPrinter &Printer,166               llvm::ArrayRef<llvm::MCInst> S, unsigned Iterations,167               unsigned Cycles);168 169  // Event handlers.170  void onCycleEnd() override { ++CurrentCycle; }171  void onEvent(const HWInstructionEvent &Event) override;172  void onReservedBuffers(const InstRef &IR,173                         llvm::ArrayRef<unsigned> Buffers) override;174 175  // print functionalities.176  void printTimeline(llvm::raw_ostream &OS) const;177  void printAverageWaitTimes(llvm::raw_ostream &OS) const;178  void printView(llvm::raw_ostream &OS) const override {179    printTimeline(OS);180    printAverageWaitTimes(OS);181  }182  StringRef getNameAsString() const override { return "TimelineView"; }183  json::Value toJSON() const override;184};185} // namespace mca186} // namespace llvm187 188#endif189