1//===--------------------- Support.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 a few helper functions used by various pipeline
11/// components.
12///
13//===----------------------------------------------------------------------===//
14
15#include "llvm/MCA/Support.h"
16#include "llvm/MC/MCSchedule.h"
17#include <numeric>
18
19namespace llvm {
20namespace mca {
21
22#define DEBUG_TYPE "llvm-mca"
23
24ResourceCycles &ResourceCycles::operator+=(const ResourceCycles &RHS) {
25  if (Denominator == RHS.Denominator)
26    Numerator += RHS.Numerator;
27  else {
28    // Create a common denominator for LHS and RHS by calculating the least
29    // common multiple from the GCD.
30    unsigned GCD = std::gcd(Denominator, RHS.Denominator);
31    unsigned LCM = (Denominator * RHS.Denominator) / GCD;
32    unsigned LHSNumerator = Numerator * (LCM / Denominator);
33    unsigned RHSNumerator = RHS.Numerator * (LCM / RHS.Denominator);
34    Numerator = LHSNumerator + RHSNumerator;
35    Denominator = LCM;
36  }
37  return *this;
38}
39
40void computeProcResourceMasks(const MCSchedModel &SM,
41                              MutableArrayRef<uint64_t> Masks) {
42  unsigned ProcResourceID = 0;
43
44  assert(Masks.size() == SM.getNumProcResourceKinds() &&
45         "Invalid number of elements");
46  // Resource at index 0 is the 'InvalidUnit'. Set an invalid mask for it.
47  Masks[0] = 0;
48
49  // Create a unique bitmask for every processor resource unit.
50  for (unsigned I = 1, E = SM.getNumProcResourceKinds(); I < E; ++I) {
51    const MCProcResourceDesc &Desc = *SM.getProcResource(I);
52    if (Desc.SubUnitsIdxBegin)
53      continue;
54    Masks[I] = 1ULL << ProcResourceID;
55    ProcResourceID++;
56  }
57
58  // Create a unique bitmask for every processor resource group.
59  for (unsigned I = 1, E = SM.getNumProcResourceKinds(); I < E; ++I) {
60    const MCProcResourceDesc &Desc = *SM.getProcResource(I);
61    if (!Desc.SubUnitsIdxBegin)
62      continue;
63    Masks[I] = 1ULL << ProcResourceID;
64    for (unsigned U = 0; U < Desc.NumUnits; ++U) {
65      uint64_t OtherMask = Masks[Desc.SubUnitsIdxBegin[U]];
66      Masks[I] |= OtherMask;
67    }
68    ProcResourceID++;
69  }
70
71#ifndef NDEBUG
72  LLVM_DEBUG(dbgs() << "\nProcessor resource masks:"
73                    << "\n");
74  for (unsigned I = 0, E = SM.getNumProcResourceKinds(); I < E; ++I) {
75    const MCProcResourceDesc &Desc = *SM.getProcResource(I);
76    LLVM_DEBUG(dbgs() << '[' << format_decimal(I,2) << "] " << " - "
77                      << format_hex(Masks[I],16) << " - "
78                      << Desc.Name << '\n');
79  }
80#endif
81}
82
83double computeBlockRThroughput(const MCSchedModel &SM, unsigned DispatchWidth,
84                               unsigned NumMicroOps,
85                               ArrayRef<unsigned> ProcResourceUsage) {
86  // The block throughput is bounded from above by the hardware dispatch
87  // throughput. That is because the DispatchWidth is an upper bound on the
88  // number of opcodes that can be part of a single dispatch group.
89  double Max = static_cast<double>(NumMicroOps) / DispatchWidth;
90
91  // The block throughput is also limited by the amount of hardware parallelism.
92  // The number of available resource units affects the resource pressure
93  // distribution, as well as how many blocks can be executed every cycle.
94  for (unsigned I = 0, E = SM.getNumProcResourceKinds(); I < E; ++I) {
95    unsigned ResourceCycles = ProcResourceUsage[I];
96    if (!ResourceCycles)
97      continue;
98
99    const MCProcResourceDesc &MCDesc = *SM.getProcResource(I);
100    double Throughput = static_cast<double>(ResourceCycles) / MCDesc.NumUnits;
101    Max = std::max(Max, Throughput);
102  }
103
104  // The block reciprocal throughput is computed as the MAX of:
105  //  - (NumMicroOps / DispatchWidth)
106  //  - (NumUnits / ResourceCycles)   for every consumed processor resource.
107  return Max;
108}
109
110} // namespace mca
111} // namespace llvm
112