cpu_info.cc 4.4 KB
Newer Older
1
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
L
liaogang 已提交
2 3 4 5 6 7 8 9 10 11 12 13 14

Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at

    http://www.apache.org/licenses/LICENSE-2.0

Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */

Y
Yi Wang 已提交
15
#include "paddle/fluid/platform/cpu_info.h"
T
tensor-tang 已提交
16 17

#ifdef PADDLE_WITH_XBYAK
T
tensor-tang 已提交
18 19
#include "xbyak/xbyak.h"
#include "xbyak/xbyak_util.h"
T
tensor-tang 已提交
20
#endif
L
liaogang 已提交
21 22 23 24

#ifdef __APPLE__
#include <sys/sysctl.h>
#include <sys/types.h>
D
dzhwinter 已提交
25
#elif defined(_WIN32)
P
peizhilin 已提交
26
#define NOMINMAX  // msvc max/min macro conflict with std::min/max
D
dzhwinter 已提交
27
#include <windows.h>
L
liaogang 已提交
28 29
#else
#include <unistd.h>
D
dzhwinter 已提交
30
#endif  // _WIN32
L
liaogang 已提交
31

32
#include <algorithm>
L
liaogang 已提交
33 34
#include "gflags/gflags.h"

35 36 37
DECLARE_double(fraction_of_cpu_memory_to_use);
DECLARE_uint64(initial_cpu_memory_in_mb);
DECLARE_double(fraction_of_cuda_pinned_memory_to_use);
38

S
sneaxiy 已提交
39 40 41 42 43 44 45
// If use_pinned_memory is true, CPUAllocator calls mlock, which
// returns pinned and locked memory as staging areas for data exchange
// between host and device.  Allocates too much would reduce the amount
// of memory available to the system for paging.  So, by default, we
// should set false to use_pinned_memory.
DEFINE_bool(use_pinned_memory, true, "If set, allocate cpu pinned memory.");

L
liaogang 已提交
46 47 48
namespace paddle {
namespace platform {

S
sneaxiy 已提交
49
size_t CpuTotalPhysicalMemory() {
L
liaogang 已提交
50 51 52 53 54 55 56 57
#ifdef __APPLE__
  int mib[2];
  mib[0] = CTL_HW;
  mib[1] = HW_MEMSIZE;
  int64_t size = 0;
  size_t len = sizeof(size);
  if (sysctl(mib, 2, &size, &len, NULL, 0) == 0) return (size_t)size;
  return 0L;
D
dzhwinter 已提交
58 59 60 61 62
#elif defined(_WIN32)
  MEMORYSTATUSEX sMeminfo;
  sMeminfo.dwLength = sizeof(sMeminfo);
  GlobalMemoryStatusEx(&sMeminfo);
  return sMeminfo.ullTotalPhys;
L
liaogang 已提交
63
#else
L
liaogang 已提交
64 65
  int64_t pages = sysconf(_SC_PHYS_PAGES);
  int64_t page_size = sysconf(_SC_PAGE_SIZE);
L
liaogang 已提交
66 67 68 69
  return pages * page_size;
#endif
}

L
liaogang 已提交
70 71 72
size_t CpuMaxAllocSize() {
  // For distributed systems, it requires configuring and limiting
  // the fraction of memory to use.
73
  return FLAGS_fraction_of_cpu_memory_to_use * CpuTotalPhysicalMemory();
L
liaogang 已提交
74 75
}

L
liaogang 已提交
76
size_t CpuMinChunkSize() {
L
liaogang 已提交
77 78
  // Allow to allocate the minimum chunk size is 4 KB.
  return 1 << 12;
L
liaogang 已提交
79 80 81
}

size_t CpuMaxChunkSize() {
82 83 84 85 86
  // Allow to allocate the maximum chunk size is roughly 3% of CPU memory,
  // or the initial_cpu_memory_in_mb.
  return std::min(
      static_cast<size_t>(CpuMaxAllocSize() / 32),
      static_cast<size_t>(FLAGS_initial_cpu_memory_in_mb * 1 << 20));
L
liaogang 已提交
87 88
}

89 90 91 92 93 94 95 96 97 98 99 100
size_t CUDAPinnedMaxAllocSize() {
  // For distributed systems, it requires configuring and limiting
  // the fraction of memory to use.
  return FLAGS_fraction_of_cuda_pinned_memory_to_use * CpuTotalPhysicalMemory();
}

size_t CUDAPinnedMinChunkSize() {
  // Allow to allocate the minimum chunk size is 64 KB.
  return 1 << 16;
}

size_t CUDAPinnedMaxChunkSize() {
C
chengduoZH 已提交
101
  // Allow to allocate the maximum chunk size is roughly 1/256 of CUDA_PINNED
102 103 104 105
  // memory.
  return CUDAPinnedMaxAllocSize() / 256;
}

T
tensor-tang 已提交
106
#ifdef PADDLE_WITH_XBYAK
T
tensor-tang 已提交
107 108 109 110 111 112
static Xbyak::util::Cpu cpu;
bool MayIUse(const cpu_isa_t cpu_isa) {
  using namespace Xbyak::util;  // NOLINT
  switch (cpu_isa) {
    case sse42:
      return cpu.has(Cpu::tSSE42);
T
tensor-tang 已提交
113 114
    case avx:
      return cpu.has(Cpu::tAVX);
T
tensor-tang 已提交
115 116
    case avx2:
      return cpu.has(Cpu::tAVX2);
T
tensor-tang 已提交
117
    case avx512f:
T
tensor-tang 已提交
118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136
      return cpu.has(Cpu::tAVX512F);
    case avx512_core:
      return true && cpu.has(Cpu::tAVX512F) && cpu.has(Cpu::tAVX512BW) &&
             cpu.has(Cpu::tAVX512VL) && cpu.has(Cpu::tAVX512DQ);
    case avx512_core_vnni:
      return true && cpu.has(Cpu::tAVX512F) && cpu.has(Cpu::tAVX512BW) &&
             cpu.has(Cpu::tAVX512VL) && cpu.has(Cpu::tAVX512DQ) &&
             cpu.has(Cpu::tAVX512_VNNI);
    case avx512_mic:
      return true && cpu.has(Cpu::tAVX512F) && cpu.has(Cpu::tAVX512CD) &&
             cpu.has(Cpu::tAVX512ER) && cpu.has(Cpu::tAVX512PF);
    case avx512_mic_4ops:
      return true && MayIUse(avx512_mic) && cpu.has(Cpu::tAVX512_4FMAPS) &&
             cpu.has(Cpu::tAVX512_4VNNIW);
    case isa_any:
      return true;
  }
  return false;
}
T
tensor-tang 已提交
137 138 139 140 141 142 143 144 145
#else
bool MayIUse(const cpu_isa_t cpu_isa) {
  if (cpu_isa == isa_any) {
    return true;
  } else {
    return false;
  }
}
#endif
T
tensor-tang 已提交
146

L
liaogang 已提交
147 148
}  // namespace platform
}  // namespace paddle