core/cpu_caps.cpp

   1
   2 #include "config.h"
   3
   4 #include "cpu_caps.h"
   5
   6 #if defined(_WIN32) && (defined(_M_ARM) || defined(_M_ARM64))
   7 #define WIN32_LEAN_AND_MEAN
   8 #include <windows.h>
   9 #ifndef PF_ARM_NEON_INSTRUCTIONS_AVAILABLE
  10 #define PF_ARM_NEON_INSTRUCTIONS_AVAILABLE 19
  11 #endif
  12 #endif
  13
  14 #if defined(HAVE_CPUID_H)
  15 #include <cpuid.h>
  16 #elif defined(HAVE_INTRIN_H)
  17 #include <intrin.h>
  18 #endif
  19
  20 #include <array>
  21 #include <cctype>
  22 #include <string>
  23
  24
  25 int CPUCapFlags{0};
  26
  27 namespace {
  28
  29 #if defined(HAVE_GCC_GET_CPUID) \
  30     && (defined(__i386__) || defined(__x86_64__) || defined(_M_IX86) || defined(_M_X64))
  31 using reg_type = unsigned int;
  32 inline std::array<reg_type,4> get_cpuid(unsigned int f)
  33 {
  34     std::array<reg_type,4> ret{};
  35     __get_cpuid(f, ret.data(), &ret[1], &ret[2], &ret[3]);
  36     return ret;
  37 }
  38 #define CAN_GET_CPUID
  39 #elif defined(HAVE_CPUID_INTRINSIC) \
  40     && (defined(__i386__) || defined(__x86_64__) || defined(_M_IX86) || defined(_M_X64))
  41 using reg_type = int;
  42 inline std::array<reg_type,4> get_cpuid(unsigned int f)
  43 {
  44     std::array<reg_type,4> ret{};
  45     (__cpuid)(ret.data(), f);
  46     return ret;
  47 }
  48 #define CAN_GET_CPUID
  49 #endif
  50
  51 } // namespace
  52
  53 al::optional<CPUInfo> GetCPUInfo()
  54 {
  55     CPUInfo ret;
  56
  57 #ifdef CAN_GET_CPUID
  58     auto cpuregs = get_cpuid(0);
  59     if(cpuregs[0] == 0)
  60         return al::nullopt;
  61
  62     const reg_type maxfunc{cpuregs[0]};
  63
  64     cpuregs = get_cpuid(0x80000000);
  65     const reg_type maxextfunc{cpuregs[0]};
  66
  67     ret.mVendor.append(reinterpret_cast<char*>(&cpuregs[1]), 4);
  68     ret.mVendor.append(reinterpret_cast<char*>(&cpuregs[3]), 4);
  69     ret.mVendor.append(reinterpret_cast<char*>(&cpuregs[2]), 4);
  70     auto iter_end = std::remove(ret.mVendor.begin(), ret.mVendor.end(), '\0');
  71     iter_end = std::unique(ret.mVendor.begin(), iter_end,
  72         [](auto&& c0, auto&& c1) { return std::isspace(c0) && std::isspace(c1); });
  73     ret.mVendor.erase(iter_end, ret.mVendor.end());
  74     if(!ret.mVendor.empty() && std::isspace(ret.mVendor.back()))
  75         ret.mVendor.pop_back();
  76     if(!ret.mVendor.empty() && std::isspace(ret.mVendor.front()))
  77         ret.mVendor.erase(ret.mVendor.begin());
  78
  79     if(maxextfunc >= 0x80000004)
  80     {
  81         cpuregs = get_cpuid(0x80000002);
  82         ret.mName.append(reinterpret_cast<char*>(cpuregs.data()), 16);
  83         cpuregs = get_cpuid(0x80000003);
  84         ret.mName.append(reinterpret_cast<char*>(cpuregs.data()), 16);
  85         cpuregs = get_cpuid(0x80000004);
  86         ret.mName.append(reinterpret_cast<char*>(cpuregs.data()), 16);
  87         iter_end = std::remove(ret.mName.begin(), ret.mName.end(), '\0');
  88         iter_end = std::unique(ret.mName.begin(), iter_end,
  89             [](auto&& c0, auto&& c1) { return std::isspace(c0) && std::isspace(c1); });
  90         ret.mName.erase(iter_end, ret.mName.end());
  91         if(!ret.mName.empty() && std::isspace(ret.mName.back()))
  92             ret.mName.pop_back();
  93         if(!ret.mName.empty() && std::isspace(ret.mName.front()))
  94             ret.mName.erase(ret.mName.begin());
  95     }
  96
  97     if(maxfunc >= 1)
  98     {
  99         cpuregs = get_cpuid(1);
 100         if((cpuregs[3]&(1<<25)))
 101             ret.mCaps |= CPU_CAP_SSE;
 102         if((ret.mCaps&CPU_CAP_SSE) && (cpuregs[3]&(1<<26)))
 103             ret.mCaps |= CPU_CAP_SSE2;
 104         if((ret.mCaps&CPU_CAP_SSE2) && (cpuregs[2]&(1<<0)))
 105             ret.mCaps |= CPU_CAP_SSE3;
 106         if((ret.mCaps&CPU_CAP_SSE3) && (cpuregs[2]&(1<<19)))
 107             ret.mCaps |= CPU_CAP_SSE4_1;
 108     }
 109
 110 #else
 111
 112     /* Assume support for whatever's supported if we can't check for it */
 113 #if defined(HAVE_SSE4_1)
 114 #warning "Assuming SSE 4.1 run-time support!"
 115     ret.mCaps |= CPU_CAP_SSE | CPU_CAP_SSE2 | CPU_CAP_SSE3 | CPU_CAP_SSE4_1;
 116 #elif defined(HAVE_SSE3)
 117 #warning "Assuming SSE 3 run-time support!"
 118     ret.mCaps |= CPU_CAP_SSE | CPU_CAP_SSE2 | CPU_CAP_SSE3;
 119 #elif defined(HAVE_SSE2)
 120 #warning "Assuming SSE 2 run-time support!"
 121     ret.mCaps |= CPU_CAP_SSE | CPU_CAP_SSE2;
 122 #elif defined(HAVE_SSE)
 123 #warning "Assuming SSE run-time support!"
 124     ret.mCaps |= CPU_CAP_SSE;
 125 #endif
 126 #endif /* CAN_GET_CPUID */
 127
 128 #ifdef HAVE_NEON
 129 #ifdef __ARM_NEON
 130     ret.mCaps |= CPU_CAP_NEON;
 131 #elif defined(_WIN32) && (defined(_M_ARM) || defined(_M_ARM64))
 132     if(IsProcessorFeaturePresent(PF_ARM_NEON_INSTRUCTIONS_AVAILABLE))
 133         ret.mCaps |= CPU_CAP_NEON;
 134 #else
 135 #warning "Assuming NEON run-time support!"
 136     ret.mCaps |= CPU_CAP_NEON;
 137 #endif
 138 #endif
 139
 140     return ret;
 141 }