1  // Copyright 2013 Dolphin Emulator Project / 2015 Citra Emulator Project
  2  // Licensed under GPLv2 or any later version
  3  // Refer to the license.txt file included.
  4  
  5  #include "common/arch.h"
  6  #if CITRA_ARCH(x86_64)
  7  
  8  #include <cstring>
  9  #include "common/common_types.h"
 10  #include "common/x64/cpu_detect.h"
 11  
 12  #ifdef _MSC_VER
 13  #include <intrin.h>
 14  #else
 15  
 16  #if defined(__DragonFly__) || defined(__FreeBSD__)
 17  // clang-format off
 18  #include <sys/types.h>
 19  #include <machine/cpufunc.h>
 20  // clang-format on
 21  #endif
 22  
 23  static inline void __cpuidex(int info[4], int function_id, int subfunction_id) {
 24  #if defined(__DragonFly__) || defined(__FreeBSD__)
 25      // Despite the name, this is just do_cpuid() with ECX as second input.
 26      cpuid_count((u_int)function_id, (u_int)subfunction_id, (u_int*)info);
 27  #else
 28      info[0] = function_id;    // eax
 29      info[2] = subfunction_id; // ecx
 30      __asm__("cpuid"
 31              : "=a"(info[0]), "=b"(info[1]), "=c"(info[2]), "=d"(info[3])
 32              : "a"(function_id), "c"(subfunction_id));
 33  #endif
 34  }
 35  
 36  static inline void __cpuid(int info[4], int function_id) {
 37      return __cpuidex(info, function_id, 0);
 38  }
 39  
 40  #define _XCR_XFEATURE_ENABLED_MASK 0
 41  static inline u64 _xgetbv(u32 index) {
 42      u32 eax, edx;
 43      __asm__ __volatile__("xgetbv" : "=a"(eax), "=d"(edx) : "c"(index));
 44      return ((u64)edx << 32) | eax;
 45  }
 46  
 47  #endif // _MSC_VER
 48  
 49  namespace Common {
 50  
 51  // Detects the various CPU features
 52  static CPUCaps Detect() {
 53      CPUCaps caps = {};
 54  
 55      // Assumes the CPU supports the CPUID instruction. Those that don't would likely not support
 56      // Citra at all anyway
 57  
 58      int cpu_id[4];
 59      std::memset(caps.brand_string, 0, sizeof(caps.brand_string));
 60  
 61      // Detect CPU's CPUID capabilities and grab CPU string
 62      __cpuid(cpu_id, 0x00000000);
 63      u32 max_std_fn = cpu_id[0]; // EAX
 64  
 65      std::memcpy(&caps.brand_string[0], &cpu_id[1], sizeof(int));
 66      std::memcpy(&caps.brand_string[4], &cpu_id[3], sizeof(int));
 67      std::memcpy(&caps.brand_string[8], &cpu_id[2], sizeof(int));
 68  
 69      __cpuid(cpu_id, 0x80000000);
 70  
 71      u32 max_ex_fn = cpu_id[0];
 72  
 73      // Set reasonable default brand string even if brand string not available
 74      strcpy(caps.cpu_string, caps.brand_string);
 75  
 76      // Detect family and other miscellaneous features
 77      if (max_std_fn >= 1) {
 78          __cpuid(cpu_id, 0x00000001);
 79  
 80          if ((cpu_id[3] >> 25) & 1)
 81              caps.sse = true;
 82          if ((cpu_id[3] >> 26) & 1)
 83              caps.sse2 = true;
 84          if ((cpu_id[2]) & 1)
 85              caps.sse3 = true;
 86          if ((cpu_id[2] >> 9) & 1)
 87              caps.ssse3 = true;
 88          if ((cpu_id[2] >> 19) & 1)
 89              caps.sse4_1 = true;
 90          if ((cpu_id[2] >> 20) & 1)
 91              caps.sse4_2 = true;
 92          if ((cpu_id[2] >> 25) & 1)
 93              caps.aes = true;
 94  
 95          // AVX support requires 3 separate checks:
 96          //  - Is the AVX bit set in CPUID?
 97          //  - Is the XSAVE bit set in CPUID?
 98          //  - XGETBV result has the XCR bit set.
 99          if (((cpu_id[2] >> 28) & 1) && ((cpu_id[2] >> 27) & 1)) {
100              if ((_xgetbv(_XCR_XFEATURE_ENABLED_MASK) & 0x6) == 0x6) {
101                  caps.avx = true;
102                  if ((cpu_id[2] >> 12) & 1)
103                      caps.fma = true;
104              }
105          }
106  
107          if (max_std_fn >= 7) {
108              __cpuidex(cpu_id, 0x00000007, 0x00000000);
109              // Can't enable AVX2 unless the XSAVE/XGETBV checks above passed
110              if ((cpu_id[1] >> 5) & 1)
111                  caps.avx2 = caps.avx;
112              if ((cpu_id[1] >> 3) & 1)
113                  caps.bmi1 = true;
114              if ((cpu_id[1] >> 8) & 1)
115                  caps.bmi2 = true;
116              // Checks for AVX512F, AVX512CD, AVX512VL, AVX512DQ, AVX512BW (Intel Skylake-X/SP)
117              if ((cpu_id[1] >> 16) & 1 && (cpu_id[1] >> 28) & 1 && (cpu_id[1] >> 31) & 1 &&
118                  (cpu_id[1] >> 17) & 1 && (cpu_id[1] >> 30) & 1) {
119                  caps.avx512 = caps.avx2;
120              }
121          }
122      }
123  
124      if (max_ex_fn >= 0x80000004) {
125          // Extract CPU model string
126          __cpuid(cpu_id, 0x80000002);
127          std::memcpy(caps.cpu_string, cpu_id, sizeof(cpu_id));
128          __cpuid(cpu_id, 0x80000003);
129          std::memcpy(caps.cpu_string + 16, cpu_id, sizeof(cpu_id));
130          __cpuid(cpu_id, 0x80000004);
131          std::memcpy(caps.cpu_string + 32, cpu_id, sizeof(cpu_id));
132      }
133  
134      if (max_ex_fn >= 0x80000001) {
135          // Check for more features
136          __cpuid(cpu_id, 0x80000001);
137          if ((cpu_id[2] >> 16) & 1)
138              caps.fma4 = true;
139      }
140  
141      return caps;
142  }
143  
144  const CPUCaps& GetCPUCaps() {
145      static CPUCaps caps = Detect();
146      return caps;
147  }
148  
149  } // namespace Common
150  
151  #endif // CITRA_ARCH(x86_64)