dolphin/Source/Core/Common/x64CPUDetect.cpp

Ignoring revisions in .git-blame-ignore-revs. Click here to bypass and see the normal blame view.

258 lines
6.1 KiB
C++
Raw Normal View History

// Copyright 2008 Dolphin Emulator Project
// SPDX-License-Identifier: GPL-2.0-or-later
2021-12-10 03:22:16 +01:00
#include "Common/CPUDetect.h"
#include <cstring>
#include <string>
#include <thread>
#include "Common/CommonTypes.h"
2015-02-23 20:40:05 +01:00
#include "Common/Intrinsics.h"
2015-02-23 20:40:05 +01:00
#ifndef _WIN32
#ifdef __FreeBSD__
2017-01-24 02:44:10 +01:00
#include <unistd.h>
#include <machine/cpufunc.h>
#include <sys/types.h>
#endif
static inline void __cpuidex(int info[4], int function_id, int subfunction_id)
{
#ifdef __FreeBSD__
// Despite the name, this is just do_cpuid() with ECX as second input.
cpuid_count((u_int)function_id, (u_int)subfunction_id, (u_int*)info);
#else
info[0] = function_id; // eax
info[2] = subfunction_id; // ecx
__asm__("cpuid"
: "=a"(info[0]), "=b"(info[1]), "=c"(info[2]), "=d"(info[3])
: "a"(function_id), "c"(subfunction_id));
#endif
}
static inline void __cpuid(int info[4], int function_id)
{
return __cpuidex(info, function_id, 0);
}
#endif // ifndef _WIN32
#ifdef _WIN32
static u64 xgetbv(u32 index)
{
return _xgetbv(index);
}
constexpr u32 XCR_XFEATURE_ENABLED_MASK = _XCR_XFEATURE_ENABLED_MASK;
#else
static u64 xgetbv(u32 index)
{
u32 eax, edx;
__asm__ __volatile__("xgetbv" : "=a"(eax), "=d"(edx) : "c"(index));
return ((u64)edx << 32) | eax;
}
constexpr u32 XCR_XFEATURE_ENABLED_MASK = 0;
#endif // ifdef _WIN32
CPUInfo cpu_info;
CPUInfo::CPUInfo()
{
Detect();
}
// Detects the various CPU features
void CPUInfo::Detect()
{
#ifdef _M_X86_64
Mode64bit = true;
OS64bit = true;
#endif
num_cores = 1;
// Set obvious defaults, for extra safety
if (Mode64bit)
{
bSSE = true;
bSSE2 = true;
bLongMode = true;
}
// Assume CPU supports the CPUID instruction. Those that don't can barely
// boot modern OS:es anyway.
int cpu_id[4];
// Detect CPU's CPUID capabilities, and grab CPU string
__cpuid(cpu_id, 0x00000000);
u32 max_std_fn = cpu_id[0]; // EAX
2015-09-15 18:13:13 +02:00
std::memcpy(&brand_string[0], &cpu_id[1], sizeof(int));
std::memcpy(&brand_string[4], &cpu_id[3], sizeof(int));
std::memcpy(&brand_string[8], &cpu_id[2], sizeof(int));
__cpuid(cpu_id, 0x80000000);
u32 max_ex_fn = cpu_id[0];
if (!strcmp(brand_string, "GenuineIntel"))
vendor = CPUVendor::Intel;
else if (!strcmp(brand_string, "AuthenticAMD"))
vendor = CPUVendor::AMD;
else
vendor = CPUVendor::Other;
// Set reasonable default brand string even if brand string not available.
strcpy(cpu_string, brand_string);
// Detect family and other misc stuff.
bool ht = false;
HTT = ht;
if (max_std_fn >= 1)
{
__cpuid(cpu_id, 0x00000001);
int family = ((cpu_id[0] >> 8) & 0xf) + ((cpu_id[0] >> 20) & 0xff);
int model = ((cpu_id[0] >> 4) & 0xf) + ((cpu_id[0] >> 12) & 0xf0);
// Detect people unfortunate enough to be running Dolphin on an Atom
if (family == 6 &&
(model == 0x1C || model == 0x26 || model == 0x27 || model == 0x35 || model == 0x36 ||
model == 0x37 || model == 0x4A || model == 0x4D || model == 0x5A || model == 0x5D))
bAtom = true;
// Detect AMD Zen1, Zen1+ and Zen2
if (family == 23)
bZen1p2 = true;
ht = (cpu_id[3] >> 28) & 1;
// AMD CPUs before Zen faked this flag and didn't actually
// implement simultaneous multithreading (SMT; Intel calls it HTT)
// but rather some weird middle-ground between 1-2 cores
HTT = ht && (vendor == CPUVendor::Intel || family >= 23);
if ((cpu_id[3] >> 25) & 1)
bSSE = true;
if ((cpu_id[3] >> 26) & 1)
bSSE2 = true;
if ((cpu_id[2]) & 1)
bSSE3 = true;
if ((cpu_id[2] >> 9) & 1)
bSSSE3 = true;
if ((cpu_id[2] >> 19) & 1)
bSSE4_1 = true;
if ((cpu_id[2] >> 20) & 1)
bSSE4_2 = true;
if ((cpu_id[2] >> 22) & 1)
bMOVBE = true;
if ((cpu_id[2] >> 25) & 1)
bAES = true;
if ((cpu_id[3] >> 24) & 1)
{
// We can use FXSAVE.
bFXSR = true;
}
// AVX support requires 3 separate checks:
// - Is the AVX bit set in CPUID?
// - Is the XSAVE bit set in CPUID?
// - XGETBV result has the XCR bit set.
if (((cpu_id[2] >> 28) & 1) && ((cpu_id[2] >> 27) & 1))
{
if ((xgetbv(XCR_XFEATURE_ENABLED_MASK) & 0x6) == 0x6)
2013-11-13 04:46:34 +01:00
{
bAVX = true;
2013-11-13 04:46:34 +01:00
if ((cpu_id[2] >> 12) & 1)
bFMA = true;
}
}
if (max_std_fn >= 7)
{
__cpuidex(cpu_id, 0x00000007, 0x00000000);
// careful; we can't enable AVX2 unless the XSAVE/XGETBV checks above passed
if ((cpu_id[1] >> 5) & 1)
bAVX2 = bAVX;
if ((cpu_id[1] >> 3) & 1)
bBMI1 = true;
if ((cpu_id[1] >> 8) & 1)
bBMI2 = true;
}
}
bFlushToZero = bSSE;
bFastBMI2 = bBMI2 && !bZen1p2;
if (max_ex_fn >= 0x80000004)
{
// Extract CPU model string
__cpuid(cpu_id, 0x80000002);
memcpy(cpu_string, cpu_id, sizeof(cpu_id));
__cpuid(cpu_id, 0x80000003);
memcpy(cpu_string + 16, cpu_id, sizeof(cpu_id));
__cpuid(cpu_id, 0x80000004);
memcpy(cpu_string + 32, cpu_id, sizeof(cpu_id));
}
if (max_ex_fn >= 0x80000001)
{
// Check for more features.
__cpuid(cpu_id, 0x80000001);
if (cpu_id[2] & 1)
bLAHFSAHF64 = true;
if ((cpu_id[2] >> 5) & 1)
bLZCNT = true;
2015-05-17 09:20:29 +02:00
if ((cpu_id[2] >> 16) & 1)
bFMA4 = true;
if ((cpu_id[3] >> 29) & 1)
bLongMode = true;
}
// this should be much more reliable and easier
// than trying to get the number of cores out of the CPUID data
// ourselves
num_cores = std::max(std::thread::hardware_concurrency(), 1u);
}
// Turn the CPU info into a string we can show
std::string CPUInfo::Summarize()
{
std::string sum(cpu_string);
sum += " (";
sum += brand_string;
sum += ")";
if (bSSE)
sum += ", SSE";
if (bSSE2)
{
sum += ", SSE2";
if (!bFlushToZero)
sum += " (but not DAZ!)";
}
if (bSSE3)
sum += ", SSE3";
if (bSSSE3)
sum += ", SSSE3";
if (bSSE4_1)
sum += ", SSE4.1";
if (bSSE4_2)
sum += ", SSE4.2";
if (HTT)
sum += ", HTT";
if (bAVX)
sum += ", AVX";
if (bAVX2)
sum += ", AVX2";
if (bBMI1)
sum += ", BMI1";
if (bBMI2)
sum += ", BMI2";
2013-11-13 04:46:34 +01:00
if (bFMA)
sum += ", FMA";
if (bAES)
sum += ", AES";
if (bMOVBE)
sum += ", MOVBE";
if (bLongMode)
sum += ", 64-bit support";
return sum;
}