mctaylors/suyu
1
0
Fork 0
forked from suyu/suyu
Code Pull requests Activity

Common: Cleanup CPU capability detection code.

Browse source
This commit is contained in:
bunnei 2015-08-12 17:42:13 -04:00
parent a1942238f5
commit 0ee00861f6
5 changed files with 144 additions and 201 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

214
src/common/x64/cpu_detect.cpp
View file

@ -1,23 +1,25 @@
// Copyright 2008 Dolphin Emulator Project
// Licensed under GPLv2+
// Copyright 2013 Dolphin Emulator Project / 2015 Citra Emulator Project
// Licensed under GPLv2 or any later version
// Refer to the license.txt file included.
#include <cstring>
#include <string>
#include <thread>
#include "common/common_types.h"
#include "cpu_detect.h"
#ifndef _WIN32
namespace Common {
#ifndef _MSC_VER
#ifdef __FreeBSD__
#include <sys/types.h>
#include <machine/cpufunc.h>
#endif
static inline void __cpuidex(int info[4], int function_id, int subfunction_id)
{
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);
@ -36,96 +38,67 @@ static inline void __cpuidex(int info[4], int function_id, int subfunction_id)
#endif
}
static inline void __cpuid(int info[4], int function_id)
{
static inline void __cpuid(int info[4], int function_id) {
return __cpuidex(info, function_id, 0);
}
#define _XCR_XFEATURE_ENABLED_MASK 0
static u64 _xgetbv(u32 index)
{
static u64 _xgetbv(u32 index) {
u32 eax, edx;
__asm__ __volatile__("xgetbv" : "=a"(eax), "=d"(edx) : "c"(index));
return ((u64)edx << 32) | eax;
}
#endif // ifndef _WIN32
namespace Common {
CPUInfo cpu_info;
CPUInfo::CPUInfo() {
Detect();
}
#endif // ifndef _MSC_VER
// Detects the various CPU features
void CPUInfo::Detect() {
memset(this, 0, sizeof(*this));
#ifdef ARCHITECTURE_X64
Mode64bit = true;
OS64bit = true;
#endif
num_cores = 1;
static CPUCaps Detect() {
CPUCaps caps = {};
// Set obvious defaults, for extra safety
if (Mode64bit) {
bSSE = true;
bSSE2 = true;
bLongMode = true;
}
caps.num_cores = std::thread::hardware_concurrency();
// Assumes the CPU supports the CPUID instruction. Those that don't would likely not support
// Citra at all anyway
// Assume CPU supports the CPUID instruction. Those that don't can barely
// boot modern OS:es anyway.
int cpu_id[4];
memset(brand_string, 0, sizeof(brand_string));
memset(caps.brand_string, 0, sizeof(caps.brand_string));
// Detect CPU's CPUID capabilities, and grab CPU string
// Detect CPU's CPUID capabilities and grab CPU string
__cpuid(cpu_id, 0x00000000);
u32 max_std_fn = cpu_id[0]; // EAX
*((int *)brand_string) = cpu_id[1];
*((int *)(brand_string + 4)) = cpu_id[3];
*((int *)(brand_string + 8)) = cpu_id[2];
std::memcpy(&caps.brand_string[0], &cpu_id[1], sizeof(int));
std::memcpy(&caps.brand_string[4], &cpu_id[3], sizeof(int));
std::memcpy(&caps.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 = VENDOR_INTEL;
else if (!strcmp(brand_string, "AuthenticAMD"))
vendor = VENDOR_AMD;
if (!strcmp(caps.brand_string, "GenuineIntel"))
caps.vendor = CPUVendor::INTEL;
else if (!strcmp(caps.brand_string, "AuthenticAMD"))
caps.vendor = CPUVendor::AMD;
else
vendor = VENDOR_OTHER;
caps.vendor = CPUVendor::OTHER;
// Set reasonable default brand string even if brand string not available.
strcpy(cpu_string, brand_string);
// Set reasonable default brand string even if brand string not available
strcpy(caps.cpu_string, caps.brand_string);
// Detect family and other misc stuff.
bool ht = false;
HTT = ht;
logical_cpu_count = 1;
// Detect family and other miscellaneous features
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;
logical_cpu_count = (cpu_id[1] >> 16) & 0xFF;
ht = (cpu_id[3] >> 28) & 1;
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] >> 25) & 1) caps.sse = true;
if ((cpu_id[3] >> 26) & 1) caps.sse2 = true;
if ((cpu_id[2]) & 1) caps.sse3 = true;
if ((cpu_id[2] >> 9) & 1) caps.ssse3 = true;
if ((cpu_id[2] >> 19) & 1) caps.sse4_1 = true;
if ((cpu_id[2] >> 20) & 1) caps.sse4_2 = true;
if ((cpu_id[2] >> 22) & 1) caps.movbe = true;
if ((cpu_id[2] >> 25) & 1) caps.aes = true;
if ((cpu_id[3] >> 24) & 1)
{
// We can use FXSAVE.
bFXSR = true;
if ((cpu_id[3] >> 24) & 1) {
caps.fxsave_fxrstor = true;
}
// AVX support requires 3 separate checks:
@ -134,95 +107,80 @@ void CPUInfo::Detect() {
// - 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) {
bAVX = true;
caps.avx = true;
if ((cpu_id[2] >> 12) & 1)
bFMA = true;
caps.fma = true;
}
}
if (max_std_fn >= 7) {
__cpuidex(cpu_id, 0x00000007, 0x00000000);
// careful; we can't enable AVX2 unless the XSAVE/XGETBV checks above passed
// Can't enable AVX2 unless the XSAVE/XGETBV checks above passed
if ((cpu_id[1] >> 5) & 1)
bAVX2 = bAVX;
caps.avx2 = caps.avx;
if ((cpu_id[1] >> 3) & 1)
bBMI1 = true;
caps.bmi1 = true;
if ((cpu_id[1] >> 8) & 1)
bBMI2 = true;
caps.bmi2 = true;
}
}
bFlushToZero = bSSE;
caps.flush_to_zero = caps.sse;
if (max_ex_fn >= 0x80000004) {
// Extract CPU model string
__cpuid(cpu_id, 0x80000002);
memcpy(cpu_string, cpu_id, sizeof(cpu_id));
std::memcpy(caps.cpu_string, cpu_id, sizeof(cpu_id));
__cpuid(cpu_id, 0x80000003);
memcpy(cpu_string + 16, cpu_id, sizeof(cpu_id));
std::memcpy(caps.cpu_string + 16, cpu_id, sizeof(cpu_id));
__cpuid(cpu_id, 0x80000004);
memcpy(cpu_string + 32, cpu_id, sizeof(cpu_id));
std::memcpy(caps.cpu_string + 32, cpu_id, sizeof(cpu_id));
}
if (max_ex_fn >= 0x80000001) {
// Check for more features.
// Check for more features
__cpuid(cpu_id, 0x80000001);
if (cpu_id[2] & 1) bLAHFSAHF64 = true;
if ((cpu_id[2] >> 5) & 1) bLZCNT = true;
if ((cpu_id[2] >> 16) & 1) bFMA4 = true;
if ((cpu_id[3] >> 29) & 1) bLongMode = true;
if (cpu_id[2] & 1) caps.lahf_sahf_64 = true;
if ((cpu_id[2] >> 5) & 1) caps.lzcnt = true;
if ((cpu_id[2] >> 16) & 1) caps.fma4 = true;
if ((cpu_id[3] >> 29) & 1) caps.long_mode = true;
}
num_cores = (logical_cpu_count == 0) ? 1 : logical_cpu_count;
if (max_ex_fn >= 0x80000008) {
// Get number of cores. This is a bit complicated. Following AMD manual here.
__cpuid(cpu_id, 0x80000008);
int apic_id_core_id_size = (cpu_id[2] >> 12) & 0xF;
if (apic_id_core_id_size == 0) {
if (ht) {
// New mechanism for modern Intel CPUs.
if (vendor == VENDOR_INTEL) {
__cpuidex(cpu_id, 0x00000004, 0x00000000);
int cores_x_package = ((cpu_id[0] >> 26) & 0x3F) + 1;
HTT = (cores_x_package < logical_cpu_count);
cores_x_package = ((logical_cpu_count % cores_x_package) == 0) ? cores_x_package : 1;
num_cores = (cores_x_package > 1) ? cores_x_package : num_cores;
logical_cpu_count /= cores_x_package;
}
}
} else {
// Use AMD's new method.
num_cores = (cpu_id[2] & 0xFF) + 1;
}
}
return caps;
}
// Turn the CPU info into a string we can show
std::string CPUInfo::Summarize() {
std::string sum(cpu_string);
const CPUCaps& GetCPUCaps() {
static CPUCaps caps = Detect();
return caps;
}
std::string GetCPUCapsString() {
auto caps = GetCPUCaps();
std::string sum(caps.cpu_string);
sum += " (";
sum += brand_string;
sum += caps.brand_string;
sum += ")";
if (bSSE) sum += ", SSE";
if (bSSE2) {
if (caps.sse) sum += ", SSE";
if (caps.sse2) {
sum += ", SSE2";
if (!bFlushToZero)
sum += " (but not DAZ!)";
if (!caps.flush_to_zero) sum += " (without 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";
if (bFMA) sum += ", FMA";
if (bAES) sum += ", AES";
if (bMOVBE) sum += ", MOVBE";
if (bLongMode) sum += ", 64-bit support";
if (caps.sse3) sum += ", SSE3";
if (caps.ssse3) sum += ", SSSE3";
if (caps.sse4_1) sum += ", SSE4.1";
if (caps.sse4_2) sum += ", SSE4.2";
if (caps.avx) sum += ", AVX";
if (caps.avx2) sum += ", AVX2";
if (caps.bmi1) sum += ", BMI1";
if (caps.bmi2) sum += ", BMI2";
if (caps.fma) sum += ", FMA";
if (caps.aes) sum += ", AES";
if (caps.movbe) sum += ", MOVBE";
if (caps.long_mode) sum += ", 64-bit support";
return sum;
}

99
src/common/x64/cpu_detect.h
View file

@ -1,81 +1,66 @@
// Copyright 2013 Dolphin Emulator Project / 2014 Citra Emulator Project
// Copyright 2013 Dolphin Emulator Project / 2015 Citra Emulator Project
// Licensed under GPLv2 or any later version
// Refer to the license.txt file included.
// Detect the CPU, so we'll know which optimizations to use
#pragma once
#include <string>
namespace Common {
enum CPUVendor
{
VENDOR_INTEL = 0,
VENDOR_AMD = 1,
VENDOR_ARM = 2,
VENDOR_OTHER = 3,
/// x86/x64 CPU vendors that may be detected by this module
enum class CPUVendor {
INTEL,
AMD,
OTHER,
};
struct CPUInfo
{
/// x86/x64 CPU capabilities that may be detected by this module
struct CPUCaps {
CPUVendor vendor;
char cpu_string[0x21];
char brand_string[0x41];
bool OS64bit;
bool CPU64bit;
bool Mode64bit;
bool HTT;
int num_cores;
int logical_cpu_count;
bool sse;
bool sse2;
bool sse3;
bool ssse3;
bool sse4_1;
bool sse4_2;
bool lzcnt;
bool avx;
bool avx2;
bool bmi1;
bool bmi2;
bool fma;
bool fma4;
bool aes;
bool bSSE;
bool bSSE2;
bool bSSE3;
bool bSSSE3;
bool bPOPCNT;
bool bSSE4_1;
bool bSSE4_2;
bool bLZCNT;
bool bSSE4A;
bool bAVX;
bool bAVX2;
bool bBMI1;
bool bBMI2;
bool bFMA;
bool bFMA4;
bool bAES;
// FXSAVE/FXRSTOR
bool bFXSR;
bool bMOVBE;
// This flag indicates that the hardware supports some mode
// in which denormal inputs _and_ outputs are automatically set to (signed) zero.
bool bFlushToZero;
bool bLAHFSAHF64;
bool bLongMode;
bool bAtom;
// Support for the FXSAVE and FXRSTOR instructions
bool fxsave_fxrstor;
// ARMv8 specific
bool bFP;
bool bASIMD;
bool bCRC32;
bool bSHA1;
bool bSHA2;
bool movbe;
// Call Detect()
explicit CPUInfo();
// This flag indicates that the hardware supports some mode in which denormal inputs and outputs
// are automatically set to (signed) zero.
bool flush_to_zero;
// Turn the cpu info into a string we can show
std::string Summarize();
// Support for LAHF and SAHF instructions in 64-bit mode
bool lahf_sahf_64;
private:
// Detects the various cpu features
void Detect();
bool long_mode;
};
extern CPUInfo cpu_info;
/**
* Gets the supported capabilities of the host CPU
* @return Reference to a CPUCaps struct with the detected host CPU capabilities
*/
const CPUCaps& GetCPUCaps();
/**
* Gets a string summary of the name and supported capabilities of the host CPU
* @return String summary
*/
std::string GetCPUCapsString();
} // namespace Common

18
src/common/x64/emitter.cpp
View file

@ -826,14 +826,14 @@ void XEmitter::BSR(int bits, X64Reg dest, OpArg src) {WriteBitSearchType(bits,de
void XEmitter::TZCNT(int bits, X64Reg dest, OpArg src)
{
CheckFlags();
if (!Common::cpu_info.bBMI1)
if (!Common::GetCPUCaps().bmi1)
ASSERT_MSG(0, "Trying to use BMI1 on a system that doesn't support it. Bad programmer.");
WriteBitSearchType(bits, dest, src, 0xBC, true);
}
void XEmitter::LZCNT(int bits, X64Reg dest, OpArg src)
{
CheckFlags();
if (!Common::cpu_info.bLZCNT)
if (!Common::GetCPUCaps().lzcnt)
ASSERT_MSG(0, "Trying to use LZCNT on a system that doesn't support it. Bad programmer.");
WriteBitSearchType(bits, dest, src, 0xBD, true);
}
@ -907,7 +907,7 @@ void XEmitter::MOVZX(int dbits, int sbits, X64Reg dest, OpArg src)
void XEmitter::MOVBE(int bits, const OpArg& dest, const OpArg& src)
{
ASSERT_MSG(Common::cpu_info.bMOVBE, "Generating MOVBE on a system that does not support it.");
ASSERT_MSG(Common::GetCPUCaps().movbe, "Generating MOVBE on a system that does not support it.");
if (bits == 8)
{
MOV(bits, dest, src);
@ -1420,7 +1420,7 @@ static int GetVEXpp(u8 opPrefix)
void XEmitter::WriteAVXOp(u8 opPrefix, u16 op, X64Reg regOp1, X64Reg regOp2, OpArg arg, int extrabytes)
{
if (!Common::cpu_info.bAVX)
if (!Common::GetCPUCaps().avx)
ASSERT_MSG(0, "Trying to use AVX on a system that doesn't support it. Bad programmer.");
int mmmmm = GetVEXmmmmm(op);
int pp = GetVEXpp(opPrefix);
@ -1445,7 +1445,7 @@ void XEmitter::WriteVEXOp(int size, u8 opPrefix, u16 op, X64Reg regOp1, X64Reg r
void XEmitter::WriteBMI1Op(int size, u8 opPrefix, u16 op, X64Reg regOp1, X64Reg regOp2, OpArg arg, int extrabytes)
{
CheckFlags();
if (!Common::cpu_info.bBMI1)
if (!Common::GetCPUCaps().bmi1)
ASSERT_MSG(0, "Trying to use BMI1 on a system that doesn't support it. Bad programmer.");
WriteVEXOp(size, opPrefix, op, regOp1, regOp2, arg, extrabytes);
}
@ -1453,7 +1453,7 @@ void XEmitter::WriteBMI1Op(int size, u8 opPrefix, u16 op, X64Reg regOp1, X64Reg
void XEmitter::WriteBMI2Op(int size, u8 opPrefix, u16 op, X64Reg regOp1, X64Reg regOp2, OpArg arg, int extrabytes)
{
CheckFlags();
if (!Common::cpu_info.bBMI2)
if (!Common::GetCPUCaps().bmi2)
ASSERT_MSG(0, "Trying to use BMI2 on a system that doesn't support it. Bad programmer.");
WriteVEXOp(size, opPrefix, op, regOp1, regOp2, arg, extrabytes);
}
@ -1647,7 +1647,7 @@ void XEmitter::UNPCKHPD(X64Reg dest, OpArg arg) {WriteSSEOp(0x66, 0x15, dest, ar
void XEmitter::MOVDDUP(X64Reg regOp, OpArg arg)
{
if (Common::cpu_info.bSSE3)
if (Common::GetCPUCaps().sse3)
{
WriteSSEOp(0xF2, 0x12, regOp, arg); //SSE3 movddup
}
@ -1737,14 +1737,14 @@ void XEmitter::PSRAD(X64Reg reg, int shift)
void XEmitter::WriteSSSE3Op(u8 opPrefix, u16 op, X64Reg regOp, OpArg arg, int extrabytes)
{
if (!Common::cpu_info.bSSSE3)
if (!Common::GetCPUCaps().ssse3)
ASSERT_MSG(0, "Trying to use SSSE3 on a system that doesn't support it. Bad programmer.");
WriteSSEOp(opPrefix, op, regOp, arg, extrabytes);
}
void XEmitter::WriteSSE41Op(u8 opPrefix, u16 op, X64Reg regOp, OpArg arg, int extrabytes)
{
if (!Common::cpu_info.bSSE4_1)
if (!Common::GetCPUCaps().sse4_1)
ASSERT_MSG(0, "Trying to use SSE4.1 on a system that doesn't support it. Bad programmer.");
WriteSSEOp(opPrefix, op, regOp, arg, extrabytes);
}

2
src/common/x64/emitter.h
View file

@ -532,7 +532,7 @@ public:
void MOVSX(int dbits, int sbits, X64Reg dest, OpArg src); //automatically uses MOVSXD if necessary
void MOVZX(int dbits, int sbits, X64Reg dest, OpArg src);
// Available only on Atom or >= Haswell so far. Test with cpu_info.bMOVBE.
// Available only on Atom or >= Haswell so far. Test with GetCPUCaps().movbe.
void MOVBE(int dbits, const OpArg& dest, const OpArg& src);
// Available only on AMD >= Phenom or Intel >= Haswell

10
src/video_core/shader/shader_jit_x64.cpp
View file

@ -223,7 +223,7 @@ void JitCompiler::Compile_DestEnable(Instruction instr,X64Reg src) {
// Not all components are enabled, so mask the result when storing to the destination register...
MOVAPS(SCRATCH, MDisp(STATE, UnitState::OutputOffset(dest)));
if (Common::cpu_info.bSSE4_1) {
if (Common::GetCPUCaps().sse4_1) {
u8 mask = ((swiz.dest_mask & 1) << 3) | ((swiz.dest_mask & 8) >> 3) | ((swiz.dest_mask & 2) << 1) | ((swiz.dest_mask & 4) >> 1);
BLENDPS(SCRATCH, R(src), mask);
} else {
@ -291,7 +291,7 @@ void JitCompiler::Compile_DP3(Instruction instr) {
Compile_SwizzleSrc(instr, 1, instr.common.src1, SRC1);
Compile_SwizzleSrc(instr, 2, instr.common.src2, SRC2);
if (Common::cpu_info.bSSE4_1) {
if (Common::GetCPUCaps().sse4_1) {
DPPS(SRC1, R(SRC2), 0x7f);
} else {
MULPS(SRC1, R(SRC2));
@ -314,7 +314,7 @@ void JitCompiler::Compile_DP4(Instruction instr) {
Compile_SwizzleSrc(instr, 1, instr.common.src1, SRC1);
Compile_SwizzleSrc(instr, 2, instr.common.src2, SRC2);
if (Common::cpu_info.bSSE4_1) {
if (Common::GetCPUCaps().sse4_1) {
DPPS(SRC1, R(SRC2), 0xff);
} else {
MULPS(SRC1, R(SRC2));
@ -341,7 +341,7 @@ void JitCompiler::Compile_MUL(Instruction instr) {
void JitCompiler::Compile_FLR(Instruction instr) {
Compile_SwizzleSrc(instr, 1, instr.common.src1, SRC1);
if (Common::cpu_info.bSSE4_1) {
if (Common::GetCPUCaps().sse4_1) {
ROUNDFLOORPS(SRC1, R(SRC1));
} else {
CVTPS2DQ(SRC1, R(SRC1));
@ -513,7 +513,7 @@ void JitCompiler::Compile_MAD(Instruction instr) {
Compile_SwizzleSrc(instr, 3, instr.mad.src3, SRC3);
}
if (Common::cpu_info.bFMA) {
if (Common::GetCPUCaps().fma) {
VFMADD213PS(SRC1, SRC2, R(SRC3));
} else {
MULPS(SRC1, R(SRC2));