Page MenuHome
Differential D8762 Diff 28333 extern/bullet2/src/LinearMath/btCpuFeatureUtility.h

Changeset View

Standalone View

View Options

extern/bullet2/src/LinearMath/btCpuFeatureUtility.h

#ifndef BT_CPU_UTILITY_H #ifndef BT_CPU_UTILITY_H
#define BT_CPU_UTILITY_H #define BT_CPU_UTILITY_H
#include "LinearMath/btScalar.h" #include "LinearMath/btScalar.h"
#include <string.h>//memset #include <string.h> //memset
#ifdef USE_SIMD #ifdef USE_SIMD
#include <emmintrin.h> #include <emmintrin.h>
#ifdef BT_ALLOW_SSE4 #ifdef BT_ALLOW_SSE4
#include <intrin.h> #include <intrin.h>
#endif //BT_ALLOW_SSE4 #endif //BT_ALLOW_SSE4
#endif //USE_SIMD #endif //USE_SIMD
#if defined BT_USE_NEON #if defined BT_USE_NEON
#define ARM_NEON_GCC_COMPATIBILITY 1 #define ARM_NEON_GCC_COMPATIBILITY 1
#include <arm_neon.h> #include <arm_neon.h>
#include <sys/types.h> #include <sys/types.h>
#include <sys/sysctl.h> //for sysctlbyname #include <sys/sysctl.h> //for sysctlbyname
#endif //BT_USE_NEON #endif //BT_USE_NEON
///Rudimentary btCpuFeatureUtility for CPU features: only report the features that Bullet actually uses (SSE4/FMA3, NEON_HPFP) ///Rudimentary btCpuFeatureUtility for CPU features: only report the features that Bullet actually uses (SSE4/FMA3, NEON_HPFP)
///We assume SSE2 in case BT_USE_SSE2 is defined in LinearMath/btScalar.h ///We assume SSE2 in case BT_USE_SSE2 is defined in LinearMath/btScalar.h
class btCpuFeatureUtility class btCpuFeatureUtility
{ {
public: public:
enum btCpuFeature enum btCpuFeature
{ {
CPU_FEATURE_FMA3=1, CPU_FEATURE_FMA3 = 1,
CPU_FEATURE_SSE4_1=2, CPU_FEATURE_SSE4_1 = 2,
CPU_FEATURE_NEON_HPFP=4 CPU_FEATURE_NEON_HPFP = 4
}; };
static int getCpuFeatures() static int getCpuFeatures()
{ {
static int capabilities = 0; static int capabilities = 0;
static bool testedCapabilities = false; static bool testedCapabilities = false;
if (0 != testedCapabilities) if (0 != testedCapabilities)
{ {
return capabilities; return capabilities;
} }
#ifdef BT_USE_NEON #ifdef BT_USE_NEON
{ {
uint32_t hasFeature = 0; uint32_t hasFeature = 0;
size_t featureSize = sizeof(hasFeature); size_t featureSize = sizeof(hasFeature);
int err = sysctlbyname("hw.optional.neon_hpfp", &hasFeature, &featureSize, NULL, 0); int err = sysctlbyname("hw.optional.neon_hpfp", &hasFeature, &featureSize, NULL, 0);
if (0 == err && hasFeature) if (0 == err && hasFeature)
capabilities |= CPU_FEATURE_NEON_HPFP; capabilities |= CPU_FEATURE_NEON_HPFP;
} }
#endif //BT_USE_NEON #endif //BT_USE_NEON
#ifdef BT_ALLOW_SSE4 #ifdef BT_ALLOW_SSE4
{ {
int cpuInfo[4]; int cpuInfo[4];
memset(cpuInfo, 0, sizeof(cpuInfo)); memset(cpuInfo, 0, sizeof(cpuInfo));
unsigned long long sseExt = 0; unsigned long long sseExt = 0;
__cpuid(cpuInfo, 1); __cpuid(cpuInfo, 1);
bool osUsesXSAVE_XRSTORE = cpuInfo[2] & (1 << 27) || false; bool osUsesXSAVE_XRSTORE = cpuInfo[2] & (1 << 27) || false;
bool cpuAVXSuport = cpuInfo[2] & (1 << 28) || false; bool cpuAVXSuport = cpuInfo[2] & (1 << 28) || false;
if (osUsesXSAVE_XRSTORE && cpuAVXSuport) if (osUsesXSAVE_XRSTORE && cpuAVXSuport)
{ {
sseExt = _xgetbv(0); sseExt = _xgetbv(0);
} }
const int OSXSAVEFlag = (1UL << 27); const int OSXSAVEFlag = (1UL << 27);
const int AVXFlag = ((1UL << 28) | OSXSAVEFlag); const int AVXFlag = ((1UL << 28) | OSXSAVEFlag);
const int FMAFlag = ((1UL << 12) | AVXFlag | OSXSAVEFlag); const int FMAFlag = ((1UL << 12) | AVXFlag | OSXSAVEFlag);
if ((cpuInfo[2] & FMAFlag) == FMAFlag && (sseExt & 6) == 6) if ((cpuInfo[2] & FMAFlag) == FMAFlag && (sseExt & 6) == 6)
{ {
capabilities |= btCpuFeatureUtility::CPU_FEATURE_FMA3; capabilities |= btCpuFeatureUtility::CPU_FEATURE_FMA3;
} }
const int SSE41Flag = (1 << 19); const int SSE41Flag = (1 << 19);
if (cpuInfo[2] & SSE41Flag) if (cpuInfo[2] & SSE41Flag)
{ {
capabilities |= btCpuFeatureUtility::CPU_FEATURE_SSE4_1; capabilities |= btCpuFeatureUtility::CPU_FEATURE_SSE4_1;
} }
} }
#endif//BT_ALLOW_SSE4 #endif //BT_ALLOW_SSE4
testedCapabilities = true; testedCapabilities = true;
return capabilities; return capabilities;
} }
}; };
#endif //BT_CPU_UTILITY_H #endif //BT_CPU_UTILITY_H