FGFS-4.1/simgear
4
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Maximize simd optimization

Browse Source
This commit is contained in:
Erik Hofman
2016-11-17 11:48:16 +01:00
parent 4a4baf1b42
commit ff7e4597e7
2 changed files with 458 additions and 202 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
simgear/math/SGVec4.hxx
View File

@@ -19,6 +19,7 @@
#define SGVec4_H
#include <iosfwd>
#include "simd.hxx"
/// 4D Vector Class
template<typename T>
@@ -92,10 +93,10 @@ public:
/// Readonly raw storage interface
const T (&data(void) const)[4]
{ return _data; }
{ return _data.ptr(); }
/// Readonly raw storage interface
T (&data(void))[4]
{ return _data; }
{ return _data.ptr(); }
/// Inplace addition
SGVec4& operator+=(const SGVec4& v)
@@ -126,7 +127,7 @@ public:
{ return SGVec4(0, 0, 0, 1); }
private:
T _data[4];
simd4_t<T> _data;
};
/// Unary +, do nothing ...

653
simgear/math/simd.hxx
View File

@@ -18,6 +18,8 @@
#ifndef __SIMD_H__
#define __SIMD_H__ 1
#include <cstring>
# ifdef __SSE__
# include <xmmintrin.h>
# endif
@@ -50,12 +52,17 @@ public:
{
vec[0] = vec[1] = vec[2] = vec[3] = f;
}
simd4_t(const __vec4_t& v)
simd4_t(const __vec4_t v)
{
vec[0] = v[0];
vec[1] = v[1];
vec[2] = v[2];
vec[3] = v[3];
std::memcpy(vec, v, sizeof(float[4]));
}
simd4_t(const simd4_t& v)
{
# ifdef __SSE__
v4 = v.v4;
#else
std::memcpy(vec, v.vec, sizeof(float[4]));
#endif
}
inline float (&ptr(void))[4] {
@@ -71,23 +78,38 @@ public:
vec[0] = vec[1] = vec[2] = vec[3] = f;
return *this;
}
inline simd4_t& operator=(const __vec4_t& v)
inline simd4_t& operator=(const __vec4_t v)
{
vec[0] = v[0];
vec[1] = v[1];
vec[2] = v[2];
vec[3] = v[3];
std::memcpy(vec, v, sizeof(float[4]));
return *this;
}
inline simd4_t& operator=(const simd4_t& v)
{
# ifdef __SSE__
v4 = v.v4;
#else
std::memcpy(vec, v.vec, sizeof(float[4]));
#endif
return *this;
}
inline simd4_t& operator=(const simd4_t& v)
inline simd4_t operator+(float f)
{
vec[0] = v[0];
vec[1] = v[1];
vec[2] = v[2];
vec[3] = v[3];
return *this;
simd4_t r(*this);
r += f;
return r;
}
inline simd4_t operator+(const __vec4_t v)
{
simd4_t r(*this);
r += v;
return r;
}
inline simd4_t operator+(const simd4_t& v)
{
simd4_t r(*this);
r += v;
return v;
}
inline simd4_t operator-()
@@ -96,36 +118,64 @@ public:
r -= vec;
return r;
}
inline simd4_t operator+(float f)
{
simd4_t r(vec);
r += f;
return r;
}
inline simd4_t operator-(float f)
{
simd4_t r(vec);
simd4_t r(*this);
r -= f;
return r;
}
inline simd4_t operator-(const __vec4_t v)
{
simd4_t r(*this);
r -= v;
return r;
}
inline simd4_t operator-(simd4_t& v)
{
simd4_t r(*this);
r -= v;
return r;
}
inline simd4_t operator*(float f)
{
simd4_t r(vec);
simd4_t r(*this);
r *= f;
return r;
}
inline simd4_t operator*(const __vec4_t v)
{
simd4_t r(*this);
r *= v;
return r;
}
inline simd4_t operator*(simd4_t& v)
{
simd4_t r(*this);
r *= v;
return r;
}
inline simd4_t operator/(float f)
{
simd4_t r(vec);
simd4_t r(*this);
r /= f;
return r;
}
inline simd4_t operator/(const __vec4_t v)
{
simd4_t r(*this);
r /= v;
return r;
}
inline simd4_t operator/(simd4_t& v)
{
simd4_t r(*this);
r /= v;
return r;
}
inline simd4_t& operator +=(float f)
inline simd4_t& operator+=(float f)
{
# ifdef __SSE__
v4 += f;
@@ -137,8 +187,26 @@ public:
# endif
return *this;
}
inline simd4_t& operator+=(const __vec4_t v)
{
simd4_t r(v);
*this += r;
return *this;
}
inline simd4_t& operator+=(const simd4_t& v)
{
# ifdef __SSE__
v4 += v.v4;
# else
vec[0] += v[0];
vec[1] += v[1];
vec[2] += v[2];
vec[3] += v[3];
#endif
return *this;
}
inline simd4_t& operator -=(float f)
inline simd4_t& operator-=(float f)
{
# ifdef __SSE__
v4 -= f;
@@ -150,6 +218,24 @@ public:
# endif
return *this;
}
inline simd4_t& operator-=(const __vec4_t v)
{
simd4_t r(v);
*this -= r;
return *this;
}
inline simd4_t& operator-=(const simd4_t& v)
{
# ifdef __SSE__
v4 -= v.v4;
# else
vec[0] -= v[0];
vec[1] -= v[1];
vec[2] -= v[2];
vec[3] -= v[3];
#endif
return *this;
}
inline simd4_t& operator *=(float f)
{
@@ -163,8 +249,26 @@ public:
# endif
return *this;
}
inline simd4_t& operator*=(const __vec4_t v)
{
simd4_t r(v);
*this *= r;
return *this;
}
inline simd4_t& operator*=(const simd4_t& v)
{
# ifdef __SSE__
v4 *= v.v4;
# else
vec[0] *= v[0];
vec[1] *= v[1];
vec[2] *= v[2];
vec[3] *= v[3];
#endif
return *this;
}
inline simd4_t& operator /=(float f)
inline simd4_t& operator/=(float f)
{
# ifdef __SSE__
v4 /= f;
@@ -176,40 +280,22 @@ public:
#endif
return *this;
}
inline simd4_t& operator +=(__vec4_t v)
inline simd4_t& operator/=(const __vec4_t v)
{
vec[0] += v[0];
vec[1] += v[1];
vec[2] += v[2];
vec[3] += v[3];
simd4_t r(v);
*this /= r;
return *this;
}
inline simd4_t& operator -=(__vec4_t v)
{
vec[0] -= v[0];
vec[1] -= v[1];
vec[2] -= v[2];
vec[3] -= v[3];
return *this;
}
inline simd4_t& operator *=(__vec4_t v)
{
vec[0] *= v[0];
vec[1] *= v[1];
vec[2] *= v[2];
vec[3] *= v[3];
return *this;
}
inline simd4_t& operator /=(__vec4_t v)
inline simd4_t& operator/=(const simd4_t& v)
{
# ifdef __SSE__
v4 /= v.v4;
# else
vec[0] /= v[0];
vec[1] /= v[1];
vec[2] /= v[2];
vec[3] /= v[3];
#endif
return *this;
}
@@ -222,6 +308,8 @@ public:
}
};
template<>
class simd4_t<double>
{
@@ -239,16 +327,21 @@ private:
public:
simd4_t() {}
simd4_t(double f)
simd4_t(double d)
{
vec[0] = vec[1] = vec[2] = vec[3] = f;
vec[0] = vec[1] = vec[2] = vec[3] = d;
}
simd4_t(const __vec4_t& v)
simd4_t(const __vec4_t v)
{
vec[0] = v[0];
vec[1] = v[1];
vec[2] = v[2];
vec[3] = v[3];
std::memcpy(vec, v, sizeof(double[4]));
}
simd4_t(const simd4_t& v)
{
# ifdef __SSE__
v4 = v.v4;
#else
std::memcpy(vec, v.vec, sizeof(double[4]));
#endif
}
inline double (&ptr(void))[4] {
@@ -259,159 +352,237 @@ public:
return vec;
}
inline simd4_t& operator=(double f)
inline simd4_t& operator=(double d)
{
vec[0] = vec[1] = vec[2] = vec[3] = f;
vec[0] = vec[1] = vec[2] = vec[3] = d;
return *this;
}
inline simd4_t& operator=(const __vec4_t& v)
inline simd4_t& operator=(const __vec4_t v)
{
vec[0] = v[0];
vec[1] = v[1];
vec[2] = v[2];
vec[3] = v[3];
std::memcpy(vec, v, sizeof(double[4]));
return *this;
}
inline simd4_t& operator=(const simd4_t& v)
{
simd4_t(r);
vec[0] = v[0];
vec[1] = v[1];
vec[2] = v[2];
vec[3] = v[3];
# ifdef __SSE__
v4 = v.v4;
#else
std::memcpy(vec, v.vec, sizeof(double[4]));
#endif
return *this;
}
inline simd4_t operator+(double d)
{
simd4_t r(*this);
r += d;
return r;
}
inline simd4_t operator+(const __vec4_t v)
{
simd4_t r(*this);
r += v;
return r;
}
inline simd4_t operator+(const simd4_t& v)
{
simd4_t r(*this);
r += v;
return v;
}
inline simd4_t operator-()
{
simd4_t r(0.0f);
r -= vec;
return r;
}
inline simd4_t operator+(double f)
inline simd4_t operator-(double d)
{
simd4_t r(vec);
r += f;
simd4_t r(*this);
r -= d;
return r;
}
inline simd4_t operator-(const __vec4_t v)
{
simd4_t r(*this);
r -= v;
return r;
}
inline simd4_t operator-(simd4_t& v)
{
simd4_t r(*this);
r -= v;
return r;
}
inline simd4_t operator-(double f)
inline simd4_t operator*(double d)
{
simd4_t r(vec);
r -= f;
simd4_t r(*this);
r *= d;
return r;
}
inline simd4_t operator*(const __vec4_t v)
{
simd4_t r(*this);
r *= v;
return r;
}
inline simd4_t operator*(simd4_t& v)
{
simd4_t r(*this);
r *= v;
return r;
}
inline simd4_t operator*(double f)
inline simd4_t operator/(double d)
{
simd4_t r(vec);
r *= f;
simd4_t r(*this);
r /= d;
return r;
}
inline simd4_t operator/(const __vec4_t v)
{
simd4_t r(*this);
r /= v;
return r;
}
inline simd4_t operator/(simd4_t& v)
{
simd4_t r(*this);
r /= v;
return r;
}
inline simd4_t operator/(double f)
{
simd4_t r(vec);
r /= f;
return r;
}
inline simd4_t& operator +=(double f)
inline simd4_t& operator+=(double d)
{
# ifdef __SSE__
v4 += f;
v4 += d;
# else
vec[0] += f;
vec[1] += f;
vec[2] += f;
vec[3] += f;
vec[0] += d;
vec[1] += d;
vec[2] += d;
vec[3] += d;
# endif
return *this;
}
inline simd4_t& operator -=(double f)
inline simd4_t& operator+=(const __vec4_t v)
{
# ifdef __SSE__
v4 -= f;
# else
vec[0] -= f;
vec[1] -= f;
vec[2] -= f;
vec[3] -= f;
# endif
simd4_t r(v);
*this += r;
return *this;
}
inline simd4_t& operator *=(double f)
inline simd4_t& operator+=(const simd4_t& v)
{
# ifdef __SSE__
v4 *= f;
v4 += v.v4;
# else
vec[0] *= f;
vec[1] *= f;
vec[2] *= f;
vec[3] *= f;
# endif
return *this;
}
inline simd4_t& operator /=(double f)
{
# ifdef __SSE__
v4 /= f;
# else
vec[0] /= f;
vec[1] /= f;
vec[2] /= f;
vec[3] /= f;
#endif
return *this;
}
inline simd4_t& operator +=(__vec4_t v)
{
vec[0] += v[0];
vec[1] += v[1];
vec[2] += v[2];
vec[3] += v[3];
#endif
return *this;
}
inline simd4_t& operator -=(__vec4_t v)
inline simd4_t& operator-=(double d)
{
# ifdef __SSE__
v4 -= d;
# else
vec[0] -= d;
vec[1] -= d;
vec[2] -= d;
vec[3] -= d;
# endif
return *this;
}
inline simd4_t& operator-=(const __vec4_t v)
{
simd4_t r(v);
*this -= r;
return *this;
}
inline simd4_t& operator-=(const simd4_t& v)
{
# ifdef __SSE__
v4 -= v.v4;
# else
vec[0] -= v[0];
vec[1] -= v[1];
vec[2] -= v[2];
vec[3] -= v[3];
#endif
return *this;
}
inline simd4_t& operator *=(__vec4_t v)
inline simd4_t& operator *=(double d)
{
# ifdef __SSE__
v4 *= d;
# else
vec[0] *= d;
vec[1] *= d;
vec[2] *= d;
vec[3] *= d;
# endif
return *this;
}
inline simd4_t& operator*=(const __vec4_t v)
{
simd4_t r(v);
*this *= r;
return *this;
}
inline simd4_t& operator*=(const simd4_t& v)
{
# ifdef __SSE__
v4 *= v.v4;
# else
vec[0] *= v[0];
vec[1] *= v[1];
vec[2] *= v[2];
vec[3] *= v[3];
#endif
return *this;
}
inline simd4_t& operator /=(__vec4_t v)
inline simd4_t& operator/=(double d)
{
# ifdef __SSE__
v4 /= d;
# else
vec[0] /= d;
vec[1] /= d;
vec[2] /= d;
vec[3] /= d;
#endif
return *this;
}
inline simd4_t& operator/=(const __vec4_t v)
{
simd4_t r(v);
*this /= r;
return *this;
}
inline simd4_t& operator/=(const simd4_t& v)
{
# ifdef __SSE__
v4 /= v.v4;
# else
vec[0] /= v[0];
vec[1] /= v[1];
vec[2] /= v[2];
vec[3] /= v[3];
#endif
return *this;
}
inline operator const double*(void) const {
inline operator const double*() const {
return vec;
}
inline operator double*(void) {
inline operator double*() {
return vec;
}
};
@@ -437,12 +608,17 @@ public:
{
vec[0] = vec[1] = vec[2] = vec[3] = i;
}
simd4_t(const __vec4_t& v)
simd4_t(const __vec4_t v)
{
vec[0] = v[0];
vec[1] = v[1];
vec[2] = v[2];
vec[3] = v[3];
std::memcpy(vec, v, sizeof(int[4]));
}
simd4_t(const simd4_t& v)
{
# ifdef __SSE__
v4 = v.v4;
#else
std::memcpy(vec, v.vec, sizeof(int[4]));
#endif
}
inline int (&ptr(void))[4] {
@@ -458,23 +634,38 @@ public:
vec[0] = vec[1] = vec[2] = vec[3] = i;
return *this;
}
inline simd4_t& operator=(const __vec4_t& v)
inline simd4_t& operator=(const __vec4_t v)
{
vec[0] = v[0];
vec[1] = v[1];
vec[2] = v[2];
vec[3] = v[3];
std::memcpy(vec, v, sizeof(int[4]));
return *this;
}
inline simd4_t& operator=(const simd4_t& v)
{
# ifdef __SSE__
v4 = v.v4;
#else
std::memcpy(vec, v.vec, sizeof(int[4]));
#endif
return *this;
}
inline simd4_t& operator=(const simd4_t& v)
inline simd4_t operator+(int i)
{
vec[0] = v[0];
vec[1] = v[1];
vec[2] = v[2];
vec[3] = v[3];
return *this;
simd4_t r(*this);
r += i;
return r;
}
inline simd4_t operator+(const __vec4_t v)
{
simd4_t r(*this);
r += v;
return r;
}
inline simd4_t operator+(const simd4_t& v)
{
simd4_t r(*this);
r += v;
return v;
}
inline simd4_t operator-()
@@ -483,36 +674,64 @@ public:
r -= vec;
return r;
}
inline simd4_t operator+(int i)
{
simd4_t r(vec);
r += i;
return r;
}
inline simd4_t operator-(int i)
{
simd4_t r(vec);
simd4_t r(*this);
r -= i;
return r;
}
inline simd4_t operator-(const __vec4_t v)
{
simd4_t r(*this);
r -= v;
return r;
}
inline simd4_t operator-(simd4_t& v)
{
simd4_t r(*this);
r -= v;
return r;
}
inline simd4_t operator*(int i)
{
simd4_t r(vec);
simd4_t r(*this);
r *= i;
return r;
}
inline simd4_t operator*(const __vec4_t v)
{
simd4_t r(*this);
r *= v;
return r;
}
inline simd4_t operator*(simd4_t& v)
{
simd4_t r(*this);
r *= v;
return r;
}
inline simd4_t operator/(int i)
{
simd4_t r(vec);
simd4_t r(*this);
r /= i;
return r;
}
inline simd4_t operator/(const __vec4_t v)
{
simd4_t r(*this);
r /= v;
return r;
}
inline simd4_t operator/(simd4_t& v)
{
simd4_t r(*this);
r /= v;
return r;
}
inline simd4_t& operator +=(int i)
inline simd4_t& operator+=(int i)
{
# ifdef __SSE__
v4 += i;
@@ -524,8 +743,26 @@ public:
# endif
return *this;
}
inline simd4_t& operator+=(const __vec4_t v)
{
simd4_t r(v);
*this += r;
return *this;
}
inline simd4_t& operator+=(const simd4_t& v)
{
# ifdef __SSE__
v4 += v.v4;
# else
vec[0] += v[0];
vec[1] += v[1];
vec[2] += v[2];
vec[3] += v[3];
#endif
return *this;
}
inline simd4_t& operator -=(int i)
inline simd4_t& operator-=(int i)
{
# ifdef __SSE__
v4 -= i;
@@ -537,6 +774,24 @@ public:
# endif
return *this;
}
inline simd4_t& operator-=(const __vec4_t v)
{
simd4_t r(v);
*this -= r;
return *this;
}
inline simd4_t& operator-=(const simd4_t& v)
{
# ifdef __SSE__
v4 -= v.v4;
# else
vec[0] -= v[0];
vec[1] -= v[1];
vec[2] -= v[2];
vec[3] -= v[3];
#endif
return *this;
}
inline simd4_t& operator *=(int i)
{
@@ -550,8 +805,26 @@ public:
# endif
return *this;
}
inline simd4_t& operator*=(const __vec4_t v)
{
simd4_t r(v);
*this *= r;
return *this;
}
inline simd4_t& operator*=(const simd4_t& v)
{
# ifdef __SSE__
v4 *= v.v4;
# else
vec[0] *= v[0];
vec[1] *= v[1];
vec[2] *= v[2];
vec[3] *= v[3];
#endif
return *this;
}
inline simd4_t& operator /=(int i)
inline simd4_t& operator/=(int i)
{
# ifdef __SSE__
v4 /= i;
@@ -563,40 +836,22 @@ public:
#endif
return *this;
}
inline simd4_t& operator +=(__vec4_t v)
inline simd4_t& operator/=(const __vec4_t v)
{
vec[0] += v[0];
vec[1] += v[1];
vec[2] += v[2];
vec[3] += v[3];
simd4_t r(v);
*this /= r;
return *this;
}
inline simd4_t& operator -=(__vec4_t v)
{
vec[0] -= v[0];
vec[1] -= v[1];
vec[2] -= v[2];
vec[3] -= v[3];
return *this;
}
inline simd4_t& operator *=(__vec4_t v)
{
vec[0] *= v[0];
vec[1] *= v[1];
vec[2] *= v[2];
vec[3] *= v[3];
return *this;
}
inline simd4_t& operator /=(__vec4_t v)
inline simd4_t& operator/=(const simd4_t& v)
{
# ifdef __SSE__
v4 /= v.v4;
# else
vec[0] /= v[0];
vec[1] /= v[1];
vec[2] /= v[2];
vec[3] /= v[3];
#endif
return *this;
}