anulax
/
glm
1
0
Fork 0
Code Issues Pull Requests Packages project_board Releases Wiki Activity

Added generic quaternion SIMD optimizations

Browse Source
master
Christophe Riccio ago%!(EXTRA string=9 years)
parent 9459aac681
commit 166099e82d
4 changed files with 274 additions and 21 deletions
Whitespace
Show all changes Ignore whitespace when comparing lines Ignore changes in amount of whitespace Ignore changes in whitespace at EOL
Split View
Diff Options
Show Stats Download Patch File Download Diff File
  1. 2
      glm/detail/type_vec4.hpp
  2. 12
      glm/gtc/quaternion.hpp
  3. 83
      glm/gtc/quaternion.inl
  4. 198
      glm/gtc/quaternion_simd.inl

2
glm/detail/type_vec4.hpp
Unescape View File

@ -57,7 +57,7 @@ namespace glm
# ifdef GLM_SWIZZLE
GLM_SWIZZLE_GEN_VEC_FROM_VEC4(T, P, tvec4, tvec2, tvec3, tvec4)
# endif//GLM_SWIZZLE
# endif//GLM_LANG
# endif
// -- Component accesses --

12
glm/gtc/quaternion.hpp
Unescape View File

@ -40,7 +40,15 @@ namespace glm
// -- Data --
T x, y, z, w;
# if GLM_HAS_UNRESTRICTED_UNIONS
union
{
struct { T x, y, z, w;};
typename detail::storage<T, sizeof(T) * 4, detail::is_aligned<P>::value>::type data;
};
# else
T x, y, z, w;
# endif
// -- Component accesses --
@ -97,6 +105,8 @@ namespace glm
template <typename U>
GLM_FUNC_DECL tquat<T, P> & operator+=(tquat<U, P> const & q);
template <typename U>
GLM_FUNC_DECL tquat<T, P> & operator-=(tquat<U, P> const & q);
template <typename U>
GLM_FUNC_DECL tquat<T, P> & operator*=(tquat<U, P> const & q);
template <typename U>
GLM_FUNC_DECL tquat<T, P> & operator*=(U s);

83
glm/gtc/quaternion.inl
Unescape View File

@ -12,12 +12,57 @@ namespace detail
template <typename T, precision P, bool Aligned>
struct compute_dot<tquat, T, P, Aligned>
{
static GLM_FUNC_QUALIFIER T call(tquat<T, P> const & x, tquat<T, P> const & y)
static GLM_FUNC_QUALIFIER T call(tquat<T, P> const& x, tquat<T, P> const& y)
{
tvec4<T, P> tmp(x.x * y.x, x.y * y.y, x.z * y.z, x.w * y.w);
return (tmp.x + tmp.y) + (tmp.z + tmp.w);
}
};
template <typename T, precision P, bool Aligned>
struct compute_quat_add
{
static tquat<T, P> call(tquat<T, P> const& q, tquat<T, P> const& p)
{
return tquat<T, P>(q.w + p.w, q.x + p.x, q.y + p.y, q.z + p.z);
}
};
template <typename T, precision P, bool Aligned>
struct compute_quat_sub
{
static tquat<T, P> call(tquat<T, P> const& q, tquat<T, P> const& p)
{
return tquat<T, P>(q.w - p.w, q.x - p.x, q.y - p.y, q.z - p.z);
}
};
template <typename T, precision P, bool Aligned>
struct compute_quat_mul_scalar
{
static tquat<T, P> call(tquat<T, P> const& q, T s)
{
return tquat<T, P>(q.w * s, q.x * s, q.y * s, q.z * s);
}
};
template <typename T, precision P, bool Aligned>
struct compute_quat_div_scalar
{
static tquat<T, P> call(tquat<T, P> const& q, T s)
{
return tquat<T, P>(q.w / s, q.x / s, q.y / s, q.z / s);
}
};
template <typename T, precision P, bool Aligned>
struct compute_quat_mul_vec4
{
static tvec4<T, P> call(tquat<T, P> const & q, tvec4<T, P> const & v)
{
return tvec4<T, P>(q * tvec3<T, P>(v), v.w);
}
};
}//namespace detail
// -- Component accesses --
@ -198,13 +243,16 @@ namespace detail
template <typename T, precision P>
template <typename U>
GLM_FUNC_QUALIFIER tquat<T, P> & tquat<T, P>::operator+=(tquat<U, P> const & q)
GLM_FUNC_QUALIFIER tquat<T, P> & tquat<T, P>::operator+=(tquat<U, P> const& q)
{
this->w += static_cast<T>(q.w);
this->x += static_cast<T>(q.x);
this->y += static_cast<T>(q.y);
this->z += static_cast<T>(q.z);
return *this;
return (*this = detail::compute_quat_add<T, P, detail::is_aligned<P>::value>::call(*this, tquat<T, P>(q)));
}
template <typename T, precision P>
template <typename U>
GLM_FUNC_QUALIFIER tquat<T, P> & tquat<T, P>::operator-=(tquat<U, P> const& q)
{
return (*this = detail::compute_quat_sub<T, P, detail::is_aligned<P>::value>::call(*this, tquat<T, P>(q)));
}
template <typename T, precision P>
@ -225,22 +273,14 @@ namespace detail
template <typename U>
GLM_FUNC_QUALIFIER tquat<T, P> & tquat<T, P>::operator*=(U s)
{
this->w *= static_cast<U>(s);
this->x *= static_cast<U>(s);
this->y *= static_cast<U>(s);
this->z *= static_cast<U>(s);
return *this;
return (*this = detail::compute_quat_mul_scalar<T, P, detail::is_aligned<P>::value>::call(*this, static_cast<U>(s)));
}
template <typename T, precision P>
template <typename U>
GLM_FUNC_QUALIFIER tquat<T, P> & tquat<T, P>::operator/=(U s)
{
this->w /= static_cast<U>(s);
this->x /= static_cast<U>(s);
this->y /= static_cast<U>(s);
this->z /= static_cast<U>(s);
return *this;
return (*this = detail::compute_quat_div_scalar<T, P, detail::is_aligned<P>::value>::call(*this, static_cast<U>(s)));
}
// -- Unary bit operators --
@ -288,9 +328,9 @@ namespace detail
}
template <typename T, precision P>
GLM_FUNC_QUALIFIER tvec4<T, P> operator*(tquat<T, P> const & q, tvec4<T, P> const & v)
GLM_FUNC_QUALIFIER tvec4<T, P> operator*(tquat<T, P> const& q, tvec4<T, P> const& v)
{
return tvec4<T, P>(q * tvec3<T, P>(v), v.w);
return detail::compute_quat_mul_vec4<T, P, detail::is_aligned<P>::value>::call(q, v);
}
template <typename T, precision P>
@ -738,3 +778,8 @@ namespace detail
return Result;
}
}//namespace glm
#if GLM_ARCH != GLM_ARCH_PURE && GLM_HAS_UNRESTRICTED_UNIONS
# include "quaternion_simd.inl"
#endif

198
glm/gtc/quaternion_simd.inl
Unescape View File

@ -0,0 +1,198 @@
/// @ref core
/// @file glm/gtc/quaternion_simd.inl
#if GLM_ARCH & GLM_ARCH_SSE2_BIT
namespace glm{
namespace detail
{
/*
template <precision P>
struct compute_quat_mul<float, P, true>
{
static tquat<float, P> call(tquat<float, P> const& q1, tquat<float, P> const& q2)
{
// SSE2 STATS: 11 shuffle, 8 mul, 8 add
// SSE4 STATS: 3 shuffle, 4 mul, 4 dpps
__m128 const mul0 = _mm_mul_ps(q1.Data, _mm_shuffle_ps(q2.Data, q2.Data, _MM_SHUFFLE(0, 1, 2, 3)));
__m128 const mul1 = _mm_mul_ps(q1.Data, _mm_shuffle_ps(q2.Data, q2.Data, _MM_SHUFFLE(1, 0, 3, 2)));
__m128 const mul2 = _mm_mul_ps(q1.Data, _mm_shuffle_ps(q2.Data, q2.Data, _MM_SHUFFLE(2, 3, 0, 1)));
__m128 const mul3 = _mm_mul_ps(q1.Data, q2.Data);
# if GLM_ARCH & GLM_ARCH_SSE41_BIT
__m128 const add0 = _mm_dp_ps(mul0, _mm_set_ps(1.0f, -1.0f, 1.0f, 1.0f), 0xff);
__m128 const add1 = _mm_dp_ps(mul1, _mm_set_ps(1.0f, 1.0f, 1.0f, -1.0f), 0xff);
__m128 const add2 = _mm_dp_ps(mul2, _mm_set_ps(1.0f, 1.0f, -1.0f, 1.0f), 0xff);
__m128 const add3 = _mm_dp_ps(mul3, _mm_set_ps(1.0f, -1.0f, -1.0f, -1.0f), 0xff);
# else
__m128 const mul4 = _mm_mul_ps(mul0, _mm_set_ps(1.0f, -1.0f, 1.0f, 1.0f));
__m128 const add0 = _mm_add_ps(mul0, _mm_movehl_ps(mul4, mul4));
__m128 const add4 = _mm_add_ss(add0, _mm_shuffle_ps(add0, add0, 1));
__m128 const mul5 = _mm_mul_ps(mul1, _mm_set_ps(1.0f, 1.0f, 1.0f, -1.0f));
__m128 const add1 = _mm_add_ps(mul1, _mm_movehl_ps(mul5, mul5));
__m128 const add5 = _mm_add_ss(add1, _mm_shuffle_ps(add1, add1, 1));
__m128 const mul6 = _mm_mul_ps(mul2, _mm_set_ps(1.0f, 1.0f, -1.0f, 1.0f));
__m128 const add2 = _mm_add_ps(mul6, _mm_movehl_ps(mul6, mul6));
__m128 const add6 = _mm_add_ss(add2, _mm_shuffle_ps(add2, add2, 1));
__m128 const mul7 = _mm_mul_ps(mul3, _mm_set_ps(1.0f, -1.0f, -1.0f, -1.0f));
__m128 const add3 = _mm_add_ps(mul3, _mm_movehl_ps(mul7, mul7));
__m128 const add7 = _mm_add_ss(add3, _mm_shuffle_ps(add3, add3, 1));
#endif
// This SIMD code is a politically correct way of doing this, but in every test I've tried it has been slower than
// the final code below. I'll keep this here for reference - maybe somebody else can do something better...
//
//__m128 xxyy = _mm_shuffle_ps(add4, add5, _MM_SHUFFLE(0, 0, 0, 0));
//__m128 zzww = _mm_shuffle_ps(add6, add7, _MM_SHUFFLE(0, 0, 0, 0));
//
//return _mm_shuffle_ps(xxyy, zzww, _MM_SHUFFLE(2, 0, 2, 0));
tquat<float, P> Result(uninitialize);
_mm_store_ss(&Result.x, add4);
_mm_store_ss(&Result.y, add5);
_mm_store_ss(&Result.z, add6);
_mm_store_ss(&Result.w, add7);
return Result;
}
};
*/
template <precision P>
struct compute_dot<tquat, float, P, true>
{
static GLM_FUNC_QUALIFIER float call(tquat<float, P> const& x, tquat<float, P> const& y)
{
return _mm_cvtss_f32(glm_vec1_dot(x.data, y.data));
}
};
template <precision P>
struct compute_quat_add<float, P, true>
{
static tquat<float, P> call(tquat<float, P> const& q, tquat<float, P> const& p)
{
tquat<float, P> Result(uninitialize);
Result.data = _mm_add_ps(q.data, p.data);
return Result;
}
};
# if GLM_ARCH & GLM_ARCH_AVX_BIT
template <precision P>
struct compute_quat_add<double, P, true>
{
static tquat<double, P> call(tquat<double, P> const & a, tquat<double, P> const & b)
{
tquat<double, P> Result(uninitialize);
Result.data = _mm256_add_pd(a.data, b.data);
return Result;
}
};
# endif
template <precision P>
struct compute_quat_sub<float, P, true>
{
static tquat<float, P> call(tquat<float, P> const& q, tquat<float, P> const& p)
{
tvec4<float, P> Result(uninitialize);
Result.data = _mm_sub_ps(q.data, p.data);
return Result;
}
};
# if GLM_ARCH & GLM_ARCH_AVX_BIT
template <precision P>
struct compute_quat_sub<double, P, true>
{
static tquat<double, P> call(tquat<double, P> const & a, tquat<double, P> const & b)
{
tquat<double, P> Result(uninitialize);
Result.data = _mm256_sub_pd(a.data, b.data);
return Result;
}
};
# endif
template <precision P>
struct compute_quat_mul_scalar<float, P, true>
{
static tquat<float, P> call(tquat<float, P> const& q, float s)
{
tvec4<float, P> Result(uninitialize);
Result.data = _mm_mul_ps(q.data, _mm_set_ps1(s));
return Result;
}
};
# if GLM_ARCH & GLM_ARCH_AVX_BIT
template <precision P>
struct compute_quat_mul_scalar<double, P, true>
{
static tquat<double, P> call(tquat<double, P> const& q, double s)
{
tquat<double, P> Result(uninitialize);
Result.data = _mm256_mul_pd(q.data, _mm_set_ps1(s));
return Result;
}
};
# endif
template <precision P>
struct compute_quat_div_scalar<float, P, true>
{
static tquat<float, P> call(tquat<float, P> const& q, float s)
{
tvec4<float, P> Result(uninitialize);
Result.data = _mm_div_ps(q.data, _mm_set_ps1(s));
return Result;
}
};
# if GLM_ARCH & GLM_ARCH_AVX_BIT
template <precision P>
struct compute_quat_div_scalar<double, P, true>
{
static tquat<double, P> call(tquat<double, P> const& q, double s)
{
tquat<double, P> Result(uninitialize);
Result.data = _mm256_div_pd(q.data, _mm_set_ps1(s));
return Result;
}
};
# endif
template <precision P>
struct compute_quat_mul_vec4<float, P, true>
{
static tvec4<float, P> call(tquat<float, P> const& q, tvec4<float, P> const& v)
{
__m128 const q_wwww = _mm_shuffle_ps(q.data, q.data, _MM_SHUFFLE(3, 3, 3, 3));
__m128 const q_swp0 = _mm_shuffle_ps(q.data, q.data, _MM_SHUFFLE(3, 0, 2, 1));
__m128 const q_swp1 = _mm_shuffle_ps(q.data, q.data, _MM_SHUFFLE(3, 1, 0, 2));
__m128 const v_swp0 = _mm_shuffle_ps(v.data, v.data, _MM_SHUFFLE(3, 0, 2, 1));
__m128 const v_swp1 = _mm_shuffle_ps(v.data, v.data, _MM_SHUFFLE(3, 1, 0, 2));
__m128 uv = _mm_sub_ps(_mm_mul_ps(q_swp0, v_swp1), _mm_mul_ps(q_swp1, v_swp0));
__m128 uv_swp0 = _mm_shuffle_ps(uv, uv, _MM_SHUFFLE(3, 0, 2, 1));
__m128 uv_swp1 = _mm_shuffle_ps(uv, uv, _MM_SHUFFLE(3, 1, 0, 2));
__m128 uuv = _mm_sub_ps(_mm_mul_ps(q_swp0, uv_swp1), _mm_mul_ps(q_swp1, uv_swp0));
__m128 const two = _mm_set1_ps(2.0f);
uv = _mm_mul_ps(uv, _mm_mul_ps(q_wwww, two));
uuv = _mm_mul_ps(uuv, two);
tvec4<float, P> Result(uninitialize);
Result.data = _mm_add_ps(v.Data, _mm_add_ps(uv, uuv));
return Result;
}
};
}//namespace detail
}//namespace glm
#endif//GLM_ARCH & GLM_ARCH_SSE2_BIT
Write Preview
Loading…
Cancel
Save