mewin/glm
1
0
Fork 0
Code Releases Activity

SIMD matrix functions optimizations

Browse Source
This commit is contained in:
Christophe Riccio
2016-05-30 18:01:06 +02:00
parent 84caa1092f
commit 688756b3e2
5 changed files with 90 additions and 48 deletions
Download Patch File Download Diff File Expand all files Collapse all files

18
glm/detail/func_matrix.inl
View File

@@ -7,6 +7,18 @@
namespace glm{
namespace detail
{
template <template <typename, precision> class matType, typename T, precision P>
struct compute_matrixCompMult
{
GLM_FUNC_QUALIFIER static matType<T, P> call(matType<T, P> const& x, matType<T, P> const& y)
{
matType<T, P> result(uninitialize);
for(length_t i = 0; i < result.length(); ++i)
result[i] = x[i] * y[i];
return result;
}
};
template <template <class, precision> class matType, typename T, precision P>
struct compute_transpose{};
@@ -347,11 +359,7 @@ namespace detail
GLM_FUNC_QUALIFIER matType<T, P> matrixCompMult(matType<T, P> const & x, matType<T, P> const & y)
{
GLM_STATIC_ASSERT(std::numeric_limits<T>::is_iec559, "'matrixCompMult' only accept floating-point inputs");
matType<T, P> result(uninitialize);
for(length_t i = 0; i < result.length(); ++i)
result[i] = x[i] * y[i];
return result;
return detail::compute_matrixCompMult<matType, T, P>::call(x, y);
}
template<typename T, precision P, template <typename, precision> class vecTypeA, template <typename, precision> class vecTypeB>

84
glm/detail/func_matrix_simd.inl
View File

@@ -1,6 +1,8 @@
/// @ref core
/// @file glm/detail/func_matrix_simd.inl
#if GLM_ARCH & GLM_ARCH_SSE2_BIT
#include "type_mat4x4.hpp"
#include "func_geometric.hpp"
#include "../simd/matrix.h"
@@ -8,17 +10,77 @@
namespace glm{
namespace detail
{
# if GLM_ARCH & GLM_ARCH_SSE2_BIT
template <precision P>
struct compute_inverse<tmat4x4, float, P>
template <precision P>
struct compute_matrixCompMult<tmat4x4, float, P>
{
GLM_FUNC_QUALIFIER static tmat4x4<float, P> call(tmat4x4<float, P> const & x, tmat4x4<float, P> const & y)
{
GLM_FUNC_QUALIFIER static tmat4x4<float, P> call(tmat4x4<float, P> const& m)
{
tmat4x4<float, P> Result(uninitialize);
glm_mat4_inverse(*reinterpret_cast<__m128 const(*)[4]>(&m[0].data), *reinterpret_cast<__m128(*)[4]>(&Result[0].data));
return Result;
}
};
# endif
tmat4x4<float, P> result(uninitialize);
glm_mat4_matrixCompMult(
*(glm_vec4 const (*)[4])&x[0].data,
*(glm_vec4 const (*)[4])&y[0].data,
*(glm_vec4(*)[4])&result[0].data);
return result;
}
};
template <precision P>
struct compute_transpose<tmat4x4, float, P>
{
GLM_FUNC_QUALIFIER static tmat4x4<float, P> call(tmat4x4<float, P> const & m)
{
tmat4x4<float, P> result(uninitialize);
glm_mat4_transpose(
*(glm_vec4 const (*)[4])&m[0].data,
*(glm_vec4(*)[4])&result[0].data);
return result;
}
};
template <precision P>
struct compute_determinant<tmat4x4, float, P>
{
GLM_FUNC_QUALIFIER static float call(tmat4x4<float, P> const& m)
{
return _mm_cvtss_f32(glm_mat4_determinant(*reinterpret_cast<__m128 const(*)[4]>(&m[0].data)));
}
};
template <precision P>
struct compute_inverse<tmat4x4, float, P>
{
GLM_FUNC_QUALIFIER static tmat4x4<float, P> call(tmat4x4<float, P> const& m)
{
tmat4x4<float, P> Result(uninitialize);
glm_mat4_inverse(*reinterpret_cast<__m128 const(*)[4]>(&m[0].data), *reinterpret_cast<__m128(*)[4]>(&Result[0].data));
return Result;
}
};
}//namespace detail
template<>
GLM_FUNC_QUALIFIER tmat4x4<float, lowp> outerProduct<float, lowp, tvec4, tvec4>(tvec4<float, lowp> const & c, tvec4<float, lowp> const & r)
{
tmat4x4<float, lowp> m(uninitialize);
glm_mat4_outerProduct(c.data, r.data, *reinterpret_cast<__m128(*)[4]>(&m[0].data));
return m;
}
template<>
GLM_FUNC_QUALIFIER tmat4x4<float, mediump> outerProduct<float, mediump, tvec4, tvec4>(tvec4<float, mediump> const & c, tvec4<float, mediump> const & r)
{
tmat4x4<float, mediump> m(uninitialize);
glm_mat4_outerProduct(c.data, r.data, *reinterpret_cast<__m128(*)[4]>(&m[0].data));
return m;
}
template<>
GLM_FUNC_QUALIFIER tmat4x4<float, highp> outerProduct<float, highp, tvec4, tvec4>(tvec4<float, highp> const & c, tvec4<float, highp> const & r)
{
tmat4x4<float, highp> m(uninitialize);
glm_mat4_outerProduct(c.data, r.data, *reinterpret_cast<__m128(*)[4]>(&m[0].data));
return m;
}
}//namespace glm
#endif

4
glm/gtx/simd_mat4.inl
View File

@@ -563,14 +563,14 @@ GLM_FUNC_QUALIFIER detail::fmat4x4SIMD outerProduct
GLM_FUNC_QUALIFIER detail::fmat4x4SIMD transpose(detail::fmat4x4SIMD const & m)
{
detail::fmat4x4SIMD result;
detail::sse_transpose_ps(&m[0].Data, &result[0].Data);
glm_mat4_transpose(&m[0].Data, &result[0].Data);
return result;
}
GLM_FUNC_QUALIFIER float determinant(detail::fmat4x4SIMD const & m)
{
float Result;
_mm_store_ss(&Result, detail::sse_det_ps(&m[0].Data));
_mm_store_ss(&Result, glm_mat4_determinant(&m[0].Data));
return Result;
}

4
glm/simd/matrix.h
View File

@@ -947,7 +947,7 @@ GLM_FUNC_QUALIFIER void glm_mat4_inverse_lowp(glm_vec4 const in[4], glm_vec4 out
out[3] = _mm_mul_ps(Inv3, Rcp0);
}
/*
GLM_FUNC_QUALIFIER void glm_f32m4_rotate(__m128 const in[4], float Angle, float const v[3], __m128 out[4])
GLM_FUNC_QUALIFIER void glm_mat4_rotate(__m128 const in[4], float Angle, float const v[3], __m128 out[4])
{
float a = glm::radians(Angle);
float c = cos(a);
@@ -1017,7 +1017,7 @@ GLM_FUNC_QUALIFIER void glm_f32m4_rotate(__m128 const in[4], float Angle, float
sse_mul_ps(in, Result, out);
}
*/
GLM_FUNC_QUALIFIER void glm_f32m4_outer(__m128 const & c, __m128 const & r, __m128 out[4])
GLM_FUNC_QUALIFIER void glm_mat4_outerProduct(__m128 const & c, __m128 const & r, __m128 out[4])
{
out[0] = _mm_mul_ps(c, _mm_shuffle_ps(r, r, _MM_SHUFFLE(0, 0, 0, 0)));
out[1] = _mm_mul_ps(c, _mm_shuffle_ps(r, r, _MM_SHUFFLE(1, 1, 1, 1)));