Scarfski/glm
1
0
Fork 0
mirror of https://github.com/g-truc/glm.git synced 2026-10-22 08:50:40 +00:00
Code Issues Packages Projects Releases Wiki Activity

SIMD dot product and normalization changes (#1427)

Browse Source 
* added aligned_highp/medp normalization
This commit is contained in:
Forward-Fax3
2026-01-01 21:43:36 +00:00
committed by GitHub
parent 8f6213d379
commit 2436fe7671
2 changed files with 104 additions and 8 deletions
Download Patch File Download Diff File Expand all files Collapse all files

39
glm/detail/func_geometric_simd.inl
View File

@@ -1,4 +1,4 @@
/// @ref core
/// @ref core
/// @file glm/detail/func_geometric_simd.inl
#include "../simd/geometric.h"
@@ -40,9 +40,7 @@ namespace detail
{
GLM_FUNC_QUALIFIER static float call(vec<3, float, Q> const& a, vec<3, float, Q> const& b)
{
vec<4, float, Q> aa = xyz0(a);
vec<4, float, Q> bb = xyz0(b);
return _mm_cvtss_f32(glm_vec1_dot(aa.data, bb.data));
return _mm_cvtss_f32(glm_vec3_dot(a.data, b.data)); // aligned as vec4
}
};
@@ -68,6 +66,28 @@ namespace detail
}
};
template<>
struct compute_normalize<4, float, aligned_lowp, true>
{
GLM_FUNC_QUALIFIER static vec<4, float, aligned_lowp> call(vec<4, float, aligned_lowp> const& v)
{
vec<4, float, aligned_lowp> Result;
Result.data = glm_vec4_normalize(v.data);
return Result;
}
};
template<>
struct compute_normalize<3, float, aligned_lowp, true>
{
GLM_FUNC_QUALIFIER static vec<3, float, aligned_lowp> call(vec<3, float, aligned_lowp> const& v)
{
vec<3, float, aligned_lowp> Result;
Result.data = glm_vec3_normalize_lowp(v.data);
return Result;
}
};
template<qualifier Q>
struct compute_normalize<4, float, Q, true>
{
@@ -79,6 +99,17 @@ namespace detail
}
};
template<qualifier Q>
struct compute_normalize<3, float, Q, true>
{
GLM_FUNC_QUALIFIER static vec<3, float, Q> call(vec<3, float, Q> const& v)
{
vec<3, float, Q> Result;
Result.data = glm_vec3_normalize(v.data);
return Result;
}
};
template<qualifier Q>
struct compute_faceforward<4, float, Q, true>
{

73
glm/simd/geometric.h
View File

@@ -1,4 +1,4 @@
/// @ref simd
/// @ref simd
/// @file glm/simd/geometric.h
#pragma once
@@ -26,7 +26,7 @@ GLM_FUNC_QUALIFIER glm_vec4 glm_vec4_distance(glm_vec4 p0, glm_vec4 p1)
GLM_FUNC_QUALIFIER glm_vec4 glm_vec4_dot(glm_vec4 v1, glm_vec4 v2)
{
# if GLM_ARCH & GLM_ARCH_AVX_BIT
# if GLM_ARCH & GLM_ARCH_SSE41_BIT
return _mm_dp_ps(v1, v2, 0xff);
# elif GLM_ARCH & GLM_ARCH_SSE3_BIT
glm_vec4 const mul0 = _mm_mul_ps(v1, v2);
@@ -43,9 +43,33 @@ GLM_FUNC_QUALIFIER glm_vec4 glm_vec4_dot(glm_vec4 v1, glm_vec4 v2)
# endif
}
GLM_FUNC_QUALIFIER glm_vec4 glm_vec3_dot(glm_vec4 v1, glm_vec4 v2)
{
# if GLM_ARCH & GLM_ARCH_SSE41_BIT
return _mm_dp_ps(v1, v2, 0x77);
# elif GLM_ARCH & GLM_ARCH_SSE3_BIT
glm_vec4 const constant0 = _mm_castsi128_ps(_mm_setr_epi32(0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0));
glm_vec4 const mul0 = _mm_mul_ps(v1, v2);
glm_vec4 const and0 = _mm_and_ps(mul0, constant0);
glm_vec4 const hadd0 = _mm_hadd_ps(and0, and0);
glm_vec4 const hadd1 = _mm_hadd_ps(hadd0, hadd0);
return hadd1;
# else
glm_vec4 const constant0 = _mm_castsi128_ps(_mm_setr_epi32(0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0));
glm_vec4 const mul0 = _mm_mul_ps(v1, v2);
glm_vec4 const and0 = _mm_and_ps(mul0, constant0);
glm_vec4 const swp0 = _mm_shuffle_ps(and0, and0, _MM_SHUFFLE(2, 3, 0, 1));
glm_vec4 const add0 = _mm_add_ps(and0, swp0);
glm_vec4 const swp1 = _mm_shuffle_ps(add0, add0, _MM_SHUFFLE(0, 1, 2, 3));
glm_vec4 const add1 = _mm_add_ps(add0, swp1);
return add1;
# endif
}
GLM_FUNC_QUALIFIER glm_vec4 glm_vec1_dot(glm_vec4 v1, glm_vec4 v2)
{
# if GLM_ARCH & GLM_ARCH_AVX_BIT
# if GLM_ARCH & GLM_ARCH_SSE41_BIT
return _mm_dp_ps(v1, v2, 0xff);
# elif GLM_ARCH & GLM_ARCH_SSE3_BIT
glm_vec4 const mul0 = _mm_mul_ps(v1, v2);
@@ -74,7 +98,7 @@ GLM_FUNC_QUALIFIER glm_vec4 glm_vec4_cross(glm_vec4 v1, glm_vec4 v2)
return sub0;
}
GLM_FUNC_QUALIFIER glm_vec4 glm_vec4_normalize(glm_vec4 v)
GLM_FUNC_QUALIFIER glm_vec4 glm_vec4_normalize_lowp(glm_vec4 v)
{
glm_vec4 const dot0 = glm_vec4_dot(v, v);
glm_vec4 const isr0 = _mm_rsqrt_ps(dot0);
@@ -82,6 +106,47 @@ GLM_FUNC_QUALIFIER glm_vec4 glm_vec4_normalize(glm_vec4 v)
return mul0;
}
GLM_FUNC_QUALIFIER glm_vec4 glm_vec3_normalize_lowp(glm_vec4 v)
{
glm_vec4 const dot0 = glm_vec3_dot(v, v);
glm_vec4 const isr0 = _mm_rsqrt_ps(dot0);
glm_vec4 const mul0 = _mm_mul_ps(v, isr0);
return mul0;
}
GLM_FUNC_QUALIFIER glm_vec4 glm_vec4_normalize(glm_vec4 v)
{
glm_vec4 const dot0 = glm_vec4_dot(v, v);
glm_vec4 const isr0 = _mm_rsqrt_ps(dot0);
// One iteration of Newton-Raphson method to improve precision
glm_vec4 const mul0 = glm_vec4_mul(dot0, isr0);
glm_vec4 const fma0 = glm_vec4_fma(mul0, isr0, _mm_set1_ps(-3.0f));
glm_vec4 const mul1 = glm_vec4_mul(isr0, _mm_set1_ps(-0.5f));
glm_vec4 const mul2 = glm_vec4_mul(mul1, v);
glm_vec4 const mul3 = glm_vec4_mul(mul2, fma0);
return mul3;
}
GLM_FUNC_QUALIFIER glm_vec4 glm_vec3_normalize(glm_vec4 v)
{
glm_vec4 const dot0 = glm_vec3_dot(v, v);
glm_vec4 const isr0 = _mm_rsqrt_ps(dot0);
// One iteration of Newton-Raphson method to improve precision
glm_vec4 const mul0 = glm_vec4_mul(dot0, isr0);
glm_vec4 const fma0 = glm_vec4_fma(mul0, isr0, _mm_set1_ps(-3.0f));
glm_vec4 const mul1 = glm_vec4_mul(isr0, _mm_set1_ps(-0.5f));
glm_vec4 const mul2 = glm_vec4_mul(mul1, v);
glm_vec4 const mul3 = glm_vec4_mul(mul2, fma0);
return mul3;
}
GLM_FUNC_QUALIFIER glm_vec4 glm_vec4_faceforward(glm_vec4 N, glm_vec4 I, glm_vec4 Nref)
{
glm_vec4 const dot0 = glm_vec4_dot(Nref, I);