Fix quat packing XYZW usage (#1204)

This commit is contained in:
Stéphane GINIER 2023-12-30 09:32:24 +01:00 committed by GitHub
parent 5ce98b7514
commit 8d337c0c65
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
4 changed files with 28 additions and 24 deletions

View File

@ -89,7 +89,7 @@ namespace glm
GLM_FUNC_DECL GLM_CONSTEXPR qua(T s, vec<3, T, Q> const& v); GLM_FUNC_DECL GLM_CONSTEXPR qua(T s, vec<3, T, Q> const& v);
# ifdef GLM_FORCE_QUAT_DATA_XYZW # ifdef GLM_FORCE_QUAT_CTOR_XYZW
GLM_FUNC_DECL GLM_CONSTEXPR qua(T x, T y, T z, T w); GLM_FUNC_DECL GLM_CONSTEXPR qua(T x, T y, T z, T w);
# else # else
GLM_FUNC_DECL GLM_CONSTEXPR qua(T w, T x, T y, T z); GLM_FUNC_DECL GLM_CONSTEXPR qua(T w, T x, T y, T z);

View File

@ -141,7 +141,11 @@ namespace detail
{} {}
template <typename T, qualifier Q> template <typename T, qualifier Q>
# ifdef GLM_FORCE_QUAT_CTOR_XYZW
GLM_FUNC_QUALIFIER GLM_CONSTEXPR qua<T, Q>::qua(T _x, T _y, T _z, T _w)
# else
GLM_FUNC_QUALIFIER GLM_CONSTEXPR qua<T, Q>::qua(T _w, T _x, T _y, T _z) GLM_FUNC_QUALIFIER GLM_CONSTEXPR qua<T, Q>::qua(T _w, T _x, T _y, T _z)
# endif
# ifdef GLM_FORCE_QUAT_DATA_WXYZ # ifdef GLM_FORCE_QUAT_DATA_WXYZ
: w(_w), x(_x), y(_y), z(_z) : w(_w), x(_x), y(_y), z(_z)
# else # else

View File

@ -161,26 +161,7 @@ namespace detail
{ {
static vec<4, float, Q> call(qua<float, Q> const& q, vec<4, float, Q> const& v) static vec<4, float, Q> call(qua<float, Q> const& q, vec<4, float, Q> const& v)
{ {
# ifdef GLM_FORCE_QUAT_DATA_XYZW # ifdef GLM_FORCE_QUAT_DATA_WXYZ
__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);
vec<4, float, Q> Result;
Result.data = _mm_add_ps(v.data, _mm_add_ps(uv, uuv));
return Result;
# else
__m128 const q_wwww = _mm_shuffle_ps(q.data, q.data, _MM_SHUFFLE(0, 0, 0, 0)); __m128 const q_wwww = _mm_shuffle_ps(q.data, q.data, _MM_SHUFFLE(0, 0, 0, 0));
__m128 const q_swp0 = _mm_shuffle_ps(q.data, q.data, _MM_SHUFFLE(0, 1, 3, 2)); __m128 const q_swp0 = _mm_shuffle_ps(q.data, q.data, _MM_SHUFFLE(0, 1, 3, 2));
__m128 const q_swp1 = _mm_shuffle_ps(q.data, q.data, _MM_SHUFFLE(0, 2, 1, 3)); __m128 const q_swp1 = _mm_shuffle_ps(q.data, q.data, _MM_SHUFFLE(0, 2, 1, 3));
@ -196,6 +177,25 @@ namespace detail
uv = _mm_mul_ps(uv, _mm_mul_ps(q_wwww, two)); uv = _mm_mul_ps(uv, _mm_mul_ps(q_wwww, two));
uuv = _mm_mul_ps(uuv, two); uuv = _mm_mul_ps(uuv, two);
vec<4, float, Q> Result;
Result.data = _mm_add_ps(v.data, _mm_add_ps(uv, uuv));
return Result;
# else
__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);
vec<4, float, Q> Result; vec<4, float, Q> Result;
Result.data = _mm_add_ps(v.data, _mm_add_ps(uv, uuv)); Result.data = _mm_add_ps(v.data, _mm_add_ps(uv, uuv));
return Result; return Result;

View File

@ -173,10 +173,10 @@ namespace detail
j = Next[i]; j = Next[i];
k = Next[j]; k = Next[j];
# ifdef GLM_FORCE_QUAT_DATA_XYZW # ifdef GLM_FORCE_QUAT_DATA_WXYZ
int off = 0;
# else
int off = 1; int off = 1;
# else
int off = 0;
# endif # endif
root = sqrt(Row[i][i] - Row[j][j] - Row[k][k] + static_cast<T>(1.0)); root = sqrt(Row[i][i] - Row[j][j] - Row[k][k] + static_cast<T>(1.0));