From a8b3118e8305fd1c668ea25e07157b625c9747ff Mon Sep 17 00:00:00 2001 From: Joursoir Date: Sat, 10 Apr 2021 17:29:02 +0000 Subject: add glm headers --- src/include/glm/detail/type_quat_simd.inl | 188 ++++++++++++++++++++++++++++++ 1 file changed, 188 insertions(+) create mode 100644 src/include/glm/detail/type_quat_simd.inl (limited to 'src/include/glm/detail/type_quat_simd.inl') diff --git a/src/include/glm/detail/type_quat_simd.inl b/src/include/glm/detail/type_quat_simd.inl new file mode 100644 index 0000000..2d4e68c --- /dev/null +++ b/src/include/glm/detail/type_quat_simd.inl @@ -0,0 +1,188 @@ +/// @ref core + +#if GLM_ARCH & GLM_ARCH_SSE2_BIT + +namespace glm{ +namespace detail +{ +/* + template + struct compute_quat_mul + { + static qua call(qua const& q1, qua 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)); + + qua Result; + _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 + struct compute_quat_add + { + static qua call(qua const& q, qua const& p) + { + qua Result; + Result.data = _mm_add_ps(q.data, p.data); + return Result; + } + }; + +# if GLM_ARCH & GLM_ARCH_AVX_BIT + template + struct compute_quat_add + { + static qua call(qua const& a, qua const& b) + { + qua Result; + Result.data = _mm256_add_pd(a.data, b.data); + return Result; + } + }; +# endif + + template + struct compute_quat_sub + { + static qua call(qua const& q, qua const& p) + { + vec<4, float, Q> Result; + Result.data = _mm_sub_ps(q.data, p.data); + return Result; + } + }; + +# if GLM_ARCH & GLM_ARCH_AVX_BIT + template + struct compute_quat_sub + { + static qua call(qua const& a, qua const& b) + { + qua Result; + Result.data = _mm256_sub_pd(a.data, b.data); + return Result; + } + }; +# endif + + template + struct compute_quat_mul_scalar + { + static qua call(qua const& q, float s) + { + vec<4, float, Q> Result; + Result.data = _mm_mul_ps(q.data, _mm_set_ps1(s)); + return Result; + } + }; + +# if GLM_ARCH & GLM_ARCH_AVX_BIT + template + struct compute_quat_mul_scalar + { + static qua call(qua const& q, double s) + { + qua Result; + Result.data = _mm256_mul_pd(q.data, _mm_set_ps1(s)); + return Result; + } + }; +# endif + + template + struct compute_quat_div_scalar + { + static qua call(qua const& q, float s) + { + vec<4, float, Q> Result; + Result.data = _mm_div_ps(q.data, _mm_set_ps1(s)); + return Result; + } + }; + +# if GLM_ARCH & GLM_ARCH_AVX_BIT + template + struct compute_quat_div_scalar + { + static qua call(qua const& q, double s) + { + qua Result; + Result.data = _mm256_div_pd(q.data, _mm_set_ps1(s)); + return Result; + } + }; +# endif + + template + struct compute_quat_mul_vec4 + { + static vec<4, float, Q> call(qua const& q, vec<4, float, Q> 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); + + vec<4, float, Q> Result; + 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 + -- cgit v1.2.3-18-g5258