Lines Matching refs:__m128
33 GLM_FUNC_QUALIFIER __m128 sse_len_ps(__m128 x)
35 __m128 dot0 = sse_dot_ps(x, x);
36 __m128 sqt0 = _mm_sqrt_ps(dot0);
41 GLM_FUNC_QUALIFIER __m128 sse_dst_ps(__m128 p0, __m128 p1)
43 __m128 sub0 = _mm_sub_ps(p0, p1);
44 __m128 len0 = sse_len_ps(sub0);
49 GLM_FUNC_QUALIFIER __m128 sse_dot_ps(__m128 v1, __m128 v2)
54 __m128 mul0 = _mm_mul_ps(v1, v2);
55 __m128 swp0 = _mm_shuffle_ps(mul0, mul0, _MM_SHUFFLE(2, 3, 0, 1));
56 __m128 add0 = _mm_add_ps(mul0, swp0);
57 __m128 swp1 = _mm_shuffle_ps(add0, add0, _MM_SHUFFLE(0, 1, 2, 3));
58 __m128 add1 = _mm_add_ps(add0, swp1);
64 GLM_FUNC_QUALIFIER __m128 sse_dot_ss(__m128 v1, __m128 v2)
66 __m128 mul0 = _mm_mul_ps(v1, v2);
67 __m128 mov0 = _mm_movehl_ps(mul0, mul0);
68 __m128 add0 = _mm_add_ps(mov0, mul0);
69 __m128 swp1 = _mm_shuffle_ps(add0, add0, 1);
70 __m128 add1 = _mm_add_ss(add0, swp1);
75 GLM_FUNC_QUALIFIER __m128 sse_xpd_ps(__m128 v1, __m128 v2)
77 __m128 swp0 = _mm_shuffle_ps(v1, v1, _MM_SHUFFLE(3, 0, 2, 1));
78 __m128 swp1 = _mm_shuffle_ps(v1, v1, _MM_SHUFFLE(3, 1, 0, 2));
79 __m128 swp2 = _mm_shuffle_ps(v2, v2, _MM_SHUFFLE(3, 0, 2, 1));
80 __m128 swp3 = _mm_shuffle_ps(v2, v2, _MM_SHUFFLE(3, 1, 0, 2));
81 __m128 mul0 = _mm_mul_ps(swp0, swp3);
82 __m128 mul1 = _mm_mul_ps(swp1, swp2);
83 __m128 sub0 = _mm_sub_ps(mul0, mul1);
88 GLM_FUNC_QUALIFIER __m128 sse_nrm_ps(__m128 v)
90 __m128 dot0 = sse_dot_ps(v, v);
91 __m128 isr0 = _mm_rsqrt_ps(dot0);
92 __m128 mul0 = _mm_mul_ps(v, isr0);
97 GLM_FUNC_QUALIFIER __m128 sse_ffd_ps(__m128 N, __m128 I, __m128 Nref)
99 //__m128 dot0 = _mm_dot_ps(v, v);
100 //__m128 neg0 = _mm_neg_ps(N);
101 //__m128 sgn0 = _mm_sgn_ps(dot0);
102 //__m128 mix0 = _mm_mix_ps(N, neg0, sgn0);
105 __m128 dot0 = sse_dot_ps(Nref, I);
106 __m128 sgn0 = sse_sgn_ps(dot0);
107 __m128 mul0 = _mm_mul_ps(sgn0, glm::detail::minus_one);
108 __m128 mul1 = _mm_mul_ps(N, mul0);
113 GLM_FUNC_QUALIFIER __m128 sse_rfe_ps(__m128 I, __m128 N)
115 __m128 dot0 = sse_dot_ps(N, I);
116 __m128 mul0 = _mm_mul_ps(N, dot0);
117 __m128 mul1 = _mm_mul_ps(mul0, glm::detail::two);
118 __m128 sub0 = _mm_sub_ps(I, mul1);
123 GLM_FUNC_QUALIFIER __m128 sse_rfa_ps(__m128 I, __m128 N, __m128 eta)
125 __m128 dot0 = sse_dot_ps(N, I);
126 __m128 mul0 = _mm_mul_ps(eta, eta);
127 __m128 mul1 = _mm_mul_ps(dot0, dot0);
128 __m128 sub0 = _mm_sub_ps(glm::detail::one, mul0);
129 __m128 sub1 = _mm_sub_ps(glm::detail::one, mul1);
130 __m128 mul2 = _mm_mul_ps(sub0, sub1);
135 __m128 sqt0 = _mm_sqrt_ps(mul2);
136 __m128 mul3 = _mm_mul_ps(eta, dot0);
137 __m128 add0 = _mm_add_ps(mul3, sqt0);
138 __m128 mul4 = _mm_mul_ps(add0, N);
139 __m128 mul5 = _mm_mul_ps(eta, I);
140 __m128 sub2 = _mm_sub_ps(mul5, mul4);