35     __m128 dot0 = sse_dot_ps(x, x);  local
36 	__m128 sqt0 = _mm_sqrt_ps(dot0);
90 	__m128 dot0 = sse_dot_ps(v, v);  local
91 	__m128 isr0 = _mm_rsqrt_ps(dot0);
99 	//__m128 dot0 = _mm_dot_ps(v, v);
101 	//__m128 sgn0 = _mm_sgn_ps(dot0);
105 	__m128 dot0 = sse_dot_ps(Nref, I);  local
106 	__m128 sgn0 = sse_sgn_ps(dot0);
115 	__m128 dot0 = sse_dot_ps(N, I);  local
116 	__m128 mul0 = _mm_mul_ps(N, dot0);
125 	__m128 dot0 = sse_dot_ps(N, I);  local
127 	__m128 mul1 = _mm_mul_ps(dot0, dot0);
136 	__m128 mul3 = _mm_mul_ps(eta, dot0);