|
|
reciRoot = MUL_F32_SIMD(MUL_F32_SIMD(LOAD_SCALAR_F32_SIMD(0.5f), reciRoot), SUB_F32_SIMD(LOAD_SCALAR_F32_SIMD(3.0f), mul));
|
|
reciRoot = MUL_F32_SIMD(MUL_F32_SIMD(LOAD_SCALAR_F32_SIMD(0.5f), reciRoot), SUB_F32_SIMD(LOAD_SCALAR_F32_SIMD(3.0f), mul));
|
|
|
root = _mm_mul_ps(_mm_add_ps(root, _mm_div_ps(this->v, root)), half);
|
|
root = _mm_mul_ps(_mm_add_ps(root, _mm_div_ps(this->v, root)), half);
|
|
|
return F32x4(MUL_F32_SIMD(this->v, this->reciprocalSquareRoot().v));
|
|
return F32x4(MUL_F32_SIMD(this->v, this->reciprocalSquareRoot().v));
|