v4sf DisplayFunctionGGBA::inv_display( v4sf L )
{
const v4sf voffset = _mm_set1_ps(L_offset);
const v4sf vmax = _mm_set1_ps(L_max);
L = _mm_max_ps(L, voffset);
L = _mm_min_ps(L, voffset+vmax);
return _mm_pow_ps((L - voffset)/(vmax-_mm_set1_ps(L_black)), _mm_set1_ps(1.0f/gamma));
}
void pow_fmath(const Mat& src, const float a, Mat & dest)
{
if (dest.empty())dest.create(src.size(), CV_32F);
int width = src.cols;
int height = src.rows;
int size = src.size().area();
int i = 0;
const float* s = src.ptr<float>(0);
float* d = dest.ptr<float>(0);
const __m128 ma = _mm_set1_ps(a);
for (i = 0; i <= size - 4; i += 4)
{
_mm_store_ps(d + i, _mm_pow_ps(_mm_load_ps(s + i), ma));
}
for (; i < size; i++)
{
d[i] = cv::pow(s[i], a);
}
}
void pow_fmath(const Mat& src1, const Mat& src2, Mat & dest)
{
if (dest.empty())dest.create(src1.size(), CV_32F);
int width = src1.cols;
int height = src1.rows;
int size = src1.size().area();
int i = 0;
const float* s1 = src1.ptr<float>(0);
const float* s2 = src2.ptr<float>(0);
float* d = dest.ptr<float>(0);
for (i = 0; i <= size - 4; i += 4)
{
_mm_store_ps(d + i, _mm_pow_ps(_mm_load_ps(s1 + i), _mm_load_ps(s2 + i)));
}
for (; i < size; i++)
{
d[i] = cv::pow(s1[i], s2[i]);
}
}
v4sf DisplayFunctionGGBA::display( v4sf pix )
{
return _mm_pow_ps(pix, _mm_set1_ps(gamma)) * (_mm_set1_ps(L_max)-_mm_set1_ps(L_black)) + _mm_set1_ps(L_offset);
}
