boost::optional<double> SimpleClean::FindPeakAVX(const double *image, size_t width, size_t height, size_t& x, size_t& y, size_t startY, size_t endY, size_t horizontalBorder, size_t verticalBorder)
{
double peakMax = std::numeric_limits<double>::min();
size_t peakIndex = 0;
__m256d mPeakMax = _mm256_set1_pd(peakMax);
size_t xiStart = horizontalBorder, xiEnd = width - horizontalBorder;
size_t yiStart = std::max(startY, verticalBorder), yiEnd = std::min(endY, height - verticalBorder);
if(xiEnd < xiStart) xiEnd = xiStart;
if(yiEnd < yiStart) yiEnd = yiStart;
for(size_t yi=yiStart; yi!=yiEnd; ++yi)
{
size_t index = yi*width + xiStart;
const double* const endPtr = image + yi*width + xiEnd - 4;
const double *i=image + index;
for(; i<endPtr; i+=4)
{
__m256d val = _mm256_loadu_pd(i);
if(AllowNegativeComponent) {
__m256d negVal = _mm256_sub_pd(_mm256_set1_pd(0.0), val);
val = _mm256_max_pd(val, negVal);
}
int mask = _mm256_movemask_pd(_mm256_cmp_pd(val, mPeakMax, _CMP_GT_OQ));
if(mask != 0)
{
for(size_t di=0; di!=4; ++di)
{
double value = i[di];
if(AllowNegativeComponent) value = std::fabs(value);
if(value > peakMax)
{
peakIndex = index+di;
peakMax = std::fabs(i[di]);
mPeakMax = _mm256_set1_pd(peakMax);
}
}
}
index+=4;
}
for(; i!=endPtr+4; ++i)
{
double value = *i;
if(AllowNegativeComponent) value = std::fabs(value);
if(value > peakMax)
{
peakIndex = index;
peakMax = std::fabs(*i);
}
++index;
}
}
x = peakIndex % width;
y = peakIndex / width;
return image[x + y*width];
}
void static
avx_test (void)
{
int i;
int d;
union256d s1;
double source[4] = {-45, -3, -34.56, 35};
int e = 0;
s1.x = _mm256_loadu_pd (source);
d = _mm256_movemask_pd (s1.x);
for (i = 0; i < 4; i++)
if (source[i] < 0)
e |= (1 << i);
if (checkVi (&d, &e, 1))
abort ();
}
