template <bool align> void LaplaceAbsSum(const uint8_t * src, size_t stride, size_t width, size_t height, uint64_t * sum)
{
assert(width > A);
if(align)
assert(Aligned(src) && Aligned(stride));
size_t bodyWidth = Simd::AlignHi(width, A) - A;
const uint8_t *src0, *src1, *src2;
v128_u8 a[3][3];
v128_u8 tailMask = ShiftLeft(K8_FF, A - width + bodyWidth);
*sum = 0;
for(size_t row = 0; row < height; ++row)
{
src0 = src + stride*(row - 1);
src1 = src0 + stride;
src2 = src1 + stride;
if(row == 0)
src0 = src1;
if(row == height - 1)
src2 = src1;
v128_u32 sums[2] = {K32_00000000, K32_00000000};
LoadNose3<align, 1>(src0 + 0, a[0]);
LoadNose3<align, 1>(src1 + 0, a[1]);
LoadNose3<align, 1>(src2 + 0, a[2]);
LaplaceAbsSum(a, sums);
for(size_t col = A; col < bodyWidth; col += A)
{
LoadBody3<align, 1>(src0 + col, a[0]);
LoadBody3<align, 1>(src1 + col, a[1]);
LoadBody3<align, 1>(src2 + col, a[2]);
LaplaceAbsSum(a, sums);
}
LoadTail3<false, 1>(src0 + width - A, a[0]);
LoadTail3<false, 1>(src1 + width - A, a[1]);
LoadTail3<false, 1>(src2 + width - A, a[2]);
SetMask3x3(a, tailMask);
LaplaceAbsSum(a, sums);
*sum += ExtractSum(vec_add(sums[0], sums[1]));
}
}
void SobelDxAbsSum(const uint8_t * src, size_t stride, size_t width, size_t height, uint64_t * sum)
{
assert(width > A);
size_t bodyWidth = Simd::AlignHi(width, A) - A;
const uint8_t *src0, *src1, *src2;
v16u8 a[3][3];
v2u64 fullSum = Zero<v2u64>();
const v16u8 K8_FF = Fill((uint8_t)0xff);
v16u8 tailMask = ShiftLeft(K8_FF, A - width + bodyWidth);
for (size_t row = 0; row < height; ++row)
{
src0 = src + stride*(row - 1);
src1 = src0 + stride;
src2 = src1 + stride;
if (row == 0)
src0 = src1;
if (row == height - 1)
src2 = src1;
v4u32 rowSum = Zero<v4u32>();
LoadNoseDx(src0 + 0, a[0]);
LoadNoseDx(src1 + 0, a[1]);
LoadNoseDx(src2 + 0, a[2]);
SobelDxAbsSum(a, rowSum);
for (size_t col = A; col < bodyWidth; col += A)
{
LoadBodyDx(src0 + col, a[0]);
LoadBodyDx(src1 + col, a[1]);
LoadBodyDx(src2 + col, a[2]);
SobelDxAbsSum(a, rowSum);
}
LoadTailDx(src0 + width - A, a[0]);
LoadTailDx(src1 + width - A, a[1]);
LoadTailDx(src2 + width - A, a[2]);
SetMask3x3(a, tailMask);
SobelDxAbsSum(a, rowSum);
fullSum = PadSum(fullSum,rowSum);
}
*sum = ExtractSum(fullSum);
}
