int operator () (const schar * src0, const uchar * mask, int * dst, int len, int cn) const
{
if (mask || (cn != 1 && cn != 2 && cn != 4))
return 0;
len *= cn;
int x = 0;
v_int32 v_sum = vx_setzero_s32();
int len0 = len & -v_int8::nlanes;
while (x < len0)
{
const int len_tmp = min(x + 256*v_int16::nlanes, len0);
v_int16 v_sum16 = vx_setzero_s16();
for (; x < len_tmp; x += v_int8::nlanes)
{
v_int16 v_src0, v_src1;
v_expand(vx_load(src0 + x), v_src0, v_src1);
v_sum16 += v_src0 + v_src1;
}
v_int32 v_half0, v_half1;
v_expand(v_sum16, v_half0, v_half1);
v_sum += v_half0 + v_half1;
}
if (x <= len - v_int16::nlanes)
{
v_int32 v_half0, v_half1;
v_expand(vx_load_expand(src0 + x), v_half0, v_half1);
v_sum += v_half0 + v_half1;
x += v_int16::nlanes;
}
if (x <= len - v_int32::nlanes)
{
v_sum += vx_load_expand_q(src0 + x);
x += v_int32::nlanes;
}
if (cn == 1)
*dst += v_reduce_sum(v_sum);
else
{
int32_t CV_DECL_ALIGNED(CV_SIMD_WIDTH) ar[v_int32::nlanes];
v_store_aligned(ar, v_sum);
for (int i = 0; i < v_int32::nlanes; ++i)
dst[i % cn] += ar[i];
}
v_cleanup();
return x / cn;
}
int operator () (const short * src0, const uchar * mask, int * dst, int len, int cn) const
{
if (mask || (cn != 1 && cn != 2 && cn != 4))
return 0;
len *= cn;
int x = 0;
v_int32 v_sum = vx_setzero_s32();
for (; x <= len - v_int16::nlanes; x += v_int16::nlanes)
{
v_int32 v_src0, v_src1;
v_expand(vx_load(src0 + x), v_src0, v_src1);
v_sum += v_src0 + v_src1;
}
if (x <= len - v_int32::nlanes)
{
v_sum += vx_load_expand(src0 + x);
x += v_int32::nlanes;
}
if (cn == 1)
*dst += v_reduce_sum(v_sum);
else
{
int32_t CV_DECL_ALIGNED(CV_SIMD_WIDTH) ar[v_int32::nlanes];
v_store_aligned(ar, v_sum);
for (int i = 0; i < v_int32::nlanes; ++i)
dst[i % cn] += ar[i];
}
v_cleanup();
return x / cn;
}
// v_expand and v_load_expand
TheTest & test_expand()
{
typedef typename V_RegTrait128<LaneType>::w_reg Rx2;
Data<R> dataA;
R a = dataA;
Data<Rx2> resB = v_load_expand(dataA.d);
Rx2 c, d;
v_expand(a, c, d);
Data<Rx2> resC = c, resD = d;
const int n = Rx2::nlanes;
for (int i = 0; i < n; ++i)
{
EXPECT_EQ(dataA[i], resB[i]);
EXPECT_EQ(dataA[i], resC[i]);
EXPECT_EQ(dataA[i + n], resD[i]);
}
return *this;
}
void exp64f( const double *_x, double *y, int n )
{
CV_INSTRUMENT_REGION();
const double* const expTab = cv::details::getExpTab64f();
const double
A5 = .99999999999999999998285227504999 / EXPPOLY_32F_A0,
A4 = .69314718055994546743029643825322 / EXPPOLY_32F_A0,
A3 = .24022650695886477918181338054308 / EXPPOLY_32F_A0,
A2 = .55504108793649567998466049042729e-1 / EXPPOLY_32F_A0,
A1 = .96180973140732918010002372686186e-2 / EXPPOLY_32F_A0,
A0 = .13369713757180123244806654839424e-2 / EXPPOLY_32F_A0;
int i = 0;
const Cv64suf* x = (const Cv64suf*)_x;
double minval = (-exp_max_val/exp_prescale);
double maxval = (exp_max_val/exp_prescale);
#if CV_SIMD_64F
const int VECSZ = v_float64::nlanes;
const v_float64 vprescale = vx_setall_f64(exp_prescale);
const v_float64 vpostscale = vx_setall_f64(exp_postscale);
const v_float64 vminval = vx_setall_f64(minval);
const v_float64 vmaxval = vx_setall_f64(maxval);
const v_float64 vA1 = vx_setall_f64(A1);
const v_float64 vA2 = vx_setall_f64(A2);
const v_float64 vA3 = vx_setall_f64(A3);
const v_float64 vA4 = vx_setall_f64(A4);
const v_float64 vA5 = vx_setall_f64(A5);
const v_int32 vidxmask = vx_setall_s32(EXPTAB_MASK);
bool y_aligned = (size_t)(void*)y % 32 == 0;
for( ; i < n; i += VECSZ*2 )
{
if( i + VECSZ*2 > n )
{
if( i == 0 || _x == y )
break;
i = n - VECSZ*2;
y_aligned = false;
}
v_float64 xf0 = vx_load(&x[i].f), xf1 = vx_load(&x[i + VECSZ].f);
xf0 = v_min(v_max(xf0, vminval), vmaxval);
xf1 = v_min(v_max(xf1, vminval), vmaxval);
xf0 *= vprescale;
xf1 *= vprescale;
v_int32 xi0 = v_round(xf0);
v_int32 xi1 = v_round(xf1);
xf0 = (xf0 - v_cvt_f64(xi0))*vpostscale;
xf1 = (xf1 - v_cvt_f64(xi1))*vpostscale;
v_float64 yf0 = v_lut(expTab, xi0 & vidxmask);
v_float64 yf1 = v_lut(expTab, xi1 & vidxmask);
v_int32 v0 = vx_setzero_s32(), v1023 = vx_setall_s32(1023), v2047 = vx_setall_s32(2047);
xi0 = v_min(v_max(v_shr<EXPTAB_SCALE>(xi0) + v1023, v0), v2047);
xi1 = v_min(v_max(v_shr<EXPTAB_SCALE>(xi1) + v1023, v0), v2047);
v_int64 xq0, xq1, dummy;
v_expand(xi0, xq0, dummy);
v_expand(xi1, xq1, dummy);
yf0 *= v_reinterpret_as_f64(v_shl<52>(xq0));
yf1 *= v_reinterpret_as_f64(v_shl<52>(xq1));
v_float64 zf0 = xf0 + vA1;
v_float64 zf1 = xf1 + vA1;
zf0 = v_fma(zf0, xf0, vA2);
zf1 = v_fma(zf1, xf1, vA2);
zf0 = v_fma(zf0, xf0, vA3);
zf1 = v_fma(zf1, xf1, vA3);
zf0 = v_fma(zf0, xf0, vA4);
zf1 = v_fma(zf1, xf1, vA4);
zf0 = v_fma(zf0, xf0, vA5);
zf1 = v_fma(zf1, xf1, vA5);
zf0 *= yf0;
zf1 *= yf1;
if( y_aligned )
{
v_store_aligned(y + i, zf0);
v_store_aligned(y + i + VECSZ, zf1);
}
else
{
v_store(y + i, zf0);
v_store(y + i + VECSZ, zf1);
}
}
vx_cleanup();
#endif
for( ; i < n; i++ )
{
double x0 = x[i].f;
x0 = std::min(std::max(x0, minval), maxval);
x0 *= exp_prescale;
Cv64suf buf;
int xi = saturate_cast<int>(x0);
x0 = (x0 - xi)*exp_postscale;
int t = (xi >> EXPTAB_SCALE) + 1023;
t = !(t & ~2047) ? t : t < 0 ? 0 : 2047;
buf.i = (int64)t << 52;
y[i] = buf.f * expTab[xi & EXPTAB_MASK] * (((((A0*x0 + A1)*x0 + A2)*x0 + A3)*x0 + A4)*x0 + A5);
}
}
void operator()(const Range &boundaries) const
{
CV_TRACE_FUNCTION();
Mat dx, dy;
AutoBuffer<short> dxMax(0), dyMax(0);
std::deque<uchar*> stack, borderPeaksLocal;
const int rowStart = max(0, boundaries.start - 1), rowEnd = min(src.rows, boundaries.end + 1);
int *_mag_p, *_mag_a, *_mag_n;
short *_dx, *_dy, *_dx_a = NULL, *_dy_a = NULL, *_dx_n = NULL, *_dy_n = NULL;
uchar *_pmap;
double scale = 1.0;
CV_TRACE_REGION("gradient")
if(needGradient)
{
if (aperture_size == 7)
{
scale = 1 / 16.0;
}
Sobel(src.rowRange(rowStart, rowEnd), dx, CV_16S, 1, 0, aperture_size, scale, 0, BORDER_REPLICATE);
Sobel(src.rowRange(rowStart, rowEnd), dy, CV_16S, 0, 1, aperture_size, scale, 0, BORDER_REPLICATE);
}
else
{
dx = src.rowRange(rowStart, rowEnd);
dy = src2.rowRange(rowStart, rowEnd);
}
CV_TRACE_REGION_NEXT("magnitude");
if(cn > 1)
{
dxMax.allocate(2 * dx.cols);
dyMax.allocate(2 * dy.cols);
_dx_a = (short*)dxMax;
_dx_n = _dx_a + dx.cols;
_dy_a = (short*)dyMax;
_dy_n = _dy_a + dy.cols;
}
// _mag_p: previous row, _mag_a: actual row, _mag_n: next row
#if CV_SIMD128
AutoBuffer<int> buffer(3 * (mapstep * cn + CV_MALLOC_SIMD128));
_mag_p = alignPtr((int*)buffer + 1, CV_MALLOC_SIMD128);
_mag_a = alignPtr(_mag_p + mapstep * cn, CV_MALLOC_SIMD128);
_mag_n = alignPtr(_mag_a + mapstep * cn, CV_MALLOC_SIMD128);
#else
AutoBuffer<int> buffer(3 * (mapstep * cn));
_mag_p = (int*)buffer + 1;
_mag_a = _mag_p + mapstep * cn;
_mag_n = _mag_a + mapstep * cn;
#endif
// For the first time when just 2 rows are filled and for left and right borders
if(rowStart == boundaries.start)
memset(_mag_n - 1, 0, mapstep * sizeof(int));
else
_mag_n[src.cols] = _mag_n[-1] = 0;
_mag_a[src.cols] = _mag_a[-1] = _mag_p[src.cols] = _mag_p[-1] = 0;
// calculate magnitude and angle of gradient, perform non-maxima suppression.
// fill the map with one of the following values:
// 0 - the pixel might belong to an edge
// 1 - the pixel can not belong to an edge
// 2 - the pixel does belong to an edge
for (int i = rowStart; i <= boundaries.end; ++i)
{
// Scroll the ring buffer
std::swap(_mag_n, _mag_a);
std::swap(_mag_n, _mag_p);
if(i < rowEnd)
{
// Next row calculation
_dx = dx.ptr<short>(i - rowStart);
_dy = dy.ptr<short>(i - rowStart);
if (L2gradient)
{
int j = 0, width = src.cols * cn;
#if CV_SIMD128
if (haveSIMD)
{
for ( ; j <= width - 8; j += 8)
{
v_int16x8 v_dx = v_load((const short*)(_dx + j));
v_int16x8 v_dy = v_load((const short*)(_dy + j));
v_int32x4 v_dxp_low, v_dxp_high;
v_int32x4 v_dyp_low, v_dyp_high;
v_expand(v_dx, v_dxp_low, v_dxp_high);
v_expand(v_dy, v_dyp_low, v_dyp_high);
v_store_aligned((int *)(_mag_n + j), v_dxp_low*v_dxp_low+v_dyp_low*v_dyp_low);
v_store_aligned((int *)(_mag_n + j + 4), v_dxp_high*v_dxp_high+v_dyp_high*v_dyp_high);
}
}
#endif
for ( ; j < width; ++j)
_mag_n[j] = int(_dx[j])*_dx[j] + int(_dy[j])*_dy[j];
}
else
{
int j = 0, width = src.cols * cn;
#if CV_SIMD128
if (haveSIMD)
{
for(; j <= width - 8; j += 8)
{
v_int16x8 v_dx = v_load((const short *)(_dx + j));
v_int16x8 v_dy = v_load((const short *)(_dy + j));
v_dx = v_reinterpret_as_s16(v_abs(v_dx));
v_dy = v_reinterpret_as_s16(v_abs(v_dy));
v_int32x4 v_dx_ml, v_dy_ml, v_dx_mh, v_dy_mh;
v_expand(v_dx, v_dx_ml, v_dx_mh);
v_expand(v_dy, v_dy_ml, v_dy_mh);
v_store_aligned((int *)(_mag_n + j), v_dx_ml + v_dy_ml);
v_store_aligned((int *)(_mag_n + j + 4), v_dx_mh + v_dy_mh);
}
}
#endif
for ( ; j < width; ++j)
_mag_n[j] = std::abs(int(_dx[j])) + std::abs(int(_dy[j]));
}
if(cn > 1)
{
std::swap(_dx_n, _dx_a);
std::swap(_dy_n, _dy_a);
for(int j = 0, jn = 0; j < src.cols; ++j, jn += cn)
{
int maxIdx = jn;
for(int k = 1; k < cn; ++k)
if(_mag_n[jn + k] > _mag_n[maxIdx]) maxIdx = jn + k;
_mag_n[j] = _mag_n[maxIdx];
_dx_n[j] = _dx[maxIdx];
_dy_n[j] = _dy[maxIdx];
}
_mag_n[src.cols] = 0;
}
// at the very beginning we do not have a complete ring
// buffer of 3 magnitude rows for non-maxima suppression
if (i <= boundaries.start)
continue;
}
else
{
memset(_mag_n - 1, 0, mapstep * sizeof(int));
if(cn > 1)
{
std::swap(_dx_n, _dx_a);
std::swap(_dy_n, _dy_a);
}
}
// From here actual src row is (i - 1)
// Set left and right border to 1
#if CV_SIMD128
if(haveSIMD)
_pmap = map.ptr<uchar>(i) + CV_MALLOC_SIMD128;
else
#endif
_pmap = map.ptr<uchar>(i) + 1;
_pmap[src.cols] =_pmap[-1] = 1;
if(cn == 1)
{
_dx = dx.ptr<short>(i - rowStart - 1);
_dy = dy.ptr<short>(i - rowStart - 1);
}
else
{
_dx = _dx_a;
_dy = _dy_a;
}
const int TG22 = 13573;
int j = 0;
#if CV_SIMD128
if (haveSIMD)
{
const v_int32x4 v_low = v_setall_s32(low);
const v_int8x16 v_one = v_setall_s8(1);
for (; j <= src.cols - 32; j += 32)
{
v_int32x4 v_m1 = v_load_aligned((const int*)(_mag_a + j));
v_int32x4 v_m2 = v_load_aligned((const int*)(_mag_a + j + 4));
v_int32x4 v_m3 = v_load_aligned((const int*)(_mag_a + j + 8));
v_int32x4 v_m4 = v_load_aligned((const int*)(_mag_a + j + 12));
v_int32x4 v_cmp1 = v_m1 > v_low;
v_int32x4 v_cmp2 = v_m2 > v_low;
v_int32x4 v_cmp3 = v_m3 > v_low;
v_int32x4 v_cmp4 = v_m4 > v_low;
v_m1 = v_load_aligned((const int*)(_mag_a + j + 16));
v_m2 = v_load_aligned((const int*)(_mag_a + j + 20));
v_m3 = v_load_aligned((const int*)(_mag_a + j + 24));
v_m4 = v_load_aligned((const int*)(_mag_a + j + 28));
v_store_aligned((signed char*)(_pmap + j), v_one);
v_store_aligned((signed char*)(_pmap + j + 16), v_one);
v_int16x8 v_cmp80 = v_pack(v_cmp1, v_cmp2);
v_int16x8 v_cmp81 = v_pack(v_cmp3, v_cmp4);
v_cmp1 = v_m1 > v_low;
v_cmp2 = v_m2 > v_low;
v_cmp3 = v_m3 > v_low;
v_cmp4 = v_m4 > v_low;
v_int8x16 v_cmp = v_pack(v_cmp80, v_cmp81);
v_cmp80 = v_pack(v_cmp1, v_cmp2);
v_cmp81 = v_pack(v_cmp3, v_cmp4);
unsigned int mask = v_signmask(v_cmp);
v_cmp = v_pack(v_cmp80, v_cmp81);
mask |= v_signmask(v_cmp) << 16;
if (mask)
{
int k = j;
do
{
int l = trailingZeros32(mask);
k += l;
mask >>= l;
int m = _mag_a[k];
short xs = _dx[k];
short ys = _dy[k];
int x = (int)std::abs(xs);
int y = (int)std::abs(ys) << 15;
int tg22x = x * TG22;
if (y < tg22x)
{
if (m > _mag_a[k - 1] && m >= _mag_a[k + 1])
{
CANNY_CHECK_SIMD(m, high, (_pmap+k), stack);
}
}
else
{
int tg67x = tg22x + (x << 16);
if (y > tg67x)
{
if (m > _mag_p[k] && m >= _mag_n[k])
{
CANNY_CHECK_SIMD(m, high, (_pmap+k), stack);
}
}
else
{
int s = (xs ^ ys) < 0 ? -1 : 1;
if(m > _mag_p[k - s] && m > _mag_n[k + s])
{
CANNY_CHECK_SIMD(m, high, (_pmap+k), stack);
}
}
}
++k;
} while((mask >>= 1));
}
}
if (j <= src.cols - 16)
{
v_int32x4 v_m1 = v_load_aligned((const int*)(_mag_a + j));
v_int32x4 v_m2 = v_load_aligned((const int*)(_mag_a + j + 4));
v_int32x4 v_m3 = v_load_aligned((const int*)(_mag_a + j + 8));
v_int32x4 v_m4 = v_load_aligned((const int*)(_mag_a + j + 12));
v_store_aligned((signed char*)(_pmap + j), v_one);
v_int32x4 v_cmp1 = v_m1 > v_low;
v_int32x4 v_cmp2 = v_m2 > v_low;
v_int32x4 v_cmp3 = v_m3 > v_low;
v_int32x4 v_cmp4 = v_m4 > v_low;
v_int16x8 v_cmp80 = v_pack(v_cmp1, v_cmp2);
v_int16x8 v_cmp81 = v_pack(v_cmp3, v_cmp4);
v_int8x16 v_cmp = v_pack(v_cmp80, v_cmp81);
unsigned int mask = v_signmask(v_cmp);
if (mask)
{
int k = j;
do
{
int l = trailingZeros32(mask);
k += l;
mask >>= l;
int m = _mag_a[k];
short xs = _dx[k];
short ys = _dy[k];
int x = (int)std::abs(xs);
int y = (int)std::abs(ys) << 15;
int tg22x = x * TG22;
if (y < tg22x)
{
if (m > _mag_a[k - 1] && m >= _mag_a[k + 1])
{
CANNY_CHECK_SIMD(m, high, (_pmap+k), stack);
}
}
else
{
int tg67x = tg22x + (x << 16);
if (y > tg67x)
{
if (m > _mag_p[k] && m >= _mag_n[k])
{
CANNY_CHECK_SIMD(m, high, (_pmap+k), stack);
}
}
else
{
int s = (xs ^ ys) < 0 ? -1 : 1;
if(m > _mag_p[k - s] && m > _mag_n[k + s])
{
CANNY_CHECK_SIMD(m, high, (_pmap+k), stack);
}
}
}
++k;
} while((mask >>= 1));
}
j += 16;
}
}
#endif
for (; j < src.cols; j++)
{
int m = _mag_a[j];
if (m > low)
{
short xs = _dx[j];
short ys = _dy[j];
int x = (int)std::abs(xs);
int y = (int)std::abs(ys) << 15;
int tg22x = x * TG22;
if (y < tg22x)
{
if (m > _mag_a[j - 1] && m >= _mag_a[j + 1])
{
CANNY_CHECK(m, high, (_pmap+j), stack);
}
}
else
{
int tg67x = tg22x + (x << 16);
if (y > tg67x)
{
if (m > _mag_p[j] && m >= _mag_n[j])
{
CANNY_CHECK(m, high, (_pmap+j), stack);
}
}
else
{
int s = (xs ^ ys) < 0 ? -1 : 1;
if(m > _mag_p[j - s] && m > _mag_n[j + s])
{
CANNY_CHECK(m, high, (_pmap+j), stack);
}
}
}
}
_pmap[j] = 1;
}
}
void spatialGradient( InputArray _src, OutputArray _dx, OutputArray _dy,
int ksize, int borderType )
{
CV_INSTRUMENT_REGION()
// Prepare InputArray src
Mat src = _src.getMat();
CV_Assert( !src.empty() );
CV_Assert( src.type() == CV_8UC1 );
CV_Assert( borderType == BORDER_DEFAULT || borderType == BORDER_REPLICATE );
// Prepare OutputArrays dx, dy
_dx.create( src.size(), CV_16SC1 );
_dy.create( src.size(), CV_16SC1 );
Mat dx = _dx.getMat(),
dy = _dy.getMat();
// TODO: Allow for other kernel sizes
CV_Assert(ksize == 3);
// Get dimensions
const int H = src.rows,
W = src.cols;
// Row, column indices
int i = 0,
j = 0;
// Handle border types
int i_top = 0, // Case for H == 1 && W == 1 && BORDER_REPLICATE
i_bottom = H - 1,
j_offl = 0, // j offset from 0th pixel to reach -1st pixel
j_offr = 0; // j offset from W-1th pixel to reach Wth pixel
if ( borderType == BORDER_DEFAULT ) // Equiv. to BORDER_REFLECT_101
{
if ( H > 1 )
{
i_top = 1;
i_bottom = H - 2;
}
if ( W > 1 )
{
j_offl = 1;
j_offr = -1;
}
}
// Pointer to row vectors
uchar *p_src, *c_src, *n_src; // previous, current, next row
short *c_dx, *c_dy;
int i_start = 0;
int j_start = 0;
#if CV_SIMD128 && CV_SSE2
if(hasSIMD128())
{
uchar *m_src;
short *n_dx, *n_dy;
// Characters in variable names have the following meanings:
// u: unsigned char
// s: signed int
//
// [row][column]
// m: offset -1
// n: offset 0
// p: offset 1
// Example: umn is offset -1 in row and offset 0 in column
for ( i = 0; i < H - 1; i += 2 )
{
if ( i == 0 ) p_src = src.ptr<uchar>(i_top);
else p_src = src.ptr<uchar>(i-1);
c_src = src.ptr<uchar>(i);
n_src = src.ptr<uchar>(i+1);
if ( i == H - 2 ) m_src = src.ptr<uchar>(i_bottom);
else m_src = src.ptr<uchar>(i+2);
c_dx = dx.ptr<short>(i);
c_dy = dy.ptr<short>(i);
n_dx = dx.ptr<short>(i+1);
n_dy = dy.ptr<short>(i+1);
v_uint8x16 v_select_m = v_uint8x16(0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0xFF);
// Process rest of columns 16-column chunks at a time
for ( j = 1; j < W - 16; j += 16 )
{
// Load top row for 3x3 Sobel filter
v_uint8x16 v_um = v_load(&p_src[j-1]);
v_uint8x16 v_up = v_load(&p_src[j+1]);
// TODO: Replace _mm_slli_si128 with hal method
v_uint8x16 v_un = v_select(v_select_m, v_uint8x16(_mm_slli_si128(v_up.val, 1)),
v_uint8x16(_mm_srli_si128(v_um.val, 1)));
v_uint16x8 v_um1, v_um2, v_un1, v_un2, v_up1, v_up2;
v_expand(v_um, v_um1, v_um2);
v_expand(v_un, v_un1, v_un2);
v_expand(v_up, v_up1, v_up2);
v_int16x8 v_s1m1 = v_reinterpret_as_s16(v_um1);
v_int16x8 v_s1m2 = v_reinterpret_as_s16(v_um2);
v_int16x8 v_s1n1 = v_reinterpret_as_s16(v_un1);
v_int16x8 v_s1n2 = v_reinterpret_as_s16(v_un2);
v_int16x8 v_s1p1 = v_reinterpret_as_s16(v_up1);
v_int16x8 v_s1p2 = v_reinterpret_as_s16(v_up2);
// Load second row for 3x3 Sobel filter
v_um = v_load(&c_src[j-1]);
v_up = v_load(&c_src[j+1]);
// TODO: Replace _mm_slli_si128 with hal method
v_un = v_select(v_select_m, v_uint8x16(_mm_slli_si128(v_up.val, 1)),
v_uint8x16(_mm_srli_si128(v_um.val, 1)));
v_expand(v_um, v_um1, v_um2);
v_expand(v_un, v_un1, v_un2);
v_expand(v_up, v_up1, v_up2);
v_int16x8 v_s2m1 = v_reinterpret_as_s16(v_um1);
v_int16x8 v_s2m2 = v_reinterpret_as_s16(v_um2);
v_int16x8 v_s2n1 = v_reinterpret_as_s16(v_un1);
v_int16x8 v_s2n2 = v_reinterpret_as_s16(v_un2);
v_int16x8 v_s2p1 = v_reinterpret_as_s16(v_up1);
v_int16x8 v_s2p2 = v_reinterpret_as_s16(v_up2);
// Load third row for 3x3 Sobel filter
v_um = v_load(&n_src[j-1]);
v_up = v_load(&n_src[j+1]);
// TODO: Replace _mm_slli_si128 with hal method
v_un = v_select(v_select_m, v_uint8x16(_mm_slli_si128(v_up.val, 1)),
v_uint8x16(_mm_srli_si128(v_um.val, 1)));
v_expand(v_um, v_um1, v_um2);
v_expand(v_un, v_un1, v_un2);
v_expand(v_up, v_up1, v_up2);
v_int16x8 v_s3m1 = v_reinterpret_as_s16(v_um1);
v_int16x8 v_s3m2 = v_reinterpret_as_s16(v_um2);
v_int16x8 v_s3n1 = v_reinterpret_as_s16(v_un1);
v_int16x8 v_s3n2 = v_reinterpret_as_s16(v_un2);
v_int16x8 v_s3p1 = v_reinterpret_as_s16(v_up1);
v_int16x8 v_s3p2 = v_reinterpret_as_s16(v_up2);
// dx & dy for rows 1, 2, 3
v_int16x8 v_sdx1, v_sdy1;
spatialGradientKernel<v_int16x8>( v_sdx1, v_sdy1,
v_s1m1, v_s1n1, v_s1p1,
v_s2m1, v_s2p1,
v_s3m1, v_s3n1, v_s3p1 );
v_int16x8 v_sdx2, v_sdy2;
spatialGradientKernel<v_int16x8>( v_sdx2, v_sdy2,
v_s1m2, v_s1n2, v_s1p2,
v_s2m2, v_s2p2,
v_s3m2, v_s3n2, v_s3p2 );
// Store
v_store(&c_dx[j], v_sdx1);
v_store(&c_dx[j+8], v_sdx2);
v_store(&c_dy[j], v_sdy1);
v_store(&c_dy[j+8], v_sdy2);
// Load fourth row for 3x3 Sobel filter
v_um = v_load(&m_src[j-1]);
v_up = v_load(&m_src[j+1]);
// TODO: Replace _mm_slli_si128 with hal method
v_un = v_select(v_select_m, v_uint8x16(_mm_slli_si128(v_up.val, 1)),
v_uint8x16(_mm_srli_si128(v_um.val, 1)));
v_expand(v_um, v_um1, v_um2);
v_expand(v_un, v_un1, v_un2);
v_expand(v_up, v_up1, v_up2);
v_int16x8 v_s4m1 = v_reinterpret_as_s16(v_um1);
v_int16x8 v_s4m2 = v_reinterpret_as_s16(v_um2);
v_int16x8 v_s4n1 = v_reinterpret_as_s16(v_un1);
v_int16x8 v_s4n2 = v_reinterpret_as_s16(v_un2);
v_int16x8 v_s4p1 = v_reinterpret_as_s16(v_up1);
v_int16x8 v_s4p2 = v_reinterpret_as_s16(v_up2);
// dx & dy for rows 2, 3, 4
spatialGradientKernel<v_int16x8>( v_sdx1, v_sdy1,
v_s2m1, v_s2n1, v_s2p1,
v_s3m1, v_s3p1,
v_s4m1, v_s4n1, v_s4p1 );
spatialGradientKernel<v_int16x8>( v_sdx2, v_sdy2,
v_s2m2, v_s2n2, v_s2p2,
v_s3m2, v_s3p2,
v_s4m2, v_s4n2, v_s4p2 );
// Store
v_store(&n_dx[j], v_sdx1);
v_store(&n_dx[j+8], v_sdx2);
v_store(&n_dy[j], v_sdy1);
v_store(&n_dy[j+8], v_sdy2);
}
}
}
i_start = i;
j_start = j;
#endif
int j_p, j_n;
uchar v00, v01, v02, v10, v11, v12, v20, v21, v22;
for ( i = 0; i < H; i++ )
{
if ( i == 0 ) p_src = src.ptr<uchar>(i_top);
else p_src = src.ptr<uchar>(i-1);
c_src = src.ptr<uchar>(i);
if ( i == H - 1 ) n_src = src.ptr<uchar>(i_bottom);
else n_src = src.ptr<uchar>(i+1);
c_dx = dx.ptr<short>(i);
c_dy = dy.ptr<short>(i);
// Process left-most column
j = 0;
j_p = j + j_offl;
j_n = 1;
if ( j_n >= W ) j_n = j + j_offr;
v00 = p_src[j_p]; v01 = p_src[j]; v02 = p_src[j_n];
v10 = c_src[j_p]; v11 = c_src[j]; v12 = c_src[j_n];
v20 = n_src[j_p]; v21 = n_src[j]; v22 = n_src[j_n];
spatialGradientKernel<short>( c_dx[0], c_dy[0], v00, v01, v02, v10,
v12, v20, v21, v22 );
v00 = v01; v10 = v11; v20 = v21;
v01 = v02; v11 = v12; v21 = v22;
// Process middle columns
j = i >= i_start ? 1 : j_start;
j_p = j - 1;
v00 = p_src[j_p]; v01 = p_src[j];
v10 = c_src[j_p]; v11 = c_src[j];
v20 = n_src[j_p]; v21 = n_src[j];
for ( ; j < W - 1; j++ )
{
// Get values for next column
j_n = j + 1; v02 = p_src[j_n]; v12 = c_src[j_n]; v22 = n_src[j_n];
spatialGradientKernel<short>( c_dx[j], c_dy[j], v00, v01, v02, v10,
v12, v20, v21, v22 );
// Move values back one column for next iteration
v00 = v01; v10 = v11; v20 = v21;
v01 = v02; v11 = v12; v21 = v22;
}
// Process right-most column
if ( j < W )
{
j_n = j + j_offr; v02 = p_src[j_n]; v12 = c_src[j_n]; v22 = n_src[j_n];
spatialGradientKernel<short>( c_dx[j], c_dy[j], v00, v01, v02, v10,
v12, v20, v21, v22 );
}
}
}
