parallelCanny(const Mat &_dx, const Mat &_dy, Mat &_map, std::deque<uchar*> &borderPeaksParallel,
int _low, int _high, bool _L2gradient) :
src(_dx), src2(_dy), map(_map), _borderPeaksParallel(borderPeaksParallel),
low(_low), high(_high), aperture_size(0), L2gradient(_L2gradient)
{
#if CV_SIMD128
haveSIMD = hasSIMD128();
if(haveSIMD)
_map.create(src.rows + 2, (int)alignSize((size_t)(src.cols + CV_MALLOC_SIMD128 + 1), CV_MALLOC_SIMD128), CV_8UC1);
else
#endif
_map.create(src.rows + 2, src.cols + 2, CV_8UC1);
map = _map;
map.row(0).setTo(1);
map.row(src.rows + 1).setTo(1);
mapstep = map.cols;
needGradient = false;
cn = src.channels();
}
void FAST_t(InputArray _img, std::vector<KeyPoint>& keypoints, int threshold, bool nonmax_suppression)
{
Mat img = _img.getMat();
const int K = patternSize/2, N = patternSize + K + 1;
int i, j, k, pixel[25];
makeOffsets(pixel, (int)img.step, patternSize);
#if CV_SIMD128
const int quarterPatternSize = patternSize/4;
v_uint8x16 delta = v_setall_u8(0x80), t = v_setall_u8((char)threshold), K16 = v_setall_u8((char)K);
bool hasSimd = hasSIMD128();
#if CV_TRY_AVX2
Ptr<opt_AVX2::FAST_t_patternSize16_AVX2> fast_t_impl_avx2;
if(CV_CPU_HAS_SUPPORT_AVX2)
fast_t_impl_avx2 = opt_AVX2::FAST_t_patternSize16_AVX2::getImpl(img.cols, threshold, nonmax_suppression, pixel);
#endif
#endif
keypoints.clear();
threshold = std::min(std::max(threshold, 0), 255);
uchar threshold_tab[512];
for( i = -255; i <= 255; i++ )
threshold_tab[i+255] = (uchar)(i < -threshold ? 1 : i > threshold ? 2 : 0);
AutoBuffer<uchar> _buf((img.cols+16)*3*(sizeof(int) + sizeof(uchar)) + 128);
uchar* buf[3];
buf[0] = _buf.data(); buf[1] = buf[0] + img.cols; buf[2] = buf[1] + img.cols;
int* cpbuf[3];
cpbuf[0] = (int*)alignPtr(buf[2] + img.cols, sizeof(int)) + 1;
cpbuf[1] = cpbuf[0] + img.cols + 1;
cpbuf[2] = cpbuf[1] + img.cols + 1;
memset(buf[0], 0, img.cols*3);
for(i = 3; i < img.rows-2; i++)
{
const uchar* ptr = img.ptr<uchar>(i) + 3;
uchar* curr = buf[(i - 3)%3];
int* cornerpos = cpbuf[(i - 3)%3];
memset(curr, 0, img.cols);
int ncorners = 0;
if( i < img.rows - 3 )
{
j = 3;
#if CV_SIMD128
if( hasSimd )
{
if( patternSize == 16 )
{
#if CV_TRY_AVX2
if (fast_t_impl_avx2)
fast_t_impl_avx2->process(j, ptr, curr, cornerpos, ncorners);
#endif
//vz if (j <= (img.cols - 27)) //it doesn't make sense using vectors for less than 8 elements
{
for (; j < img.cols - 16 - 3; j += 16, ptr += 16)
{
v_uint8x16 v = v_load(ptr);
v_int8x16 v0 = v_reinterpret_as_s8((v + t) ^ delta);
v_int8x16 v1 = v_reinterpret_as_s8((v - t) ^ delta);
v_int8x16 x0 = v_reinterpret_as_s8(v_sub_wrap(v_load(ptr + pixel[0]), delta));
v_int8x16 x1 = v_reinterpret_as_s8(v_sub_wrap(v_load(ptr + pixel[quarterPatternSize]), delta));
v_int8x16 x2 = v_reinterpret_as_s8(v_sub_wrap(v_load(ptr + pixel[2*quarterPatternSize]), delta));
v_int8x16 x3 = v_reinterpret_as_s8(v_sub_wrap(v_load(ptr + pixel[3*quarterPatternSize]), delta));
v_int8x16 m0, m1;
m0 = (v0 < x0) & (v0 < x1);
m1 = (x0 < v1) & (x1 < v1);
m0 = m0 | ((v0 < x1) & (v0 < x2));
m1 = m1 | ((x1 < v1) & (x2 < v1));
m0 = m0 | ((v0 < x2) & (v0 < x3));
m1 = m1 | ((x2 < v1) & (x3 < v1));
m0 = m0 | ((v0 < x3) & (v0 < x0));
m1 = m1 | ((x3 < v1) & (x0 < v1));
m0 = m0 | m1;
int mask = v_signmask(m0);
if( mask == 0 )
continue;
if( (mask & 255) == 0 )
{
j -= 8;
ptr -= 8;
continue;
}
v_int8x16 c0 = v_setzero_s8();
v_int8x16 c1 = v_setzero_s8();
v_uint8x16 max0 = v_setzero_u8();
v_uint8x16 max1 = v_setzero_u8();
for( k = 0; k < N; k++ )
{
v_int8x16 x = v_reinterpret_as_s8(v_load((ptr + pixel[k])) ^ delta);
m0 = v0 < x;
m1 = x < v1;
c0 = v_sub_wrap(c0, m0) & m0;
c1 = v_sub_wrap(c1, m1) & m1;
max0 = v_max(max0, v_reinterpret_as_u8(c0));
max1 = v_max(max1, v_reinterpret_as_u8(c1));
}
max0 = v_max(max0, max1);
int m = v_signmask(K16 < max0);
for( k = 0; m > 0 && k < 16; k++, m >>= 1 )
{
if(m & 1)
{
cornerpos[ncorners++] = j+k;
if(nonmax_suppression)
curr[j+k] = (uchar)cornerScore<patternSize>(ptr+k, pixel, threshold);
}
}
}
}
}
}
#endif
for( ; j < img.cols - 3; j++, ptr++ )
{
int v = ptr[0];
const uchar* tab = &threshold_tab[0] - v + 255;
int d = tab[ptr[pixel[0]]] | tab[ptr[pixel[8]]];
if( d == 0 )
continue;
d &= tab[ptr[pixel[2]]] | tab[ptr[pixel[10]]];
d &= tab[ptr[pixel[4]]] | tab[ptr[pixel[12]]];
d &= tab[ptr[pixel[6]]] | tab[ptr[pixel[14]]];
if( d == 0 )
continue;
d &= tab[ptr[pixel[1]]] | tab[ptr[pixel[9]]];
d &= tab[ptr[pixel[3]]] | tab[ptr[pixel[11]]];
d &= tab[ptr[pixel[5]]] | tab[ptr[pixel[13]]];
d &= tab[ptr[pixel[7]]] | tab[ptr[pixel[15]]];
if( d & 1 )
{
int vt = v - threshold, count = 0;
for( k = 0; k < N; k++ )
{
int x = ptr[pixel[k]];
if(x < vt)
{
if( ++count > K )
{
cornerpos[ncorners++] = j;
if(nonmax_suppression)
curr[j] = (uchar)cornerScore<patternSize>(ptr, pixel, threshold);
break;
}
}
else
count = 0;
}
}
if( d & 2 )
{
int vt = v + threshold, count = 0;
for( k = 0; k < N; k++ )
{
int x = ptr[pixel[k]];
if(x > vt)
{
if( ++count > K )
{
cornerpos[ncorners++] = j;
if(nonmax_suppression)
curr[j] = (uchar)cornerScore<patternSize>(ptr, pixel, threshold);
break;
}
}
else
count = 0;
}
}
}
}
cornerpos[-1] = ncorners;
if( i == 3 )
continue;
const uchar* prev = buf[(i - 4 + 3)%3];
const uchar* pprev = buf[(i - 5 + 3)%3];
cornerpos = cpbuf[(i - 4 + 3)%3];
ncorners = cornerpos[-1];
for( k = 0; k < ncorners; k++ )
{
j = cornerpos[k];
int score = prev[j];
if( !nonmax_suppression ||
(score > prev[j+1] && score > prev[j-1] &&
score > pprev[j-1] && score > pprev[j] && score > pprev[j+1] &&
score > curr[j-1] && score > curr[j] && score > curr[j+1]) )
{
keypoints.push_back(KeyPoint((float)j, (float)(i-1), 7.f, -1, (float)score));
}
}
}
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 );
}
}
}
