@ -130,140 +130,143 @@ void spatialGradient( InputArray _src, OutputArray _dx, OutputArray _dy,
int i_start = 0 ; int i_start = 0 ;
int j_start = 0 ; int j_start = 0 ;
# if CV_SIMD128 && CV_SSE2 # if CV_SIMD128 && CV_SSE2
uchar * m_src ; if ( hasSIMD128 ( ) )
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 ) ; uchar * m_src ;
else p_src = src . ptr < uchar > ( i - 1 ) ; short * n_dx , * n_dy ;
c_src = src . ptr < uchar > ( i ) ; // Characters in variable names have the following meanings:
n_src = src . ptr < uchar > ( i + 1 ) ; // u: unsigned char
// s: signed int
if ( i = = H - 2 ) m_src = src . ptr < uchar > ( i_bottom ) ; //
else m_src = src . ptr < uchar > ( i + 2 ) ; // [row][column]
// m: offset -1
c_dx = dx . ptr < short > ( i ) ; // n: offset 0
c_dy = dy . ptr < short > ( i ) ; // p: offset 1
n_dx = dx . ptr < short > ( i + 1 ) ; // Example: umn is offset -1 in row and offset 0 in column
n_dy = dy . ptr < short > ( i + 1 ) ; for ( i = 0 ; i < H - 1 ; i + = 2 )
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
if ( i = = 0 ) p_src = src . ptr < uchar > ( i_top ) ;
v_uint8x16 v_um = v_load ( & p_src [ j - 1 ] ) ; else p_src = src . ptr < uchar > ( i - 1 ) ;
v_uint8x16 v_up = v_load ( & p_src [ j + 1 ] ) ;
// TODO: Replace _mm_slli_si128 with hal method
c_src = src . ptr < uchar > ( i ) ;
v_uint8x16 v_un = v_select ( v_select_m , v_uint8x16 ( _mm_slli_si128 ( v_up . val , 1 ) ) , n_src = src . ptr < uchar > ( i + 1 ) ;
v_uint8x16 ( _mm_srli_si128 ( v_um . val , 1 ) ) ) ;
v_uint16x8 v_um1 , v_um2 , v_un1 , v_un2 , v_up1 , v_up2 ; if ( i = = H - 2 ) m_src = src . ptr < uchar > ( i_bottom ) ;
v_expand ( v_um , v_um1 , v_um2 ) ; else m_src = src . ptr < uchar > ( i + 2 ) ;
v_expand ( v_un , v_un1 , v_un2 ) ;
v_expand ( v_up , v_up1 , v_up2 ) ; c_dx = dx . ptr < short > ( i ) ;
v_int16x8 v_s1m1 = v_reinterpret_as_s16 ( v_um1 ) ; c_dy = dy . ptr < short > ( i ) ;
v_int16x8 v_s1m2 = v_reinterpret_as_s16 ( v_um2 ) ; n_dx = dx . ptr < short > ( i + 1 ) ;
v_int16x8 v_s1n1 = v_reinterpret_as_s16 ( v_un1 ) ; n_dy = dy . ptr < short > ( i + 1 ) ;
v_int16x8 v_s1n2 = v_reinterpret_as_s16 ( v_un2 ) ;
v_int16x8 v_s1p1 = v_reinterpret_as_s16 ( v_up1 ) ; v_uint8x16 v_select_m = v_uint8x16 ( 0 , 0 , 0 , 0 , 0 , 0 , 0 , 0 , 0 , 0 , 0 , 0 ,
v_int16x8 v_s1p2 = v_reinterpret_as_s16 ( v_up2 ) ; 0 , 0 , 0 , 0xFF ) ;
// Load second row for 3x3 Sobel filter
// Process rest of columns 16-column chunks at a time
v_um = v_load ( & c_src [ j - 1 ] ) ; for ( j = 1 ; j < W - 16 ; j + = 16 )
v_up = v_load ( & c_src [ j + 1 ] ) ; {
// TODO: Replace _mm_slli_si128 with hal method
// Load top row for 3x3 Sobel filter
v_un = v_select ( v_select_m , v_uint8x16 ( _mm_slli_si128 ( v_up . val , 1 ) ) , v_uint8x16 v_um = v_load ( & p_src [ j - 1 ] ) ;
v_uint8x16 ( _mm_srli_si128 ( v_um . val , 1 ) ) ) ; v_uint8x16 v_up = v_load ( & p_src [ j + 1 ] ) ;
v_expand ( v_um , v_um1 , v_um2 ) ; // TODO: Replace _mm_slli_si128 with hal method
v_expand ( v_un , v_un1 , v_un2 ) ; v_uint8x16 v_un = v_select ( v_select_m , v_uint8x16 ( _mm_slli_si128 ( v_up . val , 1 ) ) ,
v_expand ( v_up , v_up1 , v_up2 ) ; v_uint8x16 ( _mm_srli_si128 ( v_um . val , 1 ) ) ) ;
v_int16x8 v_s2m1 = v_reinterpret_as_s16 ( v_um1 ) ; v_uint16x8 v_um1 , v_um2 , v_un1 , v_un2 , v_up1 , v_up2 ;
v_int16x8 v_s2m2 = v_reinterpret_as_s16 ( v_um2 ) ; v_expand ( v_um , v_um1 , v_um2 ) ;
v_int16x8 v_s2n1 = v_reinterpret_as_s16 ( v_un1 ) ; v_expand ( v_un , v_un1 , v_un2 ) ;
v_int16x8 v_s2n2 = v_reinterpret_as_s16 ( v_un2 ) ; v_expand ( v_up , v_up1 , v_up2 ) ;
v_int16x8 v_s2p1 = v_reinterpret_as_s16 ( v_up1 ) ; v_int16x8 v_s1m1 = v_reinterpret_as_s16 ( v_um1 ) ;
v_int16x8 v_s2p2 = v_reinterpret_as_s16 ( v_up2 ) ; v_int16x8 v_s1m2 = v_reinterpret_as_s16 ( v_um2 ) ;
v_int16x8 v_s1n1 = v_reinterpret_as_s16 ( v_un1 ) ;
// Load third row for 3x3 Sobel filter
v_int16x8 v_s1n2 = v_reinterpret_as_s16 ( v_un2 ) ;
v_um = v_load ( & n_src [ j - 1 ] ) ; v_int16x8 v_s1p1 = v_reinterpret_as_s16 ( v_up1 ) ;
v_up = v_load ( & n_src [ j + 1 ] ) ; v_int16x8 v_s1p2 = v_reinterpret_as_s16 ( v_up2 ) ;
// TODO: Replace _mm_slli_si128 with hal method
v_un = v_select ( v_select_m , v_uint8x16 ( _mm_slli_si128 ( v_up . val , 1 ) ) , // Load second row for 3x3 Sobel filter
v_uint8x16 ( _mm_srli_si128 ( v_um . val , 1 ) ) ) ; v_um = v_load ( & c_src [ j - 1 ] ) ;
v_expand ( v_um , v_um1 , v_um2 ) ; v_up = v_load ( & c_src [ j + 1 ] ) ;
v_expand ( v_un , v_un1 , v_un2 ) ; // TODO: Replace _mm_slli_si128 with hal method
v_expand ( v_up , v_up1 , v_up2 ) ; v_un = v_select ( v_select_m , v_uint8x16 ( _mm_slli_si128 ( v_up . val , 1 ) ) ,
v_int16x8 v_s3m1 = v_reinterpret_as_s16 ( v_um1 ) ; v_uint8x16 ( _mm_srli_si128 ( v_um . val , 1 ) ) ) ;
v_int16x8 v_s3m2 = v_reinterpret_as_s16 ( v_um2 ) ; v_expand ( v_um , v_um1 , v_um2 ) ;
v_int16x8 v_s3n1 = v_reinterpret_as_s16 ( v_un1 ) ; v_expand ( v_un , v_un1 , v_un2 ) ;
v_int16x8 v_s3n2 = v_reinterpret_as_s16 ( v_un2 ) ; v_expand ( v_up , v_up1 , v_up2 ) ;
v_int16x8 v_s3p1 = v_reinterpret_as_s16 ( v_up1 ) ; v_int16x8 v_s2m1 = v_reinterpret_as_s16 ( v_um1 ) ;
v_int16x8 v_s3p2 = v_reinterpret_as_s16 ( v_up2 ) ; v_int16x8 v_s2m2 = v_reinterpret_as_s16 ( v_um2 ) ;
v_int16x8 v_s2n1 = v_reinterpret_as_s16 ( v_un1 ) ;
// dx & dy for rows 1, 2, 3
v_int16x8 v_s2n2 = v_reinterpret_as_s16 ( v_un2 ) ;
v_int16x8 v_sdx1 , v_sdy1 ; v_int16x8 v_s2p1 = v_reinterpret_as_s16 ( v_up1 ) ;
spatialGradientKernel < v_int16x8 > ( v_sdx1 , v_sdy1 , v_int16x8 v_s2p2 = v_reinterpret_as_s16 ( v_up2 ) ;
v_s1m1 , v_s1n1 , v_s1p1 ,
v_s2m1 , v_s2p1 , // Load third row for 3x3 Sobel filter
v_s3m1 , v_s3n1 , v_s3p1 ) ; v_um = v_load ( & n_src [ j - 1 ] ) ;
v_up = v_load ( & n_src [ j + 1 ] ) ;
v_int16x8 v_sdx2 , v_sdy2 ; // TODO: Replace _mm_slli_si128 with hal method
spatialGradientKernel < v_int16x8 > ( v_sdx2 , v_sdy2 , v_un = v_select ( v_select_m , v_uint8x16 ( _mm_slli_si128 ( v_up . val , 1 ) ) ,
v_s1m2 , v_s1n2 , v_s1p2 , v_uint8x16 ( _mm_srli_si128 ( v_um . val , 1 ) ) ) ;
v_s2m2 , v_s2p2 , v_expand ( v_um , v_um1 , v_um2 ) ;
v_s3m2 , v_s3n2 , v_s3p2 ) ; v_expand ( v_un , v_un1 , v_un2 ) ;
v_expand ( v_up , v_up1 , v_up2 ) ;
// Store
v_int16x8 v_s3m1 = v_reinterpret_as_s16 ( v_um1 ) ;
v_store ( & c_dx [ j ] , v_sdx1 ) ; v_int16x8 v_s3m2 = v_reinterpret_as_s16 ( v_um2 ) ;
v_store ( & c_dx [ j + 8 ] , v_sdx2 ) ; v_int16x8 v_s3n1 = v_reinterpret_as_s16 ( v_un1 ) ;
v_store ( & c_dy [ j ] , v_sdy1 ) ; v_int16x8 v_s3n2 = v_reinterpret_as_s16 ( v_un2 ) ;
v_store ( & c_dy [ j + 8 ] , v_sdy2 ) ; v_int16x8 v_s3p1 = v_reinterpret_as_s16 ( v_up1 ) ;
v_int16x8 v_s3p2 = v_reinterpret_as_s16 ( v_up2 ) ;
// Load fourth row for 3x3 Sobel filter
v_um = v_load ( & m_src [ j - 1 ] ) ; // dx & dy for rows 1, 2, 3
v_up = v_load ( & m_src [ j + 1 ] ) ; v_int16x8 v_sdx1 , v_sdy1 ;
// TODO: Replace _mm_slli_si128 with hal method
spatialGradientKernel < v_int16x8 > ( v_sdx1 , v_sdy1 ,
v_un = v_select ( v_select_m , v_uint8x16 ( _mm_slli_si128 ( v_up . val , 1 ) ) , v_s1m1 , v_s1n1 , v_s1p1 ,
v_uint8x16 ( _mm_srli_si128 ( v_um . val , 1 ) ) ) ; v_s2m1 , v_s2p1 ,
v_expand ( v_um , v_um1 , v_um2 ) ; v_s3m1 , v_s3n1 , v_s3p1 ) ;
v_expand ( v_un , v_un1 , v_un2 ) ;
v_expand ( v_up , v_up1 , v_up2 ) ; v_int16x8 v_sdx2 , v_sdy2 ;
v_int16x8 v_s4m1 = v_reinterpret_as_s16 ( v_um1 ) ; spatialGradientKernel < v_int16x8 > ( v_sdx2 , v_sdy2 ,
v_int16x8 v_s4m2 = v_reinterpret_as_s16 ( v_um2 ) ; v_s1m2 , v_s1n2 , v_s1p2 ,
v_int16x8 v_s4n1 = v_reinterpret_as_s16 ( v_un1 ) ; v_s2m2 , v_s2p2 ,
v_int16x8 v_s4n2 = v_reinterpret_as_s16 ( v_un2 ) ; v_s3m2 , v_s3n2 , v_s3p2 ) ;
v_int16x8 v_s4p1 = v_reinterpret_as_s16 ( v_up1 ) ;
v_int16x8 v_s4p2 = v_reinterpret_as_s16 ( v_up2 ) ; // Store
v_store ( & c_dx [ j ] , v_sdx1 ) ;
// dx & dy for rows 2, 3, 4
v_store ( & c_dx [ j + 8 ] , v_sdx2 ) ;
spatialGradientKernel < v_int16x8 > ( v_sdx1 , v_sdy1 , v_store ( & c_dy [ j ] , v_sdy1 ) ;
v_s2m1 , v_s2n1 , v_s2p1 , v_store ( & c_dy [ j + 8 ] , v_sdy2 ) ;
v_s3m1 , v_s3p1 ,
v_s4m1 , v_s4n1 , v_s4p1 ) ; // Load fourth row for 3x3 Sobel filter
v_um = v_load ( & m_src [ j - 1 ] ) ;
spatialGradientKernel < v_int16x8 > ( v_sdx2 , v_sdy2 , v_up = v_load ( & m_src [ j + 1 ] ) ;
v_s2m2 , v_s2n2 , v_s2p2 , // TODO: Replace _mm_slli_si128 with hal method
v_s3m2 , v_s3p2 , v_un = v_select ( v_select_m , v_uint8x16 ( _mm_slli_si128 ( v_up . val , 1 ) ) ,
v_s4m2 , v_s4n2 , v_s4p2 ) ; v_uint8x16 ( _mm_srli_si128 ( v_um . val , 1 ) ) ) ;
v_expand ( v_um , v_um1 , v_um2 ) ;
// Store
v_expand ( v_un , v_un1 , v_un2 ) ;
v_store ( & n_dx [ j ] , v_sdx1 ) ; v_expand ( v_up , v_up1 , v_up2 ) ;
v_store ( & n_dx [ j + 8 ] , v_sdx2 ) ; v_int16x8 v_s4m1 = v_reinterpret_as_s16 ( v_um1 ) ;
v_store ( & n_dy [ j ] , v_sdy1 ) ; v_int16x8 v_s4m2 = v_reinterpret_as_s16 ( v_um2 ) ;
v_store ( & n_dy [ j + 8 ] , v_sdy2 ) ; 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 ; i_start = i ;
Tomoaki Teshima
8 years ago